1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
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
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
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
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
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.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
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>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.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>
64 #include <net/af_unix.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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
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)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
239 struct list_head list;
245 struct io_restriction {
246 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
247 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
248 u8 sqe_flags_allowed;
249 u8 sqe_flags_required;
254 IO_SQ_THREAD_SHOULD_STOP = 0,
255 IO_SQ_THREAD_SHOULD_PARK,
260 atomic_t park_pending;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
276 struct callback_head *park_task_work;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
449 struct callback_head *exit_task_work;
451 struct wait_queue_head hash_wait;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work;
455 struct list_head tctx_list;
458 struct io_uring_task {
459 /* submission side */
461 struct wait_queue_head wait;
462 const struct io_ring_ctx *last;
464 struct percpu_counter inflight;
468 spinlock_t task_lock;
469 struct io_wq_work_list task_list;
470 unsigned long task_state;
471 struct callback_head task_work;
475 * First field must be the file pointer in all the
476 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
478 struct io_poll_iocb {
480 struct wait_queue_head *head;
484 struct wait_queue_entry wait;
487 struct io_poll_remove {
497 struct io_timeout_data {
498 struct io_kiocb *req;
499 struct hrtimer timer;
500 struct timespec64 ts;
501 enum hrtimer_mode mode;
506 struct sockaddr __user *addr;
507 int __user *addr_len;
509 unsigned long nofile;
529 struct list_head list;
530 /* head of the link, used by linked timeouts only */
531 struct io_kiocb *head;
534 struct io_timeout_rem {
539 struct timespec64 ts;
544 /* NOTE: kiocb has the file as the first member, so don't do it here */
552 struct sockaddr __user *addr;
559 struct user_msghdr __user *umsg;
565 struct io_buffer *kbuf;
571 struct filename *filename;
573 unsigned long nofile;
576 struct io_rsrc_update {
602 struct epoll_event event;
606 struct file *file_out;
607 struct file *file_in;
614 struct io_provide_buf {
628 const char __user *filename;
629 struct statx __user *buffer;
641 struct filename *oldpath;
642 struct filename *newpath;
650 struct filename *filename;
653 struct io_completion {
655 struct list_head list;
659 struct io_async_connect {
660 struct sockaddr_storage address;
663 struct io_async_msghdr {
664 struct iovec fast_iov[UIO_FASTIOV];
665 /* points to an allocated iov, if NULL we use fast_iov instead */
666 struct iovec *free_iov;
667 struct sockaddr __user *uaddr;
669 struct sockaddr_storage addr;
673 struct iovec fast_iov[UIO_FASTIOV];
674 const struct iovec *free_iovec;
675 struct iov_iter iter;
677 struct wait_page_queue wpq;
681 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
682 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
683 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
684 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
685 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
686 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
692 REQ_F_LINK_TIMEOUT_BIT,
694 REQ_F_NEED_CLEANUP_BIT,
696 REQ_F_BUFFER_SELECTED_BIT,
697 REQ_F_NO_FILE_TABLE_BIT,
698 REQ_F_LTIMEOUT_ACTIVE_BIT,
699 REQ_F_COMPLETE_INLINE_BIT,
702 /* not a real bit, just to check we're not overflowing the space */
708 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
709 /* drain existing IO first */
710 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
712 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
713 /* doesn't sever on completion < 0 */
714 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
716 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
717 /* IOSQE_BUFFER_SELECT */
718 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
720 /* fail rest of links */
721 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
722 /* on inflight list, should be cancelled and waited on exit reliably */
723 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
724 /* read/write uses file position */
725 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
726 /* must not punt to workers */
727 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
728 /* has or had linked timeout */
729 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
731 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
733 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
734 /* already went through poll handler */
735 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
736 /* buffer already selected */
737 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
738 /* doesn't need file table for this request */
739 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
740 /* linked timeout is active, i.e. prepared by link's head */
741 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
742 /* completion is deferred through io_comp_state */
743 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
744 /* caller should reissue async */
745 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
749 struct io_poll_iocb poll;
750 struct io_poll_iocb *double_poll;
753 struct io_task_work {
754 struct io_wq_work_node node;
755 task_work_func_t func;
759 * NOTE! Each of the iocb union members has the file pointer
760 * as the first entry in their struct definition. So you can
761 * access the file pointer through any of the sub-structs,
762 * or directly as just 'ki_filp' in this struct.
768 struct io_poll_iocb poll;
769 struct io_poll_remove poll_remove;
770 struct io_accept accept;
772 struct io_cancel cancel;
773 struct io_timeout timeout;
774 struct io_timeout_rem timeout_rem;
775 struct io_connect connect;
776 struct io_sr_msg sr_msg;
778 struct io_close close;
779 struct io_rsrc_update rsrc_update;
780 struct io_fadvise fadvise;
781 struct io_madvise madvise;
782 struct io_epoll epoll;
783 struct io_splice splice;
784 struct io_provide_buf pbuf;
785 struct io_statx statx;
786 struct io_shutdown shutdown;
787 struct io_rename rename;
788 struct io_unlink unlink;
789 /* use only after cleaning per-op data, see io_clean_op() */
790 struct io_completion compl;
793 /* opcode allocated if it needs to store data for async defer */
796 /* polled IO has completed */
802 struct io_ring_ctx *ctx;
805 struct task_struct *task;
808 struct io_kiocb *link;
809 struct percpu_ref *fixed_rsrc_refs;
812 * 1. used with ctx->iopoll_list with reads/writes
813 * 2. to track reqs with ->files (see io_op_def::file_table)
815 struct list_head inflight_entry;
817 struct io_task_work io_task_work;
818 struct callback_head task_work;
820 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
821 struct hlist_node hash_node;
822 struct async_poll *apoll;
823 struct io_wq_work work;
826 struct io_tctx_node {
827 struct list_head ctx_node;
828 struct task_struct *task;
829 struct io_ring_ctx *ctx;
832 struct io_defer_entry {
833 struct list_head list;
834 struct io_kiocb *req;
839 /* needs req->file assigned */
840 unsigned needs_file : 1;
841 /* hash wq insertion if file is a regular file */
842 unsigned hash_reg_file : 1;
843 /* unbound wq insertion if file is a non-regular file */
844 unsigned unbound_nonreg_file : 1;
845 /* opcode is not supported by this kernel */
846 unsigned not_supported : 1;
847 /* set if opcode supports polled "wait" */
849 unsigned pollout : 1;
850 /* op supports buffer selection */
851 unsigned buffer_select : 1;
852 /* must always have async data allocated */
853 unsigned needs_async_data : 1;
854 /* should block plug */
856 /* size of async data needed, if any */
857 unsigned short async_size;
860 static const struct io_op_def io_op_defs[] = {
861 [IORING_OP_NOP] = {},
862 [IORING_OP_READV] = {
864 .unbound_nonreg_file = 1,
867 .needs_async_data = 1,
869 .async_size = sizeof(struct io_async_rw),
871 [IORING_OP_WRITEV] = {
874 .unbound_nonreg_file = 1,
876 .needs_async_data = 1,
878 .async_size = sizeof(struct io_async_rw),
880 [IORING_OP_FSYNC] = {
883 [IORING_OP_READ_FIXED] = {
885 .unbound_nonreg_file = 1,
888 .async_size = sizeof(struct io_async_rw),
890 [IORING_OP_WRITE_FIXED] = {
893 .unbound_nonreg_file = 1,
896 .async_size = sizeof(struct io_async_rw),
898 [IORING_OP_POLL_ADD] = {
900 .unbound_nonreg_file = 1,
902 [IORING_OP_POLL_REMOVE] = {},
903 [IORING_OP_SYNC_FILE_RANGE] = {
906 [IORING_OP_SENDMSG] = {
908 .unbound_nonreg_file = 1,
910 .needs_async_data = 1,
911 .async_size = sizeof(struct io_async_msghdr),
913 [IORING_OP_RECVMSG] = {
915 .unbound_nonreg_file = 1,
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_async_msghdr),
921 [IORING_OP_TIMEOUT] = {
922 .needs_async_data = 1,
923 .async_size = sizeof(struct io_timeout_data),
925 [IORING_OP_TIMEOUT_REMOVE] = {
926 /* used by timeout updates' prep() */
928 [IORING_OP_ACCEPT] = {
930 .unbound_nonreg_file = 1,
933 [IORING_OP_ASYNC_CANCEL] = {},
934 [IORING_OP_LINK_TIMEOUT] = {
935 .needs_async_data = 1,
936 .async_size = sizeof(struct io_timeout_data),
938 [IORING_OP_CONNECT] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_data = 1,
943 .async_size = sizeof(struct io_async_connect),
945 [IORING_OP_FALLOCATE] = {
948 [IORING_OP_OPENAT] = {},
949 [IORING_OP_CLOSE] = {},
950 [IORING_OP_FILES_UPDATE] = {},
951 [IORING_OP_STATX] = {},
954 .unbound_nonreg_file = 1,
958 .async_size = sizeof(struct io_async_rw),
960 [IORING_OP_WRITE] = {
962 .unbound_nonreg_file = 1,
965 .async_size = sizeof(struct io_async_rw),
967 [IORING_OP_FADVISE] = {
970 [IORING_OP_MADVISE] = {},
973 .unbound_nonreg_file = 1,
978 .unbound_nonreg_file = 1,
982 [IORING_OP_OPENAT2] = {
984 [IORING_OP_EPOLL_CTL] = {
985 .unbound_nonreg_file = 1,
987 [IORING_OP_SPLICE] = {
990 .unbound_nonreg_file = 1,
992 [IORING_OP_PROVIDE_BUFFERS] = {},
993 [IORING_OP_REMOVE_BUFFERS] = {},
997 .unbound_nonreg_file = 1,
999 [IORING_OP_SHUTDOWN] = {
1002 [IORING_OP_RENAMEAT] = {},
1003 [IORING_OP_UNLINKAT] = {},
1006 static bool io_disarm_next(struct io_kiocb *req);
1007 static void io_uring_del_task_file(unsigned long index);
1008 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1009 struct task_struct *task,
1010 struct files_struct *files);
1011 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1012 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1013 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1014 struct io_ring_ctx *ctx);
1015 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1017 static bool io_rw_reissue(struct io_kiocb *req);
1018 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1019 static void io_put_req(struct io_kiocb *req);
1020 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1021 static void io_double_put_req(struct io_kiocb *req);
1022 static void io_dismantle_req(struct io_kiocb *req);
1023 static void io_put_task(struct task_struct *task, int nr);
1024 static void io_queue_next(struct io_kiocb *req);
1025 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1026 static void __io_queue_linked_timeout(struct io_kiocb *req);
1027 static void io_queue_linked_timeout(struct io_kiocb *req);
1028 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1029 struct io_uring_rsrc_update *ip,
1031 static void __io_clean_op(struct io_kiocb *req);
1032 static struct file *io_file_get(struct io_submit_state *state,
1033 struct io_kiocb *req, int fd, bool fixed);
1034 static void __io_queue_sqe(struct io_kiocb *req);
1035 static void io_rsrc_put_work(struct work_struct *work);
1037 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1038 struct iov_iter *iter, bool needs_lock);
1039 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1040 const struct iovec *fast_iov,
1041 struct iov_iter *iter, bool force);
1042 static void io_req_task_queue(struct io_kiocb *req);
1043 static void io_submit_flush_completions(struct io_comp_state *cs,
1044 struct io_ring_ctx *ctx);
1046 static struct kmem_cache *req_cachep;
1048 static const struct file_operations io_uring_fops;
1050 struct sock *io_uring_get_socket(struct file *file)
1052 #if defined(CONFIG_UNIX)
1053 if (file->f_op == &io_uring_fops) {
1054 struct io_ring_ctx *ctx = file->private_data;
1056 return ctx->ring_sock->sk;
1061 EXPORT_SYMBOL(io_uring_get_socket);
1063 #define io_for_each_link(pos, head) \
1064 for (pos = (head); pos; pos = pos->link)
1066 static inline void io_clean_op(struct io_kiocb *req)
1068 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1072 static inline void io_set_resource_node(struct io_kiocb *req)
1074 struct io_ring_ctx *ctx = req->ctx;
1076 if (!req->fixed_rsrc_refs) {
1077 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1078 percpu_ref_get(req->fixed_rsrc_refs);
1082 static bool io_match_task(struct io_kiocb *head,
1083 struct task_struct *task,
1084 struct files_struct *files)
1086 struct io_kiocb *req;
1088 if (task && head->task != task) {
1089 /* in terms of cancelation, always match if req task is dead */
1090 if (head->task->flags & PF_EXITING)
1097 io_for_each_link(req, head) {
1098 if (req->flags & REQ_F_INFLIGHT)
1104 static inline void req_set_fail_links(struct io_kiocb *req)
1106 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1107 req->flags |= REQ_F_FAIL_LINK;
1110 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1112 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1114 complete(&ctx->ref_comp);
1117 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1119 return !req->timeout.off;
1122 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1124 struct io_ring_ctx *ctx;
1127 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1132 * Use 5 bits less than the max cq entries, that should give us around
1133 * 32 entries per hash list if totally full and uniformly spread.
1135 hash_bits = ilog2(p->cq_entries);
1139 ctx->cancel_hash_bits = hash_bits;
1140 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1142 if (!ctx->cancel_hash)
1144 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1146 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1147 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1150 ctx->flags = p->flags;
1151 init_waitqueue_head(&ctx->sqo_sq_wait);
1152 INIT_LIST_HEAD(&ctx->sqd_list);
1153 init_waitqueue_head(&ctx->cq_wait);
1154 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1155 init_completion(&ctx->ref_comp);
1156 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1157 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1158 mutex_init(&ctx->uring_lock);
1159 init_waitqueue_head(&ctx->wait);
1160 spin_lock_init(&ctx->completion_lock);
1161 INIT_LIST_HEAD(&ctx->iopoll_list);
1162 INIT_LIST_HEAD(&ctx->defer_list);
1163 INIT_LIST_HEAD(&ctx->timeout_list);
1164 spin_lock_init(&ctx->inflight_lock);
1165 INIT_LIST_HEAD(&ctx->inflight_list);
1166 spin_lock_init(&ctx->rsrc_ref_lock);
1167 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1168 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1169 init_llist_head(&ctx->rsrc_put_llist);
1170 INIT_LIST_HEAD(&ctx->tctx_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1172 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1175 kfree(ctx->cancel_hash);
1180 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1182 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1183 struct io_ring_ctx *ctx = req->ctx;
1185 return seq != ctx->cached_cq_tail
1186 + READ_ONCE(ctx->cached_cq_overflow);
1192 static void io_req_track_inflight(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!(req->flags & REQ_F_INFLIGHT)) {
1197 req->flags |= REQ_F_INFLIGHT;
1199 spin_lock_irq(&ctx->inflight_lock);
1200 list_add(&req->inflight_entry, &ctx->inflight_list);
1201 spin_unlock_irq(&ctx->inflight_lock);
1205 static void io_prep_async_work(struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 struct io_ring_ctx *ctx = req->ctx;
1210 if (!req->work.creds)
1211 req->work.creds = get_current_cred();
1213 if (req->flags & REQ_F_FORCE_ASYNC)
1214 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1216 if (req->flags & REQ_F_ISREG) {
1217 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1218 io_wq_hash_work(&req->work, file_inode(req->file));
1219 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1220 if (def->unbound_nonreg_file)
1221 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 static void io_prep_async_link(struct io_kiocb *req)
1227 struct io_kiocb *cur;
1229 io_for_each_link(cur, req)
1230 io_prep_async_work(cur);
1233 static void io_queue_async_work(struct io_kiocb *req)
1235 struct io_ring_ctx *ctx = req->ctx;
1236 struct io_kiocb *link = io_prep_linked_timeout(req);
1237 struct io_uring_task *tctx = req->task->io_uring;
1240 BUG_ON(!tctx->io_wq);
1242 /* init ->work of the whole link before punting */
1243 io_prep_async_link(req);
1244 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1245 &req->work, req->flags);
1246 io_wq_enqueue(tctx->io_wq, &req->work);
1248 io_queue_linked_timeout(link);
1251 static void io_kill_timeout(struct io_kiocb *req, int status)
1253 struct io_timeout_data *io = req->async_data;
1256 ret = hrtimer_try_to_cancel(&io->timer);
1258 atomic_set(&req->ctx->cq_timeouts,
1259 atomic_read(&req->ctx->cq_timeouts) + 1);
1260 list_del_init(&req->timeout.list);
1261 io_cqring_fill_event(req, status);
1262 io_put_req_deferred(req, 1);
1266 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1269 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1270 struct io_defer_entry, list);
1272 if (req_need_defer(de->req, de->seq))
1274 list_del_init(&de->list);
1275 io_req_task_queue(de->req);
1277 } while (!list_empty(&ctx->defer_list));
1280 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1284 if (list_empty(&ctx->timeout_list))
1287 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1290 u32 events_needed, events_got;
1291 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1292 struct io_kiocb, timeout.list);
1294 if (io_is_timeout_noseq(req))
1298 * Since seq can easily wrap around over time, subtract
1299 * the last seq at which timeouts were flushed before comparing.
1300 * Assuming not more than 2^31-1 events have happened since,
1301 * these subtractions won't have wrapped, so we can check if
1302 * target is in [last_seq, current_seq] by comparing the two.
1304 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1305 events_got = seq - ctx->cq_last_tm_flush;
1306 if (events_got < events_needed)
1309 list_del_init(&req->timeout.list);
1310 io_kill_timeout(req, 0);
1311 } while (!list_empty(&ctx->timeout_list));
1313 ctx->cq_last_tm_flush = seq;
1316 static void io_commit_cqring(struct io_ring_ctx *ctx)
1318 io_flush_timeouts(ctx);
1320 /* order cqe stores with ring update */
1321 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1323 if (unlikely(!list_empty(&ctx->defer_list)))
1324 __io_queue_deferred(ctx);
1327 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1329 struct io_rings *r = ctx->rings;
1331 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1334 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1336 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1339 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1341 struct io_rings *rings = ctx->rings;
1345 * writes to the cq entry need to come after reading head; the
1346 * control dependency is enough as we're using WRITE_ONCE to
1349 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1352 tail = ctx->cached_cq_tail++;
1353 return &rings->cqes[tail & ctx->cq_mask];
1356 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1360 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1362 if (!ctx->eventfd_async)
1364 return io_wq_current_is_worker();
1367 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1369 /* see waitqueue_active() comment */
1372 if (waitqueue_active(&ctx->wait))
1373 wake_up(&ctx->wait);
1374 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1375 wake_up(&ctx->sq_data->wait);
1376 if (io_should_trigger_evfd(ctx))
1377 eventfd_signal(ctx->cq_ev_fd, 1);
1378 if (waitqueue_active(&ctx->cq_wait)) {
1379 wake_up_interruptible(&ctx->cq_wait);
1380 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1384 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1386 /* see waitqueue_active() comment */
1389 if (ctx->flags & IORING_SETUP_SQPOLL) {
1390 if (waitqueue_active(&ctx->wait))
1391 wake_up(&ctx->wait);
1393 if (io_should_trigger_evfd(ctx))
1394 eventfd_signal(ctx->cq_ev_fd, 1);
1395 if (waitqueue_active(&ctx->cq_wait)) {
1396 wake_up_interruptible(&ctx->cq_wait);
1397 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1401 /* Returns true if there are no backlogged entries after the flush */
1402 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1403 struct task_struct *tsk,
1404 struct files_struct *files)
1406 struct io_rings *rings = ctx->rings;
1407 struct io_kiocb *req, *tmp;
1408 struct io_uring_cqe *cqe;
1409 unsigned long flags;
1410 bool all_flushed, posted;
1413 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1417 spin_lock_irqsave(&ctx->completion_lock, flags);
1418 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1419 if (!io_match_task(req, tsk, files))
1422 cqe = io_get_cqring(ctx);
1426 list_move(&req->compl.list, &list);
1428 WRITE_ONCE(cqe->user_data, req->user_data);
1429 WRITE_ONCE(cqe->res, req->result);
1430 WRITE_ONCE(cqe->flags, req->compl.cflags);
1432 ctx->cached_cq_overflow++;
1433 WRITE_ONCE(ctx->rings->cq_overflow,
1434 ctx->cached_cq_overflow);
1439 all_flushed = list_empty(&ctx->cq_overflow_list);
1441 clear_bit(0, &ctx->sq_check_overflow);
1442 clear_bit(0, &ctx->cq_check_overflow);
1443 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1447 io_commit_cqring(ctx);
1448 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1450 io_cqring_ev_posted(ctx);
1452 while (!list_empty(&list)) {
1453 req = list_first_entry(&list, struct io_kiocb, compl.list);
1454 list_del(&req->compl.list);
1461 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1462 struct task_struct *tsk,
1463 struct files_struct *files)
1467 if (test_bit(0, &ctx->cq_check_overflow)) {
1468 /* iopoll syncs against uring_lock, not completion_lock */
1469 if (ctx->flags & IORING_SETUP_IOPOLL)
1470 mutex_lock(&ctx->uring_lock);
1471 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1472 if (ctx->flags & IORING_SETUP_IOPOLL)
1473 mutex_unlock(&ctx->uring_lock);
1479 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1481 struct io_ring_ctx *ctx = req->ctx;
1482 struct io_uring_cqe *cqe;
1484 trace_io_uring_complete(ctx, req->user_data, res);
1487 * If we can't get a cq entry, userspace overflowed the
1488 * submission (by quite a lot). Increment the overflow count in
1491 cqe = io_get_cqring(ctx);
1493 WRITE_ONCE(cqe->user_data, req->user_data);
1494 WRITE_ONCE(cqe->res, res);
1495 WRITE_ONCE(cqe->flags, cflags);
1496 } else if (ctx->cq_overflow_flushed ||
1497 atomic_read(&req->task->io_uring->in_idle)) {
1499 * If we're in ring overflow flush mode, or in task cancel mode,
1500 * then we cannot store the request for later flushing, we need
1501 * to drop it on the floor.
1503 ctx->cached_cq_overflow++;
1504 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1506 if (list_empty(&ctx->cq_overflow_list)) {
1507 set_bit(0, &ctx->sq_check_overflow);
1508 set_bit(0, &ctx->cq_check_overflow);
1509 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1513 req->compl.cflags = cflags;
1514 refcount_inc(&req->refs);
1515 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1519 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1521 __io_cqring_fill_event(req, res, 0);
1524 static void io_req_complete_post(struct io_kiocb *req, long res,
1525 unsigned int cflags)
1527 struct io_ring_ctx *ctx = req->ctx;
1528 unsigned long flags;
1530 spin_lock_irqsave(&ctx->completion_lock, flags);
1531 __io_cqring_fill_event(req, res, cflags);
1533 * If we're the last reference to this request, add to our locked
1536 if (refcount_dec_and_test(&req->refs)) {
1537 struct io_comp_state *cs = &ctx->submit_state.comp;
1539 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1540 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1541 io_disarm_next(req);
1543 io_req_task_queue(req->link);
1547 io_dismantle_req(req);
1548 io_put_task(req->task, 1);
1549 list_add(&req->compl.list, &cs->locked_free_list);
1550 cs->locked_free_nr++;
1552 if (!percpu_ref_tryget(&ctx->refs))
1555 io_commit_cqring(ctx);
1556 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1559 io_cqring_ev_posted(ctx);
1560 percpu_ref_put(&ctx->refs);
1564 static void io_req_complete_state(struct io_kiocb *req, long res,
1565 unsigned int cflags)
1569 req->compl.cflags = cflags;
1570 req->flags |= REQ_F_COMPLETE_INLINE;
1573 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1574 long res, unsigned cflags)
1576 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1577 io_req_complete_state(req, res, cflags);
1579 io_req_complete_post(req, res, cflags);
1582 static inline void io_req_complete(struct io_kiocb *req, long res)
1584 __io_req_complete(req, 0, res, 0);
1587 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1589 struct io_submit_state *state = &ctx->submit_state;
1590 struct io_comp_state *cs = &state->comp;
1591 struct io_kiocb *req = NULL;
1594 * If we have more than a batch's worth of requests in our IRQ side
1595 * locked cache, grab the lock and move them over to our submission
1598 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1599 spin_lock_irq(&ctx->completion_lock);
1600 list_splice_init(&cs->locked_free_list, &cs->free_list);
1601 cs->locked_free_nr = 0;
1602 spin_unlock_irq(&ctx->completion_lock);
1605 while (!list_empty(&cs->free_list)) {
1606 req = list_first_entry(&cs->free_list, struct io_kiocb,
1608 list_del(&req->compl.list);
1609 state->reqs[state->free_reqs++] = req;
1610 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1617 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1619 struct io_submit_state *state = &ctx->submit_state;
1621 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1623 if (!state->free_reqs) {
1624 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1627 if (io_flush_cached_reqs(ctx))
1630 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1634 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1635 * retry single alloc to be on the safe side.
1637 if (unlikely(ret <= 0)) {
1638 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1639 if (!state->reqs[0])
1643 state->free_reqs = ret;
1647 return state->reqs[state->free_reqs];
1650 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1657 static void io_dismantle_req(struct io_kiocb *req)
1661 if (req->async_data)
1662 kfree(req->async_data);
1664 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1665 if (req->fixed_rsrc_refs)
1666 percpu_ref_put(req->fixed_rsrc_refs);
1667 if (req->work.creds) {
1668 put_cred(req->work.creds);
1669 req->work.creds = NULL;
1672 if (req->flags & REQ_F_INFLIGHT) {
1673 struct io_ring_ctx *ctx = req->ctx;
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->inflight_lock, flags);
1677 list_del(&req->inflight_entry);
1678 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1679 req->flags &= ~REQ_F_INFLIGHT;
1683 /* must to be called somewhat shortly after putting a request */
1684 static inline void io_put_task(struct task_struct *task, int nr)
1686 struct io_uring_task *tctx = task->io_uring;
1688 percpu_counter_sub(&tctx->inflight, nr);
1689 if (unlikely(atomic_read(&tctx->in_idle)))
1690 wake_up(&tctx->wait);
1691 put_task_struct_many(task, nr);
1694 static void __io_free_req(struct io_kiocb *req)
1696 struct io_ring_ctx *ctx = req->ctx;
1698 io_dismantle_req(req);
1699 io_put_task(req->task, 1);
1701 kmem_cache_free(req_cachep, req);
1702 percpu_ref_put(&ctx->refs);
1705 static inline void io_remove_next_linked(struct io_kiocb *req)
1707 struct io_kiocb *nxt = req->link;
1709 req->link = nxt->link;
1713 static bool io_kill_linked_timeout(struct io_kiocb *req)
1714 __must_hold(&req->ctx->completion_lock)
1716 struct io_kiocb *link = req->link;
1717 bool cancelled = false;
1720 * Can happen if a linked timeout fired and link had been like
1721 * req -> link t-out -> link t-out [-> ...]
1723 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1724 struct io_timeout_data *io = link->async_data;
1727 io_remove_next_linked(req);
1728 link->timeout.head = NULL;
1729 ret = hrtimer_try_to_cancel(&io->timer);
1731 io_cqring_fill_event(link, -ECANCELED);
1732 io_put_req_deferred(link, 1);
1736 req->flags &= ~REQ_F_LINK_TIMEOUT;
1740 static void io_fail_links(struct io_kiocb *req)
1741 __must_hold(&req->ctx->completion_lock)
1743 struct io_kiocb *nxt, *link = req->link;
1750 trace_io_uring_fail_link(req, link);
1751 io_cqring_fill_event(link, -ECANCELED);
1752 io_put_req_deferred(link, 2);
1757 static bool io_disarm_next(struct io_kiocb *req)
1758 __must_hold(&req->ctx->completion_lock)
1760 bool posted = false;
1762 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1763 posted = io_kill_linked_timeout(req);
1764 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1765 posted |= (req->link != NULL);
1771 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1773 struct io_kiocb *nxt;
1776 * If LINK is set, we have dependent requests in this chain. If we
1777 * didn't fail this request, queue the first one up, moving any other
1778 * dependencies to the next request. In case of failure, fail the rest
1781 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1782 struct io_ring_ctx *ctx = req->ctx;
1783 unsigned long flags;
1786 spin_lock_irqsave(&ctx->completion_lock, flags);
1787 posted = io_disarm_next(req);
1789 io_commit_cqring(req->ctx);
1790 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1792 io_cqring_ev_posted(ctx);
1799 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1801 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1803 return __io_req_find_next(req);
1806 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1810 if (ctx->submit_state.comp.nr) {
1811 mutex_lock(&ctx->uring_lock);
1812 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1813 mutex_unlock(&ctx->uring_lock);
1815 percpu_ref_put(&ctx->refs);
1818 static bool __tctx_task_work(struct io_uring_task *tctx)
1820 struct io_ring_ctx *ctx = NULL;
1821 struct io_wq_work_list list;
1822 struct io_wq_work_node *node;
1824 if (wq_list_empty(&tctx->task_list))
1827 spin_lock_irq(&tctx->task_lock);
1828 list = tctx->task_list;
1829 INIT_WQ_LIST(&tctx->task_list);
1830 spin_unlock_irq(&tctx->task_lock);
1834 struct io_wq_work_node *next = node->next;
1835 struct io_kiocb *req;
1837 req = container_of(node, struct io_kiocb, io_task_work.node);
1838 if (req->ctx != ctx) {
1839 ctx_flush_and_put(ctx);
1841 percpu_ref_get(&ctx->refs);
1844 req->task_work.func(&req->task_work);
1848 ctx_flush_and_put(ctx);
1849 return list.first != NULL;
1852 static void tctx_task_work(struct callback_head *cb)
1854 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1856 clear_bit(0, &tctx->task_state);
1858 while (__tctx_task_work(tctx))
1862 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1863 enum task_work_notify_mode notify)
1865 struct io_uring_task *tctx = tsk->io_uring;
1866 struct io_wq_work_node *node, *prev;
1867 unsigned long flags;
1870 WARN_ON_ONCE(!tctx);
1872 spin_lock_irqsave(&tctx->task_lock, flags);
1873 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1874 spin_unlock_irqrestore(&tctx->task_lock, flags);
1876 /* task_work already pending, we're done */
1877 if (test_bit(0, &tctx->task_state) ||
1878 test_and_set_bit(0, &tctx->task_state))
1881 if (!task_work_add(tsk, &tctx->task_work, notify))
1885 * Slow path - we failed, find and delete work. if the work is not
1886 * in the list, it got run and we're fine.
1889 spin_lock_irqsave(&tctx->task_lock, flags);
1890 wq_list_for_each(node, prev, &tctx->task_list) {
1891 if (&req->io_task_work.node == node) {
1892 wq_list_del(&tctx->task_list, node, prev);
1897 spin_unlock_irqrestore(&tctx->task_lock, flags);
1898 clear_bit(0, &tctx->task_state);
1902 static int io_req_task_work_add(struct io_kiocb *req)
1904 struct task_struct *tsk = req->task;
1905 struct io_ring_ctx *ctx = req->ctx;
1906 enum task_work_notify_mode notify;
1909 if (tsk->flags & PF_EXITING)
1913 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1914 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1915 * processing task_work. There's no reliable way to tell if TWA_RESUME
1919 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1920 notify = TWA_SIGNAL;
1922 ret = io_task_work_add(tsk, req, notify);
1924 wake_up_process(tsk);
1929 static bool io_run_task_work_head(struct callback_head **work_head)
1931 struct callback_head *work, *next;
1932 bool executed = false;
1935 work = xchg(work_head, NULL);
1951 static void io_task_work_add_head(struct callback_head **work_head,
1952 struct callback_head *task_work)
1954 struct callback_head *head;
1957 head = READ_ONCE(*work_head);
1958 task_work->next = head;
1959 } while (cmpxchg(work_head, head, task_work) != head);
1962 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1963 task_work_func_t cb)
1965 init_task_work(&req->task_work, cb);
1966 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1969 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1971 struct io_ring_ctx *ctx = req->ctx;
1973 spin_lock_irq(&ctx->completion_lock);
1974 io_cqring_fill_event(req, error);
1975 io_commit_cqring(ctx);
1976 spin_unlock_irq(&ctx->completion_lock);
1978 io_cqring_ev_posted(ctx);
1979 req_set_fail_links(req);
1980 io_double_put_req(req);
1983 static void io_req_task_cancel(struct callback_head *cb)
1985 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1986 struct io_ring_ctx *ctx = req->ctx;
1988 mutex_lock(&ctx->uring_lock);
1989 __io_req_task_cancel(req, req->result);
1990 mutex_unlock(&ctx->uring_lock);
1991 percpu_ref_put(&ctx->refs);
1994 static void __io_req_task_submit(struct io_kiocb *req)
1996 struct io_ring_ctx *ctx = req->ctx;
1998 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1999 mutex_lock(&ctx->uring_lock);
2000 if (!(current->flags & PF_EXITING) && !current->in_execve)
2001 __io_queue_sqe(req);
2003 __io_req_task_cancel(req, -EFAULT);
2004 mutex_unlock(&ctx->uring_lock);
2007 static void io_req_task_submit(struct callback_head *cb)
2009 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2011 __io_req_task_submit(req);
2014 static void io_req_task_queue(struct io_kiocb *req)
2018 req->task_work.func = io_req_task_submit;
2019 ret = io_req_task_work_add(req);
2020 if (unlikely(ret)) {
2021 req->result = -ECANCELED;
2022 percpu_ref_get(&req->ctx->refs);
2023 io_req_task_work_add_fallback(req, io_req_task_cancel);
2027 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2029 percpu_ref_get(&req->ctx->refs);
2031 req->task_work.func = io_req_task_cancel;
2033 if (unlikely(io_req_task_work_add(req)))
2034 io_req_task_work_add_fallback(req, io_req_task_cancel);
2037 static inline void io_queue_next(struct io_kiocb *req)
2039 struct io_kiocb *nxt = io_req_find_next(req);
2042 io_req_task_queue(nxt);
2045 static void io_free_req(struct io_kiocb *req)
2052 struct task_struct *task;
2057 static inline void io_init_req_batch(struct req_batch *rb)
2064 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2065 struct req_batch *rb)
2068 io_put_task(rb->task, rb->task_refs);
2070 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2073 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2074 struct io_submit_state *state)
2078 if (req->task != rb->task) {
2080 io_put_task(rb->task, rb->task_refs);
2081 rb->task = req->task;
2087 io_dismantle_req(req);
2088 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2089 state->reqs[state->free_reqs++] = req;
2091 list_add(&req->compl.list, &state->comp.free_list);
2094 static void io_submit_flush_completions(struct io_comp_state *cs,
2095 struct io_ring_ctx *ctx)
2098 struct io_kiocb *req;
2099 struct req_batch rb;
2101 io_init_req_batch(&rb);
2102 spin_lock_irq(&ctx->completion_lock);
2103 for (i = 0; i < nr; i++) {
2105 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2107 io_commit_cqring(ctx);
2108 spin_unlock_irq(&ctx->completion_lock);
2110 io_cqring_ev_posted(ctx);
2111 for (i = 0; i < nr; i++) {
2114 /* submission and completion refs */
2115 if (refcount_sub_and_test(2, &req->refs))
2116 io_req_free_batch(&rb, req, &ctx->submit_state);
2119 io_req_free_batch_finish(ctx, &rb);
2124 * Drop reference to request, return next in chain (if there is one) if this
2125 * was the last reference to this request.
2127 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2129 struct io_kiocb *nxt = NULL;
2131 if (refcount_dec_and_test(&req->refs)) {
2132 nxt = io_req_find_next(req);
2138 static void io_put_req(struct io_kiocb *req)
2140 if (refcount_dec_and_test(&req->refs))
2144 static void io_put_req_deferred_cb(struct callback_head *cb)
2146 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2151 static void io_free_req_deferred(struct io_kiocb *req)
2155 req->task_work.func = io_put_req_deferred_cb;
2156 ret = io_req_task_work_add(req);
2158 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2161 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2163 if (refcount_sub_and_test(refs, &req->refs))
2164 io_free_req_deferred(req);
2167 static void io_double_put_req(struct io_kiocb *req)
2169 /* drop both submit and complete references */
2170 if (refcount_sub_and_test(2, &req->refs))
2174 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2176 /* See comment at the top of this file */
2178 return __io_cqring_events(ctx);
2181 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2183 struct io_rings *rings = ctx->rings;
2185 /* make sure SQ entry isn't read before tail */
2186 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2189 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2191 unsigned int cflags;
2193 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2194 cflags |= IORING_CQE_F_BUFFER;
2195 req->flags &= ~REQ_F_BUFFER_SELECTED;
2200 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2202 struct io_buffer *kbuf;
2204 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2205 return io_put_kbuf(req, kbuf);
2208 static inline bool io_run_task_work(void)
2211 * Not safe to run on exiting task, and the task_work handling will
2212 * not add work to such a task.
2214 if (unlikely(current->flags & PF_EXITING))
2216 if (current->task_works) {
2217 __set_current_state(TASK_RUNNING);
2226 * Find and free completed poll iocbs
2228 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2229 struct list_head *done)
2231 struct req_batch rb;
2232 struct io_kiocb *req;
2234 /* order with ->result store in io_complete_rw_iopoll() */
2237 io_init_req_batch(&rb);
2238 while (!list_empty(done)) {
2241 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2242 list_del(&req->inflight_entry);
2244 if (READ_ONCE(req->result) == -EAGAIN) {
2245 req->iopoll_completed = 0;
2246 if (io_rw_reissue(req))
2250 if (req->flags & REQ_F_BUFFER_SELECTED)
2251 cflags = io_put_rw_kbuf(req);
2253 __io_cqring_fill_event(req, req->result, cflags);
2256 if (refcount_dec_and_test(&req->refs))
2257 io_req_free_batch(&rb, req, &ctx->submit_state);
2260 io_commit_cqring(ctx);
2261 io_cqring_ev_posted_iopoll(ctx);
2262 io_req_free_batch_finish(ctx, &rb);
2265 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2268 struct io_kiocb *req, *tmp;
2274 * Only spin for completions if we don't have multiple devices hanging
2275 * off our complete list, and we're under the requested amount.
2277 spin = !ctx->poll_multi_file && *nr_events < min;
2280 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2281 struct kiocb *kiocb = &req->rw.kiocb;
2284 * Move completed and retryable entries to our local lists.
2285 * If we find a request that requires polling, break out
2286 * and complete those lists first, if we have entries there.
2288 if (READ_ONCE(req->iopoll_completed)) {
2289 list_move_tail(&req->inflight_entry, &done);
2292 if (!list_empty(&done))
2295 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2299 /* iopoll may have completed current req */
2300 if (READ_ONCE(req->iopoll_completed))
2301 list_move_tail(&req->inflight_entry, &done);
2308 if (!list_empty(&done))
2309 io_iopoll_complete(ctx, nr_events, &done);
2315 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2316 * non-spinning poll check - we'll still enter the driver poll loop, but only
2317 * as a non-spinning completion check.
2319 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2322 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2325 ret = io_do_iopoll(ctx, nr_events, min);
2328 if (*nr_events >= min)
2336 * We can't just wait for polled events to come to us, we have to actively
2337 * find and complete them.
2339 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2341 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2344 mutex_lock(&ctx->uring_lock);
2345 while (!list_empty(&ctx->iopoll_list)) {
2346 unsigned int nr_events = 0;
2348 io_do_iopoll(ctx, &nr_events, 0);
2350 /* let it sleep and repeat later if can't complete a request */
2354 * Ensure we allow local-to-the-cpu processing to take place,
2355 * in this case we need to ensure that we reap all events.
2356 * Also let task_work, etc. to progress by releasing the mutex
2358 if (need_resched()) {
2359 mutex_unlock(&ctx->uring_lock);
2361 mutex_lock(&ctx->uring_lock);
2364 mutex_unlock(&ctx->uring_lock);
2367 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2369 unsigned int nr_events = 0;
2370 int iters = 0, ret = 0;
2373 * We disallow the app entering submit/complete with polling, but we
2374 * still need to lock the ring to prevent racing with polled issue
2375 * that got punted to a workqueue.
2377 mutex_lock(&ctx->uring_lock);
2380 * Don't enter poll loop if we already have events pending.
2381 * If we do, we can potentially be spinning for commands that
2382 * already triggered a CQE (eg in error).
2384 if (test_bit(0, &ctx->cq_check_overflow))
2385 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2386 if (io_cqring_events(ctx))
2390 * If a submit got punted to a workqueue, we can have the
2391 * application entering polling for a command before it gets
2392 * issued. That app will hold the uring_lock for the duration
2393 * of the poll right here, so we need to take a breather every
2394 * now and then to ensure that the issue has a chance to add
2395 * the poll to the issued list. Otherwise we can spin here
2396 * forever, while the workqueue is stuck trying to acquire the
2399 if (!(++iters & 7)) {
2400 mutex_unlock(&ctx->uring_lock);
2402 mutex_lock(&ctx->uring_lock);
2405 ret = io_iopoll_getevents(ctx, &nr_events, min);
2409 } while (min && !nr_events && !need_resched());
2411 mutex_unlock(&ctx->uring_lock);
2415 static void kiocb_end_write(struct io_kiocb *req)
2418 * Tell lockdep we inherited freeze protection from submission
2421 if (req->flags & REQ_F_ISREG) {
2422 struct inode *inode = file_inode(req->file);
2424 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2426 file_end_write(req->file);
2430 static bool io_resubmit_prep(struct io_kiocb *req)
2432 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2434 struct iov_iter iter;
2436 /* already prepared */
2437 if (req->async_data)
2440 switch (req->opcode) {
2441 case IORING_OP_READV:
2442 case IORING_OP_READ_FIXED:
2443 case IORING_OP_READ:
2446 case IORING_OP_WRITEV:
2447 case IORING_OP_WRITE_FIXED:
2448 case IORING_OP_WRITE:
2452 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2457 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2460 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2463 static bool io_rw_should_reissue(struct io_kiocb *req)
2465 umode_t mode = file_inode(req->file)->i_mode;
2466 struct io_ring_ctx *ctx = req->ctx;
2468 if (!S_ISBLK(mode) && !S_ISREG(mode))
2470 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2471 !(ctx->flags & IORING_SETUP_IOPOLL)))
2474 * If ref is dying, we might be running poll reap from the exit work.
2475 * Don't attempt to reissue from that path, just let it fail with
2478 if (percpu_ref_is_dying(&ctx->refs))
2483 static bool io_rw_should_reissue(struct io_kiocb *req)
2489 static bool io_rw_reissue(struct io_kiocb *req)
2492 if (!io_rw_should_reissue(req))
2495 lockdep_assert_held(&req->ctx->uring_lock);
2497 if (io_resubmit_prep(req)) {
2498 refcount_inc(&req->refs);
2499 io_queue_async_work(req);
2502 req_set_fail_links(req);
2507 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2508 unsigned int issue_flags)
2512 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2513 kiocb_end_write(req);
2514 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2515 req->flags |= REQ_F_REISSUE;
2518 if (res != req->result)
2519 req_set_fail_links(req);
2520 if (req->flags & REQ_F_BUFFER_SELECTED)
2521 cflags = io_put_rw_kbuf(req);
2522 __io_req_complete(req, issue_flags, res, cflags);
2525 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2527 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2529 __io_complete_rw(req, res, res2, 0);
2532 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2534 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2537 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2538 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2539 struct io_async_rw *rw = req->async_data;
2542 iov_iter_revert(&rw->iter,
2543 req->result - iov_iter_count(&rw->iter));
2544 else if (!io_resubmit_prep(req))
2549 if (kiocb->ki_flags & IOCB_WRITE)
2550 kiocb_end_write(req);
2552 if (res != -EAGAIN && res != req->result)
2553 req_set_fail_links(req);
2555 WRITE_ONCE(req->result, res);
2556 /* order with io_poll_complete() checking ->result */
2558 WRITE_ONCE(req->iopoll_completed, 1);
2562 * After the iocb has been issued, it's safe to be found on the poll list.
2563 * Adding the kiocb to the list AFTER submission ensures that we don't
2564 * find it from a io_iopoll_getevents() thread before the issuer is done
2565 * accessing the kiocb cookie.
2567 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2569 struct io_ring_ctx *ctx = req->ctx;
2572 * Track whether we have multiple files in our lists. This will impact
2573 * how we do polling eventually, not spinning if we're on potentially
2574 * different devices.
2576 if (list_empty(&ctx->iopoll_list)) {
2577 ctx->poll_multi_file = false;
2578 } else if (!ctx->poll_multi_file) {
2579 struct io_kiocb *list_req;
2581 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2583 if (list_req->file != req->file)
2584 ctx->poll_multi_file = true;
2588 * For fast devices, IO may have already completed. If it has, add
2589 * it to the front so we find it first.
2591 if (READ_ONCE(req->iopoll_completed))
2592 list_add(&req->inflight_entry, &ctx->iopoll_list);
2594 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2597 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2598 * task context or in io worker task context. If current task context is
2599 * sq thread, we don't need to check whether should wake up sq thread.
2601 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2602 wq_has_sleeper(&ctx->sq_data->wait))
2603 wake_up(&ctx->sq_data->wait);
2606 static inline void io_state_file_put(struct io_submit_state *state)
2608 if (state->file_refs) {
2609 fput_many(state->file, state->file_refs);
2610 state->file_refs = 0;
2615 * Get as many references to a file as we have IOs left in this submission,
2616 * assuming most submissions are for one file, or at least that each file
2617 * has more than one submission.
2619 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2624 if (state->file_refs) {
2625 if (state->fd == fd) {
2629 io_state_file_put(state);
2631 state->file = fget_many(fd, state->ios_left);
2632 if (unlikely(!state->file))
2636 state->file_refs = state->ios_left - 1;
2640 static bool io_bdev_nowait(struct block_device *bdev)
2642 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2646 * If we tracked the file through the SCM inflight mechanism, we could support
2647 * any file. For now, just ensure that anything potentially problematic is done
2650 static bool io_file_supports_async(struct file *file, int rw)
2652 umode_t mode = file_inode(file)->i_mode;
2654 if (S_ISBLK(mode)) {
2655 if (IS_ENABLED(CONFIG_BLOCK) &&
2656 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2660 if (S_ISCHR(mode) || S_ISSOCK(mode))
2662 if (S_ISREG(mode)) {
2663 if (IS_ENABLED(CONFIG_BLOCK) &&
2664 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2665 file->f_op != &io_uring_fops)
2670 /* any ->read/write should understand O_NONBLOCK */
2671 if (file->f_flags & O_NONBLOCK)
2674 if (!(file->f_mode & FMODE_NOWAIT))
2678 return file->f_op->read_iter != NULL;
2680 return file->f_op->write_iter != NULL;
2683 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2685 struct io_ring_ctx *ctx = req->ctx;
2686 struct kiocb *kiocb = &req->rw.kiocb;
2687 struct file *file = req->file;
2691 if (S_ISREG(file_inode(file)->i_mode))
2692 req->flags |= REQ_F_ISREG;
2694 kiocb->ki_pos = READ_ONCE(sqe->off);
2695 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2696 req->flags |= REQ_F_CUR_POS;
2697 kiocb->ki_pos = file->f_pos;
2699 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2700 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2701 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2705 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2706 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2707 req->flags |= REQ_F_NOWAIT;
2709 ioprio = READ_ONCE(sqe->ioprio);
2711 ret = ioprio_check_cap(ioprio);
2715 kiocb->ki_ioprio = ioprio;
2717 kiocb->ki_ioprio = get_current_ioprio();
2719 if (ctx->flags & IORING_SETUP_IOPOLL) {
2720 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2721 !kiocb->ki_filp->f_op->iopoll)
2724 kiocb->ki_flags |= IOCB_HIPRI;
2725 kiocb->ki_complete = io_complete_rw_iopoll;
2726 req->iopoll_completed = 0;
2728 if (kiocb->ki_flags & IOCB_HIPRI)
2730 kiocb->ki_complete = io_complete_rw;
2733 req->rw.addr = READ_ONCE(sqe->addr);
2734 req->rw.len = READ_ONCE(sqe->len);
2735 req->buf_index = READ_ONCE(sqe->buf_index);
2739 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2745 case -ERESTARTNOINTR:
2746 case -ERESTARTNOHAND:
2747 case -ERESTART_RESTARTBLOCK:
2749 * We can't just restart the syscall, since previously
2750 * submitted sqes may already be in progress. Just fail this
2756 kiocb->ki_complete(kiocb, ret, 0);
2760 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2761 unsigned int issue_flags)
2763 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2764 struct io_async_rw *io = req->async_data;
2766 /* add previously done IO, if any */
2767 if (io && io->bytes_done > 0) {
2769 ret = io->bytes_done;
2771 ret += io->bytes_done;
2774 if (req->flags & REQ_F_CUR_POS)
2775 req->file->f_pos = kiocb->ki_pos;
2776 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2777 __io_complete_rw(req, ret, 0, issue_flags);
2779 io_rw_done(kiocb, ret);
2782 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2784 struct io_ring_ctx *ctx = req->ctx;
2785 size_t len = req->rw.len;
2786 struct io_mapped_ubuf *imu;
2787 u16 index, buf_index = req->buf_index;
2791 if (unlikely(buf_index >= ctx->nr_user_bufs))
2793 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2794 imu = &ctx->user_bufs[index];
2795 buf_addr = req->rw.addr;
2798 if (buf_addr + len < buf_addr)
2800 /* not inside the mapped region */
2801 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2805 * May not be a start of buffer, set size appropriately
2806 * and advance us to the beginning.
2808 offset = buf_addr - imu->ubuf;
2809 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2813 * Don't use iov_iter_advance() here, as it's really slow for
2814 * using the latter parts of a big fixed buffer - it iterates
2815 * over each segment manually. We can cheat a bit here, because
2818 * 1) it's a BVEC iter, we set it up
2819 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2820 * first and last bvec
2822 * So just find our index, and adjust the iterator afterwards.
2823 * If the offset is within the first bvec (or the whole first
2824 * bvec, just use iov_iter_advance(). This makes it easier
2825 * since we can just skip the first segment, which may not
2826 * be PAGE_SIZE aligned.
2828 const struct bio_vec *bvec = imu->bvec;
2830 if (offset <= bvec->bv_len) {
2831 iov_iter_advance(iter, offset);
2833 unsigned long seg_skip;
2835 /* skip first vec */
2836 offset -= bvec->bv_len;
2837 seg_skip = 1 + (offset >> PAGE_SHIFT);
2839 iter->bvec = bvec + seg_skip;
2840 iter->nr_segs -= seg_skip;
2841 iter->count -= bvec->bv_len + offset;
2842 iter->iov_offset = offset & ~PAGE_MASK;
2849 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2852 mutex_unlock(&ctx->uring_lock);
2855 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2858 * "Normal" inline submissions always hold the uring_lock, since we
2859 * grab it from the system call. Same is true for the SQPOLL offload.
2860 * The only exception is when we've detached the request and issue it
2861 * from an async worker thread, grab the lock for that case.
2864 mutex_lock(&ctx->uring_lock);
2867 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2868 int bgid, struct io_buffer *kbuf,
2871 struct io_buffer *head;
2873 if (req->flags & REQ_F_BUFFER_SELECTED)
2876 io_ring_submit_lock(req->ctx, needs_lock);
2878 lockdep_assert_held(&req->ctx->uring_lock);
2880 head = xa_load(&req->ctx->io_buffers, bgid);
2882 if (!list_empty(&head->list)) {
2883 kbuf = list_last_entry(&head->list, struct io_buffer,
2885 list_del(&kbuf->list);
2888 xa_erase(&req->ctx->io_buffers, bgid);
2890 if (*len > kbuf->len)
2893 kbuf = ERR_PTR(-ENOBUFS);
2896 io_ring_submit_unlock(req->ctx, needs_lock);
2901 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2904 struct io_buffer *kbuf;
2907 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2908 bgid = req->buf_index;
2909 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2912 req->rw.addr = (u64) (unsigned long) kbuf;
2913 req->flags |= REQ_F_BUFFER_SELECTED;
2914 return u64_to_user_ptr(kbuf->addr);
2917 #ifdef CONFIG_COMPAT
2918 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2921 struct compat_iovec __user *uiov;
2922 compat_ssize_t clen;
2926 uiov = u64_to_user_ptr(req->rw.addr);
2927 if (!access_ok(uiov, sizeof(*uiov)))
2929 if (__get_user(clen, &uiov->iov_len))
2935 buf = io_rw_buffer_select(req, &len, needs_lock);
2937 return PTR_ERR(buf);
2938 iov[0].iov_base = buf;
2939 iov[0].iov_len = (compat_size_t) len;
2944 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2947 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2951 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2954 len = iov[0].iov_len;
2957 buf = io_rw_buffer_select(req, &len, needs_lock);
2959 return PTR_ERR(buf);
2960 iov[0].iov_base = buf;
2961 iov[0].iov_len = len;
2965 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2968 if (req->flags & REQ_F_BUFFER_SELECTED) {
2969 struct io_buffer *kbuf;
2971 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2972 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2973 iov[0].iov_len = kbuf->len;
2976 if (req->rw.len != 1)
2979 #ifdef CONFIG_COMPAT
2980 if (req->ctx->compat)
2981 return io_compat_import(req, iov, needs_lock);
2984 return __io_iov_buffer_select(req, iov, needs_lock);
2987 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2988 struct iov_iter *iter, bool needs_lock)
2990 void __user *buf = u64_to_user_ptr(req->rw.addr);
2991 size_t sqe_len = req->rw.len;
2992 u8 opcode = req->opcode;
2995 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2997 return io_import_fixed(req, rw, iter);
3000 /* buffer index only valid with fixed read/write, or buffer select */
3001 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3004 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3005 if (req->flags & REQ_F_BUFFER_SELECT) {
3006 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3008 return PTR_ERR(buf);
3009 req->rw.len = sqe_len;
3012 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3017 if (req->flags & REQ_F_BUFFER_SELECT) {
3018 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3020 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3025 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3029 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3031 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3035 * For files that don't have ->read_iter() and ->write_iter(), handle them
3036 * by looping over ->read() or ->write() manually.
3038 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3040 struct kiocb *kiocb = &req->rw.kiocb;
3041 struct file *file = req->file;
3045 * Don't support polled IO through this interface, and we can't
3046 * support non-blocking either. For the latter, this just causes
3047 * the kiocb to be handled from an async context.
3049 if (kiocb->ki_flags & IOCB_HIPRI)
3051 if (kiocb->ki_flags & IOCB_NOWAIT)
3054 while (iov_iter_count(iter)) {
3058 if (!iov_iter_is_bvec(iter)) {
3059 iovec = iov_iter_iovec(iter);
3061 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3062 iovec.iov_len = req->rw.len;
3066 nr = file->f_op->read(file, iovec.iov_base,
3067 iovec.iov_len, io_kiocb_ppos(kiocb));
3069 nr = file->f_op->write(file, iovec.iov_base,
3070 iovec.iov_len, io_kiocb_ppos(kiocb));
3079 if (nr != iovec.iov_len)
3083 iov_iter_advance(iter, nr);
3089 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3090 const struct iovec *fast_iov, struct iov_iter *iter)
3092 struct io_async_rw *rw = req->async_data;
3094 memcpy(&rw->iter, iter, sizeof(*iter));
3095 rw->free_iovec = iovec;
3097 /* can only be fixed buffers, no need to do anything */
3098 if (iov_iter_is_bvec(iter))
3101 unsigned iov_off = 0;
3103 rw->iter.iov = rw->fast_iov;
3104 if (iter->iov != fast_iov) {
3105 iov_off = iter->iov - fast_iov;
3106 rw->iter.iov += iov_off;
3108 if (rw->fast_iov != fast_iov)
3109 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3110 sizeof(struct iovec) * iter->nr_segs);
3112 req->flags |= REQ_F_NEED_CLEANUP;
3116 static inline int __io_alloc_async_data(struct io_kiocb *req)
3118 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3119 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3120 return req->async_data == NULL;
3123 static int io_alloc_async_data(struct io_kiocb *req)
3125 if (!io_op_defs[req->opcode].needs_async_data)
3128 return __io_alloc_async_data(req);
3131 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3132 const struct iovec *fast_iov,
3133 struct iov_iter *iter, bool force)
3135 if (!force && !io_op_defs[req->opcode].needs_async_data)
3137 if (!req->async_data) {
3138 if (__io_alloc_async_data(req)) {
3143 io_req_map_rw(req, iovec, fast_iov, iter);
3148 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3150 struct io_async_rw *iorw = req->async_data;
3151 struct iovec *iov = iorw->fast_iov;
3154 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3155 if (unlikely(ret < 0))
3158 iorw->bytes_done = 0;
3159 iorw->free_iovec = iov;
3161 req->flags |= REQ_F_NEED_CLEANUP;
3165 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3167 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3169 return io_prep_rw(req, sqe);
3173 * This is our waitqueue callback handler, registered through lock_page_async()
3174 * when we initially tried to do the IO with the iocb armed our waitqueue.
3175 * This gets called when the page is unlocked, and we generally expect that to
3176 * happen when the page IO is completed and the page is now uptodate. This will
3177 * queue a task_work based retry of the operation, attempting to copy the data
3178 * again. If the latter fails because the page was NOT uptodate, then we will
3179 * do a thread based blocking retry of the operation. That's the unexpected
3182 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3183 int sync, void *arg)
3185 struct wait_page_queue *wpq;
3186 struct io_kiocb *req = wait->private;
3187 struct wait_page_key *key = arg;
3189 wpq = container_of(wait, struct wait_page_queue, wait);
3191 if (!wake_page_match(wpq, key))
3194 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3195 list_del_init(&wait->entry);
3197 /* submit ref gets dropped, acquire a new one */
3198 refcount_inc(&req->refs);
3199 io_req_task_queue(req);
3204 * This controls whether a given IO request should be armed for async page
3205 * based retry. If we return false here, the request is handed to the async
3206 * worker threads for retry. If we're doing buffered reads on a regular file,
3207 * we prepare a private wait_page_queue entry and retry the operation. This
3208 * will either succeed because the page is now uptodate and unlocked, or it
3209 * will register a callback when the page is unlocked at IO completion. Through
3210 * that callback, io_uring uses task_work to setup a retry of the operation.
3211 * That retry will attempt the buffered read again. The retry will generally
3212 * succeed, or in rare cases where it fails, we then fall back to using the
3213 * async worker threads for a blocking retry.
3215 static bool io_rw_should_retry(struct io_kiocb *req)
3217 struct io_async_rw *rw = req->async_data;
3218 struct wait_page_queue *wait = &rw->wpq;
3219 struct kiocb *kiocb = &req->rw.kiocb;
3221 /* never retry for NOWAIT, we just complete with -EAGAIN */
3222 if (req->flags & REQ_F_NOWAIT)
3225 /* Only for buffered IO */
3226 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3230 * just use poll if we can, and don't attempt if the fs doesn't
3231 * support callback based unlocks
3233 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3236 wait->wait.func = io_async_buf_func;
3237 wait->wait.private = req;
3238 wait->wait.flags = 0;
3239 INIT_LIST_HEAD(&wait->wait.entry);
3240 kiocb->ki_flags |= IOCB_WAITQ;
3241 kiocb->ki_flags &= ~IOCB_NOWAIT;
3242 kiocb->ki_waitq = wait;
3246 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3248 if (req->file->f_op->read_iter)
3249 return call_read_iter(req->file, &req->rw.kiocb, iter);
3250 else if (req->file->f_op->read)
3251 return loop_rw_iter(READ, req, iter);
3256 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3258 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3259 struct kiocb *kiocb = &req->rw.kiocb;
3260 struct iov_iter __iter, *iter = &__iter;
3261 struct io_async_rw *rw = req->async_data;
3262 ssize_t io_size, ret, ret2;
3263 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3269 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3273 io_size = iov_iter_count(iter);
3274 req->result = io_size;
3276 /* Ensure we clear previously set non-block flag */
3277 if (!force_nonblock)
3278 kiocb->ki_flags &= ~IOCB_NOWAIT;
3280 kiocb->ki_flags |= IOCB_NOWAIT;
3282 /* If the file doesn't support async, just async punt */
3283 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3284 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3285 return ret ?: -EAGAIN;
3288 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3289 if (unlikely(ret)) {
3294 ret = io_iter_do_read(req, iter);
3296 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3297 /* IOPOLL retry should happen for io-wq threads */
3298 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3300 /* no retry on NONBLOCK nor RWF_NOWAIT */
3301 if (req->flags & REQ_F_NOWAIT)
3303 /* some cases will consume bytes even on error returns */
3304 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3306 } else if (ret == -EIOCBQUEUED) {
3308 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3309 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3310 /* read all, failed, already did sync or don't want to retry */
3314 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3319 rw = req->async_data;
3320 /* now use our persistent iterator, if we aren't already */
3325 rw->bytes_done += ret;
3326 /* if we can retry, do so with the callbacks armed */
3327 if (!io_rw_should_retry(req)) {
3328 kiocb->ki_flags &= ~IOCB_WAITQ;
3333 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3334 * we get -EIOCBQUEUED, then we'll get a notification when the
3335 * desired page gets unlocked. We can also get a partial read
3336 * here, and if we do, then just retry at the new offset.
3338 ret = io_iter_do_read(req, iter);
3339 if (ret == -EIOCBQUEUED)
3341 /* we got some bytes, but not all. retry. */
3342 kiocb->ki_flags &= ~IOCB_WAITQ;
3343 } while (ret > 0 && ret < io_size);
3345 kiocb_done(kiocb, ret, issue_flags);
3347 /* it's faster to check here then delegate to kfree */
3353 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3355 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3357 return io_prep_rw(req, sqe);
3360 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3362 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3363 struct kiocb *kiocb = &req->rw.kiocb;
3364 struct iov_iter __iter, *iter = &__iter;
3365 struct io_async_rw *rw = req->async_data;
3366 ssize_t ret, ret2, io_size;
3367 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3373 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3377 io_size = iov_iter_count(iter);
3378 req->result = io_size;
3380 /* Ensure we clear previously set non-block flag */
3381 if (!force_nonblock)
3382 kiocb->ki_flags &= ~IOCB_NOWAIT;
3384 kiocb->ki_flags |= IOCB_NOWAIT;
3386 /* If the file doesn't support async, just async punt */
3387 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3390 /* file path doesn't support NOWAIT for non-direct_IO */
3391 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3392 (req->flags & REQ_F_ISREG))
3395 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3400 * Open-code file_start_write here to grab freeze protection,
3401 * which will be released by another thread in
3402 * io_complete_rw(). Fool lockdep by telling it the lock got
3403 * released so that it doesn't complain about the held lock when
3404 * we return to userspace.
3406 if (req->flags & REQ_F_ISREG) {
3407 sb_start_write(file_inode(req->file)->i_sb);
3408 __sb_writers_release(file_inode(req->file)->i_sb,
3411 kiocb->ki_flags |= IOCB_WRITE;
3413 if (req->file->f_op->write_iter)
3414 ret2 = call_write_iter(req->file, kiocb, iter);
3415 else if (req->file->f_op->write)
3416 ret2 = loop_rw_iter(WRITE, req, iter);
3420 if (req->flags & REQ_F_REISSUE)
3424 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3425 * retry them without IOCB_NOWAIT.
3427 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3429 /* no retry on NONBLOCK nor RWF_NOWAIT */
3430 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3432 if (!force_nonblock || ret2 != -EAGAIN) {
3433 /* IOPOLL retry should happen for io-wq threads */
3434 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3437 kiocb_done(kiocb, ret2, issue_flags);
3440 /* some cases will consume bytes even on error returns */
3441 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3442 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3443 return ret ?: -EAGAIN;
3446 /* it's reportedly faster than delegating the null check to kfree() */
3452 static int io_renameat_prep(struct io_kiocb *req,
3453 const struct io_uring_sqe *sqe)
3455 struct io_rename *ren = &req->rename;
3456 const char __user *oldf, *newf;
3458 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3461 ren->old_dfd = READ_ONCE(sqe->fd);
3462 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3463 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3464 ren->new_dfd = READ_ONCE(sqe->len);
3465 ren->flags = READ_ONCE(sqe->rename_flags);
3467 ren->oldpath = getname(oldf);
3468 if (IS_ERR(ren->oldpath))
3469 return PTR_ERR(ren->oldpath);
3471 ren->newpath = getname(newf);
3472 if (IS_ERR(ren->newpath)) {
3473 putname(ren->oldpath);
3474 return PTR_ERR(ren->newpath);
3477 req->flags |= REQ_F_NEED_CLEANUP;
3481 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3483 struct io_rename *ren = &req->rename;
3486 if (issue_flags & IO_URING_F_NONBLOCK)
3489 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3490 ren->newpath, ren->flags);
3492 req->flags &= ~REQ_F_NEED_CLEANUP;
3494 req_set_fail_links(req);
3495 io_req_complete(req, ret);
3499 static int io_unlinkat_prep(struct io_kiocb *req,
3500 const struct io_uring_sqe *sqe)
3502 struct io_unlink *un = &req->unlink;
3503 const char __user *fname;
3505 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3508 un->dfd = READ_ONCE(sqe->fd);
3510 un->flags = READ_ONCE(sqe->unlink_flags);
3511 if (un->flags & ~AT_REMOVEDIR)
3514 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3515 un->filename = getname(fname);
3516 if (IS_ERR(un->filename))
3517 return PTR_ERR(un->filename);
3519 req->flags |= REQ_F_NEED_CLEANUP;
3523 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3525 struct io_unlink *un = &req->unlink;
3528 if (issue_flags & IO_URING_F_NONBLOCK)
3531 if (un->flags & AT_REMOVEDIR)
3532 ret = do_rmdir(un->dfd, un->filename);
3534 ret = do_unlinkat(un->dfd, un->filename);
3536 req->flags &= ~REQ_F_NEED_CLEANUP;
3538 req_set_fail_links(req);
3539 io_req_complete(req, ret);
3543 static int io_shutdown_prep(struct io_kiocb *req,
3544 const struct io_uring_sqe *sqe)
3546 #if defined(CONFIG_NET)
3547 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3549 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3553 req->shutdown.how = READ_ONCE(sqe->len);
3560 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3562 #if defined(CONFIG_NET)
3563 struct socket *sock;
3566 if (issue_flags & IO_URING_F_NONBLOCK)
3569 sock = sock_from_file(req->file);
3570 if (unlikely(!sock))
3573 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3575 req_set_fail_links(req);
3576 io_req_complete(req, ret);
3583 static int __io_splice_prep(struct io_kiocb *req,
3584 const struct io_uring_sqe *sqe)
3586 struct io_splice* sp = &req->splice;
3587 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3589 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3593 sp->len = READ_ONCE(sqe->len);
3594 sp->flags = READ_ONCE(sqe->splice_flags);
3596 if (unlikely(sp->flags & ~valid_flags))
3599 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3600 (sp->flags & SPLICE_F_FD_IN_FIXED));
3603 req->flags |= REQ_F_NEED_CLEANUP;
3605 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3607 * Splice operation will be punted aync, and here need to
3608 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3610 req->work.flags |= IO_WQ_WORK_UNBOUND;
3616 static int io_tee_prep(struct io_kiocb *req,
3617 const struct io_uring_sqe *sqe)
3619 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3621 return __io_splice_prep(req, sqe);
3624 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3626 struct io_splice *sp = &req->splice;
3627 struct file *in = sp->file_in;
3628 struct file *out = sp->file_out;
3629 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3632 if (issue_flags & IO_URING_F_NONBLOCK)
3635 ret = do_tee(in, out, sp->len, flags);
3637 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3638 req->flags &= ~REQ_F_NEED_CLEANUP;
3641 req_set_fail_links(req);
3642 io_req_complete(req, ret);
3646 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3648 struct io_splice* sp = &req->splice;
3650 sp->off_in = READ_ONCE(sqe->splice_off_in);
3651 sp->off_out = READ_ONCE(sqe->off);
3652 return __io_splice_prep(req, sqe);
3655 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3657 struct io_splice *sp = &req->splice;
3658 struct file *in = sp->file_in;
3659 struct file *out = sp->file_out;
3660 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3661 loff_t *poff_in, *poff_out;
3664 if (issue_flags & IO_URING_F_NONBLOCK)
3667 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3668 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3671 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3673 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3674 req->flags &= ~REQ_F_NEED_CLEANUP;
3677 req_set_fail_links(req);
3678 io_req_complete(req, ret);
3683 * IORING_OP_NOP just posts a completion event, nothing else.
3685 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3687 struct io_ring_ctx *ctx = req->ctx;
3689 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3692 __io_req_complete(req, issue_flags, 0, 0);
3696 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3698 struct io_ring_ctx *ctx = req->ctx;
3703 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3705 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3708 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3709 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3712 req->sync.off = READ_ONCE(sqe->off);
3713 req->sync.len = READ_ONCE(sqe->len);
3717 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3719 loff_t end = req->sync.off + req->sync.len;
3722 /* fsync always requires a blocking context */
3723 if (issue_flags & IO_URING_F_NONBLOCK)
3726 ret = vfs_fsync_range(req->file, req->sync.off,
3727 end > 0 ? end : LLONG_MAX,
3728 req->sync.flags & IORING_FSYNC_DATASYNC);
3730 req_set_fail_links(req);
3731 io_req_complete(req, ret);
3735 static int io_fallocate_prep(struct io_kiocb *req,
3736 const struct io_uring_sqe *sqe)
3738 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3740 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3743 req->sync.off = READ_ONCE(sqe->off);
3744 req->sync.len = READ_ONCE(sqe->addr);
3745 req->sync.mode = READ_ONCE(sqe->len);
3749 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3753 /* fallocate always requiring blocking context */
3754 if (issue_flags & IO_URING_F_NONBLOCK)
3756 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3759 req_set_fail_links(req);
3760 io_req_complete(req, ret);
3764 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3766 const char __user *fname;
3769 if (unlikely(sqe->ioprio || sqe->buf_index))
3771 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3774 /* open.how should be already initialised */
3775 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3776 req->open.how.flags |= O_LARGEFILE;
3778 req->open.dfd = READ_ONCE(sqe->fd);
3779 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3780 req->open.filename = getname(fname);
3781 if (IS_ERR(req->open.filename)) {
3782 ret = PTR_ERR(req->open.filename);
3783 req->open.filename = NULL;
3786 req->open.nofile = rlimit(RLIMIT_NOFILE);
3787 req->flags |= REQ_F_NEED_CLEANUP;
3791 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3795 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3797 mode = READ_ONCE(sqe->len);
3798 flags = READ_ONCE(sqe->open_flags);
3799 req->open.how = build_open_how(flags, mode);
3800 return __io_openat_prep(req, sqe);
3803 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3805 struct open_how __user *how;
3809 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3811 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3812 len = READ_ONCE(sqe->len);
3813 if (len < OPEN_HOW_SIZE_VER0)
3816 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3821 return __io_openat_prep(req, sqe);
3824 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3826 struct open_flags op;
3829 bool resolve_nonblock;
3832 ret = build_open_flags(&req->open.how, &op);
3835 nonblock_set = op.open_flag & O_NONBLOCK;
3836 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3837 if (issue_flags & IO_URING_F_NONBLOCK) {
3839 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3840 * it'll always -EAGAIN
3842 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3844 op.lookup_flags |= LOOKUP_CACHED;
3845 op.open_flag |= O_NONBLOCK;
3848 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3852 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3853 /* only retry if RESOLVE_CACHED wasn't already set by application */
3854 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3855 file == ERR_PTR(-EAGAIN)) {
3857 * We could hang on to this 'fd', but seems like marginal
3858 * gain for something that is now known to be a slower path.
3859 * So just put it, and we'll get a new one when we retry.
3867 ret = PTR_ERR(file);
3869 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3870 file->f_flags &= ~O_NONBLOCK;
3871 fsnotify_open(file);
3872 fd_install(ret, file);
3875 putname(req->open.filename);
3876 req->flags &= ~REQ_F_NEED_CLEANUP;
3878 req_set_fail_links(req);
3879 io_req_complete(req, ret);
3883 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3885 return io_openat2(req, issue_flags);
3888 static int io_remove_buffers_prep(struct io_kiocb *req,
3889 const struct io_uring_sqe *sqe)
3891 struct io_provide_buf *p = &req->pbuf;
3894 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3897 tmp = READ_ONCE(sqe->fd);
3898 if (!tmp || tmp > USHRT_MAX)
3901 memset(p, 0, sizeof(*p));
3903 p->bgid = READ_ONCE(sqe->buf_group);
3907 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3908 int bgid, unsigned nbufs)
3912 /* shouldn't happen */
3916 /* the head kbuf is the list itself */
3917 while (!list_empty(&buf->list)) {
3918 struct io_buffer *nxt;
3920 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3921 list_del(&nxt->list);
3928 xa_erase(&ctx->io_buffers, bgid);
3933 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3935 struct io_provide_buf *p = &req->pbuf;
3936 struct io_ring_ctx *ctx = req->ctx;
3937 struct io_buffer *head;
3939 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3941 io_ring_submit_lock(ctx, !force_nonblock);
3943 lockdep_assert_held(&ctx->uring_lock);
3946 head = xa_load(&ctx->io_buffers, p->bgid);
3948 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3950 req_set_fail_links(req);
3952 /* need to hold the lock to complete IOPOLL requests */
3953 if (ctx->flags & IORING_SETUP_IOPOLL) {
3954 __io_req_complete(req, issue_flags, ret, 0);
3955 io_ring_submit_unlock(ctx, !force_nonblock);
3957 io_ring_submit_unlock(ctx, !force_nonblock);
3958 __io_req_complete(req, issue_flags, ret, 0);
3963 static int io_provide_buffers_prep(struct io_kiocb *req,
3964 const struct io_uring_sqe *sqe)
3967 struct io_provide_buf *p = &req->pbuf;
3970 if (sqe->ioprio || sqe->rw_flags)
3973 tmp = READ_ONCE(sqe->fd);
3974 if (!tmp || tmp > USHRT_MAX)
3977 p->addr = READ_ONCE(sqe->addr);
3978 p->len = READ_ONCE(sqe->len);
3980 size = (unsigned long)p->len * p->nbufs;
3981 if (!access_ok(u64_to_user_ptr(p->addr), size))
3984 p->bgid = READ_ONCE(sqe->buf_group);
3985 tmp = READ_ONCE(sqe->off);
3986 if (tmp > USHRT_MAX)
3992 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3994 struct io_buffer *buf;
3995 u64 addr = pbuf->addr;
3996 int i, bid = pbuf->bid;
3998 for (i = 0; i < pbuf->nbufs; i++) {
3999 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4004 buf->len = pbuf->len;
4009 INIT_LIST_HEAD(&buf->list);
4012 list_add_tail(&buf->list, &(*head)->list);
4016 return i ? i : -ENOMEM;
4019 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4021 struct io_provide_buf *p = &req->pbuf;
4022 struct io_ring_ctx *ctx = req->ctx;
4023 struct io_buffer *head, *list;
4025 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4027 io_ring_submit_lock(ctx, !force_nonblock);
4029 lockdep_assert_held(&ctx->uring_lock);
4031 list = head = xa_load(&ctx->io_buffers, p->bgid);
4033 ret = io_add_buffers(p, &head);
4034 if (ret >= 0 && !list) {
4035 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4037 __io_remove_buffers(ctx, head, p->bgid, -1U);
4040 req_set_fail_links(req);
4042 /* need to hold the lock to complete IOPOLL requests */
4043 if (ctx->flags & IORING_SETUP_IOPOLL) {
4044 __io_req_complete(req, issue_flags, ret, 0);
4045 io_ring_submit_unlock(ctx, !force_nonblock);
4047 io_ring_submit_unlock(ctx, !force_nonblock);
4048 __io_req_complete(req, issue_flags, ret, 0);
4053 static int io_epoll_ctl_prep(struct io_kiocb *req,
4054 const struct io_uring_sqe *sqe)
4056 #if defined(CONFIG_EPOLL)
4057 if (sqe->ioprio || sqe->buf_index)
4059 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4062 req->epoll.epfd = READ_ONCE(sqe->fd);
4063 req->epoll.op = READ_ONCE(sqe->len);
4064 req->epoll.fd = READ_ONCE(sqe->off);
4066 if (ep_op_has_event(req->epoll.op)) {
4067 struct epoll_event __user *ev;
4069 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4070 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4080 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4082 #if defined(CONFIG_EPOLL)
4083 struct io_epoll *ie = &req->epoll;
4085 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4087 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4088 if (force_nonblock && ret == -EAGAIN)
4092 req_set_fail_links(req);
4093 __io_req_complete(req, issue_flags, ret, 0);
4100 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4102 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4103 if (sqe->ioprio || sqe->buf_index || sqe->off)
4105 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4108 req->madvise.addr = READ_ONCE(sqe->addr);
4109 req->madvise.len = READ_ONCE(sqe->len);
4110 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4117 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4119 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4120 struct io_madvise *ma = &req->madvise;
4123 if (issue_flags & IO_URING_F_NONBLOCK)
4126 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4128 req_set_fail_links(req);
4129 io_req_complete(req, ret);
4136 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4138 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4140 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4143 req->fadvise.offset = READ_ONCE(sqe->off);
4144 req->fadvise.len = READ_ONCE(sqe->len);
4145 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4149 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4151 struct io_fadvise *fa = &req->fadvise;
4154 if (issue_flags & IO_URING_F_NONBLOCK) {
4155 switch (fa->advice) {
4156 case POSIX_FADV_NORMAL:
4157 case POSIX_FADV_RANDOM:
4158 case POSIX_FADV_SEQUENTIAL:
4165 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4167 req_set_fail_links(req);
4168 io_req_complete(req, ret);
4172 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4174 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4176 if (sqe->ioprio || sqe->buf_index)
4178 if (req->flags & REQ_F_FIXED_FILE)
4181 req->statx.dfd = READ_ONCE(sqe->fd);
4182 req->statx.mask = READ_ONCE(sqe->len);
4183 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4184 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4185 req->statx.flags = READ_ONCE(sqe->statx_flags);
4190 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4192 struct io_statx *ctx = &req->statx;
4195 if (issue_flags & IO_URING_F_NONBLOCK) {
4196 /* only need file table for an actual valid fd */
4197 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4198 req->flags |= REQ_F_NO_FILE_TABLE;
4202 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4206 req_set_fail_links(req);
4207 io_req_complete(req, ret);
4211 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4213 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4215 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4216 sqe->rw_flags || sqe->buf_index)
4218 if (req->flags & REQ_F_FIXED_FILE)
4221 req->close.fd = READ_ONCE(sqe->fd);
4225 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4227 struct files_struct *files = current->files;
4228 struct io_close *close = &req->close;
4229 struct fdtable *fdt;
4235 spin_lock(&files->file_lock);
4236 fdt = files_fdtable(files);
4237 if (close->fd >= fdt->max_fds) {
4238 spin_unlock(&files->file_lock);
4241 file = fdt->fd[close->fd];
4243 spin_unlock(&files->file_lock);
4247 if (file->f_op == &io_uring_fops) {
4248 spin_unlock(&files->file_lock);
4253 /* if the file has a flush method, be safe and punt to async */
4254 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4255 spin_unlock(&files->file_lock);
4259 ret = __close_fd_get_file(close->fd, &file);
4260 spin_unlock(&files->file_lock);
4267 /* No ->flush() or already async, safely close from here */
4268 ret = filp_close(file, current->files);
4271 req_set_fail_links(req);
4274 __io_req_complete(req, issue_flags, ret, 0);
4278 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4280 struct io_ring_ctx *ctx = req->ctx;
4282 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4284 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4287 req->sync.off = READ_ONCE(sqe->off);
4288 req->sync.len = READ_ONCE(sqe->len);
4289 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4293 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4297 /* sync_file_range always requires a blocking context */
4298 if (issue_flags & IO_URING_F_NONBLOCK)
4301 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4304 req_set_fail_links(req);
4305 io_req_complete(req, ret);
4309 #if defined(CONFIG_NET)
4310 static int io_setup_async_msg(struct io_kiocb *req,
4311 struct io_async_msghdr *kmsg)
4313 struct io_async_msghdr *async_msg = req->async_data;
4317 if (io_alloc_async_data(req)) {
4318 kfree(kmsg->free_iov);
4321 async_msg = req->async_data;
4322 req->flags |= REQ_F_NEED_CLEANUP;
4323 memcpy(async_msg, kmsg, sizeof(*kmsg));
4324 async_msg->msg.msg_name = &async_msg->addr;
4325 /* if were using fast_iov, set it to the new one */
4326 if (!async_msg->free_iov)
4327 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4332 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4333 struct io_async_msghdr *iomsg)
4335 iomsg->msg.msg_name = &iomsg->addr;
4336 iomsg->free_iov = iomsg->fast_iov;
4337 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4338 req->sr_msg.msg_flags, &iomsg->free_iov);
4341 static int io_sendmsg_prep_async(struct io_kiocb *req)
4345 if (!io_op_defs[req->opcode].needs_async_data)
4347 ret = io_sendmsg_copy_hdr(req, req->async_data);
4349 req->flags |= REQ_F_NEED_CLEANUP;
4353 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4355 struct io_sr_msg *sr = &req->sr_msg;
4357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4360 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4361 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4362 sr->len = READ_ONCE(sqe->len);
4364 #ifdef CONFIG_COMPAT
4365 if (req->ctx->compat)
4366 sr->msg_flags |= MSG_CMSG_COMPAT;
4371 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4373 struct io_async_msghdr iomsg, *kmsg;
4374 struct socket *sock;
4379 sock = sock_from_file(req->file);
4380 if (unlikely(!sock))
4383 kmsg = req->async_data;
4385 ret = io_sendmsg_copy_hdr(req, &iomsg);
4391 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4392 if (flags & MSG_DONTWAIT)
4393 req->flags |= REQ_F_NOWAIT;
4394 else if (issue_flags & IO_URING_F_NONBLOCK)
4395 flags |= MSG_DONTWAIT;
4397 if (flags & MSG_WAITALL)
4398 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4400 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4401 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4402 return io_setup_async_msg(req, kmsg);
4403 if (ret == -ERESTARTSYS)
4406 /* fast path, check for non-NULL to avoid function call */
4408 kfree(kmsg->free_iov);
4409 req->flags &= ~REQ_F_NEED_CLEANUP;
4411 req_set_fail_links(req);
4412 __io_req_complete(req, issue_flags, ret, 0);
4416 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4418 struct io_sr_msg *sr = &req->sr_msg;
4421 struct socket *sock;
4426 sock = sock_from_file(req->file);
4427 if (unlikely(!sock))
4430 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4434 msg.msg_name = NULL;
4435 msg.msg_control = NULL;
4436 msg.msg_controllen = 0;
4437 msg.msg_namelen = 0;
4439 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4440 if (flags & MSG_DONTWAIT)
4441 req->flags |= REQ_F_NOWAIT;
4442 else if (issue_flags & IO_URING_F_NONBLOCK)
4443 flags |= MSG_DONTWAIT;
4445 if (flags & MSG_WAITALL)
4446 min_ret = iov_iter_count(&msg.msg_iter);
4448 msg.msg_flags = flags;
4449 ret = sock_sendmsg(sock, &msg);
4450 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4452 if (ret == -ERESTARTSYS)
4456 req_set_fail_links(req);
4457 __io_req_complete(req, issue_flags, ret, 0);
4461 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4462 struct io_async_msghdr *iomsg)
4464 struct io_sr_msg *sr = &req->sr_msg;
4465 struct iovec __user *uiov;
4469 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4470 &iomsg->uaddr, &uiov, &iov_len);
4474 if (req->flags & REQ_F_BUFFER_SELECT) {
4477 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4479 sr->len = iomsg->fast_iov[0].iov_len;
4480 iomsg->free_iov = NULL;
4482 iomsg->free_iov = iomsg->fast_iov;
4483 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4484 &iomsg->free_iov, &iomsg->msg.msg_iter,
4493 #ifdef CONFIG_COMPAT
4494 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4495 struct io_async_msghdr *iomsg)
4497 struct compat_msghdr __user *msg_compat;
4498 struct io_sr_msg *sr = &req->sr_msg;
4499 struct compat_iovec __user *uiov;
4504 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4505 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4510 uiov = compat_ptr(ptr);
4511 if (req->flags & REQ_F_BUFFER_SELECT) {
4512 compat_ssize_t clen;
4516 if (!access_ok(uiov, sizeof(*uiov)))
4518 if (__get_user(clen, &uiov->iov_len))
4523 iomsg->free_iov = NULL;
4525 iomsg->free_iov = iomsg->fast_iov;
4526 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4527 UIO_FASTIOV, &iomsg->free_iov,
4528 &iomsg->msg.msg_iter, true);
4537 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4538 struct io_async_msghdr *iomsg)
4540 iomsg->msg.msg_name = &iomsg->addr;
4542 #ifdef CONFIG_COMPAT
4543 if (req->ctx->compat)
4544 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4547 return __io_recvmsg_copy_hdr(req, iomsg);
4550 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4553 struct io_sr_msg *sr = &req->sr_msg;
4554 struct io_buffer *kbuf;
4556 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4561 req->flags |= REQ_F_BUFFER_SELECTED;
4565 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4567 return io_put_kbuf(req, req->sr_msg.kbuf);
4570 static int io_recvmsg_prep_async(struct io_kiocb *req)
4574 if (!io_op_defs[req->opcode].needs_async_data)
4576 ret = io_recvmsg_copy_hdr(req, req->async_data);
4578 req->flags |= REQ_F_NEED_CLEANUP;
4582 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4584 struct io_sr_msg *sr = &req->sr_msg;
4586 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4589 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4590 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4591 sr->len = READ_ONCE(sqe->len);
4592 sr->bgid = READ_ONCE(sqe->buf_group);
4594 #ifdef CONFIG_COMPAT
4595 if (req->ctx->compat)
4596 sr->msg_flags |= MSG_CMSG_COMPAT;
4601 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4603 struct io_async_msghdr iomsg, *kmsg;
4604 struct socket *sock;
4605 struct io_buffer *kbuf;
4608 int ret, cflags = 0;
4609 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4611 sock = sock_from_file(req->file);
4612 if (unlikely(!sock))
4615 kmsg = req->async_data;
4617 ret = io_recvmsg_copy_hdr(req, &iomsg);
4623 if (req->flags & REQ_F_BUFFER_SELECT) {
4624 kbuf = io_recv_buffer_select(req, !force_nonblock);
4626 return PTR_ERR(kbuf);
4627 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4628 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4629 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4630 1, req->sr_msg.len);
4633 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4634 if (flags & MSG_DONTWAIT)
4635 req->flags |= REQ_F_NOWAIT;
4636 else if (force_nonblock)
4637 flags |= MSG_DONTWAIT;
4639 if (flags & MSG_WAITALL)
4640 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4642 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4643 kmsg->uaddr, flags);
4644 if (force_nonblock && ret == -EAGAIN)
4645 return io_setup_async_msg(req, kmsg);
4646 if (ret == -ERESTARTSYS)
4649 if (req->flags & REQ_F_BUFFER_SELECTED)
4650 cflags = io_put_recv_kbuf(req);
4651 /* fast path, check for non-NULL to avoid function call */
4653 kfree(kmsg->free_iov);
4654 req->flags &= ~REQ_F_NEED_CLEANUP;
4655 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4656 req_set_fail_links(req);
4657 __io_req_complete(req, issue_flags, ret, cflags);
4661 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4663 struct io_buffer *kbuf;
4664 struct io_sr_msg *sr = &req->sr_msg;
4666 void __user *buf = sr->buf;
4667 struct socket *sock;
4671 int ret, cflags = 0;
4672 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4674 sock = sock_from_file(req->file);
4675 if (unlikely(!sock))
4678 if (req->flags & REQ_F_BUFFER_SELECT) {
4679 kbuf = io_recv_buffer_select(req, !force_nonblock);
4681 return PTR_ERR(kbuf);
4682 buf = u64_to_user_ptr(kbuf->addr);
4685 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4689 msg.msg_name = NULL;
4690 msg.msg_control = NULL;
4691 msg.msg_controllen = 0;
4692 msg.msg_namelen = 0;
4693 msg.msg_iocb = NULL;
4696 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4697 if (flags & MSG_DONTWAIT)
4698 req->flags |= REQ_F_NOWAIT;
4699 else if (force_nonblock)
4700 flags |= MSG_DONTWAIT;
4702 if (flags & MSG_WAITALL)
4703 min_ret = iov_iter_count(&msg.msg_iter);
4705 ret = sock_recvmsg(sock, &msg, flags);
4706 if (force_nonblock && ret == -EAGAIN)
4708 if (ret == -ERESTARTSYS)
4711 if (req->flags & REQ_F_BUFFER_SELECTED)
4712 cflags = io_put_recv_kbuf(req);
4713 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4714 req_set_fail_links(req);
4715 __io_req_complete(req, issue_flags, ret, cflags);
4719 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4721 struct io_accept *accept = &req->accept;
4723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4725 if (sqe->ioprio || sqe->len || sqe->buf_index)
4728 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4729 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4730 accept->flags = READ_ONCE(sqe->accept_flags);
4731 accept->nofile = rlimit(RLIMIT_NOFILE);
4735 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4737 struct io_accept *accept = &req->accept;
4738 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4739 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4742 if (req->file->f_flags & O_NONBLOCK)
4743 req->flags |= REQ_F_NOWAIT;
4745 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4746 accept->addr_len, accept->flags,
4748 if (ret == -EAGAIN && force_nonblock)
4751 if (ret == -ERESTARTSYS)
4753 req_set_fail_links(req);
4755 __io_req_complete(req, issue_flags, ret, 0);
4759 static int io_connect_prep_async(struct io_kiocb *req)
4761 struct io_async_connect *io = req->async_data;
4762 struct io_connect *conn = &req->connect;
4764 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4767 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4769 struct io_connect *conn = &req->connect;
4771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4773 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4776 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4777 conn->addr_len = READ_ONCE(sqe->addr2);
4781 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4783 struct io_async_connect __io, *io;
4784 unsigned file_flags;
4786 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4788 if (req->async_data) {
4789 io = req->async_data;
4791 ret = move_addr_to_kernel(req->connect.addr,
4792 req->connect.addr_len,
4799 file_flags = force_nonblock ? O_NONBLOCK : 0;
4801 ret = __sys_connect_file(req->file, &io->address,
4802 req->connect.addr_len, file_flags);
4803 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4804 if (req->async_data)
4806 if (io_alloc_async_data(req)) {
4810 memcpy(req->async_data, &__io, sizeof(__io));
4813 if (ret == -ERESTARTSYS)
4817 req_set_fail_links(req);
4818 __io_req_complete(req, issue_flags, ret, 0);
4821 #else /* !CONFIG_NET */
4822 #define IO_NETOP_FN(op) \
4823 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4825 return -EOPNOTSUPP; \
4828 #define IO_NETOP_PREP(op) \
4830 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4832 return -EOPNOTSUPP; \
4835 #define IO_NETOP_PREP_ASYNC(op) \
4837 static int io_##op##_prep_async(struct io_kiocb *req) \
4839 return -EOPNOTSUPP; \
4842 IO_NETOP_PREP_ASYNC(sendmsg);
4843 IO_NETOP_PREP_ASYNC(recvmsg);
4844 IO_NETOP_PREP_ASYNC(connect);
4845 IO_NETOP_PREP(accept);
4848 #endif /* CONFIG_NET */
4850 struct io_poll_table {
4851 struct poll_table_struct pt;
4852 struct io_kiocb *req;
4856 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4857 __poll_t mask, task_work_func_t func)
4861 /* for instances that support it check for an event match first: */
4862 if (mask && !(mask & poll->events))
4865 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4867 list_del_init(&poll->wait.entry);
4870 req->task_work.func = func;
4871 percpu_ref_get(&req->ctx->refs);
4874 * If this fails, then the task is exiting. When a task exits, the
4875 * work gets canceled, so just cancel this request as well instead
4876 * of executing it. We can't safely execute it anyway, as we may not
4877 * have the needed state needed for it anyway.
4879 ret = io_req_task_work_add(req);
4880 if (unlikely(ret)) {
4881 WRITE_ONCE(poll->canceled, true);
4882 io_req_task_work_add_fallback(req, func);
4887 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4888 __acquires(&req->ctx->completion_lock)
4890 struct io_ring_ctx *ctx = req->ctx;
4892 if (!req->result && !READ_ONCE(poll->canceled)) {
4893 struct poll_table_struct pt = { ._key = poll->events };
4895 req->result = vfs_poll(req->file, &pt) & poll->events;
4898 spin_lock_irq(&ctx->completion_lock);
4899 if (!req->result && !READ_ONCE(poll->canceled)) {
4900 add_wait_queue(poll->head, &poll->wait);
4907 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4909 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4910 if (req->opcode == IORING_OP_POLL_ADD)
4911 return req->async_data;
4912 return req->apoll->double_poll;
4915 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4917 if (req->opcode == IORING_OP_POLL_ADD)
4919 return &req->apoll->poll;
4922 static void io_poll_remove_double(struct io_kiocb *req)
4924 struct io_poll_iocb *poll = io_poll_get_double(req);
4926 lockdep_assert_held(&req->ctx->completion_lock);
4928 if (poll && poll->head) {
4929 struct wait_queue_head *head = poll->head;
4931 spin_lock(&head->lock);
4932 list_del_init(&poll->wait.entry);
4933 if (poll->wait.private)
4934 refcount_dec(&req->refs);
4936 spin_unlock(&head->lock);
4940 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4942 struct io_ring_ctx *ctx = req->ctx;
4944 io_poll_remove_double(req);
4945 req->poll.done = true;
4946 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4947 io_commit_cqring(ctx);
4950 static void io_poll_task_func(struct callback_head *cb)
4952 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4953 struct io_ring_ctx *ctx = req->ctx;
4954 struct io_kiocb *nxt;
4956 if (io_poll_rewait(req, &req->poll)) {
4957 spin_unlock_irq(&ctx->completion_lock);
4959 hash_del(&req->hash_node);
4960 io_poll_complete(req, req->result, 0);
4961 spin_unlock_irq(&ctx->completion_lock);
4963 nxt = io_put_req_find_next(req);
4964 io_cqring_ev_posted(ctx);
4966 __io_req_task_submit(nxt);
4969 percpu_ref_put(&ctx->refs);
4972 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4973 int sync, void *key)
4975 struct io_kiocb *req = wait->private;
4976 struct io_poll_iocb *poll = io_poll_get_single(req);
4977 __poll_t mask = key_to_poll(key);
4979 /* for instances that support it check for an event match first: */
4980 if (mask && !(mask & poll->events))
4983 list_del_init(&wait->entry);
4985 if (poll && poll->head) {
4988 spin_lock(&poll->head->lock);
4989 done = list_empty(&poll->wait.entry);
4991 list_del_init(&poll->wait.entry);
4992 /* make sure double remove sees this as being gone */
4993 wait->private = NULL;
4994 spin_unlock(&poll->head->lock);
4996 /* use wait func handler, so it matches the rq type */
4997 poll->wait.func(&poll->wait, mode, sync, key);
5000 refcount_dec(&req->refs);
5004 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5005 wait_queue_func_t wake_func)
5009 poll->canceled = false;
5010 poll->events = events;
5011 INIT_LIST_HEAD(&poll->wait.entry);
5012 init_waitqueue_func_entry(&poll->wait, wake_func);
5015 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5016 struct wait_queue_head *head,
5017 struct io_poll_iocb **poll_ptr)
5019 struct io_kiocb *req = pt->req;
5022 * If poll->head is already set, it's because the file being polled
5023 * uses multiple waitqueues for poll handling (eg one for read, one
5024 * for write). Setup a separate io_poll_iocb if this happens.
5026 if (unlikely(poll->head)) {
5027 struct io_poll_iocb *poll_one = poll;
5029 /* already have a 2nd entry, fail a third attempt */
5031 pt->error = -EINVAL;
5034 /* double add on the same waitqueue head, ignore */
5035 if (poll->head == head)
5037 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5039 pt->error = -ENOMEM;
5042 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5043 refcount_inc(&req->refs);
5044 poll->wait.private = req;
5051 if (poll->events & EPOLLEXCLUSIVE)
5052 add_wait_queue_exclusive(head, &poll->wait);
5054 add_wait_queue(head, &poll->wait);
5057 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5058 struct poll_table_struct *p)
5060 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5061 struct async_poll *apoll = pt->req->apoll;
5063 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5066 static void io_async_task_func(struct callback_head *cb)
5068 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5069 struct async_poll *apoll = req->apoll;
5070 struct io_ring_ctx *ctx = req->ctx;
5072 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5074 if (io_poll_rewait(req, &apoll->poll)) {
5075 spin_unlock_irq(&ctx->completion_lock);
5076 percpu_ref_put(&ctx->refs);
5080 /* If req is still hashed, it cannot have been canceled. Don't check. */
5081 if (hash_hashed(&req->hash_node))
5082 hash_del(&req->hash_node);
5084 io_poll_remove_double(req);
5085 spin_unlock_irq(&ctx->completion_lock);
5087 if (!READ_ONCE(apoll->poll.canceled))
5088 __io_req_task_submit(req);
5090 __io_req_task_cancel(req, -ECANCELED);
5092 percpu_ref_put(&ctx->refs);
5093 kfree(apoll->double_poll);
5097 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5100 struct io_kiocb *req = wait->private;
5101 struct io_poll_iocb *poll = &req->apoll->poll;
5103 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5106 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5109 static void io_poll_req_insert(struct io_kiocb *req)
5111 struct io_ring_ctx *ctx = req->ctx;
5112 struct hlist_head *list;
5114 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5115 hlist_add_head(&req->hash_node, list);
5118 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5119 struct io_poll_iocb *poll,
5120 struct io_poll_table *ipt, __poll_t mask,
5121 wait_queue_func_t wake_func)
5122 __acquires(&ctx->completion_lock)
5124 struct io_ring_ctx *ctx = req->ctx;
5125 bool cancel = false;
5127 INIT_HLIST_NODE(&req->hash_node);
5128 io_init_poll_iocb(poll, mask, wake_func);
5129 poll->file = req->file;
5130 poll->wait.private = req;
5132 ipt->pt._key = mask;
5134 ipt->error = -EINVAL;
5136 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5138 spin_lock_irq(&ctx->completion_lock);
5139 if (likely(poll->head)) {
5140 spin_lock(&poll->head->lock);
5141 if (unlikely(list_empty(&poll->wait.entry))) {
5147 if (mask || ipt->error)
5148 list_del_init(&poll->wait.entry);
5150 WRITE_ONCE(poll->canceled, true);
5151 else if (!poll->done) /* actually waiting for an event */
5152 io_poll_req_insert(req);
5153 spin_unlock(&poll->head->lock);
5159 static bool io_arm_poll_handler(struct io_kiocb *req)
5161 const struct io_op_def *def = &io_op_defs[req->opcode];
5162 struct io_ring_ctx *ctx = req->ctx;
5163 struct async_poll *apoll;
5164 struct io_poll_table ipt;
5168 if (!req->file || !file_can_poll(req->file))
5170 if (req->flags & REQ_F_POLLED)
5174 else if (def->pollout)
5178 /* if we can't nonblock try, then no point in arming a poll handler */
5179 if (!io_file_supports_async(req->file, rw))
5182 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5183 if (unlikely(!apoll))
5185 apoll->double_poll = NULL;
5187 req->flags |= REQ_F_POLLED;
5192 mask |= POLLIN | POLLRDNORM;
5194 mask |= POLLOUT | POLLWRNORM;
5196 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5197 if ((req->opcode == IORING_OP_RECVMSG) &&
5198 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5201 mask |= POLLERR | POLLPRI;
5203 ipt.pt._qproc = io_async_queue_proc;
5205 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5207 if (ret || ipt.error) {
5208 io_poll_remove_double(req);
5209 spin_unlock_irq(&ctx->completion_lock);
5210 kfree(apoll->double_poll);
5214 spin_unlock_irq(&ctx->completion_lock);
5215 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5216 apoll->poll.events);
5220 static bool __io_poll_remove_one(struct io_kiocb *req,
5221 struct io_poll_iocb *poll)
5223 bool do_complete = false;
5225 spin_lock(&poll->head->lock);
5226 WRITE_ONCE(poll->canceled, true);
5227 if (!list_empty(&poll->wait.entry)) {
5228 list_del_init(&poll->wait.entry);
5231 spin_unlock(&poll->head->lock);
5232 hash_del(&req->hash_node);
5236 static bool io_poll_remove_one(struct io_kiocb *req)
5240 io_poll_remove_double(req);
5242 if (req->opcode == IORING_OP_POLL_ADD) {
5243 do_complete = __io_poll_remove_one(req, &req->poll);
5245 struct async_poll *apoll = req->apoll;
5247 /* non-poll requests have submit ref still */
5248 do_complete = __io_poll_remove_one(req, &apoll->poll);
5251 kfree(apoll->double_poll);
5257 io_cqring_fill_event(req, -ECANCELED);
5258 io_commit_cqring(req->ctx);
5259 req_set_fail_links(req);
5260 io_put_req_deferred(req, 1);
5267 * Returns true if we found and killed one or more poll requests
5269 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5270 struct files_struct *files)
5272 struct hlist_node *tmp;
5273 struct io_kiocb *req;
5276 spin_lock_irq(&ctx->completion_lock);
5277 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5278 struct hlist_head *list;
5280 list = &ctx->cancel_hash[i];
5281 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5282 if (io_match_task(req, tsk, files))
5283 posted += io_poll_remove_one(req);
5286 spin_unlock_irq(&ctx->completion_lock);
5289 io_cqring_ev_posted(ctx);
5294 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5296 struct hlist_head *list;
5297 struct io_kiocb *req;
5299 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5300 hlist_for_each_entry(req, list, hash_node) {
5301 if (sqe_addr != req->user_data)
5303 if (io_poll_remove_one(req))
5311 static int io_poll_remove_prep(struct io_kiocb *req,
5312 const struct io_uring_sqe *sqe)
5314 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5316 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5320 req->poll_remove.addr = READ_ONCE(sqe->addr);
5325 * Find a running poll command that matches one specified in sqe->addr,
5326 * and remove it if found.
5328 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5330 struct io_ring_ctx *ctx = req->ctx;
5333 spin_lock_irq(&ctx->completion_lock);
5334 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5335 spin_unlock_irq(&ctx->completion_lock);
5338 req_set_fail_links(req);
5339 io_req_complete(req, ret);
5343 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5346 struct io_kiocb *req = wait->private;
5347 struct io_poll_iocb *poll = &req->poll;
5349 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5352 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5353 struct poll_table_struct *p)
5355 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5357 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5360 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5362 struct io_poll_iocb *poll = &req->poll;
5365 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5367 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5370 events = READ_ONCE(sqe->poll32_events);
5372 events = swahw32(events);
5374 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5375 (events & EPOLLEXCLUSIVE);
5379 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5381 struct io_poll_iocb *poll = &req->poll;
5382 struct io_ring_ctx *ctx = req->ctx;
5383 struct io_poll_table ipt;
5386 ipt.pt._qproc = io_poll_queue_proc;
5388 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5391 if (mask) { /* no async, we'd stolen it */
5393 io_poll_complete(req, mask, 0);
5395 spin_unlock_irq(&ctx->completion_lock);
5398 io_cqring_ev_posted(ctx);
5404 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5406 struct io_timeout_data *data = container_of(timer,
5407 struct io_timeout_data, timer);
5408 struct io_kiocb *req = data->req;
5409 struct io_ring_ctx *ctx = req->ctx;
5410 unsigned long flags;
5412 spin_lock_irqsave(&ctx->completion_lock, flags);
5413 list_del_init(&req->timeout.list);
5414 atomic_set(&req->ctx->cq_timeouts,
5415 atomic_read(&req->ctx->cq_timeouts) + 1);
5417 io_cqring_fill_event(req, -ETIME);
5418 io_commit_cqring(ctx);
5419 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5421 io_cqring_ev_posted(ctx);
5422 req_set_fail_links(req);
5424 return HRTIMER_NORESTART;
5427 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5430 struct io_timeout_data *io;
5431 struct io_kiocb *req;
5434 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5435 if (user_data == req->user_data) {
5442 return ERR_PTR(ret);
5444 io = req->async_data;
5445 ret = hrtimer_try_to_cancel(&io->timer);
5447 return ERR_PTR(-EALREADY);
5448 list_del_init(&req->timeout.list);
5452 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5454 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5457 return PTR_ERR(req);
5459 req_set_fail_links(req);
5460 io_cqring_fill_event(req, -ECANCELED);
5461 io_put_req_deferred(req, 1);
5465 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5466 struct timespec64 *ts, enum hrtimer_mode mode)
5468 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5469 struct io_timeout_data *data;
5472 return PTR_ERR(req);
5474 req->timeout.off = 0; /* noseq */
5475 data = req->async_data;
5476 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5477 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5478 data->timer.function = io_timeout_fn;
5479 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5483 static int io_timeout_remove_prep(struct io_kiocb *req,
5484 const struct io_uring_sqe *sqe)
5486 struct io_timeout_rem *tr = &req->timeout_rem;
5488 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5490 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5492 if (sqe->ioprio || sqe->buf_index || sqe->len)
5495 tr->addr = READ_ONCE(sqe->addr);
5496 tr->flags = READ_ONCE(sqe->timeout_flags);
5497 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5498 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5500 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5502 } else if (tr->flags) {
5503 /* timeout removal doesn't support flags */
5510 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5512 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5517 * Remove or update an existing timeout command
5519 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5521 struct io_timeout_rem *tr = &req->timeout_rem;
5522 struct io_ring_ctx *ctx = req->ctx;
5525 spin_lock_irq(&ctx->completion_lock);
5526 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5527 ret = io_timeout_cancel(ctx, tr->addr);
5529 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5530 io_translate_timeout_mode(tr->flags));
5532 io_cqring_fill_event(req, ret);
5533 io_commit_cqring(ctx);
5534 spin_unlock_irq(&ctx->completion_lock);
5535 io_cqring_ev_posted(ctx);
5537 req_set_fail_links(req);
5542 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5543 bool is_timeout_link)
5545 struct io_timeout_data *data;
5547 u32 off = READ_ONCE(sqe->off);
5549 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5551 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5553 if (off && is_timeout_link)
5555 flags = READ_ONCE(sqe->timeout_flags);
5556 if (flags & ~IORING_TIMEOUT_ABS)
5559 req->timeout.off = off;
5561 if (!req->async_data && io_alloc_async_data(req))
5564 data = req->async_data;
5567 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5570 data->mode = io_translate_timeout_mode(flags);
5571 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5572 if (is_timeout_link)
5573 io_req_track_inflight(req);
5577 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5579 struct io_ring_ctx *ctx = req->ctx;
5580 struct io_timeout_data *data = req->async_data;
5581 struct list_head *entry;
5582 u32 tail, off = req->timeout.off;
5584 spin_lock_irq(&ctx->completion_lock);
5587 * sqe->off holds how many events that need to occur for this
5588 * timeout event to be satisfied. If it isn't set, then this is
5589 * a pure timeout request, sequence isn't used.
5591 if (io_is_timeout_noseq(req)) {
5592 entry = ctx->timeout_list.prev;
5596 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5597 req->timeout.target_seq = tail + off;
5599 /* Update the last seq here in case io_flush_timeouts() hasn't.
5600 * This is safe because ->completion_lock is held, and submissions
5601 * and completions are never mixed in the same ->completion_lock section.
5603 ctx->cq_last_tm_flush = tail;
5606 * Insertion sort, ensuring the first entry in the list is always
5607 * the one we need first.
5609 list_for_each_prev(entry, &ctx->timeout_list) {
5610 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5613 if (io_is_timeout_noseq(nxt))
5615 /* nxt.seq is behind @tail, otherwise would've been completed */
5616 if (off >= nxt->timeout.target_seq - tail)
5620 list_add(&req->timeout.list, entry);
5621 data->timer.function = io_timeout_fn;
5622 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5623 spin_unlock_irq(&ctx->completion_lock);
5627 struct io_cancel_data {
5628 struct io_ring_ctx *ctx;
5632 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5634 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5635 struct io_cancel_data *cd = data;
5637 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5640 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5641 struct io_ring_ctx *ctx)
5643 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5644 enum io_wq_cancel cancel_ret;
5647 if (!tctx || !tctx->io_wq)
5650 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5651 switch (cancel_ret) {
5652 case IO_WQ_CANCEL_OK:
5655 case IO_WQ_CANCEL_RUNNING:
5658 case IO_WQ_CANCEL_NOTFOUND:
5666 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5667 struct io_kiocb *req, __u64 sqe_addr,
5670 unsigned long flags;
5673 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5674 if (ret != -ENOENT) {
5675 spin_lock_irqsave(&ctx->completion_lock, flags);
5679 spin_lock_irqsave(&ctx->completion_lock, flags);
5680 ret = io_timeout_cancel(ctx, sqe_addr);
5683 ret = io_poll_cancel(ctx, sqe_addr);
5687 io_cqring_fill_event(req, ret);
5688 io_commit_cqring(ctx);
5689 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5690 io_cqring_ev_posted(ctx);
5693 req_set_fail_links(req);
5697 static int io_async_cancel_prep(struct io_kiocb *req,
5698 const struct io_uring_sqe *sqe)
5700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5702 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5704 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5707 req->cancel.addr = READ_ONCE(sqe->addr);
5711 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5713 struct io_ring_ctx *ctx = req->ctx;
5714 u64 sqe_addr = req->cancel.addr;
5715 struct io_tctx_node *node;
5718 /* tasks should wait for their io-wq threads, so safe w/o sync */
5719 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5720 spin_lock_irq(&ctx->completion_lock);
5723 ret = io_timeout_cancel(ctx, sqe_addr);
5726 ret = io_poll_cancel(ctx, sqe_addr);
5729 spin_unlock_irq(&ctx->completion_lock);
5731 /* slow path, try all io-wq's */
5732 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5734 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5735 struct io_uring_task *tctx = node->task->io_uring;
5737 if (!tctx || !tctx->io_wq)
5739 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5743 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5745 spin_lock_irq(&ctx->completion_lock);
5747 io_cqring_fill_event(req, ret);
5748 io_commit_cqring(ctx);
5749 spin_unlock_irq(&ctx->completion_lock);
5750 io_cqring_ev_posted(ctx);
5753 req_set_fail_links(req);
5758 static int io_rsrc_update_prep(struct io_kiocb *req,
5759 const struct io_uring_sqe *sqe)
5761 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5763 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5765 if (sqe->ioprio || sqe->rw_flags)
5768 req->rsrc_update.offset = READ_ONCE(sqe->off);
5769 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5770 if (!req->rsrc_update.nr_args)
5772 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5776 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5778 struct io_ring_ctx *ctx = req->ctx;
5779 struct io_uring_rsrc_update up;
5782 if (issue_flags & IO_URING_F_NONBLOCK)
5785 up.offset = req->rsrc_update.offset;
5786 up.data = req->rsrc_update.arg;
5788 mutex_lock(&ctx->uring_lock);
5789 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5790 mutex_unlock(&ctx->uring_lock);
5793 req_set_fail_links(req);
5794 __io_req_complete(req, issue_flags, ret, 0);
5798 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5800 switch (req->opcode) {
5803 case IORING_OP_READV:
5804 case IORING_OP_READ_FIXED:
5805 case IORING_OP_READ:
5806 return io_read_prep(req, sqe);
5807 case IORING_OP_WRITEV:
5808 case IORING_OP_WRITE_FIXED:
5809 case IORING_OP_WRITE:
5810 return io_write_prep(req, sqe);
5811 case IORING_OP_POLL_ADD:
5812 return io_poll_add_prep(req, sqe);
5813 case IORING_OP_POLL_REMOVE:
5814 return io_poll_remove_prep(req, sqe);
5815 case IORING_OP_FSYNC:
5816 return io_fsync_prep(req, sqe);
5817 case IORING_OP_SYNC_FILE_RANGE:
5818 return io_sfr_prep(req, sqe);
5819 case IORING_OP_SENDMSG:
5820 case IORING_OP_SEND:
5821 return io_sendmsg_prep(req, sqe);
5822 case IORING_OP_RECVMSG:
5823 case IORING_OP_RECV:
5824 return io_recvmsg_prep(req, sqe);
5825 case IORING_OP_CONNECT:
5826 return io_connect_prep(req, sqe);
5827 case IORING_OP_TIMEOUT:
5828 return io_timeout_prep(req, sqe, false);
5829 case IORING_OP_TIMEOUT_REMOVE:
5830 return io_timeout_remove_prep(req, sqe);
5831 case IORING_OP_ASYNC_CANCEL:
5832 return io_async_cancel_prep(req, sqe);
5833 case IORING_OP_LINK_TIMEOUT:
5834 return io_timeout_prep(req, sqe, true);
5835 case IORING_OP_ACCEPT:
5836 return io_accept_prep(req, sqe);
5837 case IORING_OP_FALLOCATE:
5838 return io_fallocate_prep(req, sqe);
5839 case IORING_OP_OPENAT:
5840 return io_openat_prep(req, sqe);
5841 case IORING_OP_CLOSE:
5842 return io_close_prep(req, sqe);
5843 case IORING_OP_FILES_UPDATE:
5844 return io_rsrc_update_prep(req, sqe);
5845 case IORING_OP_STATX:
5846 return io_statx_prep(req, sqe);
5847 case IORING_OP_FADVISE:
5848 return io_fadvise_prep(req, sqe);
5849 case IORING_OP_MADVISE:
5850 return io_madvise_prep(req, sqe);
5851 case IORING_OP_OPENAT2:
5852 return io_openat2_prep(req, sqe);
5853 case IORING_OP_EPOLL_CTL:
5854 return io_epoll_ctl_prep(req, sqe);
5855 case IORING_OP_SPLICE:
5856 return io_splice_prep(req, sqe);
5857 case IORING_OP_PROVIDE_BUFFERS:
5858 return io_provide_buffers_prep(req, sqe);
5859 case IORING_OP_REMOVE_BUFFERS:
5860 return io_remove_buffers_prep(req, sqe);
5862 return io_tee_prep(req, sqe);
5863 case IORING_OP_SHUTDOWN:
5864 return io_shutdown_prep(req, sqe);
5865 case IORING_OP_RENAMEAT:
5866 return io_renameat_prep(req, sqe);
5867 case IORING_OP_UNLINKAT:
5868 return io_unlinkat_prep(req, sqe);
5871 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5876 static int io_req_prep_async(struct io_kiocb *req)
5878 switch (req->opcode) {
5879 case IORING_OP_READV:
5880 case IORING_OP_READ_FIXED:
5881 case IORING_OP_READ:
5882 return io_rw_prep_async(req, READ);
5883 case IORING_OP_WRITEV:
5884 case IORING_OP_WRITE_FIXED:
5885 case IORING_OP_WRITE:
5886 return io_rw_prep_async(req, WRITE);
5887 case IORING_OP_SENDMSG:
5888 case IORING_OP_SEND:
5889 return io_sendmsg_prep_async(req);
5890 case IORING_OP_RECVMSG:
5891 case IORING_OP_RECV:
5892 return io_recvmsg_prep_async(req);
5893 case IORING_OP_CONNECT:
5894 return io_connect_prep_async(req);
5899 static int io_req_defer_prep(struct io_kiocb *req)
5901 if (!io_op_defs[req->opcode].needs_async_data)
5903 /* some opcodes init it during the inital prep */
5904 if (req->async_data)
5906 if (__io_alloc_async_data(req))
5908 return io_req_prep_async(req);
5911 static u32 io_get_sequence(struct io_kiocb *req)
5913 struct io_kiocb *pos;
5914 struct io_ring_ctx *ctx = req->ctx;
5915 u32 total_submitted, nr_reqs = 0;
5917 io_for_each_link(pos, req)
5920 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5921 return total_submitted - nr_reqs;
5924 static int io_req_defer(struct io_kiocb *req)
5926 struct io_ring_ctx *ctx = req->ctx;
5927 struct io_defer_entry *de;
5931 /* Still need defer if there is pending req in defer list. */
5932 if (likely(list_empty_careful(&ctx->defer_list) &&
5933 !(req->flags & REQ_F_IO_DRAIN)))
5936 seq = io_get_sequence(req);
5937 /* Still a chance to pass the sequence check */
5938 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5941 ret = io_req_defer_prep(req);
5944 io_prep_async_link(req);
5945 de = kmalloc(sizeof(*de), GFP_KERNEL);
5949 spin_lock_irq(&ctx->completion_lock);
5950 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5951 spin_unlock_irq(&ctx->completion_lock);
5953 io_queue_async_work(req);
5954 return -EIOCBQUEUED;
5957 trace_io_uring_defer(ctx, req, req->user_data);
5960 list_add_tail(&de->list, &ctx->defer_list);
5961 spin_unlock_irq(&ctx->completion_lock);
5962 return -EIOCBQUEUED;
5965 static void __io_clean_op(struct io_kiocb *req)
5967 if (req->flags & REQ_F_BUFFER_SELECTED) {
5968 switch (req->opcode) {
5969 case IORING_OP_READV:
5970 case IORING_OP_READ_FIXED:
5971 case IORING_OP_READ:
5972 kfree((void *)(unsigned long)req->rw.addr);
5974 case IORING_OP_RECVMSG:
5975 case IORING_OP_RECV:
5976 kfree(req->sr_msg.kbuf);
5979 req->flags &= ~REQ_F_BUFFER_SELECTED;
5982 if (req->flags & REQ_F_NEED_CLEANUP) {
5983 switch (req->opcode) {
5984 case IORING_OP_READV:
5985 case IORING_OP_READ_FIXED:
5986 case IORING_OP_READ:
5987 case IORING_OP_WRITEV:
5988 case IORING_OP_WRITE_FIXED:
5989 case IORING_OP_WRITE: {
5990 struct io_async_rw *io = req->async_data;
5992 kfree(io->free_iovec);
5995 case IORING_OP_RECVMSG:
5996 case IORING_OP_SENDMSG: {
5997 struct io_async_msghdr *io = req->async_data;
5999 kfree(io->free_iov);
6002 case IORING_OP_SPLICE:
6004 io_put_file(req, req->splice.file_in,
6005 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6007 case IORING_OP_OPENAT:
6008 case IORING_OP_OPENAT2:
6009 if (req->open.filename)
6010 putname(req->open.filename);
6012 case IORING_OP_RENAMEAT:
6013 putname(req->rename.oldpath);
6014 putname(req->rename.newpath);
6016 case IORING_OP_UNLINKAT:
6017 putname(req->unlink.filename);
6020 req->flags &= ~REQ_F_NEED_CLEANUP;
6024 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6026 struct io_ring_ctx *ctx = req->ctx;
6027 const struct cred *creds = NULL;
6030 if (req->work.creds && req->work.creds != current_cred())
6031 creds = override_creds(req->work.creds);
6033 switch (req->opcode) {
6035 ret = io_nop(req, issue_flags);
6037 case IORING_OP_READV:
6038 case IORING_OP_READ_FIXED:
6039 case IORING_OP_READ:
6040 ret = io_read(req, issue_flags);
6042 case IORING_OP_WRITEV:
6043 case IORING_OP_WRITE_FIXED:
6044 case IORING_OP_WRITE:
6045 ret = io_write(req, issue_flags);
6047 case IORING_OP_FSYNC:
6048 ret = io_fsync(req, issue_flags);
6050 case IORING_OP_POLL_ADD:
6051 ret = io_poll_add(req, issue_flags);
6053 case IORING_OP_POLL_REMOVE:
6054 ret = io_poll_remove(req, issue_flags);
6056 case IORING_OP_SYNC_FILE_RANGE:
6057 ret = io_sync_file_range(req, issue_flags);
6059 case IORING_OP_SENDMSG:
6060 ret = io_sendmsg(req, issue_flags);
6062 case IORING_OP_SEND:
6063 ret = io_send(req, issue_flags);
6065 case IORING_OP_RECVMSG:
6066 ret = io_recvmsg(req, issue_flags);
6068 case IORING_OP_RECV:
6069 ret = io_recv(req, issue_flags);
6071 case IORING_OP_TIMEOUT:
6072 ret = io_timeout(req, issue_flags);
6074 case IORING_OP_TIMEOUT_REMOVE:
6075 ret = io_timeout_remove(req, issue_flags);
6077 case IORING_OP_ACCEPT:
6078 ret = io_accept(req, issue_flags);
6080 case IORING_OP_CONNECT:
6081 ret = io_connect(req, issue_flags);
6083 case IORING_OP_ASYNC_CANCEL:
6084 ret = io_async_cancel(req, issue_flags);
6086 case IORING_OP_FALLOCATE:
6087 ret = io_fallocate(req, issue_flags);
6089 case IORING_OP_OPENAT:
6090 ret = io_openat(req, issue_flags);
6092 case IORING_OP_CLOSE:
6093 ret = io_close(req, issue_flags);
6095 case IORING_OP_FILES_UPDATE:
6096 ret = io_files_update(req, issue_flags);
6098 case IORING_OP_STATX:
6099 ret = io_statx(req, issue_flags);
6101 case IORING_OP_FADVISE:
6102 ret = io_fadvise(req, issue_flags);
6104 case IORING_OP_MADVISE:
6105 ret = io_madvise(req, issue_flags);
6107 case IORING_OP_OPENAT2:
6108 ret = io_openat2(req, issue_flags);
6110 case IORING_OP_EPOLL_CTL:
6111 ret = io_epoll_ctl(req, issue_flags);
6113 case IORING_OP_SPLICE:
6114 ret = io_splice(req, issue_flags);
6116 case IORING_OP_PROVIDE_BUFFERS:
6117 ret = io_provide_buffers(req, issue_flags);
6119 case IORING_OP_REMOVE_BUFFERS:
6120 ret = io_remove_buffers(req, issue_flags);
6123 ret = io_tee(req, issue_flags);
6125 case IORING_OP_SHUTDOWN:
6126 ret = io_shutdown(req, issue_flags);
6128 case IORING_OP_RENAMEAT:
6129 ret = io_renameat(req, issue_flags);
6131 case IORING_OP_UNLINKAT:
6132 ret = io_unlinkat(req, issue_flags);
6140 revert_creds(creds);
6145 /* If the op doesn't have a file, we're not polling for it */
6146 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6147 const bool in_async = io_wq_current_is_worker();
6149 /* workqueue context doesn't hold uring_lock, grab it now */
6151 mutex_lock(&ctx->uring_lock);
6153 io_iopoll_req_issued(req, in_async);
6156 mutex_unlock(&ctx->uring_lock);
6162 static void io_wq_submit_work(struct io_wq_work *work)
6164 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6165 struct io_kiocb *timeout;
6168 timeout = io_prep_linked_timeout(req);
6170 io_queue_linked_timeout(timeout);
6172 if (work->flags & IO_WQ_WORK_CANCEL)
6176 req->flags &= ~REQ_F_REISSUE;
6178 ret = io_issue_sqe(req, 0);
6180 * We can get EAGAIN for polled IO even though we're
6181 * forcing a sync submission from here, since we can't
6182 * wait for request slots on the block side.
6190 /* avoid locking problems by failing it from a clean context */
6192 /* io-wq is going to take one down */
6193 refcount_inc(&req->refs);
6194 io_req_task_queue_fail(req, ret);
6198 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6201 struct fixed_rsrc_table *table;
6203 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6204 return table->files[index & IORING_FILE_TABLE_MASK];
6207 static struct file *io_file_get(struct io_submit_state *state,
6208 struct io_kiocb *req, int fd, bool fixed)
6210 struct io_ring_ctx *ctx = req->ctx;
6214 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6216 fd = array_index_nospec(fd, ctx->nr_user_files);
6217 file = io_file_from_index(ctx, fd);
6218 io_set_resource_node(req);
6220 trace_io_uring_file_get(ctx, fd);
6221 file = __io_file_get(state, fd);
6224 if (file && unlikely(file->f_op == &io_uring_fops))
6225 io_req_track_inflight(req);
6229 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6231 struct io_timeout_data *data = container_of(timer,
6232 struct io_timeout_data, timer);
6233 struct io_kiocb *prev, *req = data->req;
6234 struct io_ring_ctx *ctx = req->ctx;
6235 unsigned long flags;
6237 spin_lock_irqsave(&ctx->completion_lock, flags);
6238 prev = req->timeout.head;
6239 req->timeout.head = NULL;
6242 * We don't expect the list to be empty, that will only happen if we
6243 * race with the completion of the linked work.
6245 if (prev && refcount_inc_not_zero(&prev->refs))
6246 io_remove_next_linked(prev);
6249 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6252 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6253 io_put_req_deferred(prev, 1);
6255 io_req_complete_post(req, -ETIME, 0);
6256 io_put_req_deferred(req, 1);
6258 return HRTIMER_NORESTART;
6261 static void __io_queue_linked_timeout(struct io_kiocb *req)
6264 * If the back reference is NULL, then our linked request finished
6265 * before we got a chance to setup the timer
6267 if (req->timeout.head) {
6268 struct io_timeout_data *data = req->async_data;
6270 data->timer.function = io_link_timeout_fn;
6271 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6276 static void io_queue_linked_timeout(struct io_kiocb *req)
6278 struct io_ring_ctx *ctx = req->ctx;
6280 spin_lock_irq(&ctx->completion_lock);
6281 __io_queue_linked_timeout(req);
6282 spin_unlock_irq(&ctx->completion_lock);
6284 /* drop submission reference */
6288 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6290 struct io_kiocb *nxt = req->link;
6292 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6293 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6296 nxt->timeout.head = req;
6297 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6298 req->flags |= REQ_F_LINK_TIMEOUT;
6302 static void __io_queue_sqe(struct io_kiocb *req)
6304 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6307 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6310 * We async punt it if the file wasn't marked NOWAIT, or if the file
6311 * doesn't support non-blocking read/write attempts
6313 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6314 if (!io_arm_poll_handler(req)) {
6316 * Queued up for async execution, worker will release
6317 * submit reference when the iocb is actually submitted.
6319 io_queue_async_work(req);
6321 } else if (likely(!ret)) {
6322 /* drop submission reference */
6323 if (req->flags & REQ_F_COMPLETE_INLINE) {
6324 struct io_ring_ctx *ctx = req->ctx;
6325 struct io_comp_state *cs = &ctx->submit_state.comp;
6327 cs->reqs[cs->nr++] = req;
6328 if (cs->nr == ARRAY_SIZE(cs->reqs))
6329 io_submit_flush_completions(cs, ctx);
6334 req_set_fail_links(req);
6336 io_req_complete(req, ret);
6339 io_queue_linked_timeout(linked_timeout);
6342 static void io_queue_sqe(struct io_kiocb *req)
6346 ret = io_req_defer(req);
6348 if (ret != -EIOCBQUEUED) {
6350 req_set_fail_links(req);
6352 io_req_complete(req, ret);
6354 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6355 ret = io_req_defer_prep(req);
6358 io_queue_async_work(req);
6360 __io_queue_sqe(req);
6365 * Check SQE restrictions (opcode and flags).
6367 * Returns 'true' if SQE is allowed, 'false' otherwise.
6369 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6370 struct io_kiocb *req,
6371 unsigned int sqe_flags)
6373 if (!ctx->restricted)
6376 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6379 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6380 ctx->restrictions.sqe_flags_required)
6383 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6384 ctx->restrictions.sqe_flags_required))
6390 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6391 const struct io_uring_sqe *sqe)
6393 struct io_submit_state *state;
6394 unsigned int sqe_flags;
6395 int personality, ret = 0;
6397 req->opcode = READ_ONCE(sqe->opcode);
6398 /* same numerical values with corresponding REQ_F_*, safe to copy */
6399 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6400 req->user_data = READ_ONCE(sqe->user_data);
6401 req->async_data = NULL;
6405 req->fixed_rsrc_refs = NULL;
6406 /* one is dropped after submission, the other at completion */
6407 refcount_set(&req->refs, 2);
6408 req->task = current;
6410 req->work.list.next = NULL;
6411 req->work.creds = NULL;
6412 req->work.flags = 0;
6414 /* enforce forwards compatibility on users */
6415 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6420 if (unlikely(req->opcode >= IORING_OP_LAST))
6423 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6426 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6427 !io_op_defs[req->opcode].buffer_select)
6430 personality = READ_ONCE(sqe->personality);
6432 req->work.creds = xa_load(&ctx->personalities, personality);
6433 if (!req->work.creds)
6435 get_cred(req->work.creds);
6437 state = &ctx->submit_state;
6440 * Plug now if we have more than 1 IO left after this, and the target
6441 * is potentially a read/write to block based storage.
6443 if (!state->plug_started && state->ios_left > 1 &&
6444 io_op_defs[req->opcode].plug) {
6445 blk_start_plug(&state->plug);
6446 state->plug_started = true;
6449 if (io_op_defs[req->opcode].needs_file) {
6450 bool fixed = req->flags & REQ_F_FIXED_FILE;
6452 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6453 if (unlikely(!req->file))
6461 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6462 const struct io_uring_sqe *sqe)
6464 struct io_submit_link *link = &ctx->submit_state.link;
6467 ret = io_init_req(ctx, req, sqe);
6468 if (unlikely(ret)) {
6471 /* fail even hard links since we don't submit */
6472 link->head->flags |= REQ_F_FAIL_LINK;
6473 io_put_req(link->head);
6474 io_req_complete(link->head, -ECANCELED);
6478 io_req_complete(req, ret);
6481 ret = io_req_prep(req, sqe);
6485 /* don't need @sqe from now on */
6486 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6487 true, ctx->flags & IORING_SETUP_SQPOLL);
6490 * If we already have a head request, queue this one for async
6491 * submittal once the head completes. If we don't have a head but
6492 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6493 * submitted sync once the chain is complete. If none of those
6494 * conditions are true (normal request), then just queue it.
6497 struct io_kiocb *head = link->head;
6500 * Taking sequential execution of a link, draining both sides
6501 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6502 * requests in the link. So, it drains the head and the
6503 * next after the link request. The last one is done via
6504 * drain_next flag to persist the effect across calls.
6506 if (req->flags & REQ_F_IO_DRAIN) {
6507 head->flags |= REQ_F_IO_DRAIN;
6508 ctx->drain_next = 1;
6510 ret = io_req_defer_prep(req);
6513 trace_io_uring_link(ctx, req, head);
6514 link->last->link = req;
6517 /* last request of a link, enqueue the link */
6518 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6523 if (unlikely(ctx->drain_next)) {
6524 req->flags |= REQ_F_IO_DRAIN;
6525 ctx->drain_next = 0;
6527 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6539 * Batched submission is done, ensure local IO is flushed out.
6541 static void io_submit_state_end(struct io_submit_state *state,
6542 struct io_ring_ctx *ctx)
6544 if (state->link.head)
6545 io_queue_sqe(state->link.head);
6547 io_submit_flush_completions(&state->comp, ctx);
6548 if (state->plug_started)
6549 blk_finish_plug(&state->plug);
6550 io_state_file_put(state);
6554 * Start submission side cache.
6556 static void io_submit_state_start(struct io_submit_state *state,
6557 unsigned int max_ios)
6559 state->plug_started = false;
6560 state->ios_left = max_ios;
6561 /* set only head, no need to init link_last in advance */
6562 state->link.head = NULL;
6565 static void io_commit_sqring(struct io_ring_ctx *ctx)
6567 struct io_rings *rings = ctx->rings;
6570 * Ensure any loads from the SQEs are done at this point,
6571 * since once we write the new head, the application could
6572 * write new data to them.
6574 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6578 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6579 * that is mapped by userspace. This means that care needs to be taken to
6580 * ensure that reads are stable, as we cannot rely on userspace always
6581 * being a good citizen. If members of the sqe are validated and then later
6582 * used, it's important that those reads are done through READ_ONCE() to
6583 * prevent a re-load down the line.
6585 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6587 u32 *sq_array = ctx->sq_array;
6591 * The cached sq head (or cq tail) serves two purposes:
6593 * 1) allows us to batch the cost of updating the user visible
6595 * 2) allows the kernel side to track the head on its own, even
6596 * though the application is the one updating it.
6598 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6599 if (likely(head < ctx->sq_entries))
6600 return &ctx->sq_sqes[head];
6602 /* drop invalid entries */
6603 ctx->cached_sq_dropped++;
6604 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6608 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6612 /* if we have a backlog and couldn't flush it all, return BUSY */
6613 if (test_bit(0, &ctx->sq_check_overflow)) {
6614 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6618 /* make sure SQ entry isn't read before tail */
6619 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6621 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6624 percpu_counter_add(¤t->io_uring->inflight, nr);
6625 refcount_add(nr, ¤t->usage);
6626 io_submit_state_start(&ctx->submit_state, nr);
6628 while (submitted < nr) {
6629 const struct io_uring_sqe *sqe;
6630 struct io_kiocb *req;
6632 req = io_alloc_req(ctx);
6633 if (unlikely(!req)) {
6635 submitted = -EAGAIN;
6638 sqe = io_get_sqe(ctx);
6639 if (unlikely(!sqe)) {
6640 kmem_cache_free(req_cachep, req);
6643 /* will complete beyond this point, count as submitted */
6645 if (io_submit_sqe(ctx, req, sqe))
6649 if (unlikely(submitted != nr)) {
6650 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6651 struct io_uring_task *tctx = current->io_uring;
6652 int unused = nr - ref_used;
6654 percpu_ref_put_many(&ctx->refs, unused);
6655 percpu_counter_sub(&tctx->inflight, unused);
6656 put_task_struct_many(current, unused);
6659 io_submit_state_end(&ctx->submit_state, ctx);
6660 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6661 io_commit_sqring(ctx);
6666 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6668 /* Tell userspace we may need a wakeup call */
6669 spin_lock_irq(&ctx->completion_lock);
6670 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6671 spin_unlock_irq(&ctx->completion_lock);
6674 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6676 spin_lock_irq(&ctx->completion_lock);
6677 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6678 spin_unlock_irq(&ctx->completion_lock);
6681 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6683 unsigned int to_submit;
6686 to_submit = io_sqring_entries(ctx);
6687 /* if we're handling multiple rings, cap submit size for fairness */
6688 if (cap_entries && to_submit > 8)
6691 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6692 unsigned nr_events = 0;
6694 mutex_lock(&ctx->uring_lock);
6695 if (!list_empty(&ctx->iopoll_list))
6696 io_do_iopoll(ctx, &nr_events, 0);
6698 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6699 !(ctx->flags & IORING_SETUP_R_DISABLED))
6700 ret = io_submit_sqes(ctx, to_submit);
6701 mutex_unlock(&ctx->uring_lock);
6704 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6705 wake_up(&ctx->sqo_sq_wait);
6710 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6712 struct io_ring_ctx *ctx;
6713 unsigned sq_thread_idle = 0;
6715 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6716 if (sq_thread_idle < ctx->sq_thread_idle)
6717 sq_thread_idle = ctx->sq_thread_idle;
6720 sqd->sq_thread_idle = sq_thread_idle;
6723 static int io_sq_thread(void *data)
6725 struct io_sq_data *sqd = data;
6726 struct io_ring_ctx *ctx;
6727 unsigned long timeout = 0;
6728 char buf[TASK_COMM_LEN];
6731 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6732 set_task_comm(current, buf);
6733 current->pf_io_worker = NULL;
6735 if (sqd->sq_cpu != -1)
6736 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6738 set_cpus_allowed_ptr(current, cpu_online_mask);
6739 current->flags |= PF_NO_SETAFFINITY;
6741 mutex_lock(&sqd->lock);
6742 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6744 bool cap_entries, sqt_spin, needs_sched;
6746 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6747 signal_pending(current)) {
6748 bool did_sig = false;
6750 mutex_unlock(&sqd->lock);
6751 if (signal_pending(current)) {
6752 struct ksignal ksig;
6754 did_sig = get_signal(&ksig);
6757 mutex_lock(&sqd->lock);
6761 io_run_task_work_head(&sqd->park_task_work);
6762 timeout = jiffies + sqd->sq_thread_idle;
6766 cap_entries = !list_is_singular(&sqd->ctx_list);
6767 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6768 const struct cred *creds = NULL;
6770 if (ctx->sq_creds != current_cred())
6771 creds = override_creds(ctx->sq_creds);
6772 ret = __io_sq_thread(ctx, cap_entries);
6774 revert_creds(creds);
6775 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6779 if (sqt_spin || !time_after(jiffies, timeout)) {
6783 timeout = jiffies + sqd->sq_thread_idle;
6788 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6789 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6790 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6791 !list_empty_careful(&ctx->iopoll_list)) {
6792 needs_sched = false;
6795 if (io_sqring_entries(ctx)) {
6796 needs_sched = false;
6801 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6802 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6803 io_ring_set_wakeup_flag(ctx);
6805 mutex_unlock(&sqd->lock);
6807 mutex_lock(&sqd->lock);
6808 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6809 io_ring_clear_wakeup_flag(ctx);
6812 finish_wait(&sqd->wait, &wait);
6813 io_run_task_work_head(&sqd->park_task_work);
6814 timeout = jiffies + sqd->sq_thread_idle;
6817 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6818 io_uring_cancel_sqpoll(ctx);
6820 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6821 io_ring_set_wakeup_flag(ctx);
6822 mutex_unlock(&sqd->lock);
6825 io_run_task_work_head(&sqd->park_task_work);
6826 complete(&sqd->exited);
6830 struct io_wait_queue {
6831 struct wait_queue_entry wq;
6832 struct io_ring_ctx *ctx;
6834 unsigned nr_timeouts;
6837 static inline bool io_should_wake(struct io_wait_queue *iowq)
6839 struct io_ring_ctx *ctx = iowq->ctx;
6842 * Wake up if we have enough events, or if a timeout occurred since we
6843 * started waiting. For timeouts, we always want to return to userspace,
6844 * regardless of event count.
6846 return io_cqring_events(ctx) >= iowq->to_wait ||
6847 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6850 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6851 int wake_flags, void *key)
6853 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6857 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6858 * the task, and the next invocation will do it.
6860 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6861 return autoremove_wake_function(curr, mode, wake_flags, key);
6865 static int io_run_task_work_sig(void)
6867 if (io_run_task_work())
6869 if (!signal_pending(current))
6871 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6872 return -ERESTARTSYS;
6876 /* when returns >0, the caller should retry */
6877 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6878 struct io_wait_queue *iowq,
6879 signed long *timeout)
6883 /* make sure we run task_work before checking for signals */
6884 ret = io_run_task_work_sig();
6885 if (ret || io_should_wake(iowq))
6887 /* let the caller flush overflows, retry */
6888 if (test_bit(0, &ctx->cq_check_overflow))
6891 *timeout = schedule_timeout(*timeout);
6892 return !*timeout ? -ETIME : 1;
6896 * Wait until events become available, if we don't already have some. The
6897 * application must reap them itself, as they reside on the shared cq ring.
6899 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6900 const sigset_t __user *sig, size_t sigsz,
6901 struct __kernel_timespec __user *uts)
6903 struct io_wait_queue iowq = {
6906 .func = io_wake_function,
6907 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6910 .to_wait = min_events,
6912 struct io_rings *rings = ctx->rings;
6913 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6917 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6918 if (io_cqring_events(ctx) >= min_events)
6920 if (!io_run_task_work())
6925 #ifdef CONFIG_COMPAT
6926 if (in_compat_syscall())
6927 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6931 ret = set_user_sigmask(sig, sigsz);
6938 struct timespec64 ts;
6940 if (get_timespec64(&ts, uts))
6942 timeout = timespec64_to_jiffies(&ts);
6945 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6946 trace_io_uring_cqring_wait(ctx, min_events);
6948 /* if we can't even flush overflow, don't wait for more */
6949 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6953 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6954 TASK_INTERRUPTIBLE);
6955 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6956 finish_wait(&ctx->wait, &iowq.wq);
6960 restore_saved_sigmask_unless(ret == -EINTR);
6962 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6965 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6967 #if defined(CONFIG_UNIX)
6968 if (ctx->ring_sock) {
6969 struct sock *sock = ctx->ring_sock->sk;
6970 struct sk_buff *skb;
6972 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6978 for (i = 0; i < ctx->nr_user_files; i++) {
6981 file = io_file_from_index(ctx, i);
6988 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6990 struct fixed_rsrc_data *data;
6992 data = container_of(ref, struct fixed_rsrc_data, refs);
6993 complete(&data->done);
6996 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6998 spin_lock_bh(&ctx->rsrc_ref_lock);
7001 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7003 spin_unlock_bh(&ctx->rsrc_ref_lock);
7006 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7007 struct fixed_rsrc_data *rsrc_data,
7008 struct fixed_rsrc_ref_node *ref_node)
7010 io_rsrc_ref_lock(ctx);
7011 rsrc_data->node = ref_node;
7012 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7013 io_rsrc_ref_unlock(ctx);
7014 percpu_ref_get(&rsrc_data->refs);
7017 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7019 struct fixed_rsrc_ref_node *ref_node = NULL;
7021 io_rsrc_ref_lock(ctx);
7022 ref_node = data->node;
7024 io_rsrc_ref_unlock(ctx);
7026 percpu_ref_kill(&ref_node->refs);
7029 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7030 struct io_ring_ctx *ctx,
7031 void (*rsrc_put)(struct io_ring_ctx *ctx,
7032 struct io_rsrc_put *prsrc))
7034 struct fixed_rsrc_ref_node *backup_node;
7040 data->quiesce = true;
7043 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7046 backup_node->rsrc_data = data;
7047 backup_node->rsrc_put = rsrc_put;
7049 io_sqe_rsrc_kill_node(ctx, data);
7050 percpu_ref_kill(&data->refs);
7051 flush_delayed_work(&ctx->rsrc_put_work);
7053 ret = wait_for_completion_interruptible(&data->done);
7057 percpu_ref_resurrect(&data->refs);
7058 io_sqe_rsrc_set_node(ctx, data, backup_node);
7060 reinit_completion(&data->done);
7061 mutex_unlock(&ctx->uring_lock);
7062 ret = io_run_task_work_sig();
7063 mutex_lock(&ctx->uring_lock);
7065 data->quiesce = false;
7068 destroy_fixed_rsrc_ref_node(backup_node);
7072 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7074 struct fixed_rsrc_data *data;
7076 data = kzalloc(sizeof(*data), GFP_KERNEL);
7080 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7081 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7086 init_completion(&data->done);
7090 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7092 percpu_ref_exit(&data->refs);
7097 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7099 struct fixed_rsrc_data *data = ctx->file_data;
7100 unsigned nr_tables, i;
7104 * percpu_ref_is_dying() is to stop parallel files unregister
7105 * Since we possibly drop uring lock later in this function to
7108 if (!data || percpu_ref_is_dying(&data->refs))
7110 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7114 __io_sqe_files_unregister(ctx);
7115 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7116 for (i = 0; i < nr_tables; i++)
7117 kfree(data->table[i].files);
7118 free_fixed_rsrc_data(data);
7119 ctx->file_data = NULL;
7120 ctx->nr_user_files = 0;
7124 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7125 __releases(&sqd->lock)
7127 WARN_ON_ONCE(sqd->thread == current);
7130 * Do the dance but not conditional clear_bit() because it'd race with
7131 * other threads incrementing park_pending and setting the bit.
7133 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7134 if (atomic_dec_return(&sqd->park_pending))
7135 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7136 mutex_unlock(&sqd->lock);
7139 static void io_sq_thread_park(struct io_sq_data *sqd)
7140 __acquires(&sqd->lock)
7142 WARN_ON_ONCE(sqd->thread == current);
7144 atomic_inc(&sqd->park_pending);
7145 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7146 mutex_lock(&sqd->lock);
7148 wake_up_process(sqd->thread);
7151 static void io_sq_thread_stop(struct io_sq_data *sqd)
7153 WARN_ON_ONCE(sqd->thread == current);
7155 mutex_lock(&sqd->lock);
7156 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7158 wake_up_process(sqd->thread);
7159 mutex_unlock(&sqd->lock);
7160 wait_for_completion(&sqd->exited);
7163 static void io_put_sq_data(struct io_sq_data *sqd)
7165 if (refcount_dec_and_test(&sqd->refs)) {
7166 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7168 io_sq_thread_stop(sqd);
7173 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7175 struct io_sq_data *sqd = ctx->sq_data;
7178 io_sq_thread_park(sqd);
7179 list_del_init(&ctx->sqd_list);
7180 io_sqd_update_thread_idle(sqd);
7181 io_sq_thread_unpark(sqd);
7183 io_put_sq_data(sqd);
7184 ctx->sq_data = NULL;
7186 put_cred(ctx->sq_creds);
7190 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7192 struct io_ring_ctx *ctx_attach;
7193 struct io_sq_data *sqd;
7196 f = fdget(p->wq_fd);
7198 return ERR_PTR(-ENXIO);
7199 if (f.file->f_op != &io_uring_fops) {
7201 return ERR_PTR(-EINVAL);
7204 ctx_attach = f.file->private_data;
7205 sqd = ctx_attach->sq_data;
7208 return ERR_PTR(-EINVAL);
7210 if (sqd->task_tgid != current->tgid) {
7212 return ERR_PTR(-EPERM);
7215 refcount_inc(&sqd->refs);
7220 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7223 struct io_sq_data *sqd;
7226 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7227 sqd = io_attach_sq_data(p);
7232 /* fall through for EPERM case, setup new sqd/task */
7233 if (PTR_ERR(sqd) != -EPERM)
7237 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7239 return ERR_PTR(-ENOMEM);
7241 atomic_set(&sqd->park_pending, 0);
7242 refcount_set(&sqd->refs, 1);
7243 INIT_LIST_HEAD(&sqd->ctx_list);
7244 mutex_init(&sqd->lock);
7245 init_waitqueue_head(&sqd->wait);
7246 init_completion(&sqd->exited);
7250 #if defined(CONFIG_UNIX)
7252 * Ensure the UNIX gc is aware of our file set, so we are certain that
7253 * the io_uring can be safely unregistered on process exit, even if we have
7254 * loops in the file referencing.
7256 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7258 struct sock *sk = ctx->ring_sock->sk;
7259 struct scm_fp_list *fpl;
7260 struct sk_buff *skb;
7263 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7267 skb = alloc_skb(0, GFP_KERNEL);
7276 fpl->user = get_uid(current_user());
7277 for (i = 0; i < nr; i++) {
7278 struct file *file = io_file_from_index(ctx, i + offset);
7282 fpl->fp[nr_files] = get_file(file);
7283 unix_inflight(fpl->user, fpl->fp[nr_files]);
7288 fpl->max = SCM_MAX_FD;
7289 fpl->count = nr_files;
7290 UNIXCB(skb).fp = fpl;
7291 skb->destructor = unix_destruct_scm;
7292 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7293 skb_queue_head(&sk->sk_receive_queue, skb);
7295 for (i = 0; i < nr_files; i++)
7306 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7307 * causes regular reference counting to break down. We rely on the UNIX
7308 * garbage collection to take care of this problem for us.
7310 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7312 unsigned left, total;
7316 left = ctx->nr_user_files;
7318 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7320 ret = __io_sqe_files_scm(ctx, this_files, total);
7324 total += this_files;
7330 while (total < ctx->nr_user_files) {
7331 struct file *file = io_file_from_index(ctx, total);
7341 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7347 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7348 unsigned nr_tables, unsigned nr_files)
7352 for (i = 0; i < nr_tables; i++) {
7353 struct fixed_rsrc_table *table = &file_data->table[i];
7354 unsigned this_files;
7356 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7357 table->files = kcalloc(this_files, sizeof(struct file *),
7361 nr_files -= this_files;
7367 for (i = 0; i < nr_tables; i++) {
7368 struct fixed_rsrc_table *table = &file_data->table[i];
7369 kfree(table->files);
7374 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7376 struct file *file = prsrc->file;
7377 #if defined(CONFIG_UNIX)
7378 struct sock *sock = ctx->ring_sock->sk;
7379 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7380 struct sk_buff *skb;
7383 __skb_queue_head_init(&list);
7386 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7387 * remove this entry and rearrange the file array.
7389 skb = skb_dequeue(head);
7391 struct scm_fp_list *fp;
7393 fp = UNIXCB(skb).fp;
7394 for (i = 0; i < fp->count; i++) {
7397 if (fp->fp[i] != file)
7400 unix_notinflight(fp->user, fp->fp[i]);
7401 left = fp->count - 1 - i;
7403 memmove(&fp->fp[i], &fp->fp[i + 1],
7404 left * sizeof(struct file *));
7411 __skb_queue_tail(&list, skb);
7421 __skb_queue_tail(&list, skb);
7423 skb = skb_dequeue(head);
7426 if (skb_peek(&list)) {
7427 spin_lock_irq(&head->lock);
7428 while ((skb = __skb_dequeue(&list)) != NULL)
7429 __skb_queue_tail(head, skb);
7430 spin_unlock_irq(&head->lock);
7437 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7439 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7440 struct io_ring_ctx *ctx = rsrc_data->ctx;
7441 struct io_rsrc_put *prsrc, *tmp;
7443 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7444 list_del(&prsrc->list);
7445 ref_node->rsrc_put(ctx, prsrc);
7449 percpu_ref_exit(&ref_node->refs);
7451 percpu_ref_put(&rsrc_data->refs);
7454 static void io_rsrc_put_work(struct work_struct *work)
7456 struct io_ring_ctx *ctx;
7457 struct llist_node *node;
7459 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7460 node = llist_del_all(&ctx->rsrc_put_llist);
7463 struct fixed_rsrc_ref_node *ref_node;
7464 struct llist_node *next = node->next;
7466 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7467 __io_rsrc_put_work(ref_node);
7472 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7475 struct fixed_rsrc_table *table;
7477 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7478 return &table->files[i & IORING_FILE_TABLE_MASK];
7481 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7483 struct fixed_rsrc_ref_node *ref_node;
7484 struct fixed_rsrc_data *data;
7485 struct io_ring_ctx *ctx;
7486 bool first_add = false;
7489 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7490 data = ref_node->rsrc_data;
7493 io_rsrc_ref_lock(ctx);
7494 ref_node->done = true;
7496 while (!list_empty(&ctx->rsrc_ref_list)) {
7497 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7498 struct fixed_rsrc_ref_node, node);
7499 /* recycle ref nodes in order */
7500 if (!ref_node->done)
7502 list_del(&ref_node->node);
7503 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7505 io_rsrc_ref_unlock(ctx);
7507 if (percpu_ref_is_dying(&data->refs))
7511 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7513 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7516 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7517 struct io_ring_ctx *ctx)
7519 struct fixed_rsrc_ref_node *ref_node;
7521 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7525 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7530 INIT_LIST_HEAD(&ref_node->node);
7531 INIT_LIST_HEAD(&ref_node->rsrc_list);
7532 ref_node->done = false;
7536 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7537 struct fixed_rsrc_ref_node *ref_node)
7539 ref_node->rsrc_data = ctx->file_data;
7540 ref_node->rsrc_put = io_ring_file_put;
7543 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7545 percpu_ref_exit(&ref_node->refs);
7550 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7553 __s32 __user *fds = (__s32 __user *) arg;
7554 unsigned nr_tables, i;
7556 int fd, ret = -ENOMEM;
7557 struct fixed_rsrc_ref_node *ref_node;
7558 struct fixed_rsrc_data *file_data;
7564 if (nr_args > IORING_MAX_FIXED_FILES)
7567 file_data = alloc_fixed_rsrc_data(ctx);
7570 ctx->file_data = file_data;
7572 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7573 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7575 if (!file_data->table)
7578 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7581 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7582 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7586 /* allow sparse sets */
7596 * Don't allow io_uring instances to be registered. If UNIX
7597 * isn't enabled, then this causes a reference cycle and this
7598 * instance can never get freed. If UNIX is enabled we'll
7599 * handle it just fine, but there's still no point in allowing
7600 * a ring fd as it doesn't support regular read/write anyway.
7602 if (file->f_op == &io_uring_fops) {
7606 *io_fixed_file_slot(file_data, i) = file;
7609 ret = io_sqe_files_scm(ctx);
7611 io_sqe_files_unregister(ctx);
7615 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7617 io_sqe_files_unregister(ctx);
7620 init_fixed_file_ref_node(ctx, ref_node);
7622 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7625 for (i = 0; i < ctx->nr_user_files; i++) {
7626 file = io_file_from_index(ctx, i);
7630 for (i = 0; i < nr_tables; i++)
7631 kfree(file_data->table[i].files);
7632 ctx->nr_user_files = 0;
7634 free_fixed_rsrc_data(ctx->file_data);
7635 ctx->file_data = NULL;
7639 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7642 #if defined(CONFIG_UNIX)
7643 struct sock *sock = ctx->ring_sock->sk;
7644 struct sk_buff_head *head = &sock->sk_receive_queue;
7645 struct sk_buff *skb;
7648 * See if we can merge this file into an existing skb SCM_RIGHTS
7649 * file set. If there's no room, fall back to allocating a new skb
7650 * and filling it in.
7652 spin_lock_irq(&head->lock);
7653 skb = skb_peek(head);
7655 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7657 if (fpl->count < SCM_MAX_FD) {
7658 __skb_unlink(skb, head);
7659 spin_unlock_irq(&head->lock);
7660 fpl->fp[fpl->count] = get_file(file);
7661 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7663 spin_lock_irq(&head->lock);
7664 __skb_queue_head(head, skb);
7669 spin_unlock_irq(&head->lock);
7676 return __io_sqe_files_scm(ctx, 1, index);
7682 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7684 struct io_rsrc_put *prsrc;
7685 struct fixed_rsrc_ref_node *ref_node = data->node;
7687 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7692 list_add(&prsrc->list, &ref_node->rsrc_list);
7697 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7700 return io_queue_rsrc_removal(data, (void *)file);
7703 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7704 struct io_uring_rsrc_update *up,
7707 struct fixed_rsrc_data *data = ctx->file_data;
7708 struct fixed_rsrc_ref_node *ref_node;
7709 struct file *file, **file_slot;
7713 bool needs_switch = false;
7715 if (check_add_overflow(up->offset, nr_args, &done))
7717 if (done > ctx->nr_user_files)
7720 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7723 init_fixed_file_ref_node(ctx, ref_node);
7725 fds = u64_to_user_ptr(up->data);
7726 for (done = 0; done < nr_args; done++) {
7728 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7732 if (fd == IORING_REGISTER_FILES_SKIP)
7735 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7736 file_slot = io_fixed_file_slot(ctx->file_data, i);
7739 err = io_queue_file_removal(data, *file_slot);
7743 needs_switch = true;
7752 * Don't allow io_uring instances to be registered. If
7753 * UNIX isn't enabled, then this causes a reference
7754 * cycle and this instance can never get freed. If UNIX
7755 * is enabled we'll handle it just fine, but there's
7756 * still no point in allowing a ring fd as it doesn't
7757 * support regular read/write anyway.
7759 if (file->f_op == &io_uring_fops) {
7765 err = io_sqe_file_register(ctx, file, i);
7775 percpu_ref_kill(&data->node->refs);
7776 io_sqe_rsrc_set_node(ctx, data, ref_node);
7778 destroy_fixed_rsrc_ref_node(ref_node);
7780 return done ? done : err;
7783 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7786 struct io_uring_rsrc_update up;
7788 if (!ctx->file_data)
7792 if (copy_from_user(&up, arg, sizeof(up)))
7797 return __io_sqe_files_update(ctx, &up, nr_args);
7800 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7802 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7804 req = io_put_req_find_next(req);
7805 return req ? &req->work : NULL;
7808 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7810 struct io_wq_hash *hash;
7811 struct io_wq_data data;
7812 unsigned int concurrency;
7814 hash = ctx->hash_map;
7816 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7818 return ERR_PTR(-ENOMEM);
7819 refcount_set(&hash->refs, 1);
7820 init_waitqueue_head(&hash->wait);
7821 ctx->hash_map = hash;
7825 data.free_work = io_free_work;
7826 data.do_work = io_wq_submit_work;
7828 /* Do QD, or 4 * CPUS, whatever is smallest */
7829 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7831 return io_wq_create(concurrency, &data);
7834 static int io_uring_alloc_task_context(struct task_struct *task,
7835 struct io_ring_ctx *ctx)
7837 struct io_uring_task *tctx;
7840 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7841 if (unlikely(!tctx))
7844 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7845 if (unlikely(ret)) {
7850 tctx->io_wq = io_init_wq_offload(ctx);
7851 if (IS_ERR(tctx->io_wq)) {
7852 ret = PTR_ERR(tctx->io_wq);
7853 percpu_counter_destroy(&tctx->inflight);
7859 init_waitqueue_head(&tctx->wait);
7861 atomic_set(&tctx->in_idle, 0);
7862 task->io_uring = tctx;
7863 spin_lock_init(&tctx->task_lock);
7864 INIT_WQ_LIST(&tctx->task_list);
7865 tctx->task_state = 0;
7866 init_task_work(&tctx->task_work, tctx_task_work);
7870 void __io_uring_free(struct task_struct *tsk)
7872 struct io_uring_task *tctx = tsk->io_uring;
7874 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7875 WARN_ON_ONCE(tctx->io_wq);
7877 percpu_counter_destroy(&tctx->inflight);
7879 tsk->io_uring = NULL;
7882 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7883 struct io_uring_params *p)
7887 /* Retain compatibility with failing for an invalid attach attempt */
7888 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7889 IORING_SETUP_ATTACH_WQ) {
7892 f = fdget(p->wq_fd);
7895 if (f.file->f_op != &io_uring_fops) {
7901 if (ctx->flags & IORING_SETUP_SQPOLL) {
7902 struct task_struct *tsk;
7903 struct io_sq_data *sqd;
7907 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7910 sqd = io_get_sq_data(p, &attached);
7916 ctx->sq_creds = get_current_cred();
7918 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7919 if (!ctx->sq_thread_idle)
7920 ctx->sq_thread_idle = HZ;
7923 io_sq_thread_park(sqd);
7924 list_add(&ctx->sqd_list, &sqd->ctx_list);
7925 io_sqd_update_thread_idle(sqd);
7926 /* don't attach to a dying SQPOLL thread, would be racy */
7927 if (attached && !sqd->thread)
7929 io_sq_thread_unpark(sqd);
7936 if (p->flags & IORING_SETUP_SQ_AFF) {
7937 int cpu = p->sq_thread_cpu;
7940 if (cpu >= nr_cpu_ids)
7942 if (!cpu_online(cpu))
7950 sqd->task_pid = current->pid;
7951 sqd->task_tgid = current->tgid;
7952 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7959 ret = io_uring_alloc_task_context(tsk, ctx);
7960 wake_up_new_task(tsk);
7963 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7964 /* Can't have SQ_AFF without SQPOLL */
7971 io_sq_thread_finish(ctx);
7974 complete(&ctx->sq_data->exited);
7978 static inline void __io_unaccount_mem(struct user_struct *user,
7979 unsigned long nr_pages)
7981 atomic_long_sub(nr_pages, &user->locked_vm);
7984 static inline int __io_account_mem(struct user_struct *user,
7985 unsigned long nr_pages)
7987 unsigned long page_limit, cur_pages, new_pages;
7989 /* Don't allow more pages than we can safely lock */
7990 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7993 cur_pages = atomic_long_read(&user->locked_vm);
7994 new_pages = cur_pages + nr_pages;
7995 if (new_pages > page_limit)
7997 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7998 new_pages) != cur_pages);
8003 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8006 __io_unaccount_mem(ctx->user, nr_pages);
8008 if (ctx->mm_account)
8009 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8012 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8017 ret = __io_account_mem(ctx->user, nr_pages);
8022 if (ctx->mm_account)
8023 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8028 static void io_mem_free(void *ptr)
8035 page = virt_to_head_page(ptr);
8036 if (put_page_testzero(page))
8037 free_compound_page(page);
8040 static void *io_mem_alloc(size_t size)
8042 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8043 __GFP_NORETRY | __GFP_ACCOUNT;
8045 return (void *) __get_free_pages(gfp_flags, get_order(size));
8048 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8051 struct io_rings *rings;
8052 size_t off, sq_array_size;
8054 off = struct_size(rings, cqes, cq_entries);
8055 if (off == SIZE_MAX)
8059 off = ALIGN(off, SMP_CACHE_BYTES);
8067 sq_array_size = array_size(sizeof(u32), sq_entries);
8068 if (sq_array_size == SIZE_MAX)
8071 if (check_add_overflow(off, sq_array_size, &off))
8077 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8081 if (!ctx->user_bufs)
8084 for (i = 0; i < ctx->nr_user_bufs; i++) {
8085 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8087 for (j = 0; j < imu->nr_bvecs; j++)
8088 unpin_user_page(imu->bvec[j].bv_page);
8090 if (imu->acct_pages)
8091 io_unaccount_mem(ctx, imu->acct_pages);
8096 kfree(ctx->user_bufs);
8097 ctx->user_bufs = NULL;
8098 ctx->nr_user_bufs = 0;
8102 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8103 void __user *arg, unsigned index)
8105 struct iovec __user *src;
8107 #ifdef CONFIG_COMPAT
8109 struct compat_iovec __user *ciovs;
8110 struct compat_iovec ciov;
8112 ciovs = (struct compat_iovec __user *) arg;
8113 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8116 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8117 dst->iov_len = ciov.iov_len;
8121 src = (struct iovec __user *) arg;
8122 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8128 * Not super efficient, but this is just a registration time. And we do cache
8129 * the last compound head, so generally we'll only do a full search if we don't
8132 * We check if the given compound head page has already been accounted, to
8133 * avoid double accounting it. This allows us to account the full size of the
8134 * page, not just the constituent pages of a huge page.
8136 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8137 int nr_pages, struct page *hpage)
8141 /* check current page array */
8142 for (i = 0; i < nr_pages; i++) {
8143 if (!PageCompound(pages[i]))
8145 if (compound_head(pages[i]) == hpage)
8149 /* check previously registered pages */
8150 for (i = 0; i < ctx->nr_user_bufs; i++) {
8151 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8153 for (j = 0; j < imu->nr_bvecs; j++) {
8154 if (!PageCompound(imu->bvec[j].bv_page))
8156 if (compound_head(imu->bvec[j].bv_page) == hpage)
8164 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8165 int nr_pages, struct io_mapped_ubuf *imu,
8166 struct page **last_hpage)
8170 for (i = 0; i < nr_pages; i++) {
8171 if (!PageCompound(pages[i])) {
8176 hpage = compound_head(pages[i]);
8177 if (hpage == *last_hpage)
8179 *last_hpage = hpage;
8180 if (headpage_already_acct(ctx, pages, i, hpage))
8182 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8186 if (!imu->acct_pages)
8189 ret = io_account_mem(ctx, imu->acct_pages);
8191 imu->acct_pages = 0;
8195 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8196 struct io_mapped_ubuf *imu,
8197 struct page **last_hpage)
8199 struct vm_area_struct **vmas = NULL;
8200 struct page **pages = NULL;
8201 unsigned long off, start, end, ubuf;
8203 int ret, pret, nr_pages, i;
8205 ubuf = (unsigned long) iov->iov_base;
8206 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8207 start = ubuf >> PAGE_SHIFT;
8208 nr_pages = end - start;
8212 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8216 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8221 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8227 mmap_read_lock(current->mm);
8228 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8230 if (pret == nr_pages) {
8231 /* don't support file backed memory */
8232 for (i = 0; i < nr_pages; i++) {
8233 struct vm_area_struct *vma = vmas[i];
8236 !is_file_hugepages(vma->vm_file)) {
8242 ret = pret < 0 ? pret : -EFAULT;
8244 mmap_read_unlock(current->mm);
8247 * if we did partial map, or found file backed vmas,
8248 * release any pages we did get
8251 unpin_user_pages(pages, pret);
8256 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8258 unpin_user_pages(pages, pret);
8263 off = ubuf & ~PAGE_MASK;
8264 size = iov->iov_len;
8265 for (i = 0; i < nr_pages; i++) {
8268 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8269 imu->bvec[i].bv_page = pages[i];
8270 imu->bvec[i].bv_len = vec_len;
8271 imu->bvec[i].bv_offset = off;
8275 /* store original address for later verification */
8277 imu->len = iov->iov_len;
8278 imu->nr_bvecs = nr_pages;
8286 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8290 if (!nr_args || nr_args > UIO_MAXIOV)
8293 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8295 if (!ctx->user_bufs)
8301 static int io_buffer_validate(struct iovec *iov)
8304 * Don't impose further limits on the size and buffer
8305 * constraints here, we'll -EINVAL later when IO is
8306 * submitted if they are wrong.
8308 if (!iov->iov_base || !iov->iov_len)
8311 /* arbitrary limit, but we need something */
8312 if (iov->iov_len > SZ_1G)
8318 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8319 unsigned int nr_args)
8323 struct page *last_hpage = NULL;
8325 ret = io_buffers_map_alloc(ctx, nr_args);
8329 for (i = 0; i < nr_args; i++) {
8330 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8332 ret = io_copy_iov(ctx, &iov, arg, i);
8336 ret = io_buffer_validate(&iov);
8340 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8344 ctx->nr_user_bufs++;
8348 io_sqe_buffers_unregister(ctx);
8353 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8355 __s32 __user *fds = arg;
8361 if (copy_from_user(&fd, fds, sizeof(*fds)))
8364 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8365 if (IS_ERR(ctx->cq_ev_fd)) {
8366 int ret = PTR_ERR(ctx->cq_ev_fd);
8367 ctx->cq_ev_fd = NULL;
8374 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8376 if (ctx->cq_ev_fd) {
8377 eventfd_ctx_put(ctx->cq_ev_fd);
8378 ctx->cq_ev_fd = NULL;
8385 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8387 struct io_buffer *buf;
8388 unsigned long index;
8390 xa_for_each(&ctx->io_buffers, index, buf)
8391 __io_remove_buffers(ctx, buf, index, -1U);
8394 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8396 struct io_kiocb *req, *nxt;
8398 list_for_each_entry_safe(req, nxt, list, compl.list) {
8399 if (tsk && req->task != tsk)
8401 list_del(&req->compl.list);
8402 kmem_cache_free(req_cachep, req);
8406 static void io_req_caches_free(struct io_ring_ctx *ctx)
8408 struct io_submit_state *submit_state = &ctx->submit_state;
8409 struct io_comp_state *cs = &ctx->submit_state.comp;
8411 mutex_lock(&ctx->uring_lock);
8413 if (submit_state->free_reqs) {
8414 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8415 submit_state->reqs);
8416 submit_state->free_reqs = 0;
8419 spin_lock_irq(&ctx->completion_lock);
8420 list_splice_init(&cs->locked_free_list, &cs->free_list);
8421 cs->locked_free_nr = 0;
8422 spin_unlock_irq(&ctx->completion_lock);
8424 io_req_cache_free(&cs->free_list, NULL);
8426 mutex_unlock(&ctx->uring_lock);
8429 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8432 * Some may use context even when all refs and requests have been put,
8433 * and they are free to do so while still holding uring_lock or
8434 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8436 mutex_lock(&ctx->uring_lock);
8437 mutex_unlock(&ctx->uring_lock);
8438 spin_lock_irq(&ctx->completion_lock);
8439 spin_unlock_irq(&ctx->completion_lock);
8441 io_sq_thread_finish(ctx);
8442 io_sqe_buffers_unregister(ctx);
8444 if (ctx->mm_account) {
8445 mmdrop(ctx->mm_account);
8446 ctx->mm_account = NULL;
8449 mutex_lock(&ctx->uring_lock);
8450 io_sqe_files_unregister(ctx);
8451 mutex_unlock(&ctx->uring_lock);
8452 io_eventfd_unregister(ctx);
8453 io_destroy_buffers(ctx);
8455 #if defined(CONFIG_UNIX)
8456 if (ctx->ring_sock) {
8457 ctx->ring_sock->file = NULL; /* so that iput() is called */
8458 sock_release(ctx->ring_sock);
8462 io_mem_free(ctx->rings);
8463 io_mem_free(ctx->sq_sqes);
8465 percpu_ref_exit(&ctx->refs);
8466 free_uid(ctx->user);
8467 io_req_caches_free(ctx);
8469 io_wq_put_hash(ctx->hash_map);
8470 kfree(ctx->cancel_hash);
8474 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8476 struct io_ring_ctx *ctx = file->private_data;
8479 poll_wait(file, &ctx->cq_wait, wait);
8481 * synchronizes with barrier from wq_has_sleeper call in
8485 if (!io_sqring_full(ctx))
8486 mask |= EPOLLOUT | EPOLLWRNORM;
8489 * Don't flush cqring overflow list here, just do a simple check.
8490 * Otherwise there could possible be ABBA deadlock:
8493 * lock(&ctx->uring_lock);
8495 * lock(&ctx->uring_lock);
8498 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8499 * pushs them to do the flush.
8501 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8502 mask |= EPOLLIN | EPOLLRDNORM;
8507 static int io_uring_fasync(int fd, struct file *file, int on)
8509 struct io_ring_ctx *ctx = file->private_data;
8511 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8514 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8516 const struct cred *creds;
8518 creds = xa_erase(&ctx->personalities, id);
8527 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8529 return io_run_task_work_head(&ctx->exit_task_work);
8532 struct io_tctx_exit {
8533 struct callback_head task_work;
8534 struct completion completion;
8535 struct io_ring_ctx *ctx;
8538 static void io_tctx_exit_cb(struct callback_head *cb)
8540 struct io_uring_task *tctx = current->io_uring;
8541 struct io_tctx_exit *work;
8543 work = container_of(cb, struct io_tctx_exit, task_work);
8545 * When @in_idle, we're in cancellation and it's racy to remove the
8546 * node. It'll be removed by the end of cancellation, just ignore it.
8548 if (!atomic_read(&tctx->in_idle))
8549 io_uring_del_task_file((unsigned long)work->ctx);
8550 complete(&work->completion);
8553 static void io_ring_exit_work(struct work_struct *work)
8555 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8556 unsigned long timeout = jiffies + HZ * 60 * 5;
8557 struct io_tctx_exit exit;
8558 struct io_tctx_node *node;
8561 /* prevent SQPOLL from submitting new requests */
8563 io_sq_thread_park(ctx->sq_data);
8564 list_del_init(&ctx->sqd_list);
8565 io_sqd_update_thread_idle(ctx->sq_data);
8566 io_sq_thread_unpark(ctx->sq_data);
8570 * If we're doing polled IO and end up having requests being
8571 * submitted async (out-of-line), then completions can come in while
8572 * we're waiting for refs to drop. We need to reap these manually,
8573 * as nobody else will be looking for them.
8576 io_uring_try_cancel_requests(ctx, NULL, NULL);
8578 WARN_ON_ONCE(time_after(jiffies, timeout));
8579 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8581 mutex_lock(&ctx->uring_lock);
8582 while (!list_empty(&ctx->tctx_list)) {
8583 WARN_ON_ONCE(time_after(jiffies, timeout));
8585 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8588 init_completion(&exit.completion);
8589 init_task_work(&exit.task_work, io_tctx_exit_cb);
8590 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8591 if (WARN_ON_ONCE(ret))
8593 wake_up_process(node->task);
8595 mutex_unlock(&ctx->uring_lock);
8596 wait_for_completion(&exit.completion);
8598 mutex_lock(&ctx->uring_lock);
8600 mutex_unlock(&ctx->uring_lock);
8602 io_ring_ctx_free(ctx);
8605 /* Returns true if we found and killed one or more timeouts */
8606 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8607 struct files_struct *files)
8609 struct io_kiocb *req, *tmp;
8612 spin_lock_irq(&ctx->completion_lock);
8613 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8614 if (io_match_task(req, tsk, files)) {
8615 io_kill_timeout(req, -ECANCELED);
8620 io_commit_cqring(ctx);
8621 spin_unlock_irq(&ctx->completion_lock);
8623 io_cqring_ev_posted(ctx);
8624 return canceled != 0;
8627 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8629 unsigned long index;
8630 struct creds *creds;
8632 mutex_lock(&ctx->uring_lock);
8633 percpu_ref_kill(&ctx->refs);
8634 /* if force is set, the ring is going away. always drop after that */
8635 ctx->cq_overflow_flushed = 1;
8637 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8638 xa_for_each(&ctx->personalities, index, creds)
8639 io_unregister_personality(ctx, index);
8640 mutex_unlock(&ctx->uring_lock);
8642 io_kill_timeouts(ctx, NULL, NULL);
8643 io_poll_remove_all(ctx, NULL, NULL);
8645 /* if we failed setting up the ctx, we might not have any rings */
8646 io_iopoll_try_reap_events(ctx);
8648 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8650 * Use system_unbound_wq to avoid spawning tons of event kworkers
8651 * if we're exiting a ton of rings at the same time. It just adds
8652 * noise and overhead, there's no discernable change in runtime
8653 * over using system_wq.
8655 queue_work(system_unbound_wq, &ctx->exit_work);
8658 static int io_uring_release(struct inode *inode, struct file *file)
8660 struct io_ring_ctx *ctx = file->private_data;
8662 file->private_data = NULL;
8663 io_ring_ctx_wait_and_kill(ctx);
8667 struct io_task_cancel {
8668 struct task_struct *task;
8669 struct files_struct *files;
8672 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8674 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8675 struct io_task_cancel *cancel = data;
8678 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8679 unsigned long flags;
8680 struct io_ring_ctx *ctx = req->ctx;
8682 /* protect against races with linked timeouts */
8683 spin_lock_irqsave(&ctx->completion_lock, flags);
8684 ret = io_match_task(req, cancel->task, cancel->files);
8685 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8687 ret = io_match_task(req, cancel->task, cancel->files);
8692 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8693 struct task_struct *task,
8694 struct files_struct *files)
8696 struct io_defer_entry *de;
8699 spin_lock_irq(&ctx->completion_lock);
8700 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8701 if (io_match_task(de->req, task, files)) {
8702 list_cut_position(&list, &ctx->defer_list, &de->list);
8706 spin_unlock_irq(&ctx->completion_lock);
8707 if (list_empty(&list))
8710 while (!list_empty(&list)) {
8711 de = list_first_entry(&list, struct io_defer_entry, list);
8712 list_del_init(&de->list);
8713 req_set_fail_links(de->req);
8714 io_put_req(de->req);
8715 io_req_complete(de->req, -ECANCELED);
8721 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8725 return req->ctx == data;
8728 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8730 struct io_tctx_node *node;
8731 enum io_wq_cancel cret;
8734 mutex_lock(&ctx->uring_lock);
8735 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8736 struct io_uring_task *tctx = node->task->io_uring;
8739 * io_wq will stay alive while we hold uring_lock, because it's
8740 * killed after ctx nodes, which requires to take the lock.
8742 if (!tctx || !tctx->io_wq)
8744 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8745 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8747 mutex_unlock(&ctx->uring_lock);
8752 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8753 struct task_struct *task,
8754 struct files_struct *files)
8756 struct io_task_cancel cancel = { .task = task, .files = files, };
8757 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8760 enum io_wq_cancel cret;
8764 ret |= io_uring_try_cancel_iowq(ctx);
8765 } else if (tctx && tctx->io_wq) {
8767 * Cancels requests of all rings, not only @ctx, but
8768 * it's fine as the task is in exit/exec.
8770 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8772 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8775 /* SQPOLL thread does its own polling */
8776 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8777 (ctx->sq_data && ctx->sq_data->thread == current)) {
8778 while (!list_empty_careful(&ctx->iopoll_list)) {
8779 io_iopoll_try_reap_events(ctx);
8784 ret |= io_cancel_defer_files(ctx, task, files);
8785 ret |= io_poll_remove_all(ctx, task, files);
8786 ret |= io_kill_timeouts(ctx, task, files);
8787 ret |= io_run_task_work();
8788 ret |= io_run_ctx_fallback(ctx);
8789 io_cqring_overflow_flush(ctx, true, task, files);
8796 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8797 struct task_struct *task,
8798 struct files_struct *files)
8800 struct io_kiocb *req;
8803 spin_lock_irq(&ctx->inflight_lock);
8804 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8805 cnt += io_match_task(req, task, files);
8806 spin_unlock_irq(&ctx->inflight_lock);
8810 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8811 struct task_struct *task,
8812 struct files_struct *files)
8814 while (!list_empty_careful(&ctx->inflight_list)) {
8818 inflight = io_uring_count_inflight(ctx, task, files);
8822 io_uring_try_cancel_requests(ctx, task, files);
8824 prepare_to_wait(&task->io_uring->wait, &wait,
8825 TASK_UNINTERRUPTIBLE);
8826 if (inflight == io_uring_count_inflight(ctx, task, files))
8828 finish_wait(&task->io_uring->wait, &wait);
8833 * Note that this task has used io_uring. We use it for cancelation purposes.
8835 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8837 struct io_uring_task *tctx = current->io_uring;
8838 struct io_tctx_node *node;
8841 if (unlikely(!tctx)) {
8842 ret = io_uring_alloc_task_context(current, ctx);
8845 tctx = current->io_uring;
8847 if (tctx->last != ctx) {
8848 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8851 node = kmalloc(sizeof(*node), GFP_KERNEL);
8855 node->task = current;
8857 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8864 mutex_lock(&ctx->uring_lock);
8865 list_add(&node->ctx_node, &ctx->tctx_list);
8866 mutex_unlock(&ctx->uring_lock);
8874 * Remove this io_uring_file -> task mapping.
8876 static void io_uring_del_task_file(unsigned long index)
8878 struct io_uring_task *tctx = current->io_uring;
8879 struct io_tctx_node *node;
8883 node = xa_erase(&tctx->xa, index);
8887 WARN_ON_ONCE(current != node->task);
8888 WARN_ON_ONCE(list_empty(&node->ctx_node));
8890 mutex_lock(&node->ctx->uring_lock);
8891 list_del(&node->ctx_node);
8892 mutex_unlock(&node->ctx->uring_lock);
8894 if (tctx->last == node->ctx)
8899 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8901 struct io_tctx_node *node;
8902 unsigned long index;
8904 xa_for_each(&tctx->xa, index, node)
8905 io_uring_del_task_file(index);
8907 io_wq_put_and_exit(tctx->io_wq);
8912 static s64 tctx_inflight(struct io_uring_task *tctx)
8914 return percpu_counter_sum(&tctx->inflight);
8917 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8919 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8920 struct io_ring_ctx *ctx = work->ctx;
8921 struct io_sq_data *sqd = ctx->sq_data;
8924 io_uring_cancel_sqpoll(ctx);
8925 complete(&work->completion);
8928 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8930 struct io_sq_data *sqd = ctx->sq_data;
8931 struct io_tctx_exit work = { .ctx = ctx, };
8932 struct task_struct *task;
8934 io_sq_thread_park(sqd);
8935 list_del_init(&ctx->sqd_list);
8936 io_sqd_update_thread_idle(sqd);
8939 init_completion(&work.completion);
8940 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8941 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8942 wake_up_process(task);
8944 io_sq_thread_unpark(sqd);
8947 wait_for_completion(&work.completion);
8950 void __io_uring_files_cancel(struct files_struct *files)
8952 struct io_uring_task *tctx = current->io_uring;
8953 struct io_tctx_node *node;
8954 unsigned long index;
8956 /* make sure overflow events are dropped */
8957 atomic_inc(&tctx->in_idle);
8958 xa_for_each(&tctx->xa, index, node) {
8959 struct io_ring_ctx *ctx = node->ctx;
8962 io_sqpoll_cancel_sync(ctx);
8965 io_uring_cancel_files(ctx, current, files);
8967 io_uring_try_cancel_requests(ctx, current, NULL);
8969 atomic_dec(&tctx->in_idle);
8972 io_uring_clean_tctx(tctx);
8975 /* should only be called by SQPOLL task */
8976 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8978 struct io_sq_data *sqd = ctx->sq_data;
8979 struct io_uring_task *tctx = current->io_uring;
8983 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8985 atomic_inc(&tctx->in_idle);
8987 /* read completions before cancelations */
8988 inflight = tctx_inflight(tctx);
8991 io_uring_try_cancel_requests(ctx, current, NULL);
8993 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8995 * If we've seen completions, retry without waiting. This
8996 * avoids a race where a completion comes in before we did
8997 * prepare_to_wait().
8999 if (inflight == tctx_inflight(tctx))
9001 finish_wait(&tctx->wait, &wait);
9003 atomic_dec(&tctx->in_idle);
9007 * Find any io_uring fd that this task has registered or done IO on, and cancel
9010 void __io_uring_task_cancel(void)
9012 struct io_uring_task *tctx = current->io_uring;
9016 /* make sure overflow events are dropped */
9017 atomic_inc(&tctx->in_idle);
9018 __io_uring_files_cancel(NULL);
9021 /* read completions before cancelations */
9022 inflight = tctx_inflight(tctx);
9025 __io_uring_files_cancel(NULL);
9027 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9030 * If we've seen completions, retry without waiting. This
9031 * avoids a race where a completion comes in before we did
9032 * prepare_to_wait().
9034 if (inflight == tctx_inflight(tctx))
9036 finish_wait(&tctx->wait, &wait);
9039 atomic_dec(&tctx->in_idle);
9041 io_uring_clean_tctx(tctx);
9042 /* all current's requests should be gone, we can kill tctx */
9043 __io_uring_free(current);
9046 static void *io_uring_validate_mmap_request(struct file *file,
9047 loff_t pgoff, size_t sz)
9049 struct io_ring_ctx *ctx = file->private_data;
9050 loff_t offset = pgoff << PAGE_SHIFT;
9055 case IORING_OFF_SQ_RING:
9056 case IORING_OFF_CQ_RING:
9059 case IORING_OFF_SQES:
9063 return ERR_PTR(-EINVAL);
9066 page = virt_to_head_page(ptr);
9067 if (sz > page_size(page))
9068 return ERR_PTR(-EINVAL);
9075 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9077 size_t sz = vma->vm_end - vma->vm_start;
9081 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9083 return PTR_ERR(ptr);
9085 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9086 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9089 #else /* !CONFIG_MMU */
9091 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9093 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9096 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9098 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9101 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9102 unsigned long addr, unsigned long len,
9103 unsigned long pgoff, unsigned long flags)
9107 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9109 return PTR_ERR(ptr);
9111 return (unsigned long) ptr;
9114 #endif /* !CONFIG_MMU */
9116 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9121 if (!io_sqring_full(ctx))
9123 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9125 if (!io_sqring_full(ctx))
9128 } while (!signal_pending(current));
9130 finish_wait(&ctx->sqo_sq_wait, &wait);
9134 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9135 struct __kernel_timespec __user **ts,
9136 const sigset_t __user **sig)
9138 struct io_uring_getevents_arg arg;
9141 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9142 * is just a pointer to the sigset_t.
9144 if (!(flags & IORING_ENTER_EXT_ARG)) {
9145 *sig = (const sigset_t __user *) argp;
9151 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9152 * timespec and sigset_t pointers if good.
9154 if (*argsz != sizeof(arg))
9156 if (copy_from_user(&arg, argp, sizeof(arg)))
9158 *sig = u64_to_user_ptr(arg.sigmask);
9159 *argsz = arg.sigmask_sz;
9160 *ts = u64_to_user_ptr(arg.ts);
9164 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9165 u32, min_complete, u32, flags, const void __user *, argp,
9168 struct io_ring_ctx *ctx;
9175 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9176 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9184 if (f.file->f_op != &io_uring_fops)
9188 ctx = f.file->private_data;
9189 if (!percpu_ref_tryget(&ctx->refs))
9193 if (ctx->flags & IORING_SETUP_R_DISABLED)
9197 * For SQ polling, the thread will do all submissions and completions.
9198 * Just return the requested submit count, and wake the thread if
9202 if (ctx->flags & IORING_SETUP_SQPOLL) {
9203 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9206 if (unlikely(ctx->sq_data->thread == NULL)) {
9209 if (flags & IORING_ENTER_SQ_WAKEUP)
9210 wake_up(&ctx->sq_data->wait);
9211 if (flags & IORING_ENTER_SQ_WAIT) {
9212 ret = io_sqpoll_wait_sq(ctx);
9216 submitted = to_submit;
9217 } else if (to_submit) {
9218 ret = io_uring_add_task_file(ctx);
9221 mutex_lock(&ctx->uring_lock);
9222 submitted = io_submit_sqes(ctx, to_submit);
9223 mutex_unlock(&ctx->uring_lock);
9225 if (submitted != to_submit)
9228 if (flags & IORING_ENTER_GETEVENTS) {
9229 const sigset_t __user *sig;
9230 struct __kernel_timespec __user *ts;
9232 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9236 min_complete = min(min_complete, ctx->cq_entries);
9239 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9240 * space applications don't need to do io completion events
9241 * polling again, they can rely on io_sq_thread to do polling
9242 * work, which can reduce cpu usage and uring_lock contention.
9244 if (ctx->flags & IORING_SETUP_IOPOLL &&
9245 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9246 ret = io_iopoll_check(ctx, min_complete);
9248 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9253 percpu_ref_put(&ctx->refs);
9256 return submitted ? submitted : ret;
9259 #ifdef CONFIG_PROC_FS
9260 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9261 const struct cred *cred)
9263 struct user_namespace *uns = seq_user_ns(m);
9264 struct group_info *gi;
9269 seq_printf(m, "%5d\n", id);
9270 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9271 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9272 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9273 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9274 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9275 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9276 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9277 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9278 seq_puts(m, "\n\tGroups:\t");
9279 gi = cred->group_info;
9280 for (g = 0; g < gi->ngroups; g++) {
9281 seq_put_decimal_ull(m, g ? " " : "",
9282 from_kgid_munged(uns, gi->gid[g]));
9284 seq_puts(m, "\n\tCapEff:\t");
9285 cap = cred->cap_effective;
9286 CAP_FOR_EACH_U32(__capi)
9287 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9292 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9294 struct io_sq_data *sq = NULL;
9299 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9300 * since fdinfo case grabs it in the opposite direction of normal use
9301 * cases. If we fail to get the lock, we just don't iterate any
9302 * structures that could be going away outside the io_uring mutex.
9304 has_lock = mutex_trylock(&ctx->uring_lock);
9306 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9312 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9313 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9314 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9315 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9316 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9319 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9321 seq_printf(m, "%5u: <none>\n", i);
9323 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9324 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9325 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9327 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9328 (unsigned int) buf->len);
9330 if (has_lock && !xa_empty(&ctx->personalities)) {
9331 unsigned long index;
9332 const struct cred *cred;
9334 seq_printf(m, "Personalities:\n");
9335 xa_for_each(&ctx->personalities, index, cred)
9336 io_uring_show_cred(m, index, cred);
9338 seq_printf(m, "PollList:\n");
9339 spin_lock_irq(&ctx->completion_lock);
9340 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9341 struct hlist_head *list = &ctx->cancel_hash[i];
9342 struct io_kiocb *req;
9344 hlist_for_each_entry(req, list, hash_node)
9345 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9346 req->task->task_works != NULL);
9348 spin_unlock_irq(&ctx->completion_lock);
9350 mutex_unlock(&ctx->uring_lock);
9353 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9355 struct io_ring_ctx *ctx = f->private_data;
9357 if (percpu_ref_tryget(&ctx->refs)) {
9358 __io_uring_show_fdinfo(ctx, m);
9359 percpu_ref_put(&ctx->refs);
9364 static const struct file_operations io_uring_fops = {
9365 .release = io_uring_release,
9366 .mmap = io_uring_mmap,
9368 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9369 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9371 .poll = io_uring_poll,
9372 .fasync = io_uring_fasync,
9373 #ifdef CONFIG_PROC_FS
9374 .show_fdinfo = io_uring_show_fdinfo,
9378 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9379 struct io_uring_params *p)
9381 struct io_rings *rings;
9382 size_t size, sq_array_offset;
9384 /* make sure these are sane, as we already accounted them */
9385 ctx->sq_entries = p->sq_entries;
9386 ctx->cq_entries = p->cq_entries;
9388 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9389 if (size == SIZE_MAX)
9392 rings = io_mem_alloc(size);
9397 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9398 rings->sq_ring_mask = p->sq_entries - 1;
9399 rings->cq_ring_mask = p->cq_entries - 1;
9400 rings->sq_ring_entries = p->sq_entries;
9401 rings->cq_ring_entries = p->cq_entries;
9402 ctx->sq_mask = rings->sq_ring_mask;
9403 ctx->cq_mask = rings->cq_ring_mask;
9405 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9406 if (size == SIZE_MAX) {
9407 io_mem_free(ctx->rings);
9412 ctx->sq_sqes = io_mem_alloc(size);
9413 if (!ctx->sq_sqes) {
9414 io_mem_free(ctx->rings);
9422 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9426 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9430 ret = io_uring_add_task_file(ctx);
9435 fd_install(fd, file);
9440 * Allocate an anonymous fd, this is what constitutes the application
9441 * visible backing of an io_uring instance. The application mmaps this
9442 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9443 * we have to tie this fd to a socket for file garbage collection purposes.
9445 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9448 #if defined(CONFIG_UNIX)
9451 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9454 return ERR_PTR(ret);
9457 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9458 O_RDWR | O_CLOEXEC);
9459 #if defined(CONFIG_UNIX)
9461 sock_release(ctx->ring_sock);
9462 ctx->ring_sock = NULL;
9464 ctx->ring_sock->file = file;
9470 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9471 struct io_uring_params __user *params)
9473 struct io_ring_ctx *ctx;
9479 if (entries > IORING_MAX_ENTRIES) {
9480 if (!(p->flags & IORING_SETUP_CLAMP))
9482 entries = IORING_MAX_ENTRIES;
9486 * Use twice as many entries for the CQ ring. It's possible for the
9487 * application to drive a higher depth than the size of the SQ ring,
9488 * since the sqes are only used at submission time. This allows for
9489 * some flexibility in overcommitting a bit. If the application has
9490 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9491 * of CQ ring entries manually.
9493 p->sq_entries = roundup_pow_of_two(entries);
9494 if (p->flags & IORING_SETUP_CQSIZE) {
9496 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9497 * to a power-of-two, if it isn't already. We do NOT impose
9498 * any cq vs sq ring sizing.
9502 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9503 if (!(p->flags & IORING_SETUP_CLAMP))
9505 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9507 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9508 if (p->cq_entries < p->sq_entries)
9511 p->cq_entries = 2 * p->sq_entries;
9514 ctx = io_ring_ctx_alloc(p);
9517 ctx->compat = in_compat_syscall();
9518 if (!capable(CAP_IPC_LOCK))
9519 ctx->user = get_uid(current_user());
9522 * This is just grabbed for accounting purposes. When a process exits,
9523 * the mm is exited and dropped before the files, hence we need to hang
9524 * on to this mm purely for the purposes of being able to unaccount
9525 * memory (locked/pinned vm). It's not used for anything else.
9527 mmgrab(current->mm);
9528 ctx->mm_account = current->mm;
9530 ret = io_allocate_scq_urings(ctx, p);
9534 ret = io_sq_offload_create(ctx, p);
9538 memset(&p->sq_off, 0, sizeof(p->sq_off));
9539 p->sq_off.head = offsetof(struct io_rings, sq.head);
9540 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9541 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9542 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9543 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9544 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9545 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9547 memset(&p->cq_off, 0, sizeof(p->cq_off));
9548 p->cq_off.head = offsetof(struct io_rings, cq.head);
9549 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9550 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9551 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9552 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9553 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9554 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9556 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9557 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9558 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9559 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9560 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9562 if (copy_to_user(params, p, sizeof(*p))) {
9567 file = io_uring_get_file(ctx);
9569 ret = PTR_ERR(file);
9574 * Install ring fd as the very last thing, so we don't risk someone
9575 * having closed it before we finish setup
9577 ret = io_uring_install_fd(ctx, file);
9579 /* fput will clean it up */
9584 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9587 io_ring_ctx_wait_and_kill(ctx);
9592 * Sets up an aio uring context, and returns the fd. Applications asks for a
9593 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9594 * params structure passed in.
9596 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9598 struct io_uring_params p;
9601 if (copy_from_user(&p, params, sizeof(p)))
9603 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9608 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9609 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9610 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9611 IORING_SETUP_R_DISABLED))
9614 return io_uring_create(entries, &p, params);
9617 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9618 struct io_uring_params __user *, params)
9620 return io_uring_setup(entries, params);
9623 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9625 struct io_uring_probe *p;
9629 size = struct_size(p, ops, nr_args);
9630 if (size == SIZE_MAX)
9632 p = kzalloc(size, GFP_KERNEL);
9637 if (copy_from_user(p, arg, size))
9640 if (memchr_inv(p, 0, size))
9643 p->last_op = IORING_OP_LAST - 1;
9644 if (nr_args > IORING_OP_LAST)
9645 nr_args = IORING_OP_LAST;
9647 for (i = 0; i < nr_args; i++) {
9649 if (!io_op_defs[i].not_supported)
9650 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9655 if (copy_to_user(arg, p, size))
9662 static int io_register_personality(struct io_ring_ctx *ctx)
9664 const struct cred *creds;
9668 creds = get_current_cred();
9670 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9671 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9678 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9679 unsigned int nr_args)
9681 struct io_uring_restriction *res;
9685 /* Restrictions allowed only if rings started disabled */
9686 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9689 /* We allow only a single restrictions registration */
9690 if (ctx->restrictions.registered)
9693 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9696 size = array_size(nr_args, sizeof(*res));
9697 if (size == SIZE_MAX)
9700 res = memdup_user(arg, size);
9702 return PTR_ERR(res);
9706 for (i = 0; i < nr_args; i++) {
9707 switch (res[i].opcode) {
9708 case IORING_RESTRICTION_REGISTER_OP:
9709 if (res[i].register_op >= IORING_REGISTER_LAST) {
9714 __set_bit(res[i].register_op,
9715 ctx->restrictions.register_op);
9717 case IORING_RESTRICTION_SQE_OP:
9718 if (res[i].sqe_op >= IORING_OP_LAST) {
9723 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9725 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9726 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9728 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9729 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9738 /* Reset all restrictions if an error happened */
9740 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9742 ctx->restrictions.registered = true;
9748 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9750 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9753 if (ctx->restrictions.registered)
9754 ctx->restricted = 1;
9756 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9757 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9758 wake_up(&ctx->sq_data->wait);
9762 static bool io_register_op_must_quiesce(int op)
9765 case IORING_UNREGISTER_FILES:
9766 case IORING_REGISTER_FILES_UPDATE:
9767 case IORING_REGISTER_PROBE:
9768 case IORING_REGISTER_PERSONALITY:
9769 case IORING_UNREGISTER_PERSONALITY:
9776 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9777 void __user *arg, unsigned nr_args)
9778 __releases(ctx->uring_lock)
9779 __acquires(ctx->uring_lock)
9784 * We're inside the ring mutex, if the ref is already dying, then
9785 * someone else killed the ctx or is already going through
9786 * io_uring_register().
9788 if (percpu_ref_is_dying(&ctx->refs))
9791 if (io_register_op_must_quiesce(opcode)) {
9792 percpu_ref_kill(&ctx->refs);
9795 * Drop uring mutex before waiting for references to exit. If
9796 * another thread is currently inside io_uring_enter() it might
9797 * need to grab the uring_lock to make progress. If we hold it
9798 * here across the drain wait, then we can deadlock. It's safe
9799 * to drop the mutex here, since no new references will come in
9800 * after we've killed the percpu ref.
9802 mutex_unlock(&ctx->uring_lock);
9804 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9807 ret = io_run_task_work_sig();
9812 mutex_lock(&ctx->uring_lock);
9815 percpu_ref_resurrect(&ctx->refs);
9820 if (ctx->restricted) {
9821 if (opcode >= IORING_REGISTER_LAST) {
9826 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9833 case IORING_REGISTER_BUFFERS:
9834 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9836 case IORING_UNREGISTER_BUFFERS:
9840 ret = io_sqe_buffers_unregister(ctx);
9842 case IORING_REGISTER_FILES:
9843 ret = io_sqe_files_register(ctx, arg, nr_args);
9845 case IORING_UNREGISTER_FILES:
9849 ret = io_sqe_files_unregister(ctx);
9851 case IORING_REGISTER_FILES_UPDATE:
9852 ret = io_sqe_files_update(ctx, arg, nr_args);
9854 case IORING_REGISTER_EVENTFD:
9855 case IORING_REGISTER_EVENTFD_ASYNC:
9859 ret = io_eventfd_register(ctx, arg);
9862 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9863 ctx->eventfd_async = 1;
9865 ctx->eventfd_async = 0;
9867 case IORING_UNREGISTER_EVENTFD:
9871 ret = io_eventfd_unregister(ctx);
9873 case IORING_REGISTER_PROBE:
9875 if (!arg || nr_args > 256)
9877 ret = io_probe(ctx, arg, nr_args);
9879 case IORING_REGISTER_PERSONALITY:
9883 ret = io_register_personality(ctx);
9885 case IORING_UNREGISTER_PERSONALITY:
9889 ret = io_unregister_personality(ctx, nr_args);
9891 case IORING_REGISTER_ENABLE_RINGS:
9895 ret = io_register_enable_rings(ctx);
9897 case IORING_REGISTER_RESTRICTIONS:
9898 ret = io_register_restrictions(ctx, arg, nr_args);
9906 if (io_register_op_must_quiesce(opcode)) {
9907 /* bring the ctx back to life */
9908 percpu_ref_reinit(&ctx->refs);
9910 reinit_completion(&ctx->ref_comp);
9915 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9916 void __user *, arg, unsigned int, nr_args)
9918 struct io_ring_ctx *ctx;
9927 if (f.file->f_op != &io_uring_fops)
9930 ctx = f.file->private_data;
9934 mutex_lock(&ctx->uring_lock);
9935 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9936 mutex_unlock(&ctx->uring_lock);
9937 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9938 ctx->cq_ev_fd != NULL, ret);
9944 static int __init io_uring_init(void)
9946 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9947 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9948 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9951 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9952 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9953 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9954 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9955 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9956 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9957 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9958 BUILD_BUG_SQE_ELEM(8, __u64, off);
9959 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9960 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9961 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9962 BUILD_BUG_SQE_ELEM(24, __u32, len);
9963 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9964 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9965 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9966 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9967 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9968 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9969 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9970 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9971 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9972 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9973 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9974 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9975 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9976 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9977 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9978 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9979 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9980 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9981 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9983 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9984 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9985 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9989 __initcall(io_uring_init);