1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
205 struct io_overflow_cqe {
206 struct io_uring_cqe cqe;
207 struct list_head list;
210 struct io_fixed_file {
211 /* file * with additional FFS_* flags */
212 unsigned long file_ptr;
216 struct list_head list;
223 struct io_file_table {
224 /* two level table */
225 struct io_fixed_file **files;
228 struct io_rsrc_node {
229 struct percpu_ref refs;
230 struct list_head node;
231 struct list_head rsrc_list;
232 struct io_rsrc_data *rsrc_data;
233 struct llist_node llist;
237 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
239 struct io_rsrc_data {
240 struct io_ring_ctx *ctx;
244 struct completion done;
249 struct list_head list;
255 struct io_restriction {
256 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
257 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
258 u8 sqe_flags_allowed;
259 u8 sqe_flags_required;
264 IO_SQ_THREAD_SHOULD_STOP = 0,
265 IO_SQ_THREAD_SHOULD_PARK,
270 atomic_t park_pending;
273 /* ctx's that are using this sqd */
274 struct list_head ctx_list;
276 struct task_struct *thread;
277 struct wait_queue_head wait;
279 unsigned sq_thread_idle;
285 struct completion exited;
286 struct callback_head *park_task_work;
289 #define IO_IOPOLL_BATCH 8
290 #define IO_COMPL_BATCH 32
291 #define IO_REQ_CACHE_SIZE 32
292 #define IO_REQ_ALLOC_BATCH 8
294 struct io_comp_state {
295 struct io_kiocb *reqs[IO_COMPL_BATCH];
297 unsigned int locked_free_nr;
298 /* inline/task_work completion list, under ->uring_lock */
299 struct list_head free_list;
300 /* IRQ completion list, under ->completion_lock */
301 struct list_head locked_free_list;
304 struct io_submit_link {
305 struct io_kiocb *head;
306 struct io_kiocb *last;
309 struct io_submit_state {
310 struct blk_plug plug;
311 struct io_submit_link link;
314 * io_kiocb alloc cache
316 void *reqs[IO_REQ_CACHE_SIZE];
317 unsigned int free_reqs;
322 * Batch completion logic
324 struct io_comp_state comp;
327 * File reference cache
331 unsigned int file_refs;
332 unsigned int ios_left;
337 struct percpu_ref refs;
338 } ____cacheline_aligned_in_smp;
342 unsigned int compat: 1;
343 unsigned int drain_next: 1;
344 unsigned int eventfd_async: 1;
345 unsigned int restricted: 1;
348 * Ring buffer of indices into array of io_uring_sqe, which is
349 * mmapped by the application using the IORING_OFF_SQES offset.
351 * This indirection could e.g. be used to assign fixed
352 * io_uring_sqe entries to operations and only submit them to
353 * the queue when needed.
355 * The kernel modifies neither the indices array nor the entries
359 unsigned cached_sq_head;
362 unsigned sq_thread_idle;
363 unsigned cached_sq_dropped;
364 unsigned cached_cq_overflow;
365 unsigned long sq_check_overflow;
367 /* hashed buffered write serialization */
368 struct io_wq_hash *hash_map;
370 struct list_head defer_list;
371 struct list_head timeout_list;
372 struct list_head cq_overflow_list;
374 struct io_uring_sqe *sq_sqes;
375 } ____cacheline_aligned_in_smp;
378 struct mutex uring_lock;
379 wait_queue_head_t wait;
380 } ____cacheline_aligned_in_smp;
382 struct io_submit_state submit_state;
384 struct io_rings *rings;
386 /* Only used for accounting purposes */
387 struct mm_struct *mm_account;
389 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
390 struct io_sq_data *sq_data; /* if using sq thread polling */
392 struct wait_queue_head sqo_sq_wait;
393 struct list_head sqd_list;
396 * If used, fixed file set. Writers must ensure that ->refs is dead,
397 * readers must ensure that ->refs is alive as long as the file* is
398 * used. Only updated through io_uring_register(2).
400 struct io_rsrc_data *file_data;
401 struct io_file_table file_table;
402 unsigned nr_user_files;
404 /* if used, fixed mapped user buffers */
405 unsigned nr_user_bufs;
406 struct io_mapped_ubuf *user_bufs;
408 struct user_struct *user;
410 struct completion ref_comp;
412 #if defined(CONFIG_UNIX)
413 struct socket *ring_sock;
416 struct xarray io_buffers;
418 struct xarray personalities;
422 unsigned cached_cq_tail;
425 atomic_t cq_timeouts;
426 unsigned cq_last_tm_flush;
427 unsigned long cq_check_overflow;
428 struct wait_queue_head cq_wait;
429 struct fasync_struct *cq_fasync;
430 struct eventfd_ctx *cq_ev_fd;
431 } ____cacheline_aligned_in_smp;
434 spinlock_t completion_lock;
437 * ->iopoll_list is protected by the ctx->uring_lock for
438 * io_uring instances that don't use IORING_SETUP_SQPOLL.
439 * For SQPOLL, only the single threaded io_sq_thread() will
440 * manipulate the list, hence no extra locking is needed there.
442 struct list_head iopoll_list;
443 struct hlist_head *cancel_hash;
444 unsigned cancel_hash_bits;
445 bool poll_multi_file;
446 } ____cacheline_aligned_in_smp;
448 struct delayed_work rsrc_put_work;
449 struct llist_head rsrc_put_llist;
450 struct list_head rsrc_ref_list;
451 spinlock_t rsrc_ref_lock;
452 struct io_rsrc_node *rsrc_node;
453 struct io_rsrc_node *rsrc_backup_node;
455 struct io_restriction restrictions;
458 struct callback_head *exit_task_work;
460 /* Keep this last, we don't need it for the fast path */
461 struct work_struct exit_work;
462 struct list_head tctx_list;
465 struct io_uring_task {
466 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 unsigned long task_state;
478 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
491 struct wait_queue_entry wait;
494 struct io_poll_update {
500 bool update_user_data;
508 struct io_timeout_data {
509 struct io_kiocb *req;
510 struct hrtimer timer;
511 struct timespec64 ts;
512 enum hrtimer_mode mode;
517 struct sockaddr __user *addr;
518 int __user *addr_len;
520 unsigned long nofile;
540 struct list_head list;
541 /* head of the link, used by linked timeouts only */
542 struct io_kiocb *head;
545 struct io_timeout_rem {
550 struct timespec64 ts;
555 /* NOTE: kiocb has the file as the first member, so don't do it here */
563 struct sockaddr __user *addr;
570 struct compat_msghdr __user *umsg_compat;
571 struct user_msghdr __user *umsg;
577 struct io_buffer *kbuf;
583 struct filename *filename;
585 unsigned long nofile;
588 struct io_rsrc_update {
614 struct epoll_event event;
618 struct file *file_out;
619 struct file *file_in;
626 struct io_provide_buf {
640 const char __user *filename;
641 struct statx __user *buffer;
653 struct filename *oldpath;
654 struct filename *newpath;
662 struct filename *filename;
665 struct io_completion {
667 struct list_head list;
671 struct io_async_connect {
672 struct sockaddr_storage address;
675 struct io_async_msghdr {
676 struct iovec fast_iov[UIO_FASTIOV];
677 /* points to an allocated iov, if NULL we use fast_iov instead */
678 struct iovec *free_iov;
679 struct sockaddr __user *uaddr;
681 struct sockaddr_storage addr;
685 struct iovec fast_iov[UIO_FASTIOV];
686 const struct iovec *free_iovec;
687 struct iov_iter iter;
689 struct wait_page_queue wpq;
693 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
694 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
695 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
696 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
697 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
698 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
704 REQ_F_LINK_TIMEOUT_BIT,
705 REQ_F_NEED_CLEANUP_BIT,
707 REQ_F_BUFFER_SELECTED_BIT,
708 REQ_F_LTIMEOUT_ACTIVE_BIT,
709 REQ_F_COMPLETE_INLINE_BIT,
711 REQ_F_DONT_REISSUE_BIT,
712 /* keep async read/write and isreg together and in order */
713 REQ_F_ASYNC_READ_BIT,
714 REQ_F_ASYNC_WRITE_BIT,
717 /* not a real bit, just to check we're not overflowing the space */
723 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
724 /* drain existing IO first */
725 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
727 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
728 /* doesn't sever on completion < 0 */
729 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
731 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
732 /* IOSQE_BUFFER_SELECT */
733 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
735 /* fail rest of links */
736 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
737 /* on inflight list, should be cancelled and waited on exit reliably */
738 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
739 /* read/write uses file position */
740 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
741 /* must not punt to workers */
742 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
743 /* has or had linked timeout */
744 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
746 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
747 /* already went through poll handler */
748 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
749 /* buffer already selected */
750 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
751 /* linked timeout is active, i.e. prepared by link's head */
752 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
753 /* completion is deferred through io_comp_state */
754 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
755 /* caller should reissue async */
756 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
757 /* don't attempt request reissue, see io_rw_reissue() */
758 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
759 /* supports async reads */
760 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
761 /* supports async writes */
762 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
764 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
768 struct io_poll_iocb poll;
769 struct io_poll_iocb *double_poll;
772 struct io_task_work {
773 struct io_wq_work_node node;
774 task_work_func_t func;
778 * NOTE! Each of the iocb union members has the file pointer
779 * as the first entry in their struct definition. So you can
780 * access the file pointer through any of the sub-structs,
781 * or directly as just 'ki_filp' in this struct.
787 struct io_poll_iocb poll;
788 struct io_poll_update poll_update;
789 struct io_accept accept;
791 struct io_cancel cancel;
792 struct io_timeout timeout;
793 struct io_timeout_rem timeout_rem;
794 struct io_connect connect;
795 struct io_sr_msg sr_msg;
797 struct io_close close;
798 struct io_rsrc_update rsrc_update;
799 struct io_fadvise fadvise;
800 struct io_madvise madvise;
801 struct io_epoll epoll;
802 struct io_splice splice;
803 struct io_provide_buf pbuf;
804 struct io_statx statx;
805 struct io_shutdown shutdown;
806 struct io_rename rename;
807 struct io_unlink unlink;
808 /* use only after cleaning per-op data, see io_clean_op() */
809 struct io_completion compl;
812 /* opcode allocated if it needs to store data for async defer */
815 /* polled IO has completed */
821 struct io_ring_ctx *ctx;
824 struct task_struct *task;
827 struct io_kiocb *link;
828 struct percpu_ref *fixed_rsrc_refs;
830 /* used with ctx->iopoll_list with reads/writes */
831 struct list_head inflight_entry;
833 struct io_task_work io_task_work;
834 struct callback_head task_work;
836 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
837 struct hlist_node hash_node;
838 struct async_poll *apoll;
839 struct io_wq_work work;
842 struct io_tctx_node {
843 struct list_head ctx_node;
844 struct task_struct *task;
845 struct io_ring_ctx *ctx;
848 struct io_defer_entry {
849 struct list_head list;
850 struct io_kiocb *req;
855 /* needs req->file assigned */
856 unsigned needs_file : 1;
857 /* hash wq insertion if file is a regular file */
858 unsigned hash_reg_file : 1;
859 /* unbound wq insertion if file is a non-regular file */
860 unsigned unbound_nonreg_file : 1;
861 /* opcode is not supported by this kernel */
862 unsigned not_supported : 1;
863 /* set if opcode supports polled "wait" */
865 unsigned pollout : 1;
866 /* op supports buffer selection */
867 unsigned buffer_select : 1;
868 /* do prep async if is going to be punted */
869 unsigned needs_async_setup : 1;
870 /* should block plug */
872 /* size of async data needed, if any */
873 unsigned short async_size;
876 static const struct io_op_def io_op_defs[] = {
877 [IORING_OP_NOP] = {},
878 [IORING_OP_READV] = {
880 .unbound_nonreg_file = 1,
883 .needs_async_setup = 1,
885 .async_size = sizeof(struct io_async_rw),
887 [IORING_OP_WRITEV] = {
890 .unbound_nonreg_file = 1,
892 .needs_async_setup = 1,
894 .async_size = sizeof(struct io_async_rw),
896 [IORING_OP_FSYNC] = {
899 [IORING_OP_READ_FIXED] = {
901 .unbound_nonreg_file = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_WRITE_FIXED] = {
909 .unbound_nonreg_file = 1,
912 .async_size = sizeof(struct io_async_rw),
914 [IORING_OP_POLL_ADD] = {
916 .unbound_nonreg_file = 1,
918 [IORING_OP_POLL_REMOVE] = {},
919 [IORING_OP_SYNC_FILE_RANGE] = {
922 [IORING_OP_SENDMSG] = {
924 .unbound_nonreg_file = 1,
926 .needs_async_setup = 1,
927 .async_size = sizeof(struct io_async_msghdr),
929 [IORING_OP_RECVMSG] = {
931 .unbound_nonreg_file = 1,
934 .needs_async_setup = 1,
935 .async_size = sizeof(struct io_async_msghdr),
937 [IORING_OP_TIMEOUT] = {
938 .async_size = sizeof(struct io_timeout_data),
940 [IORING_OP_TIMEOUT_REMOVE] = {
941 /* used by timeout updates' prep() */
943 [IORING_OP_ACCEPT] = {
945 .unbound_nonreg_file = 1,
948 [IORING_OP_ASYNC_CANCEL] = {},
949 [IORING_OP_LINK_TIMEOUT] = {
950 .async_size = sizeof(struct io_timeout_data),
952 [IORING_OP_CONNECT] = {
954 .unbound_nonreg_file = 1,
956 .needs_async_setup = 1,
957 .async_size = sizeof(struct io_async_connect),
959 [IORING_OP_FALLOCATE] = {
962 [IORING_OP_OPENAT] = {},
963 [IORING_OP_CLOSE] = {},
964 [IORING_OP_FILES_UPDATE] = {},
965 [IORING_OP_STATX] = {},
968 .unbound_nonreg_file = 1,
972 .async_size = sizeof(struct io_async_rw),
974 [IORING_OP_WRITE] = {
976 .unbound_nonreg_file = 1,
979 .async_size = sizeof(struct io_async_rw),
981 [IORING_OP_FADVISE] = {
984 [IORING_OP_MADVISE] = {},
987 .unbound_nonreg_file = 1,
992 .unbound_nonreg_file = 1,
996 [IORING_OP_OPENAT2] = {
998 [IORING_OP_EPOLL_CTL] = {
999 .unbound_nonreg_file = 1,
1001 [IORING_OP_SPLICE] = {
1004 .unbound_nonreg_file = 1,
1006 [IORING_OP_PROVIDE_BUFFERS] = {},
1007 [IORING_OP_REMOVE_BUFFERS] = {},
1011 .unbound_nonreg_file = 1,
1013 [IORING_OP_SHUTDOWN] = {
1016 [IORING_OP_RENAMEAT] = {},
1017 [IORING_OP_UNLINKAT] = {},
1020 static bool io_disarm_next(struct io_kiocb *req);
1021 static void io_uring_del_task_file(unsigned long index);
1022 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1023 struct task_struct *task,
1024 struct files_struct *files);
1025 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1026 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1028 static bool io_cqring_fill_event(struct io_kiocb *req, long res, unsigned cflags);
1029 static void io_put_req(struct io_kiocb *req);
1030 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1031 static void io_dismantle_req(struct io_kiocb *req);
1032 static void io_put_task(struct task_struct *task, int nr);
1033 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1034 static void io_queue_linked_timeout(struct io_kiocb *req);
1035 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1036 struct io_uring_rsrc_update *ip,
1038 static void io_clean_op(struct io_kiocb *req);
1039 static struct file *io_file_get(struct io_submit_state *state,
1040 struct io_kiocb *req, int fd, bool fixed);
1041 static void __io_queue_sqe(struct io_kiocb *req);
1042 static void io_rsrc_put_work(struct work_struct *work);
1044 static void io_req_task_queue(struct io_kiocb *req);
1045 static void io_submit_flush_completions(struct io_comp_state *cs,
1046 struct io_ring_ctx *ctx);
1047 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1048 static int io_req_prep_async(struct io_kiocb *req);
1050 static struct kmem_cache *req_cachep;
1052 static const struct file_operations io_uring_fops;
1054 struct sock *io_uring_get_socket(struct file *file)
1056 #if defined(CONFIG_UNIX)
1057 if (file->f_op == &io_uring_fops) {
1058 struct io_ring_ctx *ctx = file->private_data;
1060 return ctx->ring_sock->sk;
1065 EXPORT_SYMBOL(io_uring_get_socket);
1067 #define io_for_each_link(pos, head) \
1068 for (pos = (head); pos; pos = pos->link)
1070 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1072 struct io_ring_ctx *ctx = req->ctx;
1074 if (!req->fixed_rsrc_refs) {
1075 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1076 percpu_ref_get(req->fixed_rsrc_refs);
1080 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1082 bool got = percpu_ref_tryget(ref);
1084 /* already at zero, wait for ->release() */
1086 wait_for_completion(compl);
1087 percpu_ref_resurrect(ref);
1089 percpu_ref_put(ref);
1092 static bool io_match_task(struct io_kiocb *head,
1093 struct task_struct *task,
1094 struct files_struct *files)
1096 struct io_kiocb *req;
1098 if (task && head->task != task)
1103 io_for_each_link(req, head) {
1104 if (req->flags & REQ_F_INFLIGHT)
1110 static inline void req_set_fail_links(struct io_kiocb *req)
1112 if (req->flags & REQ_F_LINK)
1113 req->flags |= REQ_F_FAIL_LINK;
1116 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1118 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1120 complete(&ctx->ref_comp);
1123 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1125 return !req->timeout.off;
1128 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1130 struct io_ring_ctx *ctx;
1133 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1138 * Use 5 bits less than the max cq entries, that should give us around
1139 * 32 entries per hash list if totally full and uniformly spread.
1141 hash_bits = ilog2(p->cq_entries);
1145 ctx->cancel_hash_bits = hash_bits;
1146 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1148 if (!ctx->cancel_hash)
1150 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1152 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1153 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1156 ctx->flags = p->flags;
1157 init_waitqueue_head(&ctx->sqo_sq_wait);
1158 INIT_LIST_HEAD(&ctx->sqd_list);
1159 init_waitqueue_head(&ctx->cq_wait);
1160 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1161 init_completion(&ctx->ref_comp);
1162 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1163 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1164 mutex_init(&ctx->uring_lock);
1165 init_waitqueue_head(&ctx->wait);
1166 spin_lock_init(&ctx->completion_lock);
1167 INIT_LIST_HEAD(&ctx->iopoll_list);
1168 INIT_LIST_HEAD(&ctx->defer_list);
1169 INIT_LIST_HEAD(&ctx->timeout_list);
1170 spin_lock_init(&ctx->rsrc_ref_lock);
1171 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1172 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1173 init_llist_head(&ctx->rsrc_put_llist);
1174 INIT_LIST_HEAD(&ctx->tctx_list);
1175 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1176 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1179 kfree(ctx->cancel_hash);
1184 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1186 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1187 struct io_ring_ctx *ctx = req->ctx;
1189 return seq != ctx->cached_cq_tail
1190 + READ_ONCE(ctx->cached_cq_overflow);
1196 static void io_req_track_inflight(struct io_kiocb *req)
1198 if (!(req->flags & REQ_F_INFLIGHT)) {
1199 req->flags |= REQ_F_INFLIGHT;
1200 atomic_inc(¤t->io_uring->inflight_tracked);
1204 static void io_prep_async_work(struct io_kiocb *req)
1206 const struct io_op_def *def = &io_op_defs[req->opcode];
1207 struct io_ring_ctx *ctx = req->ctx;
1209 if (!req->work.creds)
1210 req->work.creds = get_current_cred();
1212 req->work.list.next = NULL;
1213 req->work.flags = 0;
1214 if (req->flags & REQ_F_FORCE_ASYNC)
1215 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1217 if (req->flags & REQ_F_ISREG) {
1218 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1219 io_wq_hash_work(&req->work, file_inode(req->file));
1220 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1221 if (def->unbound_nonreg_file)
1222 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 switch (req->opcode) {
1226 case IORING_OP_SPLICE:
1228 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1229 req->work.flags |= IO_WQ_WORK_UNBOUND;
1234 static void io_prep_async_link(struct io_kiocb *req)
1236 struct io_kiocb *cur;
1238 io_for_each_link(cur, req)
1239 io_prep_async_work(cur);
1242 static void io_queue_async_work(struct io_kiocb *req)
1244 struct io_ring_ctx *ctx = req->ctx;
1245 struct io_kiocb *link = io_prep_linked_timeout(req);
1246 struct io_uring_task *tctx = req->task->io_uring;
1249 BUG_ON(!tctx->io_wq);
1251 /* init ->work of the whole link before punting */
1252 io_prep_async_link(req);
1253 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1254 &req->work, req->flags);
1255 io_wq_enqueue(tctx->io_wq, &req->work);
1257 io_queue_linked_timeout(link);
1260 static void io_kill_timeout(struct io_kiocb *req, int status)
1261 __must_hold(&req->ctx->completion_lock)
1263 struct io_timeout_data *io = req->async_data;
1265 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1266 atomic_set(&req->ctx->cq_timeouts,
1267 atomic_read(&req->ctx->cq_timeouts) + 1);
1268 list_del_init(&req->timeout.list);
1269 io_cqring_fill_event(req, status, 0);
1270 io_put_req_deferred(req, 1);
1274 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1277 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1278 struct io_defer_entry, list);
1280 if (req_need_defer(de->req, de->seq))
1282 list_del_init(&de->list);
1283 io_req_task_queue(de->req);
1285 } while (!list_empty(&ctx->defer_list));
1288 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1292 if (list_empty(&ctx->timeout_list))
1295 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1298 u32 events_needed, events_got;
1299 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1300 struct io_kiocb, timeout.list);
1302 if (io_is_timeout_noseq(req))
1306 * Since seq can easily wrap around over time, subtract
1307 * the last seq at which timeouts were flushed before comparing.
1308 * Assuming not more than 2^31-1 events have happened since,
1309 * these subtractions won't have wrapped, so we can check if
1310 * target is in [last_seq, current_seq] by comparing the two.
1312 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1313 events_got = seq - ctx->cq_last_tm_flush;
1314 if (events_got < events_needed)
1317 list_del_init(&req->timeout.list);
1318 io_kill_timeout(req, 0);
1319 } while (!list_empty(&ctx->timeout_list));
1321 ctx->cq_last_tm_flush = seq;
1324 static void io_commit_cqring(struct io_ring_ctx *ctx)
1326 io_flush_timeouts(ctx);
1328 /* order cqe stores with ring update */
1329 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1331 if (unlikely(!list_empty(&ctx->defer_list)))
1332 __io_queue_deferred(ctx);
1335 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1337 struct io_rings *r = ctx->rings;
1339 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1342 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1344 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1347 static inline struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1349 struct io_rings *rings = ctx->rings;
1353 * writes to the cq entry need to come after reading head; the
1354 * control dependency is enough as we're using WRITE_ONCE to
1357 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1360 tail = ctx->cached_cq_tail++;
1361 return &rings->cqes[tail & ctx->cq_mask];
1364 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1366 if (likely(!ctx->cq_ev_fd))
1368 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1370 return !ctx->eventfd_async || io_wq_current_is_worker();
1373 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1375 /* see waitqueue_active() comment */
1378 if (waitqueue_active(&ctx->wait))
1379 wake_up(&ctx->wait);
1380 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1381 wake_up(&ctx->sq_data->wait);
1382 if (io_should_trigger_evfd(ctx))
1383 eventfd_signal(ctx->cq_ev_fd, 1);
1384 if (waitqueue_active(&ctx->cq_wait)) {
1385 wake_up_interruptible(&ctx->cq_wait);
1386 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1390 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1392 /* see waitqueue_active() comment */
1395 if (ctx->flags & IORING_SETUP_SQPOLL) {
1396 if (waitqueue_active(&ctx->wait))
1397 wake_up(&ctx->wait);
1399 if (io_should_trigger_evfd(ctx))
1400 eventfd_signal(ctx->cq_ev_fd, 1);
1401 if (waitqueue_active(&ctx->cq_wait)) {
1402 wake_up_interruptible(&ctx->cq_wait);
1403 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1407 /* Returns true if there are no backlogged entries after the flush */
1408 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1410 struct io_rings *rings = ctx->rings;
1411 unsigned long flags;
1412 bool all_flushed, posted;
1414 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1418 spin_lock_irqsave(&ctx->completion_lock, flags);
1419 while (!list_empty(&ctx->cq_overflow_list)) {
1420 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1421 struct io_overflow_cqe *ocqe;
1425 ocqe = list_first_entry(&ctx->cq_overflow_list,
1426 struct io_overflow_cqe, list);
1428 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1430 WRITE_ONCE(ctx->rings->cq_overflow,
1431 ++ctx->cached_cq_overflow);
1433 list_del(&ocqe->list);
1437 all_flushed = list_empty(&ctx->cq_overflow_list);
1439 clear_bit(0, &ctx->sq_check_overflow);
1440 clear_bit(0, &ctx->cq_check_overflow);
1441 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1445 io_commit_cqring(ctx);
1446 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1448 io_cqring_ev_posted(ctx);
1452 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1456 if (test_bit(0, &ctx->cq_check_overflow)) {
1457 /* iopoll syncs against uring_lock, not completion_lock */
1458 if (ctx->flags & IORING_SETUP_IOPOLL)
1459 mutex_lock(&ctx->uring_lock);
1460 ret = __io_cqring_overflow_flush(ctx, force);
1461 if (ctx->flags & IORING_SETUP_IOPOLL)
1462 mutex_unlock(&ctx->uring_lock);
1469 * Shamelessly stolen from the mm implementation of page reference checking,
1470 * see commit f958d7b528b1 for details.
1472 #define req_ref_zero_or_close_to_overflow(req) \
1473 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1475 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1477 return atomic_inc_not_zero(&req->refs);
1480 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1482 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1483 return atomic_sub_and_test(refs, &req->refs);
1486 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1488 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1489 return atomic_dec_and_test(&req->refs);
1492 static inline void req_ref_put(struct io_kiocb *req)
1494 WARN_ON_ONCE(req_ref_put_and_test(req));
1497 static inline void req_ref_get(struct io_kiocb *req)
1499 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1500 atomic_inc(&req->refs);
1503 static bool io_cqring_event_overflow(struct io_kiocb *req, long res,
1504 unsigned int cflags)
1506 struct io_ring_ctx *ctx = req->ctx;
1507 struct io_overflow_cqe *ocqe;
1509 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1512 * If we're in ring overflow flush mode, or in task cancel mode,
1513 * or cannot allocate an overflow entry, then we need to drop it
1516 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1519 if (list_empty(&ctx->cq_overflow_list)) {
1520 set_bit(0, &ctx->sq_check_overflow);
1521 set_bit(0, &ctx->cq_check_overflow);
1522 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1524 ocqe->cqe.user_data = req->user_data;
1525 ocqe->cqe.res = res;
1526 ocqe->cqe.flags = cflags;
1527 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1531 static inline bool __io_cqring_fill_event(struct io_kiocb *req, long res,
1532 unsigned int cflags)
1534 struct io_ring_ctx *ctx = req->ctx;
1535 struct io_uring_cqe *cqe;
1537 trace_io_uring_complete(ctx, req->user_data, res, cflags);
1540 * If we can't get a cq entry, userspace overflowed the
1541 * submission (by quite a lot). Increment the overflow count in
1544 cqe = io_get_cqring(ctx);
1546 WRITE_ONCE(cqe->user_data, req->user_data);
1547 WRITE_ONCE(cqe->res, res);
1548 WRITE_ONCE(cqe->flags, cflags);
1551 return io_cqring_event_overflow(req, res, cflags);
1554 /* not as hot to bloat with inlining */
1555 static noinline bool io_cqring_fill_event(struct io_kiocb *req, long res,
1556 unsigned int cflags)
1558 return __io_cqring_fill_event(req, res, cflags);
1561 static void io_req_complete_post(struct io_kiocb *req, long res,
1562 unsigned int cflags)
1564 struct io_ring_ctx *ctx = req->ctx;
1565 unsigned long flags;
1567 spin_lock_irqsave(&ctx->completion_lock, flags);
1568 __io_cqring_fill_event(req, res, cflags);
1570 * If we're the last reference to this request, add to our locked
1573 if (req_ref_put_and_test(req)) {
1574 struct io_comp_state *cs = &ctx->submit_state.comp;
1576 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1577 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1578 io_disarm_next(req);
1580 io_req_task_queue(req->link);
1584 io_dismantle_req(req);
1585 io_put_task(req->task, 1);
1586 list_add(&req->compl.list, &cs->locked_free_list);
1587 cs->locked_free_nr++;
1589 if (!percpu_ref_tryget(&ctx->refs))
1592 io_commit_cqring(ctx);
1593 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1596 io_cqring_ev_posted(ctx);
1597 percpu_ref_put(&ctx->refs);
1601 static inline bool io_req_needs_clean(struct io_kiocb *req)
1603 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1607 static void io_req_complete_state(struct io_kiocb *req, long res,
1608 unsigned int cflags)
1610 if (io_req_needs_clean(req))
1613 req->compl.cflags = cflags;
1614 req->flags |= REQ_F_COMPLETE_INLINE;
1617 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1618 long res, unsigned cflags)
1620 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1621 io_req_complete_state(req, res, cflags);
1623 io_req_complete_post(req, res, cflags);
1626 static inline void io_req_complete(struct io_kiocb *req, long res)
1628 __io_req_complete(req, 0, res, 0);
1631 static void io_req_complete_failed(struct io_kiocb *req, long res)
1633 req_set_fail_links(req);
1635 io_req_complete_post(req, res, 0);
1638 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1639 struct io_comp_state *cs)
1641 spin_lock_irq(&ctx->completion_lock);
1642 list_splice_init(&cs->locked_free_list, &cs->free_list);
1643 cs->locked_free_nr = 0;
1644 spin_unlock_irq(&ctx->completion_lock);
1647 /* Returns true IFF there are requests in the cache */
1648 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1650 struct io_submit_state *state = &ctx->submit_state;
1651 struct io_comp_state *cs = &state->comp;
1655 * If we have more than a batch's worth of requests in our IRQ side
1656 * locked cache, grab the lock and move them over to our submission
1659 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1660 io_flush_cached_locked_reqs(ctx, cs);
1662 nr = state->free_reqs;
1663 while (!list_empty(&cs->free_list)) {
1664 struct io_kiocb *req = list_first_entry(&cs->free_list,
1665 struct io_kiocb, compl.list);
1667 list_del(&req->compl.list);
1668 state->reqs[nr++] = req;
1669 if (nr == ARRAY_SIZE(state->reqs))
1673 state->free_reqs = nr;
1677 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1679 struct io_submit_state *state = &ctx->submit_state;
1681 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1683 if (!state->free_reqs) {
1684 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1687 if (io_flush_cached_reqs(ctx))
1690 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1694 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1695 * retry single alloc to be on the safe side.
1697 if (unlikely(ret <= 0)) {
1698 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1699 if (!state->reqs[0])
1703 state->free_reqs = ret;
1707 return state->reqs[state->free_reqs];
1710 static inline void io_put_file(struct file *file)
1716 static void io_dismantle_req(struct io_kiocb *req)
1718 unsigned int flags = req->flags;
1720 if (!(flags & REQ_F_FIXED_FILE))
1721 io_put_file(req->file);
1722 if (io_req_needs_clean(req) || (req->flags & REQ_F_INFLIGHT)) {
1724 if (req->flags & REQ_F_INFLIGHT) {
1725 struct io_uring_task *tctx = req->task->io_uring;
1727 atomic_dec(&tctx->inflight_tracked);
1728 req->flags &= ~REQ_F_INFLIGHT;
1731 if (req->fixed_rsrc_refs)
1732 percpu_ref_put(req->fixed_rsrc_refs);
1733 if (req->async_data)
1734 kfree(req->async_data);
1735 if (req->work.creds) {
1736 put_cred(req->work.creds);
1737 req->work.creds = NULL;
1741 /* must to be called somewhat shortly after putting a request */
1742 static inline void io_put_task(struct task_struct *task, int nr)
1744 struct io_uring_task *tctx = task->io_uring;
1746 percpu_counter_sub(&tctx->inflight, nr);
1747 if (unlikely(atomic_read(&tctx->in_idle)))
1748 wake_up(&tctx->wait);
1749 put_task_struct_many(task, nr);
1752 static void __io_free_req(struct io_kiocb *req)
1754 struct io_ring_ctx *ctx = req->ctx;
1756 io_dismantle_req(req);
1757 io_put_task(req->task, 1);
1759 kmem_cache_free(req_cachep, req);
1760 percpu_ref_put(&ctx->refs);
1763 static inline void io_remove_next_linked(struct io_kiocb *req)
1765 struct io_kiocb *nxt = req->link;
1767 req->link = nxt->link;
1771 static bool io_kill_linked_timeout(struct io_kiocb *req)
1772 __must_hold(&req->ctx->completion_lock)
1774 struct io_kiocb *link = req->link;
1777 * Can happen if a linked timeout fired and link had been like
1778 * req -> link t-out -> link t-out [-> ...]
1780 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1781 struct io_timeout_data *io = link->async_data;
1783 io_remove_next_linked(req);
1784 link->timeout.head = NULL;
1785 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1786 io_cqring_fill_event(link, -ECANCELED, 0);
1787 io_put_req_deferred(link, 1);
1794 static void io_fail_links(struct io_kiocb *req)
1795 __must_hold(&req->ctx->completion_lock)
1797 struct io_kiocb *nxt, *link = req->link;
1804 trace_io_uring_fail_link(req, link);
1805 io_cqring_fill_event(link, -ECANCELED, 0);
1806 io_put_req_deferred(link, 2);
1811 static bool io_disarm_next(struct io_kiocb *req)
1812 __must_hold(&req->ctx->completion_lock)
1814 bool posted = false;
1816 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1817 posted = io_kill_linked_timeout(req);
1818 if (unlikely((req->flags & REQ_F_FAIL_LINK) &&
1819 !(req->flags & REQ_F_HARDLINK))) {
1820 posted |= (req->link != NULL);
1826 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1828 struct io_kiocb *nxt;
1831 * If LINK is set, we have dependent requests in this chain. If we
1832 * didn't fail this request, queue the first one up, moving any other
1833 * dependencies to the next request. In case of failure, fail the rest
1836 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1837 struct io_ring_ctx *ctx = req->ctx;
1838 unsigned long flags;
1841 spin_lock_irqsave(&ctx->completion_lock, flags);
1842 posted = io_disarm_next(req);
1844 io_commit_cqring(req->ctx);
1845 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1847 io_cqring_ev_posted(ctx);
1854 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1856 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1858 return __io_req_find_next(req);
1861 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1865 if (ctx->submit_state.comp.nr) {
1866 mutex_lock(&ctx->uring_lock);
1867 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1868 mutex_unlock(&ctx->uring_lock);
1870 percpu_ref_put(&ctx->refs);
1873 static bool __tctx_task_work(struct io_uring_task *tctx)
1875 struct io_ring_ctx *ctx = NULL;
1876 struct io_wq_work_list list;
1877 struct io_wq_work_node *node;
1879 if (wq_list_empty(&tctx->task_list))
1882 spin_lock_irq(&tctx->task_lock);
1883 list = tctx->task_list;
1884 INIT_WQ_LIST(&tctx->task_list);
1885 spin_unlock_irq(&tctx->task_lock);
1889 struct io_wq_work_node *next = node->next;
1890 struct io_kiocb *req;
1892 req = container_of(node, struct io_kiocb, io_task_work.node);
1893 if (req->ctx != ctx) {
1894 ctx_flush_and_put(ctx);
1896 percpu_ref_get(&ctx->refs);
1899 req->task_work.func(&req->task_work);
1903 ctx_flush_and_put(ctx);
1904 return list.first != NULL;
1907 static void tctx_task_work(struct callback_head *cb)
1909 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1911 clear_bit(0, &tctx->task_state);
1913 while (__tctx_task_work(tctx))
1917 static int io_req_task_work_add(struct io_kiocb *req)
1919 struct task_struct *tsk = req->task;
1920 struct io_uring_task *tctx = tsk->io_uring;
1921 enum task_work_notify_mode notify;
1922 struct io_wq_work_node *node, *prev;
1923 unsigned long flags;
1926 if (unlikely(tsk->flags & PF_EXITING))
1929 WARN_ON_ONCE(!tctx);
1931 spin_lock_irqsave(&tctx->task_lock, flags);
1932 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1933 spin_unlock_irqrestore(&tctx->task_lock, flags);
1935 /* task_work already pending, we're done */
1936 if (test_bit(0, &tctx->task_state) ||
1937 test_and_set_bit(0, &tctx->task_state))
1941 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1942 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1943 * processing task_work. There's no reliable way to tell if TWA_RESUME
1946 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1948 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1949 wake_up_process(tsk);
1954 * Slow path - we failed, find and delete work. if the work is not
1955 * in the list, it got run and we're fine.
1957 spin_lock_irqsave(&tctx->task_lock, flags);
1958 wq_list_for_each(node, prev, &tctx->task_list) {
1959 if (&req->io_task_work.node == node) {
1960 wq_list_del(&tctx->task_list, node, prev);
1965 spin_unlock_irqrestore(&tctx->task_lock, flags);
1966 clear_bit(0, &tctx->task_state);
1970 static bool io_run_task_work_head(struct callback_head **work_head)
1972 struct callback_head *work, *next;
1973 bool executed = false;
1976 work = xchg(work_head, NULL);
1992 static void io_task_work_add_head(struct callback_head **work_head,
1993 struct callback_head *task_work)
1995 struct callback_head *head;
1998 head = READ_ONCE(*work_head);
1999 task_work->next = head;
2000 } while (cmpxchg(work_head, head, task_work) != head);
2003 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2004 task_work_func_t cb)
2006 init_task_work(&req->task_work, cb);
2007 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2010 static void io_req_task_cancel(struct callback_head *cb)
2012 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2013 struct io_ring_ctx *ctx = req->ctx;
2015 /* ctx is guaranteed to stay alive while we hold uring_lock */
2016 mutex_lock(&ctx->uring_lock);
2017 io_req_complete_failed(req, req->result);
2018 mutex_unlock(&ctx->uring_lock);
2021 static void __io_req_task_submit(struct io_kiocb *req)
2023 struct io_ring_ctx *ctx = req->ctx;
2025 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2026 mutex_lock(&ctx->uring_lock);
2027 if (!(current->flags & PF_EXITING) && !current->in_execve)
2028 __io_queue_sqe(req);
2030 io_req_complete_failed(req, -EFAULT);
2031 mutex_unlock(&ctx->uring_lock);
2034 static void io_req_task_submit(struct callback_head *cb)
2036 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2038 __io_req_task_submit(req);
2041 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2044 req->task_work.func = io_req_task_cancel;
2046 if (unlikely(io_req_task_work_add(req)))
2047 io_req_task_work_add_fallback(req, io_req_task_cancel);
2050 static void io_req_task_queue(struct io_kiocb *req)
2052 req->task_work.func = io_req_task_submit;
2054 if (unlikely(io_req_task_work_add(req)))
2055 io_req_task_queue_fail(req, -ECANCELED);
2058 static inline void io_queue_next(struct io_kiocb *req)
2060 struct io_kiocb *nxt = io_req_find_next(req);
2063 io_req_task_queue(nxt);
2066 static void io_free_req(struct io_kiocb *req)
2073 struct task_struct *task;
2078 static inline void io_init_req_batch(struct req_batch *rb)
2085 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2086 struct req_batch *rb)
2089 io_put_task(rb->task, rb->task_refs);
2091 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2094 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2095 struct io_submit_state *state)
2098 io_dismantle_req(req);
2100 if (req->task != rb->task) {
2102 io_put_task(rb->task, rb->task_refs);
2103 rb->task = req->task;
2109 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2110 state->reqs[state->free_reqs++] = req;
2112 list_add(&req->compl.list, &state->comp.free_list);
2115 static void io_submit_flush_completions(struct io_comp_state *cs,
2116 struct io_ring_ctx *ctx)
2119 struct io_kiocb *req;
2120 struct req_batch rb;
2122 io_init_req_batch(&rb);
2123 spin_lock_irq(&ctx->completion_lock);
2124 for (i = 0; i < nr; i++) {
2126 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2128 io_commit_cqring(ctx);
2129 spin_unlock_irq(&ctx->completion_lock);
2131 io_cqring_ev_posted(ctx);
2132 for (i = 0; i < nr; i++) {
2135 /* submission and completion refs */
2136 if (req_ref_sub_and_test(req, 2))
2137 io_req_free_batch(&rb, req, &ctx->submit_state);
2140 io_req_free_batch_finish(ctx, &rb);
2145 * Drop reference to request, return next in chain (if there is one) if this
2146 * was the last reference to this request.
2148 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2150 struct io_kiocb *nxt = NULL;
2152 if (req_ref_put_and_test(req)) {
2153 nxt = io_req_find_next(req);
2159 static inline void io_put_req(struct io_kiocb *req)
2161 if (req_ref_put_and_test(req))
2165 static void io_put_req_deferred_cb(struct callback_head *cb)
2167 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2172 static void io_free_req_deferred(struct io_kiocb *req)
2174 req->task_work.func = io_put_req_deferred_cb;
2175 if (unlikely(io_req_task_work_add(req)))
2176 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2179 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2181 if (req_ref_sub_and_test(req, refs))
2182 io_free_req_deferred(req);
2185 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2187 /* See comment at the top of this file */
2189 return __io_cqring_events(ctx);
2192 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2194 struct io_rings *rings = ctx->rings;
2196 /* make sure SQ entry isn't read before tail */
2197 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2200 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2202 unsigned int cflags;
2204 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2205 cflags |= IORING_CQE_F_BUFFER;
2206 req->flags &= ~REQ_F_BUFFER_SELECTED;
2211 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2213 struct io_buffer *kbuf;
2215 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2216 return io_put_kbuf(req, kbuf);
2219 static inline bool io_run_task_work(void)
2222 * Not safe to run on exiting task, and the task_work handling will
2223 * not add work to such a task.
2225 if (unlikely(current->flags & PF_EXITING))
2227 if (current->task_works) {
2228 __set_current_state(TASK_RUNNING);
2237 * Find and free completed poll iocbs
2239 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2240 struct list_head *done)
2242 struct req_batch rb;
2243 struct io_kiocb *req;
2245 /* order with ->result store in io_complete_rw_iopoll() */
2248 io_init_req_batch(&rb);
2249 while (!list_empty(done)) {
2252 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2253 list_del(&req->inflight_entry);
2255 if (READ_ONCE(req->result) == -EAGAIN &&
2256 !(req->flags & REQ_F_DONT_REISSUE)) {
2257 req->iopoll_completed = 0;
2259 io_queue_async_work(req);
2263 if (req->flags & REQ_F_BUFFER_SELECTED)
2264 cflags = io_put_rw_kbuf(req);
2266 __io_cqring_fill_event(req, req->result, cflags);
2269 if (req_ref_put_and_test(req))
2270 io_req_free_batch(&rb, req, &ctx->submit_state);
2273 io_commit_cqring(ctx);
2274 io_cqring_ev_posted_iopoll(ctx);
2275 io_req_free_batch_finish(ctx, &rb);
2278 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2281 struct io_kiocb *req, *tmp;
2287 * Only spin for completions if we don't have multiple devices hanging
2288 * off our complete list, and we're under the requested amount.
2290 spin = !ctx->poll_multi_file && *nr_events < min;
2293 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2294 struct kiocb *kiocb = &req->rw.kiocb;
2297 * Move completed and retryable entries to our local lists.
2298 * If we find a request that requires polling, break out
2299 * and complete those lists first, if we have entries there.
2301 if (READ_ONCE(req->iopoll_completed)) {
2302 list_move_tail(&req->inflight_entry, &done);
2305 if (!list_empty(&done))
2308 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2312 /* iopoll may have completed current req */
2313 if (READ_ONCE(req->iopoll_completed))
2314 list_move_tail(&req->inflight_entry, &done);
2321 if (!list_empty(&done))
2322 io_iopoll_complete(ctx, nr_events, &done);
2328 * We can't just wait for polled events to come to us, we have to actively
2329 * find and complete them.
2331 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2333 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2336 mutex_lock(&ctx->uring_lock);
2337 while (!list_empty(&ctx->iopoll_list)) {
2338 unsigned int nr_events = 0;
2340 io_do_iopoll(ctx, &nr_events, 0);
2342 /* let it sleep and repeat later if can't complete a request */
2346 * Ensure we allow local-to-the-cpu processing to take place,
2347 * in this case we need to ensure that we reap all events.
2348 * Also let task_work, etc. to progress by releasing the mutex
2350 if (need_resched()) {
2351 mutex_unlock(&ctx->uring_lock);
2353 mutex_lock(&ctx->uring_lock);
2356 mutex_unlock(&ctx->uring_lock);
2359 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2361 unsigned int nr_events = 0;
2365 * We disallow the app entering submit/complete with polling, but we
2366 * still need to lock the ring to prevent racing with polled issue
2367 * that got punted to a workqueue.
2369 mutex_lock(&ctx->uring_lock);
2371 * Don't enter poll loop if we already have events pending.
2372 * If we do, we can potentially be spinning for commands that
2373 * already triggered a CQE (eg in error).
2375 if (test_bit(0, &ctx->cq_check_overflow))
2376 __io_cqring_overflow_flush(ctx, false);
2377 if (io_cqring_events(ctx))
2381 * If a submit got punted to a workqueue, we can have the
2382 * application entering polling for a command before it gets
2383 * issued. That app will hold the uring_lock for the duration
2384 * of the poll right here, so we need to take a breather every
2385 * now and then to ensure that the issue has a chance to add
2386 * the poll to the issued list. Otherwise we can spin here
2387 * forever, while the workqueue is stuck trying to acquire the
2390 if (list_empty(&ctx->iopoll_list)) {
2391 mutex_unlock(&ctx->uring_lock);
2393 mutex_lock(&ctx->uring_lock);
2395 if (list_empty(&ctx->iopoll_list))
2398 ret = io_do_iopoll(ctx, &nr_events, min);
2399 } while (!ret && nr_events < min && !need_resched());
2401 mutex_unlock(&ctx->uring_lock);
2405 static void kiocb_end_write(struct io_kiocb *req)
2408 * Tell lockdep we inherited freeze protection from submission
2411 if (req->flags & REQ_F_ISREG) {
2412 struct super_block *sb = file_inode(req->file)->i_sb;
2414 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2420 static bool io_resubmit_prep(struct io_kiocb *req)
2422 struct io_async_rw *rw = req->async_data;
2425 return !io_req_prep_async(req);
2426 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2427 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2431 static bool io_rw_should_reissue(struct io_kiocb *req)
2433 umode_t mode = file_inode(req->file)->i_mode;
2434 struct io_ring_ctx *ctx = req->ctx;
2436 if (!S_ISBLK(mode) && !S_ISREG(mode))
2438 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2439 !(ctx->flags & IORING_SETUP_IOPOLL)))
2442 * If ref is dying, we might be running poll reap from the exit work.
2443 * Don't attempt to reissue from that path, just let it fail with
2446 if (percpu_ref_is_dying(&ctx->refs))
2451 static bool io_resubmit_prep(struct io_kiocb *req)
2455 static bool io_rw_should_reissue(struct io_kiocb *req)
2461 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2462 unsigned int issue_flags)
2466 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2467 kiocb_end_write(req);
2468 if (res != req->result) {
2469 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2470 io_rw_should_reissue(req)) {
2471 req->flags |= REQ_F_REISSUE;
2474 req_set_fail_links(req);
2476 if (req->flags & REQ_F_BUFFER_SELECTED)
2477 cflags = io_put_rw_kbuf(req);
2478 __io_req_complete(req, issue_flags, res, cflags);
2481 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2483 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2485 __io_complete_rw(req, res, res2, 0);
2488 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2490 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2492 if (kiocb->ki_flags & IOCB_WRITE)
2493 kiocb_end_write(req);
2494 if (unlikely(res != req->result)) {
2495 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2496 io_resubmit_prep(req))) {
2497 req_set_fail_links(req);
2498 req->flags |= REQ_F_DONT_REISSUE;
2502 WRITE_ONCE(req->result, res);
2503 /* order with io_iopoll_complete() checking ->result */
2505 WRITE_ONCE(req->iopoll_completed, 1);
2509 * After the iocb has been issued, it's safe to be found on the poll list.
2510 * Adding the kiocb to the list AFTER submission ensures that we don't
2511 * find it from a io_do_iopoll() thread before the issuer is done
2512 * accessing the kiocb cookie.
2514 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2516 struct io_ring_ctx *ctx = req->ctx;
2519 * Track whether we have multiple files in our lists. This will impact
2520 * how we do polling eventually, not spinning if we're on potentially
2521 * different devices.
2523 if (list_empty(&ctx->iopoll_list)) {
2524 ctx->poll_multi_file = false;
2525 } else if (!ctx->poll_multi_file) {
2526 struct io_kiocb *list_req;
2528 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2530 if (list_req->file != req->file)
2531 ctx->poll_multi_file = true;
2535 * For fast devices, IO may have already completed. If it has, add
2536 * it to the front so we find it first.
2538 if (READ_ONCE(req->iopoll_completed))
2539 list_add(&req->inflight_entry, &ctx->iopoll_list);
2541 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2544 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2545 * task context or in io worker task context. If current task context is
2546 * sq thread, we don't need to check whether should wake up sq thread.
2548 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2549 wq_has_sleeper(&ctx->sq_data->wait))
2550 wake_up(&ctx->sq_data->wait);
2553 static inline void io_state_file_put(struct io_submit_state *state)
2555 if (state->file_refs) {
2556 fput_many(state->file, state->file_refs);
2557 state->file_refs = 0;
2562 * Get as many references to a file as we have IOs left in this submission,
2563 * assuming most submissions are for one file, or at least that each file
2564 * has more than one submission.
2566 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2571 if (state->file_refs) {
2572 if (state->fd == fd) {
2576 io_state_file_put(state);
2578 state->file = fget_many(fd, state->ios_left);
2579 if (unlikely(!state->file))
2583 state->file_refs = state->ios_left - 1;
2587 static bool io_bdev_nowait(struct block_device *bdev)
2589 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2593 * If we tracked the file through the SCM inflight mechanism, we could support
2594 * any file. For now, just ensure that anything potentially problematic is done
2597 static bool __io_file_supports_async(struct file *file, int rw)
2599 umode_t mode = file_inode(file)->i_mode;
2601 if (S_ISBLK(mode)) {
2602 if (IS_ENABLED(CONFIG_BLOCK) &&
2603 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2607 if (S_ISCHR(mode) || S_ISSOCK(mode))
2609 if (S_ISREG(mode)) {
2610 if (IS_ENABLED(CONFIG_BLOCK) &&
2611 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2612 file->f_op != &io_uring_fops)
2617 /* any ->read/write should understand O_NONBLOCK */
2618 if (file->f_flags & O_NONBLOCK)
2621 if (!(file->f_mode & FMODE_NOWAIT))
2625 return file->f_op->read_iter != NULL;
2627 return file->f_op->write_iter != NULL;
2630 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2632 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2634 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2637 return __io_file_supports_async(req->file, rw);
2640 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2642 struct io_ring_ctx *ctx = req->ctx;
2643 struct kiocb *kiocb = &req->rw.kiocb;
2644 struct file *file = req->file;
2648 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2649 req->flags |= REQ_F_ISREG;
2651 kiocb->ki_pos = READ_ONCE(sqe->off);
2652 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2653 req->flags |= REQ_F_CUR_POS;
2654 kiocb->ki_pos = file->f_pos;
2656 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2657 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2658 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2662 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2663 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2664 req->flags |= REQ_F_NOWAIT;
2666 ioprio = READ_ONCE(sqe->ioprio);
2668 ret = ioprio_check_cap(ioprio);
2672 kiocb->ki_ioprio = ioprio;
2674 kiocb->ki_ioprio = get_current_ioprio();
2676 if (ctx->flags & IORING_SETUP_IOPOLL) {
2677 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2678 !kiocb->ki_filp->f_op->iopoll)
2681 kiocb->ki_flags |= IOCB_HIPRI;
2682 kiocb->ki_complete = io_complete_rw_iopoll;
2683 req->iopoll_completed = 0;
2685 if (kiocb->ki_flags & IOCB_HIPRI)
2687 kiocb->ki_complete = io_complete_rw;
2690 req->rw.addr = READ_ONCE(sqe->addr);
2691 req->rw.len = READ_ONCE(sqe->len);
2692 req->buf_index = READ_ONCE(sqe->buf_index);
2696 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2702 case -ERESTARTNOINTR:
2703 case -ERESTARTNOHAND:
2704 case -ERESTART_RESTARTBLOCK:
2706 * We can't just restart the syscall, since previously
2707 * submitted sqes may already be in progress. Just fail this
2713 kiocb->ki_complete(kiocb, ret, 0);
2717 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2718 unsigned int issue_flags)
2720 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2721 struct io_async_rw *io = req->async_data;
2722 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2724 /* add previously done IO, if any */
2725 if (io && io->bytes_done > 0) {
2727 ret = io->bytes_done;
2729 ret += io->bytes_done;
2732 if (req->flags & REQ_F_CUR_POS)
2733 req->file->f_pos = kiocb->ki_pos;
2734 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2735 __io_complete_rw(req, ret, 0, issue_flags);
2737 io_rw_done(kiocb, ret);
2739 if (check_reissue && req->flags & REQ_F_REISSUE) {
2740 req->flags &= ~REQ_F_REISSUE;
2741 if (io_resubmit_prep(req)) {
2743 io_queue_async_work(req);
2747 req_set_fail_links(req);
2748 if (req->flags & REQ_F_BUFFER_SELECTED)
2749 cflags = io_put_rw_kbuf(req);
2750 __io_req_complete(req, issue_flags, ret, cflags);
2755 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2757 struct io_ring_ctx *ctx = req->ctx;
2758 size_t len = req->rw.len;
2759 struct io_mapped_ubuf *imu;
2760 u16 index, buf_index = req->buf_index;
2761 u64 buf_end, buf_addr = req->rw.addr;
2764 if (unlikely(buf_index >= ctx->nr_user_bufs))
2766 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2767 imu = &ctx->user_bufs[index];
2768 buf_addr = req->rw.addr;
2770 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2772 /* not inside the mapped region */
2773 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2777 * May not be a start of buffer, set size appropriately
2778 * and advance us to the beginning.
2780 offset = buf_addr - imu->ubuf;
2781 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2785 * Don't use iov_iter_advance() here, as it's really slow for
2786 * using the latter parts of a big fixed buffer - it iterates
2787 * over each segment manually. We can cheat a bit here, because
2790 * 1) it's a BVEC iter, we set it up
2791 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2792 * first and last bvec
2794 * So just find our index, and adjust the iterator afterwards.
2795 * If the offset is within the first bvec (or the whole first
2796 * bvec, just use iov_iter_advance(). This makes it easier
2797 * since we can just skip the first segment, which may not
2798 * be PAGE_SIZE aligned.
2800 const struct bio_vec *bvec = imu->bvec;
2802 if (offset <= bvec->bv_len) {
2803 iov_iter_advance(iter, offset);
2805 unsigned long seg_skip;
2807 /* skip first vec */
2808 offset -= bvec->bv_len;
2809 seg_skip = 1 + (offset >> PAGE_SHIFT);
2811 iter->bvec = bvec + seg_skip;
2812 iter->nr_segs -= seg_skip;
2813 iter->count -= bvec->bv_len + offset;
2814 iter->iov_offset = offset & ~PAGE_MASK;
2821 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2824 mutex_unlock(&ctx->uring_lock);
2827 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2830 * "Normal" inline submissions always hold the uring_lock, since we
2831 * grab it from the system call. Same is true for the SQPOLL offload.
2832 * The only exception is when we've detached the request and issue it
2833 * from an async worker thread, grab the lock for that case.
2836 mutex_lock(&ctx->uring_lock);
2839 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2840 int bgid, struct io_buffer *kbuf,
2843 struct io_buffer *head;
2845 if (req->flags & REQ_F_BUFFER_SELECTED)
2848 io_ring_submit_lock(req->ctx, needs_lock);
2850 lockdep_assert_held(&req->ctx->uring_lock);
2852 head = xa_load(&req->ctx->io_buffers, bgid);
2854 if (!list_empty(&head->list)) {
2855 kbuf = list_last_entry(&head->list, struct io_buffer,
2857 list_del(&kbuf->list);
2860 xa_erase(&req->ctx->io_buffers, bgid);
2862 if (*len > kbuf->len)
2865 kbuf = ERR_PTR(-ENOBUFS);
2868 io_ring_submit_unlock(req->ctx, needs_lock);
2873 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2876 struct io_buffer *kbuf;
2879 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2880 bgid = req->buf_index;
2881 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2884 req->rw.addr = (u64) (unsigned long) kbuf;
2885 req->flags |= REQ_F_BUFFER_SELECTED;
2886 return u64_to_user_ptr(kbuf->addr);
2889 #ifdef CONFIG_COMPAT
2890 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2893 struct compat_iovec __user *uiov;
2894 compat_ssize_t clen;
2898 uiov = u64_to_user_ptr(req->rw.addr);
2899 if (!access_ok(uiov, sizeof(*uiov)))
2901 if (__get_user(clen, &uiov->iov_len))
2907 buf = io_rw_buffer_select(req, &len, needs_lock);
2909 return PTR_ERR(buf);
2910 iov[0].iov_base = buf;
2911 iov[0].iov_len = (compat_size_t) len;
2916 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2919 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2923 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2926 len = iov[0].iov_len;
2929 buf = io_rw_buffer_select(req, &len, needs_lock);
2931 return PTR_ERR(buf);
2932 iov[0].iov_base = buf;
2933 iov[0].iov_len = len;
2937 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2940 if (req->flags & REQ_F_BUFFER_SELECTED) {
2941 struct io_buffer *kbuf;
2943 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2944 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2945 iov[0].iov_len = kbuf->len;
2948 if (req->rw.len != 1)
2951 #ifdef CONFIG_COMPAT
2952 if (req->ctx->compat)
2953 return io_compat_import(req, iov, needs_lock);
2956 return __io_iov_buffer_select(req, iov, needs_lock);
2959 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2960 struct iov_iter *iter, bool needs_lock)
2962 void __user *buf = u64_to_user_ptr(req->rw.addr);
2963 size_t sqe_len = req->rw.len;
2964 u8 opcode = req->opcode;
2967 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2969 return io_import_fixed(req, rw, iter);
2972 /* buffer index only valid with fixed read/write, or buffer select */
2973 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2976 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2977 if (req->flags & REQ_F_BUFFER_SELECT) {
2978 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2980 return PTR_ERR(buf);
2981 req->rw.len = sqe_len;
2984 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2989 if (req->flags & REQ_F_BUFFER_SELECT) {
2990 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2992 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2997 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3001 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3003 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3007 * For files that don't have ->read_iter() and ->write_iter(), handle them
3008 * by looping over ->read() or ->write() manually.
3010 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3012 struct kiocb *kiocb = &req->rw.kiocb;
3013 struct file *file = req->file;
3017 * Don't support polled IO through this interface, and we can't
3018 * support non-blocking either. For the latter, this just causes
3019 * the kiocb to be handled from an async context.
3021 if (kiocb->ki_flags & IOCB_HIPRI)
3023 if (kiocb->ki_flags & IOCB_NOWAIT)
3026 while (iov_iter_count(iter)) {
3030 if (!iov_iter_is_bvec(iter)) {
3031 iovec = iov_iter_iovec(iter);
3033 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3034 iovec.iov_len = req->rw.len;
3038 nr = file->f_op->read(file, iovec.iov_base,
3039 iovec.iov_len, io_kiocb_ppos(kiocb));
3041 nr = file->f_op->write(file, iovec.iov_base,
3042 iovec.iov_len, io_kiocb_ppos(kiocb));
3051 if (nr != iovec.iov_len)
3055 iov_iter_advance(iter, nr);
3061 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3062 const struct iovec *fast_iov, struct iov_iter *iter)
3064 struct io_async_rw *rw = req->async_data;
3066 memcpy(&rw->iter, iter, sizeof(*iter));
3067 rw->free_iovec = iovec;
3069 /* can only be fixed buffers, no need to do anything */
3070 if (iov_iter_is_bvec(iter))
3073 unsigned iov_off = 0;
3075 rw->iter.iov = rw->fast_iov;
3076 if (iter->iov != fast_iov) {
3077 iov_off = iter->iov - fast_iov;
3078 rw->iter.iov += iov_off;
3080 if (rw->fast_iov != fast_iov)
3081 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3082 sizeof(struct iovec) * iter->nr_segs);
3084 req->flags |= REQ_F_NEED_CLEANUP;
3088 static inline int io_alloc_async_data(struct io_kiocb *req)
3090 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3091 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3092 return req->async_data == NULL;
3095 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3096 const struct iovec *fast_iov,
3097 struct iov_iter *iter, bool force)
3099 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3101 if (!req->async_data) {
3102 if (io_alloc_async_data(req)) {
3107 io_req_map_rw(req, iovec, fast_iov, iter);
3112 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3114 struct io_async_rw *iorw = req->async_data;
3115 struct iovec *iov = iorw->fast_iov;
3118 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3119 if (unlikely(ret < 0))
3122 iorw->bytes_done = 0;
3123 iorw->free_iovec = iov;
3125 req->flags |= REQ_F_NEED_CLEANUP;
3129 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3131 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3133 return io_prep_rw(req, sqe);
3137 * This is our waitqueue callback handler, registered through lock_page_async()
3138 * when we initially tried to do the IO with the iocb armed our waitqueue.
3139 * This gets called when the page is unlocked, and we generally expect that to
3140 * happen when the page IO is completed and the page is now uptodate. This will
3141 * queue a task_work based retry of the operation, attempting to copy the data
3142 * again. If the latter fails because the page was NOT uptodate, then we will
3143 * do a thread based blocking retry of the operation. That's the unexpected
3146 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3147 int sync, void *arg)
3149 struct wait_page_queue *wpq;
3150 struct io_kiocb *req = wait->private;
3151 struct wait_page_key *key = arg;
3153 wpq = container_of(wait, struct wait_page_queue, wait);
3155 if (!wake_page_match(wpq, key))
3158 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3159 list_del_init(&wait->entry);
3161 /* submit ref gets dropped, acquire a new one */
3163 io_req_task_queue(req);
3168 * This controls whether a given IO request should be armed for async page
3169 * based retry. If we return false here, the request is handed to the async
3170 * worker threads for retry. If we're doing buffered reads on a regular file,
3171 * we prepare a private wait_page_queue entry and retry the operation. This
3172 * will either succeed because the page is now uptodate and unlocked, or it
3173 * will register a callback when the page is unlocked at IO completion. Through
3174 * that callback, io_uring uses task_work to setup a retry of the operation.
3175 * That retry will attempt the buffered read again. The retry will generally
3176 * succeed, or in rare cases where it fails, we then fall back to using the
3177 * async worker threads for a blocking retry.
3179 static bool io_rw_should_retry(struct io_kiocb *req)
3181 struct io_async_rw *rw = req->async_data;
3182 struct wait_page_queue *wait = &rw->wpq;
3183 struct kiocb *kiocb = &req->rw.kiocb;
3185 /* never retry for NOWAIT, we just complete with -EAGAIN */
3186 if (req->flags & REQ_F_NOWAIT)
3189 /* Only for buffered IO */
3190 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3194 * just use poll if we can, and don't attempt if the fs doesn't
3195 * support callback based unlocks
3197 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3200 wait->wait.func = io_async_buf_func;
3201 wait->wait.private = req;
3202 wait->wait.flags = 0;
3203 INIT_LIST_HEAD(&wait->wait.entry);
3204 kiocb->ki_flags |= IOCB_WAITQ;
3205 kiocb->ki_flags &= ~IOCB_NOWAIT;
3206 kiocb->ki_waitq = wait;
3210 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3212 if (req->file->f_op->read_iter)
3213 return call_read_iter(req->file, &req->rw.kiocb, iter);
3214 else if (req->file->f_op->read)
3215 return loop_rw_iter(READ, req, iter);
3220 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3222 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3223 struct kiocb *kiocb = &req->rw.kiocb;
3224 struct iov_iter __iter, *iter = &__iter;
3225 struct io_async_rw *rw = req->async_data;
3226 ssize_t io_size, ret, ret2;
3227 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3233 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3237 io_size = iov_iter_count(iter);
3238 req->result = io_size;
3240 /* Ensure we clear previously set non-block flag */
3241 if (!force_nonblock)
3242 kiocb->ki_flags &= ~IOCB_NOWAIT;
3244 kiocb->ki_flags |= IOCB_NOWAIT;
3246 /* If the file doesn't support async, just async punt */
3247 if (force_nonblock && !io_file_supports_async(req, READ)) {
3248 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3249 return ret ?: -EAGAIN;
3252 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3253 if (unlikely(ret)) {
3258 ret = io_iter_do_read(req, iter);
3260 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3261 req->flags &= ~REQ_F_REISSUE;
3262 /* IOPOLL retry should happen for io-wq threads */
3263 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3265 /* no retry on NONBLOCK nor RWF_NOWAIT */
3266 if (req->flags & REQ_F_NOWAIT)
3268 /* some cases will consume bytes even on error returns */
3269 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3271 } else if (ret == -EIOCBQUEUED) {
3273 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3274 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3275 /* read all, failed, already did sync or don't want to retry */
3279 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3284 rw = req->async_data;
3285 /* now use our persistent iterator, if we aren't already */
3290 rw->bytes_done += ret;
3291 /* if we can retry, do so with the callbacks armed */
3292 if (!io_rw_should_retry(req)) {
3293 kiocb->ki_flags &= ~IOCB_WAITQ;
3298 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3299 * we get -EIOCBQUEUED, then we'll get a notification when the
3300 * desired page gets unlocked. We can also get a partial read
3301 * here, and if we do, then just retry at the new offset.
3303 ret = io_iter_do_read(req, iter);
3304 if (ret == -EIOCBQUEUED)
3306 /* we got some bytes, but not all. retry. */
3307 kiocb->ki_flags &= ~IOCB_WAITQ;
3308 } while (ret > 0 && ret < io_size);
3310 kiocb_done(kiocb, ret, issue_flags);
3312 /* it's faster to check here then delegate to kfree */
3318 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3320 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3322 return io_prep_rw(req, sqe);
3325 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3327 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3328 struct kiocb *kiocb = &req->rw.kiocb;
3329 struct iov_iter __iter, *iter = &__iter;
3330 struct io_async_rw *rw = req->async_data;
3331 ssize_t ret, ret2, io_size;
3332 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3338 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3342 io_size = iov_iter_count(iter);
3343 req->result = io_size;
3345 /* Ensure we clear previously set non-block flag */
3346 if (!force_nonblock)
3347 kiocb->ki_flags &= ~IOCB_NOWAIT;
3349 kiocb->ki_flags |= IOCB_NOWAIT;
3351 /* If the file doesn't support async, just async punt */
3352 if (force_nonblock && !io_file_supports_async(req, WRITE))
3355 /* file path doesn't support NOWAIT for non-direct_IO */
3356 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3357 (req->flags & REQ_F_ISREG))
3360 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3365 * Open-code file_start_write here to grab freeze protection,
3366 * which will be released by another thread in
3367 * io_complete_rw(). Fool lockdep by telling it the lock got
3368 * released so that it doesn't complain about the held lock when
3369 * we return to userspace.
3371 if (req->flags & REQ_F_ISREG) {
3372 sb_start_write(file_inode(req->file)->i_sb);
3373 __sb_writers_release(file_inode(req->file)->i_sb,
3376 kiocb->ki_flags |= IOCB_WRITE;
3378 if (req->file->f_op->write_iter)
3379 ret2 = call_write_iter(req->file, kiocb, iter);
3380 else if (req->file->f_op->write)
3381 ret2 = loop_rw_iter(WRITE, req, iter);
3385 if (req->flags & REQ_F_REISSUE) {
3386 req->flags &= ~REQ_F_REISSUE;
3391 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3392 * retry them without IOCB_NOWAIT.
3394 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3396 /* no retry on NONBLOCK nor RWF_NOWAIT */
3397 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3399 if (!force_nonblock || ret2 != -EAGAIN) {
3400 /* IOPOLL retry should happen for io-wq threads */
3401 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3404 kiocb_done(kiocb, ret2, issue_flags);
3407 /* some cases will consume bytes even on error returns */
3408 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3409 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3410 return ret ?: -EAGAIN;
3413 /* it's reportedly faster than delegating the null check to kfree() */
3419 static int io_renameat_prep(struct io_kiocb *req,
3420 const struct io_uring_sqe *sqe)
3422 struct io_rename *ren = &req->rename;
3423 const char __user *oldf, *newf;
3425 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3428 ren->old_dfd = READ_ONCE(sqe->fd);
3429 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3430 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3431 ren->new_dfd = READ_ONCE(sqe->len);
3432 ren->flags = READ_ONCE(sqe->rename_flags);
3434 ren->oldpath = getname(oldf);
3435 if (IS_ERR(ren->oldpath))
3436 return PTR_ERR(ren->oldpath);
3438 ren->newpath = getname(newf);
3439 if (IS_ERR(ren->newpath)) {
3440 putname(ren->oldpath);
3441 return PTR_ERR(ren->newpath);
3444 req->flags |= REQ_F_NEED_CLEANUP;
3448 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3450 struct io_rename *ren = &req->rename;
3453 if (issue_flags & IO_URING_F_NONBLOCK)
3456 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3457 ren->newpath, ren->flags);
3459 req->flags &= ~REQ_F_NEED_CLEANUP;
3461 req_set_fail_links(req);
3462 io_req_complete(req, ret);
3466 static int io_unlinkat_prep(struct io_kiocb *req,
3467 const struct io_uring_sqe *sqe)
3469 struct io_unlink *un = &req->unlink;
3470 const char __user *fname;
3472 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3475 un->dfd = READ_ONCE(sqe->fd);
3477 un->flags = READ_ONCE(sqe->unlink_flags);
3478 if (un->flags & ~AT_REMOVEDIR)
3481 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3482 un->filename = getname(fname);
3483 if (IS_ERR(un->filename))
3484 return PTR_ERR(un->filename);
3486 req->flags |= REQ_F_NEED_CLEANUP;
3490 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3492 struct io_unlink *un = &req->unlink;
3495 if (issue_flags & IO_URING_F_NONBLOCK)
3498 if (un->flags & AT_REMOVEDIR)
3499 ret = do_rmdir(un->dfd, un->filename);
3501 ret = do_unlinkat(un->dfd, un->filename);
3503 req->flags &= ~REQ_F_NEED_CLEANUP;
3505 req_set_fail_links(req);
3506 io_req_complete(req, ret);
3510 static int io_shutdown_prep(struct io_kiocb *req,
3511 const struct io_uring_sqe *sqe)
3513 #if defined(CONFIG_NET)
3514 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3516 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3520 req->shutdown.how = READ_ONCE(sqe->len);
3527 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3529 #if defined(CONFIG_NET)
3530 struct socket *sock;
3533 if (issue_flags & IO_URING_F_NONBLOCK)
3536 sock = sock_from_file(req->file);
3537 if (unlikely(!sock))
3540 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3542 req_set_fail_links(req);
3543 io_req_complete(req, ret);
3550 static int __io_splice_prep(struct io_kiocb *req,
3551 const struct io_uring_sqe *sqe)
3553 struct io_splice* sp = &req->splice;
3554 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3560 sp->len = READ_ONCE(sqe->len);
3561 sp->flags = READ_ONCE(sqe->splice_flags);
3563 if (unlikely(sp->flags & ~valid_flags))
3566 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3567 (sp->flags & SPLICE_F_FD_IN_FIXED));
3570 req->flags |= REQ_F_NEED_CLEANUP;
3574 static int io_tee_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3579 return __io_splice_prep(req, sqe);
3582 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3584 struct io_splice *sp = &req->splice;
3585 struct file *in = sp->file_in;
3586 struct file *out = sp->file_out;
3587 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3590 if (issue_flags & IO_URING_F_NONBLOCK)
3593 ret = do_tee(in, out, sp->len, flags);
3595 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3597 req->flags &= ~REQ_F_NEED_CLEANUP;
3600 req_set_fail_links(req);
3601 io_req_complete(req, ret);
3605 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3607 struct io_splice* sp = &req->splice;
3609 sp->off_in = READ_ONCE(sqe->splice_off_in);
3610 sp->off_out = READ_ONCE(sqe->off);
3611 return __io_splice_prep(req, sqe);
3614 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3616 struct io_splice *sp = &req->splice;
3617 struct file *in = sp->file_in;
3618 struct file *out = sp->file_out;
3619 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3620 loff_t *poff_in, *poff_out;
3623 if (issue_flags & IO_URING_F_NONBLOCK)
3626 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3627 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3630 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3632 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3634 req->flags &= ~REQ_F_NEED_CLEANUP;
3637 req_set_fail_links(req);
3638 io_req_complete(req, ret);
3643 * IORING_OP_NOP just posts a completion event, nothing else.
3645 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3647 struct io_ring_ctx *ctx = req->ctx;
3649 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3652 __io_req_complete(req, issue_flags, 0, 0);
3656 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3658 struct io_ring_ctx *ctx = req->ctx;
3663 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3665 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3668 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3669 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3672 req->sync.off = READ_ONCE(sqe->off);
3673 req->sync.len = READ_ONCE(sqe->len);
3677 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3679 loff_t end = req->sync.off + req->sync.len;
3682 /* fsync always requires a blocking context */
3683 if (issue_flags & IO_URING_F_NONBLOCK)
3686 ret = vfs_fsync_range(req->file, req->sync.off,
3687 end > 0 ? end : LLONG_MAX,
3688 req->sync.flags & IORING_FSYNC_DATASYNC);
3690 req_set_fail_links(req);
3691 io_req_complete(req, ret);
3695 static int io_fallocate_prep(struct io_kiocb *req,
3696 const struct io_uring_sqe *sqe)
3698 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3703 req->sync.off = READ_ONCE(sqe->off);
3704 req->sync.len = READ_ONCE(sqe->addr);
3705 req->sync.mode = READ_ONCE(sqe->len);
3709 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3713 /* fallocate always requiring blocking context */
3714 if (issue_flags & IO_URING_F_NONBLOCK)
3716 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3719 req_set_fail_links(req);
3720 io_req_complete(req, ret);
3724 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3726 const char __user *fname;
3729 if (unlikely(sqe->ioprio || sqe->buf_index))
3731 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3734 /* open.how should be already initialised */
3735 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3736 req->open.how.flags |= O_LARGEFILE;
3738 req->open.dfd = READ_ONCE(sqe->fd);
3739 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3740 req->open.filename = getname(fname);
3741 if (IS_ERR(req->open.filename)) {
3742 ret = PTR_ERR(req->open.filename);
3743 req->open.filename = NULL;
3746 req->open.nofile = rlimit(RLIMIT_NOFILE);
3747 req->flags |= REQ_F_NEED_CLEANUP;
3751 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3757 mode = READ_ONCE(sqe->len);
3758 flags = READ_ONCE(sqe->open_flags);
3759 req->open.how = build_open_how(flags, mode);
3760 return __io_openat_prep(req, sqe);
3763 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3765 struct open_how __user *how;
3769 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3771 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3772 len = READ_ONCE(sqe->len);
3773 if (len < OPEN_HOW_SIZE_VER0)
3776 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3781 return __io_openat_prep(req, sqe);
3784 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3786 struct open_flags op;
3789 bool resolve_nonblock;
3792 ret = build_open_flags(&req->open.how, &op);
3795 nonblock_set = op.open_flag & O_NONBLOCK;
3796 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3797 if (issue_flags & IO_URING_F_NONBLOCK) {
3799 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3800 * it'll always -EAGAIN
3802 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3804 op.lookup_flags |= LOOKUP_CACHED;
3805 op.open_flag |= O_NONBLOCK;
3808 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3812 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3813 /* only retry if RESOLVE_CACHED wasn't already set by application */
3814 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3815 file == ERR_PTR(-EAGAIN)) {
3817 * We could hang on to this 'fd', but seems like marginal
3818 * gain for something that is now known to be a slower path.
3819 * So just put it, and we'll get a new one when we retry.
3827 ret = PTR_ERR(file);
3829 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3830 file->f_flags &= ~O_NONBLOCK;
3831 fsnotify_open(file);
3832 fd_install(ret, file);
3835 putname(req->open.filename);
3836 req->flags &= ~REQ_F_NEED_CLEANUP;
3838 req_set_fail_links(req);
3839 __io_req_complete(req, issue_flags, ret, 0);
3843 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3845 return io_openat2(req, issue_flags);
3848 static int io_remove_buffers_prep(struct io_kiocb *req,
3849 const struct io_uring_sqe *sqe)
3851 struct io_provide_buf *p = &req->pbuf;
3854 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3857 tmp = READ_ONCE(sqe->fd);
3858 if (!tmp || tmp > USHRT_MAX)
3861 memset(p, 0, sizeof(*p));
3863 p->bgid = READ_ONCE(sqe->buf_group);
3867 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3868 int bgid, unsigned nbufs)
3872 /* shouldn't happen */
3876 /* the head kbuf is the list itself */
3877 while (!list_empty(&buf->list)) {
3878 struct io_buffer *nxt;
3880 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3881 list_del(&nxt->list);
3888 xa_erase(&ctx->io_buffers, bgid);
3893 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3895 struct io_provide_buf *p = &req->pbuf;
3896 struct io_ring_ctx *ctx = req->ctx;
3897 struct io_buffer *head;
3899 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3901 io_ring_submit_lock(ctx, !force_nonblock);
3903 lockdep_assert_held(&ctx->uring_lock);
3906 head = xa_load(&ctx->io_buffers, p->bgid);
3908 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3910 req_set_fail_links(req);
3912 /* complete before unlock, IOPOLL may need the lock */
3913 __io_req_complete(req, issue_flags, ret, 0);
3914 io_ring_submit_unlock(ctx, !force_nonblock);
3918 static int io_provide_buffers_prep(struct io_kiocb *req,
3919 const struct io_uring_sqe *sqe)
3921 unsigned long size, tmp_check;
3922 struct io_provide_buf *p = &req->pbuf;
3925 if (sqe->ioprio || sqe->rw_flags)
3928 tmp = READ_ONCE(sqe->fd);
3929 if (!tmp || tmp > USHRT_MAX)
3932 p->addr = READ_ONCE(sqe->addr);
3933 p->len = READ_ONCE(sqe->len);
3935 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3938 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3941 size = (unsigned long)p->len * p->nbufs;
3942 if (!access_ok(u64_to_user_ptr(p->addr), size))
3945 p->bgid = READ_ONCE(sqe->buf_group);
3946 tmp = READ_ONCE(sqe->off);
3947 if (tmp > USHRT_MAX)
3953 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3955 struct io_buffer *buf;
3956 u64 addr = pbuf->addr;
3957 int i, bid = pbuf->bid;
3959 for (i = 0; i < pbuf->nbufs; i++) {
3960 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3965 buf->len = pbuf->len;
3970 INIT_LIST_HEAD(&buf->list);
3973 list_add_tail(&buf->list, &(*head)->list);
3977 return i ? i : -ENOMEM;
3980 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3982 struct io_provide_buf *p = &req->pbuf;
3983 struct io_ring_ctx *ctx = req->ctx;
3984 struct io_buffer *head, *list;
3986 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3988 io_ring_submit_lock(ctx, !force_nonblock);
3990 lockdep_assert_held(&ctx->uring_lock);
3992 list = head = xa_load(&ctx->io_buffers, p->bgid);
3994 ret = io_add_buffers(p, &head);
3995 if (ret >= 0 && !list) {
3996 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
3998 __io_remove_buffers(ctx, head, p->bgid, -1U);
4001 req_set_fail_links(req);
4002 /* complete before unlock, IOPOLL may need the lock */
4003 __io_req_complete(req, issue_flags, ret, 0);
4004 io_ring_submit_unlock(ctx, !force_nonblock);
4008 static int io_epoll_ctl_prep(struct io_kiocb *req,
4009 const struct io_uring_sqe *sqe)
4011 #if defined(CONFIG_EPOLL)
4012 if (sqe->ioprio || sqe->buf_index)
4014 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4017 req->epoll.epfd = READ_ONCE(sqe->fd);
4018 req->epoll.op = READ_ONCE(sqe->len);
4019 req->epoll.fd = READ_ONCE(sqe->off);
4021 if (ep_op_has_event(req->epoll.op)) {
4022 struct epoll_event __user *ev;
4024 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4025 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4035 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4037 #if defined(CONFIG_EPOLL)
4038 struct io_epoll *ie = &req->epoll;
4040 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4042 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4043 if (force_nonblock && ret == -EAGAIN)
4047 req_set_fail_links(req);
4048 __io_req_complete(req, issue_flags, ret, 0);
4055 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4057 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4058 if (sqe->ioprio || sqe->buf_index || sqe->off)
4060 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4063 req->madvise.addr = READ_ONCE(sqe->addr);
4064 req->madvise.len = READ_ONCE(sqe->len);
4065 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4072 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4074 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4075 struct io_madvise *ma = &req->madvise;
4078 if (issue_flags & IO_URING_F_NONBLOCK)
4081 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4083 req_set_fail_links(req);
4084 io_req_complete(req, ret);
4091 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4093 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4095 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4098 req->fadvise.offset = READ_ONCE(sqe->off);
4099 req->fadvise.len = READ_ONCE(sqe->len);
4100 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4104 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4106 struct io_fadvise *fa = &req->fadvise;
4109 if (issue_flags & IO_URING_F_NONBLOCK) {
4110 switch (fa->advice) {
4111 case POSIX_FADV_NORMAL:
4112 case POSIX_FADV_RANDOM:
4113 case POSIX_FADV_SEQUENTIAL:
4120 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4122 req_set_fail_links(req);
4123 __io_req_complete(req, issue_flags, ret, 0);
4127 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4131 if (sqe->ioprio || sqe->buf_index)
4133 if (req->flags & REQ_F_FIXED_FILE)
4136 req->statx.dfd = READ_ONCE(sqe->fd);
4137 req->statx.mask = READ_ONCE(sqe->len);
4138 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4139 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4140 req->statx.flags = READ_ONCE(sqe->statx_flags);
4145 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4147 struct io_statx *ctx = &req->statx;
4150 if (issue_flags & IO_URING_F_NONBLOCK)
4153 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4157 req_set_fail_links(req);
4158 io_req_complete(req, ret);
4162 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4164 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4166 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4167 sqe->rw_flags || sqe->buf_index)
4169 if (req->flags & REQ_F_FIXED_FILE)
4172 req->close.fd = READ_ONCE(sqe->fd);
4176 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4178 struct files_struct *files = current->files;
4179 struct io_close *close = &req->close;
4180 struct fdtable *fdt;
4181 struct file *file = NULL;
4184 spin_lock(&files->file_lock);
4185 fdt = files_fdtable(files);
4186 if (close->fd >= fdt->max_fds) {
4187 spin_unlock(&files->file_lock);
4190 file = fdt->fd[close->fd];
4191 if (!file || file->f_op == &io_uring_fops) {
4192 spin_unlock(&files->file_lock);
4197 /* if the file has a flush method, be safe and punt to async */
4198 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4199 spin_unlock(&files->file_lock);
4203 ret = __close_fd_get_file(close->fd, &file);
4204 spin_unlock(&files->file_lock);
4211 /* No ->flush() or already async, safely close from here */
4212 ret = filp_close(file, current->files);
4215 req_set_fail_links(req);
4218 __io_req_complete(req, issue_flags, ret, 0);
4222 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4224 struct io_ring_ctx *ctx = req->ctx;
4226 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4228 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4231 req->sync.off = READ_ONCE(sqe->off);
4232 req->sync.len = READ_ONCE(sqe->len);
4233 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4237 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4241 /* sync_file_range always requires a blocking context */
4242 if (issue_flags & IO_URING_F_NONBLOCK)
4245 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4248 req_set_fail_links(req);
4249 io_req_complete(req, ret);
4253 #if defined(CONFIG_NET)
4254 static int io_setup_async_msg(struct io_kiocb *req,
4255 struct io_async_msghdr *kmsg)
4257 struct io_async_msghdr *async_msg = req->async_data;
4261 if (io_alloc_async_data(req)) {
4262 kfree(kmsg->free_iov);
4265 async_msg = req->async_data;
4266 req->flags |= REQ_F_NEED_CLEANUP;
4267 memcpy(async_msg, kmsg, sizeof(*kmsg));
4268 async_msg->msg.msg_name = &async_msg->addr;
4269 /* if were using fast_iov, set it to the new one */
4270 if (!async_msg->free_iov)
4271 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4276 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4277 struct io_async_msghdr *iomsg)
4279 iomsg->msg.msg_name = &iomsg->addr;
4280 iomsg->free_iov = iomsg->fast_iov;
4281 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4282 req->sr_msg.msg_flags, &iomsg->free_iov);
4285 static int io_sendmsg_prep_async(struct io_kiocb *req)
4289 ret = io_sendmsg_copy_hdr(req, req->async_data);
4291 req->flags |= REQ_F_NEED_CLEANUP;
4295 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4297 struct io_sr_msg *sr = &req->sr_msg;
4299 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4302 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4303 sr->len = READ_ONCE(sqe->len);
4304 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4305 if (sr->msg_flags & MSG_DONTWAIT)
4306 req->flags |= REQ_F_NOWAIT;
4308 #ifdef CONFIG_COMPAT
4309 if (req->ctx->compat)
4310 sr->msg_flags |= MSG_CMSG_COMPAT;
4315 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4317 struct io_async_msghdr iomsg, *kmsg;
4318 struct socket *sock;
4323 sock = sock_from_file(req->file);
4324 if (unlikely(!sock))
4327 kmsg = req->async_data;
4329 ret = io_sendmsg_copy_hdr(req, &iomsg);
4335 flags = req->sr_msg.msg_flags;
4336 if (issue_flags & IO_URING_F_NONBLOCK)
4337 flags |= MSG_DONTWAIT;
4338 if (flags & MSG_WAITALL)
4339 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4341 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4342 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4343 return io_setup_async_msg(req, kmsg);
4344 if (ret == -ERESTARTSYS)
4347 /* fast path, check for non-NULL to avoid function call */
4349 kfree(kmsg->free_iov);
4350 req->flags &= ~REQ_F_NEED_CLEANUP;
4352 req_set_fail_links(req);
4353 __io_req_complete(req, issue_flags, ret, 0);
4357 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4359 struct io_sr_msg *sr = &req->sr_msg;
4362 struct socket *sock;
4367 sock = sock_from_file(req->file);
4368 if (unlikely(!sock))
4371 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4375 msg.msg_name = NULL;
4376 msg.msg_control = NULL;
4377 msg.msg_controllen = 0;
4378 msg.msg_namelen = 0;
4380 flags = req->sr_msg.msg_flags;
4381 if (issue_flags & IO_URING_F_NONBLOCK)
4382 flags |= MSG_DONTWAIT;
4383 if (flags & MSG_WAITALL)
4384 min_ret = iov_iter_count(&msg.msg_iter);
4386 msg.msg_flags = flags;
4387 ret = sock_sendmsg(sock, &msg);
4388 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4390 if (ret == -ERESTARTSYS)
4394 req_set_fail_links(req);
4395 __io_req_complete(req, issue_flags, ret, 0);
4399 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4400 struct io_async_msghdr *iomsg)
4402 struct io_sr_msg *sr = &req->sr_msg;
4403 struct iovec __user *uiov;
4407 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4408 &iomsg->uaddr, &uiov, &iov_len);
4412 if (req->flags & REQ_F_BUFFER_SELECT) {
4415 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4417 sr->len = iomsg->fast_iov[0].iov_len;
4418 iomsg->free_iov = NULL;
4420 iomsg->free_iov = iomsg->fast_iov;
4421 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4422 &iomsg->free_iov, &iomsg->msg.msg_iter,
4431 #ifdef CONFIG_COMPAT
4432 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4433 struct io_async_msghdr *iomsg)
4435 struct io_sr_msg *sr = &req->sr_msg;
4436 struct compat_iovec __user *uiov;
4441 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4446 uiov = compat_ptr(ptr);
4447 if (req->flags & REQ_F_BUFFER_SELECT) {
4448 compat_ssize_t clen;
4452 if (!access_ok(uiov, sizeof(*uiov)))
4454 if (__get_user(clen, &uiov->iov_len))
4459 iomsg->free_iov = NULL;
4461 iomsg->free_iov = iomsg->fast_iov;
4462 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4463 UIO_FASTIOV, &iomsg->free_iov,
4464 &iomsg->msg.msg_iter, true);
4473 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4474 struct io_async_msghdr *iomsg)
4476 iomsg->msg.msg_name = &iomsg->addr;
4478 #ifdef CONFIG_COMPAT
4479 if (req->ctx->compat)
4480 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4483 return __io_recvmsg_copy_hdr(req, iomsg);
4486 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4489 struct io_sr_msg *sr = &req->sr_msg;
4490 struct io_buffer *kbuf;
4492 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4497 req->flags |= REQ_F_BUFFER_SELECTED;
4501 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4503 return io_put_kbuf(req, req->sr_msg.kbuf);
4506 static int io_recvmsg_prep_async(struct io_kiocb *req)
4510 ret = io_recvmsg_copy_hdr(req, req->async_data);
4512 req->flags |= REQ_F_NEED_CLEANUP;
4516 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4518 struct io_sr_msg *sr = &req->sr_msg;
4520 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4523 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4524 sr->len = READ_ONCE(sqe->len);
4525 sr->bgid = READ_ONCE(sqe->buf_group);
4526 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4527 if (sr->msg_flags & MSG_DONTWAIT)
4528 req->flags |= REQ_F_NOWAIT;
4530 #ifdef CONFIG_COMPAT
4531 if (req->ctx->compat)
4532 sr->msg_flags |= MSG_CMSG_COMPAT;
4537 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4539 struct io_async_msghdr iomsg, *kmsg;
4540 struct socket *sock;
4541 struct io_buffer *kbuf;
4544 int ret, cflags = 0;
4545 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4547 sock = sock_from_file(req->file);
4548 if (unlikely(!sock))
4551 kmsg = req->async_data;
4553 ret = io_recvmsg_copy_hdr(req, &iomsg);
4559 if (req->flags & REQ_F_BUFFER_SELECT) {
4560 kbuf = io_recv_buffer_select(req, !force_nonblock);
4562 return PTR_ERR(kbuf);
4563 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4564 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4565 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4566 1, req->sr_msg.len);
4569 flags = req->sr_msg.msg_flags;
4571 flags |= MSG_DONTWAIT;
4572 if (flags & MSG_WAITALL)
4573 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4575 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4576 kmsg->uaddr, flags);
4577 if (force_nonblock && ret == -EAGAIN)
4578 return io_setup_async_msg(req, kmsg);
4579 if (ret == -ERESTARTSYS)
4582 if (req->flags & REQ_F_BUFFER_SELECTED)
4583 cflags = io_put_recv_kbuf(req);
4584 /* fast path, check for non-NULL to avoid function call */
4586 kfree(kmsg->free_iov);
4587 req->flags &= ~REQ_F_NEED_CLEANUP;
4588 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4589 req_set_fail_links(req);
4590 __io_req_complete(req, issue_flags, ret, cflags);
4594 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4596 struct io_buffer *kbuf;
4597 struct io_sr_msg *sr = &req->sr_msg;
4599 void __user *buf = sr->buf;
4600 struct socket *sock;
4604 int ret, cflags = 0;
4605 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4607 sock = sock_from_file(req->file);
4608 if (unlikely(!sock))
4611 if (req->flags & REQ_F_BUFFER_SELECT) {
4612 kbuf = io_recv_buffer_select(req, !force_nonblock);
4614 return PTR_ERR(kbuf);
4615 buf = u64_to_user_ptr(kbuf->addr);
4618 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4622 msg.msg_name = NULL;
4623 msg.msg_control = NULL;
4624 msg.msg_controllen = 0;
4625 msg.msg_namelen = 0;
4626 msg.msg_iocb = NULL;
4629 flags = req->sr_msg.msg_flags;
4631 flags |= MSG_DONTWAIT;
4632 if (flags & MSG_WAITALL)
4633 min_ret = iov_iter_count(&msg.msg_iter);
4635 ret = sock_recvmsg(sock, &msg, flags);
4636 if (force_nonblock && ret == -EAGAIN)
4638 if (ret == -ERESTARTSYS)
4641 if (req->flags & REQ_F_BUFFER_SELECTED)
4642 cflags = io_put_recv_kbuf(req);
4643 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4644 req_set_fail_links(req);
4645 __io_req_complete(req, issue_flags, ret, cflags);
4649 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4651 struct io_accept *accept = &req->accept;
4653 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4655 if (sqe->ioprio || sqe->len || sqe->buf_index)
4658 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4659 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4660 accept->flags = READ_ONCE(sqe->accept_flags);
4661 accept->nofile = rlimit(RLIMIT_NOFILE);
4665 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4667 struct io_accept *accept = &req->accept;
4668 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4669 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4672 if (req->file->f_flags & O_NONBLOCK)
4673 req->flags |= REQ_F_NOWAIT;
4675 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4676 accept->addr_len, accept->flags,
4678 if (ret == -EAGAIN && force_nonblock)
4681 if (ret == -ERESTARTSYS)
4683 req_set_fail_links(req);
4685 __io_req_complete(req, issue_flags, ret, 0);
4689 static int io_connect_prep_async(struct io_kiocb *req)
4691 struct io_async_connect *io = req->async_data;
4692 struct io_connect *conn = &req->connect;
4694 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4697 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4699 struct io_connect *conn = &req->connect;
4701 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4703 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4706 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4707 conn->addr_len = READ_ONCE(sqe->addr2);
4711 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4713 struct io_async_connect __io, *io;
4714 unsigned file_flags;
4716 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4718 if (req->async_data) {
4719 io = req->async_data;
4721 ret = move_addr_to_kernel(req->connect.addr,
4722 req->connect.addr_len,
4729 file_flags = force_nonblock ? O_NONBLOCK : 0;
4731 ret = __sys_connect_file(req->file, &io->address,
4732 req->connect.addr_len, file_flags);
4733 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4734 if (req->async_data)
4736 if (io_alloc_async_data(req)) {
4740 memcpy(req->async_data, &__io, sizeof(__io));
4743 if (ret == -ERESTARTSYS)
4747 req_set_fail_links(req);
4748 __io_req_complete(req, issue_flags, ret, 0);
4751 #else /* !CONFIG_NET */
4752 #define IO_NETOP_FN(op) \
4753 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4755 return -EOPNOTSUPP; \
4758 #define IO_NETOP_PREP(op) \
4760 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4762 return -EOPNOTSUPP; \
4765 #define IO_NETOP_PREP_ASYNC(op) \
4767 static int io_##op##_prep_async(struct io_kiocb *req) \
4769 return -EOPNOTSUPP; \
4772 IO_NETOP_PREP_ASYNC(sendmsg);
4773 IO_NETOP_PREP_ASYNC(recvmsg);
4774 IO_NETOP_PREP_ASYNC(connect);
4775 IO_NETOP_PREP(accept);
4778 #endif /* CONFIG_NET */
4780 struct io_poll_table {
4781 struct poll_table_struct pt;
4782 struct io_kiocb *req;
4786 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4787 __poll_t mask, task_work_func_t func)
4791 /* for instances that support it check for an event match first: */
4792 if (mask && !(mask & poll->events))
4795 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4797 list_del_init(&poll->wait.entry);
4800 req->task_work.func = func;
4803 * If this fails, then the task is exiting. When a task exits, the
4804 * work gets canceled, so just cancel this request as well instead
4805 * of executing it. We can't safely execute it anyway, as we may not
4806 * have the needed state needed for it anyway.
4808 ret = io_req_task_work_add(req);
4809 if (unlikely(ret)) {
4810 WRITE_ONCE(poll->canceled, true);
4811 io_req_task_work_add_fallback(req, func);
4816 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4817 __acquires(&req->ctx->completion_lock)
4819 struct io_ring_ctx *ctx = req->ctx;
4821 if (!req->result && !READ_ONCE(poll->canceled)) {
4822 struct poll_table_struct pt = { ._key = poll->events };
4824 req->result = vfs_poll(req->file, &pt) & poll->events;
4827 spin_lock_irq(&ctx->completion_lock);
4828 if (!req->result && !READ_ONCE(poll->canceled)) {
4829 add_wait_queue(poll->head, &poll->wait);
4836 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4838 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4839 if (req->opcode == IORING_OP_POLL_ADD)
4840 return req->async_data;
4841 return req->apoll->double_poll;
4844 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4846 if (req->opcode == IORING_OP_POLL_ADD)
4848 return &req->apoll->poll;
4851 static void io_poll_remove_double(struct io_kiocb *req)
4852 __must_hold(&req->ctx->completion_lock)
4854 struct io_poll_iocb *poll = io_poll_get_double(req);
4856 lockdep_assert_held(&req->ctx->completion_lock);
4858 if (poll && poll->head) {
4859 struct wait_queue_head *head = poll->head;
4861 spin_lock(&head->lock);
4862 list_del_init(&poll->wait.entry);
4863 if (poll->wait.private)
4866 spin_unlock(&head->lock);
4870 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4871 __must_hold(&req->ctx->completion_lock)
4873 struct io_ring_ctx *ctx = req->ctx;
4874 unsigned flags = IORING_CQE_F_MORE;
4877 if (READ_ONCE(req->poll.canceled)) {
4879 req->poll.events |= EPOLLONESHOT;
4881 error = mangle_poll(mask);
4883 if (req->poll.events & EPOLLONESHOT)
4885 if (!io_cqring_fill_event(req, error, flags)) {
4886 io_poll_remove_waitqs(req);
4887 req->poll.done = true;
4890 io_commit_cqring(ctx);
4891 return !(flags & IORING_CQE_F_MORE);
4894 static void io_poll_task_func(struct callback_head *cb)
4896 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4897 struct io_ring_ctx *ctx = req->ctx;
4898 struct io_kiocb *nxt;
4900 if (io_poll_rewait(req, &req->poll)) {
4901 spin_unlock_irq(&ctx->completion_lock);
4905 done = io_poll_complete(req, req->result);
4907 hash_del(&req->hash_node);
4910 add_wait_queue(req->poll.head, &req->poll.wait);
4912 spin_unlock_irq(&ctx->completion_lock);
4913 io_cqring_ev_posted(ctx);
4916 nxt = io_put_req_find_next(req);
4918 __io_req_task_submit(nxt);
4923 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4924 int sync, void *key)
4926 struct io_kiocb *req = wait->private;
4927 struct io_poll_iocb *poll = io_poll_get_single(req);
4928 __poll_t mask = key_to_poll(key);
4930 /* for instances that support it check for an event match first: */
4931 if (mask && !(mask & poll->events))
4933 if (!(poll->events & EPOLLONESHOT))
4934 return poll->wait.func(&poll->wait, mode, sync, key);
4936 list_del_init(&wait->entry);
4938 if (poll && poll->head) {
4941 spin_lock(&poll->head->lock);
4942 done = list_empty(&poll->wait.entry);
4944 list_del_init(&poll->wait.entry);
4945 /* make sure double remove sees this as being gone */
4946 wait->private = NULL;
4947 spin_unlock(&poll->head->lock);
4949 /* use wait func handler, so it matches the rq type */
4950 poll->wait.func(&poll->wait, mode, sync, key);
4957 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4958 wait_queue_func_t wake_func)
4962 poll->canceled = false;
4963 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4964 /* mask in events that we always want/need */
4965 poll->events = events | IO_POLL_UNMASK;
4966 INIT_LIST_HEAD(&poll->wait.entry);
4967 init_waitqueue_func_entry(&poll->wait, wake_func);
4970 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4971 struct wait_queue_head *head,
4972 struct io_poll_iocb **poll_ptr)
4974 struct io_kiocb *req = pt->req;
4977 * If poll->head is already set, it's because the file being polled
4978 * uses multiple waitqueues for poll handling (eg one for read, one
4979 * for write). Setup a separate io_poll_iocb if this happens.
4981 if (unlikely(poll->head)) {
4982 struct io_poll_iocb *poll_one = poll;
4984 /* already have a 2nd entry, fail a third attempt */
4986 pt->error = -EINVAL;
4990 * Can't handle multishot for double wait for now, turn it
4991 * into one-shot mode.
4993 if (!(req->poll.events & EPOLLONESHOT))
4994 req->poll.events |= EPOLLONESHOT;
4995 /* double add on the same waitqueue head, ignore */
4996 if (poll->head == head)
4998 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5000 pt->error = -ENOMEM;
5003 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5005 poll->wait.private = req;
5012 if (poll->events & EPOLLEXCLUSIVE)
5013 add_wait_queue_exclusive(head, &poll->wait);
5015 add_wait_queue(head, &poll->wait);
5018 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5019 struct poll_table_struct *p)
5021 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5022 struct async_poll *apoll = pt->req->apoll;
5024 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5027 static void io_async_task_func(struct callback_head *cb)
5029 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5030 struct async_poll *apoll = req->apoll;
5031 struct io_ring_ctx *ctx = req->ctx;
5033 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5035 if (io_poll_rewait(req, &apoll->poll)) {
5036 spin_unlock_irq(&ctx->completion_lock);
5040 hash_del(&req->hash_node);
5041 io_poll_remove_double(req);
5042 spin_unlock_irq(&ctx->completion_lock);
5044 if (!READ_ONCE(apoll->poll.canceled))
5045 __io_req_task_submit(req);
5047 io_req_complete_failed(req, -ECANCELED);
5050 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5053 struct io_kiocb *req = wait->private;
5054 struct io_poll_iocb *poll = &req->apoll->poll;
5056 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5059 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5062 static void io_poll_req_insert(struct io_kiocb *req)
5064 struct io_ring_ctx *ctx = req->ctx;
5065 struct hlist_head *list;
5067 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5068 hlist_add_head(&req->hash_node, list);
5071 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5072 struct io_poll_iocb *poll,
5073 struct io_poll_table *ipt, __poll_t mask,
5074 wait_queue_func_t wake_func)
5075 __acquires(&ctx->completion_lock)
5077 struct io_ring_ctx *ctx = req->ctx;
5078 bool cancel = false;
5080 INIT_HLIST_NODE(&req->hash_node);
5081 io_init_poll_iocb(poll, mask, wake_func);
5082 poll->file = req->file;
5083 poll->wait.private = req;
5085 ipt->pt._key = mask;
5087 ipt->error = -EINVAL;
5089 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5091 spin_lock_irq(&ctx->completion_lock);
5092 if (likely(poll->head)) {
5093 spin_lock(&poll->head->lock);
5094 if (unlikely(list_empty(&poll->wait.entry))) {
5100 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5101 list_del_init(&poll->wait.entry);
5103 WRITE_ONCE(poll->canceled, true);
5104 else if (!poll->done) /* actually waiting for an event */
5105 io_poll_req_insert(req);
5106 spin_unlock(&poll->head->lock);
5112 static bool io_arm_poll_handler(struct io_kiocb *req)
5114 const struct io_op_def *def = &io_op_defs[req->opcode];
5115 struct io_ring_ctx *ctx = req->ctx;
5116 struct async_poll *apoll;
5117 struct io_poll_table ipt;
5121 if (!req->file || !file_can_poll(req->file))
5123 if (req->flags & REQ_F_POLLED)
5127 else if (def->pollout)
5131 /* if we can't nonblock try, then no point in arming a poll handler */
5132 if (!io_file_supports_async(req, rw))
5135 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5136 if (unlikely(!apoll))
5138 apoll->double_poll = NULL;
5140 req->flags |= REQ_F_POLLED;
5143 mask = EPOLLONESHOT;
5145 mask |= POLLIN | POLLRDNORM;
5147 mask |= POLLOUT | POLLWRNORM;
5149 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5150 if ((req->opcode == IORING_OP_RECVMSG) &&
5151 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5154 mask |= POLLERR | POLLPRI;
5156 ipt.pt._qproc = io_async_queue_proc;
5158 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5160 if (ret || ipt.error) {
5161 io_poll_remove_double(req);
5162 spin_unlock_irq(&ctx->completion_lock);
5165 spin_unlock_irq(&ctx->completion_lock);
5166 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5167 apoll->poll.events);
5171 static bool __io_poll_remove_one(struct io_kiocb *req,
5172 struct io_poll_iocb *poll, bool do_cancel)
5173 __must_hold(&req->ctx->completion_lock)
5175 bool do_complete = false;
5179 spin_lock(&poll->head->lock);
5181 WRITE_ONCE(poll->canceled, true);
5182 if (!list_empty(&poll->wait.entry)) {
5183 list_del_init(&poll->wait.entry);
5186 spin_unlock(&poll->head->lock);
5187 hash_del(&req->hash_node);
5191 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5192 __must_hold(&req->ctx->completion_lock)
5196 io_poll_remove_double(req);
5197 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5199 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5200 /* non-poll requests have submit ref still */
5206 static bool io_poll_remove_one(struct io_kiocb *req)
5207 __must_hold(&req->ctx->completion_lock)
5211 do_complete = io_poll_remove_waitqs(req);
5213 io_cqring_fill_event(req, -ECANCELED, 0);
5214 io_commit_cqring(req->ctx);
5215 req_set_fail_links(req);
5216 io_put_req_deferred(req, 1);
5223 * Returns true if we found and killed one or more poll requests
5225 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5226 struct files_struct *files)
5228 struct hlist_node *tmp;
5229 struct io_kiocb *req;
5232 spin_lock_irq(&ctx->completion_lock);
5233 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5234 struct hlist_head *list;
5236 list = &ctx->cancel_hash[i];
5237 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5238 if (io_match_task(req, tsk, files))
5239 posted += io_poll_remove_one(req);
5242 spin_unlock_irq(&ctx->completion_lock);
5245 io_cqring_ev_posted(ctx);
5250 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5252 __must_hold(&ctx->completion_lock)
5254 struct hlist_head *list;
5255 struct io_kiocb *req;
5257 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5258 hlist_for_each_entry(req, list, hash_node) {
5259 if (sqe_addr != req->user_data)
5261 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5268 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5270 __must_hold(&ctx->completion_lock)
5272 struct io_kiocb *req;
5274 req = io_poll_find(ctx, sqe_addr, poll_only);
5277 if (io_poll_remove_one(req))
5283 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5288 events = READ_ONCE(sqe->poll32_events);
5290 events = swahw32(events);
5292 if (!(flags & IORING_POLL_ADD_MULTI))
5293 events |= EPOLLONESHOT;
5294 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5297 static int io_poll_update_prep(struct io_kiocb *req,
5298 const struct io_uring_sqe *sqe)
5300 struct io_poll_update *upd = &req->poll_update;
5303 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5305 if (sqe->ioprio || sqe->buf_index)
5307 flags = READ_ONCE(sqe->len);
5308 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5309 IORING_POLL_ADD_MULTI))
5311 /* meaningless without update */
5312 if (flags == IORING_POLL_ADD_MULTI)
5315 upd->old_user_data = READ_ONCE(sqe->addr);
5316 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5317 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5319 upd->new_user_data = READ_ONCE(sqe->off);
5320 if (!upd->update_user_data && upd->new_user_data)
5322 if (upd->update_events)
5323 upd->events = io_poll_parse_events(sqe, flags);
5324 else if (sqe->poll32_events)
5330 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5333 struct io_kiocb *req = wait->private;
5334 struct io_poll_iocb *poll = &req->poll;
5336 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5339 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5340 struct poll_table_struct *p)
5342 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5344 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5347 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5349 struct io_poll_iocb *poll = &req->poll;
5352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5354 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5356 flags = READ_ONCE(sqe->len);
5357 if (flags & ~IORING_POLL_ADD_MULTI)
5360 poll->events = io_poll_parse_events(sqe, flags);
5364 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5366 struct io_poll_iocb *poll = &req->poll;
5367 struct io_ring_ctx *ctx = req->ctx;
5368 struct io_poll_table ipt;
5371 ipt.pt._qproc = io_poll_queue_proc;
5373 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5376 if (mask) { /* no async, we'd stolen it */
5378 io_poll_complete(req, mask);
5380 spin_unlock_irq(&ctx->completion_lock);
5383 io_cqring_ev_posted(ctx);
5384 if (poll->events & EPOLLONESHOT)
5390 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5392 struct io_ring_ctx *ctx = req->ctx;
5393 struct io_kiocb *preq;
5397 spin_lock_irq(&ctx->completion_lock);
5398 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5404 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5406 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5411 * Don't allow racy completion with singleshot, as we cannot safely
5412 * update those. For multishot, if we're racing with completion, just
5413 * let completion re-add it.
5415 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5416 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5420 /* we now have a detached poll request. reissue. */
5424 spin_unlock_irq(&ctx->completion_lock);
5425 req_set_fail_links(req);
5426 io_req_complete(req, ret);
5429 /* only mask one event flags, keep behavior flags */
5430 if (req->poll_update.update_events) {
5431 preq->poll.events &= ~0xffff;
5432 preq->poll.events |= req->poll_update.events & 0xffff;
5433 preq->poll.events |= IO_POLL_UNMASK;
5435 if (req->poll_update.update_user_data)
5436 preq->user_data = req->poll_update.new_user_data;
5437 spin_unlock_irq(&ctx->completion_lock);
5439 /* complete update request, we're done with it */
5440 io_req_complete(req, ret);
5443 ret = io_poll_add(preq, issue_flags);
5445 req_set_fail_links(preq);
5446 io_req_complete(preq, ret);
5452 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5454 struct io_timeout_data *data = container_of(timer,
5455 struct io_timeout_data, timer);
5456 struct io_kiocb *req = data->req;
5457 struct io_ring_ctx *ctx = req->ctx;
5458 unsigned long flags;
5460 spin_lock_irqsave(&ctx->completion_lock, flags);
5461 list_del_init(&req->timeout.list);
5462 atomic_set(&req->ctx->cq_timeouts,
5463 atomic_read(&req->ctx->cq_timeouts) + 1);
5465 io_cqring_fill_event(req, -ETIME, 0);
5466 io_commit_cqring(ctx);
5467 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5469 io_cqring_ev_posted(ctx);
5470 req_set_fail_links(req);
5472 return HRTIMER_NORESTART;
5475 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5477 __must_hold(&ctx->completion_lock)
5479 struct io_timeout_data *io;
5480 struct io_kiocb *req;
5483 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5484 found = user_data == req->user_data;
5489 return ERR_PTR(-ENOENT);
5491 io = req->async_data;
5492 if (hrtimer_try_to_cancel(&io->timer) == -1)
5493 return ERR_PTR(-EALREADY);
5494 list_del_init(&req->timeout.list);
5498 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5499 __must_hold(&ctx->completion_lock)
5501 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5504 return PTR_ERR(req);
5506 req_set_fail_links(req);
5507 io_cqring_fill_event(req, -ECANCELED, 0);
5508 io_put_req_deferred(req, 1);
5512 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5513 struct timespec64 *ts, enum hrtimer_mode mode)
5514 __must_hold(&ctx->completion_lock)
5516 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5517 struct io_timeout_data *data;
5520 return PTR_ERR(req);
5522 req->timeout.off = 0; /* noseq */
5523 data = req->async_data;
5524 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5525 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5526 data->timer.function = io_timeout_fn;
5527 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5531 static int io_timeout_remove_prep(struct io_kiocb *req,
5532 const struct io_uring_sqe *sqe)
5534 struct io_timeout_rem *tr = &req->timeout_rem;
5536 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5538 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5540 if (sqe->ioprio || sqe->buf_index || sqe->len)
5543 tr->addr = READ_ONCE(sqe->addr);
5544 tr->flags = READ_ONCE(sqe->timeout_flags);
5545 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5546 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5548 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5550 } else if (tr->flags) {
5551 /* timeout removal doesn't support flags */
5558 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5560 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5565 * Remove or update an existing timeout command
5567 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5569 struct io_timeout_rem *tr = &req->timeout_rem;
5570 struct io_ring_ctx *ctx = req->ctx;
5573 spin_lock_irq(&ctx->completion_lock);
5574 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5575 ret = io_timeout_cancel(ctx, tr->addr);
5577 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5578 io_translate_timeout_mode(tr->flags));
5580 io_cqring_fill_event(req, ret, 0);
5581 io_commit_cqring(ctx);
5582 spin_unlock_irq(&ctx->completion_lock);
5583 io_cqring_ev_posted(ctx);
5585 req_set_fail_links(req);
5590 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5591 bool is_timeout_link)
5593 struct io_timeout_data *data;
5595 u32 off = READ_ONCE(sqe->off);
5597 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5599 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5601 if (off && is_timeout_link)
5603 flags = READ_ONCE(sqe->timeout_flags);
5604 if (flags & ~IORING_TIMEOUT_ABS)
5607 req->timeout.off = off;
5609 if (!req->async_data && io_alloc_async_data(req))
5612 data = req->async_data;
5615 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5618 data->mode = io_translate_timeout_mode(flags);
5619 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5620 if (is_timeout_link)
5621 io_req_track_inflight(req);
5625 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5627 struct io_ring_ctx *ctx = req->ctx;
5628 struct io_timeout_data *data = req->async_data;
5629 struct list_head *entry;
5630 u32 tail, off = req->timeout.off;
5632 spin_lock_irq(&ctx->completion_lock);
5635 * sqe->off holds how many events that need to occur for this
5636 * timeout event to be satisfied. If it isn't set, then this is
5637 * a pure timeout request, sequence isn't used.
5639 if (io_is_timeout_noseq(req)) {
5640 entry = ctx->timeout_list.prev;
5644 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5645 req->timeout.target_seq = tail + off;
5647 /* Update the last seq here in case io_flush_timeouts() hasn't.
5648 * This is safe because ->completion_lock is held, and submissions
5649 * and completions are never mixed in the same ->completion_lock section.
5651 ctx->cq_last_tm_flush = tail;
5654 * Insertion sort, ensuring the first entry in the list is always
5655 * the one we need first.
5657 list_for_each_prev(entry, &ctx->timeout_list) {
5658 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5661 if (io_is_timeout_noseq(nxt))
5663 /* nxt.seq is behind @tail, otherwise would've been completed */
5664 if (off >= nxt->timeout.target_seq - tail)
5668 list_add(&req->timeout.list, entry);
5669 data->timer.function = io_timeout_fn;
5670 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5671 spin_unlock_irq(&ctx->completion_lock);
5675 struct io_cancel_data {
5676 struct io_ring_ctx *ctx;
5680 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5682 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5683 struct io_cancel_data *cd = data;
5685 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5688 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5689 struct io_ring_ctx *ctx)
5691 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5692 enum io_wq_cancel cancel_ret;
5695 if (!tctx || !tctx->io_wq)
5698 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5699 switch (cancel_ret) {
5700 case IO_WQ_CANCEL_OK:
5703 case IO_WQ_CANCEL_RUNNING:
5706 case IO_WQ_CANCEL_NOTFOUND:
5714 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5715 struct io_kiocb *req, __u64 sqe_addr,
5718 unsigned long flags;
5721 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5722 spin_lock_irqsave(&ctx->completion_lock, flags);
5725 ret = io_timeout_cancel(ctx, sqe_addr);
5728 ret = io_poll_cancel(ctx, sqe_addr, false);
5732 io_cqring_fill_event(req, ret, 0);
5733 io_commit_cqring(ctx);
5734 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5735 io_cqring_ev_posted(ctx);
5738 req_set_fail_links(req);
5741 static int io_async_cancel_prep(struct io_kiocb *req,
5742 const struct io_uring_sqe *sqe)
5744 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5746 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5748 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5751 req->cancel.addr = READ_ONCE(sqe->addr);
5755 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5757 struct io_ring_ctx *ctx = req->ctx;
5758 u64 sqe_addr = req->cancel.addr;
5759 struct io_tctx_node *node;
5762 /* tasks should wait for their io-wq threads, so safe w/o sync */
5763 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5764 spin_lock_irq(&ctx->completion_lock);
5767 ret = io_timeout_cancel(ctx, sqe_addr);
5770 ret = io_poll_cancel(ctx, sqe_addr, false);
5773 spin_unlock_irq(&ctx->completion_lock);
5775 /* slow path, try all io-wq's */
5776 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5778 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5779 struct io_uring_task *tctx = node->task->io_uring;
5781 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5785 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5787 spin_lock_irq(&ctx->completion_lock);
5789 io_cqring_fill_event(req, ret, 0);
5790 io_commit_cqring(ctx);
5791 spin_unlock_irq(&ctx->completion_lock);
5792 io_cqring_ev_posted(ctx);
5795 req_set_fail_links(req);
5800 static int io_rsrc_update_prep(struct io_kiocb *req,
5801 const struct io_uring_sqe *sqe)
5803 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5805 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5807 if (sqe->ioprio || sqe->rw_flags)
5810 req->rsrc_update.offset = READ_ONCE(sqe->off);
5811 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5812 if (!req->rsrc_update.nr_args)
5814 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5818 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5820 struct io_ring_ctx *ctx = req->ctx;
5821 struct io_uring_rsrc_update up;
5824 if (issue_flags & IO_URING_F_NONBLOCK)
5827 up.offset = req->rsrc_update.offset;
5828 up.data = req->rsrc_update.arg;
5830 mutex_lock(&ctx->uring_lock);
5831 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5832 mutex_unlock(&ctx->uring_lock);
5835 req_set_fail_links(req);
5836 __io_req_complete(req, issue_flags, ret, 0);
5840 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5842 switch (req->opcode) {
5845 case IORING_OP_READV:
5846 case IORING_OP_READ_FIXED:
5847 case IORING_OP_READ:
5848 return io_read_prep(req, sqe);
5849 case IORING_OP_WRITEV:
5850 case IORING_OP_WRITE_FIXED:
5851 case IORING_OP_WRITE:
5852 return io_write_prep(req, sqe);
5853 case IORING_OP_POLL_ADD:
5854 return io_poll_add_prep(req, sqe);
5855 case IORING_OP_POLL_REMOVE:
5856 return io_poll_update_prep(req, sqe);
5857 case IORING_OP_FSYNC:
5858 return io_fsync_prep(req, sqe);
5859 case IORING_OP_SYNC_FILE_RANGE:
5860 return io_sfr_prep(req, sqe);
5861 case IORING_OP_SENDMSG:
5862 case IORING_OP_SEND:
5863 return io_sendmsg_prep(req, sqe);
5864 case IORING_OP_RECVMSG:
5865 case IORING_OP_RECV:
5866 return io_recvmsg_prep(req, sqe);
5867 case IORING_OP_CONNECT:
5868 return io_connect_prep(req, sqe);
5869 case IORING_OP_TIMEOUT:
5870 return io_timeout_prep(req, sqe, false);
5871 case IORING_OP_TIMEOUT_REMOVE:
5872 return io_timeout_remove_prep(req, sqe);
5873 case IORING_OP_ASYNC_CANCEL:
5874 return io_async_cancel_prep(req, sqe);
5875 case IORING_OP_LINK_TIMEOUT:
5876 return io_timeout_prep(req, sqe, true);
5877 case IORING_OP_ACCEPT:
5878 return io_accept_prep(req, sqe);
5879 case IORING_OP_FALLOCATE:
5880 return io_fallocate_prep(req, sqe);
5881 case IORING_OP_OPENAT:
5882 return io_openat_prep(req, sqe);
5883 case IORING_OP_CLOSE:
5884 return io_close_prep(req, sqe);
5885 case IORING_OP_FILES_UPDATE:
5886 return io_rsrc_update_prep(req, sqe);
5887 case IORING_OP_STATX:
5888 return io_statx_prep(req, sqe);
5889 case IORING_OP_FADVISE:
5890 return io_fadvise_prep(req, sqe);
5891 case IORING_OP_MADVISE:
5892 return io_madvise_prep(req, sqe);
5893 case IORING_OP_OPENAT2:
5894 return io_openat2_prep(req, sqe);
5895 case IORING_OP_EPOLL_CTL:
5896 return io_epoll_ctl_prep(req, sqe);
5897 case IORING_OP_SPLICE:
5898 return io_splice_prep(req, sqe);
5899 case IORING_OP_PROVIDE_BUFFERS:
5900 return io_provide_buffers_prep(req, sqe);
5901 case IORING_OP_REMOVE_BUFFERS:
5902 return io_remove_buffers_prep(req, sqe);
5904 return io_tee_prep(req, sqe);
5905 case IORING_OP_SHUTDOWN:
5906 return io_shutdown_prep(req, sqe);
5907 case IORING_OP_RENAMEAT:
5908 return io_renameat_prep(req, sqe);
5909 case IORING_OP_UNLINKAT:
5910 return io_unlinkat_prep(req, sqe);
5913 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5918 static int io_req_prep_async(struct io_kiocb *req)
5920 if (!io_op_defs[req->opcode].needs_async_setup)
5922 if (WARN_ON_ONCE(req->async_data))
5924 if (io_alloc_async_data(req))
5927 switch (req->opcode) {
5928 case IORING_OP_READV:
5929 return io_rw_prep_async(req, READ);
5930 case IORING_OP_WRITEV:
5931 return io_rw_prep_async(req, WRITE);
5932 case IORING_OP_SENDMSG:
5933 return io_sendmsg_prep_async(req);
5934 case IORING_OP_RECVMSG:
5935 return io_recvmsg_prep_async(req);
5936 case IORING_OP_CONNECT:
5937 return io_connect_prep_async(req);
5939 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5944 static u32 io_get_sequence(struct io_kiocb *req)
5946 struct io_kiocb *pos;
5947 struct io_ring_ctx *ctx = req->ctx;
5948 u32 total_submitted, nr_reqs = 0;
5950 io_for_each_link(pos, req)
5953 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5954 return total_submitted - nr_reqs;
5957 static int io_req_defer(struct io_kiocb *req)
5959 struct io_ring_ctx *ctx = req->ctx;
5960 struct io_defer_entry *de;
5964 /* Still need defer if there is pending req in defer list. */
5965 if (likely(list_empty_careful(&ctx->defer_list) &&
5966 !(req->flags & REQ_F_IO_DRAIN)))
5969 seq = io_get_sequence(req);
5970 /* Still a chance to pass the sequence check */
5971 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5974 ret = io_req_prep_async(req);
5977 io_prep_async_link(req);
5978 de = kmalloc(sizeof(*de), GFP_KERNEL);
5982 spin_lock_irq(&ctx->completion_lock);
5983 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5984 spin_unlock_irq(&ctx->completion_lock);
5986 io_queue_async_work(req);
5987 return -EIOCBQUEUED;
5990 trace_io_uring_defer(ctx, req, req->user_data);
5993 list_add_tail(&de->list, &ctx->defer_list);
5994 spin_unlock_irq(&ctx->completion_lock);
5995 return -EIOCBQUEUED;
5998 static void io_clean_op(struct io_kiocb *req)
6000 if (req->flags & REQ_F_BUFFER_SELECTED) {
6001 switch (req->opcode) {
6002 case IORING_OP_READV:
6003 case IORING_OP_READ_FIXED:
6004 case IORING_OP_READ:
6005 kfree((void *)(unsigned long)req->rw.addr);
6007 case IORING_OP_RECVMSG:
6008 case IORING_OP_RECV:
6009 kfree(req->sr_msg.kbuf);
6012 req->flags &= ~REQ_F_BUFFER_SELECTED;
6015 if (req->flags & REQ_F_NEED_CLEANUP) {
6016 switch (req->opcode) {
6017 case IORING_OP_READV:
6018 case IORING_OP_READ_FIXED:
6019 case IORING_OP_READ:
6020 case IORING_OP_WRITEV:
6021 case IORING_OP_WRITE_FIXED:
6022 case IORING_OP_WRITE: {
6023 struct io_async_rw *io = req->async_data;
6025 kfree(io->free_iovec);
6028 case IORING_OP_RECVMSG:
6029 case IORING_OP_SENDMSG: {
6030 struct io_async_msghdr *io = req->async_data;
6032 kfree(io->free_iov);
6035 case IORING_OP_SPLICE:
6037 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6038 io_put_file(req->splice.file_in);
6040 case IORING_OP_OPENAT:
6041 case IORING_OP_OPENAT2:
6042 if (req->open.filename)
6043 putname(req->open.filename);
6045 case IORING_OP_RENAMEAT:
6046 putname(req->rename.oldpath);
6047 putname(req->rename.newpath);
6049 case IORING_OP_UNLINKAT:
6050 putname(req->unlink.filename);
6053 req->flags &= ~REQ_F_NEED_CLEANUP;
6055 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6056 kfree(req->apoll->double_poll);
6062 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6064 struct io_ring_ctx *ctx = req->ctx;
6065 const struct cred *creds = NULL;
6068 if (req->work.creds && req->work.creds != current_cred())
6069 creds = override_creds(req->work.creds);
6071 switch (req->opcode) {
6073 ret = io_nop(req, issue_flags);
6075 case IORING_OP_READV:
6076 case IORING_OP_READ_FIXED:
6077 case IORING_OP_READ:
6078 ret = io_read(req, issue_flags);
6080 case IORING_OP_WRITEV:
6081 case IORING_OP_WRITE_FIXED:
6082 case IORING_OP_WRITE:
6083 ret = io_write(req, issue_flags);
6085 case IORING_OP_FSYNC:
6086 ret = io_fsync(req, issue_flags);
6088 case IORING_OP_POLL_ADD:
6089 ret = io_poll_add(req, issue_flags);
6091 case IORING_OP_POLL_REMOVE:
6092 ret = io_poll_update(req, issue_flags);
6094 case IORING_OP_SYNC_FILE_RANGE:
6095 ret = io_sync_file_range(req, issue_flags);
6097 case IORING_OP_SENDMSG:
6098 ret = io_sendmsg(req, issue_flags);
6100 case IORING_OP_SEND:
6101 ret = io_send(req, issue_flags);
6103 case IORING_OP_RECVMSG:
6104 ret = io_recvmsg(req, issue_flags);
6106 case IORING_OP_RECV:
6107 ret = io_recv(req, issue_flags);
6109 case IORING_OP_TIMEOUT:
6110 ret = io_timeout(req, issue_flags);
6112 case IORING_OP_TIMEOUT_REMOVE:
6113 ret = io_timeout_remove(req, issue_flags);
6115 case IORING_OP_ACCEPT:
6116 ret = io_accept(req, issue_flags);
6118 case IORING_OP_CONNECT:
6119 ret = io_connect(req, issue_flags);
6121 case IORING_OP_ASYNC_CANCEL:
6122 ret = io_async_cancel(req, issue_flags);
6124 case IORING_OP_FALLOCATE:
6125 ret = io_fallocate(req, issue_flags);
6127 case IORING_OP_OPENAT:
6128 ret = io_openat(req, issue_flags);
6130 case IORING_OP_CLOSE:
6131 ret = io_close(req, issue_flags);
6133 case IORING_OP_FILES_UPDATE:
6134 ret = io_files_update(req, issue_flags);
6136 case IORING_OP_STATX:
6137 ret = io_statx(req, issue_flags);
6139 case IORING_OP_FADVISE:
6140 ret = io_fadvise(req, issue_flags);
6142 case IORING_OP_MADVISE:
6143 ret = io_madvise(req, issue_flags);
6145 case IORING_OP_OPENAT2:
6146 ret = io_openat2(req, issue_flags);
6148 case IORING_OP_EPOLL_CTL:
6149 ret = io_epoll_ctl(req, issue_flags);
6151 case IORING_OP_SPLICE:
6152 ret = io_splice(req, issue_flags);
6154 case IORING_OP_PROVIDE_BUFFERS:
6155 ret = io_provide_buffers(req, issue_flags);
6157 case IORING_OP_REMOVE_BUFFERS:
6158 ret = io_remove_buffers(req, issue_flags);
6161 ret = io_tee(req, issue_flags);
6163 case IORING_OP_SHUTDOWN:
6164 ret = io_shutdown(req, issue_flags);
6166 case IORING_OP_RENAMEAT:
6167 ret = io_renameat(req, issue_flags);
6169 case IORING_OP_UNLINKAT:
6170 ret = io_unlinkat(req, issue_flags);
6178 revert_creds(creds);
6183 /* If the op doesn't have a file, we're not polling for it */
6184 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6185 const bool in_async = io_wq_current_is_worker();
6187 /* workqueue context doesn't hold uring_lock, grab it now */
6189 mutex_lock(&ctx->uring_lock);
6191 io_iopoll_req_issued(req, in_async);
6194 mutex_unlock(&ctx->uring_lock);
6200 static void io_wq_submit_work(struct io_wq_work *work)
6202 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6203 struct io_kiocb *timeout;
6206 timeout = io_prep_linked_timeout(req);
6208 io_queue_linked_timeout(timeout);
6210 if (work->flags & IO_WQ_WORK_CANCEL)
6215 ret = io_issue_sqe(req, 0);
6217 * We can get EAGAIN for polled IO even though we're
6218 * forcing a sync submission from here, since we can't
6219 * wait for request slots on the block side.
6227 /* avoid locking problems by failing it from a clean context */
6229 /* io-wq is going to take one down */
6231 io_req_task_queue_fail(req, ret);
6235 #define FFS_ASYNC_READ 0x1UL
6236 #define FFS_ASYNC_WRITE 0x2UL
6238 #define FFS_ISREG 0x4UL
6240 #define FFS_ISREG 0x0UL
6242 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6244 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6247 struct io_fixed_file *table_l2;
6249 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6250 return &table_l2[i & IORING_FILE_TABLE_MASK];
6253 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6256 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6258 return (struct file *) (slot->file_ptr & FFS_MASK);
6261 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6263 unsigned long file_ptr = (unsigned long) file;
6265 if (__io_file_supports_async(file, READ))
6266 file_ptr |= FFS_ASYNC_READ;
6267 if (__io_file_supports_async(file, WRITE))
6268 file_ptr |= FFS_ASYNC_WRITE;
6269 if (S_ISREG(file_inode(file)->i_mode))
6270 file_ptr |= FFS_ISREG;
6271 file_slot->file_ptr = file_ptr;
6274 static struct file *io_file_get(struct io_submit_state *state,
6275 struct io_kiocb *req, int fd, bool fixed)
6277 struct io_ring_ctx *ctx = req->ctx;
6281 unsigned long file_ptr;
6283 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6285 fd = array_index_nospec(fd, ctx->nr_user_files);
6286 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6287 file = (struct file *) (file_ptr & FFS_MASK);
6288 file_ptr &= ~FFS_MASK;
6289 /* mask in overlapping REQ_F and FFS bits */
6290 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6291 io_req_set_rsrc_node(req);
6293 trace_io_uring_file_get(ctx, fd);
6294 file = __io_file_get(state, fd);
6296 /* we don't allow fixed io_uring files */
6297 if (file && unlikely(file->f_op == &io_uring_fops))
6298 io_req_track_inflight(req);
6304 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6306 struct io_timeout_data *data = container_of(timer,
6307 struct io_timeout_data, timer);
6308 struct io_kiocb *prev, *req = data->req;
6309 struct io_ring_ctx *ctx = req->ctx;
6310 unsigned long flags;
6312 spin_lock_irqsave(&ctx->completion_lock, flags);
6313 prev = req->timeout.head;
6314 req->timeout.head = NULL;
6317 * We don't expect the list to be empty, that will only happen if we
6318 * race with the completion of the linked work.
6320 if (prev && req_ref_inc_not_zero(prev))
6321 io_remove_next_linked(prev);
6324 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6327 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6328 io_put_req_deferred(prev, 1);
6330 io_req_complete_post(req, -ETIME, 0);
6332 io_put_req_deferred(req, 1);
6333 return HRTIMER_NORESTART;
6336 static void io_queue_linked_timeout(struct io_kiocb *req)
6338 struct io_ring_ctx *ctx = req->ctx;
6340 spin_lock_irq(&ctx->completion_lock);
6342 * If the back reference is NULL, then our linked request finished
6343 * before we got a chance to setup the timer
6345 if (req->timeout.head) {
6346 struct io_timeout_data *data = req->async_data;
6348 data->timer.function = io_link_timeout_fn;
6349 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6352 spin_unlock_irq(&ctx->completion_lock);
6353 /* drop submission reference */
6357 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6359 struct io_kiocb *nxt = req->link;
6361 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6362 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6365 nxt->timeout.head = req;
6366 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6367 req->flags |= REQ_F_LINK_TIMEOUT;
6371 static void __io_queue_sqe(struct io_kiocb *req)
6373 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6376 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6379 * We async punt it if the file wasn't marked NOWAIT, or if the file
6380 * doesn't support non-blocking read/write attempts
6383 /* drop submission reference */
6384 if (req->flags & REQ_F_COMPLETE_INLINE) {
6385 struct io_ring_ctx *ctx = req->ctx;
6386 struct io_comp_state *cs = &ctx->submit_state.comp;
6388 cs->reqs[cs->nr++] = req;
6389 if (cs->nr == ARRAY_SIZE(cs->reqs))
6390 io_submit_flush_completions(cs, ctx);
6394 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6395 if (!io_arm_poll_handler(req)) {
6397 * Queued up for async execution, worker will release
6398 * submit reference when the iocb is actually submitted.
6400 io_queue_async_work(req);
6403 io_req_complete_failed(req, ret);
6406 io_queue_linked_timeout(linked_timeout);
6409 static void io_queue_sqe(struct io_kiocb *req)
6413 ret = io_req_defer(req);
6415 if (ret != -EIOCBQUEUED) {
6417 io_req_complete_failed(req, ret);
6419 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6420 ret = io_req_prep_async(req);
6423 io_queue_async_work(req);
6425 __io_queue_sqe(req);
6430 * Check SQE restrictions (opcode and flags).
6432 * Returns 'true' if SQE is allowed, 'false' otherwise.
6434 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6435 struct io_kiocb *req,
6436 unsigned int sqe_flags)
6438 if (!ctx->restricted)
6441 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6444 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6445 ctx->restrictions.sqe_flags_required)
6448 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6449 ctx->restrictions.sqe_flags_required))
6455 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6456 const struct io_uring_sqe *sqe)
6458 struct io_submit_state *state;
6459 unsigned int sqe_flags;
6460 int personality, ret = 0;
6462 req->opcode = READ_ONCE(sqe->opcode);
6463 /* same numerical values with corresponding REQ_F_*, safe to copy */
6464 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6465 req->user_data = READ_ONCE(sqe->user_data);
6466 req->async_data = NULL;
6470 req->fixed_rsrc_refs = NULL;
6471 /* one is dropped after submission, the other at completion */
6472 atomic_set(&req->refs, 2);
6473 req->task = current;
6475 req->work.creds = NULL;
6477 /* enforce forwards compatibility on users */
6478 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6483 if (unlikely(req->opcode >= IORING_OP_LAST))
6486 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6489 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6490 !io_op_defs[req->opcode].buffer_select)
6493 personality = READ_ONCE(sqe->personality);
6495 req->work.creds = xa_load(&ctx->personalities, personality);
6496 if (!req->work.creds)
6498 get_cred(req->work.creds);
6500 state = &ctx->submit_state;
6503 * Plug now if we have more than 1 IO left after this, and the target
6504 * is potentially a read/write to block based storage.
6506 if (!state->plug_started && state->ios_left > 1 &&
6507 io_op_defs[req->opcode].plug) {
6508 blk_start_plug(&state->plug);
6509 state->plug_started = true;
6512 if (io_op_defs[req->opcode].needs_file) {
6513 bool fixed = req->flags & REQ_F_FIXED_FILE;
6515 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6516 if (unlikely(!req->file))
6524 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6525 const struct io_uring_sqe *sqe)
6527 struct io_submit_link *link = &ctx->submit_state.link;
6530 ret = io_init_req(ctx, req, sqe);
6531 if (unlikely(ret)) {
6534 /* fail even hard links since we don't submit */
6535 link->head->flags |= REQ_F_FAIL_LINK;
6536 io_req_complete_failed(link->head, -ECANCELED);
6539 io_req_complete_failed(req, ret);
6542 ret = io_req_prep(req, sqe);
6546 /* don't need @sqe from now on */
6547 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6548 true, ctx->flags & IORING_SETUP_SQPOLL);
6551 * If we already have a head request, queue this one for async
6552 * submittal once the head completes. If we don't have a head but
6553 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6554 * submitted sync once the chain is complete. If none of those
6555 * conditions are true (normal request), then just queue it.
6558 struct io_kiocb *head = link->head;
6561 * Taking sequential execution of a link, draining both sides
6562 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6563 * requests in the link. So, it drains the head and the
6564 * next after the link request. The last one is done via
6565 * drain_next flag to persist the effect across calls.
6567 if (req->flags & REQ_F_IO_DRAIN) {
6568 head->flags |= REQ_F_IO_DRAIN;
6569 ctx->drain_next = 1;
6571 ret = io_req_prep_async(req);
6574 trace_io_uring_link(ctx, req, head);
6575 link->last->link = req;
6578 /* last request of a link, enqueue the link */
6579 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6584 if (unlikely(ctx->drain_next)) {
6585 req->flags |= REQ_F_IO_DRAIN;
6586 ctx->drain_next = 0;
6588 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6600 * Batched submission is done, ensure local IO is flushed out.
6602 static void io_submit_state_end(struct io_submit_state *state,
6603 struct io_ring_ctx *ctx)
6605 if (state->link.head)
6606 io_queue_sqe(state->link.head);
6608 io_submit_flush_completions(&state->comp, ctx);
6609 if (state->plug_started)
6610 blk_finish_plug(&state->plug);
6611 io_state_file_put(state);
6615 * Start submission side cache.
6617 static void io_submit_state_start(struct io_submit_state *state,
6618 unsigned int max_ios)
6620 state->plug_started = false;
6621 state->ios_left = max_ios;
6622 /* set only head, no need to init link_last in advance */
6623 state->link.head = NULL;
6626 static void io_commit_sqring(struct io_ring_ctx *ctx)
6628 struct io_rings *rings = ctx->rings;
6631 * Ensure any loads from the SQEs are done at this point,
6632 * since once we write the new head, the application could
6633 * write new data to them.
6635 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6639 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6640 * that is mapped by userspace. This means that care needs to be taken to
6641 * ensure that reads are stable, as we cannot rely on userspace always
6642 * being a good citizen. If members of the sqe are validated and then later
6643 * used, it's important that those reads are done through READ_ONCE() to
6644 * prevent a re-load down the line.
6646 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6648 u32 *sq_array = ctx->sq_array;
6652 * The cached sq head (or cq tail) serves two purposes:
6654 * 1) allows us to batch the cost of updating the user visible
6656 * 2) allows the kernel side to track the head on its own, even
6657 * though the application is the one updating it.
6659 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6660 if (likely(head < ctx->sq_entries))
6661 return &ctx->sq_sqes[head];
6663 /* drop invalid entries */
6664 ctx->cached_sq_dropped++;
6665 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6669 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6673 /* make sure SQ entry isn't read before tail */
6674 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6676 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6679 percpu_counter_add(¤t->io_uring->inflight, nr);
6680 refcount_add(nr, ¤t->usage);
6681 io_submit_state_start(&ctx->submit_state, nr);
6683 while (submitted < nr) {
6684 const struct io_uring_sqe *sqe;
6685 struct io_kiocb *req;
6687 req = io_alloc_req(ctx);
6688 if (unlikely(!req)) {
6690 submitted = -EAGAIN;
6693 sqe = io_get_sqe(ctx);
6694 if (unlikely(!sqe)) {
6695 kmem_cache_free(req_cachep, req);
6698 /* will complete beyond this point, count as submitted */
6700 if (io_submit_sqe(ctx, req, sqe))
6704 if (unlikely(submitted != nr)) {
6705 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6706 struct io_uring_task *tctx = current->io_uring;
6707 int unused = nr - ref_used;
6709 percpu_ref_put_many(&ctx->refs, unused);
6710 percpu_counter_sub(&tctx->inflight, unused);
6711 put_task_struct_many(current, unused);
6714 io_submit_state_end(&ctx->submit_state, ctx);
6715 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6716 io_commit_sqring(ctx);
6721 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6723 /* Tell userspace we may need a wakeup call */
6724 spin_lock_irq(&ctx->completion_lock);
6725 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6726 spin_unlock_irq(&ctx->completion_lock);
6729 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6731 spin_lock_irq(&ctx->completion_lock);
6732 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6733 spin_unlock_irq(&ctx->completion_lock);
6736 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6738 unsigned int to_submit;
6741 to_submit = io_sqring_entries(ctx);
6742 /* if we're handling multiple rings, cap submit size for fairness */
6743 if (cap_entries && to_submit > 8)
6746 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6747 unsigned nr_events = 0;
6749 mutex_lock(&ctx->uring_lock);
6750 if (!list_empty(&ctx->iopoll_list))
6751 io_do_iopoll(ctx, &nr_events, 0);
6754 * Don't submit if refs are dying, good for io_uring_register(),
6755 * but also it is relied upon by io_ring_exit_work()
6757 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6758 !(ctx->flags & IORING_SETUP_R_DISABLED))
6759 ret = io_submit_sqes(ctx, to_submit);
6760 mutex_unlock(&ctx->uring_lock);
6763 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6764 wake_up(&ctx->sqo_sq_wait);
6769 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6771 struct io_ring_ctx *ctx;
6772 unsigned sq_thread_idle = 0;
6774 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6775 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6776 sqd->sq_thread_idle = sq_thread_idle;
6779 static int io_sq_thread(void *data)
6781 struct io_sq_data *sqd = data;
6782 struct io_ring_ctx *ctx;
6783 unsigned long timeout = 0;
6784 char buf[TASK_COMM_LEN];
6787 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6788 set_task_comm(current, buf);
6789 current->pf_io_worker = NULL;
6791 if (sqd->sq_cpu != -1)
6792 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6794 set_cpus_allowed_ptr(current, cpu_online_mask);
6795 current->flags |= PF_NO_SETAFFINITY;
6797 mutex_lock(&sqd->lock);
6798 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6800 bool cap_entries, sqt_spin, needs_sched;
6802 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6803 signal_pending(current)) {
6804 bool did_sig = false;
6806 mutex_unlock(&sqd->lock);
6807 if (signal_pending(current)) {
6808 struct ksignal ksig;
6810 did_sig = get_signal(&ksig);
6813 mutex_lock(&sqd->lock);
6817 io_run_task_work_head(&sqd->park_task_work);
6818 timeout = jiffies + sqd->sq_thread_idle;
6822 cap_entries = !list_is_singular(&sqd->ctx_list);
6823 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6824 const struct cred *creds = NULL;
6826 if (ctx->sq_creds != current_cred())
6827 creds = override_creds(ctx->sq_creds);
6828 ret = __io_sq_thread(ctx, cap_entries);
6830 revert_creds(creds);
6831 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6835 if (sqt_spin || !time_after(jiffies, timeout)) {
6839 timeout = jiffies + sqd->sq_thread_idle;
6844 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6845 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6846 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6847 !list_empty_careful(&ctx->iopoll_list)) {
6848 needs_sched = false;
6851 if (io_sqring_entries(ctx)) {
6852 needs_sched = false;
6857 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6858 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6859 io_ring_set_wakeup_flag(ctx);
6861 mutex_unlock(&sqd->lock);
6863 mutex_lock(&sqd->lock);
6864 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6865 io_ring_clear_wakeup_flag(ctx);
6868 finish_wait(&sqd->wait, &wait);
6869 io_run_task_work_head(&sqd->park_task_work);
6870 timeout = jiffies + sqd->sq_thread_idle;
6873 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6874 io_uring_cancel_sqpoll(ctx);
6876 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6877 io_ring_set_wakeup_flag(ctx);
6878 mutex_unlock(&sqd->lock);
6881 io_run_task_work_head(&sqd->park_task_work);
6882 complete(&sqd->exited);
6886 struct io_wait_queue {
6887 struct wait_queue_entry wq;
6888 struct io_ring_ctx *ctx;
6890 unsigned nr_timeouts;
6893 static inline bool io_should_wake(struct io_wait_queue *iowq)
6895 struct io_ring_ctx *ctx = iowq->ctx;
6898 * Wake up if we have enough events, or if a timeout occurred since we
6899 * started waiting. For timeouts, we always want to return to userspace,
6900 * regardless of event count.
6902 return io_cqring_events(ctx) >= iowq->to_wait ||
6903 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6906 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6907 int wake_flags, void *key)
6909 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6913 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6914 * the task, and the next invocation will do it.
6916 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6917 return autoremove_wake_function(curr, mode, wake_flags, key);
6921 static int io_run_task_work_sig(void)
6923 if (io_run_task_work())
6925 if (!signal_pending(current))
6927 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6928 return -ERESTARTSYS;
6932 /* when returns >0, the caller should retry */
6933 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6934 struct io_wait_queue *iowq,
6935 signed long *timeout)
6939 /* make sure we run task_work before checking for signals */
6940 ret = io_run_task_work_sig();
6941 if (ret || io_should_wake(iowq))
6943 /* let the caller flush overflows, retry */
6944 if (test_bit(0, &ctx->cq_check_overflow))
6947 *timeout = schedule_timeout(*timeout);
6948 return !*timeout ? -ETIME : 1;
6952 * Wait until events become available, if we don't already have some. The
6953 * application must reap them itself, as they reside on the shared cq ring.
6955 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6956 const sigset_t __user *sig, size_t sigsz,
6957 struct __kernel_timespec __user *uts)
6959 struct io_wait_queue iowq = {
6962 .func = io_wake_function,
6963 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6966 .to_wait = min_events,
6968 struct io_rings *rings = ctx->rings;
6969 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6973 io_cqring_overflow_flush(ctx, false);
6974 if (io_cqring_events(ctx) >= min_events)
6976 if (!io_run_task_work())
6981 #ifdef CONFIG_COMPAT
6982 if (in_compat_syscall())
6983 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6987 ret = set_user_sigmask(sig, sigsz);
6994 struct timespec64 ts;
6996 if (get_timespec64(&ts, uts))
6998 timeout = timespec64_to_jiffies(&ts);
7001 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7002 trace_io_uring_cqring_wait(ctx, min_events);
7004 /* if we can't even flush overflow, don't wait for more */
7005 if (!io_cqring_overflow_flush(ctx, false)) {
7009 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7010 TASK_INTERRUPTIBLE);
7011 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7012 finish_wait(&ctx->wait, &iowq.wq);
7016 restore_saved_sigmask_unless(ret == -EINTR);
7018 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7021 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7023 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7025 for (i = 0; i < nr_tables; i++)
7026 kfree(table->files[i]);
7027 kfree(table->files);
7028 table->files = NULL;
7031 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7033 #if defined(CONFIG_UNIX)
7034 if (ctx->ring_sock) {
7035 struct sock *sock = ctx->ring_sock->sk;
7036 struct sk_buff *skb;
7038 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7044 for (i = 0; i < ctx->nr_user_files; i++) {
7047 file = io_file_from_index(ctx, i);
7052 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7053 kfree(ctx->file_data);
7054 ctx->file_data = NULL;
7055 ctx->nr_user_files = 0;
7058 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7060 spin_lock_bh(&ctx->rsrc_ref_lock);
7063 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7065 spin_unlock_bh(&ctx->rsrc_ref_lock);
7068 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7070 percpu_ref_exit(&ref_node->refs);
7074 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7075 struct io_rsrc_data *data_to_kill)
7077 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7078 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7081 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7083 rsrc_node->rsrc_data = data_to_kill;
7084 io_rsrc_ref_lock(ctx);
7085 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7086 io_rsrc_ref_unlock(ctx);
7088 atomic_inc(&data_to_kill->refs);
7089 percpu_ref_kill(&rsrc_node->refs);
7090 ctx->rsrc_node = NULL;
7093 if (!ctx->rsrc_node) {
7094 ctx->rsrc_node = ctx->rsrc_backup_node;
7095 ctx->rsrc_backup_node = NULL;
7099 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7101 if (ctx->rsrc_backup_node)
7103 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7104 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7107 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7111 /* As we may drop ->uring_lock, other task may have started quiesce */
7115 data->quiesce = true;
7117 ret = io_rsrc_node_switch_start(ctx);
7120 io_rsrc_node_switch(ctx, data);
7122 /* kill initial ref, already quiesced if zero */
7123 if (atomic_dec_and_test(&data->refs))
7125 flush_delayed_work(&ctx->rsrc_put_work);
7126 ret = wait_for_completion_interruptible(&data->done);
7130 atomic_inc(&data->refs);
7131 /* wait for all works potentially completing data->done */
7132 flush_delayed_work(&ctx->rsrc_put_work);
7133 reinit_completion(&data->done);
7135 mutex_unlock(&ctx->uring_lock);
7136 ret = io_run_task_work_sig();
7137 mutex_lock(&ctx->uring_lock);
7139 data->quiesce = false;
7144 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7145 rsrc_put_fn *do_put)
7147 struct io_rsrc_data *data;
7149 data = kzalloc(sizeof(*data), GFP_KERNEL);
7153 atomic_set(&data->refs, 1);
7155 data->do_put = do_put;
7156 init_completion(&data->done);
7160 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7164 if (!ctx->file_data)
7166 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7168 __io_sqe_files_unregister(ctx);
7172 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7173 __releases(&sqd->lock)
7175 WARN_ON_ONCE(sqd->thread == current);
7178 * Do the dance but not conditional clear_bit() because it'd race with
7179 * other threads incrementing park_pending and setting the bit.
7181 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7182 if (atomic_dec_return(&sqd->park_pending))
7183 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7184 mutex_unlock(&sqd->lock);
7187 static void io_sq_thread_park(struct io_sq_data *sqd)
7188 __acquires(&sqd->lock)
7190 WARN_ON_ONCE(sqd->thread == current);
7192 atomic_inc(&sqd->park_pending);
7193 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7194 mutex_lock(&sqd->lock);
7196 wake_up_process(sqd->thread);
7199 static void io_sq_thread_stop(struct io_sq_data *sqd)
7201 WARN_ON_ONCE(sqd->thread == current);
7202 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7204 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7205 mutex_lock(&sqd->lock);
7207 wake_up_process(sqd->thread);
7208 mutex_unlock(&sqd->lock);
7209 wait_for_completion(&sqd->exited);
7212 static void io_put_sq_data(struct io_sq_data *sqd)
7214 if (refcount_dec_and_test(&sqd->refs)) {
7215 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7217 io_sq_thread_stop(sqd);
7222 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7224 struct io_sq_data *sqd = ctx->sq_data;
7227 io_sq_thread_park(sqd);
7228 list_del_init(&ctx->sqd_list);
7229 io_sqd_update_thread_idle(sqd);
7230 io_sq_thread_unpark(sqd);
7232 io_put_sq_data(sqd);
7233 ctx->sq_data = NULL;
7235 put_cred(ctx->sq_creds);
7239 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7241 struct io_ring_ctx *ctx_attach;
7242 struct io_sq_data *sqd;
7245 f = fdget(p->wq_fd);
7247 return ERR_PTR(-ENXIO);
7248 if (f.file->f_op != &io_uring_fops) {
7250 return ERR_PTR(-EINVAL);
7253 ctx_attach = f.file->private_data;
7254 sqd = ctx_attach->sq_data;
7257 return ERR_PTR(-EINVAL);
7259 if (sqd->task_tgid != current->tgid) {
7261 return ERR_PTR(-EPERM);
7264 refcount_inc(&sqd->refs);
7269 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7272 struct io_sq_data *sqd;
7275 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7276 sqd = io_attach_sq_data(p);
7281 /* fall through for EPERM case, setup new sqd/task */
7282 if (PTR_ERR(sqd) != -EPERM)
7286 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7288 return ERR_PTR(-ENOMEM);
7290 atomic_set(&sqd->park_pending, 0);
7291 refcount_set(&sqd->refs, 1);
7292 INIT_LIST_HEAD(&sqd->ctx_list);
7293 mutex_init(&sqd->lock);
7294 init_waitqueue_head(&sqd->wait);
7295 init_completion(&sqd->exited);
7299 #if defined(CONFIG_UNIX)
7301 * Ensure the UNIX gc is aware of our file set, so we are certain that
7302 * the io_uring can be safely unregistered on process exit, even if we have
7303 * loops in the file referencing.
7305 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7307 struct sock *sk = ctx->ring_sock->sk;
7308 struct scm_fp_list *fpl;
7309 struct sk_buff *skb;
7312 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7316 skb = alloc_skb(0, GFP_KERNEL);
7325 fpl->user = get_uid(current_user());
7326 for (i = 0; i < nr; i++) {
7327 struct file *file = io_file_from_index(ctx, i + offset);
7331 fpl->fp[nr_files] = get_file(file);
7332 unix_inflight(fpl->user, fpl->fp[nr_files]);
7337 fpl->max = SCM_MAX_FD;
7338 fpl->count = nr_files;
7339 UNIXCB(skb).fp = fpl;
7340 skb->destructor = unix_destruct_scm;
7341 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7342 skb_queue_head(&sk->sk_receive_queue, skb);
7344 for (i = 0; i < nr_files; i++)
7355 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7356 * causes regular reference counting to break down. We rely on the UNIX
7357 * garbage collection to take care of this problem for us.
7359 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7361 unsigned left, total;
7365 left = ctx->nr_user_files;
7367 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7369 ret = __io_sqe_files_scm(ctx, this_files, total);
7373 total += this_files;
7379 while (total < ctx->nr_user_files) {
7380 struct file *file = io_file_from_index(ctx, total);
7390 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7396 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7398 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7400 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7404 for (i = 0; i < nr_tables; i++) {
7405 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7407 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7409 if (!table->files[i])
7411 nr_files -= this_files;
7417 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7421 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7423 struct file *file = prsrc->file;
7424 #if defined(CONFIG_UNIX)
7425 struct sock *sock = ctx->ring_sock->sk;
7426 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7427 struct sk_buff *skb;
7430 __skb_queue_head_init(&list);
7433 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7434 * remove this entry and rearrange the file array.
7436 skb = skb_dequeue(head);
7438 struct scm_fp_list *fp;
7440 fp = UNIXCB(skb).fp;
7441 for (i = 0; i < fp->count; i++) {
7444 if (fp->fp[i] != file)
7447 unix_notinflight(fp->user, fp->fp[i]);
7448 left = fp->count - 1 - i;
7450 memmove(&fp->fp[i], &fp->fp[i + 1],
7451 left * sizeof(struct file *));
7458 __skb_queue_tail(&list, skb);
7468 __skb_queue_tail(&list, skb);
7470 skb = skb_dequeue(head);
7473 if (skb_peek(&list)) {
7474 spin_lock_irq(&head->lock);
7475 while ((skb = __skb_dequeue(&list)) != NULL)
7476 __skb_queue_tail(head, skb);
7477 spin_unlock_irq(&head->lock);
7484 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7486 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7487 struct io_ring_ctx *ctx = rsrc_data->ctx;
7488 struct io_rsrc_put *prsrc, *tmp;
7490 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7491 list_del(&prsrc->list);
7492 rsrc_data->do_put(ctx, prsrc);
7496 io_rsrc_node_destroy(ref_node);
7497 if (atomic_dec_and_test(&rsrc_data->refs))
7498 complete(&rsrc_data->done);
7501 static void io_rsrc_put_work(struct work_struct *work)
7503 struct io_ring_ctx *ctx;
7504 struct llist_node *node;
7506 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7507 node = llist_del_all(&ctx->rsrc_put_llist);
7510 struct io_rsrc_node *ref_node;
7511 struct llist_node *next = node->next;
7513 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7514 __io_rsrc_put_work(ref_node);
7519 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7521 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7522 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7523 bool first_add = false;
7525 io_rsrc_ref_lock(ctx);
7528 while (!list_empty(&ctx->rsrc_ref_list)) {
7529 node = list_first_entry(&ctx->rsrc_ref_list,
7530 struct io_rsrc_node, node);
7531 /* recycle ref nodes in order */
7534 list_del(&node->node);
7535 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7537 io_rsrc_ref_unlock(ctx);
7540 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7543 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7545 struct io_rsrc_node *ref_node;
7547 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7551 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7556 INIT_LIST_HEAD(&ref_node->node);
7557 INIT_LIST_HEAD(&ref_node->rsrc_list);
7558 ref_node->done = false;
7562 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7565 __s32 __user *fds = (__s32 __user *) arg;
7569 struct io_rsrc_data *file_data;
7575 if (nr_args > IORING_MAX_FIXED_FILES)
7577 ret = io_rsrc_node_switch_start(ctx);
7581 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put);
7584 ctx->file_data = file_data;
7586 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7589 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7590 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7594 /* allow sparse sets */
7604 * Don't allow io_uring instances to be registered. If UNIX
7605 * isn't enabled, then this causes a reference cycle and this
7606 * instance can never get freed. If UNIX is enabled we'll
7607 * handle it just fine, but there's still no point in allowing
7608 * a ring fd as it doesn't support regular read/write anyway.
7610 if (file->f_op == &io_uring_fops) {
7614 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7617 ret = io_sqe_files_scm(ctx);
7619 __io_sqe_files_unregister(ctx);
7623 io_rsrc_node_switch(ctx, NULL);
7626 for (i = 0; i < ctx->nr_user_files; i++) {
7627 file = io_file_from_index(ctx, i);
7631 io_free_file_tables(&ctx->file_table, nr_args);
7632 ctx->nr_user_files = 0;
7634 kfree(ctx->file_data);
7635 ctx->file_data = NULL;
7639 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7642 #if defined(CONFIG_UNIX)
7643 struct sock *sock = ctx->ring_sock->sk;
7644 struct sk_buff_head *head = &sock->sk_receive_queue;
7645 struct sk_buff *skb;
7648 * See if we can merge this file into an existing skb SCM_RIGHTS
7649 * file set. If there's no room, fall back to allocating a new skb
7650 * and filling it in.
7652 spin_lock_irq(&head->lock);
7653 skb = skb_peek(head);
7655 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7657 if (fpl->count < SCM_MAX_FD) {
7658 __skb_unlink(skb, head);
7659 spin_unlock_irq(&head->lock);
7660 fpl->fp[fpl->count] = get_file(file);
7661 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7663 spin_lock_irq(&head->lock);
7664 __skb_queue_head(head, skb);
7669 spin_unlock_irq(&head->lock);
7676 return __io_sqe_files_scm(ctx, 1, index);
7682 static int io_queue_rsrc_removal(struct io_rsrc_data *data,
7683 struct io_rsrc_node *node, void *rsrc)
7685 struct io_rsrc_put *prsrc;
7687 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7692 list_add(&prsrc->list, &node->rsrc_list);
7696 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7697 struct io_uring_rsrc_update *up,
7700 struct io_rsrc_data *data = ctx->file_data;
7701 struct io_fixed_file *file_slot;
7706 bool needs_switch = false;
7708 if (check_add_overflow(up->offset, nr_args, &done))
7710 if (done > ctx->nr_user_files)
7712 err = io_rsrc_node_switch_start(ctx);
7716 fds = u64_to_user_ptr(up->data);
7717 for (done = 0; done < nr_args; done++) {
7719 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7723 if (fd == IORING_REGISTER_FILES_SKIP)
7726 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7727 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7729 if (file_slot->file_ptr) {
7730 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7731 err = io_queue_rsrc_removal(data, ctx->rsrc_node, file);
7734 file_slot->file_ptr = 0;
7735 needs_switch = true;
7744 * Don't allow io_uring instances to be registered. If
7745 * UNIX isn't enabled, then this causes a reference
7746 * cycle and this instance can never get freed. If UNIX
7747 * is enabled we'll handle it just fine, but there's
7748 * still no point in allowing a ring fd as it doesn't
7749 * support regular read/write anyway.
7751 if (file->f_op == &io_uring_fops) {
7756 io_fixed_file_set(file_slot, file);
7757 err = io_sqe_file_register(ctx, file, i);
7759 file_slot->file_ptr = 0;
7767 io_rsrc_node_switch(ctx, data);
7768 return done ? done : err;
7771 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7774 struct io_uring_rsrc_update up;
7776 if (!ctx->file_data)
7780 if (copy_from_user(&up, arg, sizeof(up)))
7785 return __io_sqe_files_update(ctx, &up, nr_args);
7788 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7790 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7792 req = io_put_req_find_next(req);
7793 return req ? &req->work : NULL;
7796 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7797 struct task_struct *task)
7799 struct io_wq_hash *hash;
7800 struct io_wq_data data;
7801 unsigned int concurrency;
7803 hash = ctx->hash_map;
7805 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7807 return ERR_PTR(-ENOMEM);
7808 refcount_set(&hash->refs, 1);
7809 init_waitqueue_head(&hash->wait);
7810 ctx->hash_map = hash;
7815 data.free_work = io_free_work;
7816 data.do_work = io_wq_submit_work;
7818 /* Do QD, or 4 * CPUS, whatever is smallest */
7819 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7821 return io_wq_create(concurrency, &data);
7824 static int io_uring_alloc_task_context(struct task_struct *task,
7825 struct io_ring_ctx *ctx)
7827 struct io_uring_task *tctx;
7830 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7831 if (unlikely(!tctx))
7834 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7835 if (unlikely(ret)) {
7840 tctx->io_wq = io_init_wq_offload(ctx, task);
7841 if (IS_ERR(tctx->io_wq)) {
7842 ret = PTR_ERR(tctx->io_wq);
7843 percpu_counter_destroy(&tctx->inflight);
7849 init_waitqueue_head(&tctx->wait);
7851 atomic_set(&tctx->in_idle, 0);
7852 atomic_set(&tctx->inflight_tracked, 0);
7853 task->io_uring = tctx;
7854 spin_lock_init(&tctx->task_lock);
7855 INIT_WQ_LIST(&tctx->task_list);
7856 tctx->task_state = 0;
7857 init_task_work(&tctx->task_work, tctx_task_work);
7861 void __io_uring_free(struct task_struct *tsk)
7863 struct io_uring_task *tctx = tsk->io_uring;
7865 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7866 WARN_ON_ONCE(tctx->io_wq);
7868 percpu_counter_destroy(&tctx->inflight);
7870 tsk->io_uring = NULL;
7873 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7874 struct io_uring_params *p)
7878 /* Retain compatibility with failing for an invalid attach attempt */
7879 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7880 IORING_SETUP_ATTACH_WQ) {
7883 f = fdget(p->wq_fd);
7886 if (f.file->f_op != &io_uring_fops) {
7892 if (ctx->flags & IORING_SETUP_SQPOLL) {
7893 struct task_struct *tsk;
7894 struct io_sq_data *sqd;
7897 sqd = io_get_sq_data(p, &attached);
7903 ctx->sq_creds = get_current_cred();
7905 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7906 if (!ctx->sq_thread_idle)
7907 ctx->sq_thread_idle = HZ;
7910 io_sq_thread_park(sqd);
7911 list_add(&ctx->sqd_list, &sqd->ctx_list);
7912 io_sqd_update_thread_idle(sqd);
7913 /* don't attach to a dying SQPOLL thread, would be racy */
7914 if (attached && !sqd->thread)
7916 io_sq_thread_unpark(sqd);
7923 if (p->flags & IORING_SETUP_SQ_AFF) {
7924 int cpu = p->sq_thread_cpu;
7927 if (cpu >= nr_cpu_ids)
7929 if (!cpu_online(cpu))
7937 sqd->task_pid = current->pid;
7938 sqd->task_tgid = current->tgid;
7939 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7946 ret = io_uring_alloc_task_context(tsk, ctx);
7947 wake_up_new_task(tsk);
7950 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7951 /* Can't have SQ_AFF without SQPOLL */
7958 io_sq_thread_finish(ctx);
7961 complete(&ctx->sq_data->exited);
7965 static inline void __io_unaccount_mem(struct user_struct *user,
7966 unsigned long nr_pages)
7968 atomic_long_sub(nr_pages, &user->locked_vm);
7971 static inline int __io_account_mem(struct user_struct *user,
7972 unsigned long nr_pages)
7974 unsigned long page_limit, cur_pages, new_pages;
7976 /* Don't allow more pages than we can safely lock */
7977 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7980 cur_pages = atomic_long_read(&user->locked_vm);
7981 new_pages = cur_pages + nr_pages;
7982 if (new_pages > page_limit)
7984 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7985 new_pages) != cur_pages);
7990 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7993 __io_unaccount_mem(ctx->user, nr_pages);
7995 if (ctx->mm_account)
7996 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7999 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8004 ret = __io_account_mem(ctx->user, nr_pages);
8009 if (ctx->mm_account)
8010 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8015 static void io_mem_free(void *ptr)
8022 page = virt_to_head_page(ptr);
8023 if (put_page_testzero(page))
8024 free_compound_page(page);
8027 static void *io_mem_alloc(size_t size)
8029 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8030 __GFP_NORETRY | __GFP_ACCOUNT;
8032 return (void *) __get_free_pages(gfp_flags, get_order(size));
8035 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8038 struct io_rings *rings;
8039 size_t off, sq_array_size;
8041 off = struct_size(rings, cqes, cq_entries);
8042 if (off == SIZE_MAX)
8046 off = ALIGN(off, SMP_CACHE_BYTES);
8054 sq_array_size = array_size(sizeof(u32), sq_entries);
8055 if (sq_array_size == SIZE_MAX)
8058 if (check_add_overflow(off, sq_array_size, &off))
8064 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf *imu)
8068 for (i = 0; i < imu->nr_bvecs; i++)
8069 unpin_user_page(imu->bvec[i].bv_page);
8070 if (imu->acct_pages)
8071 io_unaccount_mem(ctx, imu->acct_pages);
8076 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8080 if (!ctx->user_bufs)
8083 for (i = 0; i < ctx->nr_user_bufs; i++)
8084 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8085 kfree(ctx->user_bufs);
8086 ctx->user_bufs = NULL;
8087 ctx->nr_user_bufs = 0;
8091 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8092 void __user *arg, unsigned index)
8094 struct iovec __user *src;
8096 #ifdef CONFIG_COMPAT
8098 struct compat_iovec __user *ciovs;
8099 struct compat_iovec ciov;
8101 ciovs = (struct compat_iovec __user *) arg;
8102 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8105 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8106 dst->iov_len = ciov.iov_len;
8110 src = (struct iovec __user *) arg;
8111 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8117 * Not super efficient, but this is just a registration time. And we do cache
8118 * the last compound head, so generally we'll only do a full search if we don't
8121 * We check if the given compound head page has already been accounted, to
8122 * avoid double accounting it. This allows us to account the full size of the
8123 * page, not just the constituent pages of a huge page.
8125 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8126 int nr_pages, struct page *hpage)
8130 /* check current page array */
8131 for (i = 0; i < nr_pages; i++) {
8132 if (!PageCompound(pages[i]))
8134 if (compound_head(pages[i]) == hpage)
8138 /* check previously registered pages */
8139 for (i = 0; i < ctx->nr_user_bufs; i++) {
8140 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8142 for (j = 0; j < imu->nr_bvecs; j++) {
8143 if (!PageCompound(imu->bvec[j].bv_page))
8145 if (compound_head(imu->bvec[j].bv_page) == hpage)
8153 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8154 int nr_pages, struct io_mapped_ubuf *imu,
8155 struct page **last_hpage)
8159 for (i = 0; i < nr_pages; i++) {
8160 if (!PageCompound(pages[i])) {
8165 hpage = compound_head(pages[i]);
8166 if (hpage == *last_hpage)
8168 *last_hpage = hpage;
8169 if (headpage_already_acct(ctx, pages, i, hpage))
8171 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8175 if (!imu->acct_pages)
8178 ret = io_account_mem(ctx, imu->acct_pages);
8180 imu->acct_pages = 0;
8184 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8185 struct io_mapped_ubuf *imu,
8186 struct page **last_hpage)
8188 struct vm_area_struct **vmas = NULL;
8189 struct page **pages = NULL;
8190 unsigned long off, start, end, ubuf;
8192 int ret, pret, nr_pages, i;
8194 ubuf = (unsigned long) iov->iov_base;
8195 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8196 start = ubuf >> PAGE_SHIFT;
8197 nr_pages = end - start;
8201 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8205 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8210 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8216 mmap_read_lock(current->mm);
8217 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8219 if (pret == nr_pages) {
8220 /* don't support file backed memory */
8221 for (i = 0; i < nr_pages; i++) {
8222 struct vm_area_struct *vma = vmas[i];
8225 !is_file_hugepages(vma->vm_file)) {
8231 ret = pret < 0 ? pret : -EFAULT;
8233 mmap_read_unlock(current->mm);
8236 * if we did partial map, or found file backed vmas,
8237 * release any pages we did get
8240 unpin_user_pages(pages, pret);
8245 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8247 unpin_user_pages(pages, pret);
8252 off = ubuf & ~PAGE_MASK;
8253 size = iov->iov_len;
8254 for (i = 0; i < nr_pages; i++) {
8257 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8258 imu->bvec[i].bv_page = pages[i];
8259 imu->bvec[i].bv_len = vec_len;
8260 imu->bvec[i].bv_offset = off;
8264 /* store original address for later verification */
8266 imu->ubuf_end = ubuf + iov->iov_len;
8267 imu->nr_bvecs = nr_pages;
8275 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8277 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8278 return ctx->user_bufs ? 0 : -ENOMEM;
8281 static int io_buffer_validate(struct iovec *iov)
8283 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8286 * Don't impose further limits on the size and buffer
8287 * constraints here, we'll -EINVAL later when IO is
8288 * submitted if they are wrong.
8290 if (!iov->iov_base || !iov->iov_len)
8293 /* arbitrary limit, but we need something */
8294 if (iov->iov_len > SZ_1G)
8297 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8303 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8304 unsigned int nr_args)
8308 struct page *last_hpage = NULL;
8312 if (!nr_args || nr_args > UIO_MAXIOV)
8314 ret = io_buffers_map_alloc(ctx, nr_args);
8318 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8319 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8321 ret = io_copy_iov(ctx, &iov, arg, i);
8324 ret = io_buffer_validate(&iov);
8327 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8333 io_sqe_buffers_unregister(ctx);
8338 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8340 __s32 __user *fds = arg;
8346 if (copy_from_user(&fd, fds, sizeof(*fds)))
8349 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8350 if (IS_ERR(ctx->cq_ev_fd)) {
8351 int ret = PTR_ERR(ctx->cq_ev_fd);
8352 ctx->cq_ev_fd = NULL;
8359 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8361 if (ctx->cq_ev_fd) {
8362 eventfd_ctx_put(ctx->cq_ev_fd);
8363 ctx->cq_ev_fd = NULL;
8370 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8372 struct io_buffer *buf;
8373 unsigned long index;
8375 xa_for_each(&ctx->io_buffers, index, buf)
8376 __io_remove_buffers(ctx, buf, index, -1U);
8379 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8381 struct io_kiocb *req, *nxt;
8383 list_for_each_entry_safe(req, nxt, list, compl.list) {
8384 if (tsk && req->task != tsk)
8386 list_del(&req->compl.list);
8387 kmem_cache_free(req_cachep, req);
8391 static void io_req_caches_free(struct io_ring_ctx *ctx)
8393 struct io_submit_state *submit_state = &ctx->submit_state;
8394 struct io_comp_state *cs = &ctx->submit_state.comp;
8396 mutex_lock(&ctx->uring_lock);
8398 if (submit_state->free_reqs) {
8399 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8400 submit_state->reqs);
8401 submit_state->free_reqs = 0;
8404 io_flush_cached_locked_reqs(ctx, cs);
8405 io_req_cache_free(&cs->free_list, NULL);
8406 mutex_unlock(&ctx->uring_lock);
8409 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8411 io_sq_thread_finish(ctx);
8412 io_sqe_buffers_unregister(ctx);
8414 if (ctx->mm_account) {
8415 mmdrop(ctx->mm_account);
8416 ctx->mm_account = NULL;
8419 mutex_lock(&ctx->uring_lock);
8420 if (ctx->file_data) {
8421 if (!atomic_dec_and_test(&ctx->file_data->refs))
8422 wait_for_completion(&ctx->file_data->done);
8423 __io_sqe_files_unregister(ctx);
8426 __io_cqring_overflow_flush(ctx, true);
8427 mutex_unlock(&ctx->uring_lock);
8428 io_eventfd_unregister(ctx);
8429 io_destroy_buffers(ctx);
8431 /* there are no registered resources left, nobody uses it */
8433 io_rsrc_node_destroy(ctx->rsrc_node);
8434 if (ctx->rsrc_backup_node)
8435 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8436 flush_delayed_work(&ctx->rsrc_put_work);
8438 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8439 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8441 #if defined(CONFIG_UNIX)
8442 if (ctx->ring_sock) {
8443 ctx->ring_sock->file = NULL; /* so that iput() is called */
8444 sock_release(ctx->ring_sock);
8448 io_mem_free(ctx->rings);
8449 io_mem_free(ctx->sq_sqes);
8451 percpu_ref_exit(&ctx->refs);
8452 free_uid(ctx->user);
8453 io_req_caches_free(ctx);
8455 io_wq_put_hash(ctx->hash_map);
8456 kfree(ctx->cancel_hash);
8460 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8462 struct io_ring_ctx *ctx = file->private_data;
8465 poll_wait(file, &ctx->cq_wait, wait);
8467 * synchronizes with barrier from wq_has_sleeper call in
8471 if (!io_sqring_full(ctx))
8472 mask |= EPOLLOUT | EPOLLWRNORM;
8475 * Don't flush cqring overflow list here, just do a simple check.
8476 * Otherwise there could possible be ABBA deadlock:
8479 * lock(&ctx->uring_lock);
8481 * lock(&ctx->uring_lock);
8484 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8485 * pushs them to do the flush.
8487 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8488 mask |= EPOLLIN | EPOLLRDNORM;
8493 static int io_uring_fasync(int fd, struct file *file, int on)
8495 struct io_ring_ctx *ctx = file->private_data;
8497 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8500 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8502 const struct cred *creds;
8504 creds = xa_erase(&ctx->personalities, id);
8513 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8515 return io_run_task_work_head(&ctx->exit_task_work);
8518 struct io_tctx_exit {
8519 struct callback_head task_work;
8520 struct completion completion;
8521 struct io_ring_ctx *ctx;
8524 static void io_tctx_exit_cb(struct callback_head *cb)
8526 struct io_uring_task *tctx = current->io_uring;
8527 struct io_tctx_exit *work;
8529 work = container_of(cb, struct io_tctx_exit, task_work);
8531 * When @in_idle, we're in cancellation and it's racy to remove the
8532 * node. It'll be removed by the end of cancellation, just ignore it.
8534 if (!atomic_read(&tctx->in_idle))
8535 io_uring_del_task_file((unsigned long)work->ctx);
8536 complete(&work->completion);
8539 static void io_ring_exit_work(struct work_struct *work)
8541 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8542 unsigned long timeout = jiffies + HZ * 60 * 5;
8543 struct io_tctx_exit exit;
8544 struct io_tctx_node *node;
8548 * If we're doing polled IO and end up having requests being
8549 * submitted async (out-of-line), then completions can come in while
8550 * we're waiting for refs to drop. We need to reap these manually,
8551 * as nobody else will be looking for them.
8554 io_uring_try_cancel_requests(ctx, NULL, NULL);
8556 WARN_ON_ONCE(time_after(jiffies, timeout));
8557 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8559 init_completion(&exit.completion);
8560 init_task_work(&exit.task_work, io_tctx_exit_cb);
8563 * Some may use context even when all refs and requests have been put,
8564 * and they are free to do so while still holding uring_lock or
8565 * completion_lock, see __io_req_task_submit(). Apart from other work,
8566 * this lock/unlock section also waits them to finish.
8568 mutex_lock(&ctx->uring_lock);
8569 while (!list_empty(&ctx->tctx_list)) {
8570 WARN_ON_ONCE(time_after(jiffies, timeout));
8572 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8574 /* don't spin on a single task if cancellation failed */
8575 list_rotate_left(&ctx->tctx_list);
8576 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8577 if (WARN_ON_ONCE(ret))
8579 wake_up_process(node->task);
8581 mutex_unlock(&ctx->uring_lock);
8582 wait_for_completion(&exit.completion);
8583 mutex_lock(&ctx->uring_lock);
8585 mutex_unlock(&ctx->uring_lock);
8586 spin_lock_irq(&ctx->completion_lock);
8587 spin_unlock_irq(&ctx->completion_lock);
8589 io_ring_ctx_free(ctx);
8592 /* Returns true if we found and killed one or more timeouts */
8593 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8594 struct files_struct *files)
8596 struct io_kiocb *req, *tmp;
8599 spin_lock_irq(&ctx->completion_lock);
8600 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8601 if (io_match_task(req, tsk, files)) {
8602 io_kill_timeout(req, -ECANCELED);
8607 io_commit_cqring(ctx);
8608 spin_unlock_irq(&ctx->completion_lock);
8610 io_cqring_ev_posted(ctx);
8611 return canceled != 0;
8614 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8616 unsigned long index;
8617 struct creds *creds;
8619 mutex_lock(&ctx->uring_lock);
8620 percpu_ref_kill(&ctx->refs);
8622 __io_cqring_overflow_flush(ctx, true);
8623 xa_for_each(&ctx->personalities, index, creds)
8624 io_unregister_personality(ctx, index);
8625 mutex_unlock(&ctx->uring_lock);
8627 io_kill_timeouts(ctx, NULL, NULL);
8628 io_poll_remove_all(ctx, NULL, NULL);
8630 /* if we failed setting up the ctx, we might not have any rings */
8631 io_iopoll_try_reap_events(ctx);
8633 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8635 * Use system_unbound_wq to avoid spawning tons of event kworkers
8636 * if we're exiting a ton of rings at the same time. It just adds
8637 * noise and overhead, there's no discernable change in runtime
8638 * over using system_wq.
8640 queue_work(system_unbound_wq, &ctx->exit_work);
8643 static int io_uring_release(struct inode *inode, struct file *file)
8645 struct io_ring_ctx *ctx = file->private_data;
8647 file->private_data = NULL;
8648 io_ring_ctx_wait_and_kill(ctx);
8652 struct io_task_cancel {
8653 struct task_struct *task;
8654 struct files_struct *files;
8657 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8659 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8660 struct io_task_cancel *cancel = data;
8663 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8664 unsigned long flags;
8665 struct io_ring_ctx *ctx = req->ctx;
8667 /* protect against races with linked timeouts */
8668 spin_lock_irqsave(&ctx->completion_lock, flags);
8669 ret = io_match_task(req, cancel->task, cancel->files);
8670 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8672 ret = io_match_task(req, cancel->task, cancel->files);
8677 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8678 struct task_struct *task,
8679 struct files_struct *files)
8681 struct io_defer_entry *de;
8684 spin_lock_irq(&ctx->completion_lock);
8685 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8686 if (io_match_task(de->req, task, files)) {
8687 list_cut_position(&list, &ctx->defer_list, &de->list);
8691 spin_unlock_irq(&ctx->completion_lock);
8692 if (list_empty(&list))
8695 while (!list_empty(&list)) {
8696 de = list_first_entry(&list, struct io_defer_entry, list);
8697 list_del_init(&de->list);
8698 io_req_complete_failed(de->req, -ECANCELED);
8704 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8706 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8708 return req->ctx == data;
8711 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8713 struct io_tctx_node *node;
8714 enum io_wq_cancel cret;
8717 mutex_lock(&ctx->uring_lock);
8718 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8719 struct io_uring_task *tctx = node->task->io_uring;
8722 * io_wq will stay alive while we hold uring_lock, because it's
8723 * killed after ctx nodes, which requires to take the lock.
8725 if (!tctx || !tctx->io_wq)
8727 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8728 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8730 mutex_unlock(&ctx->uring_lock);
8735 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8736 struct task_struct *task,
8737 struct files_struct *files)
8739 struct io_task_cancel cancel = { .task = task, .files = files, };
8740 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8743 enum io_wq_cancel cret;
8747 ret |= io_uring_try_cancel_iowq(ctx);
8748 } else if (tctx && tctx->io_wq) {
8750 * Cancels requests of all rings, not only @ctx, but
8751 * it's fine as the task is in exit/exec.
8753 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8755 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8758 /* SQPOLL thread does its own polling */
8759 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8760 (ctx->sq_data && ctx->sq_data->thread == current)) {
8761 while (!list_empty_careful(&ctx->iopoll_list)) {
8762 io_iopoll_try_reap_events(ctx);
8767 ret |= io_cancel_defer_files(ctx, task, files);
8768 ret |= io_poll_remove_all(ctx, task, files);
8769 ret |= io_kill_timeouts(ctx, task, files);
8770 ret |= io_run_task_work();
8771 ret |= io_run_ctx_fallback(ctx);
8778 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8780 struct io_uring_task *tctx = current->io_uring;
8781 struct io_tctx_node *node;
8784 if (unlikely(!tctx)) {
8785 ret = io_uring_alloc_task_context(current, ctx);
8788 tctx = current->io_uring;
8790 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8791 node = kmalloc(sizeof(*node), GFP_KERNEL);
8795 node->task = current;
8797 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8804 mutex_lock(&ctx->uring_lock);
8805 list_add(&node->ctx_node, &ctx->tctx_list);
8806 mutex_unlock(&ctx->uring_lock);
8813 * Note that this task has used io_uring. We use it for cancelation purposes.
8815 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8817 struct io_uring_task *tctx = current->io_uring;
8819 if (likely(tctx && tctx->last == ctx))
8821 return __io_uring_add_task_file(ctx);
8825 * Remove this io_uring_file -> task mapping.
8827 static void io_uring_del_task_file(unsigned long index)
8829 struct io_uring_task *tctx = current->io_uring;
8830 struct io_tctx_node *node;
8834 node = xa_erase(&tctx->xa, index);
8838 WARN_ON_ONCE(current != node->task);
8839 WARN_ON_ONCE(list_empty(&node->ctx_node));
8841 mutex_lock(&node->ctx->uring_lock);
8842 list_del(&node->ctx_node);
8843 mutex_unlock(&node->ctx->uring_lock);
8845 if (tctx->last == node->ctx)
8850 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8852 struct io_tctx_node *node;
8853 unsigned long index;
8855 xa_for_each(&tctx->xa, index, node)
8856 io_uring_del_task_file(index);
8858 io_wq_put_and_exit(tctx->io_wq);
8863 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
8866 return atomic_read(&tctx->inflight_tracked);
8867 return percpu_counter_sum(&tctx->inflight);
8870 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8872 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8873 struct io_ring_ctx *ctx = work->ctx;
8874 struct io_sq_data *sqd = ctx->sq_data;
8877 io_uring_cancel_sqpoll(ctx);
8878 complete(&work->completion);
8881 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8883 struct io_sq_data *sqd = ctx->sq_data;
8884 struct io_tctx_exit work = { .ctx = ctx, };
8885 struct task_struct *task;
8887 io_sq_thread_park(sqd);
8888 list_del_init(&ctx->sqd_list);
8889 io_sqd_update_thread_idle(sqd);
8892 init_completion(&work.completion);
8893 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8894 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8895 wake_up_process(task);
8897 io_sq_thread_unpark(sqd);
8900 wait_for_completion(&work.completion);
8903 static void io_uring_try_cancel(struct files_struct *files)
8905 struct io_uring_task *tctx = current->io_uring;
8906 struct io_tctx_node *node;
8907 unsigned long index;
8909 xa_for_each(&tctx->xa, index, node) {
8910 struct io_ring_ctx *ctx = node->ctx;
8913 io_sqpoll_cancel_sync(ctx);
8916 io_uring_try_cancel_requests(ctx, current, files);
8920 /* should only be called by SQPOLL task */
8921 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8923 struct io_sq_data *sqd = ctx->sq_data;
8924 struct io_uring_task *tctx = current->io_uring;
8928 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8930 atomic_inc(&tctx->in_idle);
8932 /* read completions before cancelations */
8933 inflight = tctx_inflight(tctx, false);
8936 io_uring_try_cancel_requests(ctx, current, NULL);
8938 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8940 * If we've seen completions, retry without waiting. This
8941 * avoids a race where a completion comes in before we did
8942 * prepare_to_wait().
8944 if (inflight == tctx_inflight(tctx, false))
8946 finish_wait(&tctx->wait, &wait);
8948 atomic_dec(&tctx->in_idle);
8952 * Find any io_uring fd that this task has registered or done IO on, and cancel
8955 void __io_uring_cancel(struct files_struct *files)
8957 struct io_uring_task *tctx = current->io_uring;
8961 /* make sure overflow events are dropped */
8962 atomic_inc(&tctx->in_idle);
8963 io_uring_try_cancel(files);
8966 /* read completions before cancelations */
8967 inflight = tctx_inflight(tctx, !!files);
8970 io_uring_try_cancel(files);
8971 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8974 * If we've seen completions, retry without waiting. This
8975 * avoids a race where a completion comes in before we did
8976 * prepare_to_wait().
8978 if (inflight == tctx_inflight(tctx, !!files))
8980 finish_wait(&tctx->wait, &wait);
8982 atomic_dec(&tctx->in_idle);
8984 io_uring_clean_tctx(tctx);
8986 /* for exec all current's requests should be gone, kill tctx */
8987 __io_uring_free(current);
8991 static void *io_uring_validate_mmap_request(struct file *file,
8992 loff_t pgoff, size_t sz)
8994 struct io_ring_ctx *ctx = file->private_data;
8995 loff_t offset = pgoff << PAGE_SHIFT;
9000 case IORING_OFF_SQ_RING:
9001 case IORING_OFF_CQ_RING:
9004 case IORING_OFF_SQES:
9008 return ERR_PTR(-EINVAL);
9011 page = virt_to_head_page(ptr);
9012 if (sz > page_size(page))
9013 return ERR_PTR(-EINVAL);
9020 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9022 size_t sz = vma->vm_end - vma->vm_start;
9026 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9028 return PTR_ERR(ptr);
9030 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9031 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9034 #else /* !CONFIG_MMU */
9036 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9038 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9041 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9043 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9046 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9047 unsigned long addr, unsigned long len,
9048 unsigned long pgoff, unsigned long flags)
9052 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9054 return PTR_ERR(ptr);
9056 return (unsigned long) ptr;
9059 #endif /* !CONFIG_MMU */
9061 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9066 if (!io_sqring_full(ctx))
9068 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9070 if (!io_sqring_full(ctx))
9073 } while (!signal_pending(current));
9075 finish_wait(&ctx->sqo_sq_wait, &wait);
9079 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9080 struct __kernel_timespec __user **ts,
9081 const sigset_t __user **sig)
9083 struct io_uring_getevents_arg arg;
9086 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9087 * is just a pointer to the sigset_t.
9089 if (!(flags & IORING_ENTER_EXT_ARG)) {
9090 *sig = (const sigset_t __user *) argp;
9096 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9097 * timespec and sigset_t pointers if good.
9099 if (*argsz != sizeof(arg))
9101 if (copy_from_user(&arg, argp, sizeof(arg)))
9103 *sig = u64_to_user_ptr(arg.sigmask);
9104 *argsz = arg.sigmask_sz;
9105 *ts = u64_to_user_ptr(arg.ts);
9109 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9110 u32, min_complete, u32, flags, const void __user *, argp,
9113 struct io_ring_ctx *ctx;
9120 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9121 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9125 if (unlikely(!f.file))
9129 if (unlikely(f.file->f_op != &io_uring_fops))
9133 ctx = f.file->private_data;
9134 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9138 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9142 * For SQ polling, the thread will do all submissions and completions.
9143 * Just return the requested submit count, and wake the thread if
9147 if (ctx->flags & IORING_SETUP_SQPOLL) {
9148 io_cqring_overflow_flush(ctx, false);
9151 if (unlikely(ctx->sq_data->thread == NULL)) {
9154 if (flags & IORING_ENTER_SQ_WAKEUP)
9155 wake_up(&ctx->sq_data->wait);
9156 if (flags & IORING_ENTER_SQ_WAIT) {
9157 ret = io_sqpoll_wait_sq(ctx);
9161 submitted = to_submit;
9162 } else if (to_submit) {
9163 ret = io_uring_add_task_file(ctx);
9166 mutex_lock(&ctx->uring_lock);
9167 submitted = io_submit_sqes(ctx, to_submit);
9168 mutex_unlock(&ctx->uring_lock);
9170 if (submitted != to_submit)
9173 if (flags & IORING_ENTER_GETEVENTS) {
9174 const sigset_t __user *sig;
9175 struct __kernel_timespec __user *ts;
9177 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9181 min_complete = min(min_complete, ctx->cq_entries);
9184 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9185 * space applications don't need to do io completion events
9186 * polling again, they can rely on io_sq_thread to do polling
9187 * work, which can reduce cpu usage and uring_lock contention.
9189 if (ctx->flags & IORING_SETUP_IOPOLL &&
9190 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9191 ret = io_iopoll_check(ctx, min_complete);
9193 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9198 percpu_ref_put(&ctx->refs);
9201 return submitted ? submitted : ret;
9204 #ifdef CONFIG_PROC_FS
9205 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9206 const struct cred *cred)
9208 struct user_namespace *uns = seq_user_ns(m);
9209 struct group_info *gi;
9214 seq_printf(m, "%5d\n", id);
9215 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9216 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9217 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9218 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9219 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9220 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9221 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9222 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9223 seq_puts(m, "\n\tGroups:\t");
9224 gi = cred->group_info;
9225 for (g = 0; g < gi->ngroups; g++) {
9226 seq_put_decimal_ull(m, g ? " " : "",
9227 from_kgid_munged(uns, gi->gid[g]));
9229 seq_puts(m, "\n\tCapEff:\t");
9230 cap = cred->cap_effective;
9231 CAP_FOR_EACH_U32(__capi)
9232 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9237 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9239 struct io_sq_data *sq = NULL;
9244 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9245 * since fdinfo case grabs it in the opposite direction of normal use
9246 * cases. If we fail to get the lock, we just don't iterate any
9247 * structures that could be going away outside the io_uring mutex.
9249 has_lock = mutex_trylock(&ctx->uring_lock);
9251 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9257 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9258 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9259 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9260 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9261 struct file *f = io_file_from_index(ctx, i);
9264 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9266 seq_printf(m, "%5u: <none>\n", i);
9268 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9269 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9270 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9271 unsigned int len = buf->ubuf_end - buf->ubuf;
9273 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9275 if (has_lock && !xa_empty(&ctx->personalities)) {
9276 unsigned long index;
9277 const struct cred *cred;
9279 seq_printf(m, "Personalities:\n");
9280 xa_for_each(&ctx->personalities, index, cred)
9281 io_uring_show_cred(m, index, cred);
9283 seq_printf(m, "PollList:\n");
9284 spin_lock_irq(&ctx->completion_lock);
9285 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9286 struct hlist_head *list = &ctx->cancel_hash[i];
9287 struct io_kiocb *req;
9289 hlist_for_each_entry(req, list, hash_node)
9290 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9291 req->task->task_works != NULL);
9293 spin_unlock_irq(&ctx->completion_lock);
9295 mutex_unlock(&ctx->uring_lock);
9298 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9300 struct io_ring_ctx *ctx = f->private_data;
9302 if (percpu_ref_tryget(&ctx->refs)) {
9303 __io_uring_show_fdinfo(ctx, m);
9304 percpu_ref_put(&ctx->refs);
9309 static const struct file_operations io_uring_fops = {
9310 .release = io_uring_release,
9311 .mmap = io_uring_mmap,
9313 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9314 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9316 .poll = io_uring_poll,
9317 .fasync = io_uring_fasync,
9318 #ifdef CONFIG_PROC_FS
9319 .show_fdinfo = io_uring_show_fdinfo,
9323 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9324 struct io_uring_params *p)
9326 struct io_rings *rings;
9327 size_t size, sq_array_offset;
9329 /* make sure these are sane, as we already accounted them */
9330 ctx->sq_entries = p->sq_entries;
9331 ctx->cq_entries = p->cq_entries;
9333 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9334 if (size == SIZE_MAX)
9337 rings = io_mem_alloc(size);
9342 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9343 rings->sq_ring_mask = p->sq_entries - 1;
9344 rings->cq_ring_mask = p->cq_entries - 1;
9345 rings->sq_ring_entries = p->sq_entries;
9346 rings->cq_ring_entries = p->cq_entries;
9347 ctx->sq_mask = rings->sq_ring_mask;
9348 ctx->cq_mask = rings->cq_ring_mask;
9350 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9351 if (size == SIZE_MAX) {
9352 io_mem_free(ctx->rings);
9357 ctx->sq_sqes = io_mem_alloc(size);
9358 if (!ctx->sq_sqes) {
9359 io_mem_free(ctx->rings);
9367 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9371 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9375 ret = io_uring_add_task_file(ctx);
9380 fd_install(fd, file);
9385 * Allocate an anonymous fd, this is what constitutes the application
9386 * visible backing of an io_uring instance. The application mmaps this
9387 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9388 * we have to tie this fd to a socket for file garbage collection purposes.
9390 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9393 #if defined(CONFIG_UNIX)
9396 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9399 return ERR_PTR(ret);
9402 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9403 O_RDWR | O_CLOEXEC);
9404 #if defined(CONFIG_UNIX)
9406 sock_release(ctx->ring_sock);
9407 ctx->ring_sock = NULL;
9409 ctx->ring_sock->file = file;
9415 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9416 struct io_uring_params __user *params)
9418 struct io_ring_ctx *ctx;
9424 if (entries > IORING_MAX_ENTRIES) {
9425 if (!(p->flags & IORING_SETUP_CLAMP))
9427 entries = IORING_MAX_ENTRIES;
9431 * Use twice as many entries for the CQ ring. It's possible for the
9432 * application to drive a higher depth than the size of the SQ ring,
9433 * since the sqes are only used at submission time. This allows for
9434 * some flexibility in overcommitting a bit. If the application has
9435 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9436 * of CQ ring entries manually.
9438 p->sq_entries = roundup_pow_of_two(entries);
9439 if (p->flags & IORING_SETUP_CQSIZE) {
9441 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9442 * to a power-of-two, if it isn't already. We do NOT impose
9443 * any cq vs sq ring sizing.
9447 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9448 if (!(p->flags & IORING_SETUP_CLAMP))
9450 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9452 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9453 if (p->cq_entries < p->sq_entries)
9456 p->cq_entries = 2 * p->sq_entries;
9459 ctx = io_ring_ctx_alloc(p);
9462 ctx->compat = in_compat_syscall();
9463 if (!capable(CAP_IPC_LOCK))
9464 ctx->user = get_uid(current_user());
9467 * This is just grabbed for accounting purposes. When a process exits,
9468 * the mm is exited and dropped before the files, hence we need to hang
9469 * on to this mm purely for the purposes of being able to unaccount
9470 * memory (locked/pinned vm). It's not used for anything else.
9472 mmgrab(current->mm);
9473 ctx->mm_account = current->mm;
9475 ret = io_allocate_scq_urings(ctx, p);
9479 ret = io_sq_offload_create(ctx, p);
9483 memset(&p->sq_off, 0, sizeof(p->sq_off));
9484 p->sq_off.head = offsetof(struct io_rings, sq.head);
9485 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9486 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9487 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9488 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9489 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9490 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9492 memset(&p->cq_off, 0, sizeof(p->cq_off));
9493 p->cq_off.head = offsetof(struct io_rings, cq.head);
9494 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9495 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9496 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9497 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9498 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9499 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9501 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9502 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9503 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9504 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9505 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9507 if (copy_to_user(params, p, sizeof(*p))) {
9512 file = io_uring_get_file(ctx);
9514 ret = PTR_ERR(file);
9519 * Install ring fd as the very last thing, so we don't risk someone
9520 * having closed it before we finish setup
9522 ret = io_uring_install_fd(ctx, file);
9524 /* fput will clean it up */
9529 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9532 io_ring_ctx_wait_and_kill(ctx);
9537 * Sets up an aio uring context, and returns the fd. Applications asks for a
9538 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9539 * params structure passed in.
9541 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9543 struct io_uring_params p;
9546 if (copy_from_user(&p, params, sizeof(p)))
9548 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9553 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9554 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9555 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9556 IORING_SETUP_R_DISABLED))
9559 return io_uring_create(entries, &p, params);
9562 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9563 struct io_uring_params __user *, params)
9565 return io_uring_setup(entries, params);
9568 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9570 struct io_uring_probe *p;
9574 size = struct_size(p, ops, nr_args);
9575 if (size == SIZE_MAX)
9577 p = kzalloc(size, GFP_KERNEL);
9582 if (copy_from_user(p, arg, size))
9585 if (memchr_inv(p, 0, size))
9588 p->last_op = IORING_OP_LAST - 1;
9589 if (nr_args > IORING_OP_LAST)
9590 nr_args = IORING_OP_LAST;
9592 for (i = 0; i < nr_args; i++) {
9594 if (!io_op_defs[i].not_supported)
9595 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9600 if (copy_to_user(arg, p, size))
9607 static int io_register_personality(struct io_ring_ctx *ctx)
9609 const struct cred *creds;
9613 creds = get_current_cred();
9615 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9616 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9623 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9624 unsigned int nr_args)
9626 struct io_uring_restriction *res;
9630 /* Restrictions allowed only if rings started disabled */
9631 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9634 /* We allow only a single restrictions registration */
9635 if (ctx->restrictions.registered)
9638 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9641 size = array_size(nr_args, sizeof(*res));
9642 if (size == SIZE_MAX)
9645 res = memdup_user(arg, size);
9647 return PTR_ERR(res);
9651 for (i = 0; i < nr_args; i++) {
9652 switch (res[i].opcode) {
9653 case IORING_RESTRICTION_REGISTER_OP:
9654 if (res[i].register_op >= IORING_REGISTER_LAST) {
9659 __set_bit(res[i].register_op,
9660 ctx->restrictions.register_op);
9662 case IORING_RESTRICTION_SQE_OP:
9663 if (res[i].sqe_op >= IORING_OP_LAST) {
9668 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9670 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9671 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9673 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9674 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9683 /* Reset all restrictions if an error happened */
9685 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9687 ctx->restrictions.registered = true;
9693 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9695 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9698 if (ctx->restrictions.registered)
9699 ctx->restricted = 1;
9701 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9702 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9703 wake_up(&ctx->sq_data->wait);
9707 static bool io_register_op_must_quiesce(int op)
9710 case IORING_REGISTER_FILES:
9711 case IORING_UNREGISTER_FILES:
9712 case IORING_REGISTER_FILES_UPDATE:
9713 case IORING_REGISTER_PROBE:
9714 case IORING_REGISTER_PERSONALITY:
9715 case IORING_UNREGISTER_PERSONALITY:
9722 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9723 void __user *arg, unsigned nr_args)
9724 __releases(ctx->uring_lock)
9725 __acquires(ctx->uring_lock)
9730 * We're inside the ring mutex, if the ref is already dying, then
9731 * someone else killed the ctx or is already going through
9732 * io_uring_register().
9734 if (percpu_ref_is_dying(&ctx->refs))
9737 if (ctx->restricted) {
9738 if (opcode >= IORING_REGISTER_LAST)
9740 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9741 if (!test_bit(opcode, ctx->restrictions.register_op))
9745 if (io_register_op_must_quiesce(opcode)) {
9746 percpu_ref_kill(&ctx->refs);
9749 * Drop uring mutex before waiting for references to exit. If
9750 * another thread is currently inside io_uring_enter() it might
9751 * need to grab the uring_lock to make progress. If we hold it
9752 * here across the drain wait, then we can deadlock. It's safe
9753 * to drop the mutex here, since no new references will come in
9754 * after we've killed the percpu ref.
9756 mutex_unlock(&ctx->uring_lock);
9758 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9761 ret = io_run_task_work_sig();
9765 mutex_lock(&ctx->uring_lock);
9768 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
9774 case IORING_REGISTER_BUFFERS:
9775 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9777 case IORING_UNREGISTER_BUFFERS:
9781 ret = io_sqe_buffers_unregister(ctx);
9783 case IORING_REGISTER_FILES:
9784 ret = io_sqe_files_register(ctx, arg, nr_args);
9786 case IORING_UNREGISTER_FILES:
9790 ret = io_sqe_files_unregister(ctx);
9792 case IORING_REGISTER_FILES_UPDATE:
9793 ret = io_sqe_files_update(ctx, arg, nr_args);
9795 case IORING_REGISTER_EVENTFD:
9796 case IORING_REGISTER_EVENTFD_ASYNC:
9800 ret = io_eventfd_register(ctx, arg);
9803 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9804 ctx->eventfd_async = 1;
9806 ctx->eventfd_async = 0;
9808 case IORING_UNREGISTER_EVENTFD:
9812 ret = io_eventfd_unregister(ctx);
9814 case IORING_REGISTER_PROBE:
9816 if (!arg || nr_args > 256)
9818 ret = io_probe(ctx, arg, nr_args);
9820 case IORING_REGISTER_PERSONALITY:
9824 ret = io_register_personality(ctx);
9826 case IORING_UNREGISTER_PERSONALITY:
9830 ret = io_unregister_personality(ctx, nr_args);
9832 case IORING_REGISTER_ENABLE_RINGS:
9836 ret = io_register_enable_rings(ctx);
9838 case IORING_REGISTER_RESTRICTIONS:
9839 ret = io_register_restrictions(ctx, arg, nr_args);
9846 if (io_register_op_must_quiesce(opcode)) {
9847 /* bring the ctx back to life */
9848 percpu_ref_reinit(&ctx->refs);
9849 reinit_completion(&ctx->ref_comp);
9854 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9855 void __user *, arg, unsigned int, nr_args)
9857 struct io_ring_ctx *ctx;
9866 if (f.file->f_op != &io_uring_fops)
9869 ctx = f.file->private_data;
9873 mutex_lock(&ctx->uring_lock);
9874 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9875 mutex_unlock(&ctx->uring_lock);
9876 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9877 ctx->cq_ev_fd != NULL, ret);
9883 static int __init io_uring_init(void)
9885 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9886 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9887 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9890 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9891 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9892 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9893 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9894 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9895 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9896 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9897 BUILD_BUG_SQE_ELEM(8, __u64, off);
9898 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9899 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9900 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9901 BUILD_BUG_SQE_ELEM(24, __u32, len);
9902 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9903 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9904 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9905 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9906 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9907 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9908 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9909 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9910 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9911 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9912 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9913 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9914 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9915 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9916 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9917 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9918 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9919 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9920 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9922 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9923 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9924 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9928 __initcall(io_uring_init);