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_COMPLETE_INLINE_BIT,
710 REQ_F_DONT_REISSUE_BIT,
713 REQ_F_ARM_LTIMEOUT_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 /* completion is deferred through io_comp_state */
754 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
755 /* caller should reissue async */
756 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
757 /* don't attempt request reissue, see io_rw_reissue() */
758 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
759 /* supports async reads */
760 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
761 /* supports async writes */
762 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
764 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
765 /* has creds assigned */
766 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
767 /* skip refcounting if not set */
768 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
769 /* there is a linked timeout that has to be armed */
770 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_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 void io_queue_linked_timeout(struct io_kiocb *req);
1050 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1051 struct io_uring_rsrc_update2 *up,
1053 static void io_clean_op(struct io_kiocb *req);
1054 static struct file *io_file_get(struct io_ring_ctx *ctx,
1055 struct io_kiocb *req, int fd, bool fixed);
1056 static void __io_queue_sqe(struct io_kiocb *req);
1057 static void io_rsrc_put_work(struct work_struct *work);
1059 static void io_req_task_queue(struct io_kiocb *req);
1060 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1061 static int io_req_prep_async(struct io_kiocb *req);
1063 static struct kmem_cache *req_cachep;
1065 static const struct file_operations io_uring_fops;
1067 struct sock *io_uring_get_socket(struct file *file)
1069 #if defined(CONFIG_UNIX)
1070 if (file->f_op == &io_uring_fops) {
1071 struct io_ring_ctx *ctx = file->private_data;
1073 return ctx->ring_sock->sk;
1078 EXPORT_SYMBOL(io_uring_get_socket);
1080 #define io_for_each_link(pos, head) \
1081 for (pos = (head); pos; pos = pos->link)
1084 * Shamelessly stolen from the mm implementation of page reference checking,
1085 * see commit f958d7b528b1 for details.
1087 #define req_ref_zero_or_close_to_overflow(req) \
1088 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1090 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1092 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1093 return atomic_inc_not_zero(&req->refs);
1096 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1098 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1101 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1102 return atomic_dec_and_test(&req->refs);
1105 static inline void req_ref_put(struct io_kiocb *req)
1107 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1108 WARN_ON_ONCE(req_ref_put_and_test(req));
1111 static inline void req_ref_get(struct io_kiocb *req)
1113 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1114 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1115 atomic_inc(&req->refs);
1118 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1120 if (!(req->flags & REQ_F_REFCOUNT)) {
1121 req->flags |= REQ_F_REFCOUNT;
1122 atomic_set(&req->refs, nr);
1126 static inline void io_req_set_refcount(struct io_kiocb *req)
1128 __io_req_set_refcount(req, 1);
1131 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1133 struct io_ring_ctx *ctx = req->ctx;
1135 if (!req->fixed_rsrc_refs) {
1136 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1137 percpu_ref_get(req->fixed_rsrc_refs);
1141 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1143 bool got = percpu_ref_tryget(ref);
1145 /* already at zero, wait for ->release() */
1147 wait_for_completion(compl);
1148 percpu_ref_resurrect(ref);
1150 percpu_ref_put(ref);
1153 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1156 struct io_kiocb *req;
1158 if (task && head->task != task)
1163 io_for_each_link(req, head) {
1164 if (req->flags & REQ_F_INFLIGHT)
1170 static inline void req_set_fail(struct io_kiocb *req)
1172 req->flags |= REQ_F_FAIL;
1175 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1177 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1179 complete(&ctx->ref_comp);
1182 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1184 return !req->timeout.off;
1187 static void io_fallback_req_func(struct work_struct *work)
1189 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1190 fallback_work.work);
1191 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1192 struct io_kiocb *req, *tmp;
1194 percpu_ref_get(&ctx->refs);
1195 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1196 req->io_task_work.func(req);
1197 percpu_ref_put(&ctx->refs);
1200 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1202 struct io_ring_ctx *ctx;
1205 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1210 * Use 5 bits less than the max cq entries, that should give us around
1211 * 32 entries per hash list if totally full and uniformly spread.
1213 hash_bits = ilog2(p->cq_entries);
1217 ctx->cancel_hash_bits = hash_bits;
1218 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1220 if (!ctx->cancel_hash)
1222 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1224 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1225 if (!ctx->dummy_ubuf)
1227 /* set invalid range, so io_import_fixed() fails meeting it */
1228 ctx->dummy_ubuf->ubuf = -1UL;
1230 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1231 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1234 ctx->flags = p->flags;
1235 init_waitqueue_head(&ctx->sqo_sq_wait);
1236 INIT_LIST_HEAD(&ctx->sqd_list);
1237 init_waitqueue_head(&ctx->poll_wait);
1238 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1239 init_completion(&ctx->ref_comp);
1240 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1241 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1242 mutex_init(&ctx->uring_lock);
1243 init_waitqueue_head(&ctx->cq_wait);
1244 spin_lock_init(&ctx->completion_lock);
1245 spin_lock_init(&ctx->timeout_lock);
1246 INIT_LIST_HEAD(&ctx->iopoll_list);
1247 INIT_LIST_HEAD(&ctx->defer_list);
1248 INIT_LIST_HEAD(&ctx->timeout_list);
1249 spin_lock_init(&ctx->rsrc_ref_lock);
1250 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1251 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1252 init_llist_head(&ctx->rsrc_put_llist);
1253 INIT_LIST_HEAD(&ctx->tctx_list);
1254 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1255 INIT_LIST_HEAD(&ctx->locked_free_list);
1256 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1259 kfree(ctx->dummy_ubuf);
1260 kfree(ctx->cancel_hash);
1265 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1267 struct io_rings *r = ctx->rings;
1269 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1273 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1275 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1276 struct io_ring_ctx *ctx = req->ctx;
1278 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1284 #define FFS_ASYNC_READ 0x1UL
1285 #define FFS_ASYNC_WRITE 0x2UL
1287 #define FFS_ISREG 0x4UL
1289 #define FFS_ISREG 0x0UL
1291 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1293 static inline bool io_req_ffs_set(struct io_kiocb *req)
1295 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1298 static void io_req_track_inflight(struct io_kiocb *req)
1300 if (!(req->flags & REQ_F_INFLIGHT)) {
1301 req->flags |= REQ_F_INFLIGHT;
1302 atomic_inc(¤t->io_uring->inflight_tracked);
1306 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1308 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1309 req->flags |= REQ_F_LINK_TIMEOUT;
1311 /* linked timeouts should have two refs once prep'ed */
1312 io_req_set_refcount(req);
1313 __io_req_set_refcount(req->link, 2);
1317 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1319 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1321 return __io_prep_linked_timeout(req);
1324 static void io_prep_async_work(struct io_kiocb *req)
1326 const struct io_op_def *def = &io_op_defs[req->opcode];
1327 struct io_ring_ctx *ctx = req->ctx;
1329 if (!(req->flags & REQ_F_CREDS)) {
1330 req->flags |= REQ_F_CREDS;
1331 req->creds = get_current_cred();
1334 req->work.list.next = NULL;
1335 req->work.flags = 0;
1336 if (req->flags & REQ_F_FORCE_ASYNC)
1337 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1339 if (req->flags & REQ_F_ISREG) {
1340 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1341 io_wq_hash_work(&req->work, file_inode(req->file));
1342 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1343 if (def->unbound_nonreg_file)
1344 req->work.flags |= IO_WQ_WORK_UNBOUND;
1347 switch (req->opcode) {
1348 case IORING_OP_SPLICE:
1350 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1351 req->work.flags |= IO_WQ_WORK_UNBOUND;
1356 static void io_prep_async_link(struct io_kiocb *req)
1358 struct io_kiocb *cur;
1360 if (req->flags & REQ_F_LINK_TIMEOUT) {
1361 struct io_ring_ctx *ctx = req->ctx;
1363 spin_lock(&ctx->completion_lock);
1364 io_for_each_link(cur, req)
1365 io_prep_async_work(cur);
1366 spin_unlock(&ctx->completion_lock);
1368 io_for_each_link(cur, req)
1369 io_prep_async_work(cur);
1373 static void io_queue_async_work(struct io_kiocb *req)
1375 struct io_ring_ctx *ctx = req->ctx;
1376 struct io_kiocb *link = io_prep_linked_timeout(req);
1377 struct io_uring_task *tctx = req->task->io_uring;
1380 BUG_ON(!tctx->io_wq);
1382 /* init ->work of the whole link before punting */
1383 io_prep_async_link(req);
1386 * Not expected to happen, but if we do have a bug where this _can_
1387 * happen, catch it here and ensure the request is marked as
1388 * canceled. That will make io-wq go through the usual work cancel
1389 * procedure rather than attempt to run this request (or create a new
1392 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1393 req->work.flags |= IO_WQ_WORK_CANCEL;
1395 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1396 &req->work, req->flags);
1397 io_wq_enqueue(tctx->io_wq, &req->work);
1399 io_queue_linked_timeout(link);
1402 static void io_kill_timeout(struct io_kiocb *req, int status)
1403 __must_hold(&req->ctx->completion_lock)
1404 __must_hold(&req->ctx->timeout_lock)
1406 struct io_timeout_data *io = req->async_data;
1408 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1409 atomic_set(&req->ctx->cq_timeouts,
1410 atomic_read(&req->ctx->cq_timeouts) + 1);
1411 list_del_init(&req->timeout.list);
1412 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1413 io_put_req_deferred(req);
1417 static void io_queue_deferred(struct io_ring_ctx *ctx)
1419 while (!list_empty(&ctx->defer_list)) {
1420 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1421 struct io_defer_entry, list);
1423 if (req_need_defer(de->req, de->seq))
1425 list_del_init(&de->list);
1426 io_req_task_queue(de->req);
1431 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1432 __must_hold(&ctx->completion_lock)
1434 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1436 spin_lock_irq(&ctx->timeout_lock);
1437 while (!list_empty(&ctx->timeout_list)) {
1438 u32 events_needed, events_got;
1439 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1440 struct io_kiocb, timeout.list);
1442 if (io_is_timeout_noseq(req))
1446 * Since seq can easily wrap around over time, subtract
1447 * the last seq at which timeouts were flushed before comparing.
1448 * Assuming not more than 2^31-1 events have happened since,
1449 * these subtractions won't have wrapped, so we can check if
1450 * target is in [last_seq, current_seq] by comparing the two.
1452 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1453 events_got = seq - ctx->cq_last_tm_flush;
1454 if (events_got < events_needed)
1457 list_del_init(&req->timeout.list);
1458 io_kill_timeout(req, 0);
1460 ctx->cq_last_tm_flush = seq;
1461 spin_unlock_irq(&ctx->timeout_lock);
1464 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1466 if (ctx->off_timeout_used)
1467 io_flush_timeouts(ctx);
1468 if (ctx->drain_active)
1469 io_queue_deferred(ctx);
1472 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1474 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1475 __io_commit_cqring_flush(ctx);
1476 /* order cqe stores with ring update */
1477 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1480 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1482 struct io_rings *r = ctx->rings;
1484 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1487 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1489 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1492 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1494 struct io_rings *rings = ctx->rings;
1495 unsigned tail, mask = ctx->cq_entries - 1;
1498 * writes to the cq entry need to come after reading head; the
1499 * control dependency is enough as we're using WRITE_ONCE to
1502 if (__io_cqring_events(ctx) == ctx->cq_entries)
1505 tail = ctx->cached_cq_tail++;
1506 return &rings->cqes[tail & mask];
1509 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1511 if (likely(!ctx->cq_ev_fd))
1513 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1515 return !ctx->eventfd_async || io_wq_current_is_worker();
1518 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1521 * wake_up_all() may seem excessive, but io_wake_function() and
1522 * io_should_wake() handle the termination of the loop and only
1523 * wake as many waiters as we need to.
1525 if (wq_has_sleeper(&ctx->cq_wait))
1526 wake_up_all(&ctx->cq_wait);
1527 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1528 wake_up(&ctx->sq_data->wait);
1529 if (io_should_trigger_evfd(ctx))
1530 eventfd_signal(ctx->cq_ev_fd, 1);
1531 if (waitqueue_active(&ctx->poll_wait)) {
1532 wake_up_interruptible(&ctx->poll_wait);
1533 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1537 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1539 if (ctx->flags & IORING_SETUP_SQPOLL) {
1540 if (wq_has_sleeper(&ctx->cq_wait))
1541 wake_up_all(&ctx->cq_wait);
1543 if (io_should_trigger_evfd(ctx))
1544 eventfd_signal(ctx->cq_ev_fd, 1);
1545 if (waitqueue_active(&ctx->poll_wait)) {
1546 wake_up_interruptible(&ctx->poll_wait);
1547 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1551 /* Returns true if there are no backlogged entries after the flush */
1552 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1554 bool all_flushed, posted;
1556 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1560 spin_lock(&ctx->completion_lock);
1561 while (!list_empty(&ctx->cq_overflow_list)) {
1562 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1563 struct io_overflow_cqe *ocqe;
1567 ocqe = list_first_entry(&ctx->cq_overflow_list,
1568 struct io_overflow_cqe, list);
1570 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1572 io_account_cq_overflow(ctx);
1575 list_del(&ocqe->list);
1579 all_flushed = list_empty(&ctx->cq_overflow_list);
1581 clear_bit(0, &ctx->check_cq_overflow);
1582 WRITE_ONCE(ctx->rings->sq_flags,
1583 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1587 io_commit_cqring(ctx);
1588 spin_unlock(&ctx->completion_lock);
1590 io_cqring_ev_posted(ctx);
1594 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1598 if (test_bit(0, &ctx->check_cq_overflow)) {
1599 /* iopoll syncs against uring_lock, not completion_lock */
1600 if (ctx->flags & IORING_SETUP_IOPOLL)
1601 mutex_lock(&ctx->uring_lock);
1602 ret = __io_cqring_overflow_flush(ctx, false);
1603 if (ctx->flags & IORING_SETUP_IOPOLL)
1604 mutex_unlock(&ctx->uring_lock);
1610 /* must to be called somewhat shortly after putting a request */
1611 static inline void io_put_task(struct task_struct *task, int nr)
1613 struct io_uring_task *tctx = task->io_uring;
1615 percpu_counter_sub(&tctx->inflight, nr);
1616 if (unlikely(atomic_read(&tctx->in_idle)))
1617 wake_up(&tctx->wait);
1618 put_task_struct_many(task, nr);
1621 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1622 long res, unsigned int cflags)
1624 struct io_overflow_cqe *ocqe;
1626 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1629 * If we're in ring overflow flush mode, or in task cancel mode,
1630 * or cannot allocate an overflow entry, then we need to drop it
1633 io_account_cq_overflow(ctx);
1636 if (list_empty(&ctx->cq_overflow_list)) {
1637 set_bit(0, &ctx->check_cq_overflow);
1638 WRITE_ONCE(ctx->rings->sq_flags,
1639 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1642 ocqe->cqe.user_data = user_data;
1643 ocqe->cqe.res = res;
1644 ocqe->cqe.flags = cflags;
1645 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1649 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1650 long res, unsigned int cflags)
1652 struct io_uring_cqe *cqe;
1654 trace_io_uring_complete(ctx, user_data, res, cflags);
1657 * If we can't get a cq entry, userspace overflowed the
1658 * submission (by quite a lot). Increment the overflow count in
1661 cqe = io_get_cqe(ctx);
1663 WRITE_ONCE(cqe->user_data, user_data);
1664 WRITE_ONCE(cqe->res, res);
1665 WRITE_ONCE(cqe->flags, cflags);
1668 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1671 /* not as hot to bloat with inlining */
1672 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1673 long res, unsigned int cflags)
1675 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1678 static void io_req_complete_post(struct io_kiocb *req, long res,
1679 unsigned int cflags)
1681 struct io_ring_ctx *ctx = req->ctx;
1683 spin_lock(&ctx->completion_lock);
1684 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1686 * If we're the last reference to this request, add to our locked
1689 if (req_ref_put_and_test(req)) {
1690 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1691 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1692 io_disarm_next(req);
1694 io_req_task_queue(req->link);
1698 io_dismantle_req(req);
1699 io_put_task(req->task, 1);
1700 list_add(&req->inflight_entry, &ctx->locked_free_list);
1701 ctx->locked_free_nr++;
1703 if (!percpu_ref_tryget(&ctx->refs))
1706 io_commit_cqring(ctx);
1707 spin_unlock(&ctx->completion_lock);
1710 io_cqring_ev_posted(ctx);
1711 percpu_ref_put(&ctx->refs);
1715 static inline bool io_req_needs_clean(struct io_kiocb *req)
1717 return req->flags & IO_REQ_CLEAN_FLAGS;
1720 static void io_req_complete_state(struct io_kiocb *req, long res,
1721 unsigned int cflags)
1723 if (io_req_needs_clean(req))
1726 req->compl.cflags = cflags;
1727 req->flags |= REQ_F_COMPLETE_INLINE;
1730 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1731 long res, unsigned cflags)
1733 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1734 io_req_complete_state(req, res, cflags);
1736 io_req_complete_post(req, res, cflags);
1739 static inline void io_req_complete(struct io_kiocb *req, long res)
1741 __io_req_complete(req, 0, res, 0);
1744 static void io_req_complete_failed(struct io_kiocb *req, long res)
1747 io_req_complete_post(req, res, 0);
1751 * Don't initialise the fields below on every allocation, but do that in
1752 * advance and keep them valid across allocations.
1754 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1758 req->async_data = NULL;
1759 /* not necessary, but safer to zero */
1763 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1764 struct io_submit_state *state)
1766 spin_lock(&ctx->completion_lock);
1767 list_splice_init(&ctx->locked_free_list, &state->free_list);
1768 ctx->locked_free_nr = 0;
1769 spin_unlock(&ctx->completion_lock);
1772 /* Returns true IFF there are requests in the cache */
1773 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1775 struct io_submit_state *state = &ctx->submit_state;
1779 * If we have more than a batch's worth of requests in our IRQ side
1780 * locked cache, grab the lock and move them over to our submission
1783 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1784 io_flush_cached_locked_reqs(ctx, state);
1786 nr = state->free_reqs;
1787 while (!list_empty(&state->free_list)) {
1788 struct io_kiocb *req = list_first_entry(&state->free_list,
1789 struct io_kiocb, inflight_entry);
1791 list_del(&req->inflight_entry);
1792 state->reqs[nr++] = req;
1793 if (nr == ARRAY_SIZE(state->reqs))
1797 state->free_reqs = nr;
1802 * A request might get retired back into the request caches even before opcode
1803 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1804 * Because of that, io_alloc_req() should be called only under ->uring_lock
1805 * and with extra caution to not get a request that is still worked on.
1807 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1808 __must_hold(&ctx->uring_lock)
1810 struct io_submit_state *state = &ctx->submit_state;
1811 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1814 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1816 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1819 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1823 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1824 * retry single alloc to be on the safe side.
1826 if (unlikely(ret <= 0)) {
1827 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1828 if (!state->reqs[0])
1833 for (i = 0; i < ret; i++)
1834 io_preinit_req(state->reqs[i], ctx);
1835 state->free_reqs = ret;
1838 return state->reqs[state->free_reqs];
1841 static inline void io_put_file(struct file *file)
1847 static void io_dismantle_req(struct io_kiocb *req)
1849 unsigned int flags = req->flags;
1851 if (io_req_needs_clean(req))
1853 if (!(flags & REQ_F_FIXED_FILE))
1854 io_put_file(req->file);
1855 if (req->fixed_rsrc_refs)
1856 percpu_ref_put(req->fixed_rsrc_refs);
1857 if (req->async_data) {
1858 kfree(req->async_data);
1859 req->async_data = NULL;
1863 static void __io_free_req(struct io_kiocb *req)
1865 struct io_ring_ctx *ctx = req->ctx;
1867 io_dismantle_req(req);
1868 io_put_task(req->task, 1);
1870 spin_lock(&ctx->completion_lock);
1871 list_add(&req->inflight_entry, &ctx->locked_free_list);
1872 ctx->locked_free_nr++;
1873 spin_unlock(&ctx->completion_lock);
1875 percpu_ref_put(&ctx->refs);
1878 static inline void io_remove_next_linked(struct io_kiocb *req)
1880 struct io_kiocb *nxt = req->link;
1882 req->link = nxt->link;
1886 static bool io_kill_linked_timeout(struct io_kiocb *req)
1887 __must_hold(&req->ctx->completion_lock)
1888 __must_hold(&req->ctx->timeout_lock)
1890 struct io_kiocb *link = req->link;
1892 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1893 struct io_timeout_data *io = link->async_data;
1895 io_remove_next_linked(req);
1896 link->timeout.head = NULL;
1897 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1898 io_cqring_fill_event(link->ctx, link->user_data,
1900 io_put_req_deferred(link);
1907 static void io_fail_links(struct io_kiocb *req)
1908 __must_hold(&req->ctx->completion_lock)
1910 struct io_kiocb *nxt, *link = req->link;
1917 trace_io_uring_fail_link(req, link);
1918 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1919 io_put_req_deferred(link);
1924 static bool io_disarm_next(struct io_kiocb *req)
1925 __must_hold(&req->ctx->completion_lock)
1927 bool posted = false;
1929 if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1930 struct io_ring_ctx *ctx = req->ctx;
1932 spin_lock_irq(&ctx->timeout_lock);
1933 posted = io_kill_linked_timeout(req);
1934 spin_unlock_irq(&ctx->timeout_lock);
1936 if (unlikely((req->flags & REQ_F_FAIL) &&
1937 !(req->flags & REQ_F_HARDLINK))) {
1938 posted |= (req->link != NULL);
1944 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1946 struct io_kiocb *nxt;
1949 * If LINK is set, we have dependent requests in this chain. If we
1950 * didn't fail this request, queue the first one up, moving any other
1951 * dependencies to the next request. In case of failure, fail the rest
1954 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1955 struct io_ring_ctx *ctx = req->ctx;
1958 spin_lock(&ctx->completion_lock);
1959 posted = io_disarm_next(req);
1961 io_commit_cqring(req->ctx);
1962 spin_unlock(&ctx->completion_lock);
1964 io_cqring_ev_posted(ctx);
1971 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1973 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1975 return __io_req_find_next(req);
1978 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1982 if (ctx->submit_state.compl_nr) {
1983 mutex_lock(&ctx->uring_lock);
1984 io_submit_flush_completions(ctx);
1985 mutex_unlock(&ctx->uring_lock);
1987 percpu_ref_put(&ctx->refs);
1990 static void tctx_task_work(struct callback_head *cb)
1992 struct io_ring_ctx *ctx = NULL;
1993 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1997 struct io_wq_work_node *node;
1999 spin_lock_irq(&tctx->task_lock);
2000 node = tctx->task_list.first;
2001 INIT_WQ_LIST(&tctx->task_list);
2003 tctx->task_running = false;
2004 spin_unlock_irq(&tctx->task_lock);
2009 struct io_wq_work_node *next = node->next;
2010 struct io_kiocb *req = container_of(node, struct io_kiocb,
2013 if (req->ctx != ctx) {
2014 ctx_flush_and_put(ctx);
2016 percpu_ref_get(&ctx->refs);
2018 req->io_task_work.func(req);
2025 ctx_flush_and_put(ctx);
2028 static void io_req_task_work_add(struct io_kiocb *req)
2030 struct task_struct *tsk = req->task;
2031 struct io_uring_task *tctx = tsk->io_uring;
2032 enum task_work_notify_mode notify;
2033 struct io_wq_work_node *node;
2034 unsigned long flags;
2037 WARN_ON_ONCE(!tctx);
2039 spin_lock_irqsave(&tctx->task_lock, flags);
2040 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2041 running = tctx->task_running;
2043 tctx->task_running = true;
2044 spin_unlock_irqrestore(&tctx->task_lock, flags);
2046 /* task_work already pending, we're done */
2051 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2052 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2053 * processing task_work. There's no reliable way to tell if TWA_RESUME
2056 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2057 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2058 wake_up_process(tsk);
2062 spin_lock_irqsave(&tctx->task_lock, flags);
2063 tctx->task_running = false;
2064 node = tctx->task_list.first;
2065 INIT_WQ_LIST(&tctx->task_list);
2066 spin_unlock_irqrestore(&tctx->task_lock, flags);
2069 req = container_of(node, struct io_kiocb, io_task_work.node);
2071 if (llist_add(&req->io_task_work.fallback_node,
2072 &req->ctx->fallback_llist))
2073 schedule_delayed_work(&req->ctx->fallback_work, 1);
2077 static void io_req_task_cancel(struct io_kiocb *req)
2079 struct io_ring_ctx *ctx = req->ctx;
2081 /* ctx is guaranteed to stay alive while we hold uring_lock */
2082 mutex_lock(&ctx->uring_lock);
2083 io_req_complete_failed(req, req->result);
2084 mutex_unlock(&ctx->uring_lock);
2087 static void io_req_task_submit(struct io_kiocb *req)
2089 struct io_ring_ctx *ctx = req->ctx;
2091 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2092 mutex_lock(&ctx->uring_lock);
2093 if (likely(!(req->task->flags & PF_EXITING)))
2094 __io_queue_sqe(req);
2096 io_req_complete_failed(req, -EFAULT);
2097 mutex_unlock(&ctx->uring_lock);
2100 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2103 req->io_task_work.func = io_req_task_cancel;
2104 io_req_task_work_add(req);
2107 static void io_req_task_queue(struct io_kiocb *req)
2109 req->io_task_work.func = io_req_task_submit;
2110 io_req_task_work_add(req);
2113 static void io_req_task_queue_reissue(struct io_kiocb *req)
2115 req->io_task_work.func = io_queue_async_work;
2116 io_req_task_work_add(req);
2119 static inline void io_queue_next(struct io_kiocb *req)
2121 struct io_kiocb *nxt = io_req_find_next(req);
2124 io_req_task_queue(nxt);
2127 static void io_free_req(struct io_kiocb *req)
2134 struct task_struct *task;
2139 static inline void io_init_req_batch(struct req_batch *rb)
2146 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2147 struct req_batch *rb)
2150 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2151 if (rb->task == current)
2152 current->io_uring->cached_refs += rb->task_refs;
2154 io_put_task(rb->task, rb->task_refs);
2157 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2158 struct io_submit_state *state)
2161 io_dismantle_req(req);
2163 if (req->task != rb->task) {
2165 io_put_task(rb->task, rb->task_refs);
2166 rb->task = req->task;
2172 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2173 state->reqs[state->free_reqs++] = req;
2175 list_add(&req->inflight_entry, &state->free_list);
2178 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2179 __must_hold(&ctx->uring_lock)
2181 struct io_submit_state *state = &ctx->submit_state;
2182 int i, nr = state->compl_nr;
2183 struct req_batch rb;
2185 spin_lock(&ctx->completion_lock);
2186 for (i = 0; i < nr; i++) {
2187 struct io_kiocb *req = state->compl_reqs[i];
2189 __io_cqring_fill_event(ctx, req->user_data, req->result,
2192 io_commit_cqring(ctx);
2193 spin_unlock(&ctx->completion_lock);
2194 io_cqring_ev_posted(ctx);
2196 io_init_req_batch(&rb);
2197 for (i = 0; i < nr; i++) {
2198 struct io_kiocb *req = state->compl_reqs[i];
2200 if (req_ref_put_and_test(req))
2201 io_req_free_batch(&rb, req, &ctx->submit_state);
2204 io_req_free_batch_finish(ctx, &rb);
2205 state->compl_nr = 0;
2209 * Drop reference to request, return next in chain (if there is one) if this
2210 * was the last reference to this request.
2212 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2214 struct io_kiocb *nxt = NULL;
2216 if (req_ref_put_and_test(req)) {
2217 nxt = io_req_find_next(req);
2223 static inline void io_put_req(struct io_kiocb *req)
2225 if (req_ref_put_and_test(req))
2229 static inline void io_put_req_deferred(struct io_kiocb *req)
2231 if (req_ref_put_and_test(req)) {
2232 req->io_task_work.func = io_free_req;
2233 io_req_task_work_add(req);
2237 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2239 /* See comment at the top of this file */
2241 return __io_cqring_events(ctx);
2244 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2246 struct io_rings *rings = ctx->rings;
2248 /* make sure SQ entry isn't read before tail */
2249 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2252 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2254 unsigned int cflags;
2256 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2257 cflags |= IORING_CQE_F_BUFFER;
2258 req->flags &= ~REQ_F_BUFFER_SELECTED;
2263 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2265 struct io_buffer *kbuf;
2267 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2268 return io_put_kbuf(req, kbuf);
2271 static inline bool io_run_task_work(void)
2273 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2274 __set_current_state(TASK_RUNNING);
2275 tracehook_notify_signal();
2283 * Find and free completed poll iocbs
2285 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2286 struct list_head *done)
2288 struct req_batch rb;
2289 struct io_kiocb *req;
2291 /* order with ->result store in io_complete_rw_iopoll() */
2294 io_init_req_batch(&rb);
2295 while (!list_empty(done)) {
2298 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2299 list_del(&req->inflight_entry);
2301 if (READ_ONCE(req->result) == -EAGAIN &&
2302 !(req->flags & REQ_F_DONT_REISSUE)) {
2303 req->iopoll_completed = 0;
2304 io_req_task_queue_reissue(req);
2308 if (req->flags & REQ_F_BUFFER_SELECTED)
2309 cflags = io_put_rw_kbuf(req);
2311 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2314 if (req_ref_put_and_test(req))
2315 io_req_free_batch(&rb, req, &ctx->submit_state);
2318 io_commit_cqring(ctx);
2319 io_cqring_ev_posted_iopoll(ctx);
2320 io_req_free_batch_finish(ctx, &rb);
2323 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2326 struct io_kiocb *req, *tmp;
2331 * Only spin for completions if we don't have multiple devices hanging
2332 * off our complete list, and we're under the requested amount.
2334 spin = !ctx->poll_multi_queue && *nr_events < min;
2336 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2337 struct kiocb *kiocb = &req->rw.kiocb;
2341 * Move completed and retryable entries to our local lists.
2342 * If we find a request that requires polling, break out
2343 * and complete those lists first, if we have entries there.
2345 if (READ_ONCE(req->iopoll_completed)) {
2346 list_move_tail(&req->inflight_entry, &done);
2349 if (!list_empty(&done))
2352 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2353 if (unlikely(ret < 0))
2358 /* iopoll may have completed current req */
2359 if (READ_ONCE(req->iopoll_completed))
2360 list_move_tail(&req->inflight_entry, &done);
2363 if (!list_empty(&done))
2364 io_iopoll_complete(ctx, nr_events, &done);
2370 * We can't just wait for polled events to come to us, we have to actively
2371 * find and complete them.
2373 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2375 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2378 mutex_lock(&ctx->uring_lock);
2379 while (!list_empty(&ctx->iopoll_list)) {
2380 unsigned int nr_events = 0;
2382 io_do_iopoll(ctx, &nr_events, 0);
2384 /* let it sleep and repeat later if can't complete a request */
2388 * Ensure we allow local-to-the-cpu processing to take place,
2389 * in this case we need to ensure that we reap all events.
2390 * Also let task_work, etc. to progress by releasing the mutex
2392 if (need_resched()) {
2393 mutex_unlock(&ctx->uring_lock);
2395 mutex_lock(&ctx->uring_lock);
2398 mutex_unlock(&ctx->uring_lock);
2401 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2403 unsigned int nr_events = 0;
2407 * We disallow the app entering submit/complete with polling, but we
2408 * still need to lock the ring to prevent racing with polled issue
2409 * that got punted to a workqueue.
2411 mutex_lock(&ctx->uring_lock);
2413 * Don't enter poll loop if we already have events pending.
2414 * If we do, we can potentially be spinning for commands that
2415 * already triggered a CQE (eg in error).
2417 if (test_bit(0, &ctx->check_cq_overflow))
2418 __io_cqring_overflow_flush(ctx, false);
2419 if (io_cqring_events(ctx))
2423 * If a submit got punted to a workqueue, we can have the
2424 * application entering polling for a command before it gets
2425 * issued. That app will hold the uring_lock for the duration
2426 * of the poll right here, so we need to take a breather every
2427 * now and then to ensure that the issue has a chance to add
2428 * the poll to the issued list. Otherwise we can spin here
2429 * forever, while the workqueue is stuck trying to acquire the
2432 if (list_empty(&ctx->iopoll_list)) {
2433 u32 tail = ctx->cached_cq_tail;
2435 mutex_unlock(&ctx->uring_lock);
2437 mutex_lock(&ctx->uring_lock);
2439 /* some requests don't go through iopoll_list */
2440 if (tail != ctx->cached_cq_tail ||
2441 list_empty(&ctx->iopoll_list))
2444 ret = io_do_iopoll(ctx, &nr_events, min);
2445 } while (!ret && nr_events < min && !need_resched());
2447 mutex_unlock(&ctx->uring_lock);
2451 static void kiocb_end_write(struct io_kiocb *req)
2454 * Tell lockdep we inherited freeze protection from submission
2457 if (req->flags & REQ_F_ISREG) {
2458 struct super_block *sb = file_inode(req->file)->i_sb;
2460 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2466 static bool io_resubmit_prep(struct io_kiocb *req)
2468 struct io_async_rw *rw = req->async_data;
2471 return !io_req_prep_async(req);
2472 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2473 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2477 static bool io_rw_should_reissue(struct io_kiocb *req)
2479 umode_t mode = file_inode(req->file)->i_mode;
2480 struct io_ring_ctx *ctx = req->ctx;
2482 if (!S_ISBLK(mode) && !S_ISREG(mode))
2484 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2485 !(ctx->flags & IORING_SETUP_IOPOLL)))
2488 * If ref is dying, we might be running poll reap from the exit work.
2489 * Don't attempt to reissue from that path, just let it fail with
2492 if (percpu_ref_is_dying(&ctx->refs))
2495 * Play it safe and assume not safe to re-import and reissue if we're
2496 * not in the original thread group (or in task context).
2498 if (!same_thread_group(req->task, current) || !in_task())
2503 static bool io_resubmit_prep(struct io_kiocb *req)
2507 static bool io_rw_should_reissue(struct io_kiocb *req)
2513 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2515 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2516 kiocb_end_write(req);
2517 if (res != req->result) {
2518 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2519 io_rw_should_reissue(req)) {
2520 req->flags |= REQ_F_REISSUE;
2529 static void io_req_task_complete(struct io_kiocb *req)
2533 if (req->flags & REQ_F_BUFFER_SELECTED)
2534 cflags = io_put_rw_kbuf(req);
2535 __io_req_complete(req, 0, req->result, cflags);
2538 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2539 unsigned int issue_flags)
2541 if (__io_complete_rw_common(req, res))
2543 io_req_task_complete(req);
2546 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2548 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2550 if (__io_complete_rw_common(req, res))
2553 req->io_task_work.func = io_req_task_complete;
2554 io_req_task_work_add(req);
2557 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2559 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2561 if (kiocb->ki_flags & IOCB_WRITE)
2562 kiocb_end_write(req);
2563 if (unlikely(res != req->result)) {
2564 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2565 io_resubmit_prep(req))) {
2567 req->flags |= REQ_F_DONT_REISSUE;
2571 WRITE_ONCE(req->result, res);
2572 /* order with io_iopoll_complete() checking ->result */
2574 WRITE_ONCE(req->iopoll_completed, 1);
2578 * After the iocb has been issued, it's safe to be found on the poll list.
2579 * Adding the kiocb to the list AFTER submission ensures that we don't
2580 * find it from a io_do_iopoll() thread before the issuer is done
2581 * accessing the kiocb cookie.
2583 static void io_iopoll_req_issued(struct io_kiocb *req)
2585 struct io_ring_ctx *ctx = req->ctx;
2586 const bool in_async = io_wq_current_is_worker();
2588 /* workqueue context doesn't hold uring_lock, grab it now */
2589 if (unlikely(in_async))
2590 mutex_lock(&ctx->uring_lock);
2593 * Track whether we have multiple files in our lists. This will impact
2594 * how we do polling eventually, not spinning if we're on potentially
2595 * different devices.
2597 if (list_empty(&ctx->iopoll_list)) {
2598 ctx->poll_multi_queue = false;
2599 } else if (!ctx->poll_multi_queue) {
2600 struct io_kiocb *list_req;
2601 unsigned int queue_num0, queue_num1;
2603 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2606 if (list_req->file != req->file) {
2607 ctx->poll_multi_queue = true;
2609 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2610 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2611 if (queue_num0 != queue_num1)
2612 ctx->poll_multi_queue = true;
2617 * For fast devices, IO may have already completed. If it has, add
2618 * it to the front so we find it first.
2620 if (READ_ONCE(req->iopoll_completed))
2621 list_add(&req->inflight_entry, &ctx->iopoll_list);
2623 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2625 if (unlikely(in_async)) {
2627 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2628 * in sq thread task context or in io worker task context. If
2629 * current task context is sq thread, we don't need to check
2630 * whether should wake up sq thread.
2632 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2633 wq_has_sleeper(&ctx->sq_data->wait))
2634 wake_up(&ctx->sq_data->wait);
2636 mutex_unlock(&ctx->uring_lock);
2640 static bool io_bdev_nowait(struct block_device *bdev)
2642 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2646 * If we tracked the file through the SCM inflight mechanism, we could support
2647 * any file. For now, just ensure that anything potentially problematic is done
2650 static bool __io_file_supports_nowait(struct file *file, int rw)
2652 umode_t mode = file_inode(file)->i_mode;
2654 if (S_ISBLK(mode)) {
2655 if (IS_ENABLED(CONFIG_BLOCK) &&
2656 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2662 if (S_ISREG(mode)) {
2663 if (IS_ENABLED(CONFIG_BLOCK) &&
2664 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2665 file->f_op != &io_uring_fops)
2670 /* any ->read/write should understand O_NONBLOCK */
2671 if (file->f_flags & O_NONBLOCK)
2674 if (!(file->f_mode & FMODE_NOWAIT))
2678 return file->f_op->read_iter != NULL;
2680 return file->f_op->write_iter != NULL;
2683 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2685 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2687 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2690 return __io_file_supports_nowait(req->file, rw);
2693 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2695 struct io_ring_ctx *ctx = req->ctx;
2696 struct kiocb *kiocb = &req->rw.kiocb;
2697 struct file *file = req->file;
2701 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2702 req->flags |= REQ_F_ISREG;
2704 kiocb->ki_pos = READ_ONCE(sqe->off);
2705 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2706 req->flags |= REQ_F_CUR_POS;
2707 kiocb->ki_pos = file->f_pos;
2709 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2710 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2711 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2715 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2716 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2717 req->flags |= REQ_F_NOWAIT;
2719 ioprio = READ_ONCE(sqe->ioprio);
2721 ret = ioprio_check_cap(ioprio);
2725 kiocb->ki_ioprio = ioprio;
2727 kiocb->ki_ioprio = get_current_ioprio();
2729 if (ctx->flags & IORING_SETUP_IOPOLL) {
2730 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2731 !kiocb->ki_filp->f_op->iopoll)
2734 kiocb->ki_flags |= IOCB_HIPRI;
2735 kiocb->ki_complete = io_complete_rw_iopoll;
2736 req->iopoll_completed = 0;
2738 if (kiocb->ki_flags & IOCB_HIPRI)
2740 kiocb->ki_complete = io_complete_rw;
2743 if (req->opcode == IORING_OP_READ_FIXED ||
2744 req->opcode == IORING_OP_WRITE_FIXED) {
2746 io_req_set_rsrc_node(req);
2749 req->rw.addr = READ_ONCE(sqe->addr);
2750 req->rw.len = READ_ONCE(sqe->len);
2751 req->buf_index = READ_ONCE(sqe->buf_index);
2755 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2761 case -ERESTARTNOINTR:
2762 case -ERESTARTNOHAND:
2763 case -ERESTART_RESTARTBLOCK:
2765 * We can't just restart the syscall, since previously
2766 * submitted sqes may already be in progress. Just fail this
2772 kiocb->ki_complete(kiocb, ret, 0);
2776 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2777 unsigned int issue_flags)
2779 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2780 struct io_async_rw *io = req->async_data;
2781 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2783 /* add previously done IO, if any */
2784 if (io && io->bytes_done > 0) {
2786 ret = io->bytes_done;
2788 ret += io->bytes_done;
2791 if (req->flags & REQ_F_CUR_POS)
2792 req->file->f_pos = kiocb->ki_pos;
2793 if (ret >= 0 && check_reissue)
2794 __io_complete_rw(req, ret, 0, issue_flags);
2796 io_rw_done(kiocb, ret);
2798 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2799 req->flags &= ~REQ_F_REISSUE;
2800 if (io_resubmit_prep(req)) {
2801 io_req_task_queue_reissue(req);
2806 if (req->flags & REQ_F_BUFFER_SELECTED)
2807 cflags = io_put_rw_kbuf(req);
2808 __io_req_complete(req, issue_flags, ret, cflags);
2813 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2814 struct io_mapped_ubuf *imu)
2816 size_t len = req->rw.len;
2817 u64 buf_end, buf_addr = req->rw.addr;
2820 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2822 /* not inside the mapped region */
2823 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2827 * May not be a start of buffer, set size appropriately
2828 * and advance us to the beginning.
2830 offset = buf_addr - imu->ubuf;
2831 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2835 * Don't use iov_iter_advance() here, as it's really slow for
2836 * using the latter parts of a big fixed buffer - it iterates
2837 * over each segment manually. We can cheat a bit here, because
2840 * 1) it's a BVEC iter, we set it up
2841 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2842 * first and last bvec
2844 * So just find our index, and adjust the iterator afterwards.
2845 * If the offset is within the first bvec (or the whole first
2846 * bvec, just use iov_iter_advance(). This makes it easier
2847 * since we can just skip the first segment, which may not
2848 * be PAGE_SIZE aligned.
2850 const struct bio_vec *bvec = imu->bvec;
2852 if (offset <= bvec->bv_len) {
2853 iov_iter_advance(iter, offset);
2855 unsigned long seg_skip;
2857 /* skip first vec */
2858 offset -= bvec->bv_len;
2859 seg_skip = 1 + (offset >> PAGE_SHIFT);
2861 iter->bvec = bvec + seg_skip;
2862 iter->nr_segs -= seg_skip;
2863 iter->count -= bvec->bv_len + offset;
2864 iter->iov_offset = offset & ~PAGE_MASK;
2871 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2873 struct io_ring_ctx *ctx = req->ctx;
2874 struct io_mapped_ubuf *imu = req->imu;
2875 u16 index, buf_index = req->buf_index;
2878 if (unlikely(buf_index >= ctx->nr_user_bufs))
2880 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2881 imu = READ_ONCE(ctx->user_bufs[index]);
2884 return __io_import_fixed(req, rw, iter, imu);
2887 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2890 mutex_unlock(&ctx->uring_lock);
2893 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2896 * "Normal" inline submissions always hold the uring_lock, since we
2897 * grab it from the system call. Same is true for the SQPOLL offload.
2898 * The only exception is when we've detached the request and issue it
2899 * from an async worker thread, grab the lock for that case.
2902 mutex_lock(&ctx->uring_lock);
2905 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2906 int bgid, struct io_buffer *kbuf,
2909 struct io_buffer *head;
2911 if (req->flags & REQ_F_BUFFER_SELECTED)
2914 io_ring_submit_lock(req->ctx, needs_lock);
2916 lockdep_assert_held(&req->ctx->uring_lock);
2918 head = xa_load(&req->ctx->io_buffers, bgid);
2920 if (!list_empty(&head->list)) {
2921 kbuf = list_last_entry(&head->list, struct io_buffer,
2923 list_del(&kbuf->list);
2926 xa_erase(&req->ctx->io_buffers, bgid);
2928 if (*len > kbuf->len)
2931 kbuf = ERR_PTR(-ENOBUFS);
2934 io_ring_submit_unlock(req->ctx, needs_lock);
2939 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2942 struct io_buffer *kbuf;
2945 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2946 bgid = req->buf_index;
2947 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2950 req->rw.addr = (u64) (unsigned long) kbuf;
2951 req->flags |= REQ_F_BUFFER_SELECTED;
2952 return u64_to_user_ptr(kbuf->addr);
2955 #ifdef CONFIG_COMPAT
2956 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2959 struct compat_iovec __user *uiov;
2960 compat_ssize_t clen;
2964 uiov = u64_to_user_ptr(req->rw.addr);
2965 if (!access_ok(uiov, sizeof(*uiov)))
2967 if (__get_user(clen, &uiov->iov_len))
2973 buf = io_rw_buffer_select(req, &len, needs_lock);
2975 return PTR_ERR(buf);
2976 iov[0].iov_base = buf;
2977 iov[0].iov_len = (compat_size_t) len;
2982 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2985 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2989 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2992 len = iov[0].iov_len;
2995 buf = io_rw_buffer_select(req, &len, needs_lock);
2997 return PTR_ERR(buf);
2998 iov[0].iov_base = buf;
2999 iov[0].iov_len = len;
3003 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3006 if (req->flags & REQ_F_BUFFER_SELECTED) {
3007 struct io_buffer *kbuf;
3009 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3010 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3011 iov[0].iov_len = kbuf->len;
3014 if (req->rw.len != 1)
3017 #ifdef CONFIG_COMPAT
3018 if (req->ctx->compat)
3019 return io_compat_import(req, iov, needs_lock);
3022 return __io_iov_buffer_select(req, iov, needs_lock);
3025 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3026 struct iov_iter *iter, bool needs_lock)
3028 void __user *buf = u64_to_user_ptr(req->rw.addr);
3029 size_t sqe_len = req->rw.len;
3030 u8 opcode = req->opcode;
3033 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3035 return io_import_fixed(req, rw, iter);
3038 /* buffer index only valid with fixed read/write, or buffer select */
3039 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3042 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3043 if (req->flags & REQ_F_BUFFER_SELECT) {
3044 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3046 return PTR_ERR(buf);
3047 req->rw.len = sqe_len;
3050 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3055 if (req->flags & REQ_F_BUFFER_SELECT) {
3056 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3058 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3063 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3067 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3069 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3073 * For files that don't have ->read_iter() and ->write_iter(), handle them
3074 * by looping over ->read() or ->write() manually.
3076 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3078 struct kiocb *kiocb = &req->rw.kiocb;
3079 struct file *file = req->file;
3083 * Don't support polled IO through this interface, and we can't
3084 * support non-blocking either. For the latter, this just causes
3085 * the kiocb to be handled from an async context.
3087 if (kiocb->ki_flags & IOCB_HIPRI)
3089 if (kiocb->ki_flags & IOCB_NOWAIT)
3092 while (iov_iter_count(iter)) {
3096 if (!iov_iter_is_bvec(iter)) {
3097 iovec = iov_iter_iovec(iter);
3099 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3100 iovec.iov_len = req->rw.len;
3104 nr = file->f_op->read(file, iovec.iov_base,
3105 iovec.iov_len, io_kiocb_ppos(kiocb));
3107 nr = file->f_op->write(file, iovec.iov_base,
3108 iovec.iov_len, io_kiocb_ppos(kiocb));
3117 if (nr != iovec.iov_len)
3121 iov_iter_advance(iter, nr);
3127 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3128 const struct iovec *fast_iov, struct iov_iter *iter)
3130 struct io_async_rw *rw = req->async_data;
3132 memcpy(&rw->iter, iter, sizeof(*iter));
3133 rw->free_iovec = iovec;
3135 /* can only be fixed buffers, no need to do anything */
3136 if (iov_iter_is_bvec(iter))
3139 unsigned iov_off = 0;
3141 rw->iter.iov = rw->fast_iov;
3142 if (iter->iov != fast_iov) {
3143 iov_off = iter->iov - fast_iov;
3144 rw->iter.iov += iov_off;
3146 if (rw->fast_iov != fast_iov)
3147 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3148 sizeof(struct iovec) * iter->nr_segs);
3150 req->flags |= REQ_F_NEED_CLEANUP;
3154 static inline int io_alloc_async_data(struct io_kiocb *req)
3156 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3157 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3158 return req->async_data == NULL;
3161 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3162 const struct iovec *fast_iov,
3163 struct iov_iter *iter, bool force)
3165 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3167 if (!req->async_data) {
3168 if (io_alloc_async_data(req)) {
3173 io_req_map_rw(req, iovec, fast_iov, iter);
3178 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3180 struct io_async_rw *iorw = req->async_data;
3181 struct iovec *iov = iorw->fast_iov;
3184 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3185 if (unlikely(ret < 0))
3188 iorw->bytes_done = 0;
3189 iorw->free_iovec = iov;
3191 req->flags |= REQ_F_NEED_CLEANUP;
3195 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3197 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3199 return io_prep_rw(req, sqe);
3203 * This is our waitqueue callback handler, registered through lock_page_async()
3204 * when we initially tried to do the IO with the iocb armed our waitqueue.
3205 * This gets called when the page is unlocked, and we generally expect that to
3206 * happen when the page IO is completed and the page is now uptodate. This will
3207 * queue a task_work based retry of the operation, attempting to copy the data
3208 * again. If the latter fails because the page was NOT uptodate, then we will
3209 * do a thread based blocking retry of the operation. That's the unexpected
3212 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3213 int sync, void *arg)
3215 struct wait_page_queue *wpq;
3216 struct io_kiocb *req = wait->private;
3217 struct wait_page_key *key = arg;
3219 wpq = container_of(wait, struct wait_page_queue, wait);
3221 if (!wake_page_match(wpq, key))
3224 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3225 list_del_init(&wait->entry);
3226 io_req_task_queue(req);
3231 * This controls whether a given IO request should be armed for async page
3232 * based retry. If we return false here, the request is handed to the async
3233 * worker threads for retry. If we're doing buffered reads on a regular file,
3234 * we prepare a private wait_page_queue entry and retry the operation. This
3235 * will either succeed because the page is now uptodate and unlocked, or it
3236 * will register a callback when the page is unlocked at IO completion. Through
3237 * that callback, io_uring uses task_work to setup a retry of the operation.
3238 * That retry will attempt the buffered read again. The retry will generally
3239 * succeed, or in rare cases where it fails, we then fall back to using the
3240 * async worker threads for a blocking retry.
3242 static bool io_rw_should_retry(struct io_kiocb *req)
3244 struct io_async_rw *rw = req->async_data;
3245 struct wait_page_queue *wait = &rw->wpq;
3246 struct kiocb *kiocb = &req->rw.kiocb;
3248 /* never retry for NOWAIT, we just complete with -EAGAIN */
3249 if (req->flags & REQ_F_NOWAIT)
3252 /* Only for buffered IO */
3253 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3257 * just use poll if we can, and don't attempt if the fs doesn't
3258 * support callback based unlocks
3260 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3263 wait->wait.func = io_async_buf_func;
3264 wait->wait.private = req;
3265 wait->wait.flags = 0;
3266 INIT_LIST_HEAD(&wait->wait.entry);
3267 kiocb->ki_flags |= IOCB_WAITQ;
3268 kiocb->ki_flags &= ~IOCB_NOWAIT;
3269 kiocb->ki_waitq = wait;
3273 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3275 if (req->file->f_op->read_iter)
3276 return call_read_iter(req->file, &req->rw.kiocb, iter);
3277 else if (req->file->f_op->read)
3278 return loop_rw_iter(READ, req, iter);
3283 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3285 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3286 struct kiocb *kiocb = &req->rw.kiocb;
3287 struct iov_iter __iter, *iter = &__iter;
3288 struct io_async_rw *rw = req->async_data;
3289 ssize_t io_size, ret, ret2;
3290 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3296 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3300 io_size = iov_iter_count(iter);
3301 req->result = io_size;
3303 /* Ensure we clear previously set non-block flag */
3304 if (!force_nonblock)
3305 kiocb->ki_flags &= ~IOCB_NOWAIT;
3307 kiocb->ki_flags |= IOCB_NOWAIT;
3309 /* If the file doesn't support async, just async punt */
3310 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3311 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3312 return ret ?: -EAGAIN;
3315 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3316 if (unlikely(ret)) {
3321 ret = io_iter_do_read(req, iter);
3323 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3324 req->flags &= ~REQ_F_REISSUE;
3325 /* IOPOLL retry should happen for io-wq threads */
3326 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3328 /* no retry on NONBLOCK nor RWF_NOWAIT */
3329 if (req->flags & REQ_F_NOWAIT)
3331 /* some cases will consume bytes even on error returns */
3332 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3334 } else if (ret == -EIOCBQUEUED) {
3336 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3337 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3338 /* read all, failed, already did sync or don't want to retry */
3342 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3347 rw = req->async_data;
3348 /* now use our persistent iterator, if we aren't already */
3353 rw->bytes_done += ret;
3354 /* if we can retry, do so with the callbacks armed */
3355 if (!io_rw_should_retry(req)) {
3356 kiocb->ki_flags &= ~IOCB_WAITQ;
3361 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3362 * we get -EIOCBQUEUED, then we'll get a notification when the
3363 * desired page gets unlocked. We can also get a partial read
3364 * here, and if we do, then just retry at the new offset.
3366 ret = io_iter_do_read(req, iter);
3367 if (ret == -EIOCBQUEUED)
3369 /* we got some bytes, but not all. retry. */
3370 kiocb->ki_flags &= ~IOCB_WAITQ;
3371 } while (ret > 0 && ret < io_size);
3373 kiocb_done(kiocb, ret, issue_flags);
3375 /* it's faster to check here then delegate to kfree */
3381 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3383 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3385 return io_prep_rw(req, sqe);
3388 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3390 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3391 struct kiocb *kiocb = &req->rw.kiocb;
3392 struct iov_iter __iter, *iter = &__iter;
3393 struct io_async_rw *rw = req->async_data;
3394 ssize_t ret, ret2, io_size;
3395 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3401 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3405 io_size = iov_iter_count(iter);
3406 req->result = io_size;
3408 /* Ensure we clear previously set non-block flag */
3409 if (!force_nonblock)
3410 kiocb->ki_flags &= ~IOCB_NOWAIT;
3412 kiocb->ki_flags |= IOCB_NOWAIT;
3414 /* If the file doesn't support async, just async punt */
3415 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3418 /* file path doesn't support NOWAIT for non-direct_IO */
3419 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3420 (req->flags & REQ_F_ISREG))
3423 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3428 * Open-code file_start_write here to grab freeze protection,
3429 * which will be released by another thread in
3430 * io_complete_rw(). Fool lockdep by telling it the lock got
3431 * released so that it doesn't complain about the held lock when
3432 * we return to userspace.
3434 if (req->flags & REQ_F_ISREG) {
3435 sb_start_write(file_inode(req->file)->i_sb);
3436 __sb_writers_release(file_inode(req->file)->i_sb,
3439 kiocb->ki_flags |= IOCB_WRITE;
3441 if (req->file->f_op->write_iter)
3442 ret2 = call_write_iter(req->file, kiocb, iter);
3443 else if (req->file->f_op->write)
3444 ret2 = loop_rw_iter(WRITE, req, iter);
3448 if (req->flags & REQ_F_REISSUE) {
3449 req->flags &= ~REQ_F_REISSUE;
3454 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3455 * retry them without IOCB_NOWAIT.
3457 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3459 /* no retry on NONBLOCK nor RWF_NOWAIT */
3460 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3462 if (!force_nonblock || ret2 != -EAGAIN) {
3463 /* IOPOLL retry should happen for io-wq threads */
3464 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3467 kiocb_done(kiocb, ret2, issue_flags);
3470 /* some cases will consume bytes even on error returns */
3471 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3472 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3473 return ret ?: -EAGAIN;
3476 /* it's reportedly faster than delegating the null check to kfree() */
3482 static int io_renameat_prep(struct io_kiocb *req,
3483 const struct io_uring_sqe *sqe)
3485 struct io_rename *ren = &req->rename;
3486 const char __user *oldf, *newf;
3488 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3490 if (sqe->ioprio || sqe->buf_index)
3492 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3495 ren->old_dfd = READ_ONCE(sqe->fd);
3496 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3497 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3498 ren->new_dfd = READ_ONCE(sqe->len);
3499 ren->flags = READ_ONCE(sqe->rename_flags);
3501 ren->oldpath = getname(oldf);
3502 if (IS_ERR(ren->oldpath))
3503 return PTR_ERR(ren->oldpath);
3505 ren->newpath = getname(newf);
3506 if (IS_ERR(ren->newpath)) {
3507 putname(ren->oldpath);
3508 return PTR_ERR(ren->newpath);
3511 req->flags |= REQ_F_NEED_CLEANUP;
3515 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3517 struct io_rename *ren = &req->rename;
3520 if (issue_flags & IO_URING_F_NONBLOCK)
3523 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3524 ren->newpath, ren->flags);
3526 req->flags &= ~REQ_F_NEED_CLEANUP;
3529 io_req_complete(req, ret);
3533 static int io_unlinkat_prep(struct io_kiocb *req,
3534 const struct io_uring_sqe *sqe)
3536 struct io_unlink *un = &req->unlink;
3537 const char __user *fname;
3539 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3541 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3543 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3546 un->dfd = READ_ONCE(sqe->fd);
3548 un->flags = READ_ONCE(sqe->unlink_flags);
3549 if (un->flags & ~AT_REMOVEDIR)
3552 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3553 un->filename = getname(fname);
3554 if (IS_ERR(un->filename))
3555 return PTR_ERR(un->filename);
3557 req->flags |= REQ_F_NEED_CLEANUP;
3561 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3563 struct io_unlink *un = &req->unlink;
3566 if (issue_flags & IO_URING_F_NONBLOCK)
3569 if (un->flags & AT_REMOVEDIR)
3570 ret = do_rmdir(un->dfd, un->filename);
3572 ret = do_unlinkat(un->dfd, un->filename);
3574 req->flags &= ~REQ_F_NEED_CLEANUP;
3577 io_req_complete(req, ret);
3581 static int io_shutdown_prep(struct io_kiocb *req,
3582 const struct io_uring_sqe *sqe)
3584 #if defined(CONFIG_NET)
3585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3587 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3591 req->shutdown.how = READ_ONCE(sqe->len);
3598 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3600 #if defined(CONFIG_NET)
3601 struct socket *sock;
3604 if (issue_flags & IO_URING_F_NONBLOCK)
3607 sock = sock_from_file(req->file);
3608 if (unlikely(!sock))
3611 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3614 io_req_complete(req, ret);
3621 static int __io_splice_prep(struct io_kiocb *req,
3622 const struct io_uring_sqe *sqe)
3624 struct io_splice *sp = &req->splice;
3625 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3627 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3631 sp->len = READ_ONCE(sqe->len);
3632 sp->flags = READ_ONCE(sqe->splice_flags);
3634 if (unlikely(sp->flags & ~valid_flags))
3637 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3638 (sp->flags & SPLICE_F_FD_IN_FIXED));
3641 req->flags |= REQ_F_NEED_CLEANUP;
3645 static int io_tee_prep(struct io_kiocb *req,
3646 const struct io_uring_sqe *sqe)
3648 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3650 return __io_splice_prep(req, sqe);
3653 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3655 struct io_splice *sp = &req->splice;
3656 struct file *in = sp->file_in;
3657 struct file *out = sp->file_out;
3658 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3661 if (issue_flags & IO_URING_F_NONBLOCK)
3664 ret = do_tee(in, out, sp->len, flags);
3666 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3668 req->flags &= ~REQ_F_NEED_CLEANUP;
3672 io_req_complete(req, ret);
3676 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3678 struct io_splice *sp = &req->splice;
3680 sp->off_in = READ_ONCE(sqe->splice_off_in);
3681 sp->off_out = READ_ONCE(sqe->off);
3682 return __io_splice_prep(req, sqe);
3685 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3687 struct io_splice *sp = &req->splice;
3688 struct file *in = sp->file_in;
3689 struct file *out = sp->file_out;
3690 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3691 loff_t *poff_in, *poff_out;
3694 if (issue_flags & IO_URING_F_NONBLOCK)
3697 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3698 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3701 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3703 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3705 req->flags &= ~REQ_F_NEED_CLEANUP;
3709 io_req_complete(req, ret);
3714 * IORING_OP_NOP just posts a completion event, nothing else.
3716 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3718 struct io_ring_ctx *ctx = req->ctx;
3720 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3723 __io_req_complete(req, issue_flags, 0, 0);
3727 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3729 struct io_ring_ctx *ctx = req->ctx;
3734 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3736 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3739 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3740 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3743 req->sync.off = READ_ONCE(sqe->off);
3744 req->sync.len = READ_ONCE(sqe->len);
3748 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3750 loff_t end = req->sync.off + req->sync.len;
3753 /* fsync always requires a blocking context */
3754 if (issue_flags & IO_URING_F_NONBLOCK)
3757 ret = vfs_fsync_range(req->file, req->sync.off,
3758 end > 0 ? end : LLONG_MAX,
3759 req->sync.flags & IORING_FSYNC_DATASYNC);
3762 io_req_complete(req, ret);
3766 static int io_fallocate_prep(struct io_kiocb *req,
3767 const struct io_uring_sqe *sqe)
3769 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3774 req->sync.off = READ_ONCE(sqe->off);
3775 req->sync.len = READ_ONCE(sqe->addr);
3776 req->sync.mode = READ_ONCE(sqe->len);
3780 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3784 /* fallocate always requiring blocking context */
3785 if (issue_flags & IO_URING_F_NONBLOCK)
3787 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3791 io_req_complete(req, ret);
3795 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3797 const char __user *fname;
3800 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 if (unlikely(sqe->ioprio || sqe->buf_index))
3804 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3807 /* open.how should be already initialised */
3808 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3809 req->open.how.flags |= O_LARGEFILE;
3811 req->open.dfd = READ_ONCE(sqe->fd);
3812 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3813 req->open.filename = getname(fname);
3814 if (IS_ERR(req->open.filename)) {
3815 ret = PTR_ERR(req->open.filename);
3816 req->open.filename = NULL;
3819 req->open.nofile = rlimit(RLIMIT_NOFILE);
3820 req->flags |= REQ_F_NEED_CLEANUP;
3824 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3826 u64 mode = READ_ONCE(sqe->len);
3827 u64 flags = READ_ONCE(sqe->open_flags);
3829 req->open.how = build_open_how(flags, mode);
3830 return __io_openat_prep(req, sqe);
3833 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3835 struct open_how __user *how;
3839 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3840 len = READ_ONCE(sqe->len);
3841 if (len < OPEN_HOW_SIZE_VER0)
3844 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3849 return __io_openat_prep(req, sqe);
3852 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3854 struct open_flags op;
3857 bool resolve_nonblock;
3860 ret = build_open_flags(&req->open.how, &op);
3863 nonblock_set = op.open_flag & O_NONBLOCK;
3864 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3865 if (issue_flags & IO_URING_F_NONBLOCK) {
3867 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3868 * it'll always -EAGAIN
3870 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3872 op.lookup_flags |= LOOKUP_CACHED;
3873 op.open_flag |= O_NONBLOCK;
3876 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3880 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3883 * We could hang on to this 'fd' on retrying, but seems like
3884 * marginal gain for something that is now known to be a slower
3885 * path. So just put it, and we'll get a new one when we retry.
3889 ret = PTR_ERR(file);
3890 /* only retry if RESOLVE_CACHED wasn't already set by application */
3891 if (ret == -EAGAIN &&
3892 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3897 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3898 file->f_flags &= ~O_NONBLOCK;
3899 fsnotify_open(file);
3900 fd_install(ret, file);
3902 putname(req->open.filename);
3903 req->flags &= ~REQ_F_NEED_CLEANUP;
3906 __io_req_complete(req, issue_flags, ret, 0);
3910 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3912 return io_openat2(req, issue_flags);
3915 static int io_remove_buffers_prep(struct io_kiocb *req,
3916 const struct io_uring_sqe *sqe)
3918 struct io_provide_buf *p = &req->pbuf;
3921 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3924 tmp = READ_ONCE(sqe->fd);
3925 if (!tmp || tmp > USHRT_MAX)
3928 memset(p, 0, sizeof(*p));
3930 p->bgid = READ_ONCE(sqe->buf_group);
3934 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3935 int bgid, unsigned nbufs)
3939 /* shouldn't happen */
3943 /* the head kbuf is the list itself */
3944 while (!list_empty(&buf->list)) {
3945 struct io_buffer *nxt;
3947 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3948 list_del(&nxt->list);
3955 xa_erase(&ctx->io_buffers, bgid);
3960 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3962 struct io_provide_buf *p = &req->pbuf;
3963 struct io_ring_ctx *ctx = req->ctx;
3964 struct io_buffer *head;
3966 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3968 io_ring_submit_lock(ctx, !force_nonblock);
3970 lockdep_assert_held(&ctx->uring_lock);
3973 head = xa_load(&ctx->io_buffers, p->bgid);
3975 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3979 /* complete before unlock, IOPOLL may need the lock */
3980 __io_req_complete(req, issue_flags, ret, 0);
3981 io_ring_submit_unlock(ctx, !force_nonblock);
3985 static int io_provide_buffers_prep(struct io_kiocb *req,
3986 const struct io_uring_sqe *sqe)
3988 unsigned long size, tmp_check;
3989 struct io_provide_buf *p = &req->pbuf;
3992 if (sqe->ioprio || sqe->rw_flags)
3995 tmp = READ_ONCE(sqe->fd);
3996 if (!tmp || tmp > USHRT_MAX)
3999 p->addr = READ_ONCE(sqe->addr);
4000 p->len = READ_ONCE(sqe->len);
4002 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4005 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4008 size = (unsigned long)p->len * p->nbufs;
4009 if (!access_ok(u64_to_user_ptr(p->addr), size))
4012 p->bgid = READ_ONCE(sqe->buf_group);
4013 tmp = READ_ONCE(sqe->off);
4014 if (tmp > USHRT_MAX)
4020 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4022 struct io_buffer *buf;
4023 u64 addr = pbuf->addr;
4024 int i, bid = pbuf->bid;
4026 for (i = 0; i < pbuf->nbufs; i++) {
4027 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4032 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4037 INIT_LIST_HEAD(&buf->list);
4040 list_add_tail(&buf->list, &(*head)->list);
4044 return i ? i : -ENOMEM;
4047 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4049 struct io_provide_buf *p = &req->pbuf;
4050 struct io_ring_ctx *ctx = req->ctx;
4051 struct io_buffer *head, *list;
4053 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4055 io_ring_submit_lock(ctx, !force_nonblock);
4057 lockdep_assert_held(&ctx->uring_lock);
4059 list = head = xa_load(&ctx->io_buffers, p->bgid);
4061 ret = io_add_buffers(p, &head);
4062 if (ret >= 0 && !list) {
4063 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4065 __io_remove_buffers(ctx, head, p->bgid, -1U);
4069 /* complete before unlock, IOPOLL may need the lock */
4070 __io_req_complete(req, issue_flags, ret, 0);
4071 io_ring_submit_unlock(ctx, !force_nonblock);
4075 static int io_epoll_ctl_prep(struct io_kiocb *req,
4076 const struct io_uring_sqe *sqe)
4078 #if defined(CONFIG_EPOLL)
4079 if (sqe->ioprio || sqe->buf_index)
4081 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4084 req->epoll.epfd = READ_ONCE(sqe->fd);
4085 req->epoll.op = READ_ONCE(sqe->len);
4086 req->epoll.fd = READ_ONCE(sqe->off);
4088 if (ep_op_has_event(req->epoll.op)) {
4089 struct epoll_event __user *ev;
4091 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4092 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4102 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4104 #if defined(CONFIG_EPOLL)
4105 struct io_epoll *ie = &req->epoll;
4107 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4109 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4110 if (force_nonblock && ret == -EAGAIN)
4115 __io_req_complete(req, issue_flags, ret, 0);
4122 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4125 if (sqe->ioprio || sqe->buf_index || sqe->off)
4127 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4130 req->madvise.addr = READ_ONCE(sqe->addr);
4131 req->madvise.len = READ_ONCE(sqe->len);
4132 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4139 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4141 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4142 struct io_madvise *ma = &req->madvise;
4145 if (issue_flags & IO_URING_F_NONBLOCK)
4148 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4151 io_req_complete(req, ret);
4158 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4160 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4162 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4165 req->fadvise.offset = READ_ONCE(sqe->off);
4166 req->fadvise.len = READ_ONCE(sqe->len);
4167 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4171 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4173 struct io_fadvise *fa = &req->fadvise;
4176 if (issue_flags & IO_URING_F_NONBLOCK) {
4177 switch (fa->advice) {
4178 case POSIX_FADV_NORMAL:
4179 case POSIX_FADV_RANDOM:
4180 case POSIX_FADV_SEQUENTIAL:
4187 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4190 __io_req_complete(req, issue_flags, ret, 0);
4194 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4196 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4198 if (sqe->ioprio || sqe->buf_index)
4200 if (req->flags & REQ_F_FIXED_FILE)
4203 req->statx.dfd = READ_ONCE(sqe->fd);
4204 req->statx.mask = READ_ONCE(sqe->len);
4205 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4206 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4207 req->statx.flags = READ_ONCE(sqe->statx_flags);
4212 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4214 struct io_statx *ctx = &req->statx;
4217 if (issue_flags & IO_URING_F_NONBLOCK)
4220 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4225 io_req_complete(req, ret);
4229 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4231 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4233 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4234 sqe->rw_flags || sqe->buf_index)
4236 if (req->flags & REQ_F_FIXED_FILE)
4239 req->close.fd = READ_ONCE(sqe->fd);
4243 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4245 struct files_struct *files = current->files;
4246 struct io_close *close = &req->close;
4247 struct fdtable *fdt;
4248 struct file *file = NULL;
4251 spin_lock(&files->file_lock);
4252 fdt = files_fdtable(files);
4253 if (close->fd >= fdt->max_fds) {
4254 spin_unlock(&files->file_lock);
4257 file = fdt->fd[close->fd];
4258 if (!file || file->f_op == &io_uring_fops) {
4259 spin_unlock(&files->file_lock);
4264 /* if the file has a flush method, be safe and punt to async */
4265 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4266 spin_unlock(&files->file_lock);
4270 ret = __close_fd_get_file(close->fd, &file);
4271 spin_unlock(&files->file_lock);
4278 /* No ->flush() or already async, safely close from here */
4279 ret = filp_close(file, current->files);
4285 __io_req_complete(req, issue_flags, ret, 0);
4289 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4291 struct io_ring_ctx *ctx = req->ctx;
4293 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4295 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4298 req->sync.off = READ_ONCE(sqe->off);
4299 req->sync.len = READ_ONCE(sqe->len);
4300 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4304 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4308 /* sync_file_range always requires a blocking context */
4309 if (issue_flags & IO_URING_F_NONBLOCK)
4312 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4316 io_req_complete(req, ret);
4320 #if defined(CONFIG_NET)
4321 static int io_setup_async_msg(struct io_kiocb *req,
4322 struct io_async_msghdr *kmsg)
4324 struct io_async_msghdr *async_msg = req->async_data;
4328 if (io_alloc_async_data(req)) {
4329 kfree(kmsg->free_iov);
4332 async_msg = req->async_data;
4333 req->flags |= REQ_F_NEED_CLEANUP;
4334 memcpy(async_msg, kmsg, sizeof(*kmsg));
4335 async_msg->msg.msg_name = &async_msg->addr;
4336 /* if were using fast_iov, set it to the new one */
4337 if (!async_msg->free_iov)
4338 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4343 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4344 struct io_async_msghdr *iomsg)
4346 iomsg->msg.msg_name = &iomsg->addr;
4347 iomsg->free_iov = iomsg->fast_iov;
4348 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4349 req->sr_msg.msg_flags, &iomsg->free_iov);
4352 static int io_sendmsg_prep_async(struct io_kiocb *req)
4356 ret = io_sendmsg_copy_hdr(req, req->async_data);
4358 req->flags |= REQ_F_NEED_CLEANUP;
4362 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4364 struct io_sr_msg *sr = &req->sr_msg;
4366 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4369 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4370 sr->len = READ_ONCE(sqe->len);
4371 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4372 if (sr->msg_flags & MSG_DONTWAIT)
4373 req->flags |= REQ_F_NOWAIT;
4375 #ifdef CONFIG_COMPAT
4376 if (req->ctx->compat)
4377 sr->msg_flags |= MSG_CMSG_COMPAT;
4382 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4384 struct io_async_msghdr iomsg, *kmsg;
4385 struct socket *sock;
4390 sock = sock_from_file(req->file);
4391 if (unlikely(!sock))
4394 kmsg = req->async_data;
4396 ret = io_sendmsg_copy_hdr(req, &iomsg);
4402 flags = req->sr_msg.msg_flags;
4403 if (issue_flags & IO_URING_F_NONBLOCK)
4404 flags |= MSG_DONTWAIT;
4405 if (flags & MSG_WAITALL)
4406 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4408 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4409 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4410 return io_setup_async_msg(req, kmsg);
4411 if (ret == -ERESTARTSYS)
4414 /* fast path, check for non-NULL to avoid function call */
4416 kfree(kmsg->free_iov);
4417 req->flags &= ~REQ_F_NEED_CLEANUP;
4420 __io_req_complete(req, issue_flags, ret, 0);
4424 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4426 struct io_sr_msg *sr = &req->sr_msg;
4429 struct socket *sock;
4434 sock = sock_from_file(req->file);
4435 if (unlikely(!sock))
4438 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4442 msg.msg_name = NULL;
4443 msg.msg_control = NULL;
4444 msg.msg_controllen = 0;
4445 msg.msg_namelen = 0;
4447 flags = req->sr_msg.msg_flags;
4448 if (issue_flags & IO_URING_F_NONBLOCK)
4449 flags |= MSG_DONTWAIT;
4450 if (flags & MSG_WAITALL)
4451 min_ret = iov_iter_count(&msg.msg_iter);
4453 msg.msg_flags = flags;
4454 ret = sock_sendmsg(sock, &msg);
4455 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4457 if (ret == -ERESTARTSYS)
4462 __io_req_complete(req, issue_flags, ret, 0);
4466 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4467 struct io_async_msghdr *iomsg)
4469 struct io_sr_msg *sr = &req->sr_msg;
4470 struct iovec __user *uiov;
4474 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4475 &iomsg->uaddr, &uiov, &iov_len);
4479 if (req->flags & REQ_F_BUFFER_SELECT) {
4482 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4484 sr->len = iomsg->fast_iov[0].iov_len;
4485 iomsg->free_iov = NULL;
4487 iomsg->free_iov = iomsg->fast_iov;
4488 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4489 &iomsg->free_iov, &iomsg->msg.msg_iter,
4498 #ifdef CONFIG_COMPAT
4499 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4500 struct io_async_msghdr *iomsg)
4502 struct io_sr_msg *sr = &req->sr_msg;
4503 struct compat_iovec __user *uiov;
4508 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4513 uiov = compat_ptr(ptr);
4514 if (req->flags & REQ_F_BUFFER_SELECT) {
4515 compat_ssize_t clen;
4519 if (!access_ok(uiov, sizeof(*uiov)))
4521 if (__get_user(clen, &uiov->iov_len))
4526 iomsg->free_iov = NULL;
4528 iomsg->free_iov = iomsg->fast_iov;
4529 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4530 UIO_FASTIOV, &iomsg->free_iov,
4531 &iomsg->msg.msg_iter, true);
4540 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4541 struct io_async_msghdr *iomsg)
4543 iomsg->msg.msg_name = &iomsg->addr;
4545 #ifdef CONFIG_COMPAT
4546 if (req->ctx->compat)
4547 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4550 return __io_recvmsg_copy_hdr(req, iomsg);
4553 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4556 struct io_sr_msg *sr = &req->sr_msg;
4557 struct io_buffer *kbuf;
4559 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4564 req->flags |= REQ_F_BUFFER_SELECTED;
4568 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4570 return io_put_kbuf(req, req->sr_msg.kbuf);
4573 static int io_recvmsg_prep_async(struct io_kiocb *req)
4577 ret = io_recvmsg_copy_hdr(req, req->async_data);
4579 req->flags |= REQ_F_NEED_CLEANUP;
4583 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4585 struct io_sr_msg *sr = &req->sr_msg;
4587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4590 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4591 sr->len = READ_ONCE(sqe->len);
4592 sr->bgid = READ_ONCE(sqe->buf_group);
4593 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4594 if (sr->msg_flags & MSG_DONTWAIT)
4595 req->flags |= REQ_F_NOWAIT;
4597 #ifdef CONFIG_COMPAT
4598 if (req->ctx->compat)
4599 sr->msg_flags |= MSG_CMSG_COMPAT;
4604 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4606 struct io_async_msghdr iomsg, *kmsg;
4607 struct socket *sock;
4608 struct io_buffer *kbuf;
4611 int ret, cflags = 0;
4612 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4614 sock = sock_from_file(req->file);
4615 if (unlikely(!sock))
4618 kmsg = req->async_data;
4620 ret = io_recvmsg_copy_hdr(req, &iomsg);
4626 if (req->flags & REQ_F_BUFFER_SELECT) {
4627 kbuf = io_recv_buffer_select(req, !force_nonblock);
4629 return PTR_ERR(kbuf);
4630 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4631 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4632 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4633 1, req->sr_msg.len);
4636 flags = req->sr_msg.msg_flags;
4638 flags |= MSG_DONTWAIT;
4639 if (flags & MSG_WAITALL)
4640 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4642 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4643 kmsg->uaddr, flags);
4644 if (force_nonblock && ret == -EAGAIN)
4645 return io_setup_async_msg(req, kmsg);
4646 if (ret == -ERESTARTSYS)
4649 if (req->flags & REQ_F_BUFFER_SELECTED)
4650 cflags = io_put_recv_kbuf(req);
4651 /* fast path, check for non-NULL to avoid function call */
4653 kfree(kmsg->free_iov);
4654 req->flags &= ~REQ_F_NEED_CLEANUP;
4655 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4657 __io_req_complete(req, issue_flags, ret, cflags);
4661 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4663 struct io_buffer *kbuf;
4664 struct io_sr_msg *sr = &req->sr_msg;
4666 void __user *buf = sr->buf;
4667 struct socket *sock;
4671 int ret, cflags = 0;
4672 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4674 sock = sock_from_file(req->file);
4675 if (unlikely(!sock))
4678 if (req->flags & REQ_F_BUFFER_SELECT) {
4679 kbuf = io_recv_buffer_select(req, !force_nonblock);
4681 return PTR_ERR(kbuf);
4682 buf = u64_to_user_ptr(kbuf->addr);
4685 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4689 msg.msg_name = NULL;
4690 msg.msg_control = NULL;
4691 msg.msg_controllen = 0;
4692 msg.msg_namelen = 0;
4693 msg.msg_iocb = NULL;
4696 flags = req->sr_msg.msg_flags;
4698 flags |= MSG_DONTWAIT;
4699 if (flags & MSG_WAITALL)
4700 min_ret = iov_iter_count(&msg.msg_iter);
4702 ret = sock_recvmsg(sock, &msg, flags);
4703 if (force_nonblock && ret == -EAGAIN)
4705 if (ret == -ERESTARTSYS)
4708 if (req->flags & REQ_F_BUFFER_SELECTED)
4709 cflags = io_put_recv_kbuf(req);
4710 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4712 __io_req_complete(req, issue_flags, ret, cflags);
4716 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_accept *accept = &req->accept;
4720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722 if (sqe->ioprio || sqe->len || sqe->buf_index)
4725 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4727 accept->flags = READ_ONCE(sqe->accept_flags);
4728 accept->nofile = rlimit(RLIMIT_NOFILE);
4732 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4734 struct io_accept *accept = &req->accept;
4735 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4736 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4739 if (req->file->f_flags & O_NONBLOCK)
4740 req->flags |= REQ_F_NOWAIT;
4742 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4743 accept->addr_len, accept->flags,
4745 if (ret == -EAGAIN && force_nonblock)
4748 if (ret == -ERESTARTSYS)
4752 __io_req_complete(req, issue_flags, ret, 0);
4756 static int io_connect_prep_async(struct io_kiocb *req)
4758 struct io_async_connect *io = req->async_data;
4759 struct io_connect *conn = &req->connect;
4761 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4764 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4766 struct io_connect *conn = &req->connect;
4768 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4770 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4773 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4774 conn->addr_len = READ_ONCE(sqe->addr2);
4778 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4780 struct io_async_connect __io, *io;
4781 unsigned file_flags;
4783 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4785 if (req->async_data) {
4786 io = req->async_data;
4788 ret = move_addr_to_kernel(req->connect.addr,
4789 req->connect.addr_len,
4796 file_flags = force_nonblock ? O_NONBLOCK : 0;
4798 ret = __sys_connect_file(req->file, &io->address,
4799 req->connect.addr_len, file_flags);
4800 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4801 if (req->async_data)
4803 if (io_alloc_async_data(req)) {
4807 memcpy(req->async_data, &__io, sizeof(__io));
4810 if (ret == -ERESTARTSYS)
4815 __io_req_complete(req, issue_flags, ret, 0);
4818 #else /* !CONFIG_NET */
4819 #define IO_NETOP_FN(op) \
4820 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4822 return -EOPNOTSUPP; \
4825 #define IO_NETOP_PREP(op) \
4827 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4829 return -EOPNOTSUPP; \
4832 #define IO_NETOP_PREP_ASYNC(op) \
4834 static int io_##op##_prep_async(struct io_kiocb *req) \
4836 return -EOPNOTSUPP; \
4839 IO_NETOP_PREP_ASYNC(sendmsg);
4840 IO_NETOP_PREP_ASYNC(recvmsg);
4841 IO_NETOP_PREP_ASYNC(connect);
4842 IO_NETOP_PREP(accept);
4845 #endif /* CONFIG_NET */
4847 struct io_poll_table {
4848 struct poll_table_struct pt;
4849 struct io_kiocb *req;
4854 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4855 __poll_t mask, io_req_tw_func_t func)
4857 /* for instances that support it check for an event match first: */
4858 if (mask && !(mask & poll->events))
4861 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4863 list_del_init(&poll->wait.entry);
4866 req->io_task_work.func = func;
4869 * If this fails, then the task is exiting. When a task exits, the
4870 * work gets canceled, so just cancel this request as well instead
4871 * of executing it. We can't safely execute it anyway, as we may not
4872 * have the needed state needed for it anyway.
4874 io_req_task_work_add(req);
4878 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4879 __acquires(&req->ctx->completion_lock)
4881 struct io_ring_ctx *ctx = req->ctx;
4883 if (unlikely(req->task->flags & PF_EXITING))
4884 WRITE_ONCE(poll->canceled, true);
4886 if (!req->result && !READ_ONCE(poll->canceled)) {
4887 struct poll_table_struct pt = { ._key = poll->events };
4889 req->result = vfs_poll(req->file, &pt) & poll->events;
4892 spin_lock(&ctx->completion_lock);
4893 if (!req->result && !READ_ONCE(poll->canceled)) {
4894 add_wait_queue(poll->head, &poll->wait);
4901 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4903 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4904 if (req->opcode == IORING_OP_POLL_ADD)
4905 return req->async_data;
4906 return req->apoll->double_poll;
4909 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4911 if (req->opcode == IORING_OP_POLL_ADD)
4913 return &req->apoll->poll;
4916 static void io_poll_remove_double(struct io_kiocb *req)
4917 __must_hold(&req->ctx->completion_lock)
4919 struct io_poll_iocb *poll = io_poll_get_double(req);
4921 lockdep_assert_held(&req->ctx->completion_lock);
4923 if (poll && poll->head) {
4924 struct wait_queue_head *head = poll->head;
4926 spin_lock_irq(&head->lock);
4927 list_del_init(&poll->wait.entry);
4928 if (poll->wait.private)
4931 spin_unlock_irq(&head->lock);
4935 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4936 __must_hold(&req->ctx->completion_lock)
4938 struct io_ring_ctx *ctx = req->ctx;
4939 unsigned flags = IORING_CQE_F_MORE;
4942 if (READ_ONCE(req->poll.canceled)) {
4944 req->poll.events |= EPOLLONESHOT;
4946 error = mangle_poll(mask);
4948 if (req->poll.events & EPOLLONESHOT)
4950 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4951 req->poll.done = true;
4954 if (flags & IORING_CQE_F_MORE)
4957 io_commit_cqring(ctx);
4958 return !(flags & IORING_CQE_F_MORE);
4961 static void io_poll_task_func(struct io_kiocb *req)
4963 struct io_ring_ctx *ctx = req->ctx;
4964 struct io_kiocb *nxt;
4966 if (io_poll_rewait(req, &req->poll)) {
4967 spin_unlock(&ctx->completion_lock);
4971 done = io_poll_complete(req, req->result);
4973 io_poll_remove_double(req);
4974 hash_del(&req->hash_node);
4977 add_wait_queue(req->poll.head, &req->poll.wait);
4979 spin_unlock(&ctx->completion_lock);
4980 io_cqring_ev_posted(ctx);
4983 nxt = io_put_req_find_next(req);
4985 io_req_task_submit(nxt);
4990 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4991 int sync, void *key)
4993 struct io_kiocb *req = wait->private;
4994 struct io_poll_iocb *poll = io_poll_get_single(req);
4995 __poll_t mask = key_to_poll(key);
4996 unsigned long flags;
4998 /* for instances that support it check for an event match first: */
4999 if (mask && !(mask & poll->events))
5001 if (!(poll->events & EPOLLONESHOT))
5002 return poll->wait.func(&poll->wait, mode, sync, key);
5004 list_del_init(&wait->entry);
5009 spin_lock_irqsave(&poll->head->lock, flags);
5010 done = list_empty(&poll->wait.entry);
5012 list_del_init(&poll->wait.entry);
5013 /* make sure double remove sees this as being gone */
5014 wait->private = NULL;
5015 spin_unlock_irqrestore(&poll->head->lock, flags);
5017 /* use wait func handler, so it matches the rq type */
5018 poll->wait.func(&poll->wait, mode, sync, key);
5025 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5026 wait_queue_func_t wake_func)
5030 poll->canceled = false;
5031 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5032 /* mask in events that we always want/need */
5033 poll->events = events | IO_POLL_UNMASK;
5034 INIT_LIST_HEAD(&poll->wait.entry);
5035 init_waitqueue_func_entry(&poll->wait, wake_func);
5038 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5039 struct wait_queue_head *head,
5040 struct io_poll_iocb **poll_ptr)
5042 struct io_kiocb *req = pt->req;
5045 * The file being polled uses multiple waitqueues for poll handling
5046 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5049 if (unlikely(pt->nr_entries)) {
5050 struct io_poll_iocb *poll_one = poll;
5052 /* already have a 2nd entry, fail a third attempt */
5054 pt->error = -EINVAL;
5058 * Can't handle multishot for double wait for now, turn it
5059 * into one-shot mode.
5061 if (!(poll_one->events & EPOLLONESHOT))
5062 poll_one->events |= EPOLLONESHOT;
5063 /* double add on the same waitqueue head, ignore */
5064 if (poll_one->head == head)
5066 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5068 pt->error = -ENOMEM;
5071 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5073 poll->wait.private = req;
5080 if (poll->events & EPOLLEXCLUSIVE)
5081 add_wait_queue_exclusive(head, &poll->wait);
5083 add_wait_queue(head, &poll->wait);
5086 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5087 struct poll_table_struct *p)
5089 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5090 struct async_poll *apoll = pt->req->apoll;
5092 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5095 static void io_async_task_func(struct io_kiocb *req)
5097 struct async_poll *apoll = req->apoll;
5098 struct io_ring_ctx *ctx = req->ctx;
5100 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5102 if (io_poll_rewait(req, &apoll->poll)) {
5103 spin_unlock(&ctx->completion_lock);
5107 hash_del(&req->hash_node);
5108 io_poll_remove_double(req);
5109 spin_unlock(&ctx->completion_lock);
5111 if (!READ_ONCE(apoll->poll.canceled))
5112 io_req_task_submit(req);
5114 io_req_complete_failed(req, -ECANCELED);
5117 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5120 struct io_kiocb *req = wait->private;
5121 struct io_poll_iocb *poll = &req->apoll->poll;
5123 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5126 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5129 static void io_poll_req_insert(struct io_kiocb *req)
5131 struct io_ring_ctx *ctx = req->ctx;
5132 struct hlist_head *list;
5134 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5135 hlist_add_head(&req->hash_node, list);
5138 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5139 struct io_poll_iocb *poll,
5140 struct io_poll_table *ipt, __poll_t mask,
5141 wait_queue_func_t wake_func)
5142 __acquires(&ctx->completion_lock)
5144 struct io_ring_ctx *ctx = req->ctx;
5145 bool cancel = false;
5147 INIT_HLIST_NODE(&req->hash_node);
5148 io_init_poll_iocb(poll, mask, wake_func);
5149 poll->file = req->file;
5150 poll->wait.private = req;
5152 ipt->pt._key = mask;
5155 ipt->nr_entries = 0;
5157 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5158 if (unlikely(!ipt->nr_entries) && !ipt->error)
5159 ipt->error = -EINVAL;
5161 spin_lock(&ctx->completion_lock);
5162 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5163 io_poll_remove_double(req);
5164 if (likely(poll->head)) {
5165 spin_lock_irq(&poll->head->lock);
5166 if (unlikely(list_empty(&poll->wait.entry))) {
5172 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5173 list_del_init(&poll->wait.entry);
5175 WRITE_ONCE(poll->canceled, true);
5176 else if (!poll->done) /* actually waiting for an event */
5177 io_poll_req_insert(req);
5178 spin_unlock_irq(&poll->head->lock);
5190 static int io_arm_poll_handler(struct io_kiocb *req)
5192 const struct io_op_def *def = &io_op_defs[req->opcode];
5193 struct io_ring_ctx *ctx = req->ctx;
5194 struct async_poll *apoll;
5195 struct io_poll_table ipt;
5196 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5199 if (!req->file || !file_can_poll(req->file))
5200 return IO_APOLL_ABORTED;
5201 if (req->flags & REQ_F_POLLED)
5202 return IO_APOLL_ABORTED;
5203 if (!def->pollin && !def->pollout)
5204 return IO_APOLL_ABORTED;
5208 mask |= POLLIN | POLLRDNORM;
5210 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211 if ((req->opcode == IORING_OP_RECVMSG) &&
5212 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5216 mask |= POLLOUT | POLLWRNORM;
5219 /* if we can't nonblock try, then no point in arming a poll handler */
5220 if (!io_file_supports_nowait(req, rw))
5221 return IO_APOLL_ABORTED;
5223 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5224 if (unlikely(!apoll))
5225 return IO_APOLL_ABORTED;
5226 apoll->double_poll = NULL;
5228 req->flags |= REQ_F_POLLED;
5229 ipt.pt._qproc = io_async_queue_proc;
5230 io_req_set_refcount(req);
5232 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5234 spin_unlock(&ctx->completion_lock);
5235 if (ret || ipt.error)
5236 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5238 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5239 mask, apoll->poll.events);
5243 static bool __io_poll_remove_one(struct io_kiocb *req,
5244 struct io_poll_iocb *poll, bool do_cancel)
5245 __must_hold(&req->ctx->completion_lock)
5247 bool do_complete = false;
5251 spin_lock_irq(&poll->head->lock);
5253 WRITE_ONCE(poll->canceled, true);
5254 if (!list_empty(&poll->wait.entry)) {
5255 list_del_init(&poll->wait.entry);
5258 spin_unlock_irq(&poll->head->lock);
5259 hash_del(&req->hash_node);
5263 static bool io_poll_remove_one(struct io_kiocb *req)
5264 __must_hold(&req->ctx->completion_lock)
5268 io_poll_remove_double(req);
5269 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5272 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5273 io_commit_cqring(req->ctx);
5275 io_put_req_deferred(req);
5281 * Returns true if we found and killed one or more poll requests
5283 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5286 struct hlist_node *tmp;
5287 struct io_kiocb *req;
5290 spin_lock(&ctx->completion_lock);
5291 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5292 struct hlist_head *list;
5294 list = &ctx->cancel_hash[i];
5295 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5296 if (io_match_task(req, tsk, cancel_all))
5297 posted += io_poll_remove_one(req);
5300 spin_unlock(&ctx->completion_lock);
5303 io_cqring_ev_posted(ctx);
5308 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5310 __must_hold(&ctx->completion_lock)
5312 struct hlist_head *list;
5313 struct io_kiocb *req;
5315 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5316 hlist_for_each_entry(req, list, hash_node) {
5317 if (sqe_addr != req->user_data)
5319 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5326 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5328 __must_hold(&ctx->completion_lock)
5330 struct io_kiocb *req;
5332 req = io_poll_find(ctx, sqe_addr, poll_only);
5335 if (io_poll_remove_one(req))
5341 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5346 events = READ_ONCE(sqe->poll32_events);
5348 events = swahw32(events);
5350 if (!(flags & IORING_POLL_ADD_MULTI))
5351 events |= EPOLLONESHOT;
5352 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5355 static int io_poll_update_prep(struct io_kiocb *req,
5356 const struct io_uring_sqe *sqe)
5358 struct io_poll_update *upd = &req->poll_update;
5361 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5363 if (sqe->ioprio || sqe->buf_index)
5365 flags = READ_ONCE(sqe->len);
5366 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5367 IORING_POLL_ADD_MULTI))
5369 /* meaningless without update */
5370 if (flags == IORING_POLL_ADD_MULTI)
5373 upd->old_user_data = READ_ONCE(sqe->addr);
5374 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5375 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5377 upd->new_user_data = READ_ONCE(sqe->off);
5378 if (!upd->update_user_data && upd->new_user_data)
5380 if (upd->update_events)
5381 upd->events = io_poll_parse_events(sqe, flags);
5382 else if (sqe->poll32_events)
5388 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5391 struct io_kiocb *req = wait->private;
5392 struct io_poll_iocb *poll = &req->poll;
5394 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5397 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5398 struct poll_table_struct *p)
5400 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5402 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5405 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5407 struct io_poll_iocb *poll = &req->poll;
5410 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5412 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5414 flags = READ_ONCE(sqe->len);
5415 if (flags & ~IORING_POLL_ADD_MULTI)
5418 io_req_set_refcount(req);
5419 poll->events = io_poll_parse_events(sqe, flags);
5423 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5425 struct io_poll_iocb *poll = &req->poll;
5426 struct io_ring_ctx *ctx = req->ctx;
5427 struct io_poll_table ipt;
5430 ipt.pt._qproc = io_poll_queue_proc;
5432 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5435 if (mask) { /* no async, we'd stolen it */
5437 io_poll_complete(req, mask);
5439 spin_unlock(&ctx->completion_lock);
5442 io_cqring_ev_posted(ctx);
5443 if (poll->events & EPOLLONESHOT)
5449 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5451 struct io_ring_ctx *ctx = req->ctx;
5452 struct io_kiocb *preq;
5456 spin_lock(&ctx->completion_lock);
5457 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5463 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5465 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5470 * Don't allow racy completion with singleshot, as we cannot safely
5471 * update those. For multishot, if we're racing with completion, just
5472 * let completion re-add it.
5474 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5475 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5479 /* we now have a detached poll request. reissue. */
5483 spin_unlock(&ctx->completion_lock);
5485 io_req_complete(req, ret);
5488 /* only mask one event flags, keep behavior flags */
5489 if (req->poll_update.update_events) {
5490 preq->poll.events &= ~0xffff;
5491 preq->poll.events |= req->poll_update.events & 0xffff;
5492 preq->poll.events |= IO_POLL_UNMASK;
5494 if (req->poll_update.update_user_data)
5495 preq->user_data = req->poll_update.new_user_data;
5496 spin_unlock(&ctx->completion_lock);
5498 /* complete update request, we're done with it */
5499 io_req_complete(req, ret);
5502 ret = io_poll_add(preq, issue_flags);
5505 io_req_complete(preq, ret);
5511 static void io_req_task_timeout(struct io_kiocb *req)
5513 struct io_ring_ctx *ctx = req->ctx;
5515 spin_lock(&ctx->completion_lock);
5516 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5517 io_commit_cqring(ctx);
5518 spin_unlock(&ctx->completion_lock);
5520 io_cqring_ev_posted(ctx);
5525 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5527 struct io_timeout_data *data = container_of(timer,
5528 struct io_timeout_data, timer);
5529 struct io_kiocb *req = data->req;
5530 struct io_ring_ctx *ctx = req->ctx;
5531 unsigned long flags;
5533 spin_lock_irqsave(&ctx->timeout_lock, flags);
5534 list_del_init(&req->timeout.list);
5535 atomic_set(&req->ctx->cq_timeouts,
5536 atomic_read(&req->ctx->cq_timeouts) + 1);
5537 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5539 req->io_task_work.func = io_req_task_timeout;
5540 io_req_task_work_add(req);
5541 return HRTIMER_NORESTART;
5544 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5546 __must_hold(&ctx->timeout_lock)
5548 struct io_timeout_data *io;
5549 struct io_kiocb *req;
5552 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5553 found = user_data == req->user_data;
5558 return ERR_PTR(-ENOENT);
5560 io = req->async_data;
5561 if (hrtimer_try_to_cancel(&io->timer) == -1)
5562 return ERR_PTR(-EALREADY);
5563 list_del_init(&req->timeout.list);
5567 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5568 __must_hold(&ctx->timeout_lock)
5570 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5573 return PTR_ERR(req);
5576 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5577 io_put_req_deferred(req);
5581 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5582 struct timespec64 *ts, enum hrtimer_mode mode)
5583 __must_hold(&ctx->timeout_lock)
5585 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5586 struct io_timeout_data *data;
5589 return PTR_ERR(req);
5591 req->timeout.off = 0; /* noseq */
5592 data = req->async_data;
5593 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5594 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5595 data->timer.function = io_timeout_fn;
5596 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5600 static int io_timeout_remove_prep(struct io_kiocb *req,
5601 const struct io_uring_sqe *sqe)
5603 struct io_timeout_rem *tr = &req->timeout_rem;
5605 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5607 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5609 if (sqe->ioprio || sqe->buf_index || sqe->len)
5612 tr->addr = READ_ONCE(sqe->addr);
5613 tr->flags = READ_ONCE(sqe->timeout_flags);
5614 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5615 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5617 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5619 } else if (tr->flags) {
5620 /* timeout removal doesn't support flags */
5627 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5629 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5634 * Remove or update an existing timeout command
5636 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5638 struct io_timeout_rem *tr = &req->timeout_rem;
5639 struct io_ring_ctx *ctx = req->ctx;
5642 spin_lock_irq(&ctx->timeout_lock);
5643 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5644 ret = io_timeout_cancel(ctx, tr->addr);
5646 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5647 io_translate_timeout_mode(tr->flags));
5648 spin_unlock_irq(&ctx->timeout_lock);
5650 spin_lock(&ctx->completion_lock);
5651 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5652 io_commit_cqring(ctx);
5653 spin_unlock(&ctx->completion_lock);
5654 io_cqring_ev_posted(ctx);
5661 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5662 bool is_timeout_link)
5664 struct io_timeout_data *data;
5666 u32 off = READ_ONCE(sqe->off);
5668 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5670 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5672 if (off && is_timeout_link)
5674 flags = READ_ONCE(sqe->timeout_flags);
5675 if (flags & ~IORING_TIMEOUT_ABS)
5678 req->timeout.off = off;
5679 if (unlikely(off && !req->ctx->off_timeout_used))
5680 req->ctx->off_timeout_used = true;
5682 if (!req->async_data && io_alloc_async_data(req))
5685 data = req->async_data;
5688 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5691 data->mode = io_translate_timeout_mode(flags);
5692 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5694 if (is_timeout_link) {
5695 struct io_submit_link *link = &req->ctx->submit_state.link;
5699 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5701 req->timeout.head = link->last;
5702 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5707 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5709 struct io_ring_ctx *ctx = req->ctx;
5710 struct io_timeout_data *data = req->async_data;
5711 struct list_head *entry;
5712 u32 tail, off = req->timeout.off;
5714 spin_lock_irq(&ctx->timeout_lock);
5717 * sqe->off holds how many events that need to occur for this
5718 * timeout event to be satisfied. If it isn't set, then this is
5719 * a pure timeout request, sequence isn't used.
5721 if (io_is_timeout_noseq(req)) {
5722 entry = ctx->timeout_list.prev;
5726 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5727 req->timeout.target_seq = tail + off;
5729 /* Update the last seq here in case io_flush_timeouts() hasn't.
5730 * This is safe because ->completion_lock is held, and submissions
5731 * and completions are never mixed in the same ->completion_lock section.
5733 ctx->cq_last_tm_flush = tail;
5736 * Insertion sort, ensuring the first entry in the list is always
5737 * the one we need first.
5739 list_for_each_prev(entry, &ctx->timeout_list) {
5740 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5743 if (io_is_timeout_noseq(nxt))
5745 /* nxt.seq is behind @tail, otherwise would've been completed */
5746 if (off >= nxt->timeout.target_seq - tail)
5750 list_add(&req->timeout.list, entry);
5751 data->timer.function = io_timeout_fn;
5752 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5753 spin_unlock_irq(&ctx->timeout_lock);
5757 struct io_cancel_data {
5758 struct io_ring_ctx *ctx;
5762 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5764 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5765 struct io_cancel_data *cd = data;
5767 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5770 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5771 struct io_ring_ctx *ctx)
5773 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5774 enum io_wq_cancel cancel_ret;
5777 if (!tctx || !tctx->io_wq)
5780 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5781 switch (cancel_ret) {
5782 case IO_WQ_CANCEL_OK:
5785 case IO_WQ_CANCEL_RUNNING:
5788 case IO_WQ_CANCEL_NOTFOUND:
5796 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5797 __acquires(&req->ctx->completion_lock)
5799 struct io_ring_ctx *ctx = req->ctx;
5802 WARN_ON_ONCE(req->task != current);
5804 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5805 spin_lock(&ctx->completion_lock);
5808 spin_lock_irq(&ctx->timeout_lock);
5809 ret = io_timeout_cancel(ctx, sqe_addr);
5810 spin_unlock_irq(&ctx->timeout_lock);
5813 return io_poll_cancel(ctx, sqe_addr, false);
5816 static int io_async_cancel_prep(struct io_kiocb *req,
5817 const struct io_uring_sqe *sqe)
5819 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5821 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5823 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5826 req->cancel.addr = READ_ONCE(sqe->addr);
5830 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5832 struct io_ring_ctx *ctx = req->ctx;
5833 u64 sqe_addr = req->cancel.addr;
5834 struct io_tctx_node *node;
5837 ret = io_try_cancel_userdata(req, sqe_addr);
5840 spin_unlock(&ctx->completion_lock);
5842 /* slow path, try all io-wq's */
5843 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5845 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5846 struct io_uring_task *tctx = node->task->io_uring;
5848 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5852 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5854 spin_lock(&ctx->completion_lock);
5856 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5857 io_commit_cqring(ctx);
5858 spin_unlock(&ctx->completion_lock);
5859 io_cqring_ev_posted(ctx);
5867 static int io_rsrc_update_prep(struct io_kiocb *req,
5868 const struct io_uring_sqe *sqe)
5870 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5872 if (sqe->ioprio || sqe->rw_flags)
5875 req->rsrc_update.offset = READ_ONCE(sqe->off);
5876 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5877 if (!req->rsrc_update.nr_args)
5879 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5883 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5885 struct io_ring_ctx *ctx = req->ctx;
5886 struct io_uring_rsrc_update2 up;
5889 if (issue_flags & IO_URING_F_NONBLOCK)
5892 up.offset = req->rsrc_update.offset;
5893 up.data = req->rsrc_update.arg;
5898 mutex_lock(&ctx->uring_lock);
5899 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5900 &up, req->rsrc_update.nr_args);
5901 mutex_unlock(&ctx->uring_lock);
5905 __io_req_complete(req, issue_flags, ret, 0);
5909 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5911 switch (req->opcode) {
5914 case IORING_OP_READV:
5915 case IORING_OP_READ_FIXED:
5916 case IORING_OP_READ:
5917 return io_read_prep(req, sqe);
5918 case IORING_OP_WRITEV:
5919 case IORING_OP_WRITE_FIXED:
5920 case IORING_OP_WRITE:
5921 return io_write_prep(req, sqe);
5922 case IORING_OP_POLL_ADD:
5923 return io_poll_add_prep(req, sqe);
5924 case IORING_OP_POLL_REMOVE:
5925 return io_poll_update_prep(req, sqe);
5926 case IORING_OP_FSYNC:
5927 return io_fsync_prep(req, sqe);
5928 case IORING_OP_SYNC_FILE_RANGE:
5929 return io_sfr_prep(req, sqe);
5930 case IORING_OP_SENDMSG:
5931 case IORING_OP_SEND:
5932 return io_sendmsg_prep(req, sqe);
5933 case IORING_OP_RECVMSG:
5934 case IORING_OP_RECV:
5935 return io_recvmsg_prep(req, sqe);
5936 case IORING_OP_CONNECT:
5937 return io_connect_prep(req, sqe);
5938 case IORING_OP_TIMEOUT:
5939 return io_timeout_prep(req, sqe, false);
5940 case IORING_OP_TIMEOUT_REMOVE:
5941 return io_timeout_remove_prep(req, sqe);
5942 case IORING_OP_ASYNC_CANCEL:
5943 return io_async_cancel_prep(req, sqe);
5944 case IORING_OP_LINK_TIMEOUT:
5945 return io_timeout_prep(req, sqe, true);
5946 case IORING_OP_ACCEPT:
5947 return io_accept_prep(req, sqe);
5948 case IORING_OP_FALLOCATE:
5949 return io_fallocate_prep(req, sqe);
5950 case IORING_OP_OPENAT:
5951 return io_openat_prep(req, sqe);
5952 case IORING_OP_CLOSE:
5953 return io_close_prep(req, sqe);
5954 case IORING_OP_FILES_UPDATE:
5955 return io_rsrc_update_prep(req, sqe);
5956 case IORING_OP_STATX:
5957 return io_statx_prep(req, sqe);
5958 case IORING_OP_FADVISE:
5959 return io_fadvise_prep(req, sqe);
5960 case IORING_OP_MADVISE:
5961 return io_madvise_prep(req, sqe);
5962 case IORING_OP_OPENAT2:
5963 return io_openat2_prep(req, sqe);
5964 case IORING_OP_EPOLL_CTL:
5965 return io_epoll_ctl_prep(req, sqe);
5966 case IORING_OP_SPLICE:
5967 return io_splice_prep(req, sqe);
5968 case IORING_OP_PROVIDE_BUFFERS:
5969 return io_provide_buffers_prep(req, sqe);
5970 case IORING_OP_REMOVE_BUFFERS:
5971 return io_remove_buffers_prep(req, sqe);
5973 return io_tee_prep(req, sqe);
5974 case IORING_OP_SHUTDOWN:
5975 return io_shutdown_prep(req, sqe);
5976 case IORING_OP_RENAMEAT:
5977 return io_renameat_prep(req, sqe);
5978 case IORING_OP_UNLINKAT:
5979 return io_unlinkat_prep(req, sqe);
5982 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5987 static int io_req_prep_async(struct io_kiocb *req)
5989 if (!io_op_defs[req->opcode].needs_async_setup)
5991 if (WARN_ON_ONCE(req->async_data))
5993 if (io_alloc_async_data(req))
5996 switch (req->opcode) {
5997 case IORING_OP_READV:
5998 return io_rw_prep_async(req, READ);
5999 case IORING_OP_WRITEV:
6000 return io_rw_prep_async(req, WRITE);
6001 case IORING_OP_SENDMSG:
6002 return io_sendmsg_prep_async(req);
6003 case IORING_OP_RECVMSG:
6004 return io_recvmsg_prep_async(req);
6005 case IORING_OP_CONNECT:
6006 return io_connect_prep_async(req);
6008 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6013 static u32 io_get_sequence(struct io_kiocb *req)
6015 u32 seq = req->ctx->cached_sq_head;
6017 /* need original cached_sq_head, but it was increased for each req */
6018 io_for_each_link(req, req)
6023 static bool io_drain_req(struct io_kiocb *req)
6025 struct io_kiocb *pos;
6026 struct io_ring_ctx *ctx = req->ctx;
6027 struct io_defer_entry *de;
6032 * If we need to drain a request in the middle of a link, drain the
6033 * head request and the next request/link after the current link.
6034 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6035 * maintained for every request of our link.
6037 if (ctx->drain_next) {
6038 req->flags |= REQ_F_IO_DRAIN;
6039 ctx->drain_next = false;
6041 /* not interested in head, start from the first linked */
6042 io_for_each_link(pos, req->link) {
6043 if (pos->flags & REQ_F_IO_DRAIN) {
6044 ctx->drain_next = true;
6045 req->flags |= REQ_F_IO_DRAIN;
6050 /* Still need defer if there is pending req in defer list. */
6051 if (likely(list_empty_careful(&ctx->defer_list) &&
6052 !(req->flags & REQ_F_IO_DRAIN))) {
6053 ctx->drain_active = false;
6057 seq = io_get_sequence(req);
6058 /* Still a chance to pass the sequence check */
6059 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6062 ret = io_req_prep_async(req);
6065 io_prep_async_link(req);
6066 de = kmalloc(sizeof(*de), GFP_KERNEL);
6070 io_req_complete_failed(req, ret);
6074 spin_lock(&ctx->completion_lock);
6075 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6076 spin_unlock(&ctx->completion_lock);
6078 io_queue_async_work(req);
6082 trace_io_uring_defer(ctx, req, req->user_data);
6085 list_add_tail(&de->list, &ctx->defer_list);
6086 spin_unlock(&ctx->completion_lock);
6090 static void io_clean_op(struct io_kiocb *req)
6092 if (req->flags & REQ_F_BUFFER_SELECTED) {
6093 switch (req->opcode) {
6094 case IORING_OP_READV:
6095 case IORING_OP_READ_FIXED:
6096 case IORING_OP_READ:
6097 kfree((void *)(unsigned long)req->rw.addr);
6099 case IORING_OP_RECVMSG:
6100 case IORING_OP_RECV:
6101 kfree(req->sr_msg.kbuf);
6106 if (req->flags & REQ_F_NEED_CLEANUP) {
6107 switch (req->opcode) {
6108 case IORING_OP_READV:
6109 case IORING_OP_READ_FIXED:
6110 case IORING_OP_READ:
6111 case IORING_OP_WRITEV:
6112 case IORING_OP_WRITE_FIXED:
6113 case IORING_OP_WRITE: {
6114 struct io_async_rw *io = req->async_data;
6116 kfree(io->free_iovec);
6119 case IORING_OP_RECVMSG:
6120 case IORING_OP_SENDMSG: {
6121 struct io_async_msghdr *io = req->async_data;
6123 kfree(io->free_iov);
6126 case IORING_OP_SPLICE:
6128 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6129 io_put_file(req->splice.file_in);
6131 case IORING_OP_OPENAT:
6132 case IORING_OP_OPENAT2:
6133 if (req->open.filename)
6134 putname(req->open.filename);
6136 case IORING_OP_RENAMEAT:
6137 putname(req->rename.oldpath);
6138 putname(req->rename.newpath);
6140 case IORING_OP_UNLINKAT:
6141 putname(req->unlink.filename);
6145 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6146 kfree(req->apoll->double_poll);
6150 if (req->flags & REQ_F_INFLIGHT) {
6151 struct io_uring_task *tctx = req->task->io_uring;
6153 atomic_dec(&tctx->inflight_tracked);
6155 if (req->flags & REQ_F_CREDS)
6156 put_cred(req->creds);
6158 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6161 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6163 struct io_ring_ctx *ctx = req->ctx;
6164 const struct cred *creds = NULL;
6167 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6168 creds = override_creds(req->creds);
6170 switch (req->opcode) {
6172 ret = io_nop(req, issue_flags);
6174 case IORING_OP_READV:
6175 case IORING_OP_READ_FIXED:
6176 case IORING_OP_READ:
6177 ret = io_read(req, issue_flags);
6179 case IORING_OP_WRITEV:
6180 case IORING_OP_WRITE_FIXED:
6181 case IORING_OP_WRITE:
6182 ret = io_write(req, issue_flags);
6184 case IORING_OP_FSYNC:
6185 ret = io_fsync(req, issue_flags);
6187 case IORING_OP_POLL_ADD:
6188 ret = io_poll_add(req, issue_flags);
6190 case IORING_OP_POLL_REMOVE:
6191 ret = io_poll_update(req, issue_flags);
6193 case IORING_OP_SYNC_FILE_RANGE:
6194 ret = io_sync_file_range(req, issue_flags);
6196 case IORING_OP_SENDMSG:
6197 ret = io_sendmsg(req, issue_flags);
6199 case IORING_OP_SEND:
6200 ret = io_send(req, issue_flags);
6202 case IORING_OP_RECVMSG:
6203 ret = io_recvmsg(req, issue_flags);
6205 case IORING_OP_RECV:
6206 ret = io_recv(req, issue_flags);
6208 case IORING_OP_TIMEOUT:
6209 ret = io_timeout(req, issue_flags);
6211 case IORING_OP_TIMEOUT_REMOVE:
6212 ret = io_timeout_remove(req, issue_flags);
6214 case IORING_OP_ACCEPT:
6215 ret = io_accept(req, issue_flags);
6217 case IORING_OP_CONNECT:
6218 ret = io_connect(req, issue_flags);
6220 case IORING_OP_ASYNC_CANCEL:
6221 ret = io_async_cancel(req, issue_flags);
6223 case IORING_OP_FALLOCATE:
6224 ret = io_fallocate(req, issue_flags);
6226 case IORING_OP_OPENAT:
6227 ret = io_openat(req, issue_flags);
6229 case IORING_OP_CLOSE:
6230 ret = io_close(req, issue_flags);
6232 case IORING_OP_FILES_UPDATE:
6233 ret = io_files_update(req, issue_flags);
6235 case IORING_OP_STATX:
6236 ret = io_statx(req, issue_flags);
6238 case IORING_OP_FADVISE:
6239 ret = io_fadvise(req, issue_flags);
6241 case IORING_OP_MADVISE:
6242 ret = io_madvise(req, issue_flags);
6244 case IORING_OP_OPENAT2:
6245 ret = io_openat2(req, issue_flags);
6247 case IORING_OP_EPOLL_CTL:
6248 ret = io_epoll_ctl(req, issue_flags);
6250 case IORING_OP_SPLICE:
6251 ret = io_splice(req, issue_flags);
6253 case IORING_OP_PROVIDE_BUFFERS:
6254 ret = io_provide_buffers(req, issue_flags);
6256 case IORING_OP_REMOVE_BUFFERS:
6257 ret = io_remove_buffers(req, issue_flags);
6260 ret = io_tee(req, issue_flags);
6262 case IORING_OP_SHUTDOWN:
6263 ret = io_shutdown(req, issue_flags);
6265 case IORING_OP_RENAMEAT:
6266 ret = io_renameat(req, issue_flags);
6268 case IORING_OP_UNLINKAT:
6269 ret = io_unlinkat(req, issue_flags);
6277 revert_creds(creds);
6280 /* If the op doesn't have a file, we're not polling for it */
6281 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6282 io_iopoll_req_issued(req);
6287 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6289 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6291 req = io_put_req_find_next(req);
6292 return req ? &req->work : NULL;
6295 static void io_wq_submit_work(struct io_wq_work *work)
6297 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6298 struct io_kiocb *timeout;
6301 /* one will be dropped by ->io_free_work() after returning to io-wq */
6302 if (!(req->flags & REQ_F_REFCOUNT))
6303 __io_req_set_refcount(req, 2);
6307 timeout = io_prep_linked_timeout(req);
6309 io_queue_linked_timeout(timeout);
6311 if (work->flags & IO_WQ_WORK_CANCEL)
6316 ret = io_issue_sqe(req, 0);
6318 * We can get EAGAIN for polled IO even though we're
6319 * forcing a sync submission from here, since we can't
6320 * wait for request slots on the block side.
6328 /* avoid locking problems by failing it from a clean context */
6330 io_req_task_queue_fail(req, ret);
6333 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6336 return &table->files[i];
6339 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6342 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6344 return (struct file *) (slot->file_ptr & FFS_MASK);
6347 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6349 unsigned long file_ptr = (unsigned long) file;
6351 if (__io_file_supports_nowait(file, READ))
6352 file_ptr |= FFS_ASYNC_READ;
6353 if (__io_file_supports_nowait(file, WRITE))
6354 file_ptr |= FFS_ASYNC_WRITE;
6355 if (S_ISREG(file_inode(file)->i_mode))
6356 file_ptr |= FFS_ISREG;
6357 file_slot->file_ptr = file_ptr;
6360 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6361 struct io_kiocb *req, int fd)
6364 unsigned long file_ptr;
6366 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6368 fd = array_index_nospec(fd, ctx->nr_user_files);
6369 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6370 file = (struct file *) (file_ptr & FFS_MASK);
6371 file_ptr &= ~FFS_MASK;
6372 /* mask in overlapping REQ_F and FFS bits */
6373 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6374 io_req_set_rsrc_node(req);
6378 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6379 struct io_kiocb *req, int fd)
6381 struct file *file = fget(fd);
6383 trace_io_uring_file_get(ctx, fd);
6385 /* we don't allow fixed io_uring files */
6386 if (file && unlikely(file->f_op == &io_uring_fops))
6387 io_req_track_inflight(req);
6391 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6392 struct io_kiocb *req, int fd, bool fixed)
6395 return io_file_get_fixed(ctx, req, fd);
6397 return io_file_get_normal(ctx, req, fd);
6400 static void io_req_task_link_timeout(struct io_kiocb *req)
6402 struct io_kiocb *prev = req->timeout.prev;
6403 struct io_ring_ctx *ctx = req->ctx;
6407 ret = io_try_cancel_userdata(req, prev->user_data);
6410 io_cqring_fill_event(ctx, req->user_data, ret, 0);
6411 io_commit_cqring(ctx);
6412 spin_unlock(&ctx->completion_lock);
6413 io_cqring_ev_posted(ctx);
6418 io_req_complete_post(req, -ETIME, 0);
6422 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6424 struct io_timeout_data *data = container_of(timer,
6425 struct io_timeout_data, timer);
6426 struct io_kiocb *prev, *req = data->req;
6427 struct io_ring_ctx *ctx = req->ctx;
6428 unsigned long flags;
6430 spin_lock_irqsave(&ctx->timeout_lock, flags);
6431 prev = req->timeout.head;
6432 req->timeout.head = NULL;
6435 * We don't expect the list to be empty, that will only happen if we
6436 * race with the completion of the linked work.
6439 io_remove_next_linked(prev);
6440 if (!req_ref_inc_not_zero(prev))
6443 req->timeout.prev = prev;
6444 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6446 req->io_task_work.func = io_req_task_link_timeout;
6447 io_req_task_work_add(req);
6448 return HRTIMER_NORESTART;
6451 static void io_queue_linked_timeout(struct io_kiocb *req)
6453 struct io_ring_ctx *ctx = req->ctx;
6455 spin_lock_irq(&ctx->timeout_lock);
6457 * If the back reference is NULL, then our linked request finished
6458 * before we got a chance to setup the timer
6460 if (req->timeout.head) {
6461 struct io_timeout_data *data = req->async_data;
6463 data->timer.function = io_link_timeout_fn;
6464 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6467 spin_unlock_irq(&ctx->timeout_lock);
6468 /* drop submission reference */
6472 static void __io_queue_sqe(struct io_kiocb *req)
6473 __must_hold(&req->ctx->uring_lock)
6475 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6479 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6482 * We async punt it if the file wasn't marked NOWAIT, or if the file
6483 * doesn't support non-blocking read/write attempts
6486 if (req->flags & REQ_F_COMPLETE_INLINE) {
6487 struct io_ring_ctx *ctx = req->ctx;
6488 struct io_submit_state *state = &ctx->submit_state;
6490 state->compl_reqs[state->compl_nr++] = req;
6491 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6492 io_submit_flush_completions(ctx);
6494 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6495 switch (io_arm_poll_handler(req)) {
6496 case IO_APOLL_READY:
6498 case IO_APOLL_ABORTED:
6500 * Queued up for async execution, worker will release
6501 * submit reference when the iocb is actually submitted.
6503 io_queue_async_work(req);
6507 io_req_complete_failed(req, ret);
6510 io_queue_linked_timeout(linked_timeout);
6513 static inline void io_queue_sqe(struct io_kiocb *req)
6514 __must_hold(&req->ctx->uring_lock)
6516 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6519 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6520 __io_queue_sqe(req);
6522 int ret = io_req_prep_async(req);
6525 io_req_complete_failed(req, ret);
6527 io_queue_async_work(req);
6532 * Check SQE restrictions (opcode and flags).
6534 * Returns 'true' if SQE is allowed, 'false' otherwise.
6536 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6537 struct io_kiocb *req,
6538 unsigned int sqe_flags)
6540 if (likely(!ctx->restricted))
6543 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6546 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6547 ctx->restrictions.sqe_flags_required)
6550 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6551 ctx->restrictions.sqe_flags_required))
6557 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6558 const struct io_uring_sqe *sqe)
6559 __must_hold(&ctx->uring_lock)
6561 struct io_submit_state *state;
6562 unsigned int sqe_flags;
6563 int personality, ret = 0;
6565 /* req is partially pre-initialised, see io_preinit_req() */
6566 req->opcode = READ_ONCE(sqe->opcode);
6567 /* same numerical values with corresponding REQ_F_*, safe to copy */
6568 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6569 req->user_data = READ_ONCE(sqe->user_data);
6571 req->fixed_rsrc_refs = NULL;
6572 req->task = current;
6574 /* enforce forwards compatibility on users */
6575 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6577 if (unlikely(req->opcode >= IORING_OP_LAST))
6579 if (!io_check_restriction(ctx, req, sqe_flags))
6582 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6583 !io_op_defs[req->opcode].buffer_select)
6585 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6586 ctx->drain_active = true;
6588 personality = READ_ONCE(sqe->personality);
6590 req->creds = xa_load(&ctx->personalities, personality);
6593 get_cred(req->creds);
6594 req->flags |= REQ_F_CREDS;
6596 state = &ctx->submit_state;
6599 * Plug now if we have more than 1 IO left after this, and the target
6600 * is potentially a read/write to block based storage.
6602 if (!state->plug_started && state->ios_left > 1 &&
6603 io_op_defs[req->opcode].plug) {
6604 blk_start_plug(&state->plug);
6605 state->plug_started = true;
6608 if (io_op_defs[req->opcode].needs_file) {
6609 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6610 (sqe_flags & IOSQE_FIXED_FILE));
6611 if (unlikely(!req->file))
6619 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6620 const struct io_uring_sqe *sqe)
6621 __must_hold(&ctx->uring_lock)
6623 struct io_submit_link *link = &ctx->submit_state.link;
6626 ret = io_init_req(ctx, req, sqe);
6627 if (unlikely(ret)) {
6630 /* fail even hard links since we don't submit */
6631 req_set_fail(link->head);
6632 io_req_complete_failed(link->head, -ECANCELED);
6635 io_req_complete_failed(req, ret);
6639 ret = io_req_prep(req, sqe);
6643 /* don't need @sqe from now on */
6644 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6646 ctx->flags & IORING_SETUP_SQPOLL);
6649 * If we already have a head request, queue this one for async
6650 * submittal once the head completes. If we don't have a head but
6651 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6652 * submitted sync once the chain is complete. If none of those
6653 * conditions are true (normal request), then just queue it.
6656 struct io_kiocb *head = link->head;
6658 ret = io_req_prep_async(req);
6661 trace_io_uring_link(ctx, req, head);
6662 link->last->link = req;
6665 /* last request of a link, enqueue the link */
6666 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6671 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6683 * Batched submission is done, ensure local IO is flushed out.
6685 static void io_submit_state_end(struct io_submit_state *state,
6686 struct io_ring_ctx *ctx)
6688 if (state->link.head)
6689 io_queue_sqe(state->link.head);
6690 if (state->compl_nr)
6691 io_submit_flush_completions(ctx);
6692 if (state->plug_started)
6693 blk_finish_plug(&state->plug);
6697 * Start submission side cache.
6699 static void io_submit_state_start(struct io_submit_state *state,
6700 unsigned int max_ios)
6702 state->plug_started = false;
6703 state->ios_left = max_ios;
6704 /* set only head, no need to init link_last in advance */
6705 state->link.head = NULL;
6708 static void io_commit_sqring(struct io_ring_ctx *ctx)
6710 struct io_rings *rings = ctx->rings;
6713 * Ensure any loads from the SQEs are done at this point,
6714 * since once we write the new head, the application could
6715 * write new data to them.
6717 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6721 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6722 * that is mapped by userspace. This means that care needs to be taken to
6723 * ensure that reads are stable, as we cannot rely on userspace always
6724 * being a good citizen. If members of the sqe are validated and then later
6725 * used, it's important that those reads are done through READ_ONCE() to
6726 * prevent a re-load down the line.
6728 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6730 unsigned head, mask = ctx->sq_entries - 1;
6731 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6734 * The cached sq head (or cq tail) serves two purposes:
6736 * 1) allows us to batch the cost of updating the user visible
6738 * 2) allows the kernel side to track the head on its own, even
6739 * though the application is the one updating it.
6741 head = READ_ONCE(ctx->sq_array[sq_idx]);
6742 if (likely(head < ctx->sq_entries))
6743 return &ctx->sq_sqes[head];
6745 /* drop invalid entries */
6747 WRITE_ONCE(ctx->rings->sq_dropped,
6748 READ_ONCE(ctx->rings->sq_dropped) + 1);
6752 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6753 __must_hold(&ctx->uring_lock)
6755 struct io_uring_task *tctx;
6758 /* make sure SQ entry isn't read before tail */
6759 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6760 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6763 tctx = current->io_uring;
6764 tctx->cached_refs -= nr;
6765 if (unlikely(tctx->cached_refs < 0)) {
6766 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6768 percpu_counter_add(&tctx->inflight, refill);
6769 refcount_add(refill, ¤t->usage);
6770 tctx->cached_refs += refill;
6772 io_submit_state_start(&ctx->submit_state, nr);
6774 while (submitted < nr) {
6775 const struct io_uring_sqe *sqe;
6776 struct io_kiocb *req;
6778 req = io_alloc_req(ctx);
6779 if (unlikely(!req)) {
6781 submitted = -EAGAIN;
6784 sqe = io_get_sqe(ctx);
6785 if (unlikely(!sqe)) {
6786 kmem_cache_free(req_cachep, req);
6789 /* will complete beyond this point, count as submitted */
6791 if (io_submit_sqe(ctx, req, sqe))
6795 if (unlikely(submitted != nr)) {
6796 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6797 int unused = nr - ref_used;
6799 current->io_uring->cached_refs += unused;
6800 percpu_ref_put_many(&ctx->refs, unused);
6803 io_submit_state_end(&ctx->submit_state, ctx);
6804 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6805 io_commit_sqring(ctx);
6810 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6812 return READ_ONCE(sqd->state);
6815 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6817 /* Tell userspace we may need a wakeup call */
6818 spin_lock(&ctx->completion_lock);
6819 WRITE_ONCE(ctx->rings->sq_flags,
6820 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6821 spin_unlock(&ctx->completion_lock);
6824 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6826 spin_lock(&ctx->completion_lock);
6827 WRITE_ONCE(ctx->rings->sq_flags,
6828 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6829 spin_unlock(&ctx->completion_lock);
6832 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6834 unsigned int to_submit;
6837 to_submit = io_sqring_entries(ctx);
6838 /* if we're handling multiple rings, cap submit size for fairness */
6839 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6840 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6842 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6843 unsigned nr_events = 0;
6844 const struct cred *creds = NULL;
6846 if (ctx->sq_creds != current_cred())
6847 creds = override_creds(ctx->sq_creds);
6849 mutex_lock(&ctx->uring_lock);
6850 if (!list_empty(&ctx->iopoll_list))
6851 io_do_iopoll(ctx, &nr_events, 0);
6854 * Don't submit if refs are dying, good for io_uring_register(),
6855 * but also it is relied upon by io_ring_exit_work()
6857 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6858 !(ctx->flags & IORING_SETUP_R_DISABLED))
6859 ret = io_submit_sqes(ctx, to_submit);
6860 mutex_unlock(&ctx->uring_lock);
6862 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6863 wake_up(&ctx->sqo_sq_wait);
6865 revert_creds(creds);
6871 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6873 struct io_ring_ctx *ctx;
6874 unsigned sq_thread_idle = 0;
6876 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6877 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6878 sqd->sq_thread_idle = sq_thread_idle;
6881 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6883 bool did_sig = false;
6884 struct ksignal ksig;
6886 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6887 signal_pending(current)) {
6888 mutex_unlock(&sqd->lock);
6889 if (signal_pending(current))
6890 did_sig = get_signal(&ksig);
6892 mutex_lock(&sqd->lock);
6894 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6897 static int io_sq_thread(void *data)
6899 struct io_sq_data *sqd = data;
6900 struct io_ring_ctx *ctx;
6901 unsigned long timeout = 0;
6902 char buf[TASK_COMM_LEN];
6905 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6906 set_task_comm(current, buf);
6908 if (sqd->sq_cpu != -1)
6909 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6911 set_cpus_allowed_ptr(current, cpu_online_mask);
6912 current->flags |= PF_NO_SETAFFINITY;
6914 mutex_lock(&sqd->lock);
6916 bool cap_entries, sqt_spin = false;
6918 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6919 if (io_sqd_handle_event(sqd))
6921 timeout = jiffies + sqd->sq_thread_idle;
6924 cap_entries = !list_is_singular(&sqd->ctx_list);
6925 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6926 int ret = __io_sq_thread(ctx, cap_entries);
6928 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6931 if (io_run_task_work())
6934 if (sqt_spin || !time_after(jiffies, timeout)) {
6937 timeout = jiffies + sqd->sq_thread_idle;
6941 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6942 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6943 bool needs_sched = true;
6945 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6946 io_ring_set_wakeup_flag(ctx);
6948 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6949 !list_empty_careful(&ctx->iopoll_list)) {
6950 needs_sched = false;
6953 if (io_sqring_entries(ctx)) {
6954 needs_sched = false;
6960 mutex_unlock(&sqd->lock);
6962 mutex_lock(&sqd->lock);
6964 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6965 io_ring_clear_wakeup_flag(ctx);
6968 finish_wait(&sqd->wait, &wait);
6969 timeout = jiffies + sqd->sq_thread_idle;
6972 io_uring_cancel_generic(true, sqd);
6974 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6975 io_ring_set_wakeup_flag(ctx);
6977 mutex_unlock(&sqd->lock);
6979 complete(&sqd->exited);
6983 struct io_wait_queue {
6984 struct wait_queue_entry wq;
6985 struct io_ring_ctx *ctx;
6987 unsigned nr_timeouts;
6990 static inline bool io_should_wake(struct io_wait_queue *iowq)
6992 struct io_ring_ctx *ctx = iowq->ctx;
6993 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
6996 * Wake up if we have enough events, or if a timeout occurred since we
6997 * started waiting. For timeouts, we always want to return to userspace,
6998 * regardless of event count.
7000 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7003 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7004 int wake_flags, void *key)
7006 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7010 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7011 * the task, and the next invocation will do it.
7013 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7014 return autoremove_wake_function(curr, mode, wake_flags, key);
7018 static int io_run_task_work_sig(void)
7020 if (io_run_task_work())
7022 if (!signal_pending(current))
7024 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7025 return -ERESTARTSYS;
7029 /* when returns >0, the caller should retry */
7030 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7031 struct io_wait_queue *iowq,
7032 signed long *timeout)
7036 /* make sure we run task_work before checking for signals */
7037 ret = io_run_task_work_sig();
7038 if (ret || io_should_wake(iowq))
7040 /* let the caller flush overflows, retry */
7041 if (test_bit(0, &ctx->check_cq_overflow))
7044 *timeout = schedule_timeout(*timeout);
7045 return !*timeout ? -ETIME : 1;
7049 * Wait until events become available, if we don't already have some. The
7050 * application must reap them itself, as they reside on the shared cq ring.
7052 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7053 const sigset_t __user *sig, size_t sigsz,
7054 struct __kernel_timespec __user *uts)
7056 struct io_wait_queue iowq;
7057 struct io_rings *rings = ctx->rings;
7058 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7062 io_cqring_overflow_flush(ctx);
7063 if (io_cqring_events(ctx) >= min_events)
7065 if (!io_run_task_work())
7070 #ifdef CONFIG_COMPAT
7071 if (in_compat_syscall())
7072 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7076 ret = set_user_sigmask(sig, sigsz);
7083 struct timespec64 ts;
7085 if (get_timespec64(&ts, uts))
7087 timeout = timespec64_to_jiffies(&ts);
7090 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7091 iowq.wq.private = current;
7092 INIT_LIST_HEAD(&iowq.wq.entry);
7094 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7095 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7097 trace_io_uring_cqring_wait(ctx, min_events);
7099 /* if we can't even flush overflow, don't wait for more */
7100 if (!io_cqring_overflow_flush(ctx)) {
7104 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7105 TASK_INTERRUPTIBLE);
7106 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7107 finish_wait(&ctx->cq_wait, &iowq.wq);
7111 restore_saved_sigmask_unless(ret == -EINTR);
7113 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7116 static void io_free_page_table(void **table, size_t size)
7118 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7120 for (i = 0; i < nr_tables; i++)
7125 static void **io_alloc_page_table(size_t size)
7127 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7128 size_t init_size = size;
7131 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7135 for (i = 0; i < nr_tables; i++) {
7136 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7138 table[i] = kzalloc(this_size, GFP_KERNEL);
7140 io_free_page_table(table, init_size);
7148 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7150 percpu_ref_exit(&ref_node->refs);
7154 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7156 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7157 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7158 unsigned long flags;
7159 bool first_add = false;
7161 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7164 while (!list_empty(&ctx->rsrc_ref_list)) {
7165 node = list_first_entry(&ctx->rsrc_ref_list,
7166 struct io_rsrc_node, node);
7167 /* recycle ref nodes in order */
7170 list_del(&node->node);
7171 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7173 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7176 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7179 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7181 struct io_rsrc_node *ref_node;
7183 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7187 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7192 INIT_LIST_HEAD(&ref_node->node);
7193 INIT_LIST_HEAD(&ref_node->rsrc_list);
7194 ref_node->done = false;
7198 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7199 struct io_rsrc_data *data_to_kill)
7201 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7202 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7205 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7207 rsrc_node->rsrc_data = data_to_kill;
7208 spin_lock_irq(&ctx->rsrc_ref_lock);
7209 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7210 spin_unlock_irq(&ctx->rsrc_ref_lock);
7212 atomic_inc(&data_to_kill->refs);
7213 percpu_ref_kill(&rsrc_node->refs);
7214 ctx->rsrc_node = NULL;
7217 if (!ctx->rsrc_node) {
7218 ctx->rsrc_node = ctx->rsrc_backup_node;
7219 ctx->rsrc_backup_node = NULL;
7223 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7225 if (ctx->rsrc_backup_node)
7227 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7228 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7231 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7235 /* As we may drop ->uring_lock, other task may have started quiesce */
7239 data->quiesce = true;
7241 ret = io_rsrc_node_switch_start(ctx);
7244 io_rsrc_node_switch(ctx, data);
7246 /* kill initial ref, already quiesced if zero */
7247 if (atomic_dec_and_test(&data->refs))
7249 mutex_unlock(&ctx->uring_lock);
7250 flush_delayed_work(&ctx->rsrc_put_work);
7251 ret = wait_for_completion_interruptible(&data->done);
7253 mutex_lock(&ctx->uring_lock);
7257 atomic_inc(&data->refs);
7258 /* wait for all works potentially completing data->done */
7259 flush_delayed_work(&ctx->rsrc_put_work);
7260 reinit_completion(&data->done);
7262 ret = io_run_task_work_sig();
7263 mutex_lock(&ctx->uring_lock);
7265 data->quiesce = false;
7270 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7272 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7273 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7275 return &data->tags[table_idx][off];
7278 static void io_rsrc_data_free(struct io_rsrc_data *data)
7280 size_t size = data->nr * sizeof(data->tags[0][0]);
7283 io_free_page_table((void **)data->tags, size);
7287 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7288 u64 __user *utags, unsigned nr,
7289 struct io_rsrc_data **pdata)
7291 struct io_rsrc_data *data;
7295 data = kzalloc(sizeof(*data), GFP_KERNEL);
7298 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7306 data->do_put = do_put;
7309 for (i = 0; i < nr; i++) {
7310 u64 *tag_slot = io_get_tag_slot(data, i);
7312 if (copy_from_user(tag_slot, &utags[i],
7318 atomic_set(&data->refs, 1);
7319 init_completion(&data->done);
7323 io_rsrc_data_free(data);
7327 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7329 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7330 return !!table->files;
7333 static void io_free_file_tables(struct io_file_table *table)
7335 kvfree(table->files);
7336 table->files = NULL;
7339 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7341 #if defined(CONFIG_UNIX)
7342 if (ctx->ring_sock) {
7343 struct sock *sock = ctx->ring_sock->sk;
7344 struct sk_buff *skb;
7346 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7352 for (i = 0; i < ctx->nr_user_files; i++) {
7355 file = io_file_from_index(ctx, i);
7360 io_free_file_tables(&ctx->file_table);
7361 io_rsrc_data_free(ctx->file_data);
7362 ctx->file_data = NULL;
7363 ctx->nr_user_files = 0;
7366 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7370 if (!ctx->file_data)
7372 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7374 __io_sqe_files_unregister(ctx);
7378 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7379 __releases(&sqd->lock)
7381 WARN_ON_ONCE(sqd->thread == current);
7384 * Do the dance but not conditional clear_bit() because it'd race with
7385 * other threads incrementing park_pending and setting the bit.
7387 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7388 if (atomic_dec_return(&sqd->park_pending))
7389 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7390 mutex_unlock(&sqd->lock);
7393 static void io_sq_thread_park(struct io_sq_data *sqd)
7394 __acquires(&sqd->lock)
7396 WARN_ON_ONCE(sqd->thread == current);
7398 atomic_inc(&sqd->park_pending);
7399 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7400 mutex_lock(&sqd->lock);
7402 wake_up_process(sqd->thread);
7405 static void io_sq_thread_stop(struct io_sq_data *sqd)
7407 WARN_ON_ONCE(sqd->thread == current);
7408 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7410 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7411 mutex_lock(&sqd->lock);
7413 wake_up_process(sqd->thread);
7414 mutex_unlock(&sqd->lock);
7415 wait_for_completion(&sqd->exited);
7418 static void io_put_sq_data(struct io_sq_data *sqd)
7420 if (refcount_dec_and_test(&sqd->refs)) {
7421 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7423 io_sq_thread_stop(sqd);
7428 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7430 struct io_sq_data *sqd = ctx->sq_data;
7433 io_sq_thread_park(sqd);
7434 list_del_init(&ctx->sqd_list);
7435 io_sqd_update_thread_idle(sqd);
7436 io_sq_thread_unpark(sqd);
7438 io_put_sq_data(sqd);
7439 ctx->sq_data = NULL;
7443 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7445 struct io_ring_ctx *ctx_attach;
7446 struct io_sq_data *sqd;
7449 f = fdget(p->wq_fd);
7451 return ERR_PTR(-ENXIO);
7452 if (f.file->f_op != &io_uring_fops) {
7454 return ERR_PTR(-EINVAL);
7457 ctx_attach = f.file->private_data;
7458 sqd = ctx_attach->sq_data;
7461 return ERR_PTR(-EINVAL);
7463 if (sqd->task_tgid != current->tgid) {
7465 return ERR_PTR(-EPERM);
7468 refcount_inc(&sqd->refs);
7473 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7476 struct io_sq_data *sqd;
7479 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7480 sqd = io_attach_sq_data(p);
7485 /* fall through for EPERM case, setup new sqd/task */
7486 if (PTR_ERR(sqd) != -EPERM)
7490 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7492 return ERR_PTR(-ENOMEM);
7494 atomic_set(&sqd->park_pending, 0);
7495 refcount_set(&sqd->refs, 1);
7496 INIT_LIST_HEAD(&sqd->ctx_list);
7497 mutex_init(&sqd->lock);
7498 init_waitqueue_head(&sqd->wait);
7499 init_completion(&sqd->exited);
7503 #if defined(CONFIG_UNIX)
7505 * Ensure the UNIX gc is aware of our file set, so we are certain that
7506 * the io_uring can be safely unregistered on process exit, even if we have
7507 * loops in the file referencing.
7509 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7511 struct sock *sk = ctx->ring_sock->sk;
7512 struct scm_fp_list *fpl;
7513 struct sk_buff *skb;
7516 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7520 skb = alloc_skb(0, GFP_KERNEL);
7529 fpl->user = get_uid(current_user());
7530 for (i = 0; i < nr; i++) {
7531 struct file *file = io_file_from_index(ctx, i + offset);
7535 fpl->fp[nr_files] = get_file(file);
7536 unix_inflight(fpl->user, fpl->fp[nr_files]);
7541 fpl->max = SCM_MAX_FD;
7542 fpl->count = nr_files;
7543 UNIXCB(skb).fp = fpl;
7544 skb->destructor = unix_destruct_scm;
7545 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7546 skb_queue_head(&sk->sk_receive_queue, skb);
7548 for (i = 0; i < nr_files; i++)
7559 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7560 * causes regular reference counting to break down. We rely on the UNIX
7561 * garbage collection to take care of this problem for us.
7563 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7565 unsigned left, total;
7569 left = ctx->nr_user_files;
7571 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7573 ret = __io_sqe_files_scm(ctx, this_files, total);
7577 total += this_files;
7583 while (total < ctx->nr_user_files) {
7584 struct file *file = io_file_from_index(ctx, total);
7594 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7600 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7602 struct file *file = prsrc->file;
7603 #if defined(CONFIG_UNIX)
7604 struct sock *sock = ctx->ring_sock->sk;
7605 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7606 struct sk_buff *skb;
7609 __skb_queue_head_init(&list);
7612 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7613 * remove this entry and rearrange the file array.
7615 skb = skb_dequeue(head);
7617 struct scm_fp_list *fp;
7619 fp = UNIXCB(skb).fp;
7620 for (i = 0; i < fp->count; i++) {
7623 if (fp->fp[i] != file)
7626 unix_notinflight(fp->user, fp->fp[i]);
7627 left = fp->count - 1 - i;
7629 memmove(&fp->fp[i], &fp->fp[i + 1],
7630 left * sizeof(struct file *));
7637 __skb_queue_tail(&list, skb);
7647 __skb_queue_tail(&list, skb);
7649 skb = skb_dequeue(head);
7652 if (skb_peek(&list)) {
7653 spin_lock_irq(&head->lock);
7654 while ((skb = __skb_dequeue(&list)) != NULL)
7655 __skb_queue_tail(head, skb);
7656 spin_unlock_irq(&head->lock);
7663 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7665 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7666 struct io_ring_ctx *ctx = rsrc_data->ctx;
7667 struct io_rsrc_put *prsrc, *tmp;
7669 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7670 list_del(&prsrc->list);
7673 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7675 io_ring_submit_lock(ctx, lock_ring);
7676 spin_lock(&ctx->completion_lock);
7677 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7679 io_commit_cqring(ctx);
7680 spin_unlock(&ctx->completion_lock);
7681 io_cqring_ev_posted(ctx);
7682 io_ring_submit_unlock(ctx, lock_ring);
7685 rsrc_data->do_put(ctx, prsrc);
7689 io_rsrc_node_destroy(ref_node);
7690 if (atomic_dec_and_test(&rsrc_data->refs))
7691 complete(&rsrc_data->done);
7694 static void io_rsrc_put_work(struct work_struct *work)
7696 struct io_ring_ctx *ctx;
7697 struct llist_node *node;
7699 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7700 node = llist_del_all(&ctx->rsrc_put_llist);
7703 struct io_rsrc_node *ref_node;
7704 struct llist_node *next = node->next;
7706 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7707 __io_rsrc_put_work(ref_node);
7712 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7713 unsigned nr_args, u64 __user *tags)
7715 __s32 __user *fds = (__s32 __user *) arg;
7724 if (nr_args > IORING_MAX_FIXED_FILES)
7726 ret = io_rsrc_node_switch_start(ctx);
7729 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7735 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7738 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7739 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7743 /* allow sparse sets */
7746 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7753 if (unlikely(!file))
7757 * Don't allow io_uring instances to be registered. If UNIX
7758 * isn't enabled, then this causes a reference cycle and this
7759 * instance can never get freed. If UNIX is enabled we'll
7760 * handle it just fine, but there's still no point in allowing
7761 * a ring fd as it doesn't support regular read/write anyway.
7763 if (file->f_op == &io_uring_fops) {
7767 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7770 ret = io_sqe_files_scm(ctx);
7772 __io_sqe_files_unregister(ctx);
7776 io_rsrc_node_switch(ctx, NULL);
7779 for (i = 0; i < ctx->nr_user_files; i++) {
7780 file = io_file_from_index(ctx, i);
7784 io_free_file_tables(&ctx->file_table);
7785 ctx->nr_user_files = 0;
7787 io_rsrc_data_free(ctx->file_data);
7788 ctx->file_data = NULL;
7792 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7795 #if defined(CONFIG_UNIX)
7796 struct sock *sock = ctx->ring_sock->sk;
7797 struct sk_buff_head *head = &sock->sk_receive_queue;
7798 struct sk_buff *skb;
7801 * See if we can merge this file into an existing skb SCM_RIGHTS
7802 * file set. If there's no room, fall back to allocating a new skb
7803 * and filling it in.
7805 spin_lock_irq(&head->lock);
7806 skb = skb_peek(head);
7808 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7810 if (fpl->count < SCM_MAX_FD) {
7811 __skb_unlink(skb, head);
7812 spin_unlock_irq(&head->lock);
7813 fpl->fp[fpl->count] = get_file(file);
7814 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7816 spin_lock_irq(&head->lock);
7817 __skb_queue_head(head, skb);
7822 spin_unlock_irq(&head->lock);
7829 return __io_sqe_files_scm(ctx, 1, index);
7835 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7836 struct io_rsrc_node *node, void *rsrc)
7838 struct io_rsrc_put *prsrc;
7840 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7844 prsrc->tag = *io_get_tag_slot(data, idx);
7846 list_add(&prsrc->list, &node->rsrc_list);
7850 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7851 struct io_uring_rsrc_update2 *up,
7854 u64 __user *tags = u64_to_user_ptr(up->tags);
7855 __s32 __user *fds = u64_to_user_ptr(up->data);
7856 struct io_rsrc_data *data = ctx->file_data;
7857 struct io_fixed_file *file_slot;
7861 bool needs_switch = false;
7863 if (!ctx->file_data)
7865 if (up->offset + nr_args > ctx->nr_user_files)
7868 for (done = 0; done < nr_args; done++) {
7871 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7872 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7876 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7880 if (fd == IORING_REGISTER_FILES_SKIP)
7883 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7884 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7886 if (file_slot->file_ptr) {
7887 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7888 err = io_queue_rsrc_removal(data, up->offset + done,
7889 ctx->rsrc_node, file);
7892 file_slot->file_ptr = 0;
7893 needs_switch = true;
7902 * Don't allow io_uring instances to be registered. If
7903 * UNIX isn't enabled, then this causes a reference
7904 * cycle and this instance can never get freed. If UNIX
7905 * is enabled we'll handle it just fine, but there's
7906 * still no point in allowing a ring fd as it doesn't
7907 * support regular read/write anyway.
7909 if (file->f_op == &io_uring_fops) {
7914 *io_get_tag_slot(data, up->offset + done) = tag;
7915 io_fixed_file_set(file_slot, file);
7916 err = io_sqe_file_register(ctx, file, i);
7918 file_slot->file_ptr = 0;
7926 io_rsrc_node_switch(ctx, data);
7927 return done ? done : err;
7930 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7931 struct task_struct *task)
7933 struct io_wq_hash *hash;
7934 struct io_wq_data data;
7935 unsigned int concurrency;
7937 mutex_lock(&ctx->uring_lock);
7938 hash = ctx->hash_map;
7940 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7942 mutex_unlock(&ctx->uring_lock);
7943 return ERR_PTR(-ENOMEM);
7945 refcount_set(&hash->refs, 1);
7946 init_waitqueue_head(&hash->wait);
7947 ctx->hash_map = hash;
7949 mutex_unlock(&ctx->uring_lock);
7953 data.free_work = io_wq_free_work;
7954 data.do_work = io_wq_submit_work;
7956 /* Do QD, or 4 * CPUS, whatever is smallest */
7957 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7959 return io_wq_create(concurrency, &data);
7962 static int io_uring_alloc_task_context(struct task_struct *task,
7963 struct io_ring_ctx *ctx)
7965 struct io_uring_task *tctx;
7968 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7969 if (unlikely(!tctx))
7972 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7973 if (unlikely(ret)) {
7978 tctx->io_wq = io_init_wq_offload(ctx, task);
7979 if (IS_ERR(tctx->io_wq)) {
7980 ret = PTR_ERR(tctx->io_wq);
7981 percpu_counter_destroy(&tctx->inflight);
7987 init_waitqueue_head(&tctx->wait);
7988 atomic_set(&tctx->in_idle, 0);
7989 atomic_set(&tctx->inflight_tracked, 0);
7990 task->io_uring = tctx;
7991 spin_lock_init(&tctx->task_lock);
7992 INIT_WQ_LIST(&tctx->task_list);
7993 init_task_work(&tctx->task_work, tctx_task_work);
7997 void __io_uring_free(struct task_struct *tsk)
7999 struct io_uring_task *tctx = tsk->io_uring;
8001 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8002 WARN_ON_ONCE(tctx->io_wq);
8003 WARN_ON_ONCE(tctx->cached_refs);
8005 percpu_counter_destroy(&tctx->inflight);
8007 tsk->io_uring = NULL;
8010 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8011 struct io_uring_params *p)
8015 /* Retain compatibility with failing for an invalid attach attempt */
8016 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8017 IORING_SETUP_ATTACH_WQ) {
8020 f = fdget(p->wq_fd);
8023 if (f.file->f_op != &io_uring_fops) {
8029 if (ctx->flags & IORING_SETUP_SQPOLL) {
8030 struct task_struct *tsk;
8031 struct io_sq_data *sqd;
8034 sqd = io_get_sq_data(p, &attached);
8040 ctx->sq_creds = get_current_cred();
8042 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8043 if (!ctx->sq_thread_idle)
8044 ctx->sq_thread_idle = HZ;
8046 io_sq_thread_park(sqd);
8047 list_add(&ctx->sqd_list, &sqd->ctx_list);
8048 io_sqd_update_thread_idle(sqd);
8049 /* don't attach to a dying SQPOLL thread, would be racy */
8050 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8051 io_sq_thread_unpark(sqd);
8058 if (p->flags & IORING_SETUP_SQ_AFF) {
8059 int cpu = p->sq_thread_cpu;
8062 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8069 sqd->task_pid = current->pid;
8070 sqd->task_tgid = current->tgid;
8071 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8078 ret = io_uring_alloc_task_context(tsk, ctx);
8079 wake_up_new_task(tsk);
8082 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8083 /* Can't have SQ_AFF without SQPOLL */
8090 complete(&ctx->sq_data->exited);
8092 io_sq_thread_finish(ctx);
8096 static inline void __io_unaccount_mem(struct user_struct *user,
8097 unsigned long nr_pages)
8099 atomic_long_sub(nr_pages, &user->locked_vm);
8102 static inline int __io_account_mem(struct user_struct *user,
8103 unsigned long nr_pages)
8105 unsigned long page_limit, cur_pages, new_pages;
8107 /* Don't allow more pages than we can safely lock */
8108 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8111 cur_pages = atomic_long_read(&user->locked_vm);
8112 new_pages = cur_pages + nr_pages;
8113 if (new_pages > page_limit)
8115 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8116 new_pages) != cur_pages);
8121 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8124 __io_unaccount_mem(ctx->user, nr_pages);
8126 if (ctx->mm_account)
8127 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8130 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8135 ret = __io_account_mem(ctx->user, nr_pages);
8140 if (ctx->mm_account)
8141 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8146 static void io_mem_free(void *ptr)
8153 page = virt_to_head_page(ptr);
8154 if (put_page_testzero(page))
8155 free_compound_page(page);
8158 static void *io_mem_alloc(size_t size)
8160 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8161 __GFP_NORETRY | __GFP_ACCOUNT;
8163 return (void *) __get_free_pages(gfp_flags, get_order(size));
8166 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8169 struct io_rings *rings;
8170 size_t off, sq_array_size;
8172 off = struct_size(rings, cqes, cq_entries);
8173 if (off == SIZE_MAX)
8177 off = ALIGN(off, SMP_CACHE_BYTES);
8185 sq_array_size = array_size(sizeof(u32), sq_entries);
8186 if (sq_array_size == SIZE_MAX)
8189 if (check_add_overflow(off, sq_array_size, &off))
8195 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8197 struct io_mapped_ubuf *imu = *slot;
8200 if (imu != ctx->dummy_ubuf) {
8201 for (i = 0; i < imu->nr_bvecs; i++)
8202 unpin_user_page(imu->bvec[i].bv_page);
8203 if (imu->acct_pages)
8204 io_unaccount_mem(ctx, imu->acct_pages);
8210 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8212 io_buffer_unmap(ctx, &prsrc->buf);
8216 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8220 for (i = 0; i < ctx->nr_user_bufs; i++)
8221 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8222 kfree(ctx->user_bufs);
8223 io_rsrc_data_free(ctx->buf_data);
8224 ctx->user_bufs = NULL;
8225 ctx->buf_data = NULL;
8226 ctx->nr_user_bufs = 0;
8229 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8236 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8238 __io_sqe_buffers_unregister(ctx);
8242 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8243 void __user *arg, unsigned index)
8245 struct iovec __user *src;
8247 #ifdef CONFIG_COMPAT
8249 struct compat_iovec __user *ciovs;
8250 struct compat_iovec ciov;
8252 ciovs = (struct compat_iovec __user *) arg;
8253 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8256 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8257 dst->iov_len = ciov.iov_len;
8261 src = (struct iovec __user *) arg;
8262 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8268 * Not super efficient, but this is just a registration time. And we do cache
8269 * the last compound head, so generally we'll only do a full search if we don't
8272 * We check if the given compound head page has already been accounted, to
8273 * avoid double accounting it. This allows us to account the full size of the
8274 * page, not just the constituent pages of a huge page.
8276 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8277 int nr_pages, struct page *hpage)
8281 /* check current page array */
8282 for (i = 0; i < nr_pages; i++) {
8283 if (!PageCompound(pages[i]))
8285 if (compound_head(pages[i]) == hpage)
8289 /* check previously registered pages */
8290 for (i = 0; i < ctx->nr_user_bufs; i++) {
8291 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8293 for (j = 0; j < imu->nr_bvecs; j++) {
8294 if (!PageCompound(imu->bvec[j].bv_page))
8296 if (compound_head(imu->bvec[j].bv_page) == hpage)
8304 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8305 int nr_pages, struct io_mapped_ubuf *imu,
8306 struct page **last_hpage)
8310 imu->acct_pages = 0;
8311 for (i = 0; i < nr_pages; i++) {
8312 if (!PageCompound(pages[i])) {
8317 hpage = compound_head(pages[i]);
8318 if (hpage == *last_hpage)
8320 *last_hpage = hpage;
8321 if (headpage_already_acct(ctx, pages, i, hpage))
8323 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8327 if (!imu->acct_pages)
8330 ret = io_account_mem(ctx, imu->acct_pages);
8332 imu->acct_pages = 0;
8336 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8337 struct io_mapped_ubuf **pimu,
8338 struct page **last_hpage)
8340 struct io_mapped_ubuf *imu = NULL;
8341 struct vm_area_struct **vmas = NULL;
8342 struct page **pages = NULL;
8343 unsigned long off, start, end, ubuf;
8345 int ret, pret, nr_pages, i;
8347 if (!iov->iov_base) {
8348 *pimu = ctx->dummy_ubuf;
8352 ubuf = (unsigned long) iov->iov_base;
8353 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8354 start = ubuf >> PAGE_SHIFT;
8355 nr_pages = end - start;
8360 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8364 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8369 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8374 mmap_read_lock(current->mm);
8375 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8377 if (pret == nr_pages) {
8378 /* don't support file backed memory */
8379 for (i = 0; i < nr_pages; i++) {
8380 struct vm_area_struct *vma = vmas[i];
8382 if (vma_is_shmem(vma))
8385 !is_file_hugepages(vma->vm_file)) {
8391 ret = pret < 0 ? pret : -EFAULT;
8393 mmap_read_unlock(current->mm);
8396 * if we did partial map, or found file backed vmas,
8397 * release any pages we did get
8400 unpin_user_pages(pages, pret);
8404 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8406 unpin_user_pages(pages, pret);
8410 off = ubuf & ~PAGE_MASK;
8411 size = iov->iov_len;
8412 for (i = 0; i < nr_pages; i++) {
8415 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8416 imu->bvec[i].bv_page = pages[i];
8417 imu->bvec[i].bv_len = vec_len;
8418 imu->bvec[i].bv_offset = off;
8422 /* store original address for later verification */
8424 imu->ubuf_end = ubuf + iov->iov_len;
8425 imu->nr_bvecs = nr_pages;
8436 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8438 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8439 return ctx->user_bufs ? 0 : -ENOMEM;
8442 static int io_buffer_validate(struct iovec *iov)
8444 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8447 * Don't impose further limits on the size and buffer
8448 * constraints here, we'll -EINVAL later when IO is
8449 * submitted if they are wrong.
8452 return iov->iov_len ? -EFAULT : 0;
8456 /* arbitrary limit, but we need something */
8457 if (iov->iov_len > SZ_1G)
8460 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8466 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8467 unsigned int nr_args, u64 __user *tags)
8469 struct page *last_hpage = NULL;
8470 struct io_rsrc_data *data;
8476 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8478 ret = io_rsrc_node_switch_start(ctx);
8481 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8484 ret = io_buffers_map_alloc(ctx, nr_args);
8486 io_rsrc_data_free(data);
8490 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8491 ret = io_copy_iov(ctx, &iov, arg, i);
8494 ret = io_buffer_validate(&iov);
8497 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8502 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8508 WARN_ON_ONCE(ctx->buf_data);
8510 ctx->buf_data = data;
8512 __io_sqe_buffers_unregister(ctx);
8514 io_rsrc_node_switch(ctx, NULL);
8518 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8519 struct io_uring_rsrc_update2 *up,
8520 unsigned int nr_args)
8522 u64 __user *tags = u64_to_user_ptr(up->tags);
8523 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8524 struct page *last_hpage = NULL;
8525 bool needs_switch = false;
8531 if (up->offset + nr_args > ctx->nr_user_bufs)
8534 for (done = 0; done < nr_args; done++) {
8535 struct io_mapped_ubuf *imu;
8536 int offset = up->offset + done;
8539 err = io_copy_iov(ctx, &iov, iovs, done);
8542 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8546 err = io_buffer_validate(&iov);
8549 if (!iov.iov_base && tag) {
8553 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8557 i = array_index_nospec(offset, ctx->nr_user_bufs);
8558 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8559 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8560 ctx->rsrc_node, ctx->user_bufs[i]);
8561 if (unlikely(err)) {
8562 io_buffer_unmap(ctx, &imu);
8565 ctx->user_bufs[i] = NULL;
8566 needs_switch = true;
8569 ctx->user_bufs[i] = imu;
8570 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8574 io_rsrc_node_switch(ctx, ctx->buf_data);
8575 return done ? done : err;
8578 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8580 __s32 __user *fds = arg;
8586 if (copy_from_user(&fd, fds, sizeof(*fds)))
8589 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8590 if (IS_ERR(ctx->cq_ev_fd)) {
8591 int ret = PTR_ERR(ctx->cq_ev_fd);
8593 ctx->cq_ev_fd = NULL;
8600 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8602 if (ctx->cq_ev_fd) {
8603 eventfd_ctx_put(ctx->cq_ev_fd);
8604 ctx->cq_ev_fd = NULL;
8611 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8613 struct io_buffer *buf;
8614 unsigned long index;
8616 xa_for_each(&ctx->io_buffers, index, buf)
8617 __io_remove_buffers(ctx, buf, index, -1U);
8620 static void io_req_cache_free(struct list_head *list)
8622 struct io_kiocb *req, *nxt;
8624 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8625 list_del(&req->inflight_entry);
8626 kmem_cache_free(req_cachep, req);
8630 static void io_req_caches_free(struct io_ring_ctx *ctx)
8632 struct io_submit_state *state = &ctx->submit_state;
8634 mutex_lock(&ctx->uring_lock);
8636 if (state->free_reqs) {
8637 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8638 state->free_reqs = 0;
8641 io_flush_cached_locked_reqs(ctx, state);
8642 io_req_cache_free(&state->free_list);
8643 mutex_unlock(&ctx->uring_lock);
8646 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8648 if (data && !atomic_dec_and_test(&data->refs))
8649 wait_for_completion(&data->done);
8652 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8654 io_sq_thread_finish(ctx);
8656 if (ctx->mm_account) {
8657 mmdrop(ctx->mm_account);
8658 ctx->mm_account = NULL;
8661 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8662 io_wait_rsrc_data(ctx->buf_data);
8663 io_wait_rsrc_data(ctx->file_data);
8665 mutex_lock(&ctx->uring_lock);
8667 __io_sqe_buffers_unregister(ctx);
8669 __io_sqe_files_unregister(ctx);
8671 __io_cqring_overflow_flush(ctx, true);
8672 mutex_unlock(&ctx->uring_lock);
8673 io_eventfd_unregister(ctx);
8674 io_destroy_buffers(ctx);
8676 put_cred(ctx->sq_creds);
8678 /* there are no registered resources left, nobody uses it */
8680 io_rsrc_node_destroy(ctx->rsrc_node);
8681 if (ctx->rsrc_backup_node)
8682 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8683 flush_delayed_work(&ctx->rsrc_put_work);
8685 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8686 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8688 #if defined(CONFIG_UNIX)
8689 if (ctx->ring_sock) {
8690 ctx->ring_sock->file = NULL; /* so that iput() is called */
8691 sock_release(ctx->ring_sock);
8695 io_mem_free(ctx->rings);
8696 io_mem_free(ctx->sq_sqes);
8698 percpu_ref_exit(&ctx->refs);
8699 free_uid(ctx->user);
8700 io_req_caches_free(ctx);
8702 io_wq_put_hash(ctx->hash_map);
8703 kfree(ctx->cancel_hash);
8704 kfree(ctx->dummy_ubuf);
8708 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8710 struct io_ring_ctx *ctx = file->private_data;
8713 poll_wait(file, &ctx->poll_wait, wait);
8715 * synchronizes with barrier from wq_has_sleeper call in
8719 if (!io_sqring_full(ctx))
8720 mask |= EPOLLOUT | EPOLLWRNORM;
8723 * Don't flush cqring overflow list here, just do a simple check.
8724 * Otherwise there could possible be ABBA deadlock:
8727 * lock(&ctx->uring_lock);
8729 * lock(&ctx->uring_lock);
8732 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8733 * pushs them to do the flush.
8735 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8736 mask |= EPOLLIN | EPOLLRDNORM;
8741 static int io_uring_fasync(int fd, struct file *file, int on)
8743 struct io_ring_ctx *ctx = file->private_data;
8745 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8748 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8750 const struct cred *creds;
8752 creds = xa_erase(&ctx->personalities, id);
8761 struct io_tctx_exit {
8762 struct callback_head task_work;
8763 struct completion completion;
8764 struct io_ring_ctx *ctx;
8767 static void io_tctx_exit_cb(struct callback_head *cb)
8769 struct io_uring_task *tctx = current->io_uring;
8770 struct io_tctx_exit *work;
8772 work = container_of(cb, struct io_tctx_exit, task_work);
8774 * When @in_idle, we're in cancellation and it's racy to remove the
8775 * node. It'll be removed by the end of cancellation, just ignore it.
8777 if (!atomic_read(&tctx->in_idle))
8778 io_uring_del_tctx_node((unsigned long)work->ctx);
8779 complete(&work->completion);
8782 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8784 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8786 return req->ctx == data;
8789 static void io_ring_exit_work(struct work_struct *work)
8791 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8792 unsigned long timeout = jiffies + HZ * 60 * 5;
8793 unsigned long interval = HZ / 20;
8794 struct io_tctx_exit exit;
8795 struct io_tctx_node *node;
8799 * If we're doing polled IO and end up having requests being
8800 * submitted async (out-of-line), then completions can come in while
8801 * we're waiting for refs to drop. We need to reap these manually,
8802 * as nobody else will be looking for them.
8805 io_uring_try_cancel_requests(ctx, NULL, true);
8807 struct io_sq_data *sqd = ctx->sq_data;
8808 struct task_struct *tsk;
8810 io_sq_thread_park(sqd);
8812 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8813 io_wq_cancel_cb(tsk->io_uring->io_wq,
8814 io_cancel_ctx_cb, ctx, true);
8815 io_sq_thread_unpark(sqd);
8818 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8819 /* there is little hope left, don't run it too often */
8822 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8824 init_completion(&exit.completion);
8825 init_task_work(&exit.task_work, io_tctx_exit_cb);
8828 * Some may use context even when all refs and requests have been put,
8829 * and they are free to do so while still holding uring_lock or
8830 * completion_lock, see io_req_task_submit(). Apart from other work,
8831 * this lock/unlock section also waits them to finish.
8833 mutex_lock(&ctx->uring_lock);
8834 while (!list_empty(&ctx->tctx_list)) {
8835 WARN_ON_ONCE(time_after(jiffies, timeout));
8837 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8839 /* don't spin on a single task if cancellation failed */
8840 list_rotate_left(&ctx->tctx_list);
8841 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8842 if (WARN_ON_ONCE(ret))
8844 wake_up_process(node->task);
8846 mutex_unlock(&ctx->uring_lock);
8847 wait_for_completion(&exit.completion);
8848 mutex_lock(&ctx->uring_lock);
8850 mutex_unlock(&ctx->uring_lock);
8851 spin_lock(&ctx->completion_lock);
8852 spin_unlock(&ctx->completion_lock);
8854 io_ring_ctx_free(ctx);
8857 /* Returns true if we found and killed one or more timeouts */
8858 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8861 struct io_kiocb *req, *tmp;
8864 spin_lock(&ctx->completion_lock);
8865 spin_lock_irq(&ctx->timeout_lock);
8866 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8867 if (io_match_task(req, tsk, cancel_all)) {
8868 io_kill_timeout(req, -ECANCELED);
8872 spin_unlock_irq(&ctx->timeout_lock);
8874 io_commit_cqring(ctx);
8875 spin_unlock(&ctx->completion_lock);
8877 io_cqring_ev_posted(ctx);
8878 return canceled != 0;
8881 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8883 unsigned long index;
8884 struct creds *creds;
8886 mutex_lock(&ctx->uring_lock);
8887 percpu_ref_kill(&ctx->refs);
8889 __io_cqring_overflow_flush(ctx, true);
8890 xa_for_each(&ctx->personalities, index, creds)
8891 io_unregister_personality(ctx, index);
8892 mutex_unlock(&ctx->uring_lock);
8894 io_kill_timeouts(ctx, NULL, true);
8895 io_poll_remove_all(ctx, NULL, true);
8897 /* if we failed setting up the ctx, we might not have any rings */
8898 io_iopoll_try_reap_events(ctx);
8900 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8902 * Use system_unbound_wq to avoid spawning tons of event kworkers
8903 * if we're exiting a ton of rings at the same time. It just adds
8904 * noise and overhead, there's no discernable change in runtime
8905 * over using system_wq.
8907 queue_work(system_unbound_wq, &ctx->exit_work);
8910 static int io_uring_release(struct inode *inode, struct file *file)
8912 struct io_ring_ctx *ctx = file->private_data;
8914 file->private_data = NULL;
8915 io_ring_ctx_wait_and_kill(ctx);
8919 struct io_task_cancel {
8920 struct task_struct *task;
8924 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8926 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8927 struct io_task_cancel *cancel = data;
8930 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8931 struct io_ring_ctx *ctx = req->ctx;
8933 /* protect against races with linked timeouts */
8934 spin_lock(&ctx->completion_lock);
8935 ret = io_match_task(req, cancel->task, cancel->all);
8936 spin_unlock(&ctx->completion_lock);
8938 ret = io_match_task(req, cancel->task, cancel->all);
8943 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8944 struct task_struct *task, bool cancel_all)
8946 struct io_defer_entry *de;
8949 spin_lock(&ctx->completion_lock);
8950 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8951 if (io_match_task(de->req, task, cancel_all)) {
8952 list_cut_position(&list, &ctx->defer_list, &de->list);
8956 spin_unlock(&ctx->completion_lock);
8957 if (list_empty(&list))
8960 while (!list_empty(&list)) {
8961 de = list_first_entry(&list, struct io_defer_entry, list);
8962 list_del_init(&de->list);
8963 io_req_complete_failed(de->req, -ECANCELED);
8969 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8971 struct io_tctx_node *node;
8972 enum io_wq_cancel cret;
8975 mutex_lock(&ctx->uring_lock);
8976 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8977 struct io_uring_task *tctx = node->task->io_uring;
8980 * io_wq will stay alive while we hold uring_lock, because it's
8981 * killed after ctx nodes, which requires to take the lock.
8983 if (!tctx || !tctx->io_wq)
8985 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8986 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8988 mutex_unlock(&ctx->uring_lock);
8993 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8994 struct task_struct *task,
8997 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8998 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9001 enum io_wq_cancel cret;
9005 ret |= io_uring_try_cancel_iowq(ctx);
9006 } else if (tctx && tctx->io_wq) {
9008 * Cancels requests of all rings, not only @ctx, but
9009 * it's fine as the task is in exit/exec.
9011 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9013 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9016 /* SQPOLL thread does its own polling */
9017 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9018 (ctx->sq_data && ctx->sq_data->thread == current)) {
9019 while (!list_empty_careful(&ctx->iopoll_list)) {
9020 io_iopoll_try_reap_events(ctx);
9025 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9026 ret |= io_poll_remove_all(ctx, task, cancel_all);
9027 ret |= io_kill_timeouts(ctx, task, cancel_all);
9029 ret |= io_run_task_work();
9036 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9038 struct io_uring_task *tctx = current->io_uring;
9039 struct io_tctx_node *node;
9042 if (unlikely(!tctx)) {
9043 ret = io_uring_alloc_task_context(current, ctx);
9046 tctx = current->io_uring;
9048 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9049 node = kmalloc(sizeof(*node), GFP_KERNEL);
9053 node->task = current;
9055 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9062 mutex_lock(&ctx->uring_lock);
9063 list_add(&node->ctx_node, &ctx->tctx_list);
9064 mutex_unlock(&ctx->uring_lock);
9071 * Note that this task has used io_uring. We use it for cancelation purposes.
9073 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9075 struct io_uring_task *tctx = current->io_uring;
9077 if (likely(tctx && tctx->last == ctx))
9079 return __io_uring_add_tctx_node(ctx);
9083 * Remove this io_uring_file -> task mapping.
9085 static void io_uring_del_tctx_node(unsigned long index)
9087 struct io_uring_task *tctx = current->io_uring;
9088 struct io_tctx_node *node;
9092 node = xa_erase(&tctx->xa, index);
9096 WARN_ON_ONCE(current != node->task);
9097 WARN_ON_ONCE(list_empty(&node->ctx_node));
9099 mutex_lock(&node->ctx->uring_lock);
9100 list_del(&node->ctx_node);
9101 mutex_unlock(&node->ctx->uring_lock);
9103 if (tctx->last == node->ctx)
9108 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9110 struct io_wq *wq = tctx->io_wq;
9111 struct io_tctx_node *node;
9112 unsigned long index;
9114 xa_for_each(&tctx->xa, index, node)
9115 io_uring_del_tctx_node(index);
9118 * Must be after io_uring_del_task_file() (removes nodes under
9119 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9122 io_wq_put_and_exit(wq);
9126 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9129 return atomic_read(&tctx->inflight_tracked);
9130 return percpu_counter_sum(&tctx->inflight);
9133 static void io_uring_drop_tctx_refs(struct task_struct *task)
9135 struct io_uring_task *tctx = task->io_uring;
9136 unsigned int refs = tctx->cached_refs;
9139 tctx->cached_refs = 0;
9140 percpu_counter_sub(&tctx->inflight, refs);
9141 put_task_struct_many(task, refs);
9146 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9147 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9149 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9151 struct io_uring_task *tctx = current->io_uring;
9152 struct io_ring_ctx *ctx;
9156 WARN_ON_ONCE(sqd && sqd->thread != current);
9158 if (!current->io_uring)
9161 io_wq_exit_start(tctx->io_wq);
9163 atomic_inc(&tctx->in_idle);
9165 io_uring_drop_tctx_refs(current);
9166 /* read completions before cancelations */
9167 inflight = tctx_inflight(tctx, !cancel_all);
9172 struct io_tctx_node *node;
9173 unsigned long index;
9175 xa_for_each(&tctx->xa, index, node) {
9176 /* sqpoll task will cancel all its requests */
9177 if (node->ctx->sq_data)
9179 io_uring_try_cancel_requests(node->ctx, current,
9183 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9184 io_uring_try_cancel_requests(ctx, current,
9188 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9189 io_uring_drop_tctx_refs(current);
9191 * If we've seen completions, retry without waiting. This
9192 * avoids a race where a completion comes in before we did
9193 * prepare_to_wait().
9195 if (inflight == tctx_inflight(tctx, !cancel_all))
9197 finish_wait(&tctx->wait, &wait);
9199 atomic_dec(&tctx->in_idle);
9201 io_uring_clean_tctx(tctx);
9203 /* for exec all current's requests should be gone, kill tctx */
9204 __io_uring_free(current);
9208 void __io_uring_cancel(bool cancel_all)
9210 io_uring_cancel_generic(cancel_all, NULL);
9213 static void *io_uring_validate_mmap_request(struct file *file,
9214 loff_t pgoff, size_t sz)
9216 struct io_ring_ctx *ctx = file->private_data;
9217 loff_t offset = pgoff << PAGE_SHIFT;
9222 case IORING_OFF_SQ_RING:
9223 case IORING_OFF_CQ_RING:
9226 case IORING_OFF_SQES:
9230 return ERR_PTR(-EINVAL);
9233 page = virt_to_head_page(ptr);
9234 if (sz > page_size(page))
9235 return ERR_PTR(-EINVAL);
9242 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9244 size_t sz = vma->vm_end - vma->vm_start;
9248 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9250 return PTR_ERR(ptr);
9252 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9253 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9256 #else /* !CONFIG_MMU */
9258 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9260 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9263 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9265 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9268 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9269 unsigned long addr, unsigned long len,
9270 unsigned long pgoff, unsigned long flags)
9274 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9276 return PTR_ERR(ptr);
9278 return (unsigned long) ptr;
9281 #endif /* !CONFIG_MMU */
9283 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9288 if (!io_sqring_full(ctx))
9290 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9292 if (!io_sqring_full(ctx))
9295 } while (!signal_pending(current));
9297 finish_wait(&ctx->sqo_sq_wait, &wait);
9301 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9302 struct __kernel_timespec __user **ts,
9303 const sigset_t __user **sig)
9305 struct io_uring_getevents_arg arg;
9308 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9309 * is just a pointer to the sigset_t.
9311 if (!(flags & IORING_ENTER_EXT_ARG)) {
9312 *sig = (const sigset_t __user *) argp;
9318 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9319 * timespec and sigset_t pointers if good.
9321 if (*argsz != sizeof(arg))
9323 if (copy_from_user(&arg, argp, sizeof(arg)))
9325 *sig = u64_to_user_ptr(arg.sigmask);
9326 *argsz = arg.sigmask_sz;
9327 *ts = u64_to_user_ptr(arg.ts);
9331 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9332 u32, min_complete, u32, flags, const void __user *, argp,
9335 struct io_ring_ctx *ctx;
9342 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9343 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9347 if (unlikely(!f.file))
9351 if (unlikely(f.file->f_op != &io_uring_fops))
9355 ctx = f.file->private_data;
9356 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9360 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9364 * For SQ polling, the thread will do all submissions and completions.
9365 * Just return the requested submit count, and wake the thread if
9369 if (ctx->flags & IORING_SETUP_SQPOLL) {
9370 io_cqring_overflow_flush(ctx);
9372 if (unlikely(ctx->sq_data->thread == NULL)) {
9376 if (flags & IORING_ENTER_SQ_WAKEUP)
9377 wake_up(&ctx->sq_data->wait);
9378 if (flags & IORING_ENTER_SQ_WAIT) {
9379 ret = io_sqpoll_wait_sq(ctx);
9383 submitted = to_submit;
9384 } else if (to_submit) {
9385 ret = io_uring_add_tctx_node(ctx);
9388 mutex_lock(&ctx->uring_lock);
9389 submitted = io_submit_sqes(ctx, to_submit);
9390 mutex_unlock(&ctx->uring_lock);
9392 if (submitted != to_submit)
9395 if (flags & IORING_ENTER_GETEVENTS) {
9396 const sigset_t __user *sig;
9397 struct __kernel_timespec __user *ts;
9399 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9403 min_complete = min(min_complete, ctx->cq_entries);
9406 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9407 * space applications don't need to do io completion events
9408 * polling again, they can rely on io_sq_thread to do polling
9409 * work, which can reduce cpu usage and uring_lock contention.
9411 if (ctx->flags & IORING_SETUP_IOPOLL &&
9412 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9413 ret = io_iopoll_check(ctx, min_complete);
9415 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9420 percpu_ref_put(&ctx->refs);
9423 return submitted ? submitted : ret;
9426 #ifdef CONFIG_PROC_FS
9427 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9428 const struct cred *cred)
9430 struct user_namespace *uns = seq_user_ns(m);
9431 struct group_info *gi;
9436 seq_printf(m, "%5d\n", id);
9437 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9438 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9439 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9440 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9441 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9442 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9443 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9444 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9445 seq_puts(m, "\n\tGroups:\t");
9446 gi = cred->group_info;
9447 for (g = 0; g < gi->ngroups; g++) {
9448 seq_put_decimal_ull(m, g ? " " : "",
9449 from_kgid_munged(uns, gi->gid[g]));
9451 seq_puts(m, "\n\tCapEff:\t");
9452 cap = cred->cap_effective;
9453 CAP_FOR_EACH_U32(__capi)
9454 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9459 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9461 struct io_sq_data *sq = NULL;
9466 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9467 * since fdinfo case grabs it in the opposite direction of normal use
9468 * cases. If we fail to get the lock, we just don't iterate any
9469 * structures that could be going away outside the io_uring mutex.
9471 has_lock = mutex_trylock(&ctx->uring_lock);
9473 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9479 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9480 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9481 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9482 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9483 struct file *f = io_file_from_index(ctx, i);
9486 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9488 seq_printf(m, "%5u: <none>\n", i);
9490 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9491 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9492 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9493 unsigned int len = buf->ubuf_end - buf->ubuf;
9495 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9497 if (has_lock && !xa_empty(&ctx->personalities)) {
9498 unsigned long index;
9499 const struct cred *cred;
9501 seq_printf(m, "Personalities:\n");
9502 xa_for_each(&ctx->personalities, index, cred)
9503 io_uring_show_cred(m, index, cred);
9505 seq_printf(m, "PollList:\n");
9506 spin_lock(&ctx->completion_lock);
9507 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9508 struct hlist_head *list = &ctx->cancel_hash[i];
9509 struct io_kiocb *req;
9511 hlist_for_each_entry(req, list, hash_node)
9512 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9513 req->task->task_works != NULL);
9515 spin_unlock(&ctx->completion_lock);
9517 mutex_unlock(&ctx->uring_lock);
9520 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9522 struct io_ring_ctx *ctx = f->private_data;
9524 if (percpu_ref_tryget(&ctx->refs)) {
9525 __io_uring_show_fdinfo(ctx, m);
9526 percpu_ref_put(&ctx->refs);
9531 static const struct file_operations io_uring_fops = {
9532 .release = io_uring_release,
9533 .mmap = io_uring_mmap,
9535 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9536 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9538 .poll = io_uring_poll,
9539 .fasync = io_uring_fasync,
9540 #ifdef CONFIG_PROC_FS
9541 .show_fdinfo = io_uring_show_fdinfo,
9545 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9546 struct io_uring_params *p)
9548 struct io_rings *rings;
9549 size_t size, sq_array_offset;
9551 /* make sure these are sane, as we already accounted them */
9552 ctx->sq_entries = p->sq_entries;
9553 ctx->cq_entries = p->cq_entries;
9555 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9556 if (size == SIZE_MAX)
9559 rings = io_mem_alloc(size);
9564 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9565 rings->sq_ring_mask = p->sq_entries - 1;
9566 rings->cq_ring_mask = p->cq_entries - 1;
9567 rings->sq_ring_entries = p->sq_entries;
9568 rings->cq_ring_entries = p->cq_entries;
9570 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9571 if (size == SIZE_MAX) {
9572 io_mem_free(ctx->rings);
9577 ctx->sq_sqes = io_mem_alloc(size);
9578 if (!ctx->sq_sqes) {
9579 io_mem_free(ctx->rings);
9587 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9591 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9595 ret = io_uring_add_tctx_node(ctx);
9600 fd_install(fd, file);
9605 * Allocate an anonymous fd, this is what constitutes the application
9606 * visible backing of an io_uring instance. The application mmaps this
9607 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9608 * we have to tie this fd to a socket for file garbage collection purposes.
9610 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9613 #if defined(CONFIG_UNIX)
9616 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9619 return ERR_PTR(ret);
9622 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9623 O_RDWR | O_CLOEXEC);
9624 #if defined(CONFIG_UNIX)
9626 sock_release(ctx->ring_sock);
9627 ctx->ring_sock = NULL;
9629 ctx->ring_sock->file = file;
9635 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9636 struct io_uring_params __user *params)
9638 struct io_ring_ctx *ctx;
9644 if (entries > IORING_MAX_ENTRIES) {
9645 if (!(p->flags & IORING_SETUP_CLAMP))
9647 entries = IORING_MAX_ENTRIES;
9651 * Use twice as many entries for the CQ ring. It's possible for the
9652 * application to drive a higher depth than the size of the SQ ring,
9653 * since the sqes are only used at submission time. This allows for
9654 * some flexibility in overcommitting a bit. If the application has
9655 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9656 * of CQ ring entries manually.
9658 p->sq_entries = roundup_pow_of_two(entries);
9659 if (p->flags & IORING_SETUP_CQSIZE) {
9661 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9662 * to a power-of-two, if it isn't already. We do NOT impose
9663 * any cq vs sq ring sizing.
9667 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9668 if (!(p->flags & IORING_SETUP_CLAMP))
9670 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9672 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9673 if (p->cq_entries < p->sq_entries)
9676 p->cq_entries = 2 * p->sq_entries;
9679 ctx = io_ring_ctx_alloc(p);
9682 ctx->compat = in_compat_syscall();
9683 if (!capable(CAP_IPC_LOCK))
9684 ctx->user = get_uid(current_user());
9687 * This is just grabbed for accounting purposes. When a process exits,
9688 * the mm is exited and dropped before the files, hence we need to hang
9689 * on to this mm purely for the purposes of being able to unaccount
9690 * memory (locked/pinned vm). It's not used for anything else.
9692 mmgrab(current->mm);
9693 ctx->mm_account = current->mm;
9695 ret = io_allocate_scq_urings(ctx, p);
9699 ret = io_sq_offload_create(ctx, p);
9702 /* always set a rsrc node */
9703 ret = io_rsrc_node_switch_start(ctx);
9706 io_rsrc_node_switch(ctx, NULL);
9708 memset(&p->sq_off, 0, sizeof(p->sq_off));
9709 p->sq_off.head = offsetof(struct io_rings, sq.head);
9710 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9711 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9712 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9713 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9714 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9715 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9717 memset(&p->cq_off, 0, sizeof(p->cq_off));
9718 p->cq_off.head = offsetof(struct io_rings, cq.head);
9719 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9720 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9721 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9722 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9723 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9724 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9726 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9727 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9728 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9729 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9730 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9731 IORING_FEAT_RSRC_TAGS;
9733 if (copy_to_user(params, p, sizeof(*p))) {
9738 file = io_uring_get_file(ctx);
9740 ret = PTR_ERR(file);
9745 * Install ring fd as the very last thing, so we don't risk someone
9746 * having closed it before we finish setup
9748 ret = io_uring_install_fd(ctx, file);
9750 /* fput will clean it up */
9755 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9758 io_ring_ctx_wait_and_kill(ctx);
9763 * Sets up an aio uring context, and returns the fd. Applications asks for a
9764 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9765 * params structure passed in.
9767 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9769 struct io_uring_params p;
9772 if (copy_from_user(&p, params, sizeof(p)))
9774 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9779 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9780 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9781 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9782 IORING_SETUP_R_DISABLED))
9785 return io_uring_create(entries, &p, params);
9788 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9789 struct io_uring_params __user *, params)
9791 return io_uring_setup(entries, params);
9794 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9796 struct io_uring_probe *p;
9800 size = struct_size(p, ops, nr_args);
9801 if (size == SIZE_MAX)
9803 p = kzalloc(size, GFP_KERNEL);
9808 if (copy_from_user(p, arg, size))
9811 if (memchr_inv(p, 0, size))
9814 p->last_op = IORING_OP_LAST - 1;
9815 if (nr_args > IORING_OP_LAST)
9816 nr_args = IORING_OP_LAST;
9818 for (i = 0; i < nr_args; i++) {
9820 if (!io_op_defs[i].not_supported)
9821 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9826 if (copy_to_user(arg, p, size))
9833 static int io_register_personality(struct io_ring_ctx *ctx)
9835 const struct cred *creds;
9839 creds = get_current_cred();
9841 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9842 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9850 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9851 unsigned int nr_args)
9853 struct io_uring_restriction *res;
9857 /* Restrictions allowed only if rings started disabled */
9858 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9861 /* We allow only a single restrictions registration */
9862 if (ctx->restrictions.registered)
9865 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9868 size = array_size(nr_args, sizeof(*res));
9869 if (size == SIZE_MAX)
9872 res = memdup_user(arg, size);
9874 return PTR_ERR(res);
9878 for (i = 0; i < nr_args; i++) {
9879 switch (res[i].opcode) {
9880 case IORING_RESTRICTION_REGISTER_OP:
9881 if (res[i].register_op >= IORING_REGISTER_LAST) {
9886 __set_bit(res[i].register_op,
9887 ctx->restrictions.register_op);
9889 case IORING_RESTRICTION_SQE_OP:
9890 if (res[i].sqe_op >= IORING_OP_LAST) {
9895 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9897 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9898 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9900 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9901 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9910 /* Reset all restrictions if an error happened */
9912 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9914 ctx->restrictions.registered = true;
9920 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9922 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9925 if (ctx->restrictions.registered)
9926 ctx->restricted = 1;
9928 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9929 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9930 wake_up(&ctx->sq_data->wait);
9934 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9935 struct io_uring_rsrc_update2 *up,
9943 if (check_add_overflow(up->offset, nr_args, &tmp))
9945 err = io_rsrc_node_switch_start(ctx);
9950 case IORING_RSRC_FILE:
9951 return __io_sqe_files_update(ctx, up, nr_args);
9952 case IORING_RSRC_BUFFER:
9953 return __io_sqe_buffers_update(ctx, up, nr_args);
9958 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9961 struct io_uring_rsrc_update2 up;
9965 memset(&up, 0, sizeof(up));
9966 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9968 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9971 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9972 unsigned size, unsigned type)
9974 struct io_uring_rsrc_update2 up;
9976 if (size != sizeof(up))
9978 if (copy_from_user(&up, arg, sizeof(up)))
9980 if (!up.nr || up.resv)
9982 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9985 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9986 unsigned int size, unsigned int type)
9988 struct io_uring_rsrc_register rr;
9990 /* keep it extendible */
9991 if (size != sizeof(rr))
9994 memset(&rr, 0, sizeof(rr));
9995 if (copy_from_user(&rr, arg, size))
9997 if (!rr.nr || rr.resv || rr.resv2)
10001 case IORING_RSRC_FILE:
10002 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10003 rr.nr, u64_to_user_ptr(rr.tags));
10004 case IORING_RSRC_BUFFER:
10005 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10006 rr.nr, u64_to_user_ptr(rr.tags));
10011 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10014 struct io_uring_task *tctx = current->io_uring;
10015 cpumask_var_t new_mask;
10018 if (!tctx || !tctx->io_wq)
10021 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10024 cpumask_clear(new_mask);
10025 if (len > cpumask_size())
10026 len = cpumask_size();
10028 if (copy_from_user(new_mask, arg, len)) {
10029 free_cpumask_var(new_mask);
10033 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10034 free_cpumask_var(new_mask);
10038 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10040 struct io_uring_task *tctx = current->io_uring;
10042 if (!tctx || !tctx->io_wq)
10045 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10048 static bool io_register_op_must_quiesce(int op)
10051 case IORING_REGISTER_BUFFERS:
10052 case IORING_UNREGISTER_BUFFERS:
10053 case IORING_REGISTER_FILES:
10054 case IORING_UNREGISTER_FILES:
10055 case IORING_REGISTER_FILES_UPDATE:
10056 case IORING_REGISTER_PROBE:
10057 case IORING_REGISTER_PERSONALITY:
10058 case IORING_UNREGISTER_PERSONALITY:
10059 case IORING_REGISTER_FILES2:
10060 case IORING_REGISTER_FILES_UPDATE2:
10061 case IORING_REGISTER_BUFFERS2:
10062 case IORING_REGISTER_BUFFERS_UPDATE:
10063 case IORING_REGISTER_IOWQ_AFF:
10064 case IORING_UNREGISTER_IOWQ_AFF:
10071 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10075 percpu_ref_kill(&ctx->refs);
10078 * Drop uring mutex before waiting for references to exit. If another
10079 * thread is currently inside io_uring_enter() it might need to grab the
10080 * uring_lock to make progress. If we hold it here across the drain
10081 * wait, then we can deadlock. It's safe to drop the mutex here, since
10082 * no new references will come in after we've killed the percpu ref.
10084 mutex_unlock(&ctx->uring_lock);
10086 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10089 ret = io_run_task_work_sig();
10090 } while (ret >= 0);
10091 mutex_lock(&ctx->uring_lock);
10094 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10098 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10099 void __user *arg, unsigned nr_args)
10100 __releases(ctx->uring_lock)
10101 __acquires(ctx->uring_lock)
10106 * We're inside the ring mutex, if the ref is already dying, then
10107 * someone else killed the ctx or is already going through
10108 * io_uring_register().
10110 if (percpu_ref_is_dying(&ctx->refs))
10113 if (ctx->restricted) {
10114 if (opcode >= IORING_REGISTER_LAST)
10116 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10117 if (!test_bit(opcode, ctx->restrictions.register_op))
10121 if (io_register_op_must_quiesce(opcode)) {
10122 ret = io_ctx_quiesce(ctx);
10128 case IORING_REGISTER_BUFFERS:
10129 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10131 case IORING_UNREGISTER_BUFFERS:
10133 if (arg || nr_args)
10135 ret = io_sqe_buffers_unregister(ctx);
10137 case IORING_REGISTER_FILES:
10138 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10140 case IORING_UNREGISTER_FILES:
10142 if (arg || nr_args)
10144 ret = io_sqe_files_unregister(ctx);
10146 case IORING_REGISTER_FILES_UPDATE:
10147 ret = io_register_files_update(ctx, arg, nr_args);
10149 case IORING_REGISTER_EVENTFD:
10150 case IORING_REGISTER_EVENTFD_ASYNC:
10154 ret = io_eventfd_register(ctx, arg);
10157 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10158 ctx->eventfd_async = 1;
10160 ctx->eventfd_async = 0;
10162 case IORING_UNREGISTER_EVENTFD:
10164 if (arg || nr_args)
10166 ret = io_eventfd_unregister(ctx);
10168 case IORING_REGISTER_PROBE:
10170 if (!arg || nr_args > 256)
10172 ret = io_probe(ctx, arg, nr_args);
10174 case IORING_REGISTER_PERSONALITY:
10176 if (arg || nr_args)
10178 ret = io_register_personality(ctx);
10180 case IORING_UNREGISTER_PERSONALITY:
10184 ret = io_unregister_personality(ctx, nr_args);
10186 case IORING_REGISTER_ENABLE_RINGS:
10188 if (arg || nr_args)
10190 ret = io_register_enable_rings(ctx);
10192 case IORING_REGISTER_RESTRICTIONS:
10193 ret = io_register_restrictions(ctx, arg, nr_args);
10195 case IORING_REGISTER_FILES2:
10196 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10198 case IORING_REGISTER_FILES_UPDATE2:
10199 ret = io_register_rsrc_update(ctx, arg, nr_args,
10202 case IORING_REGISTER_BUFFERS2:
10203 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10205 case IORING_REGISTER_BUFFERS_UPDATE:
10206 ret = io_register_rsrc_update(ctx, arg, nr_args,
10207 IORING_RSRC_BUFFER);
10209 case IORING_REGISTER_IOWQ_AFF:
10211 if (!arg || !nr_args)
10213 ret = io_register_iowq_aff(ctx, arg, nr_args);
10215 case IORING_UNREGISTER_IOWQ_AFF:
10217 if (arg || nr_args)
10219 ret = io_unregister_iowq_aff(ctx);
10226 if (io_register_op_must_quiesce(opcode)) {
10227 /* bring the ctx back to life */
10228 percpu_ref_reinit(&ctx->refs);
10229 reinit_completion(&ctx->ref_comp);
10234 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10235 void __user *, arg, unsigned int, nr_args)
10237 struct io_ring_ctx *ctx;
10246 if (f.file->f_op != &io_uring_fops)
10249 ctx = f.file->private_data;
10251 io_run_task_work();
10253 mutex_lock(&ctx->uring_lock);
10254 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10255 mutex_unlock(&ctx->uring_lock);
10256 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10257 ctx->cq_ev_fd != NULL, ret);
10263 static int __init io_uring_init(void)
10265 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10266 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10267 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10270 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10271 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10272 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10273 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10274 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10275 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10276 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10277 BUILD_BUG_SQE_ELEM(8, __u64, off);
10278 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10279 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10280 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10281 BUILD_BUG_SQE_ELEM(24, __u32, len);
10282 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10283 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10284 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10286 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10287 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10288 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10289 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10290 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10291 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10292 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10293 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10296 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10297 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10298 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10299 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10300 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10301 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10303 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10304 sizeof(struct io_uring_rsrc_update));
10305 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10306 sizeof(struct io_uring_rsrc_update2));
10307 /* should fit into one byte */
10308 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10310 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10311 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10313 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10317 __initcall(io_uring_init);