1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
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.
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
40 * Copyright (c) 2018-2019 Christoph Hellwig
41 */
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
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>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
63 #include <net/sock.h>
64 #include <net/af_unix.h>
65 #include <net/scm.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
84
85 #include <uapi/linux/io_uring.h>
86
87 #include "internal.h"
88 #include "io-wq.h"
89
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
92
93 /*
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
95 */
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
102
103 #define IORING_MAX_REG_BUFFERS (1U << 14)
104
105 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
107 IOSQE_BUFFER_SELECT)
108
109 struct io_uring {
110 u32 head ____cacheline_aligned_in_smp;
111 u32 tail ____cacheline_aligned_in_smp;
112 };
113
114 /*
115 * This data is shared with the application through the mmap at offsets
116 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 *
118 * The offsets to the member fields are published through struct
119 * io_sqring_offsets when calling io_uring_setup.
120 */
121 struct io_rings {
122 /*
123 * Head and tail offsets into the ring; the offsets need to be
124 * masked to get valid indices.
125 *
126 * The kernel controls head of the sq ring and the tail of the cq ring,
127 * and the application controls tail of the sq ring and the head of the
128 * cq ring.
129 */
130 struct io_uring sq, cq;
131 /*
132 * Bitmasks to apply to head and tail offsets (constant, equals
133 * ring_entries - 1)
134 */
135 u32 sq_ring_mask, cq_ring_mask;
136 /* Ring sizes (constant, power of 2) */
137 u32 sq_ring_entries, cq_ring_entries;
138 /*
139 * Number of invalid entries dropped by the kernel due to
140 * invalid index stored in array
141 *
142 * Written by the kernel, shouldn't be modified by the
143 * application (i.e. get number of "new events" by comparing to
144 * cached value).
145 *
146 * After a new SQ head value was read by the application this
147 * counter includes all submissions that were dropped reaching
148 * the new SQ head (and possibly more).
149 */
150 u32 sq_dropped;
151 /*
152 * Runtime SQ flags
153 *
154 * Written by the kernel, shouldn't be modified by the
155 * application.
156 *
157 * The application needs a full memory barrier before checking
158 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
159 */
160 u32 sq_flags;
161 /*
162 * Runtime CQ flags
163 *
164 * Written by the application, shouldn't be modified by the
165 * kernel.
166 */
167 u32 cq_flags;
168 /*
169 * Number of completion events lost because the queue was full;
170 * this should be avoided by the application by making sure
171 * there are not more requests pending than there is space in
172 * the completion queue.
173 *
174 * Written by the kernel, shouldn't be modified by the
175 * application (i.e. get number of "new events" by comparing to
176 * cached value).
177 *
178 * As completion events come in out of order this counter is not
179 * ordered with any other data.
180 */
181 u32 cq_overflow;
182 /*
183 * Ring buffer of completion events.
184 *
185 * The kernel writes completion events fresh every time they are
186 * produced, so the application is allowed to modify pending
187 * entries.
188 */
189 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 };
191
192 enum io_uring_cmd_flags {
193 IO_URING_F_NONBLOCK = 1,
194 IO_URING_F_COMPLETE_DEFER = 2,
195 };
196
197 struct io_mapped_ubuf {
198 u64 ubuf;
199 u64 ubuf_end;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
202 struct bio_vec bvec[];
203 };
204
205 struct io_ring_ctx;
206
207 struct io_overflow_cqe {
208 struct io_uring_cqe cqe;
209 struct list_head list;
210 };
211
212 struct io_fixed_file {
213 /* file * with additional FFS_* flags */
214 unsigned long file_ptr;
215 };
216
217 struct io_rsrc_put {
218 struct list_head list;
219 u64 tag;
220 union {
221 void *rsrc;
222 struct file *file;
223 struct io_mapped_ubuf *buf;
224 };
225 };
226
227 struct io_file_table {
228 /* two level table */
229 struct io_fixed_file **files;
230 };
231
232 struct io_rsrc_node {
233 struct percpu_ref refs;
234 struct list_head node;
235 struct list_head rsrc_list;
236 struct io_rsrc_data *rsrc_data;
237 struct llist_node llist;
238 bool done;
239 };
240
241 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
242
243 struct io_rsrc_data {
244 struct io_ring_ctx *ctx;
245
246 u64 *tags;
247 rsrc_put_fn *do_put;
248 atomic_t refs;
249 struct completion done;
250 bool quiesce;
251 };
252
253 struct io_buffer {
254 struct list_head list;
255 __u64 addr;
256 __u32 len;
257 __u16 bid;
258 };
259
260 struct io_restriction {
261 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
262 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
263 u8 sqe_flags_allowed;
264 u8 sqe_flags_required;
265 bool registered;
266 };
267
268 enum {
269 IO_SQ_THREAD_SHOULD_STOP = 0,
270 IO_SQ_THREAD_SHOULD_PARK,
271 };
272
273 struct io_sq_data {
274 refcount_t refs;
275 atomic_t park_pending;
276 struct mutex lock;
277
278 /* ctx's that are using this sqd */
279 struct list_head ctx_list;
280
281 struct task_struct *thread;
282 struct wait_queue_head wait;
283
284 unsigned sq_thread_idle;
285 int sq_cpu;
286 pid_t task_pid;
287 pid_t task_tgid;
288
289 unsigned long state;
290 struct completion exited;
291 struct callback_head *park_task_work;
292 };
293
294 #define IO_IOPOLL_BATCH 8
295 #define IO_COMPL_BATCH 32
296 #define IO_REQ_CACHE_SIZE 32
297 #define IO_REQ_ALLOC_BATCH 8
298
299 struct io_comp_state {
300 struct io_kiocb *reqs[IO_COMPL_BATCH];
301 unsigned int nr;
302 unsigned int locked_free_nr;
303 /* inline/task_work completion list, under ->uring_lock */
304 struct list_head free_list;
305 /* IRQ completion list, under ->completion_lock */
306 struct list_head locked_free_list;
307 };
308
309 struct io_submit_link {
310 struct io_kiocb *head;
311 struct io_kiocb *last;
312 };
313
314 struct io_submit_state {
315 struct blk_plug plug;
316 struct io_submit_link link;
317
318 /*
319 * io_kiocb alloc cache
320 */
321 void *reqs[IO_REQ_CACHE_SIZE];
322 unsigned int free_reqs;
323
324 bool plug_started;
325
326 /*
327 * Batch completion logic
328 */
329 struct io_comp_state comp;
330
331 /*
332 * File reference cache
333 */
334 struct file *file;
335 unsigned int fd;
336 unsigned int file_refs;
337 unsigned int ios_left;
338 };
339
340 struct io_ring_ctx {
341 struct {
342 struct percpu_ref refs;
343 } ____cacheline_aligned_in_smp;
344
345 struct {
346 unsigned int flags;
347 unsigned int compat: 1;
348 unsigned int drain_next: 1;
349 unsigned int eventfd_async: 1;
350 unsigned int restricted: 1;
351
352 /*
353 * Ring buffer of indices into array of io_uring_sqe, which is
354 * mmapped by the application using the IORING_OFF_SQES offset.
355 *
356 * This indirection could e.g. be used to assign fixed
357 * io_uring_sqe entries to operations and only submit them to
358 * the queue when needed.
359 *
360 * The kernel modifies neither the indices array nor the entries
361 * array.
362 */
363 u32 *sq_array;
364 unsigned cached_sq_head;
365 unsigned sq_entries;
366 unsigned sq_mask;
367 unsigned sq_thread_idle;
368 unsigned cached_sq_dropped;
369 unsigned cached_cq_overflow;
370 unsigned long sq_check_overflow;
371
372 /* hashed buffered write serialization */
373 struct io_wq_hash *hash_map;
374
375 struct list_head defer_list;
376 struct list_head timeout_list;
377 struct list_head cq_overflow_list;
378
379 struct io_uring_sqe *sq_sqes;
380 } ____cacheline_aligned_in_smp;
381
382 struct {
383 struct mutex uring_lock;
384 wait_queue_head_t wait;
385 } ____cacheline_aligned_in_smp;
386
387 struct io_submit_state submit_state;
388
389 struct io_rings *rings;
390
391 /* Only used for accounting purposes */
392 struct mm_struct *mm_account;
393
394 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
395 struct io_sq_data *sq_data; /* if using sq thread polling */
396
397 struct wait_queue_head sqo_sq_wait;
398 struct list_head sqd_list;
399
400 /*
401 * If used, fixed file set. Writers must ensure that ->refs is dead,
402 * readers must ensure that ->refs is alive as long as the file* is
403 * used. Only updated through io_uring_register(2).
404 */
405 struct io_rsrc_data *file_data;
406 struct io_file_table file_table;
407 unsigned nr_user_files;
408
409 /* if used, fixed mapped user buffers */
410 struct io_rsrc_data *buf_data;
411 unsigned nr_user_bufs;
412 struct io_mapped_ubuf **user_bufs;
413
414 struct user_struct *user;
415
416 struct completion ref_comp;
417
418 #if defined(CONFIG_UNIX)
419 struct socket *ring_sock;
420 #endif
421
422 struct xarray io_buffers;
423
424 struct xarray personalities;
425 u32 pers_next;
426
427 struct {
428 unsigned cached_cq_tail;
429 unsigned cq_entries;
430 unsigned cq_mask;
431 atomic_t cq_timeouts;
432 unsigned cq_last_tm_flush;
433 unsigned cq_extra;
434 unsigned long cq_check_overflow;
435 struct wait_queue_head cq_wait;
436 struct fasync_struct *cq_fasync;
437 struct eventfd_ctx *cq_ev_fd;
438 } ____cacheline_aligned_in_smp;
439
440 struct {
441 spinlock_t completion_lock;
442
443 /*
444 * ->iopoll_list is protected by the ctx->uring_lock for
445 * io_uring instances that don't use IORING_SETUP_SQPOLL.
446 * For SQPOLL, only the single threaded io_sq_thread() will
447 * manipulate the list, hence no extra locking is needed there.
448 */
449 struct list_head iopoll_list;
450 struct hlist_head *cancel_hash;
451 unsigned cancel_hash_bits;
452 bool poll_multi_file;
453 } ____cacheline_aligned_in_smp;
454
455 struct delayed_work rsrc_put_work;
456 struct llist_head rsrc_put_llist;
457 struct list_head rsrc_ref_list;
458 spinlock_t rsrc_ref_lock;
459 struct io_rsrc_node *rsrc_node;
460 struct io_rsrc_node *rsrc_backup_node;
461 struct io_mapped_ubuf *dummy_ubuf;
462
463 struct io_restriction restrictions;
464
465 /* exit task_work */
466 struct callback_head *exit_task_work;
467
468 /* Keep this last, we don't need it for the fast path */
469 struct work_struct exit_work;
470 struct list_head tctx_list;
471 };
472
473 struct io_uring_task {
474 /* submission side */
475 struct xarray xa;
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
478 struct io_wq *io_wq;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
481 atomic_t in_idle;
482
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
487 };
488
489 /*
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
492 */
493 struct io_poll_iocb {
494 struct file *file;
495 struct wait_queue_head *head;
496 __poll_t events;
497 bool done;
498 bool canceled;
499 struct wait_queue_entry wait;
500 };
501
502 struct io_poll_update {
503 struct file *file;
504 u64 old_user_data;
505 u64 new_user_data;
506 __poll_t events;
507 bool update_events;
508 bool update_user_data;
509 };
510
511 struct io_close {
512 struct file *file;
513 int fd;
514 };
515
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
521 };
522
523 struct io_accept {
524 struct file *file;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
527 int flags;
528 unsigned long nofile;
529 };
530
531 struct io_sync {
532 struct file *file;
533 loff_t len;
534 loff_t off;
535 int flags;
536 int mode;
537 };
538
539 struct io_cancel {
540 struct file *file;
541 u64 addr;
542 };
543
544 struct io_timeout {
545 struct file *file;
546 u32 off;
547 u32 target_seq;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
551 };
552
553 struct io_timeout_rem {
554 struct file *file;
555 u64 addr;
556
557 /* timeout update */
558 struct timespec64 ts;
559 u32 flags;
560 };
561
562 struct io_rw {
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
564 struct kiocb kiocb;
565 u64 addr;
566 u64 len;
567 };
568
569 struct io_connect {
570 struct file *file;
571 struct sockaddr __user *addr;
572 int addr_len;
573 };
574
575 struct io_sr_msg {
576 struct file *file;
577 union {
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
580 void __user *buf;
581 };
582 int msg_flags;
583 int bgid;
584 size_t len;
585 struct io_buffer *kbuf;
586 };
587
588 struct io_open {
589 struct file *file;
590 int dfd;
591 struct filename *filename;
592 struct open_how how;
593 unsigned long nofile;
594 };
595
596 struct io_rsrc_update {
597 struct file *file;
598 u64 arg;
599 u32 nr_args;
600 u32 offset;
601 };
602
603 struct io_fadvise {
604 struct file *file;
605 u64 offset;
606 u32 len;
607 u32 advice;
608 };
609
610 struct io_madvise {
611 struct file *file;
612 u64 addr;
613 u32 len;
614 u32 advice;
615 };
616
617 struct io_epoll {
618 struct file *file;
619 int epfd;
620 int op;
621 int fd;
622 struct epoll_event event;
623 };
624
625 struct io_splice {
626 struct file *file_out;
627 struct file *file_in;
628 loff_t off_out;
629 loff_t off_in;
630 u64 len;
631 unsigned int flags;
632 };
633
634 struct io_provide_buf {
635 struct file *file;
636 __u64 addr;
637 __u32 len;
638 __u32 bgid;
639 __u16 nbufs;
640 __u16 bid;
641 };
642
643 struct io_statx {
644 struct file *file;
645 int dfd;
646 unsigned int mask;
647 unsigned int flags;
648 const char __user *filename;
649 struct statx __user *buffer;
650 };
651
652 struct io_shutdown {
653 struct file *file;
654 int how;
655 };
656
657 struct io_rename {
658 struct file *file;
659 int old_dfd;
660 int new_dfd;
661 struct filename *oldpath;
662 struct filename *newpath;
663 int flags;
664 };
665
666 struct io_unlink {
667 struct file *file;
668 int dfd;
669 int flags;
670 struct filename *filename;
671 };
672
673 struct io_completion {
674 struct file *file;
675 struct list_head list;
676 u32 cflags;
677 };
678
679 struct io_async_connect {
680 struct sockaddr_storage address;
681 };
682
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
688 struct msghdr msg;
689 struct sockaddr_storage addr;
690 };
691
692 struct io_async_rw {
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
696 size_t bytes_done;
697 struct wait_page_queue wpq;
698 };
699
700 enum {
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
707
708 /* first byte is taken by user flags, shift it to not overlap */
709 REQ_F_FAIL_LINK_BIT = 8,
710 REQ_F_INFLIGHT_BIT,
711 REQ_F_CUR_POS_BIT,
712 REQ_F_NOWAIT_BIT,
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
715 REQ_F_POLLED_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
719 REQ_F_REISSUE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_ASYNC_READ_BIT,
723 REQ_F_ASYNC_WRITE_BIT,
724 REQ_F_ISREG_BIT,
725
726 /* not a real bit, just to check we're not overflowing the space */
727 __REQ_F_LAST_BIT,
728 };
729
730 enum {
731 /* ctx owns file */
732 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
733 /* drain existing IO first */
734 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
735 /* linked sqes */
736 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
737 /* doesn't sever on completion < 0 */
738 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
739 /* IOSQE_ASYNC */
740 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
741 /* IOSQE_BUFFER_SELECT */
742 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
743
744 /* fail rest of links */
745 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
746 /* on inflight list, should be cancelled and waited on exit reliably */
747 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
748 /* read/write uses file position */
749 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
750 /* must not punt to workers */
751 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
752 /* has or had linked timeout */
753 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
754 /* needs cleanup */
755 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
756 /* already went through poll handler */
757 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
758 /* buffer already selected */
759 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
760 /* linked timeout is active, i.e. prepared by link's head */
761 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762 /* completion is deferred through io_comp_state */
763 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
764 /* caller should reissue async */
765 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
766 /* don't attempt request reissue, see io_rw_reissue() */
767 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
768 /* supports async reads */
769 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
770 /* supports async writes */
771 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
772 /* regular file */
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
774 };
775
776 struct async_poll {
777 struct io_poll_iocb poll;
778 struct io_poll_iocb *double_poll;
779 };
780
781 struct io_task_work {
782 struct io_wq_work_node node;
783 task_work_func_t func;
784 };
785
786 /*
787 * NOTE! Each of the iocb union members has the file pointer
788 * as the first entry in their struct definition. So you can
789 * access the file pointer through any of the sub-structs,
790 * or directly as just 'ki_filp' in this struct.
791 */
792 struct io_kiocb {
793 union {
794 struct file *file;
795 struct io_rw rw;
796 struct io_poll_iocb poll;
797 struct io_poll_update poll_update;
798 struct io_accept accept;
799 struct io_sync sync;
800 struct io_cancel cancel;
801 struct io_timeout timeout;
802 struct io_timeout_rem timeout_rem;
803 struct io_connect connect;
804 struct io_sr_msg sr_msg;
805 struct io_open open;
806 struct io_close close;
807 struct io_rsrc_update rsrc_update;
808 struct io_fadvise fadvise;
809 struct io_madvise madvise;
810 struct io_epoll epoll;
811 struct io_splice splice;
812 struct io_provide_buf pbuf;
813 struct io_statx statx;
814 struct io_shutdown shutdown;
815 struct io_rename rename;
816 struct io_unlink unlink;
817 /* use only after cleaning per-op data, see io_clean_op() */
818 struct io_completion compl;
819 };
820
821 /* opcode allocated if it needs to store data for async defer */
822 void *async_data;
823 u8 opcode;
824 /* polled IO has completed */
825 u8 iopoll_completed;
826
827 u16 buf_index;
828 u32 result;
829
830 struct io_ring_ctx *ctx;
831 unsigned int flags;
832 atomic_t refs;
833 struct task_struct *task;
834 u64 user_data;
835
836 struct io_kiocb *link;
837 struct percpu_ref *fixed_rsrc_refs;
838
839 /* used with ctx->iopoll_list with reads/writes */
840 struct list_head inflight_entry;
841 union {
842 struct io_task_work io_task_work;
843 struct callback_head task_work;
844 };
845 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
846 struct hlist_node hash_node;
847 struct async_poll *apoll;
848 struct io_wq_work work;
849 /* store used ubuf, so we can prevent reloading */
850 struct io_mapped_ubuf *imu;
851 };
852
853 struct io_tctx_node {
854 struct list_head ctx_node;
855 struct task_struct *task;
856 struct io_ring_ctx *ctx;
857 };
858
859 struct io_defer_entry {
860 struct list_head list;
861 struct io_kiocb *req;
862 u32 seq;
863 };
864
865 struct io_op_def {
866 /* needs req->file assigned */
867 unsigned needs_file : 1;
868 /* hash wq insertion if file is a regular file */
869 unsigned hash_reg_file : 1;
870 /* unbound wq insertion if file is a non-regular file */
871 unsigned unbound_nonreg_file : 1;
872 /* opcode is not supported by this kernel */
873 unsigned not_supported : 1;
874 /* set if opcode supports polled "wait" */
875 unsigned pollin : 1;
876 unsigned pollout : 1;
877 /* op supports buffer selection */
878 unsigned buffer_select : 1;
879 /* do prep async if is going to be punted */
880 unsigned needs_async_setup : 1;
881 /* should block plug */
882 unsigned plug : 1;
883 /* size of async data needed, if any */
884 unsigned short async_size;
885 };
886
887 static const struct io_op_def io_op_defs[] = {
888 [IORING_OP_NOP] = {},
889 [IORING_OP_READV] = {
890 .needs_file = 1,
891 .unbound_nonreg_file = 1,
892 .pollin = 1,
893 .buffer_select = 1,
894 .needs_async_setup = 1,
895 .plug = 1,
896 .async_size = sizeof(struct io_async_rw),
897 },
898 [IORING_OP_WRITEV] = {
899 .needs_file = 1,
900 .hash_reg_file = 1,
901 .unbound_nonreg_file = 1,
902 .pollout = 1,
903 .needs_async_setup = 1,
904 .plug = 1,
905 .async_size = sizeof(struct io_async_rw),
906 },
907 [IORING_OP_FSYNC] = {
908 .needs_file = 1,
909 },
910 [IORING_OP_READ_FIXED] = {
911 .needs_file = 1,
912 .unbound_nonreg_file = 1,
913 .pollin = 1,
914 .plug = 1,
915 .async_size = sizeof(struct io_async_rw),
916 },
917 [IORING_OP_WRITE_FIXED] = {
918 .needs_file = 1,
919 .hash_reg_file = 1,
920 .unbound_nonreg_file = 1,
921 .pollout = 1,
922 .plug = 1,
923 .async_size = sizeof(struct io_async_rw),
924 },
925 [IORING_OP_POLL_ADD] = {
926 .needs_file = 1,
927 .unbound_nonreg_file = 1,
928 },
929 [IORING_OP_POLL_REMOVE] = {},
930 [IORING_OP_SYNC_FILE_RANGE] = {
931 .needs_file = 1,
932 },
933 [IORING_OP_SENDMSG] = {
934 .needs_file = 1,
935 .unbound_nonreg_file = 1,
936 .pollout = 1,
937 .needs_async_setup = 1,
938 .async_size = sizeof(struct io_async_msghdr),
939 },
940 [IORING_OP_RECVMSG] = {
941 .needs_file = 1,
942 .unbound_nonreg_file = 1,
943 .pollin = 1,
944 .buffer_select = 1,
945 .needs_async_setup = 1,
946 .async_size = sizeof(struct io_async_msghdr),
947 },
948 [IORING_OP_TIMEOUT] = {
949 .async_size = sizeof(struct io_timeout_data),
950 },
951 [IORING_OP_TIMEOUT_REMOVE] = {
952 /* used by timeout updates' prep() */
953 },
954 [IORING_OP_ACCEPT] = {
955 .needs_file = 1,
956 .unbound_nonreg_file = 1,
957 .pollin = 1,
958 },
959 [IORING_OP_ASYNC_CANCEL] = {},
960 [IORING_OP_LINK_TIMEOUT] = {
961 .async_size = sizeof(struct io_timeout_data),
962 },
963 [IORING_OP_CONNECT] = {
964 .needs_file = 1,
965 .unbound_nonreg_file = 1,
966 .pollout = 1,
967 .needs_async_setup = 1,
968 .async_size = sizeof(struct io_async_connect),
969 },
970 [IORING_OP_FALLOCATE] = {
971 .needs_file = 1,
972 },
973 [IORING_OP_OPENAT] = {},
974 [IORING_OP_CLOSE] = {},
975 [IORING_OP_FILES_UPDATE] = {},
976 [IORING_OP_STATX] = {},
977 [IORING_OP_READ] = {
978 .needs_file = 1,
979 .unbound_nonreg_file = 1,
980 .pollin = 1,
981 .buffer_select = 1,
982 .plug = 1,
983 .async_size = sizeof(struct io_async_rw),
984 },
985 [IORING_OP_WRITE] = {
986 .needs_file = 1,
987 .unbound_nonreg_file = 1,
988 .pollout = 1,
989 .plug = 1,
990 .async_size = sizeof(struct io_async_rw),
991 },
992 [IORING_OP_FADVISE] = {
993 .needs_file = 1,
994 },
995 [IORING_OP_MADVISE] = {},
996 [IORING_OP_SEND] = {
997 .needs_file = 1,
998 .unbound_nonreg_file = 1,
999 .pollout = 1,
1000 },
1001 [IORING_OP_RECV] = {
1002 .needs_file = 1,
1003 .unbound_nonreg_file = 1,
1004 .pollin = 1,
1005 .buffer_select = 1,
1006 },
1007 [IORING_OP_OPENAT2] = {
1008 },
1009 [IORING_OP_EPOLL_CTL] = {
1010 .unbound_nonreg_file = 1,
1011 },
1012 [IORING_OP_SPLICE] = {
1013 .needs_file = 1,
1014 .hash_reg_file = 1,
1015 .unbound_nonreg_file = 1,
1016 },
1017 [IORING_OP_PROVIDE_BUFFERS] = {},
1018 [IORING_OP_REMOVE_BUFFERS] = {},
1019 [IORING_OP_TEE] = {
1020 .needs_file = 1,
1021 .hash_reg_file = 1,
1022 .unbound_nonreg_file = 1,
1023 },
1024 [IORING_OP_SHUTDOWN] = {
1025 .needs_file = 1,
1026 },
1027 [IORING_OP_RENAMEAT] = {},
1028 [IORING_OP_UNLINKAT] = {},
1029 };
1030
1031 static bool io_disarm_next(struct io_kiocb *req);
1032 static void io_uring_del_task_file(unsigned long index);
1033 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1034 struct task_struct *task,
1035 struct files_struct *files);
1036 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1037 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1038
1039 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1040 long res, unsigned int cflags);
1041 static void io_put_req(struct io_kiocb *req);
1042 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1043 static void io_dismantle_req(struct io_kiocb *req);
1044 static void io_put_task(struct task_struct *task, int nr);
1045 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1046 static void io_queue_linked_timeout(struct io_kiocb *req);
1047 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1048 struct io_uring_rsrc_update2 *up,
1049 unsigned nr_args);
1050 static void io_clean_op(struct io_kiocb *req);
1051 static struct file *io_file_get(struct io_submit_state *state,
1052 struct io_kiocb *req, int fd, bool fixed);
1053 static void __io_queue_sqe(struct io_kiocb *req);
1054 static void io_rsrc_put_work(struct work_struct *work);
1055
1056 static void io_req_task_queue(struct io_kiocb *req);
1057 static void io_submit_flush_completions(struct io_comp_state *cs,
1058 struct io_ring_ctx *ctx);
1059 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1060 static int io_req_prep_async(struct io_kiocb *req);
1061
1062 static struct kmem_cache *req_cachep;
1063
1064 static const struct file_operations io_uring_fops;
1065
io_uring_get_socket(struct file * file)1066 struct sock *io_uring_get_socket(struct file *file)
1067 {
1068 #if defined(CONFIG_UNIX)
1069 if (file->f_op == &io_uring_fops) {
1070 struct io_ring_ctx *ctx = file->private_data;
1071
1072 return ctx->ring_sock->sk;
1073 }
1074 #endif
1075 return NULL;
1076 }
1077 EXPORT_SYMBOL(io_uring_get_socket);
1078
1079 #define io_for_each_link(pos, head) \
1080 for (pos = (head); pos; pos = pos->link)
1081
io_req_set_rsrc_node(struct io_kiocb * req)1082 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1083 {
1084 struct io_ring_ctx *ctx = req->ctx;
1085
1086 if (!req->fixed_rsrc_refs) {
1087 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1088 percpu_ref_get(req->fixed_rsrc_refs);
1089 }
1090 }
1091
io_refs_resurrect(struct percpu_ref * ref,struct completion * compl)1092 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1093 {
1094 bool got = percpu_ref_tryget(ref);
1095
1096 /* already at zero, wait for ->release() */
1097 if (!got)
1098 wait_for_completion(compl);
1099 percpu_ref_resurrect(ref);
1100 if (got)
1101 percpu_ref_put(ref);
1102 }
1103
io_match_task(struct io_kiocb * head,struct task_struct * task,struct files_struct * files)1104 static bool io_match_task(struct io_kiocb *head,
1105 struct task_struct *task,
1106 struct files_struct *files)
1107 {
1108 struct io_kiocb *req;
1109
1110 if (task && head->task != task)
1111 return false;
1112 if (!files)
1113 return true;
1114
1115 io_for_each_link(req, head) {
1116 if (req->flags & REQ_F_INFLIGHT)
1117 return true;
1118 }
1119 return false;
1120 }
1121
req_set_fail_links(struct io_kiocb * req)1122 static inline void req_set_fail_links(struct io_kiocb *req)
1123 {
1124 if (req->flags & REQ_F_LINK)
1125 req->flags |= REQ_F_FAIL_LINK;
1126 }
1127
io_ring_ctx_ref_free(struct percpu_ref * ref)1128 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1129 {
1130 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1131
1132 complete(&ctx->ref_comp);
1133 }
1134
io_is_timeout_noseq(struct io_kiocb * req)1135 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1136 {
1137 return !req->timeout.off;
1138 }
1139
io_ring_ctx_alloc(struct io_uring_params * p)1140 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1141 {
1142 struct io_ring_ctx *ctx;
1143 int hash_bits;
1144
1145 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1146 if (!ctx)
1147 return NULL;
1148
1149 /*
1150 * Use 5 bits less than the max cq entries, that should give us around
1151 * 32 entries per hash list if totally full and uniformly spread.
1152 */
1153 hash_bits = ilog2(p->cq_entries);
1154 hash_bits -= 5;
1155 if (hash_bits <= 0)
1156 hash_bits = 1;
1157 ctx->cancel_hash_bits = hash_bits;
1158 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1159 GFP_KERNEL);
1160 if (!ctx->cancel_hash)
1161 goto err;
1162 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1163
1164 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1165 if (!ctx->dummy_ubuf)
1166 goto err;
1167 /* set invalid range, so io_import_fixed() fails meeting it */
1168 ctx->dummy_ubuf->ubuf = -1UL;
1169
1170 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1171 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1172 goto err;
1173
1174 ctx->flags = p->flags;
1175 init_waitqueue_head(&ctx->sqo_sq_wait);
1176 INIT_LIST_HEAD(&ctx->sqd_list);
1177 init_waitqueue_head(&ctx->cq_wait);
1178 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1179 init_completion(&ctx->ref_comp);
1180 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1181 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1182 mutex_init(&ctx->uring_lock);
1183 init_waitqueue_head(&ctx->wait);
1184 spin_lock_init(&ctx->completion_lock);
1185 INIT_LIST_HEAD(&ctx->iopoll_list);
1186 INIT_LIST_HEAD(&ctx->defer_list);
1187 INIT_LIST_HEAD(&ctx->timeout_list);
1188 spin_lock_init(&ctx->rsrc_ref_lock);
1189 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1190 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1191 init_llist_head(&ctx->rsrc_put_llist);
1192 INIT_LIST_HEAD(&ctx->tctx_list);
1193 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1194 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1195 return ctx;
1196 err:
1197 kfree(ctx->dummy_ubuf);
1198 kfree(ctx->cancel_hash);
1199 kfree(ctx);
1200 return NULL;
1201 }
1202
req_need_defer(struct io_kiocb * req,u32 seq)1203 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1204 {
1205 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1206 struct io_ring_ctx *ctx = req->ctx;
1207
1208 return seq + ctx->cq_extra != ctx->cached_cq_tail
1209 + READ_ONCE(ctx->cached_cq_overflow);
1210 }
1211
1212 return false;
1213 }
1214
io_req_track_inflight(struct io_kiocb * req)1215 static void io_req_track_inflight(struct io_kiocb *req)
1216 {
1217 if (!(req->flags & REQ_F_INFLIGHT)) {
1218 req->flags |= REQ_F_INFLIGHT;
1219 atomic_inc(¤t->io_uring->inflight_tracked);
1220 }
1221 }
1222
io_prep_async_work(struct io_kiocb * req)1223 static void io_prep_async_work(struct io_kiocb *req)
1224 {
1225 const struct io_op_def *def = &io_op_defs[req->opcode];
1226 struct io_ring_ctx *ctx = req->ctx;
1227
1228 if (!req->work.creds)
1229 req->work.creds = get_current_cred();
1230
1231 req->work.list.next = NULL;
1232 req->work.flags = 0;
1233 if (req->flags & REQ_F_FORCE_ASYNC)
1234 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1235
1236 if (req->flags & REQ_F_ISREG) {
1237 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1238 io_wq_hash_work(&req->work, file_inode(req->file));
1239 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1240 if (def->unbound_nonreg_file)
1241 req->work.flags |= IO_WQ_WORK_UNBOUND;
1242 }
1243
1244 switch (req->opcode) {
1245 case IORING_OP_SPLICE:
1246 case IORING_OP_TEE:
1247 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1248 req->work.flags |= IO_WQ_WORK_UNBOUND;
1249 break;
1250 }
1251 }
1252
io_prep_async_link(struct io_kiocb * req)1253 static void io_prep_async_link(struct io_kiocb *req)
1254 {
1255 struct io_kiocb *cur;
1256
1257 io_for_each_link(cur, req)
1258 io_prep_async_work(cur);
1259 }
1260
io_queue_async_work(struct io_kiocb * req)1261 static void io_queue_async_work(struct io_kiocb *req)
1262 {
1263 struct io_ring_ctx *ctx = req->ctx;
1264 struct io_kiocb *link = io_prep_linked_timeout(req);
1265 struct io_uring_task *tctx = req->task->io_uring;
1266
1267 BUG_ON(!tctx);
1268 BUG_ON(!tctx->io_wq);
1269
1270 /* init ->work of the whole link before punting */
1271 io_prep_async_link(req);
1272 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1273 &req->work, req->flags);
1274 io_wq_enqueue(tctx->io_wq, &req->work);
1275 if (link)
1276 io_queue_linked_timeout(link);
1277 }
1278
io_kill_timeout(struct io_kiocb * req,int status)1279 static void io_kill_timeout(struct io_kiocb *req, int status)
1280 __must_hold(&req->ctx->completion_lock)
1281 {
1282 struct io_timeout_data *io = req->async_data;
1283
1284 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1285 atomic_set(&req->ctx->cq_timeouts,
1286 atomic_read(&req->ctx->cq_timeouts) + 1);
1287 list_del_init(&req->timeout.list);
1288 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1289 io_put_req_deferred(req, 1);
1290 }
1291 }
1292
__io_queue_deferred(struct io_ring_ctx * ctx)1293 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1294 {
1295 do {
1296 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1297 struct io_defer_entry, list);
1298
1299 if (req_need_defer(de->req, de->seq))
1300 break;
1301 list_del_init(&de->list);
1302 io_req_task_queue(de->req);
1303 kfree(de);
1304 } while (!list_empty(&ctx->defer_list));
1305 }
1306
io_flush_timeouts(struct io_ring_ctx * ctx)1307 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1308 {
1309 u32 seq;
1310
1311 if (list_empty(&ctx->timeout_list))
1312 return;
1313
1314 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1315
1316 do {
1317 u32 events_needed, events_got;
1318 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1319 struct io_kiocb, timeout.list);
1320
1321 if (io_is_timeout_noseq(req))
1322 break;
1323
1324 /*
1325 * Since seq can easily wrap around over time, subtract
1326 * the last seq at which timeouts were flushed before comparing.
1327 * Assuming not more than 2^31-1 events have happened since,
1328 * these subtractions won't have wrapped, so we can check if
1329 * target is in [last_seq, current_seq] by comparing the two.
1330 */
1331 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1332 events_got = seq - ctx->cq_last_tm_flush;
1333 if (events_got < events_needed)
1334 break;
1335
1336 list_del_init(&req->timeout.list);
1337 io_kill_timeout(req, 0);
1338 } while (!list_empty(&ctx->timeout_list));
1339
1340 ctx->cq_last_tm_flush = seq;
1341 }
1342
io_commit_cqring(struct io_ring_ctx * ctx)1343 static void io_commit_cqring(struct io_ring_ctx *ctx)
1344 {
1345 io_flush_timeouts(ctx);
1346
1347 /* order cqe stores with ring update */
1348 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1349
1350 if (unlikely(!list_empty(&ctx->defer_list)))
1351 __io_queue_deferred(ctx);
1352 }
1353
io_sqring_full(struct io_ring_ctx * ctx)1354 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1355 {
1356 struct io_rings *r = ctx->rings;
1357
1358 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1359 }
1360
__io_cqring_events(struct io_ring_ctx * ctx)1361 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1362 {
1363 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1364 }
1365
io_get_cqring(struct io_ring_ctx * ctx)1366 static inline struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1367 {
1368 struct io_rings *rings = ctx->rings;
1369 unsigned tail;
1370
1371 /*
1372 * writes to the cq entry need to come after reading head; the
1373 * control dependency is enough as we're using WRITE_ONCE to
1374 * fill the cq entry
1375 */
1376 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1377 return NULL;
1378
1379 tail = ctx->cached_cq_tail++;
1380 return &rings->cqes[tail & ctx->cq_mask];
1381 }
1382
io_should_trigger_evfd(struct io_ring_ctx * ctx)1383 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1384 {
1385 if (likely(!ctx->cq_ev_fd))
1386 return false;
1387 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1388 return false;
1389 return !ctx->eventfd_async || io_wq_current_is_worker();
1390 }
1391
io_cqring_ev_posted(struct io_ring_ctx * ctx)1392 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1393 {
1394 /* see waitqueue_active() comment */
1395 smp_mb();
1396
1397 if (waitqueue_active(&ctx->wait))
1398 wake_up(&ctx->wait);
1399 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1400 wake_up(&ctx->sq_data->wait);
1401 if (io_should_trigger_evfd(ctx))
1402 eventfd_signal(ctx->cq_ev_fd, 1);
1403 if (waitqueue_active(&ctx->cq_wait)) {
1404 wake_up_interruptible(&ctx->cq_wait);
1405 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1406 }
1407 }
1408
io_cqring_ev_posted_iopoll(struct io_ring_ctx * ctx)1409 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1410 {
1411 /* see waitqueue_active() comment */
1412 smp_mb();
1413
1414 if (ctx->flags & IORING_SETUP_SQPOLL) {
1415 if (waitqueue_active(&ctx->wait))
1416 wake_up(&ctx->wait);
1417 }
1418 if (io_should_trigger_evfd(ctx))
1419 eventfd_signal(ctx->cq_ev_fd, 1);
1420 if (waitqueue_active(&ctx->cq_wait)) {
1421 wake_up_interruptible(&ctx->cq_wait);
1422 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1423 }
1424 }
1425
1426 /* Returns true if there are no backlogged entries after the flush */
__io_cqring_overflow_flush(struct io_ring_ctx * ctx,bool force)1427 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1428 {
1429 struct io_rings *rings = ctx->rings;
1430 unsigned long flags;
1431 bool all_flushed, posted;
1432
1433 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1434 return false;
1435
1436 posted = false;
1437 spin_lock_irqsave(&ctx->completion_lock, flags);
1438 while (!list_empty(&ctx->cq_overflow_list)) {
1439 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1440 struct io_overflow_cqe *ocqe;
1441
1442 if (!cqe && !force)
1443 break;
1444 ocqe = list_first_entry(&ctx->cq_overflow_list,
1445 struct io_overflow_cqe, list);
1446 if (cqe)
1447 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1448 else
1449 WRITE_ONCE(ctx->rings->cq_overflow,
1450 ++ctx->cached_cq_overflow);
1451 posted = true;
1452 list_del(&ocqe->list);
1453 kfree(ocqe);
1454 }
1455
1456 all_flushed = list_empty(&ctx->cq_overflow_list);
1457 if (all_flushed) {
1458 clear_bit(0, &ctx->sq_check_overflow);
1459 clear_bit(0, &ctx->cq_check_overflow);
1460 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1461 }
1462
1463 if (posted)
1464 io_commit_cqring(ctx);
1465 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1466 if (posted)
1467 io_cqring_ev_posted(ctx);
1468 return all_flushed;
1469 }
1470
io_cqring_overflow_flush(struct io_ring_ctx * ctx,bool force)1471 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1472 {
1473 bool ret = true;
1474
1475 if (test_bit(0, &ctx->cq_check_overflow)) {
1476 /* iopoll syncs against uring_lock, not completion_lock */
1477 if (ctx->flags & IORING_SETUP_IOPOLL)
1478 mutex_lock(&ctx->uring_lock);
1479 ret = __io_cqring_overflow_flush(ctx, force);
1480 if (ctx->flags & IORING_SETUP_IOPOLL)
1481 mutex_unlock(&ctx->uring_lock);
1482 }
1483
1484 return ret;
1485 }
1486
1487 /*
1488 * Shamelessly stolen from the mm implementation of page reference checking,
1489 * see commit f958d7b528b1 for details.
1490 */
1491 #define req_ref_zero_or_close_to_overflow(req) \
1492 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1493
req_ref_inc_not_zero(struct io_kiocb * req)1494 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1495 {
1496 return atomic_inc_not_zero(&req->refs);
1497 }
1498
req_ref_sub_and_test(struct io_kiocb * req,int refs)1499 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1500 {
1501 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1502 return atomic_sub_and_test(refs, &req->refs);
1503 }
1504
req_ref_put_and_test(struct io_kiocb * req)1505 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1506 {
1507 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1508 return atomic_dec_and_test(&req->refs);
1509 }
1510
req_ref_put(struct io_kiocb * req)1511 static inline void req_ref_put(struct io_kiocb *req)
1512 {
1513 WARN_ON_ONCE(req_ref_put_and_test(req));
1514 }
1515
req_ref_get(struct io_kiocb * req)1516 static inline void req_ref_get(struct io_kiocb *req)
1517 {
1518 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1519 atomic_inc(&req->refs);
1520 }
1521
io_cqring_event_overflow(struct io_ring_ctx * ctx,u64 user_data,long res,unsigned int cflags)1522 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1523 long res, unsigned int cflags)
1524 {
1525 struct io_overflow_cqe *ocqe;
1526
1527 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1528 if (!ocqe) {
1529 /*
1530 * If we're in ring overflow flush mode, or in task cancel mode,
1531 * or cannot allocate an overflow entry, then we need to drop it
1532 * on the floor.
1533 */
1534 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1535 return false;
1536 }
1537 if (list_empty(&ctx->cq_overflow_list)) {
1538 set_bit(0, &ctx->sq_check_overflow);
1539 set_bit(0, &ctx->cq_check_overflow);
1540 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1541 }
1542 ocqe->cqe.user_data = user_data;
1543 ocqe->cqe.res = res;
1544 ocqe->cqe.flags = cflags;
1545 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1546 return true;
1547 }
1548
__io_cqring_fill_event(struct io_ring_ctx * ctx,u64 user_data,long res,unsigned int cflags)1549 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1550 long res, unsigned int cflags)
1551 {
1552 struct io_uring_cqe *cqe;
1553
1554 trace_io_uring_complete(ctx, user_data, res, cflags);
1555
1556 /*
1557 * If we can't get a cq entry, userspace overflowed the
1558 * submission (by quite a lot). Increment the overflow count in
1559 * the ring.
1560 */
1561 cqe = io_get_cqring(ctx);
1562 if (likely(cqe)) {
1563 WRITE_ONCE(cqe->user_data, user_data);
1564 WRITE_ONCE(cqe->res, res);
1565 WRITE_ONCE(cqe->flags, cflags);
1566 return true;
1567 }
1568 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1569 }
1570
1571 /* not as hot to bloat with inlining */
io_cqring_fill_event(struct io_ring_ctx * ctx,u64 user_data,long res,unsigned int cflags)1572 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1573 long res, unsigned int cflags)
1574 {
1575 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1576 }
1577
io_req_complete_post(struct io_kiocb * req,long res,unsigned int cflags)1578 static void io_req_complete_post(struct io_kiocb *req, long res,
1579 unsigned int cflags)
1580 {
1581 struct io_ring_ctx *ctx = req->ctx;
1582 unsigned long flags;
1583
1584 spin_lock_irqsave(&ctx->completion_lock, flags);
1585 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1586 /*
1587 * If we're the last reference to this request, add to our locked
1588 * free_list cache.
1589 */
1590 if (req_ref_put_and_test(req)) {
1591 struct io_comp_state *cs = &ctx->submit_state.comp;
1592
1593 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1594 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1595 io_disarm_next(req);
1596 if (req->link) {
1597 io_req_task_queue(req->link);
1598 req->link = NULL;
1599 }
1600 }
1601 io_dismantle_req(req);
1602 io_put_task(req->task, 1);
1603 list_add(&req->compl.list, &cs->locked_free_list);
1604 cs->locked_free_nr++;
1605 } else {
1606 if (!percpu_ref_tryget(&ctx->refs))
1607 req = NULL;
1608 }
1609 io_commit_cqring(ctx);
1610 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1611
1612 if (req) {
1613 io_cqring_ev_posted(ctx);
1614 percpu_ref_put(&ctx->refs);
1615 }
1616 }
1617
io_req_needs_clean(struct io_kiocb * req)1618 static inline bool io_req_needs_clean(struct io_kiocb *req)
1619 {
1620 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1621 REQ_F_POLLED | REQ_F_INFLIGHT);
1622 }
1623
io_req_complete_state(struct io_kiocb * req,long res,unsigned int cflags)1624 static void io_req_complete_state(struct io_kiocb *req, long res,
1625 unsigned int cflags)
1626 {
1627 if (io_req_needs_clean(req))
1628 io_clean_op(req);
1629 req->result = res;
1630 req->compl.cflags = cflags;
1631 req->flags |= REQ_F_COMPLETE_INLINE;
1632 }
1633
__io_req_complete(struct io_kiocb * req,unsigned issue_flags,long res,unsigned cflags)1634 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1635 long res, unsigned cflags)
1636 {
1637 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1638 io_req_complete_state(req, res, cflags);
1639 else
1640 io_req_complete_post(req, res, cflags);
1641 }
1642
io_req_complete(struct io_kiocb * req,long res)1643 static inline void io_req_complete(struct io_kiocb *req, long res)
1644 {
1645 __io_req_complete(req, 0, res, 0);
1646 }
1647
io_req_complete_failed(struct io_kiocb * req,long res)1648 static void io_req_complete_failed(struct io_kiocb *req, long res)
1649 {
1650 req_set_fail_links(req);
1651 io_put_req(req);
1652 io_req_complete_post(req, res, 0);
1653 }
1654
io_flush_cached_locked_reqs(struct io_ring_ctx * ctx,struct io_comp_state * cs)1655 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1656 struct io_comp_state *cs)
1657 {
1658 spin_lock_irq(&ctx->completion_lock);
1659 list_splice_init(&cs->locked_free_list, &cs->free_list);
1660 cs->locked_free_nr = 0;
1661 spin_unlock_irq(&ctx->completion_lock);
1662 }
1663
1664 /* Returns true IFF there are requests in the cache */
io_flush_cached_reqs(struct io_ring_ctx * ctx)1665 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1666 {
1667 struct io_submit_state *state = &ctx->submit_state;
1668 struct io_comp_state *cs = &state->comp;
1669 int nr;
1670
1671 /*
1672 * If we have more than a batch's worth of requests in our IRQ side
1673 * locked cache, grab the lock and move them over to our submission
1674 * side cache.
1675 */
1676 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1677 io_flush_cached_locked_reqs(ctx, cs);
1678
1679 nr = state->free_reqs;
1680 while (!list_empty(&cs->free_list)) {
1681 struct io_kiocb *req = list_first_entry(&cs->free_list,
1682 struct io_kiocb, compl.list);
1683
1684 list_del(&req->compl.list);
1685 state->reqs[nr++] = req;
1686 if (nr == ARRAY_SIZE(state->reqs))
1687 break;
1688 }
1689
1690 state->free_reqs = nr;
1691 return nr != 0;
1692 }
1693
io_alloc_req(struct io_ring_ctx * ctx)1694 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1695 {
1696 struct io_submit_state *state = &ctx->submit_state;
1697
1698 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1699
1700 if (!state->free_reqs) {
1701 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1702 int ret;
1703
1704 if (io_flush_cached_reqs(ctx))
1705 goto got_req;
1706
1707 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1708 state->reqs);
1709
1710 /*
1711 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1712 * retry single alloc to be on the safe side.
1713 */
1714 if (unlikely(ret <= 0)) {
1715 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1716 if (!state->reqs[0])
1717 return NULL;
1718 ret = 1;
1719 }
1720 state->free_reqs = ret;
1721 }
1722 got_req:
1723 state->free_reqs--;
1724 return state->reqs[state->free_reqs];
1725 }
1726
io_put_file(struct file * file)1727 static inline void io_put_file(struct file *file)
1728 {
1729 if (file)
1730 fput(file);
1731 }
1732
io_dismantle_req(struct io_kiocb * req)1733 static void io_dismantle_req(struct io_kiocb *req)
1734 {
1735 unsigned int flags = req->flags;
1736
1737 if (io_req_needs_clean(req))
1738 io_clean_op(req);
1739 if (!(flags & REQ_F_FIXED_FILE))
1740 io_put_file(req->file);
1741 if (req->fixed_rsrc_refs)
1742 percpu_ref_put(req->fixed_rsrc_refs);
1743 if (req->async_data)
1744 kfree(req->async_data);
1745 if (req->work.creds) {
1746 put_cred(req->work.creds);
1747 req->work.creds = NULL;
1748 }
1749 }
1750
1751 /* must to be called somewhat shortly after putting a request */
io_put_task(struct task_struct * task,int nr)1752 static inline void io_put_task(struct task_struct *task, int nr)
1753 {
1754 struct io_uring_task *tctx = task->io_uring;
1755
1756 percpu_counter_sub(&tctx->inflight, nr);
1757 if (unlikely(atomic_read(&tctx->in_idle)))
1758 wake_up(&tctx->wait);
1759 put_task_struct_many(task, nr);
1760 }
1761
__io_free_req(struct io_kiocb * req)1762 static void __io_free_req(struct io_kiocb *req)
1763 {
1764 struct io_ring_ctx *ctx = req->ctx;
1765
1766 io_dismantle_req(req);
1767 io_put_task(req->task, 1);
1768
1769 kmem_cache_free(req_cachep, req);
1770 percpu_ref_put(&ctx->refs);
1771 }
1772
io_remove_next_linked(struct io_kiocb * req)1773 static inline void io_remove_next_linked(struct io_kiocb *req)
1774 {
1775 struct io_kiocb *nxt = req->link;
1776
1777 req->link = nxt->link;
1778 nxt->link = NULL;
1779 }
1780
io_kill_linked_timeout(struct io_kiocb * req)1781 static bool io_kill_linked_timeout(struct io_kiocb *req)
1782 __must_hold(&req->ctx->completion_lock)
1783 {
1784 struct io_kiocb *link = req->link;
1785
1786 /*
1787 * Can happen if a linked timeout fired and link had been like
1788 * req -> link t-out -> link t-out [-> ...]
1789 */
1790 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1791 struct io_timeout_data *io = link->async_data;
1792
1793 io_remove_next_linked(req);
1794 link->timeout.head = NULL;
1795 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1796 io_cqring_fill_event(link->ctx, link->user_data,
1797 -ECANCELED, 0);
1798 io_put_req_deferred(link, 1);
1799 return true;
1800 }
1801 }
1802 return false;
1803 }
1804
io_fail_links(struct io_kiocb * req)1805 static void io_fail_links(struct io_kiocb *req)
1806 __must_hold(&req->ctx->completion_lock)
1807 {
1808 struct io_kiocb *nxt, *link = req->link;
1809
1810 req->link = NULL;
1811 while (link) {
1812 nxt = link->link;
1813 link->link = NULL;
1814
1815 trace_io_uring_fail_link(req, link);
1816 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1817 io_put_req_deferred(link, 2);
1818 link = nxt;
1819 }
1820 }
1821
io_disarm_next(struct io_kiocb * req)1822 static bool io_disarm_next(struct io_kiocb *req)
1823 __must_hold(&req->ctx->completion_lock)
1824 {
1825 bool posted = false;
1826
1827 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1828 posted = io_kill_linked_timeout(req);
1829 if (unlikely((req->flags & REQ_F_FAIL_LINK) &&
1830 !(req->flags & REQ_F_HARDLINK))) {
1831 posted |= (req->link != NULL);
1832 io_fail_links(req);
1833 }
1834 return posted;
1835 }
1836
__io_req_find_next(struct io_kiocb * req)1837 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1838 {
1839 struct io_kiocb *nxt;
1840
1841 /*
1842 * If LINK is set, we have dependent requests in this chain. If we
1843 * didn't fail this request, queue the first one up, moving any other
1844 * dependencies to the next request. In case of failure, fail the rest
1845 * of the chain.
1846 */
1847 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1848 struct io_ring_ctx *ctx = req->ctx;
1849 unsigned long flags;
1850 bool posted;
1851
1852 spin_lock_irqsave(&ctx->completion_lock, flags);
1853 posted = io_disarm_next(req);
1854 if (posted)
1855 io_commit_cqring(req->ctx);
1856 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1857 if (posted)
1858 io_cqring_ev_posted(ctx);
1859 }
1860 nxt = req->link;
1861 req->link = NULL;
1862 return nxt;
1863 }
1864
io_req_find_next(struct io_kiocb * req)1865 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1866 {
1867 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1868 return NULL;
1869 return __io_req_find_next(req);
1870 }
1871
ctx_flush_and_put(struct io_ring_ctx * ctx)1872 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1873 {
1874 if (!ctx)
1875 return;
1876 if (ctx->submit_state.comp.nr) {
1877 mutex_lock(&ctx->uring_lock);
1878 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1879 mutex_unlock(&ctx->uring_lock);
1880 }
1881 percpu_ref_put(&ctx->refs);
1882 }
1883
__tctx_task_work(struct io_uring_task * tctx)1884 static bool __tctx_task_work(struct io_uring_task *tctx)
1885 {
1886 struct io_ring_ctx *ctx = NULL;
1887 struct io_wq_work_list list;
1888 struct io_wq_work_node *node;
1889
1890 if (wq_list_empty(&tctx->task_list))
1891 return false;
1892
1893 spin_lock_irq(&tctx->task_lock);
1894 list = tctx->task_list;
1895 INIT_WQ_LIST(&tctx->task_list);
1896 spin_unlock_irq(&tctx->task_lock);
1897
1898 node = list.first;
1899 while (node) {
1900 struct io_wq_work_node *next = node->next;
1901 struct io_kiocb *req;
1902
1903 req = container_of(node, struct io_kiocb, io_task_work.node);
1904 if (req->ctx != ctx) {
1905 ctx_flush_and_put(ctx);
1906 ctx = req->ctx;
1907 percpu_ref_get(&ctx->refs);
1908 }
1909
1910 req->task_work.func(&req->task_work);
1911 node = next;
1912 }
1913
1914 ctx_flush_and_put(ctx);
1915 return list.first != NULL;
1916 }
1917
tctx_task_work(struct callback_head * cb)1918 static void tctx_task_work(struct callback_head *cb)
1919 {
1920 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1921
1922 clear_bit(0, &tctx->task_state);
1923
1924 while (__tctx_task_work(tctx))
1925 cond_resched();
1926 }
1927
io_req_task_work_add(struct io_kiocb * req)1928 static int io_req_task_work_add(struct io_kiocb *req)
1929 {
1930 struct task_struct *tsk = req->task;
1931 struct io_uring_task *tctx = tsk->io_uring;
1932 enum task_work_notify_mode notify;
1933 struct io_wq_work_node *node, *prev;
1934 unsigned long flags;
1935 int ret = 0;
1936
1937 if (unlikely(tsk->flags & PF_EXITING))
1938 return -ESRCH;
1939
1940 WARN_ON_ONCE(!tctx);
1941
1942 spin_lock_irqsave(&tctx->task_lock, flags);
1943 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1944 spin_unlock_irqrestore(&tctx->task_lock, flags);
1945
1946 /* task_work already pending, we're done */
1947 if (test_bit(0, &tctx->task_state) ||
1948 test_and_set_bit(0, &tctx->task_state))
1949 return 0;
1950
1951 /*
1952 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1953 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1954 * processing task_work. There's no reliable way to tell if TWA_RESUME
1955 * will do the job.
1956 */
1957 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1958
1959 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1960 wake_up_process(tsk);
1961 return 0;
1962 }
1963
1964 /*
1965 * Slow path - we failed, find and delete work. if the work is not
1966 * in the list, it got run and we're fine.
1967 */
1968 spin_lock_irqsave(&tctx->task_lock, flags);
1969 wq_list_for_each(node, prev, &tctx->task_list) {
1970 if (&req->io_task_work.node == node) {
1971 wq_list_del(&tctx->task_list, node, prev);
1972 ret = 1;
1973 break;
1974 }
1975 }
1976 spin_unlock_irqrestore(&tctx->task_lock, flags);
1977 clear_bit(0, &tctx->task_state);
1978 return ret;
1979 }
1980
io_run_task_work_head(struct callback_head ** work_head)1981 static bool io_run_task_work_head(struct callback_head **work_head)
1982 {
1983 struct callback_head *work, *next;
1984 bool executed = false;
1985
1986 do {
1987 work = xchg(work_head, NULL);
1988 if (!work)
1989 break;
1990
1991 do {
1992 next = work->next;
1993 work->func(work);
1994 work = next;
1995 cond_resched();
1996 } while (work);
1997 executed = true;
1998 } while (1);
1999
2000 return executed;
2001 }
2002
io_task_work_add_head(struct callback_head ** work_head,struct callback_head * task_work)2003 static void io_task_work_add_head(struct callback_head **work_head,
2004 struct callback_head *task_work)
2005 {
2006 struct callback_head *head;
2007
2008 do {
2009 head = READ_ONCE(*work_head);
2010 task_work->next = head;
2011 } while (cmpxchg(work_head, head, task_work) != head);
2012 }
2013
io_req_task_work_add_fallback(struct io_kiocb * req,task_work_func_t cb)2014 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2015 task_work_func_t cb)
2016 {
2017 init_task_work(&req->task_work, cb);
2018 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2019 }
2020
io_req_task_cancel(struct callback_head * cb)2021 static void io_req_task_cancel(struct callback_head *cb)
2022 {
2023 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2024 struct io_ring_ctx *ctx = req->ctx;
2025
2026 /* ctx is guaranteed to stay alive while we hold uring_lock */
2027 mutex_lock(&ctx->uring_lock);
2028 io_req_complete_failed(req, req->result);
2029 mutex_unlock(&ctx->uring_lock);
2030 }
2031
__io_req_task_submit(struct io_kiocb * req)2032 static void __io_req_task_submit(struct io_kiocb *req)
2033 {
2034 struct io_ring_ctx *ctx = req->ctx;
2035
2036 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2037 mutex_lock(&ctx->uring_lock);
2038 if (!(current->flags & PF_EXITING) && !current->in_execve)
2039 __io_queue_sqe(req);
2040 else
2041 io_req_complete_failed(req, -EFAULT);
2042 mutex_unlock(&ctx->uring_lock);
2043 }
2044
io_req_task_submit(struct callback_head * cb)2045 static void io_req_task_submit(struct callback_head *cb)
2046 {
2047 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2048
2049 __io_req_task_submit(req);
2050 }
2051
io_req_task_queue_fail(struct io_kiocb * req,int ret)2052 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2053 {
2054 req->result = ret;
2055 req->task_work.func = io_req_task_cancel;
2056
2057 if (unlikely(io_req_task_work_add(req)))
2058 io_req_task_work_add_fallback(req, io_req_task_cancel);
2059 }
2060
io_req_task_queue(struct io_kiocb * req)2061 static void io_req_task_queue(struct io_kiocb *req)
2062 {
2063 req->task_work.func = io_req_task_submit;
2064
2065 if (unlikely(io_req_task_work_add(req)))
2066 io_req_task_queue_fail(req, -ECANCELED);
2067 }
2068
io_queue_next(struct io_kiocb * req)2069 static inline void io_queue_next(struct io_kiocb *req)
2070 {
2071 struct io_kiocb *nxt = io_req_find_next(req);
2072
2073 if (nxt)
2074 io_req_task_queue(nxt);
2075 }
2076
io_free_req(struct io_kiocb * req)2077 static void io_free_req(struct io_kiocb *req)
2078 {
2079 io_queue_next(req);
2080 __io_free_req(req);
2081 }
2082
2083 struct req_batch {
2084 struct task_struct *task;
2085 int task_refs;
2086 int ctx_refs;
2087 };
2088
io_init_req_batch(struct req_batch * rb)2089 static inline void io_init_req_batch(struct req_batch *rb)
2090 {
2091 rb->task_refs = 0;
2092 rb->ctx_refs = 0;
2093 rb->task = NULL;
2094 }
2095
io_req_free_batch_finish(struct io_ring_ctx * ctx,struct req_batch * rb)2096 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2097 struct req_batch *rb)
2098 {
2099 if (rb->task)
2100 io_put_task(rb->task, rb->task_refs);
2101 if (rb->ctx_refs)
2102 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2103 }
2104
io_req_free_batch(struct req_batch * rb,struct io_kiocb * req,struct io_submit_state * state)2105 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2106 struct io_submit_state *state)
2107 {
2108 io_queue_next(req);
2109 io_dismantle_req(req);
2110
2111 if (req->task != rb->task) {
2112 if (rb->task)
2113 io_put_task(rb->task, rb->task_refs);
2114 rb->task = req->task;
2115 rb->task_refs = 0;
2116 }
2117 rb->task_refs++;
2118 rb->ctx_refs++;
2119
2120 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2121 state->reqs[state->free_reqs++] = req;
2122 else
2123 list_add(&req->compl.list, &state->comp.free_list);
2124 }
2125
io_submit_flush_completions(struct io_comp_state * cs,struct io_ring_ctx * ctx)2126 static void io_submit_flush_completions(struct io_comp_state *cs,
2127 struct io_ring_ctx *ctx)
2128 {
2129 int i, nr = cs->nr;
2130 struct io_kiocb *req;
2131 struct req_batch rb;
2132
2133 io_init_req_batch(&rb);
2134 spin_lock_irq(&ctx->completion_lock);
2135 for (i = 0; i < nr; i++) {
2136 req = cs->reqs[i];
2137 __io_cqring_fill_event(ctx, req->user_data, req->result,
2138 req->compl.cflags);
2139 }
2140 io_commit_cqring(ctx);
2141 spin_unlock_irq(&ctx->completion_lock);
2142
2143 io_cqring_ev_posted(ctx);
2144 for (i = 0; i < nr; i++) {
2145 req = cs->reqs[i];
2146
2147 /* submission and completion refs */
2148 if (req_ref_sub_and_test(req, 2))
2149 io_req_free_batch(&rb, req, &ctx->submit_state);
2150 }
2151
2152 io_req_free_batch_finish(ctx, &rb);
2153 cs->nr = 0;
2154 }
2155
2156 /*
2157 * Drop reference to request, return next in chain (if there is one) if this
2158 * was the last reference to this request.
2159 */
io_put_req_find_next(struct io_kiocb * req)2160 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2161 {
2162 struct io_kiocb *nxt = NULL;
2163
2164 if (req_ref_put_and_test(req)) {
2165 nxt = io_req_find_next(req);
2166 __io_free_req(req);
2167 }
2168 return nxt;
2169 }
2170
io_put_req(struct io_kiocb * req)2171 static inline void io_put_req(struct io_kiocb *req)
2172 {
2173 if (req_ref_put_and_test(req))
2174 io_free_req(req);
2175 }
2176
io_put_req_deferred_cb(struct callback_head * cb)2177 static void io_put_req_deferred_cb(struct callback_head *cb)
2178 {
2179 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2180
2181 io_free_req(req);
2182 }
2183
io_free_req_deferred(struct io_kiocb * req)2184 static void io_free_req_deferred(struct io_kiocb *req)
2185 {
2186 req->task_work.func = io_put_req_deferred_cb;
2187 if (unlikely(io_req_task_work_add(req)))
2188 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2189 }
2190
io_put_req_deferred(struct io_kiocb * req,int refs)2191 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2192 {
2193 if (req_ref_sub_and_test(req, refs))
2194 io_free_req_deferred(req);
2195 }
2196
io_cqring_events(struct io_ring_ctx * ctx)2197 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2198 {
2199 /* See comment at the top of this file */
2200 smp_rmb();
2201 return __io_cqring_events(ctx);
2202 }
2203
io_sqring_entries(struct io_ring_ctx * ctx)2204 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2205 {
2206 struct io_rings *rings = ctx->rings;
2207
2208 /* make sure SQ entry isn't read before tail */
2209 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2210 }
2211
io_put_kbuf(struct io_kiocb * req,struct io_buffer * kbuf)2212 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2213 {
2214 unsigned int cflags;
2215
2216 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2217 cflags |= IORING_CQE_F_BUFFER;
2218 req->flags &= ~REQ_F_BUFFER_SELECTED;
2219 kfree(kbuf);
2220 return cflags;
2221 }
2222
io_put_rw_kbuf(struct io_kiocb * req)2223 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2224 {
2225 struct io_buffer *kbuf;
2226
2227 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2228 return io_put_kbuf(req, kbuf);
2229 }
2230
io_run_task_work(void)2231 static inline bool io_run_task_work(void)
2232 {
2233 /*
2234 * Not safe to run on exiting task, and the task_work handling will
2235 * not add work to such a task.
2236 */
2237 if (unlikely(current->flags & PF_EXITING))
2238 return false;
2239 if (current->task_works) {
2240 __set_current_state(TASK_RUNNING);
2241 task_work_run();
2242 return true;
2243 }
2244
2245 return false;
2246 }
2247
2248 /*
2249 * Find and free completed poll iocbs
2250 */
io_iopoll_complete(struct io_ring_ctx * ctx,unsigned int * nr_events,struct list_head * done)2251 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2252 struct list_head *done)
2253 {
2254 struct req_batch rb;
2255 struct io_kiocb *req;
2256
2257 /* order with ->result store in io_complete_rw_iopoll() */
2258 smp_rmb();
2259
2260 io_init_req_batch(&rb);
2261 while (!list_empty(done)) {
2262 int cflags = 0;
2263
2264 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2265 list_del(&req->inflight_entry);
2266
2267 if (READ_ONCE(req->result) == -EAGAIN &&
2268 !(req->flags & REQ_F_DONT_REISSUE)) {
2269 req->iopoll_completed = 0;
2270 req_ref_get(req);
2271 io_queue_async_work(req);
2272 continue;
2273 }
2274
2275 if (req->flags & REQ_F_BUFFER_SELECTED)
2276 cflags = io_put_rw_kbuf(req);
2277
2278 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2279 (*nr_events)++;
2280
2281 if (req_ref_put_and_test(req))
2282 io_req_free_batch(&rb, req, &ctx->submit_state);
2283 }
2284
2285 io_commit_cqring(ctx);
2286 io_cqring_ev_posted_iopoll(ctx);
2287 io_req_free_batch_finish(ctx, &rb);
2288 }
2289
io_do_iopoll(struct io_ring_ctx * ctx,unsigned int * nr_events,long min)2290 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2291 long min)
2292 {
2293 struct io_kiocb *req, *tmp;
2294 LIST_HEAD(done);
2295 bool spin;
2296 int ret;
2297
2298 /*
2299 * Only spin for completions if we don't have multiple devices hanging
2300 * off our complete list, and we're under the requested amount.
2301 */
2302 spin = !ctx->poll_multi_file && *nr_events < min;
2303
2304 ret = 0;
2305 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2306 struct kiocb *kiocb = &req->rw.kiocb;
2307
2308 /*
2309 * Move completed and retryable entries to our local lists.
2310 * If we find a request that requires polling, break out
2311 * and complete those lists first, if we have entries there.
2312 */
2313 if (READ_ONCE(req->iopoll_completed)) {
2314 list_move_tail(&req->inflight_entry, &done);
2315 continue;
2316 }
2317 if (!list_empty(&done))
2318 break;
2319
2320 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2321 if (ret < 0)
2322 break;
2323
2324 /* iopoll may have completed current req */
2325 if (READ_ONCE(req->iopoll_completed))
2326 list_move_tail(&req->inflight_entry, &done);
2327
2328 if (ret && spin)
2329 spin = false;
2330 ret = 0;
2331 }
2332
2333 if (!list_empty(&done))
2334 io_iopoll_complete(ctx, nr_events, &done);
2335
2336 return ret;
2337 }
2338
2339 /*
2340 * We can't just wait for polled events to come to us, we have to actively
2341 * find and complete them.
2342 */
io_iopoll_try_reap_events(struct io_ring_ctx * ctx)2343 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2344 {
2345 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2346 return;
2347
2348 mutex_lock(&ctx->uring_lock);
2349 while (!list_empty(&ctx->iopoll_list)) {
2350 unsigned int nr_events = 0;
2351
2352 io_do_iopoll(ctx, &nr_events, 0);
2353
2354 /* let it sleep and repeat later if can't complete a request */
2355 if (nr_events == 0)
2356 break;
2357 /*
2358 * Ensure we allow local-to-the-cpu processing to take place,
2359 * in this case we need to ensure that we reap all events.
2360 * Also let task_work, etc. to progress by releasing the mutex
2361 */
2362 if (need_resched()) {
2363 mutex_unlock(&ctx->uring_lock);
2364 cond_resched();
2365 mutex_lock(&ctx->uring_lock);
2366 }
2367 }
2368 mutex_unlock(&ctx->uring_lock);
2369 }
2370
io_iopoll_check(struct io_ring_ctx * ctx,long min)2371 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2372 {
2373 unsigned int nr_events = 0;
2374 int ret = 0;
2375
2376 /*
2377 * We disallow the app entering submit/complete with polling, but we
2378 * still need to lock the ring to prevent racing with polled issue
2379 * that got punted to a workqueue.
2380 */
2381 mutex_lock(&ctx->uring_lock);
2382 /*
2383 * Don't enter poll loop if we already have events pending.
2384 * If we do, we can potentially be spinning for commands that
2385 * already triggered a CQE (eg in error).
2386 */
2387 if (test_bit(0, &ctx->cq_check_overflow))
2388 __io_cqring_overflow_flush(ctx, false);
2389 if (io_cqring_events(ctx))
2390 goto out;
2391 do {
2392 /*
2393 * If a submit got punted to a workqueue, we can have the
2394 * application entering polling for a command before it gets
2395 * issued. That app will hold the uring_lock for the duration
2396 * of the poll right here, so we need to take a breather every
2397 * now and then to ensure that the issue has a chance to add
2398 * the poll to the issued list. Otherwise we can spin here
2399 * forever, while the workqueue is stuck trying to acquire the
2400 * very same mutex.
2401 */
2402 if (list_empty(&ctx->iopoll_list)) {
2403 mutex_unlock(&ctx->uring_lock);
2404 io_run_task_work();
2405 mutex_lock(&ctx->uring_lock);
2406
2407 if (list_empty(&ctx->iopoll_list))
2408 break;
2409 }
2410 ret = io_do_iopoll(ctx, &nr_events, min);
2411 } while (!ret && nr_events < min && !need_resched());
2412 out:
2413 mutex_unlock(&ctx->uring_lock);
2414 return ret;
2415 }
2416
kiocb_end_write(struct io_kiocb * req)2417 static void kiocb_end_write(struct io_kiocb *req)
2418 {
2419 /*
2420 * Tell lockdep we inherited freeze protection from submission
2421 * thread.
2422 */
2423 if (req->flags & REQ_F_ISREG) {
2424 struct super_block *sb = file_inode(req->file)->i_sb;
2425
2426 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2427 sb_end_write(sb);
2428 }
2429 }
2430
2431 #ifdef CONFIG_BLOCK
io_resubmit_prep(struct io_kiocb * req)2432 static bool io_resubmit_prep(struct io_kiocb *req)
2433 {
2434 struct io_async_rw *rw = req->async_data;
2435
2436 if (!rw)
2437 return !io_req_prep_async(req);
2438 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2439 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2440 return true;
2441 }
2442
io_rw_should_reissue(struct io_kiocb * req)2443 static bool io_rw_should_reissue(struct io_kiocb *req)
2444 {
2445 umode_t mode = file_inode(req->file)->i_mode;
2446 struct io_ring_ctx *ctx = req->ctx;
2447
2448 if (!S_ISBLK(mode) && !S_ISREG(mode))
2449 return false;
2450 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2451 !(ctx->flags & IORING_SETUP_IOPOLL)))
2452 return false;
2453 /*
2454 * If ref is dying, we might be running poll reap from the exit work.
2455 * Don't attempt to reissue from that path, just let it fail with
2456 * -EAGAIN.
2457 */
2458 if (percpu_ref_is_dying(&ctx->refs))
2459 return false;
2460 return true;
2461 }
2462 #else
io_resubmit_prep(struct io_kiocb * req)2463 static bool io_resubmit_prep(struct io_kiocb *req)
2464 {
2465 return false;
2466 }
io_rw_should_reissue(struct io_kiocb * req)2467 static bool io_rw_should_reissue(struct io_kiocb *req)
2468 {
2469 return false;
2470 }
2471 #endif
2472
__io_complete_rw(struct io_kiocb * req,long res,long res2,unsigned int issue_flags)2473 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2474 unsigned int issue_flags)
2475 {
2476 int cflags = 0;
2477
2478 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2479 kiocb_end_write(req);
2480 if (res != req->result) {
2481 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2482 io_rw_should_reissue(req)) {
2483 req->flags |= REQ_F_REISSUE;
2484 return;
2485 }
2486 req_set_fail_links(req);
2487 }
2488 if (req->flags & REQ_F_BUFFER_SELECTED)
2489 cflags = io_put_rw_kbuf(req);
2490 __io_req_complete(req, issue_flags, res, cflags);
2491 }
2492
io_complete_rw(struct kiocb * kiocb,long res,long res2)2493 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2494 {
2495 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2496
2497 __io_complete_rw(req, res, res2, 0);
2498 }
2499
io_complete_rw_iopoll(struct kiocb * kiocb,long res,long res2)2500 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2501 {
2502 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503
2504 if (kiocb->ki_flags & IOCB_WRITE)
2505 kiocb_end_write(req);
2506 if (unlikely(res != req->result)) {
2507 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2508 io_resubmit_prep(req))) {
2509 req_set_fail_links(req);
2510 req->flags |= REQ_F_DONT_REISSUE;
2511 }
2512 }
2513
2514 WRITE_ONCE(req->result, res);
2515 /* order with io_iopoll_complete() checking ->result */
2516 smp_wmb();
2517 WRITE_ONCE(req->iopoll_completed, 1);
2518 }
2519
2520 /*
2521 * After the iocb has been issued, it's safe to be found on the poll list.
2522 * Adding the kiocb to the list AFTER submission ensures that we don't
2523 * find it from a io_do_iopoll() thread before the issuer is done
2524 * accessing the kiocb cookie.
2525 */
io_iopoll_req_issued(struct io_kiocb * req,bool in_async)2526 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2527 {
2528 struct io_ring_ctx *ctx = req->ctx;
2529
2530 /*
2531 * Track whether we have multiple files in our lists. This will impact
2532 * how we do polling eventually, not spinning if we're on potentially
2533 * different devices.
2534 */
2535 if (list_empty(&ctx->iopoll_list)) {
2536 ctx->poll_multi_file = false;
2537 } else if (!ctx->poll_multi_file) {
2538 struct io_kiocb *list_req;
2539
2540 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2541 inflight_entry);
2542 if (list_req->file != req->file)
2543 ctx->poll_multi_file = true;
2544 }
2545
2546 /*
2547 * For fast devices, IO may have already completed. If it has, add
2548 * it to the front so we find it first.
2549 */
2550 if (READ_ONCE(req->iopoll_completed))
2551 list_add(&req->inflight_entry, &ctx->iopoll_list);
2552 else
2553 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2554
2555 /*
2556 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2557 * task context or in io worker task context. If current task context is
2558 * sq thread, we don't need to check whether should wake up sq thread.
2559 */
2560 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2561 wq_has_sleeper(&ctx->sq_data->wait))
2562 wake_up(&ctx->sq_data->wait);
2563 }
2564
io_state_file_put(struct io_submit_state * state)2565 static inline void io_state_file_put(struct io_submit_state *state)
2566 {
2567 if (state->file_refs) {
2568 fput_many(state->file, state->file_refs);
2569 state->file_refs = 0;
2570 }
2571 }
2572
2573 /*
2574 * Get as many references to a file as we have IOs left in this submission,
2575 * assuming most submissions are for one file, or at least that each file
2576 * has more than one submission.
2577 */
__io_file_get(struct io_submit_state * state,int fd)2578 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2579 {
2580 if (!state)
2581 return fget(fd);
2582
2583 if (state->file_refs) {
2584 if (state->fd == fd) {
2585 state->file_refs--;
2586 return state->file;
2587 }
2588 io_state_file_put(state);
2589 }
2590 state->file = fget_many(fd, state->ios_left);
2591 if (unlikely(!state->file))
2592 return NULL;
2593
2594 state->fd = fd;
2595 state->file_refs = state->ios_left - 1;
2596 return state->file;
2597 }
2598
io_bdev_nowait(struct block_device * bdev)2599 static bool io_bdev_nowait(struct block_device *bdev)
2600 {
2601 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2602 }
2603
2604 /*
2605 * If we tracked the file through the SCM inflight mechanism, we could support
2606 * any file. For now, just ensure that anything potentially problematic is done
2607 * inline.
2608 */
__io_file_supports_async(struct file * file,int rw)2609 static bool __io_file_supports_async(struct file *file, int rw)
2610 {
2611 umode_t mode = file_inode(file)->i_mode;
2612
2613 if (S_ISBLK(mode)) {
2614 if (IS_ENABLED(CONFIG_BLOCK) &&
2615 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2616 return true;
2617 return false;
2618 }
2619 if (S_ISCHR(mode) || S_ISSOCK(mode))
2620 return true;
2621 if (S_ISREG(mode)) {
2622 if (IS_ENABLED(CONFIG_BLOCK) &&
2623 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2624 file->f_op != &io_uring_fops)
2625 return true;
2626 return false;
2627 }
2628
2629 /* any ->read/write should understand O_NONBLOCK */
2630 if (file->f_flags & O_NONBLOCK)
2631 return true;
2632
2633 if (!(file->f_mode & FMODE_NOWAIT))
2634 return false;
2635
2636 if (rw == READ)
2637 return file->f_op->read_iter != NULL;
2638
2639 return file->f_op->write_iter != NULL;
2640 }
2641
io_file_supports_async(struct io_kiocb * req,int rw)2642 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2643 {
2644 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2645 return true;
2646 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2647 return true;
2648
2649 return __io_file_supports_async(req->file, rw);
2650 }
2651
io_prep_rw(struct io_kiocb * req,const struct io_uring_sqe * sqe)2652 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2653 {
2654 struct io_ring_ctx *ctx = req->ctx;
2655 struct kiocb *kiocb = &req->rw.kiocb;
2656 struct file *file = req->file;
2657 unsigned ioprio;
2658 int ret;
2659
2660 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2661 req->flags |= REQ_F_ISREG;
2662
2663 kiocb->ki_pos = READ_ONCE(sqe->off);
2664 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2665 req->flags |= REQ_F_CUR_POS;
2666 kiocb->ki_pos = file->f_pos;
2667 }
2668 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2669 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2670 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2671 if (unlikely(ret))
2672 return ret;
2673
2674 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2675 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2676 req->flags |= REQ_F_NOWAIT;
2677
2678 ioprio = READ_ONCE(sqe->ioprio);
2679 if (ioprio) {
2680 ret = ioprio_check_cap(ioprio);
2681 if (ret)
2682 return ret;
2683
2684 kiocb->ki_ioprio = ioprio;
2685 } else
2686 kiocb->ki_ioprio = get_current_ioprio();
2687
2688 if (ctx->flags & IORING_SETUP_IOPOLL) {
2689 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2690 !kiocb->ki_filp->f_op->iopoll)
2691 return -EOPNOTSUPP;
2692
2693 kiocb->ki_flags |= IOCB_HIPRI;
2694 kiocb->ki_complete = io_complete_rw_iopoll;
2695 req->iopoll_completed = 0;
2696 } else {
2697 if (kiocb->ki_flags & IOCB_HIPRI)
2698 return -EINVAL;
2699 kiocb->ki_complete = io_complete_rw;
2700 }
2701
2702 if (req->opcode == IORING_OP_READ_FIXED ||
2703 req->opcode == IORING_OP_WRITE_FIXED) {
2704 req->imu = NULL;
2705 io_req_set_rsrc_node(req);
2706 }
2707
2708 req->rw.addr = READ_ONCE(sqe->addr);
2709 req->rw.len = READ_ONCE(sqe->len);
2710 req->buf_index = READ_ONCE(sqe->buf_index);
2711 return 0;
2712 }
2713
io_rw_done(struct kiocb * kiocb,ssize_t ret)2714 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2715 {
2716 switch (ret) {
2717 case -EIOCBQUEUED:
2718 break;
2719 case -ERESTARTSYS:
2720 case -ERESTARTNOINTR:
2721 case -ERESTARTNOHAND:
2722 case -ERESTART_RESTARTBLOCK:
2723 /*
2724 * We can't just restart the syscall, since previously
2725 * submitted sqes may already be in progress. Just fail this
2726 * IO with EINTR.
2727 */
2728 ret = -EINTR;
2729 fallthrough;
2730 default:
2731 kiocb->ki_complete(kiocb, ret, 0);
2732 }
2733 }
2734
kiocb_done(struct kiocb * kiocb,ssize_t ret,unsigned int issue_flags)2735 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2736 unsigned int issue_flags)
2737 {
2738 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2739 struct io_async_rw *io = req->async_data;
2740 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2741
2742 /* add previously done IO, if any */
2743 if (io && io->bytes_done > 0) {
2744 if (ret < 0)
2745 ret = io->bytes_done;
2746 else
2747 ret += io->bytes_done;
2748 }
2749
2750 if (req->flags & REQ_F_CUR_POS)
2751 req->file->f_pos = kiocb->ki_pos;
2752 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2753 __io_complete_rw(req, ret, 0, issue_flags);
2754 else
2755 io_rw_done(kiocb, ret);
2756
2757 if (check_reissue && req->flags & REQ_F_REISSUE) {
2758 req->flags &= ~REQ_F_REISSUE;
2759 if (io_resubmit_prep(req)) {
2760 req_ref_get(req);
2761 io_queue_async_work(req);
2762 } else {
2763 int cflags = 0;
2764
2765 req_set_fail_links(req);
2766 if (req->flags & REQ_F_BUFFER_SELECTED)
2767 cflags = io_put_rw_kbuf(req);
2768 __io_req_complete(req, issue_flags, ret, cflags);
2769 }
2770 }
2771 }
2772
__io_import_fixed(struct io_kiocb * req,int rw,struct iov_iter * iter,struct io_mapped_ubuf * imu)2773 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2774 struct io_mapped_ubuf *imu)
2775 {
2776 size_t len = req->rw.len;
2777 u64 buf_end, buf_addr = req->rw.addr;
2778 size_t offset;
2779
2780 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2781 return -EFAULT;
2782 /* not inside the mapped region */
2783 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2784 return -EFAULT;
2785
2786 /*
2787 * May not be a start of buffer, set size appropriately
2788 * and advance us to the beginning.
2789 */
2790 offset = buf_addr - imu->ubuf;
2791 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2792
2793 if (offset) {
2794 /*
2795 * Don't use iov_iter_advance() here, as it's really slow for
2796 * using the latter parts of a big fixed buffer - it iterates
2797 * over each segment manually. We can cheat a bit here, because
2798 * we know that:
2799 *
2800 * 1) it's a BVEC iter, we set it up
2801 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2802 * first and last bvec
2803 *
2804 * So just find our index, and adjust the iterator afterwards.
2805 * If the offset is within the first bvec (or the whole first
2806 * bvec, just use iov_iter_advance(). This makes it easier
2807 * since we can just skip the first segment, which may not
2808 * be PAGE_SIZE aligned.
2809 */
2810 const struct bio_vec *bvec = imu->bvec;
2811
2812 if (offset <= bvec->bv_len) {
2813 iov_iter_advance(iter, offset);
2814 } else {
2815 unsigned long seg_skip;
2816
2817 /* skip first vec */
2818 offset -= bvec->bv_len;
2819 seg_skip = 1 + (offset >> PAGE_SHIFT);
2820
2821 iter->bvec = bvec + seg_skip;
2822 iter->nr_segs -= seg_skip;
2823 iter->count -= bvec->bv_len + offset;
2824 iter->iov_offset = offset & ~PAGE_MASK;
2825 }
2826 }
2827
2828 return 0;
2829 }
2830
io_import_fixed(struct io_kiocb * req,int rw,struct iov_iter * iter)2831 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2832 {
2833 struct io_ring_ctx *ctx = req->ctx;
2834 struct io_mapped_ubuf *imu = req->imu;
2835 u16 index, buf_index = req->buf_index;
2836
2837 if (likely(!imu)) {
2838 if (unlikely(buf_index >= ctx->nr_user_bufs))
2839 return -EFAULT;
2840 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2841 imu = READ_ONCE(ctx->user_bufs[index]);
2842 req->imu = imu;
2843 }
2844 return __io_import_fixed(req, rw, iter, imu);
2845 }
2846
io_ring_submit_unlock(struct io_ring_ctx * ctx,bool needs_lock)2847 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2848 {
2849 if (needs_lock)
2850 mutex_unlock(&ctx->uring_lock);
2851 }
2852
io_ring_submit_lock(struct io_ring_ctx * ctx,bool needs_lock)2853 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2854 {
2855 /*
2856 * "Normal" inline submissions always hold the uring_lock, since we
2857 * grab it from the system call. Same is true for the SQPOLL offload.
2858 * The only exception is when we've detached the request and issue it
2859 * from an async worker thread, grab the lock for that case.
2860 */
2861 if (needs_lock)
2862 mutex_lock(&ctx->uring_lock);
2863 }
2864
io_buffer_select(struct io_kiocb * req,size_t * len,int bgid,struct io_buffer * kbuf,bool needs_lock)2865 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2866 int bgid, struct io_buffer *kbuf,
2867 bool needs_lock)
2868 {
2869 struct io_buffer *head;
2870
2871 if (req->flags & REQ_F_BUFFER_SELECTED)
2872 return kbuf;
2873
2874 io_ring_submit_lock(req->ctx, needs_lock);
2875
2876 lockdep_assert_held(&req->ctx->uring_lock);
2877
2878 head = xa_load(&req->ctx->io_buffers, bgid);
2879 if (head) {
2880 if (!list_empty(&head->list)) {
2881 kbuf = list_last_entry(&head->list, struct io_buffer,
2882 list);
2883 list_del(&kbuf->list);
2884 } else {
2885 kbuf = head;
2886 xa_erase(&req->ctx->io_buffers, bgid);
2887 }
2888 if (*len > kbuf->len)
2889 *len = kbuf->len;
2890 } else {
2891 kbuf = ERR_PTR(-ENOBUFS);
2892 }
2893
2894 io_ring_submit_unlock(req->ctx, needs_lock);
2895
2896 return kbuf;
2897 }
2898
io_rw_buffer_select(struct io_kiocb * req,size_t * len,bool needs_lock)2899 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2900 bool needs_lock)
2901 {
2902 struct io_buffer *kbuf;
2903 u16 bgid;
2904
2905 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2906 bgid = req->buf_index;
2907 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2908 if (IS_ERR(kbuf))
2909 return kbuf;
2910 req->rw.addr = (u64) (unsigned long) kbuf;
2911 req->flags |= REQ_F_BUFFER_SELECTED;
2912 return u64_to_user_ptr(kbuf->addr);
2913 }
2914
2915 #ifdef CONFIG_COMPAT
io_compat_import(struct io_kiocb * req,struct iovec * iov,bool needs_lock)2916 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2917 bool needs_lock)
2918 {
2919 struct compat_iovec __user *uiov;
2920 compat_ssize_t clen;
2921 void __user *buf;
2922 ssize_t len;
2923
2924 uiov = u64_to_user_ptr(req->rw.addr);
2925 if (!access_ok(uiov, sizeof(*uiov)))
2926 return -EFAULT;
2927 if (__get_user(clen, &uiov->iov_len))
2928 return -EFAULT;
2929 if (clen < 0)
2930 return -EINVAL;
2931
2932 len = clen;
2933 buf = io_rw_buffer_select(req, &len, needs_lock);
2934 if (IS_ERR(buf))
2935 return PTR_ERR(buf);
2936 iov[0].iov_base = buf;
2937 iov[0].iov_len = (compat_size_t) len;
2938 return 0;
2939 }
2940 #endif
2941
__io_iov_buffer_select(struct io_kiocb * req,struct iovec * iov,bool needs_lock)2942 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2943 bool needs_lock)
2944 {
2945 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2946 void __user *buf;
2947 ssize_t len;
2948
2949 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2950 return -EFAULT;
2951
2952 len = iov[0].iov_len;
2953 if (len < 0)
2954 return -EINVAL;
2955 buf = io_rw_buffer_select(req, &len, needs_lock);
2956 if (IS_ERR(buf))
2957 return PTR_ERR(buf);
2958 iov[0].iov_base = buf;
2959 iov[0].iov_len = len;
2960 return 0;
2961 }
2962
io_iov_buffer_select(struct io_kiocb * req,struct iovec * iov,bool needs_lock)2963 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 bool needs_lock)
2965 {
2966 if (req->flags & REQ_F_BUFFER_SELECTED) {
2967 struct io_buffer *kbuf;
2968
2969 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2970 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2971 iov[0].iov_len = kbuf->len;
2972 return 0;
2973 }
2974 if (req->rw.len != 1)
2975 return -EINVAL;
2976
2977 #ifdef CONFIG_COMPAT
2978 if (req->ctx->compat)
2979 return io_compat_import(req, iov, needs_lock);
2980 #endif
2981
2982 return __io_iov_buffer_select(req, iov, needs_lock);
2983 }
2984
io_import_iovec(int rw,struct io_kiocb * req,struct iovec ** iovec,struct iov_iter * iter,bool needs_lock)2985 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2986 struct iov_iter *iter, bool needs_lock)
2987 {
2988 void __user *buf = u64_to_user_ptr(req->rw.addr);
2989 size_t sqe_len = req->rw.len;
2990 u8 opcode = req->opcode;
2991 ssize_t ret;
2992
2993 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2994 *iovec = NULL;
2995 return io_import_fixed(req, rw, iter);
2996 }
2997
2998 /* buffer index only valid with fixed read/write, or buffer select */
2999 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3000 return -EINVAL;
3001
3002 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3003 if (req->flags & REQ_F_BUFFER_SELECT) {
3004 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3005 if (IS_ERR(buf))
3006 return PTR_ERR(buf);
3007 req->rw.len = sqe_len;
3008 }
3009
3010 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3011 *iovec = NULL;
3012 return ret;
3013 }
3014
3015 if (req->flags & REQ_F_BUFFER_SELECT) {
3016 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3017 if (!ret)
3018 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3019 *iovec = NULL;
3020 return ret;
3021 }
3022
3023 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3024 req->ctx->compat);
3025 }
3026
io_kiocb_ppos(struct kiocb * kiocb)3027 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3028 {
3029 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3030 }
3031
3032 /*
3033 * For files that don't have ->read_iter() and ->write_iter(), handle them
3034 * by looping over ->read() or ->write() manually.
3035 */
loop_rw_iter(int rw,struct io_kiocb * req,struct iov_iter * iter)3036 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3037 {
3038 struct kiocb *kiocb = &req->rw.kiocb;
3039 struct file *file = req->file;
3040 ssize_t ret = 0;
3041
3042 /*
3043 * Don't support polled IO through this interface, and we can't
3044 * support non-blocking either. For the latter, this just causes
3045 * the kiocb to be handled from an async context.
3046 */
3047 if (kiocb->ki_flags & IOCB_HIPRI)
3048 return -EOPNOTSUPP;
3049 if (kiocb->ki_flags & IOCB_NOWAIT)
3050 return -EAGAIN;
3051
3052 while (iov_iter_count(iter)) {
3053 struct iovec iovec;
3054 ssize_t nr;
3055
3056 if (!iov_iter_is_bvec(iter)) {
3057 iovec = iov_iter_iovec(iter);
3058 } else {
3059 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3060 iovec.iov_len = req->rw.len;
3061 }
3062
3063 if (rw == READ) {
3064 nr = file->f_op->read(file, iovec.iov_base,
3065 iovec.iov_len, io_kiocb_ppos(kiocb));
3066 } else {
3067 nr = file->f_op->write(file, iovec.iov_base,
3068 iovec.iov_len, io_kiocb_ppos(kiocb));
3069 }
3070
3071 if (nr < 0) {
3072 if (!ret)
3073 ret = nr;
3074 break;
3075 }
3076 ret += nr;
3077 if (nr != iovec.iov_len)
3078 break;
3079 req->rw.len -= nr;
3080 req->rw.addr += nr;
3081 iov_iter_advance(iter, nr);
3082 }
3083
3084 return ret;
3085 }
3086
io_req_map_rw(struct io_kiocb * req,const struct iovec * iovec,const struct iovec * fast_iov,struct iov_iter * iter)3087 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3088 const struct iovec *fast_iov, struct iov_iter *iter)
3089 {
3090 struct io_async_rw *rw = req->async_data;
3091
3092 memcpy(&rw->iter, iter, sizeof(*iter));
3093 rw->free_iovec = iovec;
3094 rw->bytes_done = 0;
3095 /* can only be fixed buffers, no need to do anything */
3096 if (iov_iter_is_bvec(iter))
3097 return;
3098 if (!iovec) {
3099 unsigned iov_off = 0;
3100
3101 rw->iter.iov = rw->fast_iov;
3102 if (iter->iov != fast_iov) {
3103 iov_off = iter->iov - fast_iov;
3104 rw->iter.iov += iov_off;
3105 }
3106 if (rw->fast_iov != fast_iov)
3107 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3108 sizeof(struct iovec) * iter->nr_segs);
3109 } else {
3110 req->flags |= REQ_F_NEED_CLEANUP;
3111 }
3112 }
3113
io_alloc_async_data(struct io_kiocb * req)3114 static inline int io_alloc_async_data(struct io_kiocb *req)
3115 {
3116 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3117 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3118 return req->async_data == NULL;
3119 }
3120
io_setup_async_rw(struct io_kiocb * req,const struct iovec * iovec,const struct iovec * fast_iov,struct iov_iter * iter,bool force)3121 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3122 const struct iovec *fast_iov,
3123 struct iov_iter *iter, bool force)
3124 {
3125 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3126 return 0;
3127 if (!req->async_data) {
3128 if (io_alloc_async_data(req)) {
3129 kfree(iovec);
3130 return -ENOMEM;
3131 }
3132
3133 io_req_map_rw(req, iovec, fast_iov, iter);
3134 }
3135 return 0;
3136 }
3137
io_rw_prep_async(struct io_kiocb * req,int rw)3138 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3139 {
3140 struct io_async_rw *iorw = req->async_data;
3141 struct iovec *iov = iorw->fast_iov;
3142 int ret;
3143
3144 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3145 if (unlikely(ret < 0))
3146 return ret;
3147
3148 iorw->bytes_done = 0;
3149 iorw->free_iovec = iov;
3150 if (iov)
3151 req->flags |= REQ_F_NEED_CLEANUP;
3152 return 0;
3153 }
3154
io_read_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3155 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3156 {
3157 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3158 return -EBADF;
3159 return io_prep_rw(req, sqe);
3160 }
3161
3162 /*
3163 * This is our waitqueue callback handler, registered through lock_page_async()
3164 * when we initially tried to do the IO with the iocb armed our waitqueue.
3165 * This gets called when the page is unlocked, and we generally expect that to
3166 * happen when the page IO is completed and the page is now uptodate. This will
3167 * queue a task_work based retry of the operation, attempting to copy the data
3168 * again. If the latter fails because the page was NOT uptodate, then we will
3169 * do a thread based blocking retry of the operation. That's the unexpected
3170 * slow path.
3171 */
io_async_buf_func(struct wait_queue_entry * wait,unsigned mode,int sync,void * arg)3172 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3173 int sync, void *arg)
3174 {
3175 struct wait_page_queue *wpq;
3176 struct io_kiocb *req = wait->private;
3177 struct wait_page_key *key = arg;
3178
3179 wpq = container_of(wait, struct wait_page_queue, wait);
3180
3181 if (!wake_page_match(wpq, key))
3182 return 0;
3183
3184 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3185 list_del_init(&wait->entry);
3186
3187 /* submit ref gets dropped, acquire a new one */
3188 req_ref_get(req);
3189 io_req_task_queue(req);
3190 return 1;
3191 }
3192
3193 /*
3194 * This controls whether a given IO request should be armed for async page
3195 * based retry. If we return false here, the request is handed to the async
3196 * worker threads for retry. If we're doing buffered reads on a regular file,
3197 * we prepare a private wait_page_queue entry and retry the operation. This
3198 * will either succeed because the page is now uptodate and unlocked, or it
3199 * will register a callback when the page is unlocked at IO completion. Through
3200 * that callback, io_uring uses task_work to setup a retry of the operation.
3201 * That retry will attempt the buffered read again. The retry will generally
3202 * succeed, or in rare cases where it fails, we then fall back to using the
3203 * async worker threads for a blocking retry.
3204 */
io_rw_should_retry(struct io_kiocb * req)3205 static bool io_rw_should_retry(struct io_kiocb *req)
3206 {
3207 struct io_async_rw *rw = req->async_data;
3208 struct wait_page_queue *wait = &rw->wpq;
3209 struct kiocb *kiocb = &req->rw.kiocb;
3210
3211 /* never retry for NOWAIT, we just complete with -EAGAIN */
3212 if (req->flags & REQ_F_NOWAIT)
3213 return false;
3214
3215 /* Only for buffered IO */
3216 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3217 return false;
3218
3219 /*
3220 * just use poll if we can, and don't attempt if the fs doesn't
3221 * support callback based unlocks
3222 */
3223 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3224 return false;
3225
3226 wait->wait.func = io_async_buf_func;
3227 wait->wait.private = req;
3228 wait->wait.flags = 0;
3229 INIT_LIST_HEAD(&wait->wait.entry);
3230 kiocb->ki_flags |= IOCB_WAITQ;
3231 kiocb->ki_flags &= ~IOCB_NOWAIT;
3232 kiocb->ki_waitq = wait;
3233 return true;
3234 }
3235
io_iter_do_read(struct io_kiocb * req,struct iov_iter * iter)3236 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3237 {
3238 if (req->file->f_op->read_iter)
3239 return call_read_iter(req->file, &req->rw.kiocb, iter);
3240 else if (req->file->f_op->read)
3241 return loop_rw_iter(READ, req, iter);
3242 else
3243 return -EINVAL;
3244 }
3245
io_read(struct io_kiocb * req,unsigned int issue_flags)3246 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3247 {
3248 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3249 struct kiocb *kiocb = &req->rw.kiocb;
3250 struct iov_iter __iter, *iter = &__iter;
3251 struct io_async_rw *rw = req->async_data;
3252 ssize_t io_size, ret, ret2;
3253 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3254
3255 if (rw) {
3256 iter = &rw->iter;
3257 iovec = NULL;
3258 } else {
3259 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3260 if (ret < 0)
3261 return ret;
3262 }
3263 io_size = iov_iter_count(iter);
3264 req->result = io_size;
3265
3266 /* Ensure we clear previously set non-block flag */
3267 if (!force_nonblock)
3268 kiocb->ki_flags &= ~IOCB_NOWAIT;
3269 else
3270 kiocb->ki_flags |= IOCB_NOWAIT;
3271
3272 /* If the file doesn't support async, just async punt */
3273 if (force_nonblock && !io_file_supports_async(req, READ)) {
3274 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3275 return ret ?: -EAGAIN;
3276 }
3277
3278 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3279 if (unlikely(ret)) {
3280 kfree(iovec);
3281 return ret;
3282 }
3283
3284 ret = io_iter_do_read(req, iter);
3285
3286 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3287 req->flags &= ~REQ_F_REISSUE;
3288 /* IOPOLL retry should happen for io-wq threads */
3289 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3290 goto done;
3291 /* no retry on NONBLOCK nor RWF_NOWAIT */
3292 if (req->flags & REQ_F_NOWAIT)
3293 goto done;
3294 /* some cases will consume bytes even on error returns */
3295 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3296 ret = 0;
3297 } else if (ret == -EIOCBQUEUED) {
3298 goto out_free;
3299 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3300 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3301 /* read all, failed, already did sync or don't want to retry */
3302 goto done;
3303 }
3304
3305 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3306 if (ret2)
3307 return ret2;
3308
3309 iovec = NULL;
3310 rw = req->async_data;
3311 /* now use our persistent iterator, if we aren't already */
3312 iter = &rw->iter;
3313
3314 do {
3315 io_size -= ret;
3316 rw->bytes_done += ret;
3317 /* if we can retry, do so with the callbacks armed */
3318 if (!io_rw_should_retry(req)) {
3319 kiocb->ki_flags &= ~IOCB_WAITQ;
3320 return -EAGAIN;
3321 }
3322
3323 /*
3324 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3325 * we get -EIOCBQUEUED, then we'll get a notification when the
3326 * desired page gets unlocked. We can also get a partial read
3327 * here, and if we do, then just retry at the new offset.
3328 */
3329 ret = io_iter_do_read(req, iter);
3330 if (ret == -EIOCBQUEUED)
3331 return 0;
3332 /* we got some bytes, but not all. retry. */
3333 kiocb->ki_flags &= ~IOCB_WAITQ;
3334 } while (ret > 0 && ret < io_size);
3335 done:
3336 kiocb_done(kiocb, ret, issue_flags);
3337 out_free:
3338 /* it's faster to check here then delegate to kfree */
3339 if (iovec)
3340 kfree(iovec);
3341 return 0;
3342 }
3343
io_write_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3344 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3345 {
3346 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3347 return -EBADF;
3348 return io_prep_rw(req, sqe);
3349 }
3350
io_write(struct io_kiocb * req,unsigned int issue_flags)3351 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3352 {
3353 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3354 struct kiocb *kiocb = &req->rw.kiocb;
3355 struct iov_iter __iter, *iter = &__iter;
3356 struct io_async_rw *rw = req->async_data;
3357 ssize_t ret, ret2, io_size;
3358 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3359
3360 if (rw) {
3361 iter = &rw->iter;
3362 iovec = NULL;
3363 } else {
3364 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3365 if (ret < 0)
3366 return ret;
3367 }
3368 io_size = iov_iter_count(iter);
3369 req->result = io_size;
3370
3371 /* Ensure we clear previously set non-block flag */
3372 if (!force_nonblock)
3373 kiocb->ki_flags &= ~IOCB_NOWAIT;
3374 else
3375 kiocb->ki_flags |= IOCB_NOWAIT;
3376
3377 /* If the file doesn't support async, just async punt */
3378 if (force_nonblock && !io_file_supports_async(req, WRITE))
3379 goto copy_iov;
3380
3381 /* file path doesn't support NOWAIT for non-direct_IO */
3382 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3383 (req->flags & REQ_F_ISREG))
3384 goto copy_iov;
3385
3386 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3387 if (unlikely(ret))
3388 goto out_free;
3389
3390 /*
3391 * Open-code file_start_write here to grab freeze protection,
3392 * which will be released by another thread in
3393 * io_complete_rw(). Fool lockdep by telling it the lock got
3394 * released so that it doesn't complain about the held lock when
3395 * we return to userspace.
3396 */
3397 if (req->flags & REQ_F_ISREG) {
3398 sb_start_write(file_inode(req->file)->i_sb);
3399 __sb_writers_release(file_inode(req->file)->i_sb,
3400 SB_FREEZE_WRITE);
3401 }
3402 kiocb->ki_flags |= IOCB_WRITE;
3403
3404 if (req->file->f_op->write_iter)
3405 ret2 = call_write_iter(req->file, kiocb, iter);
3406 else if (req->file->f_op->write)
3407 ret2 = loop_rw_iter(WRITE, req, iter);
3408 else
3409 ret2 = -EINVAL;
3410
3411 if (req->flags & REQ_F_REISSUE) {
3412 req->flags &= ~REQ_F_REISSUE;
3413 ret2 = -EAGAIN;
3414 }
3415
3416 /*
3417 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3418 * retry them without IOCB_NOWAIT.
3419 */
3420 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3421 ret2 = -EAGAIN;
3422 /* no retry on NONBLOCK nor RWF_NOWAIT */
3423 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3424 goto done;
3425 if (!force_nonblock || ret2 != -EAGAIN) {
3426 /* IOPOLL retry should happen for io-wq threads */
3427 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3428 goto copy_iov;
3429 done:
3430 kiocb_done(kiocb, ret2, issue_flags);
3431 } else {
3432 copy_iov:
3433 /* some cases will consume bytes even on error returns */
3434 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3435 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3436 return ret ?: -EAGAIN;
3437 }
3438 out_free:
3439 /* it's reportedly faster than delegating the null check to kfree() */
3440 if (iovec)
3441 kfree(iovec);
3442 return ret;
3443 }
3444
io_renameat_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3445 static int io_renameat_prep(struct io_kiocb *req,
3446 const struct io_uring_sqe *sqe)
3447 {
3448 struct io_rename *ren = &req->rename;
3449 const char __user *oldf, *newf;
3450
3451 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3452 return -EBADF;
3453
3454 ren->old_dfd = READ_ONCE(sqe->fd);
3455 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3456 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3457 ren->new_dfd = READ_ONCE(sqe->len);
3458 ren->flags = READ_ONCE(sqe->rename_flags);
3459
3460 ren->oldpath = getname(oldf);
3461 if (IS_ERR(ren->oldpath))
3462 return PTR_ERR(ren->oldpath);
3463
3464 ren->newpath = getname(newf);
3465 if (IS_ERR(ren->newpath)) {
3466 putname(ren->oldpath);
3467 return PTR_ERR(ren->newpath);
3468 }
3469
3470 req->flags |= REQ_F_NEED_CLEANUP;
3471 return 0;
3472 }
3473
io_renameat(struct io_kiocb * req,unsigned int issue_flags)3474 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3475 {
3476 struct io_rename *ren = &req->rename;
3477 int ret;
3478
3479 if (issue_flags & IO_URING_F_NONBLOCK)
3480 return -EAGAIN;
3481
3482 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3483 ren->newpath, ren->flags);
3484
3485 req->flags &= ~REQ_F_NEED_CLEANUP;
3486 if (ret < 0)
3487 req_set_fail_links(req);
3488 io_req_complete(req, ret);
3489 return 0;
3490 }
3491
io_unlinkat_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3492 static int io_unlinkat_prep(struct io_kiocb *req,
3493 const struct io_uring_sqe *sqe)
3494 {
3495 struct io_unlink *un = &req->unlink;
3496 const char __user *fname;
3497
3498 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3499 return -EBADF;
3500
3501 un->dfd = READ_ONCE(sqe->fd);
3502
3503 un->flags = READ_ONCE(sqe->unlink_flags);
3504 if (un->flags & ~AT_REMOVEDIR)
3505 return -EINVAL;
3506
3507 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3508 un->filename = getname(fname);
3509 if (IS_ERR(un->filename))
3510 return PTR_ERR(un->filename);
3511
3512 req->flags |= REQ_F_NEED_CLEANUP;
3513 return 0;
3514 }
3515
io_unlinkat(struct io_kiocb * req,unsigned int issue_flags)3516 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3517 {
3518 struct io_unlink *un = &req->unlink;
3519 int ret;
3520
3521 if (issue_flags & IO_URING_F_NONBLOCK)
3522 return -EAGAIN;
3523
3524 if (un->flags & AT_REMOVEDIR)
3525 ret = do_rmdir(un->dfd, un->filename);
3526 else
3527 ret = do_unlinkat(un->dfd, un->filename);
3528
3529 req->flags &= ~REQ_F_NEED_CLEANUP;
3530 if (ret < 0)
3531 req_set_fail_links(req);
3532 io_req_complete(req, ret);
3533 return 0;
3534 }
3535
io_shutdown_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3536 static int io_shutdown_prep(struct io_kiocb *req,
3537 const struct io_uring_sqe *sqe)
3538 {
3539 #if defined(CONFIG_NET)
3540 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3541 return -EINVAL;
3542 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3543 sqe->buf_index)
3544 return -EINVAL;
3545
3546 req->shutdown.how = READ_ONCE(sqe->len);
3547 return 0;
3548 #else
3549 return -EOPNOTSUPP;
3550 #endif
3551 }
3552
io_shutdown(struct io_kiocb * req,unsigned int issue_flags)3553 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3554 {
3555 #if defined(CONFIG_NET)
3556 struct socket *sock;
3557 int ret;
3558
3559 if (issue_flags & IO_URING_F_NONBLOCK)
3560 return -EAGAIN;
3561
3562 sock = sock_from_file(req->file);
3563 if (unlikely(!sock))
3564 return -ENOTSOCK;
3565
3566 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3567 if (ret < 0)
3568 req_set_fail_links(req);
3569 io_req_complete(req, ret);
3570 return 0;
3571 #else
3572 return -EOPNOTSUPP;
3573 #endif
3574 }
3575
__io_splice_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3576 static int __io_splice_prep(struct io_kiocb *req,
3577 const struct io_uring_sqe *sqe)
3578 {
3579 struct io_splice* sp = &req->splice;
3580 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3581
3582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3583 return -EINVAL;
3584
3585 sp->file_in = NULL;
3586 sp->len = READ_ONCE(sqe->len);
3587 sp->flags = READ_ONCE(sqe->splice_flags);
3588
3589 if (unlikely(sp->flags & ~valid_flags))
3590 return -EINVAL;
3591
3592 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3593 (sp->flags & SPLICE_F_FD_IN_FIXED));
3594 if (!sp->file_in)
3595 return -EBADF;
3596 req->flags |= REQ_F_NEED_CLEANUP;
3597 return 0;
3598 }
3599
io_tee_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3600 static int io_tee_prep(struct io_kiocb *req,
3601 const struct io_uring_sqe *sqe)
3602 {
3603 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3604 return -EINVAL;
3605 return __io_splice_prep(req, sqe);
3606 }
3607
io_tee(struct io_kiocb * req,unsigned int issue_flags)3608 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3609 {
3610 struct io_splice *sp = &req->splice;
3611 struct file *in = sp->file_in;
3612 struct file *out = sp->file_out;
3613 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3614 long ret = 0;
3615
3616 if (issue_flags & IO_URING_F_NONBLOCK)
3617 return -EAGAIN;
3618 if (sp->len)
3619 ret = do_tee(in, out, sp->len, flags);
3620
3621 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3622 io_put_file(in);
3623 req->flags &= ~REQ_F_NEED_CLEANUP;
3624
3625 if (ret != sp->len)
3626 req_set_fail_links(req);
3627 io_req_complete(req, ret);
3628 return 0;
3629 }
3630
io_splice_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3631 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3632 {
3633 struct io_splice* sp = &req->splice;
3634
3635 sp->off_in = READ_ONCE(sqe->splice_off_in);
3636 sp->off_out = READ_ONCE(sqe->off);
3637 return __io_splice_prep(req, sqe);
3638 }
3639
io_splice(struct io_kiocb * req,unsigned int issue_flags)3640 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3641 {
3642 struct io_splice *sp = &req->splice;
3643 struct file *in = sp->file_in;
3644 struct file *out = sp->file_out;
3645 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3646 loff_t *poff_in, *poff_out;
3647 long ret = 0;
3648
3649 if (issue_flags & IO_URING_F_NONBLOCK)
3650 return -EAGAIN;
3651
3652 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3653 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3654
3655 if (sp->len)
3656 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3657
3658 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3659 io_put_file(in);
3660 req->flags &= ~REQ_F_NEED_CLEANUP;
3661
3662 if (ret != sp->len)
3663 req_set_fail_links(req);
3664 io_req_complete(req, ret);
3665 return 0;
3666 }
3667
3668 /*
3669 * IORING_OP_NOP just posts a completion event, nothing else.
3670 */
io_nop(struct io_kiocb * req,unsigned int issue_flags)3671 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3672 {
3673 struct io_ring_ctx *ctx = req->ctx;
3674
3675 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3676 return -EINVAL;
3677
3678 __io_req_complete(req, issue_flags, 0, 0);
3679 return 0;
3680 }
3681
io_fsync_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3682 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3683 {
3684 struct io_ring_ctx *ctx = req->ctx;
3685
3686 if (!req->file)
3687 return -EBADF;
3688
3689 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3690 return -EINVAL;
3691 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3692 return -EINVAL;
3693
3694 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3695 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3696 return -EINVAL;
3697
3698 req->sync.off = READ_ONCE(sqe->off);
3699 req->sync.len = READ_ONCE(sqe->len);
3700 return 0;
3701 }
3702
io_fsync(struct io_kiocb * req,unsigned int issue_flags)3703 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3704 {
3705 loff_t end = req->sync.off + req->sync.len;
3706 int ret;
3707
3708 /* fsync always requires a blocking context */
3709 if (issue_flags & IO_URING_F_NONBLOCK)
3710 return -EAGAIN;
3711
3712 ret = vfs_fsync_range(req->file, req->sync.off,
3713 end > 0 ? end : LLONG_MAX,
3714 req->sync.flags & IORING_FSYNC_DATASYNC);
3715 if (ret < 0)
3716 req_set_fail_links(req);
3717 io_req_complete(req, ret);
3718 return 0;
3719 }
3720
io_fallocate_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3721 static int io_fallocate_prep(struct io_kiocb *req,
3722 const struct io_uring_sqe *sqe)
3723 {
3724 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3725 return -EINVAL;
3726 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3727 return -EINVAL;
3728
3729 req->sync.off = READ_ONCE(sqe->off);
3730 req->sync.len = READ_ONCE(sqe->addr);
3731 req->sync.mode = READ_ONCE(sqe->len);
3732 return 0;
3733 }
3734
io_fallocate(struct io_kiocb * req,unsigned int issue_flags)3735 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3736 {
3737 int ret;
3738
3739 /* fallocate always requiring blocking context */
3740 if (issue_flags & IO_URING_F_NONBLOCK)
3741 return -EAGAIN;
3742 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3743 req->sync.len);
3744 if (ret < 0)
3745 req_set_fail_links(req);
3746 io_req_complete(req, ret);
3747 return 0;
3748 }
3749
__io_openat_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3750 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3751 {
3752 const char __user *fname;
3753 int ret;
3754
3755 if (unlikely(sqe->ioprio || sqe->buf_index))
3756 return -EINVAL;
3757 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3758 return -EBADF;
3759
3760 /* open.how should be already initialised */
3761 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3762 req->open.how.flags |= O_LARGEFILE;
3763
3764 req->open.dfd = READ_ONCE(sqe->fd);
3765 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3766 req->open.filename = getname(fname);
3767 if (IS_ERR(req->open.filename)) {
3768 ret = PTR_ERR(req->open.filename);
3769 req->open.filename = NULL;
3770 return ret;
3771 }
3772 req->open.nofile = rlimit(RLIMIT_NOFILE);
3773 req->flags |= REQ_F_NEED_CLEANUP;
3774 return 0;
3775 }
3776
io_openat_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3777 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3778 {
3779 u64 flags, mode;
3780
3781 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3782 return -EINVAL;
3783 mode = READ_ONCE(sqe->len);
3784 flags = READ_ONCE(sqe->open_flags);
3785 req->open.how = build_open_how(flags, mode);
3786 return __io_openat_prep(req, sqe);
3787 }
3788
io_openat2_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3789 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3790 {
3791 struct open_how __user *how;
3792 size_t len;
3793 int ret;
3794
3795 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3796 return -EINVAL;
3797 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3798 len = READ_ONCE(sqe->len);
3799 if (len < OPEN_HOW_SIZE_VER0)
3800 return -EINVAL;
3801
3802 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3803 len);
3804 if (ret)
3805 return ret;
3806
3807 return __io_openat_prep(req, sqe);
3808 }
3809
io_openat2(struct io_kiocb * req,unsigned int issue_flags)3810 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3811 {
3812 struct open_flags op;
3813 struct file *file;
3814 bool nonblock_set;
3815 bool resolve_nonblock;
3816 int ret;
3817
3818 ret = build_open_flags(&req->open.how, &op);
3819 if (ret)
3820 goto err;
3821 nonblock_set = op.open_flag & O_NONBLOCK;
3822 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3823 if (issue_flags & IO_URING_F_NONBLOCK) {
3824 /*
3825 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3826 * it'll always -EAGAIN
3827 */
3828 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3829 return -EAGAIN;
3830 op.lookup_flags |= LOOKUP_CACHED;
3831 op.open_flag |= O_NONBLOCK;
3832 }
3833
3834 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3835 if (ret < 0)
3836 goto err;
3837
3838 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3839 /* only retry if RESOLVE_CACHED wasn't already set by application */
3840 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3841 file == ERR_PTR(-EAGAIN)) {
3842 /*
3843 * We could hang on to this 'fd', but seems like marginal
3844 * gain for something that is now known to be a slower path.
3845 * So just put it, and we'll get a new one when we retry.
3846 */
3847 put_unused_fd(ret);
3848 return -EAGAIN;
3849 }
3850
3851 if (IS_ERR(file)) {
3852 put_unused_fd(ret);
3853 ret = PTR_ERR(file);
3854 } else {
3855 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3856 file->f_flags &= ~O_NONBLOCK;
3857 fsnotify_open(file);
3858 fd_install(ret, file);
3859 }
3860 err:
3861 putname(req->open.filename);
3862 req->flags &= ~REQ_F_NEED_CLEANUP;
3863 if (ret < 0)
3864 req_set_fail_links(req);
3865 __io_req_complete(req, issue_flags, ret, 0);
3866 return 0;
3867 }
3868
io_openat(struct io_kiocb * req,unsigned int issue_flags)3869 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3870 {
3871 return io_openat2(req, issue_flags);
3872 }
3873
io_remove_buffers_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3874 static int io_remove_buffers_prep(struct io_kiocb *req,
3875 const struct io_uring_sqe *sqe)
3876 {
3877 struct io_provide_buf *p = &req->pbuf;
3878 u64 tmp;
3879
3880 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3881 return -EINVAL;
3882
3883 tmp = READ_ONCE(sqe->fd);
3884 if (!tmp || tmp > USHRT_MAX)
3885 return -EINVAL;
3886
3887 memset(p, 0, sizeof(*p));
3888 p->nbufs = tmp;
3889 p->bgid = READ_ONCE(sqe->buf_group);
3890 return 0;
3891 }
3892
__io_remove_buffers(struct io_ring_ctx * ctx,struct io_buffer * buf,int bgid,unsigned nbufs)3893 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3894 int bgid, unsigned nbufs)
3895 {
3896 unsigned i = 0;
3897
3898 /* shouldn't happen */
3899 if (!nbufs)
3900 return 0;
3901
3902 /* the head kbuf is the list itself */
3903 while (!list_empty(&buf->list)) {
3904 struct io_buffer *nxt;
3905
3906 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3907 list_del(&nxt->list);
3908 kfree(nxt);
3909 if (++i == nbufs)
3910 return i;
3911 }
3912 i++;
3913 kfree(buf);
3914 xa_erase(&ctx->io_buffers, bgid);
3915
3916 return i;
3917 }
3918
io_remove_buffers(struct io_kiocb * req,unsigned int issue_flags)3919 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3920 {
3921 struct io_provide_buf *p = &req->pbuf;
3922 struct io_ring_ctx *ctx = req->ctx;
3923 struct io_buffer *head;
3924 int ret = 0;
3925 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3926
3927 io_ring_submit_lock(ctx, !force_nonblock);
3928
3929 lockdep_assert_held(&ctx->uring_lock);
3930
3931 ret = -ENOENT;
3932 head = xa_load(&ctx->io_buffers, p->bgid);
3933 if (head)
3934 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3935 if (ret < 0)
3936 req_set_fail_links(req);
3937
3938 /* complete before unlock, IOPOLL may need the lock */
3939 __io_req_complete(req, issue_flags, ret, 0);
3940 io_ring_submit_unlock(ctx, !force_nonblock);
3941 return 0;
3942 }
3943
io_provide_buffers_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)3944 static int io_provide_buffers_prep(struct io_kiocb *req,
3945 const struct io_uring_sqe *sqe)
3946 {
3947 unsigned long size, tmp_check;
3948 struct io_provide_buf *p = &req->pbuf;
3949 u64 tmp;
3950
3951 if (sqe->ioprio || sqe->rw_flags)
3952 return -EINVAL;
3953
3954 tmp = READ_ONCE(sqe->fd);
3955 if (!tmp || tmp > USHRT_MAX)
3956 return -E2BIG;
3957 p->nbufs = tmp;
3958 p->addr = READ_ONCE(sqe->addr);
3959 p->len = READ_ONCE(sqe->len);
3960
3961 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3962 &size))
3963 return -EOVERFLOW;
3964 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3965 return -EOVERFLOW;
3966
3967 size = (unsigned long)p->len * p->nbufs;
3968 if (!access_ok(u64_to_user_ptr(p->addr), size))
3969 return -EFAULT;
3970
3971 p->bgid = READ_ONCE(sqe->buf_group);
3972 tmp = READ_ONCE(sqe->off);
3973 if (tmp > USHRT_MAX)
3974 return -E2BIG;
3975 p->bid = tmp;
3976 return 0;
3977 }
3978
io_add_buffers(struct io_provide_buf * pbuf,struct io_buffer ** head)3979 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3980 {
3981 struct io_buffer *buf;
3982 u64 addr = pbuf->addr;
3983 int i, bid = pbuf->bid;
3984
3985 for (i = 0; i < pbuf->nbufs; i++) {
3986 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3987 if (!buf)
3988 break;
3989
3990 buf->addr = addr;
3991 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
3992 buf->bid = bid;
3993 addr += pbuf->len;
3994 bid++;
3995 if (!*head) {
3996 INIT_LIST_HEAD(&buf->list);
3997 *head = buf;
3998 } else {
3999 list_add_tail(&buf->list, &(*head)->list);
4000 }
4001 }
4002
4003 return i ? i : -ENOMEM;
4004 }
4005
io_provide_buffers(struct io_kiocb * req,unsigned int issue_flags)4006 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4007 {
4008 struct io_provide_buf *p = &req->pbuf;
4009 struct io_ring_ctx *ctx = req->ctx;
4010 struct io_buffer *head, *list;
4011 int ret = 0;
4012 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4013
4014 io_ring_submit_lock(ctx, !force_nonblock);
4015
4016 lockdep_assert_held(&ctx->uring_lock);
4017
4018 list = head = xa_load(&ctx->io_buffers, p->bgid);
4019
4020 ret = io_add_buffers(p, &head);
4021 if (ret >= 0 && !list) {
4022 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4023 if (ret < 0)
4024 __io_remove_buffers(ctx, head, p->bgid, -1U);
4025 }
4026 if (ret < 0)
4027 req_set_fail_links(req);
4028 /* complete before unlock, IOPOLL may need the lock */
4029 __io_req_complete(req, issue_flags, ret, 0);
4030 io_ring_submit_unlock(ctx, !force_nonblock);
4031 return 0;
4032 }
4033
io_epoll_ctl_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4034 static int io_epoll_ctl_prep(struct io_kiocb *req,
4035 const struct io_uring_sqe *sqe)
4036 {
4037 #if defined(CONFIG_EPOLL)
4038 if (sqe->ioprio || sqe->buf_index)
4039 return -EINVAL;
4040 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4041 return -EINVAL;
4042
4043 req->epoll.epfd = READ_ONCE(sqe->fd);
4044 req->epoll.op = READ_ONCE(sqe->len);
4045 req->epoll.fd = READ_ONCE(sqe->off);
4046
4047 if (ep_op_has_event(req->epoll.op)) {
4048 struct epoll_event __user *ev;
4049
4050 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4051 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4052 return -EFAULT;
4053 }
4054
4055 return 0;
4056 #else
4057 return -EOPNOTSUPP;
4058 #endif
4059 }
4060
io_epoll_ctl(struct io_kiocb * req,unsigned int issue_flags)4061 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4062 {
4063 #if defined(CONFIG_EPOLL)
4064 struct io_epoll *ie = &req->epoll;
4065 int ret;
4066 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4067
4068 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4069 if (force_nonblock && ret == -EAGAIN)
4070 return -EAGAIN;
4071
4072 if (ret < 0)
4073 req_set_fail_links(req);
4074 __io_req_complete(req, issue_flags, ret, 0);
4075 return 0;
4076 #else
4077 return -EOPNOTSUPP;
4078 #endif
4079 }
4080
io_madvise_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4081 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4082 {
4083 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4084 if (sqe->ioprio || sqe->buf_index || sqe->off)
4085 return -EINVAL;
4086 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4087 return -EINVAL;
4088
4089 req->madvise.addr = READ_ONCE(sqe->addr);
4090 req->madvise.len = READ_ONCE(sqe->len);
4091 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4092 return 0;
4093 #else
4094 return -EOPNOTSUPP;
4095 #endif
4096 }
4097
io_madvise(struct io_kiocb * req,unsigned int issue_flags)4098 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4099 {
4100 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4101 struct io_madvise *ma = &req->madvise;
4102 int ret;
4103
4104 if (issue_flags & IO_URING_F_NONBLOCK)
4105 return -EAGAIN;
4106
4107 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4108 if (ret < 0)
4109 req_set_fail_links(req);
4110 io_req_complete(req, ret);
4111 return 0;
4112 #else
4113 return -EOPNOTSUPP;
4114 #endif
4115 }
4116
io_fadvise_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4117 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4118 {
4119 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4120 return -EINVAL;
4121 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4122 return -EINVAL;
4123
4124 req->fadvise.offset = READ_ONCE(sqe->off);
4125 req->fadvise.len = READ_ONCE(sqe->len);
4126 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4127 return 0;
4128 }
4129
io_fadvise(struct io_kiocb * req,unsigned int issue_flags)4130 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4131 {
4132 struct io_fadvise *fa = &req->fadvise;
4133 int ret;
4134
4135 if (issue_flags & IO_URING_F_NONBLOCK) {
4136 switch (fa->advice) {
4137 case POSIX_FADV_NORMAL:
4138 case POSIX_FADV_RANDOM:
4139 case POSIX_FADV_SEQUENTIAL:
4140 break;
4141 default:
4142 return -EAGAIN;
4143 }
4144 }
4145
4146 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4147 if (ret < 0)
4148 req_set_fail_links(req);
4149 __io_req_complete(req, issue_flags, ret, 0);
4150 return 0;
4151 }
4152
io_statx_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4153 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4154 {
4155 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 return -EINVAL;
4157 if (sqe->ioprio || sqe->buf_index)
4158 return -EINVAL;
4159 if (req->flags & REQ_F_FIXED_FILE)
4160 return -EBADF;
4161
4162 req->statx.dfd = READ_ONCE(sqe->fd);
4163 req->statx.mask = READ_ONCE(sqe->len);
4164 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4165 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4166 req->statx.flags = READ_ONCE(sqe->statx_flags);
4167
4168 return 0;
4169 }
4170
io_statx(struct io_kiocb * req,unsigned int issue_flags)4171 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4172 {
4173 struct io_statx *ctx = &req->statx;
4174 int ret;
4175
4176 if (issue_flags & IO_URING_F_NONBLOCK)
4177 return -EAGAIN;
4178
4179 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4180 ctx->buffer);
4181
4182 if (ret < 0)
4183 req_set_fail_links(req);
4184 io_req_complete(req, ret);
4185 return 0;
4186 }
4187
io_close_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4188 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4189 {
4190 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4191 return -EINVAL;
4192 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4193 sqe->rw_flags || sqe->buf_index)
4194 return -EINVAL;
4195 if (req->flags & REQ_F_FIXED_FILE)
4196 return -EBADF;
4197
4198 req->close.fd = READ_ONCE(sqe->fd);
4199 return 0;
4200 }
4201
io_close(struct io_kiocb * req,unsigned int issue_flags)4202 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4203 {
4204 struct files_struct *files = current->files;
4205 struct io_close *close = &req->close;
4206 struct fdtable *fdt;
4207 struct file *file = NULL;
4208 int ret = -EBADF;
4209
4210 spin_lock(&files->file_lock);
4211 fdt = files_fdtable(files);
4212 if (close->fd >= fdt->max_fds) {
4213 spin_unlock(&files->file_lock);
4214 goto err;
4215 }
4216 file = fdt->fd[close->fd];
4217 if (!file || file->f_op == &io_uring_fops) {
4218 spin_unlock(&files->file_lock);
4219 file = NULL;
4220 goto err;
4221 }
4222
4223 /* if the file has a flush method, be safe and punt to async */
4224 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4225 spin_unlock(&files->file_lock);
4226 return -EAGAIN;
4227 }
4228
4229 ret = __close_fd_get_file(close->fd, &file);
4230 spin_unlock(&files->file_lock);
4231 if (ret < 0) {
4232 if (ret == -ENOENT)
4233 ret = -EBADF;
4234 goto err;
4235 }
4236
4237 /* No ->flush() or already async, safely close from here */
4238 ret = filp_close(file, current->files);
4239 err:
4240 if (ret < 0)
4241 req_set_fail_links(req);
4242 if (file)
4243 fput(file);
4244 __io_req_complete(req, issue_flags, ret, 0);
4245 return 0;
4246 }
4247
io_sfr_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4248 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4249 {
4250 struct io_ring_ctx *ctx = req->ctx;
4251
4252 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4253 return -EINVAL;
4254 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4255 return -EINVAL;
4256
4257 req->sync.off = READ_ONCE(sqe->off);
4258 req->sync.len = READ_ONCE(sqe->len);
4259 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4260 return 0;
4261 }
4262
io_sync_file_range(struct io_kiocb * req,unsigned int issue_flags)4263 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4264 {
4265 int ret;
4266
4267 /* sync_file_range always requires a blocking context */
4268 if (issue_flags & IO_URING_F_NONBLOCK)
4269 return -EAGAIN;
4270
4271 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4272 req->sync.flags);
4273 if (ret < 0)
4274 req_set_fail_links(req);
4275 io_req_complete(req, ret);
4276 return 0;
4277 }
4278
4279 #if defined(CONFIG_NET)
io_setup_async_msg(struct io_kiocb * req,struct io_async_msghdr * kmsg)4280 static int io_setup_async_msg(struct io_kiocb *req,
4281 struct io_async_msghdr *kmsg)
4282 {
4283 struct io_async_msghdr *async_msg = req->async_data;
4284
4285 if (async_msg)
4286 return -EAGAIN;
4287 if (io_alloc_async_data(req)) {
4288 kfree(kmsg->free_iov);
4289 return -ENOMEM;
4290 }
4291 async_msg = req->async_data;
4292 req->flags |= REQ_F_NEED_CLEANUP;
4293 memcpy(async_msg, kmsg, sizeof(*kmsg));
4294 async_msg->msg.msg_name = &async_msg->addr;
4295 /* if were using fast_iov, set it to the new one */
4296 if (!async_msg->free_iov)
4297 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4298
4299 return -EAGAIN;
4300 }
4301
io_sendmsg_copy_hdr(struct io_kiocb * req,struct io_async_msghdr * iomsg)4302 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4303 struct io_async_msghdr *iomsg)
4304 {
4305 iomsg->msg.msg_name = &iomsg->addr;
4306 iomsg->free_iov = iomsg->fast_iov;
4307 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4308 req->sr_msg.msg_flags, &iomsg->free_iov);
4309 }
4310
io_sendmsg_prep_async(struct io_kiocb * req)4311 static int io_sendmsg_prep_async(struct io_kiocb *req)
4312 {
4313 int ret;
4314
4315 ret = io_sendmsg_copy_hdr(req, req->async_data);
4316 if (!ret)
4317 req->flags |= REQ_F_NEED_CLEANUP;
4318 return ret;
4319 }
4320
io_sendmsg_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4321 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4322 {
4323 struct io_sr_msg *sr = &req->sr_msg;
4324
4325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4326 return -EINVAL;
4327
4328 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4329 sr->len = READ_ONCE(sqe->len);
4330 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4331 if (sr->msg_flags & MSG_DONTWAIT)
4332 req->flags |= REQ_F_NOWAIT;
4333
4334 #ifdef CONFIG_COMPAT
4335 if (req->ctx->compat)
4336 sr->msg_flags |= MSG_CMSG_COMPAT;
4337 #endif
4338 return 0;
4339 }
4340
io_sendmsg(struct io_kiocb * req,unsigned int issue_flags)4341 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4342 {
4343 struct io_async_msghdr iomsg, *kmsg;
4344 struct socket *sock;
4345 unsigned flags;
4346 int min_ret = 0;
4347 int ret;
4348
4349 sock = sock_from_file(req->file);
4350 if (unlikely(!sock))
4351 return -ENOTSOCK;
4352
4353 kmsg = req->async_data;
4354 if (!kmsg) {
4355 ret = io_sendmsg_copy_hdr(req, &iomsg);
4356 if (ret)
4357 return ret;
4358 kmsg = &iomsg;
4359 }
4360
4361 flags = req->sr_msg.msg_flags;
4362 if (issue_flags & IO_URING_F_NONBLOCK)
4363 flags |= MSG_DONTWAIT;
4364 if (flags & MSG_WAITALL)
4365 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4366
4367 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4368 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4369 return io_setup_async_msg(req, kmsg);
4370 if (ret == -ERESTARTSYS)
4371 ret = -EINTR;
4372
4373 /* fast path, check for non-NULL to avoid function call */
4374 if (kmsg->free_iov)
4375 kfree(kmsg->free_iov);
4376 req->flags &= ~REQ_F_NEED_CLEANUP;
4377 if (ret < min_ret)
4378 req_set_fail_links(req);
4379 __io_req_complete(req, issue_flags, ret, 0);
4380 return 0;
4381 }
4382
io_send(struct io_kiocb * req,unsigned int issue_flags)4383 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4384 {
4385 struct io_sr_msg *sr = &req->sr_msg;
4386 struct msghdr msg;
4387 struct iovec iov;
4388 struct socket *sock;
4389 unsigned flags;
4390 int min_ret = 0;
4391 int ret;
4392
4393 sock = sock_from_file(req->file);
4394 if (unlikely(!sock))
4395 return -ENOTSOCK;
4396
4397 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4398 if (unlikely(ret))
4399 return ret;
4400
4401 msg.msg_name = NULL;
4402 msg.msg_control = NULL;
4403 msg.msg_controllen = 0;
4404 msg.msg_namelen = 0;
4405
4406 flags = req->sr_msg.msg_flags;
4407 if (issue_flags & IO_URING_F_NONBLOCK)
4408 flags |= MSG_DONTWAIT;
4409 if (flags & MSG_WAITALL)
4410 min_ret = iov_iter_count(&msg.msg_iter);
4411
4412 msg.msg_flags = flags;
4413 ret = sock_sendmsg(sock, &msg);
4414 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4415 return -EAGAIN;
4416 if (ret == -ERESTARTSYS)
4417 ret = -EINTR;
4418
4419 if (ret < min_ret)
4420 req_set_fail_links(req);
4421 __io_req_complete(req, issue_flags, ret, 0);
4422 return 0;
4423 }
4424
__io_recvmsg_copy_hdr(struct io_kiocb * req,struct io_async_msghdr * iomsg)4425 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4426 struct io_async_msghdr *iomsg)
4427 {
4428 struct io_sr_msg *sr = &req->sr_msg;
4429 struct iovec __user *uiov;
4430 size_t iov_len;
4431 int ret;
4432
4433 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4434 &iomsg->uaddr, &uiov, &iov_len);
4435 if (ret)
4436 return ret;
4437
4438 if (req->flags & REQ_F_BUFFER_SELECT) {
4439 if (iov_len > 1)
4440 return -EINVAL;
4441 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4442 return -EFAULT;
4443 sr->len = iomsg->fast_iov[0].iov_len;
4444 iomsg->free_iov = NULL;
4445 } else {
4446 iomsg->free_iov = iomsg->fast_iov;
4447 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4448 &iomsg->free_iov, &iomsg->msg.msg_iter,
4449 false);
4450 if (ret > 0)
4451 ret = 0;
4452 }
4453
4454 return ret;
4455 }
4456
4457 #ifdef CONFIG_COMPAT
__io_compat_recvmsg_copy_hdr(struct io_kiocb * req,struct io_async_msghdr * iomsg)4458 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4459 struct io_async_msghdr *iomsg)
4460 {
4461 struct io_sr_msg *sr = &req->sr_msg;
4462 struct compat_iovec __user *uiov;
4463 compat_uptr_t ptr;
4464 compat_size_t len;
4465 int ret;
4466
4467 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4468 &ptr, &len);
4469 if (ret)
4470 return ret;
4471
4472 uiov = compat_ptr(ptr);
4473 if (req->flags & REQ_F_BUFFER_SELECT) {
4474 compat_ssize_t clen;
4475
4476 if (len > 1)
4477 return -EINVAL;
4478 if (!access_ok(uiov, sizeof(*uiov)))
4479 return -EFAULT;
4480 if (__get_user(clen, &uiov->iov_len))
4481 return -EFAULT;
4482 if (clen < 0)
4483 return -EINVAL;
4484 sr->len = clen;
4485 iomsg->free_iov = NULL;
4486 } else {
4487 iomsg->free_iov = iomsg->fast_iov;
4488 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4489 UIO_FASTIOV, &iomsg->free_iov,
4490 &iomsg->msg.msg_iter, true);
4491 if (ret < 0)
4492 return ret;
4493 }
4494
4495 return 0;
4496 }
4497 #endif
4498
io_recvmsg_copy_hdr(struct io_kiocb * req,struct io_async_msghdr * iomsg)4499 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4500 struct io_async_msghdr *iomsg)
4501 {
4502 iomsg->msg.msg_name = &iomsg->addr;
4503
4504 #ifdef CONFIG_COMPAT
4505 if (req->ctx->compat)
4506 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4507 #endif
4508
4509 return __io_recvmsg_copy_hdr(req, iomsg);
4510 }
4511
io_recv_buffer_select(struct io_kiocb * req,bool needs_lock)4512 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4513 bool needs_lock)
4514 {
4515 struct io_sr_msg *sr = &req->sr_msg;
4516 struct io_buffer *kbuf;
4517
4518 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4519 if (IS_ERR(kbuf))
4520 return kbuf;
4521
4522 sr->kbuf = kbuf;
4523 req->flags |= REQ_F_BUFFER_SELECTED;
4524 return kbuf;
4525 }
4526
io_put_recv_kbuf(struct io_kiocb * req)4527 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4528 {
4529 return io_put_kbuf(req, req->sr_msg.kbuf);
4530 }
4531
io_recvmsg_prep_async(struct io_kiocb * req)4532 static int io_recvmsg_prep_async(struct io_kiocb *req)
4533 {
4534 int ret;
4535
4536 ret = io_recvmsg_copy_hdr(req, req->async_data);
4537 if (!ret)
4538 req->flags |= REQ_F_NEED_CLEANUP;
4539 return ret;
4540 }
4541
io_recvmsg_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4542 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4543 {
4544 struct io_sr_msg *sr = &req->sr_msg;
4545
4546 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4547 return -EINVAL;
4548
4549 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4550 sr->len = READ_ONCE(sqe->len);
4551 sr->bgid = READ_ONCE(sqe->buf_group);
4552 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4553 if (sr->msg_flags & MSG_DONTWAIT)
4554 req->flags |= REQ_F_NOWAIT;
4555
4556 #ifdef CONFIG_COMPAT
4557 if (req->ctx->compat)
4558 sr->msg_flags |= MSG_CMSG_COMPAT;
4559 #endif
4560 return 0;
4561 }
4562
io_recvmsg(struct io_kiocb * req,unsigned int issue_flags)4563 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4564 {
4565 struct io_async_msghdr iomsg, *kmsg;
4566 struct socket *sock;
4567 struct io_buffer *kbuf;
4568 unsigned flags;
4569 int min_ret = 0;
4570 int ret, cflags = 0;
4571 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4572
4573 sock = sock_from_file(req->file);
4574 if (unlikely(!sock))
4575 return -ENOTSOCK;
4576
4577 kmsg = req->async_data;
4578 if (!kmsg) {
4579 ret = io_recvmsg_copy_hdr(req, &iomsg);
4580 if (ret)
4581 return ret;
4582 kmsg = &iomsg;
4583 }
4584
4585 if (req->flags & REQ_F_BUFFER_SELECT) {
4586 kbuf = io_recv_buffer_select(req, !force_nonblock);
4587 if (IS_ERR(kbuf))
4588 return PTR_ERR(kbuf);
4589 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4590 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4591 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4592 1, req->sr_msg.len);
4593 }
4594
4595 flags = req->sr_msg.msg_flags;
4596 if (force_nonblock)
4597 flags |= MSG_DONTWAIT;
4598 if (flags & MSG_WAITALL)
4599 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4600
4601 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4602 kmsg->uaddr, flags);
4603 if (force_nonblock && ret == -EAGAIN)
4604 return io_setup_async_msg(req, kmsg);
4605 if (ret == -ERESTARTSYS)
4606 ret = -EINTR;
4607
4608 if (req->flags & REQ_F_BUFFER_SELECTED)
4609 cflags = io_put_recv_kbuf(req);
4610 /* fast path, check for non-NULL to avoid function call */
4611 if (kmsg->free_iov)
4612 kfree(kmsg->free_iov);
4613 req->flags &= ~REQ_F_NEED_CLEANUP;
4614 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4615 req_set_fail_links(req);
4616 __io_req_complete(req, issue_flags, ret, cflags);
4617 return 0;
4618 }
4619
io_recv(struct io_kiocb * req,unsigned int issue_flags)4620 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4621 {
4622 struct io_buffer *kbuf;
4623 struct io_sr_msg *sr = &req->sr_msg;
4624 struct msghdr msg;
4625 void __user *buf = sr->buf;
4626 struct socket *sock;
4627 struct iovec iov;
4628 unsigned flags;
4629 int min_ret = 0;
4630 int ret, cflags = 0;
4631 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4632
4633 sock = sock_from_file(req->file);
4634 if (unlikely(!sock))
4635 return -ENOTSOCK;
4636
4637 if (req->flags & REQ_F_BUFFER_SELECT) {
4638 kbuf = io_recv_buffer_select(req, !force_nonblock);
4639 if (IS_ERR(kbuf))
4640 return PTR_ERR(kbuf);
4641 buf = u64_to_user_ptr(kbuf->addr);
4642 }
4643
4644 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4645 if (unlikely(ret))
4646 goto out_free;
4647
4648 msg.msg_name = NULL;
4649 msg.msg_control = NULL;
4650 msg.msg_controllen = 0;
4651 msg.msg_namelen = 0;
4652 msg.msg_iocb = NULL;
4653 msg.msg_flags = 0;
4654
4655 flags = req->sr_msg.msg_flags;
4656 if (force_nonblock)
4657 flags |= MSG_DONTWAIT;
4658 if (flags & MSG_WAITALL)
4659 min_ret = iov_iter_count(&msg.msg_iter);
4660
4661 ret = sock_recvmsg(sock, &msg, flags);
4662 if (force_nonblock && ret == -EAGAIN)
4663 return -EAGAIN;
4664 if (ret == -ERESTARTSYS)
4665 ret = -EINTR;
4666 out_free:
4667 if (req->flags & REQ_F_BUFFER_SELECTED)
4668 cflags = io_put_recv_kbuf(req);
4669 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4670 req_set_fail_links(req);
4671 __io_req_complete(req, issue_flags, ret, cflags);
4672 return 0;
4673 }
4674
io_accept_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4675 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4676 {
4677 struct io_accept *accept = &req->accept;
4678
4679 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4680 return -EINVAL;
4681 if (sqe->ioprio || sqe->len || sqe->buf_index)
4682 return -EINVAL;
4683
4684 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4685 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4686 accept->flags = READ_ONCE(sqe->accept_flags);
4687 accept->nofile = rlimit(RLIMIT_NOFILE);
4688 return 0;
4689 }
4690
io_accept(struct io_kiocb * req,unsigned int issue_flags)4691 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4692 {
4693 struct io_accept *accept = &req->accept;
4694 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4695 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4696 int ret;
4697
4698 if (req->file->f_flags & O_NONBLOCK)
4699 req->flags |= REQ_F_NOWAIT;
4700
4701 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4702 accept->addr_len, accept->flags,
4703 accept->nofile);
4704 if (ret == -EAGAIN && force_nonblock)
4705 return -EAGAIN;
4706 if (ret < 0) {
4707 if (ret == -ERESTARTSYS)
4708 ret = -EINTR;
4709 req_set_fail_links(req);
4710 }
4711 __io_req_complete(req, issue_flags, ret, 0);
4712 return 0;
4713 }
4714
io_connect_prep_async(struct io_kiocb * req)4715 static int io_connect_prep_async(struct io_kiocb *req)
4716 {
4717 struct io_async_connect *io = req->async_data;
4718 struct io_connect *conn = &req->connect;
4719
4720 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4721 }
4722
io_connect_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)4723 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4724 {
4725 struct io_connect *conn = &req->connect;
4726
4727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4728 return -EINVAL;
4729 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4730 return -EINVAL;
4731
4732 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4733 conn->addr_len = READ_ONCE(sqe->addr2);
4734 return 0;
4735 }
4736
io_connect(struct io_kiocb * req,unsigned int issue_flags)4737 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4738 {
4739 struct io_async_connect __io, *io;
4740 unsigned file_flags;
4741 int ret;
4742 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4743
4744 if (req->async_data) {
4745 io = req->async_data;
4746 } else {
4747 ret = move_addr_to_kernel(req->connect.addr,
4748 req->connect.addr_len,
4749 &__io.address);
4750 if (ret)
4751 goto out;
4752 io = &__io;
4753 }
4754
4755 file_flags = force_nonblock ? O_NONBLOCK : 0;
4756
4757 ret = __sys_connect_file(req->file, &io->address,
4758 req->connect.addr_len, file_flags);
4759 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4760 if (req->async_data)
4761 return -EAGAIN;
4762 if (io_alloc_async_data(req)) {
4763 ret = -ENOMEM;
4764 goto out;
4765 }
4766 memcpy(req->async_data, &__io, sizeof(__io));
4767 return -EAGAIN;
4768 }
4769 if (ret == -ERESTARTSYS)
4770 ret = -EINTR;
4771 out:
4772 if (ret < 0)
4773 req_set_fail_links(req);
4774 __io_req_complete(req, issue_flags, ret, 0);
4775 return 0;
4776 }
4777 #else /* !CONFIG_NET */
4778 #define IO_NETOP_FN(op) \
4779 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4780 { \
4781 return -EOPNOTSUPP; \
4782 }
4783
4784 #define IO_NETOP_PREP(op) \
4785 IO_NETOP_FN(op) \
4786 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4787 { \
4788 return -EOPNOTSUPP; \
4789 } \
4790
4791 #define IO_NETOP_PREP_ASYNC(op) \
4792 IO_NETOP_PREP(op) \
4793 static int io_##op##_prep_async(struct io_kiocb *req) \
4794 { \
4795 return -EOPNOTSUPP; \
4796 }
4797
4798 IO_NETOP_PREP_ASYNC(sendmsg);
4799 IO_NETOP_PREP_ASYNC(recvmsg);
4800 IO_NETOP_PREP_ASYNC(connect);
4801 IO_NETOP_PREP(accept);
4802 IO_NETOP_FN(send);
4803 IO_NETOP_FN(recv);
4804 #endif /* CONFIG_NET */
4805
4806 struct io_poll_table {
4807 struct poll_table_struct pt;
4808 struct io_kiocb *req;
4809 int error;
4810 };
4811
__io_async_wake(struct io_kiocb * req,struct io_poll_iocb * poll,__poll_t mask,task_work_func_t func)4812 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4813 __poll_t mask, task_work_func_t func)
4814 {
4815 int ret;
4816
4817 /* for instances that support it check for an event match first: */
4818 if (mask && !(mask & poll->events))
4819 return 0;
4820
4821 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4822
4823 list_del_init(&poll->wait.entry);
4824
4825 req->result = mask;
4826 req->task_work.func = func;
4827
4828 /*
4829 * If this fails, then the task is exiting. When a task exits, the
4830 * work gets canceled, so just cancel this request as well instead
4831 * of executing it. We can't safely execute it anyway, as we may not
4832 * have the needed state needed for it anyway.
4833 */
4834 ret = io_req_task_work_add(req);
4835 if (unlikely(ret)) {
4836 WRITE_ONCE(poll->canceled, true);
4837 io_req_task_work_add_fallback(req, func);
4838 }
4839 return 1;
4840 }
4841
io_poll_rewait(struct io_kiocb * req,struct io_poll_iocb * poll)4842 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4843 __acquires(&req->ctx->completion_lock)
4844 {
4845 struct io_ring_ctx *ctx = req->ctx;
4846
4847 if (!req->result && !READ_ONCE(poll->canceled)) {
4848 struct poll_table_struct pt = { ._key = poll->events };
4849
4850 req->result = vfs_poll(req->file, &pt) & poll->events;
4851 }
4852
4853 spin_lock_irq(&ctx->completion_lock);
4854 if (!req->result && !READ_ONCE(poll->canceled)) {
4855 add_wait_queue(poll->head, &poll->wait);
4856 return true;
4857 }
4858
4859 return false;
4860 }
4861
io_poll_get_double(struct io_kiocb * req)4862 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4863 {
4864 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4865 if (req->opcode == IORING_OP_POLL_ADD)
4866 return req->async_data;
4867 return req->apoll->double_poll;
4868 }
4869
io_poll_get_single(struct io_kiocb * req)4870 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4871 {
4872 if (req->opcode == IORING_OP_POLL_ADD)
4873 return &req->poll;
4874 return &req->apoll->poll;
4875 }
4876
io_poll_remove_double(struct io_kiocb * req)4877 static void io_poll_remove_double(struct io_kiocb *req)
4878 __must_hold(&req->ctx->completion_lock)
4879 {
4880 struct io_poll_iocb *poll = io_poll_get_double(req);
4881
4882 lockdep_assert_held(&req->ctx->completion_lock);
4883
4884 if (poll && poll->head) {
4885 struct wait_queue_head *head = poll->head;
4886
4887 spin_lock(&head->lock);
4888 list_del_init(&poll->wait.entry);
4889 if (poll->wait.private)
4890 req_ref_put(req);
4891 poll->head = NULL;
4892 spin_unlock(&head->lock);
4893 }
4894 }
4895
io_poll_complete(struct io_kiocb * req,__poll_t mask)4896 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4897 __must_hold(&req->ctx->completion_lock)
4898 {
4899 struct io_ring_ctx *ctx = req->ctx;
4900 unsigned flags = IORING_CQE_F_MORE;
4901 int error;
4902
4903 if (READ_ONCE(req->poll.canceled)) {
4904 error = -ECANCELED;
4905 req->poll.events |= EPOLLONESHOT;
4906 } else {
4907 error = mangle_poll(mask);
4908 }
4909 if (req->poll.events & EPOLLONESHOT)
4910 flags = 0;
4911 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4912 io_poll_remove_waitqs(req);
4913 req->poll.done = true;
4914 flags = 0;
4915 }
4916 if (flags & IORING_CQE_F_MORE)
4917 ctx->cq_extra++;
4918
4919 io_commit_cqring(ctx);
4920 return !(flags & IORING_CQE_F_MORE);
4921 }
4922
io_poll_task_func(struct callback_head * cb)4923 static void io_poll_task_func(struct callback_head *cb)
4924 {
4925 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4926 struct io_ring_ctx *ctx = req->ctx;
4927 struct io_kiocb *nxt;
4928
4929 if (io_poll_rewait(req, &req->poll)) {
4930 spin_unlock_irq(&ctx->completion_lock);
4931 } else {
4932 bool done;
4933
4934 done = io_poll_complete(req, req->result);
4935 if (done) {
4936 hash_del(&req->hash_node);
4937 } else {
4938 req->result = 0;
4939 add_wait_queue(req->poll.head, &req->poll.wait);
4940 }
4941 spin_unlock_irq(&ctx->completion_lock);
4942 io_cqring_ev_posted(ctx);
4943
4944 if (done) {
4945 nxt = io_put_req_find_next(req);
4946 if (nxt)
4947 __io_req_task_submit(nxt);
4948 }
4949 }
4950 }
4951
io_poll_double_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)4952 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4953 int sync, void *key)
4954 {
4955 struct io_kiocb *req = wait->private;
4956 struct io_poll_iocb *poll = io_poll_get_single(req);
4957 __poll_t mask = key_to_poll(key);
4958
4959 /* for instances that support it check for an event match first: */
4960 if (mask && !(mask & poll->events))
4961 return 0;
4962 if (!(poll->events & EPOLLONESHOT))
4963 return poll->wait.func(&poll->wait, mode, sync, key);
4964
4965 list_del_init(&wait->entry);
4966
4967 if (poll && poll->head) {
4968 bool done;
4969
4970 spin_lock(&poll->head->lock);
4971 done = list_empty(&poll->wait.entry);
4972 if (!done)
4973 list_del_init(&poll->wait.entry);
4974 /* make sure double remove sees this as being gone */
4975 wait->private = NULL;
4976 spin_unlock(&poll->head->lock);
4977 if (!done) {
4978 /* use wait func handler, so it matches the rq type */
4979 poll->wait.func(&poll->wait, mode, sync, key);
4980 }
4981 }
4982 req_ref_put(req);
4983 return 1;
4984 }
4985
io_init_poll_iocb(struct io_poll_iocb * poll,__poll_t events,wait_queue_func_t wake_func)4986 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4987 wait_queue_func_t wake_func)
4988 {
4989 poll->head = NULL;
4990 poll->done = false;
4991 poll->canceled = false;
4992 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4993 /* mask in events that we always want/need */
4994 poll->events = events | IO_POLL_UNMASK;
4995 INIT_LIST_HEAD(&poll->wait.entry);
4996 init_waitqueue_func_entry(&poll->wait, wake_func);
4997 }
4998
__io_queue_proc(struct io_poll_iocb * poll,struct io_poll_table * pt,struct wait_queue_head * head,struct io_poll_iocb ** poll_ptr)4999 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5000 struct wait_queue_head *head,
5001 struct io_poll_iocb **poll_ptr)
5002 {
5003 struct io_kiocb *req = pt->req;
5004
5005 /*
5006 * If poll->head is already set, it's because the file being polled
5007 * uses multiple waitqueues for poll handling (eg one for read, one
5008 * for write). Setup a separate io_poll_iocb if this happens.
5009 */
5010 if (unlikely(poll->head)) {
5011 struct io_poll_iocb *poll_one = poll;
5012
5013 /* already have a 2nd entry, fail a third attempt */
5014 if (*poll_ptr) {
5015 pt->error = -EINVAL;
5016 return;
5017 }
5018 /*
5019 * Can't handle multishot for double wait for now, turn it
5020 * into one-shot mode.
5021 */
5022 if (!(req->poll.events & EPOLLONESHOT))
5023 req->poll.events |= EPOLLONESHOT;
5024 /* double add on the same waitqueue head, ignore */
5025 if (poll->head == head)
5026 return;
5027 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5028 if (!poll) {
5029 pt->error = -ENOMEM;
5030 return;
5031 }
5032 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5033 req_ref_get(req);
5034 poll->wait.private = req;
5035 *poll_ptr = poll;
5036 }
5037
5038 pt->error = 0;
5039 poll->head = head;
5040
5041 if (poll->events & EPOLLEXCLUSIVE)
5042 add_wait_queue_exclusive(head, &poll->wait);
5043 else
5044 add_wait_queue(head, &poll->wait);
5045 }
5046
io_async_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)5047 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5048 struct poll_table_struct *p)
5049 {
5050 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5051 struct async_poll *apoll = pt->req->apoll;
5052
5053 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5054 }
5055
io_async_task_func(struct callback_head * cb)5056 static void io_async_task_func(struct callback_head *cb)
5057 {
5058 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5059 struct async_poll *apoll = req->apoll;
5060 struct io_ring_ctx *ctx = req->ctx;
5061
5062 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5063
5064 if (io_poll_rewait(req, &apoll->poll)) {
5065 spin_unlock_irq(&ctx->completion_lock);
5066 return;
5067 }
5068
5069 hash_del(&req->hash_node);
5070 io_poll_remove_double(req);
5071 spin_unlock_irq(&ctx->completion_lock);
5072
5073 if (!READ_ONCE(apoll->poll.canceled))
5074 __io_req_task_submit(req);
5075 else
5076 io_req_complete_failed(req, -ECANCELED);
5077 }
5078
io_async_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)5079 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5080 void *key)
5081 {
5082 struct io_kiocb *req = wait->private;
5083 struct io_poll_iocb *poll = &req->apoll->poll;
5084
5085 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5086 key_to_poll(key));
5087
5088 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5089 }
5090
io_poll_req_insert(struct io_kiocb * req)5091 static void io_poll_req_insert(struct io_kiocb *req)
5092 {
5093 struct io_ring_ctx *ctx = req->ctx;
5094 struct hlist_head *list;
5095
5096 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5097 hlist_add_head(&req->hash_node, list);
5098 }
5099
__io_arm_poll_handler(struct io_kiocb * req,struct io_poll_iocb * poll,struct io_poll_table * ipt,__poll_t mask,wait_queue_func_t wake_func)5100 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5101 struct io_poll_iocb *poll,
5102 struct io_poll_table *ipt, __poll_t mask,
5103 wait_queue_func_t wake_func)
5104 __acquires(&ctx->completion_lock)
5105 {
5106 struct io_ring_ctx *ctx = req->ctx;
5107 bool cancel = false;
5108
5109 INIT_HLIST_NODE(&req->hash_node);
5110 io_init_poll_iocb(poll, mask, wake_func);
5111 poll->file = req->file;
5112 poll->wait.private = req;
5113
5114 ipt->pt._key = mask;
5115 ipt->req = req;
5116 ipt->error = -EINVAL;
5117
5118 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5119
5120 spin_lock_irq(&ctx->completion_lock);
5121 if (likely(poll->head)) {
5122 spin_lock(&poll->head->lock);
5123 if (unlikely(list_empty(&poll->wait.entry))) {
5124 if (ipt->error)
5125 cancel = true;
5126 ipt->error = 0;
5127 mask = 0;
5128 }
5129 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5130 list_del_init(&poll->wait.entry);
5131 else if (cancel)
5132 WRITE_ONCE(poll->canceled, true);
5133 else if (!poll->done) /* actually waiting for an event */
5134 io_poll_req_insert(req);
5135 spin_unlock(&poll->head->lock);
5136 }
5137
5138 return mask;
5139 }
5140
io_arm_poll_handler(struct io_kiocb * req)5141 static bool io_arm_poll_handler(struct io_kiocb *req)
5142 {
5143 const struct io_op_def *def = &io_op_defs[req->opcode];
5144 struct io_ring_ctx *ctx = req->ctx;
5145 struct async_poll *apoll;
5146 struct io_poll_table ipt;
5147 __poll_t mask, ret;
5148 int rw;
5149
5150 if (!req->file || !file_can_poll(req->file))
5151 return false;
5152 if (req->flags & REQ_F_POLLED)
5153 return false;
5154 if (def->pollin)
5155 rw = READ;
5156 else if (def->pollout)
5157 rw = WRITE;
5158 else
5159 return false;
5160 /* if we can't nonblock try, then no point in arming a poll handler */
5161 if (!io_file_supports_async(req, rw))
5162 return false;
5163
5164 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5165 if (unlikely(!apoll))
5166 return false;
5167 apoll->double_poll = NULL;
5168
5169 req->flags |= REQ_F_POLLED;
5170 req->apoll = apoll;
5171
5172 mask = EPOLLONESHOT;
5173 if (def->pollin)
5174 mask |= POLLIN | POLLRDNORM;
5175 if (def->pollout)
5176 mask |= POLLOUT | POLLWRNORM;
5177
5178 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5179 if ((req->opcode == IORING_OP_RECVMSG) &&
5180 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5181 mask &= ~POLLIN;
5182
5183 mask |= POLLERR | POLLPRI;
5184
5185 ipt.pt._qproc = io_async_queue_proc;
5186
5187 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5188 io_async_wake);
5189 if (ret || ipt.error) {
5190 io_poll_remove_double(req);
5191 spin_unlock_irq(&ctx->completion_lock);
5192 return false;
5193 }
5194 spin_unlock_irq(&ctx->completion_lock);
5195 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5196 apoll->poll.events);
5197 return true;
5198 }
5199
__io_poll_remove_one(struct io_kiocb * req,struct io_poll_iocb * poll,bool do_cancel)5200 static bool __io_poll_remove_one(struct io_kiocb *req,
5201 struct io_poll_iocb *poll, bool do_cancel)
5202 __must_hold(&req->ctx->completion_lock)
5203 {
5204 bool do_complete = false;
5205
5206 if (!poll->head)
5207 return false;
5208 spin_lock(&poll->head->lock);
5209 if (do_cancel)
5210 WRITE_ONCE(poll->canceled, true);
5211 if (!list_empty(&poll->wait.entry)) {
5212 list_del_init(&poll->wait.entry);
5213 do_complete = true;
5214 }
5215 spin_unlock(&poll->head->lock);
5216 hash_del(&req->hash_node);
5217 return do_complete;
5218 }
5219
io_poll_remove_waitqs(struct io_kiocb * req)5220 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5221 __must_hold(&req->ctx->completion_lock)
5222 {
5223 bool do_complete;
5224
5225 io_poll_remove_double(req);
5226 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5227
5228 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5229 /* non-poll requests have submit ref still */
5230 req_ref_put(req);
5231 }
5232 return do_complete;
5233 }
5234
io_poll_remove_one(struct io_kiocb * req)5235 static bool io_poll_remove_one(struct io_kiocb *req)
5236 __must_hold(&req->ctx->completion_lock)
5237 {
5238 bool do_complete;
5239
5240 do_complete = io_poll_remove_waitqs(req);
5241 if (do_complete) {
5242 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5243 io_commit_cqring(req->ctx);
5244 req_set_fail_links(req);
5245 io_put_req_deferred(req, 1);
5246 }
5247
5248 return do_complete;
5249 }
5250
5251 /*
5252 * Returns true if we found and killed one or more poll requests
5253 */
io_poll_remove_all(struct io_ring_ctx * ctx,struct task_struct * tsk,struct files_struct * files)5254 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5255 struct files_struct *files)
5256 {
5257 struct hlist_node *tmp;
5258 struct io_kiocb *req;
5259 int posted = 0, i;
5260
5261 spin_lock_irq(&ctx->completion_lock);
5262 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5263 struct hlist_head *list;
5264
5265 list = &ctx->cancel_hash[i];
5266 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5267 if (io_match_task(req, tsk, files))
5268 posted += io_poll_remove_one(req);
5269 }
5270 }
5271 spin_unlock_irq(&ctx->completion_lock);
5272
5273 if (posted)
5274 io_cqring_ev_posted(ctx);
5275
5276 return posted != 0;
5277 }
5278
io_poll_find(struct io_ring_ctx * ctx,__u64 sqe_addr,bool poll_only)5279 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5280 bool poll_only)
5281 __must_hold(&ctx->completion_lock)
5282 {
5283 struct hlist_head *list;
5284 struct io_kiocb *req;
5285
5286 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5287 hlist_for_each_entry(req, list, hash_node) {
5288 if (sqe_addr != req->user_data)
5289 continue;
5290 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5291 continue;
5292 return req;
5293 }
5294 return NULL;
5295 }
5296
io_poll_cancel(struct io_ring_ctx * ctx,__u64 sqe_addr,bool poll_only)5297 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5298 bool poll_only)
5299 __must_hold(&ctx->completion_lock)
5300 {
5301 struct io_kiocb *req;
5302
5303 req = io_poll_find(ctx, sqe_addr, poll_only);
5304 if (!req)
5305 return -ENOENT;
5306 if (io_poll_remove_one(req))
5307 return 0;
5308
5309 return -EALREADY;
5310 }
5311
io_poll_parse_events(const struct io_uring_sqe * sqe,unsigned int flags)5312 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5313 unsigned int flags)
5314 {
5315 u32 events;
5316
5317 events = READ_ONCE(sqe->poll32_events);
5318 #ifdef __BIG_ENDIAN
5319 events = swahw32(events);
5320 #endif
5321 if (!(flags & IORING_POLL_ADD_MULTI))
5322 events |= EPOLLONESHOT;
5323 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5324 }
5325
io_poll_update_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5326 static int io_poll_update_prep(struct io_kiocb *req,
5327 const struct io_uring_sqe *sqe)
5328 {
5329 struct io_poll_update *upd = &req->poll_update;
5330 u32 flags;
5331
5332 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5333 return -EINVAL;
5334 if (sqe->ioprio || sqe->buf_index)
5335 return -EINVAL;
5336 flags = READ_ONCE(sqe->len);
5337 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5338 IORING_POLL_ADD_MULTI))
5339 return -EINVAL;
5340 /* meaningless without update */
5341 if (flags == IORING_POLL_ADD_MULTI)
5342 return -EINVAL;
5343
5344 upd->old_user_data = READ_ONCE(sqe->addr);
5345 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5346 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5347
5348 upd->new_user_data = READ_ONCE(sqe->off);
5349 if (!upd->update_user_data && upd->new_user_data)
5350 return -EINVAL;
5351 if (upd->update_events)
5352 upd->events = io_poll_parse_events(sqe, flags);
5353 else if (sqe->poll32_events)
5354 return -EINVAL;
5355
5356 return 0;
5357 }
5358
io_poll_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)5359 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5360 void *key)
5361 {
5362 struct io_kiocb *req = wait->private;
5363 struct io_poll_iocb *poll = &req->poll;
5364
5365 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5366 }
5367
io_poll_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)5368 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5369 struct poll_table_struct *p)
5370 {
5371 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5372
5373 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5374 }
5375
io_poll_add_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5376 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5377 {
5378 struct io_poll_iocb *poll = &req->poll;
5379 u32 flags;
5380
5381 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5382 return -EINVAL;
5383 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5384 return -EINVAL;
5385 flags = READ_ONCE(sqe->len);
5386 if (flags & ~IORING_POLL_ADD_MULTI)
5387 return -EINVAL;
5388
5389 poll->events = io_poll_parse_events(sqe, flags);
5390 return 0;
5391 }
5392
io_poll_add(struct io_kiocb * req,unsigned int issue_flags)5393 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5394 {
5395 struct io_poll_iocb *poll = &req->poll;
5396 struct io_ring_ctx *ctx = req->ctx;
5397 struct io_poll_table ipt;
5398 __poll_t mask;
5399
5400 ipt.pt._qproc = io_poll_queue_proc;
5401
5402 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5403 io_poll_wake);
5404
5405 if (mask) { /* no async, we'd stolen it */
5406 ipt.error = 0;
5407 io_poll_complete(req, mask);
5408 }
5409 spin_unlock_irq(&ctx->completion_lock);
5410
5411 if (mask) {
5412 io_cqring_ev_posted(ctx);
5413 if (poll->events & EPOLLONESHOT)
5414 io_put_req(req);
5415 }
5416 return ipt.error;
5417 }
5418
io_poll_update(struct io_kiocb * req,unsigned int issue_flags)5419 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5420 {
5421 struct io_ring_ctx *ctx = req->ctx;
5422 struct io_kiocb *preq;
5423 bool completing;
5424 int ret;
5425
5426 spin_lock_irq(&ctx->completion_lock);
5427 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5428 if (!preq) {
5429 ret = -ENOENT;
5430 goto err;
5431 }
5432
5433 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5434 completing = true;
5435 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5436 goto err;
5437 }
5438
5439 /*
5440 * Don't allow racy completion with singleshot, as we cannot safely
5441 * update those. For multishot, if we're racing with completion, just
5442 * let completion re-add it.
5443 */
5444 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5445 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5446 ret = -EALREADY;
5447 goto err;
5448 }
5449 /* we now have a detached poll request. reissue. */
5450 ret = 0;
5451 err:
5452 if (ret < 0) {
5453 spin_unlock_irq(&ctx->completion_lock);
5454 req_set_fail_links(req);
5455 io_req_complete(req, ret);
5456 return 0;
5457 }
5458 /* only mask one event flags, keep behavior flags */
5459 if (req->poll_update.update_events) {
5460 preq->poll.events &= ~0xffff;
5461 preq->poll.events |= req->poll_update.events & 0xffff;
5462 preq->poll.events |= IO_POLL_UNMASK;
5463 }
5464 if (req->poll_update.update_user_data)
5465 preq->user_data = req->poll_update.new_user_data;
5466 spin_unlock_irq(&ctx->completion_lock);
5467
5468 /* complete update request, we're done with it */
5469 io_req_complete(req, ret);
5470
5471 if (!completing) {
5472 ret = io_poll_add(preq, issue_flags);
5473 if (ret < 0) {
5474 req_set_fail_links(preq);
5475 io_req_complete(preq, ret);
5476 }
5477 }
5478 return 0;
5479 }
5480
io_timeout_fn(struct hrtimer * timer)5481 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5482 {
5483 struct io_timeout_data *data = container_of(timer,
5484 struct io_timeout_data, timer);
5485 struct io_kiocb *req = data->req;
5486 struct io_ring_ctx *ctx = req->ctx;
5487 unsigned long flags;
5488
5489 spin_lock_irqsave(&ctx->completion_lock, flags);
5490 list_del_init(&req->timeout.list);
5491 atomic_set(&req->ctx->cq_timeouts,
5492 atomic_read(&req->ctx->cq_timeouts) + 1);
5493
5494 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5495 io_commit_cqring(ctx);
5496 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5497
5498 io_cqring_ev_posted(ctx);
5499 req_set_fail_links(req);
5500 io_put_req(req);
5501 return HRTIMER_NORESTART;
5502 }
5503
io_timeout_extract(struct io_ring_ctx * ctx,__u64 user_data)5504 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5505 __u64 user_data)
5506 __must_hold(&ctx->completion_lock)
5507 {
5508 struct io_timeout_data *io;
5509 struct io_kiocb *req;
5510 bool found = false;
5511
5512 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5513 found = user_data == req->user_data;
5514 if (found)
5515 break;
5516 }
5517 if (!found)
5518 return ERR_PTR(-ENOENT);
5519
5520 io = req->async_data;
5521 if (hrtimer_try_to_cancel(&io->timer) == -1)
5522 return ERR_PTR(-EALREADY);
5523 list_del_init(&req->timeout.list);
5524 return req;
5525 }
5526
io_timeout_cancel(struct io_ring_ctx * ctx,__u64 user_data)5527 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5528 __must_hold(&ctx->completion_lock)
5529 {
5530 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5531
5532 if (IS_ERR(req))
5533 return PTR_ERR(req);
5534
5535 req_set_fail_links(req);
5536 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5537 io_put_req_deferred(req, 1);
5538 return 0;
5539 }
5540
io_timeout_update(struct io_ring_ctx * ctx,__u64 user_data,struct timespec64 * ts,enum hrtimer_mode mode)5541 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5542 struct timespec64 *ts, enum hrtimer_mode mode)
5543 __must_hold(&ctx->completion_lock)
5544 {
5545 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5546 struct io_timeout_data *data;
5547
5548 if (IS_ERR(req))
5549 return PTR_ERR(req);
5550
5551 req->timeout.off = 0; /* noseq */
5552 data = req->async_data;
5553 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5554 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5555 data->timer.function = io_timeout_fn;
5556 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5557 return 0;
5558 }
5559
io_timeout_remove_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5560 static int io_timeout_remove_prep(struct io_kiocb *req,
5561 const struct io_uring_sqe *sqe)
5562 {
5563 struct io_timeout_rem *tr = &req->timeout_rem;
5564
5565 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5566 return -EINVAL;
5567 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5568 return -EINVAL;
5569 if (sqe->ioprio || sqe->buf_index || sqe->len)
5570 return -EINVAL;
5571
5572 tr->addr = READ_ONCE(sqe->addr);
5573 tr->flags = READ_ONCE(sqe->timeout_flags);
5574 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5575 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5576 return -EINVAL;
5577 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5578 return -EFAULT;
5579 } else if (tr->flags) {
5580 /* timeout removal doesn't support flags */
5581 return -EINVAL;
5582 }
5583
5584 return 0;
5585 }
5586
io_translate_timeout_mode(unsigned int flags)5587 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5588 {
5589 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5590 : HRTIMER_MODE_REL;
5591 }
5592
5593 /*
5594 * Remove or update an existing timeout command
5595 */
io_timeout_remove(struct io_kiocb * req,unsigned int issue_flags)5596 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5597 {
5598 struct io_timeout_rem *tr = &req->timeout_rem;
5599 struct io_ring_ctx *ctx = req->ctx;
5600 int ret;
5601
5602 spin_lock_irq(&ctx->completion_lock);
5603 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5604 ret = io_timeout_cancel(ctx, tr->addr);
5605 else
5606 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5607 io_translate_timeout_mode(tr->flags));
5608
5609 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5610 io_commit_cqring(ctx);
5611 spin_unlock_irq(&ctx->completion_lock);
5612 io_cqring_ev_posted(ctx);
5613 if (ret < 0)
5614 req_set_fail_links(req);
5615 io_put_req(req);
5616 return 0;
5617 }
5618
io_timeout_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe,bool is_timeout_link)5619 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5620 bool is_timeout_link)
5621 {
5622 struct io_timeout_data *data;
5623 unsigned flags;
5624 u32 off = READ_ONCE(sqe->off);
5625
5626 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5627 return -EINVAL;
5628 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5629 return -EINVAL;
5630 if (off && is_timeout_link)
5631 return -EINVAL;
5632 flags = READ_ONCE(sqe->timeout_flags);
5633 if (flags & ~IORING_TIMEOUT_ABS)
5634 return -EINVAL;
5635
5636 req->timeout.off = off;
5637
5638 if (!req->async_data && io_alloc_async_data(req))
5639 return -ENOMEM;
5640
5641 data = req->async_data;
5642 data->req = req;
5643
5644 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5645 return -EFAULT;
5646
5647 data->mode = io_translate_timeout_mode(flags);
5648 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5649 if (is_timeout_link)
5650 io_req_track_inflight(req);
5651 return 0;
5652 }
5653
io_timeout(struct io_kiocb * req,unsigned int issue_flags)5654 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5655 {
5656 struct io_ring_ctx *ctx = req->ctx;
5657 struct io_timeout_data *data = req->async_data;
5658 struct list_head *entry;
5659 u32 tail, off = req->timeout.off;
5660
5661 spin_lock_irq(&ctx->completion_lock);
5662
5663 /*
5664 * sqe->off holds how many events that need to occur for this
5665 * timeout event to be satisfied. If it isn't set, then this is
5666 * a pure timeout request, sequence isn't used.
5667 */
5668 if (io_is_timeout_noseq(req)) {
5669 entry = ctx->timeout_list.prev;
5670 goto add;
5671 }
5672
5673 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5674 req->timeout.target_seq = tail + off;
5675
5676 /* Update the last seq here in case io_flush_timeouts() hasn't.
5677 * This is safe because ->completion_lock is held, and submissions
5678 * and completions are never mixed in the same ->completion_lock section.
5679 */
5680 ctx->cq_last_tm_flush = tail;
5681
5682 /*
5683 * Insertion sort, ensuring the first entry in the list is always
5684 * the one we need first.
5685 */
5686 list_for_each_prev(entry, &ctx->timeout_list) {
5687 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5688 timeout.list);
5689
5690 if (io_is_timeout_noseq(nxt))
5691 continue;
5692 /* nxt.seq is behind @tail, otherwise would've been completed */
5693 if (off >= nxt->timeout.target_seq - tail)
5694 break;
5695 }
5696 add:
5697 list_add(&req->timeout.list, entry);
5698 data->timer.function = io_timeout_fn;
5699 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5700 spin_unlock_irq(&ctx->completion_lock);
5701 return 0;
5702 }
5703
5704 struct io_cancel_data {
5705 struct io_ring_ctx *ctx;
5706 u64 user_data;
5707 };
5708
io_cancel_cb(struct io_wq_work * work,void * data)5709 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5710 {
5711 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5712 struct io_cancel_data *cd = data;
5713
5714 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5715 }
5716
io_async_cancel_one(struct io_uring_task * tctx,u64 user_data,struct io_ring_ctx * ctx)5717 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5718 struct io_ring_ctx *ctx)
5719 {
5720 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5721 enum io_wq_cancel cancel_ret;
5722 int ret = 0;
5723
5724 if (!tctx || !tctx->io_wq)
5725 return -ENOENT;
5726
5727 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5728 switch (cancel_ret) {
5729 case IO_WQ_CANCEL_OK:
5730 ret = 0;
5731 break;
5732 case IO_WQ_CANCEL_RUNNING:
5733 ret = -EALREADY;
5734 break;
5735 case IO_WQ_CANCEL_NOTFOUND:
5736 ret = -ENOENT;
5737 break;
5738 }
5739
5740 return ret;
5741 }
5742
io_async_find_and_cancel(struct io_ring_ctx * ctx,struct io_kiocb * req,__u64 sqe_addr,int success_ret)5743 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5744 struct io_kiocb *req, __u64 sqe_addr,
5745 int success_ret)
5746 {
5747 unsigned long flags;
5748 int ret;
5749
5750 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5751 spin_lock_irqsave(&ctx->completion_lock, flags);
5752 if (ret != -ENOENT)
5753 goto done;
5754 ret = io_timeout_cancel(ctx, sqe_addr);
5755 if (ret != -ENOENT)
5756 goto done;
5757 ret = io_poll_cancel(ctx, sqe_addr, false);
5758 done:
5759 if (!ret)
5760 ret = success_ret;
5761 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5762 io_commit_cqring(ctx);
5763 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5764 io_cqring_ev_posted(ctx);
5765
5766 if (ret < 0)
5767 req_set_fail_links(req);
5768 }
5769
io_async_cancel_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5770 static int io_async_cancel_prep(struct io_kiocb *req,
5771 const struct io_uring_sqe *sqe)
5772 {
5773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5774 return -EINVAL;
5775 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5776 return -EINVAL;
5777 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5778 return -EINVAL;
5779
5780 req->cancel.addr = READ_ONCE(sqe->addr);
5781 return 0;
5782 }
5783
io_async_cancel(struct io_kiocb * req,unsigned int issue_flags)5784 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5785 {
5786 struct io_ring_ctx *ctx = req->ctx;
5787 u64 sqe_addr = req->cancel.addr;
5788 struct io_tctx_node *node;
5789 int ret;
5790
5791 /* tasks should wait for their io-wq threads, so safe w/o sync */
5792 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5793 spin_lock_irq(&ctx->completion_lock);
5794 if (ret != -ENOENT)
5795 goto done;
5796 ret = io_timeout_cancel(ctx, sqe_addr);
5797 if (ret != -ENOENT)
5798 goto done;
5799 ret = io_poll_cancel(ctx, sqe_addr, false);
5800 if (ret != -ENOENT)
5801 goto done;
5802 spin_unlock_irq(&ctx->completion_lock);
5803
5804 /* slow path, try all io-wq's */
5805 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5806 ret = -ENOENT;
5807 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5808 struct io_uring_task *tctx = node->task->io_uring;
5809
5810 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5811 if (ret != -ENOENT)
5812 break;
5813 }
5814 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5815
5816 spin_lock_irq(&ctx->completion_lock);
5817 done:
5818 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5819 io_commit_cqring(ctx);
5820 spin_unlock_irq(&ctx->completion_lock);
5821 io_cqring_ev_posted(ctx);
5822
5823 if (ret < 0)
5824 req_set_fail_links(req);
5825 io_put_req(req);
5826 return 0;
5827 }
5828
io_rsrc_update_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5829 static int io_rsrc_update_prep(struct io_kiocb *req,
5830 const struct io_uring_sqe *sqe)
5831 {
5832 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5833 return -EINVAL;
5834 if (sqe->ioprio || sqe->rw_flags)
5835 return -EINVAL;
5836
5837 req->rsrc_update.offset = READ_ONCE(sqe->off);
5838 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5839 if (!req->rsrc_update.nr_args)
5840 return -EINVAL;
5841 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5842 return 0;
5843 }
5844
io_files_update(struct io_kiocb * req,unsigned int issue_flags)5845 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5846 {
5847 struct io_ring_ctx *ctx = req->ctx;
5848 struct io_uring_rsrc_update2 up;
5849 int ret;
5850
5851 if (issue_flags & IO_URING_F_NONBLOCK)
5852 return -EAGAIN;
5853
5854 up.offset = req->rsrc_update.offset;
5855 up.data = req->rsrc_update.arg;
5856 up.nr = 0;
5857 up.tags = 0;
5858 up.resv = 0;
5859
5860 mutex_lock(&ctx->uring_lock);
5861 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5862 &up, req->rsrc_update.nr_args);
5863 mutex_unlock(&ctx->uring_lock);
5864
5865 if (ret < 0)
5866 req_set_fail_links(req);
5867 __io_req_complete(req, issue_flags, ret, 0);
5868 return 0;
5869 }
5870
io_req_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)5871 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5872 {
5873 switch (req->opcode) {
5874 case IORING_OP_NOP:
5875 return 0;
5876 case IORING_OP_READV:
5877 case IORING_OP_READ_FIXED:
5878 case IORING_OP_READ:
5879 return io_read_prep(req, sqe);
5880 case IORING_OP_WRITEV:
5881 case IORING_OP_WRITE_FIXED:
5882 case IORING_OP_WRITE:
5883 return io_write_prep(req, sqe);
5884 case IORING_OP_POLL_ADD:
5885 return io_poll_add_prep(req, sqe);
5886 case IORING_OP_POLL_REMOVE:
5887 return io_poll_update_prep(req, sqe);
5888 case IORING_OP_FSYNC:
5889 return io_fsync_prep(req, sqe);
5890 case IORING_OP_SYNC_FILE_RANGE:
5891 return io_sfr_prep(req, sqe);
5892 case IORING_OP_SENDMSG:
5893 case IORING_OP_SEND:
5894 return io_sendmsg_prep(req, sqe);
5895 case IORING_OP_RECVMSG:
5896 case IORING_OP_RECV:
5897 return io_recvmsg_prep(req, sqe);
5898 case IORING_OP_CONNECT:
5899 return io_connect_prep(req, sqe);
5900 case IORING_OP_TIMEOUT:
5901 return io_timeout_prep(req, sqe, false);
5902 case IORING_OP_TIMEOUT_REMOVE:
5903 return io_timeout_remove_prep(req, sqe);
5904 case IORING_OP_ASYNC_CANCEL:
5905 return io_async_cancel_prep(req, sqe);
5906 case IORING_OP_LINK_TIMEOUT:
5907 return io_timeout_prep(req, sqe, true);
5908 case IORING_OP_ACCEPT:
5909 return io_accept_prep(req, sqe);
5910 case IORING_OP_FALLOCATE:
5911 return io_fallocate_prep(req, sqe);
5912 case IORING_OP_OPENAT:
5913 return io_openat_prep(req, sqe);
5914 case IORING_OP_CLOSE:
5915 return io_close_prep(req, sqe);
5916 case IORING_OP_FILES_UPDATE:
5917 return io_rsrc_update_prep(req, sqe);
5918 case IORING_OP_STATX:
5919 return io_statx_prep(req, sqe);
5920 case IORING_OP_FADVISE:
5921 return io_fadvise_prep(req, sqe);
5922 case IORING_OP_MADVISE:
5923 return io_madvise_prep(req, sqe);
5924 case IORING_OP_OPENAT2:
5925 return io_openat2_prep(req, sqe);
5926 case IORING_OP_EPOLL_CTL:
5927 return io_epoll_ctl_prep(req, sqe);
5928 case IORING_OP_SPLICE:
5929 return io_splice_prep(req, sqe);
5930 case IORING_OP_PROVIDE_BUFFERS:
5931 return io_provide_buffers_prep(req, sqe);
5932 case IORING_OP_REMOVE_BUFFERS:
5933 return io_remove_buffers_prep(req, sqe);
5934 case IORING_OP_TEE:
5935 return io_tee_prep(req, sqe);
5936 case IORING_OP_SHUTDOWN:
5937 return io_shutdown_prep(req, sqe);
5938 case IORING_OP_RENAMEAT:
5939 return io_renameat_prep(req, sqe);
5940 case IORING_OP_UNLINKAT:
5941 return io_unlinkat_prep(req, sqe);
5942 }
5943
5944 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5945 req->opcode);
5946 return -EINVAL;
5947 }
5948
io_req_prep_async(struct io_kiocb * req)5949 static int io_req_prep_async(struct io_kiocb *req)
5950 {
5951 if (!io_op_defs[req->opcode].needs_async_setup)
5952 return 0;
5953 if (WARN_ON_ONCE(req->async_data))
5954 return -EFAULT;
5955 if (io_alloc_async_data(req))
5956 return -EAGAIN;
5957
5958 switch (req->opcode) {
5959 case IORING_OP_READV:
5960 return io_rw_prep_async(req, READ);
5961 case IORING_OP_WRITEV:
5962 return io_rw_prep_async(req, WRITE);
5963 case IORING_OP_SENDMSG:
5964 return io_sendmsg_prep_async(req);
5965 case IORING_OP_RECVMSG:
5966 return io_recvmsg_prep_async(req);
5967 case IORING_OP_CONNECT:
5968 return io_connect_prep_async(req);
5969 }
5970 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5971 req->opcode);
5972 return -EFAULT;
5973 }
5974
io_get_sequence(struct io_kiocb * req)5975 static u32 io_get_sequence(struct io_kiocb *req)
5976 {
5977 struct io_kiocb *pos;
5978 struct io_ring_ctx *ctx = req->ctx;
5979 u32 total_submitted, nr_reqs = 0;
5980
5981 io_for_each_link(pos, req)
5982 nr_reqs++;
5983
5984 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5985 return total_submitted - nr_reqs;
5986 }
5987
io_req_defer(struct io_kiocb * req)5988 static int io_req_defer(struct io_kiocb *req)
5989 {
5990 struct io_ring_ctx *ctx = req->ctx;
5991 struct io_defer_entry *de;
5992 int ret;
5993 u32 seq;
5994
5995 /* Still need defer if there is pending req in defer list. */
5996 if (likely(list_empty_careful(&ctx->defer_list) &&
5997 !(req->flags & REQ_F_IO_DRAIN)))
5998 return 0;
5999
6000 seq = io_get_sequence(req);
6001 /* Still a chance to pass the sequence check */
6002 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6003 return 0;
6004
6005 ret = io_req_prep_async(req);
6006 if (ret)
6007 return ret;
6008 io_prep_async_link(req);
6009 de = kmalloc(sizeof(*de), GFP_KERNEL);
6010 if (!de)
6011 return -ENOMEM;
6012
6013 spin_lock_irq(&ctx->completion_lock);
6014 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6015 spin_unlock_irq(&ctx->completion_lock);
6016 kfree(de);
6017 io_queue_async_work(req);
6018 return -EIOCBQUEUED;
6019 }
6020
6021 trace_io_uring_defer(ctx, req, req->user_data);
6022 de->req = req;
6023 de->seq = seq;
6024 list_add_tail(&de->list, &ctx->defer_list);
6025 spin_unlock_irq(&ctx->completion_lock);
6026 return -EIOCBQUEUED;
6027 }
6028
io_clean_op(struct io_kiocb * req)6029 static void io_clean_op(struct io_kiocb *req)
6030 {
6031 if (req->flags & REQ_F_BUFFER_SELECTED) {
6032 switch (req->opcode) {
6033 case IORING_OP_READV:
6034 case IORING_OP_READ_FIXED:
6035 case IORING_OP_READ:
6036 kfree((void *)(unsigned long)req->rw.addr);
6037 break;
6038 case IORING_OP_RECVMSG:
6039 case IORING_OP_RECV:
6040 kfree(req->sr_msg.kbuf);
6041 break;
6042 }
6043 req->flags &= ~REQ_F_BUFFER_SELECTED;
6044 }
6045
6046 if (req->flags & REQ_F_NEED_CLEANUP) {
6047 switch (req->opcode) {
6048 case IORING_OP_READV:
6049 case IORING_OP_READ_FIXED:
6050 case IORING_OP_READ:
6051 case IORING_OP_WRITEV:
6052 case IORING_OP_WRITE_FIXED:
6053 case IORING_OP_WRITE: {
6054 struct io_async_rw *io = req->async_data;
6055 if (io->free_iovec)
6056 kfree(io->free_iovec);
6057 break;
6058 }
6059 case IORING_OP_RECVMSG:
6060 case IORING_OP_SENDMSG: {
6061 struct io_async_msghdr *io = req->async_data;
6062
6063 kfree(io->free_iov);
6064 break;
6065 }
6066 case IORING_OP_SPLICE:
6067 case IORING_OP_TEE:
6068 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6069 io_put_file(req->splice.file_in);
6070 break;
6071 case IORING_OP_OPENAT:
6072 case IORING_OP_OPENAT2:
6073 if (req->open.filename)
6074 putname(req->open.filename);
6075 break;
6076 case IORING_OP_RENAMEAT:
6077 putname(req->rename.oldpath);
6078 putname(req->rename.newpath);
6079 break;
6080 case IORING_OP_UNLINKAT:
6081 putname(req->unlink.filename);
6082 break;
6083 }
6084 req->flags &= ~REQ_F_NEED_CLEANUP;
6085 }
6086 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6087 kfree(req->apoll->double_poll);
6088 kfree(req->apoll);
6089 req->apoll = NULL;
6090 }
6091 if (req->flags & REQ_F_INFLIGHT) {
6092 struct io_uring_task *tctx = req->task->io_uring;
6093
6094 atomic_dec(&tctx->inflight_tracked);
6095 req->flags &= ~REQ_F_INFLIGHT;
6096 }
6097 }
6098
io_issue_sqe(struct io_kiocb * req,unsigned int issue_flags)6099 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6100 {
6101 struct io_ring_ctx *ctx = req->ctx;
6102 const struct cred *creds = NULL;
6103 int ret;
6104
6105 if (req->work.creds && req->work.creds != current_cred())
6106 creds = override_creds(req->work.creds);
6107
6108 switch (req->opcode) {
6109 case IORING_OP_NOP:
6110 ret = io_nop(req, issue_flags);
6111 break;
6112 case IORING_OP_READV:
6113 case IORING_OP_READ_FIXED:
6114 case IORING_OP_READ:
6115 ret = io_read(req, issue_flags);
6116 break;
6117 case IORING_OP_WRITEV:
6118 case IORING_OP_WRITE_FIXED:
6119 case IORING_OP_WRITE:
6120 ret = io_write(req, issue_flags);
6121 break;
6122 case IORING_OP_FSYNC:
6123 ret = io_fsync(req, issue_flags);
6124 break;
6125 case IORING_OP_POLL_ADD:
6126 ret = io_poll_add(req, issue_flags);
6127 break;
6128 case IORING_OP_POLL_REMOVE:
6129 ret = io_poll_update(req, issue_flags);
6130 break;
6131 case IORING_OP_SYNC_FILE_RANGE:
6132 ret = io_sync_file_range(req, issue_flags);
6133 break;
6134 case IORING_OP_SENDMSG:
6135 ret = io_sendmsg(req, issue_flags);
6136 break;
6137 case IORING_OP_SEND:
6138 ret = io_send(req, issue_flags);
6139 break;
6140 case IORING_OP_RECVMSG:
6141 ret = io_recvmsg(req, issue_flags);
6142 break;
6143 case IORING_OP_RECV:
6144 ret = io_recv(req, issue_flags);
6145 break;
6146 case IORING_OP_TIMEOUT:
6147 ret = io_timeout(req, issue_flags);
6148 break;
6149 case IORING_OP_TIMEOUT_REMOVE:
6150 ret = io_timeout_remove(req, issue_flags);
6151 break;
6152 case IORING_OP_ACCEPT:
6153 ret = io_accept(req, issue_flags);
6154 break;
6155 case IORING_OP_CONNECT:
6156 ret = io_connect(req, issue_flags);
6157 break;
6158 case IORING_OP_ASYNC_CANCEL:
6159 ret = io_async_cancel(req, issue_flags);
6160 break;
6161 case IORING_OP_FALLOCATE:
6162 ret = io_fallocate(req, issue_flags);
6163 break;
6164 case IORING_OP_OPENAT:
6165 ret = io_openat(req, issue_flags);
6166 break;
6167 case IORING_OP_CLOSE:
6168 ret = io_close(req, issue_flags);
6169 break;
6170 case IORING_OP_FILES_UPDATE:
6171 ret = io_files_update(req, issue_flags);
6172 break;
6173 case IORING_OP_STATX:
6174 ret = io_statx(req, issue_flags);
6175 break;
6176 case IORING_OP_FADVISE:
6177 ret = io_fadvise(req, issue_flags);
6178 break;
6179 case IORING_OP_MADVISE:
6180 ret = io_madvise(req, issue_flags);
6181 break;
6182 case IORING_OP_OPENAT2:
6183 ret = io_openat2(req, issue_flags);
6184 break;
6185 case IORING_OP_EPOLL_CTL:
6186 ret = io_epoll_ctl(req, issue_flags);
6187 break;
6188 case IORING_OP_SPLICE:
6189 ret = io_splice(req, issue_flags);
6190 break;
6191 case IORING_OP_PROVIDE_BUFFERS:
6192 ret = io_provide_buffers(req, issue_flags);
6193 break;
6194 case IORING_OP_REMOVE_BUFFERS:
6195 ret = io_remove_buffers(req, issue_flags);
6196 break;
6197 case IORING_OP_TEE:
6198 ret = io_tee(req, issue_flags);
6199 break;
6200 case IORING_OP_SHUTDOWN:
6201 ret = io_shutdown(req, issue_flags);
6202 break;
6203 case IORING_OP_RENAMEAT:
6204 ret = io_renameat(req, issue_flags);
6205 break;
6206 case IORING_OP_UNLINKAT:
6207 ret = io_unlinkat(req, issue_flags);
6208 break;
6209 default:
6210 ret = -EINVAL;
6211 break;
6212 }
6213
6214 if (creds)
6215 revert_creds(creds);
6216
6217 if (ret)
6218 return ret;
6219
6220 /* If the op doesn't have a file, we're not polling for it */
6221 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6222 const bool in_async = io_wq_current_is_worker();
6223
6224 /* workqueue context doesn't hold uring_lock, grab it now */
6225 if (in_async)
6226 mutex_lock(&ctx->uring_lock);
6227
6228 io_iopoll_req_issued(req, in_async);
6229
6230 if (in_async)
6231 mutex_unlock(&ctx->uring_lock);
6232 }
6233
6234 return 0;
6235 }
6236
io_wq_submit_work(struct io_wq_work * work)6237 static void io_wq_submit_work(struct io_wq_work *work)
6238 {
6239 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6240 struct io_kiocb *timeout;
6241 int ret = 0;
6242
6243 timeout = io_prep_linked_timeout(req);
6244 if (timeout)
6245 io_queue_linked_timeout(timeout);
6246
6247 if (work->flags & IO_WQ_WORK_CANCEL)
6248 ret = -ECANCELED;
6249
6250 if (!ret) {
6251 do {
6252 ret = io_issue_sqe(req, 0);
6253 /*
6254 * We can get EAGAIN for polled IO even though we're
6255 * forcing a sync submission from here, since we can't
6256 * wait for request slots on the block side.
6257 */
6258 if (ret != -EAGAIN)
6259 break;
6260 cond_resched();
6261 } while (1);
6262 }
6263
6264 /* avoid locking problems by failing it from a clean context */
6265 if (ret) {
6266 /* io-wq is going to take one down */
6267 req_ref_get(req);
6268 io_req_task_queue_fail(req, ret);
6269 }
6270 }
6271
6272 #define FFS_ASYNC_READ 0x1UL
6273 #define FFS_ASYNC_WRITE 0x2UL
6274 #ifdef CONFIG_64BIT
6275 #define FFS_ISREG 0x4UL
6276 #else
6277 #define FFS_ISREG 0x0UL
6278 #endif
6279 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6280
io_fixed_file_slot(struct io_file_table * table,unsigned i)6281 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6282 unsigned i)
6283 {
6284 struct io_fixed_file *table_l2;
6285
6286 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6287 return &table_l2[i & IORING_FILE_TABLE_MASK];
6288 }
6289
io_file_from_index(struct io_ring_ctx * ctx,int index)6290 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6291 int index)
6292 {
6293 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6294
6295 return (struct file *) (slot->file_ptr & FFS_MASK);
6296 }
6297
io_fixed_file_set(struct io_fixed_file * file_slot,struct file * file)6298 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6299 {
6300 unsigned long file_ptr = (unsigned long) file;
6301
6302 if (__io_file_supports_async(file, READ))
6303 file_ptr |= FFS_ASYNC_READ;
6304 if (__io_file_supports_async(file, WRITE))
6305 file_ptr |= FFS_ASYNC_WRITE;
6306 if (S_ISREG(file_inode(file)->i_mode))
6307 file_ptr |= FFS_ISREG;
6308 file_slot->file_ptr = file_ptr;
6309 }
6310
io_file_get(struct io_submit_state * state,struct io_kiocb * req,int fd,bool fixed)6311 static struct file *io_file_get(struct io_submit_state *state,
6312 struct io_kiocb *req, int fd, bool fixed)
6313 {
6314 struct io_ring_ctx *ctx = req->ctx;
6315 struct file *file;
6316
6317 if (fixed) {
6318 unsigned long file_ptr;
6319
6320 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6321 return NULL;
6322 fd = array_index_nospec(fd, ctx->nr_user_files);
6323 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6324 file = (struct file *) (file_ptr & FFS_MASK);
6325 file_ptr &= ~FFS_MASK;
6326 /* mask in overlapping REQ_F and FFS bits */
6327 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6328 io_req_set_rsrc_node(req);
6329 } else {
6330 trace_io_uring_file_get(ctx, fd);
6331 file = __io_file_get(state, fd);
6332
6333 /* we don't allow fixed io_uring files */
6334 if (file && unlikely(file->f_op == &io_uring_fops))
6335 io_req_track_inflight(req);
6336 }
6337
6338 return file;
6339 }
6340
io_link_timeout_fn(struct hrtimer * timer)6341 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6342 {
6343 struct io_timeout_data *data = container_of(timer,
6344 struct io_timeout_data, timer);
6345 struct io_kiocb *prev, *req = data->req;
6346 struct io_ring_ctx *ctx = req->ctx;
6347 unsigned long flags;
6348
6349 spin_lock_irqsave(&ctx->completion_lock, flags);
6350 prev = req->timeout.head;
6351 req->timeout.head = NULL;
6352
6353 /*
6354 * We don't expect the list to be empty, that will only happen if we
6355 * race with the completion of the linked work.
6356 */
6357 if (prev) {
6358 io_remove_next_linked(prev);
6359 if (!req_ref_inc_not_zero(prev))
6360 prev = NULL;
6361 }
6362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6363
6364 if (prev) {
6365 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6366 io_put_req_deferred(prev, 1);
6367 io_put_req_deferred(req, 1);
6368 } else {
6369 io_req_complete_post(req, -ETIME, 0);
6370 }
6371 return HRTIMER_NORESTART;
6372 }
6373
io_queue_linked_timeout(struct io_kiocb * req)6374 static void io_queue_linked_timeout(struct io_kiocb *req)
6375 {
6376 struct io_ring_ctx *ctx = req->ctx;
6377
6378 spin_lock_irq(&ctx->completion_lock);
6379 /*
6380 * If the back reference is NULL, then our linked request finished
6381 * before we got a chance to setup the timer
6382 */
6383 if (req->timeout.head) {
6384 struct io_timeout_data *data = req->async_data;
6385
6386 data->timer.function = io_link_timeout_fn;
6387 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6388 data->mode);
6389 }
6390 spin_unlock_irq(&ctx->completion_lock);
6391 /* drop submission reference */
6392 io_put_req(req);
6393 }
6394
io_prep_linked_timeout(struct io_kiocb * req)6395 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6396 {
6397 struct io_kiocb *nxt = req->link;
6398
6399 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6400 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6401 return NULL;
6402
6403 nxt->timeout.head = req;
6404 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6405 req->flags |= REQ_F_LINK_TIMEOUT;
6406 return nxt;
6407 }
6408
__io_queue_sqe(struct io_kiocb * req)6409 static void __io_queue_sqe(struct io_kiocb *req)
6410 {
6411 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6412 int ret;
6413
6414 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6415
6416 /*
6417 * We async punt it if the file wasn't marked NOWAIT, or if the file
6418 * doesn't support non-blocking read/write attempts
6419 */
6420 if (likely(!ret)) {
6421 /* drop submission reference */
6422 if (req->flags & REQ_F_COMPLETE_INLINE) {
6423 struct io_ring_ctx *ctx = req->ctx;
6424 struct io_comp_state *cs = &ctx->submit_state.comp;
6425
6426 cs->reqs[cs->nr++] = req;
6427 if (cs->nr == ARRAY_SIZE(cs->reqs))
6428 io_submit_flush_completions(cs, ctx);
6429 } else {
6430 io_put_req(req);
6431 }
6432 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6433 if (!io_arm_poll_handler(req)) {
6434 /*
6435 * Queued up for async execution, worker will release
6436 * submit reference when the iocb is actually submitted.
6437 */
6438 io_queue_async_work(req);
6439 }
6440 } else {
6441 io_req_complete_failed(req, ret);
6442 }
6443 if (linked_timeout)
6444 io_queue_linked_timeout(linked_timeout);
6445 }
6446
io_queue_sqe(struct io_kiocb * req)6447 static void io_queue_sqe(struct io_kiocb *req)
6448 {
6449 int ret;
6450
6451 ret = io_req_defer(req);
6452 if (ret) {
6453 if (ret != -EIOCBQUEUED) {
6454 fail_req:
6455 io_req_complete_failed(req, ret);
6456 }
6457 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6458 ret = io_req_prep_async(req);
6459 if (unlikely(ret))
6460 goto fail_req;
6461 io_queue_async_work(req);
6462 } else {
6463 __io_queue_sqe(req);
6464 }
6465 }
6466
6467 /*
6468 * Check SQE restrictions (opcode and flags).
6469 *
6470 * Returns 'true' if SQE is allowed, 'false' otherwise.
6471 */
io_check_restriction(struct io_ring_ctx * ctx,struct io_kiocb * req,unsigned int sqe_flags)6472 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6473 struct io_kiocb *req,
6474 unsigned int sqe_flags)
6475 {
6476 if (!ctx->restricted)
6477 return true;
6478
6479 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6480 return false;
6481
6482 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6483 ctx->restrictions.sqe_flags_required)
6484 return false;
6485
6486 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6487 ctx->restrictions.sqe_flags_required))
6488 return false;
6489
6490 return true;
6491 }
6492
io_init_req(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)6493 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6494 const struct io_uring_sqe *sqe)
6495 {
6496 struct io_submit_state *state;
6497 unsigned int sqe_flags;
6498 int personality, ret = 0;
6499
6500 req->opcode = READ_ONCE(sqe->opcode);
6501 /* same numerical values with corresponding REQ_F_*, safe to copy */
6502 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6503 req->user_data = READ_ONCE(sqe->user_data);
6504 req->async_data = NULL;
6505 req->file = NULL;
6506 req->ctx = ctx;
6507 req->link = NULL;
6508 req->fixed_rsrc_refs = NULL;
6509 /* one is dropped after submission, the other at completion */
6510 atomic_set(&req->refs, 2);
6511 req->task = current;
6512 req->result = 0;
6513 req->work.creds = NULL;
6514
6515 /* enforce forwards compatibility on users */
6516 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6517 return -EINVAL;
6518 if (unlikely(req->opcode >= IORING_OP_LAST))
6519 return -EINVAL;
6520 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6521 return -EACCES;
6522
6523 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6524 !io_op_defs[req->opcode].buffer_select)
6525 return -EOPNOTSUPP;
6526
6527 personality = READ_ONCE(sqe->personality);
6528 if (personality) {
6529 req->work.creds = xa_load(&ctx->personalities, personality);
6530 if (!req->work.creds)
6531 return -EINVAL;
6532 get_cred(req->work.creds);
6533 }
6534 state = &ctx->submit_state;
6535
6536 /*
6537 * Plug now if we have more than 1 IO left after this, and the target
6538 * is potentially a read/write to block based storage.
6539 */
6540 if (!state->plug_started && state->ios_left > 1 &&
6541 io_op_defs[req->opcode].plug) {
6542 blk_start_plug(&state->plug);
6543 state->plug_started = true;
6544 }
6545
6546 if (io_op_defs[req->opcode].needs_file) {
6547 bool fixed = req->flags & REQ_F_FIXED_FILE;
6548
6549 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6550 if (unlikely(!req->file))
6551 ret = -EBADF;
6552 }
6553
6554 state->ios_left--;
6555 return ret;
6556 }
6557
io_submit_sqe(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)6558 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6559 const struct io_uring_sqe *sqe)
6560 {
6561 struct io_submit_link *link = &ctx->submit_state.link;
6562 int ret;
6563
6564 ret = io_init_req(ctx, req, sqe);
6565 if (unlikely(ret)) {
6566 fail_req:
6567 if (link->head) {
6568 /* fail even hard links since we don't submit */
6569 link->head->flags |= REQ_F_FAIL_LINK;
6570 io_req_complete_failed(link->head, -ECANCELED);
6571 link->head = NULL;
6572 }
6573 io_req_complete_failed(req, ret);
6574 return ret;
6575 }
6576 ret = io_req_prep(req, sqe);
6577 if (unlikely(ret))
6578 goto fail_req;
6579
6580 /* don't need @sqe from now on */
6581 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6582 true, ctx->flags & IORING_SETUP_SQPOLL);
6583
6584 /*
6585 * If we already have a head request, queue this one for async
6586 * submittal once the head completes. If we don't have a head but
6587 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6588 * submitted sync once the chain is complete. If none of those
6589 * conditions are true (normal request), then just queue it.
6590 */
6591 if (link->head) {
6592 struct io_kiocb *head = link->head;
6593
6594 /*
6595 * Taking sequential execution of a link, draining both sides
6596 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6597 * requests in the link. So, it drains the head and the
6598 * next after the link request. The last one is done via
6599 * drain_next flag to persist the effect across calls.
6600 */
6601 if (req->flags & REQ_F_IO_DRAIN) {
6602 head->flags |= REQ_F_IO_DRAIN;
6603 ctx->drain_next = 1;
6604 }
6605 ret = io_req_prep_async(req);
6606 if (unlikely(ret))
6607 goto fail_req;
6608 trace_io_uring_link(ctx, req, head);
6609 link->last->link = req;
6610 link->last = req;
6611
6612 /* last request of a link, enqueue the link */
6613 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6614 io_queue_sqe(head);
6615 link->head = NULL;
6616 }
6617 } else {
6618 if (unlikely(ctx->drain_next)) {
6619 req->flags |= REQ_F_IO_DRAIN;
6620 ctx->drain_next = 0;
6621 }
6622 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6623 link->head = req;
6624 link->last = req;
6625 } else {
6626 io_queue_sqe(req);
6627 }
6628 }
6629
6630 return 0;
6631 }
6632
6633 /*
6634 * Batched submission is done, ensure local IO is flushed out.
6635 */
io_submit_state_end(struct io_submit_state * state,struct io_ring_ctx * ctx)6636 static void io_submit_state_end(struct io_submit_state *state,
6637 struct io_ring_ctx *ctx)
6638 {
6639 if (state->link.head)
6640 io_queue_sqe(state->link.head);
6641 if (state->comp.nr)
6642 io_submit_flush_completions(&state->comp, ctx);
6643 if (state->plug_started)
6644 blk_finish_plug(&state->plug);
6645 io_state_file_put(state);
6646 }
6647
6648 /*
6649 * Start submission side cache.
6650 */
io_submit_state_start(struct io_submit_state * state,unsigned int max_ios)6651 static void io_submit_state_start(struct io_submit_state *state,
6652 unsigned int max_ios)
6653 {
6654 state->plug_started = false;
6655 state->ios_left = max_ios;
6656 /* set only head, no need to init link_last in advance */
6657 state->link.head = NULL;
6658 }
6659
io_commit_sqring(struct io_ring_ctx * ctx)6660 static void io_commit_sqring(struct io_ring_ctx *ctx)
6661 {
6662 struct io_rings *rings = ctx->rings;
6663
6664 /*
6665 * Ensure any loads from the SQEs are done at this point,
6666 * since once we write the new head, the application could
6667 * write new data to them.
6668 */
6669 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6670 }
6671
6672 /*
6673 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6674 * that is mapped by userspace. This means that care needs to be taken to
6675 * ensure that reads are stable, as we cannot rely on userspace always
6676 * being a good citizen. If members of the sqe are validated and then later
6677 * used, it's important that those reads are done through READ_ONCE() to
6678 * prevent a re-load down the line.
6679 */
io_get_sqe(struct io_ring_ctx * ctx)6680 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6681 {
6682 u32 *sq_array = ctx->sq_array;
6683 unsigned head;
6684
6685 /*
6686 * The cached sq head (or cq tail) serves two purposes:
6687 *
6688 * 1) allows us to batch the cost of updating the user visible
6689 * head updates.
6690 * 2) allows the kernel side to track the head on its own, even
6691 * though the application is the one updating it.
6692 */
6693 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6694 if (likely(head < ctx->sq_entries))
6695 return &ctx->sq_sqes[head];
6696
6697 /* drop invalid entries */
6698 ctx->cached_sq_dropped++;
6699 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6700 return NULL;
6701 }
6702
io_submit_sqes(struct io_ring_ctx * ctx,unsigned int nr)6703 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6704 {
6705 int submitted = 0;
6706
6707 /* make sure SQ entry isn't read before tail */
6708 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6709
6710 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6711 return -EAGAIN;
6712
6713 percpu_counter_add(¤t->io_uring->inflight, nr);
6714 refcount_add(nr, ¤t->usage);
6715 io_submit_state_start(&ctx->submit_state, nr);
6716
6717 while (submitted < nr) {
6718 const struct io_uring_sqe *sqe;
6719 struct io_kiocb *req;
6720
6721 req = io_alloc_req(ctx);
6722 if (unlikely(!req)) {
6723 if (!submitted)
6724 submitted = -EAGAIN;
6725 break;
6726 }
6727 sqe = io_get_sqe(ctx);
6728 if (unlikely(!sqe)) {
6729 kmem_cache_free(req_cachep, req);
6730 break;
6731 }
6732 /* will complete beyond this point, count as submitted */
6733 submitted++;
6734 if (io_submit_sqe(ctx, req, sqe))
6735 break;
6736 }
6737
6738 if (unlikely(submitted != nr)) {
6739 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6740 struct io_uring_task *tctx = current->io_uring;
6741 int unused = nr - ref_used;
6742
6743 percpu_ref_put_many(&ctx->refs, unused);
6744 percpu_counter_sub(&tctx->inflight, unused);
6745 put_task_struct_many(current, unused);
6746 }
6747
6748 io_submit_state_end(&ctx->submit_state, ctx);
6749 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6750 io_commit_sqring(ctx);
6751
6752 return submitted;
6753 }
6754
io_ring_set_wakeup_flag(struct io_ring_ctx * ctx)6755 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6756 {
6757 /* Tell userspace we may need a wakeup call */
6758 spin_lock_irq(&ctx->completion_lock);
6759 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6760 spin_unlock_irq(&ctx->completion_lock);
6761 }
6762
io_ring_clear_wakeup_flag(struct io_ring_ctx * ctx)6763 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6764 {
6765 spin_lock_irq(&ctx->completion_lock);
6766 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6767 spin_unlock_irq(&ctx->completion_lock);
6768 }
6769
__io_sq_thread(struct io_ring_ctx * ctx,bool cap_entries)6770 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6771 {
6772 unsigned int to_submit;
6773 int ret = 0;
6774
6775 to_submit = io_sqring_entries(ctx);
6776 /* if we're handling multiple rings, cap submit size for fairness */
6777 if (cap_entries && to_submit > 8)
6778 to_submit = 8;
6779
6780 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6781 unsigned nr_events = 0;
6782
6783 mutex_lock(&ctx->uring_lock);
6784 if (!list_empty(&ctx->iopoll_list))
6785 io_do_iopoll(ctx, &nr_events, 0);
6786
6787 /*
6788 * Don't submit if refs are dying, good for io_uring_register(),
6789 * but also it is relied upon by io_ring_exit_work()
6790 */
6791 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6792 !(ctx->flags & IORING_SETUP_R_DISABLED))
6793 ret = io_submit_sqes(ctx, to_submit);
6794 mutex_unlock(&ctx->uring_lock);
6795 }
6796
6797 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6798 wake_up(&ctx->sqo_sq_wait);
6799
6800 return ret;
6801 }
6802
io_sqd_update_thread_idle(struct io_sq_data * sqd)6803 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6804 {
6805 struct io_ring_ctx *ctx;
6806 unsigned sq_thread_idle = 0;
6807
6808 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6809 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6810 sqd->sq_thread_idle = sq_thread_idle;
6811 }
6812
io_sq_thread(void * data)6813 static int io_sq_thread(void *data)
6814 {
6815 struct io_sq_data *sqd = data;
6816 struct io_ring_ctx *ctx;
6817 unsigned long timeout = 0;
6818 char buf[TASK_COMM_LEN];
6819 DEFINE_WAIT(wait);
6820
6821 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6822 set_task_comm(current, buf);
6823
6824 if (sqd->sq_cpu != -1)
6825 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6826 else
6827 set_cpus_allowed_ptr(current, cpu_online_mask);
6828 current->flags |= PF_NO_SETAFFINITY;
6829
6830 mutex_lock(&sqd->lock);
6831 /* a user may had exited before the thread started */
6832 io_run_task_work_head(&sqd->park_task_work);
6833
6834 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6835 int ret;
6836 bool cap_entries, sqt_spin, needs_sched;
6837
6838 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6839 signal_pending(current)) {
6840 bool did_sig = false;
6841
6842 mutex_unlock(&sqd->lock);
6843 if (signal_pending(current)) {
6844 struct ksignal ksig;
6845
6846 did_sig = get_signal(&ksig);
6847 }
6848 cond_resched();
6849 mutex_lock(&sqd->lock);
6850 io_run_task_work();
6851 io_run_task_work_head(&sqd->park_task_work);
6852 if (did_sig)
6853 break;
6854 timeout = jiffies + sqd->sq_thread_idle;
6855 continue;
6856 }
6857 sqt_spin = false;
6858 cap_entries = !list_is_singular(&sqd->ctx_list);
6859 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6860 const struct cred *creds = NULL;
6861
6862 if (ctx->sq_creds != current_cred())
6863 creds = override_creds(ctx->sq_creds);
6864 ret = __io_sq_thread(ctx, cap_entries);
6865 if (creds)
6866 revert_creds(creds);
6867 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6868 sqt_spin = true;
6869 }
6870
6871 if (sqt_spin || !time_after(jiffies, timeout)) {
6872 io_run_task_work();
6873 cond_resched();
6874 if (sqt_spin)
6875 timeout = jiffies + sqd->sq_thread_idle;
6876 continue;
6877 }
6878
6879 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6880 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6881 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6882 io_ring_set_wakeup_flag(ctx);
6883
6884 needs_sched = true;
6885 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6886 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6887 !list_empty_careful(&ctx->iopoll_list)) {
6888 needs_sched = false;
6889 break;
6890 }
6891 if (io_sqring_entries(ctx)) {
6892 needs_sched = false;
6893 break;
6894 }
6895 }
6896
6897 if (needs_sched) {
6898 mutex_unlock(&sqd->lock);
6899 schedule();
6900 mutex_lock(&sqd->lock);
6901 }
6902 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6903 io_ring_clear_wakeup_flag(ctx);
6904 }
6905
6906 finish_wait(&sqd->wait, &wait);
6907 io_run_task_work_head(&sqd->park_task_work);
6908 timeout = jiffies + sqd->sq_thread_idle;
6909 }
6910
6911 io_uring_cancel_sqpoll(sqd);
6912 sqd->thread = NULL;
6913 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6914 io_ring_set_wakeup_flag(ctx);
6915 io_run_task_work();
6916 io_run_task_work_head(&sqd->park_task_work);
6917 mutex_unlock(&sqd->lock);
6918
6919 complete(&sqd->exited);
6920 do_exit(0);
6921 }
6922
6923 struct io_wait_queue {
6924 struct wait_queue_entry wq;
6925 struct io_ring_ctx *ctx;
6926 unsigned to_wait;
6927 unsigned nr_timeouts;
6928 };
6929
io_should_wake(struct io_wait_queue * iowq)6930 static inline bool io_should_wake(struct io_wait_queue *iowq)
6931 {
6932 struct io_ring_ctx *ctx = iowq->ctx;
6933
6934 /*
6935 * Wake up if we have enough events, or if a timeout occurred since we
6936 * started waiting. For timeouts, we always want to return to userspace,
6937 * regardless of event count.
6938 */
6939 return io_cqring_events(ctx) >= iowq->to_wait ||
6940 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6941 }
6942
io_wake_function(struct wait_queue_entry * curr,unsigned int mode,int wake_flags,void * key)6943 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6944 int wake_flags, void *key)
6945 {
6946 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6947 wq);
6948
6949 /*
6950 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6951 * the task, and the next invocation will do it.
6952 */
6953 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6954 return autoremove_wake_function(curr, mode, wake_flags, key);
6955 return -1;
6956 }
6957
io_run_task_work_sig(void)6958 static int io_run_task_work_sig(void)
6959 {
6960 if (io_run_task_work())
6961 return 1;
6962 if (!signal_pending(current))
6963 return 0;
6964 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6965 return -ERESTARTSYS;
6966 return -EINTR;
6967 }
6968
6969 /* when returns >0, the caller should retry */
io_cqring_wait_schedule(struct io_ring_ctx * ctx,struct io_wait_queue * iowq,signed long * timeout)6970 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6971 struct io_wait_queue *iowq,
6972 signed long *timeout)
6973 {
6974 int ret;
6975
6976 /* make sure we run task_work before checking for signals */
6977 ret = io_run_task_work_sig();
6978 if (ret || io_should_wake(iowq))
6979 return ret;
6980 /* let the caller flush overflows, retry */
6981 if (test_bit(0, &ctx->cq_check_overflow))
6982 return 1;
6983
6984 *timeout = schedule_timeout(*timeout);
6985 return !*timeout ? -ETIME : 1;
6986 }
6987
6988 /*
6989 * Wait until events become available, if we don't already have some. The
6990 * application must reap them itself, as they reside on the shared cq ring.
6991 */
io_cqring_wait(struct io_ring_ctx * ctx,int min_events,const sigset_t __user * sig,size_t sigsz,struct __kernel_timespec __user * uts)6992 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6993 const sigset_t __user *sig, size_t sigsz,
6994 struct __kernel_timespec __user *uts)
6995 {
6996 struct io_wait_queue iowq = {
6997 .wq = {
6998 .private = current,
6999 .func = io_wake_function,
7000 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7001 },
7002 .ctx = ctx,
7003 .to_wait = min_events,
7004 };
7005 struct io_rings *rings = ctx->rings;
7006 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7007 int ret;
7008
7009 do {
7010 io_cqring_overflow_flush(ctx, false);
7011 if (io_cqring_events(ctx) >= min_events)
7012 return 0;
7013 if (!io_run_task_work())
7014 break;
7015 } while (1);
7016
7017 if (sig) {
7018 #ifdef CONFIG_COMPAT
7019 if (in_compat_syscall())
7020 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7021 sigsz);
7022 else
7023 #endif
7024 ret = set_user_sigmask(sig, sigsz);
7025
7026 if (ret)
7027 return ret;
7028 }
7029
7030 if (uts) {
7031 struct timespec64 ts;
7032
7033 if (get_timespec64(&ts, uts))
7034 return -EFAULT;
7035 timeout = timespec64_to_jiffies(&ts);
7036 }
7037
7038 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7039 trace_io_uring_cqring_wait(ctx, min_events);
7040 do {
7041 /* if we can't even flush overflow, don't wait for more */
7042 if (!io_cqring_overflow_flush(ctx, false)) {
7043 ret = -EBUSY;
7044 break;
7045 }
7046 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7047 TASK_INTERRUPTIBLE);
7048 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7049 finish_wait(&ctx->wait, &iowq.wq);
7050 cond_resched();
7051 } while (ret > 0);
7052
7053 restore_saved_sigmask_unless(ret == -EINTR);
7054
7055 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7056 }
7057
io_free_file_tables(struct io_file_table * table,unsigned nr_files)7058 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7059 {
7060 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7061
7062 for (i = 0; i < nr_tables; i++)
7063 kfree(table->files[i]);
7064 kfree(table->files);
7065 table->files = NULL;
7066 }
7067
io_rsrc_ref_lock(struct io_ring_ctx * ctx)7068 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7069 {
7070 spin_lock_bh(&ctx->rsrc_ref_lock);
7071 }
7072
io_rsrc_ref_unlock(struct io_ring_ctx * ctx)7073 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7074 {
7075 spin_unlock_bh(&ctx->rsrc_ref_lock);
7076 }
7077
io_rsrc_node_destroy(struct io_rsrc_node * ref_node)7078 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7079 {
7080 percpu_ref_exit(&ref_node->refs);
7081 kfree(ref_node);
7082 }
7083
io_rsrc_node_switch(struct io_ring_ctx * ctx,struct io_rsrc_data * data_to_kill)7084 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7085 struct io_rsrc_data *data_to_kill)
7086 {
7087 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7088 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7089
7090 if (data_to_kill) {
7091 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7092
7093 rsrc_node->rsrc_data = data_to_kill;
7094 io_rsrc_ref_lock(ctx);
7095 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7096 io_rsrc_ref_unlock(ctx);
7097
7098 atomic_inc(&data_to_kill->refs);
7099 percpu_ref_kill(&rsrc_node->refs);
7100 ctx->rsrc_node = NULL;
7101 }
7102
7103 if (!ctx->rsrc_node) {
7104 ctx->rsrc_node = ctx->rsrc_backup_node;
7105 ctx->rsrc_backup_node = NULL;
7106 }
7107 }
7108
io_rsrc_node_switch_start(struct io_ring_ctx * ctx)7109 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7110 {
7111 if (ctx->rsrc_backup_node)
7112 return 0;
7113 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7114 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7115 }
7116
io_rsrc_ref_quiesce(struct io_rsrc_data * data,struct io_ring_ctx * ctx)7117 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7118 {
7119 int ret;
7120
7121 /* As we may drop ->uring_lock, other task may have started quiesce */
7122 if (data->quiesce)
7123 return -ENXIO;
7124
7125 data->quiesce = true;
7126 do {
7127 ret = io_rsrc_node_switch_start(ctx);
7128 if (ret)
7129 break;
7130 io_rsrc_node_switch(ctx, data);
7131
7132 /* kill initial ref, already quiesced if zero */
7133 if (atomic_dec_and_test(&data->refs))
7134 break;
7135 flush_delayed_work(&ctx->rsrc_put_work);
7136 ret = wait_for_completion_interruptible(&data->done);
7137 if (!ret)
7138 break;
7139
7140 atomic_inc(&data->refs);
7141 /* wait for all works potentially completing data->done */
7142 flush_delayed_work(&ctx->rsrc_put_work);
7143 reinit_completion(&data->done);
7144
7145 mutex_unlock(&ctx->uring_lock);
7146 ret = io_run_task_work_sig();
7147 mutex_lock(&ctx->uring_lock);
7148 } while (ret >= 0);
7149 data->quiesce = false;
7150
7151 return ret;
7152 }
7153
io_rsrc_data_free(struct io_rsrc_data * data)7154 static void io_rsrc_data_free(struct io_rsrc_data *data)
7155 {
7156 kvfree(data->tags);
7157 kfree(data);
7158 }
7159
io_rsrc_data_alloc(struct io_ring_ctx * ctx,rsrc_put_fn * do_put,unsigned nr)7160 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7161 rsrc_put_fn *do_put,
7162 unsigned nr)
7163 {
7164 struct io_rsrc_data *data;
7165
7166 data = kzalloc(sizeof(*data), GFP_KERNEL);
7167 if (!data)
7168 return NULL;
7169
7170 data->tags = kvcalloc(nr, sizeof(*data->tags), GFP_KERNEL);
7171 if (!data->tags) {
7172 kfree(data);
7173 return NULL;
7174 }
7175
7176 atomic_set(&data->refs, 1);
7177 data->ctx = ctx;
7178 data->do_put = do_put;
7179 init_completion(&data->done);
7180 return data;
7181 }
7182
__io_sqe_files_unregister(struct io_ring_ctx * ctx)7183 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7184 {
7185 #if defined(CONFIG_UNIX)
7186 if (ctx->ring_sock) {
7187 struct sock *sock = ctx->ring_sock->sk;
7188 struct sk_buff *skb;
7189
7190 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7191 kfree_skb(skb);
7192 }
7193 #else
7194 int i;
7195
7196 for (i = 0; i < ctx->nr_user_files; i++) {
7197 struct file *file;
7198
7199 file = io_file_from_index(ctx, i);
7200 if (file)
7201 fput(file);
7202 }
7203 #endif
7204 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7205 io_rsrc_data_free(ctx->file_data);
7206 ctx->file_data = NULL;
7207 ctx->nr_user_files = 0;
7208 }
7209
io_sqe_files_unregister(struct io_ring_ctx * ctx)7210 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7211 {
7212 int ret;
7213
7214 if (!ctx->file_data)
7215 return -ENXIO;
7216 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7217 if (!ret)
7218 __io_sqe_files_unregister(ctx);
7219 return ret;
7220 }
7221
io_sq_thread_unpark(struct io_sq_data * sqd)7222 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7223 __releases(&sqd->lock)
7224 {
7225 WARN_ON_ONCE(sqd->thread == current);
7226
7227 /*
7228 * Do the dance but not conditional clear_bit() because it'd race with
7229 * other threads incrementing park_pending and setting the bit.
7230 */
7231 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7232 if (atomic_dec_return(&sqd->park_pending))
7233 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7234 mutex_unlock(&sqd->lock);
7235 }
7236
io_sq_thread_park(struct io_sq_data * sqd)7237 static void io_sq_thread_park(struct io_sq_data *sqd)
7238 __acquires(&sqd->lock)
7239 {
7240 WARN_ON_ONCE(sqd->thread == current);
7241
7242 atomic_inc(&sqd->park_pending);
7243 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7244 mutex_lock(&sqd->lock);
7245 if (sqd->thread)
7246 wake_up_process(sqd->thread);
7247 }
7248
io_sq_thread_stop(struct io_sq_data * sqd)7249 static void io_sq_thread_stop(struct io_sq_data *sqd)
7250 {
7251 WARN_ON_ONCE(sqd->thread == current);
7252 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7253
7254 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7255 mutex_lock(&sqd->lock);
7256 if (sqd->thread)
7257 wake_up_process(sqd->thread);
7258 mutex_unlock(&sqd->lock);
7259 wait_for_completion(&sqd->exited);
7260 }
7261
io_put_sq_data(struct io_sq_data * sqd)7262 static void io_put_sq_data(struct io_sq_data *sqd)
7263 {
7264 if (refcount_dec_and_test(&sqd->refs)) {
7265 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7266
7267 io_sq_thread_stop(sqd);
7268 kfree(sqd);
7269 }
7270 }
7271
io_sq_thread_finish(struct io_ring_ctx * ctx)7272 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7273 {
7274 struct io_sq_data *sqd = ctx->sq_data;
7275
7276 if (sqd) {
7277 io_sq_thread_park(sqd);
7278 list_del_init(&ctx->sqd_list);
7279 io_sqd_update_thread_idle(sqd);
7280 io_sq_thread_unpark(sqd);
7281
7282 io_put_sq_data(sqd);
7283 ctx->sq_data = NULL;
7284 }
7285 }
7286
io_attach_sq_data(struct io_uring_params * p)7287 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7288 {
7289 struct io_ring_ctx *ctx_attach;
7290 struct io_sq_data *sqd;
7291 struct fd f;
7292
7293 f = fdget(p->wq_fd);
7294 if (!f.file)
7295 return ERR_PTR(-ENXIO);
7296 if (f.file->f_op != &io_uring_fops) {
7297 fdput(f);
7298 return ERR_PTR(-EINVAL);
7299 }
7300
7301 ctx_attach = f.file->private_data;
7302 sqd = ctx_attach->sq_data;
7303 if (!sqd) {
7304 fdput(f);
7305 return ERR_PTR(-EINVAL);
7306 }
7307 if (sqd->task_tgid != current->tgid) {
7308 fdput(f);
7309 return ERR_PTR(-EPERM);
7310 }
7311
7312 refcount_inc(&sqd->refs);
7313 fdput(f);
7314 return sqd;
7315 }
7316
io_get_sq_data(struct io_uring_params * p,bool * attached)7317 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7318 bool *attached)
7319 {
7320 struct io_sq_data *sqd;
7321
7322 *attached = false;
7323 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7324 sqd = io_attach_sq_data(p);
7325 if (!IS_ERR(sqd)) {
7326 *attached = true;
7327 return sqd;
7328 }
7329 /* fall through for EPERM case, setup new sqd/task */
7330 if (PTR_ERR(sqd) != -EPERM)
7331 return sqd;
7332 }
7333
7334 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7335 if (!sqd)
7336 return ERR_PTR(-ENOMEM);
7337
7338 atomic_set(&sqd->park_pending, 0);
7339 refcount_set(&sqd->refs, 1);
7340 INIT_LIST_HEAD(&sqd->ctx_list);
7341 mutex_init(&sqd->lock);
7342 init_waitqueue_head(&sqd->wait);
7343 init_completion(&sqd->exited);
7344 return sqd;
7345 }
7346
7347 #if defined(CONFIG_UNIX)
7348 /*
7349 * Ensure the UNIX gc is aware of our file set, so we are certain that
7350 * the io_uring can be safely unregistered on process exit, even if we have
7351 * loops in the file referencing.
7352 */
__io_sqe_files_scm(struct io_ring_ctx * ctx,int nr,int offset)7353 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7354 {
7355 struct sock *sk = ctx->ring_sock->sk;
7356 struct scm_fp_list *fpl;
7357 struct sk_buff *skb;
7358 int i, nr_files;
7359
7360 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7361 if (!fpl)
7362 return -ENOMEM;
7363
7364 skb = alloc_skb(0, GFP_KERNEL);
7365 if (!skb) {
7366 kfree(fpl);
7367 return -ENOMEM;
7368 }
7369
7370 skb->sk = sk;
7371
7372 nr_files = 0;
7373 fpl->user = get_uid(current_user());
7374 for (i = 0; i < nr; i++) {
7375 struct file *file = io_file_from_index(ctx, i + offset);
7376
7377 if (!file)
7378 continue;
7379 fpl->fp[nr_files] = get_file(file);
7380 unix_inflight(fpl->user, fpl->fp[nr_files]);
7381 nr_files++;
7382 }
7383
7384 if (nr_files) {
7385 fpl->max = SCM_MAX_FD;
7386 fpl->count = nr_files;
7387 UNIXCB(skb).fp = fpl;
7388 skb->destructor = unix_destruct_scm;
7389 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7390 skb_queue_head(&sk->sk_receive_queue, skb);
7391
7392 for (i = 0; i < nr_files; i++)
7393 fput(fpl->fp[i]);
7394 } else {
7395 kfree_skb(skb);
7396 kfree(fpl);
7397 }
7398
7399 return 0;
7400 }
7401
7402 /*
7403 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7404 * causes regular reference counting to break down. We rely on the UNIX
7405 * garbage collection to take care of this problem for us.
7406 */
io_sqe_files_scm(struct io_ring_ctx * ctx)7407 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7408 {
7409 unsigned left, total;
7410 int ret = 0;
7411
7412 total = 0;
7413 left = ctx->nr_user_files;
7414 while (left) {
7415 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7416
7417 ret = __io_sqe_files_scm(ctx, this_files, total);
7418 if (ret)
7419 break;
7420 left -= this_files;
7421 total += this_files;
7422 }
7423
7424 if (!ret)
7425 return 0;
7426
7427 while (total < ctx->nr_user_files) {
7428 struct file *file = io_file_from_index(ctx, total);
7429
7430 if (file)
7431 fput(file);
7432 total++;
7433 }
7434
7435 return ret;
7436 }
7437 #else
io_sqe_files_scm(struct io_ring_ctx * ctx)7438 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7439 {
7440 return 0;
7441 }
7442 #endif
7443
io_alloc_file_tables(struct io_file_table * table,unsigned nr_files)7444 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7445 {
7446 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7447
7448 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7449 if (!table->files)
7450 return false;
7451
7452 for (i = 0; i < nr_tables; i++) {
7453 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7454
7455 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7456 GFP_KERNEL);
7457 if (!table->files[i])
7458 break;
7459 nr_files -= this_files;
7460 }
7461
7462 if (i == nr_tables)
7463 return true;
7464
7465 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7466 return false;
7467 }
7468
io_rsrc_file_put(struct io_ring_ctx * ctx,struct io_rsrc_put * prsrc)7469 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7470 {
7471 struct file *file = prsrc->file;
7472 #if defined(CONFIG_UNIX)
7473 struct sock *sock = ctx->ring_sock->sk;
7474 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7475 struct sk_buff *skb;
7476 int i;
7477
7478 __skb_queue_head_init(&list);
7479
7480 /*
7481 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7482 * remove this entry and rearrange the file array.
7483 */
7484 skb = skb_dequeue(head);
7485 while (skb) {
7486 struct scm_fp_list *fp;
7487
7488 fp = UNIXCB(skb).fp;
7489 for (i = 0; i < fp->count; i++) {
7490 int left;
7491
7492 if (fp->fp[i] != file)
7493 continue;
7494
7495 unix_notinflight(fp->user, fp->fp[i]);
7496 left = fp->count - 1 - i;
7497 if (left) {
7498 memmove(&fp->fp[i], &fp->fp[i + 1],
7499 left * sizeof(struct file *));
7500 }
7501 fp->count--;
7502 if (!fp->count) {
7503 kfree_skb(skb);
7504 skb = NULL;
7505 } else {
7506 __skb_queue_tail(&list, skb);
7507 }
7508 fput(file);
7509 file = NULL;
7510 break;
7511 }
7512
7513 if (!file)
7514 break;
7515
7516 __skb_queue_tail(&list, skb);
7517
7518 skb = skb_dequeue(head);
7519 }
7520
7521 if (skb_peek(&list)) {
7522 spin_lock_irq(&head->lock);
7523 while ((skb = __skb_dequeue(&list)) != NULL)
7524 __skb_queue_tail(head, skb);
7525 spin_unlock_irq(&head->lock);
7526 }
7527 #else
7528 fput(file);
7529 #endif
7530 }
7531
__io_rsrc_put_work(struct io_rsrc_node * ref_node)7532 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7533 {
7534 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7535 struct io_ring_ctx *ctx = rsrc_data->ctx;
7536 struct io_rsrc_put *prsrc, *tmp;
7537
7538 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7539 list_del(&prsrc->list);
7540
7541 if (prsrc->tag) {
7542 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7543 unsigned long flags;
7544
7545 io_ring_submit_lock(ctx, lock_ring);
7546 spin_lock_irqsave(&ctx->completion_lock, flags);
7547 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7548 ctx->cq_extra++;
7549 io_commit_cqring(ctx);
7550 spin_unlock_irqrestore(&ctx->completion_lock, flags);
7551 io_cqring_ev_posted(ctx);
7552 io_ring_submit_unlock(ctx, lock_ring);
7553 }
7554
7555 rsrc_data->do_put(ctx, prsrc);
7556 kfree(prsrc);
7557 }
7558
7559 io_rsrc_node_destroy(ref_node);
7560 if (atomic_dec_and_test(&rsrc_data->refs))
7561 complete(&rsrc_data->done);
7562 }
7563
io_rsrc_put_work(struct work_struct * work)7564 static void io_rsrc_put_work(struct work_struct *work)
7565 {
7566 struct io_ring_ctx *ctx;
7567 struct llist_node *node;
7568
7569 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7570 node = llist_del_all(&ctx->rsrc_put_llist);
7571
7572 while (node) {
7573 struct io_rsrc_node *ref_node;
7574 struct llist_node *next = node->next;
7575
7576 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7577 __io_rsrc_put_work(ref_node);
7578 node = next;
7579 }
7580 }
7581
io_rsrc_node_ref_zero(struct percpu_ref * ref)7582 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7583 {
7584 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7585 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7586 bool first_add = false;
7587
7588 io_rsrc_ref_lock(ctx);
7589 node->done = true;
7590
7591 while (!list_empty(&ctx->rsrc_ref_list)) {
7592 node = list_first_entry(&ctx->rsrc_ref_list,
7593 struct io_rsrc_node, node);
7594 /* recycle ref nodes in order */
7595 if (!node->done)
7596 break;
7597 list_del(&node->node);
7598 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7599 }
7600 io_rsrc_ref_unlock(ctx);
7601
7602 if (first_add)
7603 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7604 }
7605
io_rsrc_node_alloc(struct io_ring_ctx * ctx)7606 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7607 {
7608 struct io_rsrc_node *ref_node;
7609
7610 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7611 if (!ref_node)
7612 return NULL;
7613
7614 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7615 0, GFP_KERNEL)) {
7616 kfree(ref_node);
7617 return NULL;
7618 }
7619 INIT_LIST_HEAD(&ref_node->node);
7620 INIT_LIST_HEAD(&ref_node->rsrc_list);
7621 ref_node->done = false;
7622 return ref_node;
7623 }
7624
io_sqe_files_register(struct io_ring_ctx * ctx,void __user * arg,unsigned nr_args,u64 __user * tags)7625 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7626 unsigned nr_args, u64 __user *tags)
7627 {
7628 __s32 __user *fds = (__s32 __user *) arg;
7629 struct file *file;
7630 int fd, ret;
7631 unsigned i;
7632 struct io_rsrc_data *file_data;
7633
7634 if (ctx->file_data)
7635 return -EBUSY;
7636 if (!nr_args)
7637 return -EINVAL;
7638 if (nr_args > IORING_MAX_FIXED_FILES)
7639 return -EMFILE;
7640 ret = io_rsrc_node_switch_start(ctx);
7641 if (ret)
7642 return ret;
7643
7644 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put, nr_args);
7645 if (!file_data)
7646 return -ENOMEM;
7647 ctx->file_data = file_data;
7648 ret = -ENOMEM;
7649 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7650 goto out_free;
7651
7652 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7653 u64 tag = 0;
7654
7655 if ((tags && copy_from_user(&tag, &tags[i], sizeof(tag))) ||
7656 copy_from_user(&fd, &fds[i], sizeof(fd))) {
7657 ret = -EFAULT;
7658 goto out_fput;
7659 }
7660 /* allow sparse sets */
7661 if (fd == -1) {
7662 ret = -EINVAL;
7663 if (unlikely(tag))
7664 goto out_fput;
7665 continue;
7666 }
7667
7668 file = fget(fd);
7669 ret = -EBADF;
7670 if (unlikely(!file))
7671 goto out_fput;
7672
7673 /*
7674 * Don't allow io_uring instances to be registered. If UNIX
7675 * isn't enabled, then this causes a reference cycle and this
7676 * instance can never get freed. If UNIX is enabled we'll
7677 * handle it just fine, but there's still no point in allowing
7678 * a ring fd as it doesn't support regular read/write anyway.
7679 */
7680 if (file->f_op == &io_uring_fops) {
7681 fput(file);
7682 goto out_fput;
7683 }
7684 ctx->file_data->tags[i] = tag;
7685 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7686 }
7687
7688 ret = io_sqe_files_scm(ctx);
7689 if (ret) {
7690 __io_sqe_files_unregister(ctx);
7691 return ret;
7692 }
7693
7694 io_rsrc_node_switch(ctx, NULL);
7695 return ret;
7696 out_fput:
7697 for (i = 0; i < ctx->nr_user_files; i++) {
7698 file = io_file_from_index(ctx, i);
7699 if (file)
7700 fput(file);
7701 }
7702 io_free_file_tables(&ctx->file_table, nr_args);
7703 ctx->nr_user_files = 0;
7704 out_free:
7705 io_rsrc_data_free(ctx->file_data);
7706 ctx->file_data = NULL;
7707 return ret;
7708 }
7709
io_sqe_file_register(struct io_ring_ctx * ctx,struct file * file,int index)7710 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7711 int index)
7712 {
7713 #if defined(CONFIG_UNIX)
7714 struct sock *sock = ctx->ring_sock->sk;
7715 struct sk_buff_head *head = &sock->sk_receive_queue;
7716 struct sk_buff *skb;
7717
7718 /*
7719 * See if we can merge this file into an existing skb SCM_RIGHTS
7720 * file set. If there's no room, fall back to allocating a new skb
7721 * and filling it in.
7722 */
7723 spin_lock_irq(&head->lock);
7724 skb = skb_peek(head);
7725 if (skb) {
7726 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7727
7728 if (fpl->count < SCM_MAX_FD) {
7729 __skb_unlink(skb, head);
7730 spin_unlock_irq(&head->lock);
7731 fpl->fp[fpl->count] = get_file(file);
7732 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7733 fpl->count++;
7734 spin_lock_irq(&head->lock);
7735 __skb_queue_head(head, skb);
7736 } else {
7737 skb = NULL;
7738 }
7739 }
7740 spin_unlock_irq(&head->lock);
7741
7742 if (skb) {
7743 fput(file);
7744 return 0;
7745 }
7746
7747 return __io_sqe_files_scm(ctx, 1, index);
7748 #else
7749 return 0;
7750 #endif
7751 }
7752
io_queue_rsrc_removal(struct io_rsrc_data * data,unsigned idx,struct io_rsrc_node * node,void * rsrc)7753 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7754 struct io_rsrc_node *node, void *rsrc)
7755 {
7756 struct io_rsrc_put *prsrc;
7757
7758 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7759 if (!prsrc)
7760 return -ENOMEM;
7761
7762 prsrc->tag = data->tags[idx];
7763 prsrc->rsrc = rsrc;
7764 list_add(&prsrc->list, &node->rsrc_list);
7765 return 0;
7766 }
7767
__io_sqe_files_update(struct io_ring_ctx * ctx,struct io_uring_rsrc_update2 * up,unsigned nr_args)7768 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7769 struct io_uring_rsrc_update2 *up,
7770 unsigned nr_args)
7771 {
7772 u64 __user *tags = u64_to_user_ptr(up->tags);
7773 __s32 __user *fds = u64_to_user_ptr(up->data);
7774 struct io_rsrc_data *data = ctx->file_data;
7775 struct io_fixed_file *file_slot;
7776 struct file *file;
7777 int fd, i, err = 0;
7778 unsigned int done;
7779 bool needs_switch = false;
7780
7781 if (!ctx->file_data)
7782 return -ENXIO;
7783 if (up->offset + nr_args > ctx->nr_user_files)
7784 return -EINVAL;
7785
7786 for (done = 0; done < nr_args; done++) {
7787 u64 tag = 0;
7788
7789 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7790 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7791 err = -EFAULT;
7792 break;
7793 }
7794 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7795 err = -EINVAL;
7796 break;
7797 }
7798 if (fd == IORING_REGISTER_FILES_SKIP)
7799 continue;
7800
7801 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7802 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7803
7804 if (file_slot->file_ptr) {
7805 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7806 err = io_queue_rsrc_removal(data, up->offset + done,
7807 ctx->rsrc_node, file);
7808 if (err)
7809 break;
7810 file_slot->file_ptr = 0;
7811 needs_switch = true;
7812 }
7813 if (fd != -1) {
7814 file = fget(fd);
7815 if (!file) {
7816 err = -EBADF;
7817 break;
7818 }
7819 /*
7820 * Don't allow io_uring instances to be registered. If
7821 * UNIX isn't enabled, then this causes a reference
7822 * cycle and this instance can never get freed. If UNIX
7823 * is enabled we'll handle it just fine, but there's
7824 * still no point in allowing a ring fd as it doesn't
7825 * support regular read/write anyway.
7826 */
7827 if (file->f_op == &io_uring_fops) {
7828 fput(file);
7829 err = -EBADF;
7830 break;
7831 }
7832 data->tags[up->offset + done] = tag;
7833 io_fixed_file_set(file_slot, file);
7834 err = io_sqe_file_register(ctx, file, i);
7835 if (err) {
7836 file_slot->file_ptr = 0;
7837 fput(file);
7838 break;
7839 }
7840 }
7841 }
7842
7843 if (needs_switch)
7844 io_rsrc_node_switch(ctx, data);
7845 return done ? done : err;
7846 }
7847
io_free_work(struct io_wq_work * work)7848 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7849 {
7850 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7851
7852 req = io_put_req_find_next(req);
7853 return req ? &req->work : NULL;
7854 }
7855
io_init_wq_offload(struct io_ring_ctx * ctx,struct task_struct * task)7856 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7857 struct task_struct *task)
7858 {
7859 struct io_wq_hash *hash;
7860 struct io_wq_data data;
7861 unsigned int concurrency;
7862
7863 hash = ctx->hash_map;
7864 if (!hash) {
7865 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7866 if (!hash)
7867 return ERR_PTR(-ENOMEM);
7868 refcount_set(&hash->refs, 1);
7869 init_waitqueue_head(&hash->wait);
7870 ctx->hash_map = hash;
7871 }
7872
7873 data.hash = hash;
7874 data.task = task;
7875 data.free_work = io_free_work;
7876 data.do_work = io_wq_submit_work;
7877
7878 /* Do QD, or 4 * CPUS, whatever is smallest */
7879 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7880
7881 return io_wq_create(concurrency, &data);
7882 }
7883
io_uring_alloc_task_context(struct task_struct * task,struct io_ring_ctx * ctx)7884 static int io_uring_alloc_task_context(struct task_struct *task,
7885 struct io_ring_ctx *ctx)
7886 {
7887 struct io_uring_task *tctx;
7888 int ret;
7889
7890 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7891 if (unlikely(!tctx))
7892 return -ENOMEM;
7893
7894 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7895 if (unlikely(ret)) {
7896 kfree(tctx);
7897 return ret;
7898 }
7899
7900 tctx->io_wq = io_init_wq_offload(ctx, task);
7901 if (IS_ERR(tctx->io_wq)) {
7902 ret = PTR_ERR(tctx->io_wq);
7903 percpu_counter_destroy(&tctx->inflight);
7904 kfree(tctx);
7905 return ret;
7906 }
7907
7908 xa_init(&tctx->xa);
7909 init_waitqueue_head(&tctx->wait);
7910 tctx->last = NULL;
7911 atomic_set(&tctx->in_idle, 0);
7912 atomic_set(&tctx->inflight_tracked, 0);
7913 task->io_uring = tctx;
7914 spin_lock_init(&tctx->task_lock);
7915 INIT_WQ_LIST(&tctx->task_list);
7916 tctx->task_state = 0;
7917 init_task_work(&tctx->task_work, tctx_task_work);
7918 return 0;
7919 }
7920
__io_uring_free(struct task_struct * tsk)7921 void __io_uring_free(struct task_struct *tsk)
7922 {
7923 struct io_uring_task *tctx = tsk->io_uring;
7924
7925 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7926 WARN_ON_ONCE(tctx->io_wq);
7927
7928 percpu_counter_destroy(&tctx->inflight);
7929 kfree(tctx);
7930 tsk->io_uring = NULL;
7931 }
7932
io_sq_offload_create(struct io_ring_ctx * ctx,struct io_uring_params * p)7933 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7934 struct io_uring_params *p)
7935 {
7936 int ret;
7937
7938 /* Retain compatibility with failing for an invalid attach attempt */
7939 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7940 IORING_SETUP_ATTACH_WQ) {
7941 struct fd f;
7942
7943 f = fdget(p->wq_fd);
7944 if (!f.file)
7945 return -ENXIO;
7946 fdput(f);
7947 if (f.file->f_op != &io_uring_fops)
7948 return -EINVAL;
7949 }
7950 if (ctx->flags & IORING_SETUP_SQPOLL) {
7951 struct task_struct *tsk;
7952 struct io_sq_data *sqd;
7953 bool attached;
7954
7955 sqd = io_get_sq_data(p, &attached);
7956 if (IS_ERR(sqd)) {
7957 ret = PTR_ERR(sqd);
7958 goto err;
7959 }
7960
7961 ctx->sq_creds = get_current_cred();
7962 ctx->sq_data = sqd;
7963 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7964 if (!ctx->sq_thread_idle)
7965 ctx->sq_thread_idle = HZ;
7966
7967 io_sq_thread_park(sqd);
7968 list_add(&ctx->sqd_list, &sqd->ctx_list);
7969 io_sqd_update_thread_idle(sqd);
7970 /* don't attach to a dying SQPOLL thread, would be racy */
7971 ret = (attached && !sqd->thread) ? -ENXIO : 0;
7972 io_sq_thread_unpark(sqd);
7973
7974 if (ret < 0)
7975 goto err;
7976 if (attached)
7977 return 0;
7978
7979 if (p->flags & IORING_SETUP_SQ_AFF) {
7980 int cpu = p->sq_thread_cpu;
7981
7982 ret = -EINVAL;
7983 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
7984 goto err_sqpoll;
7985 sqd->sq_cpu = cpu;
7986 } else {
7987 sqd->sq_cpu = -1;
7988 }
7989
7990 sqd->task_pid = current->pid;
7991 sqd->task_tgid = current->tgid;
7992 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7993 if (IS_ERR(tsk)) {
7994 ret = PTR_ERR(tsk);
7995 goto err_sqpoll;
7996 }
7997
7998 sqd->thread = tsk;
7999 ret = io_uring_alloc_task_context(tsk, ctx);
8000 wake_up_new_task(tsk);
8001 if (ret)
8002 goto err;
8003 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8004 /* Can't have SQ_AFF without SQPOLL */
8005 ret = -EINVAL;
8006 goto err;
8007 }
8008
8009 return 0;
8010 err_sqpoll:
8011 complete(&ctx->sq_data->exited);
8012 err:
8013 io_sq_thread_finish(ctx);
8014 return ret;
8015 }
8016
__io_unaccount_mem(struct user_struct * user,unsigned long nr_pages)8017 static inline void __io_unaccount_mem(struct user_struct *user,
8018 unsigned long nr_pages)
8019 {
8020 atomic_long_sub(nr_pages, &user->locked_vm);
8021 }
8022
__io_account_mem(struct user_struct * user,unsigned long nr_pages)8023 static inline int __io_account_mem(struct user_struct *user,
8024 unsigned long nr_pages)
8025 {
8026 unsigned long page_limit, cur_pages, new_pages;
8027
8028 /* Don't allow more pages than we can safely lock */
8029 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8030
8031 do {
8032 cur_pages = atomic_long_read(&user->locked_vm);
8033 new_pages = cur_pages + nr_pages;
8034 if (new_pages > page_limit)
8035 return -ENOMEM;
8036 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8037 new_pages) != cur_pages);
8038
8039 return 0;
8040 }
8041
io_unaccount_mem(struct io_ring_ctx * ctx,unsigned long nr_pages)8042 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8043 {
8044 if (ctx->user)
8045 __io_unaccount_mem(ctx->user, nr_pages);
8046
8047 if (ctx->mm_account)
8048 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8049 }
8050
io_account_mem(struct io_ring_ctx * ctx,unsigned long nr_pages)8051 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8052 {
8053 int ret;
8054
8055 if (ctx->user) {
8056 ret = __io_account_mem(ctx->user, nr_pages);
8057 if (ret)
8058 return ret;
8059 }
8060
8061 if (ctx->mm_account)
8062 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8063
8064 return 0;
8065 }
8066
io_mem_free(void * ptr)8067 static void io_mem_free(void *ptr)
8068 {
8069 struct page *page;
8070
8071 if (!ptr)
8072 return;
8073
8074 page = virt_to_head_page(ptr);
8075 if (put_page_testzero(page))
8076 free_compound_page(page);
8077 }
8078
io_mem_alloc(size_t size)8079 static void *io_mem_alloc(size_t size)
8080 {
8081 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8082 __GFP_NORETRY | __GFP_ACCOUNT;
8083
8084 return (void *) __get_free_pages(gfp_flags, get_order(size));
8085 }
8086
rings_size(unsigned sq_entries,unsigned cq_entries,size_t * sq_offset)8087 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8088 size_t *sq_offset)
8089 {
8090 struct io_rings *rings;
8091 size_t off, sq_array_size;
8092
8093 off = struct_size(rings, cqes, cq_entries);
8094 if (off == SIZE_MAX)
8095 return SIZE_MAX;
8096
8097 #ifdef CONFIG_SMP
8098 off = ALIGN(off, SMP_CACHE_BYTES);
8099 if (off == 0)
8100 return SIZE_MAX;
8101 #endif
8102
8103 if (sq_offset)
8104 *sq_offset = off;
8105
8106 sq_array_size = array_size(sizeof(u32), sq_entries);
8107 if (sq_array_size == SIZE_MAX)
8108 return SIZE_MAX;
8109
8110 if (check_add_overflow(off, sq_array_size, &off))
8111 return SIZE_MAX;
8112
8113 return off;
8114 }
8115
io_buffer_unmap(struct io_ring_ctx * ctx,struct io_mapped_ubuf ** slot)8116 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8117 {
8118 struct io_mapped_ubuf *imu = *slot;
8119 unsigned int i;
8120
8121 if (imu != ctx->dummy_ubuf) {
8122 for (i = 0; i < imu->nr_bvecs; i++)
8123 unpin_user_page(imu->bvec[i].bv_page);
8124 if (imu->acct_pages)
8125 io_unaccount_mem(ctx, imu->acct_pages);
8126 kvfree(imu);
8127 }
8128 *slot = NULL;
8129 }
8130
io_rsrc_buf_put(struct io_ring_ctx * ctx,struct io_rsrc_put * prsrc)8131 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8132 {
8133 io_buffer_unmap(ctx, &prsrc->buf);
8134 prsrc->buf = NULL;
8135 }
8136
__io_sqe_buffers_unregister(struct io_ring_ctx * ctx)8137 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8138 {
8139 unsigned int i;
8140
8141 for (i = 0; i < ctx->nr_user_bufs; i++)
8142 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8143 kfree(ctx->user_bufs);
8144 io_rsrc_data_free(ctx->buf_data);
8145 ctx->user_bufs = NULL;
8146 ctx->buf_data = NULL;
8147 ctx->nr_user_bufs = 0;
8148 }
8149
io_sqe_buffers_unregister(struct io_ring_ctx * ctx)8150 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8151 {
8152 int ret;
8153
8154 if (!ctx->buf_data)
8155 return -ENXIO;
8156
8157 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8158 if (!ret)
8159 __io_sqe_buffers_unregister(ctx);
8160 return ret;
8161 }
8162
io_copy_iov(struct io_ring_ctx * ctx,struct iovec * dst,void __user * arg,unsigned index)8163 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8164 void __user *arg, unsigned index)
8165 {
8166 struct iovec __user *src;
8167
8168 #ifdef CONFIG_COMPAT
8169 if (ctx->compat) {
8170 struct compat_iovec __user *ciovs;
8171 struct compat_iovec ciov;
8172
8173 ciovs = (struct compat_iovec __user *) arg;
8174 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8175 return -EFAULT;
8176
8177 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8178 dst->iov_len = ciov.iov_len;
8179 return 0;
8180 }
8181 #endif
8182 src = (struct iovec __user *) arg;
8183 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8184 return -EFAULT;
8185 return 0;
8186 }
8187
8188 /*
8189 * Not super efficient, but this is just a registration time. And we do cache
8190 * the last compound head, so generally we'll only do a full search if we don't
8191 * match that one.
8192 *
8193 * We check if the given compound head page has already been accounted, to
8194 * avoid double accounting it. This allows us to account the full size of the
8195 * page, not just the constituent pages of a huge page.
8196 */
headpage_already_acct(struct io_ring_ctx * ctx,struct page ** pages,int nr_pages,struct page * hpage)8197 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8198 int nr_pages, struct page *hpage)
8199 {
8200 int i, j;
8201
8202 /* check current page array */
8203 for (i = 0; i < nr_pages; i++) {
8204 if (!PageCompound(pages[i]))
8205 continue;
8206 if (compound_head(pages[i]) == hpage)
8207 return true;
8208 }
8209
8210 /* check previously registered pages */
8211 for (i = 0; i < ctx->nr_user_bufs; i++) {
8212 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8213
8214 for (j = 0; j < imu->nr_bvecs; j++) {
8215 if (!PageCompound(imu->bvec[j].bv_page))
8216 continue;
8217 if (compound_head(imu->bvec[j].bv_page) == hpage)
8218 return true;
8219 }
8220 }
8221
8222 return false;
8223 }
8224
io_buffer_account_pin(struct io_ring_ctx * ctx,struct page ** pages,int nr_pages,struct io_mapped_ubuf * imu,struct page ** last_hpage)8225 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8226 int nr_pages, struct io_mapped_ubuf *imu,
8227 struct page **last_hpage)
8228 {
8229 int i, ret;
8230
8231 for (i = 0; i < nr_pages; i++) {
8232 if (!PageCompound(pages[i])) {
8233 imu->acct_pages++;
8234 } else {
8235 struct page *hpage;
8236
8237 hpage = compound_head(pages[i]);
8238 if (hpage == *last_hpage)
8239 continue;
8240 *last_hpage = hpage;
8241 if (headpage_already_acct(ctx, pages, i, hpage))
8242 continue;
8243 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8244 }
8245 }
8246
8247 if (!imu->acct_pages)
8248 return 0;
8249
8250 ret = io_account_mem(ctx, imu->acct_pages);
8251 if (ret)
8252 imu->acct_pages = 0;
8253 return ret;
8254 }
8255
io_sqe_buffer_register(struct io_ring_ctx * ctx,struct iovec * iov,struct io_mapped_ubuf ** pimu,struct page ** last_hpage)8256 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8257 struct io_mapped_ubuf **pimu,
8258 struct page **last_hpage)
8259 {
8260 struct io_mapped_ubuf *imu = NULL;
8261 struct vm_area_struct **vmas = NULL;
8262 struct page **pages = NULL;
8263 unsigned long off, start, end, ubuf;
8264 size_t size;
8265 int ret, pret, nr_pages, i;
8266
8267 if (!iov->iov_base) {
8268 *pimu = ctx->dummy_ubuf;
8269 return 0;
8270 }
8271
8272 ubuf = (unsigned long) iov->iov_base;
8273 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8274 start = ubuf >> PAGE_SHIFT;
8275 nr_pages = end - start;
8276
8277 *pimu = NULL;
8278 ret = -ENOMEM;
8279
8280 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8281 if (!pages)
8282 goto done;
8283
8284 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8285 GFP_KERNEL);
8286 if (!vmas)
8287 goto done;
8288
8289 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8290 if (!imu)
8291 goto done;
8292
8293 ret = 0;
8294 mmap_read_lock(current->mm);
8295 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8296 pages, vmas);
8297 if (pret == nr_pages) {
8298 /* don't support file backed memory */
8299 for (i = 0; i < nr_pages; i++) {
8300 struct vm_area_struct *vma = vmas[i];
8301
8302 if (vma->vm_file &&
8303 !is_file_hugepages(vma->vm_file)) {
8304 ret = -EOPNOTSUPP;
8305 break;
8306 }
8307 }
8308 } else {
8309 ret = pret < 0 ? pret : -EFAULT;
8310 }
8311 mmap_read_unlock(current->mm);
8312 if (ret) {
8313 /*
8314 * if we did partial map, or found file backed vmas,
8315 * release any pages we did get
8316 */
8317 if (pret > 0)
8318 unpin_user_pages(pages, pret);
8319 goto done;
8320 }
8321
8322 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8323 if (ret) {
8324 unpin_user_pages(pages, pret);
8325 goto done;
8326 }
8327
8328 off = ubuf & ~PAGE_MASK;
8329 size = iov->iov_len;
8330 for (i = 0; i < nr_pages; i++) {
8331 size_t vec_len;
8332
8333 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8334 imu->bvec[i].bv_page = pages[i];
8335 imu->bvec[i].bv_len = vec_len;
8336 imu->bvec[i].bv_offset = off;
8337 off = 0;
8338 size -= vec_len;
8339 }
8340 /* store original address for later verification */
8341 imu->ubuf = ubuf;
8342 imu->ubuf_end = ubuf + iov->iov_len;
8343 imu->nr_bvecs = nr_pages;
8344 *pimu = imu;
8345 ret = 0;
8346 done:
8347 if (ret)
8348 kvfree(imu);
8349 kvfree(pages);
8350 kvfree(vmas);
8351 return ret;
8352 }
8353
io_buffers_map_alloc(struct io_ring_ctx * ctx,unsigned int nr_args)8354 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8355 {
8356 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8357 return ctx->user_bufs ? 0 : -ENOMEM;
8358 }
8359
io_buffer_validate(struct iovec * iov)8360 static int io_buffer_validate(struct iovec *iov)
8361 {
8362 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8363
8364 /*
8365 * Don't impose further limits on the size and buffer
8366 * constraints here, we'll -EINVAL later when IO is
8367 * submitted if they are wrong.
8368 */
8369 if (!iov->iov_base)
8370 return iov->iov_len ? -EFAULT : 0;
8371 if (!iov->iov_len)
8372 return -EFAULT;
8373
8374 /* arbitrary limit, but we need something */
8375 if (iov->iov_len > SZ_1G)
8376 return -EFAULT;
8377
8378 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8379 return -EOVERFLOW;
8380
8381 return 0;
8382 }
8383
io_sqe_buffers_register(struct io_ring_ctx * ctx,void __user * arg,unsigned int nr_args,u64 __user * tags)8384 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8385 unsigned int nr_args, u64 __user *tags)
8386 {
8387 struct page *last_hpage = NULL;
8388 struct io_rsrc_data *data;
8389 int i, ret;
8390 struct iovec iov;
8391
8392 if (ctx->user_bufs)
8393 return -EBUSY;
8394 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8395 return -EINVAL;
8396 ret = io_rsrc_node_switch_start(ctx);
8397 if (ret)
8398 return ret;
8399 data = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, nr_args);
8400 if (!data)
8401 return -ENOMEM;
8402 ret = io_buffers_map_alloc(ctx, nr_args);
8403 if (ret) {
8404 io_rsrc_data_free(data);
8405 return ret;
8406 }
8407
8408 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8409 u64 tag = 0;
8410
8411 if (tags && copy_from_user(&tag, &tags[i], sizeof(tag))) {
8412 ret = -EFAULT;
8413 break;
8414 }
8415 ret = io_copy_iov(ctx, &iov, arg, i);
8416 if (ret)
8417 break;
8418 ret = io_buffer_validate(&iov);
8419 if (ret)
8420 break;
8421 if (!iov.iov_base && tag) {
8422 ret = -EINVAL;
8423 break;
8424 }
8425
8426 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8427 &last_hpage);
8428 if (ret)
8429 break;
8430 data->tags[i] = tag;
8431 }
8432
8433 WARN_ON_ONCE(ctx->buf_data);
8434
8435 ctx->buf_data = data;
8436 if (ret)
8437 __io_sqe_buffers_unregister(ctx);
8438 else
8439 io_rsrc_node_switch(ctx, NULL);
8440 return ret;
8441 }
8442
__io_sqe_buffers_update(struct io_ring_ctx * ctx,struct io_uring_rsrc_update2 * up,unsigned int nr_args)8443 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8444 struct io_uring_rsrc_update2 *up,
8445 unsigned int nr_args)
8446 {
8447 u64 __user *tags = u64_to_user_ptr(up->tags);
8448 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8449 struct page *last_hpage = NULL;
8450 bool needs_switch = false;
8451 __u32 done;
8452 int i, err;
8453
8454 if (!ctx->buf_data)
8455 return -ENXIO;
8456 if (up->offset + nr_args > ctx->nr_user_bufs)
8457 return -EINVAL;
8458
8459 for (done = 0; done < nr_args; done++) {
8460 struct io_mapped_ubuf *imu;
8461 int offset = up->offset + done;
8462 u64 tag = 0;
8463
8464 err = io_copy_iov(ctx, &iov, iovs, done);
8465 if (err)
8466 break;
8467 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8468 err = -EFAULT;
8469 break;
8470 }
8471 err = io_buffer_validate(&iov);
8472 if (err)
8473 break;
8474 if (!iov.iov_base && tag) {
8475 err = -EINVAL;
8476 break;
8477 }
8478 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8479 if (err)
8480 break;
8481
8482 i = array_index_nospec(offset, ctx->nr_user_bufs);
8483 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8484 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8485 ctx->rsrc_node, ctx->user_bufs[i]);
8486 if (unlikely(err)) {
8487 io_buffer_unmap(ctx, &imu);
8488 break;
8489 }
8490 ctx->user_bufs[i] = NULL;
8491 needs_switch = true;
8492 }
8493
8494 ctx->user_bufs[i] = imu;
8495 ctx->buf_data->tags[offset] = tag;
8496 }
8497
8498 if (needs_switch)
8499 io_rsrc_node_switch(ctx, ctx->buf_data);
8500 return done ? done : err;
8501 }
8502
io_eventfd_register(struct io_ring_ctx * ctx,void __user * arg)8503 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8504 {
8505 __s32 __user *fds = arg;
8506 int fd;
8507
8508 if (ctx->cq_ev_fd)
8509 return -EBUSY;
8510
8511 if (copy_from_user(&fd, fds, sizeof(*fds)))
8512 return -EFAULT;
8513
8514 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8515 if (IS_ERR(ctx->cq_ev_fd)) {
8516 int ret = PTR_ERR(ctx->cq_ev_fd);
8517 ctx->cq_ev_fd = NULL;
8518 return ret;
8519 }
8520
8521 return 0;
8522 }
8523
io_eventfd_unregister(struct io_ring_ctx * ctx)8524 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8525 {
8526 if (ctx->cq_ev_fd) {
8527 eventfd_ctx_put(ctx->cq_ev_fd);
8528 ctx->cq_ev_fd = NULL;
8529 return 0;
8530 }
8531
8532 return -ENXIO;
8533 }
8534
io_destroy_buffers(struct io_ring_ctx * ctx)8535 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8536 {
8537 struct io_buffer *buf;
8538 unsigned long index;
8539
8540 xa_for_each(&ctx->io_buffers, index, buf)
8541 __io_remove_buffers(ctx, buf, index, -1U);
8542 }
8543
io_req_cache_free(struct list_head * list,struct task_struct * tsk)8544 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8545 {
8546 struct io_kiocb *req, *nxt;
8547
8548 list_for_each_entry_safe(req, nxt, list, compl.list) {
8549 if (tsk && req->task != tsk)
8550 continue;
8551 list_del(&req->compl.list);
8552 kmem_cache_free(req_cachep, req);
8553 }
8554 }
8555
io_req_caches_free(struct io_ring_ctx * ctx)8556 static void io_req_caches_free(struct io_ring_ctx *ctx)
8557 {
8558 struct io_submit_state *submit_state = &ctx->submit_state;
8559 struct io_comp_state *cs = &ctx->submit_state.comp;
8560
8561 mutex_lock(&ctx->uring_lock);
8562
8563 if (submit_state->free_reqs) {
8564 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8565 submit_state->reqs);
8566 submit_state->free_reqs = 0;
8567 }
8568
8569 io_flush_cached_locked_reqs(ctx, cs);
8570 io_req_cache_free(&cs->free_list, NULL);
8571 mutex_unlock(&ctx->uring_lock);
8572 }
8573
io_wait_rsrc_data(struct io_rsrc_data * data)8574 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8575 {
8576 if (!data)
8577 return false;
8578 if (!atomic_dec_and_test(&data->refs))
8579 wait_for_completion(&data->done);
8580 return true;
8581 }
8582
io_ring_ctx_free(struct io_ring_ctx * ctx)8583 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8584 {
8585 io_sq_thread_finish(ctx);
8586
8587 if (ctx->mm_account) {
8588 mmdrop(ctx->mm_account);
8589 ctx->mm_account = NULL;
8590 }
8591
8592 mutex_lock(&ctx->uring_lock);
8593 if (io_wait_rsrc_data(ctx->buf_data))
8594 __io_sqe_buffers_unregister(ctx);
8595 if (io_wait_rsrc_data(ctx->file_data))
8596 __io_sqe_files_unregister(ctx);
8597 if (ctx->rings)
8598 __io_cqring_overflow_flush(ctx, true);
8599 mutex_unlock(&ctx->uring_lock);
8600 io_eventfd_unregister(ctx);
8601 io_destroy_buffers(ctx);
8602 if (ctx->sq_creds)
8603 put_cred(ctx->sq_creds);
8604
8605 /* there are no registered resources left, nobody uses it */
8606 if (ctx->rsrc_node)
8607 io_rsrc_node_destroy(ctx->rsrc_node);
8608 if (ctx->rsrc_backup_node)
8609 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8610 flush_delayed_work(&ctx->rsrc_put_work);
8611
8612 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8613 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8614
8615 #if defined(CONFIG_UNIX)
8616 if (ctx->ring_sock) {
8617 ctx->ring_sock->file = NULL; /* so that iput() is called */
8618 sock_release(ctx->ring_sock);
8619 }
8620 #endif
8621
8622 io_mem_free(ctx->rings);
8623 io_mem_free(ctx->sq_sqes);
8624
8625 percpu_ref_exit(&ctx->refs);
8626 free_uid(ctx->user);
8627 io_req_caches_free(ctx);
8628 if (ctx->hash_map)
8629 io_wq_put_hash(ctx->hash_map);
8630 kfree(ctx->cancel_hash);
8631 kfree(ctx->dummy_ubuf);
8632 kfree(ctx);
8633 }
8634
io_uring_poll(struct file * file,poll_table * wait)8635 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8636 {
8637 struct io_ring_ctx *ctx = file->private_data;
8638 __poll_t mask = 0;
8639
8640 poll_wait(file, &ctx->cq_wait, wait);
8641 /*
8642 * synchronizes with barrier from wq_has_sleeper call in
8643 * io_commit_cqring
8644 */
8645 smp_rmb();
8646 if (!io_sqring_full(ctx))
8647 mask |= EPOLLOUT | EPOLLWRNORM;
8648
8649 /*
8650 * Don't flush cqring overflow list here, just do a simple check.
8651 * Otherwise there could possible be ABBA deadlock:
8652 * CPU0 CPU1
8653 * ---- ----
8654 * lock(&ctx->uring_lock);
8655 * lock(&ep->mtx);
8656 * lock(&ctx->uring_lock);
8657 * lock(&ep->mtx);
8658 *
8659 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8660 * pushs them to do the flush.
8661 */
8662 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8663 mask |= EPOLLIN | EPOLLRDNORM;
8664
8665 return mask;
8666 }
8667
io_uring_fasync(int fd,struct file * file,int on)8668 static int io_uring_fasync(int fd, struct file *file, int on)
8669 {
8670 struct io_ring_ctx *ctx = file->private_data;
8671
8672 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8673 }
8674
io_unregister_personality(struct io_ring_ctx * ctx,unsigned id)8675 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8676 {
8677 const struct cred *creds;
8678
8679 creds = xa_erase(&ctx->personalities, id);
8680 if (creds) {
8681 put_cred(creds);
8682 return 0;
8683 }
8684
8685 return -EINVAL;
8686 }
8687
io_run_ctx_fallback(struct io_ring_ctx * ctx)8688 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8689 {
8690 return io_run_task_work_head(&ctx->exit_task_work);
8691 }
8692
8693 struct io_tctx_exit {
8694 struct callback_head task_work;
8695 struct completion completion;
8696 struct io_ring_ctx *ctx;
8697 };
8698
io_tctx_exit_cb(struct callback_head * cb)8699 static void io_tctx_exit_cb(struct callback_head *cb)
8700 {
8701 struct io_uring_task *tctx = current->io_uring;
8702 struct io_tctx_exit *work;
8703
8704 work = container_of(cb, struct io_tctx_exit, task_work);
8705 /*
8706 * When @in_idle, we're in cancellation and it's racy to remove the
8707 * node. It'll be removed by the end of cancellation, just ignore it.
8708 */
8709 if (!atomic_read(&tctx->in_idle))
8710 io_uring_del_task_file((unsigned long)work->ctx);
8711 complete(&work->completion);
8712 }
8713
io_cancel_ctx_cb(struct io_wq_work * work,void * data)8714 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8715 {
8716 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8717
8718 return req->ctx == data;
8719 }
8720
io_ring_exit_work(struct work_struct * work)8721 static void io_ring_exit_work(struct work_struct *work)
8722 {
8723 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8724 unsigned long timeout = jiffies + HZ * 60 * 5;
8725 struct io_tctx_exit exit;
8726 struct io_tctx_node *node;
8727 int ret;
8728
8729 /*
8730 * If we're doing polled IO and end up having requests being
8731 * submitted async (out-of-line), then completions can come in while
8732 * we're waiting for refs to drop. We need to reap these manually,
8733 * as nobody else will be looking for them.
8734 */
8735 do {
8736 io_uring_try_cancel_requests(ctx, NULL, NULL);
8737 if (ctx->sq_data) {
8738 struct io_sq_data *sqd = ctx->sq_data;
8739 struct task_struct *tsk;
8740
8741 io_sq_thread_park(sqd);
8742 tsk = sqd->thread;
8743 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8744 io_wq_cancel_cb(tsk->io_uring->io_wq,
8745 io_cancel_ctx_cb, ctx, true);
8746 io_sq_thread_unpark(sqd);
8747 }
8748
8749 WARN_ON_ONCE(time_after(jiffies, timeout));
8750 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8751
8752 init_completion(&exit.completion);
8753 init_task_work(&exit.task_work, io_tctx_exit_cb);
8754 exit.ctx = ctx;
8755 /*
8756 * Some may use context even when all refs and requests have been put,
8757 * and they are free to do so while still holding uring_lock or
8758 * completion_lock, see __io_req_task_submit(). Apart from other work,
8759 * this lock/unlock section also waits them to finish.
8760 */
8761 mutex_lock(&ctx->uring_lock);
8762 while (!list_empty(&ctx->tctx_list)) {
8763 WARN_ON_ONCE(time_after(jiffies, timeout));
8764
8765 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8766 ctx_node);
8767 /* don't spin on a single task if cancellation failed */
8768 list_rotate_left(&ctx->tctx_list);
8769 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8770 if (WARN_ON_ONCE(ret))
8771 continue;
8772 wake_up_process(node->task);
8773
8774 mutex_unlock(&ctx->uring_lock);
8775 wait_for_completion(&exit.completion);
8776 mutex_lock(&ctx->uring_lock);
8777 }
8778 mutex_unlock(&ctx->uring_lock);
8779 spin_lock_irq(&ctx->completion_lock);
8780 spin_unlock_irq(&ctx->completion_lock);
8781
8782 io_ring_ctx_free(ctx);
8783 }
8784
8785 /* Returns true if we found and killed one or more timeouts */
io_kill_timeouts(struct io_ring_ctx * ctx,struct task_struct * tsk,struct files_struct * files)8786 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8787 struct files_struct *files)
8788 {
8789 struct io_kiocb *req, *tmp;
8790 int canceled = 0;
8791
8792 spin_lock_irq(&ctx->completion_lock);
8793 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8794 if (io_match_task(req, tsk, files)) {
8795 io_kill_timeout(req, -ECANCELED);
8796 canceled++;
8797 }
8798 }
8799 if (canceled != 0)
8800 io_commit_cqring(ctx);
8801 spin_unlock_irq(&ctx->completion_lock);
8802 if (canceled != 0)
8803 io_cqring_ev_posted(ctx);
8804 return canceled != 0;
8805 }
8806
io_ring_ctx_wait_and_kill(struct io_ring_ctx * ctx)8807 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8808 {
8809 unsigned long index;
8810 struct creds *creds;
8811
8812 mutex_lock(&ctx->uring_lock);
8813 percpu_ref_kill(&ctx->refs);
8814 if (ctx->rings)
8815 __io_cqring_overflow_flush(ctx, true);
8816 xa_for_each(&ctx->personalities, index, creds)
8817 io_unregister_personality(ctx, index);
8818 mutex_unlock(&ctx->uring_lock);
8819
8820 io_kill_timeouts(ctx, NULL, NULL);
8821 io_poll_remove_all(ctx, NULL, NULL);
8822
8823 /* if we failed setting up the ctx, we might not have any rings */
8824 io_iopoll_try_reap_events(ctx);
8825
8826 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8827 /*
8828 * Use system_unbound_wq to avoid spawning tons of event kworkers
8829 * if we're exiting a ton of rings at the same time. It just adds
8830 * noise and overhead, there's no discernable change in runtime
8831 * over using system_wq.
8832 */
8833 queue_work(system_unbound_wq, &ctx->exit_work);
8834 }
8835
io_uring_release(struct inode * inode,struct file * file)8836 static int io_uring_release(struct inode *inode, struct file *file)
8837 {
8838 struct io_ring_ctx *ctx = file->private_data;
8839
8840 file->private_data = NULL;
8841 io_ring_ctx_wait_and_kill(ctx);
8842 return 0;
8843 }
8844
8845 struct io_task_cancel {
8846 struct task_struct *task;
8847 struct files_struct *files;
8848 };
8849
io_cancel_task_cb(struct io_wq_work * work,void * data)8850 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8851 {
8852 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8853 struct io_task_cancel *cancel = data;
8854 bool ret;
8855
8856 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8857 unsigned long flags;
8858 struct io_ring_ctx *ctx = req->ctx;
8859
8860 /* protect against races with linked timeouts */
8861 spin_lock_irqsave(&ctx->completion_lock, flags);
8862 ret = io_match_task(req, cancel->task, cancel->files);
8863 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8864 } else {
8865 ret = io_match_task(req, cancel->task, cancel->files);
8866 }
8867 return ret;
8868 }
8869
io_cancel_defer_files(struct io_ring_ctx * ctx,struct task_struct * task,struct files_struct * files)8870 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8871 struct task_struct *task,
8872 struct files_struct *files)
8873 {
8874 struct io_defer_entry *de;
8875 LIST_HEAD(list);
8876
8877 spin_lock_irq(&ctx->completion_lock);
8878 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8879 if (io_match_task(de->req, task, files)) {
8880 list_cut_position(&list, &ctx->defer_list, &de->list);
8881 break;
8882 }
8883 }
8884 spin_unlock_irq(&ctx->completion_lock);
8885 if (list_empty(&list))
8886 return false;
8887
8888 while (!list_empty(&list)) {
8889 de = list_first_entry(&list, struct io_defer_entry, list);
8890 list_del_init(&de->list);
8891 io_req_complete_failed(de->req, -ECANCELED);
8892 kfree(de);
8893 }
8894 return true;
8895 }
8896
io_uring_try_cancel_iowq(struct io_ring_ctx * ctx)8897 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8898 {
8899 struct io_tctx_node *node;
8900 enum io_wq_cancel cret;
8901 bool ret = false;
8902
8903 mutex_lock(&ctx->uring_lock);
8904 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8905 struct io_uring_task *tctx = node->task->io_uring;
8906
8907 /*
8908 * io_wq will stay alive while we hold uring_lock, because it's
8909 * killed after ctx nodes, which requires to take the lock.
8910 */
8911 if (!tctx || !tctx->io_wq)
8912 continue;
8913 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8914 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8915 }
8916 mutex_unlock(&ctx->uring_lock);
8917
8918 return ret;
8919 }
8920
io_uring_try_cancel_requests(struct io_ring_ctx * ctx,struct task_struct * task,struct files_struct * files)8921 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8922 struct task_struct *task,
8923 struct files_struct *files)
8924 {
8925 struct io_task_cancel cancel = { .task = task, .files = files, };
8926 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8927
8928 while (1) {
8929 enum io_wq_cancel cret;
8930 bool ret = false;
8931
8932 if (!task) {
8933 ret |= io_uring_try_cancel_iowq(ctx);
8934 } else if (tctx && tctx->io_wq) {
8935 /*
8936 * Cancels requests of all rings, not only @ctx, but
8937 * it's fine as the task is in exit/exec.
8938 */
8939 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8940 &cancel, true);
8941 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8942 }
8943
8944 /* SQPOLL thread does its own polling */
8945 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8946 (ctx->sq_data && ctx->sq_data->thread == current)) {
8947 while (!list_empty_careful(&ctx->iopoll_list)) {
8948 io_iopoll_try_reap_events(ctx);
8949 ret = true;
8950 }
8951 }
8952
8953 ret |= io_cancel_defer_files(ctx, task, files);
8954 ret |= io_poll_remove_all(ctx, task, files);
8955 ret |= io_kill_timeouts(ctx, task, files);
8956 ret |= io_run_task_work();
8957 ret |= io_run_ctx_fallback(ctx);
8958 if (!ret)
8959 break;
8960 cond_resched();
8961 }
8962 }
8963
__io_uring_add_task_file(struct io_ring_ctx * ctx)8964 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8965 {
8966 struct io_uring_task *tctx = current->io_uring;
8967 struct io_tctx_node *node;
8968 int ret;
8969
8970 if (unlikely(!tctx)) {
8971 ret = io_uring_alloc_task_context(current, ctx);
8972 if (unlikely(ret))
8973 return ret;
8974 tctx = current->io_uring;
8975 }
8976 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8977 node = kmalloc(sizeof(*node), GFP_KERNEL);
8978 if (!node)
8979 return -ENOMEM;
8980 node->ctx = ctx;
8981 node->task = current;
8982
8983 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8984 node, GFP_KERNEL));
8985 if (ret) {
8986 kfree(node);
8987 return ret;
8988 }
8989
8990 mutex_lock(&ctx->uring_lock);
8991 list_add(&node->ctx_node, &ctx->tctx_list);
8992 mutex_unlock(&ctx->uring_lock);
8993 }
8994 tctx->last = ctx;
8995 return 0;
8996 }
8997
8998 /*
8999 * Note that this task has used io_uring. We use it for cancelation purposes.
9000 */
io_uring_add_task_file(struct io_ring_ctx * ctx)9001 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
9002 {
9003 struct io_uring_task *tctx = current->io_uring;
9004
9005 if (likely(tctx && tctx->last == ctx))
9006 return 0;
9007 return __io_uring_add_task_file(ctx);
9008 }
9009
9010 /*
9011 * Remove this io_uring_file -> task mapping.
9012 */
io_uring_del_task_file(unsigned long index)9013 static void io_uring_del_task_file(unsigned long index)
9014 {
9015 struct io_uring_task *tctx = current->io_uring;
9016 struct io_tctx_node *node;
9017
9018 if (!tctx)
9019 return;
9020 node = xa_erase(&tctx->xa, index);
9021 if (!node)
9022 return;
9023
9024 WARN_ON_ONCE(current != node->task);
9025 WARN_ON_ONCE(list_empty(&node->ctx_node));
9026
9027 mutex_lock(&node->ctx->uring_lock);
9028 list_del(&node->ctx_node);
9029 mutex_unlock(&node->ctx->uring_lock);
9030
9031 if (tctx->last == node->ctx)
9032 tctx->last = NULL;
9033 kfree(node);
9034 }
9035
io_uring_clean_tctx(struct io_uring_task * tctx)9036 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9037 {
9038 struct io_tctx_node *node;
9039 unsigned long index;
9040
9041 xa_for_each(&tctx->xa, index, node)
9042 io_uring_del_task_file(index);
9043 if (tctx->io_wq) {
9044 io_wq_put_and_exit(tctx->io_wq);
9045 tctx->io_wq = NULL;
9046 }
9047 }
9048
tctx_inflight(struct io_uring_task * tctx,bool tracked)9049 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9050 {
9051 if (tracked)
9052 return atomic_read(&tctx->inflight_tracked);
9053 return percpu_counter_sum(&tctx->inflight);
9054 }
9055
io_uring_try_cancel(struct files_struct * files)9056 static void io_uring_try_cancel(struct files_struct *files)
9057 {
9058 struct io_uring_task *tctx = current->io_uring;
9059 struct io_tctx_node *node;
9060 unsigned long index;
9061
9062 xa_for_each(&tctx->xa, index, node) {
9063 struct io_ring_ctx *ctx = node->ctx;
9064
9065 /* sqpoll task will cancel all its requests */
9066 if (!ctx->sq_data)
9067 io_uring_try_cancel_requests(ctx, current, files);
9068 }
9069 }
9070
9071 /* should only be called by SQPOLL task */
io_uring_cancel_sqpoll(struct io_sq_data * sqd)9072 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
9073 {
9074 struct io_uring_task *tctx = current->io_uring;
9075 struct io_ring_ctx *ctx;
9076 s64 inflight;
9077 DEFINE_WAIT(wait);
9078
9079 if (!current->io_uring)
9080 return;
9081 WARN_ON_ONCE(!sqd || sqd->thread != current);
9082
9083 atomic_inc(&tctx->in_idle);
9084 do {
9085 /* read completions before cancelations */
9086 inflight = tctx_inflight(tctx, false);
9087 if (!inflight)
9088 break;
9089 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9090 io_uring_try_cancel_requests(ctx, current, NULL);
9091
9092 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9093 /*
9094 * If we've seen completions, retry without waiting. This
9095 * avoids a race where a completion comes in before we did
9096 * prepare_to_wait().
9097 */
9098 if (inflight == tctx_inflight(tctx, false))
9099 schedule();
9100 finish_wait(&tctx->wait, &wait);
9101 } while (1);
9102 atomic_dec(&tctx->in_idle);
9103 }
9104
9105 /*
9106 * Find any io_uring fd that this task has registered or done IO on, and cancel
9107 * requests.
9108 */
__io_uring_cancel(struct files_struct * files)9109 void __io_uring_cancel(struct files_struct *files)
9110 {
9111 struct io_uring_task *tctx = current->io_uring;
9112 DEFINE_WAIT(wait);
9113 s64 inflight;
9114
9115 /* make sure overflow events are dropped */
9116 atomic_inc(&tctx->in_idle);
9117 do {
9118 /* read completions before cancelations */
9119 inflight = tctx_inflight(tctx, !!files);
9120 if (!inflight)
9121 break;
9122 io_uring_try_cancel(files);
9123 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9124
9125 /*
9126 * If we've seen completions, retry without waiting. This
9127 * avoids a race where a completion comes in before we did
9128 * prepare_to_wait().
9129 */
9130 if (inflight == tctx_inflight(tctx, !!files))
9131 schedule();
9132 finish_wait(&tctx->wait, &wait);
9133 } while (1);
9134 atomic_dec(&tctx->in_idle);
9135
9136 io_uring_clean_tctx(tctx);
9137 if (!files) {
9138 /* for exec all current's requests should be gone, kill tctx */
9139 __io_uring_free(current);
9140 }
9141 }
9142
io_uring_validate_mmap_request(struct file * file,loff_t pgoff,size_t sz)9143 static void *io_uring_validate_mmap_request(struct file *file,
9144 loff_t pgoff, size_t sz)
9145 {
9146 struct io_ring_ctx *ctx = file->private_data;
9147 loff_t offset = pgoff << PAGE_SHIFT;
9148 struct page *page;
9149 void *ptr;
9150
9151 switch (offset) {
9152 case IORING_OFF_SQ_RING:
9153 case IORING_OFF_CQ_RING:
9154 ptr = ctx->rings;
9155 break;
9156 case IORING_OFF_SQES:
9157 ptr = ctx->sq_sqes;
9158 break;
9159 default:
9160 return ERR_PTR(-EINVAL);
9161 }
9162
9163 page = virt_to_head_page(ptr);
9164 if (sz > page_size(page))
9165 return ERR_PTR(-EINVAL);
9166
9167 return ptr;
9168 }
9169
9170 #ifdef CONFIG_MMU
9171
io_uring_mmap(struct file * file,struct vm_area_struct * vma)9172 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9173 {
9174 size_t sz = vma->vm_end - vma->vm_start;
9175 unsigned long pfn;
9176 void *ptr;
9177
9178 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9179 if (IS_ERR(ptr))
9180 return PTR_ERR(ptr);
9181
9182 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9183 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9184 }
9185
9186 #else /* !CONFIG_MMU */
9187
io_uring_mmap(struct file * file,struct vm_area_struct * vma)9188 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9189 {
9190 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9191 }
9192
io_uring_nommu_mmap_capabilities(struct file * file)9193 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9194 {
9195 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9196 }
9197
io_uring_nommu_get_unmapped_area(struct file * file,unsigned long addr,unsigned long len,unsigned long pgoff,unsigned long flags)9198 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9199 unsigned long addr, unsigned long len,
9200 unsigned long pgoff, unsigned long flags)
9201 {
9202 void *ptr;
9203
9204 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9205 if (IS_ERR(ptr))
9206 return PTR_ERR(ptr);
9207
9208 return (unsigned long) ptr;
9209 }
9210
9211 #endif /* !CONFIG_MMU */
9212
io_sqpoll_wait_sq(struct io_ring_ctx * ctx)9213 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9214 {
9215 DEFINE_WAIT(wait);
9216
9217 do {
9218 if (!io_sqring_full(ctx))
9219 break;
9220 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9221
9222 if (!io_sqring_full(ctx))
9223 break;
9224 schedule();
9225 } while (!signal_pending(current));
9226
9227 finish_wait(&ctx->sqo_sq_wait, &wait);
9228 return 0;
9229 }
9230
io_get_ext_arg(unsigned flags,const void __user * argp,size_t * argsz,struct __kernel_timespec __user ** ts,const sigset_t __user ** sig)9231 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9232 struct __kernel_timespec __user **ts,
9233 const sigset_t __user **sig)
9234 {
9235 struct io_uring_getevents_arg arg;
9236
9237 /*
9238 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9239 * is just a pointer to the sigset_t.
9240 */
9241 if (!(flags & IORING_ENTER_EXT_ARG)) {
9242 *sig = (const sigset_t __user *) argp;
9243 *ts = NULL;
9244 return 0;
9245 }
9246
9247 /*
9248 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9249 * timespec and sigset_t pointers if good.
9250 */
9251 if (*argsz != sizeof(arg))
9252 return -EINVAL;
9253 if (copy_from_user(&arg, argp, sizeof(arg)))
9254 return -EFAULT;
9255 *sig = u64_to_user_ptr(arg.sigmask);
9256 *argsz = arg.sigmask_sz;
9257 *ts = u64_to_user_ptr(arg.ts);
9258 return 0;
9259 }
9260
SYSCALL_DEFINE6(io_uring_enter,unsigned int,fd,u32,to_submit,u32,min_complete,u32,flags,const void __user *,argp,size_t,argsz)9261 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9262 u32, min_complete, u32, flags, const void __user *, argp,
9263 size_t, argsz)
9264 {
9265 struct io_ring_ctx *ctx;
9266 int submitted = 0;
9267 struct fd f;
9268 long ret;
9269
9270 io_run_task_work();
9271
9272 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9273 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9274 return -EINVAL;
9275
9276 f = fdget(fd);
9277 if (unlikely(!f.file))
9278 return -EBADF;
9279
9280 ret = -EOPNOTSUPP;
9281 if (unlikely(f.file->f_op != &io_uring_fops))
9282 goto out_fput;
9283
9284 ret = -ENXIO;
9285 ctx = f.file->private_data;
9286 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9287 goto out_fput;
9288
9289 ret = -EBADFD;
9290 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9291 goto out;
9292
9293 /*
9294 * For SQ polling, the thread will do all submissions and completions.
9295 * Just return the requested submit count, and wake the thread if
9296 * we were asked to.
9297 */
9298 ret = 0;
9299 if (ctx->flags & IORING_SETUP_SQPOLL) {
9300 io_cqring_overflow_flush(ctx, false);
9301
9302 ret = -EOWNERDEAD;
9303 if (unlikely(ctx->sq_data->thread == NULL)) {
9304 goto out;
9305 }
9306 if (flags & IORING_ENTER_SQ_WAKEUP)
9307 wake_up(&ctx->sq_data->wait);
9308 if (flags & IORING_ENTER_SQ_WAIT) {
9309 ret = io_sqpoll_wait_sq(ctx);
9310 if (ret)
9311 goto out;
9312 }
9313 submitted = to_submit;
9314 } else if (to_submit) {
9315 ret = io_uring_add_task_file(ctx);
9316 if (unlikely(ret))
9317 goto out;
9318 mutex_lock(&ctx->uring_lock);
9319 submitted = io_submit_sqes(ctx, to_submit);
9320 mutex_unlock(&ctx->uring_lock);
9321
9322 if (submitted != to_submit)
9323 goto out;
9324 }
9325 if (flags & IORING_ENTER_GETEVENTS) {
9326 const sigset_t __user *sig;
9327 struct __kernel_timespec __user *ts;
9328
9329 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9330 if (unlikely(ret))
9331 goto out;
9332
9333 min_complete = min(min_complete, ctx->cq_entries);
9334
9335 /*
9336 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9337 * space applications don't need to do io completion events
9338 * polling again, they can rely on io_sq_thread to do polling
9339 * work, which can reduce cpu usage and uring_lock contention.
9340 */
9341 if (ctx->flags & IORING_SETUP_IOPOLL &&
9342 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9343 ret = io_iopoll_check(ctx, min_complete);
9344 } else {
9345 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9346 }
9347 }
9348
9349 out:
9350 percpu_ref_put(&ctx->refs);
9351 out_fput:
9352 fdput(f);
9353 return submitted ? submitted : ret;
9354 }
9355
9356 #ifdef CONFIG_PROC_FS
io_uring_show_cred(struct seq_file * m,unsigned int id,const struct cred * cred)9357 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9358 const struct cred *cred)
9359 {
9360 struct user_namespace *uns = seq_user_ns(m);
9361 struct group_info *gi;
9362 kernel_cap_t cap;
9363 unsigned __capi;
9364 int g;
9365
9366 seq_printf(m, "%5d\n", id);
9367 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9368 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9369 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9370 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9371 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9372 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9373 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9374 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9375 seq_puts(m, "\n\tGroups:\t");
9376 gi = cred->group_info;
9377 for (g = 0; g < gi->ngroups; g++) {
9378 seq_put_decimal_ull(m, g ? " " : "",
9379 from_kgid_munged(uns, gi->gid[g]));
9380 }
9381 seq_puts(m, "\n\tCapEff:\t");
9382 cap = cred->cap_effective;
9383 CAP_FOR_EACH_U32(__capi)
9384 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9385 seq_putc(m, '\n');
9386 return 0;
9387 }
9388
__io_uring_show_fdinfo(struct io_ring_ctx * ctx,struct seq_file * m)9389 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9390 {
9391 struct io_sq_data *sq = NULL;
9392 bool has_lock;
9393 int i;
9394
9395 /*
9396 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9397 * since fdinfo case grabs it in the opposite direction of normal use
9398 * cases. If we fail to get the lock, we just don't iterate any
9399 * structures that could be going away outside the io_uring mutex.
9400 */
9401 has_lock = mutex_trylock(&ctx->uring_lock);
9402
9403 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9404 sq = ctx->sq_data;
9405 if (!sq->thread)
9406 sq = NULL;
9407 }
9408
9409 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9410 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9411 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9412 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9413 struct file *f = io_file_from_index(ctx, i);
9414
9415 if (f)
9416 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9417 else
9418 seq_printf(m, "%5u: <none>\n", i);
9419 }
9420 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9421 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9422 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9423 unsigned int len = buf->ubuf_end - buf->ubuf;
9424
9425 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9426 }
9427 if (has_lock && !xa_empty(&ctx->personalities)) {
9428 unsigned long index;
9429 const struct cred *cred;
9430
9431 seq_printf(m, "Personalities:\n");
9432 xa_for_each(&ctx->personalities, index, cred)
9433 io_uring_show_cred(m, index, cred);
9434 }
9435 seq_printf(m, "PollList:\n");
9436 spin_lock_irq(&ctx->completion_lock);
9437 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9438 struct hlist_head *list = &ctx->cancel_hash[i];
9439 struct io_kiocb *req;
9440
9441 hlist_for_each_entry(req, list, hash_node)
9442 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9443 req->task->task_works != NULL);
9444 }
9445 spin_unlock_irq(&ctx->completion_lock);
9446 if (has_lock)
9447 mutex_unlock(&ctx->uring_lock);
9448 }
9449
io_uring_show_fdinfo(struct seq_file * m,struct file * f)9450 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9451 {
9452 struct io_ring_ctx *ctx = f->private_data;
9453
9454 if (percpu_ref_tryget(&ctx->refs)) {
9455 __io_uring_show_fdinfo(ctx, m);
9456 percpu_ref_put(&ctx->refs);
9457 }
9458 }
9459 #endif
9460
9461 static const struct file_operations io_uring_fops = {
9462 .release = io_uring_release,
9463 .mmap = io_uring_mmap,
9464 #ifndef CONFIG_MMU
9465 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9466 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9467 #endif
9468 .poll = io_uring_poll,
9469 .fasync = io_uring_fasync,
9470 #ifdef CONFIG_PROC_FS
9471 .show_fdinfo = io_uring_show_fdinfo,
9472 #endif
9473 };
9474
io_allocate_scq_urings(struct io_ring_ctx * ctx,struct io_uring_params * p)9475 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9476 struct io_uring_params *p)
9477 {
9478 struct io_rings *rings;
9479 size_t size, sq_array_offset;
9480
9481 /* make sure these are sane, as we already accounted them */
9482 ctx->sq_entries = p->sq_entries;
9483 ctx->cq_entries = p->cq_entries;
9484
9485 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9486 if (size == SIZE_MAX)
9487 return -EOVERFLOW;
9488
9489 rings = io_mem_alloc(size);
9490 if (!rings)
9491 return -ENOMEM;
9492
9493 ctx->rings = rings;
9494 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9495 rings->sq_ring_mask = p->sq_entries - 1;
9496 rings->cq_ring_mask = p->cq_entries - 1;
9497 rings->sq_ring_entries = p->sq_entries;
9498 rings->cq_ring_entries = p->cq_entries;
9499 ctx->sq_mask = rings->sq_ring_mask;
9500 ctx->cq_mask = rings->cq_ring_mask;
9501
9502 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9503 if (size == SIZE_MAX) {
9504 io_mem_free(ctx->rings);
9505 ctx->rings = NULL;
9506 return -EOVERFLOW;
9507 }
9508
9509 ctx->sq_sqes = io_mem_alloc(size);
9510 if (!ctx->sq_sqes) {
9511 io_mem_free(ctx->rings);
9512 ctx->rings = NULL;
9513 return -ENOMEM;
9514 }
9515
9516 return 0;
9517 }
9518
io_uring_install_fd(struct io_ring_ctx * ctx,struct file * file)9519 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9520 {
9521 int ret, fd;
9522
9523 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9524 if (fd < 0)
9525 return fd;
9526
9527 ret = io_uring_add_task_file(ctx);
9528 if (ret) {
9529 put_unused_fd(fd);
9530 return ret;
9531 }
9532 fd_install(fd, file);
9533 return fd;
9534 }
9535
9536 /*
9537 * Allocate an anonymous fd, this is what constitutes the application
9538 * visible backing of an io_uring instance. The application mmaps this
9539 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9540 * we have to tie this fd to a socket for file garbage collection purposes.
9541 */
io_uring_get_file(struct io_ring_ctx * ctx)9542 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9543 {
9544 struct file *file;
9545 #if defined(CONFIG_UNIX)
9546 int ret;
9547
9548 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9549 &ctx->ring_sock);
9550 if (ret)
9551 return ERR_PTR(ret);
9552 #endif
9553
9554 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9555 O_RDWR | O_CLOEXEC);
9556 #if defined(CONFIG_UNIX)
9557 if (IS_ERR(file)) {
9558 sock_release(ctx->ring_sock);
9559 ctx->ring_sock = NULL;
9560 } else {
9561 ctx->ring_sock->file = file;
9562 }
9563 #endif
9564 return file;
9565 }
9566
io_uring_create(unsigned entries,struct io_uring_params * p,struct io_uring_params __user * params)9567 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9568 struct io_uring_params __user *params)
9569 {
9570 struct io_ring_ctx *ctx;
9571 struct file *file;
9572 int ret;
9573
9574 if (!entries)
9575 return -EINVAL;
9576 if (entries > IORING_MAX_ENTRIES) {
9577 if (!(p->flags & IORING_SETUP_CLAMP))
9578 return -EINVAL;
9579 entries = IORING_MAX_ENTRIES;
9580 }
9581
9582 /*
9583 * Use twice as many entries for the CQ ring. It's possible for the
9584 * application to drive a higher depth than the size of the SQ ring,
9585 * since the sqes are only used at submission time. This allows for
9586 * some flexibility in overcommitting a bit. If the application has
9587 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9588 * of CQ ring entries manually.
9589 */
9590 p->sq_entries = roundup_pow_of_two(entries);
9591 if (p->flags & IORING_SETUP_CQSIZE) {
9592 /*
9593 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9594 * to a power-of-two, if it isn't already. We do NOT impose
9595 * any cq vs sq ring sizing.
9596 */
9597 if (!p->cq_entries)
9598 return -EINVAL;
9599 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9600 if (!(p->flags & IORING_SETUP_CLAMP))
9601 return -EINVAL;
9602 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9603 }
9604 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9605 if (p->cq_entries < p->sq_entries)
9606 return -EINVAL;
9607 } else {
9608 p->cq_entries = 2 * p->sq_entries;
9609 }
9610
9611 ctx = io_ring_ctx_alloc(p);
9612 if (!ctx)
9613 return -ENOMEM;
9614 ctx->compat = in_compat_syscall();
9615 if (!capable(CAP_IPC_LOCK))
9616 ctx->user = get_uid(current_user());
9617
9618 /*
9619 * This is just grabbed for accounting purposes. When a process exits,
9620 * the mm is exited and dropped before the files, hence we need to hang
9621 * on to this mm purely for the purposes of being able to unaccount
9622 * memory (locked/pinned vm). It's not used for anything else.
9623 */
9624 mmgrab(current->mm);
9625 ctx->mm_account = current->mm;
9626
9627 ret = io_allocate_scq_urings(ctx, p);
9628 if (ret)
9629 goto err;
9630
9631 ret = io_sq_offload_create(ctx, p);
9632 if (ret)
9633 goto err;
9634 /* always set a rsrc node */
9635 ret = io_rsrc_node_switch_start(ctx);
9636 if (ret)
9637 goto err;
9638 io_rsrc_node_switch(ctx, NULL);
9639
9640 memset(&p->sq_off, 0, sizeof(p->sq_off));
9641 p->sq_off.head = offsetof(struct io_rings, sq.head);
9642 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9643 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9644 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9645 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9646 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9647 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9648
9649 memset(&p->cq_off, 0, sizeof(p->cq_off));
9650 p->cq_off.head = offsetof(struct io_rings, cq.head);
9651 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9652 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9653 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9654 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9655 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9656 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9657
9658 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9659 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9660 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9661 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9662 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9663
9664 if (copy_to_user(params, p, sizeof(*p))) {
9665 ret = -EFAULT;
9666 goto err;
9667 }
9668
9669 file = io_uring_get_file(ctx);
9670 if (IS_ERR(file)) {
9671 ret = PTR_ERR(file);
9672 goto err;
9673 }
9674
9675 /*
9676 * Install ring fd as the very last thing, so we don't risk someone
9677 * having closed it before we finish setup
9678 */
9679 ret = io_uring_install_fd(ctx, file);
9680 if (ret < 0) {
9681 /* fput will clean it up */
9682 fput(file);
9683 return ret;
9684 }
9685
9686 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9687 return ret;
9688 err:
9689 io_ring_ctx_wait_and_kill(ctx);
9690 return ret;
9691 }
9692
9693 /*
9694 * Sets up an aio uring context, and returns the fd. Applications asks for a
9695 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9696 * params structure passed in.
9697 */
io_uring_setup(u32 entries,struct io_uring_params __user * params)9698 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9699 {
9700 struct io_uring_params p;
9701 int i;
9702
9703 if (copy_from_user(&p, params, sizeof(p)))
9704 return -EFAULT;
9705 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9706 if (p.resv[i])
9707 return -EINVAL;
9708 }
9709
9710 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9711 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9712 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9713 IORING_SETUP_R_DISABLED))
9714 return -EINVAL;
9715
9716 return io_uring_create(entries, &p, params);
9717 }
9718
SYSCALL_DEFINE2(io_uring_setup,u32,entries,struct io_uring_params __user *,params)9719 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9720 struct io_uring_params __user *, params)
9721 {
9722 return io_uring_setup(entries, params);
9723 }
9724
io_probe(struct io_ring_ctx * ctx,void __user * arg,unsigned nr_args)9725 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9726 {
9727 struct io_uring_probe *p;
9728 size_t size;
9729 int i, ret;
9730
9731 size = struct_size(p, ops, nr_args);
9732 if (size == SIZE_MAX)
9733 return -EOVERFLOW;
9734 p = kzalloc(size, GFP_KERNEL);
9735 if (!p)
9736 return -ENOMEM;
9737
9738 ret = -EFAULT;
9739 if (copy_from_user(p, arg, size))
9740 goto out;
9741 ret = -EINVAL;
9742 if (memchr_inv(p, 0, size))
9743 goto out;
9744
9745 p->last_op = IORING_OP_LAST - 1;
9746 if (nr_args > IORING_OP_LAST)
9747 nr_args = IORING_OP_LAST;
9748
9749 for (i = 0; i < nr_args; i++) {
9750 p->ops[i].op = i;
9751 if (!io_op_defs[i].not_supported)
9752 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9753 }
9754 p->ops_len = i;
9755
9756 ret = 0;
9757 if (copy_to_user(arg, p, size))
9758 ret = -EFAULT;
9759 out:
9760 kfree(p);
9761 return ret;
9762 }
9763
io_register_personality(struct io_ring_ctx * ctx)9764 static int io_register_personality(struct io_ring_ctx *ctx)
9765 {
9766 const struct cred *creds;
9767 u32 id;
9768 int ret;
9769
9770 creds = get_current_cred();
9771
9772 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9773 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9774 if (!ret)
9775 return id;
9776 put_cred(creds);
9777 return ret;
9778 }
9779
io_register_restrictions(struct io_ring_ctx * ctx,void __user * arg,unsigned int nr_args)9780 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9781 unsigned int nr_args)
9782 {
9783 struct io_uring_restriction *res;
9784 size_t size;
9785 int i, ret;
9786
9787 /* Restrictions allowed only if rings started disabled */
9788 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9789 return -EBADFD;
9790
9791 /* We allow only a single restrictions registration */
9792 if (ctx->restrictions.registered)
9793 return -EBUSY;
9794
9795 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9796 return -EINVAL;
9797
9798 size = array_size(nr_args, sizeof(*res));
9799 if (size == SIZE_MAX)
9800 return -EOVERFLOW;
9801
9802 res = memdup_user(arg, size);
9803 if (IS_ERR(res))
9804 return PTR_ERR(res);
9805
9806 ret = 0;
9807
9808 for (i = 0; i < nr_args; i++) {
9809 switch (res[i].opcode) {
9810 case IORING_RESTRICTION_REGISTER_OP:
9811 if (res[i].register_op >= IORING_REGISTER_LAST) {
9812 ret = -EINVAL;
9813 goto out;
9814 }
9815
9816 __set_bit(res[i].register_op,
9817 ctx->restrictions.register_op);
9818 break;
9819 case IORING_RESTRICTION_SQE_OP:
9820 if (res[i].sqe_op >= IORING_OP_LAST) {
9821 ret = -EINVAL;
9822 goto out;
9823 }
9824
9825 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9826 break;
9827 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9828 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9829 break;
9830 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9831 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9832 break;
9833 default:
9834 ret = -EINVAL;
9835 goto out;
9836 }
9837 }
9838
9839 out:
9840 /* Reset all restrictions if an error happened */
9841 if (ret != 0)
9842 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9843 else
9844 ctx->restrictions.registered = true;
9845
9846 kfree(res);
9847 return ret;
9848 }
9849
io_register_enable_rings(struct io_ring_ctx * ctx)9850 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9851 {
9852 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9853 return -EBADFD;
9854
9855 if (ctx->restrictions.registered)
9856 ctx->restricted = 1;
9857
9858 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9859 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9860 wake_up(&ctx->sq_data->wait);
9861 return 0;
9862 }
9863
__io_register_rsrc_update(struct io_ring_ctx * ctx,unsigned type,struct io_uring_rsrc_update2 * up,unsigned nr_args)9864 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9865 struct io_uring_rsrc_update2 *up,
9866 unsigned nr_args)
9867 {
9868 __u32 tmp;
9869 int err;
9870
9871 if (up->resv)
9872 return -EINVAL;
9873 if (check_add_overflow(up->offset, nr_args, &tmp))
9874 return -EOVERFLOW;
9875 err = io_rsrc_node_switch_start(ctx);
9876 if (err)
9877 return err;
9878
9879 switch (type) {
9880 case IORING_RSRC_FILE:
9881 return __io_sqe_files_update(ctx, up, nr_args);
9882 case IORING_RSRC_BUFFER:
9883 return __io_sqe_buffers_update(ctx, up, nr_args);
9884 }
9885 return -EINVAL;
9886 }
9887
io_register_files_update(struct io_ring_ctx * ctx,void __user * arg,unsigned nr_args)9888 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9889 unsigned nr_args)
9890 {
9891 struct io_uring_rsrc_update2 up;
9892
9893 if (!nr_args)
9894 return -EINVAL;
9895 memset(&up, 0, sizeof(up));
9896 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9897 return -EFAULT;
9898 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9899 }
9900
io_register_rsrc_update(struct io_ring_ctx * ctx,void __user * arg,unsigned size)9901 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9902 unsigned size)
9903 {
9904 struct io_uring_rsrc_update2 up;
9905
9906 if (size != sizeof(up))
9907 return -EINVAL;
9908 if (copy_from_user(&up, arg, sizeof(up)))
9909 return -EFAULT;
9910 if (!up.nr)
9911 return -EINVAL;
9912 return __io_register_rsrc_update(ctx, up.type, &up, up.nr);
9913 }
9914
io_register_rsrc(struct io_ring_ctx * ctx,void __user * arg,unsigned int size)9915 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9916 unsigned int size)
9917 {
9918 struct io_uring_rsrc_register rr;
9919
9920 /* keep it extendible */
9921 if (size != sizeof(rr))
9922 return -EINVAL;
9923
9924 memset(&rr, 0, sizeof(rr));
9925 if (copy_from_user(&rr, arg, size))
9926 return -EFAULT;
9927 if (!rr.nr)
9928 return -EINVAL;
9929
9930 switch (rr.type) {
9931 case IORING_RSRC_FILE:
9932 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9933 rr.nr, u64_to_user_ptr(rr.tags));
9934 case IORING_RSRC_BUFFER:
9935 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9936 rr.nr, u64_to_user_ptr(rr.tags));
9937 }
9938 return -EINVAL;
9939 }
9940
io_register_op_must_quiesce(int op)9941 static bool io_register_op_must_quiesce(int op)
9942 {
9943 switch (op) {
9944 case IORING_REGISTER_BUFFERS:
9945 case IORING_UNREGISTER_BUFFERS:
9946 case IORING_REGISTER_FILES:
9947 case IORING_UNREGISTER_FILES:
9948 case IORING_REGISTER_FILES_UPDATE:
9949 case IORING_REGISTER_PROBE:
9950 case IORING_REGISTER_PERSONALITY:
9951 case IORING_UNREGISTER_PERSONALITY:
9952 case IORING_REGISTER_RSRC:
9953 case IORING_REGISTER_RSRC_UPDATE:
9954 return false;
9955 default:
9956 return true;
9957 }
9958 }
9959
__io_uring_register(struct io_ring_ctx * ctx,unsigned opcode,void __user * arg,unsigned nr_args)9960 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9961 void __user *arg, unsigned nr_args)
9962 __releases(ctx->uring_lock)
9963 __acquires(ctx->uring_lock)
9964 {
9965 int ret;
9966
9967 /*
9968 * We're inside the ring mutex, if the ref is already dying, then
9969 * someone else killed the ctx or is already going through
9970 * io_uring_register().
9971 */
9972 if (percpu_ref_is_dying(&ctx->refs))
9973 return -ENXIO;
9974
9975 if (ctx->restricted) {
9976 if (opcode >= IORING_REGISTER_LAST)
9977 return -EINVAL;
9978 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9979 if (!test_bit(opcode, ctx->restrictions.register_op))
9980 return -EACCES;
9981 }
9982
9983 if (io_register_op_must_quiesce(opcode)) {
9984 percpu_ref_kill(&ctx->refs);
9985
9986 /*
9987 * Drop uring mutex before waiting for references to exit. If
9988 * another thread is currently inside io_uring_enter() it might
9989 * need to grab the uring_lock to make progress. If we hold it
9990 * here across the drain wait, then we can deadlock. It's safe
9991 * to drop the mutex here, since no new references will come in
9992 * after we've killed the percpu ref.
9993 */
9994 mutex_unlock(&ctx->uring_lock);
9995 do {
9996 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9997 if (!ret)
9998 break;
9999 ret = io_run_task_work_sig();
10000 if (ret < 0)
10001 break;
10002 } while (1);
10003 mutex_lock(&ctx->uring_lock);
10004
10005 if (ret) {
10006 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10007 return ret;
10008 }
10009 }
10010
10011 switch (opcode) {
10012 case IORING_REGISTER_BUFFERS:
10013 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10014 break;
10015 case IORING_UNREGISTER_BUFFERS:
10016 ret = -EINVAL;
10017 if (arg || nr_args)
10018 break;
10019 ret = io_sqe_buffers_unregister(ctx);
10020 break;
10021 case IORING_REGISTER_FILES:
10022 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10023 break;
10024 case IORING_UNREGISTER_FILES:
10025 ret = -EINVAL;
10026 if (arg || nr_args)
10027 break;
10028 ret = io_sqe_files_unregister(ctx);
10029 break;
10030 case IORING_REGISTER_FILES_UPDATE:
10031 ret = io_register_files_update(ctx, arg, nr_args);
10032 break;
10033 case IORING_REGISTER_EVENTFD:
10034 case IORING_REGISTER_EVENTFD_ASYNC:
10035 ret = -EINVAL;
10036 if (nr_args != 1)
10037 break;
10038 ret = io_eventfd_register(ctx, arg);
10039 if (ret)
10040 break;
10041 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10042 ctx->eventfd_async = 1;
10043 else
10044 ctx->eventfd_async = 0;
10045 break;
10046 case IORING_UNREGISTER_EVENTFD:
10047 ret = -EINVAL;
10048 if (arg || nr_args)
10049 break;
10050 ret = io_eventfd_unregister(ctx);
10051 break;
10052 case IORING_REGISTER_PROBE:
10053 ret = -EINVAL;
10054 if (!arg || nr_args > 256)
10055 break;
10056 ret = io_probe(ctx, arg, nr_args);
10057 break;
10058 case IORING_REGISTER_PERSONALITY:
10059 ret = -EINVAL;
10060 if (arg || nr_args)
10061 break;
10062 ret = io_register_personality(ctx);
10063 break;
10064 case IORING_UNREGISTER_PERSONALITY:
10065 ret = -EINVAL;
10066 if (arg)
10067 break;
10068 ret = io_unregister_personality(ctx, nr_args);
10069 break;
10070 case IORING_REGISTER_ENABLE_RINGS:
10071 ret = -EINVAL;
10072 if (arg || nr_args)
10073 break;
10074 ret = io_register_enable_rings(ctx);
10075 break;
10076 case IORING_REGISTER_RESTRICTIONS:
10077 ret = io_register_restrictions(ctx, arg, nr_args);
10078 break;
10079 case IORING_REGISTER_RSRC:
10080 ret = io_register_rsrc(ctx, arg, nr_args);
10081 break;
10082 case IORING_REGISTER_RSRC_UPDATE:
10083 ret = io_register_rsrc_update(ctx, arg, nr_args);
10084 break;
10085 default:
10086 ret = -EINVAL;
10087 break;
10088 }
10089
10090 if (io_register_op_must_quiesce(opcode)) {
10091 /* bring the ctx back to life */
10092 percpu_ref_reinit(&ctx->refs);
10093 reinit_completion(&ctx->ref_comp);
10094 }
10095 return ret;
10096 }
10097
SYSCALL_DEFINE4(io_uring_register,unsigned int,fd,unsigned int,opcode,void __user *,arg,unsigned int,nr_args)10098 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10099 void __user *, arg, unsigned int, nr_args)
10100 {
10101 struct io_ring_ctx *ctx;
10102 long ret = -EBADF;
10103 struct fd f;
10104
10105 f = fdget(fd);
10106 if (!f.file)
10107 return -EBADF;
10108
10109 ret = -EOPNOTSUPP;
10110 if (f.file->f_op != &io_uring_fops)
10111 goto out_fput;
10112
10113 ctx = f.file->private_data;
10114
10115 io_run_task_work();
10116
10117 mutex_lock(&ctx->uring_lock);
10118 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10119 mutex_unlock(&ctx->uring_lock);
10120 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10121 ctx->cq_ev_fd != NULL, ret);
10122 out_fput:
10123 fdput(f);
10124 return ret;
10125 }
10126
io_uring_init(void)10127 static int __init io_uring_init(void)
10128 {
10129 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10130 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10131 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10132 } while (0)
10133
10134 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10135 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10136 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10137 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10138 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10139 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10140 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10141 BUILD_BUG_SQE_ELEM(8, __u64, off);
10142 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10143 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10144 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10145 BUILD_BUG_SQE_ELEM(24, __u32, len);
10146 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10147 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10148 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10149 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10150 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10151 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10152 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10153 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10154 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10155 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10156 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10157 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10158 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10159 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10160 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10161 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10162 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10163 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10164 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10165
10166 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10167 sizeof(struct io_uring_rsrc_update));
10168 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10169 sizeof(struct io_uring_rsrc_update2));
10170 /* should fit into one byte */
10171 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10172
10173 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10174 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10175 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10176 SLAB_ACCOUNT);
10177 return 0;
10178 };
10179 __initcall(io_uring_init);
10180