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,
701 /* not a real bit, just to check we're not overflowing the space */
707 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
708 /* drain existing IO first */
709 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
711 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
712 /* doesn't sever on completion < 0 */
713 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
715 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
716 /* IOSQE_BUFFER_SELECT */
717 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
719 /* fail rest of links */
720 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
721 /* on inflight list, should be cancelled and waited on exit reliably */
722 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
723 /* read/write uses file position */
724 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
725 /* must not punt to workers */
726 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
727 /* has or had linked timeout */
728 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
730 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
732 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
733 /* already went through poll handler */
734 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
735 /* buffer already selected */
736 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
737 /* doesn't need file table for this request */
738 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
739 /* linked timeout is active, i.e. prepared by link's head */
740 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
741 /* completion is deferred through io_comp_state */
742 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
746 struct io_poll_iocb poll;
747 struct io_poll_iocb *double_poll;
750 struct io_task_work {
751 struct io_wq_work_node node;
752 task_work_func_t func;
756 * NOTE! Each of the iocb union members has the file pointer
757 * as the first entry in their struct definition. So you can
758 * access the file pointer through any of the sub-structs,
759 * or directly as just 'ki_filp' in this struct.
765 struct io_poll_iocb poll;
766 struct io_poll_remove poll_remove;
767 struct io_accept accept;
769 struct io_cancel cancel;
770 struct io_timeout timeout;
771 struct io_timeout_rem timeout_rem;
772 struct io_connect connect;
773 struct io_sr_msg sr_msg;
775 struct io_close close;
776 struct io_rsrc_update rsrc_update;
777 struct io_fadvise fadvise;
778 struct io_madvise madvise;
779 struct io_epoll epoll;
780 struct io_splice splice;
781 struct io_provide_buf pbuf;
782 struct io_statx statx;
783 struct io_shutdown shutdown;
784 struct io_rename rename;
785 struct io_unlink unlink;
786 /* use only after cleaning per-op data, see io_clean_op() */
787 struct io_completion compl;
790 /* opcode allocated if it needs to store data for async defer */
793 /* polled IO has completed */
799 struct io_ring_ctx *ctx;
802 struct task_struct *task;
805 struct io_kiocb *link;
806 struct percpu_ref *fixed_rsrc_refs;
809 * 1. used with ctx->iopoll_list with reads/writes
810 * 2. to track reqs with ->files (see io_op_def::file_table)
812 struct list_head inflight_entry;
814 struct io_task_work io_task_work;
815 struct callback_head task_work;
817 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
818 struct hlist_node hash_node;
819 struct async_poll *apoll;
820 struct io_wq_work work;
823 struct io_tctx_node {
824 struct list_head ctx_node;
825 struct task_struct *task;
826 struct io_ring_ctx *ctx;
829 struct io_defer_entry {
830 struct list_head list;
831 struct io_kiocb *req;
836 /* needs req->file assigned */
837 unsigned needs_file : 1;
838 /* hash wq insertion if file is a regular file */
839 unsigned hash_reg_file : 1;
840 /* unbound wq insertion if file is a non-regular file */
841 unsigned unbound_nonreg_file : 1;
842 /* opcode is not supported by this kernel */
843 unsigned not_supported : 1;
844 /* set if opcode supports polled "wait" */
846 unsigned pollout : 1;
847 /* op supports buffer selection */
848 unsigned buffer_select : 1;
849 /* must always have async data allocated */
850 unsigned needs_async_data : 1;
851 /* should block plug */
853 /* size of async data needed, if any */
854 unsigned short async_size;
857 static const struct io_op_def io_op_defs[] = {
858 [IORING_OP_NOP] = {},
859 [IORING_OP_READV] = {
861 .unbound_nonreg_file = 1,
864 .needs_async_data = 1,
866 .async_size = sizeof(struct io_async_rw),
868 [IORING_OP_WRITEV] = {
871 .unbound_nonreg_file = 1,
873 .needs_async_data = 1,
875 .async_size = sizeof(struct io_async_rw),
877 [IORING_OP_FSYNC] = {
880 [IORING_OP_READ_FIXED] = {
882 .unbound_nonreg_file = 1,
885 .async_size = sizeof(struct io_async_rw),
887 [IORING_OP_WRITE_FIXED] = {
890 .unbound_nonreg_file = 1,
893 .async_size = sizeof(struct io_async_rw),
895 [IORING_OP_POLL_ADD] = {
897 .unbound_nonreg_file = 1,
899 [IORING_OP_POLL_REMOVE] = {},
900 [IORING_OP_SYNC_FILE_RANGE] = {
903 [IORING_OP_SENDMSG] = {
905 .unbound_nonreg_file = 1,
907 .needs_async_data = 1,
908 .async_size = sizeof(struct io_async_msghdr),
910 [IORING_OP_RECVMSG] = {
912 .unbound_nonreg_file = 1,
915 .needs_async_data = 1,
916 .async_size = sizeof(struct io_async_msghdr),
918 [IORING_OP_TIMEOUT] = {
919 .needs_async_data = 1,
920 .async_size = sizeof(struct io_timeout_data),
922 [IORING_OP_TIMEOUT_REMOVE] = {
923 /* used by timeout updates' prep() */
925 [IORING_OP_ACCEPT] = {
927 .unbound_nonreg_file = 1,
930 [IORING_OP_ASYNC_CANCEL] = {},
931 [IORING_OP_LINK_TIMEOUT] = {
932 .needs_async_data = 1,
933 .async_size = sizeof(struct io_timeout_data),
935 [IORING_OP_CONNECT] = {
937 .unbound_nonreg_file = 1,
939 .needs_async_data = 1,
940 .async_size = sizeof(struct io_async_connect),
942 [IORING_OP_FALLOCATE] = {
945 [IORING_OP_OPENAT] = {},
946 [IORING_OP_CLOSE] = {},
947 [IORING_OP_FILES_UPDATE] = {},
948 [IORING_OP_STATX] = {},
951 .unbound_nonreg_file = 1,
955 .async_size = sizeof(struct io_async_rw),
957 [IORING_OP_WRITE] = {
959 .unbound_nonreg_file = 1,
962 .async_size = sizeof(struct io_async_rw),
964 [IORING_OP_FADVISE] = {
967 [IORING_OP_MADVISE] = {},
970 .unbound_nonreg_file = 1,
975 .unbound_nonreg_file = 1,
979 [IORING_OP_OPENAT2] = {
981 [IORING_OP_EPOLL_CTL] = {
982 .unbound_nonreg_file = 1,
984 [IORING_OP_SPLICE] = {
987 .unbound_nonreg_file = 1,
989 [IORING_OP_PROVIDE_BUFFERS] = {},
990 [IORING_OP_REMOVE_BUFFERS] = {},
994 .unbound_nonreg_file = 1,
996 [IORING_OP_SHUTDOWN] = {
999 [IORING_OP_RENAMEAT] = {},
1000 [IORING_OP_UNLINKAT] = {},
1003 static bool io_disarm_next(struct io_kiocb *req);
1004 static void io_uring_del_task_file(unsigned long index);
1005 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1006 struct task_struct *task,
1007 struct files_struct *files);
1008 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1009 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1010 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1011 struct io_ring_ctx *ctx);
1012 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1014 static bool io_rw_reissue(struct io_kiocb *req);
1015 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1016 static void io_put_req(struct io_kiocb *req);
1017 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1018 static void io_double_put_req(struct io_kiocb *req);
1019 static void io_dismantle_req(struct io_kiocb *req);
1020 static void io_put_task(struct task_struct *task, int nr);
1021 static void io_queue_next(struct io_kiocb *req);
1022 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1023 static void __io_queue_linked_timeout(struct io_kiocb *req);
1024 static void io_queue_linked_timeout(struct io_kiocb *req);
1025 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1026 struct io_uring_rsrc_update *ip,
1028 static void __io_clean_op(struct io_kiocb *req);
1029 static struct file *io_file_get(struct io_submit_state *state,
1030 struct io_kiocb *req, int fd, bool fixed);
1031 static void __io_queue_sqe(struct io_kiocb *req);
1032 static void io_rsrc_put_work(struct work_struct *work);
1034 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1035 struct iov_iter *iter, bool needs_lock);
1036 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1037 const struct iovec *fast_iov,
1038 struct iov_iter *iter, bool force);
1039 static void io_req_task_queue(struct io_kiocb *req);
1040 static void io_submit_flush_completions(struct io_comp_state *cs,
1041 struct io_ring_ctx *ctx);
1043 static struct kmem_cache *req_cachep;
1045 static const struct file_operations io_uring_fops;
1047 struct sock *io_uring_get_socket(struct file *file)
1049 #if defined(CONFIG_UNIX)
1050 if (file->f_op == &io_uring_fops) {
1051 struct io_ring_ctx *ctx = file->private_data;
1053 return ctx->ring_sock->sk;
1058 EXPORT_SYMBOL(io_uring_get_socket);
1060 #define io_for_each_link(pos, head) \
1061 for (pos = (head); pos; pos = pos->link)
1063 static inline void io_clean_op(struct io_kiocb *req)
1065 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1069 static inline void io_set_resource_node(struct io_kiocb *req)
1071 struct io_ring_ctx *ctx = req->ctx;
1073 if (!req->fixed_rsrc_refs) {
1074 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1075 percpu_ref_get(req->fixed_rsrc_refs);
1079 static bool io_match_task(struct io_kiocb *head,
1080 struct task_struct *task,
1081 struct files_struct *files)
1083 struct io_kiocb *req;
1085 if (task && head->task != task) {
1086 /* in terms of cancelation, always match if req task is dead */
1087 if (head->task->flags & PF_EXITING)
1094 io_for_each_link(req, head) {
1095 if (req->flags & REQ_F_INFLIGHT)
1097 if (req->task->files == files)
1103 static inline void req_set_fail_links(struct io_kiocb *req)
1105 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1106 req->flags |= REQ_F_FAIL_LINK;
1109 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1111 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1113 complete(&ctx->ref_comp);
1116 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1118 return !req->timeout.off;
1121 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1123 struct io_ring_ctx *ctx;
1126 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1131 * Use 5 bits less than the max cq entries, that should give us around
1132 * 32 entries per hash list if totally full and uniformly spread.
1134 hash_bits = ilog2(p->cq_entries);
1138 ctx->cancel_hash_bits = hash_bits;
1139 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1141 if (!ctx->cancel_hash)
1143 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1145 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1146 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1149 ctx->flags = p->flags;
1150 init_waitqueue_head(&ctx->sqo_sq_wait);
1151 INIT_LIST_HEAD(&ctx->sqd_list);
1152 init_waitqueue_head(&ctx->cq_wait);
1153 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1154 init_completion(&ctx->ref_comp);
1155 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1156 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1157 mutex_init(&ctx->uring_lock);
1158 init_waitqueue_head(&ctx->wait);
1159 spin_lock_init(&ctx->completion_lock);
1160 INIT_LIST_HEAD(&ctx->iopoll_list);
1161 INIT_LIST_HEAD(&ctx->defer_list);
1162 INIT_LIST_HEAD(&ctx->timeout_list);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 spin_lock_init(&ctx->rsrc_ref_lock);
1166 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1167 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1168 init_llist_head(&ctx->rsrc_put_llist);
1169 INIT_LIST_HEAD(&ctx->tctx_list);
1170 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1174 kfree(ctx->cancel_hash);
1179 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1181 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1182 struct io_ring_ctx *ctx = req->ctx;
1184 return seq != ctx->cached_cq_tail
1185 + READ_ONCE(ctx->cached_cq_overflow);
1191 static void io_req_track_inflight(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1195 if (!(req->flags & REQ_F_INFLIGHT)) {
1196 req->flags |= REQ_F_INFLIGHT;
1198 spin_lock_irq(&ctx->inflight_lock);
1199 list_add(&req->inflight_entry, &ctx->inflight_list);
1200 spin_unlock_irq(&ctx->inflight_lock);
1204 static void io_prep_async_work(struct io_kiocb *req)
1206 const struct io_op_def *def = &io_op_defs[req->opcode];
1207 struct io_ring_ctx *ctx = req->ctx;
1209 if (!req->work.creds)
1210 req->work.creds = get_current_cred();
1212 if (req->flags & REQ_F_FORCE_ASYNC)
1213 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1215 if (req->flags & REQ_F_ISREG) {
1216 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1217 io_wq_hash_work(&req->work, file_inode(req->file));
1219 if (def->unbound_nonreg_file)
1220 req->work.flags |= IO_WQ_WORK_UNBOUND;
1224 static void io_prep_async_link(struct io_kiocb *req)
1226 struct io_kiocb *cur;
1228 io_for_each_link(cur, req)
1229 io_prep_async_work(cur);
1232 static void io_queue_async_work(struct io_kiocb *req)
1234 struct io_ring_ctx *ctx = req->ctx;
1235 struct io_kiocb *link = io_prep_linked_timeout(req);
1236 struct io_uring_task *tctx = req->task->io_uring;
1239 BUG_ON(!tctx->io_wq);
1241 /* init ->work of the whole link before punting */
1242 io_prep_async_link(req);
1243 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1244 &req->work, req->flags);
1245 io_wq_enqueue(tctx->io_wq, &req->work);
1247 io_queue_linked_timeout(link);
1250 static void io_kill_timeout(struct io_kiocb *req)
1252 struct io_timeout_data *io = req->async_data;
1255 ret = hrtimer_try_to_cancel(&io->timer);
1257 atomic_set(&req->ctx->cq_timeouts,
1258 atomic_read(&req->ctx->cq_timeouts) + 1);
1259 list_del_init(&req->timeout.list);
1260 io_cqring_fill_event(req, 0);
1261 io_put_req_deferred(req, 1);
1266 * Returns true if we found and killed one or more timeouts
1268 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1269 struct files_struct *files)
1271 struct io_kiocb *req, *tmp;
1274 spin_lock_irq(&ctx->completion_lock);
1275 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1276 if (io_match_task(req, tsk, files)) {
1277 io_kill_timeout(req);
1281 spin_unlock_irq(&ctx->completion_lock);
1282 return canceled != 0;
1285 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1288 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1289 struct io_defer_entry, list);
1291 if (req_need_defer(de->req, de->seq))
1293 list_del_init(&de->list);
1294 io_req_task_queue(de->req);
1296 } while (!list_empty(&ctx->defer_list));
1299 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1303 if (list_empty(&ctx->timeout_list))
1306 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1309 u32 events_needed, events_got;
1310 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1311 struct io_kiocb, timeout.list);
1313 if (io_is_timeout_noseq(req))
1317 * Since seq can easily wrap around over time, subtract
1318 * the last seq at which timeouts were flushed before comparing.
1319 * Assuming not more than 2^31-1 events have happened since,
1320 * these subtractions won't have wrapped, so we can check if
1321 * target is in [last_seq, current_seq] by comparing the two.
1323 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1324 events_got = seq - ctx->cq_last_tm_flush;
1325 if (events_got < events_needed)
1328 list_del_init(&req->timeout.list);
1329 io_kill_timeout(req);
1330 } while (!list_empty(&ctx->timeout_list));
1332 ctx->cq_last_tm_flush = seq;
1335 static void io_commit_cqring(struct io_ring_ctx *ctx)
1337 io_flush_timeouts(ctx);
1339 /* order cqe stores with ring update */
1340 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1342 if (unlikely(!list_empty(&ctx->defer_list)))
1343 __io_queue_deferred(ctx);
1346 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1348 struct io_rings *r = ctx->rings;
1350 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1353 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1355 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1358 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1360 struct io_rings *rings = ctx->rings;
1364 * writes to the cq entry need to come after reading head; the
1365 * control dependency is enough as we're using WRITE_ONCE to
1368 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1371 tail = ctx->cached_cq_tail++;
1372 return &rings->cqes[tail & ctx->cq_mask];
1375 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1379 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1381 if (!ctx->eventfd_async)
1383 return io_wq_current_is_worker();
1386 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1388 /* see waitqueue_active() comment */
1391 if (waitqueue_active(&ctx->wait))
1392 wake_up(&ctx->wait);
1393 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1394 wake_up(&ctx->sq_data->wait);
1395 if (io_should_trigger_evfd(ctx))
1396 eventfd_signal(ctx->cq_ev_fd, 1);
1397 if (waitqueue_active(&ctx->cq_wait)) {
1398 wake_up_interruptible(&ctx->cq_wait);
1399 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1403 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1405 /* see waitqueue_active() comment */
1408 if (ctx->flags & IORING_SETUP_SQPOLL) {
1409 if (waitqueue_active(&ctx->wait))
1410 wake_up(&ctx->wait);
1412 if (io_should_trigger_evfd(ctx))
1413 eventfd_signal(ctx->cq_ev_fd, 1);
1414 if (waitqueue_active(&ctx->cq_wait)) {
1415 wake_up_interruptible(&ctx->cq_wait);
1416 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1420 /* Returns true if there are no backlogged entries after the flush */
1421 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1422 struct task_struct *tsk,
1423 struct files_struct *files)
1425 struct io_rings *rings = ctx->rings;
1426 struct io_kiocb *req, *tmp;
1427 struct io_uring_cqe *cqe;
1428 unsigned long flags;
1429 bool all_flushed, posted;
1432 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1436 spin_lock_irqsave(&ctx->completion_lock, flags);
1437 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1438 if (!io_match_task(req, tsk, files))
1441 cqe = io_get_cqring(ctx);
1445 list_move(&req->compl.list, &list);
1447 WRITE_ONCE(cqe->user_data, req->user_data);
1448 WRITE_ONCE(cqe->res, req->result);
1449 WRITE_ONCE(cqe->flags, req->compl.cflags);
1451 ctx->cached_cq_overflow++;
1452 WRITE_ONCE(ctx->rings->cq_overflow,
1453 ctx->cached_cq_overflow);
1458 all_flushed = list_empty(&ctx->cq_overflow_list);
1460 clear_bit(0, &ctx->sq_check_overflow);
1461 clear_bit(0, &ctx->cq_check_overflow);
1462 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1466 io_commit_cqring(ctx);
1467 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1469 io_cqring_ev_posted(ctx);
1471 while (!list_empty(&list)) {
1472 req = list_first_entry(&list, struct io_kiocb, compl.list);
1473 list_del(&req->compl.list);
1480 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1481 struct task_struct *tsk,
1482 struct files_struct *files)
1486 if (test_bit(0, &ctx->cq_check_overflow)) {
1487 /* iopoll syncs against uring_lock, not completion_lock */
1488 if (ctx->flags & IORING_SETUP_IOPOLL)
1489 mutex_lock(&ctx->uring_lock);
1490 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1491 if (ctx->flags & IORING_SETUP_IOPOLL)
1492 mutex_unlock(&ctx->uring_lock);
1498 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1500 struct io_ring_ctx *ctx = req->ctx;
1501 struct io_uring_cqe *cqe;
1503 trace_io_uring_complete(ctx, req->user_data, res);
1506 * If we can't get a cq entry, userspace overflowed the
1507 * submission (by quite a lot). Increment the overflow count in
1510 cqe = io_get_cqring(ctx);
1512 WRITE_ONCE(cqe->user_data, req->user_data);
1513 WRITE_ONCE(cqe->res, res);
1514 WRITE_ONCE(cqe->flags, cflags);
1515 } else if (ctx->cq_overflow_flushed ||
1516 atomic_read(&req->task->io_uring->in_idle)) {
1518 * If we're in ring overflow flush mode, or in task cancel mode,
1519 * then we cannot store the request for later flushing, we need
1520 * to drop it on the floor.
1522 ctx->cached_cq_overflow++;
1523 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1525 if (list_empty(&ctx->cq_overflow_list)) {
1526 set_bit(0, &ctx->sq_check_overflow);
1527 set_bit(0, &ctx->cq_check_overflow);
1528 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1532 req->compl.cflags = cflags;
1533 refcount_inc(&req->refs);
1534 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1538 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1540 __io_cqring_fill_event(req, res, 0);
1543 static void io_req_complete_post(struct io_kiocb *req, long res,
1544 unsigned int cflags)
1546 struct io_ring_ctx *ctx = req->ctx;
1547 unsigned long flags;
1549 spin_lock_irqsave(&ctx->completion_lock, flags);
1550 __io_cqring_fill_event(req, res, cflags);
1552 * If we're the last reference to this request, add to our locked
1555 if (refcount_dec_and_test(&req->refs)) {
1556 struct io_comp_state *cs = &ctx->submit_state.comp;
1558 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1559 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1560 io_disarm_next(req);
1562 io_req_task_queue(req->link);
1566 io_dismantle_req(req);
1567 io_put_task(req->task, 1);
1568 list_add(&req->compl.list, &cs->locked_free_list);
1569 cs->locked_free_nr++;
1571 if (!percpu_ref_tryget(&ctx->refs))
1574 io_commit_cqring(ctx);
1575 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1578 io_cqring_ev_posted(ctx);
1579 percpu_ref_put(&ctx->refs);
1583 static void io_req_complete_state(struct io_kiocb *req, long res,
1584 unsigned int cflags)
1588 req->compl.cflags = cflags;
1589 req->flags |= REQ_F_COMPLETE_INLINE;
1592 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1593 long res, unsigned cflags)
1595 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1596 io_req_complete_state(req, res, cflags);
1598 io_req_complete_post(req, res, cflags);
1601 static inline void io_req_complete(struct io_kiocb *req, long res)
1603 __io_req_complete(req, 0, res, 0);
1606 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1608 struct io_submit_state *state = &ctx->submit_state;
1609 struct io_comp_state *cs = &state->comp;
1610 struct io_kiocb *req = NULL;
1613 * If we have more than a batch's worth of requests in our IRQ side
1614 * locked cache, grab the lock and move them over to our submission
1617 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1618 spin_lock_irq(&ctx->completion_lock);
1619 list_splice_init(&cs->locked_free_list, &cs->free_list);
1620 cs->locked_free_nr = 0;
1621 spin_unlock_irq(&ctx->completion_lock);
1624 while (!list_empty(&cs->free_list)) {
1625 req = list_first_entry(&cs->free_list, struct io_kiocb,
1627 list_del(&req->compl.list);
1628 state->reqs[state->free_reqs++] = req;
1629 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1636 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1638 struct io_submit_state *state = &ctx->submit_state;
1640 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1642 if (!state->free_reqs) {
1643 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1646 if (io_flush_cached_reqs(ctx))
1649 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1653 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1654 * retry single alloc to be on the safe side.
1656 if (unlikely(ret <= 0)) {
1657 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1658 if (!state->reqs[0])
1662 state->free_reqs = ret;
1666 return state->reqs[state->free_reqs];
1669 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1676 static void io_dismantle_req(struct io_kiocb *req)
1680 if (req->async_data)
1681 kfree(req->async_data);
1683 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1684 if (req->fixed_rsrc_refs)
1685 percpu_ref_put(req->fixed_rsrc_refs);
1686 if (req->work.creds) {
1687 put_cred(req->work.creds);
1688 req->work.creds = NULL;
1691 if (req->flags & REQ_F_INFLIGHT) {
1692 struct io_ring_ctx *ctx = req->ctx;
1693 unsigned long flags;
1695 spin_lock_irqsave(&ctx->inflight_lock, flags);
1696 list_del(&req->inflight_entry);
1697 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1698 req->flags &= ~REQ_F_INFLIGHT;
1702 /* must to be called somewhat shortly after putting a request */
1703 static inline void io_put_task(struct task_struct *task, int nr)
1705 struct io_uring_task *tctx = task->io_uring;
1707 percpu_counter_sub(&tctx->inflight, nr);
1708 if (unlikely(atomic_read(&tctx->in_idle)))
1709 wake_up(&tctx->wait);
1710 put_task_struct_many(task, nr);
1713 static void __io_free_req(struct io_kiocb *req)
1715 struct io_ring_ctx *ctx = req->ctx;
1717 io_dismantle_req(req);
1718 io_put_task(req->task, 1);
1720 kmem_cache_free(req_cachep, req);
1721 percpu_ref_put(&ctx->refs);
1724 static inline void io_remove_next_linked(struct io_kiocb *req)
1726 struct io_kiocb *nxt = req->link;
1728 req->link = nxt->link;
1732 static bool io_kill_linked_timeout(struct io_kiocb *req)
1733 __must_hold(&req->ctx->completion_lock)
1735 struct io_kiocb *link = req->link;
1736 bool cancelled = false;
1739 * Can happen if a linked timeout fired and link had been like
1740 * req -> link t-out -> link t-out [-> ...]
1742 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1743 struct io_timeout_data *io = link->async_data;
1746 io_remove_next_linked(req);
1747 link->timeout.head = NULL;
1748 ret = hrtimer_try_to_cancel(&io->timer);
1750 io_cqring_fill_event(link, -ECANCELED);
1751 io_put_req_deferred(link, 1);
1755 req->flags &= ~REQ_F_LINK_TIMEOUT;
1759 static void io_fail_links(struct io_kiocb *req)
1760 __must_hold(&req->ctx->completion_lock)
1762 struct io_kiocb *nxt, *link = req->link;
1769 trace_io_uring_fail_link(req, link);
1770 io_cqring_fill_event(link, -ECANCELED);
1771 io_put_req_deferred(link, 2);
1776 static bool io_disarm_next(struct io_kiocb *req)
1777 __must_hold(&req->ctx->completion_lock)
1779 bool posted = false;
1781 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1782 posted = io_kill_linked_timeout(req);
1783 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1784 posted |= (req->link != NULL);
1790 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1792 struct io_kiocb *nxt;
1795 * If LINK is set, we have dependent requests in this chain. If we
1796 * didn't fail this request, queue the first one up, moving any other
1797 * dependencies to the next request. In case of failure, fail the rest
1800 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1801 struct io_ring_ctx *ctx = req->ctx;
1802 unsigned long flags;
1805 spin_lock_irqsave(&ctx->completion_lock, flags);
1806 posted = io_disarm_next(req);
1808 io_commit_cqring(req->ctx);
1809 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1811 io_cqring_ev_posted(ctx);
1818 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1820 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1822 return __io_req_find_next(req);
1825 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1829 if (ctx->submit_state.comp.nr) {
1830 mutex_lock(&ctx->uring_lock);
1831 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1832 mutex_unlock(&ctx->uring_lock);
1834 percpu_ref_put(&ctx->refs);
1837 static bool __tctx_task_work(struct io_uring_task *tctx)
1839 struct io_ring_ctx *ctx = NULL;
1840 struct io_wq_work_list list;
1841 struct io_wq_work_node *node;
1843 if (wq_list_empty(&tctx->task_list))
1846 spin_lock_irq(&tctx->task_lock);
1847 list = tctx->task_list;
1848 INIT_WQ_LIST(&tctx->task_list);
1849 spin_unlock_irq(&tctx->task_lock);
1853 struct io_wq_work_node *next = node->next;
1854 struct io_kiocb *req;
1856 req = container_of(node, struct io_kiocb, io_task_work.node);
1857 if (req->ctx != ctx) {
1858 ctx_flush_and_put(ctx);
1860 percpu_ref_get(&ctx->refs);
1863 req->task_work.func(&req->task_work);
1867 ctx_flush_and_put(ctx);
1868 return list.first != NULL;
1871 static void tctx_task_work(struct callback_head *cb)
1873 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1875 clear_bit(0, &tctx->task_state);
1877 while (__tctx_task_work(tctx))
1881 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1882 enum task_work_notify_mode notify)
1884 struct io_uring_task *tctx = tsk->io_uring;
1885 struct io_wq_work_node *node, *prev;
1886 unsigned long flags;
1889 WARN_ON_ONCE(!tctx);
1891 spin_lock_irqsave(&tctx->task_lock, flags);
1892 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1893 spin_unlock_irqrestore(&tctx->task_lock, flags);
1895 /* task_work already pending, we're done */
1896 if (test_bit(0, &tctx->task_state) ||
1897 test_and_set_bit(0, &tctx->task_state))
1900 if (!task_work_add(tsk, &tctx->task_work, notify))
1904 * Slow path - we failed, find and delete work. if the work is not
1905 * in the list, it got run and we're fine.
1908 spin_lock_irqsave(&tctx->task_lock, flags);
1909 wq_list_for_each(node, prev, &tctx->task_list) {
1910 if (&req->io_task_work.node == node) {
1911 wq_list_del(&tctx->task_list, node, prev);
1916 spin_unlock_irqrestore(&tctx->task_lock, flags);
1917 clear_bit(0, &tctx->task_state);
1921 static int io_req_task_work_add(struct io_kiocb *req)
1923 struct task_struct *tsk = req->task;
1924 struct io_ring_ctx *ctx = req->ctx;
1925 enum task_work_notify_mode notify;
1928 if (tsk->flags & PF_EXITING)
1932 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1933 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1934 * processing task_work. There's no reliable way to tell if TWA_RESUME
1938 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1939 notify = TWA_SIGNAL;
1941 ret = io_task_work_add(tsk, req, notify);
1943 wake_up_process(tsk);
1948 static bool io_run_task_work_head(struct callback_head **work_head)
1950 struct callback_head *work, *next;
1951 bool executed = false;
1954 work = xchg(work_head, NULL);
1970 static void io_task_work_add_head(struct callback_head **work_head,
1971 struct callback_head *task_work)
1973 struct callback_head *head;
1976 head = READ_ONCE(*work_head);
1977 task_work->next = head;
1978 } while (cmpxchg(work_head, head, task_work) != head);
1981 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1982 task_work_func_t cb)
1984 init_task_work(&req->task_work, cb);
1985 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1988 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1990 struct io_ring_ctx *ctx = req->ctx;
1992 spin_lock_irq(&ctx->completion_lock);
1993 io_cqring_fill_event(req, error);
1994 io_commit_cqring(ctx);
1995 spin_unlock_irq(&ctx->completion_lock);
1997 io_cqring_ev_posted(ctx);
1998 req_set_fail_links(req);
1999 io_double_put_req(req);
2002 static void io_req_task_cancel(struct callback_head *cb)
2004 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2005 struct io_ring_ctx *ctx = req->ctx;
2007 mutex_lock(&ctx->uring_lock);
2008 __io_req_task_cancel(req, req->result);
2009 mutex_unlock(&ctx->uring_lock);
2010 percpu_ref_put(&ctx->refs);
2013 static void __io_req_task_submit(struct io_kiocb *req)
2015 struct io_ring_ctx *ctx = req->ctx;
2017 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2018 mutex_lock(&ctx->uring_lock);
2019 if (!(current->flags & PF_EXITING) && !current->in_execve)
2020 __io_queue_sqe(req);
2022 __io_req_task_cancel(req, -EFAULT);
2023 mutex_unlock(&ctx->uring_lock);
2026 static void io_req_task_submit(struct callback_head *cb)
2028 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2030 __io_req_task_submit(req);
2033 static void io_req_task_queue(struct io_kiocb *req)
2037 req->task_work.func = io_req_task_submit;
2038 ret = io_req_task_work_add(req);
2039 if (unlikely(ret)) {
2040 req->result = -ECANCELED;
2041 percpu_ref_get(&req->ctx->refs);
2042 io_req_task_work_add_fallback(req, io_req_task_cancel);
2046 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2048 percpu_ref_get(&req->ctx->refs);
2050 req->task_work.func = io_req_task_cancel;
2052 if (unlikely(io_req_task_work_add(req)))
2053 io_req_task_work_add_fallback(req, io_req_task_cancel);
2056 static inline void io_queue_next(struct io_kiocb *req)
2058 struct io_kiocb *nxt = io_req_find_next(req);
2061 io_req_task_queue(nxt);
2064 static void io_free_req(struct io_kiocb *req)
2071 struct task_struct *task;
2076 static inline void io_init_req_batch(struct req_batch *rb)
2083 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2084 struct req_batch *rb)
2087 io_put_task(rb->task, rb->task_refs);
2089 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2092 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2093 struct io_submit_state *state)
2097 if (req->task != rb->task) {
2099 io_put_task(rb->task, rb->task_refs);
2100 rb->task = req->task;
2106 io_dismantle_req(req);
2107 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2108 state->reqs[state->free_reqs++] = req;
2110 list_add(&req->compl.list, &state->comp.free_list);
2113 static void io_submit_flush_completions(struct io_comp_state *cs,
2114 struct io_ring_ctx *ctx)
2117 struct io_kiocb *req;
2118 struct req_batch rb;
2120 io_init_req_batch(&rb);
2121 spin_lock_irq(&ctx->completion_lock);
2122 for (i = 0; i < nr; i++) {
2124 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2126 io_commit_cqring(ctx);
2127 spin_unlock_irq(&ctx->completion_lock);
2129 io_cqring_ev_posted(ctx);
2130 for (i = 0; i < nr; i++) {
2133 /* submission and completion refs */
2134 if (refcount_sub_and_test(2, &req->refs))
2135 io_req_free_batch(&rb, req, &ctx->submit_state);
2138 io_req_free_batch_finish(ctx, &rb);
2143 * Drop reference to request, return next in chain (if there is one) if this
2144 * was the last reference to this request.
2146 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2148 struct io_kiocb *nxt = NULL;
2150 if (refcount_dec_and_test(&req->refs)) {
2151 nxt = io_req_find_next(req);
2157 static void io_put_req(struct io_kiocb *req)
2159 if (refcount_dec_and_test(&req->refs))
2163 static void io_put_req_deferred_cb(struct callback_head *cb)
2165 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2170 static void io_free_req_deferred(struct io_kiocb *req)
2174 req->task_work.func = io_put_req_deferred_cb;
2175 ret = io_req_task_work_add(req);
2177 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2180 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2182 if (refcount_sub_and_test(refs, &req->refs))
2183 io_free_req_deferred(req);
2186 static void io_double_put_req(struct io_kiocb *req)
2188 /* drop both submit and complete references */
2189 if (refcount_sub_and_test(2, &req->refs))
2193 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2195 /* See comment at the top of this file */
2197 return __io_cqring_events(ctx);
2200 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2202 struct io_rings *rings = ctx->rings;
2204 /* make sure SQ entry isn't read before tail */
2205 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2208 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2210 unsigned int cflags;
2212 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2213 cflags |= IORING_CQE_F_BUFFER;
2214 req->flags &= ~REQ_F_BUFFER_SELECTED;
2219 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2221 struct io_buffer *kbuf;
2223 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2224 return io_put_kbuf(req, kbuf);
2227 static inline bool io_run_task_work(void)
2230 * Not safe to run on exiting task, and the task_work handling will
2231 * not add work to such a task.
2233 if (unlikely(current->flags & PF_EXITING))
2235 if (current->task_works) {
2236 __set_current_state(TASK_RUNNING);
2245 * Find and free completed poll iocbs
2247 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2248 struct list_head *done)
2250 struct req_batch rb;
2251 struct io_kiocb *req;
2253 /* order with ->result store in io_complete_rw_iopoll() */
2256 io_init_req_batch(&rb);
2257 while (!list_empty(done)) {
2260 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2261 list_del(&req->inflight_entry);
2263 if (READ_ONCE(req->result) == -EAGAIN) {
2264 req->iopoll_completed = 0;
2265 if (io_rw_reissue(req))
2269 if (req->flags & REQ_F_BUFFER_SELECTED)
2270 cflags = io_put_rw_kbuf(req);
2272 __io_cqring_fill_event(req, req->result, cflags);
2275 if (refcount_dec_and_test(&req->refs))
2276 io_req_free_batch(&rb, req, &ctx->submit_state);
2279 io_commit_cqring(ctx);
2280 io_cqring_ev_posted_iopoll(ctx);
2281 io_req_free_batch_finish(ctx, &rb);
2284 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2287 struct io_kiocb *req, *tmp;
2293 * Only spin for completions if we don't have multiple devices hanging
2294 * off our complete list, and we're under the requested amount.
2296 spin = !ctx->poll_multi_file && *nr_events < min;
2299 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2300 struct kiocb *kiocb = &req->rw.kiocb;
2303 * Move completed and retryable entries to our local lists.
2304 * If we find a request that requires polling, break out
2305 * and complete those lists first, if we have entries there.
2307 if (READ_ONCE(req->iopoll_completed)) {
2308 list_move_tail(&req->inflight_entry, &done);
2311 if (!list_empty(&done))
2314 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2318 /* iopoll may have completed current req */
2319 if (READ_ONCE(req->iopoll_completed))
2320 list_move_tail(&req->inflight_entry, &done);
2327 if (!list_empty(&done))
2328 io_iopoll_complete(ctx, nr_events, &done);
2334 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2335 * non-spinning poll check - we'll still enter the driver poll loop, but only
2336 * as a non-spinning completion check.
2338 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2341 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2344 ret = io_do_iopoll(ctx, nr_events, min);
2347 if (*nr_events >= min)
2355 * We can't just wait for polled events to come to us, we have to actively
2356 * find and complete them.
2358 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2360 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2363 mutex_lock(&ctx->uring_lock);
2364 while (!list_empty(&ctx->iopoll_list)) {
2365 unsigned int nr_events = 0;
2367 io_do_iopoll(ctx, &nr_events, 0);
2369 /* let it sleep and repeat later if can't complete a request */
2373 * Ensure we allow local-to-the-cpu processing to take place,
2374 * in this case we need to ensure that we reap all events.
2375 * Also let task_work, etc. to progress by releasing the mutex
2377 if (need_resched()) {
2378 mutex_unlock(&ctx->uring_lock);
2380 mutex_lock(&ctx->uring_lock);
2383 mutex_unlock(&ctx->uring_lock);
2386 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2388 unsigned int nr_events = 0;
2389 int iters = 0, ret = 0;
2392 * We disallow the app entering submit/complete with polling, but we
2393 * still need to lock the ring to prevent racing with polled issue
2394 * that got punted to a workqueue.
2396 mutex_lock(&ctx->uring_lock);
2399 * Don't enter poll loop if we already have events pending.
2400 * If we do, we can potentially be spinning for commands that
2401 * already triggered a CQE (eg in error).
2403 if (test_bit(0, &ctx->cq_check_overflow))
2404 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2405 if (io_cqring_events(ctx))
2409 * If a submit got punted to a workqueue, we can have the
2410 * application entering polling for a command before it gets
2411 * issued. That app will hold the uring_lock for the duration
2412 * of the poll right here, so we need to take a breather every
2413 * now and then to ensure that the issue has a chance to add
2414 * the poll to the issued list. Otherwise we can spin here
2415 * forever, while the workqueue is stuck trying to acquire the
2418 if (!(++iters & 7)) {
2419 mutex_unlock(&ctx->uring_lock);
2421 mutex_lock(&ctx->uring_lock);
2424 ret = io_iopoll_getevents(ctx, &nr_events, min);
2428 } while (min && !nr_events && !need_resched());
2430 mutex_unlock(&ctx->uring_lock);
2434 static void kiocb_end_write(struct io_kiocb *req)
2437 * Tell lockdep we inherited freeze protection from submission
2440 if (req->flags & REQ_F_ISREG) {
2441 struct inode *inode = file_inode(req->file);
2443 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2445 file_end_write(req->file);
2449 static bool io_resubmit_prep(struct io_kiocb *req)
2451 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2453 struct iov_iter iter;
2455 /* already prepared */
2456 if (req->async_data)
2459 switch (req->opcode) {
2460 case IORING_OP_READV:
2461 case IORING_OP_READ_FIXED:
2462 case IORING_OP_READ:
2465 case IORING_OP_WRITEV:
2466 case IORING_OP_WRITE_FIXED:
2467 case IORING_OP_WRITE:
2471 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2476 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2479 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2482 static bool io_rw_should_reissue(struct io_kiocb *req)
2484 umode_t mode = file_inode(req->file)->i_mode;
2485 struct io_ring_ctx *ctx = req->ctx;
2487 if (!S_ISBLK(mode) && !S_ISREG(mode))
2489 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2490 !(ctx->flags & IORING_SETUP_IOPOLL)))
2493 * If ref is dying, we might be running poll reap from the exit work.
2494 * Don't attempt to reissue from that path, just let it fail with
2497 if (percpu_ref_is_dying(&ctx->refs))
2503 static bool io_rw_reissue(struct io_kiocb *req)
2506 if (!io_rw_should_reissue(req))
2509 lockdep_assert_held(&req->ctx->uring_lock);
2511 if (io_resubmit_prep(req)) {
2512 refcount_inc(&req->refs);
2513 io_queue_async_work(req);
2516 req_set_fail_links(req);
2521 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2522 unsigned int issue_flags)
2526 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2527 kiocb_end_write(req);
2528 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2530 if (res != req->result)
2531 req_set_fail_links(req);
2532 if (req->flags & REQ_F_BUFFER_SELECTED)
2533 cflags = io_put_rw_kbuf(req);
2534 __io_req_complete(req, issue_flags, res, cflags);
2537 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2539 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2541 __io_complete_rw(req, res, res2, 0);
2544 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2546 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2549 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2550 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2551 struct io_async_rw *rw = req->async_data;
2554 iov_iter_revert(&rw->iter,
2555 req->result - iov_iter_count(&rw->iter));
2556 else if (!io_resubmit_prep(req))
2561 if (kiocb->ki_flags & IOCB_WRITE)
2562 kiocb_end_write(req);
2564 if (res != -EAGAIN && res != req->result)
2565 req_set_fail_links(req);
2567 WRITE_ONCE(req->result, res);
2568 /* order with io_poll_complete() checking ->result */
2570 WRITE_ONCE(req->iopoll_completed, 1);
2574 * After the iocb has been issued, it's safe to be found on the poll list.
2575 * Adding the kiocb to the list AFTER submission ensures that we don't
2576 * find it from a io_iopoll_getevents() thread before the issuer is done
2577 * accessing the kiocb cookie.
2579 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2581 struct io_ring_ctx *ctx = req->ctx;
2584 * Track whether we have multiple files in our lists. This will impact
2585 * how we do polling eventually, not spinning if we're on potentially
2586 * different devices.
2588 if (list_empty(&ctx->iopoll_list)) {
2589 ctx->poll_multi_file = false;
2590 } else if (!ctx->poll_multi_file) {
2591 struct io_kiocb *list_req;
2593 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2595 if (list_req->file != req->file)
2596 ctx->poll_multi_file = true;
2600 * For fast devices, IO may have already completed. If it has, add
2601 * it to the front so we find it first.
2603 if (READ_ONCE(req->iopoll_completed))
2604 list_add(&req->inflight_entry, &ctx->iopoll_list);
2606 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2609 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2610 * task context or in io worker task context. If current task context is
2611 * sq thread, we don't need to check whether should wake up sq thread.
2613 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2614 wq_has_sleeper(&ctx->sq_data->wait))
2615 wake_up(&ctx->sq_data->wait);
2618 static inline void io_state_file_put(struct io_submit_state *state)
2620 if (state->file_refs) {
2621 fput_many(state->file, state->file_refs);
2622 state->file_refs = 0;
2627 * Get as many references to a file as we have IOs left in this submission,
2628 * assuming most submissions are for one file, or at least that each file
2629 * has more than one submission.
2631 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2636 if (state->file_refs) {
2637 if (state->fd == fd) {
2641 io_state_file_put(state);
2643 state->file = fget_many(fd, state->ios_left);
2644 if (unlikely(!state->file))
2648 state->file_refs = state->ios_left - 1;
2652 static bool io_bdev_nowait(struct block_device *bdev)
2654 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2658 * If we tracked the file through the SCM inflight mechanism, we could support
2659 * any file. For now, just ensure that anything potentially problematic is done
2662 static bool io_file_supports_async(struct file *file, int rw)
2664 umode_t mode = file_inode(file)->i_mode;
2666 if (S_ISBLK(mode)) {
2667 if (IS_ENABLED(CONFIG_BLOCK) &&
2668 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2672 if (S_ISCHR(mode) || S_ISSOCK(mode))
2674 if (S_ISREG(mode)) {
2675 if (IS_ENABLED(CONFIG_BLOCK) &&
2676 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2677 file->f_op != &io_uring_fops)
2682 /* any ->read/write should understand O_NONBLOCK */
2683 if (file->f_flags & O_NONBLOCK)
2686 if (!(file->f_mode & FMODE_NOWAIT))
2690 return file->f_op->read_iter != NULL;
2692 return file->f_op->write_iter != NULL;
2695 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2697 struct io_ring_ctx *ctx = req->ctx;
2698 struct kiocb *kiocb = &req->rw.kiocb;
2699 struct file *file = req->file;
2703 if (S_ISREG(file_inode(file)->i_mode))
2704 req->flags |= REQ_F_ISREG;
2706 kiocb->ki_pos = READ_ONCE(sqe->off);
2707 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2708 req->flags |= REQ_F_CUR_POS;
2709 kiocb->ki_pos = file->f_pos;
2711 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2712 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2713 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2717 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2718 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2719 req->flags |= REQ_F_NOWAIT;
2721 ioprio = READ_ONCE(sqe->ioprio);
2723 ret = ioprio_check_cap(ioprio);
2727 kiocb->ki_ioprio = ioprio;
2729 kiocb->ki_ioprio = get_current_ioprio();
2731 if (ctx->flags & IORING_SETUP_IOPOLL) {
2732 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2733 !kiocb->ki_filp->f_op->iopoll)
2736 kiocb->ki_flags |= IOCB_HIPRI;
2737 kiocb->ki_complete = io_complete_rw_iopoll;
2738 req->iopoll_completed = 0;
2740 if (kiocb->ki_flags & IOCB_HIPRI)
2742 kiocb->ki_complete = io_complete_rw;
2745 req->rw.addr = READ_ONCE(sqe->addr);
2746 req->rw.len = READ_ONCE(sqe->len);
2747 req->buf_index = READ_ONCE(sqe->buf_index);
2751 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2757 case -ERESTARTNOINTR:
2758 case -ERESTARTNOHAND:
2759 case -ERESTART_RESTARTBLOCK:
2761 * We can't just restart the syscall, since previously
2762 * submitted sqes may already be in progress. Just fail this
2768 kiocb->ki_complete(kiocb, ret, 0);
2772 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2773 unsigned int issue_flags)
2775 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2776 struct io_async_rw *io = req->async_data;
2778 /* add previously done IO, if any */
2779 if (io && io->bytes_done > 0) {
2781 ret = io->bytes_done;
2783 ret += io->bytes_done;
2786 if (req->flags & REQ_F_CUR_POS)
2787 req->file->f_pos = kiocb->ki_pos;
2788 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2789 __io_complete_rw(req, ret, 0, issue_flags);
2791 io_rw_done(kiocb, ret);
2794 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2796 struct io_ring_ctx *ctx = req->ctx;
2797 size_t len = req->rw.len;
2798 struct io_mapped_ubuf *imu;
2799 u16 index, buf_index = req->buf_index;
2803 if (unlikely(buf_index >= ctx->nr_user_bufs))
2805 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2806 imu = &ctx->user_bufs[index];
2807 buf_addr = req->rw.addr;
2810 if (buf_addr + len < buf_addr)
2812 /* not inside the mapped region */
2813 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2817 * May not be a start of buffer, set size appropriately
2818 * and advance us to the beginning.
2820 offset = buf_addr - imu->ubuf;
2821 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2825 * Don't use iov_iter_advance() here, as it's really slow for
2826 * using the latter parts of a big fixed buffer - it iterates
2827 * over each segment manually. We can cheat a bit here, because
2830 * 1) it's a BVEC iter, we set it up
2831 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2832 * first and last bvec
2834 * So just find our index, and adjust the iterator afterwards.
2835 * If the offset is within the first bvec (or the whole first
2836 * bvec, just use iov_iter_advance(). This makes it easier
2837 * since we can just skip the first segment, which may not
2838 * be PAGE_SIZE aligned.
2840 const struct bio_vec *bvec = imu->bvec;
2842 if (offset <= bvec->bv_len) {
2843 iov_iter_advance(iter, offset);
2845 unsigned long seg_skip;
2847 /* skip first vec */
2848 offset -= bvec->bv_len;
2849 seg_skip = 1 + (offset >> PAGE_SHIFT);
2851 iter->bvec = bvec + seg_skip;
2852 iter->nr_segs -= seg_skip;
2853 iter->count -= bvec->bv_len + offset;
2854 iter->iov_offset = offset & ~PAGE_MASK;
2861 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2864 mutex_unlock(&ctx->uring_lock);
2867 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2870 * "Normal" inline submissions always hold the uring_lock, since we
2871 * grab it from the system call. Same is true for the SQPOLL offload.
2872 * The only exception is when we've detached the request and issue it
2873 * from an async worker thread, grab the lock for that case.
2876 mutex_lock(&ctx->uring_lock);
2879 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2880 int bgid, struct io_buffer *kbuf,
2883 struct io_buffer *head;
2885 if (req->flags & REQ_F_BUFFER_SELECTED)
2888 io_ring_submit_lock(req->ctx, needs_lock);
2890 lockdep_assert_held(&req->ctx->uring_lock);
2892 head = xa_load(&req->ctx->io_buffers, bgid);
2894 if (!list_empty(&head->list)) {
2895 kbuf = list_last_entry(&head->list, struct io_buffer,
2897 list_del(&kbuf->list);
2900 xa_erase(&req->ctx->io_buffers, bgid);
2902 if (*len > kbuf->len)
2905 kbuf = ERR_PTR(-ENOBUFS);
2908 io_ring_submit_unlock(req->ctx, needs_lock);
2913 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2916 struct io_buffer *kbuf;
2919 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2920 bgid = req->buf_index;
2921 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2924 req->rw.addr = (u64) (unsigned long) kbuf;
2925 req->flags |= REQ_F_BUFFER_SELECTED;
2926 return u64_to_user_ptr(kbuf->addr);
2929 #ifdef CONFIG_COMPAT
2930 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2933 struct compat_iovec __user *uiov;
2934 compat_ssize_t clen;
2938 uiov = u64_to_user_ptr(req->rw.addr);
2939 if (!access_ok(uiov, sizeof(*uiov)))
2941 if (__get_user(clen, &uiov->iov_len))
2947 buf = io_rw_buffer_select(req, &len, needs_lock);
2949 return PTR_ERR(buf);
2950 iov[0].iov_base = buf;
2951 iov[0].iov_len = (compat_size_t) len;
2956 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2959 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2963 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2966 len = iov[0].iov_len;
2969 buf = io_rw_buffer_select(req, &len, needs_lock);
2971 return PTR_ERR(buf);
2972 iov[0].iov_base = buf;
2973 iov[0].iov_len = len;
2977 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2980 if (req->flags & REQ_F_BUFFER_SELECTED) {
2981 struct io_buffer *kbuf;
2983 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2984 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2985 iov[0].iov_len = kbuf->len;
2988 if (req->rw.len != 1)
2991 #ifdef CONFIG_COMPAT
2992 if (req->ctx->compat)
2993 return io_compat_import(req, iov, needs_lock);
2996 return __io_iov_buffer_select(req, iov, needs_lock);
2999 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3000 struct iov_iter *iter, bool needs_lock)
3002 void __user *buf = u64_to_user_ptr(req->rw.addr);
3003 size_t sqe_len = req->rw.len;
3004 u8 opcode = req->opcode;
3007 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3009 return io_import_fixed(req, rw, iter);
3012 /* buffer index only valid with fixed read/write, or buffer select */
3013 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3016 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3017 if (req->flags & REQ_F_BUFFER_SELECT) {
3018 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3020 return PTR_ERR(buf);
3021 req->rw.len = sqe_len;
3024 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3029 if (req->flags & REQ_F_BUFFER_SELECT) {
3030 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3032 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3037 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3041 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3043 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3047 * For files that don't have ->read_iter() and ->write_iter(), handle them
3048 * by looping over ->read() or ->write() manually.
3050 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3052 struct kiocb *kiocb = &req->rw.kiocb;
3053 struct file *file = req->file;
3057 * Don't support polled IO through this interface, and we can't
3058 * support non-blocking either. For the latter, this just causes
3059 * the kiocb to be handled from an async context.
3061 if (kiocb->ki_flags & IOCB_HIPRI)
3063 if (kiocb->ki_flags & IOCB_NOWAIT)
3066 while (iov_iter_count(iter)) {
3070 if (!iov_iter_is_bvec(iter)) {
3071 iovec = iov_iter_iovec(iter);
3073 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3074 iovec.iov_len = req->rw.len;
3078 nr = file->f_op->read(file, iovec.iov_base,
3079 iovec.iov_len, io_kiocb_ppos(kiocb));
3081 nr = file->f_op->write(file, iovec.iov_base,
3082 iovec.iov_len, io_kiocb_ppos(kiocb));
3091 if (nr != iovec.iov_len)
3095 iov_iter_advance(iter, nr);
3101 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3102 const struct iovec *fast_iov, struct iov_iter *iter)
3104 struct io_async_rw *rw = req->async_data;
3106 memcpy(&rw->iter, iter, sizeof(*iter));
3107 rw->free_iovec = iovec;
3109 /* can only be fixed buffers, no need to do anything */
3110 if (iov_iter_is_bvec(iter))
3113 unsigned iov_off = 0;
3115 rw->iter.iov = rw->fast_iov;
3116 if (iter->iov != fast_iov) {
3117 iov_off = iter->iov - fast_iov;
3118 rw->iter.iov += iov_off;
3120 if (rw->fast_iov != fast_iov)
3121 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3122 sizeof(struct iovec) * iter->nr_segs);
3124 req->flags |= REQ_F_NEED_CLEANUP;
3128 static inline int __io_alloc_async_data(struct io_kiocb *req)
3130 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3131 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3132 return req->async_data == NULL;
3135 static int io_alloc_async_data(struct io_kiocb *req)
3137 if (!io_op_defs[req->opcode].needs_async_data)
3140 return __io_alloc_async_data(req);
3143 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3144 const struct iovec *fast_iov,
3145 struct iov_iter *iter, bool force)
3147 if (!force && !io_op_defs[req->opcode].needs_async_data)
3149 if (!req->async_data) {
3150 if (__io_alloc_async_data(req)) {
3155 io_req_map_rw(req, iovec, fast_iov, iter);
3160 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3162 struct io_async_rw *iorw = req->async_data;
3163 struct iovec *iov = iorw->fast_iov;
3166 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3167 if (unlikely(ret < 0))
3170 iorw->bytes_done = 0;
3171 iorw->free_iovec = iov;
3173 req->flags |= REQ_F_NEED_CLEANUP;
3177 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3179 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3181 return io_prep_rw(req, sqe);
3185 * This is our waitqueue callback handler, registered through lock_page_async()
3186 * when we initially tried to do the IO with the iocb armed our waitqueue.
3187 * This gets called when the page is unlocked, and we generally expect that to
3188 * happen when the page IO is completed and the page is now uptodate. This will
3189 * queue a task_work based retry of the operation, attempting to copy the data
3190 * again. If the latter fails because the page was NOT uptodate, then we will
3191 * do a thread based blocking retry of the operation. That's the unexpected
3194 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3195 int sync, void *arg)
3197 struct wait_page_queue *wpq;
3198 struct io_kiocb *req = wait->private;
3199 struct wait_page_key *key = arg;
3201 wpq = container_of(wait, struct wait_page_queue, wait);
3203 if (!wake_page_match(wpq, key))
3206 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3207 list_del_init(&wait->entry);
3209 /* submit ref gets dropped, acquire a new one */
3210 refcount_inc(&req->refs);
3211 io_req_task_queue(req);
3216 * This controls whether a given IO request should be armed for async page
3217 * based retry. If we return false here, the request is handed to the async
3218 * worker threads for retry. If we're doing buffered reads on a regular file,
3219 * we prepare a private wait_page_queue entry and retry the operation. This
3220 * will either succeed because the page is now uptodate and unlocked, or it
3221 * will register a callback when the page is unlocked at IO completion. Through
3222 * that callback, io_uring uses task_work to setup a retry of the operation.
3223 * That retry will attempt the buffered read again. The retry will generally
3224 * succeed, or in rare cases where it fails, we then fall back to using the
3225 * async worker threads for a blocking retry.
3227 static bool io_rw_should_retry(struct io_kiocb *req)
3229 struct io_async_rw *rw = req->async_data;
3230 struct wait_page_queue *wait = &rw->wpq;
3231 struct kiocb *kiocb = &req->rw.kiocb;
3233 /* never retry for NOWAIT, we just complete with -EAGAIN */
3234 if (req->flags & REQ_F_NOWAIT)
3237 /* Only for buffered IO */
3238 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3242 * just use poll if we can, and don't attempt if the fs doesn't
3243 * support callback based unlocks
3245 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3248 wait->wait.func = io_async_buf_func;
3249 wait->wait.private = req;
3250 wait->wait.flags = 0;
3251 INIT_LIST_HEAD(&wait->wait.entry);
3252 kiocb->ki_flags |= IOCB_WAITQ;
3253 kiocb->ki_flags &= ~IOCB_NOWAIT;
3254 kiocb->ki_waitq = wait;
3258 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3260 if (req->file->f_op->read_iter)
3261 return call_read_iter(req->file, &req->rw.kiocb, iter);
3262 else if (req->file->f_op->read)
3263 return loop_rw_iter(READ, req, iter);
3268 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3270 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3271 struct kiocb *kiocb = &req->rw.kiocb;
3272 struct iov_iter __iter, *iter = &__iter;
3273 struct io_async_rw *rw = req->async_data;
3274 ssize_t io_size, ret, ret2;
3275 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3281 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3285 io_size = iov_iter_count(iter);
3286 req->result = io_size;
3288 /* Ensure we clear previously set non-block flag */
3289 if (!force_nonblock)
3290 kiocb->ki_flags &= ~IOCB_NOWAIT;
3292 kiocb->ki_flags |= IOCB_NOWAIT;
3294 /* If the file doesn't support async, just async punt */
3295 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3296 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3297 return ret ?: -EAGAIN;
3300 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3301 if (unlikely(ret)) {
3306 ret = io_iter_do_read(req, iter);
3308 if (ret == -EIOCBQUEUED) {
3309 if (req->async_data)
3310 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3312 } else if (ret == -EAGAIN) {
3313 /* IOPOLL retry should happen for io-wq threads */
3314 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3316 /* no retry on NONBLOCK nor RWF_NOWAIT */
3317 if (req->flags & REQ_F_NOWAIT)
3319 /* some cases will consume bytes even on error returns */
3320 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3322 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3323 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3324 /* read all, failed, already did sync or don't want to retry */
3328 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3333 rw = req->async_data;
3334 /* now use our persistent iterator, if we aren't already */
3339 rw->bytes_done += ret;
3340 /* if we can retry, do so with the callbacks armed */
3341 if (!io_rw_should_retry(req)) {
3342 kiocb->ki_flags &= ~IOCB_WAITQ;
3347 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3348 * we get -EIOCBQUEUED, then we'll get a notification when the
3349 * desired page gets unlocked. We can also get a partial read
3350 * here, and if we do, then just retry at the new offset.
3352 ret = io_iter_do_read(req, iter);
3353 if (ret == -EIOCBQUEUED)
3355 /* we got some bytes, but not all. retry. */
3356 kiocb->ki_flags &= ~IOCB_WAITQ;
3357 } while (ret > 0 && ret < io_size);
3359 kiocb_done(kiocb, ret, issue_flags);
3361 /* it's faster to check here then delegate to kfree */
3367 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3369 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3371 return io_prep_rw(req, sqe);
3374 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3376 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3377 struct kiocb *kiocb = &req->rw.kiocb;
3378 struct iov_iter __iter, *iter = &__iter;
3379 struct io_async_rw *rw = req->async_data;
3380 ssize_t ret, ret2, io_size;
3381 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3387 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3391 io_size = iov_iter_count(iter);
3392 req->result = io_size;
3394 /* Ensure we clear previously set non-block flag */
3395 if (!force_nonblock)
3396 kiocb->ki_flags &= ~IOCB_NOWAIT;
3398 kiocb->ki_flags |= IOCB_NOWAIT;
3400 /* If the file doesn't support async, just async punt */
3401 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3404 /* file path doesn't support NOWAIT for non-direct_IO */
3405 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3406 (req->flags & REQ_F_ISREG))
3409 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3414 * Open-code file_start_write here to grab freeze protection,
3415 * which will be released by another thread in
3416 * io_complete_rw(). Fool lockdep by telling it the lock got
3417 * released so that it doesn't complain about the held lock when
3418 * we return to userspace.
3420 if (req->flags & REQ_F_ISREG) {
3421 sb_start_write(file_inode(req->file)->i_sb);
3422 __sb_writers_release(file_inode(req->file)->i_sb,
3425 kiocb->ki_flags |= IOCB_WRITE;
3427 if (req->file->f_op->write_iter)
3428 ret2 = call_write_iter(req->file, kiocb, iter);
3429 else if (req->file->f_op->write)
3430 ret2 = loop_rw_iter(WRITE, req, iter);
3435 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3436 * retry them without IOCB_NOWAIT.
3438 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3440 /* no retry on NONBLOCK nor RWF_NOWAIT */
3441 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3443 if (ret2 == -EIOCBQUEUED && req->async_data)
3444 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3445 if (!force_nonblock || ret2 != -EAGAIN) {
3446 /* IOPOLL retry should happen for io-wq threads */
3447 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3450 kiocb_done(kiocb, ret2, issue_flags);
3453 /* some cases will consume bytes even on error returns */
3454 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3455 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3456 return ret ?: -EAGAIN;
3459 /* it's reportedly faster than delegating the null check to kfree() */
3465 static int io_renameat_prep(struct io_kiocb *req,
3466 const struct io_uring_sqe *sqe)
3468 struct io_rename *ren = &req->rename;
3469 const char __user *oldf, *newf;
3471 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3474 ren->old_dfd = READ_ONCE(sqe->fd);
3475 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3476 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3477 ren->new_dfd = READ_ONCE(sqe->len);
3478 ren->flags = READ_ONCE(sqe->rename_flags);
3480 ren->oldpath = getname(oldf);
3481 if (IS_ERR(ren->oldpath))
3482 return PTR_ERR(ren->oldpath);
3484 ren->newpath = getname(newf);
3485 if (IS_ERR(ren->newpath)) {
3486 putname(ren->oldpath);
3487 return PTR_ERR(ren->newpath);
3490 req->flags |= REQ_F_NEED_CLEANUP;
3494 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3496 struct io_rename *ren = &req->rename;
3499 if (issue_flags & IO_URING_F_NONBLOCK)
3502 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3503 ren->newpath, ren->flags);
3505 req->flags &= ~REQ_F_NEED_CLEANUP;
3507 req_set_fail_links(req);
3508 io_req_complete(req, ret);
3512 static int io_unlinkat_prep(struct io_kiocb *req,
3513 const struct io_uring_sqe *sqe)
3515 struct io_unlink *un = &req->unlink;
3516 const char __user *fname;
3518 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3521 un->dfd = READ_ONCE(sqe->fd);
3523 un->flags = READ_ONCE(sqe->unlink_flags);
3524 if (un->flags & ~AT_REMOVEDIR)
3527 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3528 un->filename = getname(fname);
3529 if (IS_ERR(un->filename))
3530 return PTR_ERR(un->filename);
3532 req->flags |= REQ_F_NEED_CLEANUP;
3536 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3538 struct io_unlink *un = &req->unlink;
3541 if (issue_flags & IO_URING_F_NONBLOCK)
3544 if (un->flags & AT_REMOVEDIR)
3545 ret = do_rmdir(un->dfd, un->filename);
3547 ret = do_unlinkat(un->dfd, un->filename);
3549 req->flags &= ~REQ_F_NEED_CLEANUP;
3551 req_set_fail_links(req);
3552 io_req_complete(req, ret);
3556 static int io_shutdown_prep(struct io_kiocb *req,
3557 const struct io_uring_sqe *sqe)
3559 #if defined(CONFIG_NET)
3560 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3562 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3566 req->shutdown.how = READ_ONCE(sqe->len);
3573 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3575 #if defined(CONFIG_NET)
3576 struct socket *sock;
3579 if (issue_flags & IO_URING_F_NONBLOCK)
3582 sock = sock_from_file(req->file);
3583 if (unlikely(!sock))
3586 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3588 req_set_fail_links(req);
3589 io_req_complete(req, ret);
3596 static int __io_splice_prep(struct io_kiocb *req,
3597 const struct io_uring_sqe *sqe)
3599 struct io_splice* sp = &req->splice;
3600 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3602 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3606 sp->len = READ_ONCE(sqe->len);
3607 sp->flags = READ_ONCE(sqe->splice_flags);
3609 if (unlikely(sp->flags & ~valid_flags))
3612 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3613 (sp->flags & SPLICE_F_FD_IN_FIXED));
3616 req->flags |= REQ_F_NEED_CLEANUP;
3618 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3620 * Splice operation will be punted aync, and here need to
3621 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3623 req->work.flags |= IO_WQ_WORK_UNBOUND;
3629 static int io_tee_prep(struct io_kiocb *req,
3630 const struct io_uring_sqe *sqe)
3632 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3634 return __io_splice_prep(req, sqe);
3637 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3639 struct io_splice *sp = &req->splice;
3640 struct file *in = sp->file_in;
3641 struct file *out = sp->file_out;
3642 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3645 if (issue_flags & IO_URING_F_NONBLOCK)
3648 ret = do_tee(in, out, sp->len, flags);
3650 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3651 req->flags &= ~REQ_F_NEED_CLEANUP;
3654 req_set_fail_links(req);
3655 io_req_complete(req, ret);
3659 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3661 struct io_splice* sp = &req->splice;
3663 sp->off_in = READ_ONCE(sqe->splice_off_in);
3664 sp->off_out = READ_ONCE(sqe->off);
3665 return __io_splice_prep(req, sqe);
3668 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3670 struct io_splice *sp = &req->splice;
3671 struct file *in = sp->file_in;
3672 struct file *out = sp->file_out;
3673 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3674 loff_t *poff_in, *poff_out;
3677 if (issue_flags & IO_URING_F_NONBLOCK)
3680 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3681 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3684 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3686 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3687 req->flags &= ~REQ_F_NEED_CLEANUP;
3690 req_set_fail_links(req);
3691 io_req_complete(req, ret);
3696 * IORING_OP_NOP just posts a completion event, nothing else.
3698 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3700 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3705 __io_req_complete(req, issue_flags, 0, 0);
3709 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3711 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3718 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3721 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3722 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3725 req->sync.off = READ_ONCE(sqe->off);
3726 req->sync.len = READ_ONCE(sqe->len);
3730 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3732 loff_t end = req->sync.off + req->sync.len;
3735 /* fsync always requires a blocking context */
3736 if (issue_flags & IO_URING_F_NONBLOCK)
3739 ret = vfs_fsync_range(req->file, req->sync.off,
3740 end > 0 ? end : LLONG_MAX,
3741 req->sync.flags & IORING_FSYNC_DATASYNC);
3743 req_set_fail_links(req);
3744 io_req_complete(req, ret);
3748 static int io_fallocate_prep(struct io_kiocb *req,
3749 const struct io_uring_sqe *sqe)
3751 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 req->sync.off = READ_ONCE(sqe->off);
3757 req->sync.len = READ_ONCE(sqe->addr);
3758 req->sync.mode = READ_ONCE(sqe->len);
3762 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3766 /* fallocate always requiring blocking context */
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3769 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3772 req_set_fail_links(req);
3773 io_req_complete(req, ret);
3777 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 const char __user *fname;
3782 if (unlikely(sqe->ioprio || sqe->buf_index))
3784 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3787 /* open.how should be already initialised */
3788 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3789 req->open.how.flags |= O_LARGEFILE;
3791 req->open.dfd = READ_ONCE(sqe->fd);
3792 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3793 req->open.filename = getname(fname);
3794 if (IS_ERR(req->open.filename)) {
3795 ret = PTR_ERR(req->open.filename);
3796 req->open.filename = NULL;
3799 req->open.nofile = rlimit(RLIMIT_NOFILE);
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3810 mode = READ_ONCE(sqe->len);
3811 flags = READ_ONCE(sqe->open_flags);
3812 req->open.how = build_open_how(flags, mode);
3813 return __io_openat_prep(req, sqe);
3816 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3818 struct open_how __user *how;
3822 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3824 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3825 len = READ_ONCE(sqe->len);
3826 if (len < OPEN_HOW_SIZE_VER0)
3829 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3834 return __io_openat_prep(req, sqe);
3837 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3839 struct open_flags op;
3842 bool resolve_nonblock;
3845 ret = build_open_flags(&req->open.how, &op);
3848 nonblock_set = op.open_flag & O_NONBLOCK;
3849 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3850 if (issue_flags & IO_URING_F_NONBLOCK) {
3852 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3853 * it'll always -EAGAIN
3855 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3857 op.lookup_flags |= LOOKUP_CACHED;
3858 op.open_flag |= O_NONBLOCK;
3861 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3865 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3866 /* only retry if RESOLVE_CACHED wasn't already set by application */
3867 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3868 file == ERR_PTR(-EAGAIN)) {
3870 * We could hang on to this 'fd', but seems like marginal
3871 * gain for something that is now known to be a slower path.
3872 * So just put it, and we'll get a new one when we retry.
3880 ret = PTR_ERR(file);
3882 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3883 file->f_flags &= ~O_NONBLOCK;
3884 fsnotify_open(file);
3885 fd_install(ret, file);
3888 putname(req->open.filename);
3889 req->flags &= ~REQ_F_NEED_CLEANUP;
3891 req_set_fail_links(req);
3892 io_req_complete(req, ret);
3896 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3898 return io_openat2(req, issue_flags);
3901 static int io_remove_buffers_prep(struct io_kiocb *req,
3902 const struct io_uring_sqe *sqe)
3904 struct io_provide_buf *p = &req->pbuf;
3907 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3910 tmp = READ_ONCE(sqe->fd);
3911 if (!tmp || tmp > USHRT_MAX)
3914 memset(p, 0, sizeof(*p));
3916 p->bgid = READ_ONCE(sqe->buf_group);
3920 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3921 int bgid, unsigned nbufs)
3925 /* shouldn't happen */
3929 /* the head kbuf is the list itself */
3930 while (!list_empty(&buf->list)) {
3931 struct io_buffer *nxt;
3933 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3934 list_del(&nxt->list);
3941 xa_erase(&ctx->io_buffers, bgid);
3946 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3948 struct io_provide_buf *p = &req->pbuf;
3949 struct io_ring_ctx *ctx = req->ctx;
3950 struct io_buffer *head;
3952 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3954 io_ring_submit_lock(ctx, !force_nonblock);
3956 lockdep_assert_held(&ctx->uring_lock);
3959 head = xa_load(&ctx->io_buffers, p->bgid);
3961 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3963 req_set_fail_links(req);
3965 /* need to hold the lock to complete IOPOLL requests */
3966 if (ctx->flags & IORING_SETUP_IOPOLL) {
3967 __io_req_complete(req, issue_flags, ret, 0);
3968 io_ring_submit_unlock(ctx, !force_nonblock);
3970 io_ring_submit_unlock(ctx, !force_nonblock);
3971 __io_req_complete(req, issue_flags, ret, 0);
3976 static int io_provide_buffers_prep(struct io_kiocb *req,
3977 const struct io_uring_sqe *sqe)
3980 struct io_provide_buf *p = &req->pbuf;
3983 if (sqe->ioprio || sqe->rw_flags)
3986 tmp = READ_ONCE(sqe->fd);
3987 if (!tmp || tmp > USHRT_MAX)
3990 p->addr = READ_ONCE(sqe->addr);
3991 p->len = READ_ONCE(sqe->len);
3993 size = (unsigned long)p->len * p->nbufs;
3994 if (!access_ok(u64_to_user_ptr(p->addr), size))
3997 p->bgid = READ_ONCE(sqe->buf_group);
3998 tmp = READ_ONCE(sqe->off);
3999 if (tmp > USHRT_MAX)
4005 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4007 struct io_buffer *buf;
4008 u64 addr = pbuf->addr;
4009 int i, bid = pbuf->bid;
4011 for (i = 0; i < pbuf->nbufs; i++) {
4012 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4017 buf->len = pbuf->len;
4022 INIT_LIST_HEAD(&buf->list);
4025 list_add_tail(&buf->list, &(*head)->list);
4029 return i ? i : -ENOMEM;
4032 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4034 struct io_provide_buf *p = &req->pbuf;
4035 struct io_ring_ctx *ctx = req->ctx;
4036 struct io_buffer *head, *list;
4038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4040 io_ring_submit_lock(ctx, !force_nonblock);
4042 lockdep_assert_held(&ctx->uring_lock);
4044 list = head = xa_load(&ctx->io_buffers, p->bgid);
4046 ret = io_add_buffers(p, &head);
4047 if (ret >= 0 && !list) {
4048 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4050 __io_remove_buffers(ctx, head, p->bgid, -1U);
4053 req_set_fail_links(req);
4055 /* need to hold the lock to complete IOPOLL requests */
4056 if (ctx->flags & IORING_SETUP_IOPOLL) {
4057 __io_req_complete(req, issue_flags, ret, 0);
4058 io_ring_submit_unlock(ctx, !force_nonblock);
4060 io_ring_submit_unlock(ctx, !force_nonblock);
4061 __io_req_complete(req, issue_flags, ret, 0);
4066 static int io_epoll_ctl_prep(struct io_kiocb *req,
4067 const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_EPOLL)
4070 if (sqe->ioprio || sqe->buf_index)
4072 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4075 req->epoll.epfd = READ_ONCE(sqe->fd);
4076 req->epoll.op = READ_ONCE(sqe->len);
4077 req->epoll.fd = READ_ONCE(sqe->off);
4079 if (ep_op_has_event(req->epoll.op)) {
4080 struct epoll_event __user *ev;
4082 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_EPOLL)
4096 struct io_epoll *ie = &req->epoll;
4098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4100 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101 if (force_nonblock && ret == -EAGAIN)
4105 req_set_fail_links(req);
4106 __io_req_complete(req, issue_flags, ret, 0);
4113 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 if (sqe->ioprio || sqe->buf_index || sqe->off)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->madvise.addr = READ_ONCE(sqe->addr);
4122 req->madvise.len = READ_ONCE(sqe->len);
4123 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4130 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133 struct io_madvise *ma = &req->madvise;
4136 if (issue_flags & IO_URING_F_NONBLOCK)
4139 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4141 req_set_fail_links(req);
4142 io_req_complete(req, ret);
4149 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 req->fadvise.offset = READ_ONCE(sqe->off);
4157 req->fadvise.len = READ_ONCE(sqe->len);
4158 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4162 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_fadvise *fa = &req->fadvise;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 switch (fa->advice) {
4169 case POSIX_FADV_NORMAL:
4170 case POSIX_FADV_RANDOM:
4171 case POSIX_FADV_SEQUENTIAL:
4178 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4180 req_set_fail_links(req);
4181 io_req_complete(req, ret);
4185 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4187 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4189 if (sqe->ioprio || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->statx.dfd = READ_ONCE(sqe->fd);
4195 req->statx.mask = READ_ONCE(sqe->len);
4196 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198 req->statx.flags = READ_ONCE(sqe->statx_flags);
4203 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4205 struct io_statx *ctx = &req->statx;
4208 if (issue_flags & IO_URING_F_NONBLOCK) {
4209 /* only need file table for an actual valid fd */
4210 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4211 req->flags |= REQ_F_NO_FILE_TABLE;
4215 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4219 req_set_fail_links(req);
4220 io_req_complete(req, ret);
4224 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4226 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4228 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4229 sqe->rw_flags || sqe->buf_index)
4231 if (req->flags & REQ_F_FIXED_FILE)
4234 req->close.fd = READ_ONCE(sqe->fd);
4238 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4240 struct files_struct *files = current->files;
4241 struct io_close *close = &req->close;
4242 struct fdtable *fdt;
4248 spin_lock(&files->file_lock);
4249 fdt = files_fdtable(files);
4250 if (close->fd >= fdt->max_fds) {
4251 spin_unlock(&files->file_lock);
4254 file = fdt->fd[close->fd];
4256 spin_unlock(&files->file_lock);
4260 if (file->f_op == &io_uring_fops) {
4261 spin_unlock(&files->file_lock);
4266 /* if the file has a flush method, be safe and punt to async */
4267 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4268 spin_unlock(&files->file_lock);
4272 ret = __close_fd_get_file(close->fd, &file);
4273 spin_unlock(&files->file_lock);
4280 /* No ->flush() or already async, safely close from here */
4281 ret = filp_close(file, current->files);
4284 req_set_fail_links(req);
4287 __io_req_complete(req, issue_flags, ret, 0);
4291 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4293 struct io_ring_ctx *ctx = req->ctx;
4295 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4297 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4300 req->sync.off = READ_ONCE(sqe->off);
4301 req->sync.len = READ_ONCE(sqe->len);
4302 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4306 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4310 /* sync_file_range always requires a blocking context */
4311 if (issue_flags & IO_URING_F_NONBLOCK)
4314 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4317 req_set_fail_links(req);
4318 io_req_complete(req, ret);
4322 #if defined(CONFIG_NET)
4323 static int io_setup_async_msg(struct io_kiocb *req,
4324 struct io_async_msghdr *kmsg)
4326 struct io_async_msghdr *async_msg = req->async_data;
4330 if (io_alloc_async_data(req)) {
4331 kfree(kmsg->free_iov);
4334 async_msg = req->async_data;
4335 req->flags |= REQ_F_NEED_CLEANUP;
4336 memcpy(async_msg, kmsg, sizeof(*kmsg));
4337 async_msg->msg.msg_name = &async_msg->addr;
4338 /* if were using fast_iov, set it to the new one */
4339 if (!async_msg->free_iov)
4340 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4345 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4346 struct io_async_msghdr *iomsg)
4348 iomsg->msg.msg_name = &iomsg->addr;
4349 iomsg->free_iov = iomsg->fast_iov;
4350 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4351 req->sr_msg.msg_flags, &iomsg->free_iov);
4354 static int io_sendmsg_prep_async(struct io_kiocb *req)
4358 if (!io_op_defs[req->opcode].needs_async_data)
4360 ret = io_sendmsg_copy_hdr(req, req->async_data);
4362 req->flags |= REQ_F_NEED_CLEANUP;
4366 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4368 struct io_sr_msg *sr = &req->sr_msg;
4370 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4373 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4374 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4375 sr->len = READ_ONCE(sqe->len);
4377 #ifdef CONFIG_COMPAT
4378 if (req->ctx->compat)
4379 sr->msg_flags |= MSG_CMSG_COMPAT;
4384 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4386 struct io_async_msghdr iomsg, *kmsg;
4387 struct socket *sock;
4392 sock = sock_from_file(req->file);
4393 if (unlikely(!sock))
4396 kmsg = req->async_data;
4398 ret = io_sendmsg_copy_hdr(req, &iomsg);
4404 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4405 if (flags & MSG_DONTWAIT)
4406 req->flags |= REQ_F_NOWAIT;
4407 else if (issue_flags & IO_URING_F_NONBLOCK)
4408 flags |= MSG_DONTWAIT;
4410 if (flags & MSG_WAITALL)
4411 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4413 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4414 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4415 return io_setup_async_msg(req, kmsg);
4416 if (ret == -ERESTARTSYS)
4419 /* fast path, check for non-NULL to avoid function call */
4421 kfree(kmsg->free_iov);
4422 req->flags &= ~REQ_F_NEED_CLEANUP;
4424 req_set_fail_links(req);
4425 __io_req_complete(req, issue_flags, ret, 0);
4429 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4431 struct io_sr_msg *sr = &req->sr_msg;
4434 struct socket *sock;
4439 sock = sock_from_file(req->file);
4440 if (unlikely(!sock))
4443 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4447 msg.msg_name = NULL;
4448 msg.msg_control = NULL;
4449 msg.msg_controllen = 0;
4450 msg.msg_namelen = 0;
4452 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4453 if (flags & MSG_DONTWAIT)
4454 req->flags |= REQ_F_NOWAIT;
4455 else if (issue_flags & IO_URING_F_NONBLOCK)
4456 flags |= MSG_DONTWAIT;
4458 if (flags & MSG_WAITALL)
4459 min_ret = iov_iter_count(&msg.msg_iter);
4461 msg.msg_flags = flags;
4462 ret = sock_sendmsg(sock, &msg);
4463 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4465 if (ret == -ERESTARTSYS)
4469 req_set_fail_links(req);
4470 __io_req_complete(req, issue_flags, ret, 0);
4474 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4475 struct io_async_msghdr *iomsg)
4477 struct io_sr_msg *sr = &req->sr_msg;
4478 struct iovec __user *uiov;
4482 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4483 &iomsg->uaddr, &uiov, &iov_len);
4487 if (req->flags & REQ_F_BUFFER_SELECT) {
4490 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4492 sr->len = iomsg->fast_iov[0].iov_len;
4493 iomsg->free_iov = NULL;
4495 iomsg->free_iov = iomsg->fast_iov;
4496 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4497 &iomsg->free_iov, &iomsg->msg.msg_iter,
4506 #ifdef CONFIG_COMPAT
4507 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4508 struct io_async_msghdr *iomsg)
4510 struct compat_msghdr __user *msg_compat;
4511 struct io_sr_msg *sr = &req->sr_msg;
4512 struct compat_iovec __user *uiov;
4517 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4518 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4523 uiov = compat_ptr(ptr);
4524 if (req->flags & REQ_F_BUFFER_SELECT) {
4525 compat_ssize_t clen;
4529 if (!access_ok(uiov, sizeof(*uiov)))
4531 if (__get_user(clen, &uiov->iov_len))
4536 iomsg->free_iov = NULL;
4538 iomsg->free_iov = iomsg->fast_iov;
4539 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4540 UIO_FASTIOV, &iomsg->free_iov,
4541 &iomsg->msg.msg_iter, true);
4550 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4551 struct io_async_msghdr *iomsg)
4553 iomsg->msg.msg_name = &iomsg->addr;
4555 #ifdef CONFIG_COMPAT
4556 if (req->ctx->compat)
4557 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4560 return __io_recvmsg_copy_hdr(req, iomsg);
4563 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4566 struct io_sr_msg *sr = &req->sr_msg;
4567 struct io_buffer *kbuf;
4569 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4574 req->flags |= REQ_F_BUFFER_SELECTED;
4578 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4580 return io_put_kbuf(req, req->sr_msg.kbuf);
4583 static int io_recvmsg_prep_async(struct io_kiocb *req)
4587 if (!io_op_defs[req->opcode].needs_async_data)
4589 ret = io_recvmsg_copy_hdr(req, req->async_data);
4591 req->flags |= REQ_F_NEED_CLEANUP;
4595 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4597 struct io_sr_msg *sr = &req->sr_msg;
4599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4602 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4603 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4604 sr->len = READ_ONCE(sqe->len);
4605 sr->bgid = READ_ONCE(sqe->buf_group);
4607 #ifdef CONFIG_COMPAT
4608 if (req->ctx->compat)
4609 sr->msg_flags |= MSG_CMSG_COMPAT;
4614 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4616 struct io_async_msghdr iomsg, *kmsg;
4617 struct socket *sock;
4618 struct io_buffer *kbuf;
4621 int ret, cflags = 0;
4622 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4624 sock = sock_from_file(req->file);
4625 if (unlikely(!sock))
4628 kmsg = req->async_data;
4630 ret = io_recvmsg_copy_hdr(req, &iomsg);
4636 if (req->flags & REQ_F_BUFFER_SELECT) {
4637 kbuf = io_recv_buffer_select(req, !force_nonblock);
4639 return PTR_ERR(kbuf);
4640 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4641 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4642 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4643 1, req->sr_msg.len);
4646 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4647 if (flags & MSG_DONTWAIT)
4648 req->flags |= REQ_F_NOWAIT;
4649 else if (force_nonblock)
4650 flags |= MSG_DONTWAIT;
4652 if (flags & MSG_WAITALL)
4653 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4655 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4656 kmsg->uaddr, flags);
4657 if (force_nonblock && ret == -EAGAIN)
4658 return io_setup_async_msg(req, kmsg);
4659 if (ret == -ERESTARTSYS)
4662 if (req->flags & REQ_F_BUFFER_SELECTED)
4663 cflags = io_put_recv_kbuf(req);
4664 /* fast path, check for non-NULL to avoid function call */
4666 kfree(kmsg->free_iov);
4667 req->flags &= ~REQ_F_NEED_CLEANUP;
4668 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4669 req_set_fail_links(req);
4670 __io_req_complete(req, issue_flags, ret, cflags);
4674 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4676 struct io_buffer *kbuf;
4677 struct io_sr_msg *sr = &req->sr_msg;
4679 void __user *buf = sr->buf;
4680 struct socket *sock;
4684 int ret, cflags = 0;
4685 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4687 sock = sock_from_file(req->file);
4688 if (unlikely(!sock))
4691 if (req->flags & REQ_F_BUFFER_SELECT) {
4692 kbuf = io_recv_buffer_select(req, !force_nonblock);
4694 return PTR_ERR(kbuf);
4695 buf = u64_to_user_ptr(kbuf->addr);
4698 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4702 msg.msg_name = NULL;
4703 msg.msg_control = NULL;
4704 msg.msg_controllen = 0;
4705 msg.msg_namelen = 0;
4706 msg.msg_iocb = NULL;
4709 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4710 if (flags & MSG_DONTWAIT)
4711 req->flags |= REQ_F_NOWAIT;
4712 else if (force_nonblock)
4713 flags |= MSG_DONTWAIT;
4715 if (flags & MSG_WAITALL)
4716 min_ret = iov_iter_count(&msg.msg_iter);
4718 ret = sock_recvmsg(sock, &msg, flags);
4719 if (force_nonblock && ret == -EAGAIN)
4721 if (ret == -ERESTARTSYS)
4724 if (req->flags & REQ_F_BUFFER_SELECTED)
4725 cflags = io_put_recv_kbuf(req);
4726 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4727 req_set_fail_links(req);
4728 __io_req_complete(req, issue_flags, ret, cflags);
4732 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4734 struct io_accept *accept = &req->accept;
4736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4738 if (sqe->ioprio || sqe->len || sqe->buf_index)
4741 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4742 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4743 accept->flags = READ_ONCE(sqe->accept_flags);
4744 accept->nofile = rlimit(RLIMIT_NOFILE);
4748 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4750 struct io_accept *accept = &req->accept;
4751 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4752 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4755 if (req->file->f_flags & O_NONBLOCK)
4756 req->flags |= REQ_F_NOWAIT;
4758 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4759 accept->addr_len, accept->flags,
4761 if (ret == -EAGAIN && force_nonblock)
4764 if (ret == -ERESTARTSYS)
4766 req_set_fail_links(req);
4768 __io_req_complete(req, issue_flags, ret, 0);
4772 static int io_connect_prep_async(struct io_kiocb *req)
4774 struct io_async_connect *io = req->async_data;
4775 struct io_connect *conn = &req->connect;
4777 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4780 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4782 struct io_connect *conn = &req->connect;
4784 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4786 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4789 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4790 conn->addr_len = READ_ONCE(sqe->addr2);
4794 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4796 struct io_async_connect __io, *io;
4797 unsigned file_flags;
4799 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4801 if (req->async_data) {
4802 io = req->async_data;
4804 ret = move_addr_to_kernel(req->connect.addr,
4805 req->connect.addr_len,
4812 file_flags = force_nonblock ? O_NONBLOCK : 0;
4814 ret = __sys_connect_file(req->file, &io->address,
4815 req->connect.addr_len, file_flags);
4816 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4817 if (req->async_data)
4819 if (io_alloc_async_data(req)) {
4823 io = req->async_data;
4824 memcpy(req->async_data, &__io, sizeof(__io));
4827 if (ret == -ERESTARTSYS)
4831 req_set_fail_links(req);
4832 __io_req_complete(req, issue_flags, ret, 0);
4835 #else /* !CONFIG_NET */
4836 #define IO_NETOP_FN(op) \
4837 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4839 return -EOPNOTSUPP; \
4842 #define IO_NETOP_PREP(op) \
4844 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4846 return -EOPNOTSUPP; \
4849 #define IO_NETOP_PREP_ASYNC(op) \
4851 static int io_##op##_prep_async(struct io_kiocb *req) \
4853 return -EOPNOTSUPP; \
4856 IO_NETOP_PREP_ASYNC(sendmsg);
4857 IO_NETOP_PREP_ASYNC(recvmsg);
4858 IO_NETOP_PREP_ASYNC(connect);
4859 IO_NETOP_PREP(accept);
4862 #endif /* CONFIG_NET */
4864 struct io_poll_table {
4865 struct poll_table_struct pt;
4866 struct io_kiocb *req;
4870 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4871 __poll_t mask, task_work_func_t func)
4875 /* for instances that support it check for an event match first: */
4876 if (mask && !(mask & poll->events))
4879 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4881 list_del_init(&poll->wait.entry);
4884 req->task_work.func = func;
4885 percpu_ref_get(&req->ctx->refs);
4888 * If this fails, then the task is exiting. When a task exits, the
4889 * work gets canceled, so just cancel this request as well instead
4890 * of executing it. We can't safely execute it anyway, as we may not
4891 * have the needed state needed for it anyway.
4893 ret = io_req_task_work_add(req);
4894 if (unlikely(ret)) {
4895 WRITE_ONCE(poll->canceled, true);
4896 io_req_task_work_add_fallback(req, func);
4901 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4902 __acquires(&req->ctx->completion_lock)
4904 struct io_ring_ctx *ctx = req->ctx;
4906 if (!req->result && !READ_ONCE(poll->canceled)) {
4907 struct poll_table_struct pt = { ._key = poll->events };
4909 req->result = vfs_poll(req->file, &pt) & poll->events;
4912 spin_lock_irq(&ctx->completion_lock);
4913 if (!req->result && !READ_ONCE(poll->canceled)) {
4914 add_wait_queue(poll->head, &poll->wait);
4921 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4923 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4924 if (req->opcode == IORING_OP_POLL_ADD)
4925 return req->async_data;
4926 return req->apoll->double_poll;
4929 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4931 if (req->opcode == IORING_OP_POLL_ADD)
4933 return &req->apoll->poll;
4936 static void io_poll_remove_double(struct io_kiocb *req)
4938 struct io_poll_iocb *poll = io_poll_get_double(req);
4940 lockdep_assert_held(&req->ctx->completion_lock);
4942 if (poll && poll->head) {
4943 struct wait_queue_head *head = poll->head;
4945 spin_lock(&head->lock);
4946 list_del_init(&poll->wait.entry);
4947 if (poll->wait.private)
4948 refcount_dec(&req->refs);
4950 spin_unlock(&head->lock);
4954 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4956 struct io_ring_ctx *ctx = req->ctx;
4958 io_poll_remove_double(req);
4959 req->poll.done = true;
4960 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4961 io_commit_cqring(ctx);
4964 static void io_poll_task_func(struct callback_head *cb)
4966 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4967 struct io_ring_ctx *ctx = req->ctx;
4968 struct io_kiocb *nxt;
4970 if (io_poll_rewait(req, &req->poll)) {
4971 spin_unlock_irq(&ctx->completion_lock);
4973 hash_del(&req->hash_node);
4974 io_poll_complete(req, req->result, 0);
4975 spin_unlock_irq(&ctx->completion_lock);
4977 nxt = io_put_req_find_next(req);
4978 io_cqring_ev_posted(ctx);
4980 __io_req_task_submit(nxt);
4983 percpu_ref_put(&ctx->refs);
4986 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4987 int sync, void *key)
4989 struct io_kiocb *req = wait->private;
4990 struct io_poll_iocb *poll = io_poll_get_single(req);
4991 __poll_t mask = key_to_poll(key);
4993 /* for instances that support it check for an event match first: */
4994 if (mask && !(mask & poll->events))
4997 list_del_init(&wait->entry);
4999 if (poll && poll->head) {
5002 spin_lock(&poll->head->lock);
5003 done = list_empty(&poll->wait.entry);
5005 list_del_init(&poll->wait.entry);
5006 /* make sure double remove sees this as being gone */
5007 wait->private = NULL;
5008 spin_unlock(&poll->head->lock);
5010 /* use wait func handler, so it matches the rq type */
5011 poll->wait.func(&poll->wait, mode, sync, key);
5014 refcount_dec(&req->refs);
5018 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5019 wait_queue_func_t wake_func)
5023 poll->canceled = false;
5024 poll->events = events;
5025 INIT_LIST_HEAD(&poll->wait.entry);
5026 init_waitqueue_func_entry(&poll->wait, wake_func);
5029 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5030 struct wait_queue_head *head,
5031 struct io_poll_iocb **poll_ptr)
5033 struct io_kiocb *req = pt->req;
5036 * If poll->head is already set, it's because the file being polled
5037 * uses multiple waitqueues for poll handling (eg one for read, one
5038 * for write). Setup a separate io_poll_iocb if this happens.
5040 if (unlikely(poll->head)) {
5041 struct io_poll_iocb *poll_one = poll;
5043 /* already have a 2nd entry, fail a third attempt */
5045 pt->error = -EINVAL;
5048 /* double add on the same waitqueue head, ignore */
5049 if (poll->head == head)
5051 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5053 pt->error = -ENOMEM;
5056 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5057 refcount_inc(&req->refs);
5058 poll->wait.private = req;
5065 if (poll->events & EPOLLEXCLUSIVE)
5066 add_wait_queue_exclusive(head, &poll->wait);
5068 add_wait_queue(head, &poll->wait);
5071 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5072 struct poll_table_struct *p)
5074 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5075 struct async_poll *apoll = pt->req->apoll;
5077 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5080 static void io_async_task_func(struct callback_head *cb)
5082 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5083 struct async_poll *apoll = req->apoll;
5084 struct io_ring_ctx *ctx = req->ctx;
5086 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5088 if (io_poll_rewait(req, &apoll->poll)) {
5089 spin_unlock_irq(&ctx->completion_lock);
5090 percpu_ref_put(&ctx->refs);
5094 /* If req is still hashed, it cannot have been canceled. Don't check. */
5095 if (hash_hashed(&req->hash_node))
5096 hash_del(&req->hash_node);
5098 io_poll_remove_double(req);
5099 spin_unlock_irq(&ctx->completion_lock);
5101 if (!READ_ONCE(apoll->poll.canceled))
5102 __io_req_task_submit(req);
5104 __io_req_task_cancel(req, -ECANCELED);
5106 percpu_ref_put(&ctx->refs);
5107 kfree(apoll->double_poll);
5111 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5114 struct io_kiocb *req = wait->private;
5115 struct io_poll_iocb *poll = &req->apoll->poll;
5117 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5120 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5123 static void io_poll_req_insert(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 struct hlist_head *list;
5128 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5129 hlist_add_head(&req->hash_node, list);
5132 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5133 struct io_poll_iocb *poll,
5134 struct io_poll_table *ipt, __poll_t mask,
5135 wait_queue_func_t wake_func)
5136 __acquires(&ctx->completion_lock)
5138 struct io_ring_ctx *ctx = req->ctx;
5139 bool cancel = false;
5141 INIT_HLIST_NODE(&req->hash_node);
5142 io_init_poll_iocb(poll, mask, wake_func);
5143 poll->file = req->file;
5144 poll->wait.private = req;
5146 ipt->pt._key = mask;
5148 ipt->error = -EINVAL;
5150 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5152 spin_lock_irq(&ctx->completion_lock);
5153 if (likely(poll->head)) {
5154 spin_lock(&poll->head->lock);
5155 if (unlikely(list_empty(&poll->wait.entry))) {
5161 if (mask || ipt->error)
5162 list_del_init(&poll->wait.entry);
5164 WRITE_ONCE(poll->canceled, true);
5165 else if (!poll->done) /* actually waiting for an event */
5166 io_poll_req_insert(req);
5167 spin_unlock(&poll->head->lock);
5173 static bool io_arm_poll_handler(struct io_kiocb *req)
5175 const struct io_op_def *def = &io_op_defs[req->opcode];
5176 struct io_ring_ctx *ctx = req->ctx;
5177 struct async_poll *apoll;
5178 struct io_poll_table ipt;
5182 if (!req->file || !file_can_poll(req->file))
5184 if (req->flags & REQ_F_POLLED)
5188 else if (def->pollout)
5192 /* if we can't nonblock try, then no point in arming a poll handler */
5193 if (!io_file_supports_async(req->file, rw))
5196 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5197 if (unlikely(!apoll))
5199 apoll->double_poll = NULL;
5201 req->flags |= REQ_F_POLLED;
5206 mask |= POLLIN | POLLRDNORM;
5208 mask |= POLLOUT | POLLWRNORM;
5210 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211 if ((req->opcode == IORING_OP_RECVMSG) &&
5212 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5215 mask |= POLLERR | POLLPRI;
5217 ipt.pt._qproc = io_async_queue_proc;
5219 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5221 if (ret || ipt.error) {
5222 io_poll_remove_double(req);
5223 spin_unlock_irq(&ctx->completion_lock);
5224 kfree(apoll->double_poll);
5228 spin_unlock_irq(&ctx->completion_lock);
5229 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5230 apoll->poll.events);
5234 static bool __io_poll_remove_one(struct io_kiocb *req,
5235 struct io_poll_iocb *poll)
5237 bool do_complete = false;
5239 spin_lock(&poll->head->lock);
5240 WRITE_ONCE(poll->canceled, true);
5241 if (!list_empty(&poll->wait.entry)) {
5242 list_del_init(&poll->wait.entry);
5245 spin_unlock(&poll->head->lock);
5246 hash_del(&req->hash_node);
5250 static bool io_poll_remove_one(struct io_kiocb *req)
5254 io_poll_remove_double(req);
5256 if (req->opcode == IORING_OP_POLL_ADD) {
5257 do_complete = __io_poll_remove_one(req, &req->poll);
5259 struct async_poll *apoll = req->apoll;
5261 /* non-poll requests have submit ref still */
5262 do_complete = __io_poll_remove_one(req, &apoll->poll);
5265 kfree(apoll->double_poll);
5271 io_cqring_fill_event(req, -ECANCELED);
5272 io_commit_cqring(req->ctx);
5273 req_set_fail_links(req);
5274 io_put_req_deferred(req, 1);
5281 * Returns true if we found and killed one or more poll requests
5283 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5284 struct files_struct *files)
5286 struct hlist_node *tmp;
5287 struct io_kiocb *req;
5290 spin_lock_irq(&ctx->completion_lock);
5291 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5292 struct hlist_head *list;
5294 list = &ctx->cancel_hash[i];
5295 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5296 if (io_match_task(req, tsk, files))
5297 posted += io_poll_remove_one(req);
5300 spin_unlock_irq(&ctx->completion_lock);
5303 io_cqring_ev_posted(ctx);
5308 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5310 struct hlist_head *list;
5311 struct io_kiocb *req;
5313 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5314 hlist_for_each_entry(req, list, hash_node) {
5315 if (sqe_addr != req->user_data)
5317 if (io_poll_remove_one(req))
5325 static int io_poll_remove_prep(struct io_kiocb *req,
5326 const struct io_uring_sqe *sqe)
5328 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5330 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5334 req->poll_remove.addr = READ_ONCE(sqe->addr);
5339 * Find a running poll command that matches one specified in sqe->addr,
5340 * and remove it if found.
5342 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5344 struct io_ring_ctx *ctx = req->ctx;
5347 spin_lock_irq(&ctx->completion_lock);
5348 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5349 spin_unlock_irq(&ctx->completion_lock);
5352 req_set_fail_links(req);
5353 io_req_complete(req, ret);
5357 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5360 struct io_kiocb *req = wait->private;
5361 struct io_poll_iocb *poll = &req->poll;
5363 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5366 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5367 struct poll_table_struct *p)
5369 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5371 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5374 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5376 struct io_poll_iocb *poll = &req->poll;
5379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5384 events = READ_ONCE(sqe->poll32_events);
5386 events = swahw32(events);
5388 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5389 (events & EPOLLEXCLUSIVE);
5393 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5395 struct io_poll_iocb *poll = &req->poll;
5396 struct io_ring_ctx *ctx = req->ctx;
5397 struct io_poll_table ipt;
5400 ipt.pt._qproc = io_poll_queue_proc;
5402 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5405 if (mask) { /* no async, we'd stolen it */
5407 io_poll_complete(req, mask, 0);
5409 spin_unlock_irq(&ctx->completion_lock);
5412 io_cqring_ev_posted(ctx);
5418 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5420 struct io_timeout_data *data = container_of(timer,
5421 struct io_timeout_data, timer);
5422 struct io_kiocb *req = data->req;
5423 struct io_ring_ctx *ctx = req->ctx;
5424 unsigned long flags;
5426 spin_lock_irqsave(&ctx->completion_lock, flags);
5427 list_del_init(&req->timeout.list);
5428 atomic_set(&req->ctx->cq_timeouts,
5429 atomic_read(&req->ctx->cq_timeouts) + 1);
5431 io_cqring_fill_event(req, -ETIME);
5432 io_commit_cqring(ctx);
5433 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5435 io_cqring_ev_posted(ctx);
5436 req_set_fail_links(req);
5438 return HRTIMER_NORESTART;
5441 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5444 struct io_timeout_data *io;
5445 struct io_kiocb *req;
5448 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5449 if (user_data == req->user_data) {
5456 return ERR_PTR(ret);
5458 io = req->async_data;
5459 ret = hrtimer_try_to_cancel(&io->timer);
5461 return ERR_PTR(-EALREADY);
5462 list_del_init(&req->timeout.list);
5466 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5468 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5471 return PTR_ERR(req);
5473 req_set_fail_links(req);
5474 io_cqring_fill_event(req, -ECANCELED);
5475 io_put_req_deferred(req, 1);
5479 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5480 struct timespec64 *ts, enum hrtimer_mode mode)
5482 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5483 struct io_timeout_data *data;
5486 return PTR_ERR(req);
5488 req->timeout.off = 0; /* noseq */
5489 data = req->async_data;
5490 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5491 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5492 data->timer.function = io_timeout_fn;
5493 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5497 static int io_timeout_remove_prep(struct io_kiocb *req,
5498 const struct io_uring_sqe *sqe)
5500 struct io_timeout_rem *tr = &req->timeout_rem;
5502 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5504 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5506 if (sqe->ioprio || sqe->buf_index || sqe->len)
5509 tr->addr = READ_ONCE(sqe->addr);
5510 tr->flags = READ_ONCE(sqe->timeout_flags);
5511 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5512 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5514 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5516 } else if (tr->flags) {
5517 /* timeout removal doesn't support flags */
5524 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5526 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5531 * Remove or update an existing timeout command
5533 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5535 struct io_timeout_rem *tr = &req->timeout_rem;
5536 struct io_ring_ctx *ctx = req->ctx;
5539 spin_lock_irq(&ctx->completion_lock);
5540 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5541 ret = io_timeout_cancel(ctx, tr->addr);
5543 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5544 io_translate_timeout_mode(tr->flags));
5546 io_cqring_fill_event(req, ret);
5547 io_commit_cqring(ctx);
5548 spin_unlock_irq(&ctx->completion_lock);
5549 io_cqring_ev_posted(ctx);
5551 req_set_fail_links(req);
5556 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5557 bool is_timeout_link)
5559 struct io_timeout_data *data;
5561 u32 off = READ_ONCE(sqe->off);
5563 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5565 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5567 if (off && is_timeout_link)
5569 flags = READ_ONCE(sqe->timeout_flags);
5570 if (flags & ~IORING_TIMEOUT_ABS)
5573 req->timeout.off = off;
5575 if (!req->async_data && io_alloc_async_data(req))
5578 data = req->async_data;
5581 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5584 data->mode = io_translate_timeout_mode(flags);
5585 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5586 io_req_track_inflight(req);
5590 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5592 struct io_ring_ctx *ctx = req->ctx;
5593 struct io_timeout_data *data = req->async_data;
5594 struct list_head *entry;
5595 u32 tail, off = req->timeout.off;
5597 spin_lock_irq(&ctx->completion_lock);
5600 * sqe->off holds how many events that need to occur for this
5601 * timeout event to be satisfied. If it isn't set, then this is
5602 * a pure timeout request, sequence isn't used.
5604 if (io_is_timeout_noseq(req)) {
5605 entry = ctx->timeout_list.prev;
5609 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5610 req->timeout.target_seq = tail + off;
5612 /* Update the last seq here in case io_flush_timeouts() hasn't.
5613 * This is safe because ->completion_lock is held, and submissions
5614 * and completions are never mixed in the same ->completion_lock section.
5616 ctx->cq_last_tm_flush = tail;
5619 * Insertion sort, ensuring the first entry in the list is always
5620 * the one we need first.
5622 list_for_each_prev(entry, &ctx->timeout_list) {
5623 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5626 if (io_is_timeout_noseq(nxt))
5628 /* nxt.seq is behind @tail, otherwise would've been completed */
5629 if (off >= nxt->timeout.target_seq - tail)
5633 list_add(&req->timeout.list, entry);
5634 data->timer.function = io_timeout_fn;
5635 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5636 spin_unlock_irq(&ctx->completion_lock);
5640 struct io_cancel_data {
5641 struct io_ring_ctx *ctx;
5645 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5647 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5648 struct io_cancel_data *cd = data;
5650 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5653 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5654 struct io_ring_ctx *ctx)
5656 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5657 enum io_wq_cancel cancel_ret;
5660 if (!tctx || !tctx->io_wq)
5663 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5664 switch (cancel_ret) {
5665 case IO_WQ_CANCEL_OK:
5668 case IO_WQ_CANCEL_RUNNING:
5671 case IO_WQ_CANCEL_NOTFOUND:
5679 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5680 struct io_kiocb *req, __u64 sqe_addr,
5683 unsigned long flags;
5686 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5687 if (ret != -ENOENT) {
5688 spin_lock_irqsave(&ctx->completion_lock, flags);
5692 spin_lock_irqsave(&ctx->completion_lock, flags);
5693 ret = io_timeout_cancel(ctx, sqe_addr);
5696 ret = io_poll_cancel(ctx, sqe_addr);
5700 io_cqring_fill_event(req, ret);
5701 io_commit_cqring(ctx);
5702 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5703 io_cqring_ev_posted(ctx);
5706 req_set_fail_links(req);
5710 static int io_async_cancel_prep(struct io_kiocb *req,
5711 const struct io_uring_sqe *sqe)
5713 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5715 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5717 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5720 req->cancel.addr = READ_ONCE(sqe->addr);
5724 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5726 struct io_ring_ctx *ctx = req->ctx;
5727 u64 sqe_addr = req->cancel.addr;
5728 struct io_tctx_node *node;
5731 /* tasks should wait for their io-wq threads, so safe w/o sync */
5732 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5733 spin_lock_irq(&ctx->completion_lock);
5736 ret = io_timeout_cancel(ctx, sqe_addr);
5739 ret = io_poll_cancel(ctx, sqe_addr);
5742 spin_unlock_irq(&ctx->completion_lock);
5744 /* slow path, try all io-wq's */
5745 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5747 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5748 struct io_uring_task *tctx = node->task->io_uring;
5750 if (!tctx || !tctx->io_wq)
5752 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5756 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5758 spin_lock_irq(&ctx->completion_lock);
5760 io_cqring_fill_event(req, ret);
5761 io_commit_cqring(ctx);
5762 spin_unlock_irq(&ctx->completion_lock);
5763 io_cqring_ev_posted(ctx);
5766 req_set_fail_links(req);
5771 static int io_rsrc_update_prep(struct io_kiocb *req,
5772 const struct io_uring_sqe *sqe)
5774 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5776 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5778 if (sqe->ioprio || sqe->rw_flags)
5781 req->rsrc_update.offset = READ_ONCE(sqe->off);
5782 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5783 if (!req->rsrc_update.nr_args)
5785 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5789 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5791 struct io_ring_ctx *ctx = req->ctx;
5792 struct io_uring_rsrc_update up;
5795 if (issue_flags & IO_URING_F_NONBLOCK)
5798 up.offset = req->rsrc_update.offset;
5799 up.data = req->rsrc_update.arg;
5801 mutex_lock(&ctx->uring_lock);
5802 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5803 mutex_unlock(&ctx->uring_lock);
5806 req_set_fail_links(req);
5807 __io_req_complete(req, issue_flags, ret, 0);
5811 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5813 switch (req->opcode) {
5816 case IORING_OP_READV:
5817 case IORING_OP_READ_FIXED:
5818 case IORING_OP_READ:
5819 return io_read_prep(req, sqe);
5820 case IORING_OP_WRITEV:
5821 case IORING_OP_WRITE_FIXED:
5822 case IORING_OP_WRITE:
5823 return io_write_prep(req, sqe);
5824 case IORING_OP_POLL_ADD:
5825 return io_poll_add_prep(req, sqe);
5826 case IORING_OP_POLL_REMOVE:
5827 return io_poll_remove_prep(req, sqe);
5828 case IORING_OP_FSYNC:
5829 return io_fsync_prep(req, sqe);
5830 case IORING_OP_SYNC_FILE_RANGE:
5831 return io_sfr_prep(req, sqe);
5832 case IORING_OP_SENDMSG:
5833 case IORING_OP_SEND:
5834 return io_sendmsg_prep(req, sqe);
5835 case IORING_OP_RECVMSG:
5836 case IORING_OP_RECV:
5837 return io_recvmsg_prep(req, sqe);
5838 case IORING_OP_CONNECT:
5839 return io_connect_prep(req, sqe);
5840 case IORING_OP_TIMEOUT:
5841 return io_timeout_prep(req, sqe, false);
5842 case IORING_OP_TIMEOUT_REMOVE:
5843 return io_timeout_remove_prep(req, sqe);
5844 case IORING_OP_ASYNC_CANCEL:
5845 return io_async_cancel_prep(req, sqe);
5846 case IORING_OP_LINK_TIMEOUT:
5847 return io_timeout_prep(req, sqe, true);
5848 case IORING_OP_ACCEPT:
5849 return io_accept_prep(req, sqe);
5850 case IORING_OP_FALLOCATE:
5851 return io_fallocate_prep(req, sqe);
5852 case IORING_OP_OPENAT:
5853 return io_openat_prep(req, sqe);
5854 case IORING_OP_CLOSE:
5855 return io_close_prep(req, sqe);
5856 case IORING_OP_FILES_UPDATE:
5857 return io_rsrc_update_prep(req, sqe);
5858 case IORING_OP_STATX:
5859 return io_statx_prep(req, sqe);
5860 case IORING_OP_FADVISE:
5861 return io_fadvise_prep(req, sqe);
5862 case IORING_OP_MADVISE:
5863 return io_madvise_prep(req, sqe);
5864 case IORING_OP_OPENAT2:
5865 return io_openat2_prep(req, sqe);
5866 case IORING_OP_EPOLL_CTL:
5867 return io_epoll_ctl_prep(req, sqe);
5868 case IORING_OP_SPLICE:
5869 return io_splice_prep(req, sqe);
5870 case IORING_OP_PROVIDE_BUFFERS:
5871 return io_provide_buffers_prep(req, sqe);
5872 case IORING_OP_REMOVE_BUFFERS:
5873 return io_remove_buffers_prep(req, sqe);
5875 return io_tee_prep(req, sqe);
5876 case IORING_OP_SHUTDOWN:
5877 return io_shutdown_prep(req, sqe);
5878 case IORING_OP_RENAMEAT:
5879 return io_renameat_prep(req, sqe);
5880 case IORING_OP_UNLINKAT:
5881 return io_unlinkat_prep(req, sqe);
5884 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5889 static int io_req_prep_async(struct io_kiocb *req)
5891 switch (req->opcode) {
5892 case IORING_OP_READV:
5893 case IORING_OP_READ_FIXED:
5894 case IORING_OP_READ:
5895 return io_rw_prep_async(req, READ);
5896 case IORING_OP_WRITEV:
5897 case IORING_OP_WRITE_FIXED:
5898 case IORING_OP_WRITE:
5899 return io_rw_prep_async(req, WRITE);
5900 case IORING_OP_SENDMSG:
5901 case IORING_OP_SEND:
5902 return io_sendmsg_prep_async(req);
5903 case IORING_OP_RECVMSG:
5904 case IORING_OP_RECV:
5905 return io_recvmsg_prep_async(req);
5906 case IORING_OP_CONNECT:
5907 return io_connect_prep_async(req);
5912 static int io_req_defer_prep(struct io_kiocb *req)
5914 if (!io_op_defs[req->opcode].needs_async_data)
5916 /* some opcodes init it during the inital prep */
5917 if (req->async_data)
5919 if (__io_alloc_async_data(req))
5921 return io_req_prep_async(req);
5924 static u32 io_get_sequence(struct io_kiocb *req)
5926 struct io_kiocb *pos;
5927 struct io_ring_ctx *ctx = req->ctx;
5928 u32 total_submitted, nr_reqs = 0;
5930 io_for_each_link(pos, req)
5933 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5934 return total_submitted - nr_reqs;
5937 static int io_req_defer(struct io_kiocb *req)
5939 struct io_ring_ctx *ctx = req->ctx;
5940 struct io_defer_entry *de;
5944 /* Still need defer if there is pending req in defer list. */
5945 if (likely(list_empty_careful(&ctx->defer_list) &&
5946 !(req->flags & REQ_F_IO_DRAIN)))
5949 seq = io_get_sequence(req);
5950 /* Still a chance to pass the sequence check */
5951 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5954 ret = io_req_defer_prep(req);
5957 io_prep_async_link(req);
5958 de = kmalloc(sizeof(*de), GFP_KERNEL);
5962 spin_lock_irq(&ctx->completion_lock);
5963 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5964 spin_unlock_irq(&ctx->completion_lock);
5966 io_queue_async_work(req);
5967 return -EIOCBQUEUED;
5970 trace_io_uring_defer(ctx, req, req->user_data);
5973 list_add_tail(&de->list, &ctx->defer_list);
5974 spin_unlock_irq(&ctx->completion_lock);
5975 return -EIOCBQUEUED;
5978 static void __io_clean_op(struct io_kiocb *req)
5980 if (req->flags & REQ_F_BUFFER_SELECTED) {
5981 switch (req->opcode) {
5982 case IORING_OP_READV:
5983 case IORING_OP_READ_FIXED:
5984 case IORING_OP_READ:
5985 kfree((void *)(unsigned long)req->rw.addr);
5987 case IORING_OP_RECVMSG:
5988 case IORING_OP_RECV:
5989 kfree(req->sr_msg.kbuf);
5992 req->flags &= ~REQ_F_BUFFER_SELECTED;
5995 if (req->flags & REQ_F_NEED_CLEANUP) {
5996 switch (req->opcode) {
5997 case IORING_OP_READV:
5998 case IORING_OP_READ_FIXED:
5999 case IORING_OP_READ:
6000 case IORING_OP_WRITEV:
6001 case IORING_OP_WRITE_FIXED:
6002 case IORING_OP_WRITE: {
6003 struct io_async_rw *io = req->async_data;
6005 kfree(io->free_iovec);
6008 case IORING_OP_RECVMSG:
6009 case IORING_OP_SENDMSG: {
6010 struct io_async_msghdr *io = req->async_data;
6012 kfree(io->free_iov);
6015 case IORING_OP_SPLICE:
6017 io_put_file(req, req->splice.file_in,
6018 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6020 case IORING_OP_OPENAT:
6021 case IORING_OP_OPENAT2:
6022 if (req->open.filename)
6023 putname(req->open.filename);
6025 case IORING_OP_RENAMEAT:
6026 putname(req->rename.oldpath);
6027 putname(req->rename.newpath);
6029 case IORING_OP_UNLINKAT:
6030 putname(req->unlink.filename);
6033 req->flags &= ~REQ_F_NEED_CLEANUP;
6037 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6039 struct io_ring_ctx *ctx = req->ctx;
6040 const struct cred *creds = NULL;
6043 if (req->work.creds && req->work.creds != current_cred())
6044 creds = override_creds(req->work.creds);
6046 switch (req->opcode) {
6048 ret = io_nop(req, issue_flags);
6050 case IORING_OP_READV:
6051 case IORING_OP_READ_FIXED:
6052 case IORING_OP_READ:
6053 ret = io_read(req, issue_flags);
6055 case IORING_OP_WRITEV:
6056 case IORING_OP_WRITE_FIXED:
6057 case IORING_OP_WRITE:
6058 ret = io_write(req, issue_flags);
6060 case IORING_OP_FSYNC:
6061 ret = io_fsync(req, issue_flags);
6063 case IORING_OP_POLL_ADD:
6064 ret = io_poll_add(req, issue_flags);
6066 case IORING_OP_POLL_REMOVE:
6067 ret = io_poll_remove(req, issue_flags);
6069 case IORING_OP_SYNC_FILE_RANGE:
6070 ret = io_sync_file_range(req, issue_flags);
6072 case IORING_OP_SENDMSG:
6073 ret = io_sendmsg(req, issue_flags);
6075 case IORING_OP_SEND:
6076 ret = io_send(req, issue_flags);
6078 case IORING_OP_RECVMSG:
6079 ret = io_recvmsg(req, issue_flags);
6081 case IORING_OP_RECV:
6082 ret = io_recv(req, issue_flags);
6084 case IORING_OP_TIMEOUT:
6085 ret = io_timeout(req, issue_flags);
6087 case IORING_OP_TIMEOUT_REMOVE:
6088 ret = io_timeout_remove(req, issue_flags);
6090 case IORING_OP_ACCEPT:
6091 ret = io_accept(req, issue_flags);
6093 case IORING_OP_CONNECT:
6094 ret = io_connect(req, issue_flags);
6096 case IORING_OP_ASYNC_CANCEL:
6097 ret = io_async_cancel(req, issue_flags);
6099 case IORING_OP_FALLOCATE:
6100 ret = io_fallocate(req, issue_flags);
6102 case IORING_OP_OPENAT:
6103 ret = io_openat(req, issue_flags);
6105 case IORING_OP_CLOSE:
6106 ret = io_close(req, issue_flags);
6108 case IORING_OP_FILES_UPDATE:
6109 ret = io_files_update(req, issue_flags);
6111 case IORING_OP_STATX:
6112 ret = io_statx(req, issue_flags);
6114 case IORING_OP_FADVISE:
6115 ret = io_fadvise(req, issue_flags);
6117 case IORING_OP_MADVISE:
6118 ret = io_madvise(req, issue_flags);
6120 case IORING_OP_OPENAT2:
6121 ret = io_openat2(req, issue_flags);
6123 case IORING_OP_EPOLL_CTL:
6124 ret = io_epoll_ctl(req, issue_flags);
6126 case IORING_OP_SPLICE:
6127 ret = io_splice(req, issue_flags);
6129 case IORING_OP_PROVIDE_BUFFERS:
6130 ret = io_provide_buffers(req, issue_flags);
6132 case IORING_OP_REMOVE_BUFFERS:
6133 ret = io_remove_buffers(req, issue_flags);
6136 ret = io_tee(req, issue_flags);
6138 case IORING_OP_SHUTDOWN:
6139 ret = io_shutdown(req, issue_flags);
6141 case IORING_OP_RENAMEAT:
6142 ret = io_renameat(req, issue_flags);
6144 case IORING_OP_UNLINKAT:
6145 ret = io_unlinkat(req, issue_flags);
6153 revert_creds(creds);
6158 /* If the op doesn't have a file, we're not polling for it */
6159 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6160 const bool in_async = io_wq_current_is_worker();
6162 /* workqueue context doesn't hold uring_lock, grab it now */
6164 mutex_lock(&ctx->uring_lock);
6166 io_iopoll_req_issued(req, in_async);
6169 mutex_unlock(&ctx->uring_lock);
6175 static void io_wq_submit_work(struct io_wq_work *work)
6177 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6178 struct io_kiocb *timeout;
6181 timeout = io_prep_linked_timeout(req);
6183 io_queue_linked_timeout(timeout);
6185 if (work->flags & IO_WQ_WORK_CANCEL)
6190 ret = io_issue_sqe(req, 0);
6192 * We can get EAGAIN for polled IO even though we're
6193 * forcing a sync submission from here, since we can't
6194 * wait for request slots on the block side.
6202 /* avoid locking problems by failing it from a clean context */
6204 /* io-wq is going to take one down */
6205 refcount_inc(&req->refs);
6206 io_req_task_queue_fail(req, ret);
6210 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6213 struct fixed_rsrc_table *table;
6215 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6216 return table->files[index & IORING_FILE_TABLE_MASK];
6219 static struct file *io_file_get(struct io_submit_state *state,
6220 struct io_kiocb *req, int fd, bool fixed)
6222 struct io_ring_ctx *ctx = req->ctx;
6226 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6228 fd = array_index_nospec(fd, ctx->nr_user_files);
6229 file = io_file_from_index(ctx, fd);
6230 io_set_resource_node(req);
6232 trace_io_uring_file_get(ctx, fd);
6233 file = __io_file_get(state, fd);
6236 if (file && unlikely(file->f_op == &io_uring_fops))
6237 io_req_track_inflight(req);
6241 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6243 struct io_timeout_data *data = container_of(timer,
6244 struct io_timeout_data, timer);
6245 struct io_kiocb *prev, *req = data->req;
6246 struct io_ring_ctx *ctx = req->ctx;
6247 unsigned long flags;
6249 spin_lock_irqsave(&ctx->completion_lock, flags);
6250 prev = req->timeout.head;
6251 req->timeout.head = NULL;
6254 * We don't expect the list to be empty, that will only happen if we
6255 * race with the completion of the linked work.
6257 if (prev && refcount_inc_not_zero(&prev->refs))
6258 io_remove_next_linked(prev);
6261 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6264 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6265 io_put_req_deferred(prev, 1);
6267 io_req_complete_post(req, -ETIME, 0);
6268 io_put_req_deferred(req, 1);
6270 return HRTIMER_NORESTART;
6273 static void __io_queue_linked_timeout(struct io_kiocb *req)
6276 * If the back reference is NULL, then our linked request finished
6277 * before we got a chance to setup the timer
6279 if (req->timeout.head) {
6280 struct io_timeout_data *data = req->async_data;
6282 data->timer.function = io_link_timeout_fn;
6283 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6288 static void io_queue_linked_timeout(struct io_kiocb *req)
6290 struct io_ring_ctx *ctx = req->ctx;
6292 spin_lock_irq(&ctx->completion_lock);
6293 __io_queue_linked_timeout(req);
6294 spin_unlock_irq(&ctx->completion_lock);
6296 /* drop submission reference */
6300 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6302 struct io_kiocb *nxt = req->link;
6304 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6305 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6308 nxt->timeout.head = req;
6309 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6310 req->flags |= REQ_F_LINK_TIMEOUT;
6314 static void __io_queue_sqe(struct io_kiocb *req)
6316 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6319 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6322 * We async punt it if the file wasn't marked NOWAIT, or if the file
6323 * doesn't support non-blocking read/write attempts
6325 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6326 if (!io_arm_poll_handler(req)) {
6328 * Queued up for async execution, worker will release
6329 * submit reference when the iocb is actually submitted.
6331 io_queue_async_work(req);
6333 } else if (likely(!ret)) {
6334 /* drop submission reference */
6335 if (req->flags & REQ_F_COMPLETE_INLINE) {
6336 struct io_ring_ctx *ctx = req->ctx;
6337 struct io_comp_state *cs = &ctx->submit_state.comp;
6339 cs->reqs[cs->nr++] = req;
6340 if (cs->nr == ARRAY_SIZE(cs->reqs))
6341 io_submit_flush_completions(cs, ctx);
6346 req_set_fail_links(req);
6348 io_req_complete(req, ret);
6351 io_queue_linked_timeout(linked_timeout);
6354 static void io_queue_sqe(struct io_kiocb *req)
6358 ret = io_req_defer(req);
6360 if (ret != -EIOCBQUEUED) {
6362 req_set_fail_links(req);
6364 io_req_complete(req, ret);
6366 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6367 ret = io_req_defer_prep(req);
6370 io_queue_async_work(req);
6372 __io_queue_sqe(req);
6377 * Check SQE restrictions (opcode and flags).
6379 * Returns 'true' if SQE is allowed, 'false' otherwise.
6381 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6382 struct io_kiocb *req,
6383 unsigned int sqe_flags)
6385 if (!ctx->restricted)
6388 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6391 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6392 ctx->restrictions.sqe_flags_required)
6395 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6396 ctx->restrictions.sqe_flags_required))
6402 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6403 const struct io_uring_sqe *sqe)
6405 struct io_submit_state *state;
6406 unsigned int sqe_flags;
6407 int personality, ret = 0;
6409 req->opcode = READ_ONCE(sqe->opcode);
6410 /* same numerical values with corresponding REQ_F_*, safe to copy */
6411 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6412 req->user_data = READ_ONCE(sqe->user_data);
6413 req->async_data = NULL;
6417 req->fixed_rsrc_refs = NULL;
6418 /* one is dropped after submission, the other at completion */
6419 refcount_set(&req->refs, 2);
6420 req->task = current;
6422 req->work.list.next = NULL;
6423 req->work.creds = NULL;
6424 req->work.flags = 0;
6426 /* enforce forwards compatibility on users */
6427 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6432 if (unlikely(req->opcode >= IORING_OP_LAST))
6435 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6438 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6439 !io_op_defs[req->opcode].buffer_select)
6442 personality = READ_ONCE(sqe->personality);
6444 req->work.creds = xa_load(&ctx->personalities, personality);
6445 if (!req->work.creds)
6447 get_cred(req->work.creds);
6449 state = &ctx->submit_state;
6452 * Plug now if we have more than 1 IO left after this, and the target
6453 * is potentially a read/write to block based storage.
6455 if (!state->plug_started && state->ios_left > 1 &&
6456 io_op_defs[req->opcode].plug) {
6457 blk_start_plug(&state->plug);
6458 state->plug_started = true;
6461 if (io_op_defs[req->opcode].needs_file) {
6462 bool fixed = req->flags & REQ_F_FIXED_FILE;
6464 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6465 if (unlikely(!req->file))
6473 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6474 const struct io_uring_sqe *sqe)
6476 struct io_submit_link *link = &ctx->submit_state.link;
6479 ret = io_init_req(ctx, req, sqe);
6480 if (unlikely(ret)) {
6483 /* fail even hard links since we don't submit */
6484 link->head->flags |= REQ_F_FAIL_LINK;
6485 io_put_req(link->head);
6486 io_req_complete(link->head, -ECANCELED);
6490 io_req_complete(req, ret);
6493 ret = io_req_prep(req, sqe);
6497 /* don't need @sqe from now on */
6498 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6499 true, ctx->flags & IORING_SETUP_SQPOLL);
6502 * If we already have a head request, queue this one for async
6503 * submittal once the head completes. If we don't have a head but
6504 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6505 * submitted sync once the chain is complete. If none of those
6506 * conditions are true (normal request), then just queue it.
6509 struct io_kiocb *head = link->head;
6512 * Taking sequential execution of a link, draining both sides
6513 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6514 * requests in the link. So, it drains the head and the
6515 * next after the link request. The last one is done via
6516 * drain_next flag to persist the effect across calls.
6518 if (req->flags & REQ_F_IO_DRAIN) {
6519 head->flags |= REQ_F_IO_DRAIN;
6520 ctx->drain_next = 1;
6522 ret = io_req_defer_prep(req);
6525 trace_io_uring_link(ctx, req, head);
6526 link->last->link = req;
6529 /* last request of a link, enqueue the link */
6530 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6535 if (unlikely(ctx->drain_next)) {
6536 req->flags |= REQ_F_IO_DRAIN;
6537 ctx->drain_next = 0;
6539 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6551 * Batched submission is done, ensure local IO is flushed out.
6553 static void io_submit_state_end(struct io_submit_state *state,
6554 struct io_ring_ctx *ctx)
6556 if (state->link.head)
6557 io_queue_sqe(state->link.head);
6559 io_submit_flush_completions(&state->comp, ctx);
6560 if (state->plug_started)
6561 blk_finish_plug(&state->plug);
6562 io_state_file_put(state);
6566 * Start submission side cache.
6568 static void io_submit_state_start(struct io_submit_state *state,
6569 unsigned int max_ios)
6571 state->plug_started = false;
6572 state->ios_left = max_ios;
6573 /* set only head, no need to init link_last in advance */
6574 state->link.head = NULL;
6577 static void io_commit_sqring(struct io_ring_ctx *ctx)
6579 struct io_rings *rings = ctx->rings;
6582 * Ensure any loads from the SQEs are done at this point,
6583 * since once we write the new head, the application could
6584 * write new data to them.
6586 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6590 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6591 * that is mapped by userspace. This means that care needs to be taken to
6592 * ensure that reads are stable, as we cannot rely on userspace always
6593 * being a good citizen. If members of the sqe are validated and then later
6594 * used, it's important that those reads are done through READ_ONCE() to
6595 * prevent a re-load down the line.
6597 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6599 u32 *sq_array = ctx->sq_array;
6603 * The cached sq head (or cq tail) serves two purposes:
6605 * 1) allows us to batch the cost of updating the user visible
6607 * 2) allows the kernel side to track the head on its own, even
6608 * though the application is the one updating it.
6610 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6611 if (likely(head < ctx->sq_entries))
6612 return &ctx->sq_sqes[head];
6614 /* drop invalid entries */
6615 ctx->cached_sq_dropped++;
6616 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6620 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6624 /* if we have a backlog and couldn't flush it all, return BUSY */
6625 if (test_bit(0, &ctx->sq_check_overflow)) {
6626 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6630 /* make sure SQ entry isn't read before tail */
6631 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6633 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6636 percpu_counter_add(¤t->io_uring->inflight, nr);
6637 refcount_add(nr, ¤t->usage);
6638 io_submit_state_start(&ctx->submit_state, nr);
6640 while (submitted < nr) {
6641 const struct io_uring_sqe *sqe;
6642 struct io_kiocb *req;
6644 req = io_alloc_req(ctx);
6645 if (unlikely(!req)) {
6647 submitted = -EAGAIN;
6650 sqe = io_get_sqe(ctx);
6651 if (unlikely(!sqe)) {
6652 kmem_cache_free(req_cachep, req);
6655 /* will complete beyond this point, count as submitted */
6657 if (io_submit_sqe(ctx, req, sqe))
6661 if (unlikely(submitted != nr)) {
6662 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6663 struct io_uring_task *tctx = current->io_uring;
6664 int unused = nr - ref_used;
6666 percpu_ref_put_many(&ctx->refs, unused);
6667 percpu_counter_sub(&tctx->inflight, unused);
6668 put_task_struct_many(current, unused);
6671 io_submit_state_end(&ctx->submit_state, ctx);
6672 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6673 io_commit_sqring(ctx);
6678 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6680 /* Tell userspace we may need a wakeup call */
6681 spin_lock_irq(&ctx->completion_lock);
6682 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6683 spin_unlock_irq(&ctx->completion_lock);
6686 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6688 spin_lock_irq(&ctx->completion_lock);
6689 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6690 spin_unlock_irq(&ctx->completion_lock);
6693 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6695 unsigned int to_submit;
6698 to_submit = io_sqring_entries(ctx);
6699 /* if we're handling multiple rings, cap submit size for fairness */
6700 if (cap_entries && to_submit > 8)
6703 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6704 unsigned nr_events = 0;
6706 mutex_lock(&ctx->uring_lock);
6707 if (!list_empty(&ctx->iopoll_list))
6708 io_do_iopoll(ctx, &nr_events, 0);
6710 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6711 !(ctx->flags & IORING_SETUP_R_DISABLED))
6712 ret = io_submit_sqes(ctx, to_submit);
6713 mutex_unlock(&ctx->uring_lock);
6716 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6717 wake_up(&ctx->sqo_sq_wait);
6722 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6724 struct io_ring_ctx *ctx;
6725 unsigned sq_thread_idle = 0;
6727 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6728 if (sq_thread_idle < ctx->sq_thread_idle)
6729 sq_thread_idle = ctx->sq_thread_idle;
6732 sqd->sq_thread_idle = sq_thread_idle;
6735 static int io_sq_thread(void *data)
6737 struct io_sq_data *sqd = data;
6738 struct io_ring_ctx *ctx;
6739 unsigned long timeout = 0;
6740 char buf[TASK_COMM_LEN];
6743 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6744 set_task_comm(current, buf);
6745 current->pf_io_worker = NULL;
6747 if (sqd->sq_cpu != -1)
6748 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6750 set_cpus_allowed_ptr(current, cpu_online_mask);
6751 current->flags |= PF_NO_SETAFFINITY;
6753 mutex_lock(&sqd->lock);
6754 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6756 bool cap_entries, sqt_spin, needs_sched;
6758 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6759 mutex_unlock(&sqd->lock);
6761 mutex_lock(&sqd->lock);
6763 io_run_task_work_head(&sqd->park_task_work);
6764 timeout = jiffies + sqd->sq_thread_idle;
6767 if (signal_pending(current)) {
6768 struct ksignal ksig;
6770 if (!get_signal(&ksig))
6775 cap_entries = !list_is_singular(&sqd->ctx_list);
6776 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6777 const struct cred *creds = NULL;
6779 if (ctx->sq_creds != current_cred())
6780 creds = override_creds(ctx->sq_creds);
6781 ret = __io_sq_thread(ctx, cap_entries);
6783 revert_creds(creds);
6784 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6788 if (sqt_spin || !time_after(jiffies, timeout)) {
6792 timeout = jiffies + sqd->sq_thread_idle;
6797 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6798 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6799 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6800 !list_empty_careful(&ctx->iopoll_list)) {
6801 needs_sched = false;
6804 if (io_sqring_entries(ctx)) {
6805 needs_sched = false;
6810 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6811 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6812 io_ring_set_wakeup_flag(ctx);
6814 mutex_unlock(&sqd->lock);
6816 mutex_lock(&sqd->lock);
6817 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6818 io_ring_clear_wakeup_flag(ctx);
6821 finish_wait(&sqd->wait, &wait);
6822 io_run_task_work_head(&sqd->park_task_work);
6823 timeout = jiffies + sqd->sq_thread_idle;
6826 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6827 io_uring_cancel_sqpoll(ctx);
6829 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6830 io_ring_set_wakeup_flag(ctx);
6831 mutex_unlock(&sqd->lock);
6834 io_run_task_work_head(&sqd->park_task_work);
6835 complete(&sqd->exited);
6839 struct io_wait_queue {
6840 struct wait_queue_entry wq;
6841 struct io_ring_ctx *ctx;
6843 unsigned nr_timeouts;
6846 static inline bool io_should_wake(struct io_wait_queue *iowq)
6848 struct io_ring_ctx *ctx = iowq->ctx;
6851 * Wake up if we have enough events, or if a timeout occurred since we
6852 * started waiting. For timeouts, we always want to return to userspace,
6853 * regardless of event count.
6855 return io_cqring_events(ctx) >= iowq->to_wait ||
6856 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6859 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6860 int wake_flags, void *key)
6862 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6866 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6867 * the task, and the next invocation will do it.
6869 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6870 return autoremove_wake_function(curr, mode, wake_flags, key);
6874 static int io_run_task_work_sig(void)
6876 if (io_run_task_work())
6878 if (!signal_pending(current))
6880 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6881 return -ERESTARTSYS;
6885 /* when returns >0, the caller should retry */
6886 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6887 struct io_wait_queue *iowq,
6888 signed long *timeout)
6892 /* make sure we run task_work before checking for signals */
6893 ret = io_run_task_work_sig();
6894 if (ret || io_should_wake(iowq))
6896 /* let the caller flush overflows, retry */
6897 if (test_bit(0, &ctx->cq_check_overflow))
6900 *timeout = schedule_timeout(*timeout);
6901 return !*timeout ? -ETIME : 1;
6905 * Wait until events become available, if we don't already have some. The
6906 * application must reap them itself, as they reside on the shared cq ring.
6908 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6909 const sigset_t __user *sig, size_t sigsz,
6910 struct __kernel_timespec __user *uts)
6912 struct io_wait_queue iowq = {
6915 .func = io_wake_function,
6916 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6919 .to_wait = min_events,
6921 struct io_rings *rings = ctx->rings;
6922 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6926 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6927 if (io_cqring_events(ctx) >= min_events)
6929 if (!io_run_task_work())
6934 #ifdef CONFIG_COMPAT
6935 if (in_compat_syscall())
6936 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6940 ret = set_user_sigmask(sig, sigsz);
6947 struct timespec64 ts;
6949 if (get_timespec64(&ts, uts))
6951 timeout = timespec64_to_jiffies(&ts);
6954 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6955 trace_io_uring_cqring_wait(ctx, min_events);
6957 /* if we can't even flush overflow, don't wait for more */
6958 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6962 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6963 TASK_INTERRUPTIBLE);
6964 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6965 finish_wait(&ctx->wait, &iowq.wq);
6969 restore_saved_sigmask_unless(ret == -EINTR);
6971 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6974 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6976 #if defined(CONFIG_UNIX)
6977 if (ctx->ring_sock) {
6978 struct sock *sock = ctx->ring_sock->sk;
6979 struct sk_buff *skb;
6981 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6987 for (i = 0; i < ctx->nr_user_files; i++) {
6990 file = io_file_from_index(ctx, i);
6997 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6999 struct fixed_rsrc_data *data;
7001 data = container_of(ref, struct fixed_rsrc_data, refs);
7002 complete(&data->done);
7005 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7007 spin_lock_bh(&ctx->rsrc_ref_lock);
7010 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7012 spin_unlock_bh(&ctx->rsrc_ref_lock);
7015 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7016 struct fixed_rsrc_data *rsrc_data,
7017 struct fixed_rsrc_ref_node *ref_node)
7019 io_rsrc_ref_lock(ctx);
7020 rsrc_data->node = ref_node;
7021 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7022 io_rsrc_ref_unlock(ctx);
7023 percpu_ref_get(&rsrc_data->refs);
7026 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7028 struct fixed_rsrc_ref_node *ref_node = NULL;
7030 io_rsrc_ref_lock(ctx);
7031 ref_node = data->node;
7033 io_rsrc_ref_unlock(ctx);
7035 percpu_ref_kill(&ref_node->refs);
7038 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7039 struct io_ring_ctx *ctx,
7040 void (*rsrc_put)(struct io_ring_ctx *ctx,
7041 struct io_rsrc_put *prsrc))
7043 struct fixed_rsrc_ref_node *backup_node;
7049 data->quiesce = true;
7052 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7055 backup_node->rsrc_data = data;
7056 backup_node->rsrc_put = rsrc_put;
7058 io_sqe_rsrc_kill_node(ctx, data);
7059 percpu_ref_kill(&data->refs);
7060 flush_delayed_work(&ctx->rsrc_put_work);
7062 ret = wait_for_completion_interruptible(&data->done);
7066 percpu_ref_resurrect(&data->refs);
7067 io_sqe_rsrc_set_node(ctx, data, backup_node);
7069 reinit_completion(&data->done);
7070 mutex_unlock(&ctx->uring_lock);
7071 ret = io_run_task_work_sig();
7072 mutex_lock(&ctx->uring_lock);
7074 data->quiesce = false;
7077 destroy_fixed_rsrc_ref_node(backup_node);
7081 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7083 struct fixed_rsrc_data *data;
7085 data = kzalloc(sizeof(*data), GFP_KERNEL);
7089 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7090 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7095 init_completion(&data->done);
7099 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7101 percpu_ref_exit(&data->refs);
7106 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7108 struct fixed_rsrc_data *data = ctx->file_data;
7109 unsigned nr_tables, i;
7113 * percpu_ref_is_dying() is to stop parallel files unregister
7114 * Since we possibly drop uring lock later in this function to
7117 if (!data || percpu_ref_is_dying(&data->refs))
7119 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7123 __io_sqe_files_unregister(ctx);
7124 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7125 for (i = 0; i < nr_tables; i++)
7126 kfree(data->table[i].files);
7127 free_fixed_rsrc_data(data);
7128 ctx->file_data = NULL;
7129 ctx->nr_user_files = 0;
7133 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7134 __releases(&sqd->lock)
7136 WARN_ON_ONCE(sqd->thread == current);
7139 * Do the dance but not conditional clear_bit() because it'd race with
7140 * other threads incrementing park_pending and setting the bit.
7142 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7143 if (atomic_dec_return(&sqd->park_pending))
7144 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7145 mutex_unlock(&sqd->lock);
7148 static void io_sq_thread_park(struct io_sq_data *sqd)
7149 __acquires(&sqd->lock)
7151 WARN_ON_ONCE(sqd->thread == current);
7153 atomic_inc(&sqd->park_pending);
7154 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7155 mutex_lock(&sqd->lock);
7157 wake_up_process(sqd->thread);
7160 static void io_sq_thread_stop(struct io_sq_data *sqd)
7162 WARN_ON_ONCE(sqd->thread == current);
7164 mutex_lock(&sqd->lock);
7165 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7167 wake_up_process(sqd->thread);
7168 mutex_unlock(&sqd->lock);
7169 wait_for_completion(&sqd->exited);
7172 static void io_put_sq_data(struct io_sq_data *sqd)
7174 if (refcount_dec_and_test(&sqd->refs)) {
7175 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7177 io_sq_thread_stop(sqd);
7182 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7184 struct io_sq_data *sqd = ctx->sq_data;
7187 io_sq_thread_park(sqd);
7188 list_del_init(&ctx->sqd_list);
7189 io_sqd_update_thread_idle(sqd);
7190 io_sq_thread_unpark(sqd);
7192 io_put_sq_data(sqd);
7193 ctx->sq_data = NULL;
7195 put_cred(ctx->sq_creds);
7199 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7201 struct io_ring_ctx *ctx_attach;
7202 struct io_sq_data *sqd;
7205 f = fdget(p->wq_fd);
7207 return ERR_PTR(-ENXIO);
7208 if (f.file->f_op != &io_uring_fops) {
7210 return ERR_PTR(-EINVAL);
7213 ctx_attach = f.file->private_data;
7214 sqd = ctx_attach->sq_data;
7217 return ERR_PTR(-EINVAL);
7219 if (sqd->task_tgid != current->tgid) {
7221 return ERR_PTR(-EPERM);
7224 refcount_inc(&sqd->refs);
7229 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7232 struct io_sq_data *sqd;
7235 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7236 sqd = io_attach_sq_data(p);
7241 /* fall through for EPERM case, setup new sqd/task */
7242 if (PTR_ERR(sqd) != -EPERM)
7246 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7248 return ERR_PTR(-ENOMEM);
7250 atomic_set(&sqd->park_pending, 0);
7251 refcount_set(&sqd->refs, 1);
7252 INIT_LIST_HEAD(&sqd->ctx_list);
7253 mutex_init(&sqd->lock);
7254 init_waitqueue_head(&sqd->wait);
7255 init_completion(&sqd->exited);
7259 #if defined(CONFIG_UNIX)
7261 * Ensure the UNIX gc is aware of our file set, so we are certain that
7262 * the io_uring can be safely unregistered on process exit, even if we have
7263 * loops in the file referencing.
7265 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7267 struct sock *sk = ctx->ring_sock->sk;
7268 struct scm_fp_list *fpl;
7269 struct sk_buff *skb;
7272 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7276 skb = alloc_skb(0, GFP_KERNEL);
7285 fpl->user = get_uid(current_user());
7286 for (i = 0; i < nr; i++) {
7287 struct file *file = io_file_from_index(ctx, i + offset);
7291 fpl->fp[nr_files] = get_file(file);
7292 unix_inflight(fpl->user, fpl->fp[nr_files]);
7297 fpl->max = SCM_MAX_FD;
7298 fpl->count = nr_files;
7299 UNIXCB(skb).fp = fpl;
7300 skb->destructor = unix_destruct_scm;
7301 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7302 skb_queue_head(&sk->sk_receive_queue, skb);
7304 for (i = 0; i < nr_files; i++)
7315 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7316 * causes regular reference counting to break down. We rely on the UNIX
7317 * garbage collection to take care of this problem for us.
7319 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7321 unsigned left, total;
7325 left = ctx->nr_user_files;
7327 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7329 ret = __io_sqe_files_scm(ctx, this_files, total);
7333 total += this_files;
7339 while (total < ctx->nr_user_files) {
7340 struct file *file = io_file_from_index(ctx, total);
7350 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7356 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7357 unsigned nr_tables, unsigned nr_files)
7361 for (i = 0; i < nr_tables; i++) {
7362 struct fixed_rsrc_table *table = &file_data->table[i];
7363 unsigned this_files;
7365 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7366 table->files = kcalloc(this_files, sizeof(struct file *),
7370 nr_files -= this_files;
7376 for (i = 0; i < nr_tables; i++) {
7377 struct fixed_rsrc_table *table = &file_data->table[i];
7378 kfree(table->files);
7383 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7385 struct file *file = prsrc->file;
7386 #if defined(CONFIG_UNIX)
7387 struct sock *sock = ctx->ring_sock->sk;
7388 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7389 struct sk_buff *skb;
7392 __skb_queue_head_init(&list);
7395 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7396 * remove this entry and rearrange the file array.
7398 skb = skb_dequeue(head);
7400 struct scm_fp_list *fp;
7402 fp = UNIXCB(skb).fp;
7403 for (i = 0; i < fp->count; i++) {
7406 if (fp->fp[i] != file)
7409 unix_notinflight(fp->user, fp->fp[i]);
7410 left = fp->count - 1 - i;
7412 memmove(&fp->fp[i], &fp->fp[i + 1],
7413 left * sizeof(struct file *));
7420 __skb_queue_tail(&list, skb);
7430 __skb_queue_tail(&list, skb);
7432 skb = skb_dequeue(head);
7435 if (skb_peek(&list)) {
7436 spin_lock_irq(&head->lock);
7437 while ((skb = __skb_dequeue(&list)) != NULL)
7438 __skb_queue_tail(head, skb);
7439 spin_unlock_irq(&head->lock);
7446 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7448 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7449 struct io_ring_ctx *ctx = rsrc_data->ctx;
7450 struct io_rsrc_put *prsrc, *tmp;
7452 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7453 list_del(&prsrc->list);
7454 ref_node->rsrc_put(ctx, prsrc);
7458 percpu_ref_exit(&ref_node->refs);
7460 percpu_ref_put(&rsrc_data->refs);
7463 static void io_rsrc_put_work(struct work_struct *work)
7465 struct io_ring_ctx *ctx;
7466 struct llist_node *node;
7468 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7469 node = llist_del_all(&ctx->rsrc_put_llist);
7472 struct fixed_rsrc_ref_node *ref_node;
7473 struct llist_node *next = node->next;
7475 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7476 __io_rsrc_put_work(ref_node);
7481 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7484 struct fixed_rsrc_table *table;
7486 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7487 return &table->files[i & IORING_FILE_TABLE_MASK];
7490 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7492 struct fixed_rsrc_ref_node *ref_node;
7493 struct fixed_rsrc_data *data;
7494 struct io_ring_ctx *ctx;
7495 bool first_add = false;
7498 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7499 data = ref_node->rsrc_data;
7502 io_rsrc_ref_lock(ctx);
7503 ref_node->done = true;
7505 while (!list_empty(&ctx->rsrc_ref_list)) {
7506 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7507 struct fixed_rsrc_ref_node, node);
7508 /* recycle ref nodes in order */
7509 if (!ref_node->done)
7511 list_del(&ref_node->node);
7512 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7514 io_rsrc_ref_unlock(ctx);
7516 if (percpu_ref_is_dying(&data->refs))
7520 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7522 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7525 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7526 struct io_ring_ctx *ctx)
7528 struct fixed_rsrc_ref_node *ref_node;
7530 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7534 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7539 INIT_LIST_HEAD(&ref_node->node);
7540 INIT_LIST_HEAD(&ref_node->rsrc_list);
7541 ref_node->done = false;
7545 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7546 struct fixed_rsrc_ref_node *ref_node)
7548 ref_node->rsrc_data = ctx->file_data;
7549 ref_node->rsrc_put = io_ring_file_put;
7552 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7554 percpu_ref_exit(&ref_node->refs);
7559 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7562 __s32 __user *fds = (__s32 __user *) arg;
7563 unsigned nr_tables, i;
7565 int fd, ret = -ENOMEM;
7566 struct fixed_rsrc_ref_node *ref_node;
7567 struct fixed_rsrc_data *file_data;
7573 if (nr_args > IORING_MAX_FIXED_FILES)
7576 file_data = alloc_fixed_rsrc_data(ctx);
7579 ctx->file_data = file_data;
7581 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7582 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7584 if (!file_data->table)
7587 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7590 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7591 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7595 /* allow sparse sets */
7605 * Don't allow io_uring instances to be registered. If UNIX
7606 * isn't enabled, then this causes a reference cycle and this
7607 * instance can never get freed. If UNIX is enabled we'll
7608 * handle it just fine, but there's still no point in allowing
7609 * a ring fd as it doesn't support regular read/write anyway.
7611 if (file->f_op == &io_uring_fops) {
7615 *io_fixed_file_slot(file_data, i) = file;
7618 ret = io_sqe_files_scm(ctx);
7620 io_sqe_files_unregister(ctx);
7624 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7626 io_sqe_files_unregister(ctx);
7629 init_fixed_file_ref_node(ctx, ref_node);
7631 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7634 for (i = 0; i < ctx->nr_user_files; i++) {
7635 file = io_file_from_index(ctx, i);
7639 for (i = 0; i < nr_tables; i++)
7640 kfree(file_data->table[i].files);
7641 ctx->nr_user_files = 0;
7643 free_fixed_rsrc_data(ctx->file_data);
7644 ctx->file_data = NULL;
7648 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7651 #if defined(CONFIG_UNIX)
7652 struct sock *sock = ctx->ring_sock->sk;
7653 struct sk_buff_head *head = &sock->sk_receive_queue;
7654 struct sk_buff *skb;
7657 * See if we can merge this file into an existing skb SCM_RIGHTS
7658 * file set. If there's no room, fall back to allocating a new skb
7659 * and filling it in.
7661 spin_lock_irq(&head->lock);
7662 skb = skb_peek(head);
7664 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7666 if (fpl->count < SCM_MAX_FD) {
7667 __skb_unlink(skb, head);
7668 spin_unlock_irq(&head->lock);
7669 fpl->fp[fpl->count] = get_file(file);
7670 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7672 spin_lock_irq(&head->lock);
7673 __skb_queue_head(head, skb);
7678 spin_unlock_irq(&head->lock);
7685 return __io_sqe_files_scm(ctx, 1, index);
7691 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7693 struct io_rsrc_put *prsrc;
7694 struct fixed_rsrc_ref_node *ref_node = data->node;
7696 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7701 list_add(&prsrc->list, &ref_node->rsrc_list);
7706 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7709 return io_queue_rsrc_removal(data, (void *)file);
7712 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7713 struct io_uring_rsrc_update *up,
7716 struct fixed_rsrc_data *data = ctx->file_data;
7717 struct fixed_rsrc_ref_node *ref_node;
7718 struct file *file, **file_slot;
7722 bool needs_switch = false;
7724 if (check_add_overflow(up->offset, nr_args, &done))
7726 if (done > ctx->nr_user_files)
7729 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7732 init_fixed_file_ref_node(ctx, ref_node);
7734 fds = u64_to_user_ptr(up->data);
7735 for (done = 0; done < nr_args; done++) {
7737 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7741 if (fd == IORING_REGISTER_FILES_SKIP)
7744 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7745 file_slot = io_fixed_file_slot(ctx->file_data, i);
7748 err = io_queue_file_removal(data, *file_slot);
7752 needs_switch = true;
7761 * Don't allow io_uring instances to be registered. If
7762 * UNIX isn't enabled, then this causes a reference
7763 * cycle and this instance can never get freed. If UNIX
7764 * is enabled we'll handle it just fine, but there's
7765 * still no point in allowing a ring fd as it doesn't
7766 * support regular read/write anyway.
7768 if (file->f_op == &io_uring_fops) {
7774 err = io_sqe_file_register(ctx, file, i);
7784 percpu_ref_kill(&data->node->refs);
7785 io_sqe_rsrc_set_node(ctx, data, ref_node);
7787 destroy_fixed_rsrc_ref_node(ref_node);
7789 return done ? done : err;
7792 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7795 struct io_uring_rsrc_update up;
7797 if (!ctx->file_data)
7801 if (copy_from_user(&up, arg, sizeof(up)))
7806 return __io_sqe_files_update(ctx, &up, nr_args);
7809 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7811 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7813 req = io_put_req_find_next(req);
7814 return req ? &req->work : NULL;
7817 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7819 struct io_wq_hash *hash;
7820 struct io_wq_data data;
7821 unsigned int concurrency;
7823 hash = ctx->hash_map;
7825 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7827 return ERR_PTR(-ENOMEM);
7828 refcount_set(&hash->refs, 1);
7829 init_waitqueue_head(&hash->wait);
7830 ctx->hash_map = hash;
7834 data.free_work = io_free_work;
7835 data.do_work = io_wq_submit_work;
7837 /* Do QD, or 4 * CPUS, whatever is smallest */
7838 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7840 return io_wq_create(concurrency, &data);
7843 static int io_uring_alloc_task_context(struct task_struct *task,
7844 struct io_ring_ctx *ctx)
7846 struct io_uring_task *tctx;
7849 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7850 if (unlikely(!tctx))
7853 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7854 if (unlikely(ret)) {
7859 tctx->io_wq = io_init_wq_offload(ctx);
7860 if (IS_ERR(tctx->io_wq)) {
7861 ret = PTR_ERR(tctx->io_wq);
7862 percpu_counter_destroy(&tctx->inflight);
7868 init_waitqueue_head(&tctx->wait);
7870 atomic_set(&tctx->in_idle, 0);
7871 task->io_uring = tctx;
7872 spin_lock_init(&tctx->task_lock);
7873 INIT_WQ_LIST(&tctx->task_list);
7874 tctx->task_state = 0;
7875 init_task_work(&tctx->task_work, tctx_task_work);
7879 void __io_uring_free(struct task_struct *tsk)
7881 struct io_uring_task *tctx = tsk->io_uring;
7883 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7884 WARN_ON_ONCE(tctx->io_wq);
7886 percpu_counter_destroy(&tctx->inflight);
7888 tsk->io_uring = NULL;
7891 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7892 struct io_uring_params *p)
7896 /* Retain compatibility with failing for an invalid attach attempt */
7897 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7898 IORING_SETUP_ATTACH_WQ) {
7901 f = fdget(p->wq_fd);
7904 if (f.file->f_op != &io_uring_fops) {
7910 if (ctx->flags & IORING_SETUP_SQPOLL) {
7911 struct task_struct *tsk;
7912 struct io_sq_data *sqd;
7916 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7919 sqd = io_get_sq_data(p, &attached);
7925 ctx->sq_creds = get_current_cred();
7927 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7928 if (!ctx->sq_thread_idle)
7929 ctx->sq_thread_idle = HZ;
7932 io_sq_thread_park(sqd);
7933 list_add(&ctx->sqd_list, &sqd->ctx_list);
7934 io_sqd_update_thread_idle(sqd);
7935 /* don't attach to a dying SQPOLL thread, would be racy */
7936 if (attached && !sqd->thread)
7938 io_sq_thread_unpark(sqd);
7945 if (p->flags & IORING_SETUP_SQ_AFF) {
7946 int cpu = p->sq_thread_cpu;
7949 if (cpu >= nr_cpu_ids)
7951 if (!cpu_online(cpu))
7959 sqd->task_pid = current->pid;
7960 sqd->task_tgid = current->tgid;
7961 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7968 ret = io_uring_alloc_task_context(tsk, ctx);
7969 wake_up_new_task(tsk);
7972 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7973 /* Can't have SQ_AFF without SQPOLL */
7980 io_sq_thread_finish(ctx);
7983 complete(&ctx->sq_data->exited);
7987 static inline void __io_unaccount_mem(struct user_struct *user,
7988 unsigned long nr_pages)
7990 atomic_long_sub(nr_pages, &user->locked_vm);
7993 static inline int __io_account_mem(struct user_struct *user,
7994 unsigned long nr_pages)
7996 unsigned long page_limit, cur_pages, new_pages;
7998 /* Don't allow more pages than we can safely lock */
7999 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8002 cur_pages = atomic_long_read(&user->locked_vm);
8003 new_pages = cur_pages + nr_pages;
8004 if (new_pages > page_limit)
8006 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8007 new_pages) != cur_pages);
8012 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8015 __io_unaccount_mem(ctx->user, nr_pages);
8017 if (ctx->mm_account)
8018 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8021 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8026 ret = __io_account_mem(ctx->user, nr_pages);
8031 if (ctx->mm_account)
8032 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8037 static void io_mem_free(void *ptr)
8044 page = virt_to_head_page(ptr);
8045 if (put_page_testzero(page))
8046 free_compound_page(page);
8049 static void *io_mem_alloc(size_t size)
8051 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8052 __GFP_NORETRY | __GFP_ACCOUNT;
8054 return (void *) __get_free_pages(gfp_flags, get_order(size));
8057 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8060 struct io_rings *rings;
8061 size_t off, sq_array_size;
8063 off = struct_size(rings, cqes, cq_entries);
8064 if (off == SIZE_MAX)
8068 off = ALIGN(off, SMP_CACHE_BYTES);
8076 sq_array_size = array_size(sizeof(u32), sq_entries);
8077 if (sq_array_size == SIZE_MAX)
8080 if (check_add_overflow(off, sq_array_size, &off))
8086 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8090 if (!ctx->user_bufs)
8093 for (i = 0; i < ctx->nr_user_bufs; i++) {
8094 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8096 for (j = 0; j < imu->nr_bvecs; j++)
8097 unpin_user_page(imu->bvec[j].bv_page);
8099 if (imu->acct_pages)
8100 io_unaccount_mem(ctx, imu->acct_pages);
8105 kfree(ctx->user_bufs);
8106 ctx->user_bufs = NULL;
8107 ctx->nr_user_bufs = 0;
8111 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8112 void __user *arg, unsigned index)
8114 struct iovec __user *src;
8116 #ifdef CONFIG_COMPAT
8118 struct compat_iovec __user *ciovs;
8119 struct compat_iovec ciov;
8121 ciovs = (struct compat_iovec __user *) arg;
8122 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8125 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8126 dst->iov_len = ciov.iov_len;
8130 src = (struct iovec __user *) arg;
8131 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8137 * Not super efficient, but this is just a registration time. And we do cache
8138 * the last compound head, so generally we'll only do a full search if we don't
8141 * We check if the given compound head page has already been accounted, to
8142 * avoid double accounting it. This allows us to account the full size of the
8143 * page, not just the constituent pages of a huge page.
8145 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8146 int nr_pages, struct page *hpage)
8150 /* check current page array */
8151 for (i = 0; i < nr_pages; i++) {
8152 if (!PageCompound(pages[i]))
8154 if (compound_head(pages[i]) == hpage)
8158 /* check previously registered pages */
8159 for (i = 0; i < ctx->nr_user_bufs; i++) {
8160 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8162 for (j = 0; j < imu->nr_bvecs; j++) {
8163 if (!PageCompound(imu->bvec[j].bv_page))
8165 if (compound_head(imu->bvec[j].bv_page) == hpage)
8173 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8174 int nr_pages, struct io_mapped_ubuf *imu,
8175 struct page **last_hpage)
8179 for (i = 0; i < nr_pages; i++) {
8180 if (!PageCompound(pages[i])) {
8185 hpage = compound_head(pages[i]);
8186 if (hpage == *last_hpage)
8188 *last_hpage = hpage;
8189 if (headpage_already_acct(ctx, pages, i, hpage))
8191 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8195 if (!imu->acct_pages)
8198 ret = io_account_mem(ctx, imu->acct_pages);
8200 imu->acct_pages = 0;
8204 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8205 struct io_mapped_ubuf *imu,
8206 struct page **last_hpage)
8208 struct vm_area_struct **vmas = NULL;
8209 struct page **pages = NULL;
8210 unsigned long off, start, end, ubuf;
8212 int ret, pret, nr_pages, i;
8214 ubuf = (unsigned long) iov->iov_base;
8215 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8216 start = ubuf >> PAGE_SHIFT;
8217 nr_pages = end - start;
8221 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8225 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8230 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8236 mmap_read_lock(current->mm);
8237 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8239 if (pret == nr_pages) {
8240 /* don't support file backed memory */
8241 for (i = 0; i < nr_pages; i++) {
8242 struct vm_area_struct *vma = vmas[i];
8245 !is_file_hugepages(vma->vm_file)) {
8251 ret = pret < 0 ? pret : -EFAULT;
8253 mmap_read_unlock(current->mm);
8256 * if we did partial map, or found file backed vmas,
8257 * release any pages we did get
8260 unpin_user_pages(pages, pret);
8265 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8267 unpin_user_pages(pages, pret);
8272 off = ubuf & ~PAGE_MASK;
8273 size = iov->iov_len;
8274 for (i = 0; i < nr_pages; i++) {
8277 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8278 imu->bvec[i].bv_page = pages[i];
8279 imu->bvec[i].bv_len = vec_len;
8280 imu->bvec[i].bv_offset = off;
8284 /* store original address for later verification */
8286 imu->len = iov->iov_len;
8287 imu->nr_bvecs = nr_pages;
8295 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8299 if (!nr_args || nr_args > UIO_MAXIOV)
8302 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8304 if (!ctx->user_bufs)
8310 static int io_buffer_validate(struct iovec *iov)
8313 * Don't impose further limits on the size and buffer
8314 * constraints here, we'll -EINVAL later when IO is
8315 * submitted if they are wrong.
8317 if (!iov->iov_base || !iov->iov_len)
8320 /* arbitrary limit, but we need something */
8321 if (iov->iov_len > SZ_1G)
8327 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8328 unsigned int nr_args)
8332 struct page *last_hpage = NULL;
8334 ret = io_buffers_map_alloc(ctx, nr_args);
8338 for (i = 0; i < nr_args; i++) {
8339 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8341 ret = io_copy_iov(ctx, &iov, arg, i);
8345 ret = io_buffer_validate(&iov);
8349 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8353 ctx->nr_user_bufs++;
8357 io_sqe_buffers_unregister(ctx);
8362 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8364 __s32 __user *fds = arg;
8370 if (copy_from_user(&fd, fds, sizeof(*fds)))
8373 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8374 if (IS_ERR(ctx->cq_ev_fd)) {
8375 int ret = PTR_ERR(ctx->cq_ev_fd);
8376 ctx->cq_ev_fd = NULL;
8383 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8385 if (ctx->cq_ev_fd) {
8386 eventfd_ctx_put(ctx->cq_ev_fd);
8387 ctx->cq_ev_fd = NULL;
8394 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8396 struct io_buffer *buf;
8397 unsigned long index;
8399 xa_for_each(&ctx->io_buffers, index, buf)
8400 __io_remove_buffers(ctx, buf, index, -1U);
8403 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8405 struct io_kiocb *req, *nxt;
8407 list_for_each_entry_safe(req, nxt, list, compl.list) {
8408 if (tsk && req->task != tsk)
8410 list_del(&req->compl.list);
8411 kmem_cache_free(req_cachep, req);
8415 static void io_req_caches_free(struct io_ring_ctx *ctx)
8417 struct io_submit_state *submit_state = &ctx->submit_state;
8418 struct io_comp_state *cs = &ctx->submit_state.comp;
8420 mutex_lock(&ctx->uring_lock);
8422 if (submit_state->free_reqs) {
8423 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8424 submit_state->reqs);
8425 submit_state->free_reqs = 0;
8428 spin_lock_irq(&ctx->completion_lock);
8429 list_splice_init(&cs->locked_free_list, &cs->free_list);
8430 cs->locked_free_nr = 0;
8431 spin_unlock_irq(&ctx->completion_lock);
8433 io_req_cache_free(&cs->free_list, NULL);
8435 mutex_unlock(&ctx->uring_lock);
8438 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8441 * Some may use context even when all refs and requests have been put,
8442 * and they are free to do so while still holding uring_lock or
8443 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8445 mutex_lock(&ctx->uring_lock);
8446 mutex_unlock(&ctx->uring_lock);
8447 spin_lock_irq(&ctx->completion_lock);
8448 spin_unlock_irq(&ctx->completion_lock);
8450 io_sq_thread_finish(ctx);
8451 io_sqe_buffers_unregister(ctx);
8453 if (ctx->mm_account) {
8454 mmdrop(ctx->mm_account);
8455 ctx->mm_account = NULL;
8458 mutex_lock(&ctx->uring_lock);
8459 io_sqe_files_unregister(ctx);
8460 mutex_unlock(&ctx->uring_lock);
8461 io_eventfd_unregister(ctx);
8462 io_destroy_buffers(ctx);
8464 #if defined(CONFIG_UNIX)
8465 if (ctx->ring_sock) {
8466 ctx->ring_sock->file = NULL; /* so that iput() is called */
8467 sock_release(ctx->ring_sock);
8471 io_mem_free(ctx->rings);
8472 io_mem_free(ctx->sq_sqes);
8474 percpu_ref_exit(&ctx->refs);
8475 free_uid(ctx->user);
8476 io_req_caches_free(ctx);
8478 io_wq_put_hash(ctx->hash_map);
8479 kfree(ctx->cancel_hash);
8483 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8485 struct io_ring_ctx *ctx = file->private_data;
8488 poll_wait(file, &ctx->cq_wait, wait);
8490 * synchronizes with barrier from wq_has_sleeper call in
8494 if (!io_sqring_full(ctx))
8495 mask |= EPOLLOUT | EPOLLWRNORM;
8498 * Don't flush cqring overflow list here, just do a simple check.
8499 * Otherwise there could possible be ABBA deadlock:
8502 * lock(&ctx->uring_lock);
8504 * lock(&ctx->uring_lock);
8507 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8508 * pushs them to do the flush.
8510 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8511 mask |= EPOLLIN | EPOLLRDNORM;
8516 static int io_uring_fasync(int fd, struct file *file, int on)
8518 struct io_ring_ctx *ctx = file->private_data;
8520 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8523 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8525 const struct cred *creds;
8527 creds = xa_erase(&ctx->personalities, id);
8536 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8538 return io_run_task_work_head(&ctx->exit_task_work);
8541 struct io_tctx_exit {
8542 struct callback_head task_work;
8543 struct completion completion;
8544 struct io_ring_ctx *ctx;
8547 static void io_tctx_exit_cb(struct callback_head *cb)
8549 struct io_uring_task *tctx = current->io_uring;
8550 struct io_tctx_exit *work;
8552 work = container_of(cb, struct io_tctx_exit, task_work);
8554 * When @in_idle, we're in cancellation and it's racy to remove the
8555 * node. It'll be removed by the end of cancellation, just ignore it.
8557 if (!atomic_read(&tctx->in_idle))
8558 io_uring_del_task_file((unsigned long)work->ctx);
8559 complete(&work->completion);
8562 static void io_ring_exit_work(struct work_struct *work)
8564 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8565 unsigned long timeout = jiffies + HZ * 60 * 5;
8566 struct io_tctx_exit exit;
8567 struct io_tctx_node *node;
8570 /* prevent SQPOLL from submitting new requests */
8572 io_sq_thread_park(ctx->sq_data);
8573 list_del_init(&ctx->sqd_list);
8574 io_sqd_update_thread_idle(ctx->sq_data);
8575 io_sq_thread_unpark(ctx->sq_data);
8579 * If we're doing polled IO and end up having requests being
8580 * submitted async (out-of-line), then completions can come in while
8581 * we're waiting for refs to drop. We need to reap these manually,
8582 * as nobody else will be looking for them.
8585 io_uring_try_cancel_requests(ctx, NULL, NULL);
8587 WARN_ON_ONCE(time_after(jiffies, timeout));
8588 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8590 mutex_lock(&ctx->uring_lock);
8591 while (!list_empty(&ctx->tctx_list)) {
8592 WARN_ON_ONCE(time_after(jiffies, timeout));
8594 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8597 init_completion(&exit.completion);
8598 init_task_work(&exit.task_work, io_tctx_exit_cb);
8599 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8600 if (WARN_ON_ONCE(ret))
8602 wake_up_process(node->task);
8604 mutex_unlock(&ctx->uring_lock);
8605 wait_for_completion(&exit.completion);
8607 mutex_lock(&ctx->uring_lock);
8609 mutex_unlock(&ctx->uring_lock);
8611 io_ring_ctx_free(ctx);
8614 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8616 unsigned long index;
8617 struct creds *creds;
8619 mutex_lock(&ctx->uring_lock);
8620 percpu_ref_kill(&ctx->refs);
8621 /* if force is set, the ring is going away. always drop after that */
8622 ctx->cq_overflow_flushed = 1;
8624 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8625 xa_for_each(&ctx->personalities, index, creds)
8626 io_unregister_personality(ctx, index);
8627 mutex_unlock(&ctx->uring_lock);
8629 io_kill_timeouts(ctx, NULL, NULL);
8630 io_poll_remove_all(ctx, NULL, NULL);
8632 /* if we failed setting up the ctx, we might not have any rings */
8633 io_iopoll_try_reap_events(ctx);
8635 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8637 * Use system_unbound_wq to avoid spawning tons of event kworkers
8638 * if we're exiting a ton of rings at the same time. It just adds
8639 * noise and overhead, there's no discernable change in runtime
8640 * over using system_wq.
8642 queue_work(system_unbound_wq, &ctx->exit_work);
8645 static int io_uring_release(struct inode *inode, struct file *file)
8647 struct io_ring_ctx *ctx = file->private_data;
8649 file->private_data = NULL;
8650 io_ring_ctx_wait_and_kill(ctx);
8654 struct io_task_cancel {
8655 struct task_struct *task;
8656 struct files_struct *files;
8659 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8661 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8662 struct io_task_cancel *cancel = data;
8665 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8666 unsigned long flags;
8667 struct io_ring_ctx *ctx = req->ctx;
8669 /* protect against races with linked timeouts */
8670 spin_lock_irqsave(&ctx->completion_lock, flags);
8671 ret = io_match_task(req, cancel->task, cancel->files);
8672 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8674 ret = io_match_task(req, cancel->task, cancel->files);
8679 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8680 struct task_struct *task,
8681 struct files_struct *files)
8683 struct io_defer_entry *de;
8686 spin_lock_irq(&ctx->completion_lock);
8687 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8688 if (io_match_task(de->req, task, files)) {
8689 list_cut_position(&list, &ctx->defer_list, &de->list);
8693 spin_unlock_irq(&ctx->completion_lock);
8694 if (list_empty(&list))
8697 while (!list_empty(&list)) {
8698 de = list_first_entry(&list, struct io_defer_entry, list);
8699 list_del_init(&de->list);
8700 req_set_fail_links(de->req);
8701 io_put_req(de->req);
8702 io_req_complete(de->req, -ECANCELED);
8708 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8710 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8712 return req->ctx == data;
8715 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8717 struct io_tctx_node *node;
8718 enum io_wq_cancel cret;
8721 mutex_lock(&ctx->uring_lock);
8722 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8723 struct io_uring_task *tctx = node->task->io_uring;
8726 * io_wq will stay alive while we hold uring_lock, because it's
8727 * killed after ctx nodes, which requires to take the lock.
8729 if (!tctx || !tctx->io_wq)
8731 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8732 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8734 mutex_unlock(&ctx->uring_lock);
8739 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8740 struct task_struct *task,
8741 struct files_struct *files)
8743 struct io_task_cancel cancel = { .task = task, .files = files, };
8744 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8747 enum io_wq_cancel cret;
8751 ret |= io_uring_try_cancel_iowq(ctx);
8752 } else if (tctx && tctx->io_wq) {
8754 * Cancels requests of all rings, not only @ctx, but
8755 * it's fine as the task is in exit/exec.
8757 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8759 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8762 /* SQPOLL thread does its own polling */
8763 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8764 (ctx->sq_data && ctx->sq_data->thread == current)) {
8765 while (!list_empty_careful(&ctx->iopoll_list)) {
8766 io_iopoll_try_reap_events(ctx);
8771 ret |= io_cancel_defer_files(ctx, task, files);
8772 ret |= io_poll_remove_all(ctx, task, files);
8773 ret |= io_kill_timeouts(ctx, task, files);
8774 ret |= io_run_task_work();
8775 ret |= io_run_ctx_fallback(ctx);
8776 io_cqring_overflow_flush(ctx, true, task, files);
8783 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8784 struct task_struct *task,
8785 struct files_struct *files)
8787 struct io_kiocb *req;
8790 spin_lock_irq(&ctx->inflight_lock);
8791 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8792 cnt += io_match_task(req, task, files);
8793 spin_unlock_irq(&ctx->inflight_lock);
8797 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8798 struct task_struct *task,
8799 struct files_struct *files)
8801 while (!list_empty_careful(&ctx->inflight_list)) {
8805 inflight = io_uring_count_inflight(ctx, task, files);
8809 io_uring_try_cancel_requests(ctx, task, files);
8811 prepare_to_wait(&task->io_uring->wait, &wait,
8812 TASK_UNINTERRUPTIBLE);
8813 if (inflight == io_uring_count_inflight(ctx, task, files))
8815 finish_wait(&task->io_uring->wait, &wait);
8820 * Note that this task has used io_uring. We use it for cancelation purposes.
8822 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8824 struct io_uring_task *tctx = current->io_uring;
8825 struct io_tctx_node *node;
8828 if (unlikely(!tctx)) {
8829 ret = io_uring_alloc_task_context(current, ctx);
8832 tctx = current->io_uring;
8834 if (tctx->last != ctx) {
8835 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8838 node = kmalloc(sizeof(*node), GFP_KERNEL);
8842 node->task = current;
8844 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8851 mutex_lock(&ctx->uring_lock);
8852 list_add(&node->ctx_node, &ctx->tctx_list);
8853 mutex_unlock(&ctx->uring_lock);
8861 * Remove this io_uring_file -> task mapping.
8863 static void io_uring_del_task_file(unsigned long index)
8865 struct io_uring_task *tctx = current->io_uring;
8866 struct io_tctx_node *node;
8870 node = xa_erase(&tctx->xa, index);
8874 WARN_ON_ONCE(current != node->task);
8875 WARN_ON_ONCE(list_empty(&node->ctx_node));
8877 mutex_lock(&node->ctx->uring_lock);
8878 list_del(&node->ctx_node);
8879 mutex_unlock(&node->ctx->uring_lock);
8881 if (tctx->last == node->ctx)
8886 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8888 struct io_tctx_node *node;
8889 unsigned long index;
8891 xa_for_each(&tctx->xa, index, node)
8892 io_uring_del_task_file(index);
8894 io_wq_put_and_exit(tctx->io_wq);
8899 static s64 tctx_inflight(struct io_uring_task *tctx)
8901 return percpu_counter_sum(&tctx->inflight);
8904 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8906 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8907 struct io_ring_ctx *ctx = work->ctx;
8908 struct io_sq_data *sqd = ctx->sq_data;
8911 io_uring_cancel_sqpoll(ctx);
8912 complete(&work->completion);
8915 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8917 struct io_sq_data *sqd = ctx->sq_data;
8918 struct io_tctx_exit work = { .ctx = ctx, };
8919 struct task_struct *task;
8921 io_sq_thread_park(sqd);
8922 list_del_init(&ctx->sqd_list);
8923 io_sqd_update_thread_idle(sqd);
8926 init_completion(&work.completion);
8927 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8928 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8929 wake_up_process(task);
8931 io_sq_thread_unpark(sqd);
8934 wait_for_completion(&work.completion);
8937 void __io_uring_files_cancel(struct files_struct *files)
8939 struct io_uring_task *tctx = current->io_uring;
8940 struct io_tctx_node *node;
8941 unsigned long index;
8943 /* make sure overflow events are dropped */
8944 atomic_inc(&tctx->in_idle);
8945 xa_for_each(&tctx->xa, index, node) {
8946 struct io_ring_ctx *ctx = node->ctx;
8949 io_sqpoll_cancel_sync(ctx);
8952 io_uring_cancel_files(ctx, current, files);
8954 io_uring_try_cancel_requests(ctx, current, NULL);
8956 atomic_dec(&tctx->in_idle);
8959 io_uring_clean_tctx(tctx);
8962 /* should only be called by SQPOLL task */
8963 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8965 struct io_sq_data *sqd = ctx->sq_data;
8966 struct io_uring_task *tctx = current->io_uring;
8970 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8972 atomic_inc(&tctx->in_idle);
8974 /* read completions before cancelations */
8975 inflight = tctx_inflight(tctx);
8978 io_uring_try_cancel_requests(ctx, current, NULL);
8980 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8982 * If we've seen completions, retry without waiting. This
8983 * avoids a race where a completion comes in before we did
8984 * prepare_to_wait().
8986 if (inflight == tctx_inflight(tctx))
8988 finish_wait(&tctx->wait, &wait);
8990 atomic_dec(&tctx->in_idle);
8994 * Find any io_uring fd that this task has registered or done IO on, and cancel
8997 void __io_uring_task_cancel(void)
8999 struct io_uring_task *tctx = current->io_uring;
9003 /* make sure overflow events are dropped */
9004 atomic_inc(&tctx->in_idle);
9006 /* read completions before cancelations */
9007 inflight = tctx_inflight(tctx);
9010 __io_uring_files_cancel(NULL);
9012 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9015 * If we've seen completions, retry without waiting. This
9016 * avoids a race where a completion comes in before we did
9017 * prepare_to_wait().
9019 if (inflight == tctx_inflight(tctx))
9021 finish_wait(&tctx->wait, &wait);
9024 atomic_dec(&tctx->in_idle);
9026 io_uring_clean_tctx(tctx);
9027 /* all current's requests should be gone, we can kill tctx */
9028 __io_uring_free(current);
9031 static void *io_uring_validate_mmap_request(struct file *file,
9032 loff_t pgoff, size_t sz)
9034 struct io_ring_ctx *ctx = file->private_data;
9035 loff_t offset = pgoff << PAGE_SHIFT;
9040 case IORING_OFF_SQ_RING:
9041 case IORING_OFF_CQ_RING:
9044 case IORING_OFF_SQES:
9048 return ERR_PTR(-EINVAL);
9051 page = virt_to_head_page(ptr);
9052 if (sz > page_size(page))
9053 return ERR_PTR(-EINVAL);
9060 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9062 size_t sz = vma->vm_end - vma->vm_start;
9066 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9068 return PTR_ERR(ptr);
9070 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9071 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9074 #else /* !CONFIG_MMU */
9076 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9078 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9081 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9083 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9086 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9087 unsigned long addr, unsigned long len,
9088 unsigned long pgoff, unsigned long flags)
9092 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9094 return PTR_ERR(ptr);
9096 return (unsigned long) ptr;
9099 #endif /* !CONFIG_MMU */
9101 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9106 if (!io_sqring_full(ctx))
9108 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9110 if (!io_sqring_full(ctx))
9113 } while (!signal_pending(current));
9115 finish_wait(&ctx->sqo_sq_wait, &wait);
9119 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9120 struct __kernel_timespec __user **ts,
9121 const sigset_t __user **sig)
9123 struct io_uring_getevents_arg arg;
9126 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9127 * is just a pointer to the sigset_t.
9129 if (!(flags & IORING_ENTER_EXT_ARG)) {
9130 *sig = (const sigset_t __user *) argp;
9136 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9137 * timespec and sigset_t pointers if good.
9139 if (*argsz != sizeof(arg))
9141 if (copy_from_user(&arg, argp, sizeof(arg)))
9143 *sig = u64_to_user_ptr(arg.sigmask);
9144 *argsz = arg.sigmask_sz;
9145 *ts = u64_to_user_ptr(arg.ts);
9149 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9150 u32, min_complete, u32, flags, const void __user *, argp,
9153 struct io_ring_ctx *ctx;
9160 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9161 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9169 if (f.file->f_op != &io_uring_fops)
9173 ctx = f.file->private_data;
9174 if (!percpu_ref_tryget(&ctx->refs))
9178 if (ctx->flags & IORING_SETUP_R_DISABLED)
9182 * For SQ polling, the thread will do all submissions and completions.
9183 * Just return the requested submit count, and wake the thread if
9187 if (ctx->flags & IORING_SETUP_SQPOLL) {
9188 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9191 if (unlikely(ctx->sq_data->thread == NULL)) {
9194 if (flags & IORING_ENTER_SQ_WAKEUP)
9195 wake_up(&ctx->sq_data->wait);
9196 if (flags & IORING_ENTER_SQ_WAIT) {
9197 ret = io_sqpoll_wait_sq(ctx);
9201 submitted = to_submit;
9202 } else if (to_submit) {
9203 ret = io_uring_add_task_file(ctx);
9206 mutex_lock(&ctx->uring_lock);
9207 submitted = io_submit_sqes(ctx, to_submit);
9208 mutex_unlock(&ctx->uring_lock);
9210 if (submitted != to_submit)
9213 if (flags & IORING_ENTER_GETEVENTS) {
9214 const sigset_t __user *sig;
9215 struct __kernel_timespec __user *ts;
9217 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9221 min_complete = min(min_complete, ctx->cq_entries);
9224 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9225 * space applications don't need to do io completion events
9226 * polling again, they can rely on io_sq_thread to do polling
9227 * work, which can reduce cpu usage and uring_lock contention.
9229 if (ctx->flags & IORING_SETUP_IOPOLL &&
9230 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9231 ret = io_iopoll_check(ctx, min_complete);
9233 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9238 percpu_ref_put(&ctx->refs);
9241 return submitted ? submitted : ret;
9244 #ifdef CONFIG_PROC_FS
9245 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9246 const struct cred *cred)
9248 struct user_namespace *uns = seq_user_ns(m);
9249 struct group_info *gi;
9254 seq_printf(m, "%5d\n", id);
9255 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9256 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9257 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9258 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9259 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9260 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9261 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9262 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9263 seq_puts(m, "\n\tGroups:\t");
9264 gi = cred->group_info;
9265 for (g = 0; g < gi->ngroups; g++) {
9266 seq_put_decimal_ull(m, g ? " " : "",
9267 from_kgid_munged(uns, gi->gid[g]));
9269 seq_puts(m, "\n\tCapEff:\t");
9270 cap = cred->cap_effective;
9271 CAP_FOR_EACH_U32(__capi)
9272 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9277 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9279 struct io_sq_data *sq = NULL;
9284 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9285 * since fdinfo case grabs it in the opposite direction of normal use
9286 * cases. If we fail to get the lock, we just don't iterate any
9287 * structures that could be going away outside the io_uring mutex.
9289 has_lock = mutex_trylock(&ctx->uring_lock);
9291 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9297 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9298 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9299 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9300 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9301 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9304 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9306 seq_printf(m, "%5u: <none>\n", i);
9308 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9309 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9310 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9312 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9313 (unsigned int) buf->len);
9315 if (has_lock && !xa_empty(&ctx->personalities)) {
9316 unsigned long index;
9317 const struct cred *cred;
9319 seq_printf(m, "Personalities:\n");
9320 xa_for_each(&ctx->personalities, index, cred)
9321 io_uring_show_cred(m, index, cred);
9323 seq_printf(m, "PollList:\n");
9324 spin_lock_irq(&ctx->completion_lock);
9325 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9326 struct hlist_head *list = &ctx->cancel_hash[i];
9327 struct io_kiocb *req;
9329 hlist_for_each_entry(req, list, hash_node)
9330 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9331 req->task->task_works != NULL);
9333 spin_unlock_irq(&ctx->completion_lock);
9335 mutex_unlock(&ctx->uring_lock);
9338 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9340 struct io_ring_ctx *ctx = f->private_data;
9342 if (percpu_ref_tryget(&ctx->refs)) {
9343 __io_uring_show_fdinfo(ctx, m);
9344 percpu_ref_put(&ctx->refs);
9349 static const struct file_operations io_uring_fops = {
9350 .release = io_uring_release,
9351 .mmap = io_uring_mmap,
9353 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9354 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9356 .poll = io_uring_poll,
9357 .fasync = io_uring_fasync,
9358 #ifdef CONFIG_PROC_FS
9359 .show_fdinfo = io_uring_show_fdinfo,
9363 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9364 struct io_uring_params *p)
9366 struct io_rings *rings;
9367 size_t size, sq_array_offset;
9369 /* make sure these are sane, as we already accounted them */
9370 ctx->sq_entries = p->sq_entries;
9371 ctx->cq_entries = p->cq_entries;
9373 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9374 if (size == SIZE_MAX)
9377 rings = io_mem_alloc(size);
9382 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9383 rings->sq_ring_mask = p->sq_entries - 1;
9384 rings->cq_ring_mask = p->cq_entries - 1;
9385 rings->sq_ring_entries = p->sq_entries;
9386 rings->cq_ring_entries = p->cq_entries;
9387 ctx->sq_mask = rings->sq_ring_mask;
9388 ctx->cq_mask = rings->cq_ring_mask;
9390 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9391 if (size == SIZE_MAX) {
9392 io_mem_free(ctx->rings);
9397 ctx->sq_sqes = io_mem_alloc(size);
9398 if (!ctx->sq_sqes) {
9399 io_mem_free(ctx->rings);
9407 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9411 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9415 ret = io_uring_add_task_file(ctx);
9420 fd_install(fd, file);
9425 * Allocate an anonymous fd, this is what constitutes the application
9426 * visible backing of an io_uring instance. The application mmaps this
9427 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9428 * we have to tie this fd to a socket for file garbage collection purposes.
9430 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9433 #if defined(CONFIG_UNIX)
9436 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9439 return ERR_PTR(ret);
9442 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9443 O_RDWR | O_CLOEXEC);
9444 #if defined(CONFIG_UNIX)
9446 sock_release(ctx->ring_sock);
9447 ctx->ring_sock = NULL;
9449 ctx->ring_sock->file = file;
9455 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9456 struct io_uring_params __user *params)
9458 struct io_ring_ctx *ctx;
9464 if (entries > IORING_MAX_ENTRIES) {
9465 if (!(p->flags & IORING_SETUP_CLAMP))
9467 entries = IORING_MAX_ENTRIES;
9471 * Use twice as many entries for the CQ ring. It's possible for the
9472 * application to drive a higher depth than the size of the SQ ring,
9473 * since the sqes are only used at submission time. This allows for
9474 * some flexibility in overcommitting a bit. If the application has
9475 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9476 * of CQ ring entries manually.
9478 p->sq_entries = roundup_pow_of_two(entries);
9479 if (p->flags & IORING_SETUP_CQSIZE) {
9481 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9482 * to a power-of-two, if it isn't already. We do NOT impose
9483 * any cq vs sq ring sizing.
9487 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9488 if (!(p->flags & IORING_SETUP_CLAMP))
9490 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9492 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9493 if (p->cq_entries < p->sq_entries)
9496 p->cq_entries = 2 * p->sq_entries;
9499 ctx = io_ring_ctx_alloc(p);
9502 ctx->compat = in_compat_syscall();
9503 if (!capable(CAP_IPC_LOCK))
9504 ctx->user = get_uid(current_user());
9507 * This is just grabbed for accounting purposes. When a process exits,
9508 * the mm is exited and dropped before the files, hence we need to hang
9509 * on to this mm purely for the purposes of being able to unaccount
9510 * memory (locked/pinned vm). It's not used for anything else.
9512 mmgrab(current->mm);
9513 ctx->mm_account = current->mm;
9515 ret = io_allocate_scq_urings(ctx, p);
9519 ret = io_sq_offload_create(ctx, p);
9523 memset(&p->sq_off, 0, sizeof(p->sq_off));
9524 p->sq_off.head = offsetof(struct io_rings, sq.head);
9525 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9526 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9527 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9528 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9529 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9530 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9532 memset(&p->cq_off, 0, sizeof(p->cq_off));
9533 p->cq_off.head = offsetof(struct io_rings, cq.head);
9534 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9535 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9536 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9537 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9538 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9539 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9541 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9542 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9543 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9544 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9545 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9547 if (copy_to_user(params, p, sizeof(*p))) {
9552 file = io_uring_get_file(ctx);
9554 ret = PTR_ERR(file);
9559 * Install ring fd as the very last thing, so we don't risk someone
9560 * having closed it before we finish setup
9562 ret = io_uring_install_fd(ctx, file);
9564 /* fput will clean it up */
9569 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9572 io_ring_ctx_wait_and_kill(ctx);
9577 * Sets up an aio uring context, and returns the fd. Applications asks for a
9578 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9579 * params structure passed in.
9581 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9583 struct io_uring_params p;
9586 if (copy_from_user(&p, params, sizeof(p)))
9588 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9593 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9594 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9595 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9596 IORING_SETUP_R_DISABLED))
9599 return io_uring_create(entries, &p, params);
9602 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9603 struct io_uring_params __user *, params)
9605 return io_uring_setup(entries, params);
9608 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9610 struct io_uring_probe *p;
9614 size = struct_size(p, ops, nr_args);
9615 if (size == SIZE_MAX)
9617 p = kzalloc(size, GFP_KERNEL);
9622 if (copy_from_user(p, arg, size))
9625 if (memchr_inv(p, 0, size))
9628 p->last_op = IORING_OP_LAST - 1;
9629 if (nr_args > IORING_OP_LAST)
9630 nr_args = IORING_OP_LAST;
9632 for (i = 0; i < nr_args; i++) {
9634 if (!io_op_defs[i].not_supported)
9635 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9640 if (copy_to_user(arg, p, size))
9647 static int io_register_personality(struct io_ring_ctx *ctx)
9649 const struct cred *creds;
9653 creds = get_current_cred();
9655 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9656 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9663 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9664 unsigned int nr_args)
9666 struct io_uring_restriction *res;
9670 /* Restrictions allowed only if rings started disabled */
9671 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9674 /* We allow only a single restrictions registration */
9675 if (ctx->restrictions.registered)
9678 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9681 size = array_size(nr_args, sizeof(*res));
9682 if (size == SIZE_MAX)
9685 res = memdup_user(arg, size);
9687 return PTR_ERR(res);
9691 for (i = 0; i < nr_args; i++) {
9692 switch (res[i].opcode) {
9693 case IORING_RESTRICTION_REGISTER_OP:
9694 if (res[i].register_op >= IORING_REGISTER_LAST) {
9699 __set_bit(res[i].register_op,
9700 ctx->restrictions.register_op);
9702 case IORING_RESTRICTION_SQE_OP:
9703 if (res[i].sqe_op >= IORING_OP_LAST) {
9708 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9710 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9711 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9713 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9714 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9723 /* Reset all restrictions if an error happened */
9725 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9727 ctx->restrictions.registered = true;
9733 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9735 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9738 if (ctx->restrictions.registered)
9739 ctx->restricted = 1;
9741 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9742 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9743 wake_up(&ctx->sq_data->wait);
9747 static bool io_register_op_must_quiesce(int op)
9750 case IORING_UNREGISTER_FILES:
9751 case IORING_REGISTER_FILES_UPDATE:
9752 case IORING_REGISTER_PROBE:
9753 case IORING_REGISTER_PERSONALITY:
9754 case IORING_UNREGISTER_PERSONALITY:
9761 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9762 void __user *arg, unsigned nr_args)
9763 __releases(ctx->uring_lock)
9764 __acquires(ctx->uring_lock)
9769 * We're inside the ring mutex, if the ref is already dying, then
9770 * someone else killed the ctx or is already going through
9771 * io_uring_register().
9773 if (percpu_ref_is_dying(&ctx->refs))
9776 if (io_register_op_must_quiesce(opcode)) {
9777 percpu_ref_kill(&ctx->refs);
9780 * Drop uring mutex before waiting for references to exit. If
9781 * another thread is currently inside io_uring_enter() it might
9782 * need to grab the uring_lock to make progress. If we hold it
9783 * here across the drain wait, then we can deadlock. It's safe
9784 * to drop the mutex here, since no new references will come in
9785 * after we've killed the percpu ref.
9787 mutex_unlock(&ctx->uring_lock);
9789 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9792 ret = io_run_task_work_sig();
9797 mutex_lock(&ctx->uring_lock);
9800 percpu_ref_resurrect(&ctx->refs);
9805 if (ctx->restricted) {
9806 if (opcode >= IORING_REGISTER_LAST) {
9811 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9818 case IORING_REGISTER_BUFFERS:
9819 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9821 case IORING_UNREGISTER_BUFFERS:
9825 ret = io_sqe_buffers_unregister(ctx);
9827 case IORING_REGISTER_FILES:
9828 ret = io_sqe_files_register(ctx, arg, nr_args);
9830 case IORING_UNREGISTER_FILES:
9834 ret = io_sqe_files_unregister(ctx);
9836 case IORING_REGISTER_FILES_UPDATE:
9837 ret = io_sqe_files_update(ctx, arg, nr_args);
9839 case IORING_REGISTER_EVENTFD:
9840 case IORING_REGISTER_EVENTFD_ASYNC:
9844 ret = io_eventfd_register(ctx, arg);
9847 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9848 ctx->eventfd_async = 1;
9850 ctx->eventfd_async = 0;
9852 case IORING_UNREGISTER_EVENTFD:
9856 ret = io_eventfd_unregister(ctx);
9858 case IORING_REGISTER_PROBE:
9860 if (!arg || nr_args > 256)
9862 ret = io_probe(ctx, arg, nr_args);
9864 case IORING_REGISTER_PERSONALITY:
9868 ret = io_register_personality(ctx);
9870 case IORING_UNREGISTER_PERSONALITY:
9874 ret = io_unregister_personality(ctx, nr_args);
9876 case IORING_REGISTER_ENABLE_RINGS:
9880 ret = io_register_enable_rings(ctx);
9882 case IORING_REGISTER_RESTRICTIONS:
9883 ret = io_register_restrictions(ctx, arg, nr_args);
9891 if (io_register_op_must_quiesce(opcode)) {
9892 /* bring the ctx back to life */
9893 percpu_ref_reinit(&ctx->refs);
9895 reinit_completion(&ctx->ref_comp);
9900 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9901 void __user *, arg, unsigned int, nr_args)
9903 struct io_ring_ctx *ctx;
9912 if (f.file->f_op != &io_uring_fops)
9915 ctx = f.file->private_data;
9919 mutex_lock(&ctx->uring_lock);
9920 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9921 mutex_unlock(&ctx->uring_lock);
9922 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9923 ctx->cq_ev_fd != NULL, ret);
9929 static int __init io_uring_init(void)
9931 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9932 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9933 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9936 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9937 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9938 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9939 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9940 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9941 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9942 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9943 BUILD_BUG_SQE_ELEM(8, __u64, off);
9944 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9945 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9946 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9947 BUILD_BUG_SQE_ELEM(24, __u32, len);
9948 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9949 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9950 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9951 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9952 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9953 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9954 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9955 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9956 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9957 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9958 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9959 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9960 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9961 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9962 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9963 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9964 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9965 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9966 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9968 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9969 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9970 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9974 __initcall(io_uring_init);