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_cqe (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>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 /* 512 entries per page on 64-bit archs, 64 pages max */
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT 9
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head cq_overflow_list;
379 struct xarray io_buffers;
380 struct xarray personalities;
382 unsigned sq_thread_idle;
383 } ____cacheline_aligned_in_smp;
385 /* IRQ completion list, under ->completion_lock */
386 struct list_head locked_free_list;
387 unsigned int locked_free_nr;
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;
395 unsigned long check_cq_overflow;
398 unsigned cached_cq_tail;
400 struct eventfd_ctx *cq_ev_fd;
401 struct wait_queue_head poll_wait;
402 struct wait_queue_head cq_wait;
404 atomic_t cq_timeouts;
405 struct fasync_struct *cq_fasync;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
462 struct io_uring_task {
463 /* submission side */
466 struct wait_queue_head wait;
467 const struct io_ring_ctx *last;
469 struct percpu_counter inflight;
470 atomic_t inflight_tracked;
473 spinlock_t task_lock;
474 struct io_wq_work_list task_list;
475 struct callback_head task_work;
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
483 struct io_poll_iocb {
485 struct wait_queue_head *head;
489 struct wait_queue_entry wait;
492 struct io_poll_update {
498 bool update_user_data;
506 struct io_timeout_data {
507 struct io_kiocb *req;
508 struct hrtimer timer;
509 struct timespec64 ts;
510 enum hrtimer_mode mode;
515 struct sockaddr __user *addr;
516 int __user *addr_len;
518 unsigned long nofile;
538 struct list_head list;
539 /* head of the link, used by linked timeouts only */
540 struct io_kiocb *head;
541 /* for linked completions */
542 struct io_kiocb *prev;
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 {
670 struct io_async_connect {
671 struct sockaddr_storage address;
674 struct io_async_msghdr {
675 struct iovec fast_iov[UIO_FASTIOV];
676 /* points to an allocated iov, if NULL we use fast_iov instead */
677 struct iovec *free_iov;
678 struct sockaddr __user *uaddr;
680 struct sockaddr_storage addr;
684 struct iovec fast_iov[UIO_FASTIOV];
685 const struct iovec *free_iovec;
686 struct iov_iter iter;
688 struct wait_page_queue wpq;
692 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
693 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
694 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
695 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
696 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
697 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
699 /* first byte is taken by user flags, shift it to not overlap */
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,
714 /* keep async read/write and isreg together and in order */
715 REQ_F_NOWAIT_READ_BIT,
716 REQ_F_NOWAIT_WRITE_BIT,
719 /* not a real bit, just to check we're not overflowing the space */
725 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
726 /* drain existing IO first */
727 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
729 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
730 /* doesn't sever on completion < 0 */
731 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
733 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
734 /* IOSQE_BUFFER_SELECT */
735 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
737 /* fail rest of links */
738 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
739 /* on inflight list, should be cancelled and waited on exit reliably */
740 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
741 /* read/write uses file position */
742 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
743 /* must not punt to workers */
744 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
745 /* has or had linked timeout */
746 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
748 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
749 /* already went through poll handler */
750 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
751 /* buffer already selected */
752 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
753 /* linked timeout is active, i.e. prepared by link's head */
754 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
755 /* completion is deferred through io_comp_state */
756 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
757 /* caller should reissue async */
758 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
759 /* don't attempt request reissue, see io_rw_reissue() */
760 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
761 /* supports async reads */
762 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
763 /* supports async writes */
764 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
766 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
767 /* has creds assigned */
768 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
769 /* skip refcounting if not set */
770 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
774 struct io_poll_iocb poll;
775 struct io_poll_iocb *double_poll;
778 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
780 struct io_task_work {
782 struct io_wq_work_node node;
783 struct llist_node fallback_node;
785 io_req_tw_func_t func;
789 IORING_RSRC_FILE = 0,
790 IORING_RSRC_BUFFER = 1,
794 * NOTE! Each of the iocb union members has the file pointer
795 * as the first entry in their struct definition. So you can
796 * access the file pointer through any of the sub-structs,
797 * or directly as just 'ki_filp' in this struct.
803 struct io_poll_iocb poll;
804 struct io_poll_update poll_update;
805 struct io_accept accept;
807 struct io_cancel cancel;
808 struct io_timeout timeout;
809 struct io_timeout_rem timeout_rem;
810 struct io_connect connect;
811 struct io_sr_msg sr_msg;
813 struct io_close close;
814 struct io_rsrc_update rsrc_update;
815 struct io_fadvise fadvise;
816 struct io_madvise madvise;
817 struct io_epoll epoll;
818 struct io_splice splice;
819 struct io_provide_buf pbuf;
820 struct io_statx statx;
821 struct io_shutdown shutdown;
822 struct io_rename rename;
823 struct io_unlink unlink;
824 /* use only after cleaning per-op data, see io_clean_op() */
825 struct io_completion compl;
828 /* opcode allocated if it needs to store data for async defer */
831 /* polled IO has completed */
837 struct io_ring_ctx *ctx;
840 struct task_struct *task;
843 struct io_kiocb *link;
844 struct percpu_ref *fixed_rsrc_refs;
846 /* used with ctx->iopoll_list with reads/writes */
847 struct list_head inflight_entry;
848 struct io_task_work io_task_work;
849 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
850 struct hlist_node hash_node;
851 struct async_poll *apoll;
852 struct io_wq_work work;
853 const struct cred *creds;
855 /* store used ubuf, so we can prevent reloading */
856 struct io_mapped_ubuf *imu;
859 struct io_tctx_node {
860 struct list_head ctx_node;
861 struct task_struct *task;
862 struct io_ring_ctx *ctx;
865 struct io_defer_entry {
866 struct list_head list;
867 struct io_kiocb *req;
872 /* needs req->file assigned */
873 unsigned needs_file : 1;
874 /* hash wq insertion if file is a regular file */
875 unsigned hash_reg_file : 1;
876 /* unbound wq insertion if file is a non-regular file */
877 unsigned unbound_nonreg_file : 1;
878 /* opcode is not supported by this kernel */
879 unsigned not_supported : 1;
880 /* set if opcode supports polled "wait" */
882 unsigned pollout : 1;
883 /* op supports buffer selection */
884 unsigned buffer_select : 1;
885 /* do prep async if is going to be punted */
886 unsigned needs_async_setup : 1;
887 /* should block plug */
889 /* size of async data needed, if any */
890 unsigned short async_size;
893 static const struct io_op_def io_op_defs[] = {
894 [IORING_OP_NOP] = {},
895 [IORING_OP_READV] = {
897 .unbound_nonreg_file = 1,
900 .needs_async_setup = 1,
902 .async_size = sizeof(struct io_async_rw),
904 [IORING_OP_WRITEV] = {
907 .unbound_nonreg_file = 1,
909 .needs_async_setup = 1,
911 .async_size = sizeof(struct io_async_rw),
913 [IORING_OP_FSYNC] = {
916 [IORING_OP_READ_FIXED] = {
918 .unbound_nonreg_file = 1,
921 .async_size = sizeof(struct io_async_rw),
923 [IORING_OP_WRITE_FIXED] = {
926 .unbound_nonreg_file = 1,
929 .async_size = sizeof(struct io_async_rw),
931 [IORING_OP_POLL_ADD] = {
933 .unbound_nonreg_file = 1,
935 [IORING_OP_POLL_REMOVE] = {},
936 [IORING_OP_SYNC_FILE_RANGE] = {
939 [IORING_OP_SENDMSG] = {
941 .unbound_nonreg_file = 1,
943 .needs_async_setup = 1,
944 .async_size = sizeof(struct io_async_msghdr),
946 [IORING_OP_RECVMSG] = {
948 .unbound_nonreg_file = 1,
951 .needs_async_setup = 1,
952 .async_size = sizeof(struct io_async_msghdr),
954 [IORING_OP_TIMEOUT] = {
955 .async_size = sizeof(struct io_timeout_data),
957 [IORING_OP_TIMEOUT_REMOVE] = {
958 /* used by timeout updates' prep() */
960 [IORING_OP_ACCEPT] = {
962 .unbound_nonreg_file = 1,
965 [IORING_OP_ASYNC_CANCEL] = {},
966 [IORING_OP_LINK_TIMEOUT] = {
967 .async_size = sizeof(struct io_timeout_data),
969 [IORING_OP_CONNECT] = {
971 .unbound_nonreg_file = 1,
973 .needs_async_setup = 1,
974 .async_size = sizeof(struct io_async_connect),
976 [IORING_OP_FALLOCATE] = {
979 [IORING_OP_OPENAT] = {},
980 [IORING_OP_CLOSE] = {},
981 [IORING_OP_FILES_UPDATE] = {},
982 [IORING_OP_STATX] = {},
985 .unbound_nonreg_file = 1,
989 .async_size = sizeof(struct io_async_rw),
991 [IORING_OP_WRITE] = {
993 .unbound_nonreg_file = 1,
996 .async_size = sizeof(struct io_async_rw),
998 [IORING_OP_FADVISE] = {
1001 [IORING_OP_MADVISE] = {},
1002 [IORING_OP_SEND] = {
1004 .unbound_nonreg_file = 1,
1007 [IORING_OP_RECV] = {
1009 .unbound_nonreg_file = 1,
1013 [IORING_OP_OPENAT2] = {
1015 [IORING_OP_EPOLL_CTL] = {
1016 .unbound_nonreg_file = 1,
1018 [IORING_OP_SPLICE] = {
1021 .unbound_nonreg_file = 1,
1023 [IORING_OP_PROVIDE_BUFFERS] = {},
1024 [IORING_OP_REMOVE_BUFFERS] = {},
1028 .unbound_nonreg_file = 1,
1030 [IORING_OP_SHUTDOWN] = {
1033 [IORING_OP_RENAMEAT] = {},
1034 [IORING_OP_UNLINKAT] = {},
1037 static bool io_disarm_next(struct io_kiocb *req);
1038 static void io_uring_del_tctx_node(unsigned long index);
1039 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1040 struct task_struct *task,
1042 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1050 static void io_queue_linked_timeout(struct io_kiocb *req);
1051 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1052 struct io_uring_rsrc_update2 *up,
1054 static void io_clean_op(struct io_kiocb *req);
1055 static struct file *io_file_get(struct io_ring_ctx *ctx,
1056 struct io_kiocb *req, int fd, bool fixed);
1057 static void __io_queue_sqe(struct io_kiocb *req);
1058 static void io_rsrc_put_work(struct work_struct *work);
1060 static void io_req_task_queue(struct io_kiocb *req);
1061 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1062 static int io_req_prep_async(struct io_kiocb *req);
1064 static struct kmem_cache *req_cachep;
1066 static const struct file_operations io_uring_fops;
1068 struct sock *io_uring_get_socket(struct file *file)
1070 #if defined(CONFIG_UNIX)
1071 if (file->f_op == &io_uring_fops) {
1072 struct io_ring_ctx *ctx = file->private_data;
1074 return ctx->ring_sock->sk;
1079 EXPORT_SYMBOL(io_uring_get_socket);
1081 #define io_for_each_link(pos, head) \
1082 for (pos = (head); pos; pos = pos->link)
1085 * Shamelessly stolen from the mm implementation of page reference checking,
1086 * see commit f958d7b528b1 for details.
1088 #define req_ref_zero_or_close_to_overflow(req) \
1089 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1091 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1093 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1094 return atomic_inc_not_zero(&req->refs);
1097 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1099 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1102 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1103 return atomic_dec_and_test(&req->refs);
1106 static inline void req_ref_put(struct io_kiocb *req)
1108 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1109 WARN_ON_ONCE(req_ref_put_and_test(req));
1112 static inline void req_ref_get(struct io_kiocb *req)
1114 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1115 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1116 atomic_inc(&req->refs);
1119 static inline void io_req_refcount(struct io_kiocb *req)
1121 if (!(req->flags & REQ_F_REFCOUNT)) {
1122 req->flags |= REQ_F_REFCOUNT;
1123 atomic_set(&req->refs, 1);
1127 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1129 struct io_ring_ctx *ctx = req->ctx;
1131 if (!req->fixed_rsrc_refs) {
1132 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1133 percpu_ref_get(req->fixed_rsrc_refs);
1137 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1139 bool got = percpu_ref_tryget(ref);
1141 /* already at zero, wait for ->release() */
1143 wait_for_completion(compl);
1144 percpu_ref_resurrect(ref);
1146 percpu_ref_put(ref);
1149 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1152 struct io_kiocb *req;
1154 if (task && head->task != task)
1159 io_for_each_link(req, head) {
1160 if (req->flags & REQ_F_INFLIGHT)
1166 static inline void req_set_fail(struct io_kiocb *req)
1168 req->flags |= REQ_F_FAIL;
1171 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1173 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1175 complete(&ctx->ref_comp);
1178 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1180 return !req->timeout.off;
1183 static void io_fallback_req_func(struct work_struct *work)
1185 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1186 fallback_work.work);
1187 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1188 struct io_kiocb *req, *tmp;
1190 percpu_ref_get(&ctx->refs);
1191 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1192 req->io_task_work.func(req);
1193 percpu_ref_put(&ctx->refs);
1196 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1198 struct io_ring_ctx *ctx;
1201 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1206 * Use 5 bits less than the max cq entries, that should give us around
1207 * 32 entries per hash list if totally full and uniformly spread.
1209 hash_bits = ilog2(p->cq_entries);
1213 ctx->cancel_hash_bits = hash_bits;
1214 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1216 if (!ctx->cancel_hash)
1218 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1220 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1221 if (!ctx->dummy_ubuf)
1223 /* set invalid range, so io_import_fixed() fails meeting it */
1224 ctx->dummy_ubuf->ubuf = -1UL;
1226 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1227 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1230 ctx->flags = p->flags;
1231 init_waitqueue_head(&ctx->sqo_sq_wait);
1232 INIT_LIST_HEAD(&ctx->sqd_list);
1233 init_waitqueue_head(&ctx->poll_wait);
1234 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1235 init_completion(&ctx->ref_comp);
1236 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1237 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1238 mutex_init(&ctx->uring_lock);
1239 init_waitqueue_head(&ctx->cq_wait);
1240 spin_lock_init(&ctx->completion_lock);
1241 spin_lock_init(&ctx->timeout_lock);
1242 INIT_LIST_HEAD(&ctx->iopoll_list);
1243 INIT_LIST_HEAD(&ctx->defer_list);
1244 INIT_LIST_HEAD(&ctx->timeout_list);
1245 spin_lock_init(&ctx->rsrc_ref_lock);
1246 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1247 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1248 init_llist_head(&ctx->rsrc_put_llist);
1249 INIT_LIST_HEAD(&ctx->tctx_list);
1250 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1251 INIT_LIST_HEAD(&ctx->locked_free_list);
1252 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1255 kfree(ctx->dummy_ubuf);
1256 kfree(ctx->cancel_hash);
1261 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1263 struct io_rings *r = ctx->rings;
1265 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1269 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1271 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1272 struct io_ring_ctx *ctx = req->ctx;
1274 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1280 #define FFS_ASYNC_READ 0x1UL
1281 #define FFS_ASYNC_WRITE 0x2UL
1283 #define FFS_ISREG 0x4UL
1285 #define FFS_ISREG 0x0UL
1287 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1289 static inline bool io_req_ffs_set(struct io_kiocb *req)
1291 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1294 static void io_req_track_inflight(struct io_kiocb *req)
1296 if (!(req->flags & REQ_F_INFLIGHT)) {
1297 req->flags |= REQ_F_INFLIGHT;
1298 atomic_inc(¤t->io_uring->inflight_tracked);
1302 static void io_prep_async_work(struct io_kiocb *req)
1304 const struct io_op_def *def = &io_op_defs[req->opcode];
1305 struct io_ring_ctx *ctx = req->ctx;
1307 if (!(req->flags & REQ_F_CREDS)) {
1308 req->flags |= REQ_F_CREDS;
1309 req->creds = get_current_cred();
1312 req->work.list.next = NULL;
1313 req->work.flags = 0;
1314 if (req->flags & REQ_F_FORCE_ASYNC)
1315 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1317 if (req->flags & REQ_F_ISREG) {
1318 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1319 io_wq_hash_work(&req->work, file_inode(req->file));
1320 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1321 if (def->unbound_nonreg_file)
1322 req->work.flags |= IO_WQ_WORK_UNBOUND;
1325 switch (req->opcode) {
1326 case IORING_OP_SPLICE:
1328 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1329 req->work.flags |= IO_WQ_WORK_UNBOUND;
1334 static void io_prep_async_link(struct io_kiocb *req)
1336 struct io_kiocb *cur;
1338 if (req->flags & REQ_F_LINK_TIMEOUT) {
1339 struct io_ring_ctx *ctx = req->ctx;
1341 spin_lock(&ctx->completion_lock);
1342 io_for_each_link(cur, req)
1343 io_prep_async_work(cur);
1344 spin_unlock(&ctx->completion_lock);
1346 io_for_each_link(cur, req)
1347 io_prep_async_work(cur);
1351 static void io_queue_async_work(struct io_kiocb *req)
1353 struct io_ring_ctx *ctx = req->ctx;
1354 struct io_kiocb *link = io_prep_linked_timeout(req);
1355 struct io_uring_task *tctx = req->task->io_uring;
1358 BUG_ON(!tctx->io_wq);
1360 /* init ->work of the whole link before punting */
1361 io_prep_async_link(req);
1364 * Not expected to happen, but if we do have a bug where this _can_
1365 * happen, catch it here and ensure the request is marked as
1366 * canceled. That will make io-wq go through the usual work cancel
1367 * procedure rather than attempt to run this request (or create a new
1370 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1371 req->work.flags |= IO_WQ_WORK_CANCEL;
1373 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1374 &req->work, req->flags);
1375 io_wq_enqueue(tctx->io_wq, &req->work);
1377 io_queue_linked_timeout(link);
1380 static void io_kill_timeout(struct io_kiocb *req, int status)
1381 __must_hold(&req->ctx->completion_lock)
1382 __must_hold(&req->ctx->timeout_lock)
1384 struct io_timeout_data *io = req->async_data;
1386 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1387 atomic_set(&req->ctx->cq_timeouts,
1388 atomic_read(&req->ctx->cq_timeouts) + 1);
1389 list_del_init(&req->timeout.list);
1390 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1391 io_put_req_deferred(req);
1395 static void io_queue_deferred(struct io_ring_ctx *ctx)
1397 while (!list_empty(&ctx->defer_list)) {
1398 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1399 struct io_defer_entry, list);
1401 if (req_need_defer(de->req, de->seq))
1403 list_del_init(&de->list);
1404 io_req_task_queue(de->req);
1409 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1410 __must_hold(&ctx->completion_lock)
1412 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1414 spin_lock_irq(&ctx->timeout_lock);
1415 while (!list_empty(&ctx->timeout_list)) {
1416 u32 events_needed, events_got;
1417 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1418 struct io_kiocb, timeout.list);
1420 if (io_is_timeout_noseq(req))
1424 * Since seq can easily wrap around over time, subtract
1425 * the last seq at which timeouts were flushed before comparing.
1426 * Assuming not more than 2^31-1 events have happened since,
1427 * these subtractions won't have wrapped, so we can check if
1428 * target is in [last_seq, current_seq] by comparing the two.
1430 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1431 events_got = seq - ctx->cq_last_tm_flush;
1432 if (events_got < events_needed)
1435 list_del_init(&req->timeout.list);
1436 io_kill_timeout(req, 0);
1438 ctx->cq_last_tm_flush = seq;
1439 spin_unlock_irq(&ctx->timeout_lock);
1442 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1444 if (ctx->off_timeout_used)
1445 io_flush_timeouts(ctx);
1446 if (ctx->drain_active)
1447 io_queue_deferred(ctx);
1450 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1452 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1453 __io_commit_cqring_flush(ctx);
1454 /* order cqe stores with ring update */
1455 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1458 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1460 struct io_rings *r = ctx->rings;
1462 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1465 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1467 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1470 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1472 struct io_rings *rings = ctx->rings;
1473 unsigned tail, mask = ctx->cq_entries - 1;
1476 * writes to the cq entry need to come after reading head; the
1477 * control dependency is enough as we're using WRITE_ONCE to
1480 if (__io_cqring_events(ctx) == ctx->cq_entries)
1483 tail = ctx->cached_cq_tail++;
1484 return &rings->cqes[tail & mask];
1487 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1489 if (likely(!ctx->cq_ev_fd))
1491 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1493 return !ctx->eventfd_async || io_wq_current_is_worker();
1496 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1499 * wake_up_all() may seem excessive, but io_wake_function() and
1500 * io_should_wake() handle the termination of the loop and only
1501 * wake as many waiters as we need to.
1503 if (wq_has_sleeper(&ctx->cq_wait))
1504 wake_up_all(&ctx->cq_wait);
1505 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1506 wake_up(&ctx->sq_data->wait);
1507 if (io_should_trigger_evfd(ctx))
1508 eventfd_signal(ctx->cq_ev_fd, 1);
1509 if (waitqueue_active(&ctx->poll_wait)) {
1510 wake_up_interruptible(&ctx->poll_wait);
1511 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1515 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1517 if (ctx->flags & IORING_SETUP_SQPOLL) {
1518 if (wq_has_sleeper(&ctx->cq_wait))
1519 wake_up_all(&ctx->cq_wait);
1521 if (io_should_trigger_evfd(ctx))
1522 eventfd_signal(ctx->cq_ev_fd, 1);
1523 if (waitqueue_active(&ctx->poll_wait)) {
1524 wake_up_interruptible(&ctx->poll_wait);
1525 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1529 /* Returns true if there are no backlogged entries after the flush */
1530 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1532 bool all_flushed, posted;
1534 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1538 spin_lock(&ctx->completion_lock);
1539 while (!list_empty(&ctx->cq_overflow_list)) {
1540 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1541 struct io_overflow_cqe *ocqe;
1545 ocqe = list_first_entry(&ctx->cq_overflow_list,
1546 struct io_overflow_cqe, list);
1548 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1550 io_account_cq_overflow(ctx);
1553 list_del(&ocqe->list);
1557 all_flushed = list_empty(&ctx->cq_overflow_list);
1559 clear_bit(0, &ctx->check_cq_overflow);
1560 WRITE_ONCE(ctx->rings->sq_flags,
1561 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1565 io_commit_cqring(ctx);
1566 spin_unlock(&ctx->completion_lock);
1568 io_cqring_ev_posted(ctx);
1572 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1576 if (test_bit(0, &ctx->check_cq_overflow)) {
1577 /* iopoll syncs against uring_lock, not completion_lock */
1578 if (ctx->flags & IORING_SETUP_IOPOLL)
1579 mutex_lock(&ctx->uring_lock);
1580 ret = __io_cqring_overflow_flush(ctx, false);
1581 if (ctx->flags & IORING_SETUP_IOPOLL)
1582 mutex_unlock(&ctx->uring_lock);
1588 /* must to be called somewhat shortly after putting a request */
1589 static inline void io_put_task(struct task_struct *task, int nr)
1591 struct io_uring_task *tctx = task->io_uring;
1593 percpu_counter_sub(&tctx->inflight, nr);
1594 if (unlikely(atomic_read(&tctx->in_idle)))
1595 wake_up(&tctx->wait);
1596 put_task_struct_many(task, nr);
1599 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1600 long res, unsigned int cflags)
1602 struct io_overflow_cqe *ocqe;
1604 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1607 * If we're in ring overflow flush mode, or in task cancel mode,
1608 * or cannot allocate an overflow entry, then we need to drop it
1611 io_account_cq_overflow(ctx);
1614 if (list_empty(&ctx->cq_overflow_list)) {
1615 set_bit(0, &ctx->check_cq_overflow);
1616 WRITE_ONCE(ctx->rings->sq_flags,
1617 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1620 ocqe->cqe.user_data = user_data;
1621 ocqe->cqe.res = res;
1622 ocqe->cqe.flags = cflags;
1623 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1627 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1628 long res, unsigned int cflags)
1630 struct io_uring_cqe *cqe;
1632 trace_io_uring_complete(ctx, user_data, res, cflags);
1635 * If we can't get a cq entry, userspace overflowed the
1636 * submission (by quite a lot). Increment the overflow count in
1639 cqe = io_get_cqe(ctx);
1641 WRITE_ONCE(cqe->user_data, user_data);
1642 WRITE_ONCE(cqe->res, res);
1643 WRITE_ONCE(cqe->flags, cflags);
1646 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1649 /* not as hot to bloat with inlining */
1650 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1651 long res, unsigned int cflags)
1653 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1656 static void io_req_complete_post(struct io_kiocb *req, long res,
1657 unsigned int cflags)
1659 struct io_ring_ctx *ctx = req->ctx;
1661 spin_lock(&ctx->completion_lock);
1662 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1664 * If we're the last reference to this request, add to our locked
1667 if (req_ref_put_and_test(req)) {
1668 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1669 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1670 io_disarm_next(req);
1672 io_req_task_queue(req->link);
1676 io_dismantle_req(req);
1677 io_put_task(req->task, 1);
1678 list_add(&req->inflight_entry, &ctx->locked_free_list);
1679 ctx->locked_free_nr++;
1681 if (!percpu_ref_tryget(&ctx->refs))
1684 io_commit_cqring(ctx);
1685 spin_unlock(&ctx->completion_lock);
1688 io_cqring_ev_posted(ctx);
1689 percpu_ref_put(&ctx->refs);
1693 static inline bool io_req_needs_clean(struct io_kiocb *req)
1695 return req->flags & IO_REQ_CLEAN_FLAGS;
1698 static void io_req_complete_state(struct io_kiocb *req, long res,
1699 unsigned int cflags)
1701 if (io_req_needs_clean(req))
1704 req->compl.cflags = cflags;
1705 req->flags |= REQ_F_COMPLETE_INLINE;
1708 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1709 long res, unsigned cflags)
1711 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1712 io_req_complete_state(req, res, cflags);
1714 io_req_complete_post(req, res, cflags);
1717 static inline void io_req_complete(struct io_kiocb *req, long res)
1719 __io_req_complete(req, 0, res, 0);
1722 static void io_req_complete_failed(struct io_kiocb *req, long res)
1725 io_req_complete_post(req, res, 0);
1729 * Don't initialise the fields below on every allocation, but do that in
1730 * advance and keep them valid across allocations.
1732 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1736 req->async_data = NULL;
1737 /* not necessary, but safer to zero */
1741 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1742 struct io_submit_state *state)
1744 spin_lock(&ctx->completion_lock);
1745 list_splice_init(&ctx->locked_free_list, &state->free_list);
1746 ctx->locked_free_nr = 0;
1747 spin_unlock(&ctx->completion_lock);
1750 /* Returns true IFF there are requests in the cache */
1751 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1753 struct io_submit_state *state = &ctx->submit_state;
1757 * If we have more than a batch's worth of requests in our IRQ side
1758 * locked cache, grab the lock and move them over to our submission
1761 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1762 io_flush_cached_locked_reqs(ctx, state);
1764 nr = state->free_reqs;
1765 while (!list_empty(&state->free_list)) {
1766 struct io_kiocb *req = list_first_entry(&state->free_list,
1767 struct io_kiocb, inflight_entry);
1769 list_del(&req->inflight_entry);
1770 state->reqs[nr++] = req;
1771 if (nr == ARRAY_SIZE(state->reqs))
1775 state->free_reqs = nr;
1780 * A request might get retired back into the request caches even before opcode
1781 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1782 * Because of that, io_alloc_req() should be called only under ->uring_lock
1783 * and with extra caution to not get a request that is still worked on.
1785 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1786 __must_hold(&ctx->uring_lock)
1788 struct io_submit_state *state = &ctx->submit_state;
1789 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1792 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1794 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1797 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1801 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1802 * retry single alloc to be on the safe side.
1804 if (unlikely(ret <= 0)) {
1805 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1806 if (!state->reqs[0])
1811 for (i = 0; i < ret; i++)
1812 io_preinit_req(state->reqs[i], ctx);
1813 state->free_reqs = ret;
1816 return state->reqs[state->free_reqs];
1819 static inline void io_put_file(struct file *file)
1825 static void io_dismantle_req(struct io_kiocb *req)
1827 unsigned int flags = req->flags;
1829 if (io_req_needs_clean(req))
1831 if (!(flags & REQ_F_FIXED_FILE))
1832 io_put_file(req->file);
1833 if (req->fixed_rsrc_refs)
1834 percpu_ref_put(req->fixed_rsrc_refs);
1835 if (req->async_data) {
1836 kfree(req->async_data);
1837 req->async_data = NULL;
1841 static void __io_free_req(struct io_kiocb *req)
1843 struct io_ring_ctx *ctx = req->ctx;
1845 io_dismantle_req(req);
1846 io_put_task(req->task, 1);
1848 spin_lock(&ctx->completion_lock);
1849 list_add(&req->inflight_entry, &ctx->locked_free_list);
1850 ctx->locked_free_nr++;
1851 spin_unlock(&ctx->completion_lock);
1853 percpu_ref_put(&ctx->refs);
1856 static inline void io_remove_next_linked(struct io_kiocb *req)
1858 struct io_kiocb *nxt = req->link;
1860 req->link = nxt->link;
1864 static bool io_kill_linked_timeout(struct io_kiocb *req)
1865 __must_hold(&req->ctx->completion_lock)
1866 __must_hold(&req->ctx->timeout_lock)
1868 struct io_kiocb *link = req->link;
1871 * Can happen if a linked timeout fired and link had been like
1872 * req -> link t-out -> link t-out [-> ...]
1874 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1875 struct io_timeout_data *io = link->async_data;
1877 io_remove_next_linked(req);
1878 link->timeout.head = NULL;
1879 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1880 io_cqring_fill_event(link->ctx, link->user_data,
1882 io_put_req_deferred(link);
1889 static void io_fail_links(struct io_kiocb *req)
1890 __must_hold(&req->ctx->completion_lock)
1892 struct io_kiocb *nxt, *link = req->link;
1899 trace_io_uring_fail_link(req, link);
1900 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1901 io_put_req_deferred(link);
1906 static bool io_disarm_next(struct io_kiocb *req)
1907 __must_hold(&req->ctx->completion_lock)
1909 bool posted = false;
1911 if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1912 struct io_ring_ctx *ctx = req->ctx;
1914 spin_lock_irq(&ctx->timeout_lock);
1915 posted = io_kill_linked_timeout(req);
1916 spin_unlock_irq(&ctx->timeout_lock);
1918 if (unlikely((req->flags & REQ_F_FAIL) &&
1919 !(req->flags & REQ_F_HARDLINK))) {
1920 posted |= (req->link != NULL);
1926 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1928 struct io_kiocb *nxt;
1931 * If LINK is set, we have dependent requests in this chain. If we
1932 * didn't fail this request, queue the first one up, moving any other
1933 * dependencies to the next request. In case of failure, fail the rest
1936 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1937 struct io_ring_ctx *ctx = req->ctx;
1940 spin_lock(&ctx->completion_lock);
1941 posted = io_disarm_next(req);
1943 io_commit_cqring(req->ctx);
1944 spin_unlock(&ctx->completion_lock);
1946 io_cqring_ev_posted(ctx);
1953 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1955 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1957 return __io_req_find_next(req);
1960 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1964 if (ctx->submit_state.compl_nr) {
1965 mutex_lock(&ctx->uring_lock);
1966 io_submit_flush_completions(ctx);
1967 mutex_unlock(&ctx->uring_lock);
1969 percpu_ref_put(&ctx->refs);
1972 static void tctx_task_work(struct callback_head *cb)
1974 struct io_ring_ctx *ctx = NULL;
1975 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1979 struct io_wq_work_node *node;
1981 spin_lock_irq(&tctx->task_lock);
1982 node = tctx->task_list.first;
1983 INIT_WQ_LIST(&tctx->task_list);
1985 tctx->task_running = false;
1986 spin_unlock_irq(&tctx->task_lock);
1991 struct io_wq_work_node *next = node->next;
1992 struct io_kiocb *req = container_of(node, struct io_kiocb,
1995 if (req->ctx != ctx) {
1996 ctx_flush_and_put(ctx);
1998 percpu_ref_get(&ctx->refs);
2000 req->io_task_work.func(req);
2007 ctx_flush_and_put(ctx);
2010 static void io_req_task_work_add(struct io_kiocb *req)
2012 struct task_struct *tsk = req->task;
2013 struct io_uring_task *tctx = tsk->io_uring;
2014 enum task_work_notify_mode notify;
2015 struct io_wq_work_node *node;
2016 unsigned long flags;
2019 WARN_ON_ONCE(!tctx);
2021 spin_lock_irqsave(&tctx->task_lock, flags);
2022 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2023 running = tctx->task_running;
2025 tctx->task_running = true;
2026 spin_unlock_irqrestore(&tctx->task_lock, flags);
2028 /* task_work already pending, we're done */
2033 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2034 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2035 * processing task_work. There's no reliable way to tell if TWA_RESUME
2038 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2039 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2040 wake_up_process(tsk);
2044 spin_lock_irqsave(&tctx->task_lock, flags);
2045 tctx->task_running = false;
2046 node = tctx->task_list.first;
2047 INIT_WQ_LIST(&tctx->task_list);
2048 spin_unlock_irqrestore(&tctx->task_lock, flags);
2051 req = container_of(node, struct io_kiocb, io_task_work.node);
2053 if (llist_add(&req->io_task_work.fallback_node,
2054 &req->ctx->fallback_llist))
2055 schedule_delayed_work(&req->ctx->fallback_work, 1);
2059 static void io_req_task_cancel(struct io_kiocb *req)
2061 struct io_ring_ctx *ctx = req->ctx;
2063 /* ctx is guaranteed to stay alive while we hold uring_lock */
2064 mutex_lock(&ctx->uring_lock);
2065 io_req_complete_failed(req, req->result);
2066 mutex_unlock(&ctx->uring_lock);
2069 static void io_req_task_submit(struct io_kiocb *req)
2071 struct io_ring_ctx *ctx = req->ctx;
2073 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2074 mutex_lock(&ctx->uring_lock);
2075 if (likely(!(req->task->flags & PF_EXITING)))
2076 __io_queue_sqe(req);
2078 io_req_complete_failed(req, -EFAULT);
2079 mutex_unlock(&ctx->uring_lock);
2082 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2085 req->io_task_work.func = io_req_task_cancel;
2086 io_req_task_work_add(req);
2089 static void io_req_task_queue(struct io_kiocb *req)
2091 req->io_task_work.func = io_req_task_submit;
2092 io_req_task_work_add(req);
2095 static void io_req_task_queue_reissue(struct io_kiocb *req)
2097 req->io_task_work.func = io_queue_async_work;
2098 io_req_task_work_add(req);
2101 static inline void io_queue_next(struct io_kiocb *req)
2103 struct io_kiocb *nxt = io_req_find_next(req);
2106 io_req_task_queue(nxt);
2109 static void io_free_req(struct io_kiocb *req)
2116 struct task_struct *task;
2121 static inline void io_init_req_batch(struct req_batch *rb)
2128 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2129 struct req_batch *rb)
2132 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2133 if (rb->task == current)
2134 current->io_uring->cached_refs += rb->task_refs;
2136 io_put_task(rb->task, rb->task_refs);
2139 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2140 struct io_submit_state *state)
2143 io_dismantle_req(req);
2145 if (req->task != rb->task) {
2147 io_put_task(rb->task, rb->task_refs);
2148 rb->task = req->task;
2154 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2155 state->reqs[state->free_reqs++] = req;
2157 list_add(&req->inflight_entry, &state->free_list);
2160 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2161 __must_hold(&req->ctx->uring_lock)
2163 struct io_submit_state *state = &ctx->submit_state;
2164 int i, nr = state->compl_nr;
2165 struct req_batch rb;
2167 spin_lock(&ctx->completion_lock);
2168 for (i = 0; i < nr; i++) {
2169 struct io_kiocb *req = state->compl_reqs[i];
2171 __io_cqring_fill_event(ctx, req->user_data, req->result,
2174 io_commit_cqring(ctx);
2175 spin_unlock(&ctx->completion_lock);
2176 io_cqring_ev_posted(ctx);
2178 io_init_req_batch(&rb);
2179 for (i = 0; i < nr; i++) {
2180 struct io_kiocb *req = state->compl_reqs[i];
2182 if (req_ref_put_and_test(req))
2183 io_req_free_batch(&rb, req, &ctx->submit_state);
2186 io_req_free_batch_finish(ctx, &rb);
2187 state->compl_nr = 0;
2191 * Drop reference to request, return next in chain (if there is one) if this
2192 * was the last reference to this request.
2194 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2196 struct io_kiocb *nxt = NULL;
2198 if (req_ref_put_and_test(req)) {
2199 nxt = io_req_find_next(req);
2205 static inline void io_put_req(struct io_kiocb *req)
2207 if (req_ref_put_and_test(req))
2211 static inline void io_put_req_deferred(struct io_kiocb *req)
2213 if (req_ref_put_and_test(req)) {
2214 req->io_task_work.func = io_free_req;
2215 io_req_task_work_add(req);
2219 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2221 /* See comment at the top of this file */
2223 return __io_cqring_events(ctx);
2226 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2228 struct io_rings *rings = ctx->rings;
2230 /* make sure SQ entry isn't read before tail */
2231 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2234 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2236 unsigned int cflags;
2238 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2239 cflags |= IORING_CQE_F_BUFFER;
2240 req->flags &= ~REQ_F_BUFFER_SELECTED;
2245 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2247 struct io_buffer *kbuf;
2249 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2250 return io_put_kbuf(req, kbuf);
2253 static inline bool io_run_task_work(void)
2255 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2256 __set_current_state(TASK_RUNNING);
2257 tracehook_notify_signal();
2265 * Find and free completed poll iocbs
2267 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2268 struct list_head *done, bool resubmit)
2270 struct req_batch rb;
2271 struct io_kiocb *req;
2273 /* order with ->result store in io_complete_rw_iopoll() */
2276 io_init_req_batch(&rb);
2277 while (!list_empty(done)) {
2280 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2281 list_del(&req->inflight_entry);
2283 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2284 !(req->flags & REQ_F_DONT_REISSUE)) {
2285 req->iopoll_completed = 0;
2286 io_req_task_queue_reissue(req);
2290 if (req->flags & REQ_F_BUFFER_SELECTED)
2291 cflags = io_put_rw_kbuf(req);
2293 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2296 if (req_ref_put_and_test(req))
2297 io_req_free_batch(&rb, req, &ctx->submit_state);
2300 io_commit_cqring(ctx);
2301 io_cqring_ev_posted_iopoll(ctx);
2302 io_req_free_batch_finish(ctx, &rb);
2305 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2306 long min, bool resubmit)
2308 struct io_kiocb *req, *tmp;
2313 * Only spin for completions if we don't have multiple devices hanging
2314 * off our complete list, and we're under the requested amount.
2316 spin = !ctx->poll_multi_queue && *nr_events < min;
2318 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2319 struct kiocb *kiocb = &req->rw.kiocb;
2323 * Move completed and retryable entries to our local lists.
2324 * If we find a request that requires polling, break out
2325 * and complete those lists first, if we have entries there.
2327 if (READ_ONCE(req->iopoll_completed)) {
2328 list_move_tail(&req->inflight_entry, &done);
2331 if (!list_empty(&done))
2334 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2335 if (unlikely(ret < 0))
2340 /* iopoll may have completed current req */
2341 if (READ_ONCE(req->iopoll_completed))
2342 list_move_tail(&req->inflight_entry, &done);
2345 if (!list_empty(&done))
2346 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2352 * We can't just wait for polled events to come to us, we have to actively
2353 * find and complete them.
2355 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2357 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2360 mutex_lock(&ctx->uring_lock);
2361 while (!list_empty(&ctx->iopoll_list)) {
2362 unsigned int nr_events = 0;
2364 io_do_iopoll(ctx, &nr_events, 0, false);
2366 /* let it sleep and repeat later if can't complete a request */
2370 * Ensure we allow local-to-the-cpu processing to take place,
2371 * in this case we need to ensure that we reap all events.
2372 * Also let task_work, etc. to progress by releasing the mutex
2374 if (need_resched()) {
2375 mutex_unlock(&ctx->uring_lock);
2377 mutex_lock(&ctx->uring_lock);
2380 mutex_unlock(&ctx->uring_lock);
2383 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2385 unsigned int nr_events = 0;
2389 * We disallow the app entering submit/complete with polling, but we
2390 * still need to lock the ring to prevent racing with polled issue
2391 * that got punted to a workqueue.
2393 mutex_lock(&ctx->uring_lock);
2395 * Don't enter poll loop if we already have events pending.
2396 * If we do, we can potentially be spinning for commands that
2397 * already triggered a CQE (eg in error).
2399 if (test_bit(0, &ctx->check_cq_overflow))
2400 __io_cqring_overflow_flush(ctx, false);
2401 if (io_cqring_events(ctx))
2405 * If a submit got punted to a workqueue, we can have the
2406 * application entering polling for a command before it gets
2407 * issued. That app will hold the uring_lock for the duration
2408 * of the poll right here, so we need to take a breather every
2409 * now and then to ensure that the issue has a chance to add
2410 * the poll to the issued list. Otherwise we can spin here
2411 * forever, while the workqueue is stuck trying to acquire the
2414 if (list_empty(&ctx->iopoll_list)) {
2415 u32 tail = ctx->cached_cq_tail;
2417 mutex_unlock(&ctx->uring_lock);
2419 mutex_lock(&ctx->uring_lock);
2421 /* some requests don't go through iopoll_list */
2422 if (tail != ctx->cached_cq_tail ||
2423 list_empty(&ctx->iopoll_list))
2426 ret = io_do_iopoll(ctx, &nr_events, min, true);
2427 } while (!ret && nr_events < min && !need_resched());
2429 mutex_unlock(&ctx->uring_lock);
2433 static void kiocb_end_write(struct io_kiocb *req)
2436 * Tell lockdep we inherited freeze protection from submission
2439 if (req->flags & REQ_F_ISREG) {
2440 struct super_block *sb = file_inode(req->file)->i_sb;
2442 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2448 static bool io_resubmit_prep(struct io_kiocb *req)
2450 struct io_async_rw *rw = req->async_data;
2453 return !io_req_prep_async(req);
2454 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2455 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2459 static bool io_rw_should_reissue(struct io_kiocb *req)
2461 umode_t mode = file_inode(req->file)->i_mode;
2462 struct io_ring_ctx *ctx = req->ctx;
2464 if (!S_ISBLK(mode) && !S_ISREG(mode))
2466 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2467 !(ctx->flags & IORING_SETUP_IOPOLL)))
2470 * If ref is dying, we might be running poll reap from the exit work.
2471 * Don't attempt to reissue from that path, just let it fail with
2474 if (percpu_ref_is_dying(&ctx->refs))
2477 * Play it safe and assume not safe to re-import and reissue if we're
2478 * not in the original thread group (or in task context).
2480 if (!same_thread_group(req->task, current) || !in_task())
2485 static bool io_resubmit_prep(struct io_kiocb *req)
2489 static bool io_rw_should_reissue(struct io_kiocb *req)
2495 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2497 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2498 kiocb_end_write(req);
2499 if (res != req->result) {
2500 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2501 io_rw_should_reissue(req)) {
2502 req->flags |= REQ_F_REISSUE;
2511 static void io_req_task_complete(struct io_kiocb *req)
2515 if (req->flags & REQ_F_BUFFER_SELECTED)
2516 cflags = io_put_rw_kbuf(req);
2517 __io_req_complete(req, 0, req->result, cflags);
2520 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2521 unsigned int issue_flags)
2523 if (__io_complete_rw_common(req, res))
2525 io_req_task_complete(req);
2528 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2530 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2532 if (__io_complete_rw_common(req, res))
2535 req->io_task_work.func = io_req_task_complete;
2536 io_req_task_work_add(req);
2539 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2541 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2543 if (kiocb->ki_flags & IOCB_WRITE)
2544 kiocb_end_write(req);
2545 if (unlikely(res != req->result)) {
2546 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2547 io_resubmit_prep(req))) {
2549 req->flags |= REQ_F_DONT_REISSUE;
2553 WRITE_ONCE(req->result, res);
2554 /* order with io_iopoll_complete() checking ->result */
2556 WRITE_ONCE(req->iopoll_completed, 1);
2560 * After the iocb has been issued, it's safe to be found on the poll list.
2561 * Adding the kiocb to the list AFTER submission ensures that we don't
2562 * find it from a io_do_iopoll() thread before the issuer is done
2563 * accessing the kiocb cookie.
2565 static void io_iopoll_req_issued(struct io_kiocb *req)
2567 struct io_ring_ctx *ctx = req->ctx;
2568 const bool in_async = io_wq_current_is_worker();
2570 /* workqueue context doesn't hold uring_lock, grab it now */
2571 if (unlikely(in_async))
2572 mutex_lock(&ctx->uring_lock);
2575 * Track whether we have multiple files in our lists. This will impact
2576 * how we do polling eventually, not spinning if we're on potentially
2577 * different devices.
2579 if (list_empty(&ctx->iopoll_list)) {
2580 ctx->poll_multi_queue = false;
2581 } else if (!ctx->poll_multi_queue) {
2582 struct io_kiocb *list_req;
2583 unsigned int queue_num0, queue_num1;
2585 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2588 if (list_req->file != req->file) {
2589 ctx->poll_multi_queue = true;
2591 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2592 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2593 if (queue_num0 != queue_num1)
2594 ctx->poll_multi_queue = true;
2599 * For fast devices, IO may have already completed. If it has, add
2600 * it to the front so we find it first.
2602 if (READ_ONCE(req->iopoll_completed))
2603 list_add(&req->inflight_entry, &ctx->iopoll_list);
2605 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2607 if (unlikely(in_async)) {
2609 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2610 * in sq thread task context or in io worker task context. If
2611 * current task context is sq thread, we don't need to check
2612 * whether should wake up sq thread.
2614 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2615 wq_has_sleeper(&ctx->sq_data->wait))
2616 wake_up(&ctx->sq_data->wait);
2618 mutex_unlock(&ctx->uring_lock);
2622 static bool io_bdev_nowait(struct block_device *bdev)
2624 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2628 * If we tracked the file through the SCM inflight mechanism, we could support
2629 * any file. For now, just ensure that anything potentially problematic is done
2632 static bool __io_file_supports_nowait(struct file *file, int rw)
2634 umode_t mode = file_inode(file)->i_mode;
2636 if (S_ISBLK(mode)) {
2637 if (IS_ENABLED(CONFIG_BLOCK) &&
2638 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2644 if (S_ISREG(mode)) {
2645 if (IS_ENABLED(CONFIG_BLOCK) &&
2646 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2647 file->f_op != &io_uring_fops)
2652 /* any ->read/write should understand O_NONBLOCK */
2653 if (file->f_flags & O_NONBLOCK)
2656 if (!(file->f_mode & FMODE_NOWAIT))
2660 return file->f_op->read_iter != NULL;
2662 return file->f_op->write_iter != NULL;
2665 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2667 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2669 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2672 return __io_file_supports_nowait(req->file, rw);
2675 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2677 struct io_ring_ctx *ctx = req->ctx;
2678 struct kiocb *kiocb = &req->rw.kiocb;
2679 struct file *file = req->file;
2683 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2684 req->flags |= REQ_F_ISREG;
2686 kiocb->ki_pos = READ_ONCE(sqe->off);
2687 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2688 req->flags |= REQ_F_CUR_POS;
2689 kiocb->ki_pos = file->f_pos;
2691 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2692 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2693 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2697 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2698 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2699 req->flags |= REQ_F_NOWAIT;
2701 ioprio = READ_ONCE(sqe->ioprio);
2703 ret = ioprio_check_cap(ioprio);
2707 kiocb->ki_ioprio = ioprio;
2709 kiocb->ki_ioprio = get_current_ioprio();
2711 if (ctx->flags & IORING_SETUP_IOPOLL) {
2712 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2713 !kiocb->ki_filp->f_op->iopoll)
2716 kiocb->ki_flags |= IOCB_HIPRI;
2717 kiocb->ki_complete = io_complete_rw_iopoll;
2718 req->iopoll_completed = 0;
2720 if (kiocb->ki_flags & IOCB_HIPRI)
2722 kiocb->ki_complete = io_complete_rw;
2725 if (req->opcode == IORING_OP_READ_FIXED ||
2726 req->opcode == IORING_OP_WRITE_FIXED) {
2728 io_req_set_rsrc_node(req);
2731 req->rw.addr = READ_ONCE(sqe->addr);
2732 req->rw.len = READ_ONCE(sqe->len);
2733 req->buf_index = READ_ONCE(sqe->buf_index);
2737 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2743 case -ERESTARTNOINTR:
2744 case -ERESTARTNOHAND:
2745 case -ERESTART_RESTARTBLOCK:
2747 * We can't just restart the syscall, since previously
2748 * submitted sqes may already be in progress. Just fail this
2754 kiocb->ki_complete(kiocb, ret, 0);
2758 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2759 unsigned int issue_flags)
2761 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2762 struct io_async_rw *io = req->async_data;
2763 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2765 /* add previously done IO, if any */
2766 if (io && io->bytes_done > 0) {
2768 ret = io->bytes_done;
2770 ret += io->bytes_done;
2773 if (req->flags & REQ_F_CUR_POS)
2774 req->file->f_pos = kiocb->ki_pos;
2775 if (ret >= 0 && check_reissue)
2776 __io_complete_rw(req, ret, 0, issue_flags);
2778 io_rw_done(kiocb, ret);
2780 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2781 req->flags &= ~REQ_F_REISSUE;
2782 if (io_resubmit_prep(req)) {
2783 io_req_task_queue_reissue(req);
2788 if (req->flags & REQ_F_BUFFER_SELECTED)
2789 cflags = io_put_rw_kbuf(req);
2790 __io_req_complete(req, issue_flags, ret, cflags);
2795 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2796 struct io_mapped_ubuf *imu)
2798 size_t len = req->rw.len;
2799 u64 buf_end, buf_addr = req->rw.addr;
2802 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2804 /* not inside the mapped region */
2805 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2809 * May not be a start of buffer, set size appropriately
2810 * and advance us to the beginning.
2812 offset = buf_addr - imu->ubuf;
2813 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2817 * Don't use iov_iter_advance() here, as it's really slow for
2818 * using the latter parts of a big fixed buffer - it iterates
2819 * over each segment manually. We can cheat a bit here, because
2822 * 1) it's a BVEC iter, we set it up
2823 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2824 * first and last bvec
2826 * So just find our index, and adjust the iterator afterwards.
2827 * If the offset is within the first bvec (or the whole first
2828 * bvec, just use iov_iter_advance(). This makes it easier
2829 * since we can just skip the first segment, which may not
2830 * be PAGE_SIZE aligned.
2832 const struct bio_vec *bvec = imu->bvec;
2834 if (offset <= bvec->bv_len) {
2835 iov_iter_advance(iter, offset);
2837 unsigned long seg_skip;
2839 /* skip first vec */
2840 offset -= bvec->bv_len;
2841 seg_skip = 1 + (offset >> PAGE_SHIFT);
2843 iter->bvec = bvec + seg_skip;
2844 iter->nr_segs -= seg_skip;
2845 iter->count -= bvec->bv_len + offset;
2846 iter->iov_offset = offset & ~PAGE_MASK;
2853 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2855 struct io_ring_ctx *ctx = req->ctx;
2856 struct io_mapped_ubuf *imu = req->imu;
2857 u16 index, buf_index = req->buf_index;
2860 if (unlikely(buf_index >= ctx->nr_user_bufs))
2862 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2863 imu = READ_ONCE(ctx->user_bufs[index]);
2866 return __io_import_fixed(req, rw, iter, imu);
2869 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2872 mutex_unlock(&ctx->uring_lock);
2875 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2878 * "Normal" inline submissions always hold the uring_lock, since we
2879 * grab it from the system call. Same is true for the SQPOLL offload.
2880 * The only exception is when we've detached the request and issue it
2881 * from an async worker thread, grab the lock for that case.
2884 mutex_lock(&ctx->uring_lock);
2887 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2888 int bgid, struct io_buffer *kbuf,
2891 struct io_buffer *head;
2893 if (req->flags & REQ_F_BUFFER_SELECTED)
2896 io_ring_submit_lock(req->ctx, needs_lock);
2898 lockdep_assert_held(&req->ctx->uring_lock);
2900 head = xa_load(&req->ctx->io_buffers, bgid);
2902 if (!list_empty(&head->list)) {
2903 kbuf = list_last_entry(&head->list, struct io_buffer,
2905 list_del(&kbuf->list);
2908 xa_erase(&req->ctx->io_buffers, bgid);
2910 if (*len > kbuf->len)
2913 kbuf = ERR_PTR(-ENOBUFS);
2916 io_ring_submit_unlock(req->ctx, needs_lock);
2921 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2924 struct io_buffer *kbuf;
2927 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2928 bgid = req->buf_index;
2929 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2932 req->rw.addr = (u64) (unsigned long) kbuf;
2933 req->flags |= REQ_F_BUFFER_SELECTED;
2934 return u64_to_user_ptr(kbuf->addr);
2937 #ifdef CONFIG_COMPAT
2938 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2941 struct compat_iovec __user *uiov;
2942 compat_ssize_t clen;
2946 uiov = u64_to_user_ptr(req->rw.addr);
2947 if (!access_ok(uiov, sizeof(*uiov)))
2949 if (__get_user(clen, &uiov->iov_len))
2955 buf = io_rw_buffer_select(req, &len, needs_lock);
2957 return PTR_ERR(buf);
2958 iov[0].iov_base = buf;
2959 iov[0].iov_len = (compat_size_t) len;
2964 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2967 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2971 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2974 len = iov[0].iov_len;
2977 buf = io_rw_buffer_select(req, &len, needs_lock);
2979 return PTR_ERR(buf);
2980 iov[0].iov_base = buf;
2981 iov[0].iov_len = len;
2985 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2988 if (req->flags & REQ_F_BUFFER_SELECTED) {
2989 struct io_buffer *kbuf;
2991 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2992 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2993 iov[0].iov_len = kbuf->len;
2996 if (req->rw.len != 1)
2999 #ifdef CONFIG_COMPAT
3000 if (req->ctx->compat)
3001 return io_compat_import(req, iov, needs_lock);
3004 return __io_iov_buffer_select(req, iov, needs_lock);
3007 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3008 struct iov_iter *iter, bool needs_lock)
3010 void __user *buf = u64_to_user_ptr(req->rw.addr);
3011 size_t sqe_len = req->rw.len;
3012 u8 opcode = req->opcode;
3015 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3017 return io_import_fixed(req, rw, iter);
3020 /* buffer index only valid with fixed read/write, or buffer select */
3021 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3024 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3025 if (req->flags & REQ_F_BUFFER_SELECT) {
3026 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3028 return PTR_ERR(buf);
3029 req->rw.len = sqe_len;
3032 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3037 if (req->flags & REQ_F_BUFFER_SELECT) {
3038 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3040 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3045 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3049 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3051 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3055 * For files that don't have ->read_iter() and ->write_iter(), handle them
3056 * by looping over ->read() or ->write() manually.
3058 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3060 struct kiocb *kiocb = &req->rw.kiocb;
3061 struct file *file = req->file;
3065 * Don't support polled IO through this interface, and we can't
3066 * support non-blocking either. For the latter, this just causes
3067 * the kiocb to be handled from an async context.
3069 if (kiocb->ki_flags & IOCB_HIPRI)
3071 if (kiocb->ki_flags & IOCB_NOWAIT)
3074 while (iov_iter_count(iter)) {
3078 if (!iov_iter_is_bvec(iter)) {
3079 iovec = iov_iter_iovec(iter);
3081 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3082 iovec.iov_len = req->rw.len;
3086 nr = file->f_op->read(file, iovec.iov_base,
3087 iovec.iov_len, io_kiocb_ppos(kiocb));
3089 nr = file->f_op->write(file, iovec.iov_base,
3090 iovec.iov_len, io_kiocb_ppos(kiocb));
3099 if (nr != iovec.iov_len)
3103 iov_iter_advance(iter, nr);
3109 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3110 const struct iovec *fast_iov, struct iov_iter *iter)
3112 struct io_async_rw *rw = req->async_data;
3114 memcpy(&rw->iter, iter, sizeof(*iter));
3115 rw->free_iovec = iovec;
3117 /* can only be fixed buffers, no need to do anything */
3118 if (iov_iter_is_bvec(iter))
3121 unsigned iov_off = 0;
3123 rw->iter.iov = rw->fast_iov;
3124 if (iter->iov != fast_iov) {
3125 iov_off = iter->iov - fast_iov;
3126 rw->iter.iov += iov_off;
3128 if (rw->fast_iov != fast_iov)
3129 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3130 sizeof(struct iovec) * iter->nr_segs);
3132 req->flags |= REQ_F_NEED_CLEANUP;
3136 static inline int io_alloc_async_data(struct io_kiocb *req)
3138 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3139 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3140 return req->async_data == NULL;
3143 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3144 const struct iovec *fast_iov,
3145 struct iov_iter *iter, bool force)
3147 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3149 if (!req->async_data) {
3150 if (io_alloc_async_data(req)) {
3155 io_req_map_rw(req, iovec, fast_iov, iter);
3160 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3162 struct io_async_rw *iorw = req->async_data;
3163 struct iovec *iov = iorw->fast_iov;
3166 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3167 if (unlikely(ret < 0))
3170 iorw->bytes_done = 0;
3171 iorw->free_iovec = iov;
3173 req->flags |= REQ_F_NEED_CLEANUP;
3177 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3179 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3181 return io_prep_rw(req, sqe);
3185 * This is our waitqueue callback handler, registered through lock_page_async()
3186 * when we initially tried to do the IO with the iocb armed our waitqueue.
3187 * This gets called when the page is unlocked, and we generally expect that to
3188 * happen when the page IO is completed and the page is now uptodate. This will
3189 * queue a task_work based retry of the operation, attempting to copy the data
3190 * again. If the latter fails because the page was NOT uptodate, then we will
3191 * do a thread based blocking retry of the operation. That's the unexpected
3194 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3195 int sync, void *arg)
3197 struct wait_page_queue *wpq;
3198 struct io_kiocb *req = wait->private;
3199 struct wait_page_key *key = arg;
3201 wpq = container_of(wait, struct wait_page_queue, wait);
3203 if (!wake_page_match(wpq, key))
3206 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3207 list_del_init(&wait->entry);
3208 io_req_task_queue(req);
3213 * This controls whether a given IO request should be armed for async page
3214 * based retry. If we return false here, the request is handed to the async
3215 * worker threads for retry. If we're doing buffered reads on a regular file,
3216 * we prepare a private wait_page_queue entry and retry the operation. This
3217 * will either succeed because the page is now uptodate and unlocked, or it
3218 * will register a callback when the page is unlocked at IO completion. Through
3219 * that callback, io_uring uses task_work to setup a retry of the operation.
3220 * That retry will attempt the buffered read again. The retry will generally
3221 * succeed, or in rare cases where it fails, we then fall back to using the
3222 * async worker threads for a blocking retry.
3224 static bool io_rw_should_retry(struct io_kiocb *req)
3226 struct io_async_rw *rw = req->async_data;
3227 struct wait_page_queue *wait = &rw->wpq;
3228 struct kiocb *kiocb = &req->rw.kiocb;
3230 /* never retry for NOWAIT, we just complete with -EAGAIN */
3231 if (req->flags & REQ_F_NOWAIT)
3234 /* Only for buffered IO */
3235 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3239 * just use poll if we can, and don't attempt if the fs doesn't
3240 * support callback based unlocks
3242 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3245 wait->wait.func = io_async_buf_func;
3246 wait->wait.private = req;
3247 wait->wait.flags = 0;
3248 INIT_LIST_HEAD(&wait->wait.entry);
3249 kiocb->ki_flags |= IOCB_WAITQ;
3250 kiocb->ki_flags &= ~IOCB_NOWAIT;
3251 kiocb->ki_waitq = wait;
3255 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3257 if (req->file->f_op->read_iter)
3258 return call_read_iter(req->file, &req->rw.kiocb, iter);
3259 else if (req->file->f_op->read)
3260 return loop_rw_iter(READ, req, iter);
3265 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3267 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct iov_iter __iter, *iter = &__iter;
3270 struct io_async_rw *rw = req->async_data;
3271 ssize_t io_size, ret, ret2;
3272 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3278 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3282 io_size = iov_iter_count(iter);
3283 req->result = io_size;
3285 /* Ensure we clear previously set non-block flag */
3286 if (!force_nonblock)
3287 kiocb->ki_flags &= ~IOCB_NOWAIT;
3289 kiocb->ki_flags |= IOCB_NOWAIT;
3291 /* If the file doesn't support async, just async punt */
3292 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3293 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3294 return ret ?: -EAGAIN;
3297 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3298 if (unlikely(ret)) {
3303 ret = io_iter_do_read(req, iter);
3305 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3306 req->flags &= ~REQ_F_REISSUE;
3307 /* IOPOLL retry should happen for io-wq threads */
3308 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3310 /* no retry on NONBLOCK nor RWF_NOWAIT */
3311 if (req->flags & REQ_F_NOWAIT)
3313 /* some cases will consume bytes even on error returns */
3314 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3316 } else if (ret == -EIOCBQUEUED) {
3318 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3319 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3320 /* read all, failed, already did sync or don't want to retry */
3324 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3329 rw = req->async_data;
3330 /* now use our persistent iterator, if we aren't already */
3335 rw->bytes_done += ret;
3336 /* if we can retry, do so with the callbacks armed */
3337 if (!io_rw_should_retry(req)) {
3338 kiocb->ki_flags &= ~IOCB_WAITQ;
3343 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3344 * we get -EIOCBQUEUED, then we'll get a notification when the
3345 * desired page gets unlocked. We can also get a partial read
3346 * here, and if we do, then just retry at the new offset.
3348 ret = io_iter_do_read(req, iter);
3349 if (ret == -EIOCBQUEUED)
3351 /* we got some bytes, but not all. retry. */
3352 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 } while (ret > 0 && ret < io_size);
3355 kiocb_done(kiocb, ret, issue_flags);
3357 /* it's faster to check here then delegate to kfree */
3363 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3365 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3367 return io_prep_rw(req, sqe);
3370 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3372 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3373 struct kiocb *kiocb = &req->rw.kiocb;
3374 struct iov_iter __iter, *iter = &__iter;
3375 struct io_async_rw *rw = req->async_data;
3376 ssize_t ret, ret2, io_size;
3377 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3383 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3387 io_size = iov_iter_count(iter);
3388 req->result = io_size;
3390 /* Ensure we clear previously set non-block flag */
3391 if (!force_nonblock)
3392 kiocb->ki_flags &= ~IOCB_NOWAIT;
3394 kiocb->ki_flags |= IOCB_NOWAIT;
3396 /* If the file doesn't support async, just async punt */
3397 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3400 /* file path doesn't support NOWAIT for non-direct_IO */
3401 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3402 (req->flags & REQ_F_ISREG))
3405 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3410 * Open-code file_start_write here to grab freeze protection,
3411 * which will be released by another thread in
3412 * io_complete_rw(). Fool lockdep by telling it the lock got
3413 * released so that it doesn't complain about the held lock when
3414 * we return to userspace.
3416 if (req->flags & REQ_F_ISREG) {
3417 sb_start_write(file_inode(req->file)->i_sb);
3418 __sb_writers_release(file_inode(req->file)->i_sb,
3421 kiocb->ki_flags |= IOCB_WRITE;
3423 if (req->file->f_op->write_iter)
3424 ret2 = call_write_iter(req->file, kiocb, iter);
3425 else if (req->file->f_op->write)
3426 ret2 = loop_rw_iter(WRITE, req, iter);
3430 if (req->flags & REQ_F_REISSUE) {
3431 req->flags &= ~REQ_F_REISSUE;
3436 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3437 * retry them without IOCB_NOWAIT.
3439 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3441 /* no retry on NONBLOCK nor RWF_NOWAIT */
3442 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3444 if (!force_nonblock || ret2 != -EAGAIN) {
3445 /* IOPOLL retry should happen for io-wq threads */
3446 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3449 kiocb_done(kiocb, ret2, issue_flags);
3452 /* some cases will consume bytes even on error returns */
3453 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3454 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3455 return ret ?: -EAGAIN;
3458 /* it's reportedly faster than delegating the null check to kfree() */
3464 static int io_renameat_prep(struct io_kiocb *req,
3465 const struct io_uring_sqe *sqe)
3467 struct io_rename *ren = &req->rename;
3468 const char __user *oldf, *newf;
3470 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3472 if (sqe->ioprio || sqe->buf_index)
3474 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3477 ren->old_dfd = READ_ONCE(sqe->fd);
3478 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3479 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3480 ren->new_dfd = READ_ONCE(sqe->len);
3481 ren->flags = READ_ONCE(sqe->rename_flags);
3483 ren->oldpath = getname(oldf);
3484 if (IS_ERR(ren->oldpath))
3485 return PTR_ERR(ren->oldpath);
3487 ren->newpath = getname(newf);
3488 if (IS_ERR(ren->newpath)) {
3489 putname(ren->oldpath);
3490 return PTR_ERR(ren->newpath);
3493 req->flags |= REQ_F_NEED_CLEANUP;
3497 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3499 struct io_rename *ren = &req->rename;
3502 if (issue_flags & IO_URING_F_NONBLOCK)
3505 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3506 ren->newpath, ren->flags);
3508 req->flags &= ~REQ_F_NEED_CLEANUP;
3511 io_req_complete(req, ret);
3515 static int io_unlinkat_prep(struct io_kiocb *req,
3516 const struct io_uring_sqe *sqe)
3518 struct io_unlink *un = &req->unlink;
3519 const char __user *fname;
3521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3523 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3525 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3528 un->dfd = READ_ONCE(sqe->fd);
3530 un->flags = READ_ONCE(sqe->unlink_flags);
3531 if (un->flags & ~AT_REMOVEDIR)
3534 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3535 un->filename = getname(fname);
3536 if (IS_ERR(un->filename))
3537 return PTR_ERR(un->filename);
3539 req->flags |= REQ_F_NEED_CLEANUP;
3543 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3545 struct io_unlink *un = &req->unlink;
3548 if (issue_flags & IO_URING_F_NONBLOCK)
3551 if (un->flags & AT_REMOVEDIR)
3552 ret = do_rmdir(un->dfd, un->filename);
3554 ret = do_unlinkat(un->dfd, un->filename);
3556 req->flags &= ~REQ_F_NEED_CLEANUP;
3559 io_req_complete(req, ret);
3563 static int io_shutdown_prep(struct io_kiocb *req,
3564 const struct io_uring_sqe *sqe)
3566 #if defined(CONFIG_NET)
3567 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3569 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3573 req->shutdown.how = READ_ONCE(sqe->len);
3580 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3582 #if defined(CONFIG_NET)
3583 struct socket *sock;
3586 if (issue_flags & IO_URING_F_NONBLOCK)
3589 sock = sock_from_file(req->file);
3590 if (unlikely(!sock))
3593 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3596 io_req_complete(req, ret);
3603 static int __io_splice_prep(struct io_kiocb *req,
3604 const struct io_uring_sqe *sqe)
3606 struct io_splice *sp = &req->splice;
3607 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3613 sp->len = READ_ONCE(sqe->len);
3614 sp->flags = READ_ONCE(sqe->splice_flags);
3616 if (unlikely(sp->flags & ~valid_flags))
3619 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3620 (sp->flags & SPLICE_F_FD_IN_FIXED));
3623 req->flags |= REQ_F_NEED_CLEANUP;
3627 static int io_tee_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3632 return __io_splice_prep(req, sqe);
3635 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3637 struct io_splice *sp = &req->splice;
3638 struct file *in = sp->file_in;
3639 struct file *out = sp->file_out;
3640 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3643 if (issue_flags & IO_URING_F_NONBLOCK)
3646 ret = do_tee(in, out, sp->len, flags);
3648 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3650 req->flags &= ~REQ_F_NEED_CLEANUP;
3654 io_req_complete(req, ret);
3658 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3660 struct io_splice *sp = &req->splice;
3662 sp->off_in = READ_ONCE(sqe->splice_off_in);
3663 sp->off_out = READ_ONCE(sqe->off);
3664 return __io_splice_prep(req, sqe);
3667 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3669 struct io_splice *sp = &req->splice;
3670 struct file *in = sp->file_in;
3671 struct file *out = sp->file_out;
3672 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3673 loff_t *poff_in, *poff_out;
3676 if (issue_flags & IO_URING_F_NONBLOCK)
3679 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3680 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3683 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3685 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3687 req->flags &= ~REQ_F_NEED_CLEANUP;
3691 io_req_complete(req, ret);
3696 * IORING_OP_NOP just posts a completion event, nothing else.
3698 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3700 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3705 __io_req_complete(req, issue_flags, 0, 0);
3709 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3711 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3718 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3721 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3722 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3725 req->sync.off = READ_ONCE(sqe->off);
3726 req->sync.len = READ_ONCE(sqe->len);
3730 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3732 loff_t end = req->sync.off + req->sync.len;
3735 /* fsync always requires a blocking context */
3736 if (issue_flags & IO_URING_F_NONBLOCK)
3739 ret = vfs_fsync_range(req->file, req->sync.off,
3740 end > 0 ? end : LLONG_MAX,
3741 req->sync.flags & IORING_FSYNC_DATASYNC);
3744 io_req_complete(req, ret);
3748 static int io_fallocate_prep(struct io_kiocb *req,
3749 const struct io_uring_sqe *sqe)
3751 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 req->sync.off = READ_ONCE(sqe->off);
3757 req->sync.len = READ_ONCE(sqe->addr);
3758 req->sync.mode = READ_ONCE(sqe->len);
3762 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3766 /* fallocate always requiring blocking context */
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3769 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3773 io_req_complete(req, ret);
3777 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 const char __user *fname;
3782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784 if (unlikely(sqe->ioprio || sqe->buf_index))
3786 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3789 /* open.how should be already initialised */
3790 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3791 req->open.how.flags |= O_LARGEFILE;
3793 req->open.dfd = READ_ONCE(sqe->fd);
3794 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3795 req->open.filename = getname(fname);
3796 if (IS_ERR(req->open.filename)) {
3797 ret = PTR_ERR(req->open.filename);
3798 req->open.filename = NULL;
3801 req->open.nofile = rlimit(RLIMIT_NOFILE);
3802 req->flags |= REQ_F_NEED_CLEANUP;
3806 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 u64 mode = READ_ONCE(sqe->len);
3809 u64 flags = READ_ONCE(sqe->open_flags);
3811 req->open.how = build_open_how(flags, mode);
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3817 struct open_how __user *how;
3821 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3822 len = READ_ONCE(sqe->len);
3823 if (len < OPEN_HOW_SIZE_VER0)
3826 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3831 return __io_openat_prep(req, sqe);
3834 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3836 struct open_flags op;
3839 bool resolve_nonblock;
3842 ret = build_open_flags(&req->open.how, &op);
3845 nonblock_set = op.open_flag & O_NONBLOCK;
3846 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3847 if (issue_flags & IO_URING_F_NONBLOCK) {
3849 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3850 * it'll always -EAGAIN
3852 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3854 op.lookup_flags |= LOOKUP_CACHED;
3855 op.open_flag |= O_NONBLOCK;
3858 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3862 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3865 * We could hang on to this 'fd' on retrying, but seems like
3866 * marginal gain for something that is now known to be a slower
3867 * path. So just put it, and we'll get a new one when we retry.
3871 ret = PTR_ERR(file);
3872 /* only retry if RESOLVE_CACHED wasn't already set by application */
3873 if (ret == -EAGAIN &&
3874 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3879 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3880 file->f_flags &= ~O_NONBLOCK;
3881 fsnotify_open(file);
3882 fd_install(ret, file);
3884 putname(req->open.filename);
3885 req->flags &= ~REQ_F_NEED_CLEANUP;
3888 __io_req_complete(req, issue_flags, ret, 0);
3892 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3894 return io_openat2(req, issue_flags);
3897 static int io_remove_buffers_prep(struct io_kiocb *req,
3898 const struct io_uring_sqe *sqe)
3900 struct io_provide_buf *p = &req->pbuf;
3903 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3906 tmp = READ_ONCE(sqe->fd);
3907 if (!tmp || tmp > USHRT_MAX)
3910 memset(p, 0, sizeof(*p));
3912 p->bgid = READ_ONCE(sqe->buf_group);
3916 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3917 int bgid, unsigned nbufs)
3921 /* shouldn't happen */
3925 /* the head kbuf is the list itself */
3926 while (!list_empty(&buf->list)) {
3927 struct io_buffer *nxt;
3929 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3930 list_del(&nxt->list);
3937 xa_erase(&ctx->io_buffers, bgid);
3942 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3944 struct io_provide_buf *p = &req->pbuf;
3945 struct io_ring_ctx *ctx = req->ctx;
3946 struct io_buffer *head;
3948 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3950 io_ring_submit_lock(ctx, !force_nonblock);
3952 lockdep_assert_held(&ctx->uring_lock);
3955 head = xa_load(&ctx->io_buffers, p->bgid);
3957 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3961 /* complete before unlock, IOPOLL may need the lock */
3962 __io_req_complete(req, issue_flags, ret, 0);
3963 io_ring_submit_unlock(ctx, !force_nonblock);
3967 static int io_provide_buffers_prep(struct io_kiocb *req,
3968 const struct io_uring_sqe *sqe)
3970 unsigned long size, tmp_check;
3971 struct io_provide_buf *p = &req->pbuf;
3974 if (sqe->ioprio || sqe->rw_flags)
3977 tmp = READ_ONCE(sqe->fd);
3978 if (!tmp || tmp > USHRT_MAX)
3981 p->addr = READ_ONCE(sqe->addr);
3982 p->len = READ_ONCE(sqe->len);
3984 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3987 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3990 size = (unsigned long)p->len * p->nbufs;
3991 if (!access_ok(u64_to_user_ptr(p->addr), size))
3994 p->bgid = READ_ONCE(sqe->buf_group);
3995 tmp = READ_ONCE(sqe->off);
3996 if (tmp > USHRT_MAX)
4002 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4004 struct io_buffer *buf;
4005 u64 addr = pbuf->addr;
4006 int i, bid = pbuf->bid;
4008 for (i = 0; i < pbuf->nbufs; i++) {
4009 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4014 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4019 INIT_LIST_HEAD(&buf->list);
4022 list_add_tail(&buf->list, &(*head)->list);
4026 return i ? i : -ENOMEM;
4029 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4031 struct io_provide_buf *p = &req->pbuf;
4032 struct io_ring_ctx *ctx = req->ctx;
4033 struct io_buffer *head, *list;
4035 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4037 io_ring_submit_lock(ctx, !force_nonblock);
4039 lockdep_assert_held(&ctx->uring_lock);
4041 list = head = xa_load(&ctx->io_buffers, p->bgid);
4043 ret = io_add_buffers(p, &head);
4044 if (ret >= 0 && !list) {
4045 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4047 __io_remove_buffers(ctx, head, p->bgid, -1U);
4051 /* complete before unlock, IOPOLL may need the lock */
4052 __io_req_complete(req, issue_flags, ret, 0);
4053 io_ring_submit_unlock(ctx, !force_nonblock);
4057 static int io_epoll_ctl_prep(struct io_kiocb *req,
4058 const struct io_uring_sqe *sqe)
4060 #if defined(CONFIG_EPOLL)
4061 if (sqe->ioprio || sqe->buf_index)
4063 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4066 req->epoll.epfd = READ_ONCE(sqe->fd);
4067 req->epoll.op = READ_ONCE(sqe->len);
4068 req->epoll.fd = READ_ONCE(sqe->off);
4070 if (ep_op_has_event(req->epoll.op)) {
4071 struct epoll_event __user *ev;
4073 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4074 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4084 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4086 #if defined(CONFIG_EPOLL)
4087 struct io_epoll *ie = &req->epoll;
4089 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4091 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4092 if (force_nonblock && ret == -EAGAIN)
4097 __io_req_complete(req, issue_flags, ret, 0);
4104 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4106 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4107 if (sqe->ioprio || sqe->buf_index || sqe->off)
4109 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4112 req->madvise.addr = READ_ONCE(sqe->addr);
4113 req->madvise.len = READ_ONCE(sqe->len);
4114 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4121 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4123 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4124 struct io_madvise *ma = &req->madvise;
4127 if (issue_flags & IO_URING_F_NONBLOCK)
4130 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4133 io_req_complete(req, ret);
4140 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4142 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4144 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4147 req->fadvise.offset = READ_ONCE(sqe->off);
4148 req->fadvise.len = READ_ONCE(sqe->len);
4149 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4153 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4155 struct io_fadvise *fa = &req->fadvise;
4158 if (issue_flags & IO_URING_F_NONBLOCK) {
4159 switch (fa->advice) {
4160 case POSIX_FADV_NORMAL:
4161 case POSIX_FADV_RANDOM:
4162 case POSIX_FADV_SEQUENTIAL:
4169 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4172 __io_req_complete(req, issue_flags, ret, 0);
4176 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4178 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4180 if (sqe->ioprio || sqe->buf_index)
4182 if (req->flags & REQ_F_FIXED_FILE)
4185 req->statx.dfd = READ_ONCE(sqe->fd);
4186 req->statx.mask = READ_ONCE(sqe->len);
4187 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4188 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4189 req->statx.flags = READ_ONCE(sqe->statx_flags);
4194 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4196 struct io_statx *ctx = &req->statx;
4199 if (issue_flags & IO_URING_F_NONBLOCK)
4202 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4207 io_req_complete(req, ret);
4211 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4213 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4215 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4216 sqe->rw_flags || sqe->buf_index)
4218 if (req->flags & REQ_F_FIXED_FILE)
4221 req->close.fd = READ_ONCE(sqe->fd);
4225 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4227 struct files_struct *files = current->files;
4228 struct io_close *close = &req->close;
4229 struct fdtable *fdt;
4230 struct file *file = NULL;
4233 spin_lock(&files->file_lock);
4234 fdt = files_fdtable(files);
4235 if (close->fd >= fdt->max_fds) {
4236 spin_unlock(&files->file_lock);
4239 file = fdt->fd[close->fd];
4240 if (!file || file->f_op == &io_uring_fops) {
4241 spin_unlock(&files->file_lock);
4246 /* if the file has a flush method, be safe and punt to async */
4247 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4248 spin_unlock(&files->file_lock);
4252 ret = __close_fd_get_file(close->fd, &file);
4253 spin_unlock(&files->file_lock);
4260 /* No ->flush() or already async, safely close from here */
4261 ret = filp_close(file, current->files);
4267 __io_req_complete(req, issue_flags, ret, 0);
4271 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4273 struct io_ring_ctx *ctx = req->ctx;
4275 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4277 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4280 req->sync.off = READ_ONCE(sqe->off);
4281 req->sync.len = READ_ONCE(sqe->len);
4282 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4286 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4290 /* sync_file_range always requires a blocking context */
4291 if (issue_flags & IO_URING_F_NONBLOCK)
4294 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4298 io_req_complete(req, ret);
4302 #if defined(CONFIG_NET)
4303 static int io_setup_async_msg(struct io_kiocb *req,
4304 struct io_async_msghdr *kmsg)
4306 struct io_async_msghdr *async_msg = req->async_data;
4310 if (io_alloc_async_data(req)) {
4311 kfree(kmsg->free_iov);
4314 async_msg = req->async_data;
4315 req->flags |= REQ_F_NEED_CLEANUP;
4316 memcpy(async_msg, kmsg, sizeof(*kmsg));
4317 async_msg->msg.msg_name = &async_msg->addr;
4318 /* if were using fast_iov, set it to the new one */
4319 if (!async_msg->free_iov)
4320 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4325 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4326 struct io_async_msghdr *iomsg)
4328 iomsg->msg.msg_name = &iomsg->addr;
4329 iomsg->free_iov = iomsg->fast_iov;
4330 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4331 req->sr_msg.msg_flags, &iomsg->free_iov);
4334 static int io_sendmsg_prep_async(struct io_kiocb *req)
4338 ret = io_sendmsg_copy_hdr(req, req->async_data);
4340 req->flags |= REQ_F_NEED_CLEANUP;
4344 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4346 struct io_sr_msg *sr = &req->sr_msg;
4348 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4351 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4352 sr->len = READ_ONCE(sqe->len);
4353 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4354 if (sr->msg_flags & MSG_DONTWAIT)
4355 req->flags |= REQ_F_NOWAIT;
4357 #ifdef CONFIG_COMPAT
4358 if (req->ctx->compat)
4359 sr->msg_flags |= MSG_CMSG_COMPAT;
4364 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4366 struct io_async_msghdr iomsg, *kmsg;
4367 struct socket *sock;
4372 sock = sock_from_file(req->file);
4373 if (unlikely(!sock))
4376 kmsg = req->async_data;
4378 ret = io_sendmsg_copy_hdr(req, &iomsg);
4384 flags = req->sr_msg.msg_flags;
4385 if (issue_flags & IO_URING_F_NONBLOCK)
4386 flags |= MSG_DONTWAIT;
4387 if (flags & MSG_WAITALL)
4388 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4390 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4391 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4392 return io_setup_async_msg(req, kmsg);
4393 if (ret == -ERESTARTSYS)
4396 /* fast path, check for non-NULL to avoid function call */
4398 kfree(kmsg->free_iov);
4399 req->flags &= ~REQ_F_NEED_CLEANUP;
4402 __io_req_complete(req, issue_flags, ret, 0);
4406 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4408 struct io_sr_msg *sr = &req->sr_msg;
4411 struct socket *sock;
4416 sock = sock_from_file(req->file);
4417 if (unlikely(!sock))
4420 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4424 msg.msg_name = NULL;
4425 msg.msg_control = NULL;
4426 msg.msg_controllen = 0;
4427 msg.msg_namelen = 0;
4429 flags = req->sr_msg.msg_flags;
4430 if (issue_flags & IO_URING_F_NONBLOCK)
4431 flags |= MSG_DONTWAIT;
4432 if (flags & MSG_WAITALL)
4433 min_ret = iov_iter_count(&msg.msg_iter);
4435 msg.msg_flags = flags;
4436 ret = sock_sendmsg(sock, &msg);
4437 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4439 if (ret == -ERESTARTSYS)
4444 __io_req_complete(req, issue_flags, ret, 0);
4448 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4449 struct io_async_msghdr *iomsg)
4451 struct io_sr_msg *sr = &req->sr_msg;
4452 struct iovec __user *uiov;
4456 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4457 &iomsg->uaddr, &uiov, &iov_len);
4461 if (req->flags & REQ_F_BUFFER_SELECT) {
4464 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4466 sr->len = iomsg->fast_iov[0].iov_len;
4467 iomsg->free_iov = NULL;
4469 iomsg->free_iov = iomsg->fast_iov;
4470 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4471 &iomsg->free_iov, &iomsg->msg.msg_iter,
4480 #ifdef CONFIG_COMPAT
4481 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4482 struct io_async_msghdr *iomsg)
4484 struct io_sr_msg *sr = &req->sr_msg;
4485 struct compat_iovec __user *uiov;
4490 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4495 uiov = compat_ptr(ptr);
4496 if (req->flags & REQ_F_BUFFER_SELECT) {
4497 compat_ssize_t clen;
4501 if (!access_ok(uiov, sizeof(*uiov)))
4503 if (__get_user(clen, &uiov->iov_len))
4508 iomsg->free_iov = NULL;
4510 iomsg->free_iov = iomsg->fast_iov;
4511 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4512 UIO_FASTIOV, &iomsg->free_iov,
4513 &iomsg->msg.msg_iter, true);
4522 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4523 struct io_async_msghdr *iomsg)
4525 iomsg->msg.msg_name = &iomsg->addr;
4527 #ifdef CONFIG_COMPAT
4528 if (req->ctx->compat)
4529 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4532 return __io_recvmsg_copy_hdr(req, iomsg);
4535 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4538 struct io_sr_msg *sr = &req->sr_msg;
4539 struct io_buffer *kbuf;
4541 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4546 req->flags |= REQ_F_BUFFER_SELECTED;
4550 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4552 return io_put_kbuf(req, req->sr_msg.kbuf);
4555 static int io_recvmsg_prep_async(struct io_kiocb *req)
4559 ret = io_recvmsg_copy_hdr(req, req->async_data);
4561 req->flags |= REQ_F_NEED_CLEANUP;
4565 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4567 struct io_sr_msg *sr = &req->sr_msg;
4569 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4572 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4573 sr->len = READ_ONCE(sqe->len);
4574 sr->bgid = READ_ONCE(sqe->buf_group);
4575 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4576 if (sr->msg_flags & MSG_DONTWAIT)
4577 req->flags |= REQ_F_NOWAIT;
4579 #ifdef CONFIG_COMPAT
4580 if (req->ctx->compat)
4581 sr->msg_flags |= MSG_CMSG_COMPAT;
4586 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4588 struct io_async_msghdr iomsg, *kmsg;
4589 struct socket *sock;
4590 struct io_buffer *kbuf;
4593 int ret, cflags = 0;
4594 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4596 sock = sock_from_file(req->file);
4597 if (unlikely(!sock))
4600 kmsg = req->async_data;
4602 ret = io_recvmsg_copy_hdr(req, &iomsg);
4608 if (req->flags & REQ_F_BUFFER_SELECT) {
4609 kbuf = io_recv_buffer_select(req, !force_nonblock);
4611 return PTR_ERR(kbuf);
4612 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4613 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4614 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4615 1, req->sr_msg.len);
4618 flags = req->sr_msg.msg_flags;
4620 flags |= MSG_DONTWAIT;
4621 if (flags & MSG_WAITALL)
4622 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4624 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4625 kmsg->uaddr, flags);
4626 if (force_nonblock && ret == -EAGAIN)
4627 return io_setup_async_msg(req, kmsg);
4628 if (ret == -ERESTARTSYS)
4631 if (req->flags & REQ_F_BUFFER_SELECTED)
4632 cflags = io_put_recv_kbuf(req);
4633 /* fast path, check for non-NULL to avoid function call */
4635 kfree(kmsg->free_iov);
4636 req->flags &= ~REQ_F_NEED_CLEANUP;
4637 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4639 __io_req_complete(req, issue_flags, ret, cflags);
4643 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4645 struct io_buffer *kbuf;
4646 struct io_sr_msg *sr = &req->sr_msg;
4648 void __user *buf = sr->buf;
4649 struct socket *sock;
4653 int ret, cflags = 0;
4654 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4656 sock = sock_from_file(req->file);
4657 if (unlikely(!sock))
4660 if (req->flags & REQ_F_BUFFER_SELECT) {
4661 kbuf = io_recv_buffer_select(req, !force_nonblock);
4663 return PTR_ERR(kbuf);
4664 buf = u64_to_user_ptr(kbuf->addr);
4667 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4671 msg.msg_name = NULL;
4672 msg.msg_control = NULL;
4673 msg.msg_controllen = 0;
4674 msg.msg_namelen = 0;
4675 msg.msg_iocb = NULL;
4678 flags = req->sr_msg.msg_flags;
4680 flags |= MSG_DONTWAIT;
4681 if (flags & MSG_WAITALL)
4682 min_ret = iov_iter_count(&msg.msg_iter);
4684 ret = sock_recvmsg(sock, &msg, flags);
4685 if (force_nonblock && ret == -EAGAIN)
4687 if (ret == -ERESTARTSYS)
4690 if (req->flags & REQ_F_BUFFER_SELECTED)
4691 cflags = io_put_recv_kbuf(req);
4692 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4694 __io_req_complete(req, issue_flags, ret, cflags);
4698 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4700 struct io_accept *accept = &req->accept;
4702 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4704 if (sqe->ioprio || sqe->len || sqe->buf_index)
4707 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4708 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4709 accept->flags = READ_ONCE(sqe->accept_flags);
4710 accept->nofile = rlimit(RLIMIT_NOFILE);
4714 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4716 struct io_accept *accept = &req->accept;
4717 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4718 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4721 if (req->file->f_flags & O_NONBLOCK)
4722 req->flags |= REQ_F_NOWAIT;
4724 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4725 accept->addr_len, accept->flags,
4727 if (ret == -EAGAIN && force_nonblock)
4730 if (ret == -ERESTARTSYS)
4734 __io_req_complete(req, issue_flags, ret, 0);
4738 static int io_connect_prep_async(struct io_kiocb *req)
4740 struct io_async_connect *io = req->async_data;
4741 struct io_connect *conn = &req->connect;
4743 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4746 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4748 struct io_connect *conn = &req->connect;
4750 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4752 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4755 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4756 conn->addr_len = READ_ONCE(sqe->addr2);
4760 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4762 struct io_async_connect __io, *io;
4763 unsigned file_flags;
4765 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4767 if (req->async_data) {
4768 io = req->async_data;
4770 ret = move_addr_to_kernel(req->connect.addr,
4771 req->connect.addr_len,
4778 file_flags = force_nonblock ? O_NONBLOCK : 0;
4780 ret = __sys_connect_file(req->file, &io->address,
4781 req->connect.addr_len, file_flags);
4782 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4783 if (req->async_data)
4785 if (io_alloc_async_data(req)) {
4789 memcpy(req->async_data, &__io, sizeof(__io));
4792 if (ret == -ERESTARTSYS)
4797 __io_req_complete(req, issue_flags, ret, 0);
4800 #else /* !CONFIG_NET */
4801 #define IO_NETOP_FN(op) \
4802 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4804 return -EOPNOTSUPP; \
4807 #define IO_NETOP_PREP(op) \
4809 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4811 return -EOPNOTSUPP; \
4814 #define IO_NETOP_PREP_ASYNC(op) \
4816 static int io_##op##_prep_async(struct io_kiocb *req) \
4818 return -EOPNOTSUPP; \
4821 IO_NETOP_PREP_ASYNC(sendmsg);
4822 IO_NETOP_PREP_ASYNC(recvmsg);
4823 IO_NETOP_PREP_ASYNC(connect);
4824 IO_NETOP_PREP(accept);
4827 #endif /* CONFIG_NET */
4829 struct io_poll_table {
4830 struct poll_table_struct pt;
4831 struct io_kiocb *req;
4836 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4837 __poll_t mask, io_req_tw_func_t func)
4839 /* for instances that support it check for an event match first: */
4840 if (mask && !(mask & poll->events))
4843 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4845 list_del_init(&poll->wait.entry);
4848 req->io_task_work.func = func;
4851 * If this fails, then the task is exiting. When a task exits, the
4852 * work gets canceled, so just cancel this request as well instead
4853 * of executing it. We can't safely execute it anyway, as we may not
4854 * have the needed state needed for it anyway.
4856 io_req_task_work_add(req);
4860 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4861 __acquires(&req->ctx->completion_lock)
4863 struct io_ring_ctx *ctx = req->ctx;
4865 if (unlikely(req->task->flags & PF_EXITING))
4866 WRITE_ONCE(poll->canceled, true);
4868 if (!req->result && !READ_ONCE(poll->canceled)) {
4869 struct poll_table_struct pt = { ._key = poll->events };
4871 req->result = vfs_poll(req->file, &pt) & poll->events;
4874 spin_lock(&ctx->completion_lock);
4875 if (!req->result && !READ_ONCE(poll->canceled)) {
4876 add_wait_queue(poll->head, &poll->wait);
4883 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4885 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4886 if (req->opcode == IORING_OP_POLL_ADD)
4887 return req->async_data;
4888 return req->apoll->double_poll;
4891 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4893 if (req->opcode == IORING_OP_POLL_ADD)
4895 return &req->apoll->poll;
4898 static void io_poll_remove_double(struct io_kiocb *req)
4899 __must_hold(&req->ctx->completion_lock)
4901 struct io_poll_iocb *poll = io_poll_get_double(req);
4903 lockdep_assert_held(&req->ctx->completion_lock);
4905 if (poll && poll->head) {
4906 struct wait_queue_head *head = poll->head;
4908 spin_lock_irq(&head->lock);
4909 list_del_init(&poll->wait.entry);
4910 if (poll->wait.private)
4913 spin_unlock_irq(&head->lock);
4917 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4918 __must_hold(&req->ctx->completion_lock)
4920 struct io_ring_ctx *ctx = req->ctx;
4921 unsigned flags = IORING_CQE_F_MORE;
4924 if (READ_ONCE(req->poll.canceled)) {
4926 req->poll.events |= EPOLLONESHOT;
4928 error = mangle_poll(mask);
4930 if (req->poll.events & EPOLLONESHOT)
4932 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4933 req->poll.done = true;
4936 if (flags & IORING_CQE_F_MORE)
4939 io_commit_cqring(ctx);
4940 return !(flags & IORING_CQE_F_MORE);
4943 static void io_poll_task_func(struct io_kiocb *req)
4945 struct io_ring_ctx *ctx = req->ctx;
4946 struct io_kiocb *nxt;
4948 if (io_poll_rewait(req, &req->poll)) {
4949 spin_unlock(&ctx->completion_lock);
4953 done = io_poll_complete(req, req->result);
4955 io_poll_remove_double(req);
4956 hash_del(&req->hash_node);
4959 add_wait_queue(req->poll.head, &req->poll.wait);
4961 spin_unlock(&ctx->completion_lock);
4962 io_cqring_ev_posted(ctx);
4965 nxt = io_put_req_find_next(req);
4967 io_req_task_submit(nxt);
4972 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4973 int sync, void *key)
4975 struct io_kiocb *req = wait->private;
4976 struct io_poll_iocb *poll = io_poll_get_single(req);
4977 __poll_t mask = key_to_poll(key);
4978 unsigned long flags;
4980 /* for instances that support it check for an event match first: */
4981 if (mask && !(mask & poll->events))
4983 if (!(poll->events & EPOLLONESHOT))
4984 return poll->wait.func(&poll->wait, mode, sync, key);
4986 list_del_init(&wait->entry);
4991 spin_lock_irqsave(&poll->head->lock, flags);
4992 done = list_empty(&poll->wait.entry);
4994 list_del_init(&poll->wait.entry);
4995 /* make sure double remove sees this as being gone */
4996 wait->private = NULL;
4997 spin_unlock_irqrestore(&poll->head->lock, flags);
4999 /* use wait func handler, so it matches the rq type */
5000 poll->wait.func(&poll->wait, mode, sync, key);
5007 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5008 wait_queue_func_t wake_func)
5012 poll->canceled = false;
5013 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5014 /* mask in events that we always want/need */
5015 poll->events = events | IO_POLL_UNMASK;
5016 INIT_LIST_HEAD(&poll->wait.entry);
5017 init_waitqueue_func_entry(&poll->wait, wake_func);
5020 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5021 struct wait_queue_head *head,
5022 struct io_poll_iocb **poll_ptr)
5024 struct io_kiocb *req = pt->req;
5027 * The file being polled uses multiple waitqueues for poll handling
5028 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5031 if (unlikely(pt->nr_entries)) {
5032 struct io_poll_iocb *poll_one = poll;
5034 /* already have a 2nd entry, fail a third attempt */
5036 pt->error = -EINVAL;
5040 * Can't handle multishot for double wait for now, turn it
5041 * into one-shot mode.
5043 if (!(poll_one->events & EPOLLONESHOT))
5044 poll_one->events |= EPOLLONESHOT;
5045 /* double add on the same waitqueue head, ignore */
5046 if (poll_one->head == head)
5048 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5050 pt->error = -ENOMEM;
5053 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5055 poll->wait.private = req;
5062 if (poll->events & EPOLLEXCLUSIVE)
5063 add_wait_queue_exclusive(head, &poll->wait);
5065 add_wait_queue(head, &poll->wait);
5068 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5069 struct poll_table_struct *p)
5071 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5072 struct async_poll *apoll = pt->req->apoll;
5074 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5077 static void io_async_task_func(struct io_kiocb *req)
5079 struct async_poll *apoll = req->apoll;
5080 struct io_ring_ctx *ctx = req->ctx;
5082 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5084 if (io_poll_rewait(req, &apoll->poll)) {
5085 spin_unlock(&ctx->completion_lock);
5089 hash_del(&req->hash_node);
5090 io_poll_remove_double(req);
5091 spin_unlock(&ctx->completion_lock);
5093 if (!READ_ONCE(apoll->poll.canceled))
5094 io_req_task_submit(req);
5096 io_req_complete_failed(req, -ECANCELED);
5099 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5102 struct io_kiocb *req = wait->private;
5103 struct io_poll_iocb *poll = &req->apoll->poll;
5105 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5108 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5111 static void io_poll_req_insert(struct io_kiocb *req)
5113 struct io_ring_ctx *ctx = req->ctx;
5114 struct hlist_head *list;
5116 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5117 hlist_add_head(&req->hash_node, list);
5120 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5121 struct io_poll_iocb *poll,
5122 struct io_poll_table *ipt, __poll_t mask,
5123 wait_queue_func_t wake_func)
5124 __acquires(&ctx->completion_lock)
5126 struct io_ring_ctx *ctx = req->ctx;
5127 bool cancel = false;
5129 INIT_HLIST_NODE(&req->hash_node);
5130 io_init_poll_iocb(poll, mask, wake_func);
5131 poll->file = req->file;
5132 poll->wait.private = req;
5134 ipt->pt._key = mask;
5137 ipt->nr_entries = 0;
5139 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5140 if (unlikely(!ipt->nr_entries) && !ipt->error)
5141 ipt->error = -EINVAL;
5143 spin_lock(&ctx->completion_lock);
5144 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5145 io_poll_remove_double(req);
5146 if (likely(poll->head)) {
5147 spin_lock_irq(&poll->head->lock);
5148 if (unlikely(list_empty(&poll->wait.entry))) {
5154 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5155 list_del_init(&poll->wait.entry);
5157 WRITE_ONCE(poll->canceled, true);
5158 else if (!poll->done) /* actually waiting for an event */
5159 io_poll_req_insert(req);
5160 spin_unlock_irq(&poll->head->lock);
5172 static int io_arm_poll_handler(struct io_kiocb *req)
5174 const struct io_op_def *def = &io_op_defs[req->opcode];
5175 struct io_ring_ctx *ctx = req->ctx;
5176 struct async_poll *apoll;
5177 struct io_poll_table ipt;
5178 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5181 if (!req->file || !file_can_poll(req->file))
5182 return IO_APOLL_ABORTED;
5183 if (req->flags & REQ_F_POLLED)
5184 return IO_APOLL_ABORTED;
5185 if (!def->pollin && !def->pollout)
5186 return IO_APOLL_ABORTED;
5190 mask |= POLLIN | POLLRDNORM;
5192 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5193 if ((req->opcode == IORING_OP_RECVMSG) &&
5194 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5198 mask |= POLLOUT | POLLWRNORM;
5201 /* if we can't nonblock try, then no point in arming a poll handler */
5202 if (!io_file_supports_nowait(req, rw))
5203 return IO_APOLL_ABORTED;
5205 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5206 if (unlikely(!apoll))
5207 return IO_APOLL_ABORTED;
5208 apoll->double_poll = NULL;
5210 req->flags |= REQ_F_POLLED;
5211 ipt.pt._qproc = io_async_queue_proc;
5212 io_req_refcount(req);
5214 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5216 if (ret || ipt.error) {
5217 spin_unlock(&ctx->completion_lock);
5219 return IO_APOLL_READY;
5220 return IO_APOLL_ABORTED;
5222 spin_unlock(&ctx->completion_lock);
5223 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5224 mask, apoll->poll.events);
5228 static bool __io_poll_remove_one(struct io_kiocb *req,
5229 struct io_poll_iocb *poll, bool do_cancel)
5230 __must_hold(&req->ctx->completion_lock)
5232 bool do_complete = false;
5236 spin_lock_irq(&poll->head->lock);
5238 WRITE_ONCE(poll->canceled, true);
5239 if (!list_empty(&poll->wait.entry)) {
5240 list_del_init(&poll->wait.entry);
5243 spin_unlock_irq(&poll->head->lock);
5244 hash_del(&req->hash_node);
5248 static bool io_poll_remove_one(struct io_kiocb *req)
5249 __must_hold(&req->ctx->completion_lock)
5253 io_poll_remove_double(req);
5254 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5257 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5258 io_commit_cqring(req->ctx);
5260 io_put_req_deferred(req);
5266 * Returns true if we found and killed one or more poll requests
5268 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5271 struct hlist_node *tmp;
5272 struct io_kiocb *req;
5275 spin_lock(&ctx->completion_lock);
5276 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5277 struct hlist_head *list;
5279 list = &ctx->cancel_hash[i];
5280 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5281 if (io_match_task(req, tsk, cancel_all))
5282 posted += io_poll_remove_one(req);
5285 spin_unlock(&ctx->completion_lock);
5288 io_cqring_ev_posted(ctx);
5293 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5295 __must_hold(&ctx->completion_lock)
5297 struct hlist_head *list;
5298 struct io_kiocb *req;
5300 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5301 hlist_for_each_entry(req, list, hash_node) {
5302 if (sqe_addr != req->user_data)
5304 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5311 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5313 __must_hold(&ctx->completion_lock)
5315 struct io_kiocb *req;
5317 req = io_poll_find(ctx, sqe_addr, poll_only);
5320 if (io_poll_remove_one(req))
5326 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5331 events = READ_ONCE(sqe->poll32_events);
5333 events = swahw32(events);
5335 if (!(flags & IORING_POLL_ADD_MULTI))
5336 events |= EPOLLONESHOT;
5337 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5340 static int io_poll_update_prep(struct io_kiocb *req,
5341 const struct io_uring_sqe *sqe)
5343 struct io_poll_update *upd = &req->poll_update;
5346 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5348 if (sqe->ioprio || sqe->buf_index)
5350 flags = READ_ONCE(sqe->len);
5351 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5352 IORING_POLL_ADD_MULTI))
5354 /* meaningless without update */
5355 if (flags == IORING_POLL_ADD_MULTI)
5358 upd->old_user_data = READ_ONCE(sqe->addr);
5359 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5360 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5362 upd->new_user_data = READ_ONCE(sqe->off);
5363 if (!upd->update_user_data && upd->new_user_data)
5365 if (upd->update_events)
5366 upd->events = io_poll_parse_events(sqe, flags);
5367 else if (sqe->poll32_events)
5373 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5376 struct io_kiocb *req = wait->private;
5377 struct io_poll_iocb *poll = &req->poll;
5379 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5382 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5383 struct poll_table_struct *p)
5385 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5387 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5390 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5392 struct io_poll_iocb *poll = &req->poll;
5395 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5397 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5399 flags = READ_ONCE(sqe->len);
5400 if (flags & ~IORING_POLL_ADD_MULTI)
5403 io_req_refcount(req);
5404 poll->events = io_poll_parse_events(sqe, flags);
5408 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5410 struct io_poll_iocb *poll = &req->poll;
5411 struct io_ring_ctx *ctx = req->ctx;
5412 struct io_poll_table ipt;
5415 ipt.pt._qproc = io_poll_queue_proc;
5417 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5420 if (mask) { /* no async, we'd stolen it */
5422 io_poll_complete(req, mask);
5424 spin_unlock(&ctx->completion_lock);
5427 io_cqring_ev_posted(ctx);
5428 if (poll->events & EPOLLONESHOT)
5434 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5436 struct io_ring_ctx *ctx = req->ctx;
5437 struct io_kiocb *preq;
5441 spin_lock(&ctx->completion_lock);
5442 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5448 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5450 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5455 * Don't allow racy completion with singleshot, as we cannot safely
5456 * update those. For multishot, if we're racing with completion, just
5457 * let completion re-add it.
5459 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5460 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5464 /* we now have a detached poll request. reissue. */
5468 spin_unlock(&ctx->completion_lock);
5470 io_req_complete(req, ret);
5473 /* only mask one event flags, keep behavior flags */
5474 if (req->poll_update.update_events) {
5475 preq->poll.events &= ~0xffff;
5476 preq->poll.events |= req->poll_update.events & 0xffff;
5477 preq->poll.events |= IO_POLL_UNMASK;
5479 if (req->poll_update.update_user_data)
5480 preq->user_data = req->poll_update.new_user_data;
5481 spin_unlock(&ctx->completion_lock);
5483 /* complete update request, we're done with it */
5484 io_req_complete(req, ret);
5487 ret = io_poll_add(preq, issue_flags);
5490 io_req_complete(preq, ret);
5496 static void io_req_task_timeout(struct io_kiocb *req)
5498 struct io_ring_ctx *ctx = req->ctx;
5500 spin_lock(&ctx->completion_lock);
5501 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5502 io_commit_cqring(ctx);
5503 spin_unlock(&ctx->completion_lock);
5505 io_cqring_ev_posted(ctx);
5510 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5512 struct io_timeout_data *data = container_of(timer,
5513 struct io_timeout_data, timer);
5514 struct io_kiocb *req = data->req;
5515 struct io_ring_ctx *ctx = req->ctx;
5516 unsigned long flags;
5518 spin_lock_irqsave(&ctx->timeout_lock, flags);
5519 list_del_init(&req->timeout.list);
5520 atomic_set(&req->ctx->cq_timeouts,
5521 atomic_read(&req->ctx->cq_timeouts) + 1);
5522 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5524 req->io_task_work.func = io_req_task_timeout;
5525 io_req_task_work_add(req);
5526 return HRTIMER_NORESTART;
5529 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5531 __must_hold(&ctx->timeout_lock)
5533 struct io_timeout_data *io;
5534 struct io_kiocb *req;
5537 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5538 found = user_data == req->user_data;
5543 return ERR_PTR(-ENOENT);
5545 io = req->async_data;
5546 if (hrtimer_try_to_cancel(&io->timer) == -1)
5547 return ERR_PTR(-EALREADY);
5548 list_del_init(&req->timeout.list);
5552 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5553 __must_hold(&ctx->timeout_lock)
5555 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5558 return PTR_ERR(req);
5561 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5562 io_put_req_deferred(req);
5566 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5567 struct timespec64 *ts, enum hrtimer_mode mode)
5568 __must_hold(&ctx->timeout_lock)
5570 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5571 struct io_timeout_data *data;
5574 return PTR_ERR(req);
5576 req->timeout.off = 0; /* noseq */
5577 data = req->async_data;
5578 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5579 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5580 data->timer.function = io_timeout_fn;
5581 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5585 static int io_timeout_remove_prep(struct io_kiocb *req,
5586 const struct io_uring_sqe *sqe)
5588 struct io_timeout_rem *tr = &req->timeout_rem;
5590 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5592 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5594 if (sqe->ioprio || sqe->buf_index || sqe->len)
5597 tr->addr = READ_ONCE(sqe->addr);
5598 tr->flags = READ_ONCE(sqe->timeout_flags);
5599 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5600 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5602 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5604 } else if (tr->flags) {
5605 /* timeout removal doesn't support flags */
5612 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5614 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5619 * Remove or update an existing timeout command
5621 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5623 struct io_timeout_rem *tr = &req->timeout_rem;
5624 struct io_ring_ctx *ctx = req->ctx;
5627 spin_lock_irq(&ctx->timeout_lock);
5628 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5629 ret = io_timeout_cancel(ctx, tr->addr);
5631 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5632 io_translate_timeout_mode(tr->flags));
5633 spin_unlock_irq(&ctx->timeout_lock);
5635 spin_lock(&ctx->completion_lock);
5636 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5637 io_commit_cqring(ctx);
5638 spin_unlock(&ctx->completion_lock);
5639 io_cqring_ev_posted(ctx);
5646 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5647 bool is_timeout_link)
5649 struct io_timeout_data *data;
5651 u32 off = READ_ONCE(sqe->off);
5653 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5655 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5657 if (off && is_timeout_link)
5659 flags = READ_ONCE(sqe->timeout_flags);
5660 if (flags & ~IORING_TIMEOUT_ABS)
5663 req->timeout.off = off;
5664 if (unlikely(off && !req->ctx->off_timeout_used))
5665 req->ctx->off_timeout_used = true;
5667 if (!req->async_data && io_alloc_async_data(req))
5670 data = req->async_data;
5673 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5676 data->mode = io_translate_timeout_mode(flags);
5677 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5678 if (is_timeout_link)
5679 io_req_track_inflight(req);
5683 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5685 struct io_ring_ctx *ctx = req->ctx;
5686 struct io_timeout_data *data = req->async_data;
5687 struct list_head *entry;
5688 u32 tail, off = req->timeout.off;
5690 spin_lock_irq(&ctx->timeout_lock);
5693 * sqe->off holds how many events that need to occur for this
5694 * timeout event to be satisfied. If it isn't set, then this is
5695 * a pure timeout request, sequence isn't used.
5697 if (io_is_timeout_noseq(req)) {
5698 entry = ctx->timeout_list.prev;
5702 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5703 req->timeout.target_seq = tail + off;
5705 /* Update the last seq here in case io_flush_timeouts() hasn't.
5706 * This is safe because ->completion_lock is held, and submissions
5707 * and completions are never mixed in the same ->completion_lock section.
5709 ctx->cq_last_tm_flush = tail;
5712 * Insertion sort, ensuring the first entry in the list is always
5713 * the one we need first.
5715 list_for_each_prev(entry, &ctx->timeout_list) {
5716 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5719 if (io_is_timeout_noseq(nxt))
5721 /* nxt.seq is behind @tail, otherwise would've been completed */
5722 if (off >= nxt->timeout.target_seq - tail)
5726 list_add(&req->timeout.list, entry);
5727 data->timer.function = io_timeout_fn;
5728 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5729 spin_unlock_irq(&ctx->timeout_lock);
5733 struct io_cancel_data {
5734 struct io_ring_ctx *ctx;
5738 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5740 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5741 struct io_cancel_data *cd = data;
5743 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5746 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5747 struct io_ring_ctx *ctx)
5749 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5750 enum io_wq_cancel cancel_ret;
5753 if (!tctx || !tctx->io_wq)
5756 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5757 switch (cancel_ret) {
5758 case IO_WQ_CANCEL_OK:
5761 case IO_WQ_CANCEL_RUNNING:
5764 case IO_WQ_CANCEL_NOTFOUND:
5772 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5773 struct io_kiocb *req, __u64 sqe_addr,
5778 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5779 spin_lock(&ctx->completion_lock);
5782 spin_lock_irq(&ctx->timeout_lock);
5783 ret = io_timeout_cancel(ctx, sqe_addr);
5784 spin_unlock_irq(&ctx->timeout_lock);
5787 ret = io_poll_cancel(ctx, sqe_addr, false);
5791 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5792 io_commit_cqring(ctx);
5793 spin_unlock(&ctx->completion_lock);
5794 io_cqring_ev_posted(ctx);
5800 static int io_async_cancel_prep(struct io_kiocb *req,
5801 const struct io_uring_sqe *sqe)
5803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5805 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5807 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5810 req->cancel.addr = READ_ONCE(sqe->addr);
5814 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5816 struct io_ring_ctx *ctx = req->ctx;
5817 u64 sqe_addr = req->cancel.addr;
5818 struct io_tctx_node *node;
5821 /* tasks should wait for their io-wq threads, so safe w/o sync */
5822 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5823 spin_lock(&ctx->completion_lock);
5826 spin_lock_irq(&ctx->timeout_lock);
5827 ret = io_timeout_cancel(ctx, sqe_addr);
5828 spin_unlock_irq(&ctx->timeout_lock);
5831 ret = io_poll_cancel(ctx, sqe_addr, false);
5834 spin_unlock(&ctx->completion_lock);
5836 /* slow path, try all io-wq's */
5837 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5839 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5840 struct io_uring_task *tctx = node->task->io_uring;
5842 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5846 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5848 spin_lock(&ctx->completion_lock);
5850 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5851 io_commit_cqring(ctx);
5852 spin_unlock(&ctx->completion_lock);
5853 io_cqring_ev_posted(ctx);
5861 static int io_rsrc_update_prep(struct io_kiocb *req,
5862 const struct io_uring_sqe *sqe)
5864 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5866 if (sqe->ioprio || sqe->rw_flags)
5869 req->rsrc_update.offset = READ_ONCE(sqe->off);
5870 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5871 if (!req->rsrc_update.nr_args)
5873 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5877 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5879 struct io_ring_ctx *ctx = req->ctx;
5880 struct io_uring_rsrc_update2 up;
5883 if (issue_flags & IO_URING_F_NONBLOCK)
5886 up.offset = req->rsrc_update.offset;
5887 up.data = req->rsrc_update.arg;
5892 mutex_lock(&ctx->uring_lock);
5893 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5894 &up, req->rsrc_update.nr_args);
5895 mutex_unlock(&ctx->uring_lock);
5899 __io_req_complete(req, issue_flags, ret, 0);
5903 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5905 switch (req->opcode) {
5908 case IORING_OP_READV:
5909 case IORING_OP_READ_FIXED:
5910 case IORING_OP_READ:
5911 return io_read_prep(req, sqe);
5912 case IORING_OP_WRITEV:
5913 case IORING_OP_WRITE_FIXED:
5914 case IORING_OP_WRITE:
5915 return io_write_prep(req, sqe);
5916 case IORING_OP_POLL_ADD:
5917 return io_poll_add_prep(req, sqe);
5918 case IORING_OP_POLL_REMOVE:
5919 return io_poll_update_prep(req, sqe);
5920 case IORING_OP_FSYNC:
5921 return io_fsync_prep(req, sqe);
5922 case IORING_OP_SYNC_FILE_RANGE:
5923 return io_sfr_prep(req, sqe);
5924 case IORING_OP_SENDMSG:
5925 case IORING_OP_SEND:
5926 return io_sendmsg_prep(req, sqe);
5927 case IORING_OP_RECVMSG:
5928 case IORING_OP_RECV:
5929 return io_recvmsg_prep(req, sqe);
5930 case IORING_OP_CONNECT:
5931 return io_connect_prep(req, sqe);
5932 case IORING_OP_TIMEOUT:
5933 return io_timeout_prep(req, sqe, false);
5934 case IORING_OP_TIMEOUT_REMOVE:
5935 return io_timeout_remove_prep(req, sqe);
5936 case IORING_OP_ASYNC_CANCEL:
5937 return io_async_cancel_prep(req, sqe);
5938 case IORING_OP_LINK_TIMEOUT:
5939 return io_timeout_prep(req, sqe, true);
5940 case IORING_OP_ACCEPT:
5941 return io_accept_prep(req, sqe);
5942 case IORING_OP_FALLOCATE:
5943 return io_fallocate_prep(req, sqe);
5944 case IORING_OP_OPENAT:
5945 return io_openat_prep(req, sqe);
5946 case IORING_OP_CLOSE:
5947 return io_close_prep(req, sqe);
5948 case IORING_OP_FILES_UPDATE:
5949 return io_rsrc_update_prep(req, sqe);
5950 case IORING_OP_STATX:
5951 return io_statx_prep(req, sqe);
5952 case IORING_OP_FADVISE:
5953 return io_fadvise_prep(req, sqe);
5954 case IORING_OP_MADVISE:
5955 return io_madvise_prep(req, sqe);
5956 case IORING_OP_OPENAT2:
5957 return io_openat2_prep(req, sqe);
5958 case IORING_OP_EPOLL_CTL:
5959 return io_epoll_ctl_prep(req, sqe);
5960 case IORING_OP_SPLICE:
5961 return io_splice_prep(req, sqe);
5962 case IORING_OP_PROVIDE_BUFFERS:
5963 return io_provide_buffers_prep(req, sqe);
5964 case IORING_OP_REMOVE_BUFFERS:
5965 return io_remove_buffers_prep(req, sqe);
5967 return io_tee_prep(req, sqe);
5968 case IORING_OP_SHUTDOWN:
5969 return io_shutdown_prep(req, sqe);
5970 case IORING_OP_RENAMEAT:
5971 return io_renameat_prep(req, sqe);
5972 case IORING_OP_UNLINKAT:
5973 return io_unlinkat_prep(req, sqe);
5976 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5981 static int io_req_prep_async(struct io_kiocb *req)
5983 if (!io_op_defs[req->opcode].needs_async_setup)
5985 if (WARN_ON_ONCE(req->async_data))
5987 if (io_alloc_async_data(req))
5990 switch (req->opcode) {
5991 case IORING_OP_READV:
5992 return io_rw_prep_async(req, READ);
5993 case IORING_OP_WRITEV:
5994 return io_rw_prep_async(req, WRITE);
5995 case IORING_OP_SENDMSG:
5996 return io_sendmsg_prep_async(req);
5997 case IORING_OP_RECVMSG:
5998 return io_recvmsg_prep_async(req);
5999 case IORING_OP_CONNECT:
6000 return io_connect_prep_async(req);
6002 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6007 static u32 io_get_sequence(struct io_kiocb *req)
6009 u32 seq = req->ctx->cached_sq_head;
6011 /* need original cached_sq_head, but it was increased for each req */
6012 io_for_each_link(req, req)
6017 static bool io_drain_req(struct io_kiocb *req)
6019 struct io_kiocb *pos;
6020 struct io_ring_ctx *ctx = req->ctx;
6021 struct io_defer_entry *de;
6026 * If we need to drain a request in the middle of a link, drain the
6027 * head request and the next request/link after the current link.
6028 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6029 * maintained for every request of our link.
6031 if (ctx->drain_next) {
6032 req->flags |= REQ_F_IO_DRAIN;
6033 ctx->drain_next = false;
6035 /* not interested in head, start from the first linked */
6036 io_for_each_link(pos, req->link) {
6037 if (pos->flags & REQ_F_IO_DRAIN) {
6038 ctx->drain_next = true;
6039 req->flags |= REQ_F_IO_DRAIN;
6044 /* Still need defer if there is pending req in defer list. */
6045 if (likely(list_empty_careful(&ctx->defer_list) &&
6046 !(req->flags & REQ_F_IO_DRAIN))) {
6047 ctx->drain_active = false;
6051 seq = io_get_sequence(req);
6052 /* Still a chance to pass the sequence check */
6053 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6056 ret = io_req_prep_async(req);
6059 io_prep_async_link(req);
6060 de = kmalloc(sizeof(*de), GFP_KERNEL);
6064 io_req_complete_failed(req, ret);
6068 spin_lock(&ctx->completion_lock);
6069 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6070 spin_unlock(&ctx->completion_lock);
6072 io_queue_async_work(req);
6076 trace_io_uring_defer(ctx, req, req->user_data);
6079 list_add_tail(&de->list, &ctx->defer_list);
6080 spin_unlock(&ctx->completion_lock);
6084 static void io_clean_op(struct io_kiocb *req)
6086 if (req->flags & REQ_F_BUFFER_SELECTED) {
6087 switch (req->opcode) {
6088 case IORING_OP_READV:
6089 case IORING_OP_READ_FIXED:
6090 case IORING_OP_READ:
6091 kfree((void *)(unsigned long)req->rw.addr);
6093 case IORING_OP_RECVMSG:
6094 case IORING_OP_RECV:
6095 kfree(req->sr_msg.kbuf);
6100 if (req->flags & REQ_F_NEED_CLEANUP) {
6101 switch (req->opcode) {
6102 case IORING_OP_READV:
6103 case IORING_OP_READ_FIXED:
6104 case IORING_OP_READ:
6105 case IORING_OP_WRITEV:
6106 case IORING_OP_WRITE_FIXED:
6107 case IORING_OP_WRITE: {
6108 struct io_async_rw *io = req->async_data;
6110 kfree(io->free_iovec);
6113 case IORING_OP_RECVMSG:
6114 case IORING_OP_SENDMSG: {
6115 struct io_async_msghdr *io = req->async_data;
6117 kfree(io->free_iov);
6120 case IORING_OP_SPLICE:
6122 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6123 io_put_file(req->splice.file_in);
6125 case IORING_OP_OPENAT:
6126 case IORING_OP_OPENAT2:
6127 if (req->open.filename)
6128 putname(req->open.filename);
6130 case IORING_OP_RENAMEAT:
6131 putname(req->rename.oldpath);
6132 putname(req->rename.newpath);
6134 case IORING_OP_UNLINKAT:
6135 putname(req->unlink.filename);
6139 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6140 kfree(req->apoll->double_poll);
6144 if (req->flags & REQ_F_INFLIGHT) {
6145 struct io_uring_task *tctx = req->task->io_uring;
6147 atomic_dec(&tctx->inflight_tracked);
6149 if (req->flags & REQ_F_CREDS)
6150 put_cred(req->creds);
6152 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6155 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6157 struct io_ring_ctx *ctx = req->ctx;
6158 const struct cred *creds = NULL;
6161 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6162 creds = override_creds(req->creds);
6164 switch (req->opcode) {
6166 ret = io_nop(req, issue_flags);
6168 case IORING_OP_READV:
6169 case IORING_OP_READ_FIXED:
6170 case IORING_OP_READ:
6171 ret = io_read(req, issue_flags);
6173 case IORING_OP_WRITEV:
6174 case IORING_OP_WRITE_FIXED:
6175 case IORING_OP_WRITE:
6176 ret = io_write(req, issue_flags);
6178 case IORING_OP_FSYNC:
6179 ret = io_fsync(req, issue_flags);
6181 case IORING_OP_POLL_ADD:
6182 ret = io_poll_add(req, issue_flags);
6184 case IORING_OP_POLL_REMOVE:
6185 ret = io_poll_update(req, issue_flags);
6187 case IORING_OP_SYNC_FILE_RANGE:
6188 ret = io_sync_file_range(req, issue_flags);
6190 case IORING_OP_SENDMSG:
6191 ret = io_sendmsg(req, issue_flags);
6193 case IORING_OP_SEND:
6194 ret = io_send(req, issue_flags);
6196 case IORING_OP_RECVMSG:
6197 ret = io_recvmsg(req, issue_flags);
6199 case IORING_OP_RECV:
6200 ret = io_recv(req, issue_flags);
6202 case IORING_OP_TIMEOUT:
6203 ret = io_timeout(req, issue_flags);
6205 case IORING_OP_TIMEOUT_REMOVE:
6206 ret = io_timeout_remove(req, issue_flags);
6208 case IORING_OP_ACCEPT:
6209 ret = io_accept(req, issue_flags);
6211 case IORING_OP_CONNECT:
6212 ret = io_connect(req, issue_flags);
6214 case IORING_OP_ASYNC_CANCEL:
6215 ret = io_async_cancel(req, issue_flags);
6217 case IORING_OP_FALLOCATE:
6218 ret = io_fallocate(req, issue_flags);
6220 case IORING_OP_OPENAT:
6221 ret = io_openat(req, issue_flags);
6223 case IORING_OP_CLOSE:
6224 ret = io_close(req, issue_flags);
6226 case IORING_OP_FILES_UPDATE:
6227 ret = io_files_update(req, issue_flags);
6229 case IORING_OP_STATX:
6230 ret = io_statx(req, issue_flags);
6232 case IORING_OP_FADVISE:
6233 ret = io_fadvise(req, issue_flags);
6235 case IORING_OP_MADVISE:
6236 ret = io_madvise(req, issue_flags);
6238 case IORING_OP_OPENAT2:
6239 ret = io_openat2(req, issue_flags);
6241 case IORING_OP_EPOLL_CTL:
6242 ret = io_epoll_ctl(req, issue_flags);
6244 case IORING_OP_SPLICE:
6245 ret = io_splice(req, issue_flags);
6247 case IORING_OP_PROVIDE_BUFFERS:
6248 ret = io_provide_buffers(req, issue_flags);
6250 case IORING_OP_REMOVE_BUFFERS:
6251 ret = io_remove_buffers(req, issue_flags);
6254 ret = io_tee(req, issue_flags);
6256 case IORING_OP_SHUTDOWN:
6257 ret = io_shutdown(req, issue_flags);
6259 case IORING_OP_RENAMEAT:
6260 ret = io_renameat(req, issue_flags);
6262 case IORING_OP_UNLINKAT:
6263 ret = io_unlinkat(req, issue_flags);
6271 revert_creds(creds);
6274 /* If the op doesn't have a file, we're not polling for it */
6275 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6276 io_iopoll_req_issued(req);
6281 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6283 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6285 req = io_put_req_find_next(req);
6286 return req ? &req->work : NULL;
6289 static void io_wq_submit_work(struct io_wq_work *work)
6291 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6292 struct io_kiocb *timeout;
6295 io_req_refcount(req);
6296 /* will be dropped by ->io_free_work() after returning to io-wq */
6299 timeout = io_prep_linked_timeout(req);
6301 io_queue_linked_timeout(timeout);
6303 if (work->flags & IO_WQ_WORK_CANCEL)
6308 ret = io_issue_sqe(req, 0);
6310 * We can get EAGAIN for polled IO even though we're
6311 * forcing a sync submission from here, since we can't
6312 * wait for request slots on the block side.
6320 /* avoid locking problems by failing it from a clean context */
6322 io_req_task_queue_fail(req, ret);
6325 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6328 return &table->files[i];
6331 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6334 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6336 return (struct file *) (slot->file_ptr & FFS_MASK);
6339 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6341 unsigned long file_ptr = (unsigned long) file;
6343 if (__io_file_supports_nowait(file, READ))
6344 file_ptr |= FFS_ASYNC_READ;
6345 if (__io_file_supports_nowait(file, WRITE))
6346 file_ptr |= FFS_ASYNC_WRITE;
6347 if (S_ISREG(file_inode(file)->i_mode))
6348 file_ptr |= FFS_ISREG;
6349 file_slot->file_ptr = file_ptr;
6352 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6353 struct io_kiocb *req, int fd)
6356 unsigned long file_ptr;
6358 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6360 fd = array_index_nospec(fd, ctx->nr_user_files);
6361 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6362 file = (struct file *) (file_ptr & FFS_MASK);
6363 file_ptr &= ~FFS_MASK;
6364 /* mask in overlapping REQ_F and FFS bits */
6365 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6366 io_req_set_rsrc_node(req);
6370 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6371 struct io_kiocb *req, int fd)
6373 struct file *file = fget(fd);
6375 trace_io_uring_file_get(ctx, fd);
6377 /* we don't allow fixed io_uring files */
6378 if (file && unlikely(file->f_op == &io_uring_fops))
6379 io_req_track_inflight(req);
6383 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6384 struct io_kiocb *req, int fd, bool fixed)
6387 return io_file_get_fixed(ctx, req, fd);
6389 return io_file_get_normal(ctx, req, fd);
6392 static void io_req_task_link_timeout(struct io_kiocb *req)
6394 struct io_kiocb *prev = req->timeout.prev;
6395 struct io_ring_ctx *ctx = req->ctx;
6398 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6402 io_req_complete_post(req, -ETIME, 0);
6406 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6408 struct io_timeout_data *data = container_of(timer,
6409 struct io_timeout_data, timer);
6410 struct io_kiocb *prev, *req = data->req;
6411 struct io_ring_ctx *ctx = req->ctx;
6412 unsigned long flags;
6414 spin_lock_irqsave(&ctx->timeout_lock, flags);
6415 prev = req->timeout.head;
6416 req->timeout.head = NULL;
6419 * We don't expect the list to be empty, that will only happen if we
6420 * race with the completion of the linked work.
6423 io_remove_next_linked(prev);
6424 if (!req_ref_inc_not_zero(prev))
6427 req->timeout.prev = prev;
6428 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6430 req->io_task_work.func = io_req_task_link_timeout;
6431 io_req_task_work_add(req);
6432 return HRTIMER_NORESTART;
6435 static void io_queue_linked_timeout(struct io_kiocb *req)
6437 struct io_ring_ctx *ctx = req->ctx;
6439 spin_lock_irq(&ctx->timeout_lock);
6441 * If the back reference is NULL, then our linked request finished
6442 * before we got a chance to setup the timer
6444 if (req->timeout.head) {
6445 struct io_timeout_data *data = req->async_data;
6447 data->timer.function = io_link_timeout_fn;
6448 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6451 spin_unlock_irq(&ctx->timeout_lock);
6452 /* drop submission reference */
6456 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6458 struct io_kiocb *nxt = req->link;
6460 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6461 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6464 /* linked timeouts should have two refs once prep'ed */
6465 io_req_refcount(req);
6466 io_req_refcount(nxt);
6469 nxt->timeout.head = req;
6470 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6471 req->flags |= REQ_F_LINK_TIMEOUT;
6475 static void __io_queue_sqe(struct io_kiocb *req)
6476 __must_hold(&req->ctx->uring_lock)
6478 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6482 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6485 * We async punt it if the file wasn't marked NOWAIT, or if the file
6486 * doesn't support non-blocking read/write attempts
6489 if (req->flags & REQ_F_COMPLETE_INLINE) {
6490 struct io_ring_ctx *ctx = req->ctx;
6491 struct io_submit_state *state = &ctx->submit_state;
6493 state->compl_reqs[state->compl_nr++] = req;
6494 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6495 io_submit_flush_completions(ctx);
6497 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6498 switch (io_arm_poll_handler(req)) {
6499 case IO_APOLL_READY:
6501 case IO_APOLL_ABORTED:
6503 * Queued up for async execution, worker will release
6504 * submit reference when the iocb is actually submitted.
6506 io_queue_async_work(req);
6510 io_req_complete_failed(req, ret);
6513 io_queue_linked_timeout(linked_timeout);
6516 static inline void io_queue_sqe(struct io_kiocb *req)
6517 __must_hold(&req->ctx->uring_lock)
6519 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6522 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6523 __io_queue_sqe(req);
6525 int ret = io_req_prep_async(req);
6528 io_req_complete_failed(req, ret);
6530 io_queue_async_work(req);
6535 * Check SQE restrictions (opcode and flags).
6537 * Returns 'true' if SQE is allowed, 'false' otherwise.
6539 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6540 struct io_kiocb *req,
6541 unsigned int sqe_flags)
6543 if (likely(!ctx->restricted))
6546 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6549 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6550 ctx->restrictions.sqe_flags_required)
6553 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6554 ctx->restrictions.sqe_flags_required))
6560 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6561 const struct io_uring_sqe *sqe)
6562 __must_hold(&ctx->uring_lock)
6564 struct io_submit_state *state;
6565 unsigned int sqe_flags;
6566 int personality, ret = 0;
6568 /* req is partially pre-initialised, see io_preinit_req() */
6569 req->opcode = READ_ONCE(sqe->opcode);
6570 /* same numerical values with corresponding REQ_F_*, safe to copy */
6571 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6572 req->user_data = READ_ONCE(sqe->user_data);
6574 req->fixed_rsrc_refs = NULL;
6575 req->task = current;
6577 /* enforce forwards compatibility on users */
6578 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6580 if (unlikely(req->opcode >= IORING_OP_LAST))
6582 if (!io_check_restriction(ctx, req, sqe_flags))
6585 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6586 !io_op_defs[req->opcode].buffer_select)
6588 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6589 ctx->drain_active = true;
6591 personality = READ_ONCE(sqe->personality);
6593 req->creds = xa_load(&ctx->personalities, personality);
6596 get_cred(req->creds);
6597 req->flags |= REQ_F_CREDS;
6599 state = &ctx->submit_state;
6602 * Plug now if we have more than 1 IO left after this, and the target
6603 * is potentially a read/write to block based storage.
6605 if (!state->plug_started && state->ios_left > 1 &&
6606 io_op_defs[req->opcode].plug) {
6607 blk_start_plug(&state->plug);
6608 state->plug_started = true;
6611 if (io_op_defs[req->opcode].needs_file) {
6612 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6613 (sqe_flags & IOSQE_FIXED_FILE));
6614 if (unlikely(!req->file))
6622 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6623 const struct io_uring_sqe *sqe)
6624 __must_hold(&ctx->uring_lock)
6626 struct io_submit_link *link = &ctx->submit_state.link;
6629 ret = io_init_req(ctx, req, sqe);
6630 if (unlikely(ret)) {
6633 /* fail even hard links since we don't submit */
6634 req_set_fail(link->head);
6635 io_req_complete_failed(link->head, -ECANCELED);
6638 io_req_complete_failed(req, ret);
6642 ret = io_req_prep(req, sqe);
6646 /* don't need @sqe from now on */
6647 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6649 ctx->flags & IORING_SETUP_SQPOLL);
6652 * If we already have a head request, queue this one for async
6653 * submittal once the head completes. If we don't have a head but
6654 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6655 * submitted sync once the chain is complete. If none of those
6656 * conditions are true (normal request), then just queue it.
6659 struct io_kiocb *head = link->head;
6661 ret = io_req_prep_async(req);
6664 trace_io_uring_link(ctx, req, head);
6665 link->last->link = req;
6668 /* last request of a link, enqueue the link */
6669 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6674 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6686 * Batched submission is done, ensure local IO is flushed out.
6688 static void io_submit_state_end(struct io_submit_state *state,
6689 struct io_ring_ctx *ctx)
6691 if (state->link.head)
6692 io_queue_sqe(state->link.head);
6693 if (state->compl_nr)
6694 io_submit_flush_completions(ctx);
6695 if (state->plug_started)
6696 blk_finish_plug(&state->plug);
6700 * Start submission side cache.
6702 static void io_submit_state_start(struct io_submit_state *state,
6703 unsigned int max_ios)
6705 state->plug_started = false;
6706 state->ios_left = max_ios;
6707 /* set only head, no need to init link_last in advance */
6708 state->link.head = NULL;
6711 static void io_commit_sqring(struct io_ring_ctx *ctx)
6713 struct io_rings *rings = ctx->rings;
6716 * Ensure any loads from the SQEs are done at this point,
6717 * since once we write the new head, the application could
6718 * write new data to them.
6720 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6724 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6725 * that is mapped by userspace. This means that care needs to be taken to
6726 * ensure that reads are stable, as we cannot rely on userspace always
6727 * being a good citizen. If members of the sqe are validated and then later
6728 * used, it's important that those reads are done through READ_ONCE() to
6729 * prevent a re-load down the line.
6731 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6733 unsigned head, mask = ctx->sq_entries - 1;
6734 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6737 * The cached sq head (or cq tail) serves two purposes:
6739 * 1) allows us to batch the cost of updating the user visible
6741 * 2) allows the kernel side to track the head on its own, even
6742 * though the application is the one updating it.
6744 head = READ_ONCE(ctx->sq_array[sq_idx]);
6745 if (likely(head < ctx->sq_entries))
6746 return &ctx->sq_sqes[head];
6748 /* drop invalid entries */
6750 WRITE_ONCE(ctx->rings->sq_dropped,
6751 READ_ONCE(ctx->rings->sq_dropped) + 1);
6755 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6756 __must_hold(&ctx->uring_lock)
6758 struct io_uring_task *tctx;
6761 /* make sure SQ entry isn't read before tail */
6762 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6763 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6766 tctx = current->io_uring;
6767 tctx->cached_refs -= nr;
6768 if (unlikely(tctx->cached_refs < 0)) {
6769 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6771 percpu_counter_add(&tctx->inflight, refill);
6772 refcount_add(refill, ¤t->usage);
6773 tctx->cached_refs += refill;
6775 io_submit_state_start(&ctx->submit_state, nr);
6777 while (submitted < nr) {
6778 const struct io_uring_sqe *sqe;
6779 struct io_kiocb *req;
6781 req = io_alloc_req(ctx);
6782 if (unlikely(!req)) {
6784 submitted = -EAGAIN;
6787 sqe = io_get_sqe(ctx);
6788 if (unlikely(!sqe)) {
6789 kmem_cache_free(req_cachep, req);
6792 /* will complete beyond this point, count as submitted */
6794 if (io_submit_sqe(ctx, req, sqe))
6798 if (unlikely(submitted != nr)) {
6799 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6800 int unused = nr - ref_used;
6802 current->io_uring->cached_refs += unused;
6803 percpu_ref_put_many(&ctx->refs, unused);
6806 io_submit_state_end(&ctx->submit_state, ctx);
6807 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6808 io_commit_sqring(ctx);
6813 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6815 return READ_ONCE(sqd->state);
6818 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6820 /* Tell userspace we may need a wakeup call */
6821 spin_lock(&ctx->completion_lock);
6822 WRITE_ONCE(ctx->rings->sq_flags,
6823 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6824 spin_unlock(&ctx->completion_lock);
6827 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6829 spin_lock(&ctx->completion_lock);
6830 WRITE_ONCE(ctx->rings->sq_flags,
6831 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6832 spin_unlock(&ctx->completion_lock);
6835 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6837 unsigned int to_submit;
6840 to_submit = io_sqring_entries(ctx);
6841 /* if we're handling multiple rings, cap submit size for fairness */
6842 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6843 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6845 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6846 unsigned nr_events = 0;
6847 const struct cred *creds = NULL;
6849 if (ctx->sq_creds != current_cred())
6850 creds = override_creds(ctx->sq_creds);
6852 mutex_lock(&ctx->uring_lock);
6853 if (!list_empty(&ctx->iopoll_list))
6854 io_do_iopoll(ctx, &nr_events, 0, true);
6857 * Don't submit if refs are dying, good for io_uring_register(),
6858 * but also it is relied upon by io_ring_exit_work()
6860 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6861 !(ctx->flags & IORING_SETUP_R_DISABLED))
6862 ret = io_submit_sqes(ctx, to_submit);
6863 mutex_unlock(&ctx->uring_lock);
6865 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6866 wake_up(&ctx->sqo_sq_wait);
6868 revert_creds(creds);
6874 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6876 struct io_ring_ctx *ctx;
6877 unsigned sq_thread_idle = 0;
6879 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6880 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6881 sqd->sq_thread_idle = sq_thread_idle;
6884 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6886 bool did_sig = false;
6887 struct ksignal ksig;
6889 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6890 signal_pending(current)) {
6891 mutex_unlock(&sqd->lock);
6892 if (signal_pending(current))
6893 did_sig = get_signal(&ksig);
6895 mutex_lock(&sqd->lock);
6897 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6900 static int io_sq_thread(void *data)
6902 struct io_sq_data *sqd = data;
6903 struct io_ring_ctx *ctx;
6904 unsigned long timeout = 0;
6905 char buf[TASK_COMM_LEN];
6908 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6909 set_task_comm(current, buf);
6911 if (sqd->sq_cpu != -1)
6912 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6914 set_cpus_allowed_ptr(current, cpu_online_mask);
6915 current->flags |= PF_NO_SETAFFINITY;
6917 mutex_lock(&sqd->lock);
6919 bool cap_entries, sqt_spin = false;
6921 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6922 if (io_sqd_handle_event(sqd))
6924 timeout = jiffies + sqd->sq_thread_idle;
6927 cap_entries = !list_is_singular(&sqd->ctx_list);
6928 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6929 int ret = __io_sq_thread(ctx, cap_entries);
6931 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6934 if (io_run_task_work())
6937 if (sqt_spin || !time_after(jiffies, timeout)) {
6940 timeout = jiffies + sqd->sq_thread_idle;
6944 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6945 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6946 bool needs_sched = true;
6948 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6949 io_ring_set_wakeup_flag(ctx);
6951 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6952 !list_empty_careful(&ctx->iopoll_list)) {
6953 needs_sched = false;
6956 if (io_sqring_entries(ctx)) {
6957 needs_sched = false;
6963 mutex_unlock(&sqd->lock);
6965 mutex_lock(&sqd->lock);
6967 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6968 io_ring_clear_wakeup_flag(ctx);
6971 finish_wait(&sqd->wait, &wait);
6972 timeout = jiffies + sqd->sq_thread_idle;
6975 io_uring_cancel_generic(true, sqd);
6977 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6978 io_ring_set_wakeup_flag(ctx);
6980 mutex_unlock(&sqd->lock);
6982 complete(&sqd->exited);
6986 struct io_wait_queue {
6987 struct wait_queue_entry wq;
6988 struct io_ring_ctx *ctx;
6990 unsigned nr_timeouts;
6993 static inline bool io_should_wake(struct io_wait_queue *iowq)
6995 struct io_ring_ctx *ctx = iowq->ctx;
6996 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
6999 * Wake up if we have enough events, or if a timeout occurred since we
7000 * started waiting. For timeouts, we always want to return to userspace,
7001 * regardless of event count.
7003 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7006 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7007 int wake_flags, void *key)
7009 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7013 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7014 * the task, and the next invocation will do it.
7016 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7017 return autoremove_wake_function(curr, mode, wake_flags, key);
7021 static int io_run_task_work_sig(void)
7023 if (io_run_task_work())
7025 if (!signal_pending(current))
7027 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7028 return -ERESTARTSYS;
7032 /* when returns >0, the caller should retry */
7033 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7034 struct io_wait_queue *iowq,
7035 signed long *timeout)
7039 /* make sure we run task_work before checking for signals */
7040 ret = io_run_task_work_sig();
7041 if (ret || io_should_wake(iowq))
7043 /* let the caller flush overflows, retry */
7044 if (test_bit(0, &ctx->check_cq_overflow))
7047 *timeout = schedule_timeout(*timeout);
7048 return !*timeout ? -ETIME : 1;
7052 * Wait until events become available, if we don't already have some. The
7053 * application must reap them itself, as they reside on the shared cq ring.
7055 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7056 const sigset_t __user *sig, size_t sigsz,
7057 struct __kernel_timespec __user *uts)
7059 struct io_wait_queue iowq;
7060 struct io_rings *rings = ctx->rings;
7061 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7065 io_cqring_overflow_flush(ctx);
7066 if (io_cqring_events(ctx) >= min_events)
7068 if (!io_run_task_work())
7073 #ifdef CONFIG_COMPAT
7074 if (in_compat_syscall())
7075 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7079 ret = set_user_sigmask(sig, sigsz);
7086 struct timespec64 ts;
7088 if (get_timespec64(&ts, uts))
7090 timeout = timespec64_to_jiffies(&ts);
7093 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7094 iowq.wq.private = current;
7095 INIT_LIST_HEAD(&iowq.wq.entry);
7097 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7098 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7100 trace_io_uring_cqring_wait(ctx, min_events);
7102 /* if we can't even flush overflow, don't wait for more */
7103 if (!io_cqring_overflow_flush(ctx)) {
7107 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7108 TASK_INTERRUPTIBLE);
7109 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7110 finish_wait(&ctx->cq_wait, &iowq.wq);
7114 restore_saved_sigmask_unless(ret == -EINTR);
7116 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7119 static void io_free_page_table(void **table, size_t size)
7121 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7123 for (i = 0; i < nr_tables; i++)
7128 static void **io_alloc_page_table(size_t size)
7130 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7131 size_t init_size = size;
7134 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7138 for (i = 0; i < nr_tables; i++) {
7139 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7141 table[i] = kzalloc(this_size, GFP_KERNEL);
7143 io_free_page_table(table, init_size);
7151 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7153 percpu_ref_exit(&ref_node->refs);
7157 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7159 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7160 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7161 unsigned long flags;
7162 bool first_add = false;
7164 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7167 while (!list_empty(&ctx->rsrc_ref_list)) {
7168 node = list_first_entry(&ctx->rsrc_ref_list,
7169 struct io_rsrc_node, node);
7170 /* recycle ref nodes in order */
7173 list_del(&node->node);
7174 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7176 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7179 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7182 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7184 struct io_rsrc_node *ref_node;
7186 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7190 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7195 INIT_LIST_HEAD(&ref_node->node);
7196 INIT_LIST_HEAD(&ref_node->rsrc_list);
7197 ref_node->done = false;
7201 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7202 struct io_rsrc_data *data_to_kill)
7204 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7205 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7208 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7210 rsrc_node->rsrc_data = data_to_kill;
7211 spin_lock_irq(&ctx->rsrc_ref_lock);
7212 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7213 spin_unlock_irq(&ctx->rsrc_ref_lock);
7215 atomic_inc(&data_to_kill->refs);
7216 percpu_ref_kill(&rsrc_node->refs);
7217 ctx->rsrc_node = NULL;
7220 if (!ctx->rsrc_node) {
7221 ctx->rsrc_node = ctx->rsrc_backup_node;
7222 ctx->rsrc_backup_node = NULL;
7226 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7228 if (ctx->rsrc_backup_node)
7230 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7231 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7234 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7238 /* As we may drop ->uring_lock, other task may have started quiesce */
7242 data->quiesce = true;
7244 ret = io_rsrc_node_switch_start(ctx);
7247 io_rsrc_node_switch(ctx, data);
7249 /* kill initial ref, already quiesced if zero */
7250 if (atomic_dec_and_test(&data->refs))
7252 mutex_unlock(&ctx->uring_lock);
7253 flush_delayed_work(&ctx->rsrc_put_work);
7254 ret = wait_for_completion_interruptible(&data->done);
7256 mutex_lock(&ctx->uring_lock);
7260 atomic_inc(&data->refs);
7261 /* wait for all works potentially completing data->done */
7262 flush_delayed_work(&ctx->rsrc_put_work);
7263 reinit_completion(&data->done);
7265 ret = io_run_task_work_sig();
7266 mutex_lock(&ctx->uring_lock);
7268 data->quiesce = false;
7273 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7275 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7276 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7278 return &data->tags[table_idx][off];
7281 static void io_rsrc_data_free(struct io_rsrc_data *data)
7283 size_t size = data->nr * sizeof(data->tags[0][0]);
7286 io_free_page_table((void **)data->tags, size);
7290 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7291 u64 __user *utags, unsigned nr,
7292 struct io_rsrc_data **pdata)
7294 struct io_rsrc_data *data;
7298 data = kzalloc(sizeof(*data), GFP_KERNEL);
7301 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7309 data->do_put = do_put;
7312 for (i = 0; i < nr; i++) {
7313 u64 *tag_slot = io_get_tag_slot(data, i);
7315 if (copy_from_user(tag_slot, &utags[i],
7321 atomic_set(&data->refs, 1);
7322 init_completion(&data->done);
7326 io_rsrc_data_free(data);
7330 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7332 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7333 return !!table->files;
7336 static void io_free_file_tables(struct io_file_table *table)
7338 kvfree(table->files);
7339 table->files = NULL;
7342 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7344 #if defined(CONFIG_UNIX)
7345 if (ctx->ring_sock) {
7346 struct sock *sock = ctx->ring_sock->sk;
7347 struct sk_buff *skb;
7349 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7355 for (i = 0; i < ctx->nr_user_files; i++) {
7358 file = io_file_from_index(ctx, i);
7363 io_free_file_tables(&ctx->file_table);
7364 io_rsrc_data_free(ctx->file_data);
7365 ctx->file_data = NULL;
7366 ctx->nr_user_files = 0;
7369 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7373 if (!ctx->file_data)
7375 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7377 __io_sqe_files_unregister(ctx);
7381 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7382 __releases(&sqd->lock)
7384 WARN_ON_ONCE(sqd->thread == current);
7387 * Do the dance but not conditional clear_bit() because it'd race with
7388 * other threads incrementing park_pending and setting the bit.
7390 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7391 if (atomic_dec_return(&sqd->park_pending))
7392 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7393 mutex_unlock(&sqd->lock);
7396 static void io_sq_thread_park(struct io_sq_data *sqd)
7397 __acquires(&sqd->lock)
7399 WARN_ON_ONCE(sqd->thread == current);
7401 atomic_inc(&sqd->park_pending);
7402 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7403 mutex_lock(&sqd->lock);
7405 wake_up_process(sqd->thread);
7408 static void io_sq_thread_stop(struct io_sq_data *sqd)
7410 WARN_ON_ONCE(sqd->thread == current);
7411 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7413 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7414 mutex_lock(&sqd->lock);
7416 wake_up_process(sqd->thread);
7417 mutex_unlock(&sqd->lock);
7418 wait_for_completion(&sqd->exited);
7421 static void io_put_sq_data(struct io_sq_data *sqd)
7423 if (refcount_dec_and_test(&sqd->refs)) {
7424 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7426 io_sq_thread_stop(sqd);
7431 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7433 struct io_sq_data *sqd = ctx->sq_data;
7436 io_sq_thread_park(sqd);
7437 list_del_init(&ctx->sqd_list);
7438 io_sqd_update_thread_idle(sqd);
7439 io_sq_thread_unpark(sqd);
7441 io_put_sq_data(sqd);
7442 ctx->sq_data = NULL;
7446 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7448 struct io_ring_ctx *ctx_attach;
7449 struct io_sq_data *sqd;
7452 f = fdget(p->wq_fd);
7454 return ERR_PTR(-ENXIO);
7455 if (f.file->f_op != &io_uring_fops) {
7457 return ERR_PTR(-EINVAL);
7460 ctx_attach = f.file->private_data;
7461 sqd = ctx_attach->sq_data;
7464 return ERR_PTR(-EINVAL);
7466 if (sqd->task_tgid != current->tgid) {
7468 return ERR_PTR(-EPERM);
7471 refcount_inc(&sqd->refs);
7476 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7479 struct io_sq_data *sqd;
7482 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7483 sqd = io_attach_sq_data(p);
7488 /* fall through for EPERM case, setup new sqd/task */
7489 if (PTR_ERR(sqd) != -EPERM)
7493 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7495 return ERR_PTR(-ENOMEM);
7497 atomic_set(&sqd->park_pending, 0);
7498 refcount_set(&sqd->refs, 1);
7499 INIT_LIST_HEAD(&sqd->ctx_list);
7500 mutex_init(&sqd->lock);
7501 init_waitqueue_head(&sqd->wait);
7502 init_completion(&sqd->exited);
7506 #if defined(CONFIG_UNIX)
7508 * Ensure the UNIX gc is aware of our file set, so we are certain that
7509 * the io_uring can be safely unregistered on process exit, even if we have
7510 * loops in the file referencing.
7512 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7514 struct sock *sk = ctx->ring_sock->sk;
7515 struct scm_fp_list *fpl;
7516 struct sk_buff *skb;
7519 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7523 skb = alloc_skb(0, GFP_KERNEL);
7532 fpl->user = get_uid(current_user());
7533 for (i = 0; i < nr; i++) {
7534 struct file *file = io_file_from_index(ctx, i + offset);
7538 fpl->fp[nr_files] = get_file(file);
7539 unix_inflight(fpl->user, fpl->fp[nr_files]);
7544 fpl->max = SCM_MAX_FD;
7545 fpl->count = nr_files;
7546 UNIXCB(skb).fp = fpl;
7547 skb->destructor = unix_destruct_scm;
7548 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7549 skb_queue_head(&sk->sk_receive_queue, skb);
7551 for (i = 0; i < nr_files; i++)
7562 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7563 * causes regular reference counting to break down. We rely on the UNIX
7564 * garbage collection to take care of this problem for us.
7566 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7568 unsigned left, total;
7572 left = ctx->nr_user_files;
7574 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7576 ret = __io_sqe_files_scm(ctx, this_files, total);
7580 total += this_files;
7586 while (total < ctx->nr_user_files) {
7587 struct file *file = io_file_from_index(ctx, total);
7597 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7603 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7605 struct file *file = prsrc->file;
7606 #if defined(CONFIG_UNIX)
7607 struct sock *sock = ctx->ring_sock->sk;
7608 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7609 struct sk_buff *skb;
7612 __skb_queue_head_init(&list);
7615 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7616 * remove this entry and rearrange the file array.
7618 skb = skb_dequeue(head);
7620 struct scm_fp_list *fp;
7622 fp = UNIXCB(skb).fp;
7623 for (i = 0; i < fp->count; i++) {
7626 if (fp->fp[i] != file)
7629 unix_notinflight(fp->user, fp->fp[i]);
7630 left = fp->count - 1 - i;
7632 memmove(&fp->fp[i], &fp->fp[i + 1],
7633 left * sizeof(struct file *));
7640 __skb_queue_tail(&list, skb);
7650 __skb_queue_tail(&list, skb);
7652 skb = skb_dequeue(head);
7655 if (skb_peek(&list)) {
7656 spin_lock_irq(&head->lock);
7657 while ((skb = __skb_dequeue(&list)) != NULL)
7658 __skb_queue_tail(head, skb);
7659 spin_unlock_irq(&head->lock);
7666 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7668 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7669 struct io_ring_ctx *ctx = rsrc_data->ctx;
7670 struct io_rsrc_put *prsrc, *tmp;
7672 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7673 list_del(&prsrc->list);
7676 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7678 io_ring_submit_lock(ctx, lock_ring);
7679 spin_lock(&ctx->completion_lock);
7680 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7682 io_commit_cqring(ctx);
7683 spin_unlock(&ctx->completion_lock);
7684 io_cqring_ev_posted(ctx);
7685 io_ring_submit_unlock(ctx, lock_ring);
7688 rsrc_data->do_put(ctx, prsrc);
7692 io_rsrc_node_destroy(ref_node);
7693 if (atomic_dec_and_test(&rsrc_data->refs))
7694 complete(&rsrc_data->done);
7697 static void io_rsrc_put_work(struct work_struct *work)
7699 struct io_ring_ctx *ctx;
7700 struct llist_node *node;
7702 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7703 node = llist_del_all(&ctx->rsrc_put_llist);
7706 struct io_rsrc_node *ref_node;
7707 struct llist_node *next = node->next;
7709 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7710 __io_rsrc_put_work(ref_node);
7715 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7716 unsigned nr_args, u64 __user *tags)
7718 __s32 __user *fds = (__s32 __user *) arg;
7727 if (nr_args > IORING_MAX_FIXED_FILES)
7729 ret = io_rsrc_node_switch_start(ctx);
7732 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7738 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7741 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7742 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7746 /* allow sparse sets */
7749 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7756 if (unlikely(!file))
7760 * Don't allow io_uring instances to be registered. If UNIX
7761 * isn't enabled, then this causes a reference cycle and this
7762 * instance can never get freed. If UNIX is enabled we'll
7763 * handle it just fine, but there's still no point in allowing
7764 * a ring fd as it doesn't support regular read/write anyway.
7766 if (file->f_op == &io_uring_fops) {
7770 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7773 ret = io_sqe_files_scm(ctx);
7775 __io_sqe_files_unregister(ctx);
7779 io_rsrc_node_switch(ctx, NULL);
7782 for (i = 0; i < ctx->nr_user_files; i++) {
7783 file = io_file_from_index(ctx, i);
7787 io_free_file_tables(&ctx->file_table);
7788 ctx->nr_user_files = 0;
7790 io_rsrc_data_free(ctx->file_data);
7791 ctx->file_data = NULL;
7795 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7798 #if defined(CONFIG_UNIX)
7799 struct sock *sock = ctx->ring_sock->sk;
7800 struct sk_buff_head *head = &sock->sk_receive_queue;
7801 struct sk_buff *skb;
7804 * See if we can merge this file into an existing skb SCM_RIGHTS
7805 * file set. If there's no room, fall back to allocating a new skb
7806 * and filling it in.
7808 spin_lock_irq(&head->lock);
7809 skb = skb_peek(head);
7811 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7813 if (fpl->count < SCM_MAX_FD) {
7814 __skb_unlink(skb, head);
7815 spin_unlock_irq(&head->lock);
7816 fpl->fp[fpl->count] = get_file(file);
7817 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7819 spin_lock_irq(&head->lock);
7820 __skb_queue_head(head, skb);
7825 spin_unlock_irq(&head->lock);
7832 return __io_sqe_files_scm(ctx, 1, index);
7838 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7839 struct io_rsrc_node *node, void *rsrc)
7841 struct io_rsrc_put *prsrc;
7843 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7847 prsrc->tag = *io_get_tag_slot(data, idx);
7849 list_add(&prsrc->list, &node->rsrc_list);
7853 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7854 struct io_uring_rsrc_update2 *up,
7857 u64 __user *tags = u64_to_user_ptr(up->tags);
7858 __s32 __user *fds = u64_to_user_ptr(up->data);
7859 struct io_rsrc_data *data = ctx->file_data;
7860 struct io_fixed_file *file_slot;
7864 bool needs_switch = false;
7866 if (!ctx->file_data)
7868 if (up->offset + nr_args > ctx->nr_user_files)
7871 for (done = 0; done < nr_args; done++) {
7874 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7875 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7879 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7883 if (fd == IORING_REGISTER_FILES_SKIP)
7886 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7887 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7889 if (file_slot->file_ptr) {
7890 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7891 err = io_queue_rsrc_removal(data, up->offset + done,
7892 ctx->rsrc_node, file);
7895 file_slot->file_ptr = 0;
7896 needs_switch = true;
7905 * Don't allow io_uring instances to be registered. If
7906 * UNIX isn't enabled, then this causes a reference
7907 * cycle and this instance can never get freed. If UNIX
7908 * is enabled we'll handle it just fine, but there's
7909 * still no point in allowing a ring fd as it doesn't
7910 * support regular read/write anyway.
7912 if (file->f_op == &io_uring_fops) {
7917 *io_get_tag_slot(data, up->offset + done) = tag;
7918 io_fixed_file_set(file_slot, file);
7919 err = io_sqe_file_register(ctx, file, i);
7921 file_slot->file_ptr = 0;
7929 io_rsrc_node_switch(ctx, data);
7930 return done ? done : err;
7933 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7934 struct task_struct *task)
7936 struct io_wq_hash *hash;
7937 struct io_wq_data data;
7938 unsigned int concurrency;
7940 mutex_lock(&ctx->uring_lock);
7941 hash = ctx->hash_map;
7943 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7945 mutex_unlock(&ctx->uring_lock);
7946 return ERR_PTR(-ENOMEM);
7948 refcount_set(&hash->refs, 1);
7949 init_waitqueue_head(&hash->wait);
7950 ctx->hash_map = hash;
7952 mutex_unlock(&ctx->uring_lock);
7956 data.free_work = io_wq_free_work;
7957 data.do_work = io_wq_submit_work;
7959 /* Do QD, or 4 * CPUS, whatever is smallest */
7960 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7962 return io_wq_create(concurrency, &data);
7965 static int io_uring_alloc_task_context(struct task_struct *task,
7966 struct io_ring_ctx *ctx)
7968 struct io_uring_task *tctx;
7971 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7972 if (unlikely(!tctx))
7975 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7976 if (unlikely(ret)) {
7981 tctx->io_wq = io_init_wq_offload(ctx, task);
7982 if (IS_ERR(tctx->io_wq)) {
7983 ret = PTR_ERR(tctx->io_wq);
7984 percpu_counter_destroy(&tctx->inflight);
7990 init_waitqueue_head(&tctx->wait);
7991 atomic_set(&tctx->in_idle, 0);
7992 atomic_set(&tctx->inflight_tracked, 0);
7993 task->io_uring = tctx;
7994 spin_lock_init(&tctx->task_lock);
7995 INIT_WQ_LIST(&tctx->task_list);
7996 init_task_work(&tctx->task_work, tctx_task_work);
8000 void __io_uring_free(struct task_struct *tsk)
8002 struct io_uring_task *tctx = tsk->io_uring;
8004 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8005 WARN_ON_ONCE(tctx->io_wq);
8006 WARN_ON_ONCE(tctx->cached_refs);
8008 percpu_counter_destroy(&tctx->inflight);
8010 tsk->io_uring = NULL;
8013 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8014 struct io_uring_params *p)
8018 /* Retain compatibility with failing for an invalid attach attempt */
8019 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8020 IORING_SETUP_ATTACH_WQ) {
8023 f = fdget(p->wq_fd);
8026 if (f.file->f_op != &io_uring_fops) {
8032 if (ctx->flags & IORING_SETUP_SQPOLL) {
8033 struct task_struct *tsk;
8034 struct io_sq_data *sqd;
8037 sqd = io_get_sq_data(p, &attached);
8043 ctx->sq_creds = get_current_cred();
8045 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8046 if (!ctx->sq_thread_idle)
8047 ctx->sq_thread_idle = HZ;
8049 io_sq_thread_park(sqd);
8050 list_add(&ctx->sqd_list, &sqd->ctx_list);
8051 io_sqd_update_thread_idle(sqd);
8052 /* don't attach to a dying SQPOLL thread, would be racy */
8053 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8054 io_sq_thread_unpark(sqd);
8061 if (p->flags & IORING_SETUP_SQ_AFF) {
8062 int cpu = p->sq_thread_cpu;
8065 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8072 sqd->task_pid = current->pid;
8073 sqd->task_tgid = current->tgid;
8074 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8081 ret = io_uring_alloc_task_context(tsk, ctx);
8082 wake_up_new_task(tsk);
8085 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8086 /* Can't have SQ_AFF without SQPOLL */
8093 complete(&ctx->sq_data->exited);
8095 io_sq_thread_finish(ctx);
8099 static inline void __io_unaccount_mem(struct user_struct *user,
8100 unsigned long nr_pages)
8102 atomic_long_sub(nr_pages, &user->locked_vm);
8105 static inline int __io_account_mem(struct user_struct *user,
8106 unsigned long nr_pages)
8108 unsigned long page_limit, cur_pages, new_pages;
8110 /* Don't allow more pages than we can safely lock */
8111 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8114 cur_pages = atomic_long_read(&user->locked_vm);
8115 new_pages = cur_pages + nr_pages;
8116 if (new_pages > page_limit)
8118 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8119 new_pages) != cur_pages);
8124 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8127 __io_unaccount_mem(ctx->user, nr_pages);
8129 if (ctx->mm_account)
8130 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8133 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8138 ret = __io_account_mem(ctx->user, nr_pages);
8143 if (ctx->mm_account)
8144 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8149 static void io_mem_free(void *ptr)
8156 page = virt_to_head_page(ptr);
8157 if (put_page_testzero(page))
8158 free_compound_page(page);
8161 static void *io_mem_alloc(size_t size)
8163 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8164 __GFP_NORETRY | __GFP_ACCOUNT;
8166 return (void *) __get_free_pages(gfp_flags, get_order(size));
8169 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8172 struct io_rings *rings;
8173 size_t off, sq_array_size;
8175 off = struct_size(rings, cqes, cq_entries);
8176 if (off == SIZE_MAX)
8180 off = ALIGN(off, SMP_CACHE_BYTES);
8188 sq_array_size = array_size(sizeof(u32), sq_entries);
8189 if (sq_array_size == SIZE_MAX)
8192 if (check_add_overflow(off, sq_array_size, &off))
8198 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8200 struct io_mapped_ubuf *imu = *slot;
8203 if (imu != ctx->dummy_ubuf) {
8204 for (i = 0; i < imu->nr_bvecs; i++)
8205 unpin_user_page(imu->bvec[i].bv_page);
8206 if (imu->acct_pages)
8207 io_unaccount_mem(ctx, imu->acct_pages);
8213 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8215 io_buffer_unmap(ctx, &prsrc->buf);
8219 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8223 for (i = 0; i < ctx->nr_user_bufs; i++)
8224 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8225 kfree(ctx->user_bufs);
8226 io_rsrc_data_free(ctx->buf_data);
8227 ctx->user_bufs = NULL;
8228 ctx->buf_data = NULL;
8229 ctx->nr_user_bufs = 0;
8232 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8239 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8241 __io_sqe_buffers_unregister(ctx);
8245 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8246 void __user *arg, unsigned index)
8248 struct iovec __user *src;
8250 #ifdef CONFIG_COMPAT
8252 struct compat_iovec __user *ciovs;
8253 struct compat_iovec ciov;
8255 ciovs = (struct compat_iovec __user *) arg;
8256 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8259 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8260 dst->iov_len = ciov.iov_len;
8264 src = (struct iovec __user *) arg;
8265 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8271 * Not super efficient, but this is just a registration time. And we do cache
8272 * the last compound head, so generally we'll only do a full search if we don't
8275 * We check if the given compound head page has already been accounted, to
8276 * avoid double accounting it. This allows us to account the full size of the
8277 * page, not just the constituent pages of a huge page.
8279 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8280 int nr_pages, struct page *hpage)
8284 /* check current page array */
8285 for (i = 0; i < nr_pages; i++) {
8286 if (!PageCompound(pages[i]))
8288 if (compound_head(pages[i]) == hpage)
8292 /* check previously registered pages */
8293 for (i = 0; i < ctx->nr_user_bufs; i++) {
8294 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8296 for (j = 0; j < imu->nr_bvecs; j++) {
8297 if (!PageCompound(imu->bvec[j].bv_page))
8299 if (compound_head(imu->bvec[j].bv_page) == hpage)
8307 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8308 int nr_pages, struct io_mapped_ubuf *imu,
8309 struct page **last_hpage)
8313 imu->acct_pages = 0;
8314 for (i = 0; i < nr_pages; i++) {
8315 if (!PageCompound(pages[i])) {
8320 hpage = compound_head(pages[i]);
8321 if (hpage == *last_hpage)
8323 *last_hpage = hpage;
8324 if (headpage_already_acct(ctx, pages, i, hpage))
8326 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8330 if (!imu->acct_pages)
8333 ret = io_account_mem(ctx, imu->acct_pages);
8335 imu->acct_pages = 0;
8339 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8340 struct io_mapped_ubuf **pimu,
8341 struct page **last_hpage)
8343 struct io_mapped_ubuf *imu = NULL;
8344 struct vm_area_struct **vmas = NULL;
8345 struct page **pages = NULL;
8346 unsigned long off, start, end, ubuf;
8348 int ret, pret, nr_pages, i;
8350 if (!iov->iov_base) {
8351 *pimu = ctx->dummy_ubuf;
8355 ubuf = (unsigned long) iov->iov_base;
8356 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8357 start = ubuf >> PAGE_SHIFT;
8358 nr_pages = end - start;
8363 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8367 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8372 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8377 mmap_read_lock(current->mm);
8378 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8380 if (pret == nr_pages) {
8381 /* don't support file backed memory */
8382 for (i = 0; i < nr_pages; i++) {
8383 struct vm_area_struct *vma = vmas[i];
8385 if (vma_is_shmem(vma))
8388 !is_file_hugepages(vma->vm_file)) {
8394 ret = pret < 0 ? pret : -EFAULT;
8396 mmap_read_unlock(current->mm);
8399 * if we did partial map, or found file backed vmas,
8400 * release any pages we did get
8403 unpin_user_pages(pages, pret);
8407 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8409 unpin_user_pages(pages, pret);
8413 off = ubuf & ~PAGE_MASK;
8414 size = iov->iov_len;
8415 for (i = 0; i < nr_pages; i++) {
8418 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8419 imu->bvec[i].bv_page = pages[i];
8420 imu->bvec[i].bv_len = vec_len;
8421 imu->bvec[i].bv_offset = off;
8425 /* store original address for later verification */
8427 imu->ubuf_end = ubuf + iov->iov_len;
8428 imu->nr_bvecs = nr_pages;
8439 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8441 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8442 return ctx->user_bufs ? 0 : -ENOMEM;
8445 static int io_buffer_validate(struct iovec *iov)
8447 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8450 * Don't impose further limits on the size and buffer
8451 * constraints here, we'll -EINVAL later when IO is
8452 * submitted if they are wrong.
8455 return iov->iov_len ? -EFAULT : 0;
8459 /* arbitrary limit, but we need something */
8460 if (iov->iov_len > SZ_1G)
8463 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8469 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8470 unsigned int nr_args, u64 __user *tags)
8472 struct page *last_hpage = NULL;
8473 struct io_rsrc_data *data;
8479 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8481 ret = io_rsrc_node_switch_start(ctx);
8484 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8487 ret = io_buffers_map_alloc(ctx, nr_args);
8489 io_rsrc_data_free(data);
8493 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8494 ret = io_copy_iov(ctx, &iov, arg, i);
8497 ret = io_buffer_validate(&iov);
8500 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8505 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8511 WARN_ON_ONCE(ctx->buf_data);
8513 ctx->buf_data = data;
8515 __io_sqe_buffers_unregister(ctx);
8517 io_rsrc_node_switch(ctx, NULL);
8521 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8522 struct io_uring_rsrc_update2 *up,
8523 unsigned int nr_args)
8525 u64 __user *tags = u64_to_user_ptr(up->tags);
8526 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8527 struct page *last_hpage = NULL;
8528 bool needs_switch = false;
8534 if (up->offset + nr_args > ctx->nr_user_bufs)
8537 for (done = 0; done < nr_args; done++) {
8538 struct io_mapped_ubuf *imu;
8539 int offset = up->offset + done;
8542 err = io_copy_iov(ctx, &iov, iovs, done);
8545 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8549 err = io_buffer_validate(&iov);
8552 if (!iov.iov_base && tag) {
8556 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8560 i = array_index_nospec(offset, ctx->nr_user_bufs);
8561 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8562 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8563 ctx->rsrc_node, ctx->user_bufs[i]);
8564 if (unlikely(err)) {
8565 io_buffer_unmap(ctx, &imu);
8568 ctx->user_bufs[i] = NULL;
8569 needs_switch = true;
8572 ctx->user_bufs[i] = imu;
8573 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8577 io_rsrc_node_switch(ctx, ctx->buf_data);
8578 return done ? done : err;
8581 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8583 __s32 __user *fds = arg;
8589 if (copy_from_user(&fd, fds, sizeof(*fds)))
8592 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8593 if (IS_ERR(ctx->cq_ev_fd)) {
8594 int ret = PTR_ERR(ctx->cq_ev_fd);
8596 ctx->cq_ev_fd = NULL;
8603 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8605 if (ctx->cq_ev_fd) {
8606 eventfd_ctx_put(ctx->cq_ev_fd);
8607 ctx->cq_ev_fd = NULL;
8614 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8616 struct io_buffer *buf;
8617 unsigned long index;
8619 xa_for_each(&ctx->io_buffers, index, buf)
8620 __io_remove_buffers(ctx, buf, index, -1U);
8623 static void io_req_cache_free(struct list_head *list)
8625 struct io_kiocb *req, *nxt;
8627 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8628 list_del(&req->inflight_entry);
8629 kmem_cache_free(req_cachep, req);
8633 static void io_req_caches_free(struct io_ring_ctx *ctx)
8635 struct io_submit_state *state = &ctx->submit_state;
8637 mutex_lock(&ctx->uring_lock);
8639 if (state->free_reqs) {
8640 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8641 state->free_reqs = 0;
8644 io_flush_cached_locked_reqs(ctx, state);
8645 io_req_cache_free(&state->free_list);
8646 mutex_unlock(&ctx->uring_lock);
8649 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8651 if (data && !atomic_dec_and_test(&data->refs))
8652 wait_for_completion(&data->done);
8655 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8657 io_sq_thread_finish(ctx);
8659 if (ctx->mm_account) {
8660 mmdrop(ctx->mm_account);
8661 ctx->mm_account = NULL;
8664 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8665 io_wait_rsrc_data(ctx->buf_data);
8666 io_wait_rsrc_data(ctx->file_data);
8668 mutex_lock(&ctx->uring_lock);
8670 __io_sqe_buffers_unregister(ctx);
8672 __io_sqe_files_unregister(ctx);
8674 __io_cqring_overflow_flush(ctx, true);
8675 mutex_unlock(&ctx->uring_lock);
8676 io_eventfd_unregister(ctx);
8677 io_destroy_buffers(ctx);
8679 put_cred(ctx->sq_creds);
8681 /* there are no registered resources left, nobody uses it */
8683 io_rsrc_node_destroy(ctx->rsrc_node);
8684 if (ctx->rsrc_backup_node)
8685 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8686 flush_delayed_work(&ctx->rsrc_put_work);
8688 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8689 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8691 #if defined(CONFIG_UNIX)
8692 if (ctx->ring_sock) {
8693 ctx->ring_sock->file = NULL; /* so that iput() is called */
8694 sock_release(ctx->ring_sock);
8698 io_mem_free(ctx->rings);
8699 io_mem_free(ctx->sq_sqes);
8701 percpu_ref_exit(&ctx->refs);
8702 free_uid(ctx->user);
8703 io_req_caches_free(ctx);
8705 io_wq_put_hash(ctx->hash_map);
8706 kfree(ctx->cancel_hash);
8707 kfree(ctx->dummy_ubuf);
8711 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8713 struct io_ring_ctx *ctx = file->private_data;
8716 poll_wait(file, &ctx->poll_wait, wait);
8718 * synchronizes with barrier from wq_has_sleeper call in
8722 if (!io_sqring_full(ctx))
8723 mask |= EPOLLOUT | EPOLLWRNORM;
8726 * Don't flush cqring overflow list here, just do a simple check.
8727 * Otherwise there could possible be ABBA deadlock:
8730 * lock(&ctx->uring_lock);
8732 * lock(&ctx->uring_lock);
8735 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8736 * pushs them to do the flush.
8738 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8739 mask |= EPOLLIN | EPOLLRDNORM;
8744 static int io_uring_fasync(int fd, struct file *file, int on)
8746 struct io_ring_ctx *ctx = file->private_data;
8748 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8751 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8753 const struct cred *creds;
8755 creds = xa_erase(&ctx->personalities, id);
8764 struct io_tctx_exit {
8765 struct callback_head task_work;
8766 struct completion completion;
8767 struct io_ring_ctx *ctx;
8770 static void io_tctx_exit_cb(struct callback_head *cb)
8772 struct io_uring_task *tctx = current->io_uring;
8773 struct io_tctx_exit *work;
8775 work = container_of(cb, struct io_tctx_exit, task_work);
8777 * When @in_idle, we're in cancellation and it's racy to remove the
8778 * node. It'll be removed by the end of cancellation, just ignore it.
8780 if (!atomic_read(&tctx->in_idle))
8781 io_uring_del_tctx_node((unsigned long)work->ctx);
8782 complete(&work->completion);
8785 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8787 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8789 return req->ctx == data;
8792 static void io_ring_exit_work(struct work_struct *work)
8794 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8795 unsigned long timeout = jiffies + HZ * 60 * 5;
8796 unsigned long interval = HZ / 20;
8797 struct io_tctx_exit exit;
8798 struct io_tctx_node *node;
8802 * If we're doing polled IO and end up having requests being
8803 * submitted async (out-of-line), then completions can come in while
8804 * we're waiting for refs to drop. We need to reap these manually,
8805 * as nobody else will be looking for them.
8808 io_uring_try_cancel_requests(ctx, NULL, true);
8810 struct io_sq_data *sqd = ctx->sq_data;
8811 struct task_struct *tsk;
8813 io_sq_thread_park(sqd);
8815 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8816 io_wq_cancel_cb(tsk->io_uring->io_wq,
8817 io_cancel_ctx_cb, ctx, true);
8818 io_sq_thread_unpark(sqd);
8821 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8822 /* there is little hope left, don't run it too often */
8825 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8827 init_completion(&exit.completion);
8828 init_task_work(&exit.task_work, io_tctx_exit_cb);
8831 * Some may use context even when all refs and requests have been put,
8832 * and they are free to do so while still holding uring_lock or
8833 * completion_lock, see io_req_task_submit(). Apart from other work,
8834 * this lock/unlock section also waits them to finish.
8836 mutex_lock(&ctx->uring_lock);
8837 while (!list_empty(&ctx->tctx_list)) {
8838 WARN_ON_ONCE(time_after(jiffies, timeout));
8840 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8842 /* don't spin on a single task if cancellation failed */
8843 list_rotate_left(&ctx->tctx_list);
8844 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8845 if (WARN_ON_ONCE(ret))
8847 wake_up_process(node->task);
8849 mutex_unlock(&ctx->uring_lock);
8850 wait_for_completion(&exit.completion);
8851 mutex_lock(&ctx->uring_lock);
8853 mutex_unlock(&ctx->uring_lock);
8854 spin_lock(&ctx->completion_lock);
8855 spin_unlock(&ctx->completion_lock);
8857 io_ring_ctx_free(ctx);
8860 /* Returns true if we found and killed one or more timeouts */
8861 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8864 struct io_kiocb *req, *tmp;
8867 spin_lock(&ctx->completion_lock);
8868 spin_lock_irq(&ctx->timeout_lock);
8869 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8870 if (io_match_task(req, tsk, cancel_all)) {
8871 io_kill_timeout(req, -ECANCELED);
8875 spin_unlock_irq(&ctx->timeout_lock);
8877 io_commit_cqring(ctx);
8878 spin_unlock(&ctx->completion_lock);
8880 io_cqring_ev_posted(ctx);
8881 return canceled != 0;
8884 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8886 unsigned long index;
8887 struct creds *creds;
8889 mutex_lock(&ctx->uring_lock);
8890 percpu_ref_kill(&ctx->refs);
8892 __io_cqring_overflow_flush(ctx, true);
8893 xa_for_each(&ctx->personalities, index, creds)
8894 io_unregister_personality(ctx, index);
8895 mutex_unlock(&ctx->uring_lock);
8897 io_kill_timeouts(ctx, NULL, true);
8898 io_poll_remove_all(ctx, NULL, true);
8900 /* if we failed setting up the ctx, we might not have any rings */
8901 io_iopoll_try_reap_events(ctx);
8903 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8905 * Use system_unbound_wq to avoid spawning tons of event kworkers
8906 * if we're exiting a ton of rings at the same time. It just adds
8907 * noise and overhead, there's no discernable change in runtime
8908 * over using system_wq.
8910 queue_work(system_unbound_wq, &ctx->exit_work);
8913 static int io_uring_release(struct inode *inode, struct file *file)
8915 struct io_ring_ctx *ctx = file->private_data;
8917 file->private_data = NULL;
8918 io_ring_ctx_wait_and_kill(ctx);
8922 struct io_task_cancel {
8923 struct task_struct *task;
8927 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8929 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8930 struct io_task_cancel *cancel = data;
8933 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8934 struct io_ring_ctx *ctx = req->ctx;
8936 /* protect against races with linked timeouts */
8937 spin_lock(&ctx->completion_lock);
8938 ret = io_match_task(req, cancel->task, cancel->all);
8939 spin_unlock(&ctx->completion_lock);
8941 ret = io_match_task(req, cancel->task, cancel->all);
8946 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8947 struct task_struct *task, bool cancel_all)
8949 struct io_defer_entry *de;
8952 spin_lock(&ctx->completion_lock);
8953 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8954 if (io_match_task(de->req, task, cancel_all)) {
8955 list_cut_position(&list, &ctx->defer_list, &de->list);
8959 spin_unlock(&ctx->completion_lock);
8960 if (list_empty(&list))
8963 while (!list_empty(&list)) {
8964 de = list_first_entry(&list, struct io_defer_entry, list);
8965 list_del_init(&de->list);
8966 io_req_complete_failed(de->req, -ECANCELED);
8972 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8974 struct io_tctx_node *node;
8975 enum io_wq_cancel cret;
8978 mutex_lock(&ctx->uring_lock);
8979 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8980 struct io_uring_task *tctx = node->task->io_uring;
8983 * io_wq will stay alive while we hold uring_lock, because it's
8984 * killed after ctx nodes, which requires to take the lock.
8986 if (!tctx || !tctx->io_wq)
8988 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8989 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8991 mutex_unlock(&ctx->uring_lock);
8996 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8997 struct task_struct *task,
9000 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9001 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9004 enum io_wq_cancel cret;
9008 ret |= io_uring_try_cancel_iowq(ctx);
9009 } else if (tctx && tctx->io_wq) {
9011 * Cancels requests of all rings, not only @ctx, but
9012 * it's fine as the task is in exit/exec.
9014 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9016 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9019 /* SQPOLL thread does its own polling */
9020 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9021 (ctx->sq_data && ctx->sq_data->thread == current)) {
9022 while (!list_empty_careful(&ctx->iopoll_list)) {
9023 io_iopoll_try_reap_events(ctx);
9028 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9029 ret |= io_poll_remove_all(ctx, task, cancel_all);
9030 ret |= io_kill_timeouts(ctx, task, cancel_all);
9032 ret |= io_run_task_work();
9039 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9041 struct io_uring_task *tctx = current->io_uring;
9042 struct io_tctx_node *node;
9045 if (unlikely(!tctx)) {
9046 ret = io_uring_alloc_task_context(current, ctx);
9049 tctx = current->io_uring;
9051 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9052 node = kmalloc(sizeof(*node), GFP_KERNEL);
9056 node->task = current;
9058 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9065 mutex_lock(&ctx->uring_lock);
9066 list_add(&node->ctx_node, &ctx->tctx_list);
9067 mutex_unlock(&ctx->uring_lock);
9074 * Note that this task has used io_uring. We use it for cancelation purposes.
9076 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9078 struct io_uring_task *tctx = current->io_uring;
9080 if (likely(tctx && tctx->last == ctx))
9082 return __io_uring_add_tctx_node(ctx);
9086 * Remove this io_uring_file -> task mapping.
9088 static void io_uring_del_tctx_node(unsigned long index)
9090 struct io_uring_task *tctx = current->io_uring;
9091 struct io_tctx_node *node;
9095 node = xa_erase(&tctx->xa, index);
9099 WARN_ON_ONCE(current != node->task);
9100 WARN_ON_ONCE(list_empty(&node->ctx_node));
9102 mutex_lock(&node->ctx->uring_lock);
9103 list_del(&node->ctx_node);
9104 mutex_unlock(&node->ctx->uring_lock);
9106 if (tctx->last == node->ctx)
9111 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9113 struct io_wq *wq = tctx->io_wq;
9114 struct io_tctx_node *node;
9115 unsigned long index;
9117 xa_for_each(&tctx->xa, index, node)
9118 io_uring_del_tctx_node(index);
9121 * Must be after io_uring_del_task_file() (removes nodes under
9122 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9125 io_wq_put_and_exit(wq);
9129 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9132 return atomic_read(&tctx->inflight_tracked);
9133 return percpu_counter_sum(&tctx->inflight);
9136 static void io_uring_drop_tctx_refs(struct task_struct *task)
9138 struct io_uring_task *tctx = task->io_uring;
9139 unsigned int refs = tctx->cached_refs;
9142 tctx->cached_refs = 0;
9143 percpu_counter_sub(&tctx->inflight, refs);
9144 put_task_struct_many(task, refs);
9149 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9150 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9152 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9154 struct io_uring_task *tctx = current->io_uring;
9155 struct io_ring_ctx *ctx;
9159 WARN_ON_ONCE(sqd && sqd->thread != current);
9161 if (!current->io_uring)
9164 io_wq_exit_start(tctx->io_wq);
9166 atomic_inc(&tctx->in_idle);
9168 io_uring_drop_tctx_refs(current);
9169 /* read completions before cancelations */
9170 inflight = tctx_inflight(tctx, !cancel_all);
9175 struct io_tctx_node *node;
9176 unsigned long index;
9178 xa_for_each(&tctx->xa, index, node) {
9179 /* sqpoll task will cancel all its requests */
9180 if (node->ctx->sq_data)
9182 io_uring_try_cancel_requests(node->ctx, current,
9186 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9187 io_uring_try_cancel_requests(ctx, current,
9191 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9192 io_uring_drop_tctx_refs(current);
9194 * If we've seen completions, retry without waiting. This
9195 * avoids a race where a completion comes in before we did
9196 * prepare_to_wait().
9198 if (inflight == tctx_inflight(tctx, !cancel_all))
9200 finish_wait(&tctx->wait, &wait);
9202 atomic_dec(&tctx->in_idle);
9204 io_uring_clean_tctx(tctx);
9206 /* for exec all current's requests should be gone, kill tctx */
9207 __io_uring_free(current);
9211 void __io_uring_cancel(struct files_struct *files)
9213 io_uring_cancel_generic(!files, NULL);
9216 static void *io_uring_validate_mmap_request(struct file *file,
9217 loff_t pgoff, size_t sz)
9219 struct io_ring_ctx *ctx = file->private_data;
9220 loff_t offset = pgoff << PAGE_SHIFT;
9225 case IORING_OFF_SQ_RING:
9226 case IORING_OFF_CQ_RING:
9229 case IORING_OFF_SQES:
9233 return ERR_PTR(-EINVAL);
9236 page = virt_to_head_page(ptr);
9237 if (sz > page_size(page))
9238 return ERR_PTR(-EINVAL);
9245 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9247 size_t sz = vma->vm_end - vma->vm_start;
9251 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9253 return PTR_ERR(ptr);
9255 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9256 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9259 #else /* !CONFIG_MMU */
9261 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9263 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9266 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9268 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9271 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9272 unsigned long addr, unsigned long len,
9273 unsigned long pgoff, unsigned long flags)
9277 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9279 return PTR_ERR(ptr);
9281 return (unsigned long) ptr;
9284 #endif /* !CONFIG_MMU */
9286 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9291 if (!io_sqring_full(ctx))
9293 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9295 if (!io_sqring_full(ctx))
9298 } while (!signal_pending(current));
9300 finish_wait(&ctx->sqo_sq_wait, &wait);
9304 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9305 struct __kernel_timespec __user **ts,
9306 const sigset_t __user **sig)
9308 struct io_uring_getevents_arg arg;
9311 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9312 * is just a pointer to the sigset_t.
9314 if (!(flags & IORING_ENTER_EXT_ARG)) {
9315 *sig = (const sigset_t __user *) argp;
9321 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9322 * timespec and sigset_t pointers if good.
9324 if (*argsz != sizeof(arg))
9326 if (copy_from_user(&arg, argp, sizeof(arg)))
9328 *sig = u64_to_user_ptr(arg.sigmask);
9329 *argsz = arg.sigmask_sz;
9330 *ts = u64_to_user_ptr(arg.ts);
9334 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9335 u32, min_complete, u32, flags, const void __user *, argp,
9338 struct io_ring_ctx *ctx;
9345 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9346 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9350 if (unlikely(!f.file))
9354 if (unlikely(f.file->f_op != &io_uring_fops))
9358 ctx = f.file->private_data;
9359 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9363 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9367 * For SQ polling, the thread will do all submissions and completions.
9368 * Just return the requested submit count, and wake the thread if
9372 if (ctx->flags & IORING_SETUP_SQPOLL) {
9373 io_cqring_overflow_flush(ctx);
9375 if (unlikely(ctx->sq_data->thread == NULL)) {
9379 if (flags & IORING_ENTER_SQ_WAKEUP)
9380 wake_up(&ctx->sq_data->wait);
9381 if (flags & IORING_ENTER_SQ_WAIT) {
9382 ret = io_sqpoll_wait_sq(ctx);
9386 submitted = to_submit;
9387 } else if (to_submit) {
9388 ret = io_uring_add_tctx_node(ctx);
9391 mutex_lock(&ctx->uring_lock);
9392 submitted = io_submit_sqes(ctx, to_submit);
9393 mutex_unlock(&ctx->uring_lock);
9395 if (submitted != to_submit)
9398 if (flags & IORING_ENTER_GETEVENTS) {
9399 const sigset_t __user *sig;
9400 struct __kernel_timespec __user *ts;
9402 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9406 min_complete = min(min_complete, ctx->cq_entries);
9409 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9410 * space applications don't need to do io completion events
9411 * polling again, they can rely on io_sq_thread to do polling
9412 * work, which can reduce cpu usage and uring_lock contention.
9414 if (ctx->flags & IORING_SETUP_IOPOLL &&
9415 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9416 ret = io_iopoll_check(ctx, min_complete);
9418 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9423 percpu_ref_put(&ctx->refs);
9426 return submitted ? submitted : ret;
9429 #ifdef CONFIG_PROC_FS
9430 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9431 const struct cred *cred)
9433 struct user_namespace *uns = seq_user_ns(m);
9434 struct group_info *gi;
9439 seq_printf(m, "%5d\n", id);
9440 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9441 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9442 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9443 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9444 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9445 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9446 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9447 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9448 seq_puts(m, "\n\tGroups:\t");
9449 gi = cred->group_info;
9450 for (g = 0; g < gi->ngroups; g++) {
9451 seq_put_decimal_ull(m, g ? " " : "",
9452 from_kgid_munged(uns, gi->gid[g]));
9454 seq_puts(m, "\n\tCapEff:\t");
9455 cap = cred->cap_effective;
9456 CAP_FOR_EACH_U32(__capi)
9457 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9462 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9464 struct io_sq_data *sq = NULL;
9469 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9470 * since fdinfo case grabs it in the opposite direction of normal use
9471 * cases. If we fail to get the lock, we just don't iterate any
9472 * structures that could be going away outside the io_uring mutex.
9474 has_lock = mutex_trylock(&ctx->uring_lock);
9476 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9482 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9483 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9484 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9485 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9486 struct file *f = io_file_from_index(ctx, i);
9489 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9491 seq_printf(m, "%5u: <none>\n", i);
9493 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9494 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9495 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9496 unsigned int len = buf->ubuf_end - buf->ubuf;
9498 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9500 if (has_lock && !xa_empty(&ctx->personalities)) {
9501 unsigned long index;
9502 const struct cred *cred;
9504 seq_printf(m, "Personalities:\n");
9505 xa_for_each(&ctx->personalities, index, cred)
9506 io_uring_show_cred(m, index, cred);
9508 seq_printf(m, "PollList:\n");
9509 spin_lock(&ctx->completion_lock);
9510 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9511 struct hlist_head *list = &ctx->cancel_hash[i];
9512 struct io_kiocb *req;
9514 hlist_for_each_entry(req, list, hash_node)
9515 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9516 req->task->task_works != NULL);
9518 spin_unlock(&ctx->completion_lock);
9520 mutex_unlock(&ctx->uring_lock);
9523 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9525 struct io_ring_ctx *ctx = f->private_data;
9527 if (percpu_ref_tryget(&ctx->refs)) {
9528 __io_uring_show_fdinfo(ctx, m);
9529 percpu_ref_put(&ctx->refs);
9534 static const struct file_operations io_uring_fops = {
9535 .release = io_uring_release,
9536 .mmap = io_uring_mmap,
9538 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9539 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9541 .poll = io_uring_poll,
9542 .fasync = io_uring_fasync,
9543 #ifdef CONFIG_PROC_FS
9544 .show_fdinfo = io_uring_show_fdinfo,
9548 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9549 struct io_uring_params *p)
9551 struct io_rings *rings;
9552 size_t size, sq_array_offset;
9554 /* make sure these are sane, as we already accounted them */
9555 ctx->sq_entries = p->sq_entries;
9556 ctx->cq_entries = p->cq_entries;
9558 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9559 if (size == SIZE_MAX)
9562 rings = io_mem_alloc(size);
9567 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9568 rings->sq_ring_mask = p->sq_entries - 1;
9569 rings->cq_ring_mask = p->cq_entries - 1;
9570 rings->sq_ring_entries = p->sq_entries;
9571 rings->cq_ring_entries = p->cq_entries;
9573 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9574 if (size == SIZE_MAX) {
9575 io_mem_free(ctx->rings);
9580 ctx->sq_sqes = io_mem_alloc(size);
9581 if (!ctx->sq_sqes) {
9582 io_mem_free(ctx->rings);
9590 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9594 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9598 ret = io_uring_add_tctx_node(ctx);
9603 fd_install(fd, file);
9608 * Allocate an anonymous fd, this is what constitutes the application
9609 * visible backing of an io_uring instance. The application mmaps this
9610 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9611 * we have to tie this fd to a socket for file garbage collection purposes.
9613 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9616 #if defined(CONFIG_UNIX)
9619 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9622 return ERR_PTR(ret);
9625 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9626 O_RDWR | O_CLOEXEC);
9627 #if defined(CONFIG_UNIX)
9629 sock_release(ctx->ring_sock);
9630 ctx->ring_sock = NULL;
9632 ctx->ring_sock->file = file;
9638 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9639 struct io_uring_params __user *params)
9641 struct io_ring_ctx *ctx;
9647 if (entries > IORING_MAX_ENTRIES) {
9648 if (!(p->flags & IORING_SETUP_CLAMP))
9650 entries = IORING_MAX_ENTRIES;
9654 * Use twice as many entries for the CQ ring. It's possible for the
9655 * application to drive a higher depth than the size of the SQ ring,
9656 * since the sqes are only used at submission time. This allows for
9657 * some flexibility in overcommitting a bit. If the application has
9658 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9659 * of CQ ring entries manually.
9661 p->sq_entries = roundup_pow_of_two(entries);
9662 if (p->flags & IORING_SETUP_CQSIZE) {
9664 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9665 * to a power-of-two, if it isn't already. We do NOT impose
9666 * any cq vs sq ring sizing.
9670 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9671 if (!(p->flags & IORING_SETUP_CLAMP))
9673 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9675 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9676 if (p->cq_entries < p->sq_entries)
9679 p->cq_entries = 2 * p->sq_entries;
9682 ctx = io_ring_ctx_alloc(p);
9685 ctx->compat = in_compat_syscall();
9686 if (!capable(CAP_IPC_LOCK))
9687 ctx->user = get_uid(current_user());
9690 * This is just grabbed for accounting purposes. When a process exits,
9691 * the mm is exited and dropped before the files, hence we need to hang
9692 * on to this mm purely for the purposes of being able to unaccount
9693 * memory (locked/pinned vm). It's not used for anything else.
9695 mmgrab(current->mm);
9696 ctx->mm_account = current->mm;
9698 ret = io_allocate_scq_urings(ctx, p);
9702 ret = io_sq_offload_create(ctx, p);
9705 /* always set a rsrc node */
9706 ret = io_rsrc_node_switch_start(ctx);
9709 io_rsrc_node_switch(ctx, NULL);
9711 memset(&p->sq_off, 0, sizeof(p->sq_off));
9712 p->sq_off.head = offsetof(struct io_rings, sq.head);
9713 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9714 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9715 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9716 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9717 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9718 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9720 memset(&p->cq_off, 0, sizeof(p->cq_off));
9721 p->cq_off.head = offsetof(struct io_rings, cq.head);
9722 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9723 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9724 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9725 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9726 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9727 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9729 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9730 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9731 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9732 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9733 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9734 IORING_FEAT_RSRC_TAGS;
9736 if (copy_to_user(params, p, sizeof(*p))) {
9741 file = io_uring_get_file(ctx);
9743 ret = PTR_ERR(file);
9748 * Install ring fd as the very last thing, so we don't risk someone
9749 * having closed it before we finish setup
9751 ret = io_uring_install_fd(ctx, file);
9753 /* fput will clean it up */
9758 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9761 io_ring_ctx_wait_and_kill(ctx);
9766 * Sets up an aio uring context, and returns the fd. Applications asks for a
9767 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9768 * params structure passed in.
9770 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9772 struct io_uring_params p;
9775 if (copy_from_user(&p, params, sizeof(p)))
9777 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9782 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9783 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9784 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9785 IORING_SETUP_R_DISABLED))
9788 return io_uring_create(entries, &p, params);
9791 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9792 struct io_uring_params __user *, params)
9794 return io_uring_setup(entries, params);
9797 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9799 struct io_uring_probe *p;
9803 size = struct_size(p, ops, nr_args);
9804 if (size == SIZE_MAX)
9806 p = kzalloc(size, GFP_KERNEL);
9811 if (copy_from_user(p, arg, size))
9814 if (memchr_inv(p, 0, size))
9817 p->last_op = IORING_OP_LAST - 1;
9818 if (nr_args > IORING_OP_LAST)
9819 nr_args = IORING_OP_LAST;
9821 for (i = 0; i < nr_args; i++) {
9823 if (!io_op_defs[i].not_supported)
9824 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9829 if (copy_to_user(arg, p, size))
9836 static int io_register_personality(struct io_ring_ctx *ctx)
9838 const struct cred *creds;
9842 creds = get_current_cred();
9844 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9845 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9853 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9854 unsigned int nr_args)
9856 struct io_uring_restriction *res;
9860 /* Restrictions allowed only if rings started disabled */
9861 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9864 /* We allow only a single restrictions registration */
9865 if (ctx->restrictions.registered)
9868 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9871 size = array_size(nr_args, sizeof(*res));
9872 if (size == SIZE_MAX)
9875 res = memdup_user(arg, size);
9877 return PTR_ERR(res);
9881 for (i = 0; i < nr_args; i++) {
9882 switch (res[i].opcode) {
9883 case IORING_RESTRICTION_REGISTER_OP:
9884 if (res[i].register_op >= IORING_REGISTER_LAST) {
9889 __set_bit(res[i].register_op,
9890 ctx->restrictions.register_op);
9892 case IORING_RESTRICTION_SQE_OP:
9893 if (res[i].sqe_op >= IORING_OP_LAST) {
9898 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9900 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9901 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9903 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9904 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9913 /* Reset all restrictions if an error happened */
9915 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9917 ctx->restrictions.registered = true;
9923 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9925 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9928 if (ctx->restrictions.registered)
9929 ctx->restricted = 1;
9931 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9932 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9933 wake_up(&ctx->sq_data->wait);
9937 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9938 struct io_uring_rsrc_update2 *up,
9946 if (check_add_overflow(up->offset, nr_args, &tmp))
9948 err = io_rsrc_node_switch_start(ctx);
9953 case IORING_RSRC_FILE:
9954 return __io_sqe_files_update(ctx, up, nr_args);
9955 case IORING_RSRC_BUFFER:
9956 return __io_sqe_buffers_update(ctx, up, nr_args);
9961 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9964 struct io_uring_rsrc_update2 up;
9968 memset(&up, 0, sizeof(up));
9969 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9971 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9974 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9975 unsigned size, unsigned type)
9977 struct io_uring_rsrc_update2 up;
9979 if (size != sizeof(up))
9981 if (copy_from_user(&up, arg, sizeof(up)))
9983 if (!up.nr || up.resv)
9985 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9988 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9989 unsigned int size, unsigned int type)
9991 struct io_uring_rsrc_register rr;
9993 /* keep it extendible */
9994 if (size != sizeof(rr))
9997 memset(&rr, 0, sizeof(rr));
9998 if (copy_from_user(&rr, arg, size))
10000 if (!rr.nr || rr.resv || rr.resv2)
10004 case IORING_RSRC_FILE:
10005 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10006 rr.nr, u64_to_user_ptr(rr.tags));
10007 case IORING_RSRC_BUFFER:
10008 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10009 rr.nr, u64_to_user_ptr(rr.tags));
10014 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10017 struct io_uring_task *tctx = current->io_uring;
10018 cpumask_var_t new_mask;
10021 if (!tctx || !tctx->io_wq)
10024 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10027 cpumask_clear(new_mask);
10028 if (len > cpumask_size())
10029 len = cpumask_size();
10031 if (copy_from_user(new_mask, arg, len)) {
10032 free_cpumask_var(new_mask);
10036 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10037 free_cpumask_var(new_mask);
10041 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10043 struct io_uring_task *tctx = current->io_uring;
10045 if (!tctx || !tctx->io_wq)
10048 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10051 static bool io_register_op_must_quiesce(int op)
10054 case IORING_REGISTER_BUFFERS:
10055 case IORING_UNREGISTER_BUFFERS:
10056 case IORING_REGISTER_FILES:
10057 case IORING_UNREGISTER_FILES:
10058 case IORING_REGISTER_FILES_UPDATE:
10059 case IORING_REGISTER_PROBE:
10060 case IORING_REGISTER_PERSONALITY:
10061 case IORING_UNREGISTER_PERSONALITY:
10062 case IORING_REGISTER_FILES2:
10063 case IORING_REGISTER_FILES_UPDATE2:
10064 case IORING_REGISTER_BUFFERS2:
10065 case IORING_REGISTER_BUFFERS_UPDATE:
10066 case IORING_REGISTER_IOWQ_AFF:
10067 case IORING_UNREGISTER_IOWQ_AFF:
10074 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10078 percpu_ref_kill(&ctx->refs);
10081 * Drop uring mutex before waiting for references to exit. If another
10082 * thread is currently inside io_uring_enter() it might need to grab the
10083 * uring_lock to make progress. If we hold it here across the drain
10084 * wait, then we can deadlock. It's safe to drop the mutex here, since
10085 * no new references will come in after we've killed the percpu ref.
10087 mutex_unlock(&ctx->uring_lock);
10089 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10092 ret = io_run_task_work_sig();
10093 } while (ret >= 0);
10094 mutex_lock(&ctx->uring_lock);
10097 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10101 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10102 void __user *arg, unsigned nr_args)
10103 __releases(ctx->uring_lock)
10104 __acquires(ctx->uring_lock)
10109 * We're inside the ring mutex, if the ref is already dying, then
10110 * someone else killed the ctx or is already going through
10111 * io_uring_register().
10113 if (percpu_ref_is_dying(&ctx->refs))
10116 if (ctx->restricted) {
10117 if (opcode >= IORING_REGISTER_LAST)
10119 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10120 if (!test_bit(opcode, ctx->restrictions.register_op))
10124 if (io_register_op_must_quiesce(opcode)) {
10125 ret = io_ctx_quiesce(ctx);
10131 case IORING_REGISTER_BUFFERS:
10132 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10134 case IORING_UNREGISTER_BUFFERS:
10136 if (arg || nr_args)
10138 ret = io_sqe_buffers_unregister(ctx);
10140 case IORING_REGISTER_FILES:
10141 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10143 case IORING_UNREGISTER_FILES:
10145 if (arg || nr_args)
10147 ret = io_sqe_files_unregister(ctx);
10149 case IORING_REGISTER_FILES_UPDATE:
10150 ret = io_register_files_update(ctx, arg, nr_args);
10152 case IORING_REGISTER_EVENTFD:
10153 case IORING_REGISTER_EVENTFD_ASYNC:
10157 ret = io_eventfd_register(ctx, arg);
10160 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10161 ctx->eventfd_async = 1;
10163 ctx->eventfd_async = 0;
10165 case IORING_UNREGISTER_EVENTFD:
10167 if (arg || nr_args)
10169 ret = io_eventfd_unregister(ctx);
10171 case IORING_REGISTER_PROBE:
10173 if (!arg || nr_args > 256)
10175 ret = io_probe(ctx, arg, nr_args);
10177 case IORING_REGISTER_PERSONALITY:
10179 if (arg || nr_args)
10181 ret = io_register_personality(ctx);
10183 case IORING_UNREGISTER_PERSONALITY:
10187 ret = io_unregister_personality(ctx, nr_args);
10189 case IORING_REGISTER_ENABLE_RINGS:
10191 if (arg || nr_args)
10193 ret = io_register_enable_rings(ctx);
10195 case IORING_REGISTER_RESTRICTIONS:
10196 ret = io_register_restrictions(ctx, arg, nr_args);
10198 case IORING_REGISTER_FILES2:
10199 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10201 case IORING_REGISTER_FILES_UPDATE2:
10202 ret = io_register_rsrc_update(ctx, arg, nr_args,
10205 case IORING_REGISTER_BUFFERS2:
10206 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10208 case IORING_REGISTER_BUFFERS_UPDATE:
10209 ret = io_register_rsrc_update(ctx, arg, nr_args,
10210 IORING_RSRC_BUFFER);
10212 case IORING_REGISTER_IOWQ_AFF:
10214 if (!arg || !nr_args)
10216 ret = io_register_iowq_aff(ctx, arg, nr_args);
10218 case IORING_UNREGISTER_IOWQ_AFF:
10220 if (arg || nr_args)
10222 ret = io_unregister_iowq_aff(ctx);
10229 if (io_register_op_must_quiesce(opcode)) {
10230 /* bring the ctx back to life */
10231 percpu_ref_reinit(&ctx->refs);
10232 reinit_completion(&ctx->ref_comp);
10237 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10238 void __user *, arg, unsigned int, nr_args)
10240 struct io_ring_ctx *ctx;
10249 if (f.file->f_op != &io_uring_fops)
10252 ctx = f.file->private_data;
10254 io_run_task_work();
10256 mutex_lock(&ctx->uring_lock);
10257 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10258 mutex_unlock(&ctx->uring_lock);
10259 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10260 ctx->cq_ev_fd != NULL, ret);
10266 static int __init io_uring_init(void)
10268 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10269 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10270 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10273 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10274 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10275 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10276 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10277 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10278 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10279 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10280 BUILD_BUG_SQE_ELEM(8, __u64, off);
10281 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10282 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10283 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10284 BUILD_BUG_SQE_ELEM(24, __u32, len);
10285 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10286 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10287 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10288 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10289 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10290 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10291 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10292 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10293 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10296 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10299 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10300 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10301 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10302 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10303 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10304 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10306 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10307 sizeof(struct io_uring_rsrc_update));
10308 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10309 sizeof(struct io_uring_rsrc_update2));
10310 /* should fit into one byte */
10311 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10313 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10314 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10316 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10320 __initcall(io_uring_init);