1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
239 struct list_head list;
245 struct io_restriction {
246 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
247 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
248 u8 sqe_flags_allowed;
249 u8 sqe_flags_required;
254 IO_SQ_THREAD_SHOULD_STOP = 0,
255 IO_SQ_THREAD_SHOULD_PARK,
260 atomic_t park_pending;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
276 struct callback_head *park_task_work;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
449 struct callback_head *exit_task_work;
451 struct wait_queue_head hash_wait;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work;
455 struct list_head tctx_list;
458 struct io_uring_task {
459 /* submission side */
461 struct wait_queue_head wait;
462 const struct io_ring_ctx *last;
464 struct percpu_counter inflight;
468 spinlock_t task_lock;
469 struct io_wq_work_list task_list;
470 unsigned long task_state;
471 struct callback_head task_work;
475 * First field must be the file pointer in all the
476 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
478 struct io_poll_iocb {
480 struct wait_queue_head *head;
484 struct wait_queue_entry wait;
487 struct io_poll_remove {
497 struct io_timeout_data {
498 struct io_kiocb *req;
499 struct hrtimer timer;
500 struct timespec64 ts;
501 enum hrtimer_mode mode;
506 struct sockaddr __user *addr;
507 int __user *addr_len;
509 unsigned long nofile;
529 struct list_head list;
530 /* head of the link, used by linked timeouts only */
531 struct io_kiocb *head;
534 struct io_timeout_rem {
539 struct timespec64 ts;
544 /* NOTE: kiocb has the file as the first member, so don't do it here */
552 struct sockaddr __user *addr;
559 struct user_msghdr __user *umsg;
565 struct io_buffer *kbuf;
571 struct filename *filename;
573 unsigned long nofile;
576 struct io_rsrc_update {
602 struct epoll_event event;
606 struct file *file_out;
607 struct file *file_in;
614 struct io_provide_buf {
628 const char __user *filename;
629 struct statx __user *buffer;
641 struct filename *oldpath;
642 struct filename *newpath;
650 struct filename *filename;
653 struct io_completion {
655 struct list_head list;
659 struct io_async_connect {
660 struct sockaddr_storage address;
663 struct io_async_msghdr {
664 struct iovec fast_iov[UIO_FASTIOV];
665 /* points to an allocated iov, if NULL we use fast_iov instead */
666 struct iovec *free_iov;
667 struct sockaddr __user *uaddr;
669 struct sockaddr_storage addr;
673 struct iovec fast_iov[UIO_FASTIOV];
674 const struct iovec *free_iovec;
675 struct iov_iter iter;
677 struct wait_page_queue wpq;
681 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
682 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
683 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
684 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
685 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
686 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
692 REQ_F_LINK_TIMEOUT_BIT,
694 REQ_F_NEED_CLEANUP_BIT,
696 REQ_F_BUFFER_SELECTED_BIT,
697 REQ_F_NO_FILE_TABLE_BIT,
698 REQ_F_LTIMEOUT_ACTIVE_BIT,
699 REQ_F_COMPLETE_INLINE_BIT,
702 /* not a real bit, just to check we're not overflowing the space */
708 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
709 /* drain existing IO first */
710 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
712 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
713 /* doesn't sever on completion < 0 */
714 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
716 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
717 /* IOSQE_BUFFER_SELECT */
718 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
720 /* fail rest of links */
721 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
722 /* on inflight list, should be cancelled and waited on exit reliably */
723 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
724 /* read/write uses file position */
725 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
726 /* must not punt to workers */
727 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
728 /* has or had linked timeout */
729 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
731 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
733 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
734 /* already went through poll handler */
735 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
736 /* buffer already selected */
737 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
738 /* doesn't need file table for this request */
739 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
740 /* linked timeout is active, i.e. prepared by link's head */
741 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
742 /* completion is deferred through io_comp_state */
743 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
744 /* caller should reissue async */
745 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
749 struct io_poll_iocb poll;
750 struct io_poll_iocb *double_poll;
753 struct io_task_work {
754 struct io_wq_work_node node;
755 task_work_func_t func;
759 * NOTE! Each of the iocb union members has the file pointer
760 * as the first entry in their struct definition. So you can
761 * access the file pointer through any of the sub-structs,
762 * or directly as just 'ki_filp' in this struct.
768 struct io_poll_iocb poll;
769 struct io_poll_remove poll_remove;
770 struct io_accept accept;
772 struct io_cancel cancel;
773 struct io_timeout timeout;
774 struct io_timeout_rem timeout_rem;
775 struct io_connect connect;
776 struct io_sr_msg sr_msg;
778 struct io_close close;
779 struct io_rsrc_update rsrc_update;
780 struct io_fadvise fadvise;
781 struct io_madvise madvise;
782 struct io_epoll epoll;
783 struct io_splice splice;
784 struct io_provide_buf pbuf;
785 struct io_statx statx;
786 struct io_shutdown shutdown;
787 struct io_rename rename;
788 struct io_unlink unlink;
789 /* use only after cleaning per-op data, see io_clean_op() */
790 struct io_completion compl;
793 /* opcode allocated if it needs to store data for async defer */
796 /* polled IO has completed */
802 struct io_ring_ctx *ctx;
805 struct task_struct *task;
808 struct io_kiocb *link;
809 struct percpu_ref *fixed_rsrc_refs;
812 * 1. used with ctx->iopoll_list with reads/writes
813 * 2. to track reqs with ->files (see io_op_def::file_table)
815 struct list_head inflight_entry;
817 struct io_task_work io_task_work;
818 struct callback_head task_work;
820 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
821 struct hlist_node hash_node;
822 struct async_poll *apoll;
823 struct io_wq_work work;
826 struct io_tctx_node {
827 struct list_head ctx_node;
828 struct task_struct *task;
829 struct io_ring_ctx *ctx;
832 struct io_defer_entry {
833 struct list_head list;
834 struct io_kiocb *req;
839 /* needs req->file assigned */
840 unsigned needs_file : 1;
841 /* hash wq insertion if file is a regular file */
842 unsigned hash_reg_file : 1;
843 /* unbound wq insertion if file is a non-regular file */
844 unsigned unbound_nonreg_file : 1;
845 /* opcode is not supported by this kernel */
846 unsigned not_supported : 1;
847 /* set if opcode supports polled "wait" */
849 unsigned pollout : 1;
850 /* op supports buffer selection */
851 unsigned buffer_select : 1;
852 /* must always have async data allocated */
853 unsigned needs_async_data : 1;
854 /* should block plug */
856 /* size of async data needed, if any */
857 unsigned short async_size;
860 static const struct io_op_def io_op_defs[] = {
861 [IORING_OP_NOP] = {},
862 [IORING_OP_READV] = {
864 .unbound_nonreg_file = 1,
867 .needs_async_data = 1,
869 .async_size = sizeof(struct io_async_rw),
871 [IORING_OP_WRITEV] = {
874 .unbound_nonreg_file = 1,
876 .needs_async_data = 1,
878 .async_size = sizeof(struct io_async_rw),
880 [IORING_OP_FSYNC] = {
883 [IORING_OP_READ_FIXED] = {
885 .unbound_nonreg_file = 1,
888 .async_size = sizeof(struct io_async_rw),
890 [IORING_OP_WRITE_FIXED] = {
893 .unbound_nonreg_file = 1,
896 .async_size = sizeof(struct io_async_rw),
898 [IORING_OP_POLL_ADD] = {
900 .unbound_nonreg_file = 1,
902 [IORING_OP_POLL_REMOVE] = {},
903 [IORING_OP_SYNC_FILE_RANGE] = {
906 [IORING_OP_SENDMSG] = {
908 .unbound_nonreg_file = 1,
910 .needs_async_data = 1,
911 .async_size = sizeof(struct io_async_msghdr),
913 [IORING_OP_RECVMSG] = {
915 .unbound_nonreg_file = 1,
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_async_msghdr),
921 [IORING_OP_TIMEOUT] = {
922 .needs_async_data = 1,
923 .async_size = sizeof(struct io_timeout_data),
925 [IORING_OP_TIMEOUT_REMOVE] = {
926 /* used by timeout updates' prep() */
928 [IORING_OP_ACCEPT] = {
930 .unbound_nonreg_file = 1,
933 [IORING_OP_ASYNC_CANCEL] = {},
934 [IORING_OP_LINK_TIMEOUT] = {
935 .needs_async_data = 1,
936 .async_size = sizeof(struct io_timeout_data),
938 [IORING_OP_CONNECT] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_data = 1,
943 .async_size = sizeof(struct io_async_connect),
945 [IORING_OP_FALLOCATE] = {
948 [IORING_OP_OPENAT] = {},
949 [IORING_OP_CLOSE] = {},
950 [IORING_OP_FILES_UPDATE] = {},
951 [IORING_OP_STATX] = {},
954 .unbound_nonreg_file = 1,
958 .async_size = sizeof(struct io_async_rw),
960 [IORING_OP_WRITE] = {
962 .unbound_nonreg_file = 1,
965 .async_size = sizeof(struct io_async_rw),
967 [IORING_OP_FADVISE] = {
970 [IORING_OP_MADVISE] = {},
973 .unbound_nonreg_file = 1,
978 .unbound_nonreg_file = 1,
982 [IORING_OP_OPENAT2] = {
984 [IORING_OP_EPOLL_CTL] = {
985 .unbound_nonreg_file = 1,
987 [IORING_OP_SPLICE] = {
990 .unbound_nonreg_file = 1,
992 [IORING_OP_PROVIDE_BUFFERS] = {},
993 [IORING_OP_REMOVE_BUFFERS] = {},
997 .unbound_nonreg_file = 1,
999 [IORING_OP_SHUTDOWN] = {
1002 [IORING_OP_RENAMEAT] = {},
1003 [IORING_OP_UNLINKAT] = {},
1006 static bool io_disarm_next(struct io_kiocb *req);
1007 static void io_uring_del_task_file(unsigned long index);
1008 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1009 struct task_struct *task,
1010 struct files_struct *files);
1011 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1012 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1013 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1014 struct io_ring_ctx *ctx);
1015 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1017 static bool io_rw_reissue(struct io_kiocb *req);
1018 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1019 static void io_put_req(struct io_kiocb *req);
1020 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1021 static void io_double_put_req(struct io_kiocb *req);
1022 static void io_dismantle_req(struct io_kiocb *req);
1023 static void io_put_task(struct task_struct *task, int nr);
1024 static void io_queue_next(struct io_kiocb *req);
1025 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1026 static void __io_queue_linked_timeout(struct io_kiocb *req);
1027 static void io_queue_linked_timeout(struct io_kiocb *req);
1028 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1029 struct io_uring_rsrc_update *ip,
1031 static void __io_clean_op(struct io_kiocb *req);
1032 static struct file *io_file_get(struct io_submit_state *state,
1033 struct io_kiocb *req, int fd, bool fixed);
1034 static void __io_queue_sqe(struct io_kiocb *req);
1035 static void io_rsrc_put_work(struct work_struct *work);
1037 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1038 struct iov_iter *iter, bool needs_lock);
1039 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1040 const struct iovec *fast_iov,
1041 struct iov_iter *iter, bool force);
1042 static void io_req_task_queue(struct io_kiocb *req);
1043 static void io_submit_flush_completions(struct io_comp_state *cs,
1044 struct io_ring_ctx *ctx);
1046 static struct kmem_cache *req_cachep;
1048 static const struct file_operations io_uring_fops;
1050 struct sock *io_uring_get_socket(struct file *file)
1052 #if defined(CONFIG_UNIX)
1053 if (file->f_op == &io_uring_fops) {
1054 struct io_ring_ctx *ctx = file->private_data;
1056 return ctx->ring_sock->sk;
1061 EXPORT_SYMBOL(io_uring_get_socket);
1063 #define io_for_each_link(pos, head) \
1064 for (pos = (head); pos; pos = pos->link)
1066 static inline void io_clean_op(struct io_kiocb *req)
1068 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1072 static inline void io_set_resource_node(struct io_kiocb *req)
1074 struct io_ring_ctx *ctx = req->ctx;
1076 if (!req->fixed_rsrc_refs) {
1077 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1078 percpu_ref_get(req->fixed_rsrc_refs);
1082 static bool io_match_task(struct io_kiocb *head,
1083 struct task_struct *task,
1084 struct files_struct *files)
1086 struct io_kiocb *req;
1088 if (task && head->task != task) {
1089 /* in terms of cancelation, always match if req task is dead */
1090 if (head->task->flags & PF_EXITING)
1097 io_for_each_link(req, head) {
1098 if (req->flags & REQ_F_INFLIGHT)
1104 static inline void req_set_fail_links(struct io_kiocb *req)
1106 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1107 req->flags |= REQ_F_FAIL_LINK;
1110 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1112 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1114 complete(&ctx->ref_comp);
1117 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1119 return !req->timeout.off;
1122 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1124 struct io_ring_ctx *ctx;
1127 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1132 * Use 5 bits less than the max cq entries, that should give us around
1133 * 32 entries per hash list if totally full and uniformly spread.
1135 hash_bits = ilog2(p->cq_entries);
1139 ctx->cancel_hash_bits = hash_bits;
1140 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1142 if (!ctx->cancel_hash)
1144 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1146 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1147 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1150 ctx->flags = p->flags;
1151 init_waitqueue_head(&ctx->sqo_sq_wait);
1152 INIT_LIST_HEAD(&ctx->sqd_list);
1153 init_waitqueue_head(&ctx->cq_wait);
1154 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1155 init_completion(&ctx->ref_comp);
1156 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1157 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1158 mutex_init(&ctx->uring_lock);
1159 init_waitqueue_head(&ctx->wait);
1160 spin_lock_init(&ctx->completion_lock);
1161 INIT_LIST_HEAD(&ctx->iopoll_list);
1162 INIT_LIST_HEAD(&ctx->defer_list);
1163 INIT_LIST_HEAD(&ctx->timeout_list);
1164 spin_lock_init(&ctx->inflight_lock);
1165 INIT_LIST_HEAD(&ctx->inflight_list);
1166 spin_lock_init(&ctx->rsrc_ref_lock);
1167 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1168 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1169 init_llist_head(&ctx->rsrc_put_llist);
1170 INIT_LIST_HEAD(&ctx->tctx_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1172 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1175 kfree(ctx->cancel_hash);
1180 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1182 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1183 struct io_ring_ctx *ctx = req->ctx;
1185 return seq != ctx->cached_cq_tail
1186 + READ_ONCE(ctx->cached_cq_overflow);
1192 static void io_req_track_inflight(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!(req->flags & REQ_F_INFLIGHT)) {
1197 req->flags |= REQ_F_INFLIGHT;
1199 spin_lock_irq(&ctx->inflight_lock);
1200 list_add(&req->inflight_entry, &ctx->inflight_list);
1201 spin_unlock_irq(&ctx->inflight_lock);
1205 static void io_prep_async_work(struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 struct io_ring_ctx *ctx = req->ctx;
1210 if (!req->work.creds)
1211 req->work.creds = get_current_cred();
1213 if (req->flags & REQ_F_FORCE_ASYNC)
1214 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1216 if (req->flags & REQ_F_ISREG) {
1217 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1218 io_wq_hash_work(&req->work, file_inode(req->file));
1219 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1220 if (def->unbound_nonreg_file)
1221 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 static void io_prep_async_link(struct io_kiocb *req)
1227 struct io_kiocb *cur;
1229 io_for_each_link(cur, req)
1230 io_prep_async_work(cur);
1233 static void io_queue_async_work(struct io_kiocb *req)
1235 struct io_ring_ctx *ctx = req->ctx;
1236 struct io_kiocb *link = io_prep_linked_timeout(req);
1237 struct io_uring_task *tctx = req->task->io_uring;
1240 BUG_ON(!tctx->io_wq);
1242 /* init ->work of the whole link before punting */
1243 io_prep_async_link(req);
1244 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1245 &req->work, req->flags);
1246 io_wq_enqueue(tctx->io_wq, &req->work);
1248 io_queue_linked_timeout(link);
1251 static void io_kill_timeout(struct io_kiocb *req, int status)
1253 struct io_timeout_data *io = req->async_data;
1256 ret = hrtimer_try_to_cancel(&io->timer);
1258 atomic_set(&req->ctx->cq_timeouts,
1259 atomic_read(&req->ctx->cq_timeouts) + 1);
1260 list_del_init(&req->timeout.list);
1261 io_cqring_fill_event(req, status);
1262 io_put_req_deferred(req, 1);
1266 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1269 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1270 struct io_defer_entry, list);
1272 if (req_need_defer(de->req, de->seq))
1274 list_del_init(&de->list);
1275 io_req_task_queue(de->req);
1277 } while (!list_empty(&ctx->defer_list));
1280 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1284 if (list_empty(&ctx->timeout_list))
1287 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1290 u32 events_needed, events_got;
1291 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1292 struct io_kiocb, timeout.list);
1294 if (io_is_timeout_noseq(req))
1298 * Since seq can easily wrap around over time, subtract
1299 * the last seq at which timeouts were flushed before comparing.
1300 * Assuming not more than 2^31-1 events have happened since,
1301 * these subtractions won't have wrapped, so we can check if
1302 * target is in [last_seq, current_seq] by comparing the two.
1304 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1305 events_got = seq - ctx->cq_last_tm_flush;
1306 if (events_got < events_needed)
1309 list_del_init(&req->timeout.list);
1310 io_kill_timeout(req, 0);
1311 } while (!list_empty(&ctx->timeout_list));
1313 ctx->cq_last_tm_flush = seq;
1316 static void io_commit_cqring(struct io_ring_ctx *ctx)
1318 io_flush_timeouts(ctx);
1320 /* order cqe stores with ring update */
1321 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1323 if (unlikely(!list_empty(&ctx->defer_list)))
1324 __io_queue_deferred(ctx);
1327 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1329 struct io_rings *r = ctx->rings;
1331 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1334 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1336 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1339 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1341 struct io_rings *rings = ctx->rings;
1345 * writes to the cq entry need to come after reading head; the
1346 * control dependency is enough as we're using WRITE_ONCE to
1349 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1352 tail = ctx->cached_cq_tail++;
1353 return &rings->cqes[tail & ctx->cq_mask];
1356 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1360 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1362 if (!ctx->eventfd_async)
1364 return io_wq_current_is_worker();
1367 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1369 /* see waitqueue_active() comment */
1372 if (waitqueue_active(&ctx->wait))
1373 wake_up(&ctx->wait);
1374 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1375 wake_up(&ctx->sq_data->wait);
1376 if (io_should_trigger_evfd(ctx))
1377 eventfd_signal(ctx->cq_ev_fd, 1);
1378 if (waitqueue_active(&ctx->cq_wait)) {
1379 wake_up_interruptible(&ctx->cq_wait);
1380 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1384 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1386 /* see waitqueue_active() comment */
1389 if (ctx->flags & IORING_SETUP_SQPOLL) {
1390 if (waitqueue_active(&ctx->wait))
1391 wake_up(&ctx->wait);
1393 if (io_should_trigger_evfd(ctx))
1394 eventfd_signal(ctx->cq_ev_fd, 1);
1395 if (waitqueue_active(&ctx->cq_wait)) {
1396 wake_up_interruptible(&ctx->cq_wait);
1397 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1401 /* Returns true if there are no backlogged entries after the flush */
1402 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1403 struct task_struct *tsk,
1404 struct files_struct *files)
1406 struct io_rings *rings = ctx->rings;
1407 struct io_kiocb *req, *tmp;
1408 struct io_uring_cqe *cqe;
1409 unsigned long flags;
1410 bool all_flushed, posted;
1413 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1417 spin_lock_irqsave(&ctx->completion_lock, flags);
1418 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1419 if (!io_match_task(req, tsk, files))
1422 cqe = io_get_cqring(ctx);
1426 list_move(&req->compl.list, &list);
1428 WRITE_ONCE(cqe->user_data, req->user_data);
1429 WRITE_ONCE(cqe->res, req->result);
1430 WRITE_ONCE(cqe->flags, req->compl.cflags);
1432 ctx->cached_cq_overflow++;
1433 WRITE_ONCE(ctx->rings->cq_overflow,
1434 ctx->cached_cq_overflow);
1439 all_flushed = list_empty(&ctx->cq_overflow_list);
1441 clear_bit(0, &ctx->sq_check_overflow);
1442 clear_bit(0, &ctx->cq_check_overflow);
1443 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1447 io_commit_cqring(ctx);
1448 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1450 io_cqring_ev_posted(ctx);
1452 while (!list_empty(&list)) {
1453 req = list_first_entry(&list, struct io_kiocb, compl.list);
1454 list_del(&req->compl.list);
1461 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1462 struct task_struct *tsk,
1463 struct files_struct *files)
1467 if (test_bit(0, &ctx->cq_check_overflow)) {
1468 /* iopoll syncs against uring_lock, not completion_lock */
1469 if (ctx->flags & IORING_SETUP_IOPOLL)
1470 mutex_lock(&ctx->uring_lock);
1471 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1472 if (ctx->flags & IORING_SETUP_IOPOLL)
1473 mutex_unlock(&ctx->uring_lock);
1479 static void __io_cqring_fill_event(struct io_kiocb *req, long res,
1480 unsigned int cflags)
1482 struct io_ring_ctx *ctx = req->ctx;
1483 struct io_uring_cqe *cqe;
1485 trace_io_uring_complete(ctx, req->user_data, res);
1488 * If we can't get a cq entry, userspace overflowed the
1489 * submission (by quite a lot). Increment the overflow count in
1492 cqe = io_get_cqring(ctx);
1494 WRITE_ONCE(cqe->user_data, req->user_data);
1495 WRITE_ONCE(cqe->res, res);
1496 WRITE_ONCE(cqe->flags, cflags);
1497 } else if (ctx->cq_overflow_flushed ||
1498 atomic_read(&req->task->io_uring->in_idle)) {
1500 * If we're in ring overflow flush mode, or in task cancel mode,
1501 * then we cannot store the request for later flushing, we need
1502 * to drop it on the floor.
1504 ctx->cached_cq_overflow++;
1505 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1507 if (list_empty(&ctx->cq_overflow_list)) {
1508 set_bit(0, &ctx->sq_check_overflow);
1509 set_bit(0, &ctx->cq_check_overflow);
1510 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1514 req->compl.cflags = cflags;
1515 refcount_inc(&req->refs);
1516 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1520 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1522 __io_cqring_fill_event(req, res, 0);
1525 static void io_req_complete_post(struct io_kiocb *req, long res,
1526 unsigned int cflags)
1528 struct io_ring_ctx *ctx = req->ctx;
1529 unsigned long flags;
1531 spin_lock_irqsave(&ctx->completion_lock, flags);
1532 __io_cqring_fill_event(req, res, cflags);
1534 * If we're the last reference to this request, add to our locked
1537 if (refcount_dec_and_test(&req->refs)) {
1538 struct io_comp_state *cs = &ctx->submit_state.comp;
1540 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1541 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1542 io_disarm_next(req);
1544 io_req_task_queue(req->link);
1548 io_dismantle_req(req);
1549 io_put_task(req->task, 1);
1550 list_add(&req->compl.list, &cs->locked_free_list);
1551 cs->locked_free_nr++;
1553 if (!percpu_ref_tryget(&ctx->refs))
1556 io_commit_cqring(ctx);
1557 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1560 io_cqring_ev_posted(ctx);
1561 percpu_ref_put(&ctx->refs);
1565 static void io_req_complete_state(struct io_kiocb *req, long res,
1566 unsigned int cflags)
1570 req->compl.cflags = cflags;
1571 req->flags |= REQ_F_COMPLETE_INLINE;
1574 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1575 long res, unsigned cflags)
1577 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1578 io_req_complete_state(req, res, cflags);
1580 io_req_complete_post(req, res, cflags);
1583 static inline void io_req_complete(struct io_kiocb *req, long res)
1585 __io_req_complete(req, 0, res, 0);
1588 static void io_req_complete_failed(struct io_kiocb *req, long res)
1590 req_set_fail_links(req);
1592 io_req_complete_post(req, res, 0);
1595 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1597 struct io_submit_state *state = &ctx->submit_state;
1598 struct io_comp_state *cs = &state->comp;
1599 struct io_kiocb *req = NULL;
1602 * If we have more than a batch's worth of requests in our IRQ side
1603 * locked cache, grab the lock and move them over to our submission
1606 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1607 spin_lock_irq(&ctx->completion_lock);
1608 list_splice_init(&cs->locked_free_list, &cs->free_list);
1609 cs->locked_free_nr = 0;
1610 spin_unlock_irq(&ctx->completion_lock);
1613 while (!list_empty(&cs->free_list)) {
1614 req = list_first_entry(&cs->free_list, struct io_kiocb,
1616 list_del(&req->compl.list);
1617 state->reqs[state->free_reqs++] = req;
1618 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1625 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1627 struct io_submit_state *state = &ctx->submit_state;
1629 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1631 if (!state->free_reqs) {
1632 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1635 if (io_flush_cached_reqs(ctx))
1638 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1642 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1643 * retry single alloc to be on the safe side.
1645 if (unlikely(ret <= 0)) {
1646 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1647 if (!state->reqs[0])
1651 state->free_reqs = ret;
1655 return state->reqs[state->free_reqs];
1658 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1665 static void io_dismantle_req(struct io_kiocb *req)
1669 if (req->async_data)
1670 kfree(req->async_data);
1672 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1673 if (req->fixed_rsrc_refs)
1674 percpu_ref_put(req->fixed_rsrc_refs);
1675 if (req->work.creds) {
1676 put_cred(req->work.creds);
1677 req->work.creds = NULL;
1680 if (req->flags & REQ_F_INFLIGHT) {
1681 struct io_ring_ctx *ctx = req->ctx;
1682 unsigned long flags;
1684 spin_lock_irqsave(&ctx->inflight_lock, flags);
1685 list_del(&req->inflight_entry);
1686 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1687 req->flags &= ~REQ_F_INFLIGHT;
1691 /* must to be called somewhat shortly after putting a request */
1692 static inline void io_put_task(struct task_struct *task, int nr)
1694 struct io_uring_task *tctx = task->io_uring;
1696 percpu_counter_sub(&tctx->inflight, nr);
1697 if (unlikely(atomic_read(&tctx->in_idle)))
1698 wake_up(&tctx->wait);
1699 put_task_struct_many(task, nr);
1702 static void __io_free_req(struct io_kiocb *req)
1704 struct io_ring_ctx *ctx = req->ctx;
1706 io_dismantle_req(req);
1707 io_put_task(req->task, 1);
1709 kmem_cache_free(req_cachep, req);
1710 percpu_ref_put(&ctx->refs);
1713 static inline void io_remove_next_linked(struct io_kiocb *req)
1715 struct io_kiocb *nxt = req->link;
1717 req->link = nxt->link;
1721 static bool io_kill_linked_timeout(struct io_kiocb *req)
1722 __must_hold(&req->ctx->completion_lock)
1724 struct io_kiocb *link = req->link;
1725 bool cancelled = false;
1728 * Can happen if a linked timeout fired and link had been like
1729 * req -> link t-out -> link t-out [-> ...]
1731 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1732 struct io_timeout_data *io = link->async_data;
1735 io_remove_next_linked(req);
1736 link->timeout.head = NULL;
1737 ret = hrtimer_try_to_cancel(&io->timer);
1739 io_cqring_fill_event(link, -ECANCELED);
1740 io_put_req_deferred(link, 1);
1744 req->flags &= ~REQ_F_LINK_TIMEOUT;
1748 static void io_fail_links(struct io_kiocb *req)
1749 __must_hold(&req->ctx->completion_lock)
1751 struct io_kiocb *nxt, *link = req->link;
1758 trace_io_uring_fail_link(req, link);
1759 io_cqring_fill_event(link, -ECANCELED);
1760 io_put_req_deferred(link, 2);
1765 static bool io_disarm_next(struct io_kiocb *req)
1766 __must_hold(&req->ctx->completion_lock)
1768 bool posted = false;
1770 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1771 posted = io_kill_linked_timeout(req);
1772 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1773 posted |= (req->link != NULL);
1779 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1781 struct io_kiocb *nxt;
1784 * If LINK is set, we have dependent requests in this chain. If we
1785 * didn't fail this request, queue the first one up, moving any other
1786 * dependencies to the next request. In case of failure, fail the rest
1789 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1790 struct io_ring_ctx *ctx = req->ctx;
1791 unsigned long flags;
1794 spin_lock_irqsave(&ctx->completion_lock, flags);
1795 posted = io_disarm_next(req);
1797 io_commit_cqring(req->ctx);
1798 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1800 io_cqring_ev_posted(ctx);
1807 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1809 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1811 return __io_req_find_next(req);
1814 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1818 if (ctx->submit_state.comp.nr) {
1819 mutex_lock(&ctx->uring_lock);
1820 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1821 mutex_unlock(&ctx->uring_lock);
1823 percpu_ref_put(&ctx->refs);
1826 static bool __tctx_task_work(struct io_uring_task *tctx)
1828 struct io_ring_ctx *ctx = NULL;
1829 struct io_wq_work_list list;
1830 struct io_wq_work_node *node;
1832 if (wq_list_empty(&tctx->task_list))
1835 spin_lock_irq(&tctx->task_lock);
1836 list = tctx->task_list;
1837 INIT_WQ_LIST(&tctx->task_list);
1838 spin_unlock_irq(&tctx->task_lock);
1842 struct io_wq_work_node *next = node->next;
1843 struct io_kiocb *req;
1845 req = container_of(node, struct io_kiocb, io_task_work.node);
1846 if (req->ctx != ctx) {
1847 ctx_flush_and_put(ctx);
1849 percpu_ref_get(&ctx->refs);
1852 req->task_work.func(&req->task_work);
1856 ctx_flush_and_put(ctx);
1857 return list.first != NULL;
1860 static void tctx_task_work(struct callback_head *cb)
1862 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1864 clear_bit(0, &tctx->task_state);
1866 while (__tctx_task_work(tctx))
1870 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1871 enum task_work_notify_mode notify)
1873 struct io_uring_task *tctx = tsk->io_uring;
1874 struct io_wq_work_node *node, *prev;
1875 unsigned long flags;
1878 WARN_ON_ONCE(!tctx);
1880 spin_lock_irqsave(&tctx->task_lock, flags);
1881 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1882 spin_unlock_irqrestore(&tctx->task_lock, flags);
1884 /* task_work already pending, we're done */
1885 if (test_bit(0, &tctx->task_state) ||
1886 test_and_set_bit(0, &tctx->task_state))
1889 if (!task_work_add(tsk, &tctx->task_work, notify))
1893 * Slow path - we failed, find and delete work. if the work is not
1894 * in the list, it got run and we're fine.
1897 spin_lock_irqsave(&tctx->task_lock, flags);
1898 wq_list_for_each(node, prev, &tctx->task_list) {
1899 if (&req->io_task_work.node == node) {
1900 wq_list_del(&tctx->task_list, node, prev);
1905 spin_unlock_irqrestore(&tctx->task_lock, flags);
1906 clear_bit(0, &tctx->task_state);
1910 static int io_req_task_work_add(struct io_kiocb *req)
1912 struct task_struct *tsk = req->task;
1913 struct io_ring_ctx *ctx = req->ctx;
1914 enum task_work_notify_mode notify;
1917 if (tsk->flags & PF_EXITING)
1921 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1922 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1923 * processing task_work. There's no reliable way to tell if TWA_RESUME
1927 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1928 notify = TWA_SIGNAL;
1930 ret = io_task_work_add(tsk, req, notify);
1932 wake_up_process(tsk);
1937 static bool io_run_task_work_head(struct callback_head **work_head)
1939 struct callback_head *work, *next;
1940 bool executed = false;
1943 work = xchg(work_head, NULL);
1959 static void io_task_work_add_head(struct callback_head **work_head,
1960 struct callback_head *task_work)
1962 struct callback_head *head;
1965 head = READ_ONCE(*work_head);
1966 task_work->next = head;
1967 } while (cmpxchg(work_head, head, task_work) != head);
1970 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1971 task_work_func_t cb)
1973 init_task_work(&req->task_work, cb);
1974 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1977 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1979 struct io_ring_ctx *ctx = req->ctx;
1981 spin_lock_irq(&ctx->completion_lock);
1982 io_cqring_fill_event(req, error);
1983 io_commit_cqring(ctx);
1984 spin_unlock_irq(&ctx->completion_lock);
1986 io_cqring_ev_posted(ctx);
1987 req_set_fail_links(req);
1988 io_double_put_req(req);
1991 static void io_req_task_cancel(struct callback_head *cb)
1993 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1994 struct io_ring_ctx *ctx = req->ctx;
1996 /* ctx is guaranteed to stay alive while we hold uring_lock */
1997 mutex_lock(&ctx->uring_lock);
1998 __io_req_task_cancel(req, req->result);
1999 mutex_unlock(&ctx->uring_lock);
2002 static void __io_req_task_submit(struct io_kiocb *req)
2004 struct io_ring_ctx *ctx = req->ctx;
2006 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2007 mutex_lock(&ctx->uring_lock);
2008 if (!(current->flags & PF_EXITING) && !current->in_execve)
2009 __io_queue_sqe(req);
2011 __io_req_task_cancel(req, -EFAULT);
2012 mutex_unlock(&ctx->uring_lock);
2015 static void io_req_task_submit(struct callback_head *cb)
2017 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2019 __io_req_task_submit(req);
2022 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2025 req->task_work.func = io_req_task_cancel;
2027 if (unlikely(io_req_task_work_add(req)))
2028 io_req_task_work_add_fallback(req, io_req_task_cancel);
2031 static void io_req_task_queue(struct io_kiocb *req)
2033 req->task_work.func = io_req_task_submit;
2035 if (unlikely(io_req_task_work_add(req)))
2036 io_req_task_queue_fail(req, -ECANCELED);
2039 static inline void io_queue_next(struct io_kiocb *req)
2041 struct io_kiocb *nxt = io_req_find_next(req);
2044 io_req_task_queue(nxt);
2047 static void io_free_req(struct io_kiocb *req)
2054 struct task_struct *task;
2059 static inline void io_init_req_batch(struct req_batch *rb)
2066 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2067 struct req_batch *rb)
2070 io_put_task(rb->task, rb->task_refs);
2072 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2075 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2076 struct io_submit_state *state)
2080 if (req->task != rb->task) {
2082 io_put_task(rb->task, rb->task_refs);
2083 rb->task = req->task;
2089 io_dismantle_req(req);
2090 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2091 state->reqs[state->free_reqs++] = req;
2093 list_add(&req->compl.list, &state->comp.free_list);
2096 static void io_submit_flush_completions(struct io_comp_state *cs,
2097 struct io_ring_ctx *ctx)
2100 struct io_kiocb *req;
2101 struct req_batch rb;
2103 io_init_req_batch(&rb);
2104 spin_lock_irq(&ctx->completion_lock);
2105 for (i = 0; i < nr; i++) {
2107 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2109 io_commit_cqring(ctx);
2110 spin_unlock_irq(&ctx->completion_lock);
2112 io_cqring_ev_posted(ctx);
2113 for (i = 0; i < nr; i++) {
2116 /* submission and completion refs */
2117 if (refcount_sub_and_test(2, &req->refs))
2118 io_req_free_batch(&rb, req, &ctx->submit_state);
2121 io_req_free_batch_finish(ctx, &rb);
2126 * Drop reference to request, return next in chain (if there is one) if this
2127 * was the last reference to this request.
2129 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2131 struct io_kiocb *nxt = NULL;
2133 if (refcount_dec_and_test(&req->refs)) {
2134 nxt = io_req_find_next(req);
2140 static void io_put_req(struct io_kiocb *req)
2142 if (refcount_dec_and_test(&req->refs))
2146 static void io_put_req_deferred_cb(struct callback_head *cb)
2148 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2153 static void io_free_req_deferred(struct io_kiocb *req)
2157 req->task_work.func = io_put_req_deferred_cb;
2158 ret = io_req_task_work_add(req);
2160 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2163 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2165 if (refcount_sub_and_test(refs, &req->refs))
2166 io_free_req_deferred(req);
2169 static void io_double_put_req(struct io_kiocb *req)
2171 /* drop both submit and complete references */
2172 if (refcount_sub_and_test(2, &req->refs))
2176 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2178 /* See comment at the top of this file */
2180 return __io_cqring_events(ctx);
2183 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2185 struct io_rings *rings = ctx->rings;
2187 /* make sure SQ entry isn't read before tail */
2188 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2191 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2193 unsigned int cflags;
2195 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2196 cflags |= IORING_CQE_F_BUFFER;
2197 req->flags &= ~REQ_F_BUFFER_SELECTED;
2202 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2204 struct io_buffer *kbuf;
2206 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2207 return io_put_kbuf(req, kbuf);
2210 static inline bool io_run_task_work(void)
2213 * Not safe to run on exiting task, and the task_work handling will
2214 * not add work to such a task.
2216 if (unlikely(current->flags & PF_EXITING))
2218 if (current->task_works) {
2219 __set_current_state(TASK_RUNNING);
2228 * Find and free completed poll iocbs
2230 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2231 struct list_head *done)
2233 struct req_batch rb;
2234 struct io_kiocb *req;
2236 /* order with ->result store in io_complete_rw_iopoll() */
2239 io_init_req_batch(&rb);
2240 while (!list_empty(done)) {
2243 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2244 list_del(&req->inflight_entry);
2246 if (READ_ONCE(req->result) == -EAGAIN) {
2247 req->iopoll_completed = 0;
2248 if (io_rw_reissue(req))
2252 if (req->flags & REQ_F_BUFFER_SELECTED)
2253 cflags = io_put_rw_kbuf(req);
2255 __io_cqring_fill_event(req, req->result, cflags);
2258 if (refcount_dec_and_test(&req->refs))
2259 io_req_free_batch(&rb, req, &ctx->submit_state);
2262 io_commit_cqring(ctx);
2263 io_cqring_ev_posted_iopoll(ctx);
2264 io_req_free_batch_finish(ctx, &rb);
2267 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2270 struct io_kiocb *req, *tmp;
2276 * Only spin for completions if we don't have multiple devices hanging
2277 * off our complete list, and we're under the requested amount.
2279 spin = !ctx->poll_multi_file && *nr_events < min;
2282 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2283 struct kiocb *kiocb = &req->rw.kiocb;
2286 * Move completed and retryable entries to our local lists.
2287 * If we find a request that requires polling, break out
2288 * and complete those lists first, if we have entries there.
2290 if (READ_ONCE(req->iopoll_completed)) {
2291 list_move_tail(&req->inflight_entry, &done);
2294 if (!list_empty(&done))
2297 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2301 /* iopoll may have completed current req */
2302 if (READ_ONCE(req->iopoll_completed))
2303 list_move_tail(&req->inflight_entry, &done);
2310 if (!list_empty(&done))
2311 io_iopoll_complete(ctx, nr_events, &done);
2317 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2318 * non-spinning poll check - we'll still enter the driver poll loop, but only
2319 * as a non-spinning completion check.
2321 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2324 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2327 ret = io_do_iopoll(ctx, nr_events, min);
2330 if (*nr_events >= min)
2338 * We can't just wait for polled events to come to us, we have to actively
2339 * find and complete them.
2341 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2343 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2346 mutex_lock(&ctx->uring_lock);
2347 while (!list_empty(&ctx->iopoll_list)) {
2348 unsigned int nr_events = 0;
2350 io_do_iopoll(ctx, &nr_events, 0);
2352 /* let it sleep and repeat later if can't complete a request */
2356 * Ensure we allow local-to-the-cpu processing to take place,
2357 * in this case we need to ensure that we reap all events.
2358 * Also let task_work, etc. to progress by releasing the mutex
2360 if (need_resched()) {
2361 mutex_unlock(&ctx->uring_lock);
2363 mutex_lock(&ctx->uring_lock);
2366 mutex_unlock(&ctx->uring_lock);
2369 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2371 unsigned int nr_events = 0;
2372 int iters = 0, ret = 0;
2375 * We disallow the app entering submit/complete with polling, but we
2376 * still need to lock the ring to prevent racing with polled issue
2377 * that got punted to a workqueue.
2379 mutex_lock(&ctx->uring_lock);
2382 * Don't enter poll loop if we already have events pending.
2383 * If we do, we can potentially be spinning for commands that
2384 * already triggered a CQE (eg in error).
2386 if (test_bit(0, &ctx->cq_check_overflow))
2387 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2388 if (io_cqring_events(ctx))
2392 * If a submit got punted to a workqueue, we can have the
2393 * application entering polling for a command before it gets
2394 * issued. That app will hold the uring_lock for the duration
2395 * of the poll right here, so we need to take a breather every
2396 * now and then to ensure that the issue has a chance to add
2397 * the poll to the issued list. Otherwise we can spin here
2398 * forever, while the workqueue is stuck trying to acquire the
2401 if (!(++iters & 7)) {
2402 mutex_unlock(&ctx->uring_lock);
2404 mutex_lock(&ctx->uring_lock);
2407 ret = io_iopoll_getevents(ctx, &nr_events, min);
2411 } while (min && !nr_events && !need_resched());
2413 mutex_unlock(&ctx->uring_lock);
2417 static void kiocb_end_write(struct io_kiocb *req)
2420 * Tell lockdep we inherited freeze protection from submission
2423 if (req->flags & REQ_F_ISREG) {
2424 struct inode *inode = file_inode(req->file);
2426 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2428 file_end_write(req->file);
2432 static bool io_resubmit_prep(struct io_kiocb *req)
2434 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2436 struct iov_iter iter;
2438 /* already prepared */
2439 if (req->async_data)
2442 switch (req->opcode) {
2443 case IORING_OP_READV:
2444 case IORING_OP_READ_FIXED:
2445 case IORING_OP_READ:
2448 case IORING_OP_WRITEV:
2449 case IORING_OP_WRITE_FIXED:
2450 case IORING_OP_WRITE:
2454 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2459 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2462 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2465 static bool io_rw_should_reissue(struct io_kiocb *req)
2467 umode_t mode = file_inode(req->file)->i_mode;
2468 struct io_ring_ctx *ctx = req->ctx;
2470 if (!S_ISBLK(mode) && !S_ISREG(mode))
2472 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2473 !(ctx->flags & IORING_SETUP_IOPOLL)))
2476 * If ref is dying, we might be running poll reap from the exit work.
2477 * Don't attempt to reissue from that path, just let it fail with
2480 if (percpu_ref_is_dying(&ctx->refs))
2485 static bool io_rw_should_reissue(struct io_kiocb *req)
2491 static bool io_rw_reissue(struct io_kiocb *req)
2494 if (!io_rw_should_reissue(req))
2497 lockdep_assert_held(&req->ctx->uring_lock);
2499 if (io_resubmit_prep(req)) {
2500 refcount_inc(&req->refs);
2501 io_queue_async_work(req);
2504 req_set_fail_links(req);
2509 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2510 unsigned int issue_flags)
2514 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2515 kiocb_end_write(req);
2516 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2517 req->flags |= REQ_F_REISSUE;
2520 if (res != req->result)
2521 req_set_fail_links(req);
2522 if (req->flags & REQ_F_BUFFER_SELECTED)
2523 cflags = io_put_rw_kbuf(req);
2524 __io_req_complete(req, issue_flags, res, cflags);
2527 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2529 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2531 __io_complete_rw(req, res, res2, 0);
2534 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2536 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2539 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2540 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2541 struct io_async_rw *rw = req->async_data;
2544 iov_iter_revert(&rw->iter,
2545 req->result - iov_iter_count(&rw->iter));
2546 else if (!io_resubmit_prep(req))
2551 if (kiocb->ki_flags & IOCB_WRITE)
2552 kiocb_end_write(req);
2554 if (res != -EAGAIN && res != req->result)
2555 req_set_fail_links(req);
2557 WRITE_ONCE(req->result, res);
2558 /* order with io_poll_complete() checking ->result */
2560 WRITE_ONCE(req->iopoll_completed, 1);
2564 * After the iocb has been issued, it's safe to be found on the poll list.
2565 * Adding the kiocb to the list AFTER submission ensures that we don't
2566 * find it from a io_iopoll_getevents() thread before the issuer is done
2567 * accessing the kiocb cookie.
2569 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2571 struct io_ring_ctx *ctx = req->ctx;
2574 * Track whether we have multiple files in our lists. This will impact
2575 * how we do polling eventually, not spinning if we're on potentially
2576 * different devices.
2578 if (list_empty(&ctx->iopoll_list)) {
2579 ctx->poll_multi_file = false;
2580 } else if (!ctx->poll_multi_file) {
2581 struct io_kiocb *list_req;
2583 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2585 if (list_req->file != req->file)
2586 ctx->poll_multi_file = true;
2590 * For fast devices, IO may have already completed. If it has, add
2591 * it to the front so we find it first.
2593 if (READ_ONCE(req->iopoll_completed))
2594 list_add(&req->inflight_entry, &ctx->iopoll_list);
2596 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2599 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2600 * task context or in io worker task context. If current task context is
2601 * sq thread, we don't need to check whether should wake up sq thread.
2603 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2604 wq_has_sleeper(&ctx->sq_data->wait))
2605 wake_up(&ctx->sq_data->wait);
2608 static inline void io_state_file_put(struct io_submit_state *state)
2610 if (state->file_refs) {
2611 fput_many(state->file, state->file_refs);
2612 state->file_refs = 0;
2617 * Get as many references to a file as we have IOs left in this submission,
2618 * assuming most submissions are for one file, or at least that each file
2619 * has more than one submission.
2621 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2626 if (state->file_refs) {
2627 if (state->fd == fd) {
2631 io_state_file_put(state);
2633 state->file = fget_many(fd, state->ios_left);
2634 if (unlikely(!state->file))
2638 state->file_refs = state->ios_left - 1;
2642 static bool io_bdev_nowait(struct block_device *bdev)
2644 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2648 * If we tracked the file through the SCM inflight mechanism, we could support
2649 * any file. For now, just ensure that anything potentially problematic is done
2652 static bool io_file_supports_async(struct file *file, int rw)
2654 umode_t mode = file_inode(file)->i_mode;
2656 if (S_ISBLK(mode)) {
2657 if (IS_ENABLED(CONFIG_BLOCK) &&
2658 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2662 if (S_ISCHR(mode) || S_ISSOCK(mode))
2664 if (S_ISREG(mode)) {
2665 if (IS_ENABLED(CONFIG_BLOCK) &&
2666 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2667 file->f_op != &io_uring_fops)
2672 /* any ->read/write should understand O_NONBLOCK */
2673 if (file->f_flags & O_NONBLOCK)
2676 if (!(file->f_mode & FMODE_NOWAIT))
2680 return file->f_op->read_iter != NULL;
2682 return file->f_op->write_iter != NULL;
2685 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2687 struct io_ring_ctx *ctx = req->ctx;
2688 struct kiocb *kiocb = &req->rw.kiocb;
2689 struct file *file = req->file;
2693 if (S_ISREG(file_inode(file)->i_mode))
2694 req->flags |= REQ_F_ISREG;
2696 kiocb->ki_pos = READ_ONCE(sqe->off);
2697 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2698 req->flags |= REQ_F_CUR_POS;
2699 kiocb->ki_pos = file->f_pos;
2701 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2702 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2703 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2707 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2708 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2709 req->flags |= REQ_F_NOWAIT;
2711 ioprio = READ_ONCE(sqe->ioprio);
2713 ret = ioprio_check_cap(ioprio);
2717 kiocb->ki_ioprio = ioprio;
2719 kiocb->ki_ioprio = get_current_ioprio();
2721 if (ctx->flags & IORING_SETUP_IOPOLL) {
2722 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2723 !kiocb->ki_filp->f_op->iopoll)
2726 kiocb->ki_flags |= IOCB_HIPRI;
2727 kiocb->ki_complete = io_complete_rw_iopoll;
2728 req->iopoll_completed = 0;
2730 if (kiocb->ki_flags & IOCB_HIPRI)
2732 kiocb->ki_complete = io_complete_rw;
2735 req->rw.addr = READ_ONCE(sqe->addr);
2736 req->rw.len = READ_ONCE(sqe->len);
2737 req->buf_index = READ_ONCE(sqe->buf_index);
2741 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2747 case -ERESTARTNOINTR:
2748 case -ERESTARTNOHAND:
2749 case -ERESTART_RESTARTBLOCK:
2751 * We can't just restart the syscall, since previously
2752 * submitted sqes may already be in progress. Just fail this
2758 kiocb->ki_complete(kiocb, ret, 0);
2762 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2763 unsigned int issue_flags)
2765 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2766 struct io_async_rw *io = req->async_data;
2767 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2769 /* add previously done IO, if any */
2770 if (io && io->bytes_done > 0) {
2772 ret = io->bytes_done;
2774 ret += io->bytes_done;
2777 if (req->flags & REQ_F_CUR_POS)
2778 req->file->f_pos = kiocb->ki_pos;
2779 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2780 __io_complete_rw(req, ret, 0, issue_flags);
2782 io_rw_done(kiocb, ret);
2784 if (check_reissue && req->flags & REQ_F_REISSUE) {
2785 req->flags &= ~REQ_F_REISSUE;
2786 if (!io_rw_reissue(req)) {
2789 req_set_fail_links(req);
2790 if (req->flags & REQ_F_BUFFER_SELECTED)
2791 cflags = io_put_rw_kbuf(req);
2792 __io_req_complete(req, issue_flags, ret, cflags);
2797 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2799 struct io_ring_ctx *ctx = req->ctx;
2800 size_t len = req->rw.len;
2801 struct io_mapped_ubuf *imu;
2802 u16 index, buf_index = req->buf_index;
2806 if (unlikely(buf_index >= ctx->nr_user_bufs))
2808 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2809 imu = &ctx->user_bufs[index];
2810 buf_addr = req->rw.addr;
2813 if (buf_addr + len < buf_addr)
2815 /* not inside the mapped region */
2816 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2820 * May not be a start of buffer, set size appropriately
2821 * and advance us to the beginning.
2823 offset = buf_addr - imu->ubuf;
2824 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2828 * Don't use iov_iter_advance() here, as it's really slow for
2829 * using the latter parts of a big fixed buffer - it iterates
2830 * over each segment manually. We can cheat a bit here, because
2833 * 1) it's a BVEC iter, we set it up
2834 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2835 * first and last bvec
2837 * So just find our index, and adjust the iterator afterwards.
2838 * If the offset is within the first bvec (or the whole first
2839 * bvec, just use iov_iter_advance(). This makes it easier
2840 * since we can just skip the first segment, which may not
2841 * be PAGE_SIZE aligned.
2843 const struct bio_vec *bvec = imu->bvec;
2845 if (offset <= bvec->bv_len) {
2846 iov_iter_advance(iter, offset);
2848 unsigned long seg_skip;
2850 /* skip first vec */
2851 offset -= bvec->bv_len;
2852 seg_skip = 1 + (offset >> PAGE_SHIFT);
2854 iter->bvec = bvec + seg_skip;
2855 iter->nr_segs -= seg_skip;
2856 iter->count -= bvec->bv_len + offset;
2857 iter->iov_offset = offset & ~PAGE_MASK;
2864 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2867 mutex_unlock(&ctx->uring_lock);
2870 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2873 * "Normal" inline submissions always hold the uring_lock, since we
2874 * grab it from the system call. Same is true for the SQPOLL offload.
2875 * The only exception is when we've detached the request and issue it
2876 * from an async worker thread, grab the lock for that case.
2879 mutex_lock(&ctx->uring_lock);
2882 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2883 int bgid, struct io_buffer *kbuf,
2886 struct io_buffer *head;
2888 if (req->flags & REQ_F_BUFFER_SELECTED)
2891 io_ring_submit_lock(req->ctx, needs_lock);
2893 lockdep_assert_held(&req->ctx->uring_lock);
2895 head = xa_load(&req->ctx->io_buffers, bgid);
2897 if (!list_empty(&head->list)) {
2898 kbuf = list_last_entry(&head->list, struct io_buffer,
2900 list_del(&kbuf->list);
2903 xa_erase(&req->ctx->io_buffers, bgid);
2905 if (*len > kbuf->len)
2908 kbuf = ERR_PTR(-ENOBUFS);
2911 io_ring_submit_unlock(req->ctx, needs_lock);
2916 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2919 struct io_buffer *kbuf;
2922 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2923 bgid = req->buf_index;
2924 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2927 req->rw.addr = (u64) (unsigned long) kbuf;
2928 req->flags |= REQ_F_BUFFER_SELECTED;
2929 return u64_to_user_ptr(kbuf->addr);
2932 #ifdef CONFIG_COMPAT
2933 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2936 struct compat_iovec __user *uiov;
2937 compat_ssize_t clen;
2941 uiov = u64_to_user_ptr(req->rw.addr);
2942 if (!access_ok(uiov, sizeof(*uiov)))
2944 if (__get_user(clen, &uiov->iov_len))
2950 buf = io_rw_buffer_select(req, &len, needs_lock);
2952 return PTR_ERR(buf);
2953 iov[0].iov_base = buf;
2954 iov[0].iov_len = (compat_size_t) len;
2959 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2962 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2966 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2969 len = iov[0].iov_len;
2972 buf = io_rw_buffer_select(req, &len, needs_lock);
2974 return PTR_ERR(buf);
2975 iov[0].iov_base = buf;
2976 iov[0].iov_len = len;
2980 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2983 if (req->flags & REQ_F_BUFFER_SELECTED) {
2984 struct io_buffer *kbuf;
2986 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2987 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2988 iov[0].iov_len = kbuf->len;
2991 if (req->rw.len != 1)
2994 #ifdef CONFIG_COMPAT
2995 if (req->ctx->compat)
2996 return io_compat_import(req, iov, needs_lock);
2999 return __io_iov_buffer_select(req, iov, needs_lock);
3002 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3003 struct iov_iter *iter, bool needs_lock)
3005 void __user *buf = u64_to_user_ptr(req->rw.addr);
3006 size_t sqe_len = req->rw.len;
3007 u8 opcode = req->opcode;
3010 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3012 return io_import_fixed(req, rw, iter);
3015 /* buffer index only valid with fixed read/write, or buffer select */
3016 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3019 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3020 if (req->flags & REQ_F_BUFFER_SELECT) {
3021 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3023 return PTR_ERR(buf);
3024 req->rw.len = sqe_len;
3027 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3032 if (req->flags & REQ_F_BUFFER_SELECT) {
3033 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3035 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3040 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3044 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3046 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3050 * For files that don't have ->read_iter() and ->write_iter(), handle them
3051 * by looping over ->read() or ->write() manually.
3053 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3055 struct kiocb *kiocb = &req->rw.kiocb;
3056 struct file *file = req->file;
3060 * Don't support polled IO through this interface, and we can't
3061 * support non-blocking either. For the latter, this just causes
3062 * the kiocb to be handled from an async context.
3064 if (kiocb->ki_flags & IOCB_HIPRI)
3066 if (kiocb->ki_flags & IOCB_NOWAIT)
3069 while (iov_iter_count(iter)) {
3073 if (!iov_iter_is_bvec(iter)) {
3074 iovec = iov_iter_iovec(iter);
3076 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3077 iovec.iov_len = req->rw.len;
3081 nr = file->f_op->read(file, iovec.iov_base,
3082 iovec.iov_len, io_kiocb_ppos(kiocb));
3084 nr = file->f_op->write(file, iovec.iov_base,
3085 iovec.iov_len, io_kiocb_ppos(kiocb));
3094 if (nr != iovec.iov_len)
3098 iov_iter_advance(iter, nr);
3104 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3105 const struct iovec *fast_iov, struct iov_iter *iter)
3107 struct io_async_rw *rw = req->async_data;
3109 memcpy(&rw->iter, iter, sizeof(*iter));
3110 rw->free_iovec = iovec;
3112 /* can only be fixed buffers, no need to do anything */
3113 if (iov_iter_is_bvec(iter))
3116 unsigned iov_off = 0;
3118 rw->iter.iov = rw->fast_iov;
3119 if (iter->iov != fast_iov) {
3120 iov_off = iter->iov - fast_iov;
3121 rw->iter.iov += iov_off;
3123 if (rw->fast_iov != fast_iov)
3124 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3125 sizeof(struct iovec) * iter->nr_segs);
3127 req->flags |= REQ_F_NEED_CLEANUP;
3131 static inline int __io_alloc_async_data(struct io_kiocb *req)
3133 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3134 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3135 return req->async_data == NULL;
3138 static int io_alloc_async_data(struct io_kiocb *req)
3140 if (!io_op_defs[req->opcode].needs_async_data)
3143 return __io_alloc_async_data(req);
3146 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3147 const struct iovec *fast_iov,
3148 struct iov_iter *iter, bool force)
3150 if (!force && !io_op_defs[req->opcode].needs_async_data)
3152 if (!req->async_data) {
3153 if (__io_alloc_async_data(req)) {
3158 io_req_map_rw(req, iovec, fast_iov, iter);
3163 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3165 struct io_async_rw *iorw = req->async_data;
3166 struct iovec *iov = iorw->fast_iov;
3169 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3170 if (unlikely(ret < 0))
3173 iorw->bytes_done = 0;
3174 iorw->free_iovec = iov;
3176 req->flags |= REQ_F_NEED_CLEANUP;
3180 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3182 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3184 return io_prep_rw(req, sqe);
3188 * This is our waitqueue callback handler, registered through lock_page_async()
3189 * when we initially tried to do the IO with the iocb armed our waitqueue.
3190 * This gets called when the page is unlocked, and we generally expect that to
3191 * happen when the page IO is completed and the page is now uptodate. This will
3192 * queue a task_work based retry of the operation, attempting to copy the data
3193 * again. If the latter fails because the page was NOT uptodate, then we will
3194 * do a thread based blocking retry of the operation. That's the unexpected
3197 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3198 int sync, void *arg)
3200 struct wait_page_queue *wpq;
3201 struct io_kiocb *req = wait->private;
3202 struct wait_page_key *key = arg;
3204 wpq = container_of(wait, struct wait_page_queue, wait);
3206 if (!wake_page_match(wpq, key))
3209 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3210 list_del_init(&wait->entry);
3212 /* submit ref gets dropped, acquire a new one */
3213 refcount_inc(&req->refs);
3214 io_req_task_queue(req);
3219 * This controls whether a given IO request should be armed for async page
3220 * based retry. If we return false here, the request is handed to the async
3221 * worker threads for retry. If we're doing buffered reads on a regular file,
3222 * we prepare a private wait_page_queue entry and retry the operation. This
3223 * will either succeed because the page is now uptodate and unlocked, or it
3224 * will register a callback when the page is unlocked at IO completion. Through
3225 * that callback, io_uring uses task_work to setup a retry of the operation.
3226 * That retry will attempt the buffered read again. The retry will generally
3227 * succeed, or in rare cases where it fails, we then fall back to using the
3228 * async worker threads for a blocking retry.
3230 static bool io_rw_should_retry(struct io_kiocb *req)
3232 struct io_async_rw *rw = req->async_data;
3233 struct wait_page_queue *wait = &rw->wpq;
3234 struct kiocb *kiocb = &req->rw.kiocb;
3236 /* never retry for NOWAIT, we just complete with -EAGAIN */
3237 if (req->flags & REQ_F_NOWAIT)
3240 /* Only for buffered IO */
3241 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3245 * just use poll if we can, and don't attempt if the fs doesn't
3246 * support callback based unlocks
3248 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3251 wait->wait.func = io_async_buf_func;
3252 wait->wait.private = req;
3253 wait->wait.flags = 0;
3254 INIT_LIST_HEAD(&wait->wait.entry);
3255 kiocb->ki_flags |= IOCB_WAITQ;
3256 kiocb->ki_flags &= ~IOCB_NOWAIT;
3257 kiocb->ki_waitq = wait;
3261 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3263 if (req->file->f_op->read_iter)
3264 return call_read_iter(req->file, &req->rw.kiocb, iter);
3265 else if (req->file->f_op->read)
3266 return loop_rw_iter(READ, req, iter);
3271 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3273 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3274 struct kiocb *kiocb = &req->rw.kiocb;
3275 struct iov_iter __iter, *iter = &__iter;
3276 struct io_async_rw *rw = req->async_data;
3277 ssize_t io_size, ret, ret2;
3278 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3284 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3288 io_size = iov_iter_count(iter);
3289 req->result = io_size;
3291 /* Ensure we clear previously set non-block flag */
3292 if (!force_nonblock)
3293 kiocb->ki_flags &= ~IOCB_NOWAIT;
3295 kiocb->ki_flags |= IOCB_NOWAIT;
3297 /* If the file doesn't support async, just async punt */
3298 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3299 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3300 return ret ?: -EAGAIN;
3303 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3304 if (unlikely(ret)) {
3309 ret = io_iter_do_read(req, iter);
3311 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3312 req->flags &= ~REQ_F_REISSUE;
3313 /* IOPOLL retry should happen for io-wq threads */
3314 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3316 /* no retry on NONBLOCK nor RWF_NOWAIT */
3317 if (req->flags & REQ_F_NOWAIT)
3319 /* some cases will consume bytes even on error returns */
3320 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3322 } else if (ret == -EIOCBQUEUED) {
3324 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3325 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3326 /* read all, failed, already did sync or don't want to retry */
3330 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3335 rw = req->async_data;
3336 /* now use our persistent iterator, if we aren't already */
3341 rw->bytes_done += ret;
3342 /* if we can retry, do so with the callbacks armed */
3343 if (!io_rw_should_retry(req)) {
3344 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3350 * we get -EIOCBQUEUED, then we'll get a notification when the
3351 * desired page gets unlocked. We can also get a partial read
3352 * here, and if we do, then just retry at the new offset.
3354 ret = io_iter_do_read(req, iter);
3355 if (ret == -EIOCBQUEUED)
3357 /* we got some bytes, but not all. retry. */
3358 kiocb->ki_flags &= ~IOCB_WAITQ;
3359 } while (ret > 0 && ret < io_size);
3361 kiocb_done(kiocb, ret, issue_flags);
3363 /* it's faster to check here then delegate to kfree */
3369 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3371 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3373 return io_prep_rw(req, sqe);
3376 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3378 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3379 struct kiocb *kiocb = &req->rw.kiocb;
3380 struct iov_iter __iter, *iter = &__iter;
3381 struct io_async_rw *rw = req->async_data;
3382 ssize_t ret, ret2, io_size;
3383 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3389 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3393 io_size = iov_iter_count(iter);
3394 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3402 /* If the file doesn't support async, just async punt */
3403 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3406 /* file path doesn't support NOWAIT for non-direct_IO */
3407 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3408 (req->flags & REQ_F_ISREG))
3411 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3416 * Open-code file_start_write here to grab freeze protection,
3417 * which will be released by another thread in
3418 * io_complete_rw(). Fool lockdep by telling it the lock got
3419 * released so that it doesn't complain about the held lock when
3420 * we return to userspace.
3422 if (req->flags & REQ_F_ISREG) {
3423 sb_start_write(file_inode(req->file)->i_sb);
3424 __sb_writers_release(file_inode(req->file)->i_sb,
3427 kiocb->ki_flags |= IOCB_WRITE;
3429 if (req->file->f_op->write_iter)
3430 ret2 = call_write_iter(req->file, kiocb, iter);
3431 else if (req->file->f_op->write)
3432 ret2 = loop_rw_iter(WRITE, req, iter);
3436 if (req->flags & REQ_F_REISSUE) {
3437 req->flags &= ~REQ_F_REISSUE;
3442 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3443 * retry them without IOCB_NOWAIT.
3445 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3447 /* no retry on NONBLOCK nor RWF_NOWAIT */
3448 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3450 if (!force_nonblock || ret2 != -EAGAIN) {
3451 /* IOPOLL retry should happen for io-wq threads */
3452 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3455 kiocb_done(kiocb, ret2, issue_flags);
3458 /* some cases will consume bytes even on error returns */
3459 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3460 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3461 return ret ?: -EAGAIN;
3464 /* it's reportedly faster than delegating the null check to kfree() */
3470 static int io_renameat_prep(struct io_kiocb *req,
3471 const struct io_uring_sqe *sqe)
3473 struct io_rename *ren = &req->rename;
3474 const char __user *oldf, *newf;
3476 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3479 ren->old_dfd = READ_ONCE(sqe->fd);
3480 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3481 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3482 ren->new_dfd = READ_ONCE(sqe->len);
3483 ren->flags = READ_ONCE(sqe->rename_flags);
3485 ren->oldpath = getname(oldf);
3486 if (IS_ERR(ren->oldpath))
3487 return PTR_ERR(ren->oldpath);
3489 ren->newpath = getname(newf);
3490 if (IS_ERR(ren->newpath)) {
3491 putname(ren->oldpath);
3492 return PTR_ERR(ren->newpath);
3495 req->flags |= REQ_F_NEED_CLEANUP;
3499 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3501 struct io_rename *ren = &req->rename;
3504 if (issue_flags & IO_URING_F_NONBLOCK)
3507 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3508 ren->newpath, ren->flags);
3510 req->flags &= ~REQ_F_NEED_CLEANUP;
3512 req_set_fail_links(req);
3513 io_req_complete(req, ret);
3517 static int io_unlinkat_prep(struct io_kiocb *req,
3518 const struct io_uring_sqe *sqe)
3520 struct io_unlink *un = &req->unlink;
3521 const char __user *fname;
3523 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3526 un->dfd = READ_ONCE(sqe->fd);
3528 un->flags = READ_ONCE(sqe->unlink_flags);
3529 if (un->flags & ~AT_REMOVEDIR)
3532 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3533 un->filename = getname(fname);
3534 if (IS_ERR(un->filename))
3535 return PTR_ERR(un->filename);
3537 req->flags |= REQ_F_NEED_CLEANUP;
3541 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3543 struct io_unlink *un = &req->unlink;
3546 if (issue_flags & IO_URING_F_NONBLOCK)
3549 if (un->flags & AT_REMOVEDIR)
3550 ret = do_rmdir(un->dfd, un->filename);
3552 ret = do_unlinkat(un->dfd, un->filename);
3554 req->flags &= ~REQ_F_NEED_CLEANUP;
3556 req_set_fail_links(req);
3557 io_req_complete(req, ret);
3561 static int io_shutdown_prep(struct io_kiocb *req,
3562 const struct io_uring_sqe *sqe)
3564 #if defined(CONFIG_NET)
3565 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3567 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3571 req->shutdown.how = READ_ONCE(sqe->len);
3578 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3580 #if defined(CONFIG_NET)
3581 struct socket *sock;
3584 if (issue_flags & IO_URING_F_NONBLOCK)
3587 sock = sock_from_file(req->file);
3588 if (unlikely(!sock))
3591 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3593 req_set_fail_links(req);
3594 io_req_complete(req, ret);
3601 static int __io_splice_prep(struct io_kiocb *req,
3602 const struct io_uring_sqe *sqe)
3604 struct io_splice* sp = &req->splice;
3605 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3607 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3611 sp->len = READ_ONCE(sqe->len);
3612 sp->flags = READ_ONCE(sqe->splice_flags);
3614 if (unlikely(sp->flags & ~valid_flags))
3617 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3618 (sp->flags & SPLICE_F_FD_IN_FIXED));
3621 req->flags |= REQ_F_NEED_CLEANUP;
3623 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3625 * Splice operation will be punted aync, and here need to
3626 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3628 req->work.flags |= IO_WQ_WORK_UNBOUND;
3634 static int io_tee_prep(struct io_kiocb *req,
3635 const struct io_uring_sqe *sqe)
3637 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3639 return __io_splice_prep(req, sqe);
3642 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3644 struct io_splice *sp = &req->splice;
3645 struct file *in = sp->file_in;
3646 struct file *out = sp->file_out;
3647 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3650 if (issue_flags & IO_URING_F_NONBLOCK)
3653 ret = do_tee(in, out, sp->len, flags);
3655 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3656 req->flags &= ~REQ_F_NEED_CLEANUP;
3659 req_set_fail_links(req);
3660 io_req_complete(req, ret);
3664 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3666 struct io_splice* sp = &req->splice;
3668 sp->off_in = READ_ONCE(sqe->splice_off_in);
3669 sp->off_out = READ_ONCE(sqe->off);
3670 return __io_splice_prep(req, sqe);
3673 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3675 struct io_splice *sp = &req->splice;
3676 struct file *in = sp->file_in;
3677 struct file *out = sp->file_out;
3678 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3679 loff_t *poff_in, *poff_out;
3682 if (issue_flags & IO_URING_F_NONBLOCK)
3685 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3686 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3689 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3691 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3692 req->flags &= ~REQ_F_NEED_CLEANUP;
3695 req_set_fail_links(req);
3696 io_req_complete(req, ret);
3701 * IORING_OP_NOP just posts a completion event, nothing else.
3703 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3705 struct io_ring_ctx *ctx = req->ctx;
3707 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3710 __io_req_complete(req, issue_flags, 0, 0);
3714 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3716 struct io_ring_ctx *ctx = req->ctx;
3721 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3723 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3726 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3727 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3730 req->sync.off = READ_ONCE(sqe->off);
3731 req->sync.len = READ_ONCE(sqe->len);
3735 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3737 loff_t end = req->sync.off + req->sync.len;
3740 /* fsync always requires a blocking context */
3741 if (issue_flags & IO_URING_F_NONBLOCK)
3744 ret = vfs_fsync_range(req->file, req->sync.off,
3745 end > 0 ? end : LLONG_MAX,
3746 req->sync.flags & IORING_FSYNC_DATASYNC);
3748 req_set_fail_links(req);
3749 io_req_complete(req, ret);
3753 static int io_fallocate_prep(struct io_kiocb *req,
3754 const struct io_uring_sqe *sqe)
3756 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3758 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3761 req->sync.off = READ_ONCE(sqe->off);
3762 req->sync.len = READ_ONCE(sqe->addr);
3763 req->sync.mode = READ_ONCE(sqe->len);
3767 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3771 /* fallocate always requiring blocking context */
3772 if (issue_flags & IO_URING_F_NONBLOCK)
3774 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3777 req_set_fail_links(req);
3778 io_req_complete(req, ret);
3782 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3784 const char __user *fname;
3787 if (unlikely(sqe->ioprio || sqe->buf_index))
3789 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3792 /* open.how should be already initialised */
3793 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3794 req->open.how.flags |= O_LARGEFILE;
3796 req->open.dfd = READ_ONCE(sqe->fd);
3797 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3798 req->open.filename = getname(fname);
3799 if (IS_ERR(req->open.filename)) {
3800 ret = PTR_ERR(req->open.filename);
3801 req->open.filename = NULL;
3804 req->open.nofile = rlimit(RLIMIT_NOFILE);
3805 req->flags |= REQ_F_NEED_CLEANUP;
3809 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3813 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3815 mode = READ_ONCE(sqe->len);
3816 flags = READ_ONCE(sqe->open_flags);
3817 req->open.how = build_open_how(flags, mode);
3818 return __io_openat_prep(req, sqe);
3821 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3823 struct open_how __user *how;
3827 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3829 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3830 len = READ_ONCE(sqe->len);
3831 if (len < OPEN_HOW_SIZE_VER0)
3834 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3839 return __io_openat_prep(req, sqe);
3842 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3844 struct open_flags op;
3847 bool resolve_nonblock;
3850 ret = build_open_flags(&req->open.how, &op);
3853 nonblock_set = op.open_flag & O_NONBLOCK;
3854 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3855 if (issue_flags & IO_URING_F_NONBLOCK) {
3857 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3858 * it'll always -EAGAIN
3860 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3862 op.lookup_flags |= LOOKUP_CACHED;
3863 op.open_flag |= O_NONBLOCK;
3866 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3870 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3871 /* only retry if RESOLVE_CACHED wasn't already set by application */
3872 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3873 file == ERR_PTR(-EAGAIN)) {
3875 * We could hang on to this 'fd', but seems like marginal
3876 * gain for something that is now known to be a slower path.
3877 * So just put it, and we'll get a new one when we retry.
3885 ret = PTR_ERR(file);
3887 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3888 file->f_flags &= ~O_NONBLOCK;
3889 fsnotify_open(file);
3890 fd_install(ret, file);
3893 putname(req->open.filename);
3894 req->flags &= ~REQ_F_NEED_CLEANUP;
3896 req_set_fail_links(req);
3897 io_req_complete(req, ret);
3901 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3903 return io_openat2(req, issue_flags);
3906 static int io_remove_buffers_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_provide_buf *p = &req->pbuf;
3912 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3915 tmp = READ_ONCE(sqe->fd);
3916 if (!tmp || tmp > USHRT_MAX)
3919 memset(p, 0, sizeof(*p));
3921 p->bgid = READ_ONCE(sqe->buf_group);
3925 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3926 int bgid, unsigned nbufs)
3930 /* shouldn't happen */
3934 /* the head kbuf is the list itself */
3935 while (!list_empty(&buf->list)) {
3936 struct io_buffer *nxt;
3938 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3939 list_del(&nxt->list);
3946 xa_erase(&ctx->io_buffers, bgid);
3951 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3953 struct io_provide_buf *p = &req->pbuf;
3954 struct io_ring_ctx *ctx = req->ctx;
3955 struct io_buffer *head;
3957 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3959 io_ring_submit_lock(ctx, !force_nonblock);
3961 lockdep_assert_held(&ctx->uring_lock);
3964 head = xa_load(&ctx->io_buffers, p->bgid);
3966 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3968 req_set_fail_links(req);
3970 /* complete before unlock, IOPOLL may need the lock */
3971 __io_req_complete(req, issue_flags, ret, 0);
3972 io_ring_submit_unlock(ctx, !force_nonblock);
3976 static int io_provide_buffers_prep(struct io_kiocb *req,
3977 const struct io_uring_sqe *sqe)
3980 struct io_provide_buf *p = &req->pbuf;
3983 if (sqe->ioprio || sqe->rw_flags)
3986 tmp = READ_ONCE(sqe->fd);
3987 if (!tmp || tmp > USHRT_MAX)
3990 p->addr = READ_ONCE(sqe->addr);
3991 p->len = READ_ONCE(sqe->len);
3993 size = (unsigned long)p->len * p->nbufs;
3994 if (!access_ok(u64_to_user_ptr(p->addr), size))
3997 p->bgid = READ_ONCE(sqe->buf_group);
3998 tmp = READ_ONCE(sqe->off);
3999 if (tmp > USHRT_MAX)
4005 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4007 struct io_buffer *buf;
4008 u64 addr = pbuf->addr;
4009 int i, bid = pbuf->bid;
4011 for (i = 0; i < pbuf->nbufs; i++) {
4012 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4017 buf->len = pbuf->len;
4022 INIT_LIST_HEAD(&buf->list);
4025 list_add_tail(&buf->list, &(*head)->list);
4029 return i ? i : -ENOMEM;
4032 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4034 struct io_provide_buf *p = &req->pbuf;
4035 struct io_ring_ctx *ctx = req->ctx;
4036 struct io_buffer *head, *list;
4038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4040 io_ring_submit_lock(ctx, !force_nonblock);
4042 lockdep_assert_held(&ctx->uring_lock);
4044 list = head = xa_load(&ctx->io_buffers, p->bgid);
4046 ret = io_add_buffers(p, &head);
4047 if (ret >= 0 && !list) {
4048 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4050 __io_remove_buffers(ctx, head, p->bgid, -1U);
4053 req_set_fail_links(req);
4054 /* complete before unlock, IOPOLL may need the lock */
4055 __io_req_complete(req, issue_flags, ret, 0);
4056 io_ring_submit_unlock(ctx, !force_nonblock);
4060 static int io_epoll_ctl_prep(struct io_kiocb *req,
4061 const struct io_uring_sqe *sqe)
4063 #if defined(CONFIG_EPOLL)
4064 if (sqe->ioprio || sqe->buf_index)
4066 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4069 req->epoll.epfd = READ_ONCE(sqe->fd);
4070 req->epoll.op = READ_ONCE(sqe->len);
4071 req->epoll.fd = READ_ONCE(sqe->off);
4073 if (ep_op_has_event(req->epoll.op)) {
4074 struct epoll_event __user *ev;
4076 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4077 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4087 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4089 #if defined(CONFIG_EPOLL)
4090 struct io_epoll *ie = &req->epoll;
4092 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4094 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4095 if (force_nonblock && ret == -EAGAIN)
4099 req_set_fail_links(req);
4100 __io_req_complete(req, issue_flags, ret, 0);
4107 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4109 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4110 if (sqe->ioprio || sqe->buf_index || sqe->off)
4112 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4115 req->madvise.addr = READ_ONCE(sqe->addr);
4116 req->madvise.len = READ_ONCE(sqe->len);
4117 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4124 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4126 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4127 struct io_madvise *ma = &req->madvise;
4130 if (issue_flags & IO_URING_F_NONBLOCK)
4133 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4135 req_set_fail_links(req);
4136 io_req_complete(req, ret);
4143 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4145 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4147 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4150 req->fadvise.offset = READ_ONCE(sqe->off);
4151 req->fadvise.len = READ_ONCE(sqe->len);
4152 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4156 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4158 struct io_fadvise *fa = &req->fadvise;
4161 if (issue_flags & IO_URING_F_NONBLOCK) {
4162 switch (fa->advice) {
4163 case POSIX_FADV_NORMAL:
4164 case POSIX_FADV_RANDOM:
4165 case POSIX_FADV_SEQUENTIAL:
4172 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4174 req_set_fail_links(req);
4175 io_req_complete(req, ret);
4179 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4181 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4183 if (sqe->ioprio || sqe->buf_index)
4185 if (req->flags & REQ_F_FIXED_FILE)
4188 req->statx.dfd = READ_ONCE(sqe->fd);
4189 req->statx.mask = READ_ONCE(sqe->len);
4190 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4191 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4192 req->statx.flags = READ_ONCE(sqe->statx_flags);
4197 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4199 struct io_statx *ctx = &req->statx;
4202 if (issue_flags & IO_URING_F_NONBLOCK) {
4203 /* only need file table for an actual valid fd */
4204 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4205 req->flags |= REQ_F_NO_FILE_TABLE;
4209 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4213 req_set_fail_links(req);
4214 io_req_complete(req, ret);
4218 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4220 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4222 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4223 sqe->rw_flags || sqe->buf_index)
4225 if (req->flags & REQ_F_FIXED_FILE)
4228 req->close.fd = READ_ONCE(sqe->fd);
4232 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4234 struct files_struct *files = current->files;
4235 struct io_close *close = &req->close;
4236 struct fdtable *fdt;
4242 spin_lock(&files->file_lock);
4243 fdt = files_fdtable(files);
4244 if (close->fd >= fdt->max_fds) {
4245 spin_unlock(&files->file_lock);
4248 file = fdt->fd[close->fd];
4250 spin_unlock(&files->file_lock);
4254 if (file->f_op == &io_uring_fops) {
4255 spin_unlock(&files->file_lock);
4260 /* if the file has a flush method, be safe and punt to async */
4261 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4262 spin_unlock(&files->file_lock);
4266 ret = __close_fd_get_file(close->fd, &file);
4267 spin_unlock(&files->file_lock);
4274 /* No ->flush() or already async, safely close from here */
4275 ret = filp_close(file, current->files);
4278 req_set_fail_links(req);
4281 __io_req_complete(req, issue_flags, ret, 0);
4285 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4287 struct io_ring_ctx *ctx = req->ctx;
4289 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4291 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4294 req->sync.off = READ_ONCE(sqe->off);
4295 req->sync.len = READ_ONCE(sqe->len);
4296 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4300 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4304 /* sync_file_range always requires a blocking context */
4305 if (issue_flags & IO_URING_F_NONBLOCK)
4308 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4311 req_set_fail_links(req);
4312 io_req_complete(req, ret);
4316 #if defined(CONFIG_NET)
4317 static int io_setup_async_msg(struct io_kiocb *req,
4318 struct io_async_msghdr *kmsg)
4320 struct io_async_msghdr *async_msg = req->async_data;
4324 if (io_alloc_async_data(req)) {
4325 kfree(kmsg->free_iov);
4328 async_msg = req->async_data;
4329 req->flags |= REQ_F_NEED_CLEANUP;
4330 memcpy(async_msg, kmsg, sizeof(*kmsg));
4331 async_msg->msg.msg_name = &async_msg->addr;
4332 /* if were using fast_iov, set it to the new one */
4333 if (!async_msg->free_iov)
4334 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4339 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4340 struct io_async_msghdr *iomsg)
4342 iomsg->msg.msg_name = &iomsg->addr;
4343 iomsg->free_iov = iomsg->fast_iov;
4344 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4345 req->sr_msg.msg_flags, &iomsg->free_iov);
4348 static int io_sendmsg_prep_async(struct io_kiocb *req)
4352 if (!io_op_defs[req->opcode].needs_async_data)
4354 ret = io_sendmsg_copy_hdr(req, req->async_data);
4356 req->flags |= REQ_F_NEED_CLEANUP;
4360 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4362 struct io_sr_msg *sr = &req->sr_msg;
4364 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4367 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4368 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4369 sr->len = READ_ONCE(sqe->len);
4371 #ifdef CONFIG_COMPAT
4372 if (req->ctx->compat)
4373 sr->msg_flags |= MSG_CMSG_COMPAT;
4378 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4380 struct io_async_msghdr iomsg, *kmsg;
4381 struct socket *sock;
4386 sock = sock_from_file(req->file);
4387 if (unlikely(!sock))
4390 kmsg = req->async_data;
4392 ret = io_sendmsg_copy_hdr(req, &iomsg);
4398 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4399 if (flags & MSG_DONTWAIT)
4400 req->flags |= REQ_F_NOWAIT;
4401 else if (issue_flags & IO_URING_F_NONBLOCK)
4402 flags |= MSG_DONTWAIT;
4404 if (flags & MSG_WAITALL)
4405 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4407 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4408 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4409 return io_setup_async_msg(req, kmsg);
4410 if (ret == -ERESTARTSYS)
4413 /* fast path, check for non-NULL to avoid function call */
4415 kfree(kmsg->free_iov);
4416 req->flags &= ~REQ_F_NEED_CLEANUP;
4418 req_set_fail_links(req);
4419 __io_req_complete(req, issue_flags, ret, 0);
4423 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4425 struct io_sr_msg *sr = &req->sr_msg;
4428 struct socket *sock;
4433 sock = sock_from_file(req->file);
4434 if (unlikely(!sock))
4437 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4441 msg.msg_name = NULL;
4442 msg.msg_control = NULL;
4443 msg.msg_controllen = 0;
4444 msg.msg_namelen = 0;
4446 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4447 if (flags & MSG_DONTWAIT)
4448 req->flags |= REQ_F_NOWAIT;
4449 else if (issue_flags & IO_URING_F_NONBLOCK)
4450 flags |= MSG_DONTWAIT;
4452 if (flags & MSG_WAITALL)
4453 min_ret = iov_iter_count(&msg.msg_iter);
4455 msg.msg_flags = flags;
4456 ret = sock_sendmsg(sock, &msg);
4457 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4459 if (ret == -ERESTARTSYS)
4463 req_set_fail_links(req);
4464 __io_req_complete(req, issue_flags, ret, 0);
4468 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4469 struct io_async_msghdr *iomsg)
4471 struct io_sr_msg *sr = &req->sr_msg;
4472 struct iovec __user *uiov;
4476 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4477 &iomsg->uaddr, &uiov, &iov_len);
4481 if (req->flags & REQ_F_BUFFER_SELECT) {
4484 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4486 sr->len = iomsg->fast_iov[0].iov_len;
4487 iomsg->free_iov = NULL;
4489 iomsg->free_iov = iomsg->fast_iov;
4490 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4491 &iomsg->free_iov, &iomsg->msg.msg_iter,
4500 #ifdef CONFIG_COMPAT
4501 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4502 struct io_async_msghdr *iomsg)
4504 struct compat_msghdr __user *msg_compat;
4505 struct io_sr_msg *sr = &req->sr_msg;
4506 struct compat_iovec __user *uiov;
4511 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4512 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4517 uiov = compat_ptr(ptr);
4518 if (req->flags & REQ_F_BUFFER_SELECT) {
4519 compat_ssize_t clen;
4523 if (!access_ok(uiov, sizeof(*uiov)))
4525 if (__get_user(clen, &uiov->iov_len))
4530 iomsg->free_iov = NULL;
4532 iomsg->free_iov = iomsg->fast_iov;
4533 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4534 UIO_FASTIOV, &iomsg->free_iov,
4535 &iomsg->msg.msg_iter, true);
4544 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4545 struct io_async_msghdr *iomsg)
4547 iomsg->msg.msg_name = &iomsg->addr;
4549 #ifdef CONFIG_COMPAT
4550 if (req->ctx->compat)
4551 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4554 return __io_recvmsg_copy_hdr(req, iomsg);
4557 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4560 struct io_sr_msg *sr = &req->sr_msg;
4561 struct io_buffer *kbuf;
4563 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4568 req->flags |= REQ_F_BUFFER_SELECTED;
4572 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4574 return io_put_kbuf(req, req->sr_msg.kbuf);
4577 static int io_recvmsg_prep_async(struct io_kiocb *req)
4581 if (!io_op_defs[req->opcode].needs_async_data)
4583 ret = io_recvmsg_copy_hdr(req, req->async_data);
4585 req->flags |= REQ_F_NEED_CLEANUP;
4589 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4591 struct io_sr_msg *sr = &req->sr_msg;
4593 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4596 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4597 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4598 sr->len = READ_ONCE(sqe->len);
4599 sr->bgid = READ_ONCE(sqe->buf_group);
4601 #ifdef CONFIG_COMPAT
4602 if (req->ctx->compat)
4603 sr->msg_flags |= MSG_CMSG_COMPAT;
4608 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4610 struct io_async_msghdr iomsg, *kmsg;
4611 struct socket *sock;
4612 struct io_buffer *kbuf;
4615 int ret, cflags = 0;
4616 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4618 sock = sock_from_file(req->file);
4619 if (unlikely(!sock))
4622 kmsg = req->async_data;
4624 ret = io_recvmsg_copy_hdr(req, &iomsg);
4630 if (req->flags & REQ_F_BUFFER_SELECT) {
4631 kbuf = io_recv_buffer_select(req, !force_nonblock);
4633 return PTR_ERR(kbuf);
4634 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4635 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4636 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4637 1, req->sr_msg.len);
4640 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4641 if (flags & MSG_DONTWAIT)
4642 req->flags |= REQ_F_NOWAIT;
4643 else if (force_nonblock)
4644 flags |= MSG_DONTWAIT;
4646 if (flags & MSG_WAITALL)
4647 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4649 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4650 kmsg->uaddr, flags);
4651 if (force_nonblock && ret == -EAGAIN)
4652 return io_setup_async_msg(req, kmsg);
4653 if (ret == -ERESTARTSYS)
4656 if (req->flags & REQ_F_BUFFER_SELECTED)
4657 cflags = io_put_recv_kbuf(req);
4658 /* fast path, check for non-NULL to avoid function call */
4660 kfree(kmsg->free_iov);
4661 req->flags &= ~REQ_F_NEED_CLEANUP;
4662 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4663 req_set_fail_links(req);
4664 __io_req_complete(req, issue_flags, ret, cflags);
4668 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4670 struct io_buffer *kbuf;
4671 struct io_sr_msg *sr = &req->sr_msg;
4673 void __user *buf = sr->buf;
4674 struct socket *sock;
4678 int ret, cflags = 0;
4679 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4681 sock = sock_from_file(req->file);
4682 if (unlikely(!sock))
4685 if (req->flags & REQ_F_BUFFER_SELECT) {
4686 kbuf = io_recv_buffer_select(req, !force_nonblock);
4688 return PTR_ERR(kbuf);
4689 buf = u64_to_user_ptr(kbuf->addr);
4692 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4696 msg.msg_name = NULL;
4697 msg.msg_control = NULL;
4698 msg.msg_controllen = 0;
4699 msg.msg_namelen = 0;
4700 msg.msg_iocb = NULL;
4703 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4704 if (flags & MSG_DONTWAIT)
4705 req->flags |= REQ_F_NOWAIT;
4706 else if (force_nonblock)
4707 flags |= MSG_DONTWAIT;
4709 if (flags & MSG_WAITALL)
4710 min_ret = iov_iter_count(&msg.msg_iter);
4712 ret = sock_recvmsg(sock, &msg, flags);
4713 if (force_nonblock && ret == -EAGAIN)
4715 if (ret == -ERESTARTSYS)
4718 if (req->flags & REQ_F_BUFFER_SELECTED)
4719 cflags = io_put_recv_kbuf(req);
4720 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4721 req_set_fail_links(req);
4722 __io_req_complete(req, issue_flags, ret, cflags);
4726 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4728 struct io_accept *accept = &req->accept;
4730 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4732 if (sqe->ioprio || sqe->len || sqe->buf_index)
4735 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4736 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4737 accept->flags = READ_ONCE(sqe->accept_flags);
4738 accept->nofile = rlimit(RLIMIT_NOFILE);
4742 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4744 struct io_accept *accept = &req->accept;
4745 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4746 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4749 if (req->file->f_flags & O_NONBLOCK)
4750 req->flags |= REQ_F_NOWAIT;
4752 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4753 accept->addr_len, accept->flags,
4755 if (ret == -EAGAIN && force_nonblock)
4758 if (ret == -ERESTARTSYS)
4760 req_set_fail_links(req);
4762 __io_req_complete(req, issue_flags, ret, 0);
4766 static int io_connect_prep_async(struct io_kiocb *req)
4768 struct io_async_connect *io = req->async_data;
4769 struct io_connect *conn = &req->connect;
4771 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4774 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4776 struct io_connect *conn = &req->connect;
4778 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4780 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4783 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4784 conn->addr_len = READ_ONCE(sqe->addr2);
4788 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4790 struct io_async_connect __io, *io;
4791 unsigned file_flags;
4793 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4795 if (req->async_data) {
4796 io = req->async_data;
4798 ret = move_addr_to_kernel(req->connect.addr,
4799 req->connect.addr_len,
4806 file_flags = force_nonblock ? O_NONBLOCK : 0;
4808 ret = __sys_connect_file(req->file, &io->address,
4809 req->connect.addr_len, file_flags);
4810 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4811 if (req->async_data)
4813 if (io_alloc_async_data(req)) {
4817 memcpy(req->async_data, &__io, sizeof(__io));
4820 if (ret == -ERESTARTSYS)
4824 req_set_fail_links(req);
4825 __io_req_complete(req, issue_flags, ret, 0);
4828 #else /* !CONFIG_NET */
4829 #define IO_NETOP_FN(op) \
4830 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4832 return -EOPNOTSUPP; \
4835 #define IO_NETOP_PREP(op) \
4837 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4839 return -EOPNOTSUPP; \
4842 #define IO_NETOP_PREP_ASYNC(op) \
4844 static int io_##op##_prep_async(struct io_kiocb *req) \
4846 return -EOPNOTSUPP; \
4849 IO_NETOP_PREP_ASYNC(sendmsg);
4850 IO_NETOP_PREP_ASYNC(recvmsg);
4851 IO_NETOP_PREP_ASYNC(connect);
4852 IO_NETOP_PREP(accept);
4855 #endif /* CONFIG_NET */
4857 struct io_poll_table {
4858 struct poll_table_struct pt;
4859 struct io_kiocb *req;
4863 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4864 __poll_t mask, task_work_func_t func)
4868 /* for instances that support it check for an event match first: */
4869 if (mask && !(mask & poll->events))
4872 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4874 list_del_init(&poll->wait.entry);
4877 req->task_work.func = func;
4878 percpu_ref_get(&req->ctx->refs);
4881 * If this fails, then the task is exiting. When a task exits, the
4882 * work gets canceled, so just cancel this request as well instead
4883 * of executing it. We can't safely execute it anyway, as we may not
4884 * have the needed state needed for it anyway.
4886 ret = io_req_task_work_add(req);
4887 if (unlikely(ret)) {
4888 WRITE_ONCE(poll->canceled, true);
4889 io_req_task_work_add_fallback(req, func);
4894 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4895 __acquires(&req->ctx->completion_lock)
4897 struct io_ring_ctx *ctx = req->ctx;
4899 if (!req->result && !READ_ONCE(poll->canceled)) {
4900 struct poll_table_struct pt = { ._key = poll->events };
4902 req->result = vfs_poll(req->file, &pt) & poll->events;
4905 spin_lock_irq(&ctx->completion_lock);
4906 if (!req->result && !READ_ONCE(poll->canceled)) {
4907 add_wait_queue(poll->head, &poll->wait);
4914 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4916 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4917 if (req->opcode == IORING_OP_POLL_ADD)
4918 return req->async_data;
4919 return req->apoll->double_poll;
4922 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4924 if (req->opcode == IORING_OP_POLL_ADD)
4926 return &req->apoll->poll;
4929 static void io_poll_remove_double(struct io_kiocb *req)
4931 struct io_poll_iocb *poll = io_poll_get_double(req);
4933 lockdep_assert_held(&req->ctx->completion_lock);
4935 if (poll && poll->head) {
4936 struct wait_queue_head *head = poll->head;
4938 spin_lock(&head->lock);
4939 list_del_init(&poll->wait.entry);
4940 if (poll->wait.private)
4941 refcount_dec(&req->refs);
4943 spin_unlock(&head->lock);
4947 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4949 struct io_ring_ctx *ctx = req->ctx;
4951 io_poll_remove_double(req);
4952 req->poll.done = true;
4953 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4954 io_commit_cqring(ctx);
4957 static void io_poll_task_func(struct callback_head *cb)
4959 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4960 struct io_ring_ctx *ctx = req->ctx;
4961 struct io_kiocb *nxt;
4963 if (io_poll_rewait(req, &req->poll)) {
4964 spin_unlock_irq(&ctx->completion_lock);
4966 hash_del(&req->hash_node);
4967 io_poll_complete(req, req->result, 0);
4968 spin_unlock_irq(&ctx->completion_lock);
4970 nxt = io_put_req_find_next(req);
4971 io_cqring_ev_posted(ctx);
4973 __io_req_task_submit(nxt);
4976 percpu_ref_put(&ctx->refs);
4979 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4980 int sync, void *key)
4982 struct io_kiocb *req = wait->private;
4983 struct io_poll_iocb *poll = io_poll_get_single(req);
4984 __poll_t mask = key_to_poll(key);
4986 /* for instances that support it check for an event match first: */
4987 if (mask && !(mask & poll->events))
4990 list_del_init(&wait->entry);
4992 if (poll && poll->head) {
4995 spin_lock(&poll->head->lock);
4996 done = list_empty(&poll->wait.entry);
4998 list_del_init(&poll->wait.entry);
4999 /* make sure double remove sees this as being gone */
5000 wait->private = NULL;
5001 spin_unlock(&poll->head->lock);
5003 /* use wait func handler, so it matches the rq type */
5004 poll->wait.func(&poll->wait, mode, sync, key);
5007 refcount_dec(&req->refs);
5011 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5012 wait_queue_func_t wake_func)
5016 poll->canceled = false;
5017 poll->events = events;
5018 INIT_LIST_HEAD(&poll->wait.entry);
5019 init_waitqueue_func_entry(&poll->wait, wake_func);
5022 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5023 struct wait_queue_head *head,
5024 struct io_poll_iocb **poll_ptr)
5026 struct io_kiocb *req = pt->req;
5029 * If poll->head is already set, it's because the file being polled
5030 * uses multiple waitqueues for poll handling (eg one for read, one
5031 * for write). Setup a separate io_poll_iocb if this happens.
5033 if (unlikely(poll->head)) {
5034 struct io_poll_iocb *poll_one = poll;
5036 /* already have a 2nd entry, fail a third attempt */
5038 pt->error = -EINVAL;
5041 /* double add on the same waitqueue head, ignore */
5042 if (poll->head == head)
5044 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5046 pt->error = -ENOMEM;
5049 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5050 refcount_inc(&req->refs);
5051 poll->wait.private = req;
5058 if (poll->events & EPOLLEXCLUSIVE)
5059 add_wait_queue_exclusive(head, &poll->wait);
5061 add_wait_queue(head, &poll->wait);
5064 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5065 struct poll_table_struct *p)
5067 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5068 struct async_poll *apoll = pt->req->apoll;
5070 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5073 static void io_async_task_func(struct callback_head *cb)
5075 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5076 struct async_poll *apoll = req->apoll;
5077 struct io_ring_ctx *ctx = req->ctx;
5079 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5081 if (io_poll_rewait(req, &apoll->poll)) {
5082 spin_unlock_irq(&ctx->completion_lock);
5083 percpu_ref_put(&ctx->refs);
5087 /* If req is still hashed, it cannot have been canceled. Don't check. */
5088 if (hash_hashed(&req->hash_node))
5089 hash_del(&req->hash_node);
5091 io_poll_remove_double(req);
5092 spin_unlock_irq(&ctx->completion_lock);
5094 if (!READ_ONCE(apoll->poll.canceled))
5095 __io_req_task_submit(req);
5097 __io_req_task_cancel(req, -ECANCELED);
5099 percpu_ref_put(&ctx->refs);
5100 kfree(apoll->double_poll);
5104 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5107 struct io_kiocb *req = wait->private;
5108 struct io_poll_iocb *poll = &req->apoll->poll;
5110 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5113 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5116 static void io_poll_req_insert(struct io_kiocb *req)
5118 struct io_ring_ctx *ctx = req->ctx;
5119 struct hlist_head *list;
5121 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5122 hlist_add_head(&req->hash_node, list);
5125 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5126 struct io_poll_iocb *poll,
5127 struct io_poll_table *ipt, __poll_t mask,
5128 wait_queue_func_t wake_func)
5129 __acquires(&ctx->completion_lock)
5131 struct io_ring_ctx *ctx = req->ctx;
5132 bool cancel = false;
5134 INIT_HLIST_NODE(&req->hash_node);
5135 io_init_poll_iocb(poll, mask, wake_func);
5136 poll->file = req->file;
5137 poll->wait.private = req;
5139 ipt->pt._key = mask;
5141 ipt->error = -EINVAL;
5143 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5145 spin_lock_irq(&ctx->completion_lock);
5146 if (likely(poll->head)) {
5147 spin_lock(&poll->head->lock);
5148 if (unlikely(list_empty(&poll->wait.entry))) {
5154 if (mask || ipt->error)
5155 list_del_init(&poll->wait.entry);
5157 WRITE_ONCE(poll->canceled, true);
5158 else if (!poll->done) /* actually waiting for an event */
5159 io_poll_req_insert(req);
5160 spin_unlock(&poll->head->lock);
5166 static bool io_arm_poll_handler(struct io_kiocb *req)
5168 const struct io_op_def *def = &io_op_defs[req->opcode];
5169 struct io_ring_ctx *ctx = req->ctx;
5170 struct async_poll *apoll;
5171 struct io_poll_table ipt;
5175 if (!req->file || !file_can_poll(req->file))
5177 if (req->flags & REQ_F_POLLED)
5181 else if (def->pollout)
5185 /* if we can't nonblock try, then no point in arming a poll handler */
5186 if (!io_file_supports_async(req->file, rw))
5189 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5190 if (unlikely(!apoll))
5192 apoll->double_poll = NULL;
5194 req->flags |= REQ_F_POLLED;
5199 mask |= POLLIN | POLLRDNORM;
5201 mask |= POLLOUT | POLLWRNORM;
5203 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5204 if ((req->opcode == IORING_OP_RECVMSG) &&
5205 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5208 mask |= POLLERR | POLLPRI;
5210 ipt.pt._qproc = io_async_queue_proc;
5212 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5214 if (ret || ipt.error) {
5215 io_poll_remove_double(req);
5216 spin_unlock_irq(&ctx->completion_lock);
5217 kfree(apoll->double_poll);
5221 spin_unlock_irq(&ctx->completion_lock);
5222 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5223 apoll->poll.events);
5227 static bool __io_poll_remove_one(struct io_kiocb *req,
5228 struct io_poll_iocb *poll)
5230 bool do_complete = false;
5232 spin_lock(&poll->head->lock);
5233 WRITE_ONCE(poll->canceled, true);
5234 if (!list_empty(&poll->wait.entry)) {
5235 list_del_init(&poll->wait.entry);
5238 spin_unlock(&poll->head->lock);
5239 hash_del(&req->hash_node);
5243 static bool io_poll_remove_one(struct io_kiocb *req)
5247 io_poll_remove_double(req);
5249 if (req->opcode == IORING_OP_POLL_ADD) {
5250 do_complete = __io_poll_remove_one(req, &req->poll);
5252 struct async_poll *apoll = req->apoll;
5254 /* non-poll requests have submit ref still */
5255 do_complete = __io_poll_remove_one(req, &apoll->poll);
5258 kfree(apoll->double_poll);
5264 io_cqring_fill_event(req, -ECANCELED);
5265 io_commit_cqring(req->ctx);
5266 req_set_fail_links(req);
5267 io_put_req_deferred(req, 1);
5274 * Returns true if we found and killed one or more poll requests
5276 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5277 struct files_struct *files)
5279 struct hlist_node *tmp;
5280 struct io_kiocb *req;
5283 spin_lock_irq(&ctx->completion_lock);
5284 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5285 struct hlist_head *list;
5287 list = &ctx->cancel_hash[i];
5288 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5289 if (io_match_task(req, tsk, files))
5290 posted += io_poll_remove_one(req);
5293 spin_unlock_irq(&ctx->completion_lock);
5296 io_cqring_ev_posted(ctx);
5301 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5303 struct hlist_head *list;
5304 struct io_kiocb *req;
5306 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5307 hlist_for_each_entry(req, list, hash_node) {
5308 if (sqe_addr != req->user_data)
5310 if (io_poll_remove_one(req))
5318 static int io_poll_remove_prep(struct io_kiocb *req,
5319 const struct io_uring_sqe *sqe)
5321 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5323 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5327 req->poll_remove.addr = READ_ONCE(sqe->addr);
5332 * Find a running poll command that matches one specified in sqe->addr,
5333 * and remove it if found.
5335 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5337 struct io_ring_ctx *ctx = req->ctx;
5340 spin_lock_irq(&ctx->completion_lock);
5341 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5342 spin_unlock_irq(&ctx->completion_lock);
5345 req_set_fail_links(req);
5346 io_req_complete(req, ret);
5350 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5353 struct io_kiocb *req = wait->private;
5354 struct io_poll_iocb *poll = &req->poll;
5356 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5359 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5360 struct poll_table_struct *p)
5362 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5364 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5367 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5369 struct io_poll_iocb *poll = &req->poll;
5372 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5374 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5377 events = READ_ONCE(sqe->poll32_events);
5379 events = swahw32(events);
5381 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5382 (events & EPOLLEXCLUSIVE);
5386 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5388 struct io_poll_iocb *poll = &req->poll;
5389 struct io_ring_ctx *ctx = req->ctx;
5390 struct io_poll_table ipt;
5393 ipt.pt._qproc = io_poll_queue_proc;
5395 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5398 if (mask) { /* no async, we'd stolen it */
5400 io_poll_complete(req, mask, 0);
5402 spin_unlock_irq(&ctx->completion_lock);
5405 io_cqring_ev_posted(ctx);
5411 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5413 struct io_timeout_data *data = container_of(timer,
5414 struct io_timeout_data, timer);
5415 struct io_kiocb *req = data->req;
5416 struct io_ring_ctx *ctx = req->ctx;
5417 unsigned long flags;
5419 spin_lock_irqsave(&ctx->completion_lock, flags);
5420 list_del_init(&req->timeout.list);
5421 atomic_set(&req->ctx->cq_timeouts,
5422 atomic_read(&req->ctx->cq_timeouts) + 1);
5424 io_cqring_fill_event(req, -ETIME);
5425 io_commit_cqring(ctx);
5426 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5428 io_cqring_ev_posted(ctx);
5429 req_set_fail_links(req);
5431 return HRTIMER_NORESTART;
5434 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5437 struct io_timeout_data *io;
5438 struct io_kiocb *req;
5441 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5442 if (user_data == req->user_data) {
5449 return ERR_PTR(ret);
5451 io = req->async_data;
5452 ret = hrtimer_try_to_cancel(&io->timer);
5454 return ERR_PTR(-EALREADY);
5455 list_del_init(&req->timeout.list);
5459 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5461 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5464 return PTR_ERR(req);
5466 req_set_fail_links(req);
5467 io_cqring_fill_event(req, -ECANCELED);
5468 io_put_req_deferred(req, 1);
5472 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5473 struct timespec64 *ts, enum hrtimer_mode mode)
5475 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5476 struct io_timeout_data *data;
5479 return PTR_ERR(req);
5481 req->timeout.off = 0; /* noseq */
5482 data = req->async_data;
5483 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5484 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5485 data->timer.function = io_timeout_fn;
5486 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5490 static int io_timeout_remove_prep(struct io_kiocb *req,
5491 const struct io_uring_sqe *sqe)
5493 struct io_timeout_rem *tr = &req->timeout_rem;
5495 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5497 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5499 if (sqe->ioprio || sqe->buf_index || sqe->len)
5502 tr->addr = READ_ONCE(sqe->addr);
5503 tr->flags = READ_ONCE(sqe->timeout_flags);
5504 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5505 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5507 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5509 } else if (tr->flags) {
5510 /* timeout removal doesn't support flags */
5517 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5519 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5524 * Remove or update an existing timeout command
5526 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5528 struct io_timeout_rem *tr = &req->timeout_rem;
5529 struct io_ring_ctx *ctx = req->ctx;
5532 spin_lock_irq(&ctx->completion_lock);
5533 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5534 ret = io_timeout_cancel(ctx, tr->addr);
5536 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5537 io_translate_timeout_mode(tr->flags));
5539 io_cqring_fill_event(req, ret);
5540 io_commit_cqring(ctx);
5541 spin_unlock_irq(&ctx->completion_lock);
5542 io_cqring_ev_posted(ctx);
5544 req_set_fail_links(req);
5549 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5550 bool is_timeout_link)
5552 struct io_timeout_data *data;
5554 u32 off = READ_ONCE(sqe->off);
5556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5558 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5560 if (off && is_timeout_link)
5562 flags = READ_ONCE(sqe->timeout_flags);
5563 if (flags & ~IORING_TIMEOUT_ABS)
5566 req->timeout.off = off;
5568 if (!req->async_data && io_alloc_async_data(req))
5571 data = req->async_data;
5574 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5577 data->mode = io_translate_timeout_mode(flags);
5578 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5579 if (is_timeout_link)
5580 io_req_track_inflight(req);
5584 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5586 struct io_ring_ctx *ctx = req->ctx;
5587 struct io_timeout_data *data = req->async_data;
5588 struct list_head *entry;
5589 u32 tail, off = req->timeout.off;
5591 spin_lock_irq(&ctx->completion_lock);
5594 * sqe->off holds how many events that need to occur for this
5595 * timeout event to be satisfied. If it isn't set, then this is
5596 * a pure timeout request, sequence isn't used.
5598 if (io_is_timeout_noseq(req)) {
5599 entry = ctx->timeout_list.prev;
5603 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5604 req->timeout.target_seq = tail + off;
5606 /* Update the last seq here in case io_flush_timeouts() hasn't.
5607 * This is safe because ->completion_lock is held, and submissions
5608 * and completions are never mixed in the same ->completion_lock section.
5610 ctx->cq_last_tm_flush = tail;
5613 * Insertion sort, ensuring the first entry in the list is always
5614 * the one we need first.
5616 list_for_each_prev(entry, &ctx->timeout_list) {
5617 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5620 if (io_is_timeout_noseq(nxt))
5622 /* nxt.seq is behind @tail, otherwise would've been completed */
5623 if (off >= nxt->timeout.target_seq - tail)
5627 list_add(&req->timeout.list, entry);
5628 data->timer.function = io_timeout_fn;
5629 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5630 spin_unlock_irq(&ctx->completion_lock);
5634 struct io_cancel_data {
5635 struct io_ring_ctx *ctx;
5639 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5641 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5642 struct io_cancel_data *cd = data;
5644 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5647 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5648 struct io_ring_ctx *ctx)
5650 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5651 enum io_wq_cancel cancel_ret;
5654 if (!tctx || !tctx->io_wq)
5657 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5658 switch (cancel_ret) {
5659 case IO_WQ_CANCEL_OK:
5662 case IO_WQ_CANCEL_RUNNING:
5665 case IO_WQ_CANCEL_NOTFOUND:
5673 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5674 struct io_kiocb *req, __u64 sqe_addr,
5677 unsigned long flags;
5680 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5681 if (ret != -ENOENT) {
5682 spin_lock_irqsave(&ctx->completion_lock, flags);
5686 spin_lock_irqsave(&ctx->completion_lock, flags);
5687 ret = io_timeout_cancel(ctx, sqe_addr);
5690 ret = io_poll_cancel(ctx, sqe_addr);
5694 io_cqring_fill_event(req, ret);
5695 io_commit_cqring(ctx);
5696 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5697 io_cqring_ev_posted(ctx);
5700 req_set_fail_links(req);
5704 static int io_async_cancel_prep(struct io_kiocb *req,
5705 const struct io_uring_sqe *sqe)
5707 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5709 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5711 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5714 req->cancel.addr = READ_ONCE(sqe->addr);
5718 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5720 struct io_ring_ctx *ctx = req->ctx;
5721 u64 sqe_addr = req->cancel.addr;
5722 struct io_tctx_node *node;
5725 /* tasks should wait for their io-wq threads, so safe w/o sync */
5726 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5727 spin_lock_irq(&ctx->completion_lock);
5730 ret = io_timeout_cancel(ctx, sqe_addr);
5733 ret = io_poll_cancel(ctx, sqe_addr);
5736 spin_unlock_irq(&ctx->completion_lock);
5738 /* slow path, try all io-wq's */
5739 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5741 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5742 struct io_uring_task *tctx = node->task->io_uring;
5744 if (!tctx || !tctx->io_wq)
5746 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5750 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5752 spin_lock_irq(&ctx->completion_lock);
5754 io_cqring_fill_event(req, ret);
5755 io_commit_cqring(ctx);
5756 spin_unlock_irq(&ctx->completion_lock);
5757 io_cqring_ev_posted(ctx);
5760 req_set_fail_links(req);
5765 static int io_rsrc_update_prep(struct io_kiocb *req,
5766 const struct io_uring_sqe *sqe)
5768 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5770 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5772 if (sqe->ioprio || sqe->rw_flags)
5775 req->rsrc_update.offset = READ_ONCE(sqe->off);
5776 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5777 if (!req->rsrc_update.nr_args)
5779 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5783 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5785 struct io_ring_ctx *ctx = req->ctx;
5786 struct io_uring_rsrc_update up;
5789 if (issue_flags & IO_URING_F_NONBLOCK)
5792 up.offset = req->rsrc_update.offset;
5793 up.data = req->rsrc_update.arg;
5795 mutex_lock(&ctx->uring_lock);
5796 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5797 mutex_unlock(&ctx->uring_lock);
5800 req_set_fail_links(req);
5801 __io_req_complete(req, issue_flags, ret, 0);
5805 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5807 switch (req->opcode) {
5810 case IORING_OP_READV:
5811 case IORING_OP_READ_FIXED:
5812 case IORING_OP_READ:
5813 return io_read_prep(req, sqe);
5814 case IORING_OP_WRITEV:
5815 case IORING_OP_WRITE_FIXED:
5816 case IORING_OP_WRITE:
5817 return io_write_prep(req, sqe);
5818 case IORING_OP_POLL_ADD:
5819 return io_poll_add_prep(req, sqe);
5820 case IORING_OP_POLL_REMOVE:
5821 return io_poll_remove_prep(req, sqe);
5822 case IORING_OP_FSYNC:
5823 return io_fsync_prep(req, sqe);
5824 case IORING_OP_SYNC_FILE_RANGE:
5825 return io_sfr_prep(req, sqe);
5826 case IORING_OP_SENDMSG:
5827 case IORING_OP_SEND:
5828 return io_sendmsg_prep(req, sqe);
5829 case IORING_OP_RECVMSG:
5830 case IORING_OP_RECV:
5831 return io_recvmsg_prep(req, sqe);
5832 case IORING_OP_CONNECT:
5833 return io_connect_prep(req, sqe);
5834 case IORING_OP_TIMEOUT:
5835 return io_timeout_prep(req, sqe, false);
5836 case IORING_OP_TIMEOUT_REMOVE:
5837 return io_timeout_remove_prep(req, sqe);
5838 case IORING_OP_ASYNC_CANCEL:
5839 return io_async_cancel_prep(req, sqe);
5840 case IORING_OP_LINK_TIMEOUT:
5841 return io_timeout_prep(req, sqe, true);
5842 case IORING_OP_ACCEPT:
5843 return io_accept_prep(req, sqe);
5844 case IORING_OP_FALLOCATE:
5845 return io_fallocate_prep(req, sqe);
5846 case IORING_OP_OPENAT:
5847 return io_openat_prep(req, sqe);
5848 case IORING_OP_CLOSE:
5849 return io_close_prep(req, sqe);
5850 case IORING_OP_FILES_UPDATE:
5851 return io_rsrc_update_prep(req, sqe);
5852 case IORING_OP_STATX:
5853 return io_statx_prep(req, sqe);
5854 case IORING_OP_FADVISE:
5855 return io_fadvise_prep(req, sqe);
5856 case IORING_OP_MADVISE:
5857 return io_madvise_prep(req, sqe);
5858 case IORING_OP_OPENAT2:
5859 return io_openat2_prep(req, sqe);
5860 case IORING_OP_EPOLL_CTL:
5861 return io_epoll_ctl_prep(req, sqe);
5862 case IORING_OP_SPLICE:
5863 return io_splice_prep(req, sqe);
5864 case IORING_OP_PROVIDE_BUFFERS:
5865 return io_provide_buffers_prep(req, sqe);
5866 case IORING_OP_REMOVE_BUFFERS:
5867 return io_remove_buffers_prep(req, sqe);
5869 return io_tee_prep(req, sqe);
5870 case IORING_OP_SHUTDOWN:
5871 return io_shutdown_prep(req, sqe);
5872 case IORING_OP_RENAMEAT:
5873 return io_renameat_prep(req, sqe);
5874 case IORING_OP_UNLINKAT:
5875 return io_unlinkat_prep(req, sqe);
5878 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5883 static int io_req_prep_async(struct io_kiocb *req)
5885 switch (req->opcode) {
5886 case IORING_OP_READV:
5887 case IORING_OP_READ_FIXED:
5888 case IORING_OP_READ:
5889 return io_rw_prep_async(req, READ);
5890 case IORING_OP_WRITEV:
5891 case IORING_OP_WRITE_FIXED:
5892 case IORING_OP_WRITE:
5893 return io_rw_prep_async(req, WRITE);
5894 case IORING_OP_SENDMSG:
5895 case IORING_OP_SEND:
5896 return io_sendmsg_prep_async(req);
5897 case IORING_OP_RECVMSG:
5898 case IORING_OP_RECV:
5899 return io_recvmsg_prep_async(req);
5900 case IORING_OP_CONNECT:
5901 return io_connect_prep_async(req);
5906 static int io_req_defer_prep(struct io_kiocb *req)
5908 if (!io_op_defs[req->opcode].needs_async_data)
5910 /* some opcodes init it during the inital prep */
5911 if (req->async_data)
5913 if (__io_alloc_async_data(req))
5915 return io_req_prep_async(req);
5918 static u32 io_get_sequence(struct io_kiocb *req)
5920 struct io_kiocb *pos;
5921 struct io_ring_ctx *ctx = req->ctx;
5922 u32 total_submitted, nr_reqs = 0;
5924 io_for_each_link(pos, req)
5927 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5928 return total_submitted - nr_reqs;
5931 static int io_req_defer(struct io_kiocb *req)
5933 struct io_ring_ctx *ctx = req->ctx;
5934 struct io_defer_entry *de;
5938 /* Still need defer if there is pending req in defer list. */
5939 if (likely(list_empty_careful(&ctx->defer_list) &&
5940 !(req->flags & REQ_F_IO_DRAIN)))
5943 seq = io_get_sequence(req);
5944 /* Still a chance to pass the sequence check */
5945 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5948 ret = io_req_defer_prep(req);
5951 io_prep_async_link(req);
5952 de = kmalloc(sizeof(*de), GFP_KERNEL);
5956 spin_lock_irq(&ctx->completion_lock);
5957 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5958 spin_unlock_irq(&ctx->completion_lock);
5960 io_queue_async_work(req);
5961 return -EIOCBQUEUED;
5964 trace_io_uring_defer(ctx, req, req->user_data);
5967 list_add_tail(&de->list, &ctx->defer_list);
5968 spin_unlock_irq(&ctx->completion_lock);
5969 return -EIOCBQUEUED;
5972 static void __io_clean_op(struct io_kiocb *req)
5974 if (req->flags & REQ_F_BUFFER_SELECTED) {
5975 switch (req->opcode) {
5976 case IORING_OP_READV:
5977 case IORING_OP_READ_FIXED:
5978 case IORING_OP_READ:
5979 kfree((void *)(unsigned long)req->rw.addr);
5981 case IORING_OP_RECVMSG:
5982 case IORING_OP_RECV:
5983 kfree(req->sr_msg.kbuf);
5986 req->flags &= ~REQ_F_BUFFER_SELECTED;
5989 if (req->flags & REQ_F_NEED_CLEANUP) {
5990 switch (req->opcode) {
5991 case IORING_OP_READV:
5992 case IORING_OP_READ_FIXED:
5993 case IORING_OP_READ:
5994 case IORING_OP_WRITEV:
5995 case IORING_OP_WRITE_FIXED:
5996 case IORING_OP_WRITE: {
5997 struct io_async_rw *io = req->async_data;
5999 kfree(io->free_iovec);
6002 case IORING_OP_RECVMSG:
6003 case IORING_OP_SENDMSG: {
6004 struct io_async_msghdr *io = req->async_data;
6006 kfree(io->free_iov);
6009 case IORING_OP_SPLICE:
6011 io_put_file(req, req->splice.file_in,
6012 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6014 case IORING_OP_OPENAT:
6015 case IORING_OP_OPENAT2:
6016 if (req->open.filename)
6017 putname(req->open.filename);
6019 case IORING_OP_RENAMEAT:
6020 putname(req->rename.oldpath);
6021 putname(req->rename.newpath);
6023 case IORING_OP_UNLINKAT:
6024 putname(req->unlink.filename);
6027 req->flags &= ~REQ_F_NEED_CLEANUP;
6031 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6033 struct io_ring_ctx *ctx = req->ctx;
6034 const struct cred *creds = NULL;
6037 if (req->work.creds && req->work.creds != current_cred())
6038 creds = override_creds(req->work.creds);
6040 switch (req->opcode) {
6042 ret = io_nop(req, issue_flags);
6044 case IORING_OP_READV:
6045 case IORING_OP_READ_FIXED:
6046 case IORING_OP_READ:
6047 ret = io_read(req, issue_flags);
6049 case IORING_OP_WRITEV:
6050 case IORING_OP_WRITE_FIXED:
6051 case IORING_OP_WRITE:
6052 ret = io_write(req, issue_flags);
6054 case IORING_OP_FSYNC:
6055 ret = io_fsync(req, issue_flags);
6057 case IORING_OP_POLL_ADD:
6058 ret = io_poll_add(req, issue_flags);
6060 case IORING_OP_POLL_REMOVE:
6061 ret = io_poll_remove(req, issue_flags);
6063 case IORING_OP_SYNC_FILE_RANGE:
6064 ret = io_sync_file_range(req, issue_flags);
6066 case IORING_OP_SENDMSG:
6067 ret = io_sendmsg(req, issue_flags);
6069 case IORING_OP_SEND:
6070 ret = io_send(req, issue_flags);
6072 case IORING_OP_RECVMSG:
6073 ret = io_recvmsg(req, issue_flags);
6075 case IORING_OP_RECV:
6076 ret = io_recv(req, issue_flags);
6078 case IORING_OP_TIMEOUT:
6079 ret = io_timeout(req, issue_flags);
6081 case IORING_OP_TIMEOUT_REMOVE:
6082 ret = io_timeout_remove(req, issue_flags);
6084 case IORING_OP_ACCEPT:
6085 ret = io_accept(req, issue_flags);
6087 case IORING_OP_CONNECT:
6088 ret = io_connect(req, issue_flags);
6090 case IORING_OP_ASYNC_CANCEL:
6091 ret = io_async_cancel(req, issue_flags);
6093 case IORING_OP_FALLOCATE:
6094 ret = io_fallocate(req, issue_flags);
6096 case IORING_OP_OPENAT:
6097 ret = io_openat(req, issue_flags);
6099 case IORING_OP_CLOSE:
6100 ret = io_close(req, issue_flags);
6102 case IORING_OP_FILES_UPDATE:
6103 ret = io_files_update(req, issue_flags);
6105 case IORING_OP_STATX:
6106 ret = io_statx(req, issue_flags);
6108 case IORING_OP_FADVISE:
6109 ret = io_fadvise(req, issue_flags);
6111 case IORING_OP_MADVISE:
6112 ret = io_madvise(req, issue_flags);
6114 case IORING_OP_OPENAT2:
6115 ret = io_openat2(req, issue_flags);
6117 case IORING_OP_EPOLL_CTL:
6118 ret = io_epoll_ctl(req, issue_flags);
6120 case IORING_OP_SPLICE:
6121 ret = io_splice(req, issue_flags);
6123 case IORING_OP_PROVIDE_BUFFERS:
6124 ret = io_provide_buffers(req, issue_flags);
6126 case IORING_OP_REMOVE_BUFFERS:
6127 ret = io_remove_buffers(req, issue_flags);
6130 ret = io_tee(req, issue_flags);
6132 case IORING_OP_SHUTDOWN:
6133 ret = io_shutdown(req, issue_flags);
6135 case IORING_OP_RENAMEAT:
6136 ret = io_renameat(req, issue_flags);
6138 case IORING_OP_UNLINKAT:
6139 ret = io_unlinkat(req, issue_flags);
6147 revert_creds(creds);
6152 /* If the op doesn't have a file, we're not polling for it */
6153 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6154 const bool in_async = io_wq_current_is_worker();
6156 /* workqueue context doesn't hold uring_lock, grab it now */
6158 mutex_lock(&ctx->uring_lock);
6160 io_iopoll_req_issued(req, in_async);
6163 mutex_unlock(&ctx->uring_lock);
6169 static void io_wq_submit_work(struct io_wq_work *work)
6171 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6172 struct io_kiocb *timeout;
6175 timeout = io_prep_linked_timeout(req);
6177 io_queue_linked_timeout(timeout);
6179 if (work->flags & IO_WQ_WORK_CANCEL)
6184 ret = io_issue_sqe(req, 0);
6186 * We can get EAGAIN for polled IO even though we're
6187 * forcing a sync submission from here, since we can't
6188 * wait for request slots on the block side.
6196 /* avoid locking problems by failing it from a clean context */
6198 /* io-wq is going to take one down */
6199 refcount_inc(&req->refs);
6200 io_req_task_queue_fail(req, ret);
6204 static inline struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
6207 struct fixed_rsrc_table *table;
6209 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6210 return &table->files[i & IORING_FILE_TABLE_MASK];
6213 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6216 return *io_fixed_file_slot(ctx->file_data, index);
6219 static struct file *io_file_get(struct io_submit_state *state,
6220 struct io_kiocb *req, int fd, bool fixed)
6222 struct io_ring_ctx *ctx = req->ctx;
6226 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6228 fd = array_index_nospec(fd, ctx->nr_user_files);
6229 file = io_file_from_index(ctx, fd);
6230 io_set_resource_node(req);
6232 trace_io_uring_file_get(ctx, fd);
6233 file = __io_file_get(state, fd);
6236 if (file && unlikely(file->f_op == &io_uring_fops))
6237 io_req_track_inflight(req);
6241 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6243 struct io_timeout_data *data = container_of(timer,
6244 struct io_timeout_data, timer);
6245 struct io_kiocb *prev, *req = data->req;
6246 struct io_ring_ctx *ctx = req->ctx;
6247 unsigned long flags;
6249 spin_lock_irqsave(&ctx->completion_lock, flags);
6250 prev = req->timeout.head;
6251 req->timeout.head = NULL;
6254 * We don't expect the list to be empty, that will only happen if we
6255 * race with the completion of the linked work.
6257 if (prev && refcount_inc_not_zero(&prev->refs))
6258 io_remove_next_linked(prev);
6261 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6264 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6265 io_put_req_deferred(prev, 1);
6267 io_req_complete_post(req, -ETIME, 0);
6268 io_put_req_deferred(req, 1);
6270 return HRTIMER_NORESTART;
6273 static void __io_queue_linked_timeout(struct io_kiocb *req)
6276 * If the back reference is NULL, then our linked request finished
6277 * before we got a chance to setup the timer
6279 if (req->timeout.head) {
6280 struct io_timeout_data *data = req->async_data;
6282 data->timer.function = io_link_timeout_fn;
6283 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6288 static void io_queue_linked_timeout(struct io_kiocb *req)
6290 struct io_ring_ctx *ctx = req->ctx;
6292 spin_lock_irq(&ctx->completion_lock);
6293 __io_queue_linked_timeout(req);
6294 spin_unlock_irq(&ctx->completion_lock);
6296 /* drop submission reference */
6300 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6302 struct io_kiocb *nxt = req->link;
6304 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6305 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6308 nxt->timeout.head = req;
6309 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6310 req->flags |= REQ_F_LINK_TIMEOUT;
6314 static void __io_queue_sqe(struct io_kiocb *req)
6316 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6319 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6322 * We async punt it if the file wasn't marked NOWAIT, or if the file
6323 * doesn't support non-blocking read/write attempts
6325 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6326 if (!io_arm_poll_handler(req)) {
6328 * Queued up for async execution, worker will release
6329 * submit reference when the iocb is actually submitted.
6331 io_queue_async_work(req);
6333 } else if (likely(!ret)) {
6334 /* drop submission reference */
6335 if (req->flags & REQ_F_COMPLETE_INLINE) {
6336 struct io_ring_ctx *ctx = req->ctx;
6337 struct io_comp_state *cs = &ctx->submit_state.comp;
6339 cs->reqs[cs->nr++] = req;
6340 if (cs->nr == ARRAY_SIZE(cs->reqs))
6341 io_submit_flush_completions(cs, ctx);
6346 io_req_complete_failed(req, ret);
6349 io_queue_linked_timeout(linked_timeout);
6352 static void io_queue_sqe(struct io_kiocb *req)
6356 ret = io_req_defer(req);
6358 if (ret != -EIOCBQUEUED) {
6360 io_req_complete_failed(req, ret);
6362 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6363 ret = io_req_defer_prep(req);
6366 io_queue_async_work(req);
6368 __io_queue_sqe(req);
6373 * Check SQE restrictions (opcode and flags).
6375 * Returns 'true' if SQE is allowed, 'false' otherwise.
6377 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6378 struct io_kiocb *req,
6379 unsigned int sqe_flags)
6381 if (!ctx->restricted)
6384 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6387 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6388 ctx->restrictions.sqe_flags_required)
6391 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6392 ctx->restrictions.sqe_flags_required))
6398 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6399 const struct io_uring_sqe *sqe)
6401 struct io_submit_state *state;
6402 unsigned int sqe_flags;
6403 int personality, ret = 0;
6405 req->opcode = READ_ONCE(sqe->opcode);
6406 /* same numerical values with corresponding REQ_F_*, safe to copy */
6407 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6408 req->user_data = READ_ONCE(sqe->user_data);
6409 req->async_data = NULL;
6413 req->fixed_rsrc_refs = NULL;
6414 /* one is dropped after submission, the other at completion */
6415 refcount_set(&req->refs, 2);
6416 req->task = current;
6418 req->work.list.next = NULL;
6419 req->work.creds = NULL;
6420 req->work.flags = 0;
6422 /* enforce forwards compatibility on users */
6423 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6428 if (unlikely(req->opcode >= IORING_OP_LAST))
6431 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6434 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6435 !io_op_defs[req->opcode].buffer_select)
6438 personality = READ_ONCE(sqe->personality);
6440 req->work.creds = xa_load(&ctx->personalities, personality);
6441 if (!req->work.creds)
6443 get_cred(req->work.creds);
6445 state = &ctx->submit_state;
6448 * Plug now if we have more than 1 IO left after this, and the target
6449 * is potentially a read/write to block based storage.
6451 if (!state->plug_started && state->ios_left > 1 &&
6452 io_op_defs[req->opcode].plug) {
6453 blk_start_plug(&state->plug);
6454 state->plug_started = true;
6457 if (io_op_defs[req->opcode].needs_file) {
6458 bool fixed = req->flags & REQ_F_FIXED_FILE;
6460 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6461 if (unlikely(!req->file))
6469 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6470 const struct io_uring_sqe *sqe)
6472 struct io_submit_link *link = &ctx->submit_state.link;
6475 ret = io_init_req(ctx, req, sqe);
6476 if (unlikely(ret)) {
6479 /* fail even hard links since we don't submit */
6480 link->head->flags |= REQ_F_FAIL_LINK;
6481 io_req_complete_failed(link->head, -ECANCELED);
6484 io_req_complete_failed(req, ret);
6487 ret = io_req_prep(req, sqe);
6491 /* don't need @sqe from now on */
6492 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6493 true, ctx->flags & IORING_SETUP_SQPOLL);
6496 * If we already have a head request, queue this one for async
6497 * submittal once the head completes. If we don't have a head but
6498 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6499 * submitted sync once the chain is complete. If none of those
6500 * conditions are true (normal request), then just queue it.
6503 struct io_kiocb *head = link->head;
6506 * Taking sequential execution of a link, draining both sides
6507 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6508 * requests in the link. So, it drains the head and the
6509 * next after the link request. The last one is done via
6510 * drain_next flag to persist the effect across calls.
6512 if (req->flags & REQ_F_IO_DRAIN) {
6513 head->flags |= REQ_F_IO_DRAIN;
6514 ctx->drain_next = 1;
6516 ret = io_req_defer_prep(req);
6519 trace_io_uring_link(ctx, req, head);
6520 link->last->link = req;
6523 /* last request of a link, enqueue the link */
6524 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6529 if (unlikely(ctx->drain_next)) {
6530 req->flags |= REQ_F_IO_DRAIN;
6531 ctx->drain_next = 0;
6533 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6545 * Batched submission is done, ensure local IO is flushed out.
6547 static void io_submit_state_end(struct io_submit_state *state,
6548 struct io_ring_ctx *ctx)
6550 if (state->link.head)
6551 io_queue_sqe(state->link.head);
6553 io_submit_flush_completions(&state->comp, ctx);
6554 if (state->plug_started)
6555 blk_finish_plug(&state->plug);
6556 io_state_file_put(state);
6560 * Start submission side cache.
6562 static void io_submit_state_start(struct io_submit_state *state,
6563 unsigned int max_ios)
6565 state->plug_started = false;
6566 state->ios_left = max_ios;
6567 /* set only head, no need to init link_last in advance */
6568 state->link.head = NULL;
6571 static void io_commit_sqring(struct io_ring_ctx *ctx)
6573 struct io_rings *rings = ctx->rings;
6576 * Ensure any loads from the SQEs are done at this point,
6577 * since once we write the new head, the application could
6578 * write new data to them.
6580 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6584 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6585 * that is mapped by userspace. This means that care needs to be taken to
6586 * ensure that reads are stable, as we cannot rely on userspace always
6587 * being a good citizen. If members of the sqe are validated and then later
6588 * used, it's important that those reads are done through READ_ONCE() to
6589 * prevent a re-load down the line.
6591 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6593 u32 *sq_array = ctx->sq_array;
6597 * The cached sq head (or cq tail) serves two purposes:
6599 * 1) allows us to batch the cost of updating the user visible
6601 * 2) allows the kernel side to track the head on its own, even
6602 * though the application is the one updating it.
6604 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6605 if (likely(head < ctx->sq_entries))
6606 return &ctx->sq_sqes[head];
6608 /* drop invalid entries */
6609 ctx->cached_sq_dropped++;
6610 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6614 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6618 /* if we have a backlog and couldn't flush it all, return BUSY */
6619 if (test_bit(0, &ctx->sq_check_overflow)) {
6620 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6624 /* make sure SQ entry isn't read before tail */
6625 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6627 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6630 percpu_counter_add(¤t->io_uring->inflight, nr);
6631 refcount_add(nr, ¤t->usage);
6632 io_submit_state_start(&ctx->submit_state, nr);
6634 while (submitted < nr) {
6635 const struct io_uring_sqe *sqe;
6636 struct io_kiocb *req;
6638 req = io_alloc_req(ctx);
6639 if (unlikely(!req)) {
6641 submitted = -EAGAIN;
6644 sqe = io_get_sqe(ctx);
6645 if (unlikely(!sqe)) {
6646 kmem_cache_free(req_cachep, req);
6649 /* will complete beyond this point, count as submitted */
6651 if (io_submit_sqe(ctx, req, sqe))
6655 if (unlikely(submitted != nr)) {
6656 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6657 struct io_uring_task *tctx = current->io_uring;
6658 int unused = nr - ref_used;
6660 percpu_ref_put_many(&ctx->refs, unused);
6661 percpu_counter_sub(&tctx->inflight, unused);
6662 put_task_struct_many(current, unused);
6665 io_submit_state_end(&ctx->submit_state, ctx);
6666 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6667 io_commit_sqring(ctx);
6672 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6674 /* Tell userspace we may need a wakeup call */
6675 spin_lock_irq(&ctx->completion_lock);
6676 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6677 spin_unlock_irq(&ctx->completion_lock);
6680 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6682 spin_lock_irq(&ctx->completion_lock);
6683 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6684 spin_unlock_irq(&ctx->completion_lock);
6687 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6689 unsigned int to_submit;
6692 to_submit = io_sqring_entries(ctx);
6693 /* if we're handling multiple rings, cap submit size for fairness */
6694 if (cap_entries && to_submit > 8)
6697 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6698 unsigned nr_events = 0;
6700 mutex_lock(&ctx->uring_lock);
6701 if (!list_empty(&ctx->iopoll_list))
6702 io_do_iopoll(ctx, &nr_events, 0);
6704 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6705 !(ctx->flags & IORING_SETUP_R_DISABLED))
6706 ret = io_submit_sqes(ctx, to_submit);
6707 mutex_unlock(&ctx->uring_lock);
6710 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6711 wake_up(&ctx->sqo_sq_wait);
6716 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6718 struct io_ring_ctx *ctx;
6719 unsigned sq_thread_idle = 0;
6721 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6722 if (sq_thread_idle < ctx->sq_thread_idle)
6723 sq_thread_idle = ctx->sq_thread_idle;
6726 sqd->sq_thread_idle = sq_thread_idle;
6729 static int io_sq_thread(void *data)
6731 struct io_sq_data *sqd = data;
6732 struct io_ring_ctx *ctx;
6733 unsigned long timeout = 0;
6734 char buf[TASK_COMM_LEN];
6737 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6738 set_task_comm(current, buf);
6739 current->pf_io_worker = NULL;
6741 if (sqd->sq_cpu != -1)
6742 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6744 set_cpus_allowed_ptr(current, cpu_online_mask);
6745 current->flags |= PF_NO_SETAFFINITY;
6747 mutex_lock(&sqd->lock);
6748 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6750 bool cap_entries, sqt_spin, needs_sched;
6752 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6753 signal_pending(current)) {
6754 bool did_sig = false;
6756 mutex_unlock(&sqd->lock);
6757 if (signal_pending(current)) {
6758 struct ksignal ksig;
6760 did_sig = get_signal(&ksig);
6763 mutex_lock(&sqd->lock);
6767 io_run_task_work_head(&sqd->park_task_work);
6768 timeout = jiffies + sqd->sq_thread_idle;
6772 cap_entries = !list_is_singular(&sqd->ctx_list);
6773 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6774 const struct cred *creds = NULL;
6776 if (ctx->sq_creds != current_cred())
6777 creds = override_creds(ctx->sq_creds);
6778 ret = __io_sq_thread(ctx, cap_entries);
6780 revert_creds(creds);
6781 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6785 if (sqt_spin || !time_after(jiffies, timeout)) {
6789 timeout = jiffies + sqd->sq_thread_idle;
6794 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6795 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6796 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6797 !list_empty_careful(&ctx->iopoll_list)) {
6798 needs_sched = false;
6801 if (io_sqring_entries(ctx)) {
6802 needs_sched = false;
6807 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6808 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6809 io_ring_set_wakeup_flag(ctx);
6811 mutex_unlock(&sqd->lock);
6813 mutex_lock(&sqd->lock);
6814 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6815 io_ring_clear_wakeup_flag(ctx);
6818 finish_wait(&sqd->wait, &wait);
6819 io_run_task_work_head(&sqd->park_task_work);
6820 timeout = jiffies + sqd->sq_thread_idle;
6823 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6824 io_uring_cancel_sqpoll(ctx);
6826 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6827 io_ring_set_wakeup_flag(ctx);
6828 mutex_unlock(&sqd->lock);
6831 io_run_task_work_head(&sqd->park_task_work);
6832 complete(&sqd->exited);
6836 struct io_wait_queue {
6837 struct wait_queue_entry wq;
6838 struct io_ring_ctx *ctx;
6840 unsigned nr_timeouts;
6843 static inline bool io_should_wake(struct io_wait_queue *iowq)
6845 struct io_ring_ctx *ctx = iowq->ctx;
6848 * Wake up if we have enough events, or if a timeout occurred since we
6849 * started waiting. For timeouts, we always want to return to userspace,
6850 * regardless of event count.
6852 return io_cqring_events(ctx) >= iowq->to_wait ||
6853 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6856 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6857 int wake_flags, void *key)
6859 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6863 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6864 * the task, and the next invocation will do it.
6866 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6867 return autoremove_wake_function(curr, mode, wake_flags, key);
6871 static int io_run_task_work_sig(void)
6873 if (io_run_task_work())
6875 if (!signal_pending(current))
6877 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6878 return -ERESTARTSYS;
6882 /* when returns >0, the caller should retry */
6883 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6884 struct io_wait_queue *iowq,
6885 signed long *timeout)
6889 /* make sure we run task_work before checking for signals */
6890 ret = io_run_task_work_sig();
6891 if (ret || io_should_wake(iowq))
6893 /* let the caller flush overflows, retry */
6894 if (test_bit(0, &ctx->cq_check_overflow))
6897 *timeout = schedule_timeout(*timeout);
6898 return !*timeout ? -ETIME : 1;
6902 * Wait until events become available, if we don't already have some. The
6903 * application must reap them itself, as they reside on the shared cq ring.
6905 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6906 const sigset_t __user *sig, size_t sigsz,
6907 struct __kernel_timespec __user *uts)
6909 struct io_wait_queue iowq = {
6912 .func = io_wake_function,
6913 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6916 .to_wait = min_events,
6918 struct io_rings *rings = ctx->rings;
6919 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6923 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6924 if (io_cqring_events(ctx) >= min_events)
6926 if (!io_run_task_work())
6931 #ifdef CONFIG_COMPAT
6932 if (in_compat_syscall())
6933 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6937 ret = set_user_sigmask(sig, sigsz);
6944 struct timespec64 ts;
6946 if (get_timespec64(&ts, uts))
6948 timeout = timespec64_to_jiffies(&ts);
6951 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6952 trace_io_uring_cqring_wait(ctx, min_events);
6954 /* if we can't even flush overflow, don't wait for more */
6955 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6959 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6960 TASK_INTERRUPTIBLE);
6961 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6962 finish_wait(&ctx->wait, &iowq.wq);
6966 restore_saved_sigmask_unless(ret == -EINTR);
6968 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6971 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6973 #if defined(CONFIG_UNIX)
6974 if (ctx->ring_sock) {
6975 struct sock *sock = ctx->ring_sock->sk;
6976 struct sk_buff *skb;
6978 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6984 for (i = 0; i < ctx->nr_user_files; i++) {
6987 file = io_file_from_index(ctx, i);
6994 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6996 struct fixed_rsrc_data *data;
6998 data = container_of(ref, struct fixed_rsrc_data, refs);
6999 complete(&data->done);
7002 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7004 spin_lock_bh(&ctx->rsrc_ref_lock);
7007 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7009 spin_unlock_bh(&ctx->rsrc_ref_lock);
7012 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7013 struct fixed_rsrc_data *rsrc_data,
7014 struct fixed_rsrc_ref_node *ref_node)
7016 io_rsrc_ref_lock(ctx);
7017 rsrc_data->node = ref_node;
7018 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7019 io_rsrc_ref_unlock(ctx);
7020 percpu_ref_get(&rsrc_data->refs);
7023 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7025 struct fixed_rsrc_ref_node *ref_node = NULL;
7027 io_rsrc_ref_lock(ctx);
7028 ref_node = data->node;
7030 io_rsrc_ref_unlock(ctx);
7032 percpu_ref_kill(&ref_node->refs);
7035 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7036 struct io_ring_ctx *ctx,
7037 void (*rsrc_put)(struct io_ring_ctx *ctx,
7038 struct io_rsrc_put *prsrc))
7040 struct fixed_rsrc_ref_node *backup_node;
7046 data->quiesce = true;
7049 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7052 backup_node->rsrc_data = data;
7053 backup_node->rsrc_put = rsrc_put;
7055 io_sqe_rsrc_kill_node(ctx, data);
7056 percpu_ref_kill(&data->refs);
7057 flush_delayed_work(&ctx->rsrc_put_work);
7059 ret = wait_for_completion_interruptible(&data->done);
7063 percpu_ref_resurrect(&data->refs);
7064 io_sqe_rsrc_set_node(ctx, data, backup_node);
7066 reinit_completion(&data->done);
7067 mutex_unlock(&ctx->uring_lock);
7068 ret = io_run_task_work_sig();
7069 mutex_lock(&ctx->uring_lock);
7071 data->quiesce = false;
7074 destroy_fixed_rsrc_ref_node(backup_node);
7078 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7080 struct fixed_rsrc_data *data;
7082 data = kzalloc(sizeof(*data), GFP_KERNEL);
7086 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7087 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7092 init_completion(&data->done);
7096 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7098 percpu_ref_exit(&data->refs);
7103 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7105 struct fixed_rsrc_data *data = ctx->file_data;
7106 unsigned nr_tables, i;
7110 * percpu_ref_is_dying() is to stop parallel files unregister
7111 * Since we possibly drop uring lock later in this function to
7114 if (!data || percpu_ref_is_dying(&data->refs))
7116 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7120 __io_sqe_files_unregister(ctx);
7121 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7122 for (i = 0; i < nr_tables; i++)
7123 kfree(data->table[i].files);
7124 free_fixed_rsrc_data(data);
7125 ctx->file_data = NULL;
7126 ctx->nr_user_files = 0;
7130 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7131 __releases(&sqd->lock)
7133 WARN_ON_ONCE(sqd->thread == current);
7136 * Do the dance but not conditional clear_bit() because it'd race with
7137 * other threads incrementing park_pending and setting the bit.
7139 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7140 if (atomic_dec_return(&sqd->park_pending))
7141 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7142 mutex_unlock(&sqd->lock);
7145 static void io_sq_thread_park(struct io_sq_data *sqd)
7146 __acquires(&sqd->lock)
7148 WARN_ON_ONCE(sqd->thread == current);
7150 atomic_inc(&sqd->park_pending);
7151 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7152 mutex_lock(&sqd->lock);
7154 wake_up_process(sqd->thread);
7157 static void io_sq_thread_stop(struct io_sq_data *sqd)
7159 WARN_ON_ONCE(sqd->thread == current);
7161 mutex_lock(&sqd->lock);
7162 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7164 wake_up_process(sqd->thread);
7165 mutex_unlock(&sqd->lock);
7166 wait_for_completion(&sqd->exited);
7169 static void io_put_sq_data(struct io_sq_data *sqd)
7171 if (refcount_dec_and_test(&sqd->refs)) {
7172 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7174 io_sq_thread_stop(sqd);
7179 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7181 struct io_sq_data *sqd = ctx->sq_data;
7184 io_sq_thread_park(sqd);
7185 list_del_init(&ctx->sqd_list);
7186 io_sqd_update_thread_idle(sqd);
7187 io_sq_thread_unpark(sqd);
7189 io_put_sq_data(sqd);
7190 ctx->sq_data = NULL;
7192 put_cred(ctx->sq_creds);
7196 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7198 struct io_ring_ctx *ctx_attach;
7199 struct io_sq_data *sqd;
7202 f = fdget(p->wq_fd);
7204 return ERR_PTR(-ENXIO);
7205 if (f.file->f_op != &io_uring_fops) {
7207 return ERR_PTR(-EINVAL);
7210 ctx_attach = f.file->private_data;
7211 sqd = ctx_attach->sq_data;
7214 return ERR_PTR(-EINVAL);
7216 if (sqd->task_tgid != current->tgid) {
7218 return ERR_PTR(-EPERM);
7221 refcount_inc(&sqd->refs);
7226 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7229 struct io_sq_data *sqd;
7232 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7233 sqd = io_attach_sq_data(p);
7238 /* fall through for EPERM case, setup new sqd/task */
7239 if (PTR_ERR(sqd) != -EPERM)
7243 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7245 return ERR_PTR(-ENOMEM);
7247 atomic_set(&sqd->park_pending, 0);
7248 refcount_set(&sqd->refs, 1);
7249 INIT_LIST_HEAD(&sqd->ctx_list);
7250 mutex_init(&sqd->lock);
7251 init_waitqueue_head(&sqd->wait);
7252 init_completion(&sqd->exited);
7256 #if defined(CONFIG_UNIX)
7258 * Ensure the UNIX gc is aware of our file set, so we are certain that
7259 * the io_uring can be safely unregistered on process exit, even if we have
7260 * loops in the file referencing.
7262 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7264 struct sock *sk = ctx->ring_sock->sk;
7265 struct scm_fp_list *fpl;
7266 struct sk_buff *skb;
7269 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7273 skb = alloc_skb(0, GFP_KERNEL);
7282 fpl->user = get_uid(current_user());
7283 for (i = 0; i < nr; i++) {
7284 struct file *file = io_file_from_index(ctx, i + offset);
7288 fpl->fp[nr_files] = get_file(file);
7289 unix_inflight(fpl->user, fpl->fp[nr_files]);
7294 fpl->max = SCM_MAX_FD;
7295 fpl->count = nr_files;
7296 UNIXCB(skb).fp = fpl;
7297 skb->destructor = unix_destruct_scm;
7298 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7299 skb_queue_head(&sk->sk_receive_queue, skb);
7301 for (i = 0; i < nr_files; i++)
7312 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7313 * causes regular reference counting to break down. We rely on the UNIX
7314 * garbage collection to take care of this problem for us.
7316 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7318 unsigned left, total;
7322 left = ctx->nr_user_files;
7324 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7326 ret = __io_sqe_files_scm(ctx, this_files, total);
7330 total += this_files;
7336 while (total < ctx->nr_user_files) {
7337 struct file *file = io_file_from_index(ctx, total);
7347 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7353 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7354 unsigned nr_tables, unsigned nr_files)
7358 for (i = 0; i < nr_tables; i++) {
7359 struct fixed_rsrc_table *table = &file_data->table[i];
7360 unsigned this_files;
7362 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7363 table->files = kcalloc(this_files, sizeof(struct file *),
7367 nr_files -= this_files;
7373 for (i = 0; i < nr_tables; i++) {
7374 struct fixed_rsrc_table *table = &file_data->table[i];
7375 kfree(table->files);
7380 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7382 struct file *file = prsrc->file;
7383 #if defined(CONFIG_UNIX)
7384 struct sock *sock = ctx->ring_sock->sk;
7385 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7386 struct sk_buff *skb;
7389 __skb_queue_head_init(&list);
7392 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7393 * remove this entry and rearrange the file array.
7395 skb = skb_dequeue(head);
7397 struct scm_fp_list *fp;
7399 fp = UNIXCB(skb).fp;
7400 for (i = 0; i < fp->count; i++) {
7403 if (fp->fp[i] != file)
7406 unix_notinflight(fp->user, fp->fp[i]);
7407 left = fp->count - 1 - i;
7409 memmove(&fp->fp[i], &fp->fp[i + 1],
7410 left * sizeof(struct file *));
7417 __skb_queue_tail(&list, skb);
7427 __skb_queue_tail(&list, skb);
7429 skb = skb_dequeue(head);
7432 if (skb_peek(&list)) {
7433 spin_lock_irq(&head->lock);
7434 while ((skb = __skb_dequeue(&list)) != NULL)
7435 __skb_queue_tail(head, skb);
7436 spin_unlock_irq(&head->lock);
7443 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7445 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7446 struct io_ring_ctx *ctx = rsrc_data->ctx;
7447 struct io_rsrc_put *prsrc, *tmp;
7449 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7450 list_del(&prsrc->list);
7451 ref_node->rsrc_put(ctx, prsrc);
7455 percpu_ref_exit(&ref_node->refs);
7457 percpu_ref_put(&rsrc_data->refs);
7460 static void io_rsrc_put_work(struct work_struct *work)
7462 struct io_ring_ctx *ctx;
7463 struct llist_node *node;
7465 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7466 node = llist_del_all(&ctx->rsrc_put_llist);
7469 struct fixed_rsrc_ref_node *ref_node;
7470 struct llist_node *next = node->next;
7472 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7473 __io_rsrc_put_work(ref_node);
7478 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7480 struct fixed_rsrc_ref_node *ref_node;
7481 struct fixed_rsrc_data *data;
7482 struct io_ring_ctx *ctx;
7483 bool first_add = false;
7486 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7487 data = ref_node->rsrc_data;
7490 io_rsrc_ref_lock(ctx);
7491 ref_node->done = true;
7493 while (!list_empty(&ctx->rsrc_ref_list)) {
7494 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7495 struct fixed_rsrc_ref_node, node);
7496 /* recycle ref nodes in order */
7497 if (!ref_node->done)
7499 list_del(&ref_node->node);
7500 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7502 io_rsrc_ref_unlock(ctx);
7504 if (percpu_ref_is_dying(&data->refs))
7508 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7510 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7513 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7514 struct io_ring_ctx *ctx)
7516 struct fixed_rsrc_ref_node *ref_node;
7518 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7522 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7527 INIT_LIST_HEAD(&ref_node->node);
7528 INIT_LIST_HEAD(&ref_node->rsrc_list);
7529 ref_node->done = false;
7533 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7534 struct fixed_rsrc_ref_node *ref_node)
7536 ref_node->rsrc_data = ctx->file_data;
7537 ref_node->rsrc_put = io_ring_file_put;
7540 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7542 percpu_ref_exit(&ref_node->refs);
7547 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7550 __s32 __user *fds = (__s32 __user *) arg;
7551 unsigned nr_tables, i;
7553 int fd, ret = -ENOMEM;
7554 struct fixed_rsrc_ref_node *ref_node;
7555 struct fixed_rsrc_data *file_data;
7561 if (nr_args > IORING_MAX_FIXED_FILES)
7564 file_data = alloc_fixed_rsrc_data(ctx);
7567 ctx->file_data = file_data;
7569 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7570 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7572 if (!file_data->table)
7575 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7578 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7579 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7583 /* allow sparse sets */
7593 * Don't allow io_uring instances to be registered. If UNIX
7594 * isn't enabled, then this causes a reference cycle and this
7595 * instance can never get freed. If UNIX is enabled we'll
7596 * handle it just fine, but there's still no point in allowing
7597 * a ring fd as it doesn't support regular read/write anyway.
7599 if (file->f_op == &io_uring_fops) {
7603 *io_fixed_file_slot(file_data, i) = file;
7606 ret = io_sqe_files_scm(ctx);
7608 io_sqe_files_unregister(ctx);
7612 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7614 io_sqe_files_unregister(ctx);
7617 init_fixed_file_ref_node(ctx, ref_node);
7619 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7622 for (i = 0; i < ctx->nr_user_files; i++) {
7623 file = io_file_from_index(ctx, i);
7627 for (i = 0; i < nr_tables; i++)
7628 kfree(file_data->table[i].files);
7629 ctx->nr_user_files = 0;
7631 free_fixed_rsrc_data(ctx->file_data);
7632 ctx->file_data = NULL;
7636 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7639 #if defined(CONFIG_UNIX)
7640 struct sock *sock = ctx->ring_sock->sk;
7641 struct sk_buff_head *head = &sock->sk_receive_queue;
7642 struct sk_buff *skb;
7645 * See if we can merge this file into an existing skb SCM_RIGHTS
7646 * file set. If there's no room, fall back to allocating a new skb
7647 * and filling it in.
7649 spin_lock_irq(&head->lock);
7650 skb = skb_peek(head);
7652 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7654 if (fpl->count < SCM_MAX_FD) {
7655 __skb_unlink(skb, head);
7656 spin_unlock_irq(&head->lock);
7657 fpl->fp[fpl->count] = get_file(file);
7658 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7660 spin_lock_irq(&head->lock);
7661 __skb_queue_head(head, skb);
7666 spin_unlock_irq(&head->lock);
7673 return __io_sqe_files_scm(ctx, 1, index);
7679 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7681 struct io_rsrc_put *prsrc;
7682 struct fixed_rsrc_ref_node *ref_node = data->node;
7684 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7689 list_add(&prsrc->list, &ref_node->rsrc_list);
7694 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7697 return io_queue_rsrc_removal(data, (void *)file);
7700 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7701 struct io_uring_rsrc_update *up,
7704 struct fixed_rsrc_data *data = ctx->file_data;
7705 struct fixed_rsrc_ref_node *ref_node;
7706 struct file *file, **file_slot;
7710 bool needs_switch = false;
7712 if (check_add_overflow(up->offset, nr_args, &done))
7714 if (done > ctx->nr_user_files)
7717 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7720 init_fixed_file_ref_node(ctx, ref_node);
7722 fds = u64_to_user_ptr(up->data);
7723 for (done = 0; done < nr_args; done++) {
7725 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7729 if (fd == IORING_REGISTER_FILES_SKIP)
7732 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7733 file_slot = io_fixed_file_slot(ctx->file_data, i);
7736 err = io_queue_file_removal(data, *file_slot);
7740 needs_switch = true;
7749 * Don't allow io_uring instances to be registered. If
7750 * UNIX isn't enabled, then this causes a reference
7751 * cycle and this instance can never get freed. If UNIX
7752 * is enabled we'll handle it just fine, but there's
7753 * still no point in allowing a ring fd as it doesn't
7754 * support regular read/write anyway.
7756 if (file->f_op == &io_uring_fops) {
7762 err = io_sqe_file_register(ctx, file, i);
7772 percpu_ref_kill(&data->node->refs);
7773 io_sqe_rsrc_set_node(ctx, data, ref_node);
7775 destroy_fixed_rsrc_ref_node(ref_node);
7777 return done ? done : err;
7780 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7783 struct io_uring_rsrc_update up;
7785 if (!ctx->file_data)
7789 if (copy_from_user(&up, arg, sizeof(up)))
7794 return __io_sqe_files_update(ctx, &up, nr_args);
7797 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7799 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7801 req = io_put_req_find_next(req);
7802 return req ? &req->work : NULL;
7805 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7807 struct io_wq_hash *hash;
7808 struct io_wq_data data;
7809 unsigned int concurrency;
7811 hash = ctx->hash_map;
7813 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7815 return ERR_PTR(-ENOMEM);
7816 refcount_set(&hash->refs, 1);
7817 init_waitqueue_head(&hash->wait);
7818 ctx->hash_map = hash;
7822 data.free_work = io_free_work;
7823 data.do_work = io_wq_submit_work;
7825 /* Do QD, or 4 * CPUS, whatever is smallest */
7826 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7828 return io_wq_create(concurrency, &data);
7831 static int io_uring_alloc_task_context(struct task_struct *task,
7832 struct io_ring_ctx *ctx)
7834 struct io_uring_task *tctx;
7837 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7838 if (unlikely(!tctx))
7841 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7842 if (unlikely(ret)) {
7847 tctx->io_wq = io_init_wq_offload(ctx);
7848 if (IS_ERR(tctx->io_wq)) {
7849 ret = PTR_ERR(tctx->io_wq);
7850 percpu_counter_destroy(&tctx->inflight);
7856 init_waitqueue_head(&tctx->wait);
7858 atomic_set(&tctx->in_idle, 0);
7859 task->io_uring = tctx;
7860 spin_lock_init(&tctx->task_lock);
7861 INIT_WQ_LIST(&tctx->task_list);
7862 tctx->task_state = 0;
7863 init_task_work(&tctx->task_work, tctx_task_work);
7867 void __io_uring_free(struct task_struct *tsk)
7869 struct io_uring_task *tctx = tsk->io_uring;
7871 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7872 WARN_ON_ONCE(tctx->io_wq);
7874 percpu_counter_destroy(&tctx->inflight);
7876 tsk->io_uring = NULL;
7879 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7880 struct io_uring_params *p)
7884 /* Retain compatibility with failing for an invalid attach attempt */
7885 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7886 IORING_SETUP_ATTACH_WQ) {
7889 f = fdget(p->wq_fd);
7892 if (f.file->f_op != &io_uring_fops) {
7898 if (ctx->flags & IORING_SETUP_SQPOLL) {
7899 struct task_struct *tsk;
7900 struct io_sq_data *sqd;
7904 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7907 sqd = io_get_sq_data(p, &attached);
7913 ctx->sq_creds = get_current_cred();
7915 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7916 if (!ctx->sq_thread_idle)
7917 ctx->sq_thread_idle = HZ;
7920 io_sq_thread_park(sqd);
7921 list_add(&ctx->sqd_list, &sqd->ctx_list);
7922 io_sqd_update_thread_idle(sqd);
7923 /* don't attach to a dying SQPOLL thread, would be racy */
7924 if (attached && !sqd->thread)
7926 io_sq_thread_unpark(sqd);
7933 if (p->flags & IORING_SETUP_SQ_AFF) {
7934 int cpu = p->sq_thread_cpu;
7937 if (cpu >= nr_cpu_ids)
7939 if (!cpu_online(cpu))
7947 sqd->task_pid = current->pid;
7948 sqd->task_tgid = current->tgid;
7949 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7956 ret = io_uring_alloc_task_context(tsk, ctx);
7957 wake_up_new_task(tsk);
7960 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7961 /* Can't have SQ_AFF without SQPOLL */
7968 io_sq_thread_finish(ctx);
7971 complete(&ctx->sq_data->exited);
7975 static inline void __io_unaccount_mem(struct user_struct *user,
7976 unsigned long nr_pages)
7978 atomic_long_sub(nr_pages, &user->locked_vm);
7981 static inline int __io_account_mem(struct user_struct *user,
7982 unsigned long nr_pages)
7984 unsigned long page_limit, cur_pages, new_pages;
7986 /* Don't allow more pages than we can safely lock */
7987 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7990 cur_pages = atomic_long_read(&user->locked_vm);
7991 new_pages = cur_pages + nr_pages;
7992 if (new_pages > page_limit)
7994 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7995 new_pages) != cur_pages);
8000 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8003 __io_unaccount_mem(ctx->user, nr_pages);
8005 if (ctx->mm_account)
8006 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8009 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8014 ret = __io_account_mem(ctx->user, nr_pages);
8019 if (ctx->mm_account)
8020 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8025 static void io_mem_free(void *ptr)
8032 page = virt_to_head_page(ptr);
8033 if (put_page_testzero(page))
8034 free_compound_page(page);
8037 static void *io_mem_alloc(size_t size)
8039 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8040 __GFP_NORETRY | __GFP_ACCOUNT;
8042 return (void *) __get_free_pages(gfp_flags, get_order(size));
8045 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8048 struct io_rings *rings;
8049 size_t off, sq_array_size;
8051 off = struct_size(rings, cqes, cq_entries);
8052 if (off == SIZE_MAX)
8056 off = ALIGN(off, SMP_CACHE_BYTES);
8064 sq_array_size = array_size(sizeof(u32), sq_entries);
8065 if (sq_array_size == SIZE_MAX)
8068 if (check_add_overflow(off, sq_array_size, &off))
8074 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8078 if (!ctx->user_bufs)
8081 for (i = 0; i < ctx->nr_user_bufs; i++) {
8082 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8084 for (j = 0; j < imu->nr_bvecs; j++)
8085 unpin_user_page(imu->bvec[j].bv_page);
8087 if (imu->acct_pages)
8088 io_unaccount_mem(ctx, imu->acct_pages);
8093 kfree(ctx->user_bufs);
8094 ctx->user_bufs = NULL;
8095 ctx->nr_user_bufs = 0;
8099 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8100 void __user *arg, unsigned index)
8102 struct iovec __user *src;
8104 #ifdef CONFIG_COMPAT
8106 struct compat_iovec __user *ciovs;
8107 struct compat_iovec ciov;
8109 ciovs = (struct compat_iovec __user *) arg;
8110 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8113 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8114 dst->iov_len = ciov.iov_len;
8118 src = (struct iovec __user *) arg;
8119 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8125 * Not super efficient, but this is just a registration time. And we do cache
8126 * the last compound head, so generally we'll only do a full search if we don't
8129 * We check if the given compound head page has already been accounted, to
8130 * avoid double accounting it. This allows us to account the full size of the
8131 * page, not just the constituent pages of a huge page.
8133 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8134 int nr_pages, struct page *hpage)
8138 /* check current page array */
8139 for (i = 0; i < nr_pages; i++) {
8140 if (!PageCompound(pages[i]))
8142 if (compound_head(pages[i]) == hpage)
8146 /* check previously registered pages */
8147 for (i = 0; i < ctx->nr_user_bufs; i++) {
8148 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8150 for (j = 0; j < imu->nr_bvecs; j++) {
8151 if (!PageCompound(imu->bvec[j].bv_page))
8153 if (compound_head(imu->bvec[j].bv_page) == hpage)
8161 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8162 int nr_pages, struct io_mapped_ubuf *imu,
8163 struct page **last_hpage)
8167 for (i = 0; i < nr_pages; i++) {
8168 if (!PageCompound(pages[i])) {
8173 hpage = compound_head(pages[i]);
8174 if (hpage == *last_hpage)
8176 *last_hpage = hpage;
8177 if (headpage_already_acct(ctx, pages, i, hpage))
8179 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8183 if (!imu->acct_pages)
8186 ret = io_account_mem(ctx, imu->acct_pages);
8188 imu->acct_pages = 0;
8192 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8193 struct io_mapped_ubuf *imu,
8194 struct page **last_hpage)
8196 struct vm_area_struct **vmas = NULL;
8197 struct page **pages = NULL;
8198 unsigned long off, start, end, ubuf;
8200 int ret, pret, nr_pages, i;
8202 ubuf = (unsigned long) iov->iov_base;
8203 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8204 start = ubuf >> PAGE_SHIFT;
8205 nr_pages = end - start;
8209 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8213 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8218 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8224 mmap_read_lock(current->mm);
8225 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8227 if (pret == nr_pages) {
8228 /* don't support file backed memory */
8229 for (i = 0; i < nr_pages; i++) {
8230 struct vm_area_struct *vma = vmas[i];
8233 !is_file_hugepages(vma->vm_file)) {
8239 ret = pret < 0 ? pret : -EFAULT;
8241 mmap_read_unlock(current->mm);
8244 * if we did partial map, or found file backed vmas,
8245 * release any pages we did get
8248 unpin_user_pages(pages, pret);
8253 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8255 unpin_user_pages(pages, pret);
8260 off = ubuf & ~PAGE_MASK;
8261 size = iov->iov_len;
8262 for (i = 0; i < nr_pages; i++) {
8265 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8266 imu->bvec[i].bv_page = pages[i];
8267 imu->bvec[i].bv_len = vec_len;
8268 imu->bvec[i].bv_offset = off;
8272 /* store original address for later verification */
8274 imu->len = iov->iov_len;
8275 imu->nr_bvecs = nr_pages;
8283 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8287 if (!nr_args || nr_args > UIO_MAXIOV)
8290 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8292 if (!ctx->user_bufs)
8298 static int io_buffer_validate(struct iovec *iov)
8301 * Don't impose further limits on the size and buffer
8302 * constraints here, we'll -EINVAL later when IO is
8303 * submitted if they are wrong.
8305 if (!iov->iov_base || !iov->iov_len)
8308 /* arbitrary limit, but we need something */
8309 if (iov->iov_len > SZ_1G)
8315 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8316 unsigned int nr_args)
8320 struct page *last_hpage = NULL;
8322 ret = io_buffers_map_alloc(ctx, nr_args);
8326 for (i = 0; i < nr_args; i++) {
8327 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8329 ret = io_copy_iov(ctx, &iov, arg, i);
8333 ret = io_buffer_validate(&iov);
8337 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8341 ctx->nr_user_bufs++;
8345 io_sqe_buffers_unregister(ctx);
8350 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8352 __s32 __user *fds = arg;
8358 if (copy_from_user(&fd, fds, sizeof(*fds)))
8361 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8362 if (IS_ERR(ctx->cq_ev_fd)) {
8363 int ret = PTR_ERR(ctx->cq_ev_fd);
8364 ctx->cq_ev_fd = NULL;
8371 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8373 if (ctx->cq_ev_fd) {
8374 eventfd_ctx_put(ctx->cq_ev_fd);
8375 ctx->cq_ev_fd = NULL;
8382 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8384 struct io_buffer *buf;
8385 unsigned long index;
8387 xa_for_each(&ctx->io_buffers, index, buf)
8388 __io_remove_buffers(ctx, buf, index, -1U);
8391 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8393 struct io_kiocb *req, *nxt;
8395 list_for_each_entry_safe(req, nxt, list, compl.list) {
8396 if (tsk && req->task != tsk)
8398 list_del(&req->compl.list);
8399 kmem_cache_free(req_cachep, req);
8403 static void io_req_caches_free(struct io_ring_ctx *ctx)
8405 struct io_submit_state *submit_state = &ctx->submit_state;
8406 struct io_comp_state *cs = &ctx->submit_state.comp;
8408 mutex_lock(&ctx->uring_lock);
8410 if (submit_state->free_reqs) {
8411 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8412 submit_state->reqs);
8413 submit_state->free_reqs = 0;
8416 spin_lock_irq(&ctx->completion_lock);
8417 list_splice_init(&cs->locked_free_list, &cs->free_list);
8418 cs->locked_free_nr = 0;
8419 spin_unlock_irq(&ctx->completion_lock);
8421 io_req_cache_free(&cs->free_list, NULL);
8423 mutex_unlock(&ctx->uring_lock);
8426 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8429 * Some may use context even when all refs and requests have been put,
8430 * and they are free to do so while still holding uring_lock or
8431 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8433 mutex_lock(&ctx->uring_lock);
8434 mutex_unlock(&ctx->uring_lock);
8435 spin_lock_irq(&ctx->completion_lock);
8436 spin_unlock_irq(&ctx->completion_lock);
8438 io_sq_thread_finish(ctx);
8439 io_sqe_buffers_unregister(ctx);
8441 if (ctx->mm_account) {
8442 mmdrop(ctx->mm_account);
8443 ctx->mm_account = NULL;
8446 mutex_lock(&ctx->uring_lock);
8447 io_sqe_files_unregister(ctx);
8448 mutex_unlock(&ctx->uring_lock);
8449 io_eventfd_unregister(ctx);
8450 io_destroy_buffers(ctx);
8452 #if defined(CONFIG_UNIX)
8453 if (ctx->ring_sock) {
8454 ctx->ring_sock->file = NULL; /* so that iput() is called */
8455 sock_release(ctx->ring_sock);
8459 io_mem_free(ctx->rings);
8460 io_mem_free(ctx->sq_sqes);
8462 percpu_ref_exit(&ctx->refs);
8463 free_uid(ctx->user);
8464 io_req_caches_free(ctx);
8466 io_wq_put_hash(ctx->hash_map);
8467 kfree(ctx->cancel_hash);
8471 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8473 struct io_ring_ctx *ctx = file->private_data;
8476 poll_wait(file, &ctx->cq_wait, wait);
8478 * synchronizes with barrier from wq_has_sleeper call in
8482 if (!io_sqring_full(ctx))
8483 mask |= EPOLLOUT | EPOLLWRNORM;
8486 * Don't flush cqring overflow list here, just do a simple check.
8487 * Otherwise there could possible be ABBA deadlock:
8490 * lock(&ctx->uring_lock);
8492 * lock(&ctx->uring_lock);
8495 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8496 * pushs them to do the flush.
8498 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8499 mask |= EPOLLIN | EPOLLRDNORM;
8504 static int io_uring_fasync(int fd, struct file *file, int on)
8506 struct io_ring_ctx *ctx = file->private_data;
8508 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8511 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8513 const struct cred *creds;
8515 creds = xa_erase(&ctx->personalities, id);
8524 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8526 return io_run_task_work_head(&ctx->exit_task_work);
8529 struct io_tctx_exit {
8530 struct callback_head task_work;
8531 struct completion completion;
8532 struct io_ring_ctx *ctx;
8535 static void io_tctx_exit_cb(struct callback_head *cb)
8537 struct io_uring_task *tctx = current->io_uring;
8538 struct io_tctx_exit *work;
8540 work = container_of(cb, struct io_tctx_exit, task_work);
8542 * When @in_idle, we're in cancellation and it's racy to remove the
8543 * node. It'll be removed by the end of cancellation, just ignore it.
8545 if (!atomic_read(&tctx->in_idle))
8546 io_uring_del_task_file((unsigned long)work->ctx);
8547 complete(&work->completion);
8550 static void io_ring_exit_work(struct work_struct *work)
8552 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8553 unsigned long timeout = jiffies + HZ * 60 * 5;
8554 struct io_tctx_exit exit;
8555 struct io_tctx_node *node;
8558 /* prevent SQPOLL from submitting new requests */
8560 io_sq_thread_park(ctx->sq_data);
8561 list_del_init(&ctx->sqd_list);
8562 io_sqd_update_thread_idle(ctx->sq_data);
8563 io_sq_thread_unpark(ctx->sq_data);
8567 * If we're doing polled IO and end up having requests being
8568 * submitted async (out-of-line), then completions can come in while
8569 * we're waiting for refs to drop. We need to reap these manually,
8570 * as nobody else will be looking for them.
8573 io_uring_try_cancel_requests(ctx, NULL, NULL);
8575 WARN_ON_ONCE(time_after(jiffies, timeout));
8576 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8578 mutex_lock(&ctx->uring_lock);
8579 while (!list_empty(&ctx->tctx_list)) {
8580 WARN_ON_ONCE(time_after(jiffies, timeout));
8582 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8585 init_completion(&exit.completion);
8586 init_task_work(&exit.task_work, io_tctx_exit_cb);
8587 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8588 if (WARN_ON_ONCE(ret))
8590 wake_up_process(node->task);
8592 mutex_unlock(&ctx->uring_lock);
8593 wait_for_completion(&exit.completion);
8595 mutex_lock(&ctx->uring_lock);
8597 mutex_unlock(&ctx->uring_lock);
8599 io_ring_ctx_free(ctx);
8602 /* Returns true if we found and killed one or more timeouts */
8603 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8604 struct files_struct *files)
8606 struct io_kiocb *req, *tmp;
8609 spin_lock_irq(&ctx->completion_lock);
8610 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8611 if (io_match_task(req, tsk, files)) {
8612 io_kill_timeout(req, -ECANCELED);
8617 io_commit_cqring(ctx);
8618 spin_unlock_irq(&ctx->completion_lock);
8620 io_cqring_ev_posted(ctx);
8621 return canceled != 0;
8624 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8626 unsigned long index;
8627 struct creds *creds;
8629 mutex_lock(&ctx->uring_lock);
8630 percpu_ref_kill(&ctx->refs);
8631 /* if force is set, the ring is going away. always drop after that */
8632 ctx->cq_overflow_flushed = 1;
8634 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8635 xa_for_each(&ctx->personalities, index, creds)
8636 io_unregister_personality(ctx, index);
8637 mutex_unlock(&ctx->uring_lock);
8639 io_kill_timeouts(ctx, NULL, NULL);
8640 io_poll_remove_all(ctx, NULL, NULL);
8642 /* if we failed setting up the ctx, we might not have any rings */
8643 io_iopoll_try_reap_events(ctx);
8645 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8647 * Use system_unbound_wq to avoid spawning tons of event kworkers
8648 * if we're exiting a ton of rings at the same time. It just adds
8649 * noise and overhead, there's no discernable change in runtime
8650 * over using system_wq.
8652 queue_work(system_unbound_wq, &ctx->exit_work);
8655 static int io_uring_release(struct inode *inode, struct file *file)
8657 struct io_ring_ctx *ctx = file->private_data;
8659 file->private_data = NULL;
8660 io_ring_ctx_wait_and_kill(ctx);
8664 struct io_task_cancel {
8665 struct task_struct *task;
8666 struct files_struct *files;
8669 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8671 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8672 struct io_task_cancel *cancel = data;
8675 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8676 unsigned long flags;
8677 struct io_ring_ctx *ctx = req->ctx;
8679 /* protect against races with linked timeouts */
8680 spin_lock_irqsave(&ctx->completion_lock, flags);
8681 ret = io_match_task(req, cancel->task, cancel->files);
8682 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8684 ret = io_match_task(req, cancel->task, cancel->files);
8689 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8690 struct task_struct *task,
8691 struct files_struct *files)
8693 struct io_defer_entry *de;
8696 spin_lock_irq(&ctx->completion_lock);
8697 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8698 if (io_match_task(de->req, task, files)) {
8699 list_cut_position(&list, &ctx->defer_list, &de->list);
8703 spin_unlock_irq(&ctx->completion_lock);
8704 if (list_empty(&list))
8707 while (!list_empty(&list)) {
8708 de = list_first_entry(&list, struct io_defer_entry, list);
8709 list_del_init(&de->list);
8710 io_req_complete_failed(de->req, -ECANCELED);
8716 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8718 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8720 return req->ctx == data;
8723 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8725 struct io_tctx_node *node;
8726 enum io_wq_cancel cret;
8729 mutex_lock(&ctx->uring_lock);
8730 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8731 struct io_uring_task *tctx = node->task->io_uring;
8734 * io_wq will stay alive while we hold uring_lock, because it's
8735 * killed after ctx nodes, which requires to take the lock.
8737 if (!tctx || !tctx->io_wq)
8739 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8740 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8742 mutex_unlock(&ctx->uring_lock);
8747 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8748 struct task_struct *task,
8749 struct files_struct *files)
8751 struct io_task_cancel cancel = { .task = task, .files = files, };
8752 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8755 enum io_wq_cancel cret;
8759 ret |= io_uring_try_cancel_iowq(ctx);
8760 } else if (tctx && tctx->io_wq) {
8762 * Cancels requests of all rings, not only @ctx, but
8763 * it's fine as the task is in exit/exec.
8765 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8767 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8770 /* SQPOLL thread does its own polling */
8771 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8772 (ctx->sq_data && ctx->sq_data->thread == current)) {
8773 while (!list_empty_careful(&ctx->iopoll_list)) {
8774 io_iopoll_try_reap_events(ctx);
8779 ret |= io_cancel_defer_files(ctx, task, files);
8780 ret |= io_poll_remove_all(ctx, task, files);
8781 ret |= io_kill_timeouts(ctx, task, files);
8782 ret |= io_run_task_work();
8783 ret |= io_run_ctx_fallback(ctx);
8784 io_cqring_overflow_flush(ctx, true, task, files);
8791 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8792 struct task_struct *task,
8793 struct files_struct *files)
8795 struct io_kiocb *req;
8798 spin_lock_irq(&ctx->inflight_lock);
8799 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8800 cnt += io_match_task(req, task, files);
8801 spin_unlock_irq(&ctx->inflight_lock);
8805 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8806 struct task_struct *task,
8807 struct files_struct *files)
8809 while (!list_empty_careful(&ctx->inflight_list)) {
8813 inflight = io_uring_count_inflight(ctx, task, files);
8817 io_uring_try_cancel_requests(ctx, task, files);
8819 prepare_to_wait(&task->io_uring->wait, &wait,
8820 TASK_UNINTERRUPTIBLE);
8821 if (inflight == io_uring_count_inflight(ctx, task, files))
8823 finish_wait(&task->io_uring->wait, &wait);
8828 * Note that this task has used io_uring. We use it for cancelation purposes.
8830 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8832 struct io_uring_task *tctx = current->io_uring;
8833 struct io_tctx_node *node;
8836 if (unlikely(!tctx)) {
8837 ret = io_uring_alloc_task_context(current, ctx);
8840 tctx = current->io_uring;
8842 if (tctx->last != ctx) {
8843 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8846 node = kmalloc(sizeof(*node), GFP_KERNEL);
8850 node->task = current;
8852 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8859 mutex_lock(&ctx->uring_lock);
8860 list_add(&node->ctx_node, &ctx->tctx_list);
8861 mutex_unlock(&ctx->uring_lock);
8869 * Remove this io_uring_file -> task mapping.
8871 static void io_uring_del_task_file(unsigned long index)
8873 struct io_uring_task *tctx = current->io_uring;
8874 struct io_tctx_node *node;
8878 node = xa_erase(&tctx->xa, index);
8882 WARN_ON_ONCE(current != node->task);
8883 WARN_ON_ONCE(list_empty(&node->ctx_node));
8885 mutex_lock(&node->ctx->uring_lock);
8886 list_del(&node->ctx_node);
8887 mutex_unlock(&node->ctx->uring_lock);
8889 if (tctx->last == node->ctx)
8894 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8896 struct io_tctx_node *node;
8897 unsigned long index;
8899 xa_for_each(&tctx->xa, index, node)
8900 io_uring_del_task_file(index);
8902 io_wq_put_and_exit(tctx->io_wq);
8907 static s64 tctx_inflight(struct io_uring_task *tctx)
8909 return percpu_counter_sum(&tctx->inflight);
8912 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8914 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8915 struct io_ring_ctx *ctx = work->ctx;
8916 struct io_sq_data *sqd = ctx->sq_data;
8919 io_uring_cancel_sqpoll(ctx);
8920 complete(&work->completion);
8923 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8925 struct io_sq_data *sqd = ctx->sq_data;
8926 struct io_tctx_exit work = { .ctx = ctx, };
8927 struct task_struct *task;
8929 io_sq_thread_park(sqd);
8930 list_del_init(&ctx->sqd_list);
8931 io_sqd_update_thread_idle(sqd);
8934 init_completion(&work.completion);
8935 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8936 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8937 wake_up_process(task);
8939 io_sq_thread_unpark(sqd);
8942 wait_for_completion(&work.completion);
8945 void __io_uring_files_cancel(struct files_struct *files)
8947 struct io_uring_task *tctx = current->io_uring;
8948 struct io_tctx_node *node;
8949 unsigned long index;
8951 /* make sure overflow events are dropped */
8952 atomic_inc(&tctx->in_idle);
8953 xa_for_each(&tctx->xa, index, node) {
8954 struct io_ring_ctx *ctx = node->ctx;
8957 io_sqpoll_cancel_sync(ctx);
8960 io_uring_cancel_files(ctx, current, files);
8962 io_uring_try_cancel_requests(ctx, current, NULL);
8964 atomic_dec(&tctx->in_idle);
8967 io_uring_clean_tctx(tctx);
8970 /* should only be called by SQPOLL task */
8971 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8973 struct io_sq_data *sqd = ctx->sq_data;
8974 struct io_uring_task *tctx = current->io_uring;
8978 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8980 atomic_inc(&tctx->in_idle);
8982 /* read completions before cancelations */
8983 inflight = tctx_inflight(tctx);
8986 io_uring_try_cancel_requests(ctx, current, NULL);
8988 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8990 * If we've seen completions, retry without waiting. This
8991 * avoids a race where a completion comes in before we did
8992 * prepare_to_wait().
8994 if (inflight == tctx_inflight(tctx))
8996 finish_wait(&tctx->wait, &wait);
8998 atomic_dec(&tctx->in_idle);
9002 * Find any io_uring fd that this task has registered or done IO on, and cancel
9005 void __io_uring_task_cancel(void)
9007 struct io_uring_task *tctx = current->io_uring;
9011 /* make sure overflow events are dropped */
9012 atomic_inc(&tctx->in_idle);
9013 __io_uring_files_cancel(NULL);
9016 /* read completions before cancelations */
9017 inflight = tctx_inflight(tctx);
9020 __io_uring_files_cancel(NULL);
9022 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9025 * If we've seen completions, retry without waiting. This
9026 * avoids a race where a completion comes in before we did
9027 * prepare_to_wait().
9029 if (inflight == tctx_inflight(tctx))
9031 finish_wait(&tctx->wait, &wait);
9034 atomic_dec(&tctx->in_idle);
9036 io_uring_clean_tctx(tctx);
9037 /* all current's requests should be gone, we can kill tctx */
9038 __io_uring_free(current);
9041 static void *io_uring_validate_mmap_request(struct file *file,
9042 loff_t pgoff, size_t sz)
9044 struct io_ring_ctx *ctx = file->private_data;
9045 loff_t offset = pgoff << PAGE_SHIFT;
9050 case IORING_OFF_SQ_RING:
9051 case IORING_OFF_CQ_RING:
9054 case IORING_OFF_SQES:
9058 return ERR_PTR(-EINVAL);
9061 page = virt_to_head_page(ptr);
9062 if (sz > page_size(page))
9063 return ERR_PTR(-EINVAL);
9070 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9072 size_t sz = vma->vm_end - vma->vm_start;
9076 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9078 return PTR_ERR(ptr);
9080 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9081 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9084 #else /* !CONFIG_MMU */
9086 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9088 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9091 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9093 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9096 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9097 unsigned long addr, unsigned long len,
9098 unsigned long pgoff, unsigned long flags)
9102 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9104 return PTR_ERR(ptr);
9106 return (unsigned long) ptr;
9109 #endif /* !CONFIG_MMU */
9111 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9116 if (!io_sqring_full(ctx))
9118 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9120 if (!io_sqring_full(ctx))
9123 } while (!signal_pending(current));
9125 finish_wait(&ctx->sqo_sq_wait, &wait);
9129 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9130 struct __kernel_timespec __user **ts,
9131 const sigset_t __user **sig)
9133 struct io_uring_getevents_arg arg;
9136 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9137 * is just a pointer to the sigset_t.
9139 if (!(flags & IORING_ENTER_EXT_ARG)) {
9140 *sig = (const sigset_t __user *) argp;
9146 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9147 * timespec and sigset_t pointers if good.
9149 if (*argsz != sizeof(arg))
9151 if (copy_from_user(&arg, argp, sizeof(arg)))
9153 *sig = u64_to_user_ptr(arg.sigmask);
9154 *argsz = arg.sigmask_sz;
9155 *ts = u64_to_user_ptr(arg.ts);
9159 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9160 u32, min_complete, u32, flags, const void __user *, argp,
9163 struct io_ring_ctx *ctx;
9170 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9171 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9179 if (f.file->f_op != &io_uring_fops)
9183 ctx = f.file->private_data;
9184 if (!percpu_ref_tryget(&ctx->refs))
9188 if (ctx->flags & IORING_SETUP_R_DISABLED)
9192 * For SQ polling, the thread will do all submissions and completions.
9193 * Just return the requested submit count, and wake the thread if
9197 if (ctx->flags & IORING_SETUP_SQPOLL) {
9198 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9201 if (unlikely(ctx->sq_data->thread == NULL)) {
9204 if (flags & IORING_ENTER_SQ_WAKEUP)
9205 wake_up(&ctx->sq_data->wait);
9206 if (flags & IORING_ENTER_SQ_WAIT) {
9207 ret = io_sqpoll_wait_sq(ctx);
9211 submitted = to_submit;
9212 } else if (to_submit) {
9213 ret = io_uring_add_task_file(ctx);
9216 mutex_lock(&ctx->uring_lock);
9217 submitted = io_submit_sqes(ctx, to_submit);
9218 mutex_unlock(&ctx->uring_lock);
9220 if (submitted != to_submit)
9223 if (flags & IORING_ENTER_GETEVENTS) {
9224 const sigset_t __user *sig;
9225 struct __kernel_timespec __user *ts;
9227 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9231 min_complete = min(min_complete, ctx->cq_entries);
9234 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9235 * space applications don't need to do io completion events
9236 * polling again, they can rely on io_sq_thread to do polling
9237 * work, which can reduce cpu usage and uring_lock contention.
9239 if (ctx->flags & IORING_SETUP_IOPOLL &&
9240 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9241 ret = io_iopoll_check(ctx, min_complete);
9243 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9248 percpu_ref_put(&ctx->refs);
9251 return submitted ? submitted : ret;
9254 #ifdef CONFIG_PROC_FS
9255 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9256 const struct cred *cred)
9258 struct user_namespace *uns = seq_user_ns(m);
9259 struct group_info *gi;
9264 seq_printf(m, "%5d\n", id);
9265 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9266 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9267 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9268 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9269 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9270 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9271 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9272 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9273 seq_puts(m, "\n\tGroups:\t");
9274 gi = cred->group_info;
9275 for (g = 0; g < gi->ngroups; g++) {
9276 seq_put_decimal_ull(m, g ? " " : "",
9277 from_kgid_munged(uns, gi->gid[g]));
9279 seq_puts(m, "\n\tCapEff:\t");
9280 cap = cred->cap_effective;
9281 CAP_FOR_EACH_U32(__capi)
9282 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9287 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9289 struct io_sq_data *sq = NULL;
9294 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9295 * since fdinfo case grabs it in the opposite direction of normal use
9296 * cases. If we fail to get the lock, we just don't iterate any
9297 * structures that could be going away outside the io_uring mutex.
9299 has_lock = mutex_trylock(&ctx->uring_lock);
9301 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9307 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9308 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9309 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9310 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9311 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9314 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9316 seq_printf(m, "%5u: <none>\n", i);
9318 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9319 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9320 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9322 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9323 (unsigned int) buf->len);
9325 if (has_lock && !xa_empty(&ctx->personalities)) {
9326 unsigned long index;
9327 const struct cred *cred;
9329 seq_printf(m, "Personalities:\n");
9330 xa_for_each(&ctx->personalities, index, cred)
9331 io_uring_show_cred(m, index, cred);
9333 seq_printf(m, "PollList:\n");
9334 spin_lock_irq(&ctx->completion_lock);
9335 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9336 struct hlist_head *list = &ctx->cancel_hash[i];
9337 struct io_kiocb *req;
9339 hlist_for_each_entry(req, list, hash_node)
9340 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9341 req->task->task_works != NULL);
9343 spin_unlock_irq(&ctx->completion_lock);
9345 mutex_unlock(&ctx->uring_lock);
9348 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9350 struct io_ring_ctx *ctx = f->private_data;
9352 if (percpu_ref_tryget(&ctx->refs)) {
9353 __io_uring_show_fdinfo(ctx, m);
9354 percpu_ref_put(&ctx->refs);
9359 static const struct file_operations io_uring_fops = {
9360 .release = io_uring_release,
9361 .mmap = io_uring_mmap,
9363 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9364 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9366 .poll = io_uring_poll,
9367 .fasync = io_uring_fasync,
9368 #ifdef CONFIG_PROC_FS
9369 .show_fdinfo = io_uring_show_fdinfo,
9373 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9374 struct io_uring_params *p)
9376 struct io_rings *rings;
9377 size_t size, sq_array_offset;
9379 /* make sure these are sane, as we already accounted them */
9380 ctx->sq_entries = p->sq_entries;
9381 ctx->cq_entries = p->cq_entries;
9383 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9384 if (size == SIZE_MAX)
9387 rings = io_mem_alloc(size);
9392 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9393 rings->sq_ring_mask = p->sq_entries - 1;
9394 rings->cq_ring_mask = p->cq_entries - 1;
9395 rings->sq_ring_entries = p->sq_entries;
9396 rings->cq_ring_entries = p->cq_entries;
9397 ctx->sq_mask = rings->sq_ring_mask;
9398 ctx->cq_mask = rings->cq_ring_mask;
9400 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9401 if (size == SIZE_MAX) {
9402 io_mem_free(ctx->rings);
9407 ctx->sq_sqes = io_mem_alloc(size);
9408 if (!ctx->sq_sqes) {
9409 io_mem_free(ctx->rings);
9417 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9421 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9425 ret = io_uring_add_task_file(ctx);
9430 fd_install(fd, file);
9435 * Allocate an anonymous fd, this is what constitutes the application
9436 * visible backing of an io_uring instance. The application mmaps this
9437 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9438 * we have to tie this fd to a socket for file garbage collection purposes.
9440 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9443 #if defined(CONFIG_UNIX)
9446 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9449 return ERR_PTR(ret);
9452 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9453 O_RDWR | O_CLOEXEC);
9454 #if defined(CONFIG_UNIX)
9456 sock_release(ctx->ring_sock);
9457 ctx->ring_sock = NULL;
9459 ctx->ring_sock->file = file;
9465 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9466 struct io_uring_params __user *params)
9468 struct io_ring_ctx *ctx;
9474 if (entries > IORING_MAX_ENTRIES) {
9475 if (!(p->flags & IORING_SETUP_CLAMP))
9477 entries = IORING_MAX_ENTRIES;
9481 * Use twice as many entries for the CQ ring. It's possible for the
9482 * application to drive a higher depth than the size of the SQ ring,
9483 * since the sqes are only used at submission time. This allows for
9484 * some flexibility in overcommitting a bit. If the application has
9485 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9486 * of CQ ring entries manually.
9488 p->sq_entries = roundup_pow_of_two(entries);
9489 if (p->flags & IORING_SETUP_CQSIZE) {
9491 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9492 * to a power-of-two, if it isn't already. We do NOT impose
9493 * any cq vs sq ring sizing.
9497 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9498 if (!(p->flags & IORING_SETUP_CLAMP))
9500 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9502 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9503 if (p->cq_entries < p->sq_entries)
9506 p->cq_entries = 2 * p->sq_entries;
9509 ctx = io_ring_ctx_alloc(p);
9512 ctx->compat = in_compat_syscall();
9513 if (!capable(CAP_IPC_LOCK))
9514 ctx->user = get_uid(current_user());
9517 * This is just grabbed for accounting purposes. When a process exits,
9518 * the mm is exited and dropped before the files, hence we need to hang
9519 * on to this mm purely for the purposes of being able to unaccount
9520 * memory (locked/pinned vm). It's not used for anything else.
9522 mmgrab(current->mm);
9523 ctx->mm_account = current->mm;
9525 ret = io_allocate_scq_urings(ctx, p);
9529 ret = io_sq_offload_create(ctx, p);
9533 memset(&p->sq_off, 0, sizeof(p->sq_off));
9534 p->sq_off.head = offsetof(struct io_rings, sq.head);
9535 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9536 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9537 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9538 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9539 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9540 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9542 memset(&p->cq_off, 0, sizeof(p->cq_off));
9543 p->cq_off.head = offsetof(struct io_rings, cq.head);
9544 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9545 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9546 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9547 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9548 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9549 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9551 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9552 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9553 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9554 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9555 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9557 if (copy_to_user(params, p, sizeof(*p))) {
9562 file = io_uring_get_file(ctx);
9564 ret = PTR_ERR(file);
9569 * Install ring fd as the very last thing, so we don't risk someone
9570 * having closed it before we finish setup
9572 ret = io_uring_install_fd(ctx, file);
9574 /* fput will clean it up */
9579 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9582 io_ring_ctx_wait_and_kill(ctx);
9587 * Sets up an aio uring context, and returns the fd. Applications asks for a
9588 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9589 * params structure passed in.
9591 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9593 struct io_uring_params p;
9596 if (copy_from_user(&p, params, sizeof(p)))
9598 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9603 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9604 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9605 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9606 IORING_SETUP_R_DISABLED))
9609 return io_uring_create(entries, &p, params);
9612 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9613 struct io_uring_params __user *, params)
9615 return io_uring_setup(entries, params);
9618 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9620 struct io_uring_probe *p;
9624 size = struct_size(p, ops, nr_args);
9625 if (size == SIZE_MAX)
9627 p = kzalloc(size, GFP_KERNEL);
9632 if (copy_from_user(p, arg, size))
9635 if (memchr_inv(p, 0, size))
9638 p->last_op = IORING_OP_LAST - 1;
9639 if (nr_args > IORING_OP_LAST)
9640 nr_args = IORING_OP_LAST;
9642 for (i = 0; i < nr_args; i++) {
9644 if (!io_op_defs[i].not_supported)
9645 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9650 if (copy_to_user(arg, p, size))
9657 static int io_register_personality(struct io_ring_ctx *ctx)
9659 const struct cred *creds;
9663 creds = get_current_cred();
9665 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9666 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9673 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9674 unsigned int nr_args)
9676 struct io_uring_restriction *res;
9680 /* Restrictions allowed only if rings started disabled */
9681 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9684 /* We allow only a single restrictions registration */
9685 if (ctx->restrictions.registered)
9688 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9691 size = array_size(nr_args, sizeof(*res));
9692 if (size == SIZE_MAX)
9695 res = memdup_user(arg, size);
9697 return PTR_ERR(res);
9701 for (i = 0; i < nr_args; i++) {
9702 switch (res[i].opcode) {
9703 case IORING_RESTRICTION_REGISTER_OP:
9704 if (res[i].register_op >= IORING_REGISTER_LAST) {
9709 __set_bit(res[i].register_op,
9710 ctx->restrictions.register_op);
9712 case IORING_RESTRICTION_SQE_OP:
9713 if (res[i].sqe_op >= IORING_OP_LAST) {
9718 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9720 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9721 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9723 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9724 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9733 /* Reset all restrictions if an error happened */
9735 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9737 ctx->restrictions.registered = true;
9743 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9745 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9748 if (ctx->restrictions.registered)
9749 ctx->restricted = 1;
9751 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9752 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9753 wake_up(&ctx->sq_data->wait);
9757 static bool io_register_op_must_quiesce(int op)
9760 case IORING_UNREGISTER_FILES:
9761 case IORING_REGISTER_FILES_UPDATE:
9762 case IORING_REGISTER_PROBE:
9763 case IORING_REGISTER_PERSONALITY:
9764 case IORING_UNREGISTER_PERSONALITY:
9771 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9772 void __user *arg, unsigned nr_args)
9773 __releases(ctx->uring_lock)
9774 __acquires(ctx->uring_lock)
9779 * We're inside the ring mutex, if the ref is already dying, then
9780 * someone else killed the ctx or is already going through
9781 * io_uring_register().
9783 if (percpu_ref_is_dying(&ctx->refs))
9786 if (io_register_op_must_quiesce(opcode)) {
9787 percpu_ref_kill(&ctx->refs);
9790 * Drop uring mutex before waiting for references to exit. If
9791 * another thread is currently inside io_uring_enter() it might
9792 * need to grab the uring_lock to make progress. If we hold it
9793 * here across the drain wait, then we can deadlock. It's safe
9794 * to drop the mutex here, since no new references will come in
9795 * after we've killed the percpu ref.
9797 mutex_unlock(&ctx->uring_lock);
9799 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9802 ret = io_run_task_work_sig();
9807 mutex_lock(&ctx->uring_lock);
9810 percpu_ref_resurrect(&ctx->refs);
9815 if (ctx->restricted) {
9816 if (opcode >= IORING_REGISTER_LAST) {
9821 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9828 case IORING_REGISTER_BUFFERS:
9829 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9831 case IORING_UNREGISTER_BUFFERS:
9835 ret = io_sqe_buffers_unregister(ctx);
9837 case IORING_REGISTER_FILES:
9838 ret = io_sqe_files_register(ctx, arg, nr_args);
9840 case IORING_UNREGISTER_FILES:
9844 ret = io_sqe_files_unregister(ctx);
9846 case IORING_REGISTER_FILES_UPDATE:
9847 ret = io_sqe_files_update(ctx, arg, nr_args);
9849 case IORING_REGISTER_EVENTFD:
9850 case IORING_REGISTER_EVENTFD_ASYNC:
9854 ret = io_eventfd_register(ctx, arg);
9857 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9858 ctx->eventfd_async = 1;
9860 ctx->eventfd_async = 0;
9862 case IORING_UNREGISTER_EVENTFD:
9866 ret = io_eventfd_unregister(ctx);
9868 case IORING_REGISTER_PROBE:
9870 if (!arg || nr_args > 256)
9872 ret = io_probe(ctx, arg, nr_args);
9874 case IORING_REGISTER_PERSONALITY:
9878 ret = io_register_personality(ctx);
9880 case IORING_UNREGISTER_PERSONALITY:
9884 ret = io_unregister_personality(ctx, nr_args);
9886 case IORING_REGISTER_ENABLE_RINGS:
9890 ret = io_register_enable_rings(ctx);
9892 case IORING_REGISTER_RESTRICTIONS:
9893 ret = io_register_restrictions(ctx, arg, nr_args);
9901 if (io_register_op_must_quiesce(opcode)) {
9902 /* bring the ctx back to life */
9903 percpu_ref_reinit(&ctx->refs);
9905 reinit_completion(&ctx->ref_comp);
9910 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9911 void __user *, arg, unsigned int, nr_args)
9913 struct io_ring_ctx *ctx;
9922 if (f.file->f_op != &io_uring_fops)
9925 ctx = f.file->private_data;
9929 mutex_lock(&ctx->uring_lock);
9930 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9931 mutex_unlock(&ctx->uring_lock);
9932 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9933 ctx->cq_ev_fd != NULL, ret);
9939 static int __init io_uring_init(void)
9941 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9942 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9943 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9946 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9947 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9948 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9949 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9950 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9951 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9952 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9953 BUILD_BUG_SQE_ELEM(8, __u64, off);
9954 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9955 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9956 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9957 BUILD_BUG_SQE_ELEM(24, __u32, len);
9958 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9959 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9960 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9961 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9962 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9963 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9964 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9965 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9966 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9967 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9968 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9969 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9970 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9971 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9972 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9973 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9974 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9975 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9976 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9978 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9979 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9980 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9984 __initcall(io_uring_init);