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)
1101 static inline void req_set_fail_links(struct io_kiocb *req)
1103 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1104 req->flags |= REQ_F_FAIL_LINK;
1107 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1109 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1111 complete(&ctx->ref_comp);
1114 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1116 return !req->timeout.off;
1119 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1121 struct io_ring_ctx *ctx;
1124 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1129 * Use 5 bits less than the max cq entries, that should give us around
1130 * 32 entries per hash list if totally full and uniformly spread.
1132 hash_bits = ilog2(p->cq_entries);
1136 ctx->cancel_hash_bits = hash_bits;
1137 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1139 if (!ctx->cancel_hash)
1141 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1143 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1144 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1147 ctx->flags = p->flags;
1148 init_waitqueue_head(&ctx->sqo_sq_wait);
1149 INIT_LIST_HEAD(&ctx->sqd_list);
1150 init_waitqueue_head(&ctx->cq_wait);
1151 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1152 init_completion(&ctx->ref_comp);
1153 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1154 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1155 mutex_init(&ctx->uring_lock);
1156 init_waitqueue_head(&ctx->wait);
1157 spin_lock_init(&ctx->completion_lock);
1158 INIT_LIST_HEAD(&ctx->iopoll_list);
1159 INIT_LIST_HEAD(&ctx->defer_list);
1160 INIT_LIST_HEAD(&ctx->timeout_list);
1161 spin_lock_init(&ctx->inflight_lock);
1162 INIT_LIST_HEAD(&ctx->inflight_list);
1163 spin_lock_init(&ctx->rsrc_ref_lock);
1164 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1165 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1166 init_llist_head(&ctx->rsrc_put_llist);
1167 INIT_LIST_HEAD(&ctx->tctx_list);
1168 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1169 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1172 kfree(ctx->cancel_hash);
1177 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1179 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1180 struct io_ring_ctx *ctx = req->ctx;
1182 return seq != ctx->cached_cq_tail
1183 + READ_ONCE(ctx->cached_cq_overflow);
1189 static void io_req_track_inflight(struct io_kiocb *req)
1191 struct io_ring_ctx *ctx = req->ctx;
1193 if (!(req->flags & REQ_F_INFLIGHT)) {
1194 req->flags |= REQ_F_INFLIGHT;
1196 spin_lock_irq(&ctx->inflight_lock);
1197 list_add(&req->inflight_entry, &ctx->inflight_list);
1198 spin_unlock_irq(&ctx->inflight_lock);
1202 static void io_prep_async_work(struct io_kiocb *req)
1204 const struct io_op_def *def = &io_op_defs[req->opcode];
1205 struct io_ring_ctx *ctx = req->ctx;
1207 if (!req->work.creds)
1208 req->work.creds = get_current_cred();
1210 if (req->flags & REQ_F_FORCE_ASYNC)
1211 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1213 if (req->flags & REQ_F_ISREG) {
1214 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1215 io_wq_hash_work(&req->work, file_inode(req->file));
1217 if (def->unbound_nonreg_file)
1218 req->work.flags |= IO_WQ_WORK_UNBOUND;
1222 static void io_prep_async_link(struct io_kiocb *req)
1224 struct io_kiocb *cur;
1226 io_for_each_link(cur, req)
1227 io_prep_async_work(cur);
1230 static void io_queue_async_work(struct io_kiocb *req)
1232 struct io_ring_ctx *ctx = req->ctx;
1233 struct io_kiocb *link = io_prep_linked_timeout(req);
1234 struct io_uring_task *tctx = req->task->io_uring;
1237 BUG_ON(!tctx->io_wq);
1239 /* init ->work of the whole link before punting */
1240 io_prep_async_link(req);
1241 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1242 &req->work, req->flags);
1243 io_wq_enqueue(tctx->io_wq, &req->work);
1245 io_queue_linked_timeout(link);
1248 static void io_kill_timeout(struct io_kiocb *req, int status)
1250 struct io_timeout_data *io = req->async_data;
1253 ret = hrtimer_try_to_cancel(&io->timer);
1255 atomic_set(&req->ctx->cq_timeouts,
1256 atomic_read(&req->ctx->cq_timeouts) + 1);
1257 list_del_init(&req->timeout.list);
1258 io_cqring_fill_event(req, status);
1259 io_put_req_deferred(req, 1);
1263 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1266 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1267 struct io_defer_entry, list);
1269 if (req_need_defer(de->req, de->seq))
1271 list_del_init(&de->list);
1272 io_req_task_queue(de->req);
1274 } while (!list_empty(&ctx->defer_list));
1277 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1281 if (list_empty(&ctx->timeout_list))
1284 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1287 u32 events_needed, events_got;
1288 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1289 struct io_kiocb, timeout.list);
1291 if (io_is_timeout_noseq(req))
1295 * Since seq can easily wrap around over time, subtract
1296 * the last seq at which timeouts were flushed before comparing.
1297 * Assuming not more than 2^31-1 events have happened since,
1298 * these subtractions won't have wrapped, so we can check if
1299 * target is in [last_seq, current_seq] by comparing the two.
1301 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1302 events_got = seq - ctx->cq_last_tm_flush;
1303 if (events_got < events_needed)
1306 list_del_init(&req->timeout.list);
1307 io_kill_timeout(req, 0);
1308 } while (!list_empty(&ctx->timeout_list));
1310 ctx->cq_last_tm_flush = seq;
1313 static void io_commit_cqring(struct io_ring_ctx *ctx)
1315 io_flush_timeouts(ctx);
1317 /* order cqe stores with ring update */
1318 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1320 if (unlikely(!list_empty(&ctx->defer_list)))
1321 __io_queue_deferred(ctx);
1324 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1326 struct io_rings *r = ctx->rings;
1328 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1331 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1333 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1336 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1338 struct io_rings *rings = ctx->rings;
1342 * writes to the cq entry need to come after reading head; the
1343 * control dependency is enough as we're using WRITE_ONCE to
1346 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1349 tail = ctx->cached_cq_tail++;
1350 return &rings->cqes[tail & ctx->cq_mask];
1353 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1357 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1359 if (!ctx->eventfd_async)
1361 return io_wq_current_is_worker();
1364 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1366 /* see waitqueue_active() comment */
1369 if (waitqueue_active(&ctx->wait))
1370 wake_up(&ctx->wait);
1371 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1372 wake_up(&ctx->sq_data->wait);
1373 if (io_should_trigger_evfd(ctx))
1374 eventfd_signal(ctx->cq_ev_fd, 1);
1375 if (waitqueue_active(&ctx->cq_wait)) {
1376 wake_up_interruptible(&ctx->cq_wait);
1377 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1381 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1383 /* see waitqueue_active() comment */
1386 if (ctx->flags & IORING_SETUP_SQPOLL) {
1387 if (waitqueue_active(&ctx->wait))
1388 wake_up(&ctx->wait);
1390 if (io_should_trigger_evfd(ctx))
1391 eventfd_signal(ctx->cq_ev_fd, 1);
1392 if (waitqueue_active(&ctx->cq_wait)) {
1393 wake_up_interruptible(&ctx->cq_wait);
1394 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1398 /* Returns true if there are no backlogged entries after the flush */
1399 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1400 struct task_struct *tsk,
1401 struct files_struct *files)
1403 struct io_rings *rings = ctx->rings;
1404 struct io_kiocb *req, *tmp;
1405 struct io_uring_cqe *cqe;
1406 unsigned long flags;
1407 bool all_flushed, posted;
1410 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1414 spin_lock_irqsave(&ctx->completion_lock, flags);
1415 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1416 if (!io_match_task(req, tsk, files))
1419 cqe = io_get_cqring(ctx);
1423 list_move(&req->compl.list, &list);
1425 WRITE_ONCE(cqe->user_data, req->user_data);
1426 WRITE_ONCE(cqe->res, req->result);
1427 WRITE_ONCE(cqe->flags, req->compl.cflags);
1429 ctx->cached_cq_overflow++;
1430 WRITE_ONCE(ctx->rings->cq_overflow,
1431 ctx->cached_cq_overflow);
1436 all_flushed = list_empty(&ctx->cq_overflow_list);
1438 clear_bit(0, &ctx->sq_check_overflow);
1439 clear_bit(0, &ctx->cq_check_overflow);
1440 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1444 io_commit_cqring(ctx);
1445 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1447 io_cqring_ev_posted(ctx);
1449 while (!list_empty(&list)) {
1450 req = list_first_entry(&list, struct io_kiocb, compl.list);
1451 list_del(&req->compl.list);
1458 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1459 struct task_struct *tsk,
1460 struct files_struct *files)
1464 if (test_bit(0, &ctx->cq_check_overflow)) {
1465 /* iopoll syncs against uring_lock, not completion_lock */
1466 if (ctx->flags & IORING_SETUP_IOPOLL)
1467 mutex_lock(&ctx->uring_lock);
1468 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1469 if (ctx->flags & IORING_SETUP_IOPOLL)
1470 mutex_unlock(&ctx->uring_lock);
1476 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1478 struct io_ring_ctx *ctx = req->ctx;
1479 struct io_uring_cqe *cqe;
1481 trace_io_uring_complete(ctx, req->user_data, res);
1484 * If we can't get a cq entry, userspace overflowed the
1485 * submission (by quite a lot). Increment the overflow count in
1488 cqe = io_get_cqring(ctx);
1490 WRITE_ONCE(cqe->user_data, req->user_data);
1491 WRITE_ONCE(cqe->res, res);
1492 WRITE_ONCE(cqe->flags, cflags);
1493 } else if (ctx->cq_overflow_flushed ||
1494 atomic_read(&req->task->io_uring->in_idle)) {
1496 * If we're in ring overflow flush mode, or in task cancel mode,
1497 * then we cannot store the request for later flushing, we need
1498 * to drop it on the floor.
1500 ctx->cached_cq_overflow++;
1501 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1503 if (list_empty(&ctx->cq_overflow_list)) {
1504 set_bit(0, &ctx->sq_check_overflow);
1505 set_bit(0, &ctx->cq_check_overflow);
1506 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1510 req->compl.cflags = cflags;
1511 refcount_inc(&req->refs);
1512 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1516 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1518 __io_cqring_fill_event(req, res, 0);
1521 static void io_req_complete_post(struct io_kiocb *req, long res,
1522 unsigned int cflags)
1524 struct io_ring_ctx *ctx = req->ctx;
1525 unsigned long flags;
1527 spin_lock_irqsave(&ctx->completion_lock, flags);
1528 __io_cqring_fill_event(req, res, cflags);
1530 * If we're the last reference to this request, add to our locked
1533 if (refcount_dec_and_test(&req->refs)) {
1534 struct io_comp_state *cs = &ctx->submit_state.comp;
1536 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1537 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1538 io_disarm_next(req);
1540 io_req_task_queue(req->link);
1544 io_dismantle_req(req);
1545 io_put_task(req->task, 1);
1546 list_add(&req->compl.list, &cs->locked_free_list);
1547 cs->locked_free_nr++;
1549 if (!percpu_ref_tryget(&ctx->refs))
1552 io_commit_cqring(ctx);
1553 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1556 io_cqring_ev_posted(ctx);
1557 percpu_ref_put(&ctx->refs);
1561 static void io_req_complete_state(struct io_kiocb *req, long res,
1562 unsigned int cflags)
1566 req->compl.cflags = cflags;
1567 req->flags |= REQ_F_COMPLETE_INLINE;
1570 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1571 long res, unsigned cflags)
1573 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1574 io_req_complete_state(req, res, cflags);
1576 io_req_complete_post(req, res, cflags);
1579 static inline void io_req_complete(struct io_kiocb *req, long res)
1581 __io_req_complete(req, 0, res, 0);
1584 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1586 struct io_submit_state *state = &ctx->submit_state;
1587 struct io_comp_state *cs = &state->comp;
1588 struct io_kiocb *req = NULL;
1591 * If we have more than a batch's worth of requests in our IRQ side
1592 * locked cache, grab the lock and move them over to our submission
1595 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1596 spin_lock_irq(&ctx->completion_lock);
1597 list_splice_init(&cs->locked_free_list, &cs->free_list);
1598 cs->locked_free_nr = 0;
1599 spin_unlock_irq(&ctx->completion_lock);
1602 while (!list_empty(&cs->free_list)) {
1603 req = list_first_entry(&cs->free_list, struct io_kiocb,
1605 list_del(&req->compl.list);
1606 state->reqs[state->free_reqs++] = req;
1607 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1614 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1616 struct io_submit_state *state = &ctx->submit_state;
1618 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1620 if (!state->free_reqs) {
1621 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1624 if (io_flush_cached_reqs(ctx))
1627 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1631 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1632 * retry single alloc to be on the safe side.
1634 if (unlikely(ret <= 0)) {
1635 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1636 if (!state->reqs[0])
1640 state->free_reqs = ret;
1644 return state->reqs[state->free_reqs];
1647 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1654 static void io_dismantle_req(struct io_kiocb *req)
1658 if (req->async_data)
1659 kfree(req->async_data);
1661 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1662 if (req->fixed_rsrc_refs)
1663 percpu_ref_put(req->fixed_rsrc_refs);
1664 if (req->work.creds) {
1665 put_cred(req->work.creds);
1666 req->work.creds = NULL;
1669 if (req->flags & REQ_F_INFLIGHT) {
1670 struct io_ring_ctx *ctx = req->ctx;
1671 unsigned long flags;
1673 spin_lock_irqsave(&ctx->inflight_lock, flags);
1674 list_del(&req->inflight_entry);
1675 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1676 req->flags &= ~REQ_F_INFLIGHT;
1680 /* must to be called somewhat shortly after putting a request */
1681 static inline void io_put_task(struct task_struct *task, int nr)
1683 struct io_uring_task *tctx = task->io_uring;
1685 percpu_counter_sub(&tctx->inflight, nr);
1686 if (unlikely(atomic_read(&tctx->in_idle)))
1687 wake_up(&tctx->wait);
1688 put_task_struct_many(task, nr);
1691 static void __io_free_req(struct io_kiocb *req)
1693 struct io_ring_ctx *ctx = req->ctx;
1695 io_dismantle_req(req);
1696 io_put_task(req->task, 1);
1698 kmem_cache_free(req_cachep, req);
1699 percpu_ref_put(&ctx->refs);
1702 static inline void io_remove_next_linked(struct io_kiocb *req)
1704 struct io_kiocb *nxt = req->link;
1706 req->link = nxt->link;
1710 static bool io_kill_linked_timeout(struct io_kiocb *req)
1711 __must_hold(&req->ctx->completion_lock)
1713 struct io_kiocb *link = req->link;
1714 bool cancelled = false;
1717 * Can happen if a linked timeout fired and link had been like
1718 * req -> link t-out -> link t-out [-> ...]
1720 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1721 struct io_timeout_data *io = link->async_data;
1724 io_remove_next_linked(req);
1725 link->timeout.head = NULL;
1726 ret = hrtimer_try_to_cancel(&io->timer);
1728 io_cqring_fill_event(link, -ECANCELED);
1729 io_put_req_deferred(link, 1);
1733 req->flags &= ~REQ_F_LINK_TIMEOUT;
1737 static void io_fail_links(struct io_kiocb *req)
1738 __must_hold(&req->ctx->completion_lock)
1740 struct io_kiocb *nxt, *link = req->link;
1747 trace_io_uring_fail_link(req, link);
1748 io_cqring_fill_event(link, -ECANCELED);
1749 io_put_req_deferred(link, 2);
1754 static bool io_disarm_next(struct io_kiocb *req)
1755 __must_hold(&req->ctx->completion_lock)
1757 bool posted = false;
1759 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1760 posted = io_kill_linked_timeout(req);
1761 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1762 posted |= (req->link != NULL);
1768 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1770 struct io_kiocb *nxt;
1773 * If LINK is set, we have dependent requests in this chain. If we
1774 * didn't fail this request, queue the first one up, moving any other
1775 * dependencies to the next request. In case of failure, fail the rest
1778 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1779 struct io_ring_ctx *ctx = req->ctx;
1780 unsigned long flags;
1783 spin_lock_irqsave(&ctx->completion_lock, flags);
1784 posted = io_disarm_next(req);
1786 io_commit_cqring(req->ctx);
1787 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1789 io_cqring_ev_posted(ctx);
1796 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1798 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1800 return __io_req_find_next(req);
1803 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1807 if (ctx->submit_state.comp.nr) {
1808 mutex_lock(&ctx->uring_lock);
1809 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1810 mutex_unlock(&ctx->uring_lock);
1812 percpu_ref_put(&ctx->refs);
1815 static bool __tctx_task_work(struct io_uring_task *tctx)
1817 struct io_ring_ctx *ctx = NULL;
1818 struct io_wq_work_list list;
1819 struct io_wq_work_node *node;
1821 if (wq_list_empty(&tctx->task_list))
1824 spin_lock_irq(&tctx->task_lock);
1825 list = tctx->task_list;
1826 INIT_WQ_LIST(&tctx->task_list);
1827 spin_unlock_irq(&tctx->task_lock);
1831 struct io_wq_work_node *next = node->next;
1832 struct io_kiocb *req;
1834 req = container_of(node, struct io_kiocb, io_task_work.node);
1835 if (req->ctx != ctx) {
1836 ctx_flush_and_put(ctx);
1838 percpu_ref_get(&ctx->refs);
1841 req->task_work.func(&req->task_work);
1845 ctx_flush_and_put(ctx);
1846 return list.first != NULL;
1849 static void tctx_task_work(struct callback_head *cb)
1851 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1853 clear_bit(0, &tctx->task_state);
1855 while (__tctx_task_work(tctx))
1859 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1860 enum task_work_notify_mode notify)
1862 struct io_uring_task *tctx = tsk->io_uring;
1863 struct io_wq_work_node *node, *prev;
1864 unsigned long flags;
1867 WARN_ON_ONCE(!tctx);
1869 spin_lock_irqsave(&tctx->task_lock, flags);
1870 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1871 spin_unlock_irqrestore(&tctx->task_lock, flags);
1873 /* task_work already pending, we're done */
1874 if (test_bit(0, &tctx->task_state) ||
1875 test_and_set_bit(0, &tctx->task_state))
1878 if (!task_work_add(tsk, &tctx->task_work, notify))
1882 * Slow path - we failed, find and delete work. if the work is not
1883 * in the list, it got run and we're fine.
1886 spin_lock_irqsave(&tctx->task_lock, flags);
1887 wq_list_for_each(node, prev, &tctx->task_list) {
1888 if (&req->io_task_work.node == node) {
1889 wq_list_del(&tctx->task_list, node, prev);
1894 spin_unlock_irqrestore(&tctx->task_lock, flags);
1895 clear_bit(0, &tctx->task_state);
1899 static int io_req_task_work_add(struct io_kiocb *req)
1901 struct task_struct *tsk = req->task;
1902 struct io_ring_ctx *ctx = req->ctx;
1903 enum task_work_notify_mode notify;
1906 if (tsk->flags & PF_EXITING)
1910 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1911 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1912 * processing task_work. There's no reliable way to tell if TWA_RESUME
1916 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1917 notify = TWA_SIGNAL;
1919 ret = io_task_work_add(tsk, req, notify);
1921 wake_up_process(tsk);
1926 static bool io_run_task_work_head(struct callback_head **work_head)
1928 struct callback_head *work, *next;
1929 bool executed = false;
1932 work = xchg(work_head, NULL);
1948 static void io_task_work_add_head(struct callback_head **work_head,
1949 struct callback_head *task_work)
1951 struct callback_head *head;
1954 head = READ_ONCE(*work_head);
1955 task_work->next = head;
1956 } while (cmpxchg(work_head, head, task_work) != head);
1959 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1960 task_work_func_t cb)
1962 init_task_work(&req->task_work, cb);
1963 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1966 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1968 struct io_ring_ctx *ctx = req->ctx;
1970 spin_lock_irq(&ctx->completion_lock);
1971 io_cqring_fill_event(req, error);
1972 io_commit_cqring(ctx);
1973 spin_unlock_irq(&ctx->completion_lock);
1975 io_cqring_ev_posted(ctx);
1976 req_set_fail_links(req);
1977 io_double_put_req(req);
1980 static void io_req_task_cancel(struct callback_head *cb)
1982 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1983 struct io_ring_ctx *ctx = req->ctx;
1985 mutex_lock(&ctx->uring_lock);
1986 __io_req_task_cancel(req, req->result);
1987 mutex_unlock(&ctx->uring_lock);
1988 percpu_ref_put(&ctx->refs);
1991 static void __io_req_task_submit(struct io_kiocb *req)
1993 struct io_ring_ctx *ctx = req->ctx;
1995 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1996 mutex_lock(&ctx->uring_lock);
1997 if (!(current->flags & PF_EXITING) && !current->in_execve)
1998 __io_queue_sqe(req);
2000 __io_req_task_cancel(req, -EFAULT);
2001 mutex_unlock(&ctx->uring_lock);
2004 static void io_req_task_submit(struct callback_head *cb)
2006 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2008 __io_req_task_submit(req);
2011 static void io_req_task_queue(struct io_kiocb *req)
2015 req->task_work.func = io_req_task_submit;
2016 ret = io_req_task_work_add(req);
2017 if (unlikely(ret)) {
2018 req->result = -ECANCELED;
2019 percpu_ref_get(&req->ctx->refs);
2020 io_req_task_work_add_fallback(req, io_req_task_cancel);
2024 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2026 percpu_ref_get(&req->ctx->refs);
2028 req->task_work.func = io_req_task_cancel;
2030 if (unlikely(io_req_task_work_add(req)))
2031 io_req_task_work_add_fallback(req, io_req_task_cancel);
2034 static inline void io_queue_next(struct io_kiocb *req)
2036 struct io_kiocb *nxt = io_req_find_next(req);
2039 io_req_task_queue(nxt);
2042 static void io_free_req(struct io_kiocb *req)
2049 struct task_struct *task;
2054 static inline void io_init_req_batch(struct req_batch *rb)
2061 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2062 struct req_batch *rb)
2065 io_put_task(rb->task, rb->task_refs);
2067 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2070 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2071 struct io_submit_state *state)
2075 if (req->task != rb->task) {
2077 io_put_task(rb->task, rb->task_refs);
2078 rb->task = req->task;
2084 io_dismantle_req(req);
2085 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2086 state->reqs[state->free_reqs++] = req;
2088 list_add(&req->compl.list, &state->comp.free_list);
2091 static void io_submit_flush_completions(struct io_comp_state *cs,
2092 struct io_ring_ctx *ctx)
2095 struct io_kiocb *req;
2096 struct req_batch rb;
2098 io_init_req_batch(&rb);
2099 spin_lock_irq(&ctx->completion_lock);
2100 for (i = 0; i < nr; i++) {
2102 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2104 io_commit_cqring(ctx);
2105 spin_unlock_irq(&ctx->completion_lock);
2107 io_cqring_ev_posted(ctx);
2108 for (i = 0; i < nr; i++) {
2111 /* submission and completion refs */
2112 if (refcount_sub_and_test(2, &req->refs))
2113 io_req_free_batch(&rb, req, &ctx->submit_state);
2116 io_req_free_batch_finish(ctx, &rb);
2121 * Drop reference to request, return next in chain (if there is one) if this
2122 * was the last reference to this request.
2124 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2126 struct io_kiocb *nxt = NULL;
2128 if (refcount_dec_and_test(&req->refs)) {
2129 nxt = io_req_find_next(req);
2135 static void io_put_req(struct io_kiocb *req)
2137 if (refcount_dec_and_test(&req->refs))
2141 static void io_put_req_deferred_cb(struct callback_head *cb)
2143 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2148 static void io_free_req_deferred(struct io_kiocb *req)
2152 req->task_work.func = io_put_req_deferred_cb;
2153 ret = io_req_task_work_add(req);
2155 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2158 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2160 if (refcount_sub_and_test(refs, &req->refs))
2161 io_free_req_deferred(req);
2164 static void io_double_put_req(struct io_kiocb *req)
2166 /* drop both submit and complete references */
2167 if (refcount_sub_and_test(2, &req->refs))
2171 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2173 /* See comment at the top of this file */
2175 return __io_cqring_events(ctx);
2178 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2180 struct io_rings *rings = ctx->rings;
2182 /* make sure SQ entry isn't read before tail */
2183 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2186 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2188 unsigned int cflags;
2190 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2191 cflags |= IORING_CQE_F_BUFFER;
2192 req->flags &= ~REQ_F_BUFFER_SELECTED;
2197 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2199 struct io_buffer *kbuf;
2201 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2202 return io_put_kbuf(req, kbuf);
2205 static inline bool io_run_task_work(void)
2208 * Not safe to run on exiting task, and the task_work handling will
2209 * not add work to such a task.
2211 if (unlikely(current->flags & PF_EXITING))
2213 if (current->task_works) {
2214 __set_current_state(TASK_RUNNING);
2223 * Find and free completed poll iocbs
2225 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2226 struct list_head *done)
2228 struct req_batch rb;
2229 struct io_kiocb *req;
2231 /* order with ->result store in io_complete_rw_iopoll() */
2234 io_init_req_batch(&rb);
2235 while (!list_empty(done)) {
2238 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2239 list_del(&req->inflight_entry);
2241 if (READ_ONCE(req->result) == -EAGAIN) {
2242 req->iopoll_completed = 0;
2243 if (io_rw_reissue(req))
2247 if (req->flags & REQ_F_BUFFER_SELECTED)
2248 cflags = io_put_rw_kbuf(req);
2250 __io_cqring_fill_event(req, req->result, cflags);
2253 if (refcount_dec_and_test(&req->refs))
2254 io_req_free_batch(&rb, req, &ctx->submit_state);
2257 io_commit_cqring(ctx);
2258 io_cqring_ev_posted_iopoll(ctx);
2259 io_req_free_batch_finish(ctx, &rb);
2262 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2265 struct io_kiocb *req, *tmp;
2271 * Only spin for completions if we don't have multiple devices hanging
2272 * off our complete list, and we're under the requested amount.
2274 spin = !ctx->poll_multi_file && *nr_events < min;
2277 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2278 struct kiocb *kiocb = &req->rw.kiocb;
2281 * Move completed and retryable entries to our local lists.
2282 * If we find a request that requires polling, break out
2283 * and complete those lists first, if we have entries there.
2285 if (READ_ONCE(req->iopoll_completed)) {
2286 list_move_tail(&req->inflight_entry, &done);
2289 if (!list_empty(&done))
2292 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2296 /* iopoll may have completed current req */
2297 if (READ_ONCE(req->iopoll_completed))
2298 list_move_tail(&req->inflight_entry, &done);
2305 if (!list_empty(&done))
2306 io_iopoll_complete(ctx, nr_events, &done);
2312 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2313 * non-spinning poll check - we'll still enter the driver poll loop, but only
2314 * as a non-spinning completion check.
2316 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2319 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2322 ret = io_do_iopoll(ctx, nr_events, min);
2325 if (*nr_events >= min)
2333 * We can't just wait for polled events to come to us, we have to actively
2334 * find and complete them.
2336 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2338 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2341 mutex_lock(&ctx->uring_lock);
2342 while (!list_empty(&ctx->iopoll_list)) {
2343 unsigned int nr_events = 0;
2345 io_do_iopoll(ctx, &nr_events, 0);
2347 /* let it sleep and repeat later if can't complete a request */
2351 * Ensure we allow local-to-the-cpu processing to take place,
2352 * in this case we need to ensure that we reap all events.
2353 * Also let task_work, etc. to progress by releasing the mutex
2355 if (need_resched()) {
2356 mutex_unlock(&ctx->uring_lock);
2358 mutex_lock(&ctx->uring_lock);
2361 mutex_unlock(&ctx->uring_lock);
2364 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2366 unsigned int nr_events = 0;
2367 int iters = 0, ret = 0;
2370 * We disallow the app entering submit/complete with polling, but we
2371 * still need to lock the ring to prevent racing with polled issue
2372 * that got punted to a workqueue.
2374 mutex_lock(&ctx->uring_lock);
2377 * Don't enter poll loop if we already have events pending.
2378 * If we do, we can potentially be spinning for commands that
2379 * already triggered a CQE (eg in error).
2381 if (test_bit(0, &ctx->cq_check_overflow))
2382 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2383 if (io_cqring_events(ctx))
2387 * If a submit got punted to a workqueue, we can have the
2388 * application entering polling for a command before it gets
2389 * issued. That app will hold the uring_lock for the duration
2390 * of the poll right here, so we need to take a breather every
2391 * now and then to ensure that the issue has a chance to add
2392 * the poll to the issued list. Otherwise we can spin here
2393 * forever, while the workqueue is stuck trying to acquire the
2396 if (!(++iters & 7)) {
2397 mutex_unlock(&ctx->uring_lock);
2399 mutex_lock(&ctx->uring_lock);
2402 ret = io_iopoll_getevents(ctx, &nr_events, min);
2406 } while (min && !nr_events && !need_resched());
2408 mutex_unlock(&ctx->uring_lock);
2412 static void kiocb_end_write(struct io_kiocb *req)
2415 * Tell lockdep we inherited freeze protection from submission
2418 if (req->flags & REQ_F_ISREG) {
2419 struct inode *inode = file_inode(req->file);
2421 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2423 file_end_write(req->file);
2427 static bool io_resubmit_prep(struct io_kiocb *req)
2429 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2431 struct iov_iter iter;
2433 /* already prepared */
2434 if (req->async_data)
2437 switch (req->opcode) {
2438 case IORING_OP_READV:
2439 case IORING_OP_READ_FIXED:
2440 case IORING_OP_READ:
2443 case IORING_OP_WRITEV:
2444 case IORING_OP_WRITE_FIXED:
2445 case IORING_OP_WRITE:
2449 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2454 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2457 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2460 static bool io_rw_should_reissue(struct io_kiocb *req)
2462 umode_t mode = file_inode(req->file)->i_mode;
2463 struct io_ring_ctx *ctx = req->ctx;
2465 if (!S_ISBLK(mode) && !S_ISREG(mode))
2467 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2468 !(ctx->flags & IORING_SETUP_IOPOLL)))
2471 * If ref is dying, we might be running poll reap from the exit work.
2472 * Don't attempt to reissue from that path, just let it fail with
2475 if (percpu_ref_is_dying(&ctx->refs))
2481 static bool io_rw_reissue(struct io_kiocb *req)
2484 if (!io_rw_should_reissue(req))
2487 lockdep_assert_held(&req->ctx->uring_lock);
2489 if (io_resubmit_prep(req)) {
2490 refcount_inc(&req->refs);
2491 io_queue_async_work(req);
2494 req_set_fail_links(req);
2499 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2500 unsigned int issue_flags)
2504 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2505 kiocb_end_write(req);
2506 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2508 if (res != req->result)
2509 req_set_fail_links(req);
2510 if (req->flags & REQ_F_BUFFER_SELECTED)
2511 cflags = io_put_rw_kbuf(req);
2512 __io_req_complete(req, issue_flags, res, cflags);
2515 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2517 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2519 __io_complete_rw(req, res, res2, 0);
2522 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2524 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2527 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2528 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2529 struct io_async_rw *rw = req->async_data;
2532 iov_iter_revert(&rw->iter,
2533 req->result - iov_iter_count(&rw->iter));
2534 else if (!io_resubmit_prep(req))
2539 if (kiocb->ki_flags & IOCB_WRITE)
2540 kiocb_end_write(req);
2542 if (res != -EAGAIN && res != req->result)
2543 req_set_fail_links(req);
2545 WRITE_ONCE(req->result, res);
2546 /* order with io_poll_complete() checking ->result */
2548 WRITE_ONCE(req->iopoll_completed, 1);
2552 * After the iocb has been issued, it's safe to be found on the poll list.
2553 * Adding the kiocb to the list AFTER submission ensures that we don't
2554 * find it from a io_iopoll_getevents() thread before the issuer is done
2555 * accessing the kiocb cookie.
2557 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2559 struct io_ring_ctx *ctx = req->ctx;
2562 * Track whether we have multiple files in our lists. This will impact
2563 * how we do polling eventually, not spinning if we're on potentially
2564 * different devices.
2566 if (list_empty(&ctx->iopoll_list)) {
2567 ctx->poll_multi_file = false;
2568 } else if (!ctx->poll_multi_file) {
2569 struct io_kiocb *list_req;
2571 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2573 if (list_req->file != req->file)
2574 ctx->poll_multi_file = true;
2578 * For fast devices, IO may have already completed. If it has, add
2579 * it to the front so we find it first.
2581 if (READ_ONCE(req->iopoll_completed))
2582 list_add(&req->inflight_entry, &ctx->iopoll_list);
2584 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2587 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2588 * task context or in io worker task context. If current task context is
2589 * sq thread, we don't need to check whether should wake up sq thread.
2591 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2592 wq_has_sleeper(&ctx->sq_data->wait))
2593 wake_up(&ctx->sq_data->wait);
2596 static inline void io_state_file_put(struct io_submit_state *state)
2598 if (state->file_refs) {
2599 fput_many(state->file, state->file_refs);
2600 state->file_refs = 0;
2605 * Get as many references to a file as we have IOs left in this submission,
2606 * assuming most submissions are for one file, or at least that each file
2607 * has more than one submission.
2609 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2614 if (state->file_refs) {
2615 if (state->fd == fd) {
2619 io_state_file_put(state);
2621 state->file = fget_many(fd, state->ios_left);
2622 if (unlikely(!state->file))
2626 state->file_refs = state->ios_left - 1;
2630 static bool io_bdev_nowait(struct block_device *bdev)
2632 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2636 * If we tracked the file through the SCM inflight mechanism, we could support
2637 * any file. For now, just ensure that anything potentially problematic is done
2640 static bool io_file_supports_async(struct file *file, int rw)
2642 umode_t mode = file_inode(file)->i_mode;
2644 if (S_ISBLK(mode)) {
2645 if (IS_ENABLED(CONFIG_BLOCK) &&
2646 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2650 if (S_ISCHR(mode) || S_ISSOCK(mode))
2652 if (S_ISREG(mode)) {
2653 if (IS_ENABLED(CONFIG_BLOCK) &&
2654 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2655 file->f_op != &io_uring_fops)
2660 /* any ->read/write should understand O_NONBLOCK */
2661 if (file->f_flags & O_NONBLOCK)
2664 if (!(file->f_mode & FMODE_NOWAIT))
2668 return file->f_op->read_iter != NULL;
2670 return file->f_op->write_iter != NULL;
2673 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2675 struct io_ring_ctx *ctx = req->ctx;
2676 struct kiocb *kiocb = &req->rw.kiocb;
2677 struct file *file = req->file;
2681 if (S_ISREG(file_inode(file)->i_mode))
2682 req->flags |= REQ_F_ISREG;
2684 kiocb->ki_pos = READ_ONCE(sqe->off);
2685 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2686 req->flags |= REQ_F_CUR_POS;
2687 kiocb->ki_pos = file->f_pos;
2689 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2690 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2691 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2695 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2696 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2697 req->flags |= REQ_F_NOWAIT;
2699 ioprio = READ_ONCE(sqe->ioprio);
2701 ret = ioprio_check_cap(ioprio);
2705 kiocb->ki_ioprio = ioprio;
2707 kiocb->ki_ioprio = get_current_ioprio();
2709 if (ctx->flags & IORING_SETUP_IOPOLL) {
2710 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2711 !kiocb->ki_filp->f_op->iopoll)
2714 kiocb->ki_flags |= IOCB_HIPRI;
2715 kiocb->ki_complete = io_complete_rw_iopoll;
2716 req->iopoll_completed = 0;
2718 if (kiocb->ki_flags & IOCB_HIPRI)
2720 kiocb->ki_complete = io_complete_rw;
2723 req->rw.addr = READ_ONCE(sqe->addr);
2724 req->rw.len = READ_ONCE(sqe->len);
2725 req->buf_index = READ_ONCE(sqe->buf_index);
2729 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2735 case -ERESTARTNOINTR:
2736 case -ERESTARTNOHAND:
2737 case -ERESTART_RESTARTBLOCK:
2739 * We can't just restart the syscall, since previously
2740 * submitted sqes may already be in progress. Just fail this
2746 kiocb->ki_complete(kiocb, ret, 0);
2750 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2751 unsigned int issue_flags)
2753 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2754 struct io_async_rw *io = req->async_data;
2756 /* add previously done IO, if any */
2757 if (io && io->bytes_done > 0) {
2759 ret = io->bytes_done;
2761 ret += io->bytes_done;
2764 if (req->flags & REQ_F_CUR_POS)
2765 req->file->f_pos = kiocb->ki_pos;
2766 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2767 __io_complete_rw(req, ret, 0, issue_flags);
2769 io_rw_done(kiocb, ret);
2772 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2774 struct io_ring_ctx *ctx = req->ctx;
2775 size_t len = req->rw.len;
2776 struct io_mapped_ubuf *imu;
2777 u16 index, buf_index = req->buf_index;
2781 if (unlikely(buf_index >= ctx->nr_user_bufs))
2783 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2784 imu = &ctx->user_bufs[index];
2785 buf_addr = req->rw.addr;
2788 if (buf_addr + len < buf_addr)
2790 /* not inside the mapped region */
2791 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2795 * May not be a start of buffer, set size appropriately
2796 * and advance us to the beginning.
2798 offset = buf_addr - imu->ubuf;
2799 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2803 * Don't use iov_iter_advance() here, as it's really slow for
2804 * using the latter parts of a big fixed buffer - it iterates
2805 * over each segment manually. We can cheat a bit here, because
2808 * 1) it's a BVEC iter, we set it up
2809 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2810 * first and last bvec
2812 * So just find our index, and adjust the iterator afterwards.
2813 * If the offset is within the first bvec (or the whole first
2814 * bvec, just use iov_iter_advance(). This makes it easier
2815 * since we can just skip the first segment, which may not
2816 * be PAGE_SIZE aligned.
2818 const struct bio_vec *bvec = imu->bvec;
2820 if (offset <= bvec->bv_len) {
2821 iov_iter_advance(iter, offset);
2823 unsigned long seg_skip;
2825 /* skip first vec */
2826 offset -= bvec->bv_len;
2827 seg_skip = 1 + (offset >> PAGE_SHIFT);
2829 iter->bvec = bvec + seg_skip;
2830 iter->nr_segs -= seg_skip;
2831 iter->count -= bvec->bv_len + offset;
2832 iter->iov_offset = offset & ~PAGE_MASK;
2839 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2842 mutex_unlock(&ctx->uring_lock);
2845 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2848 * "Normal" inline submissions always hold the uring_lock, since we
2849 * grab it from the system call. Same is true for the SQPOLL offload.
2850 * The only exception is when we've detached the request and issue it
2851 * from an async worker thread, grab the lock for that case.
2854 mutex_lock(&ctx->uring_lock);
2857 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2858 int bgid, struct io_buffer *kbuf,
2861 struct io_buffer *head;
2863 if (req->flags & REQ_F_BUFFER_SELECTED)
2866 io_ring_submit_lock(req->ctx, needs_lock);
2868 lockdep_assert_held(&req->ctx->uring_lock);
2870 head = xa_load(&req->ctx->io_buffers, bgid);
2872 if (!list_empty(&head->list)) {
2873 kbuf = list_last_entry(&head->list, struct io_buffer,
2875 list_del(&kbuf->list);
2878 xa_erase(&req->ctx->io_buffers, bgid);
2880 if (*len > kbuf->len)
2883 kbuf = ERR_PTR(-ENOBUFS);
2886 io_ring_submit_unlock(req->ctx, needs_lock);
2891 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2894 struct io_buffer *kbuf;
2897 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2898 bgid = req->buf_index;
2899 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2902 req->rw.addr = (u64) (unsigned long) kbuf;
2903 req->flags |= REQ_F_BUFFER_SELECTED;
2904 return u64_to_user_ptr(kbuf->addr);
2907 #ifdef CONFIG_COMPAT
2908 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2911 struct compat_iovec __user *uiov;
2912 compat_ssize_t clen;
2916 uiov = u64_to_user_ptr(req->rw.addr);
2917 if (!access_ok(uiov, sizeof(*uiov)))
2919 if (__get_user(clen, &uiov->iov_len))
2925 buf = io_rw_buffer_select(req, &len, needs_lock);
2927 return PTR_ERR(buf);
2928 iov[0].iov_base = buf;
2929 iov[0].iov_len = (compat_size_t) len;
2934 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2937 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2941 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2944 len = iov[0].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 = len;
2955 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2958 if (req->flags & REQ_F_BUFFER_SELECTED) {
2959 struct io_buffer *kbuf;
2961 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2962 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2963 iov[0].iov_len = kbuf->len;
2966 if (req->rw.len != 1)
2969 #ifdef CONFIG_COMPAT
2970 if (req->ctx->compat)
2971 return io_compat_import(req, iov, needs_lock);
2974 return __io_iov_buffer_select(req, iov, needs_lock);
2977 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2978 struct iov_iter *iter, bool needs_lock)
2980 void __user *buf = u64_to_user_ptr(req->rw.addr);
2981 size_t sqe_len = req->rw.len;
2982 u8 opcode = req->opcode;
2985 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2987 return io_import_fixed(req, rw, iter);
2990 /* buffer index only valid with fixed read/write, or buffer select */
2991 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2994 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2995 if (req->flags & REQ_F_BUFFER_SELECT) {
2996 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2998 return PTR_ERR(buf);
2999 req->rw.len = sqe_len;
3002 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3007 if (req->flags & REQ_F_BUFFER_SELECT) {
3008 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3010 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3015 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3019 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3021 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3025 * For files that don't have ->read_iter() and ->write_iter(), handle them
3026 * by looping over ->read() or ->write() manually.
3028 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3030 struct kiocb *kiocb = &req->rw.kiocb;
3031 struct file *file = req->file;
3035 * Don't support polled IO through this interface, and we can't
3036 * support non-blocking either. For the latter, this just causes
3037 * the kiocb to be handled from an async context.
3039 if (kiocb->ki_flags & IOCB_HIPRI)
3041 if (kiocb->ki_flags & IOCB_NOWAIT)
3044 while (iov_iter_count(iter)) {
3048 if (!iov_iter_is_bvec(iter)) {
3049 iovec = iov_iter_iovec(iter);
3051 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3052 iovec.iov_len = req->rw.len;
3056 nr = file->f_op->read(file, iovec.iov_base,
3057 iovec.iov_len, io_kiocb_ppos(kiocb));
3059 nr = file->f_op->write(file, iovec.iov_base,
3060 iovec.iov_len, io_kiocb_ppos(kiocb));
3069 if (nr != iovec.iov_len)
3073 iov_iter_advance(iter, nr);
3079 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3080 const struct iovec *fast_iov, struct iov_iter *iter)
3082 struct io_async_rw *rw = req->async_data;
3084 memcpy(&rw->iter, iter, sizeof(*iter));
3085 rw->free_iovec = iovec;
3087 /* can only be fixed buffers, no need to do anything */
3088 if (iov_iter_is_bvec(iter))
3091 unsigned iov_off = 0;
3093 rw->iter.iov = rw->fast_iov;
3094 if (iter->iov != fast_iov) {
3095 iov_off = iter->iov - fast_iov;
3096 rw->iter.iov += iov_off;
3098 if (rw->fast_iov != fast_iov)
3099 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3100 sizeof(struct iovec) * iter->nr_segs);
3102 req->flags |= REQ_F_NEED_CLEANUP;
3106 static inline int __io_alloc_async_data(struct io_kiocb *req)
3108 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3109 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3110 return req->async_data == NULL;
3113 static int io_alloc_async_data(struct io_kiocb *req)
3115 if (!io_op_defs[req->opcode].needs_async_data)
3118 return __io_alloc_async_data(req);
3121 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3122 const struct iovec *fast_iov,
3123 struct iov_iter *iter, bool force)
3125 if (!force && !io_op_defs[req->opcode].needs_async_data)
3127 if (!req->async_data) {
3128 if (__io_alloc_async_data(req)) {
3133 io_req_map_rw(req, iovec, fast_iov, iter);
3138 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3140 struct io_async_rw *iorw = req->async_data;
3141 struct iovec *iov = iorw->fast_iov;
3144 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3145 if (unlikely(ret < 0))
3148 iorw->bytes_done = 0;
3149 iorw->free_iovec = iov;
3151 req->flags |= REQ_F_NEED_CLEANUP;
3155 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3157 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3159 return io_prep_rw(req, sqe);
3163 * This is our waitqueue callback handler, registered through lock_page_async()
3164 * when we initially tried to do the IO with the iocb armed our waitqueue.
3165 * This gets called when the page is unlocked, and we generally expect that to
3166 * happen when the page IO is completed and the page is now uptodate. This will
3167 * queue a task_work based retry of the operation, attempting to copy the data
3168 * again. If the latter fails because the page was NOT uptodate, then we will
3169 * do a thread based blocking retry of the operation. That's the unexpected
3172 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3173 int sync, void *arg)
3175 struct wait_page_queue *wpq;
3176 struct io_kiocb *req = wait->private;
3177 struct wait_page_key *key = arg;
3179 wpq = container_of(wait, struct wait_page_queue, wait);
3181 if (!wake_page_match(wpq, key))
3184 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3185 list_del_init(&wait->entry);
3187 /* submit ref gets dropped, acquire a new one */
3188 refcount_inc(&req->refs);
3189 io_req_task_queue(req);
3194 * This controls whether a given IO request should be armed for async page
3195 * based retry. If we return false here, the request is handed to the async
3196 * worker threads for retry. If we're doing buffered reads on a regular file,
3197 * we prepare a private wait_page_queue entry and retry the operation. This
3198 * will either succeed because the page is now uptodate and unlocked, or it
3199 * will register a callback when the page is unlocked at IO completion. Through
3200 * that callback, io_uring uses task_work to setup a retry of the operation.
3201 * That retry will attempt the buffered read again. The retry will generally
3202 * succeed, or in rare cases where it fails, we then fall back to using the
3203 * async worker threads for a blocking retry.
3205 static bool io_rw_should_retry(struct io_kiocb *req)
3207 struct io_async_rw *rw = req->async_data;
3208 struct wait_page_queue *wait = &rw->wpq;
3209 struct kiocb *kiocb = &req->rw.kiocb;
3211 /* never retry for NOWAIT, we just complete with -EAGAIN */
3212 if (req->flags & REQ_F_NOWAIT)
3215 /* Only for buffered IO */
3216 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3220 * just use poll if we can, and don't attempt if the fs doesn't
3221 * support callback based unlocks
3223 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3226 wait->wait.func = io_async_buf_func;
3227 wait->wait.private = req;
3228 wait->wait.flags = 0;
3229 INIT_LIST_HEAD(&wait->wait.entry);
3230 kiocb->ki_flags |= IOCB_WAITQ;
3231 kiocb->ki_flags &= ~IOCB_NOWAIT;
3232 kiocb->ki_waitq = wait;
3236 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3238 if (req->file->f_op->read_iter)
3239 return call_read_iter(req->file, &req->rw.kiocb, iter);
3240 else if (req->file->f_op->read)
3241 return loop_rw_iter(READ, req, iter);
3246 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3248 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3249 struct kiocb *kiocb = &req->rw.kiocb;
3250 struct iov_iter __iter, *iter = &__iter;
3251 struct io_async_rw *rw = req->async_data;
3252 ssize_t io_size, ret, ret2;
3253 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3259 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3263 io_size = iov_iter_count(iter);
3264 req->result = io_size;
3266 /* Ensure we clear previously set non-block flag */
3267 if (!force_nonblock)
3268 kiocb->ki_flags &= ~IOCB_NOWAIT;
3270 kiocb->ki_flags |= IOCB_NOWAIT;
3272 /* If the file doesn't support async, just async punt */
3273 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3274 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3275 return ret ?: -EAGAIN;
3278 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3279 if (unlikely(ret)) {
3284 ret = io_iter_do_read(req, iter);
3286 if (ret == -EIOCBQUEUED) {
3287 if (req->async_data)
3288 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3290 } else if (ret == -EAGAIN) {
3291 /* IOPOLL retry should happen for io-wq threads */
3292 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3294 /* no retry on NONBLOCK nor RWF_NOWAIT */
3295 if (req->flags & REQ_F_NOWAIT)
3297 /* some cases will consume bytes even on error returns */
3298 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3300 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3301 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3302 /* read all, failed, already did sync or don't want to retry */
3306 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3311 rw = req->async_data;
3312 /* now use our persistent iterator, if we aren't already */
3317 rw->bytes_done += ret;
3318 /* if we can retry, do so with the callbacks armed */
3319 if (!io_rw_should_retry(req)) {
3320 kiocb->ki_flags &= ~IOCB_WAITQ;
3325 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3326 * we get -EIOCBQUEUED, then we'll get a notification when the
3327 * desired page gets unlocked. We can also get a partial read
3328 * here, and if we do, then just retry at the new offset.
3330 ret = io_iter_do_read(req, iter);
3331 if (ret == -EIOCBQUEUED)
3333 /* we got some bytes, but not all. retry. */
3334 kiocb->ki_flags &= ~IOCB_WAITQ;
3335 } while (ret > 0 && ret < io_size);
3337 kiocb_done(kiocb, ret, issue_flags);
3339 /* it's faster to check here then delegate to kfree */
3345 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3347 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3349 return io_prep_rw(req, sqe);
3352 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3354 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3355 struct kiocb *kiocb = &req->rw.kiocb;
3356 struct iov_iter __iter, *iter = &__iter;
3357 struct io_async_rw *rw = req->async_data;
3358 ssize_t ret, ret2, io_size;
3359 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3365 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3369 io_size = iov_iter_count(iter);
3370 req->result = io_size;
3372 /* Ensure we clear previously set non-block flag */
3373 if (!force_nonblock)
3374 kiocb->ki_flags &= ~IOCB_NOWAIT;
3376 kiocb->ki_flags |= IOCB_NOWAIT;
3378 /* If the file doesn't support async, just async punt */
3379 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3382 /* file path doesn't support NOWAIT for non-direct_IO */
3383 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3384 (req->flags & REQ_F_ISREG))
3387 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3392 * Open-code file_start_write here to grab freeze protection,
3393 * which will be released by another thread in
3394 * io_complete_rw(). Fool lockdep by telling it the lock got
3395 * released so that it doesn't complain about the held lock when
3396 * we return to userspace.
3398 if (req->flags & REQ_F_ISREG) {
3399 sb_start_write(file_inode(req->file)->i_sb);
3400 __sb_writers_release(file_inode(req->file)->i_sb,
3403 kiocb->ki_flags |= IOCB_WRITE;
3405 if (req->file->f_op->write_iter)
3406 ret2 = call_write_iter(req->file, kiocb, iter);
3407 else if (req->file->f_op->write)
3408 ret2 = loop_rw_iter(WRITE, req, iter);
3413 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3414 * retry them without IOCB_NOWAIT.
3416 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3418 /* no retry on NONBLOCK nor RWF_NOWAIT */
3419 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3421 if (ret2 == -EIOCBQUEUED && req->async_data)
3422 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3423 if (!force_nonblock || ret2 != -EAGAIN) {
3424 /* IOPOLL retry should happen for io-wq threads */
3425 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3428 kiocb_done(kiocb, ret2, issue_flags);
3431 /* some cases will consume bytes even on error returns */
3432 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3433 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3434 return ret ?: -EAGAIN;
3437 /* it's reportedly faster than delegating the null check to kfree() */
3443 static int io_renameat_prep(struct io_kiocb *req,
3444 const struct io_uring_sqe *sqe)
3446 struct io_rename *ren = &req->rename;
3447 const char __user *oldf, *newf;
3449 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3452 ren->old_dfd = READ_ONCE(sqe->fd);
3453 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3454 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3455 ren->new_dfd = READ_ONCE(sqe->len);
3456 ren->flags = READ_ONCE(sqe->rename_flags);
3458 ren->oldpath = getname(oldf);
3459 if (IS_ERR(ren->oldpath))
3460 return PTR_ERR(ren->oldpath);
3462 ren->newpath = getname(newf);
3463 if (IS_ERR(ren->newpath)) {
3464 putname(ren->oldpath);
3465 return PTR_ERR(ren->newpath);
3468 req->flags |= REQ_F_NEED_CLEANUP;
3472 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3474 struct io_rename *ren = &req->rename;
3477 if (issue_flags & IO_URING_F_NONBLOCK)
3480 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3481 ren->newpath, ren->flags);
3483 req->flags &= ~REQ_F_NEED_CLEANUP;
3485 req_set_fail_links(req);
3486 io_req_complete(req, ret);
3490 static int io_unlinkat_prep(struct io_kiocb *req,
3491 const struct io_uring_sqe *sqe)
3493 struct io_unlink *un = &req->unlink;
3494 const char __user *fname;
3496 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3499 un->dfd = READ_ONCE(sqe->fd);
3501 un->flags = READ_ONCE(sqe->unlink_flags);
3502 if (un->flags & ~AT_REMOVEDIR)
3505 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3506 un->filename = getname(fname);
3507 if (IS_ERR(un->filename))
3508 return PTR_ERR(un->filename);
3510 req->flags |= REQ_F_NEED_CLEANUP;
3514 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3516 struct io_unlink *un = &req->unlink;
3519 if (issue_flags & IO_URING_F_NONBLOCK)
3522 if (un->flags & AT_REMOVEDIR)
3523 ret = do_rmdir(un->dfd, un->filename);
3525 ret = do_unlinkat(un->dfd, un->filename);
3527 req->flags &= ~REQ_F_NEED_CLEANUP;
3529 req_set_fail_links(req);
3530 io_req_complete(req, ret);
3534 static int io_shutdown_prep(struct io_kiocb *req,
3535 const struct io_uring_sqe *sqe)
3537 #if defined(CONFIG_NET)
3538 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3540 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3544 req->shutdown.how = READ_ONCE(sqe->len);
3551 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3553 #if defined(CONFIG_NET)
3554 struct socket *sock;
3557 if (issue_flags & IO_URING_F_NONBLOCK)
3560 sock = sock_from_file(req->file);
3561 if (unlikely(!sock))
3564 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3566 req_set_fail_links(req);
3567 io_req_complete(req, ret);
3574 static int __io_splice_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 struct io_splice* sp = &req->splice;
3578 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3584 sp->len = READ_ONCE(sqe->len);
3585 sp->flags = READ_ONCE(sqe->splice_flags);
3587 if (unlikely(sp->flags & ~valid_flags))
3590 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3591 (sp->flags & SPLICE_F_FD_IN_FIXED));
3594 req->flags |= REQ_F_NEED_CLEANUP;
3596 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3598 * Splice operation will be punted aync, and here need to
3599 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3601 req->work.flags |= IO_WQ_WORK_UNBOUND;
3607 static int io_tee_prep(struct io_kiocb *req,
3608 const struct io_uring_sqe *sqe)
3610 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3612 return __io_splice_prep(req, sqe);
3615 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3617 struct io_splice *sp = &req->splice;
3618 struct file *in = sp->file_in;
3619 struct file *out = sp->file_out;
3620 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3623 if (issue_flags & IO_URING_F_NONBLOCK)
3626 ret = do_tee(in, out, sp->len, flags);
3628 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3629 req->flags &= ~REQ_F_NEED_CLEANUP;
3632 req_set_fail_links(req);
3633 io_req_complete(req, ret);
3637 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3639 struct io_splice* sp = &req->splice;
3641 sp->off_in = READ_ONCE(sqe->splice_off_in);
3642 sp->off_out = READ_ONCE(sqe->off);
3643 return __io_splice_prep(req, sqe);
3646 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3648 struct io_splice *sp = &req->splice;
3649 struct file *in = sp->file_in;
3650 struct file *out = sp->file_out;
3651 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3652 loff_t *poff_in, *poff_out;
3655 if (issue_flags & IO_URING_F_NONBLOCK)
3658 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3659 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3662 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3664 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3665 req->flags &= ~REQ_F_NEED_CLEANUP;
3668 req_set_fail_links(req);
3669 io_req_complete(req, ret);
3674 * IORING_OP_NOP just posts a completion event, nothing else.
3676 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3678 struct io_ring_ctx *ctx = req->ctx;
3680 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3683 __io_req_complete(req, issue_flags, 0, 0);
3687 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3689 struct io_ring_ctx *ctx = req->ctx;
3694 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3696 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3699 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3700 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3703 req->sync.off = READ_ONCE(sqe->off);
3704 req->sync.len = READ_ONCE(sqe->len);
3708 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3710 loff_t end = req->sync.off + req->sync.len;
3713 /* fsync always requires a blocking context */
3714 if (issue_flags & IO_URING_F_NONBLOCK)
3717 ret = vfs_fsync_range(req->file, req->sync.off,
3718 end > 0 ? end : LLONG_MAX,
3719 req->sync.flags & IORING_FSYNC_DATASYNC);
3721 req_set_fail_links(req);
3722 io_req_complete(req, ret);
3726 static int io_fallocate_prep(struct io_kiocb *req,
3727 const struct io_uring_sqe *sqe)
3729 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3734 req->sync.off = READ_ONCE(sqe->off);
3735 req->sync.len = READ_ONCE(sqe->addr);
3736 req->sync.mode = READ_ONCE(sqe->len);
3740 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3744 /* fallocate always requiring blocking context */
3745 if (issue_flags & IO_URING_F_NONBLOCK)
3747 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3750 req_set_fail_links(req);
3751 io_req_complete(req, ret);
3755 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3757 const char __user *fname;
3760 if (unlikely(sqe->ioprio || sqe->buf_index))
3762 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3765 /* open.how should be already initialised */
3766 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3767 req->open.how.flags |= O_LARGEFILE;
3769 req->open.dfd = READ_ONCE(sqe->fd);
3770 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3771 req->open.filename = getname(fname);
3772 if (IS_ERR(req->open.filename)) {
3773 ret = PTR_ERR(req->open.filename);
3774 req->open.filename = NULL;
3777 req->open.nofile = rlimit(RLIMIT_NOFILE);
3778 req->flags |= REQ_F_NEED_CLEANUP;
3782 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3786 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3788 mode = READ_ONCE(sqe->len);
3789 flags = READ_ONCE(sqe->open_flags);
3790 req->open.how = build_open_how(flags, mode);
3791 return __io_openat_prep(req, sqe);
3794 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796 struct open_how __user *how;
3800 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3803 len = READ_ONCE(sqe->len);
3804 if (len < OPEN_HOW_SIZE_VER0)
3807 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3817 struct open_flags op;
3820 bool resolve_nonblock;
3823 ret = build_open_flags(&req->open.how, &op);
3826 nonblock_set = op.open_flag & O_NONBLOCK;
3827 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3828 if (issue_flags & IO_URING_F_NONBLOCK) {
3830 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3831 * it'll always -EAGAIN
3833 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3835 op.lookup_flags |= LOOKUP_CACHED;
3836 op.open_flag |= O_NONBLOCK;
3839 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3843 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3844 /* only retry if RESOLVE_CACHED wasn't already set by application */
3845 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3846 file == ERR_PTR(-EAGAIN)) {
3848 * We could hang on to this 'fd', but seems like marginal
3849 * gain for something that is now known to be a slower path.
3850 * So just put it, and we'll get a new one when we retry.
3858 ret = PTR_ERR(file);
3860 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3861 file->f_flags &= ~O_NONBLOCK;
3862 fsnotify_open(file);
3863 fd_install(ret, file);
3866 putname(req->open.filename);
3867 req->flags &= ~REQ_F_NEED_CLEANUP;
3869 req_set_fail_links(req);
3870 io_req_complete(req, ret);
3874 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3876 return io_openat2(req, issue_flags);
3879 static int io_remove_buffers_prep(struct io_kiocb *req,
3880 const struct io_uring_sqe *sqe)
3882 struct io_provide_buf *p = &req->pbuf;
3885 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3888 tmp = READ_ONCE(sqe->fd);
3889 if (!tmp || tmp > USHRT_MAX)
3892 memset(p, 0, sizeof(*p));
3894 p->bgid = READ_ONCE(sqe->buf_group);
3898 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3899 int bgid, unsigned nbufs)
3903 /* shouldn't happen */
3907 /* the head kbuf is the list itself */
3908 while (!list_empty(&buf->list)) {
3909 struct io_buffer *nxt;
3911 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3912 list_del(&nxt->list);
3919 xa_erase(&ctx->io_buffers, bgid);
3924 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3926 struct io_provide_buf *p = &req->pbuf;
3927 struct io_ring_ctx *ctx = req->ctx;
3928 struct io_buffer *head;
3930 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3932 io_ring_submit_lock(ctx, !force_nonblock);
3934 lockdep_assert_held(&ctx->uring_lock);
3937 head = xa_load(&ctx->io_buffers, p->bgid);
3939 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3941 req_set_fail_links(req);
3943 /* need to hold the lock to complete IOPOLL requests */
3944 if (ctx->flags & IORING_SETUP_IOPOLL) {
3945 __io_req_complete(req, issue_flags, ret, 0);
3946 io_ring_submit_unlock(ctx, !force_nonblock);
3948 io_ring_submit_unlock(ctx, !force_nonblock);
3949 __io_req_complete(req, issue_flags, ret, 0);
3954 static int io_provide_buffers_prep(struct io_kiocb *req,
3955 const struct io_uring_sqe *sqe)
3958 struct io_provide_buf *p = &req->pbuf;
3961 if (sqe->ioprio || sqe->rw_flags)
3964 tmp = READ_ONCE(sqe->fd);
3965 if (!tmp || tmp > USHRT_MAX)
3968 p->addr = READ_ONCE(sqe->addr);
3969 p->len = READ_ONCE(sqe->len);
3971 size = (unsigned long)p->len * p->nbufs;
3972 if (!access_ok(u64_to_user_ptr(p->addr), size))
3975 p->bgid = READ_ONCE(sqe->buf_group);
3976 tmp = READ_ONCE(sqe->off);
3977 if (tmp > USHRT_MAX)
3983 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3985 struct io_buffer *buf;
3986 u64 addr = pbuf->addr;
3987 int i, bid = pbuf->bid;
3989 for (i = 0; i < pbuf->nbufs; i++) {
3990 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3995 buf->len = pbuf->len;
4000 INIT_LIST_HEAD(&buf->list);
4003 list_add_tail(&buf->list, &(*head)->list);
4007 return i ? i : -ENOMEM;
4010 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4012 struct io_provide_buf *p = &req->pbuf;
4013 struct io_ring_ctx *ctx = req->ctx;
4014 struct io_buffer *head, *list;
4016 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4018 io_ring_submit_lock(ctx, !force_nonblock);
4020 lockdep_assert_held(&ctx->uring_lock);
4022 list = head = xa_load(&ctx->io_buffers, p->bgid);
4024 ret = io_add_buffers(p, &head);
4025 if (ret >= 0 && !list) {
4026 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4028 __io_remove_buffers(ctx, head, p->bgid, -1U);
4031 req_set_fail_links(req);
4033 /* need to hold the lock to complete IOPOLL requests */
4034 if (ctx->flags & IORING_SETUP_IOPOLL) {
4035 __io_req_complete(req, issue_flags, ret, 0);
4036 io_ring_submit_unlock(ctx, !force_nonblock);
4038 io_ring_submit_unlock(ctx, !force_nonblock);
4039 __io_req_complete(req, issue_flags, ret, 0);
4044 static int io_epoll_ctl_prep(struct io_kiocb *req,
4045 const struct io_uring_sqe *sqe)
4047 #if defined(CONFIG_EPOLL)
4048 if (sqe->ioprio || sqe->buf_index)
4050 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4053 req->epoll.epfd = READ_ONCE(sqe->fd);
4054 req->epoll.op = READ_ONCE(sqe->len);
4055 req->epoll.fd = READ_ONCE(sqe->off);
4057 if (ep_op_has_event(req->epoll.op)) {
4058 struct epoll_event __user *ev;
4060 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4061 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4071 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4073 #if defined(CONFIG_EPOLL)
4074 struct io_epoll *ie = &req->epoll;
4076 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4078 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4079 if (force_nonblock && ret == -EAGAIN)
4083 req_set_fail_links(req);
4084 __io_req_complete(req, issue_flags, ret, 0);
4091 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4093 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4094 if (sqe->ioprio || sqe->buf_index || sqe->off)
4096 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4099 req->madvise.addr = READ_ONCE(sqe->addr);
4100 req->madvise.len = READ_ONCE(sqe->len);
4101 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4108 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4110 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4111 struct io_madvise *ma = &req->madvise;
4114 if (issue_flags & IO_URING_F_NONBLOCK)
4117 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4119 req_set_fail_links(req);
4120 io_req_complete(req, ret);
4127 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4131 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4134 req->fadvise.offset = READ_ONCE(sqe->off);
4135 req->fadvise.len = READ_ONCE(sqe->len);
4136 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4140 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4142 struct io_fadvise *fa = &req->fadvise;
4145 if (issue_flags & IO_URING_F_NONBLOCK) {
4146 switch (fa->advice) {
4147 case POSIX_FADV_NORMAL:
4148 case POSIX_FADV_RANDOM:
4149 case POSIX_FADV_SEQUENTIAL:
4156 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4158 req_set_fail_links(req);
4159 io_req_complete(req, ret);
4163 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4165 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4167 if (sqe->ioprio || sqe->buf_index)
4169 if (req->flags & REQ_F_FIXED_FILE)
4172 req->statx.dfd = READ_ONCE(sqe->fd);
4173 req->statx.mask = READ_ONCE(sqe->len);
4174 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4175 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4176 req->statx.flags = READ_ONCE(sqe->statx_flags);
4181 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4183 struct io_statx *ctx = &req->statx;
4186 if (issue_flags & IO_URING_F_NONBLOCK) {
4187 /* only need file table for an actual valid fd */
4188 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4189 req->flags |= REQ_F_NO_FILE_TABLE;
4193 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4197 req_set_fail_links(req);
4198 io_req_complete(req, ret);
4202 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4204 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4206 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4207 sqe->rw_flags || sqe->buf_index)
4209 if (req->flags & REQ_F_FIXED_FILE)
4212 req->close.fd = READ_ONCE(sqe->fd);
4216 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4218 struct files_struct *files = current->files;
4219 struct io_close *close = &req->close;
4220 struct fdtable *fdt;
4226 spin_lock(&files->file_lock);
4227 fdt = files_fdtable(files);
4228 if (close->fd >= fdt->max_fds) {
4229 spin_unlock(&files->file_lock);
4232 file = fdt->fd[close->fd];
4234 spin_unlock(&files->file_lock);
4238 if (file->f_op == &io_uring_fops) {
4239 spin_unlock(&files->file_lock);
4244 /* if the file has a flush method, be safe and punt to async */
4245 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4246 spin_unlock(&files->file_lock);
4250 ret = __close_fd_get_file(close->fd, &file);
4251 spin_unlock(&files->file_lock);
4258 /* No ->flush() or already async, safely close from here */
4259 ret = filp_close(file, current->files);
4262 req_set_fail_links(req);
4265 __io_req_complete(req, issue_flags, ret, 0);
4269 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4271 struct io_ring_ctx *ctx = req->ctx;
4273 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4275 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4278 req->sync.off = READ_ONCE(sqe->off);
4279 req->sync.len = READ_ONCE(sqe->len);
4280 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4284 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4288 /* sync_file_range always requires a blocking context */
4289 if (issue_flags & IO_URING_F_NONBLOCK)
4292 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4295 req_set_fail_links(req);
4296 io_req_complete(req, ret);
4300 #if defined(CONFIG_NET)
4301 static int io_setup_async_msg(struct io_kiocb *req,
4302 struct io_async_msghdr *kmsg)
4304 struct io_async_msghdr *async_msg = req->async_data;
4308 if (io_alloc_async_data(req)) {
4309 kfree(kmsg->free_iov);
4312 async_msg = req->async_data;
4313 req->flags |= REQ_F_NEED_CLEANUP;
4314 memcpy(async_msg, kmsg, sizeof(*kmsg));
4315 async_msg->msg.msg_name = &async_msg->addr;
4316 /* if were using fast_iov, set it to the new one */
4317 if (!async_msg->free_iov)
4318 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4323 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4324 struct io_async_msghdr *iomsg)
4326 iomsg->msg.msg_name = &iomsg->addr;
4327 iomsg->free_iov = iomsg->fast_iov;
4328 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4329 req->sr_msg.msg_flags, &iomsg->free_iov);
4332 static int io_sendmsg_prep_async(struct io_kiocb *req)
4336 if (!io_op_defs[req->opcode].needs_async_data)
4338 ret = io_sendmsg_copy_hdr(req, req->async_data);
4340 req->flags |= REQ_F_NEED_CLEANUP;
4344 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4346 struct io_sr_msg *sr = &req->sr_msg;
4348 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4351 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4352 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4353 sr->len = READ_ONCE(sqe->len);
4355 #ifdef CONFIG_COMPAT
4356 if (req->ctx->compat)
4357 sr->msg_flags |= MSG_CMSG_COMPAT;
4362 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4364 struct io_async_msghdr iomsg, *kmsg;
4365 struct socket *sock;
4370 sock = sock_from_file(req->file);
4371 if (unlikely(!sock))
4374 kmsg = req->async_data;
4376 ret = io_sendmsg_copy_hdr(req, &iomsg);
4382 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4383 if (flags & MSG_DONTWAIT)
4384 req->flags |= REQ_F_NOWAIT;
4385 else if (issue_flags & IO_URING_F_NONBLOCK)
4386 flags |= MSG_DONTWAIT;
4388 if (flags & MSG_WAITALL)
4389 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4391 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4392 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4393 return io_setup_async_msg(req, kmsg);
4394 if (ret == -ERESTARTSYS)
4397 /* fast path, check for non-NULL to avoid function call */
4399 kfree(kmsg->free_iov);
4400 req->flags &= ~REQ_F_NEED_CLEANUP;
4402 req_set_fail_links(req);
4403 __io_req_complete(req, issue_flags, ret, 0);
4407 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4409 struct io_sr_msg *sr = &req->sr_msg;
4412 struct socket *sock;
4417 sock = sock_from_file(req->file);
4418 if (unlikely(!sock))
4421 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4425 msg.msg_name = NULL;
4426 msg.msg_control = NULL;
4427 msg.msg_controllen = 0;
4428 msg.msg_namelen = 0;
4430 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4431 if (flags & MSG_DONTWAIT)
4432 req->flags |= REQ_F_NOWAIT;
4433 else if (issue_flags & IO_URING_F_NONBLOCK)
4434 flags |= MSG_DONTWAIT;
4436 if (flags & MSG_WAITALL)
4437 min_ret = iov_iter_count(&msg.msg_iter);
4439 msg.msg_flags = flags;
4440 ret = sock_sendmsg(sock, &msg);
4441 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4443 if (ret == -ERESTARTSYS)
4447 req_set_fail_links(req);
4448 __io_req_complete(req, issue_flags, ret, 0);
4452 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4453 struct io_async_msghdr *iomsg)
4455 struct io_sr_msg *sr = &req->sr_msg;
4456 struct iovec __user *uiov;
4460 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4461 &iomsg->uaddr, &uiov, &iov_len);
4465 if (req->flags & REQ_F_BUFFER_SELECT) {
4468 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4470 sr->len = iomsg->fast_iov[0].iov_len;
4471 iomsg->free_iov = NULL;
4473 iomsg->free_iov = iomsg->fast_iov;
4474 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4475 &iomsg->free_iov, &iomsg->msg.msg_iter,
4484 #ifdef CONFIG_COMPAT
4485 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4486 struct io_async_msghdr *iomsg)
4488 struct compat_msghdr __user *msg_compat;
4489 struct io_sr_msg *sr = &req->sr_msg;
4490 struct compat_iovec __user *uiov;
4495 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4496 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4501 uiov = compat_ptr(ptr);
4502 if (req->flags & REQ_F_BUFFER_SELECT) {
4503 compat_ssize_t clen;
4507 if (!access_ok(uiov, sizeof(*uiov)))
4509 if (__get_user(clen, &uiov->iov_len))
4514 iomsg->free_iov = NULL;
4516 iomsg->free_iov = iomsg->fast_iov;
4517 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4518 UIO_FASTIOV, &iomsg->free_iov,
4519 &iomsg->msg.msg_iter, true);
4528 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4529 struct io_async_msghdr *iomsg)
4531 iomsg->msg.msg_name = &iomsg->addr;
4533 #ifdef CONFIG_COMPAT
4534 if (req->ctx->compat)
4535 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4538 return __io_recvmsg_copy_hdr(req, iomsg);
4541 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4544 struct io_sr_msg *sr = &req->sr_msg;
4545 struct io_buffer *kbuf;
4547 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4552 req->flags |= REQ_F_BUFFER_SELECTED;
4556 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4558 return io_put_kbuf(req, req->sr_msg.kbuf);
4561 static int io_recvmsg_prep_async(struct io_kiocb *req)
4565 if (!io_op_defs[req->opcode].needs_async_data)
4567 ret = io_recvmsg_copy_hdr(req, req->async_data);
4569 req->flags |= REQ_F_NEED_CLEANUP;
4573 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4575 struct io_sr_msg *sr = &req->sr_msg;
4577 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4580 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4581 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4582 sr->len = READ_ONCE(sqe->len);
4583 sr->bgid = READ_ONCE(sqe->buf_group);
4585 #ifdef CONFIG_COMPAT
4586 if (req->ctx->compat)
4587 sr->msg_flags |= MSG_CMSG_COMPAT;
4592 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4594 struct io_async_msghdr iomsg, *kmsg;
4595 struct socket *sock;
4596 struct io_buffer *kbuf;
4599 int ret, cflags = 0;
4600 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4602 sock = sock_from_file(req->file);
4603 if (unlikely(!sock))
4606 kmsg = req->async_data;
4608 ret = io_recvmsg_copy_hdr(req, &iomsg);
4614 if (req->flags & REQ_F_BUFFER_SELECT) {
4615 kbuf = io_recv_buffer_select(req, !force_nonblock);
4617 return PTR_ERR(kbuf);
4618 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4619 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4620 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4621 1, req->sr_msg.len);
4624 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4625 if (flags & MSG_DONTWAIT)
4626 req->flags |= REQ_F_NOWAIT;
4627 else if (force_nonblock)
4628 flags |= MSG_DONTWAIT;
4630 if (flags & MSG_WAITALL)
4631 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4633 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4634 kmsg->uaddr, flags);
4635 if (force_nonblock && ret == -EAGAIN)
4636 return io_setup_async_msg(req, kmsg);
4637 if (ret == -ERESTARTSYS)
4640 if (req->flags & REQ_F_BUFFER_SELECTED)
4641 cflags = io_put_recv_kbuf(req);
4642 /* fast path, check for non-NULL to avoid function call */
4644 kfree(kmsg->free_iov);
4645 req->flags &= ~REQ_F_NEED_CLEANUP;
4646 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4647 req_set_fail_links(req);
4648 __io_req_complete(req, issue_flags, ret, cflags);
4652 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4654 struct io_buffer *kbuf;
4655 struct io_sr_msg *sr = &req->sr_msg;
4657 void __user *buf = sr->buf;
4658 struct socket *sock;
4662 int ret, cflags = 0;
4663 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4665 sock = sock_from_file(req->file);
4666 if (unlikely(!sock))
4669 if (req->flags & REQ_F_BUFFER_SELECT) {
4670 kbuf = io_recv_buffer_select(req, !force_nonblock);
4672 return PTR_ERR(kbuf);
4673 buf = u64_to_user_ptr(kbuf->addr);
4676 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4680 msg.msg_name = NULL;
4681 msg.msg_control = NULL;
4682 msg.msg_controllen = 0;
4683 msg.msg_namelen = 0;
4684 msg.msg_iocb = NULL;
4687 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4688 if (flags & MSG_DONTWAIT)
4689 req->flags |= REQ_F_NOWAIT;
4690 else if (force_nonblock)
4691 flags |= MSG_DONTWAIT;
4693 if (flags & MSG_WAITALL)
4694 min_ret = iov_iter_count(&msg.msg_iter);
4696 ret = sock_recvmsg(sock, &msg, flags);
4697 if (force_nonblock && ret == -EAGAIN)
4699 if (ret == -ERESTARTSYS)
4702 if (req->flags & REQ_F_BUFFER_SELECTED)
4703 cflags = io_put_recv_kbuf(req);
4704 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4705 req_set_fail_links(req);
4706 __io_req_complete(req, issue_flags, ret, cflags);
4710 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4712 struct io_accept *accept = &req->accept;
4714 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4716 if (sqe->ioprio || sqe->len || sqe->buf_index)
4719 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4720 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4721 accept->flags = READ_ONCE(sqe->accept_flags);
4722 accept->nofile = rlimit(RLIMIT_NOFILE);
4726 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4728 struct io_accept *accept = &req->accept;
4729 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4730 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4733 if (req->file->f_flags & O_NONBLOCK)
4734 req->flags |= REQ_F_NOWAIT;
4736 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4737 accept->addr_len, accept->flags,
4739 if (ret == -EAGAIN && force_nonblock)
4742 if (ret == -ERESTARTSYS)
4744 req_set_fail_links(req);
4746 __io_req_complete(req, issue_flags, ret, 0);
4750 static int io_connect_prep_async(struct io_kiocb *req)
4752 struct io_async_connect *io = req->async_data;
4753 struct io_connect *conn = &req->connect;
4755 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4758 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4760 struct io_connect *conn = &req->connect;
4762 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4764 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4767 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4768 conn->addr_len = READ_ONCE(sqe->addr2);
4772 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4774 struct io_async_connect __io, *io;
4775 unsigned file_flags;
4777 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4779 if (req->async_data) {
4780 io = req->async_data;
4782 ret = move_addr_to_kernel(req->connect.addr,
4783 req->connect.addr_len,
4790 file_flags = force_nonblock ? O_NONBLOCK : 0;
4792 ret = __sys_connect_file(req->file, &io->address,
4793 req->connect.addr_len, file_flags);
4794 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4795 if (req->async_data)
4797 if (io_alloc_async_data(req)) {
4801 memcpy(req->async_data, &__io, sizeof(__io));
4804 if (ret == -ERESTARTSYS)
4808 req_set_fail_links(req);
4809 __io_req_complete(req, issue_flags, ret, 0);
4812 #else /* !CONFIG_NET */
4813 #define IO_NETOP_FN(op) \
4814 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4816 return -EOPNOTSUPP; \
4819 #define IO_NETOP_PREP(op) \
4821 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4823 return -EOPNOTSUPP; \
4826 #define IO_NETOP_PREP_ASYNC(op) \
4828 static int io_##op##_prep_async(struct io_kiocb *req) \
4830 return -EOPNOTSUPP; \
4833 IO_NETOP_PREP_ASYNC(sendmsg);
4834 IO_NETOP_PREP_ASYNC(recvmsg);
4835 IO_NETOP_PREP_ASYNC(connect);
4836 IO_NETOP_PREP(accept);
4839 #endif /* CONFIG_NET */
4841 struct io_poll_table {
4842 struct poll_table_struct pt;
4843 struct io_kiocb *req;
4847 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4848 __poll_t mask, task_work_func_t func)
4852 /* for instances that support it check for an event match first: */
4853 if (mask && !(mask & poll->events))
4856 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4858 list_del_init(&poll->wait.entry);
4861 req->task_work.func = func;
4862 percpu_ref_get(&req->ctx->refs);
4865 * If this fails, then the task is exiting. When a task exits, the
4866 * work gets canceled, so just cancel this request as well instead
4867 * of executing it. We can't safely execute it anyway, as we may not
4868 * have the needed state needed for it anyway.
4870 ret = io_req_task_work_add(req);
4871 if (unlikely(ret)) {
4872 WRITE_ONCE(poll->canceled, true);
4873 io_req_task_work_add_fallback(req, func);
4878 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4879 __acquires(&req->ctx->completion_lock)
4881 struct io_ring_ctx *ctx = req->ctx;
4883 if (!req->result && !READ_ONCE(poll->canceled)) {
4884 struct poll_table_struct pt = { ._key = poll->events };
4886 req->result = vfs_poll(req->file, &pt) & poll->events;
4889 spin_lock_irq(&ctx->completion_lock);
4890 if (!req->result && !READ_ONCE(poll->canceled)) {
4891 add_wait_queue(poll->head, &poll->wait);
4898 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4900 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4901 if (req->opcode == IORING_OP_POLL_ADD)
4902 return req->async_data;
4903 return req->apoll->double_poll;
4906 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4908 if (req->opcode == IORING_OP_POLL_ADD)
4910 return &req->apoll->poll;
4913 static void io_poll_remove_double(struct io_kiocb *req)
4915 struct io_poll_iocb *poll = io_poll_get_double(req);
4917 lockdep_assert_held(&req->ctx->completion_lock);
4919 if (poll && poll->head) {
4920 struct wait_queue_head *head = poll->head;
4922 spin_lock(&head->lock);
4923 list_del_init(&poll->wait.entry);
4924 if (poll->wait.private)
4925 refcount_dec(&req->refs);
4927 spin_unlock(&head->lock);
4931 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4933 struct io_ring_ctx *ctx = req->ctx;
4935 io_poll_remove_double(req);
4936 req->poll.done = true;
4937 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4938 io_commit_cqring(ctx);
4941 static void io_poll_task_func(struct callback_head *cb)
4943 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4944 struct io_ring_ctx *ctx = req->ctx;
4945 struct io_kiocb *nxt;
4947 if (io_poll_rewait(req, &req->poll)) {
4948 spin_unlock_irq(&ctx->completion_lock);
4950 hash_del(&req->hash_node);
4951 io_poll_complete(req, req->result, 0);
4952 spin_unlock_irq(&ctx->completion_lock);
4954 nxt = io_put_req_find_next(req);
4955 io_cqring_ev_posted(ctx);
4957 __io_req_task_submit(nxt);
4960 percpu_ref_put(&ctx->refs);
4963 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4964 int sync, void *key)
4966 struct io_kiocb *req = wait->private;
4967 struct io_poll_iocb *poll = io_poll_get_single(req);
4968 __poll_t mask = key_to_poll(key);
4970 /* for instances that support it check for an event match first: */
4971 if (mask && !(mask & poll->events))
4974 list_del_init(&wait->entry);
4976 if (poll && poll->head) {
4979 spin_lock(&poll->head->lock);
4980 done = list_empty(&poll->wait.entry);
4982 list_del_init(&poll->wait.entry);
4983 /* make sure double remove sees this as being gone */
4984 wait->private = NULL;
4985 spin_unlock(&poll->head->lock);
4987 /* use wait func handler, so it matches the rq type */
4988 poll->wait.func(&poll->wait, mode, sync, key);
4991 refcount_dec(&req->refs);
4995 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4996 wait_queue_func_t wake_func)
5000 poll->canceled = false;
5001 poll->events = events;
5002 INIT_LIST_HEAD(&poll->wait.entry);
5003 init_waitqueue_func_entry(&poll->wait, wake_func);
5006 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5007 struct wait_queue_head *head,
5008 struct io_poll_iocb **poll_ptr)
5010 struct io_kiocb *req = pt->req;
5013 * If poll->head is already set, it's because the file being polled
5014 * uses multiple waitqueues for poll handling (eg one for read, one
5015 * for write). Setup a separate io_poll_iocb if this happens.
5017 if (unlikely(poll->head)) {
5018 struct io_poll_iocb *poll_one = poll;
5020 /* already have a 2nd entry, fail a third attempt */
5022 pt->error = -EINVAL;
5025 /* double add on the same waitqueue head, ignore */
5026 if (poll->head == head)
5028 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5030 pt->error = -ENOMEM;
5033 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5034 refcount_inc(&req->refs);
5035 poll->wait.private = req;
5042 if (poll->events & EPOLLEXCLUSIVE)
5043 add_wait_queue_exclusive(head, &poll->wait);
5045 add_wait_queue(head, &poll->wait);
5048 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5049 struct poll_table_struct *p)
5051 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5052 struct async_poll *apoll = pt->req->apoll;
5054 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5057 static void io_async_task_func(struct callback_head *cb)
5059 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5060 struct async_poll *apoll = req->apoll;
5061 struct io_ring_ctx *ctx = req->ctx;
5063 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5065 if (io_poll_rewait(req, &apoll->poll)) {
5066 spin_unlock_irq(&ctx->completion_lock);
5067 percpu_ref_put(&ctx->refs);
5071 /* If req is still hashed, it cannot have been canceled. Don't check. */
5072 if (hash_hashed(&req->hash_node))
5073 hash_del(&req->hash_node);
5075 io_poll_remove_double(req);
5076 spin_unlock_irq(&ctx->completion_lock);
5078 if (!READ_ONCE(apoll->poll.canceled))
5079 __io_req_task_submit(req);
5081 __io_req_task_cancel(req, -ECANCELED);
5083 percpu_ref_put(&ctx->refs);
5084 kfree(apoll->double_poll);
5088 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5091 struct io_kiocb *req = wait->private;
5092 struct io_poll_iocb *poll = &req->apoll->poll;
5094 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5097 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5100 static void io_poll_req_insert(struct io_kiocb *req)
5102 struct io_ring_ctx *ctx = req->ctx;
5103 struct hlist_head *list;
5105 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5106 hlist_add_head(&req->hash_node, list);
5109 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5110 struct io_poll_iocb *poll,
5111 struct io_poll_table *ipt, __poll_t mask,
5112 wait_queue_func_t wake_func)
5113 __acquires(&ctx->completion_lock)
5115 struct io_ring_ctx *ctx = req->ctx;
5116 bool cancel = false;
5118 INIT_HLIST_NODE(&req->hash_node);
5119 io_init_poll_iocb(poll, mask, wake_func);
5120 poll->file = req->file;
5121 poll->wait.private = req;
5123 ipt->pt._key = mask;
5125 ipt->error = -EINVAL;
5127 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5129 spin_lock_irq(&ctx->completion_lock);
5130 if (likely(poll->head)) {
5131 spin_lock(&poll->head->lock);
5132 if (unlikely(list_empty(&poll->wait.entry))) {
5138 if (mask || ipt->error)
5139 list_del_init(&poll->wait.entry);
5141 WRITE_ONCE(poll->canceled, true);
5142 else if (!poll->done) /* actually waiting for an event */
5143 io_poll_req_insert(req);
5144 spin_unlock(&poll->head->lock);
5150 static bool io_arm_poll_handler(struct io_kiocb *req)
5152 const struct io_op_def *def = &io_op_defs[req->opcode];
5153 struct io_ring_ctx *ctx = req->ctx;
5154 struct async_poll *apoll;
5155 struct io_poll_table ipt;
5159 if (!req->file || !file_can_poll(req->file))
5161 if (req->flags & REQ_F_POLLED)
5165 else if (def->pollout)
5169 /* if we can't nonblock try, then no point in arming a poll handler */
5170 if (!io_file_supports_async(req->file, rw))
5173 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5174 if (unlikely(!apoll))
5176 apoll->double_poll = NULL;
5178 req->flags |= REQ_F_POLLED;
5183 mask |= POLLIN | POLLRDNORM;
5185 mask |= POLLOUT | POLLWRNORM;
5187 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5188 if ((req->opcode == IORING_OP_RECVMSG) &&
5189 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5192 mask |= POLLERR | POLLPRI;
5194 ipt.pt._qproc = io_async_queue_proc;
5196 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5198 if (ret || ipt.error) {
5199 io_poll_remove_double(req);
5200 spin_unlock_irq(&ctx->completion_lock);
5201 kfree(apoll->double_poll);
5205 spin_unlock_irq(&ctx->completion_lock);
5206 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5207 apoll->poll.events);
5211 static bool __io_poll_remove_one(struct io_kiocb *req,
5212 struct io_poll_iocb *poll)
5214 bool do_complete = false;
5216 spin_lock(&poll->head->lock);
5217 WRITE_ONCE(poll->canceled, true);
5218 if (!list_empty(&poll->wait.entry)) {
5219 list_del_init(&poll->wait.entry);
5222 spin_unlock(&poll->head->lock);
5223 hash_del(&req->hash_node);
5227 static bool io_poll_remove_one(struct io_kiocb *req)
5231 io_poll_remove_double(req);
5233 if (req->opcode == IORING_OP_POLL_ADD) {
5234 do_complete = __io_poll_remove_one(req, &req->poll);
5236 struct async_poll *apoll = req->apoll;
5238 /* non-poll requests have submit ref still */
5239 do_complete = __io_poll_remove_one(req, &apoll->poll);
5242 kfree(apoll->double_poll);
5248 io_cqring_fill_event(req, -ECANCELED);
5249 io_commit_cqring(req->ctx);
5250 req_set_fail_links(req);
5251 io_put_req_deferred(req, 1);
5258 * Returns true if we found and killed one or more poll requests
5260 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5261 struct files_struct *files)
5263 struct hlist_node *tmp;
5264 struct io_kiocb *req;
5267 spin_lock_irq(&ctx->completion_lock);
5268 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5269 struct hlist_head *list;
5271 list = &ctx->cancel_hash[i];
5272 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5273 if (io_match_task(req, tsk, files))
5274 posted += io_poll_remove_one(req);
5277 spin_unlock_irq(&ctx->completion_lock);
5280 io_cqring_ev_posted(ctx);
5285 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5287 struct hlist_head *list;
5288 struct io_kiocb *req;
5290 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5291 hlist_for_each_entry(req, list, hash_node) {
5292 if (sqe_addr != req->user_data)
5294 if (io_poll_remove_one(req))
5302 static int io_poll_remove_prep(struct io_kiocb *req,
5303 const struct io_uring_sqe *sqe)
5305 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5307 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5311 req->poll_remove.addr = READ_ONCE(sqe->addr);
5316 * Find a running poll command that matches one specified in sqe->addr,
5317 * and remove it if found.
5319 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5321 struct io_ring_ctx *ctx = req->ctx;
5324 spin_lock_irq(&ctx->completion_lock);
5325 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5326 spin_unlock_irq(&ctx->completion_lock);
5329 req_set_fail_links(req);
5330 io_req_complete(req, ret);
5334 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5337 struct io_kiocb *req = wait->private;
5338 struct io_poll_iocb *poll = &req->poll;
5340 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5343 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5344 struct poll_table_struct *p)
5346 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5348 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5351 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5353 struct io_poll_iocb *poll = &req->poll;
5356 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5358 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5361 events = READ_ONCE(sqe->poll32_events);
5363 events = swahw32(events);
5365 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5366 (events & EPOLLEXCLUSIVE);
5370 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5372 struct io_poll_iocb *poll = &req->poll;
5373 struct io_ring_ctx *ctx = req->ctx;
5374 struct io_poll_table ipt;
5377 ipt.pt._qproc = io_poll_queue_proc;
5379 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5382 if (mask) { /* no async, we'd stolen it */
5384 io_poll_complete(req, mask, 0);
5386 spin_unlock_irq(&ctx->completion_lock);
5389 io_cqring_ev_posted(ctx);
5395 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5397 struct io_timeout_data *data = container_of(timer,
5398 struct io_timeout_data, timer);
5399 struct io_kiocb *req = data->req;
5400 struct io_ring_ctx *ctx = req->ctx;
5401 unsigned long flags;
5403 spin_lock_irqsave(&ctx->completion_lock, flags);
5404 list_del_init(&req->timeout.list);
5405 atomic_set(&req->ctx->cq_timeouts,
5406 atomic_read(&req->ctx->cq_timeouts) + 1);
5408 io_cqring_fill_event(req, -ETIME);
5409 io_commit_cqring(ctx);
5410 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5412 io_cqring_ev_posted(ctx);
5413 req_set_fail_links(req);
5415 return HRTIMER_NORESTART;
5418 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5421 struct io_timeout_data *io;
5422 struct io_kiocb *req;
5425 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5426 if (user_data == req->user_data) {
5433 return ERR_PTR(ret);
5435 io = req->async_data;
5436 ret = hrtimer_try_to_cancel(&io->timer);
5438 return ERR_PTR(-EALREADY);
5439 list_del_init(&req->timeout.list);
5443 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5445 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5448 return PTR_ERR(req);
5450 req_set_fail_links(req);
5451 io_cqring_fill_event(req, -ECANCELED);
5452 io_put_req_deferred(req, 1);
5456 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5457 struct timespec64 *ts, enum hrtimer_mode mode)
5459 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5460 struct io_timeout_data *data;
5463 return PTR_ERR(req);
5465 req->timeout.off = 0; /* noseq */
5466 data = req->async_data;
5467 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5468 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5469 data->timer.function = io_timeout_fn;
5470 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5474 static int io_timeout_remove_prep(struct io_kiocb *req,
5475 const struct io_uring_sqe *sqe)
5477 struct io_timeout_rem *tr = &req->timeout_rem;
5479 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5481 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5483 if (sqe->ioprio || sqe->buf_index || sqe->len)
5486 tr->addr = READ_ONCE(sqe->addr);
5487 tr->flags = READ_ONCE(sqe->timeout_flags);
5488 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5489 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5491 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5493 } else if (tr->flags) {
5494 /* timeout removal doesn't support flags */
5501 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5503 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5508 * Remove or update an existing timeout command
5510 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5512 struct io_timeout_rem *tr = &req->timeout_rem;
5513 struct io_ring_ctx *ctx = req->ctx;
5516 spin_lock_irq(&ctx->completion_lock);
5517 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5518 ret = io_timeout_cancel(ctx, tr->addr);
5520 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5521 io_translate_timeout_mode(tr->flags));
5523 io_cqring_fill_event(req, ret);
5524 io_commit_cqring(ctx);
5525 spin_unlock_irq(&ctx->completion_lock);
5526 io_cqring_ev_posted(ctx);
5528 req_set_fail_links(req);
5533 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5534 bool is_timeout_link)
5536 struct io_timeout_data *data;
5538 u32 off = READ_ONCE(sqe->off);
5540 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5542 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5544 if (off && is_timeout_link)
5546 flags = READ_ONCE(sqe->timeout_flags);
5547 if (flags & ~IORING_TIMEOUT_ABS)
5550 req->timeout.off = off;
5552 if (!req->async_data && io_alloc_async_data(req))
5555 data = req->async_data;
5558 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5561 data->mode = io_translate_timeout_mode(flags);
5562 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5563 if (is_timeout_link)
5564 io_req_track_inflight(req);
5568 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5570 struct io_ring_ctx *ctx = req->ctx;
5571 struct io_timeout_data *data = req->async_data;
5572 struct list_head *entry;
5573 u32 tail, off = req->timeout.off;
5575 spin_lock_irq(&ctx->completion_lock);
5578 * sqe->off holds how many events that need to occur for this
5579 * timeout event to be satisfied. If it isn't set, then this is
5580 * a pure timeout request, sequence isn't used.
5582 if (io_is_timeout_noseq(req)) {
5583 entry = ctx->timeout_list.prev;
5587 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5588 req->timeout.target_seq = tail + off;
5590 /* Update the last seq here in case io_flush_timeouts() hasn't.
5591 * This is safe because ->completion_lock is held, and submissions
5592 * and completions are never mixed in the same ->completion_lock section.
5594 ctx->cq_last_tm_flush = tail;
5597 * Insertion sort, ensuring the first entry in the list is always
5598 * the one we need first.
5600 list_for_each_prev(entry, &ctx->timeout_list) {
5601 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5604 if (io_is_timeout_noseq(nxt))
5606 /* nxt.seq is behind @tail, otherwise would've been completed */
5607 if (off >= nxt->timeout.target_seq - tail)
5611 list_add(&req->timeout.list, entry);
5612 data->timer.function = io_timeout_fn;
5613 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5614 spin_unlock_irq(&ctx->completion_lock);
5618 struct io_cancel_data {
5619 struct io_ring_ctx *ctx;
5623 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5625 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5626 struct io_cancel_data *cd = data;
5628 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5631 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5632 struct io_ring_ctx *ctx)
5634 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5635 enum io_wq_cancel cancel_ret;
5638 if (!tctx || !tctx->io_wq)
5641 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5642 switch (cancel_ret) {
5643 case IO_WQ_CANCEL_OK:
5646 case IO_WQ_CANCEL_RUNNING:
5649 case IO_WQ_CANCEL_NOTFOUND:
5657 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5658 struct io_kiocb *req, __u64 sqe_addr,
5661 unsigned long flags;
5664 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5665 if (ret != -ENOENT) {
5666 spin_lock_irqsave(&ctx->completion_lock, flags);
5670 spin_lock_irqsave(&ctx->completion_lock, flags);
5671 ret = io_timeout_cancel(ctx, sqe_addr);
5674 ret = io_poll_cancel(ctx, sqe_addr);
5678 io_cqring_fill_event(req, ret);
5679 io_commit_cqring(ctx);
5680 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5681 io_cqring_ev_posted(ctx);
5684 req_set_fail_links(req);
5688 static int io_async_cancel_prep(struct io_kiocb *req,
5689 const struct io_uring_sqe *sqe)
5691 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5693 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5695 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5698 req->cancel.addr = READ_ONCE(sqe->addr);
5702 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5704 struct io_ring_ctx *ctx = req->ctx;
5705 u64 sqe_addr = req->cancel.addr;
5706 struct io_tctx_node *node;
5709 /* tasks should wait for their io-wq threads, so safe w/o sync */
5710 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5711 spin_lock_irq(&ctx->completion_lock);
5714 ret = io_timeout_cancel(ctx, sqe_addr);
5717 ret = io_poll_cancel(ctx, sqe_addr);
5720 spin_unlock_irq(&ctx->completion_lock);
5722 /* slow path, try all io-wq's */
5723 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5725 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5726 struct io_uring_task *tctx = node->task->io_uring;
5728 if (!tctx || !tctx->io_wq)
5730 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5734 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5736 spin_lock_irq(&ctx->completion_lock);
5738 io_cqring_fill_event(req, ret);
5739 io_commit_cqring(ctx);
5740 spin_unlock_irq(&ctx->completion_lock);
5741 io_cqring_ev_posted(ctx);
5744 req_set_fail_links(req);
5749 static int io_rsrc_update_prep(struct io_kiocb *req,
5750 const struct io_uring_sqe *sqe)
5752 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5754 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5756 if (sqe->ioprio || sqe->rw_flags)
5759 req->rsrc_update.offset = READ_ONCE(sqe->off);
5760 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5761 if (!req->rsrc_update.nr_args)
5763 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5767 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5769 struct io_ring_ctx *ctx = req->ctx;
5770 struct io_uring_rsrc_update up;
5773 if (issue_flags & IO_URING_F_NONBLOCK)
5776 up.offset = req->rsrc_update.offset;
5777 up.data = req->rsrc_update.arg;
5779 mutex_lock(&ctx->uring_lock);
5780 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5781 mutex_unlock(&ctx->uring_lock);
5784 req_set_fail_links(req);
5785 __io_req_complete(req, issue_flags, ret, 0);
5789 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5791 switch (req->opcode) {
5794 case IORING_OP_READV:
5795 case IORING_OP_READ_FIXED:
5796 case IORING_OP_READ:
5797 return io_read_prep(req, sqe);
5798 case IORING_OP_WRITEV:
5799 case IORING_OP_WRITE_FIXED:
5800 case IORING_OP_WRITE:
5801 return io_write_prep(req, sqe);
5802 case IORING_OP_POLL_ADD:
5803 return io_poll_add_prep(req, sqe);
5804 case IORING_OP_POLL_REMOVE:
5805 return io_poll_remove_prep(req, sqe);
5806 case IORING_OP_FSYNC:
5807 return io_fsync_prep(req, sqe);
5808 case IORING_OP_SYNC_FILE_RANGE:
5809 return io_sfr_prep(req, sqe);
5810 case IORING_OP_SENDMSG:
5811 case IORING_OP_SEND:
5812 return io_sendmsg_prep(req, sqe);
5813 case IORING_OP_RECVMSG:
5814 case IORING_OP_RECV:
5815 return io_recvmsg_prep(req, sqe);
5816 case IORING_OP_CONNECT:
5817 return io_connect_prep(req, sqe);
5818 case IORING_OP_TIMEOUT:
5819 return io_timeout_prep(req, sqe, false);
5820 case IORING_OP_TIMEOUT_REMOVE:
5821 return io_timeout_remove_prep(req, sqe);
5822 case IORING_OP_ASYNC_CANCEL:
5823 return io_async_cancel_prep(req, sqe);
5824 case IORING_OP_LINK_TIMEOUT:
5825 return io_timeout_prep(req, sqe, true);
5826 case IORING_OP_ACCEPT:
5827 return io_accept_prep(req, sqe);
5828 case IORING_OP_FALLOCATE:
5829 return io_fallocate_prep(req, sqe);
5830 case IORING_OP_OPENAT:
5831 return io_openat_prep(req, sqe);
5832 case IORING_OP_CLOSE:
5833 return io_close_prep(req, sqe);
5834 case IORING_OP_FILES_UPDATE:
5835 return io_rsrc_update_prep(req, sqe);
5836 case IORING_OP_STATX:
5837 return io_statx_prep(req, sqe);
5838 case IORING_OP_FADVISE:
5839 return io_fadvise_prep(req, sqe);
5840 case IORING_OP_MADVISE:
5841 return io_madvise_prep(req, sqe);
5842 case IORING_OP_OPENAT2:
5843 return io_openat2_prep(req, sqe);
5844 case IORING_OP_EPOLL_CTL:
5845 return io_epoll_ctl_prep(req, sqe);
5846 case IORING_OP_SPLICE:
5847 return io_splice_prep(req, sqe);
5848 case IORING_OP_PROVIDE_BUFFERS:
5849 return io_provide_buffers_prep(req, sqe);
5850 case IORING_OP_REMOVE_BUFFERS:
5851 return io_remove_buffers_prep(req, sqe);
5853 return io_tee_prep(req, sqe);
5854 case IORING_OP_SHUTDOWN:
5855 return io_shutdown_prep(req, sqe);
5856 case IORING_OP_RENAMEAT:
5857 return io_renameat_prep(req, sqe);
5858 case IORING_OP_UNLINKAT:
5859 return io_unlinkat_prep(req, sqe);
5862 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5867 static int io_req_prep_async(struct io_kiocb *req)
5869 switch (req->opcode) {
5870 case IORING_OP_READV:
5871 case IORING_OP_READ_FIXED:
5872 case IORING_OP_READ:
5873 return io_rw_prep_async(req, READ);
5874 case IORING_OP_WRITEV:
5875 case IORING_OP_WRITE_FIXED:
5876 case IORING_OP_WRITE:
5877 return io_rw_prep_async(req, WRITE);
5878 case IORING_OP_SENDMSG:
5879 case IORING_OP_SEND:
5880 return io_sendmsg_prep_async(req);
5881 case IORING_OP_RECVMSG:
5882 case IORING_OP_RECV:
5883 return io_recvmsg_prep_async(req);
5884 case IORING_OP_CONNECT:
5885 return io_connect_prep_async(req);
5890 static int io_req_defer_prep(struct io_kiocb *req)
5892 if (!io_op_defs[req->opcode].needs_async_data)
5894 /* some opcodes init it during the inital prep */
5895 if (req->async_data)
5897 if (__io_alloc_async_data(req))
5899 return io_req_prep_async(req);
5902 static u32 io_get_sequence(struct io_kiocb *req)
5904 struct io_kiocb *pos;
5905 struct io_ring_ctx *ctx = req->ctx;
5906 u32 total_submitted, nr_reqs = 0;
5908 io_for_each_link(pos, req)
5911 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5912 return total_submitted - nr_reqs;
5915 static int io_req_defer(struct io_kiocb *req)
5917 struct io_ring_ctx *ctx = req->ctx;
5918 struct io_defer_entry *de;
5922 /* Still need defer if there is pending req in defer list. */
5923 if (likely(list_empty_careful(&ctx->defer_list) &&
5924 !(req->flags & REQ_F_IO_DRAIN)))
5927 seq = io_get_sequence(req);
5928 /* Still a chance to pass the sequence check */
5929 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5932 ret = io_req_defer_prep(req);
5935 io_prep_async_link(req);
5936 de = kmalloc(sizeof(*de), GFP_KERNEL);
5940 spin_lock_irq(&ctx->completion_lock);
5941 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5942 spin_unlock_irq(&ctx->completion_lock);
5944 io_queue_async_work(req);
5945 return -EIOCBQUEUED;
5948 trace_io_uring_defer(ctx, req, req->user_data);
5951 list_add_tail(&de->list, &ctx->defer_list);
5952 spin_unlock_irq(&ctx->completion_lock);
5953 return -EIOCBQUEUED;
5956 static void __io_clean_op(struct io_kiocb *req)
5958 if (req->flags & REQ_F_BUFFER_SELECTED) {
5959 switch (req->opcode) {
5960 case IORING_OP_READV:
5961 case IORING_OP_READ_FIXED:
5962 case IORING_OP_READ:
5963 kfree((void *)(unsigned long)req->rw.addr);
5965 case IORING_OP_RECVMSG:
5966 case IORING_OP_RECV:
5967 kfree(req->sr_msg.kbuf);
5970 req->flags &= ~REQ_F_BUFFER_SELECTED;
5973 if (req->flags & REQ_F_NEED_CLEANUP) {
5974 switch (req->opcode) {
5975 case IORING_OP_READV:
5976 case IORING_OP_READ_FIXED:
5977 case IORING_OP_READ:
5978 case IORING_OP_WRITEV:
5979 case IORING_OP_WRITE_FIXED:
5980 case IORING_OP_WRITE: {
5981 struct io_async_rw *io = req->async_data;
5983 kfree(io->free_iovec);
5986 case IORING_OP_RECVMSG:
5987 case IORING_OP_SENDMSG: {
5988 struct io_async_msghdr *io = req->async_data;
5990 kfree(io->free_iov);
5993 case IORING_OP_SPLICE:
5995 io_put_file(req, req->splice.file_in,
5996 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5998 case IORING_OP_OPENAT:
5999 case IORING_OP_OPENAT2:
6000 if (req->open.filename)
6001 putname(req->open.filename);
6003 case IORING_OP_RENAMEAT:
6004 putname(req->rename.oldpath);
6005 putname(req->rename.newpath);
6007 case IORING_OP_UNLINKAT:
6008 putname(req->unlink.filename);
6011 req->flags &= ~REQ_F_NEED_CLEANUP;
6015 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6017 struct io_ring_ctx *ctx = req->ctx;
6018 const struct cred *creds = NULL;
6021 if (req->work.creds && req->work.creds != current_cred())
6022 creds = override_creds(req->work.creds);
6024 switch (req->opcode) {
6026 ret = io_nop(req, issue_flags);
6028 case IORING_OP_READV:
6029 case IORING_OP_READ_FIXED:
6030 case IORING_OP_READ:
6031 ret = io_read(req, issue_flags);
6033 case IORING_OP_WRITEV:
6034 case IORING_OP_WRITE_FIXED:
6035 case IORING_OP_WRITE:
6036 ret = io_write(req, issue_flags);
6038 case IORING_OP_FSYNC:
6039 ret = io_fsync(req, issue_flags);
6041 case IORING_OP_POLL_ADD:
6042 ret = io_poll_add(req, issue_flags);
6044 case IORING_OP_POLL_REMOVE:
6045 ret = io_poll_remove(req, issue_flags);
6047 case IORING_OP_SYNC_FILE_RANGE:
6048 ret = io_sync_file_range(req, issue_flags);
6050 case IORING_OP_SENDMSG:
6051 ret = io_sendmsg(req, issue_flags);
6053 case IORING_OP_SEND:
6054 ret = io_send(req, issue_flags);
6056 case IORING_OP_RECVMSG:
6057 ret = io_recvmsg(req, issue_flags);
6059 case IORING_OP_RECV:
6060 ret = io_recv(req, issue_flags);
6062 case IORING_OP_TIMEOUT:
6063 ret = io_timeout(req, issue_flags);
6065 case IORING_OP_TIMEOUT_REMOVE:
6066 ret = io_timeout_remove(req, issue_flags);
6068 case IORING_OP_ACCEPT:
6069 ret = io_accept(req, issue_flags);
6071 case IORING_OP_CONNECT:
6072 ret = io_connect(req, issue_flags);
6074 case IORING_OP_ASYNC_CANCEL:
6075 ret = io_async_cancel(req, issue_flags);
6077 case IORING_OP_FALLOCATE:
6078 ret = io_fallocate(req, issue_flags);
6080 case IORING_OP_OPENAT:
6081 ret = io_openat(req, issue_flags);
6083 case IORING_OP_CLOSE:
6084 ret = io_close(req, issue_flags);
6086 case IORING_OP_FILES_UPDATE:
6087 ret = io_files_update(req, issue_flags);
6089 case IORING_OP_STATX:
6090 ret = io_statx(req, issue_flags);
6092 case IORING_OP_FADVISE:
6093 ret = io_fadvise(req, issue_flags);
6095 case IORING_OP_MADVISE:
6096 ret = io_madvise(req, issue_flags);
6098 case IORING_OP_OPENAT2:
6099 ret = io_openat2(req, issue_flags);
6101 case IORING_OP_EPOLL_CTL:
6102 ret = io_epoll_ctl(req, issue_flags);
6104 case IORING_OP_SPLICE:
6105 ret = io_splice(req, issue_flags);
6107 case IORING_OP_PROVIDE_BUFFERS:
6108 ret = io_provide_buffers(req, issue_flags);
6110 case IORING_OP_REMOVE_BUFFERS:
6111 ret = io_remove_buffers(req, issue_flags);
6114 ret = io_tee(req, issue_flags);
6116 case IORING_OP_SHUTDOWN:
6117 ret = io_shutdown(req, issue_flags);
6119 case IORING_OP_RENAMEAT:
6120 ret = io_renameat(req, issue_flags);
6122 case IORING_OP_UNLINKAT:
6123 ret = io_unlinkat(req, issue_flags);
6131 revert_creds(creds);
6136 /* If the op doesn't have a file, we're not polling for it */
6137 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6138 const bool in_async = io_wq_current_is_worker();
6140 /* workqueue context doesn't hold uring_lock, grab it now */
6142 mutex_lock(&ctx->uring_lock);
6144 io_iopoll_req_issued(req, in_async);
6147 mutex_unlock(&ctx->uring_lock);
6153 static void io_wq_submit_work(struct io_wq_work *work)
6155 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6156 struct io_kiocb *timeout;
6159 timeout = io_prep_linked_timeout(req);
6161 io_queue_linked_timeout(timeout);
6163 if (work->flags & IO_WQ_WORK_CANCEL)
6168 ret = io_issue_sqe(req, 0);
6170 * We can get EAGAIN for polled IO even though we're
6171 * forcing a sync submission from here, since we can't
6172 * wait for request slots on the block side.
6180 /* avoid locking problems by failing it from a clean context */
6182 /* io-wq is going to take one down */
6183 refcount_inc(&req->refs);
6184 io_req_task_queue_fail(req, ret);
6188 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6191 struct fixed_rsrc_table *table;
6193 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6194 return table->files[index & IORING_FILE_TABLE_MASK];
6197 static struct file *io_file_get(struct io_submit_state *state,
6198 struct io_kiocb *req, int fd, bool fixed)
6200 struct io_ring_ctx *ctx = req->ctx;
6204 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6206 fd = array_index_nospec(fd, ctx->nr_user_files);
6207 file = io_file_from_index(ctx, fd);
6208 io_set_resource_node(req);
6210 trace_io_uring_file_get(ctx, fd);
6211 file = __io_file_get(state, fd);
6214 if (file && unlikely(file->f_op == &io_uring_fops))
6215 io_req_track_inflight(req);
6219 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6221 struct io_timeout_data *data = container_of(timer,
6222 struct io_timeout_data, timer);
6223 struct io_kiocb *prev, *req = data->req;
6224 struct io_ring_ctx *ctx = req->ctx;
6225 unsigned long flags;
6227 spin_lock_irqsave(&ctx->completion_lock, flags);
6228 prev = req->timeout.head;
6229 req->timeout.head = NULL;
6232 * We don't expect the list to be empty, that will only happen if we
6233 * race with the completion of the linked work.
6235 if (prev && refcount_inc_not_zero(&prev->refs))
6236 io_remove_next_linked(prev);
6239 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6242 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6243 io_put_req_deferred(prev, 1);
6245 io_req_complete_post(req, -ETIME, 0);
6246 io_put_req_deferred(req, 1);
6248 return HRTIMER_NORESTART;
6251 static void __io_queue_linked_timeout(struct io_kiocb *req)
6254 * If the back reference is NULL, then our linked request finished
6255 * before we got a chance to setup the timer
6257 if (req->timeout.head) {
6258 struct io_timeout_data *data = req->async_data;
6260 data->timer.function = io_link_timeout_fn;
6261 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6266 static void io_queue_linked_timeout(struct io_kiocb *req)
6268 struct io_ring_ctx *ctx = req->ctx;
6270 spin_lock_irq(&ctx->completion_lock);
6271 __io_queue_linked_timeout(req);
6272 spin_unlock_irq(&ctx->completion_lock);
6274 /* drop submission reference */
6278 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6280 struct io_kiocb *nxt = req->link;
6282 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6283 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6286 nxt->timeout.head = req;
6287 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6288 req->flags |= REQ_F_LINK_TIMEOUT;
6292 static void __io_queue_sqe(struct io_kiocb *req)
6294 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6297 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6300 * We async punt it if the file wasn't marked NOWAIT, or if the file
6301 * doesn't support non-blocking read/write attempts
6303 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6304 if (!io_arm_poll_handler(req)) {
6306 * Queued up for async execution, worker will release
6307 * submit reference when the iocb is actually submitted.
6309 io_queue_async_work(req);
6311 } else if (likely(!ret)) {
6312 /* drop submission reference */
6313 if (req->flags & REQ_F_COMPLETE_INLINE) {
6314 struct io_ring_ctx *ctx = req->ctx;
6315 struct io_comp_state *cs = &ctx->submit_state.comp;
6317 cs->reqs[cs->nr++] = req;
6318 if (cs->nr == ARRAY_SIZE(cs->reqs))
6319 io_submit_flush_completions(cs, ctx);
6324 req_set_fail_links(req);
6326 io_req_complete(req, ret);
6329 io_queue_linked_timeout(linked_timeout);
6332 static void io_queue_sqe(struct io_kiocb *req)
6336 ret = io_req_defer(req);
6338 if (ret != -EIOCBQUEUED) {
6340 req_set_fail_links(req);
6342 io_req_complete(req, ret);
6344 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6345 ret = io_req_defer_prep(req);
6348 io_queue_async_work(req);
6350 __io_queue_sqe(req);
6355 * Check SQE restrictions (opcode and flags).
6357 * Returns 'true' if SQE is allowed, 'false' otherwise.
6359 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6360 struct io_kiocb *req,
6361 unsigned int sqe_flags)
6363 if (!ctx->restricted)
6366 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6369 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6370 ctx->restrictions.sqe_flags_required)
6373 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6374 ctx->restrictions.sqe_flags_required))
6380 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6381 const struct io_uring_sqe *sqe)
6383 struct io_submit_state *state;
6384 unsigned int sqe_flags;
6385 int personality, ret = 0;
6387 req->opcode = READ_ONCE(sqe->opcode);
6388 /* same numerical values with corresponding REQ_F_*, safe to copy */
6389 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6390 req->user_data = READ_ONCE(sqe->user_data);
6391 req->async_data = NULL;
6395 req->fixed_rsrc_refs = NULL;
6396 /* one is dropped after submission, the other at completion */
6397 refcount_set(&req->refs, 2);
6398 req->task = current;
6400 req->work.list.next = NULL;
6401 req->work.creds = NULL;
6402 req->work.flags = 0;
6404 /* enforce forwards compatibility on users */
6405 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6410 if (unlikely(req->opcode >= IORING_OP_LAST))
6413 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6416 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6417 !io_op_defs[req->opcode].buffer_select)
6420 personality = READ_ONCE(sqe->personality);
6422 req->work.creds = xa_load(&ctx->personalities, personality);
6423 if (!req->work.creds)
6425 get_cred(req->work.creds);
6427 state = &ctx->submit_state;
6430 * Plug now if we have more than 1 IO left after this, and the target
6431 * is potentially a read/write to block based storage.
6433 if (!state->plug_started && state->ios_left > 1 &&
6434 io_op_defs[req->opcode].plug) {
6435 blk_start_plug(&state->plug);
6436 state->plug_started = true;
6439 if (io_op_defs[req->opcode].needs_file) {
6440 bool fixed = req->flags & REQ_F_FIXED_FILE;
6442 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6443 if (unlikely(!req->file))
6451 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6452 const struct io_uring_sqe *sqe)
6454 struct io_submit_link *link = &ctx->submit_state.link;
6457 ret = io_init_req(ctx, req, sqe);
6458 if (unlikely(ret)) {
6461 /* fail even hard links since we don't submit */
6462 link->head->flags |= REQ_F_FAIL_LINK;
6463 io_put_req(link->head);
6464 io_req_complete(link->head, -ECANCELED);
6468 io_req_complete(req, ret);
6471 ret = io_req_prep(req, sqe);
6475 /* don't need @sqe from now on */
6476 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6477 true, ctx->flags & IORING_SETUP_SQPOLL);
6480 * If we already have a head request, queue this one for async
6481 * submittal once the head completes. If we don't have a head but
6482 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6483 * submitted sync once the chain is complete. If none of those
6484 * conditions are true (normal request), then just queue it.
6487 struct io_kiocb *head = link->head;
6490 * Taking sequential execution of a link, draining both sides
6491 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6492 * requests in the link. So, it drains the head and the
6493 * next after the link request. The last one is done via
6494 * drain_next flag to persist the effect across calls.
6496 if (req->flags & REQ_F_IO_DRAIN) {
6497 head->flags |= REQ_F_IO_DRAIN;
6498 ctx->drain_next = 1;
6500 ret = io_req_defer_prep(req);
6503 trace_io_uring_link(ctx, req, head);
6504 link->last->link = req;
6507 /* last request of a link, enqueue the link */
6508 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6513 if (unlikely(ctx->drain_next)) {
6514 req->flags |= REQ_F_IO_DRAIN;
6515 ctx->drain_next = 0;
6517 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6529 * Batched submission is done, ensure local IO is flushed out.
6531 static void io_submit_state_end(struct io_submit_state *state,
6532 struct io_ring_ctx *ctx)
6534 if (state->link.head)
6535 io_queue_sqe(state->link.head);
6537 io_submit_flush_completions(&state->comp, ctx);
6538 if (state->plug_started)
6539 blk_finish_plug(&state->plug);
6540 io_state_file_put(state);
6544 * Start submission side cache.
6546 static void io_submit_state_start(struct io_submit_state *state,
6547 unsigned int max_ios)
6549 state->plug_started = false;
6550 state->ios_left = max_ios;
6551 /* set only head, no need to init link_last in advance */
6552 state->link.head = NULL;
6555 static void io_commit_sqring(struct io_ring_ctx *ctx)
6557 struct io_rings *rings = ctx->rings;
6560 * Ensure any loads from the SQEs are done at this point,
6561 * since once we write the new head, the application could
6562 * write new data to them.
6564 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6568 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6569 * that is mapped by userspace. This means that care needs to be taken to
6570 * ensure that reads are stable, as we cannot rely on userspace always
6571 * being a good citizen. If members of the sqe are validated and then later
6572 * used, it's important that those reads are done through READ_ONCE() to
6573 * prevent a re-load down the line.
6575 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6577 u32 *sq_array = ctx->sq_array;
6581 * The cached sq head (or cq tail) serves two purposes:
6583 * 1) allows us to batch the cost of updating the user visible
6585 * 2) allows the kernel side to track the head on its own, even
6586 * though the application is the one updating it.
6588 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6589 if (likely(head < ctx->sq_entries))
6590 return &ctx->sq_sqes[head];
6592 /* drop invalid entries */
6593 ctx->cached_sq_dropped++;
6594 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6598 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6602 /* if we have a backlog and couldn't flush it all, return BUSY */
6603 if (test_bit(0, &ctx->sq_check_overflow)) {
6604 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6608 /* make sure SQ entry isn't read before tail */
6609 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6611 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6614 percpu_counter_add(¤t->io_uring->inflight, nr);
6615 refcount_add(nr, ¤t->usage);
6616 io_submit_state_start(&ctx->submit_state, nr);
6618 while (submitted < nr) {
6619 const struct io_uring_sqe *sqe;
6620 struct io_kiocb *req;
6622 req = io_alloc_req(ctx);
6623 if (unlikely(!req)) {
6625 submitted = -EAGAIN;
6628 sqe = io_get_sqe(ctx);
6629 if (unlikely(!sqe)) {
6630 kmem_cache_free(req_cachep, req);
6633 /* will complete beyond this point, count as submitted */
6635 if (io_submit_sqe(ctx, req, sqe))
6639 if (unlikely(submitted != nr)) {
6640 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6641 struct io_uring_task *tctx = current->io_uring;
6642 int unused = nr - ref_used;
6644 percpu_ref_put_many(&ctx->refs, unused);
6645 percpu_counter_sub(&tctx->inflight, unused);
6646 put_task_struct_many(current, unused);
6649 io_submit_state_end(&ctx->submit_state, ctx);
6650 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6651 io_commit_sqring(ctx);
6656 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6658 /* Tell userspace we may need a wakeup call */
6659 spin_lock_irq(&ctx->completion_lock);
6660 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6661 spin_unlock_irq(&ctx->completion_lock);
6664 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6666 spin_lock_irq(&ctx->completion_lock);
6667 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6668 spin_unlock_irq(&ctx->completion_lock);
6671 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6673 unsigned int to_submit;
6676 to_submit = io_sqring_entries(ctx);
6677 /* if we're handling multiple rings, cap submit size for fairness */
6678 if (cap_entries && to_submit > 8)
6681 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6682 unsigned nr_events = 0;
6684 mutex_lock(&ctx->uring_lock);
6685 if (!list_empty(&ctx->iopoll_list))
6686 io_do_iopoll(ctx, &nr_events, 0);
6688 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6689 !(ctx->flags & IORING_SETUP_R_DISABLED))
6690 ret = io_submit_sqes(ctx, to_submit);
6691 mutex_unlock(&ctx->uring_lock);
6694 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6695 wake_up(&ctx->sqo_sq_wait);
6700 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6702 struct io_ring_ctx *ctx;
6703 unsigned sq_thread_idle = 0;
6705 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6706 if (sq_thread_idle < ctx->sq_thread_idle)
6707 sq_thread_idle = ctx->sq_thread_idle;
6710 sqd->sq_thread_idle = sq_thread_idle;
6713 static int io_sq_thread(void *data)
6715 struct io_sq_data *sqd = data;
6716 struct io_ring_ctx *ctx;
6717 unsigned long timeout = 0;
6718 char buf[TASK_COMM_LEN];
6721 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6722 set_task_comm(current, buf);
6723 current->pf_io_worker = NULL;
6725 if (sqd->sq_cpu != -1)
6726 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6728 set_cpus_allowed_ptr(current, cpu_online_mask);
6729 current->flags |= PF_NO_SETAFFINITY;
6731 mutex_lock(&sqd->lock);
6732 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6734 bool cap_entries, sqt_spin, needs_sched;
6736 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6737 mutex_unlock(&sqd->lock);
6739 mutex_lock(&sqd->lock);
6741 io_run_task_work_head(&sqd->park_task_work);
6742 timeout = jiffies + sqd->sq_thread_idle;
6745 if (signal_pending(current)) {
6746 struct ksignal ksig;
6748 if (!get_signal(&ksig))
6753 cap_entries = !list_is_singular(&sqd->ctx_list);
6754 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6755 const struct cred *creds = NULL;
6757 if (ctx->sq_creds != current_cred())
6758 creds = override_creds(ctx->sq_creds);
6759 ret = __io_sq_thread(ctx, cap_entries);
6761 revert_creds(creds);
6762 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6766 if (sqt_spin || !time_after(jiffies, timeout)) {
6770 timeout = jiffies + sqd->sq_thread_idle;
6775 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6776 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6777 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6778 !list_empty_careful(&ctx->iopoll_list)) {
6779 needs_sched = false;
6782 if (io_sqring_entries(ctx)) {
6783 needs_sched = false;
6788 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6789 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6790 io_ring_set_wakeup_flag(ctx);
6792 mutex_unlock(&sqd->lock);
6794 mutex_lock(&sqd->lock);
6795 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6796 io_ring_clear_wakeup_flag(ctx);
6799 finish_wait(&sqd->wait, &wait);
6800 io_run_task_work_head(&sqd->park_task_work);
6801 timeout = jiffies + sqd->sq_thread_idle;
6804 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6805 io_uring_cancel_sqpoll(ctx);
6807 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6808 io_ring_set_wakeup_flag(ctx);
6809 mutex_unlock(&sqd->lock);
6812 io_run_task_work_head(&sqd->park_task_work);
6813 complete(&sqd->exited);
6817 struct io_wait_queue {
6818 struct wait_queue_entry wq;
6819 struct io_ring_ctx *ctx;
6821 unsigned nr_timeouts;
6824 static inline bool io_should_wake(struct io_wait_queue *iowq)
6826 struct io_ring_ctx *ctx = iowq->ctx;
6829 * Wake up if we have enough events, or if a timeout occurred since we
6830 * started waiting. For timeouts, we always want to return to userspace,
6831 * regardless of event count.
6833 return io_cqring_events(ctx) >= iowq->to_wait ||
6834 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6837 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6838 int wake_flags, void *key)
6840 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6844 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6845 * the task, and the next invocation will do it.
6847 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6848 return autoremove_wake_function(curr, mode, wake_flags, key);
6852 static int io_run_task_work_sig(void)
6854 if (io_run_task_work())
6856 if (!signal_pending(current))
6858 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6859 return -ERESTARTSYS;
6863 /* when returns >0, the caller should retry */
6864 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6865 struct io_wait_queue *iowq,
6866 signed long *timeout)
6870 /* make sure we run task_work before checking for signals */
6871 ret = io_run_task_work_sig();
6872 if (ret || io_should_wake(iowq))
6874 /* let the caller flush overflows, retry */
6875 if (test_bit(0, &ctx->cq_check_overflow))
6878 *timeout = schedule_timeout(*timeout);
6879 return !*timeout ? -ETIME : 1;
6883 * Wait until events become available, if we don't already have some. The
6884 * application must reap them itself, as they reside on the shared cq ring.
6886 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6887 const sigset_t __user *sig, size_t sigsz,
6888 struct __kernel_timespec __user *uts)
6890 struct io_wait_queue iowq = {
6893 .func = io_wake_function,
6894 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6897 .to_wait = min_events,
6899 struct io_rings *rings = ctx->rings;
6900 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6904 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6905 if (io_cqring_events(ctx) >= min_events)
6907 if (!io_run_task_work())
6912 #ifdef CONFIG_COMPAT
6913 if (in_compat_syscall())
6914 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6918 ret = set_user_sigmask(sig, sigsz);
6925 struct timespec64 ts;
6927 if (get_timespec64(&ts, uts))
6929 timeout = timespec64_to_jiffies(&ts);
6932 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6933 trace_io_uring_cqring_wait(ctx, min_events);
6935 /* if we can't even flush overflow, don't wait for more */
6936 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6940 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6941 TASK_INTERRUPTIBLE);
6942 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6943 finish_wait(&ctx->wait, &iowq.wq);
6947 restore_saved_sigmask_unless(ret == -EINTR);
6949 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6952 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6954 #if defined(CONFIG_UNIX)
6955 if (ctx->ring_sock) {
6956 struct sock *sock = ctx->ring_sock->sk;
6957 struct sk_buff *skb;
6959 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6965 for (i = 0; i < ctx->nr_user_files; i++) {
6968 file = io_file_from_index(ctx, i);
6975 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6977 struct fixed_rsrc_data *data;
6979 data = container_of(ref, struct fixed_rsrc_data, refs);
6980 complete(&data->done);
6983 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6985 spin_lock_bh(&ctx->rsrc_ref_lock);
6988 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6990 spin_unlock_bh(&ctx->rsrc_ref_lock);
6993 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6994 struct fixed_rsrc_data *rsrc_data,
6995 struct fixed_rsrc_ref_node *ref_node)
6997 io_rsrc_ref_lock(ctx);
6998 rsrc_data->node = ref_node;
6999 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7000 io_rsrc_ref_unlock(ctx);
7001 percpu_ref_get(&rsrc_data->refs);
7004 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7006 struct fixed_rsrc_ref_node *ref_node = NULL;
7008 io_rsrc_ref_lock(ctx);
7009 ref_node = data->node;
7011 io_rsrc_ref_unlock(ctx);
7013 percpu_ref_kill(&ref_node->refs);
7016 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7017 struct io_ring_ctx *ctx,
7018 void (*rsrc_put)(struct io_ring_ctx *ctx,
7019 struct io_rsrc_put *prsrc))
7021 struct fixed_rsrc_ref_node *backup_node;
7027 data->quiesce = true;
7030 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7033 backup_node->rsrc_data = data;
7034 backup_node->rsrc_put = rsrc_put;
7036 io_sqe_rsrc_kill_node(ctx, data);
7037 percpu_ref_kill(&data->refs);
7038 flush_delayed_work(&ctx->rsrc_put_work);
7040 ret = wait_for_completion_interruptible(&data->done);
7044 percpu_ref_resurrect(&data->refs);
7045 io_sqe_rsrc_set_node(ctx, data, backup_node);
7047 reinit_completion(&data->done);
7048 mutex_unlock(&ctx->uring_lock);
7049 ret = io_run_task_work_sig();
7050 mutex_lock(&ctx->uring_lock);
7052 data->quiesce = false;
7055 destroy_fixed_rsrc_ref_node(backup_node);
7059 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7061 struct fixed_rsrc_data *data;
7063 data = kzalloc(sizeof(*data), GFP_KERNEL);
7067 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7068 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7073 init_completion(&data->done);
7077 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7079 percpu_ref_exit(&data->refs);
7084 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7086 struct fixed_rsrc_data *data = ctx->file_data;
7087 unsigned nr_tables, i;
7091 * percpu_ref_is_dying() is to stop parallel files unregister
7092 * Since we possibly drop uring lock later in this function to
7095 if (!data || percpu_ref_is_dying(&data->refs))
7097 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7101 __io_sqe_files_unregister(ctx);
7102 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7103 for (i = 0; i < nr_tables; i++)
7104 kfree(data->table[i].files);
7105 free_fixed_rsrc_data(data);
7106 ctx->file_data = NULL;
7107 ctx->nr_user_files = 0;
7111 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7112 __releases(&sqd->lock)
7114 WARN_ON_ONCE(sqd->thread == current);
7117 * Do the dance but not conditional clear_bit() because it'd race with
7118 * other threads incrementing park_pending and setting the bit.
7120 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7121 if (atomic_dec_return(&sqd->park_pending))
7122 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7123 mutex_unlock(&sqd->lock);
7126 static void io_sq_thread_park(struct io_sq_data *sqd)
7127 __acquires(&sqd->lock)
7129 WARN_ON_ONCE(sqd->thread == current);
7131 atomic_inc(&sqd->park_pending);
7132 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7133 mutex_lock(&sqd->lock);
7135 wake_up_process(sqd->thread);
7138 static void io_sq_thread_stop(struct io_sq_data *sqd)
7140 WARN_ON_ONCE(sqd->thread == current);
7142 mutex_lock(&sqd->lock);
7143 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7145 wake_up_process(sqd->thread);
7146 mutex_unlock(&sqd->lock);
7147 wait_for_completion(&sqd->exited);
7150 static void io_put_sq_data(struct io_sq_data *sqd)
7152 if (refcount_dec_and_test(&sqd->refs)) {
7153 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7155 io_sq_thread_stop(sqd);
7160 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7162 struct io_sq_data *sqd = ctx->sq_data;
7165 io_sq_thread_park(sqd);
7166 list_del_init(&ctx->sqd_list);
7167 io_sqd_update_thread_idle(sqd);
7168 io_sq_thread_unpark(sqd);
7170 io_put_sq_data(sqd);
7171 ctx->sq_data = NULL;
7173 put_cred(ctx->sq_creds);
7177 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7179 struct io_ring_ctx *ctx_attach;
7180 struct io_sq_data *sqd;
7183 f = fdget(p->wq_fd);
7185 return ERR_PTR(-ENXIO);
7186 if (f.file->f_op != &io_uring_fops) {
7188 return ERR_PTR(-EINVAL);
7191 ctx_attach = f.file->private_data;
7192 sqd = ctx_attach->sq_data;
7195 return ERR_PTR(-EINVAL);
7197 if (sqd->task_tgid != current->tgid) {
7199 return ERR_PTR(-EPERM);
7202 refcount_inc(&sqd->refs);
7207 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7210 struct io_sq_data *sqd;
7213 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7214 sqd = io_attach_sq_data(p);
7219 /* fall through for EPERM case, setup new sqd/task */
7220 if (PTR_ERR(sqd) != -EPERM)
7224 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7226 return ERR_PTR(-ENOMEM);
7228 atomic_set(&sqd->park_pending, 0);
7229 refcount_set(&sqd->refs, 1);
7230 INIT_LIST_HEAD(&sqd->ctx_list);
7231 mutex_init(&sqd->lock);
7232 init_waitqueue_head(&sqd->wait);
7233 init_completion(&sqd->exited);
7237 #if defined(CONFIG_UNIX)
7239 * Ensure the UNIX gc is aware of our file set, so we are certain that
7240 * the io_uring can be safely unregistered on process exit, even if we have
7241 * loops in the file referencing.
7243 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7245 struct sock *sk = ctx->ring_sock->sk;
7246 struct scm_fp_list *fpl;
7247 struct sk_buff *skb;
7250 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7254 skb = alloc_skb(0, GFP_KERNEL);
7263 fpl->user = get_uid(current_user());
7264 for (i = 0; i < nr; i++) {
7265 struct file *file = io_file_from_index(ctx, i + offset);
7269 fpl->fp[nr_files] = get_file(file);
7270 unix_inflight(fpl->user, fpl->fp[nr_files]);
7275 fpl->max = SCM_MAX_FD;
7276 fpl->count = nr_files;
7277 UNIXCB(skb).fp = fpl;
7278 skb->destructor = unix_destruct_scm;
7279 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7280 skb_queue_head(&sk->sk_receive_queue, skb);
7282 for (i = 0; i < nr_files; i++)
7293 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7294 * causes regular reference counting to break down. We rely on the UNIX
7295 * garbage collection to take care of this problem for us.
7297 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7299 unsigned left, total;
7303 left = ctx->nr_user_files;
7305 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7307 ret = __io_sqe_files_scm(ctx, this_files, total);
7311 total += this_files;
7317 while (total < ctx->nr_user_files) {
7318 struct file *file = io_file_from_index(ctx, total);
7328 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7334 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7335 unsigned nr_tables, unsigned nr_files)
7339 for (i = 0; i < nr_tables; i++) {
7340 struct fixed_rsrc_table *table = &file_data->table[i];
7341 unsigned this_files;
7343 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7344 table->files = kcalloc(this_files, sizeof(struct file *),
7348 nr_files -= this_files;
7354 for (i = 0; i < nr_tables; i++) {
7355 struct fixed_rsrc_table *table = &file_data->table[i];
7356 kfree(table->files);
7361 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7363 struct file *file = prsrc->file;
7364 #if defined(CONFIG_UNIX)
7365 struct sock *sock = ctx->ring_sock->sk;
7366 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7367 struct sk_buff *skb;
7370 __skb_queue_head_init(&list);
7373 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7374 * remove this entry and rearrange the file array.
7376 skb = skb_dequeue(head);
7378 struct scm_fp_list *fp;
7380 fp = UNIXCB(skb).fp;
7381 for (i = 0; i < fp->count; i++) {
7384 if (fp->fp[i] != file)
7387 unix_notinflight(fp->user, fp->fp[i]);
7388 left = fp->count - 1 - i;
7390 memmove(&fp->fp[i], &fp->fp[i + 1],
7391 left * sizeof(struct file *));
7398 __skb_queue_tail(&list, skb);
7408 __skb_queue_tail(&list, skb);
7410 skb = skb_dequeue(head);
7413 if (skb_peek(&list)) {
7414 spin_lock_irq(&head->lock);
7415 while ((skb = __skb_dequeue(&list)) != NULL)
7416 __skb_queue_tail(head, skb);
7417 spin_unlock_irq(&head->lock);
7424 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7426 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7427 struct io_ring_ctx *ctx = rsrc_data->ctx;
7428 struct io_rsrc_put *prsrc, *tmp;
7430 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7431 list_del(&prsrc->list);
7432 ref_node->rsrc_put(ctx, prsrc);
7436 percpu_ref_exit(&ref_node->refs);
7438 percpu_ref_put(&rsrc_data->refs);
7441 static void io_rsrc_put_work(struct work_struct *work)
7443 struct io_ring_ctx *ctx;
7444 struct llist_node *node;
7446 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7447 node = llist_del_all(&ctx->rsrc_put_llist);
7450 struct fixed_rsrc_ref_node *ref_node;
7451 struct llist_node *next = node->next;
7453 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7454 __io_rsrc_put_work(ref_node);
7459 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7462 struct fixed_rsrc_table *table;
7464 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7465 return &table->files[i & IORING_FILE_TABLE_MASK];
7468 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7470 struct fixed_rsrc_ref_node *ref_node;
7471 struct fixed_rsrc_data *data;
7472 struct io_ring_ctx *ctx;
7473 bool first_add = false;
7476 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7477 data = ref_node->rsrc_data;
7480 io_rsrc_ref_lock(ctx);
7481 ref_node->done = true;
7483 while (!list_empty(&ctx->rsrc_ref_list)) {
7484 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7485 struct fixed_rsrc_ref_node, node);
7486 /* recycle ref nodes in order */
7487 if (!ref_node->done)
7489 list_del(&ref_node->node);
7490 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7492 io_rsrc_ref_unlock(ctx);
7494 if (percpu_ref_is_dying(&data->refs))
7498 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7500 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7503 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7504 struct io_ring_ctx *ctx)
7506 struct fixed_rsrc_ref_node *ref_node;
7508 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7512 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7517 INIT_LIST_HEAD(&ref_node->node);
7518 INIT_LIST_HEAD(&ref_node->rsrc_list);
7519 ref_node->done = false;
7523 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7524 struct fixed_rsrc_ref_node *ref_node)
7526 ref_node->rsrc_data = ctx->file_data;
7527 ref_node->rsrc_put = io_ring_file_put;
7530 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7532 percpu_ref_exit(&ref_node->refs);
7537 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7540 __s32 __user *fds = (__s32 __user *) arg;
7541 unsigned nr_tables, i;
7543 int fd, ret = -ENOMEM;
7544 struct fixed_rsrc_ref_node *ref_node;
7545 struct fixed_rsrc_data *file_data;
7551 if (nr_args > IORING_MAX_FIXED_FILES)
7554 file_data = alloc_fixed_rsrc_data(ctx);
7557 ctx->file_data = file_data;
7559 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7560 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7562 if (!file_data->table)
7565 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7568 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7569 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7573 /* allow sparse sets */
7583 * Don't allow io_uring instances to be registered. If UNIX
7584 * isn't enabled, then this causes a reference cycle and this
7585 * instance can never get freed. If UNIX is enabled we'll
7586 * handle it just fine, but there's still no point in allowing
7587 * a ring fd as it doesn't support regular read/write anyway.
7589 if (file->f_op == &io_uring_fops) {
7593 *io_fixed_file_slot(file_data, i) = file;
7596 ret = io_sqe_files_scm(ctx);
7598 io_sqe_files_unregister(ctx);
7602 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7604 io_sqe_files_unregister(ctx);
7607 init_fixed_file_ref_node(ctx, ref_node);
7609 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7612 for (i = 0; i < ctx->nr_user_files; i++) {
7613 file = io_file_from_index(ctx, i);
7617 for (i = 0; i < nr_tables; i++)
7618 kfree(file_data->table[i].files);
7619 ctx->nr_user_files = 0;
7621 free_fixed_rsrc_data(ctx->file_data);
7622 ctx->file_data = NULL;
7626 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7629 #if defined(CONFIG_UNIX)
7630 struct sock *sock = ctx->ring_sock->sk;
7631 struct sk_buff_head *head = &sock->sk_receive_queue;
7632 struct sk_buff *skb;
7635 * See if we can merge this file into an existing skb SCM_RIGHTS
7636 * file set. If there's no room, fall back to allocating a new skb
7637 * and filling it in.
7639 spin_lock_irq(&head->lock);
7640 skb = skb_peek(head);
7642 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7644 if (fpl->count < SCM_MAX_FD) {
7645 __skb_unlink(skb, head);
7646 spin_unlock_irq(&head->lock);
7647 fpl->fp[fpl->count] = get_file(file);
7648 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7650 spin_lock_irq(&head->lock);
7651 __skb_queue_head(head, skb);
7656 spin_unlock_irq(&head->lock);
7663 return __io_sqe_files_scm(ctx, 1, index);
7669 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7671 struct io_rsrc_put *prsrc;
7672 struct fixed_rsrc_ref_node *ref_node = data->node;
7674 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7679 list_add(&prsrc->list, &ref_node->rsrc_list);
7684 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7687 return io_queue_rsrc_removal(data, (void *)file);
7690 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7691 struct io_uring_rsrc_update *up,
7694 struct fixed_rsrc_data *data = ctx->file_data;
7695 struct fixed_rsrc_ref_node *ref_node;
7696 struct file *file, **file_slot;
7700 bool needs_switch = false;
7702 if (check_add_overflow(up->offset, nr_args, &done))
7704 if (done > ctx->nr_user_files)
7707 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7710 init_fixed_file_ref_node(ctx, ref_node);
7712 fds = u64_to_user_ptr(up->data);
7713 for (done = 0; done < nr_args; done++) {
7715 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7719 if (fd == IORING_REGISTER_FILES_SKIP)
7722 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7723 file_slot = io_fixed_file_slot(ctx->file_data, i);
7726 err = io_queue_file_removal(data, *file_slot);
7730 needs_switch = true;
7739 * Don't allow io_uring instances to be registered. If
7740 * UNIX isn't enabled, then this causes a reference
7741 * cycle and this instance can never get freed. If UNIX
7742 * is enabled we'll handle it just fine, but there's
7743 * still no point in allowing a ring fd as it doesn't
7744 * support regular read/write anyway.
7746 if (file->f_op == &io_uring_fops) {
7752 err = io_sqe_file_register(ctx, file, i);
7762 percpu_ref_kill(&data->node->refs);
7763 io_sqe_rsrc_set_node(ctx, data, ref_node);
7765 destroy_fixed_rsrc_ref_node(ref_node);
7767 return done ? done : err;
7770 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7773 struct io_uring_rsrc_update up;
7775 if (!ctx->file_data)
7779 if (copy_from_user(&up, arg, sizeof(up)))
7784 return __io_sqe_files_update(ctx, &up, nr_args);
7787 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7789 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7791 req = io_put_req_find_next(req);
7792 return req ? &req->work : NULL;
7795 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7797 struct io_wq_hash *hash;
7798 struct io_wq_data data;
7799 unsigned int concurrency;
7801 hash = ctx->hash_map;
7803 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7805 return ERR_PTR(-ENOMEM);
7806 refcount_set(&hash->refs, 1);
7807 init_waitqueue_head(&hash->wait);
7808 ctx->hash_map = hash;
7812 data.free_work = io_free_work;
7813 data.do_work = io_wq_submit_work;
7815 /* Do QD, or 4 * CPUS, whatever is smallest */
7816 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7818 return io_wq_create(concurrency, &data);
7821 static int io_uring_alloc_task_context(struct task_struct *task,
7822 struct io_ring_ctx *ctx)
7824 struct io_uring_task *tctx;
7827 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7828 if (unlikely(!tctx))
7831 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7832 if (unlikely(ret)) {
7837 tctx->io_wq = io_init_wq_offload(ctx);
7838 if (IS_ERR(tctx->io_wq)) {
7839 ret = PTR_ERR(tctx->io_wq);
7840 percpu_counter_destroy(&tctx->inflight);
7846 init_waitqueue_head(&tctx->wait);
7848 atomic_set(&tctx->in_idle, 0);
7849 task->io_uring = tctx;
7850 spin_lock_init(&tctx->task_lock);
7851 INIT_WQ_LIST(&tctx->task_list);
7852 tctx->task_state = 0;
7853 init_task_work(&tctx->task_work, tctx_task_work);
7857 void __io_uring_free(struct task_struct *tsk)
7859 struct io_uring_task *tctx = tsk->io_uring;
7861 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7862 WARN_ON_ONCE(tctx->io_wq);
7864 percpu_counter_destroy(&tctx->inflight);
7866 tsk->io_uring = NULL;
7869 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7870 struct io_uring_params *p)
7874 /* Retain compatibility with failing for an invalid attach attempt */
7875 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7876 IORING_SETUP_ATTACH_WQ) {
7879 f = fdget(p->wq_fd);
7882 if (f.file->f_op != &io_uring_fops) {
7888 if (ctx->flags & IORING_SETUP_SQPOLL) {
7889 struct task_struct *tsk;
7890 struct io_sq_data *sqd;
7894 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7897 sqd = io_get_sq_data(p, &attached);
7903 ctx->sq_creds = get_current_cred();
7905 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7906 if (!ctx->sq_thread_idle)
7907 ctx->sq_thread_idle = HZ;
7910 io_sq_thread_park(sqd);
7911 list_add(&ctx->sqd_list, &sqd->ctx_list);
7912 io_sqd_update_thread_idle(sqd);
7913 /* don't attach to a dying SQPOLL thread, would be racy */
7914 if (attached && !sqd->thread)
7916 io_sq_thread_unpark(sqd);
7923 if (p->flags & IORING_SETUP_SQ_AFF) {
7924 int cpu = p->sq_thread_cpu;
7927 if (cpu >= nr_cpu_ids)
7929 if (!cpu_online(cpu))
7937 sqd->task_pid = current->pid;
7938 sqd->task_tgid = current->tgid;
7939 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7946 ret = io_uring_alloc_task_context(tsk, ctx);
7947 wake_up_new_task(tsk);
7950 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7951 /* Can't have SQ_AFF without SQPOLL */
7958 io_sq_thread_finish(ctx);
7961 complete(&ctx->sq_data->exited);
7965 static inline void __io_unaccount_mem(struct user_struct *user,
7966 unsigned long nr_pages)
7968 atomic_long_sub(nr_pages, &user->locked_vm);
7971 static inline int __io_account_mem(struct user_struct *user,
7972 unsigned long nr_pages)
7974 unsigned long page_limit, cur_pages, new_pages;
7976 /* Don't allow more pages than we can safely lock */
7977 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7980 cur_pages = atomic_long_read(&user->locked_vm);
7981 new_pages = cur_pages + nr_pages;
7982 if (new_pages > page_limit)
7984 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7985 new_pages) != cur_pages);
7990 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7993 __io_unaccount_mem(ctx->user, nr_pages);
7995 if (ctx->mm_account)
7996 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7999 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8004 ret = __io_account_mem(ctx->user, nr_pages);
8009 if (ctx->mm_account)
8010 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8015 static void io_mem_free(void *ptr)
8022 page = virt_to_head_page(ptr);
8023 if (put_page_testzero(page))
8024 free_compound_page(page);
8027 static void *io_mem_alloc(size_t size)
8029 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8030 __GFP_NORETRY | __GFP_ACCOUNT;
8032 return (void *) __get_free_pages(gfp_flags, get_order(size));
8035 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8038 struct io_rings *rings;
8039 size_t off, sq_array_size;
8041 off = struct_size(rings, cqes, cq_entries);
8042 if (off == SIZE_MAX)
8046 off = ALIGN(off, SMP_CACHE_BYTES);
8054 sq_array_size = array_size(sizeof(u32), sq_entries);
8055 if (sq_array_size == SIZE_MAX)
8058 if (check_add_overflow(off, sq_array_size, &off))
8064 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8068 if (!ctx->user_bufs)
8071 for (i = 0; i < ctx->nr_user_bufs; i++) {
8072 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8074 for (j = 0; j < imu->nr_bvecs; j++)
8075 unpin_user_page(imu->bvec[j].bv_page);
8077 if (imu->acct_pages)
8078 io_unaccount_mem(ctx, imu->acct_pages);
8083 kfree(ctx->user_bufs);
8084 ctx->user_bufs = NULL;
8085 ctx->nr_user_bufs = 0;
8089 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8090 void __user *arg, unsigned index)
8092 struct iovec __user *src;
8094 #ifdef CONFIG_COMPAT
8096 struct compat_iovec __user *ciovs;
8097 struct compat_iovec ciov;
8099 ciovs = (struct compat_iovec __user *) arg;
8100 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8103 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8104 dst->iov_len = ciov.iov_len;
8108 src = (struct iovec __user *) arg;
8109 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8115 * Not super efficient, but this is just a registration time. And we do cache
8116 * the last compound head, so generally we'll only do a full search if we don't
8119 * We check if the given compound head page has already been accounted, to
8120 * avoid double accounting it. This allows us to account the full size of the
8121 * page, not just the constituent pages of a huge page.
8123 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8124 int nr_pages, struct page *hpage)
8128 /* check current page array */
8129 for (i = 0; i < nr_pages; i++) {
8130 if (!PageCompound(pages[i]))
8132 if (compound_head(pages[i]) == hpage)
8136 /* check previously registered pages */
8137 for (i = 0; i < ctx->nr_user_bufs; i++) {
8138 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8140 for (j = 0; j < imu->nr_bvecs; j++) {
8141 if (!PageCompound(imu->bvec[j].bv_page))
8143 if (compound_head(imu->bvec[j].bv_page) == hpage)
8151 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8152 int nr_pages, struct io_mapped_ubuf *imu,
8153 struct page **last_hpage)
8157 for (i = 0; i < nr_pages; i++) {
8158 if (!PageCompound(pages[i])) {
8163 hpage = compound_head(pages[i]);
8164 if (hpage == *last_hpage)
8166 *last_hpage = hpage;
8167 if (headpage_already_acct(ctx, pages, i, hpage))
8169 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8173 if (!imu->acct_pages)
8176 ret = io_account_mem(ctx, imu->acct_pages);
8178 imu->acct_pages = 0;
8182 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8183 struct io_mapped_ubuf *imu,
8184 struct page **last_hpage)
8186 struct vm_area_struct **vmas = NULL;
8187 struct page **pages = NULL;
8188 unsigned long off, start, end, ubuf;
8190 int ret, pret, nr_pages, i;
8192 ubuf = (unsigned long) iov->iov_base;
8193 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8194 start = ubuf >> PAGE_SHIFT;
8195 nr_pages = end - start;
8199 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8203 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8208 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8214 mmap_read_lock(current->mm);
8215 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8217 if (pret == nr_pages) {
8218 /* don't support file backed memory */
8219 for (i = 0; i < nr_pages; i++) {
8220 struct vm_area_struct *vma = vmas[i];
8223 !is_file_hugepages(vma->vm_file)) {
8229 ret = pret < 0 ? pret : -EFAULT;
8231 mmap_read_unlock(current->mm);
8234 * if we did partial map, or found file backed vmas,
8235 * release any pages we did get
8238 unpin_user_pages(pages, pret);
8243 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8245 unpin_user_pages(pages, pret);
8250 off = ubuf & ~PAGE_MASK;
8251 size = iov->iov_len;
8252 for (i = 0; i < nr_pages; i++) {
8255 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8256 imu->bvec[i].bv_page = pages[i];
8257 imu->bvec[i].bv_len = vec_len;
8258 imu->bvec[i].bv_offset = off;
8262 /* store original address for later verification */
8264 imu->len = iov->iov_len;
8265 imu->nr_bvecs = nr_pages;
8273 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8277 if (!nr_args || nr_args > UIO_MAXIOV)
8280 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8282 if (!ctx->user_bufs)
8288 static int io_buffer_validate(struct iovec *iov)
8291 * Don't impose further limits on the size and buffer
8292 * constraints here, we'll -EINVAL later when IO is
8293 * submitted if they are wrong.
8295 if (!iov->iov_base || !iov->iov_len)
8298 /* arbitrary limit, but we need something */
8299 if (iov->iov_len > SZ_1G)
8305 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8306 unsigned int nr_args)
8310 struct page *last_hpage = NULL;
8312 ret = io_buffers_map_alloc(ctx, nr_args);
8316 for (i = 0; i < nr_args; i++) {
8317 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8319 ret = io_copy_iov(ctx, &iov, arg, i);
8323 ret = io_buffer_validate(&iov);
8327 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8331 ctx->nr_user_bufs++;
8335 io_sqe_buffers_unregister(ctx);
8340 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8342 __s32 __user *fds = arg;
8348 if (copy_from_user(&fd, fds, sizeof(*fds)))
8351 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8352 if (IS_ERR(ctx->cq_ev_fd)) {
8353 int ret = PTR_ERR(ctx->cq_ev_fd);
8354 ctx->cq_ev_fd = NULL;
8361 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8363 if (ctx->cq_ev_fd) {
8364 eventfd_ctx_put(ctx->cq_ev_fd);
8365 ctx->cq_ev_fd = NULL;
8372 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8374 struct io_buffer *buf;
8375 unsigned long index;
8377 xa_for_each(&ctx->io_buffers, index, buf)
8378 __io_remove_buffers(ctx, buf, index, -1U);
8381 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8383 struct io_kiocb *req, *nxt;
8385 list_for_each_entry_safe(req, nxt, list, compl.list) {
8386 if (tsk && req->task != tsk)
8388 list_del(&req->compl.list);
8389 kmem_cache_free(req_cachep, req);
8393 static void io_req_caches_free(struct io_ring_ctx *ctx)
8395 struct io_submit_state *submit_state = &ctx->submit_state;
8396 struct io_comp_state *cs = &ctx->submit_state.comp;
8398 mutex_lock(&ctx->uring_lock);
8400 if (submit_state->free_reqs) {
8401 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8402 submit_state->reqs);
8403 submit_state->free_reqs = 0;
8406 spin_lock_irq(&ctx->completion_lock);
8407 list_splice_init(&cs->locked_free_list, &cs->free_list);
8408 cs->locked_free_nr = 0;
8409 spin_unlock_irq(&ctx->completion_lock);
8411 io_req_cache_free(&cs->free_list, NULL);
8413 mutex_unlock(&ctx->uring_lock);
8416 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8419 * Some may use context even when all refs and requests have been put,
8420 * and they are free to do so while still holding uring_lock or
8421 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8423 mutex_lock(&ctx->uring_lock);
8424 mutex_unlock(&ctx->uring_lock);
8425 spin_lock_irq(&ctx->completion_lock);
8426 spin_unlock_irq(&ctx->completion_lock);
8428 io_sq_thread_finish(ctx);
8429 io_sqe_buffers_unregister(ctx);
8431 if (ctx->mm_account) {
8432 mmdrop(ctx->mm_account);
8433 ctx->mm_account = NULL;
8436 mutex_lock(&ctx->uring_lock);
8437 io_sqe_files_unregister(ctx);
8438 mutex_unlock(&ctx->uring_lock);
8439 io_eventfd_unregister(ctx);
8440 io_destroy_buffers(ctx);
8442 #if defined(CONFIG_UNIX)
8443 if (ctx->ring_sock) {
8444 ctx->ring_sock->file = NULL; /* so that iput() is called */
8445 sock_release(ctx->ring_sock);
8449 io_mem_free(ctx->rings);
8450 io_mem_free(ctx->sq_sqes);
8452 percpu_ref_exit(&ctx->refs);
8453 free_uid(ctx->user);
8454 io_req_caches_free(ctx);
8456 io_wq_put_hash(ctx->hash_map);
8457 kfree(ctx->cancel_hash);
8461 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8463 struct io_ring_ctx *ctx = file->private_data;
8466 poll_wait(file, &ctx->cq_wait, wait);
8468 * synchronizes with barrier from wq_has_sleeper call in
8472 if (!io_sqring_full(ctx))
8473 mask |= EPOLLOUT | EPOLLWRNORM;
8476 * Don't flush cqring overflow list here, just do a simple check.
8477 * Otherwise there could possible be ABBA deadlock:
8480 * lock(&ctx->uring_lock);
8482 * lock(&ctx->uring_lock);
8485 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8486 * pushs them to do the flush.
8488 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8489 mask |= EPOLLIN | EPOLLRDNORM;
8494 static int io_uring_fasync(int fd, struct file *file, int on)
8496 struct io_ring_ctx *ctx = file->private_data;
8498 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8501 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8503 const struct cred *creds;
8505 creds = xa_erase(&ctx->personalities, id);
8514 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8516 return io_run_task_work_head(&ctx->exit_task_work);
8519 struct io_tctx_exit {
8520 struct callback_head task_work;
8521 struct completion completion;
8522 struct io_ring_ctx *ctx;
8525 static void io_tctx_exit_cb(struct callback_head *cb)
8527 struct io_uring_task *tctx = current->io_uring;
8528 struct io_tctx_exit *work;
8530 work = container_of(cb, struct io_tctx_exit, task_work);
8532 * When @in_idle, we're in cancellation and it's racy to remove the
8533 * node. It'll be removed by the end of cancellation, just ignore it.
8535 if (!atomic_read(&tctx->in_idle))
8536 io_uring_del_task_file((unsigned long)work->ctx);
8537 complete(&work->completion);
8540 static void io_ring_exit_work(struct work_struct *work)
8542 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8543 unsigned long timeout = jiffies + HZ * 60 * 5;
8544 struct io_tctx_exit exit;
8545 struct io_tctx_node *node;
8548 /* prevent SQPOLL from submitting new requests */
8550 io_sq_thread_park(ctx->sq_data);
8551 list_del_init(&ctx->sqd_list);
8552 io_sqd_update_thread_idle(ctx->sq_data);
8553 io_sq_thread_unpark(ctx->sq_data);
8557 * If we're doing polled IO and end up having requests being
8558 * submitted async (out-of-line), then completions can come in while
8559 * we're waiting for refs to drop. We need to reap these manually,
8560 * as nobody else will be looking for them.
8563 io_uring_try_cancel_requests(ctx, NULL, NULL);
8565 WARN_ON_ONCE(time_after(jiffies, timeout));
8566 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8568 mutex_lock(&ctx->uring_lock);
8569 while (!list_empty(&ctx->tctx_list)) {
8570 WARN_ON_ONCE(time_after(jiffies, timeout));
8572 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8575 init_completion(&exit.completion);
8576 init_task_work(&exit.task_work, io_tctx_exit_cb);
8577 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8578 if (WARN_ON_ONCE(ret))
8580 wake_up_process(node->task);
8582 mutex_unlock(&ctx->uring_lock);
8583 wait_for_completion(&exit.completion);
8585 mutex_lock(&ctx->uring_lock);
8587 mutex_unlock(&ctx->uring_lock);
8589 io_ring_ctx_free(ctx);
8592 /* Returns true if we found and killed one or more timeouts */
8593 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8594 struct files_struct *files)
8596 struct io_kiocb *req, *tmp;
8599 spin_lock_irq(&ctx->completion_lock);
8600 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8601 if (io_match_task(req, tsk, files)) {
8602 io_kill_timeout(req, -ECANCELED);
8606 io_commit_cqring(ctx);
8607 spin_unlock_irq(&ctx->completion_lock);
8610 io_cqring_ev_posted(ctx);
8611 return canceled != 0;
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);