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 /* do prep async if is going to be punted */
853 unsigned needs_async_setup : 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_setup = 1,
869 .async_size = sizeof(struct io_async_rw),
871 [IORING_OP_WRITEV] = {
874 .unbound_nonreg_file = 1,
876 .needs_async_setup = 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_setup = 1,
911 .async_size = sizeof(struct io_async_msghdr),
913 [IORING_OP_RECVMSG] = {
915 .unbound_nonreg_file = 1,
918 .needs_async_setup = 1,
919 .async_size = sizeof(struct io_async_msghdr),
921 [IORING_OP_TIMEOUT] = {
922 .async_size = sizeof(struct io_timeout_data),
924 [IORING_OP_TIMEOUT_REMOVE] = {
925 /* used by timeout updates' prep() */
927 [IORING_OP_ACCEPT] = {
929 .unbound_nonreg_file = 1,
932 [IORING_OP_ASYNC_CANCEL] = {},
933 [IORING_OP_LINK_TIMEOUT] = {
934 .async_size = sizeof(struct io_timeout_data),
936 [IORING_OP_CONNECT] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
941 .async_size = sizeof(struct io_async_connect),
943 [IORING_OP_FALLOCATE] = {
946 [IORING_OP_OPENAT] = {},
947 [IORING_OP_CLOSE] = {},
948 [IORING_OP_FILES_UPDATE] = {},
949 [IORING_OP_STATX] = {},
952 .unbound_nonreg_file = 1,
956 .async_size = sizeof(struct io_async_rw),
958 [IORING_OP_WRITE] = {
960 .unbound_nonreg_file = 1,
963 .async_size = sizeof(struct io_async_rw),
965 [IORING_OP_FADVISE] = {
968 [IORING_OP_MADVISE] = {},
971 .unbound_nonreg_file = 1,
976 .unbound_nonreg_file = 1,
980 [IORING_OP_OPENAT2] = {
982 [IORING_OP_EPOLL_CTL] = {
983 .unbound_nonreg_file = 1,
985 [IORING_OP_SPLICE] = {
988 .unbound_nonreg_file = 1,
990 [IORING_OP_PROVIDE_BUFFERS] = {},
991 [IORING_OP_REMOVE_BUFFERS] = {},
995 .unbound_nonreg_file = 1,
997 [IORING_OP_SHUTDOWN] = {
1000 [IORING_OP_RENAMEAT] = {},
1001 [IORING_OP_UNLINKAT] = {},
1004 static bool io_disarm_next(struct io_kiocb *req);
1005 static void io_uring_del_task_file(unsigned long index);
1006 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1007 struct task_struct *task,
1008 struct files_struct *files);
1009 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1010 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1011 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1012 struct io_ring_ctx *ctx);
1013 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1015 static bool io_rw_reissue(struct io_kiocb *req);
1016 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1017 static void io_put_req(struct io_kiocb *req);
1018 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1019 static void io_double_put_req(struct io_kiocb *req);
1020 static void io_dismantle_req(struct io_kiocb *req);
1021 static void io_put_task(struct task_struct *task, int nr);
1022 static void io_queue_next(struct io_kiocb *req);
1023 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1024 static void __io_queue_linked_timeout(struct io_kiocb *req);
1025 static void io_queue_linked_timeout(struct io_kiocb *req);
1026 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1027 struct io_uring_rsrc_update *ip,
1029 static void __io_clean_op(struct io_kiocb *req);
1030 static struct file *io_file_get(struct io_submit_state *state,
1031 struct io_kiocb *req, int fd, bool fixed);
1032 static void __io_queue_sqe(struct io_kiocb *req);
1033 static void io_rsrc_put_work(struct work_struct *work);
1035 static void io_req_task_queue(struct io_kiocb *req);
1036 static void io_submit_flush_completions(struct io_comp_state *cs,
1037 struct io_ring_ctx *ctx);
1038 static int io_req_prep_async(struct io_kiocb *req);
1040 static struct kmem_cache *req_cachep;
1042 static const struct file_operations io_uring_fops;
1044 struct sock *io_uring_get_socket(struct file *file)
1046 #if defined(CONFIG_UNIX)
1047 if (file->f_op == &io_uring_fops) {
1048 struct io_ring_ctx *ctx = file->private_data;
1050 return ctx->ring_sock->sk;
1055 EXPORT_SYMBOL(io_uring_get_socket);
1057 #define io_for_each_link(pos, head) \
1058 for (pos = (head); pos; pos = pos->link)
1060 static inline void io_clean_op(struct io_kiocb *req)
1062 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1066 static inline void io_set_resource_node(struct io_kiocb *req)
1068 struct io_ring_ctx *ctx = req->ctx;
1070 if (!req->fixed_rsrc_refs) {
1071 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1072 percpu_ref_get(req->fixed_rsrc_refs);
1076 static bool io_match_task(struct io_kiocb *head,
1077 struct task_struct *task,
1078 struct files_struct *files)
1080 struct io_kiocb *req;
1082 if (task && head->task != task) {
1083 /* in terms of cancelation, always match if req task is dead */
1084 if (head->task->flags & PF_EXITING)
1091 io_for_each_link(req, head) {
1092 if (req->flags & REQ_F_INFLIGHT)
1098 static inline void req_set_fail_links(struct io_kiocb *req)
1100 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1101 req->flags |= REQ_F_FAIL_LINK;
1104 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1106 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1108 complete(&ctx->ref_comp);
1111 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1113 return !req->timeout.off;
1116 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1118 struct io_ring_ctx *ctx;
1121 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1126 * Use 5 bits less than the max cq entries, that should give us around
1127 * 32 entries per hash list if totally full and uniformly spread.
1129 hash_bits = ilog2(p->cq_entries);
1133 ctx->cancel_hash_bits = hash_bits;
1134 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1136 if (!ctx->cancel_hash)
1138 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1140 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1141 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1144 ctx->flags = p->flags;
1145 init_waitqueue_head(&ctx->sqo_sq_wait);
1146 INIT_LIST_HEAD(&ctx->sqd_list);
1147 init_waitqueue_head(&ctx->cq_wait);
1148 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1149 init_completion(&ctx->ref_comp);
1150 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1151 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1152 mutex_init(&ctx->uring_lock);
1153 init_waitqueue_head(&ctx->wait);
1154 spin_lock_init(&ctx->completion_lock);
1155 INIT_LIST_HEAD(&ctx->iopoll_list);
1156 INIT_LIST_HEAD(&ctx->defer_list);
1157 INIT_LIST_HEAD(&ctx->timeout_list);
1158 spin_lock_init(&ctx->inflight_lock);
1159 INIT_LIST_HEAD(&ctx->inflight_list);
1160 spin_lock_init(&ctx->rsrc_ref_lock);
1161 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1162 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1163 init_llist_head(&ctx->rsrc_put_llist);
1164 INIT_LIST_HEAD(&ctx->tctx_list);
1165 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1166 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1169 kfree(ctx->cancel_hash);
1174 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1176 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1177 struct io_ring_ctx *ctx = req->ctx;
1179 return seq != ctx->cached_cq_tail
1180 + READ_ONCE(ctx->cached_cq_overflow);
1186 static void io_req_track_inflight(struct io_kiocb *req)
1188 struct io_ring_ctx *ctx = req->ctx;
1190 if (!(req->flags & REQ_F_INFLIGHT)) {
1191 req->flags |= REQ_F_INFLIGHT;
1193 spin_lock_irq(&ctx->inflight_lock);
1194 list_add(&req->inflight_entry, &ctx->inflight_list);
1195 spin_unlock_irq(&ctx->inflight_lock);
1199 static void io_prep_async_work(struct io_kiocb *req)
1201 const struct io_op_def *def = &io_op_defs[req->opcode];
1202 struct io_ring_ctx *ctx = req->ctx;
1204 if (!req->work.creds)
1205 req->work.creds = get_current_cred();
1207 if (req->flags & REQ_F_FORCE_ASYNC)
1208 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1210 if (req->flags & REQ_F_ISREG) {
1211 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1212 io_wq_hash_work(&req->work, file_inode(req->file));
1213 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1214 if (def->unbound_nonreg_file)
1215 req->work.flags |= IO_WQ_WORK_UNBOUND;
1219 static void io_prep_async_link(struct io_kiocb *req)
1221 struct io_kiocb *cur;
1223 io_for_each_link(cur, req)
1224 io_prep_async_work(cur);
1227 static void io_queue_async_work(struct io_kiocb *req)
1229 struct io_ring_ctx *ctx = req->ctx;
1230 struct io_kiocb *link = io_prep_linked_timeout(req);
1231 struct io_uring_task *tctx = req->task->io_uring;
1234 BUG_ON(!tctx->io_wq);
1236 /* init ->work of the whole link before punting */
1237 io_prep_async_link(req);
1238 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1239 &req->work, req->flags);
1240 io_wq_enqueue(tctx->io_wq, &req->work);
1242 io_queue_linked_timeout(link);
1245 static void io_kill_timeout(struct io_kiocb *req, int status)
1247 struct io_timeout_data *io = req->async_data;
1250 ret = hrtimer_try_to_cancel(&io->timer);
1252 atomic_set(&req->ctx->cq_timeouts,
1253 atomic_read(&req->ctx->cq_timeouts) + 1);
1254 list_del_init(&req->timeout.list);
1255 io_cqring_fill_event(req, status);
1256 io_put_req_deferred(req, 1);
1260 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1263 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1264 struct io_defer_entry, list);
1266 if (req_need_defer(de->req, de->seq))
1268 list_del_init(&de->list);
1269 io_req_task_queue(de->req);
1271 } while (!list_empty(&ctx->defer_list));
1274 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1278 if (list_empty(&ctx->timeout_list))
1281 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1284 u32 events_needed, events_got;
1285 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1286 struct io_kiocb, timeout.list);
1288 if (io_is_timeout_noseq(req))
1292 * Since seq can easily wrap around over time, subtract
1293 * the last seq at which timeouts were flushed before comparing.
1294 * Assuming not more than 2^31-1 events have happened since,
1295 * these subtractions won't have wrapped, so we can check if
1296 * target is in [last_seq, current_seq] by comparing the two.
1298 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1299 events_got = seq - ctx->cq_last_tm_flush;
1300 if (events_got < events_needed)
1303 list_del_init(&req->timeout.list);
1304 io_kill_timeout(req, 0);
1305 } while (!list_empty(&ctx->timeout_list));
1307 ctx->cq_last_tm_flush = seq;
1310 static void io_commit_cqring(struct io_ring_ctx *ctx)
1312 io_flush_timeouts(ctx);
1314 /* order cqe stores with ring update */
1315 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1317 if (unlikely(!list_empty(&ctx->defer_list)))
1318 __io_queue_deferred(ctx);
1321 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1323 struct io_rings *r = ctx->rings;
1325 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1328 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1330 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1333 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1335 struct io_rings *rings = ctx->rings;
1339 * writes to the cq entry need to come after reading head; the
1340 * control dependency is enough as we're using WRITE_ONCE to
1343 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1346 tail = ctx->cached_cq_tail++;
1347 return &rings->cqes[tail & ctx->cq_mask];
1350 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1354 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1356 if (!ctx->eventfd_async)
1358 return io_wq_current_is_worker();
1361 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1363 /* see waitqueue_active() comment */
1366 if (waitqueue_active(&ctx->wait))
1367 wake_up(&ctx->wait);
1368 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1369 wake_up(&ctx->sq_data->wait);
1370 if (io_should_trigger_evfd(ctx))
1371 eventfd_signal(ctx->cq_ev_fd, 1);
1372 if (waitqueue_active(&ctx->cq_wait)) {
1373 wake_up_interruptible(&ctx->cq_wait);
1374 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1378 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1380 /* see waitqueue_active() comment */
1383 if (ctx->flags & IORING_SETUP_SQPOLL) {
1384 if (waitqueue_active(&ctx->wait))
1385 wake_up(&ctx->wait);
1387 if (io_should_trigger_evfd(ctx))
1388 eventfd_signal(ctx->cq_ev_fd, 1);
1389 if (waitqueue_active(&ctx->cq_wait)) {
1390 wake_up_interruptible(&ctx->cq_wait);
1391 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1395 /* Returns true if there are no backlogged entries after the flush */
1396 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1397 struct task_struct *tsk,
1398 struct files_struct *files)
1400 struct io_rings *rings = ctx->rings;
1401 struct io_kiocb *req, *tmp;
1402 struct io_uring_cqe *cqe;
1403 unsigned long flags;
1404 bool all_flushed, posted;
1407 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1411 spin_lock_irqsave(&ctx->completion_lock, flags);
1412 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1413 if (!io_match_task(req, tsk, files))
1416 cqe = io_get_cqring(ctx);
1420 list_move(&req->compl.list, &list);
1422 WRITE_ONCE(cqe->user_data, req->user_data);
1423 WRITE_ONCE(cqe->res, req->result);
1424 WRITE_ONCE(cqe->flags, req->compl.cflags);
1426 ctx->cached_cq_overflow++;
1427 WRITE_ONCE(ctx->rings->cq_overflow,
1428 ctx->cached_cq_overflow);
1433 all_flushed = list_empty(&ctx->cq_overflow_list);
1435 clear_bit(0, &ctx->sq_check_overflow);
1436 clear_bit(0, &ctx->cq_check_overflow);
1437 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1441 io_commit_cqring(ctx);
1442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1444 io_cqring_ev_posted(ctx);
1446 while (!list_empty(&list)) {
1447 req = list_first_entry(&list, struct io_kiocb, compl.list);
1448 list_del(&req->compl.list);
1455 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1456 struct task_struct *tsk,
1457 struct files_struct *files)
1461 if (test_bit(0, &ctx->cq_check_overflow)) {
1462 /* iopoll syncs against uring_lock, not completion_lock */
1463 if (ctx->flags & IORING_SETUP_IOPOLL)
1464 mutex_lock(&ctx->uring_lock);
1465 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1466 if (ctx->flags & IORING_SETUP_IOPOLL)
1467 mutex_unlock(&ctx->uring_lock);
1474 * Shamelessly stolen from the mm implementation of page reference checking,
1475 * see commit f958d7b528b1 for details.
1477 #define req_ref_zero_or_close_to_overflow(req) \
1478 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1480 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1482 return atomic_inc_not_zero(&req->refs);
1485 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1487 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1488 return atomic_sub_and_test(refs, &req->refs);
1491 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1493 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1494 return atomic_dec_and_test(&req->refs);
1497 static inline void req_ref_put(struct io_kiocb *req)
1499 WARN_ON_ONCE(req_ref_put_and_test(req));
1502 static inline void req_ref_get(struct io_kiocb *req)
1504 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1505 atomic_inc(&req->refs);
1508 static void __io_cqring_fill_event(struct io_kiocb *req, long res,
1509 unsigned int cflags)
1511 struct io_ring_ctx *ctx = req->ctx;
1512 struct io_uring_cqe *cqe;
1514 trace_io_uring_complete(ctx, req->user_data, res);
1517 * If we can't get a cq entry, userspace overflowed the
1518 * submission (by quite a lot). Increment the overflow count in
1521 cqe = io_get_cqring(ctx);
1523 WRITE_ONCE(cqe->user_data, req->user_data);
1524 WRITE_ONCE(cqe->res, res);
1525 WRITE_ONCE(cqe->flags, cflags);
1526 } else if (ctx->cq_overflow_flushed ||
1527 atomic_read(&req->task->io_uring->in_idle)) {
1529 * If we're in ring overflow flush mode, or in task cancel mode,
1530 * then we cannot store the request for later flushing, we need
1531 * to drop it on the floor.
1533 ctx->cached_cq_overflow++;
1534 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1536 if (list_empty(&ctx->cq_overflow_list)) {
1537 set_bit(0, &ctx->sq_check_overflow);
1538 set_bit(0, &ctx->cq_check_overflow);
1539 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1543 req->compl.cflags = cflags;
1545 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1549 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1551 __io_cqring_fill_event(req, res, 0);
1554 static void io_req_complete_post(struct io_kiocb *req, long res,
1555 unsigned int cflags)
1557 struct io_ring_ctx *ctx = req->ctx;
1558 unsigned long flags;
1560 spin_lock_irqsave(&ctx->completion_lock, flags);
1561 __io_cqring_fill_event(req, res, cflags);
1563 * If we're the last reference to this request, add to our locked
1566 if (req_ref_put_and_test(req)) {
1567 struct io_comp_state *cs = &ctx->submit_state.comp;
1569 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1570 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1571 io_disarm_next(req);
1573 io_req_task_queue(req->link);
1577 io_dismantle_req(req);
1578 io_put_task(req->task, 1);
1579 list_add(&req->compl.list, &cs->locked_free_list);
1580 cs->locked_free_nr++;
1582 if (!percpu_ref_tryget(&ctx->refs))
1585 io_commit_cqring(ctx);
1586 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1589 io_cqring_ev_posted(ctx);
1590 percpu_ref_put(&ctx->refs);
1594 static void io_req_complete_state(struct io_kiocb *req, long res,
1595 unsigned int cflags)
1599 req->compl.cflags = cflags;
1600 req->flags |= REQ_F_COMPLETE_INLINE;
1603 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1604 long res, unsigned cflags)
1606 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1607 io_req_complete_state(req, res, cflags);
1609 io_req_complete_post(req, res, cflags);
1612 static inline void io_req_complete(struct io_kiocb *req, long res)
1614 __io_req_complete(req, 0, res, 0);
1617 static void io_req_complete_failed(struct io_kiocb *req, long res)
1619 req_set_fail_links(req);
1621 io_req_complete_post(req, res, 0);
1624 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1626 struct io_submit_state *state = &ctx->submit_state;
1627 struct io_comp_state *cs = &state->comp;
1628 struct io_kiocb *req = NULL;
1631 * If we have more than a batch's worth of requests in our IRQ side
1632 * locked cache, grab the lock and move them over to our submission
1635 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1636 spin_lock_irq(&ctx->completion_lock);
1637 list_splice_init(&cs->locked_free_list, &cs->free_list);
1638 cs->locked_free_nr = 0;
1639 spin_unlock_irq(&ctx->completion_lock);
1642 while (!list_empty(&cs->free_list)) {
1643 req = list_first_entry(&cs->free_list, struct io_kiocb,
1645 list_del(&req->compl.list);
1646 state->reqs[state->free_reqs++] = req;
1647 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1654 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1656 struct io_submit_state *state = &ctx->submit_state;
1658 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1660 if (!state->free_reqs) {
1661 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1664 if (io_flush_cached_reqs(ctx))
1667 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1671 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1672 * retry single alloc to be on the safe side.
1674 if (unlikely(ret <= 0)) {
1675 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1676 if (!state->reqs[0])
1680 state->free_reqs = ret;
1684 return state->reqs[state->free_reqs];
1687 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1694 static void io_dismantle_req(struct io_kiocb *req)
1698 if (req->async_data)
1699 kfree(req->async_data);
1701 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1702 if (req->fixed_rsrc_refs)
1703 percpu_ref_put(req->fixed_rsrc_refs);
1704 if (req->work.creds) {
1705 put_cred(req->work.creds);
1706 req->work.creds = NULL;
1709 if (req->flags & REQ_F_INFLIGHT) {
1710 struct io_ring_ctx *ctx = req->ctx;
1711 unsigned long flags;
1713 spin_lock_irqsave(&ctx->inflight_lock, flags);
1714 list_del(&req->inflight_entry);
1715 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1716 req->flags &= ~REQ_F_INFLIGHT;
1720 /* must to be called somewhat shortly after putting a request */
1721 static inline void io_put_task(struct task_struct *task, int nr)
1723 struct io_uring_task *tctx = task->io_uring;
1725 percpu_counter_sub(&tctx->inflight, nr);
1726 if (unlikely(atomic_read(&tctx->in_idle)))
1727 wake_up(&tctx->wait);
1728 put_task_struct_many(task, nr);
1731 static void __io_free_req(struct io_kiocb *req)
1733 struct io_ring_ctx *ctx = req->ctx;
1735 io_dismantle_req(req);
1736 io_put_task(req->task, 1);
1738 kmem_cache_free(req_cachep, req);
1739 percpu_ref_put(&ctx->refs);
1742 static inline void io_remove_next_linked(struct io_kiocb *req)
1744 struct io_kiocb *nxt = req->link;
1746 req->link = nxt->link;
1750 static bool io_kill_linked_timeout(struct io_kiocb *req)
1751 __must_hold(&req->ctx->completion_lock)
1753 struct io_kiocb *link = req->link;
1754 bool cancelled = false;
1757 * Can happen if a linked timeout fired and link had been like
1758 * req -> link t-out -> link t-out [-> ...]
1760 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1761 struct io_timeout_data *io = link->async_data;
1764 io_remove_next_linked(req);
1765 link->timeout.head = NULL;
1766 ret = hrtimer_try_to_cancel(&io->timer);
1768 io_cqring_fill_event(link, -ECANCELED);
1769 io_put_req_deferred(link, 1);
1773 req->flags &= ~REQ_F_LINK_TIMEOUT;
1777 static void io_fail_links(struct io_kiocb *req)
1778 __must_hold(&req->ctx->completion_lock)
1780 struct io_kiocb *nxt, *link = req->link;
1787 trace_io_uring_fail_link(req, link);
1788 io_cqring_fill_event(link, -ECANCELED);
1789 io_put_req_deferred(link, 2);
1794 static bool io_disarm_next(struct io_kiocb *req)
1795 __must_hold(&req->ctx->completion_lock)
1797 bool posted = false;
1799 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1800 posted = io_kill_linked_timeout(req);
1801 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1802 posted |= (req->link != NULL);
1808 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1810 struct io_kiocb *nxt;
1813 * If LINK is set, we have dependent requests in this chain. If we
1814 * didn't fail this request, queue the first one up, moving any other
1815 * dependencies to the next request. In case of failure, fail the rest
1818 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1819 struct io_ring_ctx *ctx = req->ctx;
1820 unsigned long flags;
1823 spin_lock_irqsave(&ctx->completion_lock, flags);
1824 posted = io_disarm_next(req);
1826 io_commit_cqring(req->ctx);
1827 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1829 io_cqring_ev_posted(ctx);
1836 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1838 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1840 return __io_req_find_next(req);
1843 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1847 if (ctx->submit_state.comp.nr) {
1848 mutex_lock(&ctx->uring_lock);
1849 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1850 mutex_unlock(&ctx->uring_lock);
1852 percpu_ref_put(&ctx->refs);
1855 static bool __tctx_task_work(struct io_uring_task *tctx)
1857 struct io_ring_ctx *ctx = NULL;
1858 struct io_wq_work_list list;
1859 struct io_wq_work_node *node;
1861 if (wq_list_empty(&tctx->task_list))
1864 spin_lock_irq(&tctx->task_lock);
1865 list = tctx->task_list;
1866 INIT_WQ_LIST(&tctx->task_list);
1867 spin_unlock_irq(&tctx->task_lock);
1871 struct io_wq_work_node *next = node->next;
1872 struct io_kiocb *req;
1874 req = container_of(node, struct io_kiocb, io_task_work.node);
1875 if (req->ctx != ctx) {
1876 ctx_flush_and_put(ctx);
1878 percpu_ref_get(&ctx->refs);
1881 req->task_work.func(&req->task_work);
1885 ctx_flush_and_put(ctx);
1886 return list.first != NULL;
1889 static void tctx_task_work(struct callback_head *cb)
1891 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1893 clear_bit(0, &tctx->task_state);
1895 while (__tctx_task_work(tctx))
1899 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1900 enum task_work_notify_mode notify)
1902 struct io_uring_task *tctx = tsk->io_uring;
1903 struct io_wq_work_node *node, *prev;
1904 unsigned long flags;
1907 WARN_ON_ONCE(!tctx);
1909 spin_lock_irqsave(&tctx->task_lock, flags);
1910 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1911 spin_unlock_irqrestore(&tctx->task_lock, flags);
1913 /* task_work already pending, we're done */
1914 if (test_bit(0, &tctx->task_state) ||
1915 test_and_set_bit(0, &tctx->task_state))
1918 if (!task_work_add(tsk, &tctx->task_work, notify))
1922 * Slow path - we failed, find and delete work. if the work is not
1923 * in the list, it got run and we're fine.
1926 spin_lock_irqsave(&tctx->task_lock, flags);
1927 wq_list_for_each(node, prev, &tctx->task_list) {
1928 if (&req->io_task_work.node == node) {
1929 wq_list_del(&tctx->task_list, node, prev);
1934 spin_unlock_irqrestore(&tctx->task_lock, flags);
1935 clear_bit(0, &tctx->task_state);
1939 static int io_req_task_work_add(struct io_kiocb *req)
1941 struct task_struct *tsk = req->task;
1942 struct io_ring_ctx *ctx = req->ctx;
1943 enum task_work_notify_mode notify;
1946 if (tsk->flags & PF_EXITING)
1950 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1951 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1952 * processing task_work. There's no reliable way to tell if TWA_RESUME
1956 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1957 notify = TWA_SIGNAL;
1959 ret = io_task_work_add(tsk, req, notify);
1961 wake_up_process(tsk);
1966 static bool io_run_task_work_head(struct callback_head **work_head)
1968 struct callback_head *work, *next;
1969 bool executed = false;
1972 work = xchg(work_head, NULL);
1988 static void io_task_work_add_head(struct callback_head **work_head,
1989 struct callback_head *task_work)
1991 struct callback_head *head;
1994 head = READ_ONCE(*work_head);
1995 task_work->next = head;
1996 } while (cmpxchg(work_head, head, task_work) != head);
1999 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2000 task_work_func_t cb)
2002 init_task_work(&req->task_work, cb);
2003 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2006 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2008 struct io_ring_ctx *ctx = req->ctx;
2010 spin_lock_irq(&ctx->completion_lock);
2011 io_cqring_fill_event(req, error);
2012 io_commit_cqring(ctx);
2013 spin_unlock_irq(&ctx->completion_lock);
2015 io_cqring_ev_posted(ctx);
2016 req_set_fail_links(req);
2017 io_double_put_req(req);
2020 static void io_req_task_cancel(struct callback_head *cb)
2022 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2023 struct io_ring_ctx *ctx = req->ctx;
2025 /* ctx is guaranteed to stay alive while we hold uring_lock */
2026 mutex_lock(&ctx->uring_lock);
2027 __io_req_task_cancel(req, req->result);
2028 mutex_unlock(&ctx->uring_lock);
2031 static void __io_req_task_submit(struct io_kiocb *req)
2033 struct io_ring_ctx *ctx = req->ctx;
2035 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2036 mutex_lock(&ctx->uring_lock);
2037 if (!(current->flags & PF_EXITING) && !current->in_execve)
2038 __io_queue_sqe(req);
2040 __io_req_task_cancel(req, -EFAULT);
2041 mutex_unlock(&ctx->uring_lock);
2044 static void io_req_task_submit(struct callback_head *cb)
2046 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2048 __io_req_task_submit(req);
2051 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2054 req->task_work.func = io_req_task_cancel;
2056 if (unlikely(io_req_task_work_add(req)))
2057 io_req_task_work_add_fallback(req, io_req_task_cancel);
2060 static void io_req_task_queue(struct io_kiocb *req)
2062 req->task_work.func = io_req_task_submit;
2064 if (unlikely(io_req_task_work_add(req)))
2065 io_req_task_queue_fail(req, -ECANCELED);
2068 static inline void io_queue_next(struct io_kiocb *req)
2070 struct io_kiocb *nxt = io_req_find_next(req);
2073 io_req_task_queue(nxt);
2076 static void io_free_req(struct io_kiocb *req)
2083 struct task_struct *task;
2088 static inline void io_init_req_batch(struct req_batch *rb)
2095 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2096 struct req_batch *rb)
2099 io_put_task(rb->task, rb->task_refs);
2101 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2104 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2105 struct io_submit_state *state)
2109 if (req->task != rb->task) {
2111 io_put_task(rb->task, rb->task_refs);
2112 rb->task = req->task;
2118 io_dismantle_req(req);
2119 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2120 state->reqs[state->free_reqs++] = req;
2122 list_add(&req->compl.list, &state->comp.free_list);
2125 static void io_submit_flush_completions(struct io_comp_state *cs,
2126 struct io_ring_ctx *ctx)
2129 struct io_kiocb *req;
2130 struct req_batch rb;
2132 io_init_req_batch(&rb);
2133 spin_lock_irq(&ctx->completion_lock);
2134 for (i = 0; i < nr; i++) {
2136 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2138 io_commit_cqring(ctx);
2139 spin_unlock_irq(&ctx->completion_lock);
2141 io_cqring_ev_posted(ctx);
2142 for (i = 0; i < nr; i++) {
2145 /* submission and completion refs */
2146 if (req_ref_sub_and_test(req, 2))
2147 io_req_free_batch(&rb, req, &ctx->submit_state);
2150 io_req_free_batch_finish(ctx, &rb);
2155 * Drop reference to request, return next in chain (if there is one) if this
2156 * was the last reference to this request.
2158 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2160 struct io_kiocb *nxt = NULL;
2162 if (req_ref_put_and_test(req)) {
2163 nxt = io_req_find_next(req);
2169 static void io_put_req(struct io_kiocb *req)
2171 if (req_ref_put_and_test(req))
2175 static void io_put_req_deferred_cb(struct callback_head *cb)
2177 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2182 static void io_free_req_deferred(struct io_kiocb *req)
2186 req->task_work.func = io_put_req_deferred_cb;
2187 ret = io_req_task_work_add(req);
2189 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2192 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2194 if (req_ref_sub_and_test(req, refs))
2195 io_free_req_deferred(req);
2198 static void io_double_put_req(struct io_kiocb *req)
2200 /* drop both submit and complete references */
2201 if (req_ref_sub_and_test(req, 2))
2205 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2207 /* See comment at the top of this file */
2209 return __io_cqring_events(ctx);
2212 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2214 struct io_rings *rings = ctx->rings;
2216 /* make sure SQ entry isn't read before tail */
2217 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2220 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2222 unsigned int cflags;
2224 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2225 cflags |= IORING_CQE_F_BUFFER;
2226 req->flags &= ~REQ_F_BUFFER_SELECTED;
2231 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2233 struct io_buffer *kbuf;
2235 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2236 return io_put_kbuf(req, kbuf);
2239 static inline bool io_run_task_work(void)
2242 * Not safe to run on exiting task, and the task_work handling will
2243 * not add work to such a task.
2245 if (unlikely(current->flags & PF_EXITING))
2247 if (current->task_works) {
2248 __set_current_state(TASK_RUNNING);
2257 * Find and free completed poll iocbs
2259 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2260 struct list_head *done)
2262 struct req_batch rb;
2263 struct io_kiocb *req;
2265 /* order with ->result store in io_complete_rw_iopoll() */
2268 io_init_req_batch(&rb);
2269 while (!list_empty(done)) {
2272 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2273 list_del(&req->inflight_entry);
2275 if (READ_ONCE(req->result) == -EAGAIN) {
2276 req->iopoll_completed = 0;
2277 if (io_rw_reissue(req))
2281 if (req->flags & REQ_F_BUFFER_SELECTED)
2282 cflags = io_put_rw_kbuf(req);
2284 __io_cqring_fill_event(req, req->result, cflags);
2287 if (req_ref_put_and_test(req))
2288 io_req_free_batch(&rb, req, &ctx->submit_state);
2291 io_commit_cqring(ctx);
2292 io_cqring_ev_posted_iopoll(ctx);
2293 io_req_free_batch_finish(ctx, &rb);
2296 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2299 struct io_kiocb *req, *tmp;
2305 * Only spin for completions if we don't have multiple devices hanging
2306 * off our complete list, and we're under the requested amount.
2308 spin = !ctx->poll_multi_file && *nr_events < min;
2311 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2312 struct kiocb *kiocb = &req->rw.kiocb;
2315 * Move completed and retryable entries to our local lists.
2316 * If we find a request that requires polling, break out
2317 * and complete those lists first, if we have entries there.
2319 if (READ_ONCE(req->iopoll_completed)) {
2320 list_move_tail(&req->inflight_entry, &done);
2323 if (!list_empty(&done))
2326 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2330 /* iopoll may have completed current req */
2331 if (READ_ONCE(req->iopoll_completed))
2332 list_move_tail(&req->inflight_entry, &done);
2339 if (!list_empty(&done))
2340 io_iopoll_complete(ctx, nr_events, &done);
2346 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2347 * non-spinning poll check - we'll still enter the driver poll loop, but only
2348 * as a non-spinning completion check.
2350 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2353 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2356 ret = io_do_iopoll(ctx, nr_events, min);
2359 if (*nr_events >= min)
2367 * We can't just wait for polled events to come to us, we have to actively
2368 * find and complete them.
2370 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2372 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2375 mutex_lock(&ctx->uring_lock);
2376 while (!list_empty(&ctx->iopoll_list)) {
2377 unsigned int nr_events = 0;
2379 io_do_iopoll(ctx, &nr_events, 0);
2381 /* let it sleep and repeat later if can't complete a request */
2385 * Ensure we allow local-to-the-cpu processing to take place,
2386 * in this case we need to ensure that we reap all events.
2387 * Also let task_work, etc. to progress by releasing the mutex
2389 if (need_resched()) {
2390 mutex_unlock(&ctx->uring_lock);
2392 mutex_lock(&ctx->uring_lock);
2395 mutex_unlock(&ctx->uring_lock);
2398 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2400 unsigned int nr_events = 0;
2401 int iters = 0, ret = 0;
2404 * We disallow the app entering submit/complete with polling, but we
2405 * still need to lock the ring to prevent racing with polled issue
2406 * that got punted to a workqueue.
2408 mutex_lock(&ctx->uring_lock);
2411 * Don't enter poll loop if we already have events pending.
2412 * If we do, we can potentially be spinning for commands that
2413 * already triggered a CQE (eg in error).
2415 if (test_bit(0, &ctx->cq_check_overflow))
2416 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2417 if (io_cqring_events(ctx))
2421 * If a submit got punted to a workqueue, we can have the
2422 * application entering polling for a command before it gets
2423 * issued. That app will hold the uring_lock for the duration
2424 * of the poll right here, so we need to take a breather every
2425 * now and then to ensure that the issue has a chance to add
2426 * the poll to the issued list. Otherwise we can spin here
2427 * forever, while the workqueue is stuck trying to acquire the
2430 if (!(++iters & 7)) {
2431 mutex_unlock(&ctx->uring_lock);
2433 mutex_lock(&ctx->uring_lock);
2436 ret = io_iopoll_getevents(ctx, &nr_events, min);
2440 } while (min && !nr_events && !need_resched());
2442 mutex_unlock(&ctx->uring_lock);
2446 static void kiocb_end_write(struct io_kiocb *req)
2449 * Tell lockdep we inherited freeze protection from submission
2452 if (req->flags & REQ_F_ISREG) {
2453 struct inode *inode = file_inode(req->file);
2455 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2457 file_end_write(req->file);
2461 static bool io_resubmit_prep(struct io_kiocb *req)
2463 /* either already prepared or successfully done */
2464 return req->async_data || !io_req_prep_async(req);
2467 static bool io_rw_should_reissue(struct io_kiocb *req)
2469 umode_t mode = file_inode(req->file)->i_mode;
2470 struct io_ring_ctx *ctx = req->ctx;
2472 if (!S_ISBLK(mode) && !S_ISREG(mode))
2474 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2475 !(ctx->flags & IORING_SETUP_IOPOLL)))
2478 * If ref is dying, we might be running poll reap from the exit work.
2479 * Don't attempt to reissue from that path, just let it fail with
2482 if (percpu_ref_is_dying(&ctx->refs))
2487 static bool io_rw_should_reissue(struct io_kiocb *req)
2493 static bool io_rw_reissue(struct io_kiocb *req)
2496 if (!io_rw_should_reissue(req))
2499 lockdep_assert_held(&req->ctx->uring_lock);
2501 if (io_resubmit_prep(req)) {
2503 io_queue_async_work(req);
2506 req_set_fail_links(req);
2511 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2512 unsigned int issue_flags)
2516 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2517 kiocb_end_write(req);
2518 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2519 req->flags |= REQ_F_REISSUE;
2522 if (res != req->result)
2523 req_set_fail_links(req);
2524 if (req->flags & REQ_F_BUFFER_SELECTED)
2525 cflags = io_put_rw_kbuf(req);
2526 __io_req_complete(req, issue_flags, res, cflags);
2529 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2531 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2533 __io_complete_rw(req, res, res2, 0);
2536 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2538 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2541 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2542 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2543 struct io_async_rw *rw = req->async_data;
2546 iov_iter_revert(&rw->iter,
2547 req->result - iov_iter_count(&rw->iter));
2548 else if (!io_resubmit_prep(req))
2553 if (kiocb->ki_flags & IOCB_WRITE)
2554 kiocb_end_write(req);
2556 if (res != -EAGAIN && res != req->result)
2557 req_set_fail_links(req);
2559 WRITE_ONCE(req->result, res);
2560 /* order with io_poll_complete() checking ->result */
2562 WRITE_ONCE(req->iopoll_completed, 1);
2566 * After the iocb has been issued, it's safe to be found on the poll list.
2567 * Adding the kiocb to the list AFTER submission ensures that we don't
2568 * find it from a io_iopoll_getevents() thread before the issuer is done
2569 * accessing the kiocb cookie.
2571 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2573 struct io_ring_ctx *ctx = req->ctx;
2576 * Track whether we have multiple files in our lists. This will impact
2577 * how we do polling eventually, not spinning if we're on potentially
2578 * different devices.
2580 if (list_empty(&ctx->iopoll_list)) {
2581 ctx->poll_multi_file = false;
2582 } else if (!ctx->poll_multi_file) {
2583 struct io_kiocb *list_req;
2585 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2587 if (list_req->file != req->file)
2588 ctx->poll_multi_file = true;
2592 * For fast devices, IO may have already completed. If it has, add
2593 * it to the front so we find it first.
2595 if (READ_ONCE(req->iopoll_completed))
2596 list_add(&req->inflight_entry, &ctx->iopoll_list);
2598 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2601 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2602 * task context or in io worker task context. If current task context is
2603 * sq thread, we don't need to check whether should wake up sq thread.
2605 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2606 wq_has_sleeper(&ctx->sq_data->wait))
2607 wake_up(&ctx->sq_data->wait);
2610 static inline void io_state_file_put(struct io_submit_state *state)
2612 if (state->file_refs) {
2613 fput_many(state->file, state->file_refs);
2614 state->file_refs = 0;
2619 * Get as many references to a file as we have IOs left in this submission,
2620 * assuming most submissions are for one file, or at least that each file
2621 * has more than one submission.
2623 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2628 if (state->file_refs) {
2629 if (state->fd == fd) {
2633 io_state_file_put(state);
2635 state->file = fget_many(fd, state->ios_left);
2636 if (unlikely(!state->file))
2640 state->file_refs = state->ios_left - 1;
2644 static bool io_bdev_nowait(struct block_device *bdev)
2646 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2650 * If we tracked the file through the SCM inflight mechanism, we could support
2651 * any file. For now, just ensure that anything potentially problematic is done
2654 static bool io_file_supports_async(struct file *file, int rw)
2656 umode_t mode = file_inode(file)->i_mode;
2658 if (S_ISBLK(mode)) {
2659 if (IS_ENABLED(CONFIG_BLOCK) &&
2660 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2664 if (S_ISCHR(mode) || S_ISSOCK(mode))
2666 if (S_ISREG(mode)) {
2667 if (IS_ENABLED(CONFIG_BLOCK) &&
2668 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2669 file->f_op != &io_uring_fops)
2674 /* any ->read/write should understand O_NONBLOCK */
2675 if (file->f_flags & O_NONBLOCK)
2678 if (!(file->f_mode & FMODE_NOWAIT))
2682 return file->f_op->read_iter != NULL;
2684 return file->f_op->write_iter != NULL;
2687 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2689 struct io_ring_ctx *ctx = req->ctx;
2690 struct kiocb *kiocb = &req->rw.kiocb;
2691 struct file *file = req->file;
2695 if (S_ISREG(file_inode(file)->i_mode))
2696 req->flags |= REQ_F_ISREG;
2698 kiocb->ki_pos = READ_ONCE(sqe->off);
2699 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2700 req->flags |= REQ_F_CUR_POS;
2701 kiocb->ki_pos = file->f_pos;
2703 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2704 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2705 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2709 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2710 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2711 req->flags |= REQ_F_NOWAIT;
2713 ioprio = READ_ONCE(sqe->ioprio);
2715 ret = ioprio_check_cap(ioprio);
2719 kiocb->ki_ioprio = ioprio;
2721 kiocb->ki_ioprio = get_current_ioprio();
2723 if (ctx->flags & IORING_SETUP_IOPOLL) {
2724 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2725 !kiocb->ki_filp->f_op->iopoll)
2728 kiocb->ki_flags |= IOCB_HIPRI;
2729 kiocb->ki_complete = io_complete_rw_iopoll;
2730 req->iopoll_completed = 0;
2732 if (kiocb->ki_flags & IOCB_HIPRI)
2734 kiocb->ki_complete = io_complete_rw;
2737 req->rw.addr = READ_ONCE(sqe->addr);
2738 req->rw.len = READ_ONCE(sqe->len);
2739 req->buf_index = READ_ONCE(sqe->buf_index);
2743 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2749 case -ERESTARTNOINTR:
2750 case -ERESTARTNOHAND:
2751 case -ERESTART_RESTARTBLOCK:
2753 * We can't just restart the syscall, since previously
2754 * submitted sqes may already be in progress. Just fail this
2760 kiocb->ki_complete(kiocb, ret, 0);
2764 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2765 unsigned int issue_flags)
2767 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2768 struct io_async_rw *io = req->async_data;
2769 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2771 /* add previously done IO, if any */
2772 if (io && io->bytes_done > 0) {
2774 ret = io->bytes_done;
2776 ret += io->bytes_done;
2779 if (req->flags & REQ_F_CUR_POS)
2780 req->file->f_pos = kiocb->ki_pos;
2781 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2782 __io_complete_rw(req, ret, 0, issue_flags);
2784 io_rw_done(kiocb, ret);
2786 if (check_reissue && req->flags & REQ_F_REISSUE) {
2787 req->flags &= ~REQ_F_REISSUE;
2788 if (!io_rw_reissue(req)) {
2791 req_set_fail_links(req);
2792 if (req->flags & REQ_F_BUFFER_SELECTED)
2793 cflags = io_put_rw_kbuf(req);
2794 __io_req_complete(req, issue_flags, ret, cflags);
2799 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2801 struct io_ring_ctx *ctx = req->ctx;
2802 size_t len = req->rw.len;
2803 struct io_mapped_ubuf *imu;
2804 u16 index, buf_index = req->buf_index;
2808 if (unlikely(buf_index >= ctx->nr_user_bufs))
2810 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2811 imu = &ctx->user_bufs[index];
2812 buf_addr = req->rw.addr;
2815 if (buf_addr + len < buf_addr)
2817 /* not inside the mapped region */
2818 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2822 * May not be a start of buffer, set size appropriately
2823 * and advance us to the beginning.
2825 offset = buf_addr - imu->ubuf;
2826 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2830 * Don't use iov_iter_advance() here, as it's really slow for
2831 * using the latter parts of a big fixed buffer - it iterates
2832 * over each segment manually. We can cheat a bit here, because
2835 * 1) it's a BVEC iter, we set it up
2836 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2837 * first and last bvec
2839 * So just find our index, and adjust the iterator afterwards.
2840 * If the offset is within the first bvec (or the whole first
2841 * bvec, just use iov_iter_advance(). This makes it easier
2842 * since we can just skip the first segment, which may not
2843 * be PAGE_SIZE aligned.
2845 const struct bio_vec *bvec = imu->bvec;
2847 if (offset <= bvec->bv_len) {
2848 iov_iter_advance(iter, offset);
2850 unsigned long seg_skip;
2852 /* skip first vec */
2853 offset -= bvec->bv_len;
2854 seg_skip = 1 + (offset >> PAGE_SHIFT);
2856 iter->bvec = bvec + seg_skip;
2857 iter->nr_segs -= seg_skip;
2858 iter->count -= bvec->bv_len + offset;
2859 iter->iov_offset = offset & ~PAGE_MASK;
2866 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2869 mutex_unlock(&ctx->uring_lock);
2872 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2875 * "Normal" inline submissions always hold the uring_lock, since we
2876 * grab it from the system call. Same is true for the SQPOLL offload.
2877 * The only exception is when we've detached the request and issue it
2878 * from an async worker thread, grab the lock for that case.
2881 mutex_lock(&ctx->uring_lock);
2884 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2885 int bgid, struct io_buffer *kbuf,
2888 struct io_buffer *head;
2890 if (req->flags & REQ_F_BUFFER_SELECTED)
2893 io_ring_submit_lock(req->ctx, needs_lock);
2895 lockdep_assert_held(&req->ctx->uring_lock);
2897 head = xa_load(&req->ctx->io_buffers, bgid);
2899 if (!list_empty(&head->list)) {
2900 kbuf = list_last_entry(&head->list, struct io_buffer,
2902 list_del(&kbuf->list);
2905 xa_erase(&req->ctx->io_buffers, bgid);
2907 if (*len > kbuf->len)
2910 kbuf = ERR_PTR(-ENOBUFS);
2913 io_ring_submit_unlock(req->ctx, needs_lock);
2918 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2921 struct io_buffer *kbuf;
2924 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2925 bgid = req->buf_index;
2926 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2929 req->rw.addr = (u64) (unsigned long) kbuf;
2930 req->flags |= REQ_F_BUFFER_SELECTED;
2931 return u64_to_user_ptr(kbuf->addr);
2934 #ifdef CONFIG_COMPAT
2935 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2938 struct compat_iovec __user *uiov;
2939 compat_ssize_t clen;
2943 uiov = u64_to_user_ptr(req->rw.addr);
2944 if (!access_ok(uiov, sizeof(*uiov)))
2946 if (__get_user(clen, &uiov->iov_len))
2952 buf = io_rw_buffer_select(req, &len, needs_lock);
2954 return PTR_ERR(buf);
2955 iov[0].iov_base = buf;
2956 iov[0].iov_len = (compat_size_t) len;
2961 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2968 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2971 len = iov[0].iov_len;
2974 buf = io_rw_buffer_select(req, &len, needs_lock);
2976 return PTR_ERR(buf);
2977 iov[0].iov_base = buf;
2978 iov[0].iov_len = len;
2982 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2985 if (req->flags & REQ_F_BUFFER_SELECTED) {
2986 struct io_buffer *kbuf;
2988 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2989 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2990 iov[0].iov_len = kbuf->len;
2993 if (req->rw.len != 1)
2996 #ifdef CONFIG_COMPAT
2997 if (req->ctx->compat)
2998 return io_compat_import(req, iov, needs_lock);
3001 return __io_iov_buffer_select(req, iov, needs_lock);
3004 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3005 struct iov_iter *iter, bool needs_lock)
3007 void __user *buf = u64_to_user_ptr(req->rw.addr);
3008 size_t sqe_len = req->rw.len;
3009 u8 opcode = req->opcode;
3012 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3014 return io_import_fixed(req, rw, iter);
3017 /* buffer index only valid with fixed read/write, or buffer select */
3018 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3021 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3022 if (req->flags & REQ_F_BUFFER_SELECT) {
3023 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3025 return PTR_ERR(buf);
3026 req->rw.len = sqe_len;
3029 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3034 if (req->flags & REQ_F_BUFFER_SELECT) {
3035 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3037 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3042 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3046 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3048 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3052 * For files that don't have ->read_iter() and ->write_iter(), handle them
3053 * by looping over ->read() or ->write() manually.
3055 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3057 struct kiocb *kiocb = &req->rw.kiocb;
3058 struct file *file = req->file;
3062 * Don't support polled IO through this interface, and we can't
3063 * support non-blocking either. For the latter, this just causes
3064 * the kiocb to be handled from an async context.
3066 if (kiocb->ki_flags & IOCB_HIPRI)
3068 if (kiocb->ki_flags & IOCB_NOWAIT)
3071 while (iov_iter_count(iter)) {
3075 if (!iov_iter_is_bvec(iter)) {
3076 iovec = iov_iter_iovec(iter);
3078 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3079 iovec.iov_len = req->rw.len;
3083 nr = file->f_op->read(file, iovec.iov_base,
3084 iovec.iov_len, io_kiocb_ppos(kiocb));
3086 nr = file->f_op->write(file, iovec.iov_base,
3087 iovec.iov_len, io_kiocb_ppos(kiocb));
3096 if (nr != iovec.iov_len)
3100 iov_iter_advance(iter, nr);
3106 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3107 const struct iovec *fast_iov, struct iov_iter *iter)
3109 struct io_async_rw *rw = req->async_data;
3111 memcpy(&rw->iter, iter, sizeof(*iter));
3112 rw->free_iovec = iovec;
3114 /* can only be fixed buffers, no need to do anything */
3115 if (iov_iter_is_bvec(iter))
3118 unsigned iov_off = 0;
3120 rw->iter.iov = rw->fast_iov;
3121 if (iter->iov != fast_iov) {
3122 iov_off = iter->iov - fast_iov;
3123 rw->iter.iov += iov_off;
3125 if (rw->fast_iov != fast_iov)
3126 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3127 sizeof(struct iovec) * iter->nr_segs);
3129 req->flags |= REQ_F_NEED_CLEANUP;
3133 static inline int io_alloc_async_data(struct io_kiocb *req)
3135 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3136 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3137 return req->async_data == NULL;
3140 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3141 const struct iovec *fast_iov,
3142 struct iov_iter *iter, bool force)
3144 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3146 if (!req->async_data) {
3147 if (io_alloc_async_data(req)) {
3152 io_req_map_rw(req, iovec, fast_iov, iter);
3157 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3159 struct io_async_rw *iorw = req->async_data;
3160 struct iovec *iov = iorw->fast_iov;
3163 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3164 if (unlikely(ret < 0))
3167 iorw->bytes_done = 0;
3168 iorw->free_iovec = iov;
3170 req->flags |= REQ_F_NEED_CLEANUP;
3174 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3176 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3178 return io_prep_rw(req, sqe);
3182 * This is our waitqueue callback handler, registered through lock_page_async()
3183 * when we initially tried to do the IO with the iocb armed our waitqueue.
3184 * This gets called when the page is unlocked, and we generally expect that to
3185 * happen when the page IO is completed and the page is now uptodate. This will
3186 * queue a task_work based retry of the operation, attempting to copy the data
3187 * again. If the latter fails because the page was NOT uptodate, then we will
3188 * do a thread based blocking retry of the operation. That's the unexpected
3191 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3192 int sync, void *arg)
3194 struct wait_page_queue *wpq;
3195 struct io_kiocb *req = wait->private;
3196 struct wait_page_key *key = arg;
3198 wpq = container_of(wait, struct wait_page_queue, wait);
3200 if (!wake_page_match(wpq, key))
3203 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3204 list_del_init(&wait->entry);
3206 /* submit ref gets dropped, acquire a new one */
3208 io_req_task_queue(req);
3213 * This controls whether a given IO request should be armed for async page
3214 * based retry. If we return false here, the request is handed to the async
3215 * worker threads for retry. If we're doing buffered reads on a regular file,
3216 * we prepare a private wait_page_queue entry and retry the operation. This
3217 * will either succeed because the page is now uptodate and unlocked, or it
3218 * will register a callback when the page is unlocked at IO completion. Through
3219 * that callback, io_uring uses task_work to setup a retry of the operation.
3220 * That retry will attempt the buffered read again. The retry will generally
3221 * succeed, or in rare cases where it fails, we then fall back to using the
3222 * async worker threads for a blocking retry.
3224 static bool io_rw_should_retry(struct io_kiocb *req)
3226 struct io_async_rw *rw = req->async_data;
3227 struct wait_page_queue *wait = &rw->wpq;
3228 struct kiocb *kiocb = &req->rw.kiocb;
3230 /* never retry for NOWAIT, we just complete with -EAGAIN */
3231 if (req->flags & REQ_F_NOWAIT)
3234 /* Only for buffered IO */
3235 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3239 * just use poll if we can, and don't attempt if the fs doesn't
3240 * support callback based unlocks
3242 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3245 wait->wait.func = io_async_buf_func;
3246 wait->wait.private = req;
3247 wait->wait.flags = 0;
3248 INIT_LIST_HEAD(&wait->wait.entry);
3249 kiocb->ki_flags |= IOCB_WAITQ;
3250 kiocb->ki_flags &= ~IOCB_NOWAIT;
3251 kiocb->ki_waitq = wait;
3255 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3257 if (req->file->f_op->read_iter)
3258 return call_read_iter(req->file, &req->rw.kiocb, iter);
3259 else if (req->file->f_op->read)
3260 return loop_rw_iter(READ, req, iter);
3265 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3267 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct iov_iter __iter, *iter = &__iter;
3270 struct io_async_rw *rw = req->async_data;
3271 ssize_t io_size, ret, ret2;
3272 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3278 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3282 io_size = iov_iter_count(iter);
3283 req->result = io_size;
3285 /* Ensure we clear previously set non-block flag */
3286 if (!force_nonblock)
3287 kiocb->ki_flags &= ~IOCB_NOWAIT;
3289 kiocb->ki_flags |= IOCB_NOWAIT;
3291 /* If the file doesn't support async, just async punt */
3292 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3293 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3294 return ret ?: -EAGAIN;
3297 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3298 if (unlikely(ret)) {
3303 ret = io_iter_do_read(req, iter);
3305 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3306 req->flags &= ~REQ_F_REISSUE;
3307 /* IOPOLL retry should happen for io-wq threads */
3308 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3310 /* no retry on NONBLOCK nor RWF_NOWAIT */
3311 if (req->flags & REQ_F_NOWAIT)
3313 /* some cases will consume bytes even on error returns */
3314 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3316 } else if (ret == -EIOCBQUEUED) {
3318 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3319 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3320 /* read all, failed, already did sync or don't want to retry */
3324 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3329 rw = req->async_data;
3330 /* now use our persistent iterator, if we aren't already */
3335 rw->bytes_done += ret;
3336 /* if we can retry, do so with the callbacks armed */
3337 if (!io_rw_should_retry(req)) {
3338 kiocb->ki_flags &= ~IOCB_WAITQ;
3343 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3344 * we get -EIOCBQUEUED, then we'll get a notification when the
3345 * desired page gets unlocked. We can also get a partial read
3346 * here, and if we do, then just retry at the new offset.
3348 ret = io_iter_do_read(req, iter);
3349 if (ret == -EIOCBQUEUED)
3351 /* we got some bytes, but not all. retry. */
3352 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 } while (ret > 0 && ret < io_size);
3355 kiocb_done(kiocb, ret, issue_flags);
3357 /* it's faster to check here then delegate to kfree */
3363 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3365 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3367 return io_prep_rw(req, sqe);
3370 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3372 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3373 struct kiocb *kiocb = &req->rw.kiocb;
3374 struct iov_iter __iter, *iter = &__iter;
3375 struct io_async_rw *rw = req->async_data;
3376 ssize_t ret, ret2, io_size;
3377 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3383 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3387 io_size = iov_iter_count(iter);
3388 req->result = io_size;
3390 /* Ensure we clear previously set non-block flag */
3391 if (!force_nonblock)
3392 kiocb->ki_flags &= ~IOCB_NOWAIT;
3394 kiocb->ki_flags |= IOCB_NOWAIT;
3396 /* If the file doesn't support async, just async punt */
3397 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3400 /* file path doesn't support NOWAIT for non-direct_IO */
3401 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3402 (req->flags & REQ_F_ISREG))
3405 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3410 * Open-code file_start_write here to grab freeze protection,
3411 * which will be released by another thread in
3412 * io_complete_rw(). Fool lockdep by telling it the lock got
3413 * released so that it doesn't complain about the held lock when
3414 * we return to userspace.
3416 if (req->flags & REQ_F_ISREG) {
3417 sb_start_write(file_inode(req->file)->i_sb);
3418 __sb_writers_release(file_inode(req->file)->i_sb,
3421 kiocb->ki_flags |= IOCB_WRITE;
3423 if (req->file->f_op->write_iter)
3424 ret2 = call_write_iter(req->file, kiocb, iter);
3425 else if (req->file->f_op->write)
3426 ret2 = loop_rw_iter(WRITE, req, iter);
3430 if (req->flags & REQ_F_REISSUE) {
3431 req->flags &= ~REQ_F_REISSUE;
3436 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3437 * retry them without IOCB_NOWAIT.
3439 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3441 /* no retry on NONBLOCK nor RWF_NOWAIT */
3442 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3444 if (!force_nonblock || ret2 != -EAGAIN) {
3445 /* IOPOLL retry should happen for io-wq threads */
3446 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3449 kiocb_done(kiocb, ret2, issue_flags);
3452 /* some cases will consume bytes even on error returns */
3453 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3454 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3455 return ret ?: -EAGAIN;
3458 /* it's reportedly faster than delegating the null check to kfree() */
3464 static int io_renameat_prep(struct io_kiocb *req,
3465 const struct io_uring_sqe *sqe)
3467 struct io_rename *ren = &req->rename;
3468 const char __user *oldf, *newf;
3470 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3473 ren->old_dfd = READ_ONCE(sqe->fd);
3474 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3475 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3476 ren->new_dfd = READ_ONCE(sqe->len);
3477 ren->flags = READ_ONCE(sqe->rename_flags);
3479 ren->oldpath = getname(oldf);
3480 if (IS_ERR(ren->oldpath))
3481 return PTR_ERR(ren->oldpath);
3483 ren->newpath = getname(newf);
3484 if (IS_ERR(ren->newpath)) {
3485 putname(ren->oldpath);
3486 return PTR_ERR(ren->newpath);
3489 req->flags |= REQ_F_NEED_CLEANUP;
3493 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3495 struct io_rename *ren = &req->rename;
3498 if (issue_flags & IO_URING_F_NONBLOCK)
3501 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3502 ren->newpath, ren->flags);
3504 req->flags &= ~REQ_F_NEED_CLEANUP;
3506 req_set_fail_links(req);
3507 io_req_complete(req, ret);
3511 static int io_unlinkat_prep(struct io_kiocb *req,
3512 const struct io_uring_sqe *sqe)
3514 struct io_unlink *un = &req->unlink;
3515 const char __user *fname;
3517 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3520 un->dfd = READ_ONCE(sqe->fd);
3522 un->flags = READ_ONCE(sqe->unlink_flags);
3523 if (un->flags & ~AT_REMOVEDIR)
3526 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3527 un->filename = getname(fname);
3528 if (IS_ERR(un->filename))
3529 return PTR_ERR(un->filename);
3531 req->flags |= REQ_F_NEED_CLEANUP;
3535 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3537 struct io_unlink *un = &req->unlink;
3540 if (issue_flags & IO_URING_F_NONBLOCK)
3543 if (un->flags & AT_REMOVEDIR)
3544 ret = do_rmdir(un->dfd, un->filename);
3546 ret = do_unlinkat(un->dfd, un->filename);
3548 req->flags &= ~REQ_F_NEED_CLEANUP;
3550 req_set_fail_links(req);
3551 io_req_complete(req, ret);
3555 static int io_shutdown_prep(struct io_kiocb *req,
3556 const struct io_uring_sqe *sqe)
3558 #if defined(CONFIG_NET)
3559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3561 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3565 req->shutdown.how = READ_ONCE(sqe->len);
3572 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3574 #if defined(CONFIG_NET)
3575 struct socket *sock;
3578 if (issue_flags & IO_URING_F_NONBLOCK)
3581 sock = sock_from_file(req->file);
3582 if (unlikely(!sock))
3585 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3587 req_set_fail_links(req);
3588 io_req_complete(req, ret);
3595 static int __io_splice_prep(struct io_kiocb *req,
3596 const struct io_uring_sqe *sqe)
3598 struct io_splice* sp = &req->splice;
3599 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3601 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3605 sp->len = READ_ONCE(sqe->len);
3606 sp->flags = READ_ONCE(sqe->splice_flags);
3608 if (unlikely(sp->flags & ~valid_flags))
3611 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3612 (sp->flags & SPLICE_F_FD_IN_FIXED));
3615 req->flags |= REQ_F_NEED_CLEANUP;
3617 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3619 * Splice operation will be punted aync, and here need to
3620 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3622 req->work.flags |= IO_WQ_WORK_UNBOUND;
3628 static int io_tee_prep(struct io_kiocb *req,
3629 const struct io_uring_sqe *sqe)
3631 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3633 return __io_splice_prep(req, sqe);
3636 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3638 struct io_splice *sp = &req->splice;
3639 struct file *in = sp->file_in;
3640 struct file *out = sp->file_out;
3641 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3644 if (issue_flags & IO_URING_F_NONBLOCK)
3647 ret = do_tee(in, out, sp->len, flags);
3649 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3650 req->flags &= ~REQ_F_NEED_CLEANUP;
3653 req_set_fail_links(req);
3654 io_req_complete(req, ret);
3658 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3660 struct io_splice* sp = &req->splice;
3662 sp->off_in = READ_ONCE(sqe->splice_off_in);
3663 sp->off_out = READ_ONCE(sqe->off);
3664 return __io_splice_prep(req, sqe);
3667 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3669 struct io_splice *sp = &req->splice;
3670 struct file *in = sp->file_in;
3671 struct file *out = sp->file_out;
3672 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3673 loff_t *poff_in, *poff_out;
3676 if (issue_flags & IO_URING_F_NONBLOCK)
3679 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3680 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3683 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3685 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3686 req->flags &= ~REQ_F_NEED_CLEANUP;
3689 req_set_fail_links(req);
3690 io_req_complete(req, ret);
3695 * IORING_OP_NOP just posts a completion event, nothing else.
3697 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3699 struct io_ring_ctx *ctx = req->ctx;
3701 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3704 __io_req_complete(req, issue_flags, 0, 0);
3708 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3710 struct io_ring_ctx *ctx = req->ctx;
3715 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3717 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3720 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3721 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3724 req->sync.off = READ_ONCE(sqe->off);
3725 req->sync.len = READ_ONCE(sqe->len);
3729 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3731 loff_t end = req->sync.off + req->sync.len;
3734 /* fsync always requires a blocking context */
3735 if (issue_flags & IO_URING_F_NONBLOCK)
3738 ret = vfs_fsync_range(req->file, req->sync.off,
3739 end > 0 ? end : LLONG_MAX,
3740 req->sync.flags & IORING_FSYNC_DATASYNC);
3742 req_set_fail_links(req);
3743 io_req_complete(req, ret);
3747 static int io_fallocate_prep(struct io_kiocb *req,
3748 const struct io_uring_sqe *sqe)
3750 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3752 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3755 req->sync.off = READ_ONCE(sqe->off);
3756 req->sync.len = READ_ONCE(sqe->addr);
3757 req->sync.mode = READ_ONCE(sqe->len);
3761 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3765 /* fallocate always requiring blocking context */
3766 if (issue_flags & IO_URING_F_NONBLOCK)
3768 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3771 req_set_fail_links(req);
3772 io_req_complete(req, ret);
3776 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3778 const char __user *fname;
3781 if (unlikely(sqe->ioprio || sqe->buf_index))
3783 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3786 /* open.how should be already initialised */
3787 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3788 req->open.how.flags |= O_LARGEFILE;
3790 req->open.dfd = READ_ONCE(sqe->fd);
3791 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 req->open.filename = getname(fname);
3793 if (IS_ERR(req->open.filename)) {
3794 ret = PTR_ERR(req->open.filename);
3795 req->open.filename = NULL;
3798 req->open.nofile = rlimit(RLIMIT_NOFILE);
3799 req->flags |= REQ_F_NEED_CLEANUP;
3803 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3807 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3809 mode = READ_ONCE(sqe->len);
3810 flags = READ_ONCE(sqe->open_flags);
3811 req->open.how = build_open_how(flags, mode);
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3817 struct open_how __user *how;
3821 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3823 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3824 len = READ_ONCE(sqe->len);
3825 if (len < OPEN_HOW_SIZE_VER0)
3828 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3833 return __io_openat_prep(req, sqe);
3836 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3838 struct open_flags op;
3841 bool resolve_nonblock;
3844 ret = build_open_flags(&req->open.how, &op);
3847 nonblock_set = op.open_flag & O_NONBLOCK;
3848 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3849 if (issue_flags & IO_URING_F_NONBLOCK) {
3851 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3852 * it'll always -EAGAIN
3854 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3856 op.lookup_flags |= LOOKUP_CACHED;
3857 op.open_flag |= O_NONBLOCK;
3860 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3864 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3865 /* only retry if RESOLVE_CACHED wasn't already set by application */
3866 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3867 file == ERR_PTR(-EAGAIN)) {
3869 * We could hang on to this 'fd', but seems like marginal
3870 * gain for something that is now known to be a slower path.
3871 * So just put it, and we'll get a new one when we retry.
3879 ret = PTR_ERR(file);
3881 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3882 file->f_flags &= ~O_NONBLOCK;
3883 fsnotify_open(file);
3884 fd_install(ret, file);
3887 putname(req->open.filename);
3888 req->flags &= ~REQ_F_NEED_CLEANUP;
3890 req_set_fail_links(req);
3891 io_req_complete(req, ret);
3895 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3897 return io_openat2(req, issue_flags);
3900 static int io_remove_buffers_prep(struct io_kiocb *req,
3901 const struct io_uring_sqe *sqe)
3903 struct io_provide_buf *p = &req->pbuf;
3906 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3909 tmp = READ_ONCE(sqe->fd);
3910 if (!tmp || tmp > USHRT_MAX)
3913 memset(p, 0, sizeof(*p));
3915 p->bgid = READ_ONCE(sqe->buf_group);
3919 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3920 int bgid, unsigned nbufs)
3924 /* shouldn't happen */
3928 /* the head kbuf is the list itself */
3929 while (!list_empty(&buf->list)) {
3930 struct io_buffer *nxt;
3932 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3933 list_del(&nxt->list);
3940 xa_erase(&ctx->io_buffers, bgid);
3945 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3947 struct io_provide_buf *p = &req->pbuf;
3948 struct io_ring_ctx *ctx = req->ctx;
3949 struct io_buffer *head;
3951 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3953 io_ring_submit_lock(ctx, !force_nonblock);
3955 lockdep_assert_held(&ctx->uring_lock);
3958 head = xa_load(&ctx->io_buffers, p->bgid);
3960 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3962 req_set_fail_links(req);
3964 /* complete before unlock, IOPOLL may need the lock */
3965 __io_req_complete(req, issue_flags, ret, 0);
3966 io_ring_submit_unlock(ctx, !force_nonblock);
3970 static int io_provide_buffers_prep(struct io_kiocb *req,
3971 const struct io_uring_sqe *sqe)
3974 struct io_provide_buf *p = &req->pbuf;
3977 if (sqe->ioprio || sqe->rw_flags)
3980 tmp = READ_ONCE(sqe->fd);
3981 if (!tmp || tmp > USHRT_MAX)
3984 p->addr = READ_ONCE(sqe->addr);
3985 p->len = READ_ONCE(sqe->len);
3987 size = (unsigned long)p->len * p->nbufs;
3988 if (!access_ok(u64_to_user_ptr(p->addr), size))
3991 p->bgid = READ_ONCE(sqe->buf_group);
3992 tmp = READ_ONCE(sqe->off);
3993 if (tmp > USHRT_MAX)
3999 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4001 struct io_buffer *buf;
4002 u64 addr = pbuf->addr;
4003 int i, bid = pbuf->bid;
4005 for (i = 0; i < pbuf->nbufs; i++) {
4006 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4011 buf->len = pbuf->len;
4016 INIT_LIST_HEAD(&buf->list);
4019 list_add_tail(&buf->list, &(*head)->list);
4023 return i ? i : -ENOMEM;
4026 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4028 struct io_provide_buf *p = &req->pbuf;
4029 struct io_ring_ctx *ctx = req->ctx;
4030 struct io_buffer *head, *list;
4032 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4034 io_ring_submit_lock(ctx, !force_nonblock);
4036 lockdep_assert_held(&ctx->uring_lock);
4038 list = head = xa_load(&ctx->io_buffers, p->bgid);
4040 ret = io_add_buffers(p, &head);
4041 if (ret >= 0 && !list) {
4042 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4044 __io_remove_buffers(ctx, head, p->bgid, -1U);
4047 req_set_fail_links(req);
4048 /* complete before unlock, IOPOLL may need the lock */
4049 __io_req_complete(req, issue_flags, ret, 0);
4050 io_ring_submit_unlock(ctx, !force_nonblock);
4054 static int io_epoll_ctl_prep(struct io_kiocb *req,
4055 const struct io_uring_sqe *sqe)
4057 #if defined(CONFIG_EPOLL)
4058 if (sqe->ioprio || sqe->buf_index)
4060 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4063 req->epoll.epfd = READ_ONCE(sqe->fd);
4064 req->epoll.op = READ_ONCE(sqe->len);
4065 req->epoll.fd = READ_ONCE(sqe->off);
4067 if (ep_op_has_event(req->epoll.op)) {
4068 struct epoll_event __user *ev;
4070 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4071 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4081 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4083 #if defined(CONFIG_EPOLL)
4084 struct io_epoll *ie = &req->epoll;
4086 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4088 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4089 if (force_nonblock && ret == -EAGAIN)
4093 req_set_fail_links(req);
4094 __io_req_complete(req, issue_flags, ret, 0);
4101 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4103 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4104 if (sqe->ioprio || sqe->buf_index || sqe->off)
4106 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4109 req->madvise.addr = READ_ONCE(sqe->addr);
4110 req->madvise.len = READ_ONCE(sqe->len);
4111 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4118 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4120 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4121 struct io_madvise *ma = &req->madvise;
4124 if (issue_flags & IO_URING_F_NONBLOCK)
4127 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4129 req_set_fail_links(req);
4130 io_req_complete(req, ret);
4137 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4141 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4144 req->fadvise.offset = READ_ONCE(sqe->off);
4145 req->fadvise.len = READ_ONCE(sqe->len);
4146 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4150 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4152 struct io_fadvise *fa = &req->fadvise;
4155 if (issue_flags & IO_URING_F_NONBLOCK) {
4156 switch (fa->advice) {
4157 case POSIX_FADV_NORMAL:
4158 case POSIX_FADV_RANDOM:
4159 case POSIX_FADV_SEQUENTIAL:
4166 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4168 req_set_fail_links(req);
4169 io_req_complete(req, ret);
4173 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4175 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4177 if (sqe->ioprio || sqe->buf_index)
4179 if (req->flags & REQ_F_FIXED_FILE)
4182 req->statx.dfd = READ_ONCE(sqe->fd);
4183 req->statx.mask = READ_ONCE(sqe->len);
4184 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4185 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4186 req->statx.flags = READ_ONCE(sqe->statx_flags);
4191 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4193 struct io_statx *ctx = &req->statx;
4196 if (issue_flags & IO_URING_F_NONBLOCK) {
4197 /* only need file table for an actual valid fd */
4198 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4199 req->flags |= REQ_F_NO_FILE_TABLE;
4203 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4207 req_set_fail_links(req);
4208 io_req_complete(req, ret);
4212 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4214 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4216 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4217 sqe->rw_flags || sqe->buf_index)
4219 if (req->flags & REQ_F_FIXED_FILE)
4222 req->close.fd = READ_ONCE(sqe->fd);
4226 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4228 struct files_struct *files = current->files;
4229 struct io_close *close = &req->close;
4230 struct fdtable *fdt;
4236 spin_lock(&files->file_lock);
4237 fdt = files_fdtable(files);
4238 if (close->fd >= fdt->max_fds) {
4239 spin_unlock(&files->file_lock);
4242 file = fdt->fd[close->fd];
4244 spin_unlock(&files->file_lock);
4248 if (file->f_op == &io_uring_fops) {
4249 spin_unlock(&files->file_lock);
4254 /* if the file has a flush method, be safe and punt to async */
4255 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4256 spin_unlock(&files->file_lock);
4260 ret = __close_fd_get_file(close->fd, &file);
4261 spin_unlock(&files->file_lock);
4268 /* No ->flush() or already async, safely close from here */
4269 ret = filp_close(file, current->files);
4272 req_set_fail_links(req);
4275 __io_req_complete(req, issue_flags, ret, 0);
4279 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4281 struct io_ring_ctx *ctx = req->ctx;
4283 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4285 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4288 req->sync.off = READ_ONCE(sqe->off);
4289 req->sync.len = READ_ONCE(sqe->len);
4290 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4294 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4298 /* sync_file_range always requires a blocking context */
4299 if (issue_flags & IO_URING_F_NONBLOCK)
4302 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4305 req_set_fail_links(req);
4306 io_req_complete(req, ret);
4310 #if defined(CONFIG_NET)
4311 static int io_setup_async_msg(struct io_kiocb *req,
4312 struct io_async_msghdr *kmsg)
4314 struct io_async_msghdr *async_msg = req->async_data;
4318 if (io_alloc_async_data(req)) {
4319 kfree(kmsg->free_iov);
4322 async_msg = req->async_data;
4323 req->flags |= REQ_F_NEED_CLEANUP;
4324 memcpy(async_msg, kmsg, sizeof(*kmsg));
4325 async_msg->msg.msg_name = &async_msg->addr;
4326 /* if were using fast_iov, set it to the new one */
4327 if (!async_msg->free_iov)
4328 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4333 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4334 struct io_async_msghdr *iomsg)
4336 iomsg->msg.msg_name = &iomsg->addr;
4337 iomsg->free_iov = iomsg->fast_iov;
4338 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4339 req->sr_msg.msg_flags, &iomsg->free_iov);
4342 static int io_sendmsg_prep_async(struct io_kiocb *req)
4346 ret = io_sendmsg_copy_hdr(req, req->async_data);
4348 req->flags |= REQ_F_NEED_CLEANUP;
4352 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4354 struct io_sr_msg *sr = &req->sr_msg;
4356 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4359 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4360 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4361 sr->len = READ_ONCE(sqe->len);
4363 #ifdef CONFIG_COMPAT
4364 if (req->ctx->compat)
4365 sr->msg_flags |= MSG_CMSG_COMPAT;
4370 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4372 struct io_async_msghdr iomsg, *kmsg;
4373 struct socket *sock;
4378 sock = sock_from_file(req->file);
4379 if (unlikely(!sock))
4382 kmsg = req->async_data;
4384 ret = io_sendmsg_copy_hdr(req, &iomsg);
4390 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4391 if (flags & MSG_DONTWAIT)
4392 req->flags |= REQ_F_NOWAIT;
4393 else if (issue_flags & IO_URING_F_NONBLOCK)
4394 flags |= MSG_DONTWAIT;
4396 if (flags & MSG_WAITALL)
4397 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4399 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4400 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4401 return io_setup_async_msg(req, kmsg);
4402 if (ret == -ERESTARTSYS)
4405 /* fast path, check for non-NULL to avoid function call */
4407 kfree(kmsg->free_iov);
4408 req->flags &= ~REQ_F_NEED_CLEANUP;
4410 req_set_fail_links(req);
4411 __io_req_complete(req, issue_flags, ret, 0);
4415 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4417 struct io_sr_msg *sr = &req->sr_msg;
4420 struct socket *sock;
4425 sock = sock_from_file(req->file);
4426 if (unlikely(!sock))
4429 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4433 msg.msg_name = NULL;
4434 msg.msg_control = NULL;
4435 msg.msg_controllen = 0;
4436 msg.msg_namelen = 0;
4438 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4439 if (flags & MSG_DONTWAIT)
4440 req->flags |= REQ_F_NOWAIT;
4441 else if (issue_flags & IO_URING_F_NONBLOCK)
4442 flags |= MSG_DONTWAIT;
4444 if (flags & MSG_WAITALL)
4445 min_ret = iov_iter_count(&msg.msg_iter);
4447 msg.msg_flags = flags;
4448 ret = sock_sendmsg(sock, &msg);
4449 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4451 if (ret == -ERESTARTSYS)
4455 req_set_fail_links(req);
4456 __io_req_complete(req, issue_flags, ret, 0);
4460 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4461 struct io_async_msghdr *iomsg)
4463 struct io_sr_msg *sr = &req->sr_msg;
4464 struct iovec __user *uiov;
4468 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4469 &iomsg->uaddr, &uiov, &iov_len);
4473 if (req->flags & REQ_F_BUFFER_SELECT) {
4476 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4478 sr->len = iomsg->fast_iov[0].iov_len;
4479 iomsg->free_iov = NULL;
4481 iomsg->free_iov = iomsg->fast_iov;
4482 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4483 &iomsg->free_iov, &iomsg->msg.msg_iter,
4492 #ifdef CONFIG_COMPAT
4493 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4494 struct io_async_msghdr *iomsg)
4496 struct compat_msghdr __user *msg_compat;
4497 struct io_sr_msg *sr = &req->sr_msg;
4498 struct compat_iovec __user *uiov;
4503 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4504 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4509 uiov = compat_ptr(ptr);
4510 if (req->flags & REQ_F_BUFFER_SELECT) {
4511 compat_ssize_t clen;
4515 if (!access_ok(uiov, sizeof(*uiov)))
4517 if (__get_user(clen, &uiov->iov_len))
4522 iomsg->free_iov = NULL;
4524 iomsg->free_iov = iomsg->fast_iov;
4525 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4526 UIO_FASTIOV, &iomsg->free_iov,
4527 &iomsg->msg.msg_iter, true);
4536 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4537 struct io_async_msghdr *iomsg)
4539 iomsg->msg.msg_name = &iomsg->addr;
4541 #ifdef CONFIG_COMPAT
4542 if (req->ctx->compat)
4543 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4546 return __io_recvmsg_copy_hdr(req, iomsg);
4549 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4552 struct io_sr_msg *sr = &req->sr_msg;
4553 struct io_buffer *kbuf;
4555 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4560 req->flags |= REQ_F_BUFFER_SELECTED;
4564 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4566 return io_put_kbuf(req, req->sr_msg.kbuf);
4569 static int io_recvmsg_prep_async(struct io_kiocb *req)
4573 ret = io_recvmsg_copy_hdr(req, req->async_data);
4575 req->flags |= REQ_F_NEED_CLEANUP;
4579 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4581 struct io_sr_msg *sr = &req->sr_msg;
4583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4586 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4587 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4588 sr->len = READ_ONCE(sqe->len);
4589 sr->bgid = READ_ONCE(sqe->buf_group);
4591 #ifdef CONFIG_COMPAT
4592 if (req->ctx->compat)
4593 sr->msg_flags |= MSG_CMSG_COMPAT;
4598 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4600 struct io_async_msghdr iomsg, *kmsg;
4601 struct socket *sock;
4602 struct io_buffer *kbuf;
4605 int ret, cflags = 0;
4606 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4608 sock = sock_from_file(req->file);
4609 if (unlikely(!sock))
4612 kmsg = req->async_data;
4614 ret = io_recvmsg_copy_hdr(req, &iomsg);
4620 if (req->flags & REQ_F_BUFFER_SELECT) {
4621 kbuf = io_recv_buffer_select(req, !force_nonblock);
4623 return PTR_ERR(kbuf);
4624 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4625 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4626 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4627 1, req->sr_msg.len);
4630 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4631 if (flags & MSG_DONTWAIT)
4632 req->flags |= REQ_F_NOWAIT;
4633 else if (force_nonblock)
4634 flags |= MSG_DONTWAIT;
4636 if (flags & MSG_WAITALL)
4637 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4639 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4640 kmsg->uaddr, flags);
4641 if (force_nonblock && ret == -EAGAIN)
4642 return io_setup_async_msg(req, kmsg);
4643 if (ret == -ERESTARTSYS)
4646 if (req->flags & REQ_F_BUFFER_SELECTED)
4647 cflags = io_put_recv_kbuf(req);
4648 /* fast path, check for non-NULL to avoid function call */
4650 kfree(kmsg->free_iov);
4651 req->flags &= ~REQ_F_NEED_CLEANUP;
4652 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4653 req_set_fail_links(req);
4654 __io_req_complete(req, issue_flags, ret, cflags);
4658 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4660 struct io_buffer *kbuf;
4661 struct io_sr_msg *sr = &req->sr_msg;
4663 void __user *buf = sr->buf;
4664 struct socket *sock;
4668 int ret, cflags = 0;
4669 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4671 sock = sock_from_file(req->file);
4672 if (unlikely(!sock))
4675 if (req->flags & REQ_F_BUFFER_SELECT) {
4676 kbuf = io_recv_buffer_select(req, !force_nonblock);
4678 return PTR_ERR(kbuf);
4679 buf = u64_to_user_ptr(kbuf->addr);
4682 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4686 msg.msg_name = NULL;
4687 msg.msg_control = NULL;
4688 msg.msg_controllen = 0;
4689 msg.msg_namelen = 0;
4690 msg.msg_iocb = NULL;
4693 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4694 if (flags & MSG_DONTWAIT)
4695 req->flags |= REQ_F_NOWAIT;
4696 else if (force_nonblock)
4697 flags |= MSG_DONTWAIT;
4699 if (flags & MSG_WAITALL)
4700 min_ret = iov_iter_count(&msg.msg_iter);
4702 ret = sock_recvmsg(sock, &msg, flags);
4703 if (force_nonblock && ret == -EAGAIN)
4705 if (ret == -ERESTARTSYS)
4708 if (req->flags & REQ_F_BUFFER_SELECTED)
4709 cflags = io_put_recv_kbuf(req);
4710 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4711 req_set_fail_links(req);
4712 __io_req_complete(req, issue_flags, ret, cflags);
4716 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_accept *accept = &req->accept;
4720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722 if (sqe->ioprio || sqe->len || sqe->buf_index)
4725 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4727 accept->flags = READ_ONCE(sqe->accept_flags);
4728 accept->nofile = rlimit(RLIMIT_NOFILE);
4732 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4734 struct io_accept *accept = &req->accept;
4735 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4736 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4739 if (req->file->f_flags & O_NONBLOCK)
4740 req->flags |= REQ_F_NOWAIT;
4742 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4743 accept->addr_len, accept->flags,
4745 if (ret == -EAGAIN && force_nonblock)
4748 if (ret == -ERESTARTSYS)
4750 req_set_fail_links(req);
4752 __io_req_complete(req, issue_flags, ret, 0);
4756 static int io_connect_prep_async(struct io_kiocb *req)
4758 struct io_async_connect *io = req->async_data;
4759 struct io_connect *conn = &req->connect;
4761 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4764 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4766 struct io_connect *conn = &req->connect;
4768 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4770 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4773 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4774 conn->addr_len = READ_ONCE(sqe->addr2);
4778 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4780 struct io_async_connect __io, *io;
4781 unsigned file_flags;
4783 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4785 if (req->async_data) {
4786 io = req->async_data;
4788 ret = move_addr_to_kernel(req->connect.addr,
4789 req->connect.addr_len,
4796 file_flags = force_nonblock ? O_NONBLOCK : 0;
4798 ret = __sys_connect_file(req->file, &io->address,
4799 req->connect.addr_len, file_flags);
4800 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4801 if (req->async_data)
4803 if (io_alloc_async_data(req)) {
4807 memcpy(req->async_data, &__io, sizeof(__io));
4810 if (ret == -ERESTARTSYS)
4814 req_set_fail_links(req);
4815 __io_req_complete(req, issue_flags, ret, 0);
4818 #else /* !CONFIG_NET */
4819 #define IO_NETOP_FN(op) \
4820 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4822 return -EOPNOTSUPP; \
4825 #define IO_NETOP_PREP(op) \
4827 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4829 return -EOPNOTSUPP; \
4832 #define IO_NETOP_PREP_ASYNC(op) \
4834 static int io_##op##_prep_async(struct io_kiocb *req) \
4836 return -EOPNOTSUPP; \
4839 IO_NETOP_PREP_ASYNC(sendmsg);
4840 IO_NETOP_PREP_ASYNC(recvmsg);
4841 IO_NETOP_PREP_ASYNC(connect);
4842 IO_NETOP_PREP(accept);
4845 #endif /* CONFIG_NET */
4847 struct io_poll_table {
4848 struct poll_table_struct pt;
4849 struct io_kiocb *req;
4853 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4854 __poll_t mask, task_work_func_t func)
4858 /* for instances that support it check for an event match first: */
4859 if (mask && !(mask & poll->events))
4862 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4864 list_del_init(&poll->wait.entry);
4867 req->task_work.func = func;
4868 percpu_ref_get(&req->ctx->refs);
4871 * If this fails, then the task is exiting. When a task exits, the
4872 * work gets canceled, so just cancel this request as well instead
4873 * of executing it. We can't safely execute it anyway, as we may not
4874 * have the needed state needed for it anyway.
4876 ret = io_req_task_work_add(req);
4877 if (unlikely(ret)) {
4878 WRITE_ONCE(poll->canceled, true);
4879 io_req_task_work_add_fallback(req, func);
4884 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4885 __acquires(&req->ctx->completion_lock)
4887 struct io_ring_ctx *ctx = req->ctx;
4889 if (!req->result && !READ_ONCE(poll->canceled)) {
4890 struct poll_table_struct pt = { ._key = poll->events };
4892 req->result = vfs_poll(req->file, &pt) & poll->events;
4895 spin_lock_irq(&ctx->completion_lock);
4896 if (!req->result && !READ_ONCE(poll->canceled)) {
4897 add_wait_queue(poll->head, &poll->wait);
4904 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4906 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4907 if (req->opcode == IORING_OP_POLL_ADD)
4908 return req->async_data;
4909 return req->apoll->double_poll;
4912 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4914 if (req->opcode == IORING_OP_POLL_ADD)
4916 return &req->apoll->poll;
4919 static void io_poll_remove_double(struct io_kiocb *req)
4921 struct io_poll_iocb *poll = io_poll_get_double(req);
4923 lockdep_assert_held(&req->ctx->completion_lock);
4925 if (poll && poll->head) {
4926 struct wait_queue_head *head = poll->head;
4928 spin_lock(&head->lock);
4929 list_del_init(&poll->wait.entry);
4930 if (poll->wait.private)
4933 spin_unlock(&head->lock);
4937 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4939 struct io_ring_ctx *ctx = req->ctx;
4941 io_poll_remove_double(req);
4942 req->poll.done = true;
4943 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4944 io_commit_cqring(ctx);
4947 static void io_poll_task_func(struct callback_head *cb)
4949 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4950 struct io_ring_ctx *ctx = req->ctx;
4951 struct io_kiocb *nxt;
4953 if (io_poll_rewait(req, &req->poll)) {
4954 spin_unlock_irq(&ctx->completion_lock);
4956 hash_del(&req->hash_node);
4957 io_poll_complete(req, req->result, 0);
4958 spin_unlock_irq(&ctx->completion_lock);
4960 nxt = io_put_req_find_next(req);
4961 io_cqring_ev_posted(ctx);
4963 __io_req_task_submit(nxt);
4966 percpu_ref_put(&ctx->refs);
4969 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4970 int sync, void *key)
4972 struct io_kiocb *req = wait->private;
4973 struct io_poll_iocb *poll = io_poll_get_single(req);
4974 __poll_t mask = key_to_poll(key);
4976 /* for instances that support it check for an event match first: */
4977 if (mask && !(mask & poll->events))
4980 list_del_init(&wait->entry);
4982 if (poll && poll->head) {
4985 spin_lock(&poll->head->lock);
4986 done = list_empty(&poll->wait.entry);
4988 list_del_init(&poll->wait.entry);
4989 /* make sure double remove sees this as being gone */
4990 wait->private = NULL;
4991 spin_unlock(&poll->head->lock);
4993 /* use wait func handler, so it matches the rq type */
4994 poll->wait.func(&poll->wait, mode, sync, key);
5001 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5002 wait_queue_func_t wake_func)
5006 poll->canceled = false;
5007 poll->events = events;
5008 INIT_LIST_HEAD(&poll->wait.entry);
5009 init_waitqueue_func_entry(&poll->wait, wake_func);
5012 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5013 struct wait_queue_head *head,
5014 struct io_poll_iocb **poll_ptr)
5016 struct io_kiocb *req = pt->req;
5019 * If poll->head is already set, it's because the file being polled
5020 * uses multiple waitqueues for poll handling (eg one for read, one
5021 * for write). Setup a separate io_poll_iocb if this happens.
5023 if (unlikely(poll->head)) {
5024 struct io_poll_iocb *poll_one = poll;
5026 /* already have a 2nd entry, fail a third attempt */
5028 pt->error = -EINVAL;
5031 /* double add on the same waitqueue head, ignore */
5032 if (poll->head == head)
5034 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5036 pt->error = -ENOMEM;
5039 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5041 poll->wait.private = req;
5048 if (poll->events & EPOLLEXCLUSIVE)
5049 add_wait_queue_exclusive(head, &poll->wait);
5051 add_wait_queue(head, &poll->wait);
5054 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5055 struct poll_table_struct *p)
5057 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5058 struct async_poll *apoll = pt->req->apoll;
5060 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5063 static void io_async_task_func(struct callback_head *cb)
5065 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5066 struct async_poll *apoll = req->apoll;
5067 struct io_ring_ctx *ctx = req->ctx;
5069 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5071 if (io_poll_rewait(req, &apoll->poll)) {
5072 spin_unlock_irq(&ctx->completion_lock);
5073 percpu_ref_put(&ctx->refs);
5077 /* If req is still hashed, it cannot have been canceled. Don't check. */
5078 if (hash_hashed(&req->hash_node))
5079 hash_del(&req->hash_node);
5081 io_poll_remove_double(req);
5082 spin_unlock_irq(&ctx->completion_lock);
5084 if (!READ_ONCE(apoll->poll.canceled))
5085 __io_req_task_submit(req);
5087 __io_req_task_cancel(req, -ECANCELED);
5089 percpu_ref_put(&ctx->refs);
5090 kfree(apoll->double_poll);
5094 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5097 struct io_kiocb *req = wait->private;
5098 struct io_poll_iocb *poll = &req->apoll->poll;
5100 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5103 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5106 static void io_poll_req_insert(struct io_kiocb *req)
5108 struct io_ring_ctx *ctx = req->ctx;
5109 struct hlist_head *list;
5111 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5112 hlist_add_head(&req->hash_node, list);
5115 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5116 struct io_poll_iocb *poll,
5117 struct io_poll_table *ipt, __poll_t mask,
5118 wait_queue_func_t wake_func)
5119 __acquires(&ctx->completion_lock)
5121 struct io_ring_ctx *ctx = req->ctx;
5122 bool cancel = false;
5124 INIT_HLIST_NODE(&req->hash_node);
5125 io_init_poll_iocb(poll, mask, wake_func);
5126 poll->file = req->file;
5127 poll->wait.private = req;
5129 ipt->pt._key = mask;
5131 ipt->error = -EINVAL;
5133 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5135 spin_lock_irq(&ctx->completion_lock);
5136 if (likely(poll->head)) {
5137 spin_lock(&poll->head->lock);
5138 if (unlikely(list_empty(&poll->wait.entry))) {
5144 if (mask || ipt->error)
5145 list_del_init(&poll->wait.entry);
5147 WRITE_ONCE(poll->canceled, true);
5148 else if (!poll->done) /* actually waiting for an event */
5149 io_poll_req_insert(req);
5150 spin_unlock(&poll->head->lock);
5156 static bool io_arm_poll_handler(struct io_kiocb *req)
5158 const struct io_op_def *def = &io_op_defs[req->opcode];
5159 struct io_ring_ctx *ctx = req->ctx;
5160 struct async_poll *apoll;
5161 struct io_poll_table ipt;
5165 if (!req->file || !file_can_poll(req->file))
5167 if (req->flags & REQ_F_POLLED)
5171 else if (def->pollout)
5175 /* if we can't nonblock try, then no point in arming a poll handler */
5176 if (!io_file_supports_async(req->file, rw))
5179 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5180 if (unlikely(!apoll))
5182 apoll->double_poll = NULL;
5184 req->flags |= REQ_F_POLLED;
5189 mask |= POLLIN | POLLRDNORM;
5191 mask |= POLLOUT | POLLWRNORM;
5193 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5194 if ((req->opcode == IORING_OP_RECVMSG) &&
5195 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5198 mask |= POLLERR | POLLPRI;
5200 ipt.pt._qproc = io_async_queue_proc;
5202 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5204 if (ret || ipt.error) {
5205 io_poll_remove_double(req);
5206 spin_unlock_irq(&ctx->completion_lock);
5207 kfree(apoll->double_poll);
5211 spin_unlock_irq(&ctx->completion_lock);
5212 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5213 apoll->poll.events);
5217 static bool __io_poll_remove_one(struct io_kiocb *req,
5218 struct io_poll_iocb *poll)
5220 bool do_complete = false;
5222 spin_lock(&poll->head->lock);
5223 WRITE_ONCE(poll->canceled, true);
5224 if (!list_empty(&poll->wait.entry)) {
5225 list_del_init(&poll->wait.entry);
5228 spin_unlock(&poll->head->lock);
5229 hash_del(&req->hash_node);
5233 static bool io_poll_remove_one(struct io_kiocb *req)
5237 io_poll_remove_double(req);
5239 if (req->opcode == IORING_OP_POLL_ADD) {
5240 do_complete = __io_poll_remove_one(req, &req->poll);
5242 struct async_poll *apoll = req->apoll;
5244 /* non-poll requests have submit ref still */
5245 do_complete = __io_poll_remove_one(req, &apoll->poll);
5248 kfree(apoll->double_poll);
5254 io_cqring_fill_event(req, -ECANCELED);
5255 io_commit_cqring(req->ctx);
5256 req_set_fail_links(req);
5257 io_put_req_deferred(req, 1);
5264 * Returns true if we found and killed one or more poll requests
5266 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5267 struct files_struct *files)
5269 struct hlist_node *tmp;
5270 struct io_kiocb *req;
5273 spin_lock_irq(&ctx->completion_lock);
5274 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5275 struct hlist_head *list;
5277 list = &ctx->cancel_hash[i];
5278 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5279 if (io_match_task(req, tsk, files))
5280 posted += io_poll_remove_one(req);
5283 spin_unlock_irq(&ctx->completion_lock);
5286 io_cqring_ev_posted(ctx);
5291 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5293 struct hlist_head *list;
5294 struct io_kiocb *req;
5296 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5297 hlist_for_each_entry(req, list, hash_node) {
5298 if (sqe_addr != req->user_data)
5300 if (io_poll_remove_one(req))
5308 static int io_poll_remove_prep(struct io_kiocb *req,
5309 const struct io_uring_sqe *sqe)
5311 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5313 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5317 req->poll_remove.addr = READ_ONCE(sqe->addr);
5322 * Find a running poll command that matches one specified in sqe->addr,
5323 * and remove it if found.
5325 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5327 struct io_ring_ctx *ctx = req->ctx;
5330 spin_lock_irq(&ctx->completion_lock);
5331 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5332 spin_unlock_irq(&ctx->completion_lock);
5335 req_set_fail_links(req);
5336 io_req_complete(req, ret);
5340 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5343 struct io_kiocb *req = wait->private;
5344 struct io_poll_iocb *poll = &req->poll;
5346 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5349 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5350 struct poll_table_struct *p)
5352 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5354 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5357 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5359 struct io_poll_iocb *poll = &req->poll;
5362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5364 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5367 events = READ_ONCE(sqe->poll32_events);
5369 events = swahw32(events);
5371 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5372 (events & EPOLLEXCLUSIVE);
5376 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5378 struct io_poll_iocb *poll = &req->poll;
5379 struct io_ring_ctx *ctx = req->ctx;
5380 struct io_poll_table ipt;
5383 ipt.pt._qproc = io_poll_queue_proc;
5385 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5388 if (mask) { /* no async, we'd stolen it */
5390 io_poll_complete(req, mask, 0);
5392 spin_unlock_irq(&ctx->completion_lock);
5395 io_cqring_ev_posted(ctx);
5401 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5403 struct io_timeout_data *data = container_of(timer,
5404 struct io_timeout_data, timer);
5405 struct io_kiocb *req = data->req;
5406 struct io_ring_ctx *ctx = req->ctx;
5407 unsigned long flags;
5409 spin_lock_irqsave(&ctx->completion_lock, flags);
5410 list_del_init(&req->timeout.list);
5411 atomic_set(&req->ctx->cq_timeouts,
5412 atomic_read(&req->ctx->cq_timeouts) + 1);
5414 io_cqring_fill_event(req, -ETIME);
5415 io_commit_cqring(ctx);
5416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5418 io_cqring_ev_posted(ctx);
5419 req_set_fail_links(req);
5421 return HRTIMER_NORESTART;
5424 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5427 struct io_timeout_data *io;
5428 struct io_kiocb *req;
5431 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5432 if (user_data == req->user_data) {
5439 return ERR_PTR(ret);
5441 io = req->async_data;
5442 ret = hrtimer_try_to_cancel(&io->timer);
5444 return ERR_PTR(-EALREADY);
5445 list_del_init(&req->timeout.list);
5449 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5451 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5454 return PTR_ERR(req);
5456 req_set_fail_links(req);
5457 io_cqring_fill_event(req, -ECANCELED);
5458 io_put_req_deferred(req, 1);
5462 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5463 struct timespec64 *ts, enum hrtimer_mode mode)
5465 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5466 struct io_timeout_data *data;
5469 return PTR_ERR(req);
5471 req->timeout.off = 0; /* noseq */
5472 data = req->async_data;
5473 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5474 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5475 data->timer.function = io_timeout_fn;
5476 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5480 static int io_timeout_remove_prep(struct io_kiocb *req,
5481 const struct io_uring_sqe *sqe)
5483 struct io_timeout_rem *tr = &req->timeout_rem;
5485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5487 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5489 if (sqe->ioprio || sqe->buf_index || sqe->len)
5492 tr->addr = READ_ONCE(sqe->addr);
5493 tr->flags = READ_ONCE(sqe->timeout_flags);
5494 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5495 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5497 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5499 } else if (tr->flags) {
5500 /* timeout removal doesn't support flags */
5507 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5509 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5514 * Remove or update an existing timeout command
5516 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5518 struct io_timeout_rem *tr = &req->timeout_rem;
5519 struct io_ring_ctx *ctx = req->ctx;
5522 spin_lock_irq(&ctx->completion_lock);
5523 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5524 ret = io_timeout_cancel(ctx, tr->addr);
5526 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5527 io_translate_timeout_mode(tr->flags));
5529 io_cqring_fill_event(req, ret);
5530 io_commit_cqring(ctx);
5531 spin_unlock_irq(&ctx->completion_lock);
5532 io_cqring_ev_posted(ctx);
5534 req_set_fail_links(req);
5539 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5540 bool is_timeout_link)
5542 struct io_timeout_data *data;
5544 u32 off = READ_ONCE(sqe->off);
5546 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5548 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5550 if (off && is_timeout_link)
5552 flags = READ_ONCE(sqe->timeout_flags);
5553 if (flags & ~IORING_TIMEOUT_ABS)
5556 req->timeout.off = off;
5558 if (!req->async_data && io_alloc_async_data(req))
5561 data = req->async_data;
5564 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5567 data->mode = io_translate_timeout_mode(flags);
5568 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5569 if (is_timeout_link)
5570 io_req_track_inflight(req);
5574 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5576 struct io_ring_ctx *ctx = req->ctx;
5577 struct io_timeout_data *data = req->async_data;
5578 struct list_head *entry;
5579 u32 tail, off = req->timeout.off;
5581 spin_lock_irq(&ctx->completion_lock);
5584 * sqe->off holds how many events that need to occur for this
5585 * timeout event to be satisfied. If it isn't set, then this is
5586 * a pure timeout request, sequence isn't used.
5588 if (io_is_timeout_noseq(req)) {
5589 entry = ctx->timeout_list.prev;
5593 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5594 req->timeout.target_seq = tail + off;
5596 /* Update the last seq here in case io_flush_timeouts() hasn't.
5597 * This is safe because ->completion_lock is held, and submissions
5598 * and completions are never mixed in the same ->completion_lock section.
5600 ctx->cq_last_tm_flush = tail;
5603 * Insertion sort, ensuring the first entry in the list is always
5604 * the one we need first.
5606 list_for_each_prev(entry, &ctx->timeout_list) {
5607 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5610 if (io_is_timeout_noseq(nxt))
5612 /* nxt.seq is behind @tail, otherwise would've been completed */
5613 if (off >= nxt->timeout.target_seq - tail)
5617 list_add(&req->timeout.list, entry);
5618 data->timer.function = io_timeout_fn;
5619 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5620 spin_unlock_irq(&ctx->completion_lock);
5624 struct io_cancel_data {
5625 struct io_ring_ctx *ctx;
5629 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5631 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5632 struct io_cancel_data *cd = data;
5634 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5637 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5638 struct io_ring_ctx *ctx)
5640 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5641 enum io_wq_cancel cancel_ret;
5644 if (!tctx || !tctx->io_wq)
5647 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5648 switch (cancel_ret) {
5649 case IO_WQ_CANCEL_OK:
5652 case IO_WQ_CANCEL_RUNNING:
5655 case IO_WQ_CANCEL_NOTFOUND:
5663 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5664 struct io_kiocb *req, __u64 sqe_addr,
5667 unsigned long flags;
5670 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5671 if (ret != -ENOENT) {
5672 spin_lock_irqsave(&ctx->completion_lock, flags);
5676 spin_lock_irqsave(&ctx->completion_lock, flags);
5677 ret = io_timeout_cancel(ctx, sqe_addr);
5680 ret = io_poll_cancel(ctx, sqe_addr);
5684 io_cqring_fill_event(req, ret);
5685 io_commit_cqring(ctx);
5686 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5687 io_cqring_ev_posted(ctx);
5690 req_set_fail_links(req);
5694 static int io_async_cancel_prep(struct io_kiocb *req,
5695 const struct io_uring_sqe *sqe)
5697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5699 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5701 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5704 req->cancel.addr = READ_ONCE(sqe->addr);
5708 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5710 struct io_ring_ctx *ctx = req->ctx;
5711 u64 sqe_addr = req->cancel.addr;
5712 struct io_tctx_node *node;
5715 /* tasks should wait for their io-wq threads, so safe w/o sync */
5716 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5717 spin_lock_irq(&ctx->completion_lock);
5720 ret = io_timeout_cancel(ctx, sqe_addr);
5723 ret = io_poll_cancel(ctx, sqe_addr);
5726 spin_unlock_irq(&ctx->completion_lock);
5728 /* slow path, try all io-wq's */
5729 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5731 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5732 struct io_uring_task *tctx = node->task->io_uring;
5734 if (!tctx || !tctx->io_wq)
5736 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5740 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5742 spin_lock_irq(&ctx->completion_lock);
5744 io_cqring_fill_event(req, ret);
5745 io_commit_cqring(ctx);
5746 spin_unlock_irq(&ctx->completion_lock);
5747 io_cqring_ev_posted(ctx);
5750 req_set_fail_links(req);
5755 static int io_rsrc_update_prep(struct io_kiocb *req,
5756 const struct io_uring_sqe *sqe)
5758 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5760 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5762 if (sqe->ioprio || sqe->rw_flags)
5765 req->rsrc_update.offset = READ_ONCE(sqe->off);
5766 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5767 if (!req->rsrc_update.nr_args)
5769 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5773 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5775 struct io_ring_ctx *ctx = req->ctx;
5776 struct io_uring_rsrc_update up;
5779 if (issue_flags & IO_URING_F_NONBLOCK)
5782 up.offset = req->rsrc_update.offset;
5783 up.data = req->rsrc_update.arg;
5785 mutex_lock(&ctx->uring_lock);
5786 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5787 mutex_unlock(&ctx->uring_lock);
5790 req_set_fail_links(req);
5791 __io_req_complete(req, issue_flags, ret, 0);
5795 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5797 switch (req->opcode) {
5800 case IORING_OP_READV:
5801 case IORING_OP_READ_FIXED:
5802 case IORING_OP_READ:
5803 return io_read_prep(req, sqe);
5804 case IORING_OP_WRITEV:
5805 case IORING_OP_WRITE_FIXED:
5806 case IORING_OP_WRITE:
5807 return io_write_prep(req, sqe);
5808 case IORING_OP_POLL_ADD:
5809 return io_poll_add_prep(req, sqe);
5810 case IORING_OP_POLL_REMOVE:
5811 return io_poll_remove_prep(req, sqe);
5812 case IORING_OP_FSYNC:
5813 return io_fsync_prep(req, sqe);
5814 case IORING_OP_SYNC_FILE_RANGE:
5815 return io_sfr_prep(req, sqe);
5816 case IORING_OP_SENDMSG:
5817 case IORING_OP_SEND:
5818 return io_sendmsg_prep(req, sqe);
5819 case IORING_OP_RECVMSG:
5820 case IORING_OP_RECV:
5821 return io_recvmsg_prep(req, sqe);
5822 case IORING_OP_CONNECT:
5823 return io_connect_prep(req, sqe);
5824 case IORING_OP_TIMEOUT:
5825 return io_timeout_prep(req, sqe, false);
5826 case IORING_OP_TIMEOUT_REMOVE:
5827 return io_timeout_remove_prep(req, sqe);
5828 case IORING_OP_ASYNC_CANCEL:
5829 return io_async_cancel_prep(req, sqe);
5830 case IORING_OP_LINK_TIMEOUT:
5831 return io_timeout_prep(req, sqe, true);
5832 case IORING_OP_ACCEPT:
5833 return io_accept_prep(req, sqe);
5834 case IORING_OP_FALLOCATE:
5835 return io_fallocate_prep(req, sqe);
5836 case IORING_OP_OPENAT:
5837 return io_openat_prep(req, sqe);
5838 case IORING_OP_CLOSE:
5839 return io_close_prep(req, sqe);
5840 case IORING_OP_FILES_UPDATE:
5841 return io_rsrc_update_prep(req, sqe);
5842 case IORING_OP_STATX:
5843 return io_statx_prep(req, sqe);
5844 case IORING_OP_FADVISE:
5845 return io_fadvise_prep(req, sqe);
5846 case IORING_OP_MADVISE:
5847 return io_madvise_prep(req, sqe);
5848 case IORING_OP_OPENAT2:
5849 return io_openat2_prep(req, sqe);
5850 case IORING_OP_EPOLL_CTL:
5851 return io_epoll_ctl_prep(req, sqe);
5852 case IORING_OP_SPLICE:
5853 return io_splice_prep(req, sqe);
5854 case IORING_OP_PROVIDE_BUFFERS:
5855 return io_provide_buffers_prep(req, sqe);
5856 case IORING_OP_REMOVE_BUFFERS:
5857 return io_remove_buffers_prep(req, sqe);
5859 return io_tee_prep(req, sqe);
5860 case IORING_OP_SHUTDOWN:
5861 return io_shutdown_prep(req, sqe);
5862 case IORING_OP_RENAMEAT:
5863 return io_renameat_prep(req, sqe);
5864 case IORING_OP_UNLINKAT:
5865 return io_unlinkat_prep(req, sqe);
5868 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5873 static int io_req_prep_async(struct io_kiocb *req)
5875 if (!io_op_defs[req->opcode].needs_async_setup)
5877 if (WARN_ON_ONCE(req->async_data))
5879 if (io_alloc_async_data(req))
5882 switch (req->opcode) {
5883 case IORING_OP_READV:
5884 return io_rw_prep_async(req, READ);
5885 case IORING_OP_WRITEV:
5886 return io_rw_prep_async(req, WRITE);
5887 case IORING_OP_SENDMSG:
5888 return io_sendmsg_prep_async(req);
5889 case IORING_OP_RECVMSG:
5890 return io_recvmsg_prep_async(req);
5891 case IORING_OP_CONNECT:
5892 return io_connect_prep_async(req);
5894 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5899 static u32 io_get_sequence(struct io_kiocb *req)
5901 struct io_kiocb *pos;
5902 struct io_ring_ctx *ctx = req->ctx;
5903 u32 total_submitted, nr_reqs = 0;
5905 io_for_each_link(pos, req)
5908 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5909 return total_submitted - nr_reqs;
5912 static int io_req_defer(struct io_kiocb *req)
5914 struct io_ring_ctx *ctx = req->ctx;
5915 struct io_defer_entry *de;
5919 /* Still need defer if there is pending req in defer list. */
5920 if (likely(list_empty_careful(&ctx->defer_list) &&
5921 !(req->flags & REQ_F_IO_DRAIN)))
5924 seq = io_get_sequence(req);
5925 /* Still a chance to pass the sequence check */
5926 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5929 ret = io_req_prep_async(req);
5932 io_prep_async_link(req);
5933 de = kmalloc(sizeof(*de), GFP_KERNEL);
5937 spin_lock_irq(&ctx->completion_lock);
5938 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5939 spin_unlock_irq(&ctx->completion_lock);
5941 io_queue_async_work(req);
5942 return -EIOCBQUEUED;
5945 trace_io_uring_defer(ctx, req, req->user_data);
5948 list_add_tail(&de->list, &ctx->defer_list);
5949 spin_unlock_irq(&ctx->completion_lock);
5950 return -EIOCBQUEUED;
5953 static void __io_clean_op(struct io_kiocb *req)
5955 if (req->flags & REQ_F_BUFFER_SELECTED) {
5956 switch (req->opcode) {
5957 case IORING_OP_READV:
5958 case IORING_OP_READ_FIXED:
5959 case IORING_OP_READ:
5960 kfree((void *)(unsigned long)req->rw.addr);
5962 case IORING_OP_RECVMSG:
5963 case IORING_OP_RECV:
5964 kfree(req->sr_msg.kbuf);
5967 req->flags &= ~REQ_F_BUFFER_SELECTED;
5970 if (req->flags & REQ_F_NEED_CLEANUP) {
5971 switch (req->opcode) {
5972 case IORING_OP_READV:
5973 case IORING_OP_READ_FIXED:
5974 case IORING_OP_READ:
5975 case IORING_OP_WRITEV:
5976 case IORING_OP_WRITE_FIXED:
5977 case IORING_OP_WRITE: {
5978 struct io_async_rw *io = req->async_data;
5980 kfree(io->free_iovec);
5983 case IORING_OP_RECVMSG:
5984 case IORING_OP_SENDMSG: {
5985 struct io_async_msghdr *io = req->async_data;
5987 kfree(io->free_iov);
5990 case IORING_OP_SPLICE:
5992 io_put_file(req, req->splice.file_in,
5993 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5995 case IORING_OP_OPENAT:
5996 case IORING_OP_OPENAT2:
5997 if (req->open.filename)
5998 putname(req->open.filename);
6000 case IORING_OP_RENAMEAT:
6001 putname(req->rename.oldpath);
6002 putname(req->rename.newpath);
6004 case IORING_OP_UNLINKAT:
6005 putname(req->unlink.filename);
6008 req->flags &= ~REQ_F_NEED_CLEANUP;
6012 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6014 struct io_ring_ctx *ctx = req->ctx;
6015 const struct cred *creds = NULL;
6018 if (req->work.creds && req->work.creds != current_cred())
6019 creds = override_creds(req->work.creds);
6021 switch (req->opcode) {
6023 ret = io_nop(req, issue_flags);
6025 case IORING_OP_READV:
6026 case IORING_OP_READ_FIXED:
6027 case IORING_OP_READ:
6028 ret = io_read(req, issue_flags);
6030 case IORING_OP_WRITEV:
6031 case IORING_OP_WRITE_FIXED:
6032 case IORING_OP_WRITE:
6033 ret = io_write(req, issue_flags);
6035 case IORING_OP_FSYNC:
6036 ret = io_fsync(req, issue_flags);
6038 case IORING_OP_POLL_ADD:
6039 ret = io_poll_add(req, issue_flags);
6041 case IORING_OP_POLL_REMOVE:
6042 ret = io_poll_remove(req, issue_flags);
6044 case IORING_OP_SYNC_FILE_RANGE:
6045 ret = io_sync_file_range(req, issue_flags);
6047 case IORING_OP_SENDMSG:
6048 ret = io_sendmsg(req, issue_flags);
6050 case IORING_OP_SEND:
6051 ret = io_send(req, issue_flags);
6053 case IORING_OP_RECVMSG:
6054 ret = io_recvmsg(req, issue_flags);
6056 case IORING_OP_RECV:
6057 ret = io_recv(req, issue_flags);
6059 case IORING_OP_TIMEOUT:
6060 ret = io_timeout(req, issue_flags);
6062 case IORING_OP_TIMEOUT_REMOVE:
6063 ret = io_timeout_remove(req, issue_flags);
6065 case IORING_OP_ACCEPT:
6066 ret = io_accept(req, issue_flags);
6068 case IORING_OP_CONNECT:
6069 ret = io_connect(req, issue_flags);
6071 case IORING_OP_ASYNC_CANCEL:
6072 ret = io_async_cancel(req, issue_flags);
6074 case IORING_OP_FALLOCATE:
6075 ret = io_fallocate(req, issue_flags);
6077 case IORING_OP_OPENAT:
6078 ret = io_openat(req, issue_flags);
6080 case IORING_OP_CLOSE:
6081 ret = io_close(req, issue_flags);
6083 case IORING_OP_FILES_UPDATE:
6084 ret = io_files_update(req, issue_flags);
6086 case IORING_OP_STATX:
6087 ret = io_statx(req, issue_flags);
6089 case IORING_OP_FADVISE:
6090 ret = io_fadvise(req, issue_flags);
6092 case IORING_OP_MADVISE:
6093 ret = io_madvise(req, issue_flags);
6095 case IORING_OP_OPENAT2:
6096 ret = io_openat2(req, issue_flags);
6098 case IORING_OP_EPOLL_CTL:
6099 ret = io_epoll_ctl(req, issue_flags);
6101 case IORING_OP_SPLICE:
6102 ret = io_splice(req, issue_flags);
6104 case IORING_OP_PROVIDE_BUFFERS:
6105 ret = io_provide_buffers(req, issue_flags);
6107 case IORING_OP_REMOVE_BUFFERS:
6108 ret = io_remove_buffers(req, issue_flags);
6111 ret = io_tee(req, issue_flags);
6113 case IORING_OP_SHUTDOWN:
6114 ret = io_shutdown(req, issue_flags);
6116 case IORING_OP_RENAMEAT:
6117 ret = io_renameat(req, issue_flags);
6119 case IORING_OP_UNLINKAT:
6120 ret = io_unlinkat(req, issue_flags);
6128 revert_creds(creds);
6133 /* If the op doesn't have a file, we're not polling for it */
6134 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6135 const bool in_async = io_wq_current_is_worker();
6137 /* workqueue context doesn't hold uring_lock, grab it now */
6139 mutex_lock(&ctx->uring_lock);
6141 io_iopoll_req_issued(req, in_async);
6144 mutex_unlock(&ctx->uring_lock);
6150 static void io_wq_submit_work(struct io_wq_work *work)
6152 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6153 struct io_kiocb *timeout;
6156 timeout = io_prep_linked_timeout(req);
6158 io_queue_linked_timeout(timeout);
6160 if (work->flags & IO_WQ_WORK_CANCEL)
6165 ret = io_issue_sqe(req, 0);
6167 * We can get EAGAIN for polled IO even though we're
6168 * forcing a sync submission from here, since we can't
6169 * wait for request slots on the block side.
6177 /* avoid locking problems by failing it from a clean context */
6179 /* io-wq is going to take one down */
6181 io_req_task_queue_fail(req, ret);
6185 static inline struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
6188 struct fixed_rsrc_table *table;
6190 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6191 return &table->files[i & IORING_FILE_TABLE_MASK];
6194 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6197 return *io_fixed_file_slot(ctx->file_data, index);
6200 static struct file *io_file_get(struct io_submit_state *state,
6201 struct io_kiocb *req, int fd, bool fixed)
6203 struct io_ring_ctx *ctx = req->ctx;
6207 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6209 fd = array_index_nospec(fd, ctx->nr_user_files);
6210 file = io_file_from_index(ctx, fd);
6211 io_set_resource_node(req);
6213 trace_io_uring_file_get(ctx, fd);
6214 file = __io_file_get(state, fd);
6217 if (file && unlikely(file->f_op == &io_uring_fops))
6218 io_req_track_inflight(req);
6222 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6224 struct io_timeout_data *data = container_of(timer,
6225 struct io_timeout_data, timer);
6226 struct io_kiocb *prev, *req = data->req;
6227 struct io_ring_ctx *ctx = req->ctx;
6228 unsigned long flags;
6230 spin_lock_irqsave(&ctx->completion_lock, flags);
6231 prev = req->timeout.head;
6232 req->timeout.head = NULL;
6235 * We don't expect the list to be empty, that will only happen if we
6236 * race with the completion of the linked work.
6238 if (prev && req_ref_inc_not_zero(prev))
6239 io_remove_next_linked(prev);
6242 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6245 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6246 io_put_req_deferred(prev, 1);
6248 io_req_complete_post(req, -ETIME, 0);
6249 io_put_req_deferred(req, 1);
6251 return HRTIMER_NORESTART;
6254 static void __io_queue_linked_timeout(struct io_kiocb *req)
6257 * If the back reference is NULL, then our linked request finished
6258 * before we got a chance to setup the timer
6260 if (req->timeout.head) {
6261 struct io_timeout_data *data = req->async_data;
6263 data->timer.function = io_link_timeout_fn;
6264 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6269 static void io_queue_linked_timeout(struct io_kiocb *req)
6271 struct io_ring_ctx *ctx = req->ctx;
6273 spin_lock_irq(&ctx->completion_lock);
6274 __io_queue_linked_timeout(req);
6275 spin_unlock_irq(&ctx->completion_lock);
6277 /* drop submission reference */
6281 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6283 struct io_kiocb *nxt = req->link;
6285 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6286 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6289 nxt->timeout.head = req;
6290 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6291 req->flags |= REQ_F_LINK_TIMEOUT;
6295 static void __io_queue_sqe(struct io_kiocb *req)
6297 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6300 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6303 * We async punt it if the file wasn't marked NOWAIT, or if the file
6304 * doesn't support non-blocking read/write attempts
6306 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6307 if (!io_arm_poll_handler(req)) {
6309 * Queued up for async execution, worker will release
6310 * submit reference when the iocb is actually submitted.
6312 io_queue_async_work(req);
6314 } else if (likely(!ret)) {
6315 /* drop submission reference */
6316 if (req->flags & REQ_F_COMPLETE_INLINE) {
6317 struct io_ring_ctx *ctx = req->ctx;
6318 struct io_comp_state *cs = &ctx->submit_state.comp;
6320 cs->reqs[cs->nr++] = req;
6321 if (cs->nr == ARRAY_SIZE(cs->reqs))
6322 io_submit_flush_completions(cs, ctx);
6327 io_req_complete_failed(req, ret);
6330 io_queue_linked_timeout(linked_timeout);
6333 static void io_queue_sqe(struct io_kiocb *req)
6337 ret = io_req_defer(req);
6339 if (ret != -EIOCBQUEUED) {
6341 io_req_complete_failed(req, ret);
6343 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6344 ret = io_req_prep_async(req);
6347 io_queue_async_work(req);
6349 __io_queue_sqe(req);
6354 * Check SQE restrictions (opcode and flags).
6356 * Returns 'true' if SQE is allowed, 'false' otherwise.
6358 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6359 struct io_kiocb *req,
6360 unsigned int sqe_flags)
6362 if (!ctx->restricted)
6365 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6368 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6369 ctx->restrictions.sqe_flags_required)
6372 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6373 ctx->restrictions.sqe_flags_required))
6379 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6380 const struct io_uring_sqe *sqe)
6382 struct io_submit_state *state;
6383 unsigned int sqe_flags;
6384 int personality, ret = 0;
6386 req->opcode = READ_ONCE(sqe->opcode);
6387 /* same numerical values with corresponding REQ_F_*, safe to copy */
6388 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6389 req->user_data = READ_ONCE(sqe->user_data);
6390 req->async_data = NULL;
6394 req->fixed_rsrc_refs = NULL;
6395 /* one is dropped after submission, the other at completion */
6396 atomic_set(&req->refs, 2);
6397 req->task = current;
6399 req->work.list.next = NULL;
6400 req->work.creds = NULL;
6401 req->work.flags = 0;
6403 /* enforce forwards compatibility on users */
6404 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6409 if (unlikely(req->opcode >= IORING_OP_LAST))
6412 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6415 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6416 !io_op_defs[req->opcode].buffer_select)
6419 personality = READ_ONCE(sqe->personality);
6421 req->work.creds = xa_load(&ctx->personalities, personality);
6422 if (!req->work.creds)
6424 get_cred(req->work.creds);
6426 state = &ctx->submit_state;
6429 * Plug now if we have more than 1 IO left after this, and the target
6430 * is potentially a read/write to block based storage.
6432 if (!state->plug_started && state->ios_left > 1 &&
6433 io_op_defs[req->opcode].plug) {
6434 blk_start_plug(&state->plug);
6435 state->plug_started = true;
6438 if (io_op_defs[req->opcode].needs_file) {
6439 bool fixed = req->flags & REQ_F_FIXED_FILE;
6441 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6442 if (unlikely(!req->file))
6450 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6451 const struct io_uring_sqe *sqe)
6453 struct io_submit_link *link = &ctx->submit_state.link;
6456 ret = io_init_req(ctx, req, sqe);
6457 if (unlikely(ret)) {
6460 /* fail even hard links since we don't submit */
6461 link->head->flags |= REQ_F_FAIL_LINK;
6462 io_req_complete_failed(link->head, -ECANCELED);
6465 io_req_complete_failed(req, ret);
6468 ret = io_req_prep(req, sqe);
6472 /* don't need @sqe from now on */
6473 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6474 true, ctx->flags & IORING_SETUP_SQPOLL);
6477 * If we already have a head request, queue this one for async
6478 * submittal once the head completes. If we don't have a head but
6479 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6480 * submitted sync once the chain is complete. If none of those
6481 * conditions are true (normal request), then just queue it.
6484 struct io_kiocb *head = link->head;
6487 * Taking sequential execution of a link, draining both sides
6488 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6489 * requests in the link. So, it drains the head and the
6490 * next after the link request. The last one is done via
6491 * drain_next flag to persist the effect across calls.
6493 if (req->flags & REQ_F_IO_DRAIN) {
6494 head->flags |= REQ_F_IO_DRAIN;
6495 ctx->drain_next = 1;
6497 ret = io_req_prep_async(req);
6500 trace_io_uring_link(ctx, req, head);
6501 link->last->link = req;
6504 /* last request of a link, enqueue the link */
6505 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6510 if (unlikely(ctx->drain_next)) {
6511 req->flags |= REQ_F_IO_DRAIN;
6512 ctx->drain_next = 0;
6514 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6526 * Batched submission is done, ensure local IO is flushed out.
6528 static void io_submit_state_end(struct io_submit_state *state,
6529 struct io_ring_ctx *ctx)
6531 if (state->link.head)
6532 io_queue_sqe(state->link.head);
6534 io_submit_flush_completions(&state->comp, ctx);
6535 if (state->plug_started)
6536 blk_finish_plug(&state->plug);
6537 io_state_file_put(state);
6541 * Start submission side cache.
6543 static void io_submit_state_start(struct io_submit_state *state,
6544 unsigned int max_ios)
6546 state->plug_started = false;
6547 state->ios_left = max_ios;
6548 /* set only head, no need to init link_last in advance */
6549 state->link.head = NULL;
6552 static void io_commit_sqring(struct io_ring_ctx *ctx)
6554 struct io_rings *rings = ctx->rings;
6557 * Ensure any loads from the SQEs are done at this point,
6558 * since once we write the new head, the application could
6559 * write new data to them.
6561 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6565 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6566 * that is mapped by userspace. This means that care needs to be taken to
6567 * ensure that reads are stable, as we cannot rely on userspace always
6568 * being a good citizen. If members of the sqe are validated and then later
6569 * used, it's important that those reads are done through READ_ONCE() to
6570 * prevent a re-load down the line.
6572 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6574 u32 *sq_array = ctx->sq_array;
6578 * The cached sq head (or cq tail) serves two purposes:
6580 * 1) allows us to batch the cost of updating the user visible
6582 * 2) allows the kernel side to track the head on its own, even
6583 * though the application is the one updating it.
6585 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6586 if (likely(head < ctx->sq_entries))
6587 return &ctx->sq_sqes[head];
6589 /* drop invalid entries */
6590 ctx->cached_sq_dropped++;
6591 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6595 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6599 /* if we have a backlog and couldn't flush it all, return BUSY */
6600 if (test_bit(0, &ctx->sq_check_overflow)) {
6601 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6605 /* make sure SQ entry isn't read before tail */
6606 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6608 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6611 percpu_counter_add(¤t->io_uring->inflight, nr);
6612 refcount_add(nr, ¤t->usage);
6613 io_submit_state_start(&ctx->submit_state, nr);
6615 while (submitted < nr) {
6616 const struct io_uring_sqe *sqe;
6617 struct io_kiocb *req;
6619 req = io_alloc_req(ctx);
6620 if (unlikely(!req)) {
6622 submitted = -EAGAIN;
6625 sqe = io_get_sqe(ctx);
6626 if (unlikely(!sqe)) {
6627 kmem_cache_free(req_cachep, req);
6630 /* will complete beyond this point, count as submitted */
6632 if (io_submit_sqe(ctx, req, sqe))
6636 if (unlikely(submitted != nr)) {
6637 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6638 struct io_uring_task *tctx = current->io_uring;
6639 int unused = nr - ref_used;
6641 percpu_ref_put_many(&ctx->refs, unused);
6642 percpu_counter_sub(&tctx->inflight, unused);
6643 put_task_struct_many(current, unused);
6646 io_submit_state_end(&ctx->submit_state, ctx);
6647 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6648 io_commit_sqring(ctx);
6653 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6655 /* Tell userspace we may need a wakeup call */
6656 spin_lock_irq(&ctx->completion_lock);
6657 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6658 spin_unlock_irq(&ctx->completion_lock);
6661 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6663 spin_lock_irq(&ctx->completion_lock);
6664 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6665 spin_unlock_irq(&ctx->completion_lock);
6668 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6670 unsigned int to_submit;
6673 to_submit = io_sqring_entries(ctx);
6674 /* if we're handling multiple rings, cap submit size for fairness */
6675 if (cap_entries && to_submit > 8)
6678 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6679 unsigned nr_events = 0;
6681 mutex_lock(&ctx->uring_lock);
6682 if (!list_empty(&ctx->iopoll_list))
6683 io_do_iopoll(ctx, &nr_events, 0);
6685 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6686 !(ctx->flags & IORING_SETUP_R_DISABLED))
6687 ret = io_submit_sqes(ctx, to_submit);
6688 mutex_unlock(&ctx->uring_lock);
6691 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6692 wake_up(&ctx->sqo_sq_wait);
6697 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6699 struct io_ring_ctx *ctx;
6700 unsigned sq_thread_idle = 0;
6702 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6703 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6704 sqd->sq_thread_idle = sq_thread_idle;
6707 static int io_sq_thread(void *data)
6709 struct io_sq_data *sqd = data;
6710 struct io_ring_ctx *ctx;
6711 unsigned long timeout = 0;
6712 char buf[TASK_COMM_LEN];
6715 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6716 set_task_comm(current, buf);
6717 current->pf_io_worker = NULL;
6719 if (sqd->sq_cpu != -1)
6720 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6722 set_cpus_allowed_ptr(current, cpu_online_mask);
6723 current->flags |= PF_NO_SETAFFINITY;
6725 mutex_lock(&sqd->lock);
6726 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6728 bool cap_entries, sqt_spin, needs_sched;
6730 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6731 signal_pending(current)) {
6732 bool did_sig = false;
6734 mutex_unlock(&sqd->lock);
6735 if (signal_pending(current)) {
6736 struct ksignal ksig;
6738 did_sig = get_signal(&ksig);
6741 mutex_lock(&sqd->lock);
6745 io_run_task_work_head(&sqd->park_task_work);
6746 timeout = jiffies + sqd->sq_thread_idle;
6750 cap_entries = !list_is_singular(&sqd->ctx_list);
6751 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6752 const struct cred *creds = NULL;
6754 if (ctx->sq_creds != current_cred())
6755 creds = override_creds(ctx->sq_creds);
6756 ret = __io_sq_thread(ctx, cap_entries);
6758 revert_creds(creds);
6759 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6763 if (sqt_spin || !time_after(jiffies, timeout)) {
6767 timeout = jiffies + sqd->sq_thread_idle;
6772 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6773 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6774 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6775 !list_empty_careful(&ctx->iopoll_list)) {
6776 needs_sched = false;
6779 if (io_sqring_entries(ctx)) {
6780 needs_sched = false;
6785 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6786 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6787 io_ring_set_wakeup_flag(ctx);
6789 mutex_unlock(&sqd->lock);
6791 mutex_lock(&sqd->lock);
6792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6793 io_ring_clear_wakeup_flag(ctx);
6796 finish_wait(&sqd->wait, &wait);
6797 io_run_task_work_head(&sqd->park_task_work);
6798 timeout = jiffies + sqd->sq_thread_idle;
6801 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6802 io_uring_cancel_sqpoll(ctx);
6804 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6805 io_ring_set_wakeup_flag(ctx);
6806 mutex_unlock(&sqd->lock);
6809 io_run_task_work_head(&sqd->park_task_work);
6810 complete(&sqd->exited);
6814 struct io_wait_queue {
6815 struct wait_queue_entry wq;
6816 struct io_ring_ctx *ctx;
6818 unsigned nr_timeouts;
6821 static inline bool io_should_wake(struct io_wait_queue *iowq)
6823 struct io_ring_ctx *ctx = iowq->ctx;
6826 * Wake up if we have enough events, or if a timeout occurred since we
6827 * started waiting. For timeouts, we always want to return to userspace,
6828 * regardless of event count.
6830 return io_cqring_events(ctx) >= iowq->to_wait ||
6831 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6834 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6835 int wake_flags, void *key)
6837 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6841 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6842 * the task, and the next invocation will do it.
6844 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6845 return autoremove_wake_function(curr, mode, wake_flags, key);
6849 static int io_run_task_work_sig(void)
6851 if (io_run_task_work())
6853 if (!signal_pending(current))
6855 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6856 return -ERESTARTSYS;
6860 /* when returns >0, the caller should retry */
6861 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6862 struct io_wait_queue *iowq,
6863 signed long *timeout)
6867 /* make sure we run task_work before checking for signals */
6868 ret = io_run_task_work_sig();
6869 if (ret || io_should_wake(iowq))
6871 /* let the caller flush overflows, retry */
6872 if (test_bit(0, &ctx->cq_check_overflow))
6875 *timeout = schedule_timeout(*timeout);
6876 return !*timeout ? -ETIME : 1;
6880 * Wait until events become available, if we don't already have some. The
6881 * application must reap them itself, as they reside on the shared cq ring.
6883 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6884 const sigset_t __user *sig, size_t sigsz,
6885 struct __kernel_timespec __user *uts)
6887 struct io_wait_queue iowq = {
6890 .func = io_wake_function,
6891 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6894 .to_wait = min_events,
6896 struct io_rings *rings = ctx->rings;
6897 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6901 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6902 if (io_cqring_events(ctx) >= min_events)
6904 if (!io_run_task_work())
6909 #ifdef CONFIG_COMPAT
6910 if (in_compat_syscall())
6911 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6915 ret = set_user_sigmask(sig, sigsz);
6922 struct timespec64 ts;
6924 if (get_timespec64(&ts, uts))
6926 timeout = timespec64_to_jiffies(&ts);
6929 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6930 trace_io_uring_cqring_wait(ctx, min_events);
6932 /* if we can't even flush overflow, don't wait for more */
6933 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6937 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6938 TASK_INTERRUPTIBLE);
6939 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6940 finish_wait(&ctx->wait, &iowq.wq);
6944 restore_saved_sigmask_unless(ret == -EINTR);
6946 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6949 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6951 #if defined(CONFIG_UNIX)
6952 if (ctx->ring_sock) {
6953 struct sock *sock = ctx->ring_sock->sk;
6954 struct sk_buff *skb;
6956 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6962 for (i = 0; i < ctx->nr_user_files; i++) {
6965 file = io_file_from_index(ctx, i);
6972 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6974 struct fixed_rsrc_data *data;
6976 data = container_of(ref, struct fixed_rsrc_data, refs);
6977 complete(&data->done);
6980 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6982 spin_lock_bh(&ctx->rsrc_ref_lock);
6985 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6987 spin_unlock_bh(&ctx->rsrc_ref_lock);
6990 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6991 struct fixed_rsrc_data *rsrc_data,
6992 struct fixed_rsrc_ref_node *ref_node)
6994 io_rsrc_ref_lock(ctx);
6995 rsrc_data->node = ref_node;
6996 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6997 io_rsrc_ref_unlock(ctx);
6998 percpu_ref_get(&rsrc_data->refs);
7001 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7003 struct fixed_rsrc_ref_node *ref_node = NULL;
7005 io_rsrc_ref_lock(ctx);
7006 ref_node = data->node;
7008 io_rsrc_ref_unlock(ctx);
7010 percpu_ref_kill(&ref_node->refs);
7013 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7014 struct io_ring_ctx *ctx,
7015 void (*rsrc_put)(struct io_ring_ctx *ctx,
7016 struct io_rsrc_put *prsrc))
7018 struct fixed_rsrc_ref_node *backup_node;
7024 data->quiesce = true;
7027 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7030 backup_node->rsrc_data = data;
7031 backup_node->rsrc_put = rsrc_put;
7033 io_sqe_rsrc_kill_node(ctx, data);
7034 percpu_ref_kill(&data->refs);
7035 flush_delayed_work(&ctx->rsrc_put_work);
7037 ret = wait_for_completion_interruptible(&data->done);
7041 percpu_ref_resurrect(&data->refs);
7042 io_sqe_rsrc_set_node(ctx, data, backup_node);
7044 reinit_completion(&data->done);
7045 mutex_unlock(&ctx->uring_lock);
7046 ret = io_run_task_work_sig();
7047 mutex_lock(&ctx->uring_lock);
7049 data->quiesce = false;
7052 destroy_fixed_rsrc_ref_node(backup_node);
7056 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7058 struct fixed_rsrc_data *data;
7060 data = kzalloc(sizeof(*data), GFP_KERNEL);
7064 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7065 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7070 init_completion(&data->done);
7074 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7076 percpu_ref_exit(&data->refs);
7081 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7083 struct fixed_rsrc_data *data = ctx->file_data;
7084 unsigned nr_tables, i;
7088 * percpu_ref_is_dying() is to stop parallel files unregister
7089 * Since we possibly drop uring lock later in this function to
7092 if (!data || percpu_ref_is_dying(&data->refs))
7094 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7098 __io_sqe_files_unregister(ctx);
7099 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7100 for (i = 0; i < nr_tables; i++)
7101 kfree(data->table[i].files);
7102 free_fixed_rsrc_data(data);
7103 ctx->file_data = NULL;
7104 ctx->nr_user_files = 0;
7108 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7109 __releases(&sqd->lock)
7111 WARN_ON_ONCE(sqd->thread == current);
7114 * Do the dance but not conditional clear_bit() because it'd race with
7115 * other threads incrementing park_pending and setting the bit.
7117 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7118 if (atomic_dec_return(&sqd->park_pending))
7119 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7120 mutex_unlock(&sqd->lock);
7123 static void io_sq_thread_park(struct io_sq_data *sqd)
7124 __acquires(&sqd->lock)
7126 WARN_ON_ONCE(sqd->thread == current);
7128 atomic_inc(&sqd->park_pending);
7129 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7130 mutex_lock(&sqd->lock);
7132 wake_up_process(sqd->thread);
7135 static void io_sq_thread_stop(struct io_sq_data *sqd)
7137 WARN_ON_ONCE(sqd->thread == current);
7139 mutex_lock(&sqd->lock);
7140 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7142 wake_up_process(sqd->thread);
7143 mutex_unlock(&sqd->lock);
7144 wait_for_completion(&sqd->exited);
7147 static void io_put_sq_data(struct io_sq_data *sqd)
7149 if (refcount_dec_and_test(&sqd->refs)) {
7150 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7152 io_sq_thread_stop(sqd);
7157 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7159 struct io_sq_data *sqd = ctx->sq_data;
7162 io_sq_thread_park(sqd);
7163 list_del_init(&ctx->sqd_list);
7164 io_sqd_update_thread_idle(sqd);
7165 io_sq_thread_unpark(sqd);
7167 io_put_sq_data(sqd);
7168 ctx->sq_data = NULL;
7170 put_cred(ctx->sq_creds);
7174 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7176 struct io_ring_ctx *ctx_attach;
7177 struct io_sq_data *sqd;
7180 f = fdget(p->wq_fd);
7182 return ERR_PTR(-ENXIO);
7183 if (f.file->f_op != &io_uring_fops) {
7185 return ERR_PTR(-EINVAL);
7188 ctx_attach = f.file->private_data;
7189 sqd = ctx_attach->sq_data;
7192 return ERR_PTR(-EINVAL);
7194 if (sqd->task_tgid != current->tgid) {
7196 return ERR_PTR(-EPERM);
7199 refcount_inc(&sqd->refs);
7204 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7207 struct io_sq_data *sqd;
7210 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7211 sqd = io_attach_sq_data(p);
7216 /* fall through for EPERM case, setup new sqd/task */
7217 if (PTR_ERR(sqd) != -EPERM)
7221 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7223 return ERR_PTR(-ENOMEM);
7225 atomic_set(&sqd->park_pending, 0);
7226 refcount_set(&sqd->refs, 1);
7227 INIT_LIST_HEAD(&sqd->ctx_list);
7228 mutex_init(&sqd->lock);
7229 init_waitqueue_head(&sqd->wait);
7230 init_completion(&sqd->exited);
7234 #if defined(CONFIG_UNIX)
7236 * Ensure the UNIX gc is aware of our file set, so we are certain that
7237 * the io_uring can be safely unregistered on process exit, even if we have
7238 * loops in the file referencing.
7240 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7242 struct sock *sk = ctx->ring_sock->sk;
7243 struct scm_fp_list *fpl;
7244 struct sk_buff *skb;
7247 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7251 skb = alloc_skb(0, GFP_KERNEL);
7260 fpl->user = get_uid(current_user());
7261 for (i = 0; i < nr; i++) {
7262 struct file *file = io_file_from_index(ctx, i + offset);
7266 fpl->fp[nr_files] = get_file(file);
7267 unix_inflight(fpl->user, fpl->fp[nr_files]);
7272 fpl->max = SCM_MAX_FD;
7273 fpl->count = nr_files;
7274 UNIXCB(skb).fp = fpl;
7275 skb->destructor = unix_destruct_scm;
7276 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7277 skb_queue_head(&sk->sk_receive_queue, skb);
7279 for (i = 0; i < nr_files; i++)
7290 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7291 * causes regular reference counting to break down. We rely on the UNIX
7292 * garbage collection to take care of this problem for us.
7294 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7296 unsigned left, total;
7300 left = ctx->nr_user_files;
7302 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7304 ret = __io_sqe_files_scm(ctx, this_files, total);
7308 total += this_files;
7314 while (total < ctx->nr_user_files) {
7315 struct file *file = io_file_from_index(ctx, total);
7325 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7331 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7332 unsigned nr_tables, unsigned nr_files)
7336 for (i = 0; i < nr_tables; i++) {
7337 struct fixed_rsrc_table *table = &file_data->table[i];
7338 unsigned this_files;
7340 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7341 table->files = kcalloc(this_files, sizeof(struct file *),
7345 nr_files -= this_files;
7351 for (i = 0; i < nr_tables; i++) {
7352 struct fixed_rsrc_table *table = &file_data->table[i];
7353 kfree(table->files);
7358 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7360 struct file *file = prsrc->file;
7361 #if defined(CONFIG_UNIX)
7362 struct sock *sock = ctx->ring_sock->sk;
7363 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7364 struct sk_buff *skb;
7367 __skb_queue_head_init(&list);
7370 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7371 * remove this entry and rearrange the file array.
7373 skb = skb_dequeue(head);
7375 struct scm_fp_list *fp;
7377 fp = UNIXCB(skb).fp;
7378 for (i = 0; i < fp->count; i++) {
7381 if (fp->fp[i] != file)
7384 unix_notinflight(fp->user, fp->fp[i]);
7385 left = fp->count - 1 - i;
7387 memmove(&fp->fp[i], &fp->fp[i + 1],
7388 left * sizeof(struct file *));
7395 __skb_queue_tail(&list, skb);
7405 __skb_queue_tail(&list, skb);
7407 skb = skb_dequeue(head);
7410 if (skb_peek(&list)) {
7411 spin_lock_irq(&head->lock);
7412 while ((skb = __skb_dequeue(&list)) != NULL)
7413 __skb_queue_tail(head, skb);
7414 spin_unlock_irq(&head->lock);
7421 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7423 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7424 struct io_ring_ctx *ctx = rsrc_data->ctx;
7425 struct io_rsrc_put *prsrc, *tmp;
7427 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7428 list_del(&prsrc->list);
7429 ref_node->rsrc_put(ctx, prsrc);
7433 percpu_ref_exit(&ref_node->refs);
7435 percpu_ref_put(&rsrc_data->refs);
7438 static void io_rsrc_put_work(struct work_struct *work)
7440 struct io_ring_ctx *ctx;
7441 struct llist_node *node;
7443 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7444 node = llist_del_all(&ctx->rsrc_put_llist);
7447 struct fixed_rsrc_ref_node *ref_node;
7448 struct llist_node *next = node->next;
7450 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7451 __io_rsrc_put_work(ref_node);
7456 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7458 struct fixed_rsrc_ref_node *ref_node;
7459 struct fixed_rsrc_data *data;
7460 struct io_ring_ctx *ctx;
7461 bool first_add = false;
7464 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7465 data = ref_node->rsrc_data;
7468 io_rsrc_ref_lock(ctx);
7469 ref_node->done = true;
7471 while (!list_empty(&ctx->rsrc_ref_list)) {
7472 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7473 struct fixed_rsrc_ref_node, node);
7474 /* recycle ref nodes in order */
7475 if (!ref_node->done)
7477 list_del(&ref_node->node);
7478 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7480 io_rsrc_ref_unlock(ctx);
7482 if (percpu_ref_is_dying(&data->refs))
7486 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7488 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7491 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7492 struct io_ring_ctx *ctx)
7494 struct fixed_rsrc_ref_node *ref_node;
7496 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7500 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7505 INIT_LIST_HEAD(&ref_node->node);
7506 INIT_LIST_HEAD(&ref_node->rsrc_list);
7507 ref_node->done = false;
7511 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7512 struct fixed_rsrc_ref_node *ref_node)
7514 ref_node->rsrc_data = ctx->file_data;
7515 ref_node->rsrc_put = io_ring_file_put;
7518 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7520 percpu_ref_exit(&ref_node->refs);
7525 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7528 __s32 __user *fds = (__s32 __user *) arg;
7529 unsigned nr_tables, i;
7531 int fd, ret = -ENOMEM;
7532 struct fixed_rsrc_ref_node *ref_node;
7533 struct fixed_rsrc_data *file_data;
7539 if (nr_args > IORING_MAX_FIXED_FILES)
7542 file_data = alloc_fixed_rsrc_data(ctx);
7545 ctx->file_data = file_data;
7547 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7548 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7550 if (!file_data->table)
7553 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7556 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7557 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7561 /* allow sparse sets */
7571 * Don't allow io_uring instances to be registered. If UNIX
7572 * isn't enabled, then this causes a reference cycle and this
7573 * instance can never get freed. If UNIX is enabled we'll
7574 * handle it just fine, but there's still no point in allowing
7575 * a ring fd as it doesn't support regular read/write anyway.
7577 if (file->f_op == &io_uring_fops) {
7581 *io_fixed_file_slot(file_data, i) = file;
7584 ret = io_sqe_files_scm(ctx);
7586 io_sqe_files_unregister(ctx);
7590 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7592 io_sqe_files_unregister(ctx);
7595 init_fixed_file_ref_node(ctx, ref_node);
7597 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7600 for (i = 0; i < ctx->nr_user_files; i++) {
7601 file = io_file_from_index(ctx, i);
7605 for (i = 0; i < nr_tables; i++)
7606 kfree(file_data->table[i].files);
7607 ctx->nr_user_files = 0;
7609 free_fixed_rsrc_data(ctx->file_data);
7610 ctx->file_data = NULL;
7614 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7617 #if defined(CONFIG_UNIX)
7618 struct sock *sock = ctx->ring_sock->sk;
7619 struct sk_buff_head *head = &sock->sk_receive_queue;
7620 struct sk_buff *skb;
7623 * See if we can merge this file into an existing skb SCM_RIGHTS
7624 * file set. If there's no room, fall back to allocating a new skb
7625 * and filling it in.
7627 spin_lock_irq(&head->lock);
7628 skb = skb_peek(head);
7630 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7632 if (fpl->count < SCM_MAX_FD) {
7633 __skb_unlink(skb, head);
7634 spin_unlock_irq(&head->lock);
7635 fpl->fp[fpl->count] = get_file(file);
7636 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7638 spin_lock_irq(&head->lock);
7639 __skb_queue_head(head, skb);
7644 spin_unlock_irq(&head->lock);
7651 return __io_sqe_files_scm(ctx, 1, index);
7657 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7659 struct io_rsrc_put *prsrc;
7660 struct fixed_rsrc_ref_node *ref_node = data->node;
7662 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7667 list_add(&prsrc->list, &ref_node->rsrc_list);
7672 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7675 return io_queue_rsrc_removal(data, (void *)file);
7678 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7679 struct io_uring_rsrc_update *up,
7682 struct fixed_rsrc_data *data = ctx->file_data;
7683 struct fixed_rsrc_ref_node *ref_node;
7684 struct file *file, **file_slot;
7688 bool needs_switch = false;
7690 if (check_add_overflow(up->offset, nr_args, &done))
7692 if (done > ctx->nr_user_files)
7695 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7698 init_fixed_file_ref_node(ctx, ref_node);
7700 fds = u64_to_user_ptr(up->data);
7701 for (done = 0; done < nr_args; done++) {
7703 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7707 if (fd == IORING_REGISTER_FILES_SKIP)
7710 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7711 file_slot = io_fixed_file_slot(ctx->file_data, i);
7714 err = io_queue_file_removal(data, *file_slot);
7718 needs_switch = true;
7727 * Don't allow io_uring instances to be registered. If
7728 * UNIX isn't enabled, then this causes a reference
7729 * cycle and this instance can never get freed. If UNIX
7730 * is enabled we'll handle it just fine, but there's
7731 * still no point in allowing a ring fd as it doesn't
7732 * support regular read/write anyway.
7734 if (file->f_op == &io_uring_fops) {
7740 err = io_sqe_file_register(ctx, file, i);
7750 percpu_ref_kill(&data->node->refs);
7751 io_sqe_rsrc_set_node(ctx, data, ref_node);
7753 destroy_fixed_rsrc_ref_node(ref_node);
7755 return done ? done : err;
7758 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7761 struct io_uring_rsrc_update up;
7763 if (!ctx->file_data)
7767 if (copy_from_user(&up, arg, sizeof(up)))
7772 return __io_sqe_files_update(ctx, &up, nr_args);
7775 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7779 req = io_put_req_find_next(req);
7780 return req ? &req->work : NULL;
7783 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7785 struct io_wq_hash *hash;
7786 struct io_wq_data data;
7787 unsigned int concurrency;
7789 hash = ctx->hash_map;
7791 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7793 return ERR_PTR(-ENOMEM);
7794 refcount_set(&hash->refs, 1);
7795 init_waitqueue_head(&hash->wait);
7796 ctx->hash_map = hash;
7800 data.free_work = io_free_work;
7801 data.do_work = io_wq_submit_work;
7803 /* Do QD, or 4 * CPUS, whatever is smallest */
7804 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7806 return io_wq_create(concurrency, &data);
7809 static int io_uring_alloc_task_context(struct task_struct *task,
7810 struct io_ring_ctx *ctx)
7812 struct io_uring_task *tctx;
7815 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7816 if (unlikely(!tctx))
7819 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7820 if (unlikely(ret)) {
7825 tctx->io_wq = io_init_wq_offload(ctx);
7826 if (IS_ERR(tctx->io_wq)) {
7827 ret = PTR_ERR(tctx->io_wq);
7828 percpu_counter_destroy(&tctx->inflight);
7834 init_waitqueue_head(&tctx->wait);
7836 atomic_set(&tctx->in_idle, 0);
7837 task->io_uring = tctx;
7838 spin_lock_init(&tctx->task_lock);
7839 INIT_WQ_LIST(&tctx->task_list);
7840 tctx->task_state = 0;
7841 init_task_work(&tctx->task_work, tctx_task_work);
7845 void __io_uring_free(struct task_struct *tsk)
7847 struct io_uring_task *tctx = tsk->io_uring;
7849 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7850 WARN_ON_ONCE(tctx->io_wq);
7852 percpu_counter_destroy(&tctx->inflight);
7854 tsk->io_uring = NULL;
7857 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7858 struct io_uring_params *p)
7862 /* Retain compatibility with failing for an invalid attach attempt */
7863 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7864 IORING_SETUP_ATTACH_WQ) {
7867 f = fdget(p->wq_fd);
7870 if (f.file->f_op != &io_uring_fops) {
7876 if (ctx->flags & IORING_SETUP_SQPOLL) {
7877 struct task_struct *tsk;
7878 struct io_sq_data *sqd;
7882 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7885 sqd = io_get_sq_data(p, &attached);
7891 ctx->sq_creds = get_current_cred();
7893 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7894 if (!ctx->sq_thread_idle)
7895 ctx->sq_thread_idle = HZ;
7898 io_sq_thread_park(sqd);
7899 list_add(&ctx->sqd_list, &sqd->ctx_list);
7900 io_sqd_update_thread_idle(sqd);
7901 /* don't attach to a dying SQPOLL thread, would be racy */
7902 if (attached && !sqd->thread)
7904 io_sq_thread_unpark(sqd);
7911 if (p->flags & IORING_SETUP_SQ_AFF) {
7912 int cpu = p->sq_thread_cpu;
7915 if (cpu >= nr_cpu_ids)
7917 if (!cpu_online(cpu))
7925 sqd->task_pid = current->pid;
7926 sqd->task_tgid = current->tgid;
7927 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7934 ret = io_uring_alloc_task_context(tsk, ctx);
7935 wake_up_new_task(tsk);
7938 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7939 /* Can't have SQ_AFF without SQPOLL */
7946 io_sq_thread_finish(ctx);
7949 complete(&ctx->sq_data->exited);
7953 static inline void __io_unaccount_mem(struct user_struct *user,
7954 unsigned long nr_pages)
7956 atomic_long_sub(nr_pages, &user->locked_vm);
7959 static inline int __io_account_mem(struct user_struct *user,
7960 unsigned long nr_pages)
7962 unsigned long page_limit, cur_pages, new_pages;
7964 /* Don't allow more pages than we can safely lock */
7965 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7968 cur_pages = atomic_long_read(&user->locked_vm);
7969 new_pages = cur_pages + nr_pages;
7970 if (new_pages > page_limit)
7972 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7973 new_pages) != cur_pages);
7978 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7981 __io_unaccount_mem(ctx->user, nr_pages);
7983 if (ctx->mm_account)
7984 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7987 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7992 ret = __io_account_mem(ctx->user, nr_pages);
7997 if (ctx->mm_account)
7998 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8003 static void io_mem_free(void *ptr)
8010 page = virt_to_head_page(ptr);
8011 if (put_page_testzero(page))
8012 free_compound_page(page);
8015 static void *io_mem_alloc(size_t size)
8017 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8018 __GFP_NORETRY | __GFP_ACCOUNT;
8020 return (void *) __get_free_pages(gfp_flags, get_order(size));
8023 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8026 struct io_rings *rings;
8027 size_t off, sq_array_size;
8029 off = struct_size(rings, cqes, cq_entries);
8030 if (off == SIZE_MAX)
8034 off = ALIGN(off, SMP_CACHE_BYTES);
8042 sq_array_size = array_size(sizeof(u32), sq_entries);
8043 if (sq_array_size == SIZE_MAX)
8046 if (check_add_overflow(off, sq_array_size, &off))
8052 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8056 if (!ctx->user_bufs)
8059 for (i = 0; i < ctx->nr_user_bufs; i++) {
8060 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8062 for (j = 0; j < imu->nr_bvecs; j++)
8063 unpin_user_page(imu->bvec[j].bv_page);
8065 if (imu->acct_pages)
8066 io_unaccount_mem(ctx, imu->acct_pages);
8071 kfree(ctx->user_bufs);
8072 ctx->user_bufs = NULL;
8073 ctx->nr_user_bufs = 0;
8077 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8078 void __user *arg, unsigned index)
8080 struct iovec __user *src;
8082 #ifdef CONFIG_COMPAT
8084 struct compat_iovec __user *ciovs;
8085 struct compat_iovec ciov;
8087 ciovs = (struct compat_iovec __user *) arg;
8088 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8091 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8092 dst->iov_len = ciov.iov_len;
8096 src = (struct iovec __user *) arg;
8097 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8103 * Not super efficient, but this is just a registration time. And we do cache
8104 * the last compound head, so generally we'll only do a full search if we don't
8107 * We check if the given compound head page has already been accounted, to
8108 * avoid double accounting it. This allows us to account the full size of the
8109 * page, not just the constituent pages of a huge page.
8111 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8112 int nr_pages, struct page *hpage)
8116 /* check current page array */
8117 for (i = 0; i < nr_pages; i++) {
8118 if (!PageCompound(pages[i]))
8120 if (compound_head(pages[i]) == hpage)
8124 /* check previously registered pages */
8125 for (i = 0; i < ctx->nr_user_bufs; i++) {
8126 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8128 for (j = 0; j < imu->nr_bvecs; j++) {
8129 if (!PageCompound(imu->bvec[j].bv_page))
8131 if (compound_head(imu->bvec[j].bv_page) == hpage)
8139 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8140 int nr_pages, struct io_mapped_ubuf *imu,
8141 struct page **last_hpage)
8145 for (i = 0; i < nr_pages; i++) {
8146 if (!PageCompound(pages[i])) {
8151 hpage = compound_head(pages[i]);
8152 if (hpage == *last_hpage)
8154 *last_hpage = hpage;
8155 if (headpage_already_acct(ctx, pages, i, hpage))
8157 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8161 if (!imu->acct_pages)
8164 ret = io_account_mem(ctx, imu->acct_pages);
8166 imu->acct_pages = 0;
8170 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8171 struct io_mapped_ubuf *imu,
8172 struct page **last_hpage)
8174 struct vm_area_struct **vmas = NULL;
8175 struct page **pages = NULL;
8176 unsigned long off, start, end, ubuf;
8178 int ret, pret, nr_pages, i;
8180 ubuf = (unsigned long) iov->iov_base;
8181 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8182 start = ubuf >> PAGE_SHIFT;
8183 nr_pages = end - start;
8187 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8191 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8196 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8202 mmap_read_lock(current->mm);
8203 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8205 if (pret == nr_pages) {
8206 /* don't support file backed memory */
8207 for (i = 0; i < nr_pages; i++) {
8208 struct vm_area_struct *vma = vmas[i];
8211 !is_file_hugepages(vma->vm_file)) {
8217 ret = pret < 0 ? pret : -EFAULT;
8219 mmap_read_unlock(current->mm);
8222 * if we did partial map, or found file backed vmas,
8223 * release any pages we did get
8226 unpin_user_pages(pages, pret);
8231 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8233 unpin_user_pages(pages, pret);
8238 off = ubuf & ~PAGE_MASK;
8239 size = iov->iov_len;
8240 for (i = 0; i < nr_pages; i++) {
8243 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8244 imu->bvec[i].bv_page = pages[i];
8245 imu->bvec[i].bv_len = vec_len;
8246 imu->bvec[i].bv_offset = off;
8250 /* store original address for later verification */
8252 imu->len = iov->iov_len;
8253 imu->nr_bvecs = nr_pages;
8261 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8265 if (!nr_args || nr_args > UIO_MAXIOV)
8268 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8270 if (!ctx->user_bufs)
8276 static int io_buffer_validate(struct iovec *iov)
8279 * Don't impose further limits on the size and buffer
8280 * constraints here, we'll -EINVAL later when IO is
8281 * submitted if they are wrong.
8283 if (!iov->iov_base || !iov->iov_len)
8286 /* arbitrary limit, but we need something */
8287 if (iov->iov_len > SZ_1G)
8293 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8294 unsigned int nr_args)
8298 struct page *last_hpage = NULL;
8300 ret = io_buffers_map_alloc(ctx, nr_args);
8304 for (i = 0; i < nr_args; i++) {
8305 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8307 ret = io_copy_iov(ctx, &iov, arg, i);
8311 ret = io_buffer_validate(&iov);
8315 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8319 ctx->nr_user_bufs++;
8323 io_sqe_buffers_unregister(ctx);
8328 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8330 __s32 __user *fds = arg;
8336 if (copy_from_user(&fd, fds, sizeof(*fds)))
8339 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8340 if (IS_ERR(ctx->cq_ev_fd)) {
8341 int ret = PTR_ERR(ctx->cq_ev_fd);
8342 ctx->cq_ev_fd = NULL;
8349 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8351 if (ctx->cq_ev_fd) {
8352 eventfd_ctx_put(ctx->cq_ev_fd);
8353 ctx->cq_ev_fd = NULL;
8360 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8362 struct io_buffer *buf;
8363 unsigned long index;
8365 xa_for_each(&ctx->io_buffers, index, buf)
8366 __io_remove_buffers(ctx, buf, index, -1U);
8369 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8371 struct io_kiocb *req, *nxt;
8373 list_for_each_entry_safe(req, nxt, list, compl.list) {
8374 if (tsk && req->task != tsk)
8376 list_del(&req->compl.list);
8377 kmem_cache_free(req_cachep, req);
8381 static void io_req_caches_free(struct io_ring_ctx *ctx)
8383 struct io_submit_state *submit_state = &ctx->submit_state;
8384 struct io_comp_state *cs = &ctx->submit_state.comp;
8386 mutex_lock(&ctx->uring_lock);
8388 if (submit_state->free_reqs) {
8389 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8390 submit_state->reqs);
8391 submit_state->free_reqs = 0;
8394 spin_lock_irq(&ctx->completion_lock);
8395 list_splice_init(&cs->locked_free_list, &cs->free_list);
8396 cs->locked_free_nr = 0;
8397 spin_unlock_irq(&ctx->completion_lock);
8399 io_req_cache_free(&cs->free_list, NULL);
8401 mutex_unlock(&ctx->uring_lock);
8404 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8407 * Some may use context even when all refs and requests have been put,
8408 * and they are free to do so while still holding uring_lock or
8409 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8411 mutex_lock(&ctx->uring_lock);
8412 mutex_unlock(&ctx->uring_lock);
8413 spin_lock_irq(&ctx->completion_lock);
8414 spin_unlock_irq(&ctx->completion_lock);
8416 io_sq_thread_finish(ctx);
8417 io_sqe_buffers_unregister(ctx);
8419 if (ctx->mm_account) {
8420 mmdrop(ctx->mm_account);
8421 ctx->mm_account = NULL;
8424 mutex_lock(&ctx->uring_lock);
8425 io_sqe_files_unregister(ctx);
8426 mutex_unlock(&ctx->uring_lock);
8427 io_eventfd_unregister(ctx);
8428 io_destroy_buffers(ctx);
8430 #if defined(CONFIG_UNIX)
8431 if (ctx->ring_sock) {
8432 ctx->ring_sock->file = NULL; /* so that iput() is called */
8433 sock_release(ctx->ring_sock);
8437 io_mem_free(ctx->rings);
8438 io_mem_free(ctx->sq_sqes);
8440 percpu_ref_exit(&ctx->refs);
8441 free_uid(ctx->user);
8442 io_req_caches_free(ctx);
8444 io_wq_put_hash(ctx->hash_map);
8445 kfree(ctx->cancel_hash);
8449 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8451 struct io_ring_ctx *ctx = file->private_data;
8454 poll_wait(file, &ctx->cq_wait, wait);
8456 * synchronizes with barrier from wq_has_sleeper call in
8460 if (!io_sqring_full(ctx))
8461 mask |= EPOLLOUT | EPOLLWRNORM;
8464 * Don't flush cqring overflow list here, just do a simple check.
8465 * Otherwise there could possible be ABBA deadlock:
8468 * lock(&ctx->uring_lock);
8470 * lock(&ctx->uring_lock);
8473 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8474 * pushs them to do the flush.
8476 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8477 mask |= EPOLLIN | EPOLLRDNORM;
8482 static int io_uring_fasync(int fd, struct file *file, int on)
8484 struct io_ring_ctx *ctx = file->private_data;
8486 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8489 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8491 const struct cred *creds;
8493 creds = xa_erase(&ctx->personalities, id);
8502 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8504 return io_run_task_work_head(&ctx->exit_task_work);
8507 struct io_tctx_exit {
8508 struct callback_head task_work;
8509 struct completion completion;
8510 struct io_ring_ctx *ctx;
8513 static void io_tctx_exit_cb(struct callback_head *cb)
8515 struct io_uring_task *tctx = current->io_uring;
8516 struct io_tctx_exit *work;
8518 work = container_of(cb, struct io_tctx_exit, task_work);
8520 * When @in_idle, we're in cancellation and it's racy to remove the
8521 * node. It'll be removed by the end of cancellation, just ignore it.
8523 if (!atomic_read(&tctx->in_idle))
8524 io_uring_del_task_file((unsigned long)work->ctx);
8525 complete(&work->completion);
8528 static void io_ring_exit_work(struct work_struct *work)
8530 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8531 unsigned long timeout = jiffies + HZ * 60 * 5;
8532 struct io_tctx_exit exit;
8533 struct io_tctx_node *node;
8536 /* prevent SQPOLL from submitting new requests */
8538 io_sq_thread_park(ctx->sq_data);
8539 list_del_init(&ctx->sqd_list);
8540 io_sqd_update_thread_idle(ctx->sq_data);
8541 io_sq_thread_unpark(ctx->sq_data);
8545 * If we're doing polled IO and end up having requests being
8546 * submitted async (out-of-line), then completions can come in while
8547 * we're waiting for refs to drop. We need to reap these manually,
8548 * as nobody else will be looking for them.
8551 io_uring_try_cancel_requests(ctx, NULL, NULL);
8553 WARN_ON_ONCE(time_after(jiffies, timeout));
8554 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8556 mutex_lock(&ctx->uring_lock);
8557 while (!list_empty(&ctx->tctx_list)) {
8558 WARN_ON_ONCE(time_after(jiffies, timeout));
8560 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8563 init_completion(&exit.completion);
8564 init_task_work(&exit.task_work, io_tctx_exit_cb);
8565 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8566 if (WARN_ON_ONCE(ret))
8568 wake_up_process(node->task);
8570 mutex_unlock(&ctx->uring_lock);
8571 wait_for_completion(&exit.completion);
8573 mutex_lock(&ctx->uring_lock);
8575 mutex_unlock(&ctx->uring_lock);
8577 io_ring_ctx_free(ctx);
8580 /* Returns true if we found and killed one or more timeouts */
8581 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8582 struct files_struct *files)
8584 struct io_kiocb *req, *tmp;
8587 spin_lock_irq(&ctx->completion_lock);
8588 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8589 if (io_match_task(req, tsk, files)) {
8590 io_kill_timeout(req, -ECANCELED);
8595 io_commit_cqring(ctx);
8596 spin_unlock_irq(&ctx->completion_lock);
8598 io_cqring_ev_posted(ctx);
8599 return canceled != 0;
8602 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8604 unsigned long index;
8605 struct creds *creds;
8607 mutex_lock(&ctx->uring_lock);
8608 percpu_ref_kill(&ctx->refs);
8609 /* if force is set, the ring is going away. always drop after that */
8610 ctx->cq_overflow_flushed = 1;
8612 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8613 xa_for_each(&ctx->personalities, index, creds)
8614 io_unregister_personality(ctx, index);
8615 mutex_unlock(&ctx->uring_lock);
8617 io_kill_timeouts(ctx, NULL, NULL);
8618 io_poll_remove_all(ctx, NULL, NULL);
8620 /* if we failed setting up the ctx, we might not have any rings */
8621 io_iopoll_try_reap_events(ctx);
8623 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8625 * Use system_unbound_wq to avoid spawning tons of event kworkers
8626 * if we're exiting a ton of rings at the same time. It just adds
8627 * noise and overhead, there's no discernable change in runtime
8628 * over using system_wq.
8630 queue_work(system_unbound_wq, &ctx->exit_work);
8633 static int io_uring_release(struct inode *inode, struct file *file)
8635 struct io_ring_ctx *ctx = file->private_data;
8637 file->private_data = NULL;
8638 io_ring_ctx_wait_and_kill(ctx);
8642 struct io_task_cancel {
8643 struct task_struct *task;
8644 struct files_struct *files;
8647 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8649 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8650 struct io_task_cancel *cancel = data;
8653 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8654 unsigned long flags;
8655 struct io_ring_ctx *ctx = req->ctx;
8657 /* protect against races with linked timeouts */
8658 spin_lock_irqsave(&ctx->completion_lock, flags);
8659 ret = io_match_task(req, cancel->task, cancel->files);
8660 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8662 ret = io_match_task(req, cancel->task, cancel->files);
8667 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8668 struct task_struct *task,
8669 struct files_struct *files)
8671 struct io_defer_entry *de;
8674 spin_lock_irq(&ctx->completion_lock);
8675 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8676 if (io_match_task(de->req, task, files)) {
8677 list_cut_position(&list, &ctx->defer_list, &de->list);
8681 spin_unlock_irq(&ctx->completion_lock);
8682 if (list_empty(&list))
8685 while (!list_empty(&list)) {
8686 de = list_first_entry(&list, struct io_defer_entry, list);
8687 list_del_init(&de->list);
8688 io_req_complete_failed(de->req, -ECANCELED);
8694 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8696 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8698 return req->ctx == data;
8701 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8703 struct io_tctx_node *node;
8704 enum io_wq_cancel cret;
8707 mutex_lock(&ctx->uring_lock);
8708 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8709 struct io_uring_task *tctx = node->task->io_uring;
8712 * io_wq will stay alive while we hold uring_lock, because it's
8713 * killed after ctx nodes, which requires to take the lock.
8715 if (!tctx || !tctx->io_wq)
8717 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8718 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8720 mutex_unlock(&ctx->uring_lock);
8725 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8726 struct task_struct *task,
8727 struct files_struct *files)
8729 struct io_task_cancel cancel = { .task = task, .files = files, };
8730 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8733 enum io_wq_cancel cret;
8737 ret |= io_uring_try_cancel_iowq(ctx);
8738 } else if (tctx && tctx->io_wq) {
8740 * Cancels requests of all rings, not only @ctx, but
8741 * it's fine as the task is in exit/exec.
8743 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8745 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8748 /* SQPOLL thread does its own polling */
8749 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8750 (ctx->sq_data && ctx->sq_data->thread == current)) {
8751 while (!list_empty_careful(&ctx->iopoll_list)) {
8752 io_iopoll_try_reap_events(ctx);
8757 ret |= io_cancel_defer_files(ctx, task, files);
8758 ret |= io_poll_remove_all(ctx, task, files);
8759 ret |= io_kill_timeouts(ctx, task, files);
8760 ret |= io_run_task_work();
8761 ret |= io_run_ctx_fallback(ctx);
8762 io_cqring_overflow_flush(ctx, true, task, files);
8769 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8770 struct task_struct *task,
8771 struct files_struct *files)
8773 struct io_kiocb *req;
8776 spin_lock_irq(&ctx->inflight_lock);
8777 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8778 cnt += io_match_task(req, task, files);
8779 spin_unlock_irq(&ctx->inflight_lock);
8783 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8784 struct task_struct *task,
8785 struct files_struct *files)
8787 while (!list_empty_careful(&ctx->inflight_list)) {
8791 inflight = io_uring_count_inflight(ctx, task, files);
8795 io_uring_try_cancel_requests(ctx, task, files);
8797 prepare_to_wait(&task->io_uring->wait, &wait,
8798 TASK_UNINTERRUPTIBLE);
8799 if (inflight == io_uring_count_inflight(ctx, task, files))
8801 finish_wait(&task->io_uring->wait, &wait);
8806 * Note that this task has used io_uring. We use it for cancelation purposes.
8808 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8810 struct io_uring_task *tctx = current->io_uring;
8811 struct io_tctx_node *node;
8814 if (unlikely(!tctx)) {
8815 ret = io_uring_alloc_task_context(current, ctx);
8818 tctx = current->io_uring;
8820 if (tctx->last != ctx) {
8821 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8824 node = kmalloc(sizeof(*node), GFP_KERNEL);
8828 node->task = current;
8830 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8837 mutex_lock(&ctx->uring_lock);
8838 list_add(&node->ctx_node, &ctx->tctx_list);
8839 mutex_unlock(&ctx->uring_lock);
8847 * Remove this io_uring_file -> task mapping.
8849 static void io_uring_del_task_file(unsigned long index)
8851 struct io_uring_task *tctx = current->io_uring;
8852 struct io_tctx_node *node;
8856 node = xa_erase(&tctx->xa, index);
8860 WARN_ON_ONCE(current != node->task);
8861 WARN_ON_ONCE(list_empty(&node->ctx_node));
8863 mutex_lock(&node->ctx->uring_lock);
8864 list_del(&node->ctx_node);
8865 mutex_unlock(&node->ctx->uring_lock);
8867 if (tctx->last == node->ctx)
8872 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8874 struct io_tctx_node *node;
8875 unsigned long index;
8877 xa_for_each(&tctx->xa, index, node)
8878 io_uring_del_task_file(index);
8880 io_wq_put_and_exit(tctx->io_wq);
8885 static s64 tctx_inflight(struct io_uring_task *tctx)
8887 return percpu_counter_sum(&tctx->inflight);
8890 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8892 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8893 struct io_ring_ctx *ctx = work->ctx;
8894 struct io_sq_data *sqd = ctx->sq_data;
8897 io_uring_cancel_sqpoll(ctx);
8898 complete(&work->completion);
8901 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8903 struct io_sq_data *sqd = ctx->sq_data;
8904 struct io_tctx_exit work = { .ctx = ctx, };
8905 struct task_struct *task;
8907 io_sq_thread_park(sqd);
8908 list_del_init(&ctx->sqd_list);
8909 io_sqd_update_thread_idle(sqd);
8912 init_completion(&work.completion);
8913 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8914 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8915 wake_up_process(task);
8917 io_sq_thread_unpark(sqd);
8920 wait_for_completion(&work.completion);
8923 void __io_uring_files_cancel(struct files_struct *files)
8925 struct io_uring_task *tctx = current->io_uring;
8926 struct io_tctx_node *node;
8927 unsigned long index;
8929 /* make sure overflow events are dropped */
8930 atomic_inc(&tctx->in_idle);
8931 xa_for_each(&tctx->xa, index, node) {
8932 struct io_ring_ctx *ctx = node->ctx;
8935 io_sqpoll_cancel_sync(ctx);
8938 io_uring_cancel_files(ctx, current, files);
8940 io_uring_try_cancel_requests(ctx, current, NULL);
8942 atomic_dec(&tctx->in_idle);
8945 io_uring_clean_tctx(tctx);
8948 /* should only be called by SQPOLL task */
8949 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8951 struct io_sq_data *sqd = ctx->sq_data;
8952 struct io_uring_task *tctx = current->io_uring;
8956 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8958 atomic_inc(&tctx->in_idle);
8960 /* read completions before cancelations */
8961 inflight = tctx_inflight(tctx);
8964 io_uring_try_cancel_requests(ctx, current, NULL);
8966 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8968 * If we've seen completions, retry without waiting. This
8969 * avoids a race where a completion comes in before we did
8970 * prepare_to_wait().
8972 if (inflight == tctx_inflight(tctx))
8974 finish_wait(&tctx->wait, &wait);
8976 atomic_dec(&tctx->in_idle);
8980 * Find any io_uring fd that this task has registered or done IO on, and cancel
8983 void __io_uring_task_cancel(void)
8985 struct io_uring_task *tctx = current->io_uring;
8989 /* make sure overflow events are dropped */
8990 atomic_inc(&tctx->in_idle);
8991 __io_uring_files_cancel(NULL);
8994 /* read completions before cancelations */
8995 inflight = tctx_inflight(tctx);
8998 __io_uring_files_cancel(NULL);
9000 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9003 * If we've seen completions, retry without waiting. This
9004 * avoids a race where a completion comes in before we did
9005 * prepare_to_wait().
9007 if (inflight == tctx_inflight(tctx))
9009 finish_wait(&tctx->wait, &wait);
9012 atomic_dec(&tctx->in_idle);
9014 io_uring_clean_tctx(tctx);
9015 /* all current's requests should be gone, we can kill tctx */
9016 __io_uring_free(current);
9019 static void *io_uring_validate_mmap_request(struct file *file,
9020 loff_t pgoff, size_t sz)
9022 struct io_ring_ctx *ctx = file->private_data;
9023 loff_t offset = pgoff << PAGE_SHIFT;
9028 case IORING_OFF_SQ_RING:
9029 case IORING_OFF_CQ_RING:
9032 case IORING_OFF_SQES:
9036 return ERR_PTR(-EINVAL);
9039 page = virt_to_head_page(ptr);
9040 if (sz > page_size(page))
9041 return ERR_PTR(-EINVAL);
9048 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9050 size_t sz = vma->vm_end - vma->vm_start;
9054 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9056 return PTR_ERR(ptr);
9058 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9059 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9062 #else /* !CONFIG_MMU */
9064 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9066 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9069 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9071 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9074 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9075 unsigned long addr, unsigned long len,
9076 unsigned long pgoff, unsigned long flags)
9080 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9082 return PTR_ERR(ptr);
9084 return (unsigned long) ptr;
9087 #endif /* !CONFIG_MMU */
9089 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9094 if (!io_sqring_full(ctx))
9096 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9098 if (!io_sqring_full(ctx))
9101 } while (!signal_pending(current));
9103 finish_wait(&ctx->sqo_sq_wait, &wait);
9107 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9108 struct __kernel_timespec __user **ts,
9109 const sigset_t __user **sig)
9111 struct io_uring_getevents_arg arg;
9114 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9115 * is just a pointer to the sigset_t.
9117 if (!(flags & IORING_ENTER_EXT_ARG)) {
9118 *sig = (const sigset_t __user *) argp;
9124 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9125 * timespec and sigset_t pointers if good.
9127 if (*argsz != sizeof(arg))
9129 if (copy_from_user(&arg, argp, sizeof(arg)))
9131 *sig = u64_to_user_ptr(arg.sigmask);
9132 *argsz = arg.sigmask_sz;
9133 *ts = u64_to_user_ptr(arg.ts);
9137 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9138 u32, min_complete, u32, flags, const void __user *, argp,
9141 struct io_ring_ctx *ctx;
9148 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9149 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9157 if (f.file->f_op != &io_uring_fops)
9161 ctx = f.file->private_data;
9162 if (!percpu_ref_tryget(&ctx->refs))
9166 if (ctx->flags & IORING_SETUP_R_DISABLED)
9170 * For SQ polling, the thread will do all submissions and completions.
9171 * Just return the requested submit count, and wake the thread if
9175 if (ctx->flags & IORING_SETUP_SQPOLL) {
9176 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9179 if (unlikely(ctx->sq_data->thread == NULL)) {
9182 if (flags & IORING_ENTER_SQ_WAKEUP)
9183 wake_up(&ctx->sq_data->wait);
9184 if (flags & IORING_ENTER_SQ_WAIT) {
9185 ret = io_sqpoll_wait_sq(ctx);
9189 submitted = to_submit;
9190 } else if (to_submit) {
9191 ret = io_uring_add_task_file(ctx);
9194 mutex_lock(&ctx->uring_lock);
9195 submitted = io_submit_sqes(ctx, to_submit);
9196 mutex_unlock(&ctx->uring_lock);
9198 if (submitted != to_submit)
9201 if (flags & IORING_ENTER_GETEVENTS) {
9202 const sigset_t __user *sig;
9203 struct __kernel_timespec __user *ts;
9205 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9209 min_complete = min(min_complete, ctx->cq_entries);
9212 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9213 * space applications don't need to do io completion events
9214 * polling again, they can rely on io_sq_thread to do polling
9215 * work, which can reduce cpu usage and uring_lock contention.
9217 if (ctx->flags & IORING_SETUP_IOPOLL &&
9218 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9219 ret = io_iopoll_check(ctx, min_complete);
9221 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9226 percpu_ref_put(&ctx->refs);
9229 return submitted ? submitted : ret;
9232 #ifdef CONFIG_PROC_FS
9233 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9234 const struct cred *cred)
9236 struct user_namespace *uns = seq_user_ns(m);
9237 struct group_info *gi;
9242 seq_printf(m, "%5d\n", id);
9243 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9244 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9245 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9246 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9247 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9248 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9249 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9250 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9251 seq_puts(m, "\n\tGroups:\t");
9252 gi = cred->group_info;
9253 for (g = 0; g < gi->ngroups; g++) {
9254 seq_put_decimal_ull(m, g ? " " : "",
9255 from_kgid_munged(uns, gi->gid[g]));
9257 seq_puts(m, "\n\tCapEff:\t");
9258 cap = cred->cap_effective;
9259 CAP_FOR_EACH_U32(__capi)
9260 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9265 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9267 struct io_sq_data *sq = NULL;
9272 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9273 * since fdinfo case grabs it in the opposite direction of normal use
9274 * cases. If we fail to get the lock, we just don't iterate any
9275 * structures that could be going away outside the io_uring mutex.
9277 has_lock = mutex_trylock(&ctx->uring_lock);
9279 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9285 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9286 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9287 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9288 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9289 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9292 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9294 seq_printf(m, "%5u: <none>\n", i);
9296 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9297 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9298 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9300 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9301 (unsigned int) buf->len);
9303 if (has_lock && !xa_empty(&ctx->personalities)) {
9304 unsigned long index;
9305 const struct cred *cred;
9307 seq_printf(m, "Personalities:\n");
9308 xa_for_each(&ctx->personalities, index, cred)
9309 io_uring_show_cred(m, index, cred);
9311 seq_printf(m, "PollList:\n");
9312 spin_lock_irq(&ctx->completion_lock);
9313 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9314 struct hlist_head *list = &ctx->cancel_hash[i];
9315 struct io_kiocb *req;
9317 hlist_for_each_entry(req, list, hash_node)
9318 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9319 req->task->task_works != NULL);
9321 spin_unlock_irq(&ctx->completion_lock);
9323 mutex_unlock(&ctx->uring_lock);
9326 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9328 struct io_ring_ctx *ctx = f->private_data;
9330 if (percpu_ref_tryget(&ctx->refs)) {
9331 __io_uring_show_fdinfo(ctx, m);
9332 percpu_ref_put(&ctx->refs);
9337 static const struct file_operations io_uring_fops = {
9338 .release = io_uring_release,
9339 .mmap = io_uring_mmap,
9341 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9342 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9344 .poll = io_uring_poll,
9345 .fasync = io_uring_fasync,
9346 #ifdef CONFIG_PROC_FS
9347 .show_fdinfo = io_uring_show_fdinfo,
9351 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9352 struct io_uring_params *p)
9354 struct io_rings *rings;
9355 size_t size, sq_array_offset;
9357 /* make sure these are sane, as we already accounted them */
9358 ctx->sq_entries = p->sq_entries;
9359 ctx->cq_entries = p->cq_entries;
9361 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9362 if (size == SIZE_MAX)
9365 rings = io_mem_alloc(size);
9370 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9371 rings->sq_ring_mask = p->sq_entries - 1;
9372 rings->cq_ring_mask = p->cq_entries - 1;
9373 rings->sq_ring_entries = p->sq_entries;
9374 rings->cq_ring_entries = p->cq_entries;
9375 ctx->sq_mask = rings->sq_ring_mask;
9376 ctx->cq_mask = rings->cq_ring_mask;
9378 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9379 if (size == SIZE_MAX) {
9380 io_mem_free(ctx->rings);
9385 ctx->sq_sqes = io_mem_alloc(size);
9386 if (!ctx->sq_sqes) {
9387 io_mem_free(ctx->rings);
9395 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9399 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9403 ret = io_uring_add_task_file(ctx);
9408 fd_install(fd, file);
9413 * Allocate an anonymous fd, this is what constitutes the application
9414 * visible backing of an io_uring instance. The application mmaps this
9415 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9416 * we have to tie this fd to a socket for file garbage collection purposes.
9418 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9421 #if defined(CONFIG_UNIX)
9424 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9427 return ERR_PTR(ret);
9430 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9431 O_RDWR | O_CLOEXEC);
9432 #if defined(CONFIG_UNIX)
9434 sock_release(ctx->ring_sock);
9435 ctx->ring_sock = NULL;
9437 ctx->ring_sock->file = file;
9443 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9444 struct io_uring_params __user *params)
9446 struct io_ring_ctx *ctx;
9452 if (entries > IORING_MAX_ENTRIES) {
9453 if (!(p->flags & IORING_SETUP_CLAMP))
9455 entries = IORING_MAX_ENTRIES;
9459 * Use twice as many entries for the CQ ring. It's possible for the
9460 * application to drive a higher depth than the size of the SQ ring,
9461 * since the sqes are only used at submission time. This allows for
9462 * some flexibility in overcommitting a bit. If the application has
9463 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9464 * of CQ ring entries manually.
9466 p->sq_entries = roundup_pow_of_two(entries);
9467 if (p->flags & IORING_SETUP_CQSIZE) {
9469 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9470 * to a power-of-two, if it isn't already. We do NOT impose
9471 * any cq vs sq ring sizing.
9475 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9476 if (!(p->flags & IORING_SETUP_CLAMP))
9478 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9480 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9481 if (p->cq_entries < p->sq_entries)
9484 p->cq_entries = 2 * p->sq_entries;
9487 ctx = io_ring_ctx_alloc(p);
9490 ctx->compat = in_compat_syscall();
9491 if (!capable(CAP_IPC_LOCK))
9492 ctx->user = get_uid(current_user());
9495 * This is just grabbed for accounting purposes. When a process exits,
9496 * the mm is exited and dropped before the files, hence we need to hang
9497 * on to this mm purely for the purposes of being able to unaccount
9498 * memory (locked/pinned vm). It's not used for anything else.
9500 mmgrab(current->mm);
9501 ctx->mm_account = current->mm;
9503 ret = io_allocate_scq_urings(ctx, p);
9507 ret = io_sq_offload_create(ctx, p);
9511 memset(&p->sq_off, 0, sizeof(p->sq_off));
9512 p->sq_off.head = offsetof(struct io_rings, sq.head);
9513 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9514 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9515 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9516 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9517 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9518 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9520 memset(&p->cq_off, 0, sizeof(p->cq_off));
9521 p->cq_off.head = offsetof(struct io_rings, cq.head);
9522 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9523 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9524 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9525 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9526 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9527 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9529 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9530 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9531 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9532 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9533 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9535 if (copy_to_user(params, p, sizeof(*p))) {
9540 file = io_uring_get_file(ctx);
9542 ret = PTR_ERR(file);
9547 * Install ring fd as the very last thing, so we don't risk someone
9548 * having closed it before we finish setup
9550 ret = io_uring_install_fd(ctx, file);
9552 /* fput will clean it up */
9557 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9560 io_ring_ctx_wait_and_kill(ctx);
9565 * Sets up an aio uring context, and returns the fd. Applications asks for a
9566 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9567 * params structure passed in.
9569 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9571 struct io_uring_params p;
9574 if (copy_from_user(&p, params, sizeof(p)))
9576 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9581 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9582 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9583 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9584 IORING_SETUP_R_DISABLED))
9587 return io_uring_create(entries, &p, params);
9590 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9591 struct io_uring_params __user *, params)
9593 return io_uring_setup(entries, params);
9596 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9598 struct io_uring_probe *p;
9602 size = struct_size(p, ops, nr_args);
9603 if (size == SIZE_MAX)
9605 p = kzalloc(size, GFP_KERNEL);
9610 if (copy_from_user(p, arg, size))
9613 if (memchr_inv(p, 0, size))
9616 p->last_op = IORING_OP_LAST - 1;
9617 if (nr_args > IORING_OP_LAST)
9618 nr_args = IORING_OP_LAST;
9620 for (i = 0; i < nr_args; i++) {
9622 if (!io_op_defs[i].not_supported)
9623 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9628 if (copy_to_user(arg, p, size))
9635 static int io_register_personality(struct io_ring_ctx *ctx)
9637 const struct cred *creds;
9641 creds = get_current_cred();
9643 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9644 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9651 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9652 unsigned int nr_args)
9654 struct io_uring_restriction *res;
9658 /* Restrictions allowed only if rings started disabled */
9659 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9662 /* We allow only a single restrictions registration */
9663 if (ctx->restrictions.registered)
9666 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9669 size = array_size(nr_args, sizeof(*res));
9670 if (size == SIZE_MAX)
9673 res = memdup_user(arg, size);
9675 return PTR_ERR(res);
9679 for (i = 0; i < nr_args; i++) {
9680 switch (res[i].opcode) {
9681 case IORING_RESTRICTION_REGISTER_OP:
9682 if (res[i].register_op >= IORING_REGISTER_LAST) {
9687 __set_bit(res[i].register_op,
9688 ctx->restrictions.register_op);
9690 case IORING_RESTRICTION_SQE_OP:
9691 if (res[i].sqe_op >= IORING_OP_LAST) {
9696 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9698 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9699 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9701 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9702 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9711 /* Reset all restrictions if an error happened */
9713 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9715 ctx->restrictions.registered = true;
9721 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9723 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9726 if (ctx->restrictions.registered)
9727 ctx->restricted = 1;
9729 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9730 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9731 wake_up(&ctx->sq_data->wait);
9735 static bool io_register_op_must_quiesce(int op)
9738 case IORING_UNREGISTER_FILES:
9739 case IORING_REGISTER_FILES_UPDATE:
9740 case IORING_REGISTER_PROBE:
9741 case IORING_REGISTER_PERSONALITY:
9742 case IORING_UNREGISTER_PERSONALITY:
9749 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9750 void __user *arg, unsigned nr_args)
9751 __releases(ctx->uring_lock)
9752 __acquires(ctx->uring_lock)
9757 * We're inside the ring mutex, if the ref is already dying, then
9758 * someone else killed the ctx or is already going through
9759 * io_uring_register().
9761 if (percpu_ref_is_dying(&ctx->refs))
9764 if (io_register_op_must_quiesce(opcode)) {
9765 percpu_ref_kill(&ctx->refs);
9768 * Drop uring mutex before waiting for references to exit. If
9769 * another thread is currently inside io_uring_enter() it might
9770 * need to grab the uring_lock to make progress. If we hold it
9771 * here across the drain wait, then we can deadlock. It's safe
9772 * to drop the mutex here, since no new references will come in
9773 * after we've killed the percpu ref.
9775 mutex_unlock(&ctx->uring_lock);
9777 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9780 ret = io_run_task_work_sig();
9785 mutex_lock(&ctx->uring_lock);
9788 percpu_ref_resurrect(&ctx->refs);
9793 if (ctx->restricted) {
9794 if (opcode >= IORING_REGISTER_LAST) {
9799 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9806 case IORING_REGISTER_BUFFERS:
9807 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9809 case IORING_UNREGISTER_BUFFERS:
9813 ret = io_sqe_buffers_unregister(ctx);
9815 case IORING_REGISTER_FILES:
9816 ret = io_sqe_files_register(ctx, arg, nr_args);
9818 case IORING_UNREGISTER_FILES:
9822 ret = io_sqe_files_unregister(ctx);
9824 case IORING_REGISTER_FILES_UPDATE:
9825 ret = io_sqe_files_update(ctx, arg, nr_args);
9827 case IORING_REGISTER_EVENTFD:
9828 case IORING_REGISTER_EVENTFD_ASYNC:
9832 ret = io_eventfd_register(ctx, arg);
9835 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9836 ctx->eventfd_async = 1;
9838 ctx->eventfd_async = 0;
9840 case IORING_UNREGISTER_EVENTFD:
9844 ret = io_eventfd_unregister(ctx);
9846 case IORING_REGISTER_PROBE:
9848 if (!arg || nr_args > 256)
9850 ret = io_probe(ctx, arg, nr_args);
9852 case IORING_REGISTER_PERSONALITY:
9856 ret = io_register_personality(ctx);
9858 case IORING_UNREGISTER_PERSONALITY:
9862 ret = io_unregister_personality(ctx, nr_args);
9864 case IORING_REGISTER_ENABLE_RINGS:
9868 ret = io_register_enable_rings(ctx);
9870 case IORING_REGISTER_RESTRICTIONS:
9871 ret = io_register_restrictions(ctx, arg, nr_args);
9879 if (io_register_op_must_quiesce(opcode)) {
9880 /* bring the ctx back to life */
9881 percpu_ref_reinit(&ctx->refs);
9883 reinit_completion(&ctx->ref_comp);
9888 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9889 void __user *, arg, unsigned int, nr_args)
9891 struct io_ring_ctx *ctx;
9900 if (f.file->f_op != &io_uring_fops)
9903 ctx = f.file->private_data;
9907 mutex_lock(&ctx->uring_lock);
9908 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9909 mutex_unlock(&ctx->uring_lock);
9910 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9911 ctx->cq_ev_fd != NULL, ret);
9917 static int __init io_uring_init(void)
9919 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9920 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9921 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9924 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9925 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9926 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9927 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9928 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9929 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9930 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9931 BUILD_BUG_SQE_ELEM(8, __u64, off);
9932 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9933 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9934 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9935 BUILD_BUG_SQE_ELEM(24, __u32, len);
9936 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9937 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9938 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9939 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9940 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9941 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9942 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9943 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9944 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9945 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9946 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9947 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9948 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9949 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9950 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9951 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9952 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9953 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9954 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9956 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9957 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9958 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9962 __initcall(io_uring_init);