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;
205 struct io_overflow_cqe {
206 struct io_uring_cqe cqe;
207 struct list_head list;
211 struct list_head list;
218 struct fixed_rsrc_table {
222 struct io_rsrc_node {
223 struct percpu_ref refs;
224 struct list_head node;
225 struct list_head rsrc_list;
226 struct io_rsrc_data *rsrc_data;
227 struct llist_node llist;
231 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
233 struct io_rsrc_data {
234 struct fixed_rsrc_table *table;
235 struct io_ring_ctx *ctx;
238 struct percpu_ref refs;
239 struct completion done;
244 struct list_head list;
250 struct io_restriction {
251 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
252 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
253 u8 sqe_flags_allowed;
254 u8 sqe_flags_required;
259 IO_SQ_THREAD_SHOULD_STOP = 0,
260 IO_SQ_THREAD_SHOULD_PARK,
265 atomic_t park_pending;
268 /* ctx's that are using this sqd */
269 struct list_head ctx_list;
271 struct task_struct *thread;
272 struct wait_queue_head wait;
274 unsigned sq_thread_idle;
280 struct completion exited;
281 struct callback_head *park_task_work;
284 #define IO_IOPOLL_BATCH 8
285 #define IO_COMPL_BATCH 32
286 #define IO_REQ_CACHE_SIZE 32
287 #define IO_REQ_ALLOC_BATCH 8
289 struct io_comp_state {
290 struct io_kiocb *reqs[IO_COMPL_BATCH];
292 unsigned int locked_free_nr;
293 /* inline/task_work completion list, under ->uring_lock */
294 struct list_head free_list;
295 /* IRQ completion list, under ->completion_lock */
296 struct list_head locked_free_list;
299 struct io_submit_link {
300 struct io_kiocb *head;
301 struct io_kiocb *last;
304 struct io_submit_state {
305 struct blk_plug plug;
306 struct io_submit_link link;
309 * io_kiocb alloc cache
311 void *reqs[IO_REQ_CACHE_SIZE];
312 unsigned int free_reqs;
317 * Batch completion logic
319 struct io_comp_state comp;
322 * File reference cache
326 unsigned int file_refs;
327 unsigned int ios_left;
332 struct percpu_ref refs;
333 } ____cacheline_aligned_in_smp;
337 unsigned int compat: 1;
338 unsigned int cq_overflow_flushed: 1;
339 unsigned int drain_next: 1;
340 unsigned int eventfd_async: 1;
341 unsigned int restricted: 1;
344 * Ring buffer of indices into array of io_uring_sqe, which is
345 * mmapped by the application using the IORING_OFF_SQES offset.
347 * This indirection could e.g. be used to assign fixed
348 * io_uring_sqe entries to operations and only submit them to
349 * the queue when needed.
351 * The kernel modifies neither the indices array nor the entries
355 unsigned cached_sq_head;
358 unsigned sq_thread_idle;
359 unsigned cached_sq_dropped;
360 unsigned cached_cq_overflow;
361 unsigned long sq_check_overflow;
363 /* hashed buffered write serialization */
364 struct io_wq_hash *hash_map;
366 struct list_head defer_list;
367 struct list_head timeout_list;
368 struct list_head cq_overflow_list;
370 struct io_uring_sqe *sq_sqes;
371 } ____cacheline_aligned_in_smp;
374 struct mutex uring_lock;
375 wait_queue_head_t wait;
376 } ____cacheline_aligned_in_smp;
378 struct io_submit_state submit_state;
380 struct io_rings *rings;
382 /* Only used for accounting purposes */
383 struct mm_struct *mm_account;
385 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
386 struct io_sq_data *sq_data; /* if using sq thread polling */
388 struct wait_queue_head sqo_sq_wait;
389 struct list_head sqd_list;
392 * If used, fixed file set. Writers must ensure that ->refs is dead,
393 * readers must ensure that ->refs is alive as long as the file* is
394 * used. Only updated through io_uring_register(2).
396 struct io_rsrc_data *file_data;
397 unsigned nr_user_files;
399 /* if used, fixed mapped user buffers */
400 unsigned nr_user_bufs;
401 struct io_mapped_ubuf *user_bufs;
403 struct user_struct *user;
405 struct completion ref_comp;
407 #if defined(CONFIG_UNIX)
408 struct socket *ring_sock;
411 struct xarray io_buffers;
413 struct xarray personalities;
417 unsigned cached_cq_tail;
420 atomic_t cq_timeouts;
421 unsigned cq_last_tm_flush;
422 unsigned long cq_check_overflow;
423 struct wait_queue_head cq_wait;
424 struct fasync_struct *cq_fasync;
425 struct eventfd_ctx *cq_ev_fd;
426 } ____cacheline_aligned_in_smp;
429 spinlock_t completion_lock;
432 * ->iopoll_list is protected by the ctx->uring_lock for
433 * io_uring instances that don't use IORING_SETUP_SQPOLL.
434 * For SQPOLL, only the single threaded io_sq_thread() will
435 * manipulate the list, hence no extra locking is needed there.
437 struct list_head iopoll_list;
438 struct hlist_head *cancel_hash;
439 unsigned cancel_hash_bits;
440 bool poll_multi_file;
442 spinlock_t inflight_lock;
443 struct list_head inflight_list;
444 } ____cacheline_aligned_in_smp;
446 struct delayed_work rsrc_put_work;
447 struct llist_head rsrc_put_llist;
448 struct list_head rsrc_ref_list;
449 spinlock_t rsrc_ref_lock;
450 struct io_rsrc_node *rsrc_node;
451 struct io_rsrc_node *rsrc_backup_node;
453 struct io_restriction restrictions;
456 struct callback_head *exit_task_work;
458 /* Keep this last, we don't need it for the fast path */
459 struct work_struct exit_work;
460 struct list_head tctx_list;
463 struct io_uring_task {
464 /* submission side */
466 struct wait_queue_head wait;
467 const struct io_ring_ctx *last;
469 struct percpu_counter inflight;
472 spinlock_t task_lock;
473 struct io_wq_work_list task_list;
474 unsigned long task_state;
475 struct callback_head task_work;
479 * First field must be the file pointer in all the
480 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
482 struct io_poll_iocb {
484 struct wait_queue_head *head;
489 bool update_user_data;
491 struct wait_queue_entry wait;
499 struct io_poll_remove {
509 struct io_timeout_data {
510 struct io_kiocb *req;
511 struct hrtimer timer;
512 struct timespec64 ts;
513 enum hrtimer_mode mode;
518 struct sockaddr __user *addr;
519 int __user *addr_len;
521 unsigned long nofile;
541 struct list_head list;
542 /* head of the link, used by linked timeouts only */
543 struct io_kiocb *head;
546 struct io_timeout_rem {
551 struct timespec64 ts;
556 /* NOTE: kiocb has the file as the first member, so don't do it here */
564 struct sockaddr __user *addr;
571 struct user_msghdr __user *umsg;
577 struct io_buffer *kbuf;
583 struct filename *filename;
585 unsigned long nofile;
588 struct io_rsrc_update {
614 struct epoll_event event;
618 struct file *file_out;
619 struct file *file_in;
626 struct io_provide_buf {
640 const char __user *filename;
641 struct statx __user *buffer;
653 struct filename *oldpath;
654 struct filename *newpath;
662 struct filename *filename;
665 struct io_completion {
667 struct list_head list;
671 struct io_async_connect {
672 struct sockaddr_storage address;
675 struct io_async_msghdr {
676 struct iovec fast_iov[UIO_FASTIOV];
677 /* points to an allocated iov, if NULL we use fast_iov instead */
678 struct iovec *free_iov;
679 struct sockaddr __user *uaddr;
681 struct sockaddr_storage addr;
685 struct iovec fast_iov[UIO_FASTIOV];
686 const struct iovec *free_iovec;
687 struct iov_iter iter;
689 struct wait_page_queue wpq;
693 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
694 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
695 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
696 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
697 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
698 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
704 REQ_F_LINK_TIMEOUT_BIT,
705 REQ_F_NEED_CLEANUP_BIT,
707 REQ_F_BUFFER_SELECTED_BIT,
708 REQ_F_LTIMEOUT_ACTIVE_BIT,
709 REQ_F_COMPLETE_INLINE_BIT,
711 REQ_F_DONT_REISSUE_BIT,
712 /* keep async read/write and isreg together and in order */
713 REQ_F_ASYNC_READ_BIT,
714 REQ_F_ASYNC_WRITE_BIT,
717 /* not a real bit, just to check we're not overflowing the space */
723 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
724 /* drain existing IO first */
725 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
727 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
728 /* doesn't sever on completion < 0 */
729 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
731 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
732 /* IOSQE_BUFFER_SELECT */
733 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
735 /* fail rest of links */
736 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
737 /* on inflight list, should be cancelled and waited on exit reliably */
738 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
739 /* read/write uses file position */
740 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
741 /* must not punt to workers */
742 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
743 /* has or had linked timeout */
744 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
746 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
747 /* already went through poll handler */
748 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
749 /* buffer already selected */
750 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
751 /* linked timeout is active, i.e. prepared by link's head */
752 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
753 /* completion is deferred through io_comp_state */
754 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
755 /* caller should reissue async */
756 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
757 /* don't attempt request reissue, see io_rw_reissue() */
758 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
759 /* supports async reads */
760 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
761 /* supports async writes */
762 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
764 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
768 struct io_poll_iocb poll;
769 struct io_poll_iocb *double_poll;
772 struct io_task_work {
773 struct io_wq_work_node node;
774 task_work_func_t func;
778 * NOTE! Each of the iocb union members has the file pointer
779 * as the first entry in their struct definition. So you can
780 * access the file pointer through any of the sub-structs,
781 * or directly as just 'ki_filp' in this struct.
787 struct io_poll_iocb poll;
788 struct io_poll_remove poll_remove;
789 struct io_accept accept;
791 struct io_cancel cancel;
792 struct io_timeout timeout;
793 struct io_timeout_rem timeout_rem;
794 struct io_connect connect;
795 struct io_sr_msg sr_msg;
797 struct io_close close;
798 struct io_rsrc_update rsrc_update;
799 struct io_fadvise fadvise;
800 struct io_madvise madvise;
801 struct io_epoll epoll;
802 struct io_splice splice;
803 struct io_provide_buf pbuf;
804 struct io_statx statx;
805 struct io_shutdown shutdown;
806 struct io_rename rename;
807 struct io_unlink unlink;
808 /* use only after cleaning per-op data, see io_clean_op() */
809 struct io_completion compl;
812 /* opcode allocated if it needs to store data for async defer */
815 /* polled IO has completed */
821 struct io_ring_ctx *ctx;
824 struct task_struct *task;
827 struct io_kiocb *link;
828 struct percpu_ref *fixed_rsrc_refs;
831 * 1. used with ctx->iopoll_list with reads/writes
832 * 2. to track reqs with ->files (see io_op_def::file_table)
834 struct list_head inflight_entry;
836 struct io_task_work io_task_work;
837 struct callback_head task_work;
839 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
840 struct hlist_node hash_node;
841 struct async_poll *apoll;
842 struct io_wq_work work;
845 struct io_tctx_node {
846 struct list_head ctx_node;
847 struct task_struct *task;
848 struct io_ring_ctx *ctx;
851 struct io_defer_entry {
852 struct list_head list;
853 struct io_kiocb *req;
858 /* needs req->file assigned */
859 unsigned needs_file : 1;
860 /* hash wq insertion if file is a regular file */
861 unsigned hash_reg_file : 1;
862 /* unbound wq insertion if file is a non-regular file */
863 unsigned unbound_nonreg_file : 1;
864 /* opcode is not supported by this kernel */
865 unsigned not_supported : 1;
866 /* set if opcode supports polled "wait" */
868 unsigned pollout : 1;
869 /* op supports buffer selection */
870 unsigned buffer_select : 1;
871 /* do prep async if is going to be punted */
872 unsigned needs_async_setup : 1;
873 /* should block plug */
875 /* size of async data needed, if any */
876 unsigned short async_size;
879 static const struct io_op_def io_op_defs[] = {
880 [IORING_OP_NOP] = {},
881 [IORING_OP_READV] = {
883 .unbound_nonreg_file = 1,
886 .needs_async_setup = 1,
888 .async_size = sizeof(struct io_async_rw),
890 [IORING_OP_WRITEV] = {
893 .unbound_nonreg_file = 1,
895 .needs_async_setup = 1,
897 .async_size = sizeof(struct io_async_rw),
899 [IORING_OP_FSYNC] = {
902 [IORING_OP_READ_FIXED] = {
904 .unbound_nonreg_file = 1,
907 .async_size = sizeof(struct io_async_rw),
909 [IORING_OP_WRITE_FIXED] = {
912 .unbound_nonreg_file = 1,
915 .async_size = sizeof(struct io_async_rw),
917 [IORING_OP_POLL_ADD] = {
919 .unbound_nonreg_file = 1,
921 [IORING_OP_POLL_REMOVE] = {},
922 [IORING_OP_SYNC_FILE_RANGE] = {
925 [IORING_OP_SENDMSG] = {
927 .unbound_nonreg_file = 1,
929 .needs_async_setup = 1,
930 .async_size = sizeof(struct io_async_msghdr),
932 [IORING_OP_RECVMSG] = {
934 .unbound_nonreg_file = 1,
937 .needs_async_setup = 1,
938 .async_size = sizeof(struct io_async_msghdr),
940 [IORING_OP_TIMEOUT] = {
941 .async_size = sizeof(struct io_timeout_data),
943 [IORING_OP_TIMEOUT_REMOVE] = {
944 /* used by timeout updates' prep() */
946 [IORING_OP_ACCEPT] = {
948 .unbound_nonreg_file = 1,
951 [IORING_OP_ASYNC_CANCEL] = {},
952 [IORING_OP_LINK_TIMEOUT] = {
953 .async_size = sizeof(struct io_timeout_data),
955 [IORING_OP_CONNECT] = {
957 .unbound_nonreg_file = 1,
959 .needs_async_setup = 1,
960 .async_size = sizeof(struct io_async_connect),
962 [IORING_OP_FALLOCATE] = {
965 [IORING_OP_OPENAT] = {},
966 [IORING_OP_CLOSE] = {},
967 [IORING_OP_FILES_UPDATE] = {},
968 [IORING_OP_STATX] = {},
971 .unbound_nonreg_file = 1,
975 .async_size = sizeof(struct io_async_rw),
977 [IORING_OP_WRITE] = {
979 .unbound_nonreg_file = 1,
982 .async_size = sizeof(struct io_async_rw),
984 [IORING_OP_FADVISE] = {
987 [IORING_OP_MADVISE] = {},
990 .unbound_nonreg_file = 1,
995 .unbound_nonreg_file = 1,
999 [IORING_OP_OPENAT2] = {
1001 [IORING_OP_EPOLL_CTL] = {
1002 .unbound_nonreg_file = 1,
1004 [IORING_OP_SPLICE] = {
1007 .unbound_nonreg_file = 1,
1009 [IORING_OP_PROVIDE_BUFFERS] = {},
1010 [IORING_OP_REMOVE_BUFFERS] = {},
1014 .unbound_nonreg_file = 1,
1016 [IORING_OP_SHUTDOWN] = {
1019 [IORING_OP_RENAMEAT] = {},
1020 [IORING_OP_UNLINKAT] = {},
1023 static bool io_disarm_next(struct io_kiocb *req);
1024 static void io_uring_del_task_file(unsigned long index);
1025 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1026 struct task_struct *task,
1027 struct files_struct *files);
1028 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1029 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1030 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1032 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1033 static void io_put_req(struct io_kiocb *req);
1034 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1035 static void io_dismantle_req(struct io_kiocb *req);
1036 static void io_put_task(struct task_struct *task, int nr);
1037 static void io_queue_next(struct io_kiocb *req);
1038 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1039 static void io_queue_linked_timeout(struct io_kiocb *req);
1040 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1041 struct io_uring_rsrc_update *ip,
1043 static void io_clean_op(struct io_kiocb *req);
1044 static struct file *io_file_get(struct io_submit_state *state,
1045 struct io_kiocb *req, int fd, bool fixed);
1046 static void __io_queue_sqe(struct io_kiocb *req);
1047 static void io_rsrc_put_work(struct work_struct *work);
1049 static void io_req_task_queue(struct io_kiocb *req);
1050 static void io_submit_flush_completions(struct io_comp_state *cs,
1051 struct io_ring_ctx *ctx);
1052 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1053 static int io_req_prep_async(struct io_kiocb *req);
1055 static struct kmem_cache *req_cachep;
1057 static const struct file_operations io_uring_fops;
1059 struct sock *io_uring_get_socket(struct file *file)
1061 #if defined(CONFIG_UNIX)
1062 if (file->f_op == &io_uring_fops) {
1063 struct io_ring_ctx *ctx = file->private_data;
1065 return ctx->ring_sock->sk;
1070 EXPORT_SYMBOL(io_uring_get_socket);
1072 #define io_for_each_link(pos, head) \
1073 for (pos = (head); pos; pos = pos->link)
1075 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1077 struct io_ring_ctx *ctx = req->ctx;
1079 if (!req->fixed_rsrc_refs) {
1080 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1081 percpu_ref_get(req->fixed_rsrc_refs);
1085 static bool io_match_task(struct io_kiocb *head,
1086 struct task_struct *task,
1087 struct files_struct *files)
1089 struct io_kiocb *req;
1091 if (task && head->task != task)
1096 io_for_each_link(req, head) {
1097 if (req->flags & REQ_F_INFLIGHT)
1103 static inline void req_set_fail_links(struct io_kiocb *req)
1105 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1106 req->flags |= REQ_F_FAIL_LINK;
1109 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1111 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1113 complete(&ctx->ref_comp);
1116 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1118 return !req->timeout.off;
1121 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1123 struct io_ring_ctx *ctx;
1126 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1131 * Use 5 bits less than the max cq entries, that should give us around
1132 * 32 entries per hash list if totally full and uniformly spread.
1134 hash_bits = ilog2(p->cq_entries);
1138 ctx->cancel_hash_bits = hash_bits;
1139 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1141 if (!ctx->cancel_hash)
1143 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1145 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1146 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1149 ctx->flags = p->flags;
1150 init_waitqueue_head(&ctx->sqo_sq_wait);
1151 INIT_LIST_HEAD(&ctx->sqd_list);
1152 init_waitqueue_head(&ctx->cq_wait);
1153 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1154 init_completion(&ctx->ref_comp);
1155 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1156 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1157 mutex_init(&ctx->uring_lock);
1158 init_waitqueue_head(&ctx->wait);
1159 spin_lock_init(&ctx->completion_lock);
1160 INIT_LIST_HEAD(&ctx->iopoll_list);
1161 INIT_LIST_HEAD(&ctx->defer_list);
1162 INIT_LIST_HEAD(&ctx->timeout_list);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 spin_lock_init(&ctx->rsrc_ref_lock);
1166 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1167 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1168 init_llist_head(&ctx->rsrc_put_llist);
1169 INIT_LIST_HEAD(&ctx->tctx_list);
1170 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1174 kfree(ctx->cancel_hash);
1179 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1181 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1182 struct io_ring_ctx *ctx = req->ctx;
1184 return seq != ctx->cached_cq_tail
1185 + READ_ONCE(ctx->cached_cq_overflow);
1191 static void io_req_track_inflight(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1195 if (!(req->flags & REQ_F_INFLIGHT)) {
1196 req->flags |= REQ_F_INFLIGHT;
1198 spin_lock_irq(&ctx->inflight_lock);
1199 list_add(&req->inflight_entry, &ctx->inflight_list);
1200 spin_unlock_irq(&ctx->inflight_lock);
1204 static void io_prep_async_work(struct io_kiocb *req)
1206 const struct io_op_def *def = &io_op_defs[req->opcode];
1207 struct io_ring_ctx *ctx = req->ctx;
1209 if (!req->work.creds)
1210 req->work.creds = get_current_cred();
1212 req->work.list.next = NULL;
1213 req->work.flags = 0;
1214 if (req->flags & REQ_F_FORCE_ASYNC)
1215 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1217 if (req->flags & REQ_F_ISREG) {
1218 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1219 io_wq_hash_work(&req->work, file_inode(req->file));
1220 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1221 if (def->unbound_nonreg_file)
1222 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 switch (req->opcode) {
1226 case IORING_OP_SPLICE:
1229 * Splice operation will be punted aync, and here need to
1230 * modify io_wq_work.flags, so initialize io_wq_work firstly.
1232 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1233 req->work.flags |= IO_WQ_WORK_UNBOUND;
1238 static void io_prep_async_link(struct io_kiocb *req)
1240 struct io_kiocb *cur;
1242 io_for_each_link(cur, req)
1243 io_prep_async_work(cur);
1246 static void io_queue_async_work(struct io_kiocb *req)
1248 struct io_ring_ctx *ctx = req->ctx;
1249 struct io_kiocb *link = io_prep_linked_timeout(req);
1250 struct io_uring_task *tctx = req->task->io_uring;
1253 BUG_ON(!tctx->io_wq);
1255 /* init ->work of the whole link before punting */
1256 io_prep_async_link(req);
1257 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1258 &req->work, req->flags);
1259 io_wq_enqueue(tctx->io_wq, &req->work);
1261 io_queue_linked_timeout(link);
1264 static void io_kill_timeout(struct io_kiocb *req, int status)
1266 struct io_timeout_data *io = req->async_data;
1269 ret = hrtimer_try_to_cancel(&io->timer);
1271 atomic_set(&req->ctx->cq_timeouts,
1272 atomic_read(&req->ctx->cq_timeouts) + 1);
1273 list_del_init(&req->timeout.list);
1274 io_cqring_fill_event(req, status);
1275 io_put_req_deferred(req, 1);
1279 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1282 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1283 struct io_defer_entry, list);
1285 if (req_need_defer(de->req, de->seq))
1287 list_del_init(&de->list);
1288 io_req_task_queue(de->req);
1290 } while (!list_empty(&ctx->defer_list));
1293 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1297 if (list_empty(&ctx->timeout_list))
1300 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1303 u32 events_needed, events_got;
1304 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1305 struct io_kiocb, timeout.list);
1307 if (io_is_timeout_noseq(req))
1311 * Since seq can easily wrap around over time, subtract
1312 * the last seq at which timeouts were flushed before comparing.
1313 * Assuming not more than 2^31-1 events have happened since,
1314 * these subtractions won't have wrapped, so we can check if
1315 * target is in [last_seq, current_seq] by comparing the two.
1317 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1318 events_got = seq - ctx->cq_last_tm_flush;
1319 if (events_got < events_needed)
1322 list_del_init(&req->timeout.list);
1323 io_kill_timeout(req, 0);
1324 } while (!list_empty(&ctx->timeout_list));
1326 ctx->cq_last_tm_flush = seq;
1329 static void io_commit_cqring(struct io_ring_ctx *ctx)
1331 io_flush_timeouts(ctx);
1333 /* order cqe stores with ring update */
1334 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1336 if (unlikely(!list_empty(&ctx->defer_list)))
1337 __io_queue_deferred(ctx);
1340 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1342 struct io_rings *r = ctx->rings;
1344 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1347 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1349 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1352 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1354 struct io_rings *rings = ctx->rings;
1358 * writes to the cq entry need to come after reading head; the
1359 * control dependency is enough as we're using WRITE_ONCE to
1362 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1365 tail = ctx->cached_cq_tail++;
1366 return &rings->cqes[tail & ctx->cq_mask];
1369 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1373 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1375 if (!ctx->eventfd_async)
1377 return io_wq_current_is_worker();
1380 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1382 /* see waitqueue_active() comment */
1385 if (waitqueue_active(&ctx->wait))
1386 wake_up(&ctx->wait);
1387 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1388 wake_up(&ctx->sq_data->wait);
1389 if (io_should_trigger_evfd(ctx))
1390 eventfd_signal(ctx->cq_ev_fd, 1);
1391 if (waitqueue_active(&ctx->cq_wait)) {
1392 wake_up_interruptible(&ctx->cq_wait);
1393 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1397 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1399 /* see waitqueue_active() comment */
1402 if (ctx->flags & IORING_SETUP_SQPOLL) {
1403 if (waitqueue_active(&ctx->wait))
1404 wake_up(&ctx->wait);
1406 if (io_should_trigger_evfd(ctx))
1407 eventfd_signal(ctx->cq_ev_fd, 1);
1408 if (waitqueue_active(&ctx->cq_wait)) {
1409 wake_up_interruptible(&ctx->cq_wait);
1410 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1414 /* Returns true if there are no backlogged entries after the flush */
1415 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1417 struct io_rings *rings = ctx->rings;
1418 unsigned long flags;
1419 bool all_flushed, posted;
1421 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1425 spin_lock_irqsave(&ctx->completion_lock, flags);
1426 while (!list_empty(&ctx->cq_overflow_list)) {
1427 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1428 struct io_overflow_cqe *ocqe;
1432 ocqe = list_first_entry(&ctx->cq_overflow_list,
1433 struct io_overflow_cqe, list);
1435 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1437 WRITE_ONCE(ctx->rings->cq_overflow,
1438 ++ctx->cached_cq_overflow);
1440 list_del(&ocqe->list);
1444 all_flushed = list_empty(&ctx->cq_overflow_list);
1446 clear_bit(0, &ctx->sq_check_overflow);
1447 clear_bit(0, &ctx->cq_check_overflow);
1448 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1452 io_commit_cqring(ctx);
1453 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1455 io_cqring_ev_posted(ctx);
1459 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1463 if (test_bit(0, &ctx->cq_check_overflow)) {
1464 /* iopoll syncs against uring_lock, not completion_lock */
1465 if (ctx->flags & IORING_SETUP_IOPOLL)
1466 mutex_lock(&ctx->uring_lock);
1467 ret = __io_cqring_overflow_flush(ctx, force);
1468 if (ctx->flags & IORING_SETUP_IOPOLL)
1469 mutex_unlock(&ctx->uring_lock);
1476 * Shamelessly stolen from the mm implementation of page reference checking,
1477 * see commit f958d7b528b1 for details.
1479 #define req_ref_zero_or_close_to_overflow(req) \
1480 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1482 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1484 return atomic_inc_not_zero(&req->refs);
1487 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1489 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1490 return atomic_sub_and_test(refs, &req->refs);
1493 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1495 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1496 return atomic_dec_and_test(&req->refs);
1499 static inline void req_ref_put(struct io_kiocb *req)
1501 WARN_ON_ONCE(req_ref_put_and_test(req));
1504 static inline void req_ref_get(struct io_kiocb *req)
1506 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1507 atomic_inc(&req->refs);
1510 static bool __io_cqring_fill_event(struct io_kiocb *req, long res,
1511 unsigned int cflags)
1513 struct io_ring_ctx *ctx = req->ctx;
1514 struct io_uring_cqe *cqe;
1516 trace_io_uring_complete(ctx, req->user_data, res, cflags);
1519 * If we can't get a cq entry, userspace overflowed the
1520 * submission (by quite a lot). Increment the overflow count in
1523 cqe = io_get_cqring(ctx);
1525 WRITE_ONCE(cqe->user_data, req->user_data);
1526 WRITE_ONCE(cqe->res, res);
1527 WRITE_ONCE(cqe->flags, cflags);
1530 if (!ctx->cq_overflow_flushed &&
1531 !atomic_read(&req->task->io_uring->in_idle)) {
1532 struct io_overflow_cqe *ocqe;
1534 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1537 if (list_empty(&ctx->cq_overflow_list)) {
1538 set_bit(0, &ctx->sq_check_overflow);
1539 set_bit(0, &ctx->cq_check_overflow);
1540 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1542 ocqe->cqe.user_data = req->user_data;
1543 ocqe->cqe.res = res;
1544 ocqe->cqe.flags = cflags;
1545 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1550 * If we're in ring overflow flush mode, or in task cancel mode,
1551 * or cannot allocate an overflow entry, then we need to drop it
1554 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1558 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1560 __io_cqring_fill_event(req, res, 0);
1563 static void io_req_complete_post(struct io_kiocb *req, long res,
1564 unsigned int cflags)
1566 struct io_ring_ctx *ctx = req->ctx;
1567 unsigned long flags;
1569 spin_lock_irqsave(&ctx->completion_lock, flags);
1570 __io_cqring_fill_event(req, res, cflags);
1572 * If we're the last reference to this request, add to our locked
1575 if (req_ref_put_and_test(req)) {
1576 struct io_comp_state *cs = &ctx->submit_state.comp;
1578 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1579 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1580 io_disarm_next(req);
1582 io_req_task_queue(req->link);
1586 io_dismantle_req(req);
1587 io_put_task(req->task, 1);
1588 list_add(&req->compl.list, &cs->locked_free_list);
1589 cs->locked_free_nr++;
1591 if (!percpu_ref_tryget(&ctx->refs))
1594 io_commit_cqring(ctx);
1595 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1598 io_cqring_ev_posted(ctx);
1599 percpu_ref_put(&ctx->refs);
1603 static void io_req_complete_state(struct io_kiocb *req, long res,
1604 unsigned int cflags)
1606 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1609 req->compl.cflags = cflags;
1610 req->flags |= REQ_F_COMPLETE_INLINE;
1613 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1614 long res, unsigned cflags)
1616 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1617 io_req_complete_state(req, res, cflags);
1619 io_req_complete_post(req, res, cflags);
1622 static inline void io_req_complete(struct io_kiocb *req, long res)
1624 __io_req_complete(req, 0, res, 0);
1627 static void io_req_complete_failed(struct io_kiocb *req, long res)
1629 req_set_fail_links(req);
1631 io_req_complete_post(req, res, 0);
1634 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1635 struct io_comp_state *cs)
1637 spin_lock_irq(&ctx->completion_lock);
1638 list_splice_init(&cs->locked_free_list, &cs->free_list);
1639 cs->locked_free_nr = 0;
1640 spin_unlock_irq(&ctx->completion_lock);
1643 /* Returns true IFF there are requests in the cache */
1644 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1646 struct io_submit_state *state = &ctx->submit_state;
1647 struct io_comp_state *cs = &state->comp;
1651 * If we have more than a batch's worth of requests in our IRQ side
1652 * locked cache, grab the lock and move them over to our submission
1655 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1656 io_flush_cached_locked_reqs(ctx, cs);
1658 nr = state->free_reqs;
1659 while (!list_empty(&cs->free_list)) {
1660 struct io_kiocb *req = list_first_entry(&cs->free_list,
1661 struct io_kiocb, compl.list);
1663 list_del(&req->compl.list);
1664 state->reqs[nr++] = req;
1665 if (nr == ARRAY_SIZE(state->reqs))
1669 state->free_reqs = nr;
1673 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1675 struct io_submit_state *state = &ctx->submit_state;
1677 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1679 if (!state->free_reqs) {
1680 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1683 if (io_flush_cached_reqs(ctx))
1686 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1690 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1691 * retry single alloc to be on the safe side.
1693 if (unlikely(ret <= 0)) {
1694 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1695 if (!state->reqs[0])
1699 state->free_reqs = ret;
1703 return state->reqs[state->free_reqs];
1706 static inline void io_put_file(struct file *file)
1712 static void io_dismantle_req(struct io_kiocb *req)
1714 unsigned int flags = req->flags;
1716 if (!(flags & REQ_F_FIXED_FILE))
1717 io_put_file(req->file);
1718 if (flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1722 if (req->flags & REQ_F_INFLIGHT) {
1723 struct io_ring_ctx *ctx = req->ctx;
1724 unsigned long flags;
1726 spin_lock_irqsave(&ctx->inflight_lock, flags);
1727 list_del(&req->inflight_entry);
1728 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1729 req->flags &= ~REQ_F_INFLIGHT;
1732 if (req->fixed_rsrc_refs)
1733 percpu_ref_put(req->fixed_rsrc_refs);
1734 if (req->async_data)
1735 kfree(req->async_data);
1736 if (req->work.creds) {
1737 put_cred(req->work.creds);
1738 req->work.creds = NULL;
1742 /* must to be called somewhat shortly after putting a request */
1743 static inline void io_put_task(struct task_struct *task, int nr)
1745 struct io_uring_task *tctx = task->io_uring;
1747 percpu_counter_sub(&tctx->inflight, nr);
1748 if (unlikely(atomic_read(&tctx->in_idle)))
1749 wake_up(&tctx->wait);
1750 put_task_struct_many(task, nr);
1753 static void __io_free_req(struct io_kiocb *req)
1755 struct io_ring_ctx *ctx = req->ctx;
1757 io_dismantle_req(req);
1758 io_put_task(req->task, 1);
1760 kmem_cache_free(req_cachep, req);
1761 percpu_ref_put(&ctx->refs);
1764 static inline void io_remove_next_linked(struct io_kiocb *req)
1766 struct io_kiocb *nxt = req->link;
1768 req->link = nxt->link;
1772 static bool io_kill_linked_timeout(struct io_kiocb *req)
1773 __must_hold(&req->ctx->completion_lock)
1775 struct io_kiocb *link = req->link;
1778 * Can happen if a linked timeout fired and link had been like
1779 * req -> link t-out -> link t-out [-> ...]
1781 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1782 struct io_timeout_data *io = link->async_data;
1785 io_remove_next_linked(req);
1786 link->timeout.head = NULL;
1787 ret = hrtimer_try_to_cancel(&io->timer);
1789 io_cqring_fill_event(link, -ECANCELED);
1790 io_put_req_deferred(link, 1);
1797 static void io_fail_links(struct io_kiocb *req)
1798 __must_hold(&req->ctx->completion_lock)
1800 struct io_kiocb *nxt, *link = req->link;
1807 trace_io_uring_fail_link(req, link);
1808 io_cqring_fill_event(link, -ECANCELED);
1809 io_put_req_deferred(link, 2);
1814 static bool io_disarm_next(struct io_kiocb *req)
1815 __must_hold(&req->ctx->completion_lock)
1817 bool posted = false;
1819 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1820 posted = io_kill_linked_timeout(req);
1821 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1822 posted |= (req->link != NULL);
1828 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1830 struct io_kiocb *nxt;
1833 * If LINK is set, we have dependent requests in this chain. If we
1834 * didn't fail this request, queue the first one up, moving any other
1835 * dependencies to the next request. In case of failure, fail the rest
1838 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1839 struct io_ring_ctx *ctx = req->ctx;
1840 unsigned long flags;
1843 spin_lock_irqsave(&ctx->completion_lock, flags);
1844 posted = io_disarm_next(req);
1846 io_commit_cqring(req->ctx);
1847 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1849 io_cqring_ev_posted(ctx);
1856 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1858 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1860 return __io_req_find_next(req);
1863 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1867 if (ctx->submit_state.comp.nr) {
1868 mutex_lock(&ctx->uring_lock);
1869 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1870 mutex_unlock(&ctx->uring_lock);
1872 percpu_ref_put(&ctx->refs);
1875 static bool __tctx_task_work(struct io_uring_task *tctx)
1877 struct io_ring_ctx *ctx = NULL;
1878 struct io_wq_work_list list;
1879 struct io_wq_work_node *node;
1881 if (wq_list_empty(&tctx->task_list))
1884 spin_lock_irq(&tctx->task_lock);
1885 list = tctx->task_list;
1886 INIT_WQ_LIST(&tctx->task_list);
1887 spin_unlock_irq(&tctx->task_lock);
1891 struct io_wq_work_node *next = node->next;
1892 struct io_kiocb *req;
1894 req = container_of(node, struct io_kiocb, io_task_work.node);
1895 if (req->ctx != ctx) {
1896 ctx_flush_and_put(ctx);
1898 percpu_ref_get(&ctx->refs);
1901 req->task_work.func(&req->task_work);
1905 ctx_flush_and_put(ctx);
1906 return list.first != NULL;
1909 static void tctx_task_work(struct callback_head *cb)
1911 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1913 clear_bit(0, &tctx->task_state);
1915 while (__tctx_task_work(tctx))
1919 static int io_req_task_work_add(struct io_kiocb *req)
1921 struct task_struct *tsk = req->task;
1922 struct io_uring_task *tctx = tsk->io_uring;
1923 enum task_work_notify_mode notify;
1924 struct io_wq_work_node *node, *prev;
1925 unsigned long flags;
1928 if (unlikely(tsk->flags & PF_EXITING))
1931 WARN_ON_ONCE(!tctx);
1933 spin_lock_irqsave(&tctx->task_lock, flags);
1934 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1935 spin_unlock_irqrestore(&tctx->task_lock, flags);
1937 /* task_work already pending, we're done */
1938 if (test_bit(0, &tctx->task_state) ||
1939 test_and_set_bit(0, &tctx->task_state))
1943 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1944 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1945 * processing task_work. There's no reliable way to tell if TWA_RESUME
1948 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1950 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1951 wake_up_process(tsk);
1956 * Slow path - we failed, find and delete work. if the work is not
1957 * in the list, it got run and we're fine.
1959 spin_lock_irqsave(&tctx->task_lock, flags);
1960 wq_list_for_each(node, prev, &tctx->task_list) {
1961 if (&req->io_task_work.node == node) {
1962 wq_list_del(&tctx->task_list, node, prev);
1967 spin_unlock_irqrestore(&tctx->task_lock, flags);
1968 clear_bit(0, &tctx->task_state);
1972 static bool io_run_task_work_head(struct callback_head **work_head)
1974 struct callback_head *work, *next;
1975 bool executed = false;
1978 work = xchg(work_head, NULL);
1994 static void io_task_work_add_head(struct callback_head **work_head,
1995 struct callback_head *task_work)
1997 struct callback_head *head;
2000 head = READ_ONCE(*work_head);
2001 task_work->next = head;
2002 } while (cmpxchg(work_head, head, task_work) != head);
2005 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2006 task_work_func_t cb)
2008 init_task_work(&req->task_work, cb);
2009 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2012 static void io_req_task_cancel(struct callback_head *cb)
2014 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2015 struct io_ring_ctx *ctx = req->ctx;
2017 /* ctx is guaranteed to stay alive while we hold uring_lock */
2018 mutex_lock(&ctx->uring_lock);
2019 io_req_complete_failed(req, req->result);
2020 mutex_unlock(&ctx->uring_lock);
2023 static void __io_req_task_submit(struct io_kiocb *req)
2025 struct io_ring_ctx *ctx = req->ctx;
2027 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2028 mutex_lock(&ctx->uring_lock);
2029 if (!(current->flags & PF_EXITING) && !current->in_execve)
2030 __io_queue_sqe(req);
2032 io_req_complete_failed(req, -EFAULT);
2033 mutex_unlock(&ctx->uring_lock);
2036 static void io_req_task_submit(struct callback_head *cb)
2038 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2040 __io_req_task_submit(req);
2043 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2046 req->task_work.func = io_req_task_cancel;
2048 if (unlikely(io_req_task_work_add(req)))
2049 io_req_task_work_add_fallback(req, io_req_task_cancel);
2052 static void io_req_task_queue(struct io_kiocb *req)
2054 req->task_work.func = io_req_task_submit;
2056 if (unlikely(io_req_task_work_add(req)))
2057 io_req_task_queue_fail(req, -ECANCELED);
2060 static inline void io_queue_next(struct io_kiocb *req)
2062 struct io_kiocb *nxt = io_req_find_next(req);
2065 io_req_task_queue(nxt);
2068 static void io_free_req(struct io_kiocb *req)
2075 struct task_struct *task;
2080 static inline void io_init_req_batch(struct req_batch *rb)
2087 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2088 struct req_batch *rb)
2091 io_put_task(rb->task, rb->task_refs);
2093 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2096 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2097 struct io_submit_state *state)
2100 io_dismantle_req(req);
2102 if (req->task != rb->task) {
2104 io_put_task(rb->task, rb->task_refs);
2105 rb->task = req->task;
2111 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2112 state->reqs[state->free_reqs++] = req;
2114 list_add(&req->compl.list, &state->comp.free_list);
2117 static void io_submit_flush_completions(struct io_comp_state *cs,
2118 struct io_ring_ctx *ctx)
2121 struct io_kiocb *req;
2122 struct req_batch rb;
2124 io_init_req_batch(&rb);
2125 spin_lock_irq(&ctx->completion_lock);
2126 for (i = 0; i < nr; i++) {
2128 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2130 io_commit_cqring(ctx);
2131 spin_unlock_irq(&ctx->completion_lock);
2133 io_cqring_ev_posted(ctx);
2134 for (i = 0; i < nr; i++) {
2137 /* submission and completion refs */
2138 if (req_ref_sub_and_test(req, 2))
2139 io_req_free_batch(&rb, req, &ctx->submit_state);
2142 io_req_free_batch_finish(ctx, &rb);
2147 * Drop reference to request, return next in chain (if there is one) if this
2148 * was the last reference to this request.
2150 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2152 struct io_kiocb *nxt = NULL;
2154 if (req_ref_put_and_test(req)) {
2155 nxt = io_req_find_next(req);
2161 static inline void io_put_req(struct io_kiocb *req)
2163 if (req_ref_put_and_test(req))
2167 static void io_put_req_deferred_cb(struct callback_head *cb)
2169 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2174 static void io_free_req_deferred(struct io_kiocb *req)
2176 req->task_work.func = io_put_req_deferred_cb;
2177 if (unlikely(io_req_task_work_add(req)))
2178 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2181 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2183 if (req_ref_sub_and_test(req, refs))
2184 io_free_req_deferred(req);
2187 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2189 /* See comment at the top of this file */
2191 return __io_cqring_events(ctx);
2194 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2196 struct io_rings *rings = ctx->rings;
2198 /* make sure SQ entry isn't read before tail */
2199 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2202 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2204 unsigned int cflags;
2206 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2207 cflags |= IORING_CQE_F_BUFFER;
2208 req->flags &= ~REQ_F_BUFFER_SELECTED;
2213 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2215 struct io_buffer *kbuf;
2217 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2218 return io_put_kbuf(req, kbuf);
2221 static inline bool io_run_task_work(void)
2224 * Not safe to run on exiting task, and the task_work handling will
2225 * not add work to such a task.
2227 if (unlikely(current->flags & PF_EXITING))
2229 if (current->task_works) {
2230 __set_current_state(TASK_RUNNING);
2239 * Find and free completed poll iocbs
2241 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2242 struct list_head *done)
2244 struct req_batch rb;
2245 struct io_kiocb *req;
2247 /* order with ->result store in io_complete_rw_iopoll() */
2250 io_init_req_batch(&rb);
2251 while (!list_empty(done)) {
2254 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2255 list_del(&req->inflight_entry);
2257 if (READ_ONCE(req->result) == -EAGAIN &&
2258 !(req->flags & REQ_F_DONT_REISSUE)) {
2259 req->iopoll_completed = 0;
2261 io_queue_async_work(req);
2265 if (req->flags & REQ_F_BUFFER_SELECTED)
2266 cflags = io_put_rw_kbuf(req);
2268 __io_cqring_fill_event(req, req->result, cflags);
2271 if (req_ref_put_and_test(req))
2272 io_req_free_batch(&rb, req, &ctx->submit_state);
2275 io_commit_cqring(ctx);
2276 io_cqring_ev_posted_iopoll(ctx);
2277 io_req_free_batch_finish(ctx, &rb);
2280 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2283 struct io_kiocb *req, *tmp;
2289 * Only spin for completions if we don't have multiple devices hanging
2290 * off our complete list, and we're under the requested amount.
2292 spin = !ctx->poll_multi_file && *nr_events < min;
2295 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2296 struct kiocb *kiocb = &req->rw.kiocb;
2299 * Move completed and retryable entries to our local lists.
2300 * If we find a request that requires polling, break out
2301 * and complete those lists first, if we have entries there.
2303 if (READ_ONCE(req->iopoll_completed)) {
2304 list_move_tail(&req->inflight_entry, &done);
2307 if (!list_empty(&done))
2310 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2314 /* iopoll may have completed current req */
2315 if (READ_ONCE(req->iopoll_completed))
2316 list_move_tail(&req->inflight_entry, &done);
2323 if (!list_empty(&done))
2324 io_iopoll_complete(ctx, nr_events, &done);
2330 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2331 * non-spinning poll check - we'll still enter the driver poll loop, but only
2332 * as a non-spinning completion check.
2334 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2337 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2340 ret = io_do_iopoll(ctx, nr_events, min);
2343 if (*nr_events >= min)
2351 * We can't just wait for polled events to come to us, we have to actively
2352 * find and complete them.
2354 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2356 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2359 mutex_lock(&ctx->uring_lock);
2360 while (!list_empty(&ctx->iopoll_list)) {
2361 unsigned int nr_events = 0;
2363 io_do_iopoll(ctx, &nr_events, 0);
2365 /* let it sleep and repeat later if can't complete a request */
2369 * Ensure we allow local-to-the-cpu processing to take place,
2370 * in this case we need to ensure that we reap all events.
2371 * Also let task_work, etc. to progress by releasing the mutex
2373 if (need_resched()) {
2374 mutex_unlock(&ctx->uring_lock);
2376 mutex_lock(&ctx->uring_lock);
2379 mutex_unlock(&ctx->uring_lock);
2382 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2384 unsigned int nr_events = 0;
2385 int iters = 0, ret = 0;
2388 * We disallow the app entering submit/complete with polling, but we
2389 * still need to lock the ring to prevent racing with polled issue
2390 * that got punted to a workqueue.
2392 mutex_lock(&ctx->uring_lock);
2395 * Don't enter poll loop if we already have events pending.
2396 * If we do, we can potentially be spinning for commands that
2397 * already triggered a CQE (eg in error).
2399 if (test_bit(0, &ctx->cq_check_overflow))
2400 __io_cqring_overflow_flush(ctx, false);
2401 if (io_cqring_events(ctx))
2405 * If a submit got punted to a workqueue, we can have the
2406 * application entering polling for a command before it gets
2407 * issued. That app will hold the uring_lock for the duration
2408 * of the poll right here, so we need to take a breather every
2409 * now and then to ensure that the issue has a chance to add
2410 * the poll to the issued list. Otherwise we can spin here
2411 * forever, while the workqueue is stuck trying to acquire the
2414 if (!(++iters & 7)) {
2415 mutex_unlock(&ctx->uring_lock);
2417 mutex_lock(&ctx->uring_lock);
2420 ret = io_iopoll_getevents(ctx, &nr_events, min);
2424 } while (min && !nr_events && !need_resched());
2426 mutex_unlock(&ctx->uring_lock);
2430 static void kiocb_end_write(struct io_kiocb *req)
2433 * Tell lockdep we inherited freeze protection from submission
2436 if (req->flags & REQ_F_ISREG) {
2437 struct super_block *sb = file_inode(req->file)->i_sb;
2439 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2445 static bool io_resubmit_prep(struct io_kiocb *req)
2447 struct io_async_rw *rw = req->async_data;
2450 return !io_req_prep_async(req);
2451 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2452 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2456 static bool io_rw_should_reissue(struct io_kiocb *req)
2458 umode_t mode = file_inode(req->file)->i_mode;
2459 struct io_ring_ctx *ctx = req->ctx;
2461 if (!S_ISBLK(mode) && !S_ISREG(mode))
2463 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2464 !(ctx->flags & IORING_SETUP_IOPOLL)))
2467 * If ref is dying, we might be running poll reap from the exit work.
2468 * Don't attempt to reissue from that path, just let it fail with
2471 if (percpu_ref_is_dying(&ctx->refs))
2476 static bool io_rw_should_reissue(struct io_kiocb *req)
2482 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2483 unsigned int issue_flags)
2487 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2488 kiocb_end_write(req);
2489 if (res != req->result) {
2490 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2491 io_rw_should_reissue(req)) {
2492 req->flags |= REQ_F_REISSUE;
2495 req_set_fail_links(req);
2497 if (req->flags & REQ_F_BUFFER_SELECTED)
2498 cflags = io_put_rw_kbuf(req);
2499 __io_req_complete(req, issue_flags, res, cflags);
2502 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2504 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 __io_complete_rw(req, res, res2, 0);
2509 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2511 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2513 if (kiocb->ki_flags & IOCB_WRITE)
2514 kiocb_end_write(req);
2515 if (unlikely(res != req->result)) {
2519 if (res == -EAGAIN && io_rw_should_reissue(req) &&
2520 io_resubmit_prep(req))
2524 req_set_fail_links(req);
2525 req->flags |= REQ_F_DONT_REISSUE;
2529 WRITE_ONCE(req->result, res);
2530 /* order with io_iopoll_complete() checking ->result */
2532 WRITE_ONCE(req->iopoll_completed, 1);
2536 * After the iocb has been issued, it's safe to be found on the poll list.
2537 * Adding the kiocb to the list AFTER submission ensures that we don't
2538 * find it from a io_iopoll_getevents() thread before the issuer is done
2539 * accessing the kiocb cookie.
2541 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2543 struct io_ring_ctx *ctx = req->ctx;
2546 * Track whether we have multiple files in our lists. This will impact
2547 * how we do polling eventually, not spinning if we're on potentially
2548 * different devices.
2550 if (list_empty(&ctx->iopoll_list)) {
2551 ctx->poll_multi_file = false;
2552 } else if (!ctx->poll_multi_file) {
2553 struct io_kiocb *list_req;
2555 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2557 if (list_req->file != req->file)
2558 ctx->poll_multi_file = true;
2562 * For fast devices, IO may have already completed. If it has, add
2563 * it to the front so we find it first.
2565 if (READ_ONCE(req->iopoll_completed))
2566 list_add(&req->inflight_entry, &ctx->iopoll_list);
2568 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2571 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2572 * task context or in io worker task context. If current task context is
2573 * sq thread, we don't need to check whether should wake up sq thread.
2575 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2576 wq_has_sleeper(&ctx->sq_data->wait))
2577 wake_up(&ctx->sq_data->wait);
2580 static inline void io_state_file_put(struct io_submit_state *state)
2582 if (state->file_refs) {
2583 fput_many(state->file, state->file_refs);
2584 state->file_refs = 0;
2589 * Get as many references to a file as we have IOs left in this submission,
2590 * assuming most submissions are for one file, or at least that each file
2591 * has more than one submission.
2593 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2598 if (state->file_refs) {
2599 if (state->fd == fd) {
2603 io_state_file_put(state);
2605 state->file = fget_many(fd, state->ios_left);
2606 if (unlikely(!state->file))
2610 state->file_refs = state->ios_left - 1;
2614 static bool io_bdev_nowait(struct block_device *bdev)
2616 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2620 * If we tracked the file through the SCM inflight mechanism, we could support
2621 * any file. For now, just ensure that anything potentially problematic is done
2624 static bool __io_file_supports_async(struct file *file, int rw)
2626 umode_t mode = file_inode(file)->i_mode;
2628 if (S_ISBLK(mode)) {
2629 if (IS_ENABLED(CONFIG_BLOCK) &&
2630 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2634 if (S_ISCHR(mode) || S_ISSOCK(mode))
2636 if (S_ISREG(mode)) {
2637 if (IS_ENABLED(CONFIG_BLOCK) &&
2638 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2639 file->f_op != &io_uring_fops)
2644 /* any ->read/write should understand O_NONBLOCK */
2645 if (file->f_flags & O_NONBLOCK)
2648 if (!(file->f_mode & FMODE_NOWAIT))
2652 return file->f_op->read_iter != NULL;
2654 return file->f_op->write_iter != NULL;
2657 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2659 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2661 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2664 return __io_file_supports_async(req->file, rw);
2667 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2669 struct io_ring_ctx *ctx = req->ctx;
2670 struct kiocb *kiocb = &req->rw.kiocb;
2671 struct file *file = req->file;
2675 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2676 req->flags |= REQ_F_ISREG;
2678 kiocb->ki_pos = READ_ONCE(sqe->off);
2679 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2680 req->flags |= REQ_F_CUR_POS;
2681 kiocb->ki_pos = file->f_pos;
2683 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2684 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2685 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2689 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2690 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2691 req->flags |= REQ_F_NOWAIT;
2693 ioprio = READ_ONCE(sqe->ioprio);
2695 ret = ioprio_check_cap(ioprio);
2699 kiocb->ki_ioprio = ioprio;
2701 kiocb->ki_ioprio = get_current_ioprio();
2703 if (ctx->flags & IORING_SETUP_IOPOLL) {
2704 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2705 !kiocb->ki_filp->f_op->iopoll)
2708 kiocb->ki_flags |= IOCB_HIPRI;
2709 kiocb->ki_complete = io_complete_rw_iopoll;
2710 req->iopoll_completed = 0;
2712 if (kiocb->ki_flags & IOCB_HIPRI)
2714 kiocb->ki_complete = io_complete_rw;
2717 req->rw.addr = READ_ONCE(sqe->addr);
2718 req->rw.len = READ_ONCE(sqe->len);
2719 req->buf_index = READ_ONCE(sqe->buf_index);
2723 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2729 case -ERESTARTNOINTR:
2730 case -ERESTARTNOHAND:
2731 case -ERESTART_RESTARTBLOCK:
2733 * We can't just restart the syscall, since previously
2734 * submitted sqes may already be in progress. Just fail this
2740 kiocb->ki_complete(kiocb, ret, 0);
2744 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2745 unsigned int issue_flags)
2747 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2748 struct io_async_rw *io = req->async_data;
2749 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2751 /* add previously done IO, if any */
2752 if (io && io->bytes_done > 0) {
2754 ret = io->bytes_done;
2756 ret += io->bytes_done;
2759 if (req->flags & REQ_F_CUR_POS)
2760 req->file->f_pos = kiocb->ki_pos;
2761 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2762 __io_complete_rw(req, ret, 0, issue_flags);
2764 io_rw_done(kiocb, ret);
2766 if (check_reissue && req->flags & REQ_F_REISSUE) {
2767 req->flags &= ~REQ_F_REISSUE;
2768 if (!io_resubmit_prep(req)) {
2770 io_queue_async_work(req);
2774 req_set_fail_links(req);
2775 if (req->flags & REQ_F_BUFFER_SELECTED)
2776 cflags = io_put_rw_kbuf(req);
2777 __io_req_complete(req, issue_flags, ret, cflags);
2782 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2784 struct io_ring_ctx *ctx = req->ctx;
2785 size_t len = req->rw.len;
2786 struct io_mapped_ubuf *imu;
2787 u16 index, buf_index = req->buf_index;
2788 u64 buf_end, buf_addr = req->rw.addr;
2791 if (unlikely(buf_index >= ctx->nr_user_bufs))
2793 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2794 imu = &ctx->user_bufs[index];
2795 buf_addr = req->rw.addr;
2797 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2799 /* not inside the mapped region */
2800 if (buf_addr < imu->ubuf || buf_end > imu->ubuf + imu->len)
2804 * May not be a start of buffer, set size appropriately
2805 * and advance us to the beginning.
2807 offset = buf_addr - imu->ubuf;
2808 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2812 * Don't use iov_iter_advance() here, as it's really slow for
2813 * using the latter parts of a big fixed buffer - it iterates
2814 * over each segment manually. We can cheat a bit here, because
2817 * 1) it's a BVEC iter, we set it up
2818 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2819 * first and last bvec
2821 * So just find our index, and adjust the iterator afterwards.
2822 * If the offset is within the first bvec (or the whole first
2823 * bvec, just use iov_iter_advance(). This makes it easier
2824 * since we can just skip the first segment, which may not
2825 * be PAGE_SIZE aligned.
2827 const struct bio_vec *bvec = imu->bvec;
2829 if (offset <= bvec->bv_len) {
2830 iov_iter_advance(iter, offset);
2832 unsigned long seg_skip;
2834 /* skip first vec */
2835 offset -= bvec->bv_len;
2836 seg_skip = 1 + (offset >> PAGE_SHIFT);
2838 iter->bvec = bvec + seg_skip;
2839 iter->nr_segs -= seg_skip;
2840 iter->count -= bvec->bv_len + offset;
2841 iter->iov_offset = offset & ~PAGE_MASK;
2848 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2851 mutex_unlock(&ctx->uring_lock);
2854 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2857 * "Normal" inline submissions always hold the uring_lock, since we
2858 * grab it from the system call. Same is true for the SQPOLL offload.
2859 * The only exception is when we've detached the request and issue it
2860 * from an async worker thread, grab the lock for that case.
2863 mutex_lock(&ctx->uring_lock);
2866 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2867 int bgid, struct io_buffer *kbuf,
2870 struct io_buffer *head;
2872 if (req->flags & REQ_F_BUFFER_SELECTED)
2875 io_ring_submit_lock(req->ctx, needs_lock);
2877 lockdep_assert_held(&req->ctx->uring_lock);
2879 head = xa_load(&req->ctx->io_buffers, bgid);
2881 if (!list_empty(&head->list)) {
2882 kbuf = list_last_entry(&head->list, struct io_buffer,
2884 list_del(&kbuf->list);
2887 xa_erase(&req->ctx->io_buffers, bgid);
2889 if (*len > kbuf->len)
2892 kbuf = ERR_PTR(-ENOBUFS);
2895 io_ring_submit_unlock(req->ctx, needs_lock);
2900 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2903 struct io_buffer *kbuf;
2906 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2907 bgid = req->buf_index;
2908 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2911 req->rw.addr = (u64) (unsigned long) kbuf;
2912 req->flags |= REQ_F_BUFFER_SELECTED;
2913 return u64_to_user_ptr(kbuf->addr);
2916 #ifdef CONFIG_COMPAT
2917 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2920 struct compat_iovec __user *uiov;
2921 compat_ssize_t clen;
2925 uiov = u64_to_user_ptr(req->rw.addr);
2926 if (!access_ok(uiov, sizeof(*uiov)))
2928 if (__get_user(clen, &uiov->iov_len))
2934 buf = io_rw_buffer_select(req, &len, needs_lock);
2936 return PTR_ERR(buf);
2937 iov[0].iov_base = buf;
2938 iov[0].iov_len = (compat_size_t) len;
2943 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2946 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2950 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2953 len = iov[0].iov_len;
2956 buf = io_rw_buffer_select(req, &len, needs_lock);
2958 return PTR_ERR(buf);
2959 iov[0].iov_base = buf;
2960 iov[0].iov_len = len;
2964 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2967 if (req->flags & REQ_F_BUFFER_SELECTED) {
2968 struct io_buffer *kbuf;
2970 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2971 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2972 iov[0].iov_len = kbuf->len;
2975 if (req->rw.len != 1)
2978 #ifdef CONFIG_COMPAT
2979 if (req->ctx->compat)
2980 return io_compat_import(req, iov, needs_lock);
2983 return __io_iov_buffer_select(req, iov, needs_lock);
2986 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2987 struct iov_iter *iter, bool needs_lock)
2989 void __user *buf = u64_to_user_ptr(req->rw.addr);
2990 size_t sqe_len = req->rw.len;
2991 u8 opcode = req->opcode;
2994 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2996 return io_import_fixed(req, rw, iter);
2999 /* buffer index only valid with fixed read/write, or buffer select */
3000 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3003 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3004 if (req->flags & REQ_F_BUFFER_SELECT) {
3005 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3007 return PTR_ERR(buf);
3008 req->rw.len = sqe_len;
3011 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3016 if (req->flags & REQ_F_BUFFER_SELECT) {
3017 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3019 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3024 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3028 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3030 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3034 * For files that don't have ->read_iter() and ->write_iter(), handle them
3035 * by looping over ->read() or ->write() manually.
3037 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3039 struct kiocb *kiocb = &req->rw.kiocb;
3040 struct file *file = req->file;
3044 * Don't support polled IO through this interface, and we can't
3045 * support non-blocking either. For the latter, this just causes
3046 * the kiocb to be handled from an async context.
3048 if (kiocb->ki_flags & IOCB_HIPRI)
3050 if (kiocb->ki_flags & IOCB_NOWAIT)
3053 while (iov_iter_count(iter)) {
3057 if (!iov_iter_is_bvec(iter)) {
3058 iovec = iov_iter_iovec(iter);
3060 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3061 iovec.iov_len = req->rw.len;
3065 nr = file->f_op->read(file, iovec.iov_base,
3066 iovec.iov_len, io_kiocb_ppos(kiocb));
3068 nr = file->f_op->write(file, iovec.iov_base,
3069 iovec.iov_len, io_kiocb_ppos(kiocb));
3078 if (nr != iovec.iov_len)
3082 iov_iter_advance(iter, nr);
3088 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3089 const struct iovec *fast_iov, struct iov_iter *iter)
3091 struct io_async_rw *rw = req->async_data;
3093 memcpy(&rw->iter, iter, sizeof(*iter));
3094 rw->free_iovec = iovec;
3096 /* can only be fixed buffers, no need to do anything */
3097 if (iov_iter_is_bvec(iter))
3100 unsigned iov_off = 0;
3102 rw->iter.iov = rw->fast_iov;
3103 if (iter->iov != fast_iov) {
3104 iov_off = iter->iov - fast_iov;
3105 rw->iter.iov += iov_off;
3107 if (rw->fast_iov != fast_iov)
3108 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3109 sizeof(struct iovec) * iter->nr_segs);
3111 req->flags |= REQ_F_NEED_CLEANUP;
3115 static inline int io_alloc_async_data(struct io_kiocb *req)
3117 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3118 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3119 return req->async_data == NULL;
3122 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3123 const struct iovec *fast_iov,
3124 struct iov_iter *iter, bool force)
3126 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3128 if (!req->async_data) {
3129 if (io_alloc_async_data(req)) {
3134 io_req_map_rw(req, iovec, fast_iov, iter);
3139 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3141 struct io_async_rw *iorw = req->async_data;
3142 struct iovec *iov = iorw->fast_iov;
3145 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3146 if (unlikely(ret < 0))
3149 iorw->bytes_done = 0;
3150 iorw->free_iovec = iov;
3152 req->flags |= REQ_F_NEED_CLEANUP;
3156 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3158 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3160 return io_prep_rw(req, sqe);
3164 * This is our waitqueue callback handler, registered through lock_page_async()
3165 * when we initially tried to do the IO with the iocb armed our waitqueue.
3166 * This gets called when the page is unlocked, and we generally expect that to
3167 * happen when the page IO is completed and the page is now uptodate. This will
3168 * queue a task_work based retry of the operation, attempting to copy the data
3169 * again. If the latter fails because the page was NOT uptodate, then we will
3170 * do a thread based blocking retry of the operation. That's the unexpected
3173 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3174 int sync, void *arg)
3176 struct wait_page_queue *wpq;
3177 struct io_kiocb *req = wait->private;
3178 struct wait_page_key *key = arg;
3180 wpq = container_of(wait, struct wait_page_queue, wait);
3182 if (!wake_page_match(wpq, key))
3185 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3186 list_del_init(&wait->entry);
3188 /* submit ref gets dropped, acquire a new one */
3190 io_req_task_queue(req);
3195 * This controls whether a given IO request should be armed for async page
3196 * based retry. If we return false here, the request is handed to the async
3197 * worker threads for retry. If we're doing buffered reads on a regular file,
3198 * we prepare a private wait_page_queue entry and retry the operation. This
3199 * will either succeed because the page is now uptodate and unlocked, or it
3200 * will register a callback when the page is unlocked at IO completion. Through
3201 * that callback, io_uring uses task_work to setup a retry of the operation.
3202 * That retry will attempt the buffered read again. The retry will generally
3203 * succeed, or in rare cases where it fails, we then fall back to using the
3204 * async worker threads for a blocking retry.
3206 static bool io_rw_should_retry(struct io_kiocb *req)
3208 struct io_async_rw *rw = req->async_data;
3209 struct wait_page_queue *wait = &rw->wpq;
3210 struct kiocb *kiocb = &req->rw.kiocb;
3212 /* never retry for NOWAIT, we just complete with -EAGAIN */
3213 if (req->flags & REQ_F_NOWAIT)
3216 /* Only for buffered IO */
3217 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3221 * just use poll if we can, and don't attempt if the fs doesn't
3222 * support callback based unlocks
3224 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3227 wait->wait.func = io_async_buf_func;
3228 wait->wait.private = req;
3229 wait->wait.flags = 0;
3230 INIT_LIST_HEAD(&wait->wait.entry);
3231 kiocb->ki_flags |= IOCB_WAITQ;
3232 kiocb->ki_flags &= ~IOCB_NOWAIT;
3233 kiocb->ki_waitq = wait;
3237 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3239 if (req->file->f_op->read_iter)
3240 return call_read_iter(req->file, &req->rw.kiocb, iter);
3241 else if (req->file->f_op->read)
3242 return loop_rw_iter(READ, req, iter);
3247 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3249 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3250 struct kiocb *kiocb = &req->rw.kiocb;
3251 struct iov_iter __iter, *iter = &__iter;
3252 struct io_async_rw *rw = req->async_data;
3253 ssize_t io_size, ret, ret2;
3254 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3260 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3264 io_size = iov_iter_count(iter);
3265 req->result = io_size;
3267 /* Ensure we clear previously set non-block flag */
3268 if (!force_nonblock)
3269 kiocb->ki_flags &= ~IOCB_NOWAIT;
3271 kiocb->ki_flags |= IOCB_NOWAIT;
3273 /* If the file doesn't support async, just async punt */
3274 if (force_nonblock && !io_file_supports_async(req, READ)) {
3275 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3276 return ret ?: -EAGAIN;
3279 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3280 if (unlikely(ret)) {
3285 ret = io_iter_do_read(req, iter);
3287 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3288 req->flags &= ~REQ_F_REISSUE;
3289 /* IOPOLL retry should happen for io-wq threads */
3290 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3292 /* no retry on NONBLOCK nor RWF_NOWAIT */
3293 if (req->flags & REQ_F_NOWAIT)
3295 /* some cases will consume bytes even on error returns */
3296 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3298 } else if (ret == -EIOCBQUEUED) {
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, 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);
3412 if (req->flags & REQ_F_REISSUE) {
3413 req->flags &= ~REQ_F_REISSUE;
3418 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3419 * retry them without IOCB_NOWAIT.
3421 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3423 /* no retry on NONBLOCK nor RWF_NOWAIT */
3424 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3426 if (!force_nonblock || ret2 != -EAGAIN) {
3427 /* IOPOLL retry should happen for io-wq threads */
3428 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3431 kiocb_done(kiocb, ret2, issue_flags);
3434 /* some cases will consume bytes even on error returns */
3435 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3436 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3437 return ret ?: -EAGAIN;
3440 /* it's reportedly faster than delegating the null check to kfree() */
3446 static int io_renameat_prep(struct io_kiocb *req,
3447 const struct io_uring_sqe *sqe)
3449 struct io_rename *ren = &req->rename;
3450 const char __user *oldf, *newf;
3452 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3455 ren->old_dfd = READ_ONCE(sqe->fd);
3456 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3457 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3458 ren->new_dfd = READ_ONCE(sqe->len);
3459 ren->flags = READ_ONCE(sqe->rename_flags);
3461 ren->oldpath = getname(oldf);
3462 if (IS_ERR(ren->oldpath))
3463 return PTR_ERR(ren->oldpath);
3465 ren->newpath = getname(newf);
3466 if (IS_ERR(ren->newpath)) {
3467 putname(ren->oldpath);
3468 return PTR_ERR(ren->newpath);
3471 req->flags |= REQ_F_NEED_CLEANUP;
3475 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3477 struct io_rename *ren = &req->rename;
3480 if (issue_flags & IO_URING_F_NONBLOCK)
3483 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3484 ren->newpath, ren->flags);
3486 req->flags &= ~REQ_F_NEED_CLEANUP;
3488 req_set_fail_links(req);
3489 io_req_complete(req, ret);
3493 static int io_unlinkat_prep(struct io_kiocb *req,
3494 const struct io_uring_sqe *sqe)
3496 struct io_unlink *un = &req->unlink;
3497 const char __user *fname;
3499 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3502 un->dfd = READ_ONCE(sqe->fd);
3504 un->flags = READ_ONCE(sqe->unlink_flags);
3505 if (un->flags & ~AT_REMOVEDIR)
3508 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3509 un->filename = getname(fname);
3510 if (IS_ERR(un->filename))
3511 return PTR_ERR(un->filename);
3513 req->flags |= REQ_F_NEED_CLEANUP;
3517 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3519 struct io_unlink *un = &req->unlink;
3522 if (issue_flags & IO_URING_F_NONBLOCK)
3525 if (un->flags & AT_REMOVEDIR)
3526 ret = do_rmdir(un->dfd, un->filename);
3528 ret = do_unlinkat(un->dfd, un->filename);
3530 req->flags &= ~REQ_F_NEED_CLEANUP;
3532 req_set_fail_links(req);
3533 io_req_complete(req, ret);
3537 static int io_shutdown_prep(struct io_kiocb *req,
3538 const struct io_uring_sqe *sqe)
3540 #if defined(CONFIG_NET)
3541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3543 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3547 req->shutdown.how = READ_ONCE(sqe->len);
3554 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3556 #if defined(CONFIG_NET)
3557 struct socket *sock;
3560 if (issue_flags & IO_URING_F_NONBLOCK)
3563 sock = sock_from_file(req->file);
3564 if (unlikely(!sock))
3567 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3569 req_set_fail_links(req);
3570 io_req_complete(req, ret);
3577 static int __io_splice_prep(struct io_kiocb *req,
3578 const struct io_uring_sqe *sqe)
3580 struct io_splice* sp = &req->splice;
3581 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3587 sp->len = READ_ONCE(sqe->len);
3588 sp->flags = READ_ONCE(sqe->splice_flags);
3590 if (unlikely(sp->flags & ~valid_flags))
3593 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3594 (sp->flags & SPLICE_F_FD_IN_FIXED));
3597 req->flags |= REQ_F_NEED_CLEANUP;
3601 static int io_tee_prep(struct io_kiocb *req,
3602 const struct io_uring_sqe *sqe)
3604 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3606 return __io_splice_prep(req, sqe);
3609 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3611 struct io_splice *sp = &req->splice;
3612 struct file *in = sp->file_in;
3613 struct file *out = sp->file_out;
3614 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3617 if (issue_flags & IO_URING_F_NONBLOCK)
3620 ret = do_tee(in, out, sp->len, flags);
3622 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3624 req->flags &= ~REQ_F_NEED_CLEANUP;
3627 req_set_fail_links(req);
3628 io_req_complete(req, ret);
3632 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3634 struct io_splice* sp = &req->splice;
3636 sp->off_in = READ_ONCE(sqe->splice_off_in);
3637 sp->off_out = READ_ONCE(sqe->off);
3638 return __io_splice_prep(req, sqe);
3641 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3643 struct io_splice *sp = &req->splice;
3644 struct file *in = sp->file_in;
3645 struct file *out = sp->file_out;
3646 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3647 loff_t *poff_in, *poff_out;
3650 if (issue_flags & IO_URING_F_NONBLOCK)
3653 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3654 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3657 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3659 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3661 req->flags &= ~REQ_F_NEED_CLEANUP;
3664 req_set_fail_links(req);
3665 io_req_complete(req, ret);
3670 * IORING_OP_NOP just posts a completion event, nothing else.
3672 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3674 struct io_ring_ctx *ctx = req->ctx;
3676 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3679 __io_req_complete(req, issue_flags, 0, 0);
3683 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3685 struct io_ring_ctx *ctx = req->ctx;
3690 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3692 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3695 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3696 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3699 req->sync.off = READ_ONCE(sqe->off);
3700 req->sync.len = READ_ONCE(sqe->len);
3704 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3706 loff_t end = req->sync.off + req->sync.len;
3709 /* fsync always requires a blocking context */
3710 if (issue_flags & IO_URING_F_NONBLOCK)
3713 ret = vfs_fsync_range(req->file, req->sync.off,
3714 end > 0 ? end : LLONG_MAX,
3715 req->sync.flags & IORING_FSYNC_DATASYNC);
3717 req_set_fail_links(req);
3718 io_req_complete(req, ret);
3722 static int io_fallocate_prep(struct io_kiocb *req,
3723 const struct io_uring_sqe *sqe)
3725 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3730 req->sync.off = READ_ONCE(sqe->off);
3731 req->sync.len = READ_ONCE(sqe->addr);
3732 req->sync.mode = READ_ONCE(sqe->len);
3736 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3740 /* fallocate always requiring blocking context */
3741 if (issue_flags & IO_URING_F_NONBLOCK)
3743 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3746 req_set_fail_links(req);
3747 io_req_complete(req, ret);
3751 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3753 const char __user *fname;
3756 if (unlikely(sqe->ioprio || sqe->buf_index))
3758 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3761 /* open.how should be already initialised */
3762 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3763 req->open.how.flags |= O_LARGEFILE;
3765 req->open.dfd = READ_ONCE(sqe->fd);
3766 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3767 req->open.filename = getname(fname);
3768 if (IS_ERR(req->open.filename)) {
3769 ret = PTR_ERR(req->open.filename);
3770 req->open.filename = NULL;
3773 req->open.nofile = rlimit(RLIMIT_NOFILE);
3774 req->flags |= REQ_F_NEED_CLEANUP;
3778 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784 mode = READ_ONCE(sqe->len);
3785 flags = READ_ONCE(sqe->open_flags);
3786 req->open.how = build_open_how(flags, mode);
3787 return __io_openat_prep(req, sqe);
3790 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3792 struct open_how __user *how;
3796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3798 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3799 len = READ_ONCE(sqe->len);
3800 if (len < OPEN_HOW_SIZE_VER0)
3803 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3808 return __io_openat_prep(req, sqe);
3811 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3813 struct open_flags op;
3816 bool resolve_nonblock;
3819 ret = build_open_flags(&req->open.how, &op);
3822 nonblock_set = op.open_flag & O_NONBLOCK;
3823 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3824 if (issue_flags & IO_URING_F_NONBLOCK) {
3826 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3827 * it'll always -EAGAIN
3829 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3831 op.lookup_flags |= LOOKUP_CACHED;
3832 op.open_flag |= O_NONBLOCK;
3835 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3839 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3840 /* only retry if RESOLVE_CACHED wasn't already set by application */
3841 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3842 file == ERR_PTR(-EAGAIN)) {
3844 * We could hang on to this 'fd', but seems like marginal
3845 * gain for something that is now known to be a slower path.
3846 * So just put it, and we'll get a new one when we retry.
3854 ret = PTR_ERR(file);
3856 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3857 file->f_flags &= ~O_NONBLOCK;
3858 fsnotify_open(file);
3859 fd_install(ret, file);
3862 putname(req->open.filename);
3863 req->flags &= ~REQ_F_NEED_CLEANUP;
3865 req_set_fail_links(req);
3866 io_req_complete(req, ret);
3870 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3872 return io_openat2(req, issue_flags);
3875 static int io_remove_buffers_prep(struct io_kiocb *req,
3876 const struct io_uring_sqe *sqe)
3878 struct io_provide_buf *p = &req->pbuf;
3881 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3884 tmp = READ_ONCE(sqe->fd);
3885 if (!tmp || tmp > USHRT_MAX)
3888 memset(p, 0, sizeof(*p));
3890 p->bgid = READ_ONCE(sqe->buf_group);
3894 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3895 int bgid, unsigned nbufs)
3899 /* shouldn't happen */
3903 /* the head kbuf is the list itself */
3904 while (!list_empty(&buf->list)) {
3905 struct io_buffer *nxt;
3907 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3908 list_del(&nxt->list);
3915 xa_erase(&ctx->io_buffers, bgid);
3920 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3922 struct io_provide_buf *p = &req->pbuf;
3923 struct io_ring_ctx *ctx = req->ctx;
3924 struct io_buffer *head;
3926 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3928 io_ring_submit_lock(ctx, !force_nonblock);
3930 lockdep_assert_held(&ctx->uring_lock);
3933 head = xa_load(&ctx->io_buffers, p->bgid);
3935 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3937 req_set_fail_links(req);
3939 /* complete before unlock, IOPOLL may need the lock */
3940 __io_req_complete(req, issue_flags, ret, 0);
3941 io_ring_submit_unlock(ctx, !force_nonblock);
3945 static int io_provide_buffers_prep(struct io_kiocb *req,
3946 const struct io_uring_sqe *sqe)
3949 struct io_provide_buf *p = &req->pbuf;
3952 if (sqe->ioprio || sqe->rw_flags)
3955 tmp = READ_ONCE(sqe->fd);
3956 if (!tmp || tmp > USHRT_MAX)
3959 p->addr = READ_ONCE(sqe->addr);
3960 p->len = READ_ONCE(sqe->len);
3962 size = (unsigned long)p->len * p->nbufs;
3963 if (!access_ok(u64_to_user_ptr(p->addr), size))
3966 p->bgid = READ_ONCE(sqe->buf_group);
3967 tmp = READ_ONCE(sqe->off);
3968 if (tmp > USHRT_MAX)
3974 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3976 struct io_buffer *buf;
3977 u64 addr = pbuf->addr;
3978 int i, bid = pbuf->bid;
3980 for (i = 0; i < pbuf->nbufs; i++) {
3981 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3986 buf->len = pbuf->len;
3991 INIT_LIST_HEAD(&buf->list);
3994 list_add_tail(&buf->list, &(*head)->list);
3998 return i ? i : -ENOMEM;
4001 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4003 struct io_provide_buf *p = &req->pbuf;
4004 struct io_ring_ctx *ctx = req->ctx;
4005 struct io_buffer *head, *list;
4007 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4009 io_ring_submit_lock(ctx, !force_nonblock);
4011 lockdep_assert_held(&ctx->uring_lock);
4013 list = head = xa_load(&ctx->io_buffers, p->bgid);
4015 ret = io_add_buffers(p, &head);
4016 if (ret >= 0 && !list) {
4017 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4019 __io_remove_buffers(ctx, head, p->bgid, -1U);
4022 req_set_fail_links(req);
4023 /* complete before unlock, IOPOLL may need the lock */
4024 __io_req_complete(req, issue_flags, ret, 0);
4025 io_ring_submit_unlock(ctx, !force_nonblock);
4029 static int io_epoll_ctl_prep(struct io_kiocb *req,
4030 const struct io_uring_sqe *sqe)
4032 #if defined(CONFIG_EPOLL)
4033 if (sqe->ioprio || sqe->buf_index)
4035 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4038 req->epoll.epfd = READ_ONCE(sqe->fd);
4039 req->epoll.op = READ_ONCE(sqe->len);
4040 req->epoll.fd = READ_ONCE(sqe->off);
4042 if (ep_op_has_event(req->epoll.op)) {
4043 struct epoll_event __user *ev;
4045 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4046 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4056 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4058 #if defined(CONFIG_EPOLL)
4059 struct io_epoll *ie = &req->epoll;
4061 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4063 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4064 if (force_nonblock && ret == -EAGAIN)
4068 req_set_fail_links(req);
4069 __io_req_complete(req, issue_flags, ret, 0);
4076 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4078 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4079 if (sqe->ioprio || sqe->buf_index || sqe->off)
4081 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4084 req->madvise.addr = READ_ONCE(sqe->addr);
4085 req->madvise.len = READ_ONCE(sqe->len);
4086 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4093 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4096 struct io_madvise *ma = &req->madvise;
4099 if (issue_flags & IO_URING_F_NONBLOCK)
4102 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4104 req_set_fail_links(req);
4105 io_req_complete(req, ret);
4112 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4114 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4116 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4119 req->fadvise.offset = READ_ONCE(sqe->off);
4120 req->fadvise.len = READ_ONCE(sqe->len);
4121 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4125 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4127 struct io_fadvise *fa = &req->fadvise;
4130 if (issue_flags & IO_URING_F_NONBLOCK) {
4131 switch (fa->advice) {
4132 case POSIX_FADV_NORMAL:
4133 case POSIX_FADV_RANDOM:
4134 case POSIX_FADV_SEQUENTIAL:
4141 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4143 req_set_fail_links(req);
4144 io_req_complete(req, ret);
4148 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4150 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4152 if (sqe->ioprio || sqe->buf_index)
4154 if (req->flags & REQ_F_FIXED_FILE)
4157 req->statx.dfd = READ_ONCE(sqe->fd);
4158 req->statx.mask = READ_ONCE(sqe->len);
4159 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4160 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4161 req->statx.flags = READ_ONCE(sqe->statx_flags);
4166 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4168 struct io_statx *ctx = &req->statx;
4171 if (issue_flags & IO_URING_F_NONBLOCK)
4174 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4178 req_set_fail_links(req);
4179 io_req_complete(req, ret);
4183 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4187 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4188 sqe->rw_flags || sqe->buf_index)
4190 if (req->flags & REQ_F_FIXED_FILE)
4193 req->close.fd = READ_ONCE(sqe->fd);
4197 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4199 struct files_struct *files = current->files;
4200 struct io_close *close = &req->close;
4201 struct fdtable *fdt;
4207 spin_lock(&files->file_lock);
4208 fdt = files_fdtable(files);
4209 if (close->fd >= fdt->max_fds) {
4210 spin_unlock(&files->file_lock);
4213 file = fdt->fd[close->fd];
4215 spin_unlock(&files->file_lock);
4219 if (file->f_op == &io_uring_fops) {
4220 spin_unlock(&files->file_lock);
4225 /* if the file has a flush method, be safe and punt to async */
4226 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4227 spin_unlock(&files->file_lock);
4231 ret = __close_fd_get_file(close->fd, &file);
4232 spin_unlock(&files->file_lock);
4239 /* No ->flush() or already async, safely close from here */
4240 ret = filp_close(file, current->files);
4243 req_set_fail_links(req);
4246 __io_req_complete(req, issue_flags, ret, 0);
4250 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4252 struct io_ring_ctx *ctx = req->ctx;
4254 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4256 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4259 req->sync.off = READ_ONCE(sqe->off);
4260 req->sync.len = READ_ONCE(sqe->len);
4261 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4265 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4269 /* sync_file_range always requires a blocking context */
4270 if (issue_flags & IO_URING_F_NONBLOCK)
4273 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4276 req_set_fail_links(req);
4277 io_req_complete(req, ret);
4281 #if defined(CONFIG_NET)
4282 static int io_setup_async_msg(struct io_kiocb *req,
4283 struct io_async_msghdr *kmsg)
4285 struct io_async_msghdr *async_msg = req->async_data;
4289 if (io_alloc_async_data(req)) {
4290 kfree(kmsg->free_iov);
4293 async_msg = req->async_data;
4294 req->flags |= REQ_F_NEED_CLEANUP;
4295 memcpy(async_msg, kmsg, sizeof(*kmsg));
4296 async_msg->msg.msg_name = &async_msg->addr;
4297 /* if were using fast_iov, set it to the new one */
4298 if (!async_msg->free_iov)
4299 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4304 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4305 struct io_async_msghdr *iomsg)
4307 iomsg->msg.msg_name = &iomsg->addr;
4308 iomsg->free_iov = iomsg->fast_iov;
4309 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4310 req->sr_msg.msg_flags, &iomsg->free_iov);
4313 static int io_sendmsg_prep_async(struct io_kiocb *req)
4317 ret = io_sendmsg_copy_hdr(req, req->async_data);
4319 req->flags |= REQ_F_NEED_CLEANUP;
4323 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4325 struct io_sr_msg *sr = &req->sr_msg;
4327 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4330 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4331 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4332 sr->len = READ_ONCE(sqe->len);
4334 #ifdef CONFIG_COMPAT
4335 if (req->ctx->compat)
4336 sr->msg_flags |= MSG_CMSG_COMPAT;
4341 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4343 struct io_async_msghdr iomsg, *kmsg;
4344 struct socket *sock;
4349 sock = sock_from_file(req->file);
4350 if (unlikely(!sock))
4353 kmsg = req->async_data;
4355 ret = io_sendmsg_copy_hdr(req, &iomsg);
4361 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4362 if (flags & MSG_DONTWAIT)
4363 req->flags |= REQ_F_NOWAIT;
4364 else if (issue_flags & IO_URING_F_NONBLOCK)
4365 flags |= MSG_DONTWAIT;
4367 if (flags & MSG_WAITALL)
4368 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4370 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4371 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4372 return io_setup_async_msg(req, kmsg);
4373 if (ret == -ERESTARTSYS)
4376 /* fast path, check for non-NULL to avoid function call */
4378 kfree(kmsg->free_iov);
4379 req->flags &= ~REQ_F_NEED_CLEANUP;
4381 req_set_fail_links(req);
4382 __io_req_complete(req, issue_flags, ret, 0);
4386 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4388 struct io_sr_msg *sr = &req->sr_msg;
4391 struct socket *sock;
4396 sock = sock_from_file(req->file);
4397 if (unlikely(!sock))
4400 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4404 msg.msg_name = NULL;
4405 msg.msg_control = NULL;
4406 msg.msg_controllen = 0;
4407 msg.msg_namelen = 0;
4409 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4410 if (flags & MSG_DONTWAIT)
4411 req->flags |= REQ_F_NOWAIT;
4412 else if (issue_flags & IO_URING_F_NONBLOCK)
4413 flags |= MSG_DONTWAIT;
4415 if (flags & MSG_WAITALL)
4416 min_ret = iov_iter_count(&msg.msg_iter);
4418 msg.msg_flags = flags;
4419 ret = sock_sendmsg(sock, &msg);
4420 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4422 if (ret == -ERESTARTSYS)
4426 req_set_fail_links(req);
4427 __io_req_complete(req, issue_flags, ret, 0);
4431 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4432 struct io_async_msghdr *iomsg)
4434 struct io_sr_msg *sr = &req->sr_msg;
4435 struct iovec __user *uiov;
4439 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4440 &iomsg->uaddr, &uiov, &iov_len);
4444 if (req->flags & REQ_F_BUFFER_SELECT) {
4447 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4449 sr->len = iomsg->fast_iov[0].iov_len;
4450 iomsg->free_iov = NULL;
4452 iomsg->free_iov = iomsg->fast_iov;
4453 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4454 &iomsg->free_iov, &iomsg->msg.msg_iter,
4463 #ifdef CONFIG_COMPAT
4464 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4465 struct io_async_msghdr *iomsg)
4467 struct compat_msghdr __user *msg_compat;
4468 struct io_sr_msg *sr = &req->sr_msg;
4469 struct compat_iovec __user *uiov;
4474 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4475 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4480 uiov = compat_ptr(ptr);
4481 if (req->flags & REQ_F_BUFFER_SELECT) {
4482 compat_ssize_t clen;
4486 if (!access_ok(uiov, sizeof(*uiov)))
4488 if (__get_user(clen, &uiov->iov_len))
4493 iomsg->free_iov = NULL;
4495 iomsg->free_iov = iomsg->fast_iov;
4496 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4497 UIO_FASTIOV, &iomsg->free_iov,
4498 &iomsg->msg.msg_iter, true);
4507 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4508 struct io_async_msghdr *iomsg)
4510 iomsg->msg.msg_name = &iomsg->addr;
4512 #ifdef CONFIG_COMPAT
4513 if (req->ctx->compat)
4514 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4517 return __io_recvmsg_copy_hdr(req, iomsg);
4520 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4523 struct io_sr_msg *sr = &req->sr_msg;
4524 struct io_buffer *kbuf;
4526 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4531 req->flags |= REQ_F_BUFFER_SELECTED;
4535 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4537 return io_put_kbuf(req, req->sr_msg.kbuf);
4540 static int io_recvmsg_prep_async(struct io_kiocb *req)
4544 ret = io_recvmsg_copy_hdr(req, req->async_data);
4546 req->flags |= REQ_F_NEED_CLEANUP;
4550 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4552 struct io_sr_msg *sr = &req->sr_msg;
4554 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4557 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4558 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4559 sr->len = READ_ONCE(sqe->len);
4560 sr->bgid = READ_ONCE(sqe->buf_group);
4562 #ifdef CONFIG_COMPAT
4563 if (req->ctx->compat)
4564 sr->msg_flags |= MSG_CMSG_COMPAT;
4569 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4571 struct io_async_msghdr iomsg, *kmsg;
4572 struct socket *sock;
4573 struct io_buffer *kbuf;
4576 int ret, cflags = 0;
4577 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4579 sock = sock_from_file(req->file);
4580 if (unlikely(!sock))
4583 kmsg = req->async_data;
4585 ret = io_recvmsg_copy_hdr(req, &iomsg);
4591 if (req->flags & REQ_F_BUFFER_SELECT) {
4592 kbuf = io_recv_buffer_select(req, !force_nonblock);
4594 return PTR_ERR(kbuf);
4595 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4596 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4597 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4598 1, req->sr_msg.len);
4601 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4602 if (flags & MSG_DONTWAIT)
4603 req->flags |= REQ_F_NOWAIT;
4604 else if (force_nonblock)
4605 flags |= MSG_DONTWAIT;
4607 if (flags & MSG_WAITALL)
4608 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4610 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4611 kmsg->uaddr, flags);
4612 if (force_nonblock && ret == -EAGAIN)
4613 return io_setup_async_msg(req, kmsg);
4614 if (ret == -ERESTARTSYS)
4617 if (req->flags & REQ_F_BUFFER_SELECTED)
4618 cflags = io_put_recv_kbuf(req);
4619 /* fast path, check for non-NULL to avoid function call */
4621 kfree(kmsg->free_iov);
4622 req->flags &= ~REQ_F_NEED_CLEANUP;
4623 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4624 req_set_fail_links(req);
4625 __io_req_complete(req, issue_flags, ret, cflags);
4629 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4631 struct io_buffer *kbuf;
4632 struct io_sr_msg *sr = &req->sr_msg;
4634 void __user *buf = sr->buf;
4635 struct socket *sock;
4639 int ret, cflags = 0;
4640 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4642 sock = sock_from_file(req->file);
4643 if (unlikely(!sock))
4646 if (req->flags & REQ_F_BUFFER_SELECT) {
4647 kbuf = io_recv_buffer_select(req, !force_nonblock);
4649 return PTR_ERR(kbuf);
4650 buf = u64_to_user_ptr(kbuf->addr);
4653 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4657 msg.msg_name = NULL;
4658 msg.msg_control = NULL;
4659 msg.msg_controllen = 0;
4660 msg.msg_namelen = 0;
4661 msg.msg_iocb = NULL;
4664 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4665 if (flags & MSG_DONTWAIT)
4666 req->flags |= REQ_F_NOWAIT;
4667 else if (force_nonblock)
4668 flags |= MSG_DONTWAIT;
4670 if (flags & MSG_WAITALL)
4671 min_ret = iov_iter_count(&msg.msg_iter);
4673 ret = sock_recvmsg(sock, &msg, flags);
4674 if (force_nonblock && ret == -EAGAIN)
4676 if (ret == -ERESTARTSYS)
4679 if (req->flags & REQ_F_BUFFER_SELECTED)
4680 cflags = io_put_recv_kbuf(req);
4681 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4682 req_set_fail_links(req);
4683 __io_req_complete(req, issue_flags, ret, cflags);
4687 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4689 struct io_accept *accept = &req->accept;
4691 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4693 if (sqe->ioprio || sqe->len || sqe->buf_index)
4696 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4697 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4698 accept->flags = READ_ONCE(sqe->accept_flags);
4699 accept->nofile = rlimit(RLIMIT_NOFILE);
4703 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4705 struct io_accept *accept = &req->accept;
4706 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4707 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4710 if (req->file->f_flags & O_NONBLOCK)
4711 req->flags |= REQ_F_NOWAIT;
4713 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4714 accept->addr_len, accept->flags,
4716 if (ret == -EAGAIN && force_nonblock)
4719 if (ret == -ERESTARTSYS)
4721 req_set_fail_links(req);
4723 __io_req_complete(req, issue_flags, ret, 0);
4727 static int io_connect_prep_async(struct io_kiocb *req)
4729 struct io_async_connect *io = req->async_data;
4730 struct io_connect *conn = &req->connect;
4732 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4735 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4737 struct io_connect *conn = &req->connect;
4739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4741 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4744 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4745 conn->addr_len = READ_ONCE(sqe->addr2);
4749 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4751 struct io_async_connect __io, *io;
4752 unsigned file_flags;
4754 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4756 if (req->async_data) {
4757 io = req->async_data;
4759 ret = move_addr_to_kernel(req->connect.addr,
4760 req->connect.addr_len,
4767 file_flags = force_nonblock ? O_NONBLOCK : 0;
4769 ret = __sys_connect_file(req->file, &io->address,
4770 req->connect.addr_len, file_flags);
4771 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4772 if (req->async_data)
4774 if (io_alloc_async_data(req)) {
4778 memcpy(req->async_data, &__io, sizeof(__io));
4781 if (ret == -ERESTARTSYS)
4785 req_set_fail_links(req);
4786 __io_req_complete(req, issue_flags, ret, 0);
4789 #else /* !CONFIG_NET */
4790 #define IO_NETOP_FN(op) \
4791 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4793 return -EOPNOTSUPP; \
4796 #define IO_NETOP_PREP(op) \
4798 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4800 return -EOPNOTSUPP; \
4803 #define IO_NETOP_PREP_ASYNC(op) \
4805 static int io_##op##_prep_async(struct io_kiocb *req) \
4807 return -EOPNOTSUPP; \
4810 IO_NETOP_PREP_ASYNC(sendmsg);
4811 IO_NETOP_PREP_ASYNC(recvmsg);
4812 IO_NETOP_PREP_ASYNC(connect);
4813 IO_NETOP_PREP(accept);
4816 #endif /* CONFIG_NET */
4818 struct io_poll_table {
4819 struct poll_table_struct pt;
4820 struct io_kiocb *req;
4824 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4825 __poll_t mask, task_work_func_t func)
4829 /* for instances that support it check for an event match first: */
4830 if (mask && !(mask & poll->events))
4833 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4835 list_del_init(&poll->wait.entry);
4838 req->task_work.func = func;
4841 * If this fails, then the task is exiting. When a task exits, the
4842 * work gets canceled, so just cancel this request as well instead
4843 * of executing it. We can't safely execute it anyway, as we may not
4844 * have the needed state needed for it anyway.
4846 ret = io_req_task_work_add(req);
4847 if (unlikely(ret)) {
4848 WRITE_ONCE(poll->canceled, true);
4849 io_req_task_work_add_fallback(req, func);
4854 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4855 __acquires(&req->ctx->completion_lock)
4857 struct io_ring_ctx *ctx = req->ctx;
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 struct poll_table_struct pt = { ._key = poll->events };
4862 req->result = vfs_poll(req->file, &pt) & poll->events;
4865 spin_lock_irq(&ctx->completion_lock);
4866 if (!req->result && !READ_ONCE(poll->canceled)) {
4867 add_wait_queue(poll->head, &poll->wait);
4874 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4876 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4877 if (req->opcode == IORING_OP_POLL_ADD)
4878 return req->async_data;
4879 return req->apoll->double_poll;
4882 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4884 if (req->opcode == IORING_OP_POLL_ADD)
4886 return &req->apoll->poll;
4889 static void io_poll_remove_double(struct io_kiocb *req)
4891 struct io_poll_iocb *poll = io_poll_get_double(req);
4893 lockdep_assert_held(&req->ctx->completion_lock);
4895 if (poll && poll->head) {
4896 struct wait_queue_head *head = poll->head;
4898 spin_lock(&head->lock);
4899 list_del_init(&poll->wait.entry);
4900 if (poll->wait.private)
4903 spin_unlock(&head->lock);
4907 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4909 struct io_ring_ctx *ctx = req->ctx;
4910 unsigned flags = IORING_CQE_F_MORE;
4912 if (!error && req->poll.canceled) {
4914 req->poll.events |= EPOLLONESHOT;
4917 error = mangle_poll(mask);
4918 if (req->poll.events & EPOLLONESHOT)
4920 if (!__io_cqring_fill_event(req, error, flags)) {
4921 io_poll_remove_waitqs(req);
4922 req->poll.done = true;
4925 io_commit_cqring(ctx);
4926 return !(flags & IORING_CQE_F_MORE);
4929 static void io_poll_task_func(struct callback_head *cb)
4931 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4932 struct io_ring_ctx *ctx = req->ctx;
4933 struct io_kiocb *nxt;
4935 if (io_poll_rewait(req, &req->poll)) {
4936 spin_unlock_irq(&ctx->completion_lock);
4940 post_ev = done = io_poll_complete(req, req->result, 0);
4942 hash_del(&req->hash_node);
4943 } else if (!(req->poll.events & EPOLLONESHOT)) {
4946 add_wait_queue(req->poll.head, &req->poll.wait);
4948 spin_unlock_irq(&ctx->completion_lock);
4951 io_cqring_ev_posted(ctx);
4953 nxt = io_put_req_find_next(req);
4955 __io_req_task_submit(nxt);
4960 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4961 int sync, void *key)
4963 struct io_kiocb *req = wait->private;
4964 struct io_poll_iocb *poll = io_poll_get_single(req);
4965 __poll_t mask = key_to_poll(key);
4967 /* for instances that support it check for an event match first: */
4968 if (mask && !(mask & poll->events))
4970 if (!(poll->events & EPOLLONESHOT))
4971 return poll->wait.func(&poll->wait, mode, sync, key);
4973 list_del_init(&wait->entry);
4975 if (poll && poll->head) {
4978 spin_lock(&poll->head->lock);
4979 done = list_empty(&poll->wait.entry);
4981 list_del_init(&poll->wait.entry);
4982 /* make sure double remove sees this as being gone */
4983 wait->private = NULL;
4984 spin_unlock(&poll->head->lock);
4986 /* use wait func handler, so it matches the rq type */
4987 poll->wait.func(&poll->wait, mode, sync, key);
4994 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4995 wait_queue_func_t wake_func)
4999 poll->canceled = false;
5000 poll->update_events = poll->update_user_data = false;
5001 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5002 /* mask in events that we always want/need */
5003 poll->events = events | IO_POLL_UNMASK;
5004 INIT_LIST_HEAD(&poll->wait.entry);
5005 init_waitqueue_func_entry(&poll->wait, wake_func);
5008 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5009 struct wait_queue_head *head,
5010 struct io_poll_iocb **poll_ptr)
5012 struct io_kiocb *req = pt->req;
5015 * If poll->head is already set, it's because the file being polled
5016 * uses multiple waitqueues for poll handling (eg one for read, one
5017 * for write). Setup a separate io_poll_iocb if this happens.
5019 if (unlikely(poll->head)) {
5020 struct io_poll_iocb *poll_one = poll;
5022 /* already have a 2nd entry, fail a third attempt */
5024 pt->error = -EINVAL;
5027 /* double add on the same waitqueue head, ignore */
5028 if (poll->head == head)
5030 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5032 pt->error = -ENOMEM;
5035 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5037 poll->wait.private = req;
5044 if (poll->events & EPOLLEXCLUSIVE)
5045 add_wait_queue_exclusive(head, &poll->wait);
5047 add_wait_queue(head, &poll->wait);
5050 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5051 struct poll_table_struct *p)
5053 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5054 struct async_poll *apoll = pt->req->apoll;
5056 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5059 static void io_async_task_func(struct callback_head *cb)
5061 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5062 struct async_poll *apoll = req->apoll;
5063 struct io_ring_ctx *ctx = req->ctx;
5065 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5067 if (io_poll_rewait(req, &apoll->poll)) {
5068 spin_unlock_irq(&ctx->completion_lock);
5072 /* If req is still hashed, it cannot have been canceled. Don't check. */
5073 if (hash_hashed(&req->hash_node))
5074 hash_del(&req->hash_node);
5076 io_poll_remove_double(req);
5077 spin_unlock_irq(&ctx->completion_lock);
5079 if (!READ_ONCE(apoll->poll.canceled))
5080 __io_req_task_submit(req);
5082 io_req_complete_failed(req, -ECANCELED);
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 && (poll->events & EPOLLONESHOT)) || 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, rw))
5173 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5174 if (unlikely(!apoll))
5176 apoll->double_poll = NULL;
5178 req->flags |= REQ_F_POLLED;
5181 mask = EPOLLONESHOT;
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, bool do_cancel)
5214 bool do_complete = false;
5218 spin_lock(&poll->head->lock);
5220 WRITE_ONCE(poll->canceled, true);
5221 if (!list_empty(&poll->wait.entry)) {
5222 list_del_init(&poll->wait.entry);
5225 spin_unlock(&poll->head->lock);
5226 hash_del(&req->hash_node);
5230 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5234 io_poll_remove_double(req);
5236 if (req->opcode == IORING_OP_POLL_ADD) {
5237 do_complete = __io_poll_remove_one(req, &req->poll, true);
5239 struct async_poll *apoll = req->apoll;
5241 /* non-poll requests have submit ref still */
5242 do_complete = __io_poll_remove_one(req, &apoll->poll, true);
5245 kfree(apoll->double_poll);
5253 static bool io_poll_remove_one(struct io_kiocb *req)
5257 do_complete = io_poll_remove_waitqs(req);
5259 io_cqring_fill_event(req, -ECANCELED);
5260 io_commit_cqring(req->ctx);
5261 req_set_fail_links(req);
5262 io_put_req_deferred(req, 1);
5269 * Returns true if we found and killed one or more poll requests
5271 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5272 struct files_struct *files)
5274 struct hlist_node *tmp;
5275 struct io_kiocb *req;
5278 spin_lock_irq(&ctx->completion_lock);
5279 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5280 struct hlist_head *list;
5282 list = &ctx->cancel_hash[i];
5283 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5284 if (io_match_task(req, tsk, files))
5285 posted += io_poll_remove_one(req);
5288 spin_unlock_irq(&ctx->completion_lock);
5291 io_cqring_ev_posted(ctx);
5296 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr)
5298 struct hlist_head *list;
5299 struct io_kiocb *req;
5301 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5302 hlist_for_each_entry(req, list, hash_node) {
5303 if (sqe_addr != req->user_data)
5311 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5313 struct io_kiocb *req;
5315 req = io_poll_find(ctx, sqe_addr);
5318 if (io_poll_remove_one(req))
5324 static int io_poll_remove_prep(struct io_kiocb *req,
5325 const struct io_uring_sqe *sqe)
5327 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5329 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5333 req->poll_remove.addr = READ_ONCE(sqe->addr);
5338 * Find a running poll command that matches one specified in sqe->addr,
5339 * and remove it if found.
5341 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5343 struct io_ring_ctx *ctx = req->ctx;
5346 spin_lock_irq(&ctx->completion_lock);
5347 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5348 spin_unlock_irq(&ctx->completion_lock);
5351 req_set_fail_links(req);
5352 io_req_complete(req, ret);
5356 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5359 struct io_kiocb *req = wait->private;
5360 struct io_poll_iocb *poll = &req->poll;
5362 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5365 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5366 struct poll_table_struct *p)
5368 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5370 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5373 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5375 struct io_poll_iocb *poll = &req->poll;
5378 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5380 if (sqe->ioprio || sqe->buf_index)
5382 flags = READ_ONCE(sqe->len);
5383 if (flags & ~(IORING_POLL_ADD_MULTI | IORING_POLL_UPDATE_EVENTS |
5384 IORING_POLL_UPDATE_USER_DATA))
5386 events = READ_ONCE(sqe->poll32_events);
5388 events = swahw32(events);
5390 if (!(flags & IORING_POLL_ADD_MULTI))
5391 events |= EPOLLONESHOT;
5392 poll->update_events = poll->update_user_data = false;
5393 if (flags & IORING_POLL_UPDATE_EVENTS) {
5394 poll->update_events = true;
5395 poll->old_user_data = READ_ONCE(sqe->addr);
5397 if (flags & IORING_POLL_UPDATE_USER_DATA) {
5398 poll->update_user_data = true;
5399 poll->new_user_data = READ_ONCE(sqe->off);
5401 if (!(poll->update_events || poll->update_user_data) &&
5402 (sqe->off || sqe->addr))
5404 poll->events = demangle_poll(events) |
5405 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5409 static int __io_poll_add(struct io_kiocb *req)
5411 struct io_poll_iocb *poll = &req->poll;
5412 struct io_ring_ctx *ctx = req->ctx;
5413 struct io_poll_table ipt;
5416 ipt.pt._qproc = io_poll_queue_proc;
5418 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5421 if (mask) { /* no async, we'd stolen it */
5423 io_poll_complete(req, mask, 0);
5425 spin_unlock_irq(&ctx->completion_lock);
5428 io_cqring_ev_posted(ctx);
5429 if (poll->events & EPOLLONESHOT)
5435 static int io_poll_update(struct io_kiocb *req)
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct io_kiocb *preq;
5441 spin_lock_irq(&ctx->completion_lock);
5442 preq = io_poll_find(ctx, req->poll.old_user_data);
5446 } else if (preq->opcode != IORING_OP_POLL_ADD) {
5447 /* don't allow internal poll updates */
5451 if (!__io_poll_remove_one(preq, &preq->poll, false)) {
5452 if (preq->poll.events & EPOLLONESHOT) {
5457 /* we now have a detached poll request. reissue. */
5460 spin_unlock_irq(&ctx->completion_lock);
5462 req_set_fail_links(req);
5463 io_req_complete(req, ret);
5466 /* only mask one event flags, keep behavior flags */
5467 if (req->poll.update_events) {
5468 preq->poll.events &= ~0xffff;
5469 preq->poll.events |= req->poll.events & 0xffff;
5470 preq->poll.events |= IO_POLL_UNMASK;
5472 if (req->poll.update_user_data)
5473 preq->user_data = req->poll.new_user_data;
5475 /* complete update request, we're done with it */
5476 io_req_complete(req, ret);
5478 ret = __io_poll_add(preq);
5480 req_set_fail_links(preq);
5481 io_req_complete(preq, ret);
5486 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5488 if (!req->poll.update_events && !req->poll.update_user_data)
5489 return __io_poll_add(req);
5490 return io_poll_update(req);
5493 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5495 struct io_timeout_data *data = container_of(timer,
5496 struct io_timeout_data, timer);
5497 struct io_kiocb *req = data->req;
5498 struct io_ring_ctx *ctx = req->ctx;
5499 unsigned long flags;
5501 spin_lock_irqsave(&ctx->completion_lock, flags);
5502 list_del_init(&req->timeout.list);
5503 atomic_set(&req->ctx->cq_timeouts,
5504 atomic_read(&req->ctx->cq_timeouts) + 1);
5506 io_cqring_fill_event(req, -ETIME);
5507 io_commit_cqring(ctx);
5508 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5510 io_cqring_ev_posted(ctx);
5511 req_set_fail_links(req);
5513 return HRTIMER_NORESTART;
5516 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5519 struct io_timeout_data *io;
5520 struct io_kiocb *req;
5523 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5524 if (user_data == req->user_data) {
5531 return ERR_PTR(ret);
5533 io = req->async_data;
5534 ret = hrtimer_try_to_cancel(&io->timer);
5536 return ERR_PTR(-EALREADY);
5537 list_del_init(&req->timeout.list);
5541 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5543 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5546 return PTR_ERR(req);
5548 req_set_fail_links(req);
5549 io_cqring_fill_event(req, -ECANCELED);
5550 io_put_req_deferred(req, 1);
5554 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5555 struct timespec64 *ts, enum hrtimer_mode mode)
5557 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5558 struct io_timeout_data *data;
5561 return PTR_ERR(req);
5563 req->timeout.off = 0; /* noseq */
5564 data = req->async_data;
5565 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5566 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5567 data->timer.function = io_timeout_fn;
5568 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5572 static int io_timeout_remove_prep(struct io_kiocb *req,
5573 const struct io_uring_sqe *sqe)
5575 struct io_timeout_rem *tr = &req->timeout_rem;
5577 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5579 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5581 if (sqe->ioprio || sqe->buf_index || sqe->len)
5584 tr->addr = READ_ONCE(sqe->addr);
5585 tr->flags = READ_ONCE(sqe->timeout_flags);
5586 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5587 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5589 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5591 } else if (tr->flags) {
5592 /* timeout removal doesn't support flags */
5599 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5601 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5606 * Remove or update an existing timeout command
5608 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5610 struct io_timeout_rem *tr = &req->timeout_rem;
5611 struct io_ring_ctx *ctx = req->ctx;
5614 spin_lock_irq(&ctx->completion_lock);
5615 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5616 ret = io_timeout_cancel(ctx, tr->addr);
5618 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5619 io_translate_timeout_mode(tr->flags));
5621 io_cqring_fill_event(req, ret);
5622 io_commit_cqring(ctx);
5623 spin_unlock_irq(&ctx->completion_lock);
5624 io_cqring_ev_posted(ctx);
5626 req_set_fail_links(req);
5631 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5632 bool is_timeout_link)
5634 struct io_timeout_data *data;
5636 u32 off = READ_ONCE(sqe->off);
5638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5640 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5642 if (off && is_timeout_link)
5644 flags = READ_ONCE(sqe->timeout_flags);
5645 if (flags & ~IORING_TIMEOUT_ABS)
5648 req->timeout.off = off;
5650 if (!req->async_data && io_alloc_async_data(req))
5653 data = req->async_data;
5656 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5659 data->mode = io_translate_timeout_mode(flags);
5660 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5661 if (is_timeout_link)
5662 io_req_track_inflight(req);
5666 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5668 struct io_ring_ctx *ctx = req->ctx;
5669 struct io_timeout_data *data = req->async_data;
5670 struct list_head *entry;
5671 u32 tail, off = req->timeout.off;
5673 spin_lock_irq(&ctx->completion_lock);
5676 * sqe->off holds how many events that need to occur for this
5677 * timeout event to be satisfied. If it isn't set, then this is
5678 * a pure timeout request, sequence isn't used.
5680 if (io_is_timeout_noseq(req)) {
5681 entry = ctx->timeout_list.prev;
5685 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5686 req->timeout.target_seq = tail + off;
5688 /* Update the last seq here in case io_flush_timeouts() hasn't.
5689 * This is safe because ->completion_lock is held, and submissions
5690 * and completions are never mixed in the same ->completion_lock section.
5692 ctx->cq_last_tm_flush = tail;
5695 * Insertion sort, ensuring the first entry in the list is always
5696 * the one we need first.
5698 list_for_each_prev(entry, &ctx->timeout_list) {
5699 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5702 if (io_is_timeout_noseq(nxt))
5704 /* nxt.seq is behind @tail, otherwise would've been completed */
5705 if (off >= nxt->timeout.target_seq - tail)
5709 list_add(&req->timeout.list, entry);
5710 data->timer.function = io_timeout_fn;
5711 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5712 spin_unlock_irq(&ctx->completion_lock);
5716 struct io_cancel_data {
5717 struct io_ring_ctx *ctx;
5721 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5724 struct io_cancel_data *cd = data;
5726 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5729 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5730 struct io_ring_ctx *ctx)
5732 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5733 enum io_wq_cancel cancel_ret;
5736 if (!tctx || !tctx->io_wq)
5739 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5740 switch (cancel_ret) {
5741 case IO_WQ_CANCEL_OK:
5744 case IO_WQ_CANCEL_RUNNING:
5747 case IO_WQ_CANCEL_NOTFOUND:
5755 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5756 struct io_kiocb *req, __u64 sqe_addr,
5759 unsigned long flags;
5762 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5763 if (ret != -ENOENT) {
5764 spin_lock_irqsave(&ctx->completion_lock, flags);
5768 spin_lock_irqsave(&ctx->completion_lock, flags);
5769 ret = io_timeout_cancel(ctx, sqe_addr);
5772 ret = io_poll_cancel(ctx, sqe_addr);
5776 io_cqring_fill_event(req, ret);
5777 io_commit_cqring(ctx);
5778 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5779 io_cqring_ev_posted(ctx);
5782 req_set_fail_links(req);
5786 static int io_async_cancel_prep(struct io_kiocb *req,
5787 const struct io_uring_sqe *sqe)
5789 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5791 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5793 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5796 req->cancel.addr = READ_ONCE(sqe->addr);
5800 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5802 struct io_ring_ctx *ctx = req->ctx;
5803 u64 sqe_addr = req->cancel.addr;
5804 struct io_tctx_node *node;
5807 /* tasks should wait for their io-wq threads, so safe w/o sync */
5808 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5809 spin_lock_irq(&ctx->completion_lock);
5812 ret = io_timeout_cancel(ctx, sqe_addr);
5815 ret = io_poll_cancel(ctx, sqe_addr);
5818 spin_unlock_irq(&ctx->completion_lock);
5820 /* slow path, try all io-wq's */
5821 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5823 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5824 struct io_uring_task *tctx = node->task->io_uring;
5826 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5830 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5832 spin_lock_irq(&ctx->completion_lock);
5834 io_cqring_fill_event(req, ret);
5835 io_commit_cqring(ctx);
5836 spin_unlock_irq(&ctx->completion_lock);
5837 io_cqring_ev_posted(ctx);
5840 req_set_fail_links(req);
5845 static int io_rsrc_update_prep(struct io_kiocb *req,
5846 const struct io_uring_sqe *sqe)
5848 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5850 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5852 if (sqe->ioprio || sqe->rw_flags)
5855 req->rsrc_update.offset = READ_ONCE(sqe->off);
5856 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5857 if (!req->rsrc_update.nr_args)
5859 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5863 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5865 struct io_ring_ctx *ctx = req->ctx;
5866 struct io_uring_rsrc_update up;
5869 if (issue_flags & IO_URING_F_NONBLOCK)
5872 up.offset = req->rsrc_update.offset;
5873 up.data = req->rsrc_update.arg;
5875 mutex_lock(&ctx->uring_lock);
5876 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5877 mutex_unlock(&ctx->uring_lock);
5880 req_set_fail_links(req);
5881 __io_req_complete(req, issue_flags, ret, 0);
5885 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5887 switch (req->opcode) {
5890 case IORING_OP_READV:
5891 case IORING_OP_READ_FIXED:
5892 case IORING_OP_READ:
5893 return io_read_prep(req, sqe);
5894 case IORING_OP_WRITEV:
5895 case IORING_OP_WRITE_FIXED:
5896 case IORING_OP_WRITE:
5897 return io_write_prep(req, sqe);
5898 case IORING_OP_POLL_ADD:
5899 return io_poll_add_prep(req, sqe);
5900 case IORING_OP_POLL_REMOVE:
5901 return io_poll_remove_prep(req, sqe);
5902 case IORING_OP_FSYNC:
5903 return io_fsync_prep(req, sqe);
5904 case IORING_OP_SYNC_FILE_RANGE:
5905 return io_sfr_prep(req, sqe);
5906 case IORING_OP_SENDMSG:
5907 case IORING_OP_SEND:
5908 return io_sendmsg_prep(req, sqe);
5909 case IORING_OP_RECVMSG:
5910 case IORING_OP_RECV:
5911 return io_recvmsg_prep(req, sqe);
5912 case IORING_OP_CONNECT:
5913 return io_connect_prep(req, sqe);
5914 case IORING_OP_TIMEOUT:
5915 return io_timeout_prep(req, sqe, false);
5916 case IORING_OP_TIMEOUT_REMOVE:
5917 return io_timeout_remove_prep(req, sqe);
5918 case IORING_OP_ASYNC_CANCEL:
5919 return io_async_cancel_prep(req, sqe);
5920 case IORING_OP_LINK_TIMEOUT:
5921 return io_timeout_prep(req, sqe, true);
5922 case IORING_OP_ACCEPT:
5923 return io_accept_prep(req, sqe);
5924 case IORING_OP_FALLOCATE:
5925 return io_fallocate_prep(req, sqe);
5926 case IORING_OP_OPENAT:
5927 return io_openat_prep(req, sqe);
5928 case IORING_OP_CLOSE:
5929 return io_close_prep(req, sqe);
5930 case IORING_OP_FILES_UPDATE:
5931 return io_rsrc_update_prep(req, sqe);
5932 case IORING_OP_STATX:
5933 return io_statx_prep(req, sqe);
5934 case IORING_OP_FADVISE:
5935 return io_fadvise_prep(req, sqe);
5936 case IORING_OP_MADVISE:
5937 return io_madvise_prep(req, sqe);
5938 case IORING_OP_OPENAT2:
5939 return io_openat2_prep(req, sqe);
5940 case IORING_OP_EPOLL_CTL:
5941 return io_epoll_ctl_prep(req, sqe);
5942 case IORING_OP_SPLICE:
5943 return io_splice_prep(req, sqe);
5944 case IORING_OP_PROVIDE_BUFFERS:
5945 return io_provide_buffers_prep(req, sqe);
5946 case IORING_OP_REMOVE_BUFFERS:
5947 return io_remove_buffers_prep(req, sqe);
5949 return io_tee_prep(req, sqe);
5950 case IORING_OP_SHUTDOWN:
5951 return io_shutdown_prep(req, sqe);
5952 case IORING_OP_RENAMEAT:
5953 return io_renameat_prep(req, sqe);
5954 case IORING_OP_UNLINKAT:
5955 return io_unlinkat_prep(req, sqe);
5958 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5963 static int io_req_prep_async(struct io_kiocb *req)
5965 if (!io_op_defs[req->opcode].needs_async_setup)
5967 if (WARN_ON_ONCE(req->async_data))
5969 if (io_alloc_async_data(req))
5972 switch (req->opcode) {
5973 case IORING_OP_READV:
5974 return io_rw_prep_async(req, READ);
5975 case IORING_OP_WRITEV:
5976 return io_rw_prep_async(req, WRITE);
5977 case IORING_OP_SENDMSG:
5978 return io_sendmsg_prep_async(req);
5979 case IORING_OP_RECVMSG:
5980 return io_recvmsg_prep_async(req);
5981 case IORING_OP_CONNECT:
5982 return io_connect_prep_async(req);
5984 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5989 static u32 io_get_sequence(struct io_kiocb *req)
5991 struct io_kiocb *pos;
5992 struct io_ring_ctx *ctx = req->ctx;
5993 u32 total_submitted, nr_reqs = 0;
5995 io_for_each_link(pos, req)
5998 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5999 return total_submitted - nr_reqs;
6002 static int io_req_defer(struct io_kiocb *req)
6004 struct io_ring_ctx *ctx = req->ctx;
6005 struct io_defer_entry *de;
6009 /* Still need defer if there is pending req in defer list. */
6010 if (likely(list_empty_careful(&ctx->defer_list) &&
6011 !(req->flags & REQ_F_IO_DRAIN)))
6014 seq = io_get_sequence(req);
6015 /* Still a chance to pass the sequence check */
6016 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6019 ret = io_req_prep_async(req);
6022 io_prep_async_link(req);
6023 de = kmalloc(sizeof(*de), GFP_KERNEL);
6027 spin_lock_irq(&ctx->completion_lock);
6028 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6029 spin_unlock_irq(&ctx->completion_lock);
6031 io_queue_async_work(req);
6032 return -EIOCBQUEUED;
6035 trace_io_uring_defer(ctx, req, req->user_data);
6038 list_add_tail(&de->list, &ctx->defer_list);
6039 spin_unlock_irq(&ctx->completion_lock);
6040 return -EIOCBQUEUED;
6043 static void io_clean_op(struct io_kiocb *req)
6045 if (req->flags & REQ_F_BUFFER_SELECTED) {
6046 switch (req->opcode) {
6047 case IORING_OP_READV:
6048 case IORING_OP_READ_FIXED:
6049 case IORING_OP_READ:
6050 kfree((void *)(unsigned long)req->rw.addr);
6052 case IORING_OP_RECVMSG:
6053 case IORING_OP_RECV:
6054 kfree(req->sr_msg.kbuf);
6057 req->flags &= ~REQ_F_BUFFER_SELECTED;
6060 if (req->flags & REQ_F_NEED_CLEANUP) {
6061 switch (req->opcode) {
6062 case IORING_OP_READV:
6063 case IORING_OP_READ_FIXED:
6064 case IORING_OP_READ:
6065 case IORING_OP_WRITEV:
6066 case IORING_OP_WRITE_FIXED:
6067 case IORING_OP_WRITE: {
6068 struct io_async_rw *io = req->async_data;
6070 kfree(io->free_iovec);
6073 case IORING_OP_RECVMSG:
6074 case IORING_OP_SENDMSG: {
6075 struct io_async_msghdr *io = req->async_data;
6077 kfree(io->free_iov);
6080 case IORING_OP_SPLICE:
6082 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6083 io_put_file(req->splice.file_in);
6085 case IORING_OP_OPENAT:
6086 case IORING_OP_OPENAT2:
6087 if (req->open.filename)
6088 putname(req->open.filename);
6090 case IORING_OP_RENAMEAT:
6091 putname(req->rename.oldpath);
6092 putname(req->rename.newpath);
6094 case IORING_OP_UNLINKAT:
6095 putname(req->unlink.filename);
6098 req->flags &= ~REQ_F_NEED_CLEANUP;
6102 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6104 struct io_ring_ctx *ctx = req->ctx;
6105 const struct cred *creds = NULL;
6108 if (req->work.creds && req->work.creds != current_cred())
6109 creds = override_creds(req->work.creds);
6111 switch (req->opcode) {
6113 ret = io_nop(req, issue_flags);
6115 case IORING_OP_READV:
6116 case IORING_OP_READ_FIXED:
6117 case IORING_OP_READ:
6118 ret = io_read(req, issue_flags);
6120 case IORING_OP_WRITEV:
6121 case IORING_OP_WRITE_FIXED:
6122 case IORING_OP_WRITE:
6123 ret = io_write(req, issue_flags);
6125 case IORING_OP_FSYNC:
6126 ret = io_fsync(req, issue_flags);
6128 case IORING_OP_POLL_ADD:
6129 ret = io_poll_add(req, issue_flags);
6131 case IORING_OP_POLL_REMOVE:
6132 ret = io_poll_remove(req, issue_flags);
6134 case IORING_OP_SYNC_FILE_RANGE:
6135 ret = io_sync_file_range(req, issue_flags);
6137 case IORING_OP_SENDMSG:
6138 ret = io_sendmsg(req, issue_flags);
6140 case IORING_OP_SEND:
6141 ret = io_send(req, issue_flags);
6143 case IORING_OP_RECVMSG:
6144 ret = io_recvmsg(req, issue_flags);
6146 case IORING_OP_RECV:
6147 ret = io_recv(req, issue_flags);
6149 case IORING_OP_TIMEOUT:
6150 ret = io_timeout(req, issue_flags);
6152 case IORING_OP_TIMEOUT_REMOVE:
6153 ret = io_timeout_remove(req, issue_flags);
6155 case IORING_OP_ACCEPT:
6156 ret = io_accept(req, issue_flags);
6158 case IORING_OP_CONNECT:
6159 ret = io_connect(req, issue_flags);
6161 case IORING_OP_ASYNC_CANCEL:
6162 ret = io_async_cancel(req, issue_flags);
6164 case IORING_OP_FALLOCATE:
6165 ret = io_fallocate(req, issue_flags);
6167 case IORING_OP_OPENAT:
6168 ret = io_openat(req, issue_flags);
6170 case IORING_OP_CLOSE:
6171 ret = io_close(req, issue_flags);
6173 case IORING_OP_FILES_UPDATE:
6174 ret = io_files_update(req, issue_flags);
6176 case IORING_OP_STATX:
6177 ret = io_statx(req, issue_flags);
6179 case IORING_OP_FADVISE:
6180 ret = io_fadvise(req, issue_flags);
6182 case IORING_OP_MADVISE:
6183 ret = io_madvise(req, issue_flags);
6185 case IORING_OP_OPENAT2:
6186 ret = io_openat2(req, issue_flags);
6188 case IORING_OP_EPOLL_CTL:
6189 ret = io_epoll_ctl(req, issue_flags);
6191 case IORING_OP_SPLICE:
6192 ret = io_splice(req, issue_flags);
6194 case IORING_OP_PROVIDE_BUFFERS:
6195 ret = io_provide_buffers(req, issue_flags);
6197 case IORING_OP_REMOVE_BUFFERS:
6198 ret = io_remove_buffers(req, issue_flags);
6201 ret = io_tee(req, issue_flags);
6203 case IORING_OP_SHUTDOWN:
6204 ret = io_shutdown(req, issue_flags);
6206 case IORING_OP_RENAMEAT:
6207 ret = io_renameat(req, issue_flags);
6209 case IORING_OP_UNLINKAT:
6210 ret = io_unlinkat(req, issue_flags);
6218 revert_creds(creds);
6223 /* If the op doesn't have a file, we're not polling for it */
6224 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6225 const bool in_async = io_wq_current_is_worker();
6227 /* workqueue context doesn't hold uring_lock, grab it now */
6229 mutex_lock(&ctx->uring_lock);
6231 io_iopoll_req_issued(req, in_async);
6234 mutex_unlock(&ctx->uring_lock);
6240 static void io_wq_submit_work(struct io_wq_work *work)
6242 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6243 struct io_kiocb *timeout;
6246 timeout = io_prep_linked_timeout(req);
6248 io_queue_linked_timeout(timeout);
6250 if (work->flags & IO_WQ_WORK_CANCEL)
6255 ret = io_issue_sqe(req, 0);
6257 * We can get EAGAIN for polled IO even though we're
6258 * forcing a sync submission from here, since we can't
6259 * wait for request slots on the block side.
6267 /* avoid locking problems by failing it from a clean context */
6269 /* io-wq is going to take one down */
6271 io_req_task_queue_fail(req, ret);
6275 #define FFS_ASYNC_READ 0x1UL
6276 #define FFS_ASYNC_WRITE 0x2UL
6278 #define FFS_ISREG 0x4UL
6280 #define FFS_ISREG 0x0UL
6282 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6284 static inline struct file **io_fixed_file_slot(struct io_rsrc_data *file_data,
6287 struct fixed_rsrc_table *table;
6289 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6290 return &table->files[i & IORING_FILE_TABLE_MASK];
6293 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6296 struct file **file_slot = io_fixed_file_slot(ctx->file_data, index);
6298 return (struct file *) ((unsigned long) *file_slot & FFS_MASK);
6301 static struct file *io_file_get(struct io_submit_state *state,
6302 struct io_kiocb *req, int fd, bool fixed)
6304 struct io_ring_ctx *ctx = req->ctx;
6308 unsigned long file_ptr;
6310 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6312 fd = array_index_nospec(fd, ctx->nr_user_files);
6313 file_ptr = (unsigned long) *io_fixed_file_slot(ctx->file_data, fd);
6314 file = (struct file *) (file_ptr & FFS_MASK);
6315 file_ptr &= ~FFS_MASK;
6316 /* mask in overlapping REQ_F and FFS bits */
6317 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6318 io_req_set_rsrc_node(req);
6320 trace_io_uring_file_get(ctx, fd);
6321 file = __io_file_get(state, fd);
6323 /* we don't allow fixed io_uring files */
6324 if (file && unlikely(file->f_op == &io_uring_fops))
6325 io_req_track_inflight(req);
6331 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6333 struct io_timeout_data *data = container_of(timer,
6334 struct io_timeout_data, timer);
6335 struct io_kiocb *prev, *req = data->req;
6336 struct io_ring_ctx *ctx = req->ctx;
6337 unsigned long flags;
6339 spin_lock_irqsave(&ctx->completion_lock, flags);
6340 prev = req->timeout.head;
6341 req->timeout.head = NULL;
6344 * We don't expect the list to be empty, that will only happen if we
6345 * race with the completion of the linked work.
6347 if (prev && req_ref_inc_not_zero(prev))
6348 io_remove_next_linked(prev);
6351 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6354 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6355 io_put_req_deferred(prev, 1);
6357 io_req_complete_post(req, -ETIME, 0);
6358 io_put_req_deferred(req, 1);
6360 return HRTIMER_NORESTART;
6363 static void io_queue_linked_timeout(struct io_kiocb *req)
6365 struct io_ring_ctx *ctx = req->ctx;
6367 spin_lock_irq(&ctx->completion_lock);
6369 * If the back reference is NULL, then our linked request finished
6370 * before we got a chance to setup the timer
6372 if (req->timeout.head) {
6373 struct io_timeout_data *data = req->async_data;
6375 data->timer.function = io_link_timeout_fn;
6376 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6379 spin_unlock_irq(&ctx->completion_lock);
6380 /* drop submission reference */
6384 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6386 struct io_kiocb *nxt = req->link;
6388 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6389 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6392 nxt->timeout.head = req;
6393 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6394 req->flags |= REQ_F_LINK_TIMEOUT;
6398 static void __io_queue_sqe(struct io_kiocb *req)
6400 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6403 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6406 * We async punt it if the file wasn't marked NOWAIT, or if the file
6407 * doesn't support non-blocking read/write attempts
6410 /* drop submission reference */
6411 if (req->flags & REQ_F_COMPLETE_INLINE) {
6412 struct io_ring_ctx *ctx = req->ctx;
6413 struct io_comp_state *cs = &ctx->submit_state.comp;
6415 cs->reqs[cs->nr++] = req;
6416 if (cs->nr == ARRAY_SIZE(cs->reqs))
6417 io_submit_flush_completions(cs, ctx);
6421 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6422 if (!io_arm_poll_handler(req)) {
6424 * Queued up for async execution, worker will release
6425 * submit reference when the iocb is actually submitted.
6427 io_queue_async_work(req);
6430 io_req_complete_failed(req, ret);
6433 io_queue_linked_timeout(linked_timeout);
6436 static void io_queue_sqe(struct io_kiocb *req)
6440 ret = io_req_defer(req);
6442 if (ret != -EIOCBQUEUED) {
6444 io_req_complete_failed(req, ret);
6446 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6447 ret = io_req_prep_async(req);
6450 io_queue_async_work(req);
6452 __io_queue_sqe(req);
6457 * Check SQE restrictions (opcode and flags).
6459 * Returns 'true' if SQE is allowed, 'false' otherwise.
6461 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6462 struct io_kiocb *req,
6463 unsigned int sqe_flags)
6465 if (!ctx->restricted)
6468 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6471 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6472 ctx->restrictions.sqe_flags_required)
6475 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6476 ctx->restrictions.sqe_flags_required))
6482 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6483 const struct io_uring_sqe *sqe)
6485 struct io_submit_state *state;
6486 unsigned int sqe_flags;
6487 int personality, ret = 0;
6489 req->opcode = READ_ONCE(sqe->opcode);
6490 /* same numerical values with corresponding REQ_F_*, safe to copy */
6491 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6492 req->user_data = READ_ONCE(sqe->user_data);
6493 req->async_data = NULL;
6497 req->fixed_rsrc_refs = NULL;
6498 /* one is dropped after submission, the other at completion */
6499 atomic_set(&req->refs, 2);
6500 req->task = current;
6502 req->work.creds = NULL;
6504 /* enforce forwards compatibility on users */
6505 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6510 if (unlikely(req->opcode >= IORING_OP_LAST))
6513 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6516 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6517 !io_op_defs[req->opcode].buffer_select)
6520 personality = READ_ONCE(sqe->personality);
6522 req->work.creds = xa_load(&ctx->personalities, personality);
6523 if (!req->work.creds)
6525 get_cred(req->work.creds);
6527 state = &ctx->submit_state;
6530 * Plug now if we have more than 1 IO left after this, and the target
6531 * is potentially a read/write to block based storage.
6533 if (!state->plug_started && state->ios_left > 1 &&
6534 io_op_defs[req->opcode].plug) {
6535 blk_start_plug(&state->plug);
6536 state->plug_started = true;
6539 if (io_op_defs[req->opcode].needs_file) {
6540 bool fixed = req->flags & REQ_F_FIXED_FILE;
6542 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6543 if (unlikely(!req->file))
6551 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6552 const struct io_uring_sqe *sqe)
6554 struct io_submit_link *link = &ctx->submit_state.link;
6557 ret = io_init_req(ctx, req, sqe);
6558 if (unlikely(ret)) {
6561 /* fail even hard links since we don't submit */
6562 link->head->flags |= REQ_F_FAIL_LINK;
6563 io_req_complete_failed(link->head, -ECANCELED);
6566 io_req_complete_failed(req, ret);
6569 ret = io_req_prep(req, sqe);
6573 /* don't need @sqe from now on */
6574 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6575 true, ctx->flags & IORING_SETUP_SQPOLL);
6578 * If we already have a head request, queue this one for async
6579 * submittal once the head completes. If we don't have a head but
6580 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6581 * submitted sync once the chain is complete. If none of those
6582 * conditions are true (normal request), then just queue it.
6585 struct io_kiocb *head = link->head;
6588 * Taking sequential execution of a link, draining both sides
6589 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6590 * requests in the link. So, it drains the head and the
6591 * next after the link request. The last one is done via
6592 * drain_next flag to persist the effect across calls.
6594 if (req->flags & REQ_F_IO_DRAIN) {
6595 head->flags |= REQ_F_IO_DRAIN;
6596 ctx->drain_next = 1;
6598 ret = io_req_prep_async(req);
6601 trace_io_uring_link(ctx, req, head);
6602 link->last->link = req;
6605 /* last request of a link, enqueue the link */
6606 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6611 if (unlikely(ctx->drain_next)) {
6612 req->flags |= REQ_F_IO_DRAIN;
6613 ctx->drain_next = 0;
6615 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6627 * Batched submission is done, ensure local IO is flushed out.
6629 static void io_submit_state_end(struct io_submit_state *state,
6630 struct io_ring_ctx *ctx)
6632 if (state->link.head)
6633 io_queue_sqe(state->link.head);
6635 io_submit_flush_completions(&state->comp, ctx);
6636 if (state->plug_started)
6637 blk_finish_plug(&state->plug);
6638 io_state_file_put(state);
6642 * Start submission side cache.
6644 static void io_submit_state_start(struct io_submit_state *state,
6645 unsigned int max_ios)
6647 state->plug_started = false;
6648 state->ios_left = max_ios;
6649 /* set only head, no need to init link_last in advance */
6650 state->link.head = NULL;
6653 static void io_commit_sqring(struct io_ring_ctx *ctx)
6655 struct io_rings *rings = ctx->rings;
6658 * Ensure any loads from the SQEs are done at this point,
6659 * since once we write the new head, the application could
6660 * write new data to them.
6662 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6666 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6667 * that is mapped by userspace. This means that care needs to be taken to
6668 * ensure that reads are stable, as we cannot rely on userspace always
6669 * being a good citizen. If members of the sqe are validated and then later
6670 * used, it's important that those reads are done through READ_ONCE() to
6671 * prevent a re-load down the line.
6673 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6675 u32 *sq_array = ctx->sq_array;
6679 * The cached sq head (or cq tail) serves two purposes:
6681 * 1) allows us to batch the cost of updating the user visible
6683 * 2) allows the kernel side to track the head on its own, even
6684 * though the application is the one updating it.
6686 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6687 if (likely(head < ctx->sq_entries))
6688 return &ctx->sq_sqes[head];
6690 /* drop invalid entries */
6691 ctx->cached_sq_dropped++;
6692 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6696 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6700 /* if we have a backlog and couldn't flush it all, return BUSY */
6701 if (test_bit(0, &ctx->sq_check_overflow)) {
6702 if (!__io_cqring_overflow_flush(ctx, false))
6706 /* make sure SQ entry isn't read before tail */
6707 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6709 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6712 percpu_counter_add(¤t->io_uring->inflight, nr);
6713 refcount_add(nr, ¤t->usage);
6714 io_submit_state_start(&ctx->submit_state, nr);
6716 while (submitted < nr) {
6717 const struct io_uring_sqe *sqe;
6718 struct io_kiocb *req;
6720 req = io_alloc_req(ctx);
6721 if (unlikely(!req)) {
6723 submitted = -EAGAIN;
6726 sqe = io_get_sqe(ctx);
6727 if (unlikely(!sqe)) {
6728 kmem_cache_free(req_cachep, req);
6731 /* will complete beyond this point, count as submitted */
6733 if (io_submit_sqe(ctx, req, sqe))
6737 if (unlikely(submitted != nr)) {
6738 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6739 struct io_uring_task *tctx = current->io_uring;
6740 int unused = nr - ref_used;
6742 percpu_ref_put_many(&ctx->refs, unused);
6743 percpu_counter_sub(&tctx->inflight, unused);
6744 put_task_struct_many(current, unused);
6747 io_submit_state_end(&ctx->submit_state, ctx);
6748 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6749 io_commit_sqring(ctx);
6754 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6756 /* Tell userspace we may need a wakeup call */
6757 spin_lock_irq(&ctx->completion_lock);
6758 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6759 spin_unlock_irq(&ctx->completion_lock);
6762 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6764 spin_lock_irq(&ctx->completion_lock);
6765 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6766 spin_unlock_irq(&ctx->completion_lock);
6769 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6771 unsigned int to_submit;
6774 to_submit = io_sqring_entries(ctx);
6775 /* if we're handling multiple rings, cap submit size for fairness */
6776 if (cap_entries && to_submit > 8)
6779 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6780 unsigned nr_events = 0;
6782 mutex_lock(&ctx->uring_lock);
6783 if (!list_empty(&ctx->iopoll_list))
6784 io_do_iopoll(ctx, &nr_events, 0);
6786 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6787 !(ctx->flags & IORING_SETUP_R_DISABLED))
6788 ret = io_submit_sqes(ctx, to_submit);
6789 mutex_unlock(&ctx->uring_lock);
6792 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6793 wake_up(&ctx->sqo_sq_wait);
6798 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6800 struct io_ring_ctx *ctx;
6801 unsigned sq_thread_idle = 0;
6803 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6804 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6805 sqd->sq_thread_idle = sq_thread_idle;
6808 static int io_sq_thread(void *data)
6810 struct io_sq_data *sqd = data;
6811 struct io_ring_ctx *ctx;
6812 unsigned long timeout = 0;
6813 char buf[TASK_COMM_LEN];
6816 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6817 set_task_comm(current, buf);
6818 current->pf_io_worker = NULL;
6820 if (sqd->sq_cpu != -1)
6821 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6823 set_cpus_allowed_ptr(current, cpu_online_mask);
6824 current->flags |= PF_NO_SETAFFINITY;
6826 mutex_lock(&sqd->lock);
6827 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6829 bool cap_entries, sqt_spin, needs_sched;
6831 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6832 signal_pending(current)) {
6833 bool did_sig = false;
6835 mutex_unlock(&sqd->lock);
6836 if (signal_pending(current)) {
6837 struct ksignal ksig;
6839 did_sig = get_signal(&ksig);
6842 mutex_lock(&sqd->lock);
6846 io_run_task_work_head(&sqd->park_task_work);
6847 timeout = jiffies + sqd->sq_thread_idle;
6851 cap_entries = !list_is_singular(&sqd->ctx_list);
6852 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6853 const struct cred *creds = NULL;
6855 if (ctx->sq_creds != current_cred())
6856 creds = override_creds(ctx->sq_creds);
6857 ret = __io_sq_thread(ctx, cap_entries);
6859 revert_creds(creds);
6860 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6864 if (sqt_spin || !time_after(jiffies, timeout)) {
6868 timeout = jiffies + sqd->sq_thread_idle;
6873 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6874 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6875 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6876 !list_empty_careful(&ctx->iopoll_list)) {
6877 needs_sched = false;
6880 if (io_sqring_entries(ctx)) {
6881 needs_sched = false;
6886 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6887 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6888 io_ring_set_wakeup_flag(ctx);
6890 mutex_unlock(&sqd->lock);
6892 mutex_lock(&sqd->lock);
6893 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6894 io_ring_clear_wakeup_flag(ctx);
6897 finish_wait(&sqd->wait, &wait);
6898 io_run_task_work_head(&sqd->park_task_work);
6899 timeout = jiffies + sqd->sq_thread_idle;
6902 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6903 io_uring_cancel_sqpoll(ctx);
6905 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6906 io_ring_set_wakeup_flag(ctx);
6907 mutex_unlock(&sqd->lock);
6910 io_run_task_work_head(&sqd->park_task_work);
6911 complete(&sqd->exited);
6915 struct io_wait_queue {
6916 struct wait_queue_entry wq;
6917 struct io_ring_ctx *ctx;
6919 unsigned nr_timeouts;
6922 static inline bool io_should_wake(struct io_wait_queue *iowq)
6924 struct io_ring_ctx *ctx = iowq->ctx;
6927 * Wake up if we have enough events, or if a timeout occurred since we
6928 * started waiting. For timeouts, we always want to return to userspace,
6929 * regardless of event count.
6931 return io_cqring_events(ctx) >= iowq->to_wait ||
6932 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6935 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6936 int wake_flags, void *key)
6938 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6942 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6943 * the task, and the next invocation will do it.
6945 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6946 return autoremove_wake_function(curr, mode, wake_flags, key);
6950 static int io_run_task_work_sig(void)
6952 if (io_run_task_work())
6954 if (!signal_pending(current))
6956 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6957 return -ERESTARTSYS;
6961 /* when returns >0, the caller should retry */
6962 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6963 struct io_wait_queue *iowq,
6964 signed long *timeout)
6968 /* make sure we run task_work before checking for signals */
6969 ret = io_run_task_work_sig();
6970 if (ret || io_should_wake(iowq))
6972 /* let the caller flush overflows, retry */
6973 if (test_bit(0, &ctx->cq_check_overflow))
6976 *timeout = schedule_timeout(*timeout);
6977 return !*timeout ? -ETIME : 1;
6981 * Wait until events become available, if we don't already have some. The
6982 * application must reap them itself, as they reside on the shared cq ring.
6984 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6985 const sigset_t __user *sig, size_t sigsz,
6986 struct __kernel_timespec __user *uts)
6988 struct io_wait_queue iowq = {
6991 .func = io_wake_function,
6992 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6995 .to_wait = min_events,
6997 struct io_rings *rings = ctx->rings;
6998 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7002 io_cqring_overflow_flush(ctx, false);
7003 if (io_cqring_events(ctx) >= min_events)
7005 if (!io_run_task_work())
7010 #ifdef CONFIG_COMPAT
7011 if (in_compat_syscall())
7012 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7016 ret = set_user_sigmask(sig, sigsz);
7023 struct timespec64 ts;
7025 if (get_timespec64(&ts, uts))
7027 timeout = timespec64_to_jiffies(&ts);
7030 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7031 trace_io_uring_cqring_wait(ctx, min_events);
7033 /* if we can't even flush overflow, don't wait for more */
7034 if (!io_cqring_overflow_flush(ctx, false)) {
7038 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7039 TASK_INTERRUPTIBLE);
7040 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7041 finish_wait(&ctx->wait, &iowq.wq);
7045 restore_saved_sigmask_unless(ret == -EINTR);
7047 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7050 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7052 #if defined(CONFIG_UNIX)
7053 if (ctx->ring_sock) {
7054 struct sock *sock = ctx->ring_sock->sk;
7055 struct sk_buff *skb;
7057 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7063 for (i = 0; i < ctx->nr_user_files; i++) {
7066 file = io_file_from_index(ctx, i);
7073 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7075 struct io_rsrc_data *data = container_of(ref, struct io_rsrc_data, refs);
7077 complete(&data->done);
7080 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7082 spin_lock_bh(&ctx->rsrc_ref_lock);
7085 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7087 spin_unlock_bh(&ctx->rsrc_ref_lock);
7090 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7092 percpu_ref_exit(&ref_node->refs);
7096 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7097 struct io_rsrc_data *data_to_kill)
7099 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7100 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7103 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7105 rsrc_node->rsrc_data = data_to_kill;
7106 io_rsrc_ref_lock(ctx);
7107 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7108 io_rsrc_ref_unlock(ctx);
7110 percpu_ref_get(&data_to_kill->refs);
7111 percpu_ref_kill(&rsrc_node->refs);
7112 ctx->rsrc_node = NULL;
7115 if (!ctx->rsrc_node) {
7116 ctx->rsrc_node = ctx->rsrc_backup_node;
7117 ctx->rsrc_backup_node = NULL;
7121 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7123 if (ctx->rsrc_backup_node)
7125 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7126 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7129 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7133 /* As we may drop ->uring_lock, other task may have started quiesce */
7137 data->quiesce = true;
7139 ret = io_rsrc_node_switch_start(ctx);
7142 io_rsrc_node_switch(ctx, data);
7144 percpu_ref_kill(&data->refs);
7145 flush_delayed_work(&ctx->rsrc_put_work);
7147 ret = wait_for_completion_interruptible(&data->done);
7151 percpu_ref_resurrect(&data->refs);
7152 reinit_completion(&data->done);
7154 mutex_unlock(&ctx->uring_lock);
7155 ret = io_run_task_work_sig();
7156 mutex_lock(&ctx->uring_lock);
7158 data->quiesce = false;
7163 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7164 rsrc_put_fn *do_put)
7166 struct io_rsrc_data *data;
7168 data = kzalloc(sizeof(*data), GFP_KERNEL);
7172 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7173 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7178 data->do_put = do_put;
7179 init_completion(&data->done);
7183 static void io_rsrc_data_free(struct io_rsrc_data *data)
7185 percpu_ref_exit(&data->refs);
7190 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7192 struct io_rsrc_data *data = ctx->file_data;
7193 unsigned nr_tables, i;
7198 ret = io_rsrc_ref_quiesce(data, ctx);
7202 __io_sqe_files_unregister(ctx);
7203 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7204 for (i = 0; i < nr_tables; i++)
7205 kfree(data->table[i].files);
7206 io_rsrc_data_free(data);
7207 ctx->file_data = NULL;
7208 ctx->nr_user_files = 0;
7212 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7213 __releases(&sqd->lock)
7215 WARN_ON_ONCE(sqd->thread == current);
7218 * Do the dance but not conditional clear_bit() because it'd race with
7219 * other threads incrementing park_pending and setting the bit.
7221 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7222 if (atomic_dec_return(&sqd->park_pending))
7223 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7224 mutex_unlock(&sqd->lock);
7227 static void io_sq_thread_park(struct io_sq_data *sqd)
7228 __acquires(&sqd->lock)
7230 WARN_ON_ONCE(sqd->thread == current);
7232 atomic_inc(&sqd->park_pending);
7233 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7234 mutex_lock(&sqd->lock);
7236 wake_up_process(sqd->thread);
7239 static void io_sq_thread_stop(struct io_sq_data *sqd)
7241 WARN_ON_ONCE(sqd->thread == current);
7243 mutex_lock(&sqd->lock);
7244 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7246 wake_up_process(sqd->thread);
7247 mutex_unlock(&sqd->lock);
7248 wait_for_completion(&sqd->exited);
7251 static void io_put_sq_data(struct io_sq_data *sqd)
7253 if (refcount_dec_and_test(&sqd->refs)) {
7254 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7256 io_sq_thread_stop(sqd);
7261 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7263 struct io_sq_data *sqd = ctx->sq_data;
7266 io_sq_thread_park(sqd);
7267 list_del_init(&ctx->sqd_list);
7268 io_sqd_update_thread_idle(sqd);
7269 io_sq_thread_unpark(sqd);
7271 io_put_sq_data(sqd);
7272 ctx->sq_data = NULL;
7274 put_cred(ctx->sq_creds);
7278 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7280 struct io_ring_ctx *ctx_attach;
7281 struct io_sq_data *sqd;
7284 f = fdget(p->wq_fd);
7286 return ERR_PTR(-ENXIO);
7287 if (f.file->f_op != &io_uring_fops) {
7289 return ERR_PTR(-EINVAL);
7292 ctx_attach = f.file->private_data;
7293 sqd = ctx_attach->sq_data;
7296 return ERR_PTR(-EINVAL);
7298 if (sqd->task_tgid != current->tgid) {
7300 return ERR_PTR(-EPERM);
7303 refcount_inc(&sqd->refs);
7308 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7311 struct io_sq_data *sqd;
7314 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7315 sqd = io_attach_sq_data(p);
7320 /* fall through for EPERM case, setup new sqd/task */
7321 if (PTR_ERR(sqd) != -EPERM)
7325 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7327 return ERR_PTR(-ENOMEM);
7329 atomic_set(&sqd->park_pending, 0);
7330 refcount_set(&sqd->refs, 1);
7331 INIT_LIST_HEAD(&sqd->ctx_list);
7332 mutex_init(&sqd->lock);
7333 init_waitqueue_head(&sqd->wait);
7334 init_completion(&sqd->exited);
7338 #if defined(CONFIG_UNIX)
7340 * Ensure the UNIX gc is aware of our file set, so we are certain that
7341 * the io_uring can be safely unregistered on process exit, even if we have
7342 * loops in the file referencing.
7344 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7346 struct sock *sk = ctx->ring_sock->sk;
7347 struct scm_fp_list *fpl;
7348 struct sk_buff *skb;
7351 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7355 skb = alloc_skb(0, GFP_KERNEL);
7364 fpl->user = get_uid(current_user());
7365 for (i = 0; i < nr; i++) {
7366 struct file *file = io_file_from_index(ctx, i + offset);
7370 fpl->fp[nr_files] = get_file(file);
7371 unix_inflight(fpl->user, fpl->fp[nr_files]);
7376 fpl->max = SCM_MAX_FD;
7377 fpl->count = nr_files;
7378 UNIXCB(skb).fp = fpl;
7379 skb->destructor = unix_destruct_scm;
7380 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7381 skb_queue_head(&sk->sk_receive_queue, skb);
7383 for (i = 0; i < nr_files; i++)
7394 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7395 * causes regular reference counting to break down. We rely on the UNIX
7396 * garbage collection to take care of this problem for us.
7398 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7400 unsigned left, total;
7404 left = ctx->nr_user_files;
7406 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7408 ret = __io_sqe_files_scm(ctx, this_files, total);
7412 total += this_files;
7418 while (total < ctx->nr_user_files) {
7419 struct file *file = io_file_from_index(ctx, total);
7429 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7435 static int io_sqe_alloc_file_tables(struct io_rsrc_data *file_data,
7436 unsigned nr_tables, unsigned nr_files)
7440 for (i = 0; i < nr_tables; i++) {
7441 struct fixed_rsrc_table *table = &file_data->table[i];
7442 unsigned this_files;
7444 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7445 table->files = kcalloc(this_files, sizeof(struct file *),
7449 nr_files -= this_files;
7455 for (i = 0; i < nr_tables; i++) {
7456 struct fixed_rsrc_table *table = &file_data->table[i];
7457 kfree(table->files);
7462 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7464 struct file *file = prsrc->file;
7465 #if defined(CONFIG_UNIX)
7466 struct sock *sock = ctx->ring_sock->sk;
7467 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7468 struct sk_buff *skb;
7471 __skb_queue_head_init(&list);
7474 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7475 * remove this entry and rearrange the file array.
7477 skb = skb_dequeue(head);
7479 struct scm_fp_list *fp;
7481 fp = UNIXCB(skb).fp;
7482 for (i = 0; i < fp->count; i++) {
7485 if (fp->fp[i] != file)
7488 unix_notinflight(fp->user, fp->fp[i]);
7489 left = fp->count - 1 - i;
7491 memmove(&fp->fp[i], &fp->fp[i + 1],
7492 left * sizeof(struct file *));
7499 __skb_queue_tail(&list, skb);
7509 __skb_queue_tail(&list, skb);
7511 skb = skb_dequeue(head);
7514 if (skb_peek(&list)) {
7515 spin_lock_irq(&head->lock);
7516 while ((skb = __skb_dequeue(&list)) != NULL)
7517 __skb_queue_tail(head, skb);
7518 spin_unlock_irq(&head->lock);
7525 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7527 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7528 struct io_ring_ctx *ctx = rsrc_data->ctx;
7529 struct io_rsrc_put *prsrc, *tmp;
7531 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7532 list_del(&prsrc->list);
7533 rsrc_data->do_put(ctx, prsrc);
7537 io_rsrc_node_destroy(ref_node);
7538 percpu_ref_put(&rsrc_data->refs);
7541 static void io_rsrc_put_work(struct work_struct *work)
7543 struct io_ring_ctx *ctx;
7544 struct llist_node *node;
7546 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7547 node = llist_del_all(&ctx->rsrc_put_llist);
7550 struct io_rsrc_node *ref_node;
7551 struct llist_node *next = node->next;
7553 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7554 __io_rsrc_put_work(ref_node);
7559 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7561 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7562 struct io_rsrc_data *data = node->rsrc_data;
7563 struct io_ring_ctx *ctx = data->ctx;
7564 bool first_add = false;
7567 io_rsrc_ref_lock(ctx);
7570 while (!list_empty(&ctx->rsrc_ref_list)) {
7571 node = list_first_entry(&ctx->rsrc_ref_list,
7572 struct io_rsrc_node, node);
7573 /* recycle ref nodes in order */
7576 list_del(&node->node);
7577 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7579 io_rsrc_ref_unlock(ctx);
7581 delay = percpu_ref_is_dying(&data->refs) ? 0 : HZ;
7582 if (first_add || !delay)
7583 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7586 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7588 struct io_rsrc_node *ref_node;
7590 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7594 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7599 INIT_LIST_HEAD(&ref_node->node);
7600 INIT_LIST_HEAD(&ref_node->rsrc_list);
7601 ref_node->done = false;
7605 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7608 __s32 __user *fds = (__s32 __user *) arg;
7609 unsigned nr_tables, i;
7612 struct io_rsrc_data *file_data;
7618 if (nr_args > IORING_MAX_FIXED_FILES)
7620 ret = io_rsrc_node_switch_start(ctx);
7624 file_data = io_rsrc_data_alloc(ctx, io_ring_file_put);
7627 ctx->file_data = file_data;
7630 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7631 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7633 if (!file_data->table)
7636 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7639 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7640 unsigned long file_ptr;
7642 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7646 /* allow sparse sets */
7656 * Don't allow io_uring instances to be registered. If UNIX
7657 * isn't enabled, then this causes a reference cycle and this
7658 * instance can never get freed. If UNIX is enabled we'll
7659 * handle it just fine, but there's still no point in allowing
7660 * a ring fd as it doesn't support regular read/write anyway.
7662 if (file->f_op == &io_uring_fops) {
7666 file_ptr = (unsigned long) file;
7667 if (__io_file_supports_async(file, READ))
7668 file_ptr |= FFS_ASYNC_READ;
7669 if (__io_file_supports_async(file, WRITE))
7670 file_ptr |= FFS_ASYNC_WRITE;
7671 if (S_ISREG(file_inode(file)->i_mode))
7672 file_ptr |= FFS_ISREG;
7673 *io_fixed_file_slot(file_data, i) = (struct file *) file_ptr;
7676 ret = io_sqe_files_scm(ctx);
7678 io_sqe_files_unregister(ctx);
7682 io_rsrc_node_switch(ctx, NULL);
7685 for (i = 0; i < ctx->nr_user_files; i++) {
7686 file = io_file_from_index(ctx, i);
7690 for (i = 0; i < nr_tables; i++)
7691 kfree(file_data->table[i].files);
7692 ctx->nr_user_files = 0;
7694 io_rsrc_data_free(ctx->file_data);
7695 ctx->file_data = NULL;
7699 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7702 #if defined(CONFIG_UNIX)
7703 struct sock *sock = ctx->ring_sock->sk;
7704 struct sk_buff_head *head = &sock->sk_receive_queue;
7705 struct sk_buff *skb;
7708 * See if we can merge this file into an existing skb SCM_RIGHTS
7709 * file set. If there's no room, fall back to allocating a new skb
7710 * and filling it in.
7712 spin_lock_irq(&head->lock);
7713 skb = skb_peek(head);
7715 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7717 if (fpl->count < SCM_MAX_FD) {
7718 __skb_unlink(skb, head);
7719 spin_unlock_irq(&head->lock);
7720 fpl->fp[fpl->count] = get_file(file);
7721 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7723 spin_lock_irq(&head->lock);
7724 __skb_queue_head(head, skb);
7729 spin_unlock_irq(&head->lock);
7736 return __io_sqe_files_scm(ctx, 1, index);
7742 static int io_queue_rsrc_removal(struct io_rsrc_data *data,
7743 struct io_rsrc_node *node, void *rsrc)
7745 struct io_rsrc_put *prsrc;
7747 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7752 list_add(&prsrc->list, &node->rsrc_list);
7756 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7757 struct io_uring_rsrc_update *up,
7760 struct io_rsrc_data *data = ctx->file_data;
7761 struct file *file, **file_slot;
7765 bool needs_switch = false;
7767 if (check_add_overflow(up->offset, nr_args, &done))
7769 if (done > ctx->nr_user_files)
7771 err = io_rsrc_node_switch_start(ctx);
7775 fds = u64_to_user_ptr(up->data);
7776 for (done = 0; done < nr_args; done++) {
7778 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7782 if (fd == IORING_REGISTER_FILES_SKIP)
7785 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7786 file_slot = io_fixed_file_slot(ctx->file_data, i);
7789 file = (struct file *) ((unsigned long) *file_slot & FFS_MASK);
7790 err = io_queue_rsrc_removal(data, ctx->rsrc_node, file);
7794 needs_switch = true;
7803 * Don't allow io_uring instances to be registered. If
7804 * UNIX isn't enabled, then this causes a reference
7805 * cycle and this instance can never get freed. If UNIX
7806 * is enabled we'll handle it just fine, but there's
7807 * still no point in allowing a ring fd as it doesn't
7808 * support regular read/write anyway.
7810 if (file->f_op == &io_uring_fops) {
7816 err = io_sqe_file_register(ctx, file, i);
7826 io_rsrc_node_switch(ctx, data);
7827 return done ? done : err;
7830 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7833 struct io_uring_rsrc_update up;
7835 if (!ctx->file_data)
7839 if (copy_from_user(&up, arg, sizeof(up)))
7844 return __io_sqe_files_update(ctx, &up, nr_args);
7847 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7849 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7851 req = io_put_req_find_next(req);
7852 return req ? &req->work : NULL;
7855 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7856 struct task_struct *task)
7858 struct io_wq_hash *hash;
7859 struct io_wq_data data;
7860 unsigned int concurrency;
7862 hash = ctx->hash_map;
7864 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7866 return ERR_PTR(-ENOMEM);
7867 refcount_set(&hash->refs, 1);
7868 init_waitqueue_head(&hash->wait);
7869 ctx->hash_map = hash;
7874 data.free_work = io_free_work;
7875 data.do_work = io_wq_submit_work;
7877 /* Do QD, or 4 * CPUS, whatever is smallest */
7878 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7880 return io_wq_create(concurrency, &data);
7883 static int io_uring_alloc_task_context(struct task_struct *task,
7884 struct io_ring_ctx *ctx)
7886 struct io_uring_task *tctx;
7889 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7890 if (unlikely(!tctx))
7893 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7894 if (unlikely(ret)) {
7899 tctx->io_wq = io_init_wq_offload(ctx, task);
7900 if (IS_ERR(tctx->io_wq)) {
7901 ret = PTR_ERR(tctx->io_wq);
7902 percpu_counter_destroy(&tctx->inflight);
7908 init_waitqueue_head(&tctx->wait);
7910 atomic_set(&tctx->in_idle, 0);
7911 task->io_uring = tctx;
7912 spin_lock_init(&tctx->task_lock);
7913 INIT_WQ_LIST(&tctx->task_list);
7914 tctx->task_state = 0;
7915 init_task_work(&tctx->task_work, tctx_task_work);
7919 void __io_uring_free(struct task_struct *tsk)
7921 struct io_uring_task *tctx = tsk->io_uring;
7923 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7924 WARN_ON_ONCE(tctx->io_wq);
7926 percpu_counter_destroy(&tctx->inflight);
7928 tsk->io_uring = NULL;
7931 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7932 struct io_uring_params *p)
7936 /* Retain compatibility with failing for an invalid attach attempt */
7937 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7938 IORING_SETUP_ATTACH_WQ) {
7941 f = fdget(p->wq_fd);
7944 if (f.file->f_op != &io_uring_fops) {
7950 if (ctx->flags & IORING_SETUP_SQPOLL) {
7951 struct task_struct *tsk;
7952 struct io_sq_data *sqd;
7955 sqd = io_get_sq_data(p, &attached);
7961 ctx->sq_creds = get_current_cred();
7963 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7964 if (!ctx->sq_thread_idle)
7965 ctx->sq_thread_idle = HZ;
7968 io_sq_thread_park(sqd);
7969 list_add(&ctx->sqd_list, &sqd->ctx_list);
7970 io_sqd_update_thread_idle(sqd);
7971 /* don't attach to a dying SQPOLL thread, would be racy */
7972 if (attached && !sqd->thread)
7974 io_sq_thread_unpark(sqd);
7981 if (p->flags & IORING_SETUP_SQ_AFF) {
7982 int cpu = p->sq_thread_cpu;
7985 if (cpu >= nr_cpu_ids)
7987 if (!cpu_online(cpu))
7995 sqd->task_pid = current->pid;
7996 sqd->task_tgid = current->tgid;
7997 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8004 ret = io_uring_alloc_task_context(tsk, ctx);
8005 wake_up_new_task(tsk);
8008 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8009 /* Can't have SQ_AFF without SQPOLL */
8016 io_sq_thread_finish(ctx);
8019 complete(&ctx->sq_data->exited);
8023 static inline void __io_unaccount_mem(struct user_struct *user,
8024 unsigned long nr_pages)
8026 atomic_long_sub(nr_pages, &user->locked_vm);
8029 static inline int __io_account_mem(struct user_struct *user,
8030 unsigned long nr_pages)
8032 unsigned long page_limit, cur_pages, new_pages;
8034 /* Don't allow more pages than we can safely lock */
8035 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8038 cur_pages = atomic_long_read(&user->locked_vm);
8039 new_pages = cur_pages + nr_pages;
8040 if (new_pages > page_limit)
8042 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8043 new_pages) != cur_pages);
8048 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8051 __io_unaccount_mem(ctx->user, nr_pages);
8053 if (ctx->mm_account)
8054 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8057 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8062 ret = __io_account_mem(ctx->user, nr_pages);
8067 if (ctx->mm_account)
8068 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8073 static void io_mem_free(void *ptr)
8080 page = virt_to_head_page(ptr);
8081 if (put_page_testzero(page))
8082 free_compound_page(page);
8085 static void *io_mem_alloc(size_t size)
8087 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8088 __GFP_NORETRY | __GFP_ACCOUNT;
8090 return (void *) __get_free_pages(gfp_flags, get_order(size));
8093 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8096 struct io_rings *rings;
8097 size_t off, sq_array_size;
8099 off = struct_size(rings, cqes, cq_entries);
8100 if (off == SIZE_MAX)
8104 off = ALIGN(off, SMP_CACHE_BYTES);
8112 sq_array_size = array_size(sizeof(u32), sq_entries);
8113 if (sq_array_size == SIZE_MAX)
8116 if (check_add_overflow(off, sq_array_size, &off))
8122 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8126 if (!ctx->user_bufs)
8129 for (i = 0; i < ctx->nr_user_bufs; i++) {
8130 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8132 for (j = 0; j < imu->nr_bvecs; j++)
8133 unpin_user_page(imu->bvec[j].bv_page);
8135 if (imu->acct_pages)
8136 io_unaccount_mem(ctx, imu->acct_pages);
8141 kfree(ctx->user_bufs);
8142 ctx->user_bufs = NULL;
8143 ctx->nr_user_bufs = 0;
8147 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8148 void __user *arg, unsigned index)
8150 struct iovec __user *src;
8152 #ifdef CONFIG_COMPAT
8154 struct compat_iovec __user *ciovs;
8155 struct compat_iovec ciov;
8157 ciovs = (struct compat_iovec __user *) arg;
8158 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8161 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8162 dst->iov_len = ciov.iov_len;
8166 src = (struct iovec __user *) arg;
8167 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8173 * Not super efficient, but this is just a registration time. And we do cache
8174 * the last compound head, so generally we'll only do a full search if we don't
8177 * We check if the given compound head page has already been accounted, to
8178 * avoid double accounting it. This allows us to account the full size of the
8179 * page, not just the constituent pages of a huge page.
8181 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8182 int nr_pages, struct page *hpage)
8186 /* check current page array */
8187 for (i = 0; i < nr_pages; i++) {
8188 if (!PageCompound(pages[i]))
8190 if (compound_head(pages[i]) == hpage)
8194 /* check previously registered pages */
8195 for (i = 0; i < ctx->nr_user_bufs; i++) {
8196 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8198 for (j = 0; j < imu->nr_bvecs; j++) {
8199 if (!PageCompound(imu->bvec[j].bv_page))
8201 if (compound_head(imu->bvec[j].bv_page) == hpage)
8209 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8210 int nr_pages, struct io_mapped_ubuf *imu,
8211 struct page **last_hpage)
8215 for (i = 0; i < nr_pages; i++) {
8216 if (!PageCompound(pages[i])) {
8221 hpage = compound_head(pages[i]);
8222 if (hpage == *last_hpage)
8224 *last_hpage = hpage;
8225 if (headpage_already_acct(ctx, pages, i, hpage))
8227 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8231 if (!imu->acct_pages)
8234 ret = io_account_mem(ctx, imu->acct_pages);
8236 imu->acct_pages = 0;
8240 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8241 struct io_mapped_ubuf *imu,
8242 struct page **last_hpage)
8244 struct vm_area_struct **vmas = NULL;
8245 struct page **pages = NULL;
8246 unsigned long off, start, end, ubuf;
8248 int ret, pret, nr_pages, i;
8250 ubuf = (unsigned long) iov->iov_base;
8251 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8252 start = ubuf >> PAGE_SHIFT;
8253 nr_pages = end - start;
8257 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8261 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8266 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8272 mmap_read_lock(current->mm);
8273 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8275 if (pret == nr_pages) {
8276 /* don't support file backed memory */
8277 for (i = 0; i < nr_pages; i++) {
8278 struct vm_area_struct *vma = vmas[i];
8281 !is_file_hugepages(vma->vm_file)) {
8287 ret = pret < 0 ? pret : -EFAULT;
8289 mmap_read_unlock(current->mm);
8292 * if we did partial map, or found file backed vmas,
8293 * release any pages we did get
8296 unpin_user_pages(pages, pret);
8301 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8303 unpin_user_pages(pages, pret);
8308 off = ubuf & ~PAGE_MASK;
8309 size = iov->iov_len;
8310 for (i = 0; i < nr_pages; i++) {
8313 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8314 imu->bvec[i].bv_page = pages[i];
8315 imu->bvec[i].bv_len = vec_len;
8316 imu->bvec[i].bv_offset = off;
8320 /* store original address for later verification */
8322 imu->len = iov->iov_len;
8323 imu->nr_bvecs = nr_pages;
8331 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8335 if (!nr_args || nr_args > UIO_MAXIOV)
8338 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8340 if (!ctx->user_bufs)
8346 static int io_buffer_validate(struct iovec *iov)
8348 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8351 * Don't impose further limits on the size and buffer
8352 * constraints here, we'll -EINVAL later when IO is
8353 * submitted if they are wrong.
8355 if (!iov->iov_base || !iov->iov_len)
8358 /* arbitrary limit, but we need something */
8359 if (iov->iov_len > SZ_1G)
8362 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8368 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8369 unsigned int nr_args)
8373 struct page *last_hpage = NULL;
8375 ret = io_buffers_map_alloc(ctx, nr_args);
8379 for (i = 0; i < nr_args; i++) {
8380 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8382 ret = io_copy_iov(ctx, &iov, arg, i);
8386 ret = io_buffer_validate(&iov);
8390 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8394 ctx->nr_user_bufs++;
8398 io_sqe_buffers_unregister(ctx);
8403 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8405 __s32 __user *fds = arg;
8411 if (copy_from_user(&fd, fds, sizeof(*fds)))
8414 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8415 if (IS_ERR(ctx->cq_ev_fd)) {
8416 int ret = PTR_ERR(ctx->cq_ev_fd);
8417 ctx->cq_ev_fd = NULL;
8424 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8426 if (ctx->cq_ev_fd) {
8427 eventfd_ctx_put(ctx->cq_ev_fd);
8428 ctx->cq_ev_fd = NULL;
8435 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8437 struct io_buffer *buf;
8438 unsigned long index;
8440 xa_for_each(&ctx->io_buffers, index, buf)
8441 __io_remove_buffers(ctx, buf, index, -1U);
8444 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8446 struct io_kiocb *req, *nxt;
8448 list_for_each_entry_safe(req, nxt, list, compl.list) {
8449 if (tsk && req->task != tsk)
8451 list_del(&req->compl.list);
8452 kmem_cache_free(req_cachep, req);
8456 static void io_req_caches_free(struct io_ring_ctx *ctx)
8458 struct io_submit_state *submit_state = &ctx->submit_state;
8459 struct io_comp_state *cs = &ctx->submit_state.comp;
8461 mutex_lock(&ctx->uring_lock);
8463 if (submit_state->free_reqs) {
8464 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8465 submit_state->reqs);
8466 submit_state->free_reqs = 0;
8469 io_flush_cached_locked_reqs(ctx, cs);
8470 io_req_cache_free(&cs->free_list, NULL);
8471 mutex_unlock(&ctx->uring_lock);
8474 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8476 io_sq_thread_finish(ctx);
8477 io_sqe_buffers_unregister(ctx);
8479 if (ctx->mm_account) {
8480 mmdrop(ctx->mm_account);
8481 ctx->mm_account = NULL;
8484 mutex_lock(&ctx->uring_lock);
8485 io_sqe_files_unregister(ctx);
8486 mutex_unlock(&ctx->uring_lock);
8487 io_eventfd_unregister(ctx);
8488 io_destroy_buffers(ctx);
8490 /* there are no registered resources left, nobody uses it */
8492 io_rsrc_node_destroy(ctx->rsrc_node);
8493 if (ctx->rsrc_backup_node)
8494 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8495 flush_delayed_work(&ctx->rsrc_put_work);
8497 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8498 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8500 #if defined(CONFIG_UNIX)
8501 if (ctx->ring_sock) {
8502 ctx->ring_sock->file = NULL; /* so that iput() is called */
8503 sock_release(ctx->ring_sock);
8507 io_mem_free(ctx->rings);
8508 io_mem_free(ctx->sq_sqes);
8510 percpu_ref_exit(&ctx->refs);
8511 free_uid(ctx->user);
8512 io_req_caches_free(ctx);
8514 io_wq_put_hash(ctx->hash_map);
8515 kfree(ctx->cancel_hash);
8519 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8521 struct io_ring_ctx *ctx = file->private_data;
8524 poll_wait(file, &ctx->cq_wait, wait);
8526 * synchronizes with barrier from wq_has_sleeper call in
8530 if (!io_sqring_full(ctx))
8531 mask |= EPOLLOUT | EPOLLWRNORM;
8534 * Don't flush cqring overflow list here, just do a simple check.
8535 * Otherwise there could possible be ABBA deadlock:
8538 * lock(&ctx->uring_lock);
8540 * lock(&ctx->uring_lock);
8543 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8544 * pushs them to do the flush.
8546 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8547 mask |= EPOLLIN | EPOLLRDNORM;
8552 static int io_uring_fasync(int fd, struct file *file, int on)
8554 struct io_ring_ctx *ctx = file->private_data;
8556 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8559 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8561 const struct cred *creds;
8563 creds = xa_erase(&ctx->personalities, id);
8572 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8574 return io_run_task_work_head(&ctx->exit_task_work);
8577 struct io_tctx_exit {
8578 struct callback_head task_work;
8579 struct completion completion;
8580 struct io_ring_ctx *ctx;
8583 static void io_tctx_exit_cb(struct callback_head *cb)
8585 struct io_uring_task *tctx = current->io_uring;
8586 struct io_tctx_exit *work;
8588 work = container_of(cb, struct io_tctx_exit, task_work);
8590 * When @in_idle, we're in cancellation and it's racy to remove the
8591 * node. It'll be removed by the end of cancellation, just ignore it.
8593 if (!atomic_read(&tctx->in_idle))
8594 io_uring_del_task_file((unsigned long)work->ctx);
8595 complete(&work->completion);
8598 static void io_ring_exit_work(struct work_struct *work)
8600 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8601 unsigned long timeout = jiffies + HZ * 60 * 5;
8602 struct io_tctx_exit exit;
8603 struct io_tctx_node *node;
8606 /* prevent SQPOLL from submitting new requests */
8608 io_sq_thread_park(ctx->sq_data);
8609 list_del_init(&ctx->sqd_list);
8610 io_sqd_update_thread_idle(ctx->sq_data);
8611 io_sq_thread_unpark(ctx->sq_data);
8615 * If we're doing polled IO and end up having requests being
8616 * submitted async (out-of-line), then completions can come in while
8617 * we're waiting for refs to drop. We need to reap these manually,
8618 * as nobody else will be looking for them.
8621 io_uring_try_cancel_requests(ctx, NULL, NULL);
8623 WARN_ON_ONCE(time_after(jiffies, timeout));
8624 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8627 * Some may use context even when all refs and requests have been put,
8628 * and they are free to do so while still holding uring_lock or
8629 * completion_lock, see __io_req_task_submit(). Apart from other work,
8630 * this lock/unlock section also waits them to finish.
8632 mutex_lock(&ctx->uring_lock);
8633 while (!list_empty(&ctx->tctx_list)) {
8634 WARN_ON_ONCE(time_after(jiffies, timeout));
8636 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8639 init_completion(&exit.completion);
8640 init_task_work(&exit.task_work, io_tctx_exit_cb);
8641 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8642 if (WARN_ON_ONCE(ret))
8644 wake_up_process(node->task);
8646 mutex_unlock(&ctx->uring_lock);
8647 wait_for_completion(&exit.completion);
8649 mutex_lock(&ctx->uring_lock);
8651 mutex_unlock(&ctx->uring_lock);
8652 spin_lock_irq(&ctx->completion_lock);
8653 spin_unlock_irq(&ctx->completion_lock);
8655 io_ring_ctx_free(ctx);
8658 /* Returns true if we found and killed one or more timeouts */
8659 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8660 struct files_struct *files)
8662 struct io_kiocb *req, *tmp;
8665 spin_lock_irq(&ctx->completion_lock);
8666 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8667 if (io_match_task(req, tsk, files)) {
8668 io_kill_timeout(req, -ECANCELED);
8673 io_commit_cqring(ctx);
8674 spin_unlock_irq(&ctx->completion_lock);
8676 io_cqring_ev_posted(ctx);
8677 return canceled != 0;
8680 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8682 unsigned long index;
8683 struct creds *creds;
8685 mutex_lock(&ctx->uring_lock);
8686 percpu_ref_kill(&ctx->refs);
8687 /* if force is set, the ring is going away. always drop after that */
8688 ctx->cq_overflow_flushed = 1;
8690 __io_cqring_overflow_flush(ctx, true);
8691 xa_for_each(&ctx->personalities, index, creds)
8692 io_unregister_personality(ctx, index);
8693 mutex_unlock(&ctx->uring_lock);
8695 io_kill_timeouts(ctx, NULL, NULL);
8696 io_poll_remove_all(ctx, NULL, NULL);
8698 /* if we failed setting up the ctx, we might not have any rings */
8699 io_iopoll_try_reap_events(ctx);
8701 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8703 * Use system_unbound_wq to avoid spawning tons of event kworkers
8704 * if we're exiting a ton of rings at the same time. It just adds
8705 * noise and overhead, there's no discernable change in runtime
8706 * over using system_wq.
8708 queue_work(system_unbound_wq, &ctx->exit_work);
8711 static int io_uring_release(struct inode *inode, struct file *file)
8713 struct io_ring_ctx *ctx = file->private_data;
8715 file->private_data = NULL;
8716 io_ring_ctx_wait_and_kill(ctx);
8720 struct io_task_cancel {
8721 struct task_struct *task;
8722 struct files_struct *files;
8725 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8727 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8728 struct io_task_cancel *cancel = data;
8731 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8732 unsigned long flags;
8733 struct io_ring_ctx *ctx = req->ctx;
8735 /* protect against races with linked timeouts */
8736 spin_lock_irqsave(&ctx->completion_lock, flags);
8737 ret = io_match_task(req, cancel->task, cancel->files);
8738 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8740 ret = io_match_task(req, cancel->task, cancel->files);
8745 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8746 struct task_struct *task,
8747 struct files_struct *files)
8749 struct io_defer_entry *de;
8752 spin_lock_irq(&ctx->completion_lock);
8753 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8754 if (io_match_task(de->req, task, files)) {
8755 list_cut_position(&list, &ctx->defer_list, &de->list);
8759 spin_unlock_irq(&ctx->completion_lock);
8760 if (list_empty(&list))
8763 while (!list_empty(&list)) {
8764 de = list_first_entry(&list, struct io_defer_entry, list);
8765 list_del_init(&de->list);
8766 io_req_complete_failed(de->req, -ECANCELED);
8772 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8774 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8776 return req->ctx == data;
8779 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8781 struct io_tctx_node *node;
8782 enum io_wq_cancel cret;
8785 mutex_lock(&ctx->uring_lock);
8786 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8787 struct io_uring_task *tctx = node->task->io_uring;
8790 * io_wq will stay alive while we hold uring_lock, because it's
8791 * killed after ctx nodes, which requires to take the lock.
8793 if (!tctx || !tctx->io_wq)
8795 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8796 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8798 mutex_unlock(&ctx->uring_lock);
8803 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8804 struct task_struct *task,
8805 struct files_struct *files)
8807 struct io_task_cancel cancel = { .task = task, .files = files, };
8808 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8811 enum io_wq_cancel cret;
8815 ret |= io_uring_try_cancel_iowq(ctx);
8816 } else if (tctx && tctx->io_wq) {
8818 * Cancels requests of all rings, not only @ctx, but
8819 * it's fine as the task is in exit/exec.
8821 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8823 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8826 /* SQPOLL thread does its own polling */
8827 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8828 (ctx->sq_data && ctx->sq_data->thread == current)) {
8829 while (!list_empty_careful(&ctx->iopoll_list)) {
8830 io_iopoll_try_reap_events(ctx);
8835 ret |= io_cancel_defer_files(ctx, task, files);
8836 ret |= io_poll_remove_all(ctx, task, files);
8837 ret |= io_kill_timeouts(ctx, task, files);
8838 ret |= io_run_task_work();
8839 ret |= io_run_ctx_fallback(ctx);
8846 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8847 struct task_struct *task,
8848 struct files_struct *files)
8850 struct io_kiocb *req;
8853 spin_lock_irq(&ctx->inflight_lock);
8854 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8855 cnt += io_match_task(req, task, files);
8856 spin_unlock_irq(&ctx->inflight_lock);
8860 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8861 struct task_struct *task,
8862 struct files_struct *files)
8864 while (!list_empty_careful(&ctx->inflight_list)) {
8868 inflight = io_uring_count_inflight(ctx, task, files);
8872 io_uring_try_cancel_requests(ctx, task, files);
8874 prepare_to_wait(&task->io_uring->wait, &wait,
8875 TASK_UNINTERRUPTIBLE);
8876 if (inflight == io_uring_count_inflight(ctx, task, files))
8878 finish_wait(&task->io_uring->wait, &wait);
8882 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8884 struct io_uring_task *tctx = current->io_uring;
8885 struct io_tctx_node *node;
8888 if (unlikely(!tctx)) {
8889 ret = io_uring_alloc_task_context(current, ctx);
8892 tctx = current->io_uring;
8894 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8895 node = kmalloc(sizeof(*node), GFP_KERNEL);
8899 node->task = current;
8901 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8908 mutex_lock(&ctx->uring_lock);
8909 list_add(&node->ctx_node, &ctx->tctx_list);
8910 mutex_unlock(&ctx->uring_lock);
8917 * Note that this task has used io_uring. We use it for cancelation purposes.
8919 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8921 struct io_uring_task *tctx = current->io_uring;
8923 if (likely(tctx && tctx->last == ctx))
8925 return __io_uring_add_task_file(ctx);
8929 * Remove this io_uring_file -> task mapping.
8931 static void io_uring_del_task_file(unsigned long index)
8933 struct io_uring_task *tctx = current->io_uring;
8934 struct io_tctx_node *node;
8938 node = xa_erase(&tctx->xa, index);
8942 WARN_ON_ONCE(current != node->task);
8943 WARN_ON_ONCE(list_empty(&node->ctx_node));
8945 mutex_lock(&node->ctx->uring_lock);
8946 list_del(&node->ctx_node);
8947 mutex_unlock(&node->ctx->uring_lock);
8949 if (tctx->last == node->ctx)
8954 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8956 struct io_tctx_node *node;
8957 unsigned long index;
8959 xa_for_each(&tctx->xa, index, node)
8960 io_uring_del_task_file(index);
8962 io_wq_put_and_exit(tctx->io_wq);
8967 static s64 tctx_inflight(struct io_uring_task *tctx)
8969 return percpu_counter_sum(&tctx->inflight);
8972 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8974 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8975 struct io_ring_ctx *ctx = work->ctx;
8976 struct io_sq_data *sqd = ctx->sq_data;
8979 io_uring_cancel_sqpoll(ctx);
8980 complete(&work->completion);
8983 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8985 struct io_sq_data *sqd = ctx->sq_data;
8986 struct io_tctx_exit work = { .ctx = ctx, };
8987 struct task_struct *task;
8989 io_sq_thread_park(sqd);
8990 list_del_init(&ctx->sqd_list);
8991 io_sqd_update_thread_idle(sqd);
8994 init_completion(&work.completion);
8995 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8996 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8997 wake_up_process(task);
8999 io_sq_thread_unpark(sqd);
9002 wait_for_completion(&work.completion);
9005 void __io_uring_files_cancel(struct files_struct *files)
9007 struct io_uring_task *tctx = current->io_uring;
9008 struct io_tctx_node *node;
9009 unsigned long index;
9011 /* make sure overflow events are dropped */
9012 atomic_inc(&tctx->in_idle);
9013 xa_for_each(&tctx->xa, index, node) {
9014 struct io_ring_ctx *ctx = node->ctx;
9017 io_sqpoll_cancel_sync(ctx);
9020 io_uring_cancel_files(ctx, current, files);
9022 io_uring_try_cancel_requests(ctx, current, NULL);
9024 atomic_dec(&tctx->in_idle);
9027 io_uring_clean_tctx(tctx);
9030 /* should only be called by SQPOLL task */
9031 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9033 struct io_sq_data *sqd = ctx->sq_data;
9034 struct io_uring_task *tctx = current->io_uring;
9038 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
9040 atomic_inc(&tctx->in_idle);
9042 /* read completions before cancelations */
9043 inflight = tctx_inflight(tctx);
9046 io_uring_try_cancel_requests(ctx, current, NULL);
9048 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9050 * If we've seen completions, retry without waiting. This
9051 * avoids a race where a completion comes in before we did
9052 * prepare_to_wait().
9054 if (inflight == tctx_inflight(tctx))
9056 finish_wait(&tctx->wait, &wait);
9058 atomic_dec(&tctx->in_idle);
9062 * Find any io_uring fd that this task has registered or done IO on, and cancel
9065 void __io_uring_task_cancel(void)
9067 struct io_uring_task *tctx = current->io_uring;
9071 /* make sure overflow events are dropped */
9072 atomic_inc(&tctx->in_idle);
9073 __io_uring_files_cancel(NULL);
9076 /* read completions before cancelations */
9077 inflight = tctx_inflight(tctx);
9080 __io_uring_files_cancel(NULL);
9082 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9085 * If we've seen completions, retry without waiting. This
9086 * avoids a race where a completion comes in before we did
9087 * prepare_to_wait().
9089 if (inflight == tctx_inflight(tctx))
9091 finish_wait(&tctx->wait, &wait);
9094 atomic_dec(&tctx->in_idle);
9096 io_uring_clean_tctx(tctx);
9097 /* all current's requests should be gone, we can kill tctx */
9098 __io_uring_free(current);
9101 static void *io_uring_validate_mmap_request(struct file *file,
9102 loff_t pgoff, size_t sz)
9104 struct io_ring_ctx *ctx = file->private_data;
9105 loff_t offset = pgoff << PAGE_SHIFT;
9110 case IORING_OFF_SQ_RING:
9111 case IORING_OFF_CQ_RING:
9114 case IORING_OFF_SQES:
9118 return ERR_PTR(-EINVAL);
9121 page = virt_to_head_page(ptr);
9122 if (sz > page_size(page))
9123 return ERR_PTR(-EINVAL);
9130 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9132 size_t sz = vma->vm_end - vma->vm_start;
9136 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9138 return PTR_ERR(ptr);
9140 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9141 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9144 #else /* !CONFIG_MMU */
9146 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9148 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9151 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9153 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9156 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9157 unsigned long addr, unsigned long len,
9158 unsigned long pgoff, unsigned long flags)
9162 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9164 return PTR_ERR(ptr);
9166 return (unsigned long) ptr;
9169 #endif /* !CONFIG_MMU */
9171 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9176 if (!io_sqring_full(ctx))
9178 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9180 if (!io_sqring_full(ctx))
9183 } while (!signal_pending(current));
9185 finish_wait(&ctx->sqo_sq_wait, &wait);
9189 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9190 struct __kernel_timespec __user **ts,
9191 const sigset_t __user **sig)
9193 struct io_uring_getevents_arg arg;
9196 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9197 * is just a pointer to the sigset_t.
9199 if (!(flags & IORING_ENTER_EXT_ARG)) {
9200 *sig = (const sigset_t __user *) argp;
9206 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9207 * timespec and sigset_t pointers if good.
9209 if (*argsz != sizeof(arg))
9211 if (copy_from_user(&arg, argp, sizeof(arg)))
9213 *sig = u64_to_user_ptr(arg.sigmask);
9214 *argsz = arg.sigmask_sz;
9215 *ts = u64_to_user_ptr(arg.ts);
9219 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9220 u32, min_complete, u32, flags, const void __user *, argp,
9223 struct io_ring_ctx *ctx;
9230 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9231 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9235 if (unlikely(!f.file))
9239 if (unlikely(f.file->f_op != &io_uring_fops))
9243 ctx = f.file->private_data;
9244 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9248 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9252 * For SQ polling, the thread will do all submissions and completions.
9253 * Just return the requested submit count, and wake the thread if
9257 if (ctx->flags & IORING_SETUP_SQPOLL) {
9258 io_cqring_overflow_flush(ctx, false);
9261 if (unlikely(ctx->sq_data->thread == NULL)) {
9264 if (flags & IORING_ENTER_SQ_WAKEUP)
9265 wake_up(&ctx->sq_data->wait);
9266 if (flags & IORING_ENTER_SQ_WAIT) {
9267 ret = io_sqpoll_wait_sq(ctx);
9271 submitted = to_submit;
9272 } else if (to_submit) {
9273 ret = io_uring_add_task_file(ctx);
9276 mutex_lock(&ctx->uring_lock);
9277 submitted = io_submit_sqes(ctx, to_submit);
9278 mutex_unlock(&ctx->uring_lock);
9280 if (submitted != to_submit)
9283 if (flags & IORING_ENTER_GETEVENTS) {
9284 const sigset_t __user *sig;
9285 struct __kernel_timespec __user *ts;
9287 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9291 min_complete = min(min_complete, ctx->cq_entries);
9294 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9295 * space applications don't need to do io completion events
9296 * polling again, they can rely on io_sq_thread to do polling
9297 * work, which can reduce cpu usage and uring_lock contention.
9299 if (ctx->flags & IORING_SETUP_IOPOLL &&
9300 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9301 ret = io_iopoll_check(ctx, min_complete);
9303 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9308 percpu_ref_put(&ctx->refs);
9311 return submitted ? submitted : ret;
9314 #ifdef CONFIG_PROC_FS
9315 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9316 const struct cred *cred)
9318 struct user_namespace *uns = seq_user_ns(m);
9319 struct group_info *gi;
9324 seq_printf(m, "%5d\n", id);
9325 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9326 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9327 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9328 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9329 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9330 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9331 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9332 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9333 seq_puts(m, "\n\tGroups:\t");
9334 gi = cred->group_info;
9335 for (g = 0; g < gi->ngroups; g++) {
9336 seq_put_decimal_ull(m, g ? " " : "",
9337 from_kgid_munged(uns, gi->gid[g]));
9339 seq_puts(m, "\n\tCapEff:\t");
9340 cap = cred->cap_effective;
9341 CAP_FOR_EACH_U32(__capi)
9342 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9347 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9349 struct io_sq_data *sq = NULL;
9354 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9355 * since fdinfo case grabs it in the opposite direction of normal use
9356 * cases. If we fail to get the lock, we just don't iterate any
9357 * structures that could be going away outside the io_uring mutex.
9359 has_lock = mutex_trylock(&ctx->uring_lock);
9361 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9367 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9368 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9369 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9370 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9371 struct file *f = io_file_from_index(ctx, i);
9374 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9376 seq_printf(m, "%5u: <none>\n", i);
9378 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9379 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9380 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9382 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9383 (unsigned int) buf->len);
9385 if (has_lock && !xa_empty(&ctx->personalities)) {
9386 unsigned long index;
9387 const struct cred *cred;
9389 seq_printf(m, "Personalities:\n");
9390 xa_for_each(&ctx->personalities, index, cred)
9391 io_uring_show_cred(m, index, cred);
9393 seq_printf(m, "PollList:\n");
9394 spin_lock_irq(&ctx->completion_lock);
9395 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9396 struct hlist_head *list = &ctx->cancel_hash[i];
9397 struct io_kiocb *req;
9399 hlist_for_each_entry(req, list, hash_node)
9400 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9401 req->task->task_works != NULL);
9403 spin_unlock_irq(&ctx->completion_lock);
9405 mutex_unlock(&ctx->uring_lock);
9408 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9410 struct io_ring_ctx *ctx = f->private_data;
9412 if (percpu_ref_tryget(&ctx->refs)) {
9413 __io_uring_show_fdinfo(ctx, m);
9414 percpu_ref_put(&ctx->refs);
9419 static const struct file_operations io_uring_fops = {
9420 .release = io_uring_release,
9421 .mmap = io_uring_mmap,
9423 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9424 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9426 .poll = io_uring_poll,
9427 .fasync = io_uring_fasync,
9428 #ifdef CONFIG_PROC_FS
9429 .show_fdinfo = io_uring_show_fdinfo,
9433 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9434 struct io_uring_params *p)
9436 struct io_rings *rings;
9437 size_t size, sq_array_offset;
9439 /* make sure these are sane, as we already accounted them */
9440 ctx->sq_entries = p->sq_entries;
9441 ctx->cq_entries = p->cq_entries;
9443 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9444 if (size == SIZE_MAX)
9447 rings = io_mem_alloc(size);
9452 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9453 rings->sq_ring_mask = p->sq_entries - 1;
9454 rings->cq_ring_mask = p->cq_entries - 1;
9455 rings->sq_ring_entries = p->sq_entries;
9456 rings->cq_ring_entries = p->cq_entries;
9457 ctx->sq_mask = rings->sq_ring_mask;
9458 ctx->cq_mask = rings->cq_ring_mask;
9460 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9461 if (size == SIZE_MAX) {
9462 io_mem_free(ctx->rings);
9467 ctx->sq_sqes = io_mem_alloc(size);
9468 if (!ctx->sq_sqes) {
9469 io_mem_free(ctx->rings);
9477 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9481 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9485 ret = io_uring_add_task_file(ctx);
9490 fd_install(fd, file);
9495 * Allocate an anonymous fd, this is what constitutes the application
9496 * visible backing of an io_uring instance. The application mmaps this
9497 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9498 * we have to tie this fd to a socket for file garbage collection purposes.
9500 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9503 #if defined(CONFIG_UNIX)
9506 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9509 return ERR_PTR(ret);
9512 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9513 O_RDWR | O_CLOEXEC);
9514 #if defined(CONFIG_UNIX)
9516 sock_release(ctx->ring_sock);
9517 ctx->ring_sock = NULL;
9519 ctx->ring_sock->file = file;
9525 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9526 struct io_uring_params __user *params)
9528 struct io_ring_ctx *ctx;
9534 if (entries > IORING_MAX_ENTRIES) {
9535 if (!(p->flags & IORING_SETUP_CLAMP))
9537 entries = IORING_MAX_ENTRIES;
9541 * Use twice as many entries for the CQ ring. It's possible for the
9542 * application to drive a higher depth than the size of the SQ ring,
9543 * since the sqes are only used at submission time. This allows for
9544 * some flexibility in overcommitting a bit. If the application has
9545 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9546 * of CQ ring entries manually.
9548 p->sq_entries = roundup_pow_of_two(entries);
9549 if (p->flags & IORING_SETUP_CQSIZE) {
9551 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9552 * to a power-of-two, if it isn't already. We do NOT impose
9553 * any cq vs sq ring sizing.
9557 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9558 if (!(p->flags & IORING_SETUP_CLAMP))
9560 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9562 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9563 if (p->cq_entries < p->sq_entries)
9566 p->cq_entries = 2 * p->sq_entries;
9569 ctx = io_ring_ctx_alloc(p);
9572 ctx->compat = in_compat_syscall();
9573 if (!capable(CAP_IPC_LOCK))
9574 ctx->user = get_uid(current_user());
9577 * This is just grabbed for accounting purposes. When a process exits,
9578 * the mm is exited and dropped before the files, hence we need to hang
9579 * on to this mm purely for the purposes of being able to unaccount
9580 * memory (locked/pinned vm). It's not used for anything else.
9582 mmgrab(current->mm);
9583 ctx->mm_account = current->mm;
9585 ret = io_allocate_scq_urings(ctx, p);
9589 ret = io_sq_offload_create(ctx, p);
9593 memset(&p->sq_off, 0, sizeof(p->sq_off));
9594 p->sq_off.head = offsetof(struct io_rings, sq.head);
9595 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9596 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9597 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9598 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9599 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9600 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9602 memset(&p->cq_off, 0, sizeof(p->cq_off));
9603 p->cq_off.head = offsetof(struct io_rings, cq.head);
9604 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9605 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9606 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9607 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9608 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9609 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9611 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9612 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9613 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9614 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9615 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9617 if (copy_to_user(params, p, sizeof(*p))) {
9622 file = io_uring_get_file(ctx);
9624 ret = PTR_ERR(file);
9629 * Install ring fd as the very last thing, so we don't risk someone
9630 * having closed it before we finish setup
9632 ret = io_uring_install_fd(ctx, file);
9634 /* fput will clean it up */
9639 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9642 io_ring_ctx_wait_and_kill(ctx);
9647 * Sets up an aio uring context, and returns the fd. Applications asks for a
9648 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9649 * params structure passed in.
9651 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9653 struct io_uring_params p;
9656 if (copy_from_user(&p, params, sizeof(p)))
9658 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9663 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9664 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9665 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9666 IORING_SETUP_R_DISABLED))
9669 return io_uring_create(entries, &p, params);
9672 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9673 struct io_uring_params __user *, params)
9675 return io_uring_setup(entries, params);
9678 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9680 struct io_uring_probe *p;
9684 size = struct_size(p, ops, nr_args);
9685 if (size == SIZE_MAX)
9687 p = kzalloc(size, GFP_KERNEL);
9692 if (copy_from_user(p, arg, size))
9695 if (memchr_inv(p, 0, size))
9698 p->last_op = IORING_OP_LAST - 1;
9699 if (nr_args > IORING_OP_LAST)
9700 nr_args = IORING_OP_LAST;
9702 for (i = 0; i < nr_args; i++) {
9704 if (!io_op_defs[i].not_supported)
9705 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9710 if (copy_to_user(arg, p, size))
9717 static int io_register_personality(struct io_ring_ctx *ctx)
9719 const struct cred *creds;
9723 creds = get_current_cred();
9725 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9726 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9733 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9734 unsigned int nr_args)
9736 struct io_uring_restriction *res;
9740 /* Restrictions allowed only if rings started disabled */
9741 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9744 /* We allow only a single restrictions registration */
9745 if (ctx->restrictions.registered)
9748 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9751 size = array_size(nr_args, sizeof(*res));
9752 if (size == SIZE_MAX)
9755 res = memdup_user(arg, size);
9757 return PTR_ERR(res);
9761 for (i = 0; i < nr_args; i++) {
9762 switch (res[i].opcode) {
9763 case IORING_RESTRICTION_REGISTER_OP:
9764 if (res[i].register_op >= IORING_REGISTER_LAST) {
9769 __set_bit(res[i].register_op,
9770 ctx->restrictions.register_op);
9772 case IORING_RESTRICTION_SQE_OP:
9773 if (res[i].sqe_op >= IORING_OP_LAST) {
9778 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9780 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9781 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9783 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9784 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9793 /* Reset all restrictions if an error happened */
9795 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9797 ctx->restrictions.registered = true;
9803 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9805 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9808 if (ctx->restrictions.registered)
9809 ctx->restricted = 1;
9811 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9812 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9813 wake_up(&ctx->sq_data->wait);
9817 static bool io_register_op_must_quiesce(int op)
9820 case IORING_UNREGISTER_FILES:
9821 case IORING_REGISTER_FILES_UPDATE:
9822 case IORING_REGISTER_PROBE:
9823 case IORING_REGISTER_PERSONALITY:
9824 case IORING_UNREGISTER_PERSONALITY:
9831 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9832 void __user *arg, unsigned nr_args)
9833 __releases(ctx->uring_lock)
9834 __acquires(ctx->uring_lock)
9839 * We're inside the ring mutex, if the ref is already dying, then
9840 * someone else killed the ctx or is already going through
9841 * io_uring_register().
9843 if (percpu_ref_is_dying(&ctx->refs))
9846 if (io_register_op_must_quiesce(opcode)) {
9847 percpu_ref_kill(&ctx->refs);
9850 * Drop uring mutex before waiting for references to exit. If
9851 * another thread is currently inside io_uring_enter() it might
9852 * need to grab the uring_lock to make progress. If we hold it
9853 * here across the drain wait, then we can deadlock. It's safe
9854 * to drop the mutex here, since no new references will come in
9855 * after we've killed the percpu ref.
9857 mutex_unlock(&ctx->uring_lock);
9859 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9862 ret = io_run_task_work_sig();
9867 mutex_lock(&ctx->uring_lock);
9870 percpu_ref_resurrect(&ctx->refs);
9875 if (ctx->restricted) {
9876 if (opcode >= IORING_REGISTER_LAST) {
9881 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9888 case IORING_REGISTER_BUFFERS:
9889 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9891 case IORING_UNREGISTER_BUFFERS:
9895 ret = io_sqe_buffers_unregister(ctx);
9897 case IORING_REGISTER_FILES:
9898 ret = io_sqe_files_register(ctx, arg, nr_args);
9900 case IORING_UNREGISTER_FILES:
9904 ret = io_sqe_files_unregister(ctx);
9906 case IORING_REGISTER_FILES_UPDATE:
9907 ret = io_sqe_files_update(ctx, arg, nr_args);
9909 case IORING_REGISTER_EVENTFD:
9910 case IORING_REGISTER_EVENTFD_ASYNC:
9914 ret = io_eventfd_register(ctx, arg);
9917 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9918 ctx->eventfd_async = 1;
9920 ctx->eventfd_async = 0;
9922 case IORING_UNREGISTER_EVENTFD:
9926 ret = io_eventfd_unregister(ctx);
9928 case IORING_REGISTER_PROBE:
9930 if (!arg || nr_args > 256)
9932 ret = io_probe(ctx, arg, nr_args);
9934 case IORING_REGISTER_PERSONALITY:
9938 ret = io_register_personality(ctx);
9940 case IORING_UNREGISTER_PERSONALITY:
9944 ret = io_unregister_personality(ctx, nr_args);
9946 case IORING_REGISTER_ENABLE_RINGS:
9950 ret = io_register_enable_rings(ctx);
9952 case IORING_REGISTER_RESTRICTIONS:
9953 ret = io_register_restrictions(ctx, arg, nr_args);
9961 if (io_register_op_must_quiesce(opcode)) {
9962 /* bring the ctx back to life */
9963 percpu_ref_reinit(&ctx->refs);
9965 reinit_completion(&ctx->ref_comp);
9970 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9971 void __user *, arg, unsigned int, nr_args)
9973 struct io_ring_ctx *ctx;
9982 if (f.file->f_op != &io_uring_fops)
9985 ctx = f.file->private_data;
9989 mutex_lock(&ctx->uring_lock);
9990 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9991 mutex_unlock(&ctx->uring_lock);
9992 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9993 ctx->cq_ev_fd != NULL, ret);
9999 static int __init io_uring_init(void)
10001 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10002 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10003 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10006 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10007 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10008 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10009 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10010 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10011 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10012 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10013 BUILD_BUG_SQE_ELEM(8, __u64, off);
10014 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10015 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10016 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10017 BUILD_BUG_SQE_ELEM(24, __u32, len);
10018 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10019 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10020 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10021 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10022 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10023 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10024 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10025 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10026 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10027 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10028 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10029 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10030 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10031 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10032 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10033 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10034 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10035 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10036 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10038 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10039 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10040 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10044 __initcall(io_uring_init);