1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
239 struct list_head list;
245 struct io_restriction {
246 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
247 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
248 u8 sqe_flags_allowed;
249 u8 sqe_flags_required;
254 IO_SQ_THREAD_SHOULD_STOP = 0,
255 IO_SQ_THREAD_SHOULD_PARK,
260 atomic_t park_pending;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
276 struct callback_head *park_task_work;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
449 struct callback_head *exit_task_work;
451 struct wait_queue_head hash_wait;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work;
455 struct list_head tctx_list;
458 struct io_uring_task {
459 /* submission side */
461 struct wait_queue_head wait;
462 const struct io_ring_ctx *last;
464 struct percpu_counter inflight;
468 spinlock_t task_lock;
469 struct io_wq_work_list task_list;
470 unsigned long task_state;
471 struct callback_head task_work;
475 * First field must be the file pointer in all the
476 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
478 struct io_poll_iocb {
480 struct wait_queue_head *head;
484 struct wait_queue_entry wait;
487 struct io_poll_remove {
497 struct io_timeout_data {
498 struct io_kiocb *req;
499 struct hrtimer timer;
500 struct timespec64 ts;
501 enum hrtimer_mode mode;
506 struct sockaddr __user *addr;
507 int __user *addr_len;
509 unsigned long nofile;
529 struct list_head list;
530 /* head of the link, used by linked timeouts only */
531 struct io_kiocb *head;
534 struct io_timeout_rem {
539 struct timespec64 ts;
544 /* NOTE: kiocb has the file as the first member, so don't do it here */
552 struct sockaddr __user *addr;
559 struct user_msghdr __user *umsg;
565 struct io_buffer *kbuf;
571 struct filename *filename;
573 unsigned long nofile;
576 struct io_rsrc_update {
602 struct epoll_event event;
606 struct file *file_out;
607 struct file *file_in;
614 struct io_provide_buf {
628 const char __user *filename;
629 struct statx __user *buffer;
641 struct filename *oldpath;
642 struct filename *newpath;
650 struct filename *filename;
653 struct io_completion {
655 struct list_head list;
659 struct io_async_connect {
660 struct sockaddr_storage address;
663 struct io_async_msghdr {
664 struct iovec fast_iov[UIO_FASTIOV];
665 /* points to an allocated iov, if NULL we use fast_iov instead */
666 struct iovec *free_iov;
667 struct sockaddr __user *uaddr;
669 struct sockaddr_storage addr;
673 struct iovec fast_iov[UIO_FASTIOV];
674 const struct iovec *free_iovec;
675 struct iov_iter iter;
677 struct wait_page_queue wpq;
681 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
682 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
683 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
684 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
685 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
686 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
692 REQ_F_LINK_TIMEOUT_BIT,
694 REQ_F_NEED_CLEANUP_BIT,
696 REQ_F_BUFFER_SELECTED_BIT,
697 REQ_F_NO_FILE_TABLE_BIT,
698 REQ_F_LTIMEOUT_ACTIVE_BIT,
699 REQ_F_COMPLETE_INLINE_BIT,
701 /* not a real bit, just to check we're not overflowing the space */
707 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
708 /* drain existing IO first */
709 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
711 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
712 /* doesn't sever on completion < 0 */
713 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
715 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
716 /* IOSQE_BUFFER_SELECT */
717 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
719 /* fail rest of links */
720 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
721 /* on inflight list, should be cancelled and waited on exit reliably */
722 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
723 /* read/write uses file position */
724 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
725 /* must not punt to workers */
726 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
727 /* has or had linked timeout */
728 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
730 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
732 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
733 /* already went through poll handler */
734 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
735 /* buffer already selected */
736 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
737 /* doesn't need file table for this request */
738 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
739 /* linked timeout is active, i.e. prepared by link's head */
740 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
741 /* completion is deferred through io_comp_state */
742 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
746 struct io_poll_iocb poll;
747 struct io_poll_iocb *double_poll;
750 struct io_task_work {
751 struct io_wq_work_node node;
752 task_work_func_t func;
756 * NOTE! Each of the iocb union members has the file pointer
757 * as the first entry in their struct definition. So you can
758 * access the file pointer through any of the sub-structs,
759 * or directly as just 'ki_filp' in this struct.
765 struct io_poll_iocb poll;
766 struct io_poll_remove poll_remove;
767 struct io_accept accept;
769 struct io_cancel cancel;
770 struct io_timeout timeout;
771 struct io_timeout_rem timeout_rem;
772 struct io_connect connect;
773 struct io_sr_msg sr_msg;
775 struct io_close close;
776 struct io_rsrc_update rsrc_update;
777 struct io_fadvise fadvise;
778 struct io_madvise madvise;
779 struct io_epoll epoll;
780 struct io_splice splice;
781 struct io_provide_buf pbuf;
782 struct io_statx statx;
783 struct io_shutdown shutdown;
784 struct io_rename rename;
785 struct io_unlink unlink;
786 /* use only after cleaning per-op data, see io_clean_op() */
787 struct io_completion compl;
790 /* opcode allocated if it needs to store data for async defer */
793 /* polled IO has completed */
799 struct io_ring_ctx *ctx;
802 struct task_struct *task;
805 struct io_kiocb *link;
806 struct percpu_ref *fixed_rsrc_refs;
809 * 1. used with ctx->iopoll_list with reads/writes
810 * 2. to track reqs with ->files (see io_op_def::file_table)
812 struct list_head inflight_entry;
814 struct io_task_work io_task_work;
815 struct callback_head task_work;
817 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
818 struct hlist_node hash_node;
819 struct async_poll *apoll;
820 struct io_wq_work work;
823 struct io_tctx_node {
824 struct list_head ctx_node;
825 struct task_struct *task;
826 struct io_ring_ctx *ctx;
829 struct io_defer_entry {
830 struct list_head list;
831 struct io_kiocb *req;
836 /* needs req->file assigned */
837 unsigned needs_file : 1;
838 /* hash wq insertion if file is a regular file */
839 unsigned hash_reg_file : 1;
840 /* unbound wq insertion if file is a non-regular file */
841 unsigned unbound_nonreg_file : 1;
842 /* opcode is not supported by this kernel */
843 unsigned not_supported : 1;
844 /* set if opcode supports polled "wait" */
846 unsigned pollout : 1;
847 /* op supports buffer selection */
848 unsigned buffer_select : 1;
849 /* must always have async data allocated */
850 unsigned needs_async_data : 1;
851 /* should block plug */
853 /* size of async data needed, if any */
854 unsigned short async_size;
857 static const struct io_op_def io_op_defs[] = {
858 [IORING_OP_NOP] = {},
859 [IORING_OP_READV] = {
861 .unbound_nonreg_file = 1,
864 .needs_async_data = 1,
866 .async_size = sizeof(struct io_async_rw),
868 [IORING_OP_WRITEV] = {
871 .unbound_nonreg_file = 1,
873 .needs_async_data = 1,
875 .async_size = sizeof(struct io_async_rw),
877 [IORING_OP_FSYNC] = {
880 [IORING_OP_READ_FIXED] = {
882 .unbound_nonreg_file = 1,
885 .async_size = sizeof(struct io_async_rw),
887 [IORING_OP_WRITE_FIXED] = {
890 .unbound_nonreg_file = 1,
893 .async_size = sizeof(struct io_async_rw),
895 [IORING_OP_POLL_ADD] = {
897 .unbound_nonreg_file = 1,
899 [IORING_OP_POLL_REMOVE] = {},
900 [IORING_OP_SYNC_FILE_RANGE] = {
903 [IORING_OP_SENDMSG] = {
905 .unbound_nonreg_file = 1,
907 .needs_async_data = 1,
908 .async_size = sizeof(struct io_async_msghdr),
910 [IORING_OP_RECVMSG] = {
912 .unbound_nonreg_file = 1,
915 .needs_async_data = 1,
916 .async_size = sizeof(struct io_async_msghdr),
918 [IORING_OP_TIMEOUT] = {
919 .needs_async_data = 1,
920 .async_size = sizeof(struct io_timeout_data),
922 [IORING_OP_TIMEOUT_REMOVE] = {
923 /* used by timeout updates' prep() */
925 [IORING_OP_ACCEPT] = {
927 .unbound_nonreg_file = 1,
930 [IORING_OP_ASYNC_CANCEL] = {},
931 [IORING_OP_LINK_TIMEOUT] = {
932 .needs_async_data = 1,
933 .async_size = sizeof(struct io_timeout_data),
935 [IORING_OP_CONNECT] = {
937 .unbound_nonreg_file = 1,
939 .needs_async_data = 1,
940 .async_size = sizeof(struct io_async_connect),
942 [IORING_OP_FALLOCATE] = {
945 [IORING_OP_OPENAT] = {},
946 [IORING_OP_CLOSE] = {},
947 [IORING_OP_FILES_UPDATE] = {},
948 [IORING_OP_STATX] = {},
951 .unbound_nonreg_file = 1,
955 .async_size = sizeof(struct io_async_rw),
957 [IORING_OP_WRITE] = {
959 .unbound_nonreg_file = 1,
962 .async_size = sizeof(struct io_async_rw),
964 [IORING_OP_FADVISE] = {
967 [IORING_OP_MADVISE] = {},
970 .unbound_nonreg_file = 1,
975 .unbound_nonreg_file = 1,
979 [IORING_OP_OPENAT2] = {
981 [IORING_OP_EPOLL_CTL] = {
982 .unbound_nonreg_file = 1,
984 [IORING_OP_SPLICE] = {
987 .unbound_nonreg_file = 1,
989 [IORING_OP_PROVIDE_BUFFERS] = {},
990 [IORING_OP_REMOVE_BUFFERS] = {},
994 .unbound_nonreg_file = 1,
996 [IORING_OP_SHUTDOWN] = {
999 [IORING_OP_RENAMEAT] = {},
1000 [IORING_OP_UNLINKAT] = {},
1003 static bool io_disarm_next(struct io_kiocb *req);
1004 static void io_uring_del_task_file(unsigned long index);
1005 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1006 struct task_struct *task,
1007 struct files_struct *files);
1008 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1009 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1010 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1011 struct io_ring_ctx *ctx);
1012 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1014 static bool io_rw_reissue(struct io_kiocb *req);
1015 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1016 static void io_put_req(struct io_kiocb *req);
1017 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1018 static void io_double_put_req(struct io_kiocb *req);
1019 static void io_dismantle_req(struct io_kiocb *req);
1020 static void io_put_task(struct task_struct *task, int nr);
1021 static void io_queue_next(struct io_kiocb *req);
1022 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1023 static void __io_queue_linked_timeout(struct io_kiocb *req);
1024 static void io_queue_linked_timeout(struct io_kiocb *req);
1025 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1026 struct io_uring_rsrc_update *ip,
1028 static void __io_clean_op(struct io_kiocb *req);
1029 static struct file *io_file_get(struct io_submit_state *state,
1030 struct io_kiocb *req, int fd, bool fixed);
1031 static void __io_queue_sqe(struct io_kiocb *req);
1032 static void io_rsrc_put_work(struct work_struct *work);
1034 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1035 struct iov_iter *iter, bool needs_lock);
1036 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1037 const struct iovec *fast_iov,
1038 struct iov_iter *iter, bool force);
1039 static void io_req_task_queue(struct io_kiocb *req);
1040 static void io_submit_flush_completions(struct io_comp_state *cs,
1041 struct io_ring_ctx *ctx);
1043 static struct kmem_cache *req_cachep;
1045 static const struct file_operations io_uring_fops;
1047 struct sock *io_uring_get_socket(struct file *file)
1049 #if defined(CONFIG_UNIX)
1050 if (file->f_op == &io_uring_fops) {
1051 struct io_ring_ctx *ctx = file->private_data;
1053 return ctx->ring_sock->sk;
1058 EXPORT_SYMBOL(io_uring_get_socket);
1060 #define io_for_each_link(pos, head) \
1061 for (pos = (head); pos; pos = pos->link)
1063 static inline void io_clean_op(struct io_kiocb *req)
1065 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1069 static inline void io_set_resource_node(struct io_kiocb *req)
1071 struct io_ring_ctx *ctx = req->ctx;
1073 if (!req->fixed_rsrc_refs) {
1074 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1075 percpu_ref_get(req->fixed_rsrc_refs);
1079 static bool io_match_task(struct io_kiocb *head,
1080 struct task_struct *task,
1081 struct files_struct *files)
1083 struct io_kiocb *req;
1085 if (task && head->task != task) {
1086 /* in terms of cancelation, always match if req task is dead */
1087 if (head->task->flags & PF_EXITING)
1094 io_for_each_link(req, head) {
1095 if (req->flags & REQ_F_INFLIGHT)
1097 if (req->task->files == files)
1103 static inline void req_set_fail_links(struct io_kiocb *req)
1105 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1106 req->flags |= REQ_F_FAIL_LINK;
1109 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1111 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1113 complete(&ctx->ref_comp);
1116 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1118 return !req->timeout.off;
1121 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1123 struct io_ring_ctx *ctx;
1126 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1131 * Use 5 bits less than the max cq entries, that should give us around
1132 * 32 entries per hash list if totally full and uniformly spread.
1134 hash_bits = ilog2(p->cq_entries);
1138 ctx->cancel_hash_bits = hash_bits;
1139 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1141 if (!ctx->cancel_hash)
1143 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1145 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1146 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1149 ctx->flags = p->flags;
1150 init_waitqueue_head(&ctx->sqo_sq_wait);
1151 INIT_LIST_HEAD(&ctx->sqd_list);
1152 init_waitqueue_head(&ctx->cq_wait);
1153 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1154 init_completion(&ctx->ref_comp);
1155 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1156 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1157 mutex_init(&ctx->uring_lock);
1158 init_waitqueue_head(&ctx->wait);
1159 spin_lock_init(&ctx->completion_lock);
1160 INIT_LIST_HEAD(&ctx->iopoll_list);
1161 INIT_LIST_HEAD(&ctx->defer_list);
1162 INIT_LIST_HEAD(&ctx->timeout_list);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 spin_lock_init(&ctx->rsrc_ref_lock);
1166 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1167 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1168 init_llist_head(&ctx->rsrc_put_llist);
1169 INIT_LIST_HEAD(&ctx->tctx_list);
1170 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1174 kfree(ctx->cancel_hash);
1179 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1181 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1182 struct io_ring_ctx *ctx = req->ctx;
1184 return seq != ctx->cached_cq_tail
1185 + READ_ONCE(ctx->cached_cq_overflow);
1191 static void io_req_track_inflight(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1195 if (!(req->flags & REQ_F_INFLIGHT)) {
1196 req->flags |= REQ_F_INFLIGHT;
1198 spin_lock_irq(&ctx->inflight_lock);
1199 list_add(&req->inflight_entry, &ctx->inflight_list);
1200 spin_unlock_irq(&ctx->inflight_lock);
1204 static void io_prep_async_work(struct io_kiocb *req)
1206 const struct io_op_def *def = &io_op_defs[req->opcode];
1207 struct io_ring_ctx *ctx = req->ctx;
1209 if (!req->work.creds)
1210 req->work.creds = get_current_cred();
1212 if (req->flags & REQ_F_FORCE_ASYNC)
1213 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1215 if (req->flags & REQ_F_ISREG) {
1216 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1217 io_wq_hash_work(&req->work, file_inode(req->file));
1219 if (def->unbound_nonreg_file)
1220 req->work.flags |= IO_WQ_WORK_UNBOUND;
1224 static void io_prep_async_link(struct io_kiocb *req)
1226 struct io_kiocb *cur;
1228 io_for_each_link(cur, req)
1229 io_prep_async_work(cur);
1232 static void io_queue_async_work(struct io_kiocb *req)
1234 struct io_ring_ctx *ctx = req->ctx;
1235 struct io_kiocb *link = io_prep_linked_timeout(req);
1236 struct io_uring_task *tctx = req->task->io_uring;
1239 BUG_ON(!tctx->io_wq);
1241 /* init ->work of the whole link before punting */
1242 io_prep_async_link(req);
1243 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1244 &req->work, req->flags);
1245 io_wq_enqueue(tctx->io_wq, &req->work);
1247 io_queue_linked_timeout(link);
1250 static void io_kill_timeout(struct io_kiocb *req, int status)
1252 struct io_timeout_data *io = req->async_data;
1255 ret = hrtimer_try_to_cancel(&io->timer);
1257 atomic_set(&req->ctx->cq_timeouts,
1258 atomic_read(&req->ctx->cq_timeouts) + 1);
1259 list_del_init(&req->timeout.list);
1260 io_cqring_fill_event(req, status);
1261 io_put_req_deferred(req, 1);
1265 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1268 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1269 struct io_defer_entry, list);
1271 if (req_need_defer(de->req, de->seq))
1273 list_del_init(&de->list);
1274 io_req_task_queue(de->req);
1276 } while (!list_empty(&ctx->defer_list));
1279 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1283 if (list_empty(&ctx->timeout_list))
1286 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1289 u32 events_needed, events_got;
1290 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1291 struct io_kiocb, timeout.list);
1293 if (io_is_timeout_noseq(req))
1297 * Since seq can easily wrap around over time, subtract
1298 * the last seq at which timeouts were flushed before comparing.
1299 * Assuming not more than 2^31-1 events have happened since,
1300 * these subtractions won't have wrapped, so we can check if
1301 * target is in [last_seq, current_seq] by comparing the two.
1303 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1304 events_got = seq - ctx->cq_last_tm_flush;
1305 if (events_got < events_needed)
1308 list_del_init(&req->timeout.list);
1309 io_kill_timeout(req, 0);
1310 } while (!list_empty(&ctx->timeout_list));
1312 ctx->cq_last_tm_flush = seq;
1315 static void io_commit_cqring(struct io_ring_ctx *ctx)
1317 io_flush_timeouts(ctx);
1319 /* order cqe stores with ring update */
1320 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1322 if (unlikely(!list_empty(&ctx->defer_list)))
1323 __io_queue_deferred(ctx);
1326 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1328 struct io_rings *r = ctx->rings;
1330 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1333 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1335 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1338 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1340 struct io_rings *rings = ctx->rings;
1344 * writes to the cq entry need to come after reading head; the
1345 * control dependency is enough as we're using WRITE_ONCE to
1348 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1351 tail = ctx->cached_cq_tail++;
1352 return &rings->cqes[tail & ctx->cq_mask];
1355 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1359 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1361 if (!ctx->eventfd_async)
1363 return io_wq_current_is_worker();
1366 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1368 /* see waitqueue_active() comment */
1371 if (waitqueue_active(&ctx->wait))
1372 wake_up(&ctx->wait);
1373 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1374 wake_up(&ctx->sq_data->wait);
1375 if (io_should_trigger_evfd(ctx))
1376 eventfd_signal(ctx->cq_ev_fd, 1);
1377 if (waitqueue_active(&ctx->cq_wait)) {
1378 wake_up_interruptible(&ctx->cq_wait);
1379 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1383 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1385 /* see waitqueue_active() comment */
1388 if (ctx->flags & IORING_SETUP_SQPOLL) {
1389 if (waitqueue_active(&ctx->wait))
1390 wake_up(&ctx->wait);
1392 if (io_should_trigger_evfd(ctx))
1393 eventfd_signal(ctx->cq_ev_fd, 1);
1394 if (waitqueue_active(&ctx->cq_wait)) {
1395 wake_up_interruptible(&ctx->cq_wait);
1396 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1400 /* Returns true if there are no backlogged entries after the flush */
1401 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1402 struct task_struct *tsk,
1403 struct files_struct *files)
1405 struct io_rings *rings = ctx->rings;
1406 struct io_kiocb *req, *tmp;
1407 struct io_uring_cqe *cqe;
1408 unsigned long flags;
1409 bool all_flushed, posted;
1412 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1416 spin_lock_irqsave(&ctx->completion_lock, flags);
1417 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1418 if (!io_match_task(req, tsk, files))
1421 cqe = io_get_cqring(ctx);
1425 list_move(&req->compl.list, &list);
1427 WRITE_ONCE(cqe->user_data, req->user_data);
1428 WRITE_ONCE(cqe->res, req->result);
1429 WRITE_ONCE(cqe->flags, req->compl.cflags);
1431 ctx->cached_cq_overflow++;
1432 WRITE_ONCE(ctx->rings->cq_overflow,
1433 ctx->cached_cq_overflow);
1438 all_flushed = list_empty(&ctx->cq_overflow_list);
1440 clear_bit(0, &ctx->sq_check_overflow);
1441 clear_bit(0, &ctx->cq_check_overflow);
1442 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1446 io_commit_cqring(ctx);
1447 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1449 io_cqring_ev_posted(ctx);
1451 while (!list_empty(&list)) {
1452 req = list_first_entry(&list, struct io_kiocb, compl.list);
1453 list_del(&req->compl.list);
1460 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1461 struct task_struct *tsk,
1462 struct files_struct *files)
1466 if (test_bit(0, &ctx->cq_check_overflow)) {
1467 /* iopoll syncs against uring_lock, not completion_lock */
1468 if (ctx->flags & IORING_SETUP_IOPOLL)
1469 mutex_lock(&ctx->uring_lock);
1470 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1471 if (ctx->flags & IORING_SETUP_IOPOLL)
1472 mutex_unlock(&ctx->uring_lock);
1478 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1480 struct io_ring_ctx *ctx = req->ctx;
1481 struct io_uring_cqe *cqe;
1483 trace_io_uring_complete(ctx, req->user_data, res);
1486 * If we can't get a cq entry, userspace overflowed the
1487 * submission (by quite a lot). Increment the overflow count in
1490 cqe = io_get_cqring(ctx);
1492 WRITE_ONCE(cqe->user_data, req->user_data);
1493 WRITE_ONCE(cqe->res, res);
1494 WRITE_ONCE(cqe->flags, cflags);
1495 } else if (ctx->cq_overflow_flushed ||
1496 atomic_read(&req->task->io_uring->in_idle)) {
1498 * If we're in ring overflow flush mode, or in task cancel mode,
1499 * then we cannot store the request for later flushing, we need
1500 * to drop it on the floor.
1502 ctx->cached_cq_overflow++;
1503 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1505 if (list_empty(&ctx->cq_overflow_list)) {
1506 set_bit(0, &ctx->sq_check_overflow);
1507 set_bit(0, &ctx->cq_check_overflow);
1508 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1512 req->compl.cflags = cflags;
1513 refcount_inc(&req->refs);
1514 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1518 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1520 __io_cqring_fill_event(req, res, 0);
1523 static void io_req_complete_post(struct io_kiocb *req, long res,
1524 unsigned int cflags)
1526 struct io_ring_ctx *ctx = req->ctx;
1527 unsigned long flags;
1529 spin_lock_irqsave(&ctx->completion_lock, flags);
1530 __io_cqring_fill_event(req, res, cflags);
1532 * If we're the last reference to this request, add to our locked
1535 if (refcount_dec_and_test(&req->refs)) {
1536 struct io_comp_state *cs = &ctx->submit_state.comp;
1538 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1539 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1540 io_disarm_next(req);
1542 io_req_task_queue(req->link);
1546 io_dismantle_req(req);
1547 io_put_task(req->task, 1);
1548 list_add(&req->compl.list, &cs->locked_free_list);
1549 cs->locked_free_nr++;
1551 if (!percpu_ref_tryget(&ctx->refs))
1554 io_commit_cqring(ctx);
1555 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1558 io_cqring_ev_posted(ctx);
1559 percpu_ref_put(&ctx->refs);
1563 static void io_req_complete_state(struct io_kiocb *req, long res,
1564 unsigned int cflags)
1568 req->compl.cflags = cflags;
1569 req->flags |= REQ_F_COMPLETE_INLINE;
1572 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1573 long res, unsigned cflags)
1575 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1576 io_req_complete_state(req, res, cflags);
1578 io_req_complete_post(req, res, cflags);
1581 static inline void io_req_complete(struct io_kiocb *req, long res)
1583 __io_req_complete(req, 0, res, 0);
1586 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1588 struct io_submit_state *state = &ctx->submit_state;
1589 struct io_comp_state *cs = &state->comp;
1590 struct io_kiocb *req = NULL;
1593 * If we have more than a batch's worth of requests in our IRQ side
1594 * locked cache, grab the lock and move them over to our submission
1597 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1598 spin_lock_irq(&ctx->completion_lock);
1599 list_splice_init(&cs->locked_free_list, &cs->free_list);
1600 cs->locked_free_nr = 0;
1601 spin_unlock_irq(&ctx->completion_lock);
1604 while (!list_empty(&cs->free_list)) {
1605 req = list_first_entry(&cs->free_list, struct io_kiocb,
1607 list_del(&req->compl.list);
1608 state->reqs[state->free_reqs++] = req;
1609 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1616 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1618 struct io_submit_state *state = &ctx->submit_state;
1620 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1622 if (!state->free_reqs) {
1623 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1626 if (io_flush_cached_reqs(ctx))
1629 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1633 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1634 * retry single alloc to be on the safe side.
1636 if (unlikely(ret <= 0)) {
1637 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1638 if (!state->reqs[0])
1642 state->free_reqs = ret;
1646 return state->reqs[state->free_reqs];
1649 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1656 static void io_dismantle_req(struct io_kiocb *req)
1660 if (req->async_data)
1661 kfree(req->async_data);
1663 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1664 if (req->fixed_rsrc_refs)
1665 percpu_ref_put(req->fixed_rsrc_refs);
1666 if (req->work.creds) {
1667 put_cred(req->work.creds);
1668 req->work.creds = NULL;
1671 if (req->flags & REQ_F_INFLIGHT) {
1672 struct io_ring_ctx *ctx = req->ctx;
1673 unsigned long flags;
1675 spin_lock_irqsave(&ctx->inflight_lock, flags);
1676 list_del(&req->inflight_entry);
1677 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1678 req->flags &= ~REQ_F_INFLIGHT;
1682 /* must to be called somewhat shortly after putting a request */
1683 static inline void io_put_task(struct task_struct *task, int nr)
1685 struct io_uring_task *tctx = task->io_uring;
1687 percpu_counter_sub(&tctx->inflight, nr);
1688 if (unlikely(atomic_read(&tctx->in_idle)))
1689 wake_up(&tctx->wait);
1690 put_task_struct_many(task, nr);
1693 static void __io_free_req(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1697 io_dismantle_req(req);
1698 io_put_task(req->task, 1);
1700 kmem_cache_free(req_cachep, req);
1701 percpu_ref_put(&ctx->refs);
1704 static inline void io_remove_next_linked(struct io_kiocb *req)
1706 struct io_kiocb *nxt = req->link;
1708 req->link = nxt->link;
1712 static bool io_kill_linked_timeout(struct io_kiocb *req)
1713 __must_hold(&req->ctx->completion_lock)
1715 struct io_kiocb *link = req->link;
1716 bool cancelled = false;
1719 * Can happen if a linked timeout fired and link had been like
1720 * req -> link t-out -> link t-out [-> ...]
1722 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1723 struct io_timeout_data *io = link->async_data;
1726 io_remove_next_linked(req);
1727 link->timeout.head = NULL;
1728 ret = hrtimer_try_to_cancel(&io->timer);
1730 io_cqring_fill_event(link, -ECANCELED);
1731 io_put_req_deferred(link, 1);
1735 req->flags &= ~REQ_F_LINK_TIMEOUT;
1739 static void io_fail_links(struct io_kiocb *req)
1740 __must_hold(&req->ctx->completion_lock)
1742 struct io_kiocb *nxt, *link = req->link;
1749 trace_io_uring_fail_link(req, link);
1750 io_cqring_fill_event(link, -ECANCELED);
1751 io_put_req_deferred(link, 2);
1756 static bool io_disarm_next(struct io_kiocb *req)
1757 __must_hold(&req->ctx->completion_lock)
1759 bool posted = false;
1761 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1762 posted = io_kill_linked_timeout(req);
1763 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1764 posted |= (req->link != NULL);
1770 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1772 struct io_kiocb *nxt;
1775 * If LINK is set, we have dependent requests in this chain. If we
1776 * didn't fail this request, queue the first one up, moving any other
1777 * dependencies to the next request. In case of failure, fail the rest
1780 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1781 struct io_ring_ctx *ctx = req->ctx;
1782 unsigned long flags;
1785 spin_lock_irqsave(&ctx->completion_lock, flags);
1786 posted = io_disarm_next(req);
1788 io_commit_cqring(req->ctx);
1789 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1791 io_cqring_ev_posted(ctx);
1798 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1800 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1802 return __io_req_find_next(req);
1805 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1809 if (ctx->submit_state.comp.nr) {
1810 mutex_lock(&ctx->uring_lock);
1811 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1812 mutex_unlock(&ctx->uring_lock);
1814 percpu_ref_put(&ctx->refs);
1817 static bool __tctx_task_work(struct io_uring_task *tctx)
1819 struct io_ring_ctx *ctx = NULL;
1820 struct io_wq_work_list list;
1821 struct io_wq_work_node *node;
1823 if (wq_list_empty(&tctx->task_list))
1826 spin_lock_irq(&tctx->task_lock);
1827 list = tctx->task_list;
1828 INIT_WQ_LIST(&tctx->task_list);
1829 spin_unlock_irq(&tctx->task_lock);
1833 struct io_wq_work_node *next = node->next;
1834 struct io_kiocb *req;
1836 req = container_of(node, struct io_kiocb, io_task_work.node);
1837 if (req->ctx != ctx) {
1838 ctx_flush_and_put(ctx);
1840 percpu_ref_get(&ctx->refs);
1843 req->task_work.func(&req->task_work);
1847 ctx_flush_and_put(ctx);
1848 return list.first != NULL;
1851 static void tctx_task_work(struct callback_head *cb)
1853 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1855 clear_bit(0, &tctx->task_state);
1857 while (__tctx_task_work(tctx))
1861 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1862 enum task_work_notify_mode notify)
1864 struct io_uring_task *tctx = tsk->io_uring;
1865 struct io_wq_work_node *node, *prev;
1866 unsigned long flags;
1869 WARN_ON_ONCE(!tctx);
1871 spin_lock_irqsave(&tctx->task_lock, flags);
1872 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1873 spin_unlock_irqrestore(&tctx->task_lock, flags);
1875 /* task_work already pending, we're done */
1876 if (test_bit(0, &tctx->task_state) ||
1877 test_and_set_bit(0, &tctx->task_state))
1880 if (!task_work_add(tsk, &tctx->task_work, notify))
1884 * Slow path - we failed, find and delete work. if the work is not
1885 * in the list, it got run and we're fine.
1888 spin_lock_irqsave(&tctx->task_lock, flags);
1889 wq_list_for_each(node, prev, &tctx->task_list) {
1890 if (&req->io_task_work.node == node) {
1891 wq_list_del(&tctx->task_list, node, prev);
1896 spin_unlock_irqrestore(&tctx->task_lock, flags);
1897 clear_bit(0, &tctx->task_state);
1901 static int io_req_task_work_add(struct io_kiocb *req)
1903 struct task_struct *tsk = req->task;
1904 struct io_ring_ctx *ctx = req->ctx;
1905 enum task_work_notify_mode notify;
1908 if (tsk->flags & PF_EXITING)
1912 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1913 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1914 * processing task_work. There's no reliable way to tell if TWA_RESUME
1918 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1919 notify = TWA_SIGNAL;
1921 ret = io_task_work_add(tsk, req, notify);
1923 wake_up_process(tsk);
1928 static bool io_run_task_work_head(struct callback_head **work_head)
1930 struct callback_head *work, *next;
1931 bool executed = false;
1934 work = xchg(work_head, NULL);
1950 static void io_task_work_add_head(struct callback_head **work_head,
1951 struct callback_head *task_work)
1953 struct callback_head *head;
1956 head = READ_ONCE(*work_head);
1957 task_work->next = head;
1958 } while (cmpxchg(work_head, head, task_work) != head);
1961 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1962 task_work_func_t cb)
1964 init_task_work(&req->task_work, cb);
1965 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1968 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1970 struct io_ring_ctx *ctx = req->ctx;
1972 spin_lock_irq(&ctx->completion_lock);
1973 io_cqring_fill_event(req, error);
1974 io_commit_cqring(ctx);
1975 spin_unlock_irq(&ctx->completion_lock);
1977 io_cqring_ev_posted(ctx);
1978 req_set_fail_links(req);
1979 io_double_put_req(req);
1982 static void io_req_task_cancel(struct callback_head *cb)
1984 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1985 struct io_ring_ctx *ctx = req->ctx;
1987 mutex_lock(&ctx->uring_lock);
1988 __io_req_task_cancel(req, req->result);
1989 mutex_unlock(&ctx->uring_lock);
1990 percpu_ref_put(&ctx->refs);
1993 static void __io_req_task_submit(struct io_kiocb *req)
1995 struct io_ring_ctx *ctx = req->ctx;
1997 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1998 mutex_lock(&ctx->uring_lock);
1999 if (!(current->flags & PF_EXITING) && !current->in_execve)
2000 __io_queue_sqe(req);
2002 __io_req_task_cancel(req, -EFAULT);
2003 mutex_unlock(&ctx->uring_lock);
2006 static void io_req_task_submit(struct callback_head *cb)
2008 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2010 __io_req_task_submit(req);
2013 static void io_req_task_queue(struct io_kiocb *req)
2017 req->task_work.func = io_req_task_submit;
2018 ret = io_req_task_work_add(req);
2019 if (unlikely(ret)) {
2020 req->result = -ECANCELED;
2021 percpu_ref_get(&req->ctx->refs);
2022 io_req_task_work_add_fallback(req, io_req_task_cancel);
2026 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2028 percpu_ref_get(&req->ctx->refs);
2030 req->task_work.func = io_req_task_cancel;
2032 if (unlikely(io_req_task_work_add(req)))
2033 io_req_task_work_add_fallback(req, io_req_task_cancel);
2036 static inline void io_queue_next(struct io_kiocb *req)
2038 struct io_kiocb *nxt = io_req_find_next(req);
2041 io_req_task_queue(nxt);
2044 static void io_free_req(struct io_kiocb *req)
2051 struct task_struct *task;
2056 static inline void io_init_req_batch(struct req_batch *rb)
2063 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2064 struct req_batch *rb)
2067 io_put_task(rb->task, rb->task_refs);
2069 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2072 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2073 struct io_submit_state *state)
2077 if (req->task != rb->task) {
2079 io_put_task(rb->task, rb->task_refs);
2080 rb->task = req->task;
2086 io_dismantle_req(req);
2087 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2088 state->reqs[state->free_reqs++] = req;
2090 list_add(&req->compl.list, &state->comp.free_list);
2093 static void io_submit_flush_completions(struct io_comp_state *cs,
2094 struct io_ring_ctx *ctx)
2097 struct io_kiocb *req;
2098 struct req_batch rb;
2100 io_init_req_batch(&rb);
2101 spin_lock_irq(&ctx->completion_lock);
2102 for (i = 0; i < nr; i++) {
2104 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2106 io_commit_cqring(ctx);
2107 spin_unlock_irq(&ctx->completion_lock);
2109 io_cqring_ev_posted(ctx);
2110 for (i = 0; i < nr; i++) {
2113 /* submission and completion refs */
2114 if (refcount_sub_and_test(2, &req->refs))
2115 io_req_free_batch(&rb, req, &ctx->submit_state);
2118 io_req_free_batch_finish(ctx, &rb);
2123 * Drop reference to request, return next in chain (if there is one) if this
2124 * was the last reference to this request.
2126 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2128 struct io_kiocb *nxt = NULL;
2130 if (refcount_dec_and_test(&req->refs)) {
2131 nxt = io_req_find_next(req);
2137 static void io_put_req(struct io_kiocb *req)
2139 if (refcount_dec_and_test(&req->refs))
2143 static void io_put_req_deferred_cb(struct callback_head *cb)
2145 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2150 static void io_free_req_deferred(struct io_kiocb *req)
2154 req->task_work.func = io_put_req_deferred_cb;
2155 ret = io_req_task_work_add(req);
2157 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2160 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2162 if (refcount_sub_and_test(refs, &req->refs))
2163 io_free_req_deferred(req);
2166 static void io_double_put_req(struct io_kiocb *req)
2168 /* drop both submit and complete references */
2169 if (refcount_sub_and_test(2, &req->refs))
2173 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2175 /* See comment at the top of this file */
2177 return __io_cqring_events(ctx);
2180 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2182 struct io_rings *rings = ctx->rings;
2184 /* make sure SQ entry isn't read before tail */
2185 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2188 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2190 unsigned int cflags;
2192 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2193 cflags |= IORING_CQE_F_BUFFER;
2194 req->flags &= ~REQ_F_BUFFER_SELECTED;
2199 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2201 struct io_buffer *kbuf;
2203 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2204 return io_put_kbuf(req, kbuf);
2207 static inline bool io_run_task_work(void)
2210 * Not safe to run on exiting task, and the task_work handling will
2211 * not add work to such a task.
2213 if (unlikely(current->flags & PF_EXITING))
2215 if (current->task_works) {
2216 __set_current_state(TASK_RUNNING);
2225 * Find and free completed poll iocbs
2227 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2228 struct list_head *done)
2230 struct req_batch rb;
2231 struct io_kiocb *req;
2233 /* order with ->result store in io_complete_rw_iopoll() */
2236 io_init_req_batch(&rb);
2237 while (!list_empty(done)) {
2240 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2241 list_del(&req->inflight_entry);
2243 if (READ_ONCE(req->result) == -EAGAIN) {
2244 req->iopoll_completed = 0;
2245 if (io_rw_reissue(req))
2249 if (req->flags & REQ_F_BUFFER_SELECTED)
2250 cflags = io_put_rw_kbuf(req);
2252 __io_cqring_fill_event(req, req->result, cflags);
2255 if (refcount_dec_and_test(&req->refs))
2256 io_req_free_batch(&rb, req, &ctx->submit_state);
2259 io_commit_cqring(ctx);
2260 io_cqring_ev_posted_iopoll(ctx);
2261 io_req_free_batch_finish(ctx, &rb);
2264 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2267 struct io_kiocb *req, *tmp;
2273 * Only spin for completions if we don't have multiple devices hanging
2274 * off our complete list, and we're under the requested amount.
2276 spin = !ctx->poll_multi_file && *nr_events < min;
2279 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2280 struct kiocb *kiocb = &req->rw.kiocb;
2283 * Move completed and retryable entries to our local lists.
2284 * If we find a request that requires polling, break out
2285 * and complete those lists first, if we have entries there.
2287 if (READ_ONCE(req->iopoll_completed)) {
2288 list_move_tail(&req->inflight_entry, &done);
2291 if (!list_empty(&done))
2294 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2298 /* iopoll may have completed current req */
2299 if (READ_ONCE(req->iopoll_completed))
2300 list_move_tail(&req->inflight_entry, &done);
2307 if (!list_empty(&done))
2308 io_iopoll_complete(ctx, nr_events, &done);
2314 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2315 * non-spinning poll check - we'll still enter the driver poll loop, but only
2316 * as a non-spinning completion check.
2318 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2321 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2324 ret = io_do_iopoll(ctx, nr_events, min);
2327 if (*nr_events >= min)
2335 * We can't just wait for polled events to come to us, we have to actively
2336 * find and complete them.
2338 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2340 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2343 mutex_lock(&ctx->uring_lock);
2344 while (!list_empty(&ctx->iopoll_list)) {
2345 unsigned int nr_events = 0;
2347 io_do_iopoll(ctx, &nr_events, 0);
2349 /* let it sleep and repeat later if can't complete a request */
2353 * Ensure we allow local-to-the-cpu processing to take place,
2354 * in this case we need to ensure that we reap all events.
2355 * Also let task_work, etc. to progress by releasing the mutex
2357 if (need_resched()) {
2358 mutex_unlock(&ctx->uring_lock);
2360 mutex_lock(&ctx->uring_lock);
2363 mutex_unlock(&ctx->uring_lock);
2366 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2368 unsigned int nr_events = 0;
2369 int iters = 0, ret = 0;
2372 * We disallow the app entering submit/complete with polling, but we
2373 * still need to lock the ring to prevent racing with polled issue
2374 * that got punted to a workqueue.
2376 mutex_lock(&ctx->uring_lock);
2379 * Don't enter poll loop if we already have events pending.
2380 * If we do, we can potentially be spinning for commands that
2381 * already triggered a CQE (eg in error).
2383 if (test_bit(0, &ctx->cq_check_overflow))
2384 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2385 if (io_cqring_events(ctx))
2389 * If a submit got punted to a workqueue, we can have the
2390 * application entering polling for a command before it gets
2391 * issued. That app will hold the uring_lock for the duration
2392 * of the poll right here, so we need to take a breather every
2393 * now and then to ensure that the issue has a chance to add
2394 * the poll to the issued list. Otherwise we can spin here
2395 * forever, while the workqueue is stuck trying to acquire the
2398 if (!(++iters & 7)) {
2399 mutex_unlock(&ctx->uring_lock);
2401 mutex_lock(&ctx->uring_lock);
2404 ret = io_iopoll_getevents(ctx, &nr_events, min);
2408 } while (min && !nr_events && !need_resched());
2410 mutex_unlock(&ctx->uring_lock);
2414 static void kiocb_end_write(struct io_kiocb *req)
2417 * Tell lockdep we inherited freeze protection from submission
2420 if (req->flags & REQ_F_ISREG) {
2421 struct inode *inode = file_inode(req->file);
2423 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2425 file_end_write(req->file);
2429 static bool io_resubmit_prep(struct io_kiocb *req)
2431 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2433 struct iov_iter iter;
2435 /* already prepared */
2436 if (req->async_data)
2439 switch (req->opcode) {
2440 case IORING_OP_READV:
2441 case IORING_OP_READ_FIXED:
2442 case IORING_OP_READ:
2445 case IORING_OP_WRITEV:
2446 case IORING_OP_WRITE_FIXED:
2447 case IORING_OP_WRITE:
2451 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2456 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2459 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2462 static bool io_rw_should_reissue(struct io_kiocb *req)
2464 umode_t mode = file_inode(req->file)->i_mode;
2465 struct io_ring_ctx *ctx = req->ctx;
2467 if (!S_ISBLK(mode) && !S_ISREG(mode))
2469 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2470 !(ctx->flags & IORING_SETUP_IOPOLL)))
2473 * If ref is dying, we might be running poll reap from the exit work.
2474 * Don't attempt to reissue from that path, just let it fail with
2477 if (percpu_ref_is_dying(&ctx->refs))
2483 static bool io_rw_reissue(struct io_kiocb *req)
2486 if (!io_rw_should_reissue(req))
2489 lockdep_assert_held(&req->ctx->uring_lock);
2491 if (io_resubmit_prep(req)) {
2492 refcount_inc(&req->refs);
2493 io_queue_async_work(req);
2496 req_set_fail_links(req);
2501 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2502 unsigned int issue_flags)
2506 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2507 kiocb_end_write(req);
2508 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2510 if (res != req->result)
2511 req_set_fail_links(req);
2512 if (req->flags & REQ_F_BUFFER_SELECTED)
2513 cflags = io_put_rw_kbuf(req);
2514 __io_req_complete(req, issue_flags, res, cflags);
2517 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2519 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2521 __io_complete_rw(req, res, res2, 0);
2524 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2526 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2529 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2530 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2531 struct io_async_rw *rw = req->async_data;
2534 iov_iter_revert(&rw->iter,
2535 req->result - iov_iter_count(&rw->iter));
2536 else if (!io_resubmit_prep(req))
2541 if (kiocb->ki_flags & IOCB_WRITE)
2542 kiocb_end_write(req);
2544 if (res != -EAGAIN && res != req->result)
2545 req_set_fail_links(req);
2547 WRITE_ONCE(req->result, res);
2548 /* order with io_poll_complete() checking ->result */
2550 WRITE_ONCE(req->iopoll_completed, 1);
2554 * After the iocb has been issued, it's safe to be found on the poll list.
2555 * Adding the kiocb to the list AFTER submission ensures that we don't
2556 * find it from a io_iopoll_getevents() thread before the issuer is done
2557 * accessing the kiocb cookie.
2559 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2561 struct io_ring_ctx *ctx = req->ctx;
2564 * Track whether we have multiple files in our lists. This will impact
2565 * how we do polling eventually, not spinning if we're on potentially
2566 * different devices.
2568 if (list_empty(&ctx->iopoll_list)) {
2569 ctx->poll_multi_file = false;
2570 } else if (!ctx->poll_multi_file) {
2571 struct io_kiocb *list_req;
2573 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2575 if (list_req->file != req->file)
2576 ctx->poll_multi_file = true;
2580 * For fast devices, IO may have already completed. If it has, add
2581 * it to the front so we find it first.
2583 if (READ_ONCE(req->iopoll_completed))
2584 list_add(&req->inflight_entry, &ctx->iopoll_list);
2586 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2589 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2590 * task context or in io worker task context. If current task context is
2591 * sq thread, we don't need to check whether should wake up sq thread.
2593 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2594 wq_has_sleeper(&ctx->sq_data->wait))
2595 wake_up(&ctx->sq_data->wait);
2598 static inline void io_state_file_put(struct io_submit_state *state)
2600 if (state->file_refs) {
2601 fput_many(state->file, state->file_refs);
2602 state->file_refs = 0;
2607 * Get as many references to a file as we have IOs left in this submission,
2608 * assuming most submissions are for one file, or at least that each file
2609 * has more than one submission.
2611 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2616 if (state->file_refs) {
2617 if (state->fd == fd) {
2621 io_state_file_put(state);
2623 state->file = fget_many(fd, state->ios_left);
2624 if (unlikely(!state->file))
2628 state->file_refs = state->ios_left - 1;
2632 static bool io_bdev_nowait(struct block_device *bdev)
2634 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2638 * If we tracked the file through the SCM inflight mechanism, we could support
2639 * any file. For now, just ensure that anything potentially problematic is done
2642 static bool io_file_supports_async(struct file *file, int rw)
2644 umode_t mode = file_inode(file)->i_mode;
2646 if (S_ISBLK(mode)) {
2647 if (IS_ENABLED(CONFIG_BLOCK) &&
2648 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2652 if (S_ISCHR(mode) || S_ISSOCK(mode))
2654 if (S_ISREG(mode)) {
2655 if (IS_ENABLED(CONFIG_BLOCK) &&
2656 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2657 file->f_op != &io_uring_fops)
2662 /* any ->read/write should understand O_NONBLOCK */
2663 if (file->f_flags & O_NONBLOCK)
2666 if (!(file->f_mode & FMODE_NOWAIT))
2670 return file->f_op->read_iter != NULL;
2672 return file->f_op->write_iter != NULL;
2675 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2677 struct io_ring_ctx *ctx = req->ctx;
2678 struct kiocb *kiocb = &req->rw.kiocb;
2679 struct file *file = req->file;
2683 if (S_ISREG(file_inode(file)->i_mode))
2684 req->flags |= REQ_F_ISREG;
2686 kiocb->ki_pos = READ_ONCE(sqe->off);
2687 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2688 req->flags |= REQ_F_CUR_POS;
2689 kiocb->ki_pos = file->f_pos;
2691 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2692 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2693 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2697 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2698 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2699 req->flags |= REQ_F_NOWAIT;
2701 ioprio = READ_ONCE(sqe->ioprio);
2703 ret = ioprio_check_cap(ioprio);
2707 kiocb->ki_ioprio = ioprio;
2709 kiocb->ki_ioprio = get_current_ioprio();
2711 if (ctx->flags & IORING_SETUP_IOPOLL) {
2712 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2713 !kiocb->ki_filp->f_op->iopoll)
2716 kiocb->ki_flags |= IOCB_HIPRI;
2717 kiocb->ki_complete = io_complete_rw_iopoll;
2718 req->iopoll_completed = 0;
2720 if (kiocb->ki_flags & IOCB_HIPRI)
2722 kiocb->ki_complete = io_complete_rw;
2725 req->rw.addr = READ_ONCE(sqe->addr);
2726 req->rw.len = READ_ONCE(sqe->len);
2727 req->buf_index = READ_ONCE(sqe->buf_index);
2731 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2737 case -ERESTARTNOINTR:
2738 case -ERESTARTNOHAND:
2739 case -ERESTART_RESTARTBLOCK:
2741 * We can't just restart the syscall, since previously
2742 * submitted sqes may already be in progress. Just fail this
2748 kiocb->ki_complete(kiocb, ret, 0);
2752 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2753 unsigned int issue_flags)
2755 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2756 struct io_async_rw *io = req->async_data;
2758 /* add previously done IO, if any */
2759 if (io && io->bytes_done > 0) {
2761 ret = io->bytes_done;
2763 ret += io->bytes_done;
2766 if (req->flags & REQ_F_CUR_POS)
2767 req->file->f_pos = kiocb->ki_pos;
2768 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2769 __io_complete_rw(req, ret, 0, issue_flags);
2771 io_rw_done(kiocb, ret);
2774 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2776 struct io_ring_ctx *ctx = req->ctx;
2777 size_t len = req->rw.len;
2778 struct io_mapped_ubuf *imu;
2779 u16 index, buf_index = req->buf_index;
2783 if (unlikely(buf_index >= ctx->nr_user_bufs))
2785 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2786 imu = &ctx->user_bufs[index];
2787 buf_addr = req->rw.addr;
2790 if (buf_addr + len < buf_addr)
2792 /* not inside the mapped region */
2793 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2797 * May not be a start of buffer, set size appropriately
2798 * and advance us to the beginning.
2800 offset = buf_addr - imu->ubuf;
2801 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2805 * Don't use iov_iter_advance() here, as it's really slow for
2806 * using the latter parts of a big fixed buffer - it iterates
2807 * over each segment manually. We can cheat a bit here, because
2810 * 1) it's a BVEC iter, we set it up
2811 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2812 * first and last bvec
2814 * So just find our index, and adjust the iterator afterwards.
2815 * If the offset is within the first bvec (or the whole first
2816 * bvec, just use iov_iter_advance(). This makes it easier
2817 * since we can just skip the first segment, which may not
2818 * be PAGE_SIZE aligned.
2820 const struct bio_vec *bvec = imu->bvec;
2822 if (offset <= bvec->bv_len) {
2823 iov_iter_advance(iter, offset);
2825 unsigned long seg_skip;
2827 /* skip first vec */
2828 offset -= bvec->bv_len;
2829 seg_skip = 1 + (offset >> PAGE_SHIFT);
2831 iter->bvec = bvec + seg_skip;
2832 iter->nr_segs -= seg_skip;
2833 iter->count -= bvec->bv_len + offset;
2834 iter->iov_offset = offset & ~PAGE_MASK;
2841 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2844 mutex_unlock(&ctx->uring_lock);
2847 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2850 * "Normal" inline submissions always hold the uring_lock, since we
2851 * grab it from the system call. Same is true for the SQPOLL offload.
2852 * The only exception is when we've detached the request and issue it
2853 * from an async worker thread, grab the lock for that case.
2856 mutex_lock(&ctx->uring_lock);
2859 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2860 int bgid, struct io_buffer *kbuf,
2863 struct io_buffer *head;
2865 if (req->flags & REQ_F_BUFFER_SELECTED)
2868 io_ring_submit_lock(req->ctx, needs_lock);
2870 lockdep_assert_held(&req->ctx->uring_lock);
2872 head = xa_load(&req->ctx->io_buffers, bgid);
2874 if (!list_empty(&head->list)) {
2875 kbuf = list_last_entry(&head->list, struct io_buffer,
2877 list_del(&kbuf->list);
2880 xa_erase(&req->ctx->io_buffers, bgid);
2882 if (*len > kbuf->len)
2885 kbuf = ERR_PTR(-ENOBUFS);
2888 io_ring_submit_unlock(req->ctx, needs_lock);
2893 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2896 struct io_buffer *kbuf;
2899 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2900 bgid = req->buf_index;
2901 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2904 req->rw.addr = (u64) (unsigned long) kbuf;
2905 req->flags |= REQ_F_BUFFER_SELECTED;
2906 return u64_to_user_ptr(kbuf->addr);
2909 #ifdef CONFIG_COMPAT
2910 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2913 struct compat_iovec __user *uiov;
2914 compat_ssize_t clen;
2918 uiov = u64_to_user_ptr(req->rw.addr);
2919 if (!access_ok(uiov, sizeof(*uiov)))
2921 if (__get_user(clen, &uiov->iov_len))
2927 buf = io_rw_buffer_select(req, &len, needs_lock);
2929 return PTR_ERR(buf);
2930 iov[0].iov_base = buf;
2931 iov[0].iov_len = (compat_size_t) len;
2936 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2939 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2943 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2946 len = iov[0].iov_len;
2949 buf = io_rw_buffer_select(req, &len, needs_lock);
2951 return PTR_ERR(buf);
2952 iov[0].iov_base = buf;
2953 iov[0].iov_len = len;
2957 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2960 if (req->flags & REQ_F_BUFFER_SELECTED) {
2961 struct io_buffer *kbuf;
2963 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2964 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2965 iov[0].iov_len = kbuf->len;
2968 if (req->rw.len != 1)
2971 #ifdef CONFIG_COMPAT
2972 if (req->ctx->compat)
2973 return io_compat_import(req, iov, needs_lock);
2976 return __io_iov_buffer_select(req, iov, needs_lock);
2979 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2980 struct iov_iter *iter, bool needs_lock)
2982 void __user *buf = u64_to_user_ptr(req->rw.addr);
2983 size_t sqe_len = req->rw.len;
2984 u8 opcode = req->opcode;
2987 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2989 return io_import_fixed(req, rw, iter);
2992 /* buffer index only valid with fixed read/write, or buffer select */
2993 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2996 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2997 if (req->flags & REQ_F_BUFFER_SELECT) {
2998 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3000 return PTR_ERR(buf);
3001 req->rw.len = sqe_len;
3004 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3009 if (req->flags & REQ_F_BUFFER_SELECT) {
3010 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3012 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3017 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3021 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3023 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3027 * For files that don't have ->read_iter() and ->write_iter(), handle them
3028 * by looping over ->read() or ->write() manually.
3030 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3032 struct kiocb *kiocb = &req->rw.kiocb;
3033 struct file *file = req->file;
3037 * Don't support polled IO through this interface, and we can't
3038 * support non-blocking either. For the latter, this just causes
3039 * the kiocb to be handled from an async context.
3041 if (kiocb->ki_flags & IOCB_HIPRI)
3043 if (kiocb->ki_flags & IOCB_NOWAIT)
3046 while (iov_iter_count(iter)) {
3050 if (!iov_iter_is_bvec(iter)) {
3051 iovec = iov_iter_iovec(iter);
3053 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3054 iovec.iov_len = req->rw.len;
3058 nr = file->f_op->read(file, iovec.iov_base,
3059 iovec.iov_len, io_kiocb_ppos(kiocb));
3061 nr = file->f_op->write(file, iovec.iov_base,
3062 iovec.iov_len, io_kiocb_ppos(kiocb));
3071 if (nr != iovec.iov_len)
3075 iov_iter_advance(iter, nr);
3081 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3082 const struct iovec *fast_iov, struct iov_iter *iter)
3084 struct io_async_rw *rw = req->async_data;
3086 memcpy(&rw->iter, iter, sizeof(*iter));
3087 rw->free_iovec = iovec;
3089 /* can only be fixed buffers, no need to do anything */
3090 if (iov_iter_is_bvec(iter))
3093 unsigned iov_off = 0;
3095 rw->iter.iov = rw->fast_iov;
3096 if (iter->iov != fast_iov) {
3097 iov_off = iter->iov - fast_iov;
3098 rw->iter.iov += iov_off;
3100 if (rw->fast_iov != fast_iov)
3101 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3102 sizeof(struct iovec) * iter->nr_segs);
3104 req->flags |= REQ_F_NEED_CLEANUP;
3108 static inline int __io_alloc_async_data(struct io_kiocb *req)
3110 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3111 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3112 return req->async_data == NULL;
3115 static int io_alloc_async_data(struct io_kiocb *req)
3117 if (!io_op_defs[req->opcode].needs_async_data)
3120 return __io_alloc_async_data(req);
3123 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3124 const struct iovec *fast_iov,
3125 struct iov_iter *iter, bool force)
3127 if (!force && !io_op_defs[req->opcode].needs_async_data)
3129 if (!req->async_data) {
3130 if (__io_alloc_async_data(req)) {
3135 io_req_map_rw(req, iovec, fast_iov, iter);
3140 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3142 struct io_async_rw *iorw = req->async_data;
3143 struct iovec *iov = iorw->fast_iov;
3146 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3147 if (unlikely(ret < 0))
3150 iorw->bytes_done = 0;
3151 iorw->free_iovec = iov;
3153 req->flags |= REQ_F_NEED_CLEANUP;
3157 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3159 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3161 return io_prep_rw(req, sqe);
3165 * This is our waitqueue callback handler, registered through lock_page_async()
3166 * when we initially tried to do the IO with the iocb armed our waitqueue.
3167 * This gets called when the page is unlocked, and we generally expect that to
3168 * happen when the page IO is completed and the page is now uptodate. This will
3169 * queue a task_work based retry of the operation, attempting to copy the data
3170 * again. If the latter fails because the page was NOT uptodate, then we will
3171 * do a thread based blocking retry of the operation. That's the unexpected
3174 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3175 int sync, void *arg)
3177 struct wait_page_queue *wpq;
3178 struct io_kiocb *req = wait->private;
3179 struct wait_page_key *key = arg;
3181 wpq = container_of(wait, struct wait_page_queue, wait);
3183 if (!wake_page_match(wpq, key))
3186 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3187 list_del_init(&wait->entry);
3189 /* submit ref gets dropped, acquire a new one */
3190 refcount_inc(&req->refs);
3191 io_req_task_queue(req);
3196 * This controls whether a given IO request should be armed for async page
3197 * based retry. If we return false here, the request is handed to the async
3198 * worker threads for retry. If we're doing buffered reads on a regular file,
3199 * we prepare a private wait_page_queue entry and retry the operation. This
3200 * will either succeed because the page is now uptodate and unlocked, or it
3201 * will register a callback when the page is unlocked at IO completion. Through
3202 * that callback, io_uring uses task_work to setup a retry of the operation.
3203 * That retry will attempt the buffered read again. The retry will generally
3204 * succeed, or in rare cases where it fails, we then fall back to using the
3205 * async worker threads for a blocking retry.
3207 static bool io_rw_should_retry(struct io_kiocb *req)
3209 struct io_async_rw *rw = req->async_data;
3210 struct wait_page_queue *wait = &rw->wpq;
3211 struct kiocb *kiocb = &req->rw.kiocb;
3213 /* never retry for NOWAIT, we just complete with -EAGAIN */
3214 if (req->flags & REQ_F_NOWAIT)
3217 /* Only for buffered IO */
3218 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3222 * just use poll if we can, and don't attempt if the fs doesn't
3223 * support callback based unlocks
3225 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3228 wait->wait.func = io_async_buf_func;
3229 wait->wait.private = req;
3230 wait->wait.flags = 0;
3231 INIT_LIST_HEAD(&wait->wait.entry);
3232 kiocb->ki_flags |= IOCB_WAITQ;
3233 kiocb->ki_flags &= ~IOCB_NOWAIT;
3234 kiocb->ki_waitq = wait;
3238 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3240 if (req->file->f_op->read_iter)
3241 return call_read_iter(req->file, &req->rw.kiocb, iter);
3242 else if (req->file->f_op->read)
3243 return loop_rw_iter(READ, req, iter);
3248 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3250 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3251 struct kiocb *kiocb = &req->rw.kiocb;
3252 struct iov_iter __iter, *iter = &__iter;
3253 struct io_async_rw *rw = req->async_data;
3254 ssize_t io_size, ret, ret2;
3255 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3261 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3265 io_size = iov_iter_count(iter);
3266 req->result = io_size;
3268 /* Ensure we clear previously set non-block flag */
3269 if (!force_nonblock)
3270 kiocb->ki_flags &= ~IOCB_NOWAIT;
3272 kiocb->ki_flags |= IOCB_NOWAIT;
3274 /* If the file doesn't support async, just async punt */
3275 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3276 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3277 return ret ?: -EAGAIN;
3280 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3281 if (unlikely(ret)) {
3286 ret = io_iter_do_read(req, iter);
3288 if (ret == -EIOCBQUEUED) {
3289 if (req->async_data)
3290 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3292 } else if (ret == -EAGAIN) {
3293 /* IOPOLL retry should happen for io-wq threads */
3294 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3296 /* no retry on NONBLOCK nor RWF_NOWAIT */
3297 if (req->flags & REQ_F_NOWAIT)
3299 /* some cases will consume bytes even on error returns */
3300 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3302 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3303 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3304 /* read all, failed, already did sync or don't want to retry */
3308 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3313 rw = req->async_data;
3314 /* now use our persistent iterator, if we aren't already */
3319 rw->bytes_done += ret;
3320 /* if we can retry, do so with the callbacks armed */
3321 if (!io_rw_should_retry(req)) {
3322 kiocb->ki_flags &= ~IOCB_WAITQ;
3327 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3328 * we get -EIOCBQUEUED, then we'll get a notification when the
3329 * desired page gets unlocked. We can also get a partial read
3330 * here, and if we do, then just retry at the new offset.
3332 ret = io_iter_do_read(req, iter);
3333 if (ret == -EIOCBQUEUED)
3335 /* we got some bytes, but not all. retry. */
3336 kiocb->ki_flags &= ~IOCB_WAITQ;
3337 } while (ret > 0 && ret < io_size);
3339 kiocb_done(kiocb, ret, issue_flags);
3341 /* it's faster to check here then delegate to kfree */
3347 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3349 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3351 return io_prep_rw(req, sqe);
3354 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3356 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3357 struct kiocb *kiocb = &req->rw.kiocb;
3358 struct iov_iter __iter, *iter = &__iter;
3359 struct io_async_rw *rw = req->async_data;
3360 ssize_t ret, ret2, io_size;
3361 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3367 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3371 io_size = iov_iter_count(iter);
3372 req->result = io_size;
3374 /* Ensure we clear previously set non-block flag */
3375 if (!force_nonblock)
3376 kiocb->ki_flags &= ~IOCB_NOWAIT;
3378 kiocb->ki_flags |= IOCB_NOWAIT;
3380 /* If the file doesn't support async, just async punt */
3381 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3384 /* file path doesn't support NOWAIT for non-direct_IO */
3385 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3386 (req->flags & REQ_F_ISREG))
3389 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3394 * Open-code file_start_write here to grab freeze protection,
3395 * which will be released by another thread in
3396 * io_complete_rw(). Fool lockdep by telling it the lock got
3397 * released so that it doesn't complain about the held lock when
3398 * we return to userspace.
3400 if (req->flags & REQ_F_ISREG) {
3401 sb_start_write(file_inode(req->file)->i_sb);
3402 __sb_writers_release(file_inode(req->file)->i_sb,
3405 kiocb->ki_flags |= IOCB_WRITE;
3407 if (req->file->f_op->write_iter)
3408 ret2 = call_write_iter(req->file, kiocb, iter);
3409 else if (req->file->f_op->write)
3410 ret2 = loop_rw_iter(WRITE, req, iter);
3415 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3416 * retry them without IOCB_NOWAIT.
3418 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3420 /* no retry on NONBLOCK nor RWF_NOWAIT */
3421 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3423 if (ret2 == -EIOCBQUEUED && req->async_data)
3424 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3425 if (!force_nonblock || ret2 != -EAGAIN) {
3426 /* IOPOLL retry should happen for io-wq threads */
3427 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3430 kiocb_done(kiocb, ret2, issue_flags);
3433 /* some cases will consume bytes even on error returns */
3434 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3435 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3436 return ret ?: -EAGAIN;
3439 /* it's reportedly faster than delegating the null check to kfree() */
3445 static int io_renameat_prep(struct io_kiocb *req,
3446 const struct io_uring_sqe *sqe)
3448 struct io_rename *ren = &req->rename;
3449 const char __user *oldf, *newf;
3451 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3454 ren->old_dfd = READ_ONCE(sqe->fd);
3455 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3456 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3457 ren->new_dfd = READ_ONCE(sqe->len);
3458 ren->flags = READ_ONCE(sqe->rename_flags);
3460 ren->oldpath = getname(oldf);
3461 if (IS_ERR(ren->oldpath))
3462 return PTR_ERR(ren->oldpath);
3464 ren->newpath = getname(newf);
3465 if (IS_ERR(ren->newpath)) {
3466 putname(ren->oldpath);
3467 return PTR_ERR(ren->newpath);
3470 req->flags |= REQ_F_NEED_CLEANUP;
3474 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3476 struct io_rename *ren = &req->rename;
3479 if (issue_flags & IO_URING_F_NONBLOCK)
3482 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3483 ren->newpath, ren->flags);
3485 req->flags &= ~REQ_F_NEED_CLEANUP;
3487 req_set_fail_links(req);
3488 io_req_complete(req, ret);
3492 static int io_unlinkat_prep(struct io_kiocb *req,
3493 const struct io_uring_sqe *sqe)
3495 struct io_unlink *un = &req->unlink;
3496 const char __user *fname;
3498 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3501 un->dfd = READ_ONCE(sqe->fd);
3503 un->flags = READ_ONCE(sqe->unlink_flags);
3504 if (un->flags & ~AT_REMOVEDIR)
3507 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3508 un->filename = getname(fname);
3509 if (IS_ERR(un->filename))
3510 return PTR_ERR(un->filename);
3512 req->flags |= REQ_F_NEED_CLEANUP;
3516 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3518 struct io_unlink *un = &req->unlink;
3521 if (issue_flags & IO_URING_F_NONBLOCK)
3524 if (un->flags & AT_REMOVEDIR)
3525 ret = do_rmdir(un->dfd, un->filename);
3527 ret = do_unlinkat(un->dfd, un->filename);
3529 req->flags &= ~REQ_F_NEED_CLEANUP;
3531 req_set_fail_links(req);
3532 io_req_complete(req, ret);
3536 static int io_shutdown_prep(struct io_kiocb *req,
3537 const struct io_uring_sqe *sqe)
3539 #if defined(CONFIG_NET)
3540 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3542 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3546 req->shutdown.how = READ_ONCE(sqe->len);
3553 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3555 #if defined(CONFIG_NET)
3556 struct socket *sock;
3559 if (issue_flags & IO_URING_F_NONBLOCK)
3562 sock = sock_from_file(req->file);
3563 if (unlikely(!sock))
3566 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3568 req_set_fail_links(req);
3569 io_req_complete(req, ret);
3576 static int __io_splice_prep(struct io_kiocb *req,
3577 const struct io_uring_sqe *sqe)
3579 struct io_splice* sp = &req->splice;
3580 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3586 sp->len = READ_ONCE(sqe->len);
3587 sp->flags = READ_ONCE(sqe->splice_flags);
3589 if (unlikely(sp->flags & ~valid_flags))
3592 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3593 (sp->flags & SPLICE_F_FD_IN_FIXED));
3596 req->flags |= REQ_F_NEED_CLEANUP;
3598 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3600 * Splice operation will be punted aync, and here need to
3601 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3603 req->work.flags |= IO_WQ_WORK_UNBOUND;
3609 static int io_tee_prep(struct io_kiocb *req,
3610 const struct io_uring_sqe *sqe)
3612 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3614 return __io_splice_prep(req, sqe);
3617 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3619 struct io_splice *sp = &req->splice;
3620 struct file *in = sp->file_in;
3621 struct file *out = sp->file_out;
3622 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3625 if (issue_flags & IO_URING_F_NONBLOCK)
3628 ret = do_tee(in, out, sp->len, flags);
3630 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3631 req->flags &= ~REQ_F_NEED_CLEANUP;
3634 req_set_fail_links(req);
3635 io_req_complete(req, ret);
3639 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3641 struct io_splice* sp = &req->splice;
3643 sp->off_in = READ_ONCE(sqe->splice_off_in);
3644 sp->off_out = READ_ONCE(sqe->off);
3645 return __io_splice_prep(req, sqe);
3648 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3650 struct io_splice *sp = &req->splice;
3651 struct file *in = sp->file_in;
3652 struct file *out = sp->file_out;
3653 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3654 loff_t *poff_in, *poff_out;
3657 if (issue_flags & IO_URING_F_NONBLOCK)
3660 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3661 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3664 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3666 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3667 req->flags &= ~REQ_F_NEED_CLEANUP;
3670 req_set_fail_links(req);
3671 io_req_complete(req, ret);
3676 * IORING_OP_NOP just posts a completion event, nothing else.
3678 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3680 struct io_ring_ctx *ctx = req->ctx;
3682 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3685 __io_req_complete(req, issue_flags, 0, 0);
3689 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3691 struct io_ring_ctx *ctx = req->ctx;
3696 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3698 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3701 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3702 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3705 req->sync.off = READ_ONCE(sqe->off);
3706 req->sync.len = READ_ONCE(sqe->len);
3710 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3712 loff_t end = req->sync.off + req->sync.len;
3715 /* fsync always requires a blocking context */
3716 if (issue_flags & IO_URING_F_NONBLOCK)
3719 ret = vfs_fsync_range(req->file, req->sync.off,
3720 end > 0 ? end : LLONG_MAX,
3721 req->sync.flags & IORING_FSYNC_DATASYNC);
3723 req_set_fail_links(req);
3724 io_req_complete(req, ret);
3728 static int io_fallocate_prep(struct io_kiocb *req,
3729 const struct io_uring_sqe *sqe)
3731 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3736 req->sync.off = READ_ONCE(sqe->off);
3737 req->sync.len = READ_ONCE(sqe->addr);
3738 req->sync.mode = READ_ONCE(sqe->len);
3742 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3746 /* fallocate always requiring blocking context */
3747 if (issue_flags & IO_URING_F_NONBLOCK)
3749 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3752 req_set_fail_links(req);
3753 io_req_complete(req, ret);
3757 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3759 const char __user *fname;
3762 if (unlikely(sqe->ioprio || sqe->buf_index))
3764 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3767 /* open.how should be already initialised */
3768 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3769 req->open.how.flags |= O_LARGEFILE;
3771 req->open.dfd = READ_ONCE(sqe->fd);
3772 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3773 req->open.filename = getname(fname);
3774 if (IS_ERR(req->open.filename)) {
3775 ret = PTR_ERR(req->open.filename);
3776 req->open.filename = NULL;
3779 req->open.nofile = rlimit(RLIMIT_NOFILE);
3780 req->flags |= REQ_F_NEED_CLEANUP;
3784 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3788 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3790 mode = READ_ONCE(sqe->len);
3791 flags = READ_ONCE(sqe->open_flags);
3792 req->open.how = build_open_how(flags, mode);
3793 return __io_openat_prep(req, sqe);
3796 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3798 struct open_how __user *how;
3802 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3804 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3805 len = READ_ONCE(sqe->len);
3806 if (len < OPEN_HOW_SIZE_VER0)
3809 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3814 return __io_openat_prep(req, sqe);
3817 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3819 struct open_flags op;
3822 bool resolve_nonblock;
3825 ret = build_open_flags(&req->open.how, &op);
3828 nonblock_set = op.open_flag & O_NONBLOCK;
3829 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3830 if (issue_flags & IO_URING_F_NONBLOCK) {
3832 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3833 * it'll always -EAGAIN
3835 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3837 op.lookup_flags |= LOOKUP_CACHED;
3838 op.open_flag |= O_NONBLOCK;
3841 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3845 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3846 /* only retry if RESOLVE_CACHED wasn't already set by application */
3847 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3848 file == ERR_PTR(-EAGAIN)) {
3850 * We could hang on to this 'fd', but seems like marginal
3851 * gain for something that is now known to be a slower path.
3852 * So just put it, and we'll get a new one when we retry.
3860 ret = PTR_ERR(file);
3862 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3863 file->f_flags &= ~O_NONBLOCK;
3864 fsnotify_open(file);
3865 fd_install(ret, file);
3868 putname(req->open.filename);
3869 req->flags &= ~REQ_F_NEED_CLEANUP;
3871 req_set_fail_links(req);
3872 io_req_complete(req, ret);
3876 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3878 return io_openat2(req, issue_flags);
3881 static int io_remove_buffers_prep(struct io_kiocb *req,
3882 const struct io_uring_sqe *sqe)
3884 struct io_provide_buf *p = &req->pbuf;
3887 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3890 tmp = READ_ONCE(sqe->fd);
3891 if (!tmp || tmp > USHRT_MAX)
3894 memset(p, 0, sizeof(*p));
3896 p->bgid = READ_ONCE(sqe->buf_group);
3900 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3901 int bgid, unsigned nbufs)
3905 /* shouldn't happen */
3909 /* the head kbuf is the list itself */
3910 while (!list_empty(&buf->list)) {
3911 struct io_buffer *nxt;
3913 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3914 list_del(&nxt->list);
3921 xa_erase(&ctx->io_buffers, bgid);
3926 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3928 struct io_provide_buf *p = &req->pbuf;
3929 struct io_ring_ctx *ctx = req->ctx;
3930 struct io_buffer *head;
3932 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3934 io_ring_submit_lock(ctx, !force_nonblock);
3936 lockdep_assert_held(&ctx->uring_lock);
3939 head = xa_load(&ctx->io_buffers, p->bgid);
3941 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3943 req_set_fail_links(req);
3945 /* need to hold the lock to complete IOPOLL requests */
3946 if (ctx->flags & IORING_SETUP_IOPOLL) {
3947 __io_req_complete(req, issue_flags, ret, 0);
3948 io_ring_submit_unlock(ctx, !force_nonblock);
3950 io_ring_submit_unlock(ctx, !force_nonblock);
3951 __io_req_complete(req, issue_flags, ret, 0);
3956 static int io_provide_buffers_prep(struct io_kiocb *req,
3957 const struct io_uring_sqe *sqe)
3960 struct io_provide_buf *p = &req->pbuf;
3963 if (sqe->ioprio || sqe->rw_flags)
3966 tmp = READ_ONCE(sqe->fd);
3967 if (!tmp || tmp > USHRT_MAX)
3970 p->addr = READ_ONCE(sqe->addr);
3971 p->len = READ_ONCE(sqe->len);
3973 size = (unsigned long)p->len * p->nbufs;
3974 if (!access_ok(u64_to_user_ptr(p->addr), size))
3977 p->bgid = READ_ONCE(sqe->buf_group);
3978 tmp = READ_ONCE(sqe->off);
3979 if (tmp > USHRT_MAX)
3985 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3987 struct io_buffer *buf;
3988 u64 addr = pbuf->addr;
3989 int i, bid = pbuf->bid;
3991 for (i = 0; i < pbuf->nbufs; i++) {
3992 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3997 buf->len = pbuf->len;
4002 INIT_LIST_HEAD(&buf->list);
4005 list_add_tail(&buf->list, &(*head)->list);
4009 return i ? i : -ENOMEM;
4012 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4014 struct io_provide_buf *p = &req->pbuf;
4015 struct io_ring_ctx *ctx = req->ctx;
4016 struct io_buffer *head, *list;
4018 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4020 io_ring_submit_lock(ctx, !force_nonblock);
4022 lockdep_assert_held(&ctx->uring_lock);
4024 list = head = xa_load(&ctx->io_buffers, p->bgid);
4026 ret = io_add_buffers(p, &head);
4027 if (ret >= 0 && !list) {
4028 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4030 __io_remove_buffers(ctx, head, p->bgid, -1U);
4033 req_set_fail_links(req);
4035 /* need to hold the lock to complete IOPOLL requests */
4036 if (ctx->flags & IORING_SETUP_IOPOLL) {
4037 __io_req_complete(req, issue_flags, ret, 0);
4038 io_ring_submit_unlock(ctx, !force_nonblock);
4040 io_ring_submit_unlock(ctx, !force_nonblock);
4041 __io_req_complete(req, issue_flags, ret, 0);
4046 static int io_epoll_ctl_prep(struct io_kiocb *req,
4047 const struct io_uring_sqe *sqe)
4049 #if defined(CONFIG_EPOLL)
4050 if (sqe->ioprio || sqe->buf_index)
4052 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4055 req->epoll.epfd = READ_ONCE(sqe->fd);
4056 req->epoll.op = READ_ONCE(sqe->len);
4057 req->epoll.fd = READ_ONCE(sqe->off);
4059 if (ep_op_has_event(req->epoll.op)) {
4060 struct epoll_event __user *ev;
4062 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4063 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4073 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4075 #if defined(CONFIG_EPOLL)
4076 struct io_epoll *ie = &req->epoll;
4078 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4080 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4081 if (force_nonblock && ret == -EAGAIN)
4085 req_set_fail_links(req);
4086 __io_req_complete(req, issue_flags, ret, 0);
4093 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4095 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4096 if (sqe->ioprio || sqe->buf_index || sqe->off)
4098 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4101 req->madvise.addr = READ_ONCE(sqe->addr);
4102 req->madvise.len = READ_ONCE(sqe->len);
4103 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4110 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4112 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4113 struct io_madvise *ma = &req->madvise;
4116 if (issue_flags & IO_URING_F_NONBLOCK)
4119 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4121 req_set_fail_links(req);
4122 io_req_complete(req, ret);
4129 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4136 req->fadvise.offset = READ_ONCE(sqe->off);
4137 req->fadvise.len = READ_ONCE(sqe->len);
4138 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4142 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4144 struct io_fadvise *fa = &req->fadvise;
4147 if (issue_flags & IO_URING_F_NONBLOCK) {
4148 switch (fa->advice) {
4149 case POSIX_FADV_NORMAL:
4150 case POSIX_FADV_RANDOM:
4151 case POSIX_FADV_SEQUENTIAL:
4158 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4160 req_set_fail_links(req);
4161 io_req_complete(req, ret);
4165 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4167 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4169 if (sqe->ioprio || sqe->buf_index)
4171 if (req->flags & REQ_F_FIXED_FILE)
4174 req->statx.dfd = READ_ONCE(sqe->fd);
4175 req->statx.mask = READ_ONCE(sqe->len);
4176 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4177 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4178 req->statx.flags = READ_ONCE(sqe->statx_flags);
4183 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4185 struct io_statx *ctx = &req->statx;
4188 if (issue_flags & IO_URING_F_NONBLOCK) {
4189 /* only need file table for an actual valid fd */
4190 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4191 req->flags |= REQ_F_NO_FILE_TABLE;
4195 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4199 req_set_fail_links(req);
4200 io_req_complete(req, ret);
4204 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4206 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4208 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4209 sqe->rw_flags || sqe->buf_index)
4211 if (req->flags & REQ_F_FIXED_FILE)
4214 req->close.fd = READ_ONCE(sqe->fd);
4218 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4220 struct files_struct *files = current->files;
4221 struct io_close *close = &req->close;
4222 struct fdtable *fdt;
4228 spin_lock(&files->file_lock);
4229 fdt = files_fdtable(files);
4230 if (close->fd >= fdt->max_fds) {
4231 spin_unlock(&files->file_lock);
4234 file = fdt->fd[close->fd];
4236 spin_unlock(&files->file_lock);
4240 if (file->f_op == &io_uring_fops) {
4241 spin_unlock(&files->file_lock);
4246 /* if the file has a flush method, be safe and punt to async */
4247 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4248 spin_unlock(&files->file_lock);
4252 ret = __close_fd_get_file(close->fd, &file);
4253 spin_unlock(&files->file_lock);
4260 /* No ->flush() or already async, safely close from here */
4261 ret = filp_close(file, current->files);
4264 req_set_fail_links(req);
4267 __io_req_complete(req, issue_flags, ret, 0);
4271 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4273 struct io_ring_ctx *ctx = req->ctx;
4275 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4277 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4280 req->sync.off = READ_ONCE(sqe->off);
4281 req->sync.len = READ_ONCE(sqe->len);
4282 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4286 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4290 /* sync_file_range always requires a blocking context */
4291 if (issue_flags & IO_URING_F_NONBLOCK)
4294 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4297 req_set_fail_links(req);
4298 io_req_complete(req, ret);
4302 #if defined(CONFIG_NET)
4303 static int io_setup_async_msg(struct io_kiocb *req,
4304 struct io_async_msghdr *kmsg)
4306 struct io_async_msghdr *async_msg = req->async_data;
4310 if (io_alloc_async_data(req)) {
4311 kfree(kmsg->free_iov);
4314 async_msg = req->async_data;
4315 req->flags |= REQ_F_NEED_CLEANUP;
4316 memcpy(async_msg, kmsg, sizeof(*kmsg));
4317 async_msg->msg.msg_name = &async_msg->addr;
4318 /* if were using fast_iov, set it to the new one */
4319 if (!async_msg->free_iov)
4320 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4325 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4326 struct io_async_msghdr *iomsg)
4328 iomsg->msg.msg_name = &iomsg->addr;
4329 iomsg->free_iov = iomsg->fast_iov;
4330 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4331 req->sr_msg.msg_flags, &iomsg->free_iov);
4334 static int io_sendmsg_prep_async(struct io_kiocb *req)
4338 if (!io_op_defs[req->opcode].needs_async_data)
4340 ret = io_sendmsg_copy_hdr(req, req->async_data);
4342 req->flags |= REQ_F_NEED_CLEANUP;
4346 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4348 struct io_sr_msg *sr = &req->sr_msg;
4350 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4353 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4354 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4355 sr->len = READ_ONCE(sqe->len);
4357 #ifdef CONFIG_COMPAT
4358 if (req->ctx->compat)
4359 sr->msg_flags |= MSG_CMSG_COMPAT;
4364 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4366 struct io_async_msghdr iomsg, *kmsg;
4367 struct socket *sock;
4372 sock = sock_from_file(req->file);
4373 if (unlikely(!sock))
4376 kmsg = req->async_data;
4378 ret = io_sendmsg_copy_hdr(req, &iomsg);
4384 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4385 if (flags & MSG_DONTWAIT)
4386 req->flags |= REQ_F_NOWAIT;
4387 else if (issue_flags & IO_URING_F_NONBLOCK)
4388 flags |= MSG_DONTWAIT;
4390 if (flags & MSG_WAITALL)
4391 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4393 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4394 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4395 return io_setup_async_msg(req, kmsg);
4396 if (ret == -ERESTARTSYS)
4399 /* fast path, check for non-NULL to avoid function call */
4401 kfree(kmsg->free_iov);
4402 req->flags &= ~REQ_F_NEED_CLEANUP;
4404 req_set_fail_links(req);
4405 __io_req_complete(req, issue_flags, ret, 0);
4409 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4411 struct io_sr_msg *sr = &req->sr_msg;
4414 struct socket *sock;
4419 sock = sock_from_file(req->file);
4420 if (unlikely(!sock))
4423 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4427 msg.msg_name = NULL;
4428 msg.msg_control = NULL;
4429 msg.msg_controllen = 0;
4430 msg.msg_namelen = 0;
4432 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4433 if (flags & MSG_DONTWAIT)
4434 req->flags |= REQ_F_NOWAIT;
4435 else if (issue_flags & IO_URING_F_NONBLOCK)
4436 flags |= MSG_DONTWAIT;
4438 if (flags & MSG_WAITALL)
4439 min_ret = iov_iter_count(&msg.msg_iter);
4441 msg.msg_flags = flags;
4442 ret = sock_sendmsg(sock, &msg);
4443 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4445 if (ret == -ERESTARTSYS)
4449 req_set_fail_links(req);
4450 __io_req_complete(req, issue_flags, ret, 0);
4454 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4455 struct io_async_msghdr *iomsg)
4457 struct io_sr_msg *sr = &req->sr_msg;
4458 struct iovec __user *uiov;
4462 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4463 &iomsg->uaddr, &uiov, &iov_len);
4467 if (req->flags & REQ_F_BUFFER_SELECT) {
4470 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4472 sr->len = iomsg->fast_iov[0].iov_len;
4473 iomsg->free_iov = NULL;
4475 iomsg->free_iov = iomsg->fast_iov;
4476 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4477 &iomsg->free_iov, &iomsg->msg.msg_iter,
4486 #ifdef CONFIG_COMPAT
4487 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4488 struct io_async_msghdr *iomsg)
4490 struct compat_msghdr __user *msg_compat;
4491 struct io_sr_msg *sr = &req->sr_msg;
4492 struct compat_iovec __user *uiov;
4497 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4498 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4503 uiov = compat_ptr(ptr);
4504 if (req->flags & REQ_F_BUFFER_SELECT) {
4505 compat_ssize_t clen;
4509 if (!access_ok(uiov, sizeof(*uiov)))
4511 if (__get_user(clen, &uiov->iov_len))
4516 iomsg->free_iov = NULL;
4518 iomsg->free_iov = iomsg->fast_iov;
4519 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4520 UIO_FASTIOV, &iomsg->free_iov,
4521 &iomsg->msg.msg_iter, true);
4530 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4531 struct io_async_msghdr *iomsg)
4533 iomsg->msg.msg_name = &iomsg->addr;
4535 #ifdef CONFIG_COMPAT
4536 if (req->ctx->compat)
4537 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4540 return __io_recvmsg_copy_hdr(req, iomsg);
4543 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4546 struct io_sr_msg *sr = &req->sr_msg;
4547 struct io_buffer *kbuf;
4549 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4554 req->flags |= REQ_F_BUFFER_SELECTED;
4558 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4560 return io_put_kbuf(req, req->sr_msg.kbuf);
4563 static int io_recvmsg_prep_async(struct io_kiocb *req)
4567 if (!io_op_defs[req->opcode].needs_async_data)
4569 ret = io_recvmsg_copy_hdr(req, req->async_data);
4571 req->flags |= REQ_F_NEED_CLEANUP;
4575 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4577 struct io_sr_msg *sr = &req->sr_msg;
4579 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4582 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4583 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4584 sr->len = READ_ONCE(sqe->len);
4585 sr->bgid = READ_ONCE(sqe->buf_group);
4587 #ifdef CONFIG_COMPAT
4588 if (req->ctx->compat)
4589 sr->msg_flags |= MSG_CMSG_COMPAT;
4594 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4596 struct io_async_msghdr iomsg, *kmsg;
4597 struct socket *sock;
4598 struct io_buffer *kbuf;
4601 int ret, cflags = 0;
4602 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4604 sock = sock_from_file(req->file);
4605 if (unlikely(!sock))
4608 kmsg = req->async_data;
4610 ret = io_recvmsg_copy_hdr(req, &iomsg);
4616 if (req->flags & REQ_F_BUFFER_SELECT) {
4617 kbuf = io_recv_buffer_select(req, !force_nonblock);
4619 return PTR_ERR(kbuf);
4620 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4621 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4622 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4623 1, req->sr_msg.len);
4626 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4627 if (flags & MSG_DONTWAIT)
4628 req->flags |= REQ_F_NOWAIT;
4629 else if (force_nonblock)
4630 flags |= MSG_DONTWAIT;
4632 if (flags & MSG_WAITALL)
4633 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4635 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4636 kmsg->uaddr, flags);
4637 if (force_nonblock && ret == -EAGAIN)
4638 return io_setup_async_msg(req, kmsg);
4639 if (ret == -ERESTARTSYS)
4642 if (req->flags & REQ_F_BUFFER_SELECTED)
4643 cflags = io_put_recv_kbuf(req);
4644 /* fast path, check for non-NULL to avoid function call */
4646 kfree(kmsg->free_iov);
4647 req->flags &= ~REQ_F_NEED_CLEANUP;
4648 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4649 req_set_fail_links(req);
4650 __io_req_complete(req, issue_flags, ret, cflags);
4654 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4656 struct io_buffer *kbuf;
4657 struct io_sr_msg *sr = &req->sr_msg;
4659 void __user *buf = sr->buf;
4660 struct socket *sock;
4664 int ret, cflags = 0;
4665 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4667 sock = sock_from_file(req->file);
4668 if (unlikely(!sock))
4671 if (req->flags & REQ_F_BUFFER_SELECT) {
4672 kbuf = io_recv_buffer_select(req, !force_nonblock);
4674 return PTR_ERR(kbuf);
4675 buf = u64_to_user_ptr(kbuf->addr);
4678 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4682 msg.msg_name = NULL;
4683 msg.msg_control = NULL;
4684 msg.msg_controllen = 0;
4685 msg.msg_namelen = 0;
4686 msg.msg_iocb = NULL;
4689 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4690 if (flags & MSG_DONTWAIT)
4691 req->flags |= REQ_F_NOWAIT;
4692 else if (force_nonblock)
4693 flags |= MSG_DONTWAIT;
4695 if (flags & MSG_WAITALL)
4696 min_ret = iov_iter_count(&msg.msg_iter);
4698 ret = sock_recvmsg(sock, &msg, flags);
4699 if (force_nonblock && ret == -EAGAIN)
4701 if (ret == -ERESTARTSYS)
4704 if (req->flags & REQ_F_BUFFER_SELECTED)
4705 cflags = io_put_recv_kbuf(req);
4706 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4707 req_set_fail_links(req);
4708 __io_req_complete(req, issue_flags, ret, cflags);
4712 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4714 struct io_accept *accept = &req->accept;
4716 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4718 if (sqe->ioprio || sqe->len || sqe->buf_index)
4721 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4722 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4723 accept->flags = READ_ONCE(sqe->accept_flags);
4724 accept->nofile = rlimit(RLIMIT_NOFILE);
4728 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4730 struct io_accept *accept = &req->accept;
4731 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4732 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4735 if (req->file->f_flags & O_NONBLOCK)
4736 req->flags |= REQ_F_NOWAIT;
4738 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4739 accept->addr_len, accept->flags,
4741 if (ret == -EAGAIN && force_nonblock)
4744 if (ret == -ERESTARTSYS)
4746 req_set_fail_links(req);
4748 __io_req_complete(req, issue_flags, ret, 0);
4752 static int io_connect_prep_async(struct io_kiocb *req)
4754 struct io_async_connect *io = req->async_data;
4755 struct io_connect *conn = &req->connect;
4757 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4760 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4762 struct io_connect *conn = &req->connect;
4764 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4766 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4769 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4770 conn->addr_len = READ_ONCE(sqe->addr2);
4774 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4776 struct io_async_connect __io, *io;
4777 unsigned file_flags;
4779 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4781 if (req->async_data) {
4782 io = req->async_data;
4784 ret = move_addr_to_kernel(req->connect.addr,
4785 req->connect.addr_len,
4792 file_flags = force_nonblock ? O_NONBLOCK : 0;
4794 ret = __sys_connect_file(req->file, &io->address,
4795 req->connect.addr_len, file_flags);
4796 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4797 if (req->async_data)
4799 if (io_alloc_async_data(req)) {
4803 io = req->async_data;
4804 memcpy(req->async_data, &__io, sizeof(__io));
4807 if (ret == -ERESTARTSYS)
4811 req_set_fail_links(req);
4812 __io_req_complete(req, issue_flags, ret, 0);
4815 #else /* !CONFIG_NET */
4816 #define IO_NETOP_FN(op) \
4817 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4819 return -EOPNOTSUPP; \
4822 #define IO_NETOP_PREP(op) \
4824 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4826 return -EOPNOTSUPP; \
4829 #define IO_NETOP_PREP_ASYNC(op) \
4831 static int io_##op##_prep_async(struct io_kiocb *req) \
4833 return -EOPNOTSUPP; \
4836 IO_NETOP_PREP_ASYNC(sendmsg);
4837 IO_NETOP_PREP_ASYNC(recvmsg);
4838 IO_NETOP_PREP_ASYNC(connect);
4839 IO_NETOP_PREP(accept);
4842 #endif /* CONFIG_NET */
4844 struct io_poll_table {
4845 struct poll_table_struct pt;
4846 struct io_kiocb *req;
4850 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4851 __poll_t mask, task_work_func_t func)
4855 /* for instances that support it check for an event match first: */
4856 if (mask && !(mask & poll->events))
4859 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4861 list_del_init(&poll->wait.entry);
4864 req->task_work.func = func;
4865 percpu_ref_get(&req->ctx->refs);
4868 * If this fails, then the task is exiting. When a task exits, the
4869 * work gets canceled, so just cancel this request as well instead
4870 * of executing it. We can't safely execute it anyway, as we may not
4871 * have the needed state needed for it anyway.
4873 ret = io_req_task_work_add(req);
4874 if (unlikely(ret)) {
4875 WRITE_ONCE(poll->canceled, true);
4876 io_req_task_work_add_fallback(req, func);
4881 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4882 __acquires(&req->ctx->completion_lock)
4884 struct io_ring_ctx *ctx = req->ctx;
4886 if (!req->result && !READ_ONCE(poll->canceled)) {
4887 struct poll_table_struct pt = { ._key = poll->events };
4889 req->result = vfs_poll(req->file, &pt) & poll->events;
4892 spin_lock_irq(&ctx->completion_lock);
4893 if (!req->result && !READ_ONCE(poll->canceled)) {
4894 add_wait_queue(poll->head, &poll->wait);
4901 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4903 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4904 if (req->opcode == IORING_OP_POLL_ADD)
4905 return req->async_data;
4906 return req->apoll->double_poll;
4909 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4911 if (req->opcode == IORING_OP_POLL_ADD)
4913 return &req->apoll->poll;
4916 static void io_poll_remove_double(struct io_kiocb *req)
4918 struct io_poll_iocb *poll = io_poll_get_double(req);
4920 lockdep_assert_held(&req->ctx->completion_lock);
4922 if (poll && poll->head) {
4923 struct wait_queue_head *head = poll->head;
4925 spin_lock(&head->lock);
4926 list_del_init(&poll->wait.entry);
4927 if (poll->wait.private)
4928 refcount_dec(&req->refs);
4930 spin_unlock(&head->lock);
4934 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4936 struct io_ring_ctx *ctx = req->ctx;
4938 io_poll_remove_double(req);
4939 req->poll.done = true;
4940 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4941 io_commit_cqring(ctx);
4944 static void io_poll_task_func(struct callback_head *cb)
4946 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4947 struct io_ring_ctx *ctx = req->ctx;
4948 struct io_kiocb *nxt;
4950 if (io_poll_rewait(req, &req->poll)) {
4951 spin_unlock_irq(&ctx->completion_lock);
4953 hash_del(&req->hash_node);
4954 io_poll_complete(req, req->result, 0);
4955 spin_unlock_irq(&ctx->completion_lock);
4957 nxt = io_put_req_find_next(req);
4958 io_cqring_ev_posted(ctx);
4960 __io_req_task_submit(nxt);
4963 percpu_ref_put(&ctx->refs);
4966 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4967 int sync, void *key)
4969 struct io_kiocb *req = wait->private;
4970 struct io_poll_iocb *poll = io_poll_get_single(req);
4971 __poll_t mask = key_to_poll(key);
4973 /* for instances that support it check for an event match first: */
4974 if (mask && !(mask & poll->events))
4977 list_del_init(&wait->entry);
4979 if (poll && poll->head) {
4982 spin_lock(&poll->head->lock);
4983 done = list_empty(&poll->wait.entry);
4985 list_del_init(&poll->wait.entry);
4986 /* make sure double remove sees this as being gone */
4987 wait->private = NULL;
4988 spin_unlock(&poll->head->lock);
4990 /* use wait func handler, so it matches the rq type */
4991 poll->wait.func(&poll->wait, mode, sync, key);
4994 refcount_dec(&req->refs);
4998 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4999 wait_queue_func_t wake_func)
5003 poll->canceled = false;
5004 poll->events = events;
5005 INIT_LIST_HEAD(&poll->wait.entry);
5006 init_waitqueue_func_entry(&poll->wait, wake_func);
5009 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5010 struct wait_queue_head *head,
5011 struct io_poll_iocb **poll_ptr)
5013 struct io_kiocb *req = pt->req;
5016 * If poll->head is already set, it's because the file being polled
5017 * uses multiple waitqueues for poll handling (eg one for read, one
5018 * for write). Setup a separate io_poll_iocb if this happens.
5020 if (unlikely(poll->head)) {
5021 struct io_poll_iocb *poll_one = poll;
5023 /* already have a 2nd entry, fail a third attempt */
5025 pt->error = -EINVAL;
5028 /* double add on the same waitqueue head, ignore */
5029 if (poll->head == head)
5031 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5033 pt->error = -ENOMEM;
5036 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5037 refcount_inc(&req->refs);
5038 poll->wait.private = req;
5045 if (poll->events & EPOLLEXCLUSIVE)
5046 add_wait_queue_exclusive(head, &poll->wait);
5048 add_wait_queue(head, &poll->wait);
5051 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5052 struct poll_table_struct *p)
5054 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5055 struct async_poll *apoll = pt->req->apoll;
5057 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5060 static void io_async_task_func(struct callback_head *cb)
5062 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5063 struct async_poll *apoll = req->apoll;
5064 struct io_ring_ctx *ctx = req->ctx;
5066 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5068 if (io_poll_rewait(req, &apoll->poll)) {
5069 spin_unlock_irq(&ctx->completion_lock);
5070 percpu_ref_put(&ctx->refs);
5074 /* If req is still hashed, it cannot have been canceled. Don't check. */
5075 if (hash_hashed(&req->hash_node))
5076 hash_del(&req->hash_node);
5078 io_poll_remove_double(req);
5079 spin_unlock_irq(&ctx->completion_lock);
5081 if (!READ_ONCE(apoll->poll.canceled))
5082 __io_req_task_submit(req);
5084 __io_req_task_cancel(req, -ECANCELED);
5086 percpu_ref_put(&ctx->refs);
5087 kfree(apoll->double_poll);
5091 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5094 struct io_kiocb *req = wait->private;
5095 struct io_poll_iocb *poll = &req->apoll->poll;
5097 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5100 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5103 static void io_poll_req_insert(struct io_kiocb *req)
5105 struct io_ring_ctx *ctx = req->ctx;
5106 struct hlist_head *list;
5108 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5109 hlist_add_head(&req->hash_node, list);
5112 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5113 struct io_poll_iocb *poll,
5114 struct io_poll_table *ipt, __poll_t mask,
5115 wait_queue_func_t wake_func)
5116 __acquires(&ctx->completion_lock)
5118 struct io_ring_ctx *ctx = req->ctx;
5119 bool cancel = false;
5121 INIT_HLIST_NODE(&req->hash_node);
5122 io_init_poll_iocb(poll, mask, wake_func);
5123 poll->file = req->file;
5124 poll->wait.private = req;
5126 ipt->pt._key = mask;
5128 ipt->error = -EINVAL;
5130 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5132 spin_lock_irq(&ctx->completion_lock);
5133 if (likely(poll->head)) {
5134 spin_lock(&poll->head->lock);
5135 if (unlikely(list_empty(&poll->wait.entry))) {
5141 if (mask || ipt->error)
5142 list_del_init(&poll->wait.entry);
5144 WRITE_ONCE(poll->canceled, true);
5145 else if (!poll->done) /* actually waiting for an event */
5146 io_poll_req_insert(req);
5147 spin_unlock(&poll->head->lock);
5153 static bool io_arm_poll_handler(struct io_kiocb *req)
5155 const struct io_op_def *def = &io_op_defs[req->opcode];
5156 struct io_ring_ctx *ctx = req->ctx;
5157 struct async_poll *apoll;
5158 struct io_poll_table ipt;
5162 if (!req->file || !file_can_poll(req->file))
5164 if (req->flags & REQ_F_POLLED)
5168 else if (def->pollout)
5172 /* if we can't nonblock try, then no point in arming a poll handler */
5173 if (!io_file_supports_async(req->file, rw))
5176 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5177 if (unlikely(!apoll))
5179 apoll->double_poll = NULL;
5181 req->flags |= REQ_F_POLLED;
5186 mask |= POLLIN | POLLRDNORM;
5188 mask |= POLLOUT | POLLWRNORM;
5190 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5191 if ((req->opcode == IORING_OP_RECVMSG) &&
5192 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5195 mask |= POLLERR | POLLPRI;
5197 ipt.pt._qproc = io_async_queue_proc;
5199 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5201 if (ret || ipt.error) {
5202 io_poll_remove_double(req);
5203 spin_unlock_irq(&ctx->completion_lock);
5204 kfree(apoll->double_poll);
5208 spin_unlock_irq(&ctx->completion_lock);
5209 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5210 apoll->poll.events);
5214 static bool __io_poll_remove_one(struct io_kiocb *req,
5215 struct io_poll_iocb *poll)
5217 bool do_complete = false;
5219 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_one(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);
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);
5245 kfree(apoll->double_poll);
5251 io_cqring_fill_event(req, -ECANCELED);
5252 io_commit_cqring(req->ctx);
5253 req_set_fail_links(req);
5254 io_put_req_deferred(req, 1);
5261 * Returns true if we found and killed one or more poll requests
5263 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5264 struct files_struct *files)
5266 struct hlist_node *tmp;
5267 struct io_kiocb *req;
5270 spin_lock_irq(&ctx->completion_lock);
5271 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5272 struct hlist_head *list;
5274 list = &ctx->cancel_hash[i];
5275 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5276 if (io_match_task(req, tsk, files))
5277 posted += io_poll_remove_one(req);
5280 spin_unlock_irq(&ctx->completion_lock);
5283 io_cqring_ev_posted(ctx);
5288 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5290 struct hlist_head *list;
5291 struct io_kiocb *req;
5293 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5294 hlist_for_each_entry(req, list, hash_node) {
5295 if (sqe_addr != req->user_data)
5297 if (io_poll_remove_one(req))
5305 static int io_poll_remove_prep(struct io_kiocb *req,
5306 const struct io_uring_sqe *sqe)
5308 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5310 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5314 req->poll_remove.addr = READ_ONCE(sqe->addr);
5319 * Find a running poll command that matches one specified in sqe->addr,
5320 * and remove it if found.
5322 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5324 struct io_ring_ctx *ctx = req->ctx;
5327 spin_lock_irq(&ctx->completion_lock);
5328 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5329 spin_unlock_irq(&ctx->completion_lock);
5332 req_set_fail_links(req);
5333 io_req_complete(req, ret);
5337 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5340 struct io_kiocb *req = wait->private;
5341 struct io_poll_iocb *poll = &req->poll;
5343 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5346 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5347 struct poll_table_struct *p)
5349 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5351 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5354 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5356 struct io_poll_iocb *poll = &req->poll;
5359 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5361 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5364 events = READ_ONCE(sqe->poll32_events);
5366 events = swahw32(events);
5368 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5369 (events & EPOLLEXCLUSIVE);
5373 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5375 struct io_poll_iocb *poll = &req->poll;
5376 struct io_ring_ctx *ctx = req->ctx;
5377 struct io_poll_table ipt;
5380 ipt.pt._qproc = io_poll_queue_proc;
5382 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5385 if (mask) { /* no async, we'd stolen it */
5387 io_poll_complete(req, mask, 0);
5389 spin_unlock_irq(&ctx->completion_lock);
5392 io_cqring_ev_posted(ctx);
5398 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5400 struct io_timeout_data *data = container_of(timer,
5401 struct io_timeout_data, timer);
5402 struct io_kiocb *req = data->req;
5403 struct io_ring_ctx *ctx = req->ctx;
5404 unsigned long flags;
5406 spin_lock_irqsave(&ctx->completion_lock, flags);
5407 list_del_init(&req->timeout.list);
5408 atomic_set(&req->ctx->cq_timeouts,
5409 atomic_read(&req->ctx->cq_timeouts) + 1);
5411 io_cqring_fill_event(req, -ETIME);
5412 io_commit_cqring(ctx);
5413 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5415 io_cqring_ev_posted(ctx);
5416 req_set_fail_links(req);
5418 return HRTIMER_NORESTART;
5421 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5424 struct io_timeout_data *io;
5425 struct io_kiocb *req;
5428 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5429 if (user_data == req->user_data) {
5436 return ERR_PTR(ret);
5438 io = req->async_data;
5439 ret = hrtimer_try_to_cancel(&io->timer);
5441 return ERR_PTR(-EALREADY);
5442 list_del_init(&req->timeout.list);
5446 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5448 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5451 return PTR_ERR(req);
5453 req_set_fail_links(req);
5454 io_cqring_fill_event(req, -ECANCELED);
5455 io_put_req_deferred(req, 1);
5459 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5460 struct timespec64 *ts, enum hrtimer_mode mode)
5462 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5463 struct io_timeout_data *data;
5466 return PTR_ERR(req);
5468 req->timeout.off = 0; /* noseq */
5469 data = req->async_data;
5470 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5471 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5472 data->timer.function = io_timeout_fn;
5473 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5477 static int io_timeout_remove_prep(struct io_kiocb *req,
5478 const struct io_uring_sqe *sqe)
5480 struct io_timeout_rem *tr = &req->timeout_rem;
5482 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5484 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5486 if (sqe->ioprio || sqe->buf_index || sqe->len)
5489 tr->addr = READ_ONCE(sqe->addr);
5490 tr->flags = READ_ONCE(sqe->timeout_flags);
5491 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5492 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5494 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5496 } else if (tr->flags) {
5497 /* timeout removal doesn't support flags */
5504 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5506 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5511 * Remove or update an existing timeout command
5513 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5515 struct io_timeout_rem *tr = &req->timeout_rem;
5516 struct io_ring_ctx *ctx = req->ctx;
5519 spin_lock_irq(&ctx->completion_lock);
5520 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5521 ret = io_timeout_cancel(ctx, tr->addr);
5523 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5524 io_translate_timeout_mode(tr->flags));
5526 io_cqring_fill_event(req, ret);
5527 io_commit_cqring(ctx);
5528 spin_unlock_irq(&ctx->completion_lock);
5529 io_cqring_ev_posted(ctx);
5531 req_set_fail_links(req);
5536 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5537 bool is_timeout_link)
5539 struct io_timeout_data *data;
5541 u32 off = READ_ONCE(sqe->off);
5543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5545 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5547 if (off && is_timeout_link)
5549 flags = READ_ONCE(sqe->timeout_flags);
5550 if (flags & ~IORING_TIMEOUT_ABS)
5553 req->timeout.off = off;
5555 if (!req->async_data && io_alloc_async_data(req))
5558 data = req->async_data;
5561 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5564 data->mode = io_translate_timeout_mode(flags);
5565 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5566 if (is_timeout_link)
5567 io_req_track_inflight(req);
5571 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5573 struct io_ring_ctx *ctx = req->ctx;
5574 struct io_timeout_data *data = req->async_data;
5575 struct list_head *entry;
5576 u32 tail, off = req->timeout.off;
5578 spin_lock_irq(&ctx->completion_lock);
5581 * sqe->off holds how many events that need to occur for this
5582 * timeout event to be satisfied. If it isn't set, then this is
5583 * a pure timeout request, sequence isn't used.
5585 if (io_is_timeout_noseq(req)) {
5586 entry = ctx->timeout_list.prev;
5590 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5591 req->timeout.target_seq = tail + off;
5593 /* Update the last seq here in case io_flush_timeouts() hasn't.
5594 * This is safe because ->completion_lock is held, and submissions
5595 * and completions are never mixed in the same ->completion_lock section.
5597 ctx->cq_last_tm_flush = tail;
5600 * Insertion sort, ensuring the first entry in the list is always
5601 * the one we need first.
5603 list_for_each_prev(entry, &ctx->timeout_list) {
5604 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5607 if (io_is_timeout_noseq(nxt))
5609 /* nxt.seq is behind @tail, otherwise would've been completed */
5610 if (off >= nxt->timeout.target_seq - tail)
5614 list_add(&req->timeout.list, entry);
5615 data->timer.function = io_timeout_fn;
5616 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5617 spin_unlock_irq(&ctx->completion_lock);
5621 struct io_cancel_data {
5622 struct io_ring_ctx *ctx;
5626 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5628 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5629 struct io_cancel_data *cd = data;
5631 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5634 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5635 struct io_ring_ctx *ctx)
5637 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5638 enum io_wq_cancel cancel_ret;
5641 if (!tctx || !tctx->io_wq)
5644 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5645 switch (cancel_ret) {
5646 case IO_WQ_CANCEL_OK:
5649 case IO_WQ_CANCEL_RUNNING:
5652 case IO_WQ_CANCEL_NOTFOUND:
5660 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5661 struct io_kiocb *req, __u64 sqe_addr,
5664 unsigned long flags;
5667 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5668 if (ret != -ENOENT) {
5669 spin_lock_irqsave(&ctx->completion_lock, flags);
5673 spin_lock_irqsave(&ctx->completion_lock, flags);
5674 ret = io_timeout_cancel(ctx, sqe_addr);
5677 ret = io_poll_cancel(ctx, sqe_addr);
5681 io_cqring_fill_event(req, ret);
5682 io_commit_cqring(ctx);
5683 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5684 io_cqring_ev_posted(ctx);
5687 req_set_fail_links(req);
5691 static int io_async_cancel_prep(struct io_kiocb *req,
5692 const struct io_uring_sqe *sqe)
5694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5696 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5698 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5701 req->cancel.addr = READ_ONCE(sqe->addr);
5705 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5707 struct io_ring_ctx *ctx = req->ctx;
5708 u64 sqe_addr = req->cancel.addr;
5709 struct io_tctx_node *node;
5712 /* tasks should wait for their io-wq threads, so safe w/o sync */
5713 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5714 spin_lock_irq(&ctx->completion_lock);
5717 ret = io_timeout_cancel(ctx, sqe_addr);
5720 ret = io_poll_cancel(ctx, sqe_addr);
5723 spin_unlock_irq(&ctx->completion_lock);
5725 /* slow path, try all io-wq's */
5726 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5728 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5729 struct io_uring_task *tctx = node->task->io_uring;
5731 if (!tctx || !tctx->io_wq)
5733 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5737 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5739 spin_lock_irq(&ctx->completion_lock);
5741 io_cqring_fill_event(req, ret);
5742 io_commit_cqring(ctx);
5743 spin_unlock_irq(&ctx->completion_lock);
5744 io_cqring_ev_posted(ctx);
5747 req_set_fail_links(req);
5752 static int io_rsrc_update_prep(struct io_kiocb *req,
5753 const struct io_uring_sqe *sqe)
5755 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5757 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5759 if (sqe->ioprio || sqe->rw_flags)
5762 req->rsrc_update.offset = READ_ONCE(sqe->off);
5763 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5764 if (!req->rsrc_update.nr_args)
5766 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5770 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5772 struct io_ring_ctx *ctx = req->ctx;
5773 struct io_uring_rsrc_update up;
5776 if (issue_flags & IO_URING_F_NONBLOCK)
5779 up.offset = req->rsrc_update.offset;
5780 up.data = req->rsrc_update.arg;
5782 mutex_lock(&ctx->uring_lock);
5783 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5784 mutex_unlock(&ctx->uring_lock);
5787 req_set_fail_links(req);
5788 __io_req_complete(req, issue_flags, ret, 0);
5792 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5794 switch (req->opcode) {
5797 case IORING_OP_READV:
5798 case IORING_OP_READ_FIXED:
5799 case IORING_OP_READ:
5800 return io_read_prep(req, sqe);
5801 case IORING_OP_WRITEV:
5802 case IORING_OP_WRITE_FIXED:
5803 case IORING_OP_WRITE:
5804 return io_write_prep(req, sqe);
5805 case IORING_OP_POLL_ADD:
5806 return io_poll_add_prep(req, sqe);
5807 case IORING_OP_POLL_REMOVE:
5808 return io_poll_remove_prep(req, sqe);
5809 case IORING_OP_FSYNC:
5810 return io_fsync_prep(req, sqe);
5811 case IORING_OP_SYNC_FILE_RANGE:
5812 return io_sfr_prep(req, sqe);
5813 case IORING_OP_SENDMSG:
5814 case IORING_OP_SEND:
5815 return io_sendmsg_prep(req, sqe);
5816 case IORING_OP_RECVMSG:
5817 case IORING_OP_RECV:
5818 return io_recvmsg_prep(req, sqe);
5819 case IORING_OP_CONNECT:
5820 return io_connect_prep(req, sqe);
5821 case IORING_OP_TIMEOUT:
5822 return io_timeout_prep(req, sqe, false);
5823 case IORING_OP_TIMEOUT_REMOVE:
5824 return io_timeout_remove_prep(req, sqe);
5825 case IORING_OP_ASYNC_CANCEL:
5826 return io_async_cancel_prep(req, sqe);
5827 case IORING_OP_LINK_TIMEOUT:
5828 return io_timeout_prep(req, sqe, true);
5829 case IORING_OP_ACCEPT:
5830 return io_accept_prep(req, sqe);
5831 case IORING_OP_FALLOCATE:
5832 return io_fallocate_prep(req, sqe);
5833 case IORING_OP_OPENAT:
5834 return io_openat_prep(req, sqe);
5835 case IORING_OP_CLOSE:
5836 return io_close_prep(req, sqe);
5837 case IORING_OP_FILES_UPDATE:
5838 return io_rsrc_update_prep(req, sqe);
5839 case IORING_OP_STATX:
5840 return io_statx_prep(req, sqe);
5841 case IORING_OP_FADVISE:
5842 return io_fadvise_prep(req, sqe);
5843 case IORING_OP_MADVISE:
5844 return io_madvise_prep(req, sqe);
5845 case IORING_OP_OPENAT2:
5846 return io_openat2_prep(req, sqe);
5847 case IORING_OP_EPOLL_CTL:
5848 return io_epoll_ctl_prep(req, sqe);
5849 case IORING_OP_SPLICE:
5850 return io_splice_prep(req, sqe);
5851 case IORING_OP_PROVIDE_BUFFERS:
5852 return io_provide_buffers_prep(req, sqe);
5853 case IORING_OP_REMOVE_BUFFERS:
5854 return io_remove_buffers_prep(req, sqe);
5856 return io_tee_prep(req, sqe);
5857 case IORING_OP_SHUTDOWN:
5858 return io_shutdown_prep(req, sqe);
5859 case IORING_OP_RENAMEAT:
5860 return io_renameat_prep(req, sqe);
5861 case IORING_OP_UNLINKAT:
5862 return io_unlinkat_prep(req, sqe);
5865 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5870 static int io_req_prep_async(struct io_kiocb *req)
5872 switch (req->opcode) {
5873 case IORING_OP_READV:
5874 case IORING_OP_READ_FIXED:
5875 case IORING_OP_READ:
5876 return io_rw_prep_async(req, READ);
5877 case IORING_OP_WRITEV:
5878 case IORING_OP_WRITE_FIXED:
5879 case IORING_OP_WRITE:
5880 return io_rw_prep_async(req, WRITE);
5881 case IORING_OP_SENDMSG:
5882 case IORING_OP_SEND:
5883 return io_sendmsg_prep_async(req);
5884 case IORING_OP_RECVMSG:
5885 case IORING_OP_RECV:
5886 return io_recvmsg_prep_async(req);
5887 case IORING_OP_CONNECT:
5888 return io_connect_prep_async(req);
5893 static int io_req_defer_prep(struct io_kiocb *req)
5895 if (!io_op_defs[req->opcode].needs_async_data)
5897 /* some opcodes init it during the inital prep */
5898 if (req->async_data)
5900 if (__io_alloc_async_data(req))
5902 return io_req_prep_async(req);
5905 static u32 io_get_sequence(struct io_kiocb *req)
5907 struct io_kiocb *pos;
5908 struct io_ring_ctx *ctx = req->ctx;
5909 u32 total_submitted, nr_reqs = 0;
5911 io_for_each_link(pos, req)
5914 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5915 return total_submitted - nr_reqs;
5918 static int io_req_defer(struct io_kiocb *req)
5920 struct io_ring_ctx *ctx = req->ctx;
5921 struct io_defer_entry *de;
5925 /* Still need defer if there is pending req in defer list. */
5926 if (likely(list_empty_careful(&ctx->defer_list) &&
5927 !(req->flags & REQ_F_IO_DRAIN)))
5930 seq = io_get_sequence(req);
5931 /* Still a chance to pass the sequence check */
5932 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5935 ret = io_req_defer_prep(req);
5938 io_prep_async_link(req);
5939 de = kmalloc(sizeof(*de), GFP_KERNEL);
5943 spin_lock_irq(&ctx->completion_lock);
5944 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5945 spin_unlock_irq(&ctx->completion_lock);
5947 io_queue_async_work(req);
5948 return -EIOCBQUEUED;
5951 trace_io_uring_defer(ctx, req, req->user_data);
5954 list_add_tail(&de->list, &ctx->defer_list);
5955 spin_unlock_irq(&ctx->completion_lock);
5956 return -EIOCBQUEUED;
5959 static void __io_clean_op(struct io_kiocb *req)
5961 if (req->flags & REQ_F_BUFFER_SELECTED) {
5962 switch (req->opcode) {
5963 case IORING_OP_READV:
5964 case IORING_OP_READ_FIXED:
5965 case IORING_OP_READ:
5966 kfree((void *)(unsigned long)req->rw.addr);
5968 case IORING_OP_RECVMSG:
5969 case IORING_OP_RECV:
5970 kfree(req->sr_msg.kbuf);
5973 req->flags &= ~REQ_F_BUFFER_SELECTED;
5976 if (req->flags & REQ_F_NEED_CLEANUP) {
5977 switch (req->opcode) {
5978 case IORING_OP_READV:
5979 case IORING_OP_READ_FIXED:
5980 case IORING_OP_READ:
5981 case IORING_OP_WRITEV:
5982 case IORING_OP_WRITE_FIXED:
5983 case IORING_OP_WRITE: {
5984 struct io_async_rw *io = req->async_data;
5986 kfree(io->free_iovec);
5989 case IORING_OP_RECVMSG:
5990 case IORING_OP_SENDMSG: {
5991 struct io_async_msghdr *io = req->async_data;
5993 kfree(io->free_iov);
5996 case IORING_OP_SPLICE:
5998 io_put_file(req, req->splice.file_in,
5999 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6001 case IORING_OP_OPENAT:
6002 case IORING_OP_OPENAT2:
6003 if (req->open.filename)
6004 putname(req->open.filename);
6006 case IORING_OP_RENAMEAT:
6007 putname(req->rename.oldpath);
6008 putname(req->rename.newpath);
6010 case IORING_OP_UNLINKAT:
6011 putname(req->unlink.filename);
6014 req->flags &= ~REQ_F_NEED_CLEANUP;
6018 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6020 struct io_ring_ctx *ctx = req->ctx;
6021 const struct cred *creds = NULL;
6024 if (req->work.creds && req->work.creds != current_cred())
6025 creds = override_creds(req->work.creds);
6027 switch (req->opcode) {
6029 ret = io_nop(req, issue_flags);
6031 case IORING_OP_READV:
6032 case IORING_OP_READ_FIXED:
6033 case IORING_OP_READ:
6034 ret = io_read(req, issue_flags);
6036 case IORING_OP_WRITEV:
6037 case IORING_OP_WRITE_FIXED:
6038 case IORING_OP_WRITE:
6039 ret = io_write(req, issue_flags);
6041 case IORING_OP_FSYNC:
6042 ret = io_fsync(req, issue_flags);
6044 case IORING_OP_POLL_ADD:
6045 ret = io_poll_add(req, issue_flags);
6047 case IORING_OP_POLL_REMOVE:
6048 ret = io_poll_remove(req, issue_flags);
6050 case IORING_OP_SYNC_FILE_RANGE:
6051 ret = io_sync_file_range(req, issue_flags);
6053 case IORING_OP_SENDMSG:
6054 ret = io_sendmsg(req, issue_flags);
6056 case IORING_OP_SEND:
6057 ret = io_send(req, issue_flags);
6059 case IORING_OP_RECVMSG:
6060 ret = io_recvmsg(req, issue_flags);
6062 case IORING_OP_RECV:
6063 ret = io_recv(req, issue_flags);
6065 case IORING_OP_TIMEOUT:
6066 ret = io_timeout(req, issue_flags);
6068 case IORING_OP_TIMEOUT_REMOVE:
6069 ret = io_timeout_remove(req, issue_flags);
6071 case IORING_OP_ACCEPT:
6072 ret = io_accept(req, issue_flags);
6074 case IORING_OP_CONNECT:
6075 ret = io_connect(req, issue_flags);
6077 case IORING_OP_ASYNC_CANCEL:
6078 ret = io_async_cancel(req, issue_flags);
6080 case IORING_OP_FALLOCATE:
6081 ret = io_fallocate(req, issue_flags);
6083 case IORING_OP_OPENAT:
6084 ret = io_openat(req, issue_flags);
6086 case IORING_OP_CLOSE:
6087 ret = io_close(req, issue_flags);
6089 case IORING_OP_FILES_UPDATE:
6090 ret = io_files_update(req, issue_flags);
6092 case IORING_OP_STATX:
6093 ret = io_statx(req, issue_flags);
6095 case IORING_OP_FADVISE:
6096 ret = io_fadvise(req, issue_flags);
6098 case IORING_OP_MADVISE:
6099 ret = io_madvise(req, issue_flags);
6101 case IORING_OP_OPENAT2:
6102 ret = io_openat2(req, issue_flags);
6104 case IORING_OP_EPOLL_CTL:
6105 ret = io_epoll_ctl(req, issue_flags);
6107 case IORING_OP_SPLICE:
6108 ret = io_splice(req, issue_flags);
6110 case IORING_OP_PROVIDE_BUFFERS:
6111 ret = io_provide_buffers(req, issue_flags);
6113 case IORING_OP_REMOVE_BUFFERS:
6114 ret = io_remove_buffers(req, issue_flags);
6117 ret = io_tee(req, issue_flags);
6119 case IORING_OP_SHUTDOWN:
6120 ret = io_shutdown(req, issue_flags);
6122 case IORING_OP_RENAMEAT:
6123 ret = io_renameat(req, issue_flags);
6125 case IORING_OP_UNLINKAT:
6126 ret = io_unlinkat(req, issue_flags);
6134 revert_creds(creds);
6139 /* If the op doesn't have a file, we're not polling for it */
6140 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6141 const bool in_async = io_wq_current_is_worker();
6143 /* workqueue context doesn't hold uring_lock, grab it now */
6145 mutex_lock(&ctx->uring_lock);
6147 io_iopoll_req_issued(req, in_async);
6150 mutex_unlock(&ctx->uring_lock);
6156 static void io_wq_submit_work(struct io_wq_work *work)
6158 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6159 struct io_kiocb *timeout;
6162 timeout = io_prep_linked_timeout(req);
6164 io_queue_linked_timeout(timeout);
6166 if (work->flags & IO_WQ_WORK_CANCEL)
6171 ret = io_issue_sqe(req, 0);
6173 * We can get EAGAIN for polled IO even though we're
6174 * forcing a sync submission from here, since we can't
6175 * wait for request slots on the block side.
6183 /* avoid locking problems by failing it from a clean context */
6185 /* io-wq is going to take one down */
6186 refcount_inc(&req->refs);
6187 io_req_task_queue_fail(req, ret);
6191 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6194 struct fixed_rsrc_table *table;
6196 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6197 return table->files[index & IORING_FILE_TABLE_MASK];
6200 static struct file *io_file_get(struct io_submit_state *state,
6201 struct io_kiocb *req, int fd, bool fixed)
6203 struct io_ring_ctx *ctx = req->ctx;
6207 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6209 fd = array_index_nospec(fd, ctx->nr_user_files);
6210 file = io_file_from_index(ctx, fd);
6211 io_set_resource_node(req);
6213 trace_io_uring_file_get(ctx, fd);
6214 file = __io_file_get(state, fd);
6217 if (file && unlikely(file->f_op == &io_uring_fops))
6218 io_req_track_inflight(req);
6222 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6224 struct io_timeout_data *data = container_of(timer,
6225 struct io_timeout_data, timer);
6226 struct io_kiocb *prev, *req = data->req;
6227 struct io_ring_ctx *ctx = req->ctx;
6228 unsigned long flags;
6230 spin_lock_irqsave(&ctx->completion_lock, flags);
6231 prev = req->timeout.head;
6232 req->timeout.head = NULL;
6235 * We don't expect the list to be empty, that will only happen if we
6236 * race with the completion of the linked work.
6238 if (prev && refcount_inc_not_zero(&prev->refs))
6239 io_remove_next_linked(prev);
6242 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6245 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6246 io_put_req_deferred(prev, 1);
6248 io_req_complete_post(req, -ETIME, 0);
6249 io_put_req_deferred(req, 1);
6251 return HRTIMER_NORESTART;
6254 static void __io_queue_linked_timeout(struct io_kiocb *req)
6257 * If the back reference is NULL, then our linked request finished
6258 * before we got a chance to setup the timer
6260 if (req->timeout.head) {
6261 struct io_timeout_data *data = req->async_data;
6263 data->timer.function = io_link_timeout_fn;
6264 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6269 static void io_queue_linked_timeout(struct io_kiocb *req)
6271 struct io_ring_ctx *ctx = req->ctx;
6273 spin_lock_irq(&ctx->completion_lock);
6274 __io_queue_linked_timeout(req);
6275 spin_unlock_irq(&ctx->completion_lock);
6277 /* drop submission reference */
6281 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6283 struct io_kiocb *nxt = req->link;
6285 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6286 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6289 nxt->timeout.head = req;
6290 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6291 req->flags |= REQ_F_LINK_TIMEOUT;
6295 static void __io_queue_sqe(struct io_kiocb *req)
6297 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6300 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6303 * We async punt it if the file wasn't marked NOWAIT, or if the file
6304 * doesn't support non-blocking read/write attempts
6306 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6307 if (!io_arm_poll_handler(req)) {
6309 * Queued up for async execution, worker will release
6310 * submit reference when the iocb is actually submitted.
6312 io_queue_async_work(req);
6314 } else if (likely(!ret)) {
6315 /* drop submission reference */
6316 if (req->flags & REQ_F_COMPLETE_INLINE) {
6317 struct io_ring_ctx *ctx = req->ctx;
6318 struct io_comp_state *cs = &ctx->submit_state.comp;
6320 cs->reqs[cs->nr++] = req;
6321 if (cs->nr == ARRAY_SIZE(cs->reqs))
6322 io_submit_flush_completions(cs, ctx);
6327 req_set_fail_links(req);
6329 io_req_complete(req, ret);
6332 io_queue_linked_timeout(linked_timeout);
6335 static void io_queue_sqe(struct io_kiocb *req)
6339 ret = io_req_defer(req);
6341 if (ret != -EIOCBQUEUED) {
6343 req_set_fail_links(req);
6345 io_req_complete(req, ret);
6347 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6348 ret = io_req_defer_prep(req);
6351 io_queue_async_work(req);
6353 __io_queue_sqe(req);
6358 * Check SQE restrictions (opcode and flags).
6360 * Returns 'true' if SQE is allowed, 'false' otherwise.
6362 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6363 struct io_kiocb *req,
6364 unsigned int sqe_flags)
6366 if (!ctx->restricted)
6369 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6372 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6373 ctx->restrictions.sqe_flags_required)
6376 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6377 ctx->restrictions.sqe_flags_required))
6383 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6384 const struct io_uring_sqe *sqe)
6386 struct io_submit_state *state;
6387 unsigned int sqe_flags;
6388 int personality, ret = 0;
6390 req->opcode = READ_ONCE(sqe->opcode);
6391 /* same numerical values with corresponding REQ_F_*, safe to copy */
6392 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6393 req->user_data = READ_ONCE(sqe->user_data);
6394 req->async_data = NULL;
6398 req->fixed_rsrc_refs = NULL;
6399 /* one is dropped after submission, the other at completion */
6400 refcount_set(&req->refs, 2);
6401 req->task = current;
6403 req->work.list.next = NULL;
6404 req->work.creds = NULL;
6405 req->work.flags = 0;
6407 /* enforce forwards compatibility on users */
6408 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6413 if (unlikely(req->opcode >= IORING_OP_LAST))
6416 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6419 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6420 !io_op_defs[req->opcode].buffer_select)
6423 personality = READ_ONCE(sqe->personality);
6425 req->work.creds = xa_load(&ctx->personalities, personality);
6426 if (!req->work.creds)
6428 get_cred(req->work.creds);
6430 state = &ctx->submit_state;
6433 * Plug now if we have more than 1 IO left after this, and the target
6434 * is potentially a read/write to block based storage.
6436 if (!state->plug_started && state->ios_left > 1 &&
6437 io_op_defs[req->opcode].plug) {
6438 blk_start_plug(&state->plug);
6439 state->plug_started = true;
6442 if (io_op_defs[req->opcode].needs_file) {
6443 bool fixed = req->flags & REQ_F_FIXED_FILE;
6445 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6446 if (unlikely(!req->file))
6454 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6455 const struct io_uring_sqe *sqe)
6457 struct io_submit_link *link = &ctx->submit_state.link;
6460 ret = io_init_req(ctx, req, sqe);
6461 if (unlikely(ret)) {
6464 /* fail even hard links since we don't submit */
6465 link->head->flags |= REQ_F_FAIL_LINK;
6466 io_put_req(link->head);
6467 io_req_complete(link->head, -ECANCELED);
6471 io_req_complete(req, ret);
6474 ret = io_req_prep(req, sqe);
6478 /* don't need @sqe from now on */
6479 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6480 true, ctx->flags & IORING_SETUP_SQPOLL);
6483 * If we already have a head request, queue this one for async
6484 * submittal once the head completes. If we don't have a head but
6485 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6486 * submitted sync once the chain is complete. If none of those
6487 * conditions are true (normal request), then just queue it.
6490 struct io_kiocb *head = link->head;
6493 * Taking sequential execution of a link, draining both sides
6494 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6495 * requests in the link. So, it drains the head and the
6496 * next after the link request. The last one is done via
6497 * drain_next flag to persist the effect across calls.
6499 if (req->flags & REQ_F_IO_DRAIN) {
6500 head->flags |= REQ_F_IO_DRAIN;
6501 ctx->drain_next = 1;
6503 ret = io_req_defer_prep(req);
6506 trace_io_uring_link(ctx, req, head);
6507 link->last->link = req;
6510 /* last request of a link, enqueue the link */
6511 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6516 if (unlikely(ctx->drain_next)) {
6517 req->flags |= REQ_F_IO_DRAIN;
6518 ctx->drain_next = 0;
6520 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6532 * Batched submission is done, ensure local IO is flushed out.
6534 static void io_submit_state_end(struct io_submit_state *state,
6535 struct io_ring_ctx *ctx)
6537 if (state->link.head)
6538 io_queue_sqe(state->link.head);
6540 io_submit_flush_completions(&state->comp, ctx);
6541 if (state->plug_started)
6542 blk_finish_plug(&state->plug);
6543 io_state_file_put(state);
6547 * Start submission side cache.
6549 static void io_submit_state_start(struct io_submit_state *state,
6550 unsigned int max_ios)
6552 state->plug_started = false;
6553 state->ios_left = max_ios;
6554 /* set only head, no need to init link_last in advance */
6555 state->link.head = NULL;
6558 static void io_commit_sqring(struct io_ring_ctx *ctx)
6560 struct io_rings *rings = ctx->rings;
6563 * Ensure any loads from the SQEs are done at this point,
6564 * since once we write the new head, the application could
6565 * write new data to them.
6567 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6571 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6572 * that is mapped by userspace. This means that care needs to be taken to
6573 * ensure that reads are stable, as we cannot rely on userspace always
6574 * being a good citizen. If members of the sqe are validated and then later
6575 * used, it's important that those reads are done through READ_ONCE() to
6576 * prevent a re-load down the line.
6578 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6580 u32 *sq_array = ctx->sq_array;
6584 * The cached sq head (or cq tail) serves two purposes:
6586 * 1) allows us to batch the cost of updating the user visible
6588 * 2) allows the kernel side to track the head on its own, even
6589 * though the application is the one updating it.
6591 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6592 if (likely(head < ctx->sq_entries))
6593 return &ctx->sq_sqes[head];
6595 /* drop invalid entries */
6596 ctx->cached_sq_dropped++;
6597 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6601 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6605 /* if we have a backlog and couldn't flush it all, return BUSY */
6606 if (test_bit(0, &ctx->sq_check_overflow)) {
6607 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6611 /* make sure SQ entry isn't read before tail */
6612 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6614 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6617 percpu_counter_add(¤t->io_uring->inflight, nr);
6618 refcount_add(nr, ¤t->usage);
6619 io_submit_state_start(&ctx->submit_state, nr);
6621 while (submitted < nr) {
6622 const struct io_uring_sqe *sqe;
6623 struct io_kiocb *req;
6625 req = io_alloc_req(ctx);
6626 if (unlikely(!req)) {
6628 submitted = -EAGAIN;
6631 sqe = io_get_sqe(ctx);
6632 if (unlikely(!sqe)) {
6633 kmem_cache_free(req_cachep, req);
6636 /* will complete beyond this point, count as submitted */
6638 if (io_submit_sqe(ctx, req, sqe))
6642 if (unlikely(submitted != nr)) {
6643 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6644 struct io_uring_task *tctx = current->io_uring;
6645 int unused = nr - ref_used;
6647 percpu_ref_put_many(&ctx->refs, unused);
6648 percpu_counter_sub(&tctx->inflight, unused);
6649 put_task_struct_many(current, unused);
6652 io_submit_state_end(&ctx->submit_state, ctx);
6653 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6654 io_commit_sqring(ctx);
6659 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6661 /* Tell userspace we may need a wakeup call */
6662 spin_lock_irq(&ctx->completion_lock);
6663 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6664 spin_unlock_irq(&ctx->completion_lock);
6667 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6669 spin_lock_irq(&ctx->completion_lock);
6670 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6671 spin_unlock_irq(&ctx->completion_lock);
6674 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6676 unsigned int to_submit;
6679 to_submit = io_sqring_entries(ctx);
6680 /* if we're handling multiple rings, cap submit size for fairness */
6681 if (cap_entries && to_submit > 8)
6684 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6685 unsigned nr_events = 0;
6687 mutex_lock(&ctx->uring_lock);
6688 if (!list_empty(&ctx->iopoll_list))
6689 io_do_iopoll(ctx, &nr_events, 0);
6691 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6692 !(ctx->flags & IORING_SETUP_R_DISABLED))
6693 ret = io_submit_sqes(ctx, to_submit);
6694 mutex_unlock(&ctx->uring_lock);
6697 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6698 wake_up(&ctx->sqo_sq_wait);
6703 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6705 struct io_ring_ctx *ctx;
6706 unsigned sq_thread_idle = 0;
6708 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6709 if (sq_thread_idle < ctx->sq_thread_idle)
6710 sq_thread_idle = ctx->sq_thread_idle;
6713 sqd->sq_thread_idle = sq_thread_idle;
6716 static int io_sq_thread(void *data)
6718 struct io_sq_data *sqd = data;
6719 struct io_ring_ctx *ctx;
6720 unsigned long timeout = 0;
6721 char buf[TASK_COMM_LEN];
6724 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6725 set_task_comm(current, buf);
6726 current->pf_io_worker = NULL;
6728 if (sqd->sq_cpu != -1)
6729 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6731 set_cpus_allowed_ptr(current, cpu_online_mask);
6732 current->flags |= PF_NO_SETAFFINITY;
6734 mutex_lock(&sqd->lock);
6735 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6737 bool cap_entries, sqt_spin, needs_sched;
6739 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6740 mutex_unlock(&sqd->lock);
6742 mutex_lock(&sqd->lock);
6744 io_run_task_work_head(&sqd->park_task_work);
6745 timeout = jiffies + sqd->sq_thread_idle;
6748 if (signal_pending(current)) {
6749 struct ksignal ksig;
6751 if (!get_signal(&ksig))
6756 cap_entries = !list_is_singular(&sqd->ctx_list);
6757 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6758 const struct cred *creds = NULL;
6760 if (ctx->sq_creds != current_cred())
6761 creds = override_creds(ctx->sq_creds);
6762 ret = __io_sq_thread(ctx, cap_entries);
6764 revert_creds(creds);
6765 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6769 if (sqt_spin || !time_after(jiffies, timeout)) {
6773 timeout = jiffies + sqd->sq_thread_idle;
6778 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6779 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6780 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6781 !list_empty_careful(&ctx->iopoll_list)) {
6782 needs_sched = false;
6785 if (io_sqring_entries(ctx)) {
6786 needs_sched = false;
6791 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6793 io_ring_set_wakeup_flag(ctx);
6795 mutex_unlock(&sqd->lock);
6797 mutex_lock(&sqd->lock);
6798 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6799 io_ring_clear_wakeup_flag(ctx);
6802 finish_wait(&sqd->wait, &wait);
6803 io_run_task_work_head(&sqd->park_task_work);
6804 timeout = jiffies + sqd->sq_thread_idle;
6807 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6808 io_uring_cancel_sqpoll(ctx);
6810 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6811 io_ring_set_wakeup_flag(ctx);
6812 mutex_unlock(&sqd->lock);
6815 io_run_task_work_head(&sqd->park_task_work);
6816 complete(&sqd->exited);
6820 struct io_wait_queue {
6821 struct wait_queue_entry wq;
6822 struct io_ring_ctx *ctx;
6824 unsigned nr_timeouts;
6827 static inline bool io_should_wake(struct io_wait_queue *iowq)
6829 struct io_ring_ctx *ctx = iowq->ctx;
6832 * Wake up if we have enough events, or if a timeout occurred since we
6833 * started waiting. For timeouts, we always want to return to userspace,
6834 * regardless of event count.
6836 return io_cqring_events(ctx) >= iowq->to_wait ||
6837 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6840 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6841 int wake_flags, void *key)
6843 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6847 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6848 * the task, and the next invocation will do it.
6850 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6851 return autoremove_wake_function(curr, mode, wake_flags, key);
6855 static int io_run_task_work_sig(void)
6857 if (io_run_task_work())
6859 if (!signal_pending(current))
6861 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6862 return -ERESTARTSYS;
6866 /* when returns >0, the caller should retry */
6867 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6868 struct io_wait_queue *iowq,
6869 signed long *timeout)
6873 /* make sure we run task_work before checking for signals */
6874 ret = io_run_task_work_sig();
6875 if (ret || io_should_wake(iowq))
6877 /* let the caller flush overflows, retry */
6878 if (test_bit(0, &ctx->cq_check_overflow))
6881 *timeout = schedule_timeout(*timeout);
6882 return !*timeout ? -ETIME : 1;
6886 * Wait until events become available, if we don't already have some. The
6887 * application must reap them itself, as they reside on the shared cq ring.
6889 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6890 const sigset_t __user *sig, size_t sigsz,
6891 struct __kernel_timespec __user *uts)
6893 struct io_wait_queue iowq = {
6896 .func = io_wake_function,
6897 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6900 .to_wait = min_events,
6902 struct io_rings *rings = ctx->rings;
6903 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6907 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6908 if (io_cqring_events(ctx) >= min_events)
6910 if (!io_run_task_work())
6915 #ifdef CONFIG_COMPAT
6916 if (in_compat_syscall())
6917 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6921 ret = set_user_sigmask(sig, sigsz);
6928 struct timespec64 ts;
6930 if (get_timespec64(&ts, uts))
6932 timeout = timespec64_to_jiffies(&ts);
6935 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6936 trace_io_uring_cqring_wait(ctx, min_events);
6938 /* if we can't even flush overflow, don't wait for more */
6939 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6943 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6944 TASK_INTERRUPTIBLE);
6945 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6946 finish_wait(&ctx->wait, &iowq.wq);
6950 restore_saved_sigmask_unless(ret == -EINTR);
6952 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6955 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6957 #if defined(CONFIG_UNIX)
6958 if (ctx->ring_sock) {
6959 struct sock *sock = ctx->ring_sock->sk;
6960 struct sk_buff *skb;
6962 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6968 for (i = 0; i < ctx->nr_user_files; i++) {
6971 file = io_file_from_index(ctx, i);
6978 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6980 struct fixed_rsrc_data *data;
6982 data = container_of(ref, struct fixed_rsrc_data, refs);
6983 complete(&data->done);
6986 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6988 spin_lock_bh(&ctx->rsrc_ref_lock);
6991 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6993 spin_unlock_bh(&ctx->rsrc_ref_lock);
6996 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6997 struct fixed_rsrc_data *rsrc_data,
6998 struct fixed_rsrc_ref_node *ref_node)
7000 io_rsrc_ref_lock(ctx);
7001 rsrc_data->node = ref_node;
7002 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7003 io_rsrc_ref_unlock(ctx);
7004 percpu_ref_get(&rsrc_data->refs);
7007 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7009 struct fixed_rsrc_ref_node *ref_node = NULL;
7011 io_rsrc_ref_lock(ctx);
7012 ref_node = data->node;
7014 io_rsrc_ref_unlock(ctx);
7016 percpu_ref_kill(&ref_node->refs);
7019 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7020 struct io_ring_ctx *ctx,
7021 void (*rsrc_put)(struct io_ring_ctx *ctx,
7022 struct io_rsrc_put *prsrc))
7024 struct fixed_rsrc_ref_node *backup_node;
7030 data->quiesce = true;
7033 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7036 backup_node->rsrc_data = data;
7037 backup_node->rsrc_put = rsrc_put;
7039 io_sqe_rsrc_kill_node(ctx, data);
7040 percpu_ref_kill(&data->refs);
7041 flush_delayed_work(&ctx->rsrc_put_work);
7043 ret = wait_for_completion_interruptible(&data->done);
7047 percpu_ref_resurrect(&data->refs);
7048 io_sqe_rsrc_set_node(ctx, data, backup_node);
7050 reinit_completion(&data->done);
7051 mutex_unlock(&ctx->uring_lock);
7052 ret = io_run_task_work_sig();
7053 mutex_lock(&ctx->uring_lock);
7055 data->quiesce = false;
7058 destroy_fixed_rsrc_ref_node(backup_node);
7062 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7064 struct fixed_rsrc_data *data;
7066 data = kzalloc(sizeof(*data), GFP_KERNEL);
7070 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7071 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7076 init_completion(&data->done);
7080 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7082 percpu_ref_exit(&data->refs);
7087 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7089 struct fixed_rsrc_data *data = ctx->file_data;
7090 unsigned nr_tables, i;
7094 * percpu_ref_is_dying() is to stop parallel files unregister
7095 * Since we possibly drop uring lock later in this function to
7098 if (!data || percpu_ref_is_dying(&data->refs))
7100 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7104 __io_sqe_files_unregister(ctx);
7105 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7106 for (i = 0; i < nr_tables; i++)
7107 kfree(data->table[i].files);
7108 free_fixed_rsrc_data(data);
7109 ctx->file_data = NULL;
7110 ctx->nr_user_files = 0;
7114 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7115 __releases(&sqd->lock)
7117 WARN_ON_ONCE(sqd->thread == current);
7120 * Do the dance but not conditional clear_bit() because it'd race with
7121 * other threads incrementing park_pending and setting the bit.
7123 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7124 if (atomic_dec_return(&sqd->park_pending))
7125 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7126 mutex_unlock(&sqd->lock);
7129 static void io_sq_thread_park(struct io_sq_data *sqd)
7130 __acquires(&sqd->lock)
7132 WARN_ON_ONCE(sqd->thread == current);
7134 atomic_inc(&sqd->park_pending);
7135 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7136 mutex_lock(&sqd->lock);
7138 wake_up_process(sqd->thread);
7141 static void io_sq_thread_stop(struct io_sq_data *sqd)
7143 WARN_ON_ONCE(sqd->thread == current);
7145 mutex_lock(&sqd->lock);
7146 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7148 wake_up_process(sqd->thread);
7149 mutex_unlock(&sqd->lock);
7150 wait_for_completion(&sqd->exited);
7153 static void io_put_sq_data(struct io_sq_data *sqd)
7155 if (refcount_dec_and_test(&sqd->refs)) {
7156 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7158 io_sq_thread_stop(sqd);
7163 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7165 struct io_sq_data *sqd = ctx->sq_data;
7168 io_sq_thread_park(sqd);
7169 list_del_init(&ctx->sqd_list);
7170 io_sqd_update_thread_idle(sqd);
7171 io_sq_thread_unpark(sqd);
7173 io_put_sq_data(sqd);
7174 ctx->sq_data = NULL;
7176 put_cred(ctx->sq_creds);
7180 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7182 struct io_ring_ctx *ctx_attach;
7183 struct io_sq_data *sqd;
7186 f = fdget(p->wq_fd);
7188 return ERR_PTR(-ENXIO);
7189 if (f.file->f_op != &io_uring_fops) {
7191 return ERR_PTR(-EINVAL);
7194 ctx_attach = f.file->private_data;
7195 sqd = ctx_attach->sq_data;
7198 return ERR_PTR(-EINVAL);
7200 if (sqd->task_tgid != current->tgid) {
7202 return ERR_PTR(-EPERM);
7205 refcount_inc(&sqd->refs);
7210 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7213 struct io_sq_data *sqd;
7216 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7217 sqd = io_attach_sq_data(p);
7222 /* fall through for EPERM case, setup new sqd/task */
7223 if (PTR_ERR(sqd) != -EPERM)
7227 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7229 return ERR_PTR(-ENOMEM);
7231 atomic_set(&sqd->park_pending, 0);
7232 refcount_set(&sqd->refs, 1);
7233 INIT_LIST_HEAD(&sqd->ctx_list);
7234 mutex_init(&sqd->lock);
7235 init_waitqueue_head(&sqd->wait);
7236 init_completion(&sqd->exited);
7240 #if defined(CONFIG_UNIX)
7242 * Ensure the UNIX gc is aware of our file set, so we are certain that
7243 * the io_uring can be safely unregistered on process exit, even if we have
7244 * loops in the file referencing.
7246 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7248 struct sock *sk = ctx->ring_sock->sk;
7249 struct scm_fp_list *fpl;
7250 struct sk_buff *skb;
7253 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7257 skb = alloc_skb(0, GFP_KERNEL);
7266 fpl->user = get_uid(current_user());
7267 for (i = 0; i < nr; i++) {
7268 struct file *file = io_file_from_index(ctx, i + offset);
7272 fpl->fp[nr_files] = get_file(file);
7273 unix_inflight(fpl->user, fpl->fp[nr_files]);
7278 fpl->max = SCM_MAX_FD;
7279 fpl->count = nr_files;
7280 UNIXCB(skb).fp = fpl;
7281 skb->destructor = unix_destruct_scm;
7282 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7283 skb_queue_head(&sk->sk_receive_queue, skb);
7285 for (i = 0; i < nr_files; i++)
7296 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7297 * causes regular reference counting to break down. We rely on the UNIX
7298 * garbage collection to take care of this problem for us.
7300 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7302 unsigned left, total;
7306 left = ctx->nr_user_files;
7308 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7310 ret = __io_sqe_files_scm(ctx, this_files, total);
7314 total += this_files;
7320 while (total < ctx->nr_user_files) {
7321 struct file *file = io_file_from_index(ctx, total);
7331 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7337 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7338 unsigned nr_tables, unsigned nr_files)
7342 for (i = 0; i < nr_tables; i++) {
7343 struct fixed_rsrc_table *table = &file_data->table[i];
7344 unsigned this_files;
7346 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7347 table->files = kcalloc(this_files, sizeof(struct file *),
7351 nr_files -= this_files;
7357 for (i = 0; i < nr_tables; i++) {
7358 struct fixed_rsrc_table *table = &file_data->table[i];
7359 kfree(table->files);
7364 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7366 struct file *file = prsrc->file;
7367 #if defined(CONFIG_UNIX)
7368 struct sock *sock = ctx->ring_sock->sk;
7369 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7370 struct sk_buff *skb;
7373 __skb_queue_head_init(&list);
7376 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7377 * remove this entry and rearrange the file array.
7379 skb = skb_dequeue(head);
7381 struct scm_fp_list *fp;
7383 fp = UNIXCB(skb).fp;
7384 for (i = 0; i < fp->count; i++) {
7387 if (fp->fp[i] != file)
7390 unix_notinflight(fp->user, fp->fp[i]);
7391 left = fp->count - 1 - i;
7393 memmove(&fp->fp[i], &fp->fp[i + 1],
7394 left * sizeof(struct file *));
7401 __skb_queue_tail(&list, skb);
7411 __skb_queue_tail(&list, skb);
7413 skb = skb_dequeue(head);
7416 if (skb_peek(&list)) {
7417 spin_lock_irq(&head->lock);
7418 while ((skb = __skb_dequeue(&list)) != NULL)
7419 __skb_queue_tail(head, skb);
7420 spin_unlock_irq(&head->lock);
7427 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7429 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7430 struct io_ring_ctx *ctx = rsrc_data->ctx;
7431 struct io_rsrc_put *prsrc, *tmp;
7433 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7434 list_del(&prsrc->list);
7435 ref_node->rsrc_put(ctx, prsrc);
7439 percpu_ref_exit(&ref_node->refs);
7441 percpu_ref_put(&rsrc_data->refs);
7444 static void io_rsrc_put_work(struct work_struct *work)
7446 struct io_ring_ctx *ctx;
7447 struct llist_node *node;
7449 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7450 node = llist_del_all(&ctx->rsrc_put_llist);
7453 struct fixed_rsrc_ref_node *ref_node;
7454 struct llist_node *next = node->next;
7456 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7457 __io_rsrc_put_work(ref_node);
7462 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7465 struct fixed_rsrc_table *table;
7467 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7468 return &table->files[i & IORING_FILE_TABLE_MASK];
7471 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7473 struct fixed_rsrc_ref_node *ref_node;
7474 struct fixed_rsrc_data *data;
7475 struct io_ring_ctx *ctx;
7476 bool first_add = false;
7479 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7480 data = ref_node->rsrc_data;
7483 io_rsrc_ref_lock(ctx);
7484 ref_node->done = true;
7486 while (!list_empty(&ctx->rsrc_ref_list)) {
7487 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7488 struct fixed_rsrc_ref_node, node);
7489 /* recycle ref nodes in order */
7490 if (!ref_node->done)
7492 list_del(&ref_node->node);
7493 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7495 io_rsrc_ref_unlock(ctx);
7497 if (percpu_ref_is_dying(&data->refs))
7501 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7503 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7506 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7507 struct io_ring_ctx *ctx)
7509 struct fixed_rsrc_ref_node *ref_node;
7511 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7515 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7520 INIT_LIST_HEAD(&ref_node->node);
7521 INIT_LIST_HEAD(&ref_node->rsrc_list);
7522 ref_node->done = false;
7526 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7527 struct fixed_rsrc_ref_node *ref_node)
7529 ref_node->rsrc_data = ctx->file_data;
7530 ref_node->rsrc_put = io_ring_file_put;
7533 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7535 percpu_ref_exit(&ref_node->refs);
7540 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7543 __s32 __user *fds = (__s32 __user *) arg;
7544 unsigned nr_tables, i;
7546 int fd, ret = -ENOMEM;
7547 struct fixed_rsrc_ref_node *ref_node;
7548 struct fixed_rsrc_data *file_data;
7554 if (nr_args > IORING_MAX_FIXED_FILES)
7557 file_data = alloc_fixed_rsrc_data(ctx);
7560 ctx->file_data = file_data;
7562 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7563 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7565 if (!file_data->table)
7568 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7571 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7572 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7576 /* allow sparse sets */
7586 * Don't allow io_uring instances to be registered. If UNIX
7587 * isn't enabled, then this causes a reference cycle and this
7588 * instance can never get freed. If UNIX is enabled we'll
7589 * handle it just fine, but there's still no point in allowing
7590 * a ring fd as it doesn't support regular read/write anyway.
7592 if (file->f_op == &io_uring_fops) {
7596 *io_fixed_file_slot(file_data, i) = file;
7599 ret = io_sqe_files_scm(ctx);
7601 io_sqe_files_unregister(ctx);
7605 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7607 io_sqe_files_unregister(ctx);
7610 init_fixed_file_ref_node(ctx, ref_node);
7612 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7615 for (i = 0; i < ctx->nr_user_files; i++) {
7616 file = io_file_from_index(ctx, i);
7620 for (i = 0; i < nr_tables; i++)
7621 kfree(file_data->table[i].files);
7622 ctx->nr_user_files = 0;
7624 free_fixed_rsrc_data(ctx->file_data);
7625 ctx->file_data = NULL;
7629 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7632 #if defined(CONFIG_UNIX)
7633 struct sock *sock = ctx->ring_sock->sk;
7634 struct sk_buff_head *head = &sock->sk_receive_queue;
7635 struct sk_buff *skb;
7638 * See if we can merge this file into an existing skb SCM_RIGHTS
7639 * file set. If there's no room, fall back to allocating a new skb
7640 * and filling it in.
7642 spin_lock_irq(&head->lock);
7643 skb = skb_peek(head);
7645 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7647 if (fpl->count < SCM_MAX_FD) {
7648 __skb_unlink(skb, head);
7649 spin_unlock_irq(&head->lock);
7650 fpl->fp[fpl->count] = get_file(file);
7651 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7653 spin_lock_irq(&head->lock);
7654 __skb_queue_head(head, skb);
7659 spin_unlock_irq(&head->lock);
7666 return __io_sqe_files_scm(ctx, 1, index);
7672 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7674 struct io_rsrc_put *prsrc;
7675 struct fixed_rsrc_ref_node *ref_node = data->node;
7677 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7682 list_add(&prsrc->list, &ref_node->rsrc_list);
7687 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7690 return io_queue_rsrc_removal(data, (void *)file);
7693 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7694 struct io_uring_rsrc_update *up,
7697 struct fixed_rsrc_data *data = ctx->file_data;
7698 struct fixed_rsrc_ref_node *ref_node;
7699 struct file *file, **file_slot;
7703 bool needs_switch = false;
7705 if (check_add_overflow(up->offset, nr_args, &done))
7707 if (done > ctx->nr_user_files)
7710 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7713 init_fixed_file_ref_node(ctx, ref_node);
7715 fds = u64_to_user_ptr(up->data);
7716 for (done = 0; done < nr_args; done++) {
7718 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7722 if (fd == IORING_REGISTER_FILES_SKIP)
7725 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7726 file_slot = io_fixed_file_slot(ctx->file_data, i);
7729 err = io_queue_file_removal(data, *file_slot);
7733 needs_switch = true;
7742 * Don't allow io_uring instances to be registered. If
7743 * UNIX isn't enabled, then this causes a reference
7744 * cycle and this instance can never get freed. If UNIX
7745 * is enabled we'll handle it just fine, but there's
7746 * still no point in allowing a ring fd as it doesn't
7747 * support regular read/write anyway.
7749 if (file->f_op == &io_uring_fops) {
7755 err = io_sqe_file_register(ctx, file, i);
7765 percpu_ref_kill(&data->node->refs);
7766 io_sqe_rsrc_set_node(ctx, data, ref_node);
7768 destroy_fixed_rsrc_ref_node(ref_node);
7770 return done ? done : err;
7773 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7776 struct io_uring_rsrc_update up;
7778 if (!ctx->file_data)
7782 if (copy_from_user(&up, arg, sizeof(up)))
7787 return __io_sqe_files_update(ctx, &up, nr_args);
7790 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7792 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7794 req = io_put_req_find_next(req);
7795 return req ? &req->work : NULL;
7798 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7800 struct io_wq_hash *hash;
7801 struct io_wq_data data;
7802 unsigned int concurrency;
7804 hash = ctx->hash_map;
7806 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7808 return ERR_PTR(-ENOMEM);
7809 refcount_set(&hash->refs, 1);
7810 init_waitqueue_head(&hash->wait);
7811 ctx->hash_map = hash;
7815 data.free_work = io_free_work;
7816 data.do_work = io_wq_submit_work;
7818 /* Do QD, or 4 * CPUS, whatever is smallest */
7819 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7821 return io_wq_create(concurrency, &data);
7824 static int io_uring_alloc_task_context(struct task_struct *task,
7825 struct io_ring_ctx *ctx)
7827 struct io_uring_task *tctx;
7830 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7831 if (unlikely(!tctx))
7834 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7835 if (unlikely(ret)) {
7840 tctx->io_wq = io_init_wq_offload(ctx);
7841 if (IS_ERR(tctx->io_wq)) {
7842 ret = PTR_ERR(tctx->io_wq);
7843 percpu_counter_destroy(&tctx->inflight);
7849 init_waitqueue_head(&tctx->wait);
7851 atomic_set(&tctx->in_idle, 0);
7852 task->io_uring = tctx;
7853 spin_lock_init(&tctx->task_lock);
7854 INIT_WQ_LIST(&tctx->task_list);
7855 tctx->task_state = 0;
7856 init_task_work(&tctx->task_work, tctx_task_work);
7860 void __io_uring_free(struct task_struct *tsk)
7862 struct io_uring_task *tctx = tsk->io_uring;
7864 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7865 WARN_ON_ONCE(tctx->io_wq);
7867 percpu_counter_destroy(&tctx->inflight);
7869 tsk->io_uring = NULL;
7872 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7873 struct io_uring_params *p)
7877 /* Retain compatibility with failing for an invalid attach attempt */
7878 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7879 IORING_SETUP_ATTACH_WQ) {
7882 f = fdget(p->wq_fd);
7885 if (f.file->f_op != &io_uring_fops) {
7891 if (ctx->flags & IORING_SETUP_SQPOLL) {
7892 struct task_struct *tsk;
7893 struct io_sq_data *sqd;
7897 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7900 sqd = io_get_sq_data(p, &attached);
7906 ctx->sq_creds = get_current_cred();
7908 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7909 if (!ctx->sq_thread_idle)
7910 ctx->sq_thread_idle = HZ;
7913 io_sq_thread_park(sqd);
7914 list_add(&ctx->sqd_list, &sqd->ctx_list);
7915 io_sqd_update_thread_idle(sqd);
7916 /* don't attach to a dying SQPOLL thread, would be racy */
7917 if (attached && !sqd->thread)
7919 io_sq_thread_unpark(sqd);
7926 if (p->flags & IORING_SETUP_SQ_AFF) {
7927 int cpu = p->sq_thread_cpu;
7930 if (cpu >= nr_cpu_ids)
7932 if (!cpu_online(cpu))
7940 sqd->task_pid = current->pid;
7941 sqd->task_tgid = current->tgid;
7942 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7949 ret = io_uring_alloc_task_context(tsk, ctx);
7950 wake_up_new_task(tsk);
7953 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7954 /* Can't have SQ_AFF without SQPOLL */
7961 io_sq_thread_finish(ctx);
7964 complete(&ctx->sq_data->exited);
7968 static inline void __io_unaccount_mem(struct user_struct *user,
7969 unsigned long nr_pages)
7971 atomic_long_sub(nr_pages, &user->locked_vm);
7974 static inline int __io_account_mem(struct user_struct *user,
7975 unsigned long nr_pages)
7977 unsigned long page_limit, cur_pages, new_pages;
7979 /* Don't allow more pages than we can safely lock */
7980 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7983 cur_pages = atomic_long_read(&user->locked_vm);
7984 new_pages = cur_pages + nr_pages;
7985 if (new_pages > page_limit)
7987 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7988 new_pages) != cur_pages);
7993 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7996 __io_unaccount_mem(ctx->user, nr_pages);
7998 if (ctx->mm_account)
7999 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8002 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8007 ret = __io_account_mem(ctx->user, nr_pages);
8012 if (ctx->mm_account)
8013 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8018 static void io_mem_free(void *ptr)
8025 page = virt_to_head_page(ptr);
8026 if (put_page_testzero(page))
8027 free_compound_page(page);
8030 static void *io_mem_alloc(size_t size)
8032 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8033 __GFP_NORETRY | __GFP_ACCOUNT;
8035 return (void *) __get_free_pages(gfp_flags, get_order(size));
8038 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8041 struct io_rings *rings;
8042 size_t off, sq_array_size;
8044 off = struct_size(rings, cqes, cq_entries);
8045 if (off == SIZE_MAX)
8049 off = ALIGN(off, SMP_CACHE_BYTES);
8057 sq_array_size = array_size(sizeof(u32), sq_entries);
8058 if (sq_array_size == SIZE_MAX)
8061 if (check_add_overflow(off, sq_array_size, &off))
8067 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8071 if (!ctx->user_bufs)
8074 for (i = 0; i < ctx->nr_user_bufs; i++) {
8075 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8077 for (j = 0; j < imu->nr_bvecs; j++)
8078 unpin_user_page(imu->bvec[j].bv_page);
8080 if (imu->acct_pages)
8081 io_unaccount_mem(ctx, imu->acct_pages);
8086 kfree(ctx->user_bufs);
8087 ctx->user_bufs = NULL;
8088 ctx->nr_user_bufs = 0;
8092 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8093 void __user *arg, unsigned index)
8095 struct iovec __user *src;
8097 #ifdef CONFIG_COMPAT
8099 struct compat_iovec __user *ciovs;
8100 struct compat_iovec ciov;
8102 ciovs = (struct compat_iovec __user *) arg;
8103 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8106 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8107 dst->iov_len = ciov.iov_len;
8111 src = (struct iovec __user *) arg;
8112 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8118 * Not super efficient, but this is just a registration time. And we do cache
8119 * the last compound head, so generally we'll only do a full search if we don't
8122 * We check if the given compound head page has already been accounted, to
8123 * avoid double accounting it. This allows us to account the full size of the
8124 * page, not just the constituent pages of a huge page.
8126 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8127 int nr_pages, struct page *hpage)
8131 /* check current page array */
8132 for (i = 0; i < nr_pages; i++) {
8133 if (!PageCompound(pages[i]))
8135 if (compound_head(pages[i]) == hpage)
8139 /* check previously registered pages */
8140 for (i = 0; i < ctx->nr_user_bufs; i++) {
8141 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8143 for (j = 0; j < imu->nr_bvecs; j++) {
8144 if (!PageCompound(imu->bvec[j].bv_page))
8146 if (compound_head(imu->bvec[j].bv_page) == hpage)
8154 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8155 int nr_pages, struct io_mapped_ubuf *imu,
8156 struct page **last_hpage)
8160 for (i = 0; i < nr_pages; i++) {
8161 if (!PageCompound(pages[i])) {
8166 hpage = compound_head(pages[i]);
8167 if (hpage == *last_hpage)
8169 *last_hpage = hpage;
8170 if (headpage_already_acct(ctx, pages, i, hpage))
8172 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8176 if (!imu->acct_pages)
8179 ret = io_account_mem(ctx, imu->acct_pages);
8181 imu->acct_pages = 0;
8185 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8186 struct io_mapped_ubuf *imu,
8187 struct page **last_hpage)
8189 struct vm_area_struct **vmas = NULL;
8190 struct page **pages = NULL;
8191 unsigned long off, start, end, ubuf;
8193 int ret, pret, nr_pages, i;
8195 ubuf = (unsigned long) iov->iov_base;
8196 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8197 start = ubuf >> PAGE_SHIFT;
8198 nr_pages = end - start;
8202 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8206 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8211 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8217 mmap_read_lock(current->mm);
8218 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8220 if (pret == nr_pages) {
8221 /* don't support file backed memory */
8222 for (i = 0; i < nr_pages; i++) {
8223 struct vm_area_struct *vma = vmas[i];
8226 !is_file_hugepages(vma->vm_file)) {
8232 ret = pret < 0 ? pret : -EFAULT;
8234 mmap_read_unlock(current->mm);
8237 * if we did partial map, or found file backed vmas,
8238 * release any pages we did get
8241 unpin_user_pages(pages, pret);
8246 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8248 unpin_user_pages(pages, pret);
8253 off = ubuf & ~PAGE_MASK;
8254 size = iov->iov_len;
8255 for (i = 0; i < nr_pages; i++) {
8258 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8259 imu->bvec[i].bv_page = pages[i];
8260 imu->bvec[i].bv_len = vec_len;
8261 imu->bvec[i].bv_offset = off;
8265 /* store original address for later verification */
8267 imu->len = iov->iov_len;
8268 imu->nr_bvecs = nr_pages;
8276 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8280 if (!nr_args || nr_args > UIO_MAXIOV)
8283 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8285 if (!ctx->user_bufs)
8291 static int io_buffer_validate(struct iovec *iov)
8294 * Don't impose further limits on the size and buffer
8295 * constraints here, we'll -EINVAL later when IO is
8296 * submitted if they are wrong.
8298 if (!iov->iov_base || !iov->iov_len)
8301 /* arbitrary limit, but we need something */
8302 if (iov->iov_len > SZ_1G)
8308 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8309 unsigned int nr_args)
8313 struct page *last_hpage = NULL;
8315 ret = io_buffers_map_alloc(ctx, nr_args);
8319 for (i = 0; i < nr_args; i++) {
8320 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8322 ret = io_copy_iov(ctx, &iov, arg, i);
8326 ret = io_buffer_validate(&iov);
8330 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8334 ctx->nr_user_bufs++;
8338 io_sqe_buffers_unregister(ctx);
8343 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8345 __s32 __user *fds = arg;
8351 if (copy_from_user(&fd, fds, sizeof(*fds)))
8354 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8355 if (IS_ERR(ctx->cq_ev_fd)) {
8356 int ret = PTR_ERR(ctx->cq_ev_fd);
8357 ctx->cq_ev_fd = NULL;
8364 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8366 if (ctx->cq_ev_fd) {
8367 eventfd_ctx_put(ctx->cq_ev_fd);
8368 ctx->cq_ev_fd = NULL;
8375 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8377 struct io_buffer *buf;
8378 unsigned long index;
8380 xa_for_each(&ctx->io_buffers, index, buf)
8381 __io_remove_buffers(ctx, buf, index, -1U);
8384 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8386 struct io_kiocb *req, *nxt;
8388 list_for_each_entry_safe(req, nxt, list, compl.list) {
8389 if (tsk && req->task != tsk)
8391 list_del(&req->compl.list);
8392 kmem_cache_free(req_cachep, req);
8396 static void io_req_caches_free(struct io_ring_ctx *ctx)
8398 struct io_submit_state *submit_state = &ctx->submit_state;
8399 struct io_comp_state *cs = &ctx->submit_state.comp;
8401 mutex_lock(&ctx->uring_lock);
8403 if (submit_state->free_reqs) {
8404 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8405 submit_state->reqs);
8406 submit_state->free_reqs = 0;
8409 spin_lock_irq(&ctx->completion_lock);
8410 list_splice_init(&cs->locked_free_list, &cs->free_list);
8411 cs->locked_free_nr = 0;
8412 spin_unlock_irq(&ctx->completion_lock);
8414 io_req_cache_free(&cs->free_list, NULL);
8416 mutex_unlock(&ctx->uring_lock);
8419 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8422 * Some may use context even when all refs and requests have been put,
8423 * and they are free to do so while still holding uring_lock or
8424 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8426 mutex_lock(&ctx->uring_lock);
8427 mutex_unlock(&ctx->uring_lock);
8428 spin_lock_irq(&ctx->completion_lock);
8429 spin_unlock_irq(&ctx->completion_lock);
8431 io_sq_thread_finish(ctx);
8432 io_sqe_buffers_unregister(ctx);
8434 if (ctx->mm_account) {
8435 mmdrop(ctx->mm_account);
8436 ctx->mm_account = NULL;
8439 mutex_lock(&ctx->uring_lock);
8440 io_sqe_files_unregister(ctx);
8441 mutex_unlock(&ctx->uring_lock);
8442 io_eventfd_unregister(ctx);
8443 io_destroy_buffers(ctx);
8445 #if defined(CONFIG_UNIX)
8446 if (ctx->ring_sock) {
8447 ctx->ring_sock->file = NULL; /* so that iput() is called */
8448 sock_release(ctx->ring_sock);
8452 io_mem_free(ctx->rings);
8453 io_mem_free(ctx->sq_sqes);
8455 percpu_ref_exit(&ctx->refs);
8456 free_uid(ctx->user);
8457 io_req_caches_free(ctx);
8459 io_wq_put_hash(ctx->hash_map);
8460 kfree(ctx->cancel_hash);
8464 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8466 struct io_ring_ctx *ctx = file->private_data;
8469 poll_wait(file, &ctx->cq_wait, wait);
8471 * synchronizes with barrier from wq_has_sleeper call in
8475 if (!io_sqring_full(ctx))
8476 mask |= EPOLLOUT | EPOLLWRNORM;
8479 * Don't flush cqring overflow list here, just do a simple check.
8480 * Otherwise there could possible be ABBA deadlock:
8483 * lock(&ctx->uring_lock);
8485 * lock(&ctx->uring_lock);
8488 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8489 * pushs them to do the flush.
8491 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8492 mask |= EPOLLIN | EPOLLRDNORM;
8497 static int io_uring_fasync(int fd, struct file *file, int on)
8499 struct io_ring_ctx *ctx = file->private_data;
8501 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8504 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8506 const struct cred *creds;
8508 creds = xa_erase(&ctx->personalities, id);
8517 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8519 return io_run_task_work_head(&ctx->exit_task_work);
8522 struct io_tctx_exit {
8523 struct callback_head task_work;
8524 struct completion completion;
8525 struct io_ring_ctx *ctx;
8528 static void io_tctx_exit_cb(struct callback_head *cb)
8530 struct io_uring_task *tctx = current->io_uring;
8531 struct io_tctx_exit *work;
8533 work = container_of(cb, struct io_tctx_exit, task_work);
8535 * When @in_idle, we're in cancellation and it's racy to remove the
8536 * node. It'll be removed by the end of cancellation, just ignore it.
8538 if (!atomic_read(&tctx->in_idle))
8539 io_uring_del_task_file((unsigned long)work->ctx);
8540 complete(&work->completion);
8543 static void io_ring_exit_work(struct work_struct *work)
8545 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8546 unsigned long timeout = jiffies + HZ * 60 * 5;
8547 struct io_tctx_exit exit;
8548 struct io_tctx_node *node;
8551 /* prevent SQPOLL from submitting new requests */
8553 io_sq_thread_park(ctx->sq_data);
8554 list_del_init(&ctx->sqd_list);
8555 io_sqd_update_thread_idle(ctx->sq_data);
8556 io_sq_thread_unpark(ctx->sq_data);
8560 * If we're doing polled IO and end up having requests being
8561 * submitted async (out-of-line), then completions can come in while
8562 * we're waiting for refs to drop. We need to reap these manually,
8563 * as nobody else will be looking for them.
8566 io_uring_try_cancel_requests(ctx, NULL, NULL);
8568 WARN_ON_ONCE(time_after(jiffies, timeout));
8569 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8571 mutex_lock(&ctx->uring_lock);
8572 while (!list_empty(&ctx->tctx_list)) {
8573 WARN_ON_ONCE(time_after(jiffies, timeout));
8575 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8578 init_completion(&exit.completion);
8579 init_task_work(&exit.task_work, io_tctx_exit_cb);
8580 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8581 if (WARN_ON_ONCE(ret))
8583 wake_up_process(node->task);
8585 mutex_unlock(&ctx->uring_lock);
8586 wait_for_completion(&exit.completion);
8588 mutex_lock(&ctx->uring_lock);
8590 mutex_unlock(&ctx->uring_lock);
8592 io_ring_ctx_free(ctx);
8595 /* Returns true if we found and killed one or more timeouts */
8596 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8597 struct files_struct *files)
8599 struct io_kiocb *req, *tmp;
8602 spin_lock_irq(&ctx->completion_lock);
8603 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8604 if (io_match_task(req, tsk, files)) {
8605 io_kill_timeout(req, -ECANCELED);
8609 io_commit_cqring(ctx);
8610 spin_unlock_irq(&ctx->completion_lock);
8613 io_cqring_ev_posted(ctx);
8614 return canceled != 0;
8617 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8619 unsigned long index;
8620 struct creds *creds;
8622 mutex_lock(&ctx->uring_lock);
8623 percpu_ref_kill(&ctx->refs);
8624 /* if force is set, the ring is going away. always drop after that */
8625 ctx->cq_overflow_flushed = 1;
8627 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8628 xa_for_each(&ctx->personalities, index, creds)
8629 io_unregister_personality(ctx, index);
8630 mutex_unlock(&ctx->uring_lock);
8632 io_kill_timeouts(ctx, NULL, NULL);
8633 io_poll_remove_all(ctx, NULL, NULL);
8635 /* if we failed setting up the ctx, we might not have any rings */
8636 io_iopoll_try_reap_events(ctx);
8638 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8640 * Use system_unbound_wq to avoid spawning tons of event kworkers
8641 * if we're exiting a ton of rings at the same time. It just adds
8642 * noise and overhead, there's no discernable change in runtime
8643 * over using system_wq.
8645 queue_work(system_unbound_wq, &ctx->exit_work);
8648 static int io_uring_release(struct inode *inode, struct file *file)
8650 struct io_ring_ctx *ctx = file->private_data;
8652 file->private_data = NULL;
8653 io_ring_ctx_wait_and_kill(ctx);
8657 struct io_task_cancel {
8658 struct task_struct *task;
8659 struct files_struct *files;
8662 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8664 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8665 struct io_task_cancel *cancel = data;
8668 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8669 unsigned long flags;
8670 struct io_ring_ctx *ctx = req->ctx;
8672 /* protect against races with linked timeouts */
8673 spin_lock_irqsave(&ctx->completion_lock, flags);
8674 ret = io_match_task(req, cancel->task, cancel->files);
8675 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8677 ret = io_match_task(req, cancel->task, cancel->files);
8682 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8683 struct task_struct *task,
8684 struct files_struct *files)
8686 struct io_defer_entry *de;
8689 spin_lock_irq(&ctx->completion_lock);
8690 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8691 if (io_match_task(de->req, task, files)) {
8692 list_cut_position(&list, &ctx->defer_list, &de->list);
8696 spin_unlock_irq(&ctx->completion_lock);
8697 if (list_empty(&list))
8700 while (!list_empty(&list)) {
8701 de = list_first_entry(&list, struct io_defer_entry, list);
8702 list_del_init(&de->list);
8703 req_set_fail_links(de->req);
8704 io_put_req(de->req);
8705 io_req_complete(de->req, -ECANCELED);
8711 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8713 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8715 return req->ctx == data;
8718 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8720 struct io_tctx_node *node;
8721 enum io_wq_cancel cret;
8724 mutex_lock(&ctx->uring_lock);
8725 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8726 struct io_uring_task *tctx = node->task->io_uring;
8729 * io_wq will stay alive while we hold uring_lock, because it's
8730 * killed after ctx nodes, which requires to take the lock.
8732 if (!tctx || !tctx->io_wq)
8734 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8735 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8737 mutex_unlock(&ctx->uring_lock);
8742 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8743 struct task_struct *task,
8744 struct files_struct *files)
8746 struct io_task_cancel cancel = { .task = task, .files = files, };
8747 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8750 enum io_wq_cancel cret;
8754 ret |= io_uring_try_cancel_iowq(ctx);
8755 } else if (tctx && tctx->io_wq) {
8757 * Cancels requests of all rings, not only @ctx, but
8758 * it's fine as the task is in exit/exec.
8760 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8762 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8765 /* SQPOLL thread does its own polling */
8766 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8767 (ctx->sq_data && ctx->sq_data->thread == current)) {
8768 while (!list_empty_careful(&ctx->iopoll_list)) {
8769 io_iopoll_try_reap_events(ctx);
8774 ret |= io_cancel_defer_files(ctx, task, files);
8775 ret |= io_poll_remove_all(ctx, task, files);
8776 ret |= io_kill_timeouts(ctx, task, files);
8777 ret |= io_run_task_work();
8778 ret |= io_run_ctx_fallback(ctx);
8779 io_cqring_overflow_flush(ctx, true, task, files);
8786 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8787 struct task_struct *task,
8788 struct files_struct *files)
8790 struct io_kiocb *req;
8793 spin_lock_irq(&ctx->inflight_lock);
8794 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8795 cnt += io_match_task(req, task, files);
8796 spin_unlock_irq(&ctx->inflight_lock);
8800 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8801 struct task_struct *task,
8802 struct files_struct *files)
8804 while (!list_empty_careful(&ctx->inflight_list)) {
8808 inflight = io_uring_count_inflight(ctx, task, files);
8812 io_uring_try_cancel_requests(ctx, task, files);
8814 prepare_to_wait(&task->io_uring->wait, &wait,
8815 TASK_UNINTERRUPTIBLE);
8816 if (inflight == io_uring_count_inflight(ctx, task, files))
8818 finish_wait(&task->io_uring->wait, &wait);
8823 * Note that this task has used io_uring. We use it for cancelation purposes.
8825 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8827 struct io_uring_task *tctx = current->io_uring;
8828 struct io_tctx_node *node;
8831 if (unlikely(!tctx)) {
8832 ret = io_uring_alloc_task_context(current, ctx);
8835 tctx = current->io_uring;
8837 if (tctx->last != ctx) {
8838 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8841 node = kmalloc(sizeof(*node), GFP_KERNEL);
8845 node->task = current;
8847 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8854 mutex_lock(&ctx->uring_lock);
8855 list_add(&node->ctx_node, &ctx->tctx_list);
8856 mutex_unlock(&ctx->uring_lock);
8864 * Remove this io_uring_file -> task mapping.
8866 static void io_uring_del_task_file(unsigned long index)
8868 struct io_uring_task *tctx = current->io_uring;
8869 struct io_tctx_node *node;
8873 node = xa_erase(&tctx->xa, index);
8877 WARN_ON_ONCE(current != node->task);
8878 WARN_ON_ONCE(list_empty(&node->ctx_node));
8880 mutex_lock(&node->ctx->uring_lock);
8881 list_del(&node->ctx_node);
8882 mutex_unlock(&node->ctx->uring_lock);
8884 if (tctx->last == node->ctx)
8889 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8891 struct io_tctx_node *node;
8892 unsigned long index;
8894 xa_for_each(&tctx->xa, index, node)
8895 io_uring_del_task_file(index);
8897 io_wq_put_and_exit(tctx->io_wq);
8902 static s64 tctx_inflight(struct io_uring_task *tctx)
8904 return percpu_counter_sum(&tctx->inflight);
8907 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8909 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8910 struct io_ring_ctx *ctx = work->ctx;
8911 struct io_sq_data *sqd = ctx->sq_data;
8914 io_uring_cancel_sqpoll(ctx);
8915 complete(&work->completion);
8918 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8920 struct io_sq_data *sqd = ctx->sq_data;
8921 struct io_tctx_exit work = { .ctx = ctx, };
8922 struct task_struct *task;
8924 io_sq_thread_park(sqd);
8925 list_del_init(&ctx->sqd_list);
8926 io_sqd_update_thread_idle(sqd);
8929 init_completion(&work.completion);
8930 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8931 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8932 wake_up_process(task);
8934 io_sq_thread_unpark(sqd);
8937 wait_for_completion(&work.completion);
8940 void __io_uring_files_cancel(struct files_struct *files)
8942 struct io_uring_task *tctx = current->io_uring;
8943 struct io_tctx_node *node;
8944 unsigned long index;
8946 /* make sure overflow events are dropped */
8947 atomic_inc(&tctx->in_idle);
8948 xa_for_each(&tctx->xa, index, node) {
8949 struct io_ring_ctx *ctx = node->ctx;
8952 io_sqpoll_cancel_sync(ctx);
8955 io_uring_cancel_files(ctx, current, files);
8957 io_uring_try_cancel_requests(ctx, current, NULL);
8959 atomic_dec(&tctx->in_idle);
8962 io_uring_clean_tctx(tctx);
8965 /* should only be called by SQPOLL task */
8966 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8968 struct io_sq_data *sqd = ctx->sq_data;
8969 struct io_uring_task *tctx = current->io_uring;
8973 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8975 atomic_inc(&tctx->in_idle);
8977 /* read completions before cancelations */
8978 inflight = tctx_inflight(tctx);
8981 io_uring_try_cancel_requests(ctx, current, NULL);
8983 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8985 * If we've seen completions, retry without waiting. This
8986 * avoids a race where a completion comes in before we did
8987 * prepare_to_wait().
8989 if (inflight == tctx_inflight(tctx))
8991 finish_wait(&tctx->wait, &wait);
8993 atomic_dec(&tctx->in_idle);
8997 * Find any io_uring fd that this task has registered or done IO on, and cancel
9000 void __io_uring_task_cancel(void)
9002 struct io_uring_task *tctx = current->io_uring;
9006 /* make sure overflow events are dropped */
9007 atomic_inc(&tctx->in_idle);
9009 /* read completions before cancelations */
9010 inflight = tctx_inflight(tctx);
9013 __io_uring_files_cancel(NULL);
9015 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9018 * If we've seen completions, retry without waiting. This
9019 * avoids a race where a completion comes in before we did
9020 * prepare_to_wait().
9022 if (inflight == tctx_inflight(tctx))
9024 finish_wait(&tctx->wait, &wait);
9027 atomic_dec(&tctx->in_idle);
9029 io_uring_clean_tctx(tctx);
9030 /* all current's requests should be gone, we can kill tctx */
9031 __io_uring_free(current);
9034 static void *io_uring_validate_mmap_request(struct file *file,
9035 loff_t pgoff, size_t sz)
9037 struct io_ring_ctx *ctx = file->private_data;
9038 loff_t offset = pgoff << PAGE_SHIFT;
9043 case IORING_OFF_SQ_RING:
9044 case IORING_OFF_CQ_RING:
9047 case IORING_OFF_SQES:
9051 return ERR_PTR(-EINVAL);
9054 page = virt_to_head_page(ptr);
9055 if (sz > page_size(page))
9056 return ERR_PTR(-EINVAL);
9063 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9065 size_t sz = vma->vm_end - vma->vm_start;
9069 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9071 return PTR_ERR(ptr);
9073 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9074 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9077 #else /* !CONFIG_MMU */
9079 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9081 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9084 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9086 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9089 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9090 unsigned long addr, unsigned long len,
9091 unsigned long pgoff, unsigned long flags)
9095 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9097 return PTR_ERR(ptr);
9099 return (unsigned long) ptr;
9102 #endif /* !CONFIG_MMU */
9104 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9109 if (!io_sqring_full(ctx))
9111 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9113 if (!io_sqring_full(ctx))
9116 } while (!signal_pending(current));
9118 finish_wait(&ctx->sqo_sq_wait, &wait);
9122 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9123 struct __kernel_timespec __user **ts,
9124 const sigset_t __user **sig)
9126 struct io_uring_getevents_arg arg;
9129 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9130 * is just a pointer to the sigset_t.
9132 if (!(flags & IORING_ENTER_EXT_ARG)) {
9133 *sig = (const sigset_t __user *) argp;
9139 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9140 * timespec and sigset_t pointers if good.
9142 if (*argsz != sizeof(arg))
9144 if (copy_from_user(&arg, argp, sizeof(arg)))
9146 *sig = u64_to_user_ptr(arg.sigmask);
9147 *argsz = arg.sigmask_sz;
9148 *ts = u64_to_user_ptr(arg.ts);
9152 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9153 u32, min_complete, u32, flags, const void __user *, argp,
9156 struct io_ring_ctx *ctx;
9163 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9164 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9172 if (f.file->f_op != &io_uring_fops)
9176 ctx = f.file->private_data;
9177 if (!percpu_ref_tryget(&ctx->refs))
9181 if (ctx->flags & IORING_SETUP_R_DISABLED)
9185 * For SQ polling, the thread will do all submissions and completions.
9186 * Just return the requested submit count, and wake the thread if
9190 if (ctx->flags & IORING_SETUP_SQPOLL) {
9191 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9194 if (unlikely(ctx->sq_data->thread == NULL)) {
9197 if (flags & IORING_ENTER_SQ_WAKEUP)
9198 wake_up(&ctx->sq_data->wait);
9199 if (flags & IORING_ENTER_SQ_WAIT) {
9200 ret = io_sqpoll_wait_sq(ctx);
9204 submitted = to_submit;
9205 } else if (to_submit) {
9206 ret = io_uring_add_task_file(ctx);
9209 mutex_lock(&ctx->uring_lock);
9210 submitted = io_submit_sqes(ctx, to_submit);
9211 mutex_unlock(&ctx->uring_lock);
9213 if (submitted != to_submit)
9216 if (flags & IORING_ENTER_GETEVENTS) {
9217 const sigset_t __user *sig;
9218 struct __kernel_timespec __user *ts;
9220 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9224 min_complete = min(min_complete, ctx->cq_entries);
9227 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9228 * space applications don't need to do io completion events
9229 * polling again, they can rely on io_sq_thread to do polling
9230 * work, which can reduce cpu usage and uring_lock contention.
9232 if (ctx->flags & IORING_SETUP_IOPOLL &&
9233 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9234 ret = io_iopoll_check(ctx, min_complete);
9236 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9241 percpu_ref_put(&ctx->refs);
9244 return submitted ? submitted : ret;
9247 #ifdef CONFIG_PROC_FS
9248 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9249 const struct cred *cred)
9251 struct user_namespace *uns = seq_user_ns(m);
9252 struct group_info *gi;
9257 seq_printf(m, "%5d\n", id);
9258 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9259 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9260 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9261 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9262 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9263 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9264 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9265 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9266 seq_puts(m, "\n\tGroups:\t");
9267 gi = cred->group_info;
9268 for (g = 0; g < gi->ngroups; g++) {
9269 seq_put_decimal_ull(m, g ? " " : "",
9270 from_kgid_munged(uns, gi->gid[g]));
9272 seq_puts(m, "\n\tCapEff:\t");
9273 cap = cred->cap_effective;
9274 CAP_FOR_EACH_U32(__capi)
9275 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9280 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9282 struct io_sq_data *sq = NULL;
9287 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9288 * since fdinfo case grabs it in the opposite direction of normal use
9289 * cases. If we fail to get the lock, we just don't iterate any
9290 * structures that could be going away outside the io_uring mutex.
9292 has_lock = mutex_trylock(&ctx->uring_lock);
9294 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9300 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9301 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9302 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9303 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9304 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9307 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9309 seq_printf(m, "%5u: <none>\n", i);
9311 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9312 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9313 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9315 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9316 (unsigned int) buf->len);
9318 if (has_lock && !xa_empty(&ctx->personalities)) {
9319 unsigned long index;
9320 const struct cred *cred;
9322 seq_printf(m, "Personalities:\n");
9323 xa_for_each(&ctx->personalities, index, cred)
9324 io_uring_show_cred(m, index, cred);
9326 seq_printf(m, "PollList:\n");
9327 spin_lock_irq(&ctx->completion_lock);
9328 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9329 struct hlist_head *list = &ctx->cancel_hash[i];
9330 struct io_kiocb *req;
9332 hlist_for_each_entry(req, list, hash_node)
9333 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9334 req->task->task_works != NULL);
9336 spin_unlock_irq(&ctx->completion_lock);
9338 mutex_unlock(&ctx->uring_lock);
9341 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9343 struct io_ring_ctx *ctx = f->private_data;
9345 if (percpu_ref_tryget(&ctx->refs)) {
9346 __io_uring_show_fdinfo(ctx, m);
9347 percpu_ref_put(&ctx->refs);
9352 static const struct file_operations io_uring_fops = {
9353 .release = io_uring_release,
9354 .mmap = io_uring_mmap,
9356 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9357 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9359 .poll = io_uring_poll,
9360 .fasync = io_uring_fasync,
9361 #ifdef CONFIG_PROC_FS
9362 .show_fdinfo = io_uring_show_fdinfo,
9366 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9367 struct io_uring_params *p)
9369 struct io_rings *rings;
9370 size_t size, sq_array_offset;
9372 /* make sure these are sane, as we already accounted them */
9373 ctx->sq_entries = p->sq_entries;
9374 ctx->cq_entries = p->cq_entries;
9376 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9377 if (size == SIZE_MAX)
9380 rings = io_mem_alloc(size);
9385 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9386 rings->sq_ring_mask = p->sq_entries - 1;
9387 rings->cq_ring_mask = p->cq_entries - 1;
9388 rings->sq_ring_entries = p->sq_entries;
9389 rings->cq_ring_entries = p->cq_entries;
9390 ctx->sq_mask = rings->sq_ring_mask;
9391 ctx->cq_mask = rings->cq_ring_mask;
9393 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9394 if (size == SIZE_MAX) {
9395 io_mem_free(ctx->rings);
9400 ctx->sq_sqes = io_mem_alloc(size);
9401 if (!ctx->sq_sqes) {
9402 io_mem_free(ctx->rings);
9410 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9414 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9418 ret = io_uring_add_task_file(ctx);
9423 fd_install(fd, file);
9428 * Allocate an anonymous fd, this is what constitutes the application
9429 * visible backing of an io_uring instance. The application mmaps this
9430 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9431 * we have to tie this fd to a socket for file garbage collection purposes.
9433 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9436 #if defined(CONFIG_UNIX)
9439 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9442 return ERR_PTR(ret);
9445 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9446 O_RDWR | O_CLOEXEC);
9447 #if defined(CONFIG_UNIX)
9449 sock_release(ctx->ring_sock);
9450 ctx->ring_sock = NULL;
9452 ctx->ring_sock->file = file;
9458 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9459 struct io_uring_params __user *params)
9461 struct io_ring_ctx *ctx;
9467 if (entries > IORING_MAX_ENTRIES) {
9468 if (!(p->flags & IORING_SETUP_CLAMP))
9470 entries = IORING_MAX_ENTRIES;
9474 * Use twice as many entries for the CQ ring. It's possible for the
9475 * application to drive a higher depth than the size of the SQ ring,
9476 * since the sqes are only used at submission time. This allows for
9477 * some flexibility in overcommitting a bit. If the application has
9478 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9479 * of CQ ring entries manually.
9481 p->sq_entries = roundup_pow_of_two(entries);
9482 if (p->flags & IORING_SETUP_CQSIZE) {
9484 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9485 * to a power-of-two, if it isn't already. We do NOT impose
9486 * any cq vs sq ring sizing.
9490 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9491 if (!(p->flags & IORING_SETUP_CLAMP))
9493 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9495 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9496 if (p->cq_entries < p->sq_entries)
9499 p->cq_entries = 2 * p->sq_entries;
9502 ctx = io_ring_ctx_alloc(p);
9505 ctx->compat = in_compat_syscall();
9506 if (!capable(CAP_IPC_LOCK))
9507 ctx->user = get_uid(current_user());
9510 * This is just grabbed for accounting purposes. When a process exits,
9511 * the mm is exited and dropped before the files, hence we need to hang
9512 * on to this mm purely for the purposes of being able to unaccount
9513 * memory (locked/pinned vm). It's not used for anything else.
9515 mmgrab(current->mm);
9516 ctx->mm_account = current->mm;
9518 ret = io_allocate_scq_urings(ctx, p);
9522 ret = io_sq_offload_create(ctx, p);
9526 memset(&p->sq_off, 0, sizeof(p->sq_off));
9527 p->sq_off.head = offsetof(struct io_rings, sq.head);
9528 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9529 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9530 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9531 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9532 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9533 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9535 memset(&p->cq_off, 0, sizeof(p->cq_off));
9536 p->cq_off.head = offsetof(struct io_rings, cq.head);
9537 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9538 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9539 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9540 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9541 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9542 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9544 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9545 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9546 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9547 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9548 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9550 if (copy_to_user(params, p, sizeof(*p))) {
9555 file = io_uring_get_file(ctx);
9557 ret = PTR_ERR(file);
9562 * Install ring fd as the very last thing, so we don't risk someone
9563 * having closed it before we finish setup
9565 ret = io_uring_install_fd(ctx, file);
9567 /* fput will clean it up */
9572 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9575 io_ring_ctx_wait_and_kill(ctx);
9580 * Sets up an aio uring context, and returns the fd. Applications asks for a
9581 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9582 * params structure passed in.
9584 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9586 struct io_uring_params p;
9589 if (copy_from_user(&p, params, sizeof(p)))
9591 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9596 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9597 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9598 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9599 IORING_SETUP_R_DISABLED))
9602 return io_uring_create(entries, &p, params);
9605 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9606 struct io_uring_params __user *, params)
9608 return io_uring_setup(entries, params);
9611 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9613 struct io_uring_probe *p;
9617 size = struct_size(p, ops, nr_args);
9618 if (size == SIZE_MAX)
9620 p = kzalloc(size, GFP_KERNEL);
9625 if (copy_from_user(p, arg, size))
9628 if (memchr_inv(p, 0, size))
9631 p->last_op = IORING_OP_LAST - 1;
9632 if (nr_args > IORING_OP_LAST)
9633 nr_args = IORING_OP_LAST;
9635 for (i = 0; i < nr_args; i++) {
9637 if (!io_op_defs[i].not_supported)
9638 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9643 if (copy_to_user(arg, p, size))
9650 static int io_register_personality(struct io_ring_ctx *ctx)
9652 const struct cred *creds;
9656 creds = get_current_cred();
9658 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9659 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9666 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9667 unsigned int nr_args)
9669 struct io_uring_restriction *res;
9673 /* Restrictions allowed only if rings started disabled */
9674 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9677 /* We allow only a single restrictions registration */
9678 if (ctx->restrictions.registered)
9681 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9684 size = array_size(nr_args, sizeof(*res));
9685 if (size == SIZE_MAX)
9688 res = memdup_user(arg, size);
9690 return PTR_ERR(res);
9694 for (i = 0; i < nr_args; i++) {
9695 switch (res[i].opcode) {
9696 case IORING_RESTRICTION_REGISTER_OP:
9697 if (res[i].register_op >= IORING_REGISTER_LAST) {
9702 __set_bit(res[i].register_op,
9703 ctx->restrictions.register_op);
9705 case IORING_RESTRICTION_SQE_OP:
9706 if (res[i].sqe_op >= IORING_OP_LAST) {
9711 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9713 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9714 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9716 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9717 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9726 /* Reset all restrictions if an error happened */
9728 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9730 ctx->restrictions.registered = true;
9736 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9738 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9741 if (ctx->restrictions.registered)
9742 ctx->restricted = 1;
9744 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9745 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9746 wake_up(&ctx->sq_data->wait);
9750 static bool io_register_op_must_quiesce(int op)
9753 case IORING_UNREGISTER_FILES:
9754 case IORING_REGISTER_FILES_UPDATE:
9755 case IORING_REGISTER_PROBE:
9756 case IORING_REGISTER_PERSONALITY:
9757 case IORING_UNREGISTER_PERSONALITY:
9764 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9765 void __user *arg, unsigned nr_args)
9766 __releases(ctx->uring_lock)
9767 __acquires(ctx->uring_lock)
9772 * We're inside the ring mutex, if the ref is already dying, then
9773 * someone else killed the ctx or is already going through
9774 * io_uring_register().
9776 if (percpu_ref_is_dying(&ctx->refs))
9779 if (io_register_op_must_quiesce(opcode)) {
9780 percpu_ref_kill(&ctx->refs);
9783 * Drop uring mutex before waiting for references to exit. If
9784 * another thread is currently inside io_uring_enter() it might
9785 * need to grab the uring_lock to make progress. If we hold it
9786 * here across the drain wait, then we can deadlock. It's safe
9787 * to drop the mutex here, since no new references will come in
9788 * after we've killed the percpu ref.
9790 mutex_unlock(&ctx->uring_lock);
9792 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9795 ret = io_run_task_work_sig();
9800 mutex_lock(&ctx->uring_lock);
9803 percpu_ref_resurrect(&ctx->refs);
9808 if (ctx->restricted) {
9809 if (opcode >= IORING_REGISTER_LAST) {
9814 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9821 case IORING_REGISTER_BUFFERS:
9822 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9824 case IORING_UNREGISTER_BUFFERS:
9828 ret = io_sqe_buffers_unregister(ctx);
9830 case IORING_REGISTER_FILES:
9831 ret = io_sqe_files_register(ctx, arg, nr_args);
9833 case IORING_UNREGISTER_FILES:
9837 ret = io_sqe_files_unregister(ctx);
9839 case IORING_REGISTER_FILES_UPDATE:
9840 ret = io_sqe_files_update(ctx, arg, nr_args);
9842 case IORING_REGISTER_EVENTFD:
9843 case IORING_REGISTER_EVENTFD_ASYNC:
9847 ret = io_eventfd_register(ctx, arg);
9850 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9851 ctx->eventfd_async = 1;
9853 ctx->eventfd_async = 0;
9855 case IORING_UNREGISTER_EVENTFD:
9859 ret = io_eventfd_unregister(ctx);
9861 case IORING_REGISTER_PROBE:
9863 if (!arg || nr_args > 256)
9865 ret = io_probe(ctx, arg, nr_args);
9867 case IORING_REGISTER_PERSONALITY:
9871 ret = io_register_personality(ctx);
9873 case IORING_UNREGISTER_PERSONALITY:
9877 ret = io_unregister_personality(ctx, nr_args);
9879 case IORING_REGISTER_ENABLE_RINGS:
9883 ret = io_register_enable_rings(ctx);
9885 case IORING_REGISTER_RESTRICTIONS:
9886 ret = io_register_restrictions(ctx, arg, nr_args);
9894 if (io_register_op_must_quiesce(opcode)) {
9895 /* bring the ctx back to life */
9896 percpu_ref_reinit(&ctx->refs);
9898 reinit_completion(&ctx->ref_comp);
9903 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9904 void __user *, arg, unsigned int, nr_args)
9906 struct io_ring_ctx *ctx;
9915 if (f.file->f_op != &io_uring_fops)
9918 ctx = f.file->private_data;
9922 mutex_lock(&ctx->uring_lock);
9923 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9924 mutex_unlock(&ctx->uring_lock);
9925 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9926 ctx->cq_ev_fd != NULL, ret);
9932 static int __init io_uring_init(void)
9934 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9935 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9936 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9939 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9940 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9941 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9942 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9943 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9944 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9945 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9946 BUILD_BUG_SQE_ELEM(8, __u64, off);
9947 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9948 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9949 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9950 BUILD_BUG_SQE_ELEM(24, __u32, len);
9951 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9952 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9953 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9954 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9955 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9956 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9957 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9958 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9959 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9960 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9961 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9962 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9963 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9964 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9965 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9966 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9967 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9968 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9969 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9971 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9972 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9973 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9977 __initcall(io_uring_init);