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 /* requests with any of those set should undergo io_disarm_next() */
1038 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1040 static bool io_disarm_next(struct io_kiocb *req);
1041 static void io_uring_del_tctx_node(unsigned long index);
1042 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1043 struct task_struct *task,
1045 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1047 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1048 long res, unsigned int cflags);
1049 static void io_put_req(struct io_kiocb *req);
1050 static void io_put_req_deferred(struct io_kiocb *req);
1051 static void io_dismantle_req(struct io_kiocb *req);
1052 static void io_queue_linked_timeout(struct io_kiocb *req);
1053 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1054 struct io_uring_rsrc_update2 *up,
1056 static void io_clean_op(struct io_kiocb *req);
1057 static struct file *io_file_get(struct io_ring_ctx *ctx,
1058 struct io_kiocb *req, int fd, bool fixed);
1059 static void __io_queue_sqe(struct io_kiocb *req);
1060 static void io_rsrc_put_work(struct work_struct *work);
1062 static void io_req_task_queue(struct io_kiocb *req);
1063 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1064 static int io_req_prep_async(struct io_kiocb *req);
1066 static struct kmem_cache *req_cachep;
1068 static const struct file_operations io_uring_fops;
1070 struct sock *io_uring_get_socket(struct file *file)
1072 #if defined(CONFIG_UNIX)
1073 if (file->f_op == &io_uring_fops) {
1074 struct io_ring_ctx *ctx = file->private_data;
1076 return ctx->ring_sock->sk;
1081 EXPORT_SYMBOL(io_uring_get_socket);
1083 #define io_for_each_link(pos, head) \
1084 for (pos = (head); pos; pos = pos->link)
1087 * Shamelessly stolen from the mm implementation of page reference checking,
1088 * see commit f958d7b528b1 for details.
1090 #define req_ref_zero_or_close_to_overflow(req) \
1091 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1093 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1095 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1096 return atomic_inc_not_zero(&req->refs);
1099 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1101 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1104 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1105 return atomic_dec_and_test(&req->refs);
1108 static inline void req_ref_put(struct io_kiocb *req)
1110 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1111 WARN_ON_ONCE(req_ref_put_and_test(req));
1114 static inline void req_ref_get(struct io_kiocb *req)
1116 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1117 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1118 atomic_inc(&req->refs);
1121 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1123 if (!(req->flags & REQ_F_REFCOUNT)) {
1124 req->flags |= REQ_F_REFCOUNT;
1125 atomic_set(&req->refs, nr);
1129 static inline void io_req_set_refcount(struct io_kiocb *req)
1131 __io_req_set_refcount(req, 1);
1134 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1136 struct io_ring_ctx *ctx = req->ctx;
1138 if (!req->fixed_rsrc_refs) {
1139 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1140 percpu_ref_get(req->fixed_rsrc_refs);
1144 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1146 bool got = percpu_ref_tryget(ref);
1148 /* already at zero, wait for ->release() */
1150 wait_for_completion(compl);
1151 percpu_ref_resurrect(ref);
1153 percpu_ref_put(ref);
1156 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1159 struct io_kiocb *req;
1161 if (task && head->task != task)
1166 io_for_each_link(req, head) {
1167 if (req->flags & REQ_F_INFLIGHT)
1173 static inline void req_set_fail(struct io_kiocb *req)
1175 req->flags |= REQ_F_FAIL;
1178 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1180 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1182 complete(&ctx->ref_comp);
1185 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1187 return !req->timeout.off;
1190 static void io_fallback_req_func(struct work_struct *work)
1192 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1193 fallback_work.work);
1194 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1195 struct io_kiocb *req, *tmp;
1197 percpu_ref_get(&ctx->refs);
1198 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1199 req->io_task_work.func(req);
1200 percpu_ref_put(&ctx->refs);
1203 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1205 struct io_ring_ctx *ctx;
1208 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1213 * Use 5 bits less than the max cq entries, that should give us around
1214 * 32 entries per hash list if totally full and uniformly spread.
1216 hash_bits = ilog2(p->cq_entries);
1220 ctx->cancel_hash_bits = hash_bits;
1221 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1223 if (!ctx->cancel_hash)
1225 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1227 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1228 if (!ctx->dummy_ubuf)
1230 /* set invalid range, so io_import_fixed() fails meeting it */
1231 ctx->dummy_ubuf->ubuf = -1UL;
1233 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1234 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1237 ctx->flags = p->flags;
1238 init_waitqueue_head(&ctx->sqo_sq_wait);
1239 INIT_LIST_HEAD(&ctx->sqd_list);
1240 init_waitqueue_head(&ctx->poll_wait);
1241 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1242 init_completion(&ctx->ref_comp);
1243 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1244 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1245 mutex_init(&ctx->uring_lock);
1246 init_waitqueue_head(&ctx->cq_wait);
1247 spin_lock_init(&ctx->completion_lock);
1248 spin_lock_init(&ctx->timeout_lock);
1249 INIT_LIST_HEAD(&ctx->iopoll_list);
1250 INIT_LIST_HEAD(&ctx->defer_list);
1251 INIT_LIST_HEAD(&ctx->timeout_list);
1252 spin_lock_init(&ctx->rsrc_ref_lock);
1253 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1254 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1255 init_llist_head(&ctx->rsrc_put_llist);
1256 INIT_LIST_HEAD(&ctx->tctx_list);
1257 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1258 INIT_LIST_HEAD(&ctx->locked_free_list);
1259 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1262 kfree(ctx->dummy_ubuf);
1263 kfree(ctx->cancel_hash);
1268 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1270 struct io_rings *r = ctx->rings;
1272 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1276 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1278 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1279 struct io_ring_ctx *ctx = req->ctx;
1281 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1287 #define FFS_ASYNC_READ 0x1UL
1288 #define FFS_ASYNC_WRITE 0x2UL
1290 #define FFS_ISREG 0x4UL
1292 #define FFS_ISREG 0x0UL
1294 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1296 static inline bool io_req_ffs_set(struct io_kiocb *req)
1298 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1301 static void io_req_track_inflight(struct io_kiocb *req)
1303 if (!(req->flags & REQ_F_INFLIGHT)) {
1304 req->flags |= REQ_F_INFLIGHT;
1305 atomic_inc(¤t->io_uring->inflight_tracked);
1309 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1311 req->flags &= ~REQ_F_LINK_TIMEOUT;
1314 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1316 if (WARN_ON_ONCE(!req->link))
1319 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1320 req->flags |= REQ_F_LINK_TIMEOUT;
1322 /* linked timeouts should have two refs once prep'ed */
1323 io_req_set_refcount(req);
1324 __io_req_set_refcount(req->link, 2);
1328 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1330 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1332 return __io_prep_linked_timeout(req);
1335 static void io_prep_async_work(struct io_kiocb *req)
1337 const struct io_op_def *def = &io_op_defs[req->opcode];
1338 struct io_ring_ctx *ctx = req->ctx;
1340 if (!(req->flags & REQ_F_CREDS)) {
1341 req->flags |= REQ_F_CREDS;
1342 req->creds = get_current_cred();
1345 req->work.list.next = NULL;
1346 req->work.flags = 0;
1347 if (req->flags & REQ_F_FORCE_ASYNC)
1348 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1350 if (req->flags & REQ_F_ISREG) {
1351 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1352 io_wq_hash_work(&req->work, file_inode(req->file));
1353 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1354 if (def->unbound_nonreg_file)
1355 req->work.flags |= IO_WQ_WORK_UNBOUND;
1358 switch (req->opcode) {
1359 case IORING_OP_SPLICE:
1361 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1362 req->work.flags |= IO_WQ_WORK_UNBOUND;
1367 static void io_prep_async_link(struct io_kiocb *req)
1369 struct io_kiocb *cur;
1371 if (req->flags & REQ_F_LINK_TIMEOUT) {
1372 struct io_ring_ctx *ctx = req->ctx;
1374 spin_lock(&ctx->completion_lock);
1375 io_for_each_link(cur, req)
1376 io_prep_async_work(cur);
1377 spin_unlock(&ctx->completion_lock);
1379 io_for_each_link(cur, req)
1380 io_prep_async_work(cur);
1384 static void io_queue_async_work(struct io_kiocb *req)
1386 struct io_ring_ctx *ctx = req->ctx;
1387 struct io_kiocb *link = io_prep_linked_timeout(req);
1388 struct io_uring_task *tctx = req->task->io_uring;
1391 BUG_ON(!tctx->io_wq);
1393 /* init ->work of the whole link before punting */
1394 io_prep_async_link(req);
1397 * Not expected to happen, but if we do have a bug where this _can_
1398 * happen, catch it here and ensure the request is marked as
1399 * canceled. That will make io-wq go through the usual work cancel
1400 * procedure rather than attempt to run this request (or create a new
1403 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1404 req->work.flags |= IO_WQ_WORK_CANCEL;
1406 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1407 &req->work, req->flags);
1408 io_wq_enqueue(tctx->io_wq, &req->work);
1410 io_queue_linked_timeout(link);
1413 static void io_kill_timeout(struct io_kiocb *req, int status)
1414 __must_hold(&req->ctx->completion_lock)
1415 __must_hold(&req->ctx->timeout_lock)
1417 struct io_timeout_data *io = req->async_data;
1419 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1420 atomic_set(&req->ctx->cq_timeouts,
1421 atomic_read(&req->ctx->cq_timeouts) + 1);
1422 list_del_init(&req->timeout.list);
1423 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1424 io_put_req_deferred(req);
1428 static void io_queue_deferred(struct io_ring_ctx *ctx)
1430 while (!list_empty(&ctx->defer_list)) {
1431 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1432 struct io_defer_entry, list);
1434 if (req_need_defer(de->req, de->seq))
1436 list_del_init(&de->list);
1437 io_req_task_queue(de->req);
1442 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1443 __must_hold(&ctx->completion_lock)
1445 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1447 spin_lock_irq(&ctx->timeout_lock);
1448 while (!list_empty(&ctx->timeout_list)) {
1449 u32 events_needed, events_got;
1450 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1451 struct io_kiocb, timeout.list);
1453 if (io_is_timeout_noseq(req))
1457 * Since seq can easily wrap around over time, subtract
1458 * the last seq at which timeouts were flushed before comparing.
1459 * Assuming not more than 2^31-1 events have happened since,
1460 * these subtractions won't have wrapped, so we can check if
1461 * target is in [last_seq, current_seq] by comparing the two.
1463 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1464 events_got = seq - ctx->cq_last_tm_flush;
1465 if (events_got < events_needed)
1468 list_del_init(&req->timeout.list);
1469 io_kill_timeout(req, 0);
1471 ctx->cq_last_tm_flush = seq;
1472 spin_unlock_irq(&ctx->timeout_lock);
1475 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1477 if (ctx->off_timeout_used)
1478 io_flush_timeouts(ctx);
1479 if (ctx->drain_active)
1480 io_queue_deferred(ctx);
1483 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1485 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1486 __io_commit_cqring_flush(ctx);
1487 /* order cqe stores with ring update */
1488 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1491 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1493 struct io_rings *r = ctx->rings;
1495 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1498 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1500 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1503 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1505 struct io_rings *rings = ctx->rings;
1506 unsigned tail, mask = ctx->cq_entries - 1;
1509 * writes to the cq entry need to come after reading head; the
1510 * control dependency is enough as we're using WRITE_ONCE to
1513 if (__io_cqring_events(ctx) == ctx->cq_entries)
1516 tail = ctx->cached_cq_tail++;
1517 return &rings->cqes[tail & mask];
1520 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1522 if (likely(!ctx->cq_ev_fd))
1524 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1526 return !ctx->eventfd_async || io_wq_current_is_worker();
1529 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1532 * wake_up_all() may seem excessive, but io_wake_function() and
1533 * io_should_wake() handle the termination of the loop and only
1534 * wake as many waiters as we need to.
1536 if (wq_has_sleeper(&ctx->cq_wait))
1537 wake_up_all(&ctx->cq_wait);
1538 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1539 wake_up(&ctx->sq_data->wait);
1540 if (io_should_trigger_evfd(ctx))
1541 eventfd_signal(ctx->cq_ev_fd, 1);
1542 if (waitqueue_active(&ctx->poll_wait)) {
1543 wake_up_interruptible(&ctx->poll_wait);
1544 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1548 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1550 if (ctx->flags & IORING_SETUP_SQPOLL) {
1551 if (wq_has_sleeper(&ctx->cq_wait))
1552 wake_up_all(&ctx->cq_wait);
1554 if (io_should_trigger_evfd(ctx))
1555 eventfd_signal(ctx->cq_ev_fd, 1);
1556 if (waitqueue_active(&ctx->poll_wait)) {
1557 wake_up_interruptible(&ctx->poll_wait);
1558 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1562 /* Returns true if there are no backlogged entries after the flush */
1563 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1565 bool all_flushed, posted;
1567 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1571 spin_lock(&ctx->completion_lock);
1572 while (!list_empty(&ctx->cq_overflow_list)) {
1573 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1574 struct io_overflow_cqe *ocqe;
1578 ocqe = list_first_entry(&ctx->cq_overflow_list,
1579 struct io_overflow_cqe, list);
1581 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1583 io_account_cq_overflow(ctx);
1586 list_del(&ocqe->list);
1590 all_flushed = list_empty(&ctx->cq_overflow_list);
1592 clear_bit(0, &ctx->check_cq_overflow);
1593 WRITE_ONCE(ctx->rings->sq_flags,
1594 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1598 io_commit_cqring(ctx);
1599 spin_unlock(&ctx->completion_lock);
1601 io_cqring_ev_posted(ctx);
1605 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1609 if (test_bit(0, &ctx->check_cq_overflow)) {
1610 /* iopoll syncs against uring_lock, not completion_lock */
1611 if (ctx->flags & IORING_SETUP_IOPOLL)
1612 mutex_lock(&ctx->uring_lock);
1613 ret = __io_cqring_overflow_flush(ctx, false);
1614 if (ctx->flags & IORING_SETUP_IOPOLL)
1615 mutex_unlock(&ctx->uring_lock);
1621 /* must to be called somewhat shortly after putting a request */
1622 static inline void io_put_task(struct task_struct *task, int nr)
1624 struct io_uring_task *tctx = task->io_uring;
1626 percpu_counter_sub(&tctx->inflight, nr);
1627 if (unlikely(atomic_read(&tctx->in_idle)))
1628 wake_up(&tctx->wait);
1629 put_task_struct_many(task, nr);
1632 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1633 long res, unsigned int cflags)
1635 struct io_overflow_cqe *ocqe;
1637 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1640 * If we're in ring overflow flush mode, or in task cancel mode,
1641 * or cannot allocate an overflow entry, then we need to drop it
1644 io_account_cq_overflow(ctx);
1647 if (list_empty(&ctx->cq_overflow_list)) {
1648 set_bit(0, &ctx->check_cq_overflow);
1649 WRITE_ONCE(ctx->rings->sq_flags,
1650 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1653 ocqe->cqe.user_data = user_data;
1654 ocqe->cqe.res = res;
1655 ocqe->cqe.flags = cflags;
1656 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1660 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1661 long res, unsigned int cflags)
1663 struct io_uring_cqe *cqe;
1665 trace_io_uring_complete(ctx, user_data, res, cflags);
1668 * If we can't get a cq entry, userspace overflowed the
1669 * submission (by quite a lot). Increment the overflow count in
1672 cqe = io_get_cqe(ctx);
1674 WRITE_ONCE(cqe->user_data, user_data);
1675 WRITE_ONCE(cqe->res, res);
1676 WRITE_ONCE(cqe->flags, cflags);
1679 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1682 /* not as hot to bloat with inlining */
1683 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1684 long res, unsigned int cflags)
1686 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1689 static void io_req_complete_post(struct io_kiocb *req, long res,
1690 unsigned int cflags)
1692 struct io_ring_ctx *ctx = req->ctx;
1694 spin_lock(&ctx->completion_lock);
1695 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1697 * If we're the last reference to this request, add to our locked
1700 if (req_ref_put_and_test(req)) {
1701 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1702 if (req->flags & IO_DISARM_MASK)
1703 io_disarm_next(req);
1705 io_req_task_queue(req->link);
1709 io_dismantle_req(req);
1710 io_put_task(req->task, 1);
1711 list_add(&req->inflight_entry, &ctx->locked_free_list);
1712 ctx->locked_free_nr++;
1714 if (!percpu_ref_tryget(&ctx->refs))
1717 io_commit_cqring(ctx);
1718 spin_unlock(&ctx->completion_lock);
1721 io_cqring_ev_posted(ctx);
1722 percpu_ref_put(&ctx->refs);
1726 static inline bool io_req_needs_clean(struct io_kiocb *req)
1728 return req->flags & IO_REQ_CLEAN_FLAGS;
1731 static void io_req_complete_state(struct io_kiocb *req, long res,
1732 unsigned int cflags)
1734 if (io_req_needs_clean(req))
1737 req->compl.cflags = cflags;
1738 req->flags |= REQ_F_COMPLETE_INLINE;
1741 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1742 long res, unsigned cflags)
1744 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1745 io_req_complete_state(req, res, cflags);
1747 io_req_complete_post(req, res, cflags);
1750 static inline void io_req_complete(struct io_kiocb *req, long res)
1752 __io_req_complete(req, 0, res, 0);
1755 static void io_req_complete_failed(struct io_kiocb *req, long res)
1758 io_req_complete_post(req, res, 0);
1762 * Don't initialise the fields below on every allocation, but do that in
1763 * advance and keep them valid across allocations.
1765 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1769 req->async_data = NULL;
1770 /* not necessary, but safer to zero */
1774 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1775 struct io_submit_state *state)
1777 spin_lock(&ctx->completion_lock);
1778 list_splice_init(&ctx->locked_free_list, &state->free_list);
1779 ctx->locked_free_nr = 0;
1780 spin_unlock(&ctx->completion_lock);
1783 /* Returns true IFF there are requests in the cache */
1784 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1786 struct io_submit_state *state = &ctx->submit_state;
1790 * If we have more than a batch's worth of requests in our IRQ side
1791 * locked cache, grab the lock and move them over to our submission
1794 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1795 io_flush_cached_locked_reqs(ctx, state);
1797 nr = state->free_reqs;
1798 while (!list_empty(&state->free_list)) {
1799 struct io_kiocb *req = list_first_entry(&state->free_list,
1800 struct io_kiocb, inflight_entry);
1802 list_del(&req->inflight_entry);
1803 state->reqs[nr++] = req;
1804 if (nr == ARRAY_SIZE(state->reqs))
1808 state->free_reqs = nr;
1813 * A request might get retired back into the request caches even before opcode
1814 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1815 * Because of that, io_alloc_req() should be called only under ->uring_lock
1816 * and with extra caution to not get a request that is still worked on.
1818 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1819 __must_hold(&ctx->uring_lock)
1821 struct io_submit_state *state = &ctx->submit_state;
1822 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1825 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1827 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1830 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1834 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1835 * retry single alloc to be on the safe side.
1837 if (unlikely(ret <= 0)) {
1838 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1839 if (!state->reqs[0])
1844 for (i = 0; i < ret; i++)
1845 io_preinit_req(state->reqs[i], ctx);
1846 state->free_reqs = ret;
1849 return state->reqs[state->free_reqs];
1852 static inline void io_put_file(struct file *file)
1858 static void io_dismantle_req(struct io_kiocb *req)
1860 unsigned int flags = req->flags;
1862 if (io_req_needs_clean(req))
1864 if (!(flags & REQ_F_FIXED_FILE))
1865 io_put_file(req->file);
1866 if (req->fixed_rsrc_refs)
1867 percpu_ref_put(req->fixed_rsrc_refs);
1868 if (req->async_data) {
1869 kfree(req->async_data);
1870 req->async_data = NULL;
1874 static void __io_free_req(struct io_kiocb *req)
1876 struct io_ring_ctx *ctx = req->ctx;
1878 io_dismantle_req(req);
1879 io_put_task(req->task, 1);
1881 spin_lock(&ctx->completion_lock);
1882 list_add(&req->inflight_entry, &ctx->locked_free_list);
1883 ctx->locked_free_nr++;
1884 spin_unlock(&ctx->completion_lock);
1886 percpu_ref_put(&ctx->refs);
1889 static inline void io_remove_next_linked(struct io_kiocb *req)
1891 struct io_kiocb *nxt = req->link;
1893 req->link = nxt->link;
1897 static bool io_kill_linked_timeout(struct io_kiocb *req)
1898 __must_hold(&req->ctx->completion_lock)
1899 __must_hold(&req->ctx->timeout_lock)
1901 struct io_kiocb *link = req->link;
1903 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1904 struct io_timeout_data *io = link->async_data;
1906 io_remove_next_linked(req);
1907 link->timeout.head = NULL;
1908 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1909 io_cqring_fill_event(link->ctx, link->user_data,
1911 io_put_req_deferred(link);
1918 static void io_fail_links(struct io_kiocb *req)
1919 __must_hold(&req->ctx->completion_lock)
1921 struct io_kiocb *nxt, *link = req->link;
1928 trace_io_uring_fail_link(req, link);
1929 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1930 io_put_req_deferred(link);
1935 static bool io_disarm_next(struct io_kiocb *req)
1936 __must_hold(&req->ctx->completion_lock)
1938 bool posted = false;
1940 if (req->flags & REQ_F_ARM_LTIMEOUT) {
1941 struct io_kiocb *link = req->link;
1943 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1944 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1945 io_remove_next_linked(req);
1946 io_cqring_fill_event(link->ctx, link->user_data,
1948 io_put_req_deferred(link);
1951 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
1952 struct io_ring_ctx *ctx = req->ctx;
1954 spin_lock_irq(&ctx->timeout_lock);
1955 posted = io_kill_linked_timeout(req);
1956 spin_unlock_irq(&ctx->timeout_lock);
1958 if (unlikely((req->flags & REQ_F_FAIL) &&
1959 !(req->flags & REQ_F_HARDLINK))) {
1960 posted |= (req->link != NULL);
1966 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1968 struct io_kiocb *nxt;
1971 * If LINK is set, we have dependent requests in this chain. If we
1972 * didn't fail this request, queue the first one up, moving any other
1973 * dependencies to the next request. In case of failure, fail the rest
1976 if (req->flags & IO_DISARM_MASK) {
1977 struct io_ring_ctx *ctx = req->ctx;
1980 spin_lock(&ctx->completion_lock);
1981 posted = io_disarm_next(req);
1983 io_commit_cqring(req->ctx);
1984 spin_unlock(&ctx->completion_lock);
1986 io_cqring_ev_posted(ctx);
1993 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1995 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1997 return __io_req_find_next(req);
2000 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2004 if (ctx->submit_state.compl_nr) {
2005 mutex_lock(&ctx->uring_lock);
2006 io_submit_flush_completions(ctx);
2007 mutex_unlock(&ctx->uring_lock);
2009 percpu_ref_put(&ctx->refs);
2012 static void tctx_task_work(struct callback_head *cb)
2014 struct io_ring_ctx *ctx = NULL;
2015 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2019 struct io_wq_work_node *node;
2021 spin_lock_irq(&tctx->task_lock);
2022 node = tctx->task_list.first;
2023 INIT_WQ_LIST(&tctx->task_list);
2025 tctx->task_running = false;
2026 spin_unlock_irq(&tctx->task_lock);
2031 struct io_wq_work_node *next = node->next;
2032 struct io_kiocb *req = container_of(node, struct io_kiocb,
2035 if (req->ctx != ctx) {
2036 ctx_flush_and_put(ctx);
2038 percpu_ref_get(&ctx->refs);
2040 req->io_task_work.func(req);
2047 ctx_flush_and_put(ctx);
2050 static void io_req_task_work_add(struct io_kiocb *req)
2052 struct task_struct *tsk = req->task;
2053 struct io_uring_task *tctx = tsk->io_uring;
2054 enum task_work_notify_mode notify;
2055 struct io_wq_work_node *node;
2056 unsigned long flags;
2059 WARN_ON_ONCE(!tctx);
2061 spin_lock_irqsave(&tctx->task_lock, flags);
2062 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2063 running = tctx->task_running;
2065 tctx->task_running = true;
2066 spin_unlock_irqrestore(&tctx->task_lock, flags);
2068 /* task_work already pending, we're done */
2073 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2074 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2075 * processing task_work. There's no reliable way to tell if TWA_RESUME
2078 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2079 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2080 wake_up_process(tsk);
2084 spin_lock_irqsave(&tctx->task_lock, flags);
2085 tctx->task_running = false;
2086 node = tctx->task_list.first;
2087 INIT_WQ_LIST(&tctx->task_list);
2088 spin_unlock_irqrestore(&tctx->task_lock, flags);
2091 req = container_of(node, struct io_kiocb, io_task_work.node);
2093 if (llist_add(&req->io_task_work.fallback_node,
2094 &req->ctx->fallback_llist))
2095 schedule_delayed_work(&req->ctx->fallback_work, 1);
2099 static void io_req_task_cancel(struct io_kiocb *req)
2101 struct io_ring_ctx *ctx = req->ctx;
2103 /* ctx is guaranteed to stay alive while we hold uring_lock */
2104 mutex_lock(&ctx->uring_lock);
2105 io_req_complete_failed(req, req->result);
2106 mutex_unlock(&ctx->uring_lock);
2109 static void io_req_task_submit(struct io_kiocb *req)
2111 struct io_ring_ctx *ctx = req->ctx;
2113 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2114 mutex_lock(&ctx->uring_lock);
2115 /* req->task == current here, checking PF_EXITING is safe */
2116 if (likely(!(req->task->flags & PF_EXITING)))
2117 __io_queue_sqe(req);
2119 io_req_complete_failed(req, -EFAULT);
2120 mutex_unlock(&ctx->uring_lock);
2123 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2126 req->io_task_work.func = io_req_task_cancel;
2127 io_req_task_work_add(req);
2130 static void io_req_task_queue(struct io_kiocb *req)
2132 req->io_task_work.func = io_req_task_submit;
2133 io_req_task_work_add(req);
2136 static void io_req_task_queue_reissue(struct io_kiocb *req)
2138 req->io_task_work.func = io_queue_async_work;
2139 io_req_task_work_add(req);
2142 static inline void io_queue_next(struct io_kiocb *req)
2144 struct io_kiocb *nxt = io_req_find_next(req);
2147 io_req_task_queue(nxt);
2150 static void io_free_req(struct io_kiocb *req)
2157 struct task_struct *task;
2162 static inline void io_init_req_batch(struct req_batch *rb)
2169 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2170 struct req_batch *rb)
2173 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2174 if (rb->task == current)
2175 current->io_uring->cached_refs += rb->task_refs;
2177 io_put_task(rb->task, rb->task_refs);
2180 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2181 struct io_submit_state *state)
2184 io_dismantle_req(req);
2186 if (req->task != rb->task) {
2188 io_put_task(rb->task, rb->task_refs);
2189 rb->task = req->task;
2195 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2196 state->reqs[state->free_reqs++] = req;
2198 list_add(&req->inflight_entry, &state->free_list);
2201 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2202 __must_hold(&ctx->uring_lock)
2204 struct io_submit_state *state = &ctx->submit_state;
2205 int i, nr = state->compl_nr;
2206 struct req_batch rb;
2208 spin_lock(&ctx->completion_lock);
2209 for (i = 0; i < nr; i++) {
2210 struct io_kiocb *req = state->compl_reqs[i];
2212 __io_cqring_fill_event(ctx, req->user_data, req->result,
2215 io_commit_cqring(ctx);
2216 spin_unlock(&ctx->completion_lock);
2217 io_cqring_ev_posted(ctx);
2219 io_init_req_batch(&rb);
2220 for (i = 0; i < nr; i++) {
2221 struct io_kiocb *req = state->compl_reqs[i];
2223 if (req_ref_put_and_test(req))
2224 io_req_free_batch(&rb, req, &ctx->submit_state);
2227 io_req_free_batch_finish(ctx, &rb);
2228 state->compl_nr = 0;
2232 * Drop reference to request, return next in chain (if there is one) if this
2233 * was the last reference to this request.
2235 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2237 struct io_kiocb *nxt = NULL;
2239 if (req_ref_put_and_test(req)) {
2240 nxt = io_req_find_next(req);
2246 static inline void io_put_req(struct io_kiocb *req)
2248 if (req_ref_put_and_test(req))
2252 static inline void io_put_req_deferred(struct io_kiocb *req)
2254 if (req_ref_put_and_test(req)) {
2255 req->io_task_work.func = io_free_req;
2256 io_req_task_work_add(req);
2260 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2262 /* See comment at the top of this file */
2264 return __io_cqring_events(ctx);
2267 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2269 struct io_rings *rings = ctx->rings;
2271 /* make sure SQ entry isn't read before tail */
2272 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2275 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2277 unsigned int cflags;
2279 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2280 cflags |= IORING_CQE_F_BUFFER;
2281 req->flags &= ~REQ_F_BUFFER_SELECTED;
2286 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2288 struct io_buffer *kbuf;
2290 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2292 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2293 return io_put_kbuf(req, kbuf);
2296 static inline bool io_run_task_work(void)
2298 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2299 __set_current_state(TASK_RUNNING);
2300 tracehook_notify_signal();
2308 * Find and free completed poll iocbs
2310 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2311 struct list_head *done)
2313 struct req_batch rb;
2314 struct io_kiocb *req;
2316 /* order with ->result store in io_complete_rw_iopoll() */
2319 io_init_req_batch(&rb);
2320 while (!list_empty(done)) {
2321 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2322 list_del(&req->inflight_entry);
2324 if (READ_ONCE(req->result) == -EAGAIN &&
2325 !(req->flags & REQ_F_DONT_REISSUE)) {
2326 req->iopoll_completed = 0;
2327 io_req_task_queue_reissue(req);
2331 __io_cqring_fill_event(ctx, req->user_data, req->result,
2332 io_put_rw_kbuf(req));
2335 if (req_ref_put_and_test(req))
2336 io_req_free_batch(&rb, req, &ctx->submit_state);
2339 io_commit_cqring(ctx);
2340 io_cqring_ev_posted_iopoll(ctx);
2341 io_req_free_batch_finish(ctx, &rb);
2344 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2347 struct io_kiocb *req, *tmp;
2352 * Only spin for completions if we don't have multiple devices hanging
2353 * off our complete list, and we're under the requested amount.
2355 spin = !ctx->poll_multi_queue && *nr_events < min;
2357 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2358 struct kiocb *kiocb = &req->rw.kiocb;
2362 * Move completed and retryable entries to our local lists.
2363 * If we find a request that requires polling, break out
2364 * and complete those lists first, if we have entries there.
2366 if (READ_ONCE(req->iopoll_completed)) {
2367 list_move_tail(&req->inflight_entry, &done);
2370 if (!list_empty(&done))
2373 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2374 if (unlikely(ret < 0))
2379 /* iopoll may have completed current req */
2380 if (READ_ONCE(req->iopoll_completed))
2381 list_move_tail(&req->inflight_entry, &done);
2384 if (!list_empty(&done))
2385 io_iopoll_complete(ctx, nr_events, &done);
2391 * We can't just wait for polled events to come to us, we have to actively
2392 * find and complete them.
2394 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2396 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2399 mutex_lock(&ctx->uring_lock);
2400 while (!list_empty(&ctx->iopoll_list)) {
2401 unsigned int nr_events = 0;
2403 io_do_iopoll(ctx, &nr_events, 0);
2405 /* let it sleep and repeat later if can't complete a request */
2409 * Ensure we allow local-to-the-cpu processing to take place,
2410 * in this case we need to ensure that we reap all events.
2411 * Also let task_work, etc. to progress by releasing the mutex
2413 if (need_resched()) {
2414 mutex_unlock(&ctx->uring_lock);
2416 mutex_lock(&ctx->uring_lock);
2419 mutex_unlock(&ctx->uring_lock);
2422 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2424 unsigned int nr_events = 0;
2428 * We disallow the app entering submit/complete with polling, but we
2429 * still need to lock the ring to prevent racing with polled issue
2430 * that got punted to a workqueue.
2432 mutex_lock(&ctx->uring_lock);
2434 * Don't enter poll loop if we already have events pending.
2435 * If we do, we can potentially be spinning for commands that
2436 * already triggered a CQE (eg in error).
2438 if (test_bit(0, &ctx->check_cq_overflow))
2439 __io_cqring_overflow_flush(ctx, false);
2440 if (io_cqring_events(ctx))
2444 * If a submit got punted to a workqueue, we can have the
2445 * application entering polling for a command before it gets
2446 * issued. That app will hold the uring_lock for the duration
2447 * of the poll right here, so we need to take a breather every
2448 * now and then to ensure that the issue has a chance to add
2449 * the poll to the issued list. Otherwise we can spin here
2450 * forever, while the workqueue is stuck trying to acquire the
2453 if (list_empty(&ctx->iopoll_list)) {
2454 u32 tail = ctx->cached_cq_tail;
2456 mutex_unlock(&ctx->uring_lock);
2458 mutex_lock(&ctx->uring_lock);
2460 /* some requests don't go through iopoll_list */
2461 if (tail != ctx->cached_cq_tail ||
2462 list_empty(&ctx->iopoll_list))
2465 ret = io_do_iopoll(ctx, &nr_events, min);
2466 } while (!ret && nr_events < min && !need_resched());
2468 mutex_unlock(&ctx->uring_lock);
2472 static void kiocb_end_write(struct io_kiocb *req)
2475 * Tell lockdep we inherited freeze protection from submission
2478 if (req->flags & REQ_F_ISREG) {
2479 struct super_block *sb = file_inode(req->file)->i_sb;
2481 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2487 static bool io_resubmit_prep(struct io_kiocb *req)
2489 struct io_async_rw *rw = req->async_data;
2492 return !io_req_prep_async(req);
2493 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2494 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2498 static bool io_rw_should_reissue(struct io_kiocb *req)
2500 umode_t mode = file_inode(req->file)->i_mode;
2501 struct io_ring_ctx *ctx = req->ctx;
2503 if (!S_ISBLK(mode) && !S_ISREG(mode))
2505 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2506 !(ctx->flags & IORING_SETUP_IOPOLL)))
2509 * If ref is dying, we might be running poll reap from the exit work.
2510 * Don't attempt to reissue from that path, just let it fail with
2513 if (percpu_ref_is_dying(&ctx->refs))
2516 * Play it safe and assume not safe to re-import and reissue if we're
2517 * not in the original thread group (or in task context).
2519 if (!same_thread_group(req->task, current) || !in_task())
2524 static bool io_resubmit_prep(struct io_kiocb *req)
2528 static bool io_rw_should_reissue(struct io_kiocb *req)
2534 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2536 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2537 kiocb_end_write(req);
2538 if (res != req->result) {
2539 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2540 io_rw_should_reissue(req)) {
2541 req->flags |= REQ_F_REISSUE;
2550 static void io_req_task_complete(struct io_kiocb *req)
2552 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2555 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2556 unsigned int issue_flags)
2558 if (__io_complete_rw_common(req, res))
2560 io_req_task_complete(req);
2563 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2565 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2567 if (__io_complete_rw_common(req, res))
2570 req->io_task_work.func = io_req_task_complete;
2571 io_req_task_work_add(req);
2574 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2576 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2578 if (kiocb->ki_flags & IOCB_WRITE)
2579 kiocb_end_write(req);
2580 if (unlikely(res != req->result)) {
2581 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2582 io_resubmit_prep(req))) {
2584 req->flags |= REQ_F_DONT_REISSUE;
2588 WRITE_ONCE(req->result, res);
2589 /* order with io_iopoll_complete() checking ->result */
2591 WRITE_ONCE(req->iopoll_completed, 1);
2595 * After the iocb has been issued, it's safe to be found on the poll list.
2596 * Adding the kiocb to the list AFTER submission ensures that we don't
2597 * find it from a io_do_iopoll() thread before the issuer is done
2598 * accessing the kiocb cookie.
2600 static void io_iopoll_req_issued(struct io_kiocb *req)
2602 struct io_ring_ctx *ctx = req->ctx;
2603 const bool in_async = io_wq_current_is_worker();
2605 /* workqueue context doesn't hold uring_lock, grab it now */
2606 if (unlikely(in_async))
2607 mutex_lock(&ctx->uring_lock);
2610 * Track whether we have multiple files in our lists. This will impact
2611 * how we do polling eventually, not spinning if we're on potentially
2612 * different devices.
2614 if (list_empty(&ctx->iopoll_list)) {
2615 ctx->poll_multi_queue = false;
2616 } else if (!ctx->poll_multi_queue) {
2617 struct io_kiocb *list_req;
2618 unsigned int queue_num0, queue_num1;
2620 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2623 if (list_req->file != req->file) {
2624 ctx->poll_multi_queue = true;
2626 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2627 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2628 if (queue_num0 != queue_num1)
2629 ctx->poll_multi_queue = true;
2634 * For fast devices, IO may have already completed. If it has, add
2635 * it to the front so we find it first.
2637 if (READ_ONCE(req->iopoll_completed))
2638 list_add(&req->inflight_entry, &ctx->iopoll_list);
2640 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2642 if (unlikely(in_async)) {
2644 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2645 * in sq thread task context or in io worker task context. If
2646 * current task context is sq thread, we don't need to check
2647 * whether should wake up sq thread.
2649 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2650 wq_has_sleeper(&ctx->sq_data->wait))
2651 wake_up(&ctx->sq_data->wait);
2653 mutex_unlock(&ctx->uring_lock);
2657 static bool io_bdev_nowait(struct block_device *bdev)
2659 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2663 * If we tracked the file through the SCM inflight mechanism, we could support
2664 * any file. For now, just ensure that anything potentially problematic is done
2667 static bool __io_file_supports_nowait(struct file *file, int rw)
2669 umode_t mode = file_inode(file)->i_mode;
2671 if (S_ISBLK(mode)) {
2672 if (IS_ENABLED(CONFIG_BLOCK) &&
2673 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2679 if (S_ISREG(mode)) {
2680 if (IS_ENABLED(CONFIG_BLOCK) &&
2681 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2682 file->f_op != &io_uring_fops)
2687 /* any ->read/write should understand O_NONBLOCK */
2688 if (file->f_flags & O_NONBLOCK)
2691 if (!(file->f_mode & FMODE_NOWAIT))
2695 return file->f_op->read_iter != NULL;
2697 return file->f_op->write_iter != NULL;
2700 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2702 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2704 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2707 return __io_file_supports_nowait(req->file, rw);
2710 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2712 struct io_ring_ctx *ctx = req->ctx;
2713 struct kiocb *kiocb = &req->rw.kiocb;
2714 struct file *file = req->file;
2718 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2719 req->flags |= REQ_F_ISREG;
2721 kiocb->ki_pos = READ_ONCE(sqe->off);
2722 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2723 req->flags |= REQ_F_CUR_POS;
2724 kiocb->ki_pos = file->f_pos;
2726 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2727 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2728 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2732 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2733 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2734 req->flags |= REQ_F_NOWAIT;
2736 ioprio = READ_ONCE(sqe->ioprio);
2738 ret = ioprio_check_cap(ioprio);
2742 kiocb->ki_ioprio = ioprio;
2744 kiocb->ki_ioprio = get_current_ioprio();
2746 if (ctx->flags & IORING_SETUP_IOPOLL) {
2747 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2748 !kiocb->ki_filp->f_op->iopoll)
2751 kiocb->ki_flags |= IOCB_HIPRI;
2752 kiocb->ki_complete = io_complete_rw_iopoll;
2753 req->iopoll_completed = 0;
2755 if (kiocb->ki_flags & IOCB_HIPRI)
2757 kiocb->ki_complete = io_complete_rw;
2760 if (req->opcode == IORING_OP_READ_FIXED ||
2761 req->opcode == IORING_OP_WRITE_FIXED) {
2763 io_req_set_rsrc_node(req);
2766 req->rw.addr = READ_ONCE(sqe->addr);
2767 req->rw.len = READ_ONCE(sqe->len);
2768 req->buf_index = READ_ONCE(sqe->buf_index);
2772 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2778 case -ERESTARTNOINTR:
2779 case -ERESTARTNOHAND:
2780 case -ERESTART_RESTARTBLOCK:
2782 * We can't just restart the syscall, since previously
2783 * submitted sqes may already be in progress. Just fail this
2789 kiocb->ki_complete(kiocb, ret, 0);
2793 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2794 unsigned int issue_flags)
2796 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2797 struct io_async_rw *io = req->async_data;
2798 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2800 /* add previously done IO, if any */
2801 if (io && io->bytes_done > 0) {
2803 ret = io->bytes_done;
2805 ret += io->bytes_done;
2808 if (req->flags & REQ_F_CUR_POS)
2809 req->file->f_pos = kiocb->ki_pos;
2810 if (ret >= 0 && check_reissue)
2811 __io_complete_rw(req, ret, 0, issue_flags);
2813 io_rw_done(kiocb, ret);
2815 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2816 req->flags &= ~REQ_F_REISSUE;
2817 if (io_resubmit_prep(req)) {
2818 io_req_task_queue_reissue(req);
2821 __io_req_complete(req, issue_flags, ret,
2822 io_put_rw_kbuf(req));
2827 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2828 struct io_mapped_ubuf *imu)
2830 size_t len = req->rw.len;
2831 u64 buf_end, buf_addr = req->rw.addr;
2834 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2836 /* not inside the mapped region */
2837 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2841 * May not be a start of buffer, set size appropriately
2842 * and advance us to the beginning.
2844 offset = buf_addr - imu->ubuf;
2845 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2849 * Don't use iov_iter_advance() here, as it's really slow for
2850 * using the latter parts of a big fixed buffer - it iterates
2851 * over each segment manually. We can cheat a bit here, because
2854 * 1) it's a BVEC iter, we set it up
2855 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2856 * first and last bvec
2858 * So just find our index, and adjust the iterator afterwards.
2859 * If the offset is within the first bvec (or the whole first
2860 * bvec, just use iov_iter_advance(). This makes it easier
2861 * since we can just skip the first segment, which may not
2862 * be PAGE_SIZE aligned.
2864 const struct bio_vec *bvec = imu->bvec;
2866 if (offset <= bvec->bv_len) {
2867 iov_iter_advance(iter, offset);
2869 unsigned long seg_skip;
2871 /* skip first vec */
2872 offset -= bvec->bv_len;
2873 seg_skip = 1 + (offset >> PAGE_SHIFT);
2875 iter->bvec = bvec + seg_skip;
2876 iter->nr_segs -= seg_skip;
2877 iter->count -= bvec->bv_len + offset;
2878 iter->iov_offset = offset & ~PAGE_MASK;
2885 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2887 struct io_ring_ctx *ctx = req->ctx;
2888 struct io_mapped_ubuf *imu = req->imu;
2889 u16 index, buf_index = req->buf_index;
2892 if (unlikely(buf_index >= ctx->nr_user_bufs))
2894 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2895 imu = READ_ONCE(ctx->user_bufs[index]);
2898 return __io_import_fixed(req, rw, iter, imu);
2901 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2904 mutex_unlock(&ctx->uring_lock);
2907 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2910 * "Normal" inline submissions always hold the uring_lock, since we
2911 * grab it from the system call. Same is true for the SQPOLL offload.
2912 * The only exception is when we've detached the request and issue it
2913 * from an async worker thread, grab the lock for that case.
2916 mutex_lock(&ctx->uring_lock);
2919 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2920 int bgid, struct io_buffer *kbuf,
2923 struct io_buffer *head;
2925 if (req->flags & REQ_F_BUFFER_SELECTED)
2928 io_ring_submit_lock(req->ctx, needs_lock);
2930 lockdep_assert_held(&req->ctx->uring_lock);
2932 head = xa_load(&req->ctx->io_buffers, bgid);
2934 if (!list_empty(&head->list)) {
2935 kbuf = list_last_entry(&head->list, struct io_buffer,
2937 list_del(&kbuf->list);
2940 xa_erase(&req->ctx->io_buffers, bgid);
2942 if (*len > kbuf->len)
2945 kbuf = ERR_PTR(-ENOBUFS);
2948 io_ring_submit_unlock(req->ctx, needs_lock);
2953 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2956 struct io_buffer *kbuf;
2959 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2960 bgid = req->buf_index;
2961 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2964 req->rw.addr = (u64) (unsigned long) kbuf;
2965 req->flags |= REQ_F_BUFFER_SELECTED;
2966 return u64_to_user_ptr(kbuf->addr);
2969 #ifdef CONFIG_COMPAT
2970 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2973 struct compat_iovec __user *uiov;
2974 compat_ssize_t clen;
2978 uiov = u64_to_user_ptr(req->rw.addr);
2979 if (!access_ok(uiov, sizeof(*uiov)))
2981 if (__get_user(clen, &uiov->iov_len))
2987 buf = io_rw_buffer_select(req, &len, needs_lock);
2989 return PTR_ERR(buf);
2990 iov[0].iov_base = buf;
2991 iov[0].iov_len = (compat_size_t) len;
2996 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2999 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3003 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3006 len = iov[0].iov_len;
3009 buf = io_rw_buffer_select(req, &len, needs_lock);
3011 return PTR_ERR(buf);
3012 iov[0].iov_base = buf;
3013 iov[0].iov_len = len;
3017 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3020 if (req->flags & REQ_F_BUFFER_SELECTED) {
3021 struct io_buffer *kbuf;
3023 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3024 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3025 iov[0].iov_len = kbuf->len;
3028 if (req->rw.len != 1)
3031 #ifdef CONFIG_COMPAT
3032 if (req->ctx->compat)
3033 return io_compat_import(req, iov, needs_lock);
3036 return __io_iov_buffer_select(req, iov, needs_lock);
3039 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3040 struct iov_iter *iter, bool needs_lock)
3042 void __user *buf = u64_to_user_ptr(req->rw.addr);
3043 size_t sqe_len = req->rw.len;
3044 u8 opcode = req->opcode;
3047 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3049 return io_import_fixed(req, rw, iter);
3052 /* buffer index only valid with fixed read/write, or buffer select */
3053 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3056 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3057 if (req->flags & REQ_F_BUFFER_SELECT) {
3058 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3060 return PTR_ERR(buf);
3061 req->rw.len = sqe_len;
3064 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3069 if (req->flags & REQ_F_BUFFER_SELECT) {
3070 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3072 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3077 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3081 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3083 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3087 * For files that don't have ->read_iter() and ->write_iter(), handle them
3088 * by looping over ->read() or ->write() manually.
3090 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3092 struct kiocb *kiocb = &req->rw.kiocb;
3093 struct file *file = req->file;
3097 * Don't support polled IO through this interface, and we can't
3098 * support non-blocking either. For the latter, this just causes
3099 * the kiocb to be handled from an async context.
3101 if (kiocb->ki_flags & IOCB_HIPRI)
3103 if (kiocb->ki_flags & IOCB_NOWAIT)
3106 while (iov_iter_count(iter)) {
3110 if (!iov_iter_is_bvec(iter)) {
3111 iovec = iov_iter_iovec(iter);
3113 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3114 iovec.iov_len = req->rw.len;
3118 nr = file->f_op->read(file, iovec.iov_base,
3119 iovec.iov_len, io_kiocb_ppos(kiocb));
3121 nr = file->f_op->write(file, iovec.iov_base,
3122 iovec.iov_len, io_kiocb_ppos(kiocb));
3131 if (nr != iovec.iov_len)
3135 iov_iter_advance(iter, nr);
3141 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3142 const struct iovec *fast_iov, struct iov_iter *iter)
3144 struct io_async_rw *rw = req->async_data;
3146 memcpy(&rw->iter, iter, sizeof(*iter));
3147 rw->free_iovec = iovec;
3149 /* can only be fixed buffers, no need to do anything */
3150 if (iov_iter_is_bvec(iter))
3153 unsigned iov_off = 0;
3155 rw->iter.iov = rw->fast_iov;
3156 if (iter->iov != fast_iov) {
3157 iov_off = iter->iov - fast_iov;
3158 rw->iter.iov += iov_off;
3160 if (rw->fast_iov != fast_iov)
3161 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3162 sizeof(struct iovec) * iter->nr_segs);
3164 req->flags |= REQ_F_NEED_CLEANUP;
3168 static inline int io_alloc_async_data(struct io_kiocb *req)
3170 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3171 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3172 return req->async_data == NULL;
3175 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3176 const struct iovec *fast_iov,
3177 struct iov_iter *iter, bool force)
3179 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3181 if (!req->async_data) {
3182 if (io_alloc_async_data(req)) {
3187 io_req_map_rw(req, iovec, fast_iov, iter);
3192 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3194 struct io_async_rw *iorw = req->async_data;
3195 struct iovec *iov = iorw->fast_iov;
3198 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3199 if (unlikely(ret < 0))
3202 iorw->bytes_done = 0;
3203 iorw->free_iovec = iov;
3205 req->flags |= REQ_F_NEED_CLEANUP;
3209 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3211 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3213 return io_prep_rw(req, sqe);
3217 * This is our waitqueue callback handler, registered through lock_page_async()
3218 * when we initially tried to do the IO with the iocb armed our waitqueue.
3219 * This gets called when the page is unlocked, and we generally expect that to
3220 * happen when the page IO is completed and the page is now uptodate. This will
3221 * queue a task_work based retry of the operation, attempting to copy the data
3222 * again. If the latter fails because the page was NOT uptodate, then we will
3223 * do a thread based blocking retry of the operation. That's the unexpected
3226 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3227 int sync, void *arg)
3229 struct wait_page_queue *wpq;
3230 struct io_kiocb *req = wait->private;
3231 struct wait_page_key *key = arg;
3233 wpq = container_of(wait, struct wait_page_queue, wait);
3235 if (!wake_page_match(wpq, key))
3238 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3239 list_del_init(&wait->entry);
3240 io_req_task_queue(req);
3245 * This controls whether a given IO request should be armed for async page
3246 * based retry. If we return false here, the request is handed to the async
3247 * worker threads for retry. If we're doing buffered reads on a regular file,
3248 * we prepare a private wait_page_queue entry and retry the operation. This
3249 * will either succeed because the page is now uptodate and unlocked, or it
3250 * will register a callback when the page is unlocked at IO completion. Through
3251 * that callback, io_uring uses task_work to setup a retry of the operation.
3252 * That retry will attempt the buffered read again. The retry will generally
3253 * succeed, or in rare cases where it fails, we then fall back to using the
3254 * async worker threads for a blocking retry.
3256 static bool io_rw_should_retry(struct io_kiocb *req)
3258 struct io_async_rw *rw = req->async_data;
3259 struct wait_page_queue *wait = &rw->wpq;
3260 struct kiocb *kiocb = &req->rw.kiocb;
3262 /* never retry for NOWAIT, we just complete with -EAGAIN */
3263 if (req->flags & REQ_F_NOWAIT)
3266 /* Only for buffered IO */
3267 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3271 * just use poll if we can, and don't attempt if the fs doesn't
3272 * support callback based unlocks
3274 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3277 wait->wait.func = io_async_buf_func;
3278 wait->wait.private = req;
3279 wait->wait.flags = 0;
3280 INIT_LIST_HEAD(&wait->wait.entry);
3281 kiocb->ki_flags |= IOCB_WAITQ;
3282 kiocb->ki_flags &= ~IOCB_NOWAIT;
3283 kiocb->ki_waitq = wait;
3287 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3289 if (req->file->f_op->read_iter)
3290 return call_read_iter(req->file, &req->rw.kiocb, iter);
3291 else if (req->file->f_op->read)
3292 return loop_rw_iter(READ, req, iter);
3297 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3299 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3300 struct kiocb *kiocb = &req->rw.kiocb;
3301 struct iov_iter __iter, *iter = &__iter;
3302 struct io_async_rw *rw = req->async_data;
3303 ssize_t io_size, ret, ret2;
3304 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3310 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3314 io_size = iov_iter_count(iter);
3315 req->result = io_size;
3317 /* Ensure we clear previously set non-block flag */
3318 if (!force_nonblock)
3319 kiocb->ki_flags &= ~IOCB_NOWAIT;
3321 kiocb->ki_flags |= IOCB_NOWAIT;
3323 /* If the file doesn't support async, just async punt */
3324 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3325 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3326 return ret ?: -EAGAIN;
3329 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3330 if (unlikely(ret)) {
3335 ret = io_iter_do_read(req, iter);
3337 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3338 req->flags &= ~REQ_F_REISSUE;
3339 /* IOPOLL retry should happen for io-wq threads */
3340 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3342 /* no retry on NONBLOCK nor RWF_NOWAIT */
3343 if (req->flags & REQ_F_NOWAIT)
3345 /* some cases will consume bytes even on error returns */
3346 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3348 } else if (ret == -EIOCBQUEUED) {
3350 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3351 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3352 /* read all, failed, already did sync or don't want to retry */
3356 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3361 rw = req->async_data;
3362 /* now use our persistent iterator, if we aren't already */
3367 rw->bytes_done += ret;
3368 /* if we can retry, do so with the callbacks armed */
3369 if (!io_rw_should_retry(req)) {
3370 kiocb->ki_flags &= ~IOCB_WAITQ;
3375 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3376 * we get -EIOCBQUEUED, then we'll get a notification when the
3377 * desired page gets unlocked. We can also get a partial read
3378 * here, and if we do, then just retry at the new offset.
3380 ret = io_iter_do_read(req, iter);
3381 if (ret == -EIOCBQUEUED)
3383 /* we got some bytes, but not all. retry. */
3384 kiocb->ki_flags &= ~IOCB_WAITQ;
3385 } while (ret > 0 && ret < io_size);
3387 kiocb_done(kiocb, ret, issue_flags);
3389 /* it's faster to check here then delegate to kfree */
3395 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3397 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3399 return io_prep_rw(req, sqe);
3402 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3404 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3405 struct kiocb *kiocb = &req->rw.kiocb;
3406 struct iov_iter __iter, *iter = &__iter;
3407 struct io_async_rw *rw = req->async_data;
3408 ssize_t ret, ret2, io_size;
3409 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3415 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3419 io_size = iov_iter_count(iter);
3420 req->result = io_size;
3422 /* Ensure we clear previously set non-block flag */
3423 if (!force_nonblock)
3424 kiocb->ki_flags &= ~IOCB_NOWAIT;
3426 kiocb->ki_flags |= IOCB_NOWAIT;
3428 /* If the file doesn't support async, just async punt */
3429 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3432 /* file path doesn't support NOWAIT for non-direct_IO */
3433 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3434 (req->flags & REQ_F_ISREG))
3437 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3442 * Open-code file_start_write here to grab freeze protection,
3443 * which will be released by another thread in
3444 * io_complete_rw(). Fool lockdep by telling it the lock got
3445 * released so that it doesn't complain about the held lock when
3446 * we return to userspace.
3448 if (req->flags & REQ_F_ISREG) {
3449 sb_start_write(file_inode(req->file)->i_sb);
3450 __sb_writers_release(file_inode(req->file)->i_sb,
3453 kiocb->ki_flags |= IOCB_WRITE;
3455 if (req->file->f_op->write_iter)
3456 ret2 = call_write_iter(req->file, kiocb, iter);
3457 else if (req->file->f_op->write)
3458 ret2 = loop_rw_iter(WRITE, req, iter);
3462 if (req->flags & REQ_F_REISSUE) {
3463 req->flags &= ~REQ_F_REISSUE;
3468 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3469 * retry them without IOCB_NOWAIT.
3471 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3473 /* no retry on NONBLOCK nor RWF_NOWAIT */
3474 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3476 if (!force_nonblock || ret2 != -EAGAIN) {
3477 /* IOPOLL retry should happen for io-wq threads */
3478 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3481 kiocb_done(kiocb, ret2, issue_flags);
3484 /* some cases will consume bytes even on error returns */
3485 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3486 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3487 return ret ?: -EAGAIN;
3490 /* it's reportedly faster than delegating the null check to kfree() */
3496 static int io_renameat_prep(struct io_kiocb *req,
3497 const struct io_uring_sqe *sqe)
3499 struct io_rename *ren = &req->rename;
3500 const char __user *oldf, *newf;
3502 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3504 if (sqe->ioprio || sqe->buf_index)
3506 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3509 ren->old_dfd = READ_ONCE(sqe->fd);
3510 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3511 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3512 ren->new_dfd = READ_ONCE(sqe->len);
3513 ren->flags = READ_ONCE(sqe->rename_flags);
3515 ren->oldpath = getname(oldf);
3516 if (IS_ERR(ren->oldpath))
3517 return PTR_ERR(ren->oldpath);
3519 ren->newpath = getname(newf);
3520 if (IS_ERR(ren->newpath)) {
3521 putname(ren->oldpath);
3522 return PTR_ERR(ren->newpath);
3525 req->flags |= REQ_F_NEED_CLEANUP;
3529 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3531 struct io_rename *ren = &req->rename;
3534 if (issue_flags & IO_URING_F_NONBLOCK)
3537 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3538 ren->newpath, ren->flags);
3540 req->flags &= ~REQ_F_NEED_CLEANUP;
3543 io_req_complete(req, ret);
3547 static int io_unlinkat_prep(struct io_kiocb *req,
3548 const struct io_uring_sqe *sqe)
3550 struct io_unlink *un = &req->unlink;
3551 const char __user *fname;
3553 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3555 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3557 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3560 un->dfd = READ_ONCE(sqe->fd);
3562 un->flags = READ_ONCE(sqe->unlink_flags);
3563 if (un->flags & ~AT_REMOVEDIR)
3566 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3567 un->filename = getname(fname);
3568 if (IS_ERR(un->filename))
3569 return PTR_ERR(un->filename);
3571 req->flags |= REQ_F_NEED_CLEANUP;
3575 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3577 struct io_unlink *un = &req->unlink;
3580 if (issue_flags & IO_URING_F_NONBLOCK)
3583 if (un->flags & AT_REMOVEDIR)
3584 ret = do_rmdir(un->dfd, un->filename);
3586 ret = do_unlinkat(un->dfd, un->filename);
3588 req->flags &= ~REQ_F_NEED_CLEANUP;
3591 io_req_complete(req, ret);
3595 static int io_shutdown_prep(struct io_kiocb *req,
3596 const struct io_uring_sqe *sqe)
3598 #if defined(CONFIG_NET)
3599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3601 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3605 req->shutdown.how = READ_ONCE(sqe->len);
3612 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3614 #if defined(CONFIG_NET)
3615 struct socket *sock;
3618 if (issue_flags & IO_URING_F_NONBLOCK)
3621 sock = sock_from_file(req->file);
3622 if (unlikely(!sock))
3625 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3628 io_req_complete(req, ret);
3635 static int __io_splice_prep(struct io_kiocb *req,
3636 const struct io_uring_sqe *sqe)
3638 struct io_splice *sp = &req->splice;
3639 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3645 sp->len = READ_ONCE(sqe->len);
3646 sp->flags = READ_ONCE(sqe->splice_flags);
3648 if (unlikely(sp->flags & ~valid_flags))
3651 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3652 (sp->flags & SPLICE_F_FD_IN_FIXED));
3655 req->flags |= REQ_F_NEED_CLEANUP;
3659 static int io_tee_prep(struct io_kiocb *req,
3660 const struct io_uring_sqe *sqe)
3662 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3664 return __io_splice_prep(req, sqe);
3667 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3669 struct io_splice *sp = &req->splice;
3670 struct file *in = sp->file_in;
3671 struct file *out = sp->file_out;
3672 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3675 if (issue_flags & IO_URING_F_NONBLOCK)
3678 ret = do_tee(in, out, sp->len, flags);
3680 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3682 req->flags &= ~REQ_F_NEED_CLEANUP;
3686 io_req_complete(req, ret);
3690 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3692 struct io_splice *sp = &req->splice;
3694 sp->off_in = READ_ONCE(sqe->splice_off_in);
3695 sp->off_out = READ_ONCE(sqe->off);
3696 return __io_splice_prep(req, sqe);
3699 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3701 struct io_splice *sp = &req->splice;
3702 struct file *in = sp->file_in;
3703 struct file *out = sp->file_out;
3704 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3705 loff_t *poff_in, *poff_out;
3708 if (issue_flags & IO_URING_F_NONBLOCK)
3711 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3712 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3715 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3717 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3719 req->flags &= ~REQ_F_NEED_CLEANUP;
3723 io_req_complete(req, ret);
3728 * IORING_OP_NOP just posts a completion event, nothing else.
3730 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3732 struct io_ring_ctx *ctx = req->ctx;
3734 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3737 __io_req_complete(req, issue_flags, 0, 0);
3741 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3743 struct io_ring_ctx *ctx = req->ctx;
3748 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3750 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3753 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3754 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3757 req->sync.off = READ_ONCE(sqe->off);
3758 req->sync.len = READ_ONCE(sqe->len);
3762 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3764 loff_t end = req->sync.off + req->sync.len;
3767 /* fsync always requires a blocking context */
3768 if (issue_flags & IO_URING_F_NONBLOCK)
3771 ret = vfs_fsync_range(req->file, req->sync.off,
3772 end > 0 ? end : LLONG_MAX,
3773 req->sync.flags & IORING_FSYNC_DATASYNC);
3776 io_req_complete(req, ret);
3780 static int io_fallocate_prep(struct io_kiocb *req,
3781 const struct io_uring_sqe *sqe)
3783 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3785 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3788 req->sync.off = READ_ONCE(sqe->off);
3789 req->sync.len = READ_ONCE(sqe->addr);
3790 req->sync.mode = READ_ONCE(sqe->len);
3794 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3798 /* fallocate always requiring blocking context */
3799 if (issue_flags & IO_URING_F_NONBLOCK)
3801 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3805 io_req_complete(req, ret);
3809 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3811 const char __user *fname;
3814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 if (unlikely(sqe->ioprio || sqe->buf_index))
3818 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3821 /* open.how should be already initialised */
3822 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3823 req->open.how.flags |= O_LARGEFILE;
3825 req->open.dfd = READ_ONCE(sqe->fd);
3826 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3827 req->open.filename = getname(fname);
3828 if (IS_ERR(req->open.filename)) {
3829 ret = PTR_ERR(req->open.filename);
3830 req->open.filename = NULL;
3833 req->open.nofile = rlimit(RLIMIT_NOFILE);
3834 req->flags |= REQ_F_NEED_CLEANUP;
3838 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3840 u64 mode = READ_ONCE(sqe->len);
3841 u64 flags = READ_ONCE(sqe->open_flags);
3843 req->open.how = build_open_how(flags, mode);
3844 return __io_openat_prep(req, sqe);
3847 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3849 struct open_how __user *how;
3853 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3854 len = READ_ONCE(sqe->len);
3855 if (len < OPEN_HOW_SIZE_VER0)
3858 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3863 return __io_openat_prep(req, sqe);
3866 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3868 struct open_flags op;
3871 bool resolve_nonblock;
3874 ret = build_open_flags(&req->open.how, &op);
3877 nonblock_set = op.open_flag & O_NONBLOCK;
3878 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3879 if (issue_flags & IO_URING_F_NONBLOCK) {
3881 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3882 * it'll always -EAGAIN
3884 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3886 op.lookup_flags |= LOOKUP_CACHED;
3887 op.open_flag |= O_NONBLOCK;
3890 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3894 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3897 * We could hang on to this 'fd' on retrying, but seems like
3898 * marginal gain for something that is now known to be a slower
3899 * path. So just put it, and we'll get a new one when we retry.
3903 ret = PTR_ERR(file);
3904 /* only retry if RESOLVE_CACHED wasn't already set by application */
3905 if (ret == -EAGAIN &&
3906 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3911 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3912 file->f_flags &= ~O_NONBLOCK;
3913 fsnotify_open(file);
3914 fd_install(ret, file);
3916 putname(req->open.filename);
3917 req->flags &= ~REQ_F_NEED_CLEANUP;
3920 __io_req_complete(req, issue_flags, ret, 0);
3924 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3926 return io_openat2(req, issue_flags);
3929 static int io_remove_buffers_prep(struct io_kiocb *req,
3930 const struct io_uring_sqe *sqe)
3932 struct io_provide_buf *p = &req->pbuf;
3935 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3938 tmp = READ_ONCE(sqe->fd);
3939 if (!tmp || tmp > USHRT_MAX)
3942 memset(p, 0, sizeof(*p));
3944 p->bgid = READ_ONCE(sqe->buf_group);
3948 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3949 int bgid, unsigned nbufs)
3953 /* shouldn't happen */
3957 /* the head kbuf is the list itself */
3958 while (!list_empty(&buf->list)) {
3959 struct io_buffer *nxt;
3961 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3962 list_del(&nxt->list);
3969 xa_erase(&ctx->io_buffers, bgid);
3974 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3976 struct io_provide_buf *p = &req->pbuf;
3977 struct io_ring_ctx *ctx = req->ctx;
3978 struct io_buffer *head;
3980 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3982 io_ring_submit_lock(ctx, !force_nonblock);
3984 lockdep_assert_held(&ctx->uring_lock);
3987 head = xa_load(&ctx->io_buffers, p->bgid);
3989 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3993 /* complete before unlock, IOPOLL may need the lock */
3994 __io_req_complete(req, issue_flags, ret, 0);
3995 io_ring_submit_unlock(ctx, !force_nonblock);
3999 static int io_provide_buffers_prep(struct io_kiocb *req,
4000 const struct io_uring_sqe *sqe)
4002 unsigned long size, tmp_check;
4003 struct io_provide_buf *p = &req->pbuf;
4006 if (sqe->ioprio || sqe->rw_flags)
4009 tmp = READ_ONCE(sqe->fd);
4010 if (!tmp || tmp > USHRT_MAX)
4013 p->addr = READ_ONCE(sqe->addr);
4014 p->len = READ_ONCE(sqe->len);
4016 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4019 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4022 size = (unsigned long)p->len * p->nbufs;
4023 if (!access_ok(u64_to_user_ptr(p->addr), size))
4026 p->bgid = READ_ONCE(sqe->buf_group);
4027 tmp = READ_ONCE(sqe->off);
4028 if (tmp > USHRT_MAX)
4034 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4036 struct io_buffer *buf;
4037 u64 addr = pbuf->addr;
4038 int i, bid = pbuf->bid;
4040 for (i = 0; i < pbuf->nbufs; i++) {
4041 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4046 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4051 INIT_LIST_HEAD(&buf->list);
4054 list_add_tail(&buf->list, &(*head)->list);
4058 return i ? i : -ENOMEM;
4061 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4063 struct io_provide_buf *p = &req->pbuf;
4064 struct io_ring_ctx *ctx = req->ctx;
4065 struct io_buffer *head, *list;
4067 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4069 io_ring_submit_lock(ctx, !force_nonblock);
4071 lockdep_assert_held(&ctx->uring_lock);
4073 list = head = xa_load(&ctx->io_buffers, p->bgid);
4075 ret = io_add_buffers(p, &head);
4076 if (ret >= 0 && !list) {
4077 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4079 __io_remove_buffers(ctx, head, p->bgid, -1U);
4083 /* complete before unlock, IOPOLL may need the lock */
4084 __io_req_complete(req, issue_flags, ret, 0);
4085 io_ring_submit_unlock(ctx, !force_nonblock);
4089 static int io_epoll_ctl_prep(struct io_kiocb *req,
4090 const struct io_uring_sqe *sqe)
4092 #if defined(CONFIG_EPOLL)
4093 if (sqe->ioprio || sqe->buf_index)
4095 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4098 req->epoll.epfd = READ_ONCE(sqe->fd);
4099 req->epoll.op = READ_ONCE(sqe->len);
4100 req->epoll.fd = READ_ONCE(sqe->off);
4102 if (ep_op_has_event(req->epoll.op)) {
4103 struct epoll_event __user *ev;
4105 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4106 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4116 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4118 #if defined(CONFIG_EPOLL)
4119 struct io_epoll *ie = &req->epoll;
4121 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4123 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4124 if (force_nonblock && ret == -EAGAIN)
4129 __io_req_complete(req, issue_flags, ret, 0);
4136 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4138 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4139 if (sqe->ioprio || sqe->buf_index || sqe->off)
4141 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4144 req->madvise.addr = READ_ONCE(sqe->addr);
4145 req->madvise.len = READ_ONCE(sqe->len);
4146 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4153 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4155 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4156 struct io_madvise *ma = &req->madvise;
4159 if (issue_flags & IO_URING_F_NONBLOCK)
4162 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4165 io_req_complete(req, ret);
4172 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4174 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4176 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4179 req->fadvise.offset = READ_ONCE(sqe->off);
4180 req->fadvise.len = READ_ONCE(sqe->len);
4181 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4185 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4187 struct io_fadvise *fa = &req->fadvise;
4190 if (issue_flags & IO_URING_F_NONBLOCK) {
4191 switch (fa->advice) {
4192 case POSIX_FADV_NORMAL:
4193 case POSIX_FADV_RANDOM:
4194 case POSIX_FADV_SEQUENTIAL:
4201 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4204 __io_req_complete(req, issue_flags, ret, 0);
4208 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4210 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4212 if (sqe->ioprio || sqe->buf_index)
4214 if (req->flags & REQ_F_FIXED_FILE)
4217 req->statx.dfd = READ_ONCE(sqe->fd);
4218 req->statx.mask = READ_ONCE(sqe->len);
4219 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4220 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4221 req->statx.flags = READ_ONCE(sqe->statx_flags);
4226 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4228 struct io_statx *ctx = &req->statx;
4231 if (issue_flags & IO_URING_F_NONBLOCK)
4234 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4239 io_req_complete(req, ret);
4243 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4245 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4247 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4248 sqe->rw_flags || sqe->buf_index)
4250 if (req->flags & REQ_F_FIXED_FILE)
4253 req->close.fd = READ_ONCE(sqe->fd);
4257 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4259 struct files_struct *files = current->files;
4260 struct io_close *close = &req->close;
4261 struct fdtable *fdt;
4262 struct file *file = NULL;
4265 spin_lock(&files->file_lock);
4266 fdt = files_fdtable(files);
4267 if (close->fd >= fdt->max_fds) {
4268 spin_unlock(&files->file_lock);
4271 file = fdt->fd[close->fd];
4272 if (!file || file->f_op == &io_uring_fops) {
4273 spin_unlock(&files->file_lock);
4278 /* if the file has a flush method, be safe and punt to async */
4279 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4280 spin_unlock(&files->file_lock);
4284 ret = __close_fd_get_file(close->fd, &file);
4285 spin_unlock(&files->file_lock);
4292 /* No ->flush() or already async, safely close from here */
4293 ret = filp_close(file, current->files);
4299 __io_req_complete(req, issue_flags, ret, 0);
4303 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4305 struct io_ring_ctx *ctx = req->ctx;
4307 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4309 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4312 req->sync.off = READ_ONCE(sqe->off);
4313 req->sync.len = READ_ONCE(sqe->len);
4314 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4318 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4322 /* sync_file_range always requires a blocking context */
4323 if (issue_flags & IO_URING_F_NONBLOCK)
4326 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4330 io_req_complete(req, ret);
4334 #if defined(CONFIG_NET)
4335 static int io_setup_async_msg(struct io_kiocb *req,
4336 struct io_async_msghdr *kmsg)
4338 struct io_async_msghdr *async_msg = req->async_data;
4342 if (io_alloc_async_data(req)) {
4343 kfree(kmsg->free_iov);
4346 async_msg = req->async_data;
4347 req->flags |= REQ_F_NEED_CLEANUP;
4348 memcpy(async_msg, kmsg, sizeof(*kmsg));
4349 async_msg->msg.msg_name = &async_msg->addr;
4350 /* if were using fast_iov, set it to the new one */
4351 if (!async_msg->free_iov)
4352 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4357 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4358 struct io_async_msghdr *iomsg)
4360 iomsg->msg.msg_name = &iomsg->addr;
4361 iomsg->free_iov = iomsg->fast_iov;
4362 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4363 req->sr_msg.msg_flags, &iomsg->free_iov);
4366 static int io_sendmsg_prep_async(struct io_kiocb *req)
4370 ret = io_sendmsg_copy_hdr(req, req->async_data);
4372 req->flags |= REQ_F_NEED_CLEANUP;
4376 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4378 struct io_sr_msg *sr = &req->sr_msg;
4380 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4383 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4384 sr->len = READ_ONCE(sqe->len);
4385 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4386 if (sr->msg_flags & MSG_DONTWAIT)
4387 req->flags |= REQ_F_NOWAIT;
4389 #ifdef CONFIG_COMPAT
4390 if (req->ctx->compat)
4391 sr->msg_flags |= MSG_CMSG_COMPAT;
4396 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4398 struct io_async_msghdr iomsg, *kmsg;
4399 struct socket *sock;
4404 sock = sock_from_file(req->file);
4405 if (unlikely(!sock))
4408 kmsg = req->async_data;
4410 ret = io_sendmsg_copy_hdr(req, &iomsg);
4416 flags = req->sr_msg.msg_flags;
4417 if (issue_flags & IO_URING_F_NONBLOCK)
4418 flags |= MSG_DONTWAIT;
4419 if (flags & MSG_WAITALL)
4420 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4422 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4423 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4424 return io_setup_async_msg(req, kmsg);
4425 if (ret == -ERESTARTSYS)
4428 /* fast path, check for non-NULL to avoid function call */
4430 kfree(kmsg->free_iov);
4431 req->flags &= ~REQ_F_NEED_CLEANUP;
4434 __io_req_complete(req, issue_flags, ret, 0);
4438 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4440 struct io_sr_msg *sr = &req->sr_msg;
4443 struct socket *sock;
4448 sock = sock_from_file(req->file);
4449 if (unlikely(!sock))
4452 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4456 msg.msg_name = NULL;
4457 msg.msg_control = NULL;
4458 msg.msg_controllen = 0;
4459 msg.msg_namelen = 0;
4461 flags = req->sr_msg.msg_flags;
4462 if (issue_flags & IO_URING_F_NONBLOCK)
4463 flags |= MSG_DONTWAIT;
4464 if (flags & MSG_WAITALL)
4465 min_ret = iov_iter_count(&msg.msg_iter);
4467 msg.msg_flags = flags;
4468 ret = sock_sendmsg(sock, &msg);
4469 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4471 if (ret == -ERESTARTSYS)
4476 __io_req_complete(req, issue_flags, ret, 0);
4480 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4481 struct io_async_msghdr *iomsg)
4483 struct io_sr_msg *sr = &req->sr_msg;
4484 struct iovec __user *uiov;
4488 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4489 &iomsg->uaddr, &uiov, &iov_len);
4493 if (req->flags & REQ_F_BUFFER_SELECT) {
4496 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4498 sr->len = iomsg->fast_iov[0].iov_len;
4499 iomsg->free_iov = NULL;
4501 iomsg->free_iov = iomsg->fast_iov;
4502 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4503 &iomsg->free_iov, &iomsg->msg.msg_iter,
4512 #ifdef CONFIG_COMPAT
4513 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4514 struct io_async_msghdr *iomsg)
4516 struct io_sr_msg *sr = &req->sr_msg;
4517 struct compat_iovec __user *uiov;
4522 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4527 uiov = compat_ptr(ptr);
4528 if (req->flags & REQ_F_BUFFER_SELECT) {
4529 compat_ssize_t clen;
4533 if (!access_ok(uiov, sizeof(*uiov)))
4535 if (__get_user(clen, &uiov->iov_len))
4540 iomsg->free_iov = NULL;
4542 iomsg->free_iov = iomsg->fast_iov;
4543 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4544 UIO_FASTIOV, &iomsg->free_iov,
4545 &iomsg->msg.msg_iter, true);
4554 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4555 struct io_async_msghdr *iomsg)
4557 iomsg->msg.msg_name = &iomsg->addr;
4559 #ifdef CONFIG_COMPAT
4560 if (req->ctx->compat)
4561 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4564 return __io_recvmsg_copy_hdr(req, iomsg);
4567 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4570 struct io_sr_msg *sr = &req->sr_msg;
4571 struct io_buffer *kbuf;
4573 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4578 req->flags |= REQ_F_BUFFER_SELECTED;
4582 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4584 return io_put_kbuf(req, req->sr_msg.kbuf);
4587 static int io_recvmsg_prep_async(struct io_kiocb *req)
4591 ret = io_recvmsg_copy_hdr(req, req->async_data);
4593 req->flags |= REQ_F_NEED_CLEANUP;
4597 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4599 struct io_sr_msg *sr = &req->sr_msg;
4601 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4604 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4605 sr->len = READ_ONCE(sqe->len);
4606 sr->bgid = READ_ONCE(sqe->buf_group);
4607 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4608 if (sr->msg_flags & MSG_DONTWAIT)
4609 req->flags |= REQ_F_NOWAIT;
4611 #ifdef CONFIG_COMPAT
4612 if (req->ctx->compat)
4613 sr->msg_flags |= MSG_CMSG_COMPAT;
4618 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4620 struct io_async_msghdr iomsg, *kmsg;
4621 struct socket *sock;
4622 struct io_buffer *kbuf;
4625 int ret, cflags = 0;
4626 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4628 sock = sock_from_file(req->file);
4629 if (unlikely(!sock))
4632 kmsg = req->async_data;
4634 ret = io_recvmsg_copy_hdr(req, &iomsg);
4640 if (req->flags & REQ_F_BUFFER_SELECT) {
4641 kbuf = io_recv_buffer_select(req, !force_nonblock);
4643 return PTR_ERR(kbuf);
4644 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4645 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4646 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4647 1, req->sr_msg.len);
4650 flags = req->sr_msg.msg_flags;
4652 flags |= MSG_DONTWAIT;
4653 if (flags & MSG_WAITALL)
4654 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4656 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4657 kmsg->uaddr, flags);
4658 if (force_nonblock && ret == -EAGAIN)
4659 return io_setup_async_msg(req, kmsg);
4660 if (ret == -ERESTARTSYS)
4663 if (req->flags & REQ_F_BUFFER_SELECTED)
4664 cflags = io_put_recv_kbuf(req);
4665 /* fast path, check for non-NULL to avoid function call */
4667 kfree(kmsg->free_iov);
4668 req->flags &= ~REQ_F_NEED_CLEANUP;
4669 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4671 __io_req_complete(req, issue_flags, ret, cflags);
4675 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4677 struct io_buffer *kbuf;
4678 struct io_sr_msg *sr = &req->sr_msg;
4680 void __user *buf = sr->buf;
4681 struct socket *sock;
4685 int ret, cflags = 0;
4686 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4688 sock = sock_from_file(req->file);
4689 if (unlikely(!sock))
4692 if (req->flags & REQ_F_BUFFER_SELECT) {
4693 kbuf = io_recv_buffer_select(req, !force_nonblock);
4695 return PTR_ERR(kbuf);
4696 buf = u64_to_user_ptr(kbuf->addr);
4699 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4703 msg.msg_name = NULL;
4704 msg.msg_control = NULL;
4705 msg.msg_controllen = 0;
4706 msg.msg_namelen = 0;
4707 msg.msg_iocb = NULL;
4710 flags = req->sr_msg.msg_flags;
4712 flags |= MSG_DONTWAIT;
4713 if (flags & MSG_WAITALL)
4714 min_ret = iov_iter_count(&msg.msg_iter);
4716 ret = sock_recvmsg(sock, &msg, flags);
4717 if (force_nonblock && ret == -EAGAIN)
4719 if (ret == -ERESTARTSYS)
4722 if (req->flags & REQ_F_BUFFER_SELECTED)
4723 cflags = io_put_recv_kbuf(req);
4724 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4726 __io_req_complete(req, issue_flags, ret, cflags);
4730 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4732 struct io_accept *accept = &req->accept;
4734 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4736 if (sqe->ioprio || sqe->len || sqe->buf_index)
4739 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4740 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4741 accept->flags = READ_ONCE(sqe->accept_flags);
4742 accept->nofile = rlimit(RLIMIT_NOFILE);
4746 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4748 struct io_accept *accept = &req->accept;
4749 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4750 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4753 if (req->file->f_flags & O_NONBLOCK)
4754 req->flags |= REQ_F_NOWAIT;
4756 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4757 accept->addr_len, accept->flags,
4759 if (ret == -EAGAIN && force_nonblock)
4762 if (ret == -ERESTARTSYS)
4766 __io_req_complete(req, issue_flags, ret, 0);
4770 static int io_connect_prep_async(struct io_kiocb *req)
4772 struct io_async_connect *io = req->async_data;
4773 struct io_connect *conn = &req->connect;
4775 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4778 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4780 struct io_connect *conn = &req->connect;
4782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4784 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4787 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4788 conn->addr_len = READ_ONCE(sqe->addr2);
4792 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4794 struct io_async_connect __io, *io;
4795 unsigned file_flags;
4797 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4799 if (req->async_data) {
4800 io = req->async_data;
4802 ret = move_addr_to_kernel(req->connect.addr,
4803 req->connect.addr_len,
4810 file_flags = force_nonblock ? O_NONBLOCK : 0;
4812 ret = __sys_connect_file(req->file, &io->address,
4813 req->connect.addr_len, file_flags);
4814 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4815 if (req->async_data)
4817 if (io_alloc_async_data(req)) {
4821 memcpy(req->async_data, &__io, sizeof(__io));
4824 if (ret == -ERESTARTSYS)
4829 __io_req_complete(req, issue_flags, ret, 0);
4832 #else /* !CONFIG_NET */
4833 #define IO_NETOP_FN(op) \
4834 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4836 return -EOPNOTSUPP; \
4839 #define IO_NETOP_PREP(op) \
4841 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4843 return -EOPNOTSUPP; \
4846 #define IO_NETOP_PREP_ASYNC(op) \
4848 static int io_##op##_prep_async(struct io_kiocb *req) \
4850 return -EOPNOTSUPP; \
4853 IO_NETOP_PREP_ASYNC(sendmsg);
4854 IO_NETOP_PREP_ASYNC(recvmsg);
4855 IO_NETOP_PREP_ASYNC(connect);
4856 IO_NETOP_PREP(accept);
4859 #endif /* CONFIG_NET */
4861 struct io_poll_table {
4862 struct poll_table_struct pt;
4863 struct io_kiocb *req;
4868 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4869 __poll_t mask, io_req_tw_func_t func)
4871 /* for instances that support it check for an event match first: */
4872 if (mask && !(mask & poll->events))
4875 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4877 list_del_init(&poll->wait.entry);
4880 req->io_task_work.func = func;
4883 * If this fails, then the task is exiting. When a task exits, the
4884 * work gets canceled, so just cancel this request as well instead
4885 * of executing it. We can't safely execute it anyway, as we may not
4886 * have the needed state needed for it anyway.
4888 io_req_task_work_add(req);
4892 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4893 __acquires(&req->ctx->completion_lock)
4895 struct io_ring_ctx *ctx = req->ctx;
4897 /* req->task == current here, checking PF_EXITING is safe */
4898 if (unlikely(req->task->flags & PF_EXITING))
4899 WRITE_ONCE(poll->canceled, true);
4901 if (!req->result && !READ_ONCE(poll->canceled)) {
4902 struct poll_table_struct pt = { ._key = poll->events };
4904 req->result = vfs_poll(req->file, &pt) & poll->events;
4907 spin_lock(&ctx->completion_lock);
4908 if (!req->result && !READ_ONCE(poll->canceled)) {
4909 add_wait_queue(poll->head, &poll->wait);
4916 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4918 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4919 if (req->opcode == IORING_OP_POLL_ADD)
4920 return req->async_data;
4921 return req->apoll->double_poll;
4924 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4926 if (req->opcode == IORING_OP_POLL_ADD)
4928 return &req->apoll->poll;
4931 static void io_poll_remove_double(struct io_kiocb *req)
4932 __must_hold(&req->ctx->completion_lock)
4934 struct io_poll_iocb *poll = io_poll_get_double(req);
4936 lockdep_assert_held(&req->ctx->completion_lock);
4938 if (poll && poll->head) {
4939 struct wait_queue_head *head = poll->head;
4941 spin_lock_irq(&head->lock);
4942 list_del_init(&poll->wait.entry);
4943 if (poll->wait.private)
4946 spin_unlock_irq(&head->lock);
4950 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4951 __must_hold(&req->ctx->completion_lock)
4953 struct io_ring_ctx *ctx = req->ctx;
4954 unsigned flags = IORING_CQE_F_MORE;
4957 if (READ_ONCE(req->poll.canceled)) {
4959 req->poll.events |= EPOLLONESHOT;
4961 error = mangle_poll(mask);
4963 if (req->poll.events & EPOLLONESHOT)
4965 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4966 req->poll.done = true;
4969 if (flags & IORING_CQE_F_MORE)
4972 io_commit_cqring(ctx);
4973 return !(flags & IORING_CQE_F_MORE);
4976 static void io_poll_task_func(struct io_kiocb *req)
4978 struct io_ring_ctx *ctx = req->ctx;
4979 struct io_kiocb *nxt;
4981 if (io_poll_rewait(req, &req->poll)) {
4982 spin_unlock(&ctx->completion_lock);
4986 done = io_poll_complete(req, req->result);
4988 io_poll_remove_double(req);
4989 hash_del(&req->hash_node);
4992 add_wait_queue(req->poll.head, &req->poll.wait);
4994 spin_unlock(&ctx->completion_lock);
4995 io_cqring_ev_posted(ctx);
4998 nxt = io_put_req_find_next(req);
5000 io_req_task_submit(nxt);
5005 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5006 int sync, void *key)
5008 struct io_kiocb *req = wait->private;
5009 struct io_poll_iocb *poll = io_poll_get_single(req);
5010 __poll_t mask = key_to_poll(key);
5011 unsigned long flags;
5013 /* for instances that support it check for an event match first: */
5014 if (mask && !(mask & poll->events))
5016 if (!(poll->events & EPOLLONESHOT))
5017 return poll->wait.func(&poll->wait, mode, sync, key);
5019 list_del_init(&wait->entry);
5024 spin_lock_irqsave(&poll->head->lock, flags);
5025 done = list_empty(&poll->wait.entry);
5027 list_del_init(&poll->wait.entry);
5028 /* make sure double remove sees this as being gone */
5029 wait->private = NULL;
5030 spin_unlock_irqrestore(&poll->head->lock, flags);
5032 /* use wait func handler, so it matches the rq type */
5033 poll->wait.func(&poll->wait, mode, sync, key);
5040 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5041 wait_queue_func_t wake_func)
5045 poll->canceled = false;
5046 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5047 /* mask in events that we always want/need */
5048 poll->events = events | IO_POLL_UNMASK;
5049 INIT_LIST_HEAD(&poll->wait.entry);
5050 init_waitqueue_func_entry(&poll->wait, wake_func);
5053 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5054 struct wait_queue_head *head,
5055 struct io_poll_iocb **poll_ptr)
5057 struct io_kiocb *req = pt->req;
5060 * The file being polled uses multiple waitqueues for poll handling
5061 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5064 if (unlikely(pt->nr_entries)) {
5065 struct io_poll_iocb *poll_one = poll;
5067 /* double add on the same waitqueue head, ignore */
5068 if (poll_one->head == head)
5070 /* already have a 2nd entry, fail a third attempt */
5072 if ((*poll_ptr)->head == head)
5074 pt->error = -EINVAL;
5078 * Can't handle multishot for double wait for now, turn it
5079 * into one-shot mode.
5081 if (!(poll_one->events & EPOLLONESHOT))
5082 poll_one->events |= EPOLLONESHOT;
5083 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5085 pt->error = -ENOMEM;
5088 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5090 poll->wait.private = req;
5097 if (poll->events & EPOLLEXCLUSIVE)
5098 add_wait_queue_exclusive(head, &poll->wait);
5100 add_wait_queue(head, &poll->wait);
5103 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5104 struct poll_table_struct *p)
5106 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5107 struct async_poll *apoll = pt->req->apoll;
5109 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5112 static void io_async_task_func(struct io_kiocb *req)
5114 struct async_poll *apoll = req->apoll;
5115 struct io_ring_ctx *ctx = req->ctx;
5117 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5119 if (io_poll_rewait(req, &apoll->poll)) {
5120 spin_unlock(&ctx->completion_lock);
5124 hash_del(&req->hash_node);
5125 io_poll_remove_double(req);
5126 spin_unlock(&ctx->completion_lock);
5128 if (!READ_ONCE(apoll->poll.canceled))
5129 io_req_task_submit(req);
5131 io_req_complete_failed(req, -ECANCELED);
5134 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5137 struct io_kiocb *req = wait->private;
5138 struct io_poll_iocb *poll = &req->apoll->poll;
5140 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5143 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5146 static void io_poll_req_insert(struct io_kiocb *req)
5148 struct io_ring_ctx *ctx = req->ctx;
5149 struct hlist_head *list;
5151 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5152 hlist_add_head(&req->hash_node, list);
5155 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5156 struct io_poll_iocb *poll,
5157 struct io_poll_table *ipt, __poll_t mask,
5158 wait_queue_func_t wake_func)
5159 __acquires(&ctx->completion_lock)
5161 struct io_ring_ctx *ctx = req->ctx;
5162 bool cancel = false;
5164 INIT_HLIST_NODE(&req->hash_node);
5165 io_init_poll_iocb(poll, mask, wake_func);
5166 poll->file = req->file;
5167 poll->wait.private = req;
5169 ipt->pt._key = mask;
5172 ipt->nr_entries = 0;
5174 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5175 if (unlikely(!ipt->nr_entries) && !ipt->error)
5176 ipt->error = -EINVAL;
5178 spin_lock(&ctx->completion_lock);
5179 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5180 io_poll_remove_double(req);
5181 if (likely(poll->head)) {
5182 spin_lock_irq(&poll->head->lock);
5183 if (unlikely(list_empty(&poll->wait.entry))) {
5189 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5190 list_del_init(&poll->wait.entry);
5192 WRITE_ONCE(poll->canceled, true);
5193 else if (!poll->done) /* actually waiting for an event */
5194 io_poll_req_insert(req);
5195 spin_unlock_irq(&poll->head->lock);
5207 static int io_arm_poll_handler(struct io_kiocb *req)
5209 const struct io_op_def *def = &io_op_defs[req->opcode];
5210 struct io_ring_ctx *ctx = req->ctx;
5211 struct async_poll *apoll;
5212 struct io_poll_table ipt;
5213 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5216 if (!req->file || !file_can_poll(req->file))
5217 return IO_APOLL_ABORTED;
5218 if (req->flags & REQ_F_POLLED)
5219 return IO_APOLL_ABORTED;
5220 if (!def->pollin && !def->pollout)
5221 return IO_APOLL_ABORTED;
5225 mask |= POLLIN | POLLRDNORM;
5227 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5228 if ((req->opcode == IORING_OP_RECVMSG) &&
5229 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5233 mask |= POLLOUT | POLLWRNORM;
5236 /* if we can't nonblock try, then no point in arming a poll handler */
5237 if (!io_file_supports_nowait(req, rw))
5238 return IO_APOLL_ABORTED;
5240 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5241 if (unlikely(!apoll))
5242 return IO_APOLL_ABORTED;
5243 apoll->double_poll = NULL;
5245 req->flags |= REQ_F_POLLED;
5246 ipt.pt._qproc = io_async_queue_proc;
5247 io_req_set_refcount(req);
5249 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5251 spin_unlock(&ctx->completion_lock);
5252 if (ret || ipt.error)
5253 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5255 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5256 mask, apoll->poll.events);
5260 static bool __io_poll_remove_one(struct io_kiocb *req,
5261 struct io_poll_iocb *poll, bool do_cancel)
5262 __must_hold(&req->ctx->completion_lock)
5264 bool do_complete = false;
5268 spin_lock_irq(&poll->head->lock);
5270 WRITE_ONCE(poll->canceled, true);
5271 if (!list_empty(&poll->wait.entry)) {
5272 list_del_init(&poll->wait.entry);
5275 spin_unlock_irq(&poll->head->lock);
5276 hash_del(&req->hash_node);
5280 static bool io_poll_remove_one(struct io_kiocb *req)
5281 __must_hold(&req->ctx->completion_lock)
5285 io_poll_remove_double(req);
5286 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5289 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5290 io_commit_cqring(req->ctx);
5292 io_put_req_deferred(req);
5298 * Returns true if we found and killed one or more poll requests
5300 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5303 struct hlist_node *tmp;
5304 struct io_kiocb *req;
5307 spin_lock(&ctx->completion_lock);
5308 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5309 struct hlist_head *list;
5311 list = &ctx->cancel_hash[i];
5312 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5313 if (io_match_task(req, tsk, cancel_all))
5314 posted += io_poll_remove_one(req);
5317 spin_unlock(&ctx->completion_lock);
5320 io_cqring_ev_posted(ctx);
5325 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5327 __must_hold(&ctx->completion_lock)
5329 struct hlist_head *list;
5330 struct io_kiocb *req;
5332 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5333 hlist_for_each_entry(req, list, hash_node) {
5334 if (sqe_addr != req->user_data)
5336 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5343 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5345 __must_hold(&ctx->completion_lock)
5347 struct io_kiocb *req;
5349 req = io_poll_find(ctx, sqe_addr, poll_only);
5352 if (io_poll_remove_one(req))
5358 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5363 events = READ_ONCE(sqe->poll32_events);
5365 events = swahw32(events);
5367 if (!(flags & IORING_POLL_ADD_MULTI))
5368 events |= EPOLLONESHOT;
5369 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5372 static int io_poll_update_prep(struct io_kiocb *req,
5373 const struct io_uring_sqe *sqe)
5375 struct io_poll_update *upd = &req->poll_update;
5378 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5380 if (sqe->ioprio || sqe->buf_index)
5382 flags = READ_ONCE(sqe->len);
5383 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5384 IORING_POLL_ADD_MULTI))
5386 /* meaningless without update */
5387 if (flags == IORING_POLL_ADD_MULTI)
5390 upd->old_user_data = READ_ONCE(sqe->addr);
5391 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5392 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5394 upd->new_user_data = READ_ONCE(sqe->off);
5395 if (!upd->update_user_data && upd->new_user_data)
5397 if (upd->update_events)
5398 upd->events = io_poll_parse_events(sqe, flags);
5399 else if (sqe->poll32_events)
5405 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5408 struct io_kiocb *req = wait->private;
5409 struct io_poll_iocb *poll = &req->poll;
5411 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5414 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5415 struct poll_table_struct *p)
5417 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5419 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5422 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5424 struct io_poll_iocb *poll = &req->poll;
5427 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5429 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5431 flags = READ_ONCE(sqe->len);
5432 if (flags & ~IORING_POLL_ADD_MULTI)
5435 io_req_set_refcount(req);
5436 poll->events = io_poll_parse_events(sqe, flags);
5440 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5442 struct io_poll_iocb *poll = &req->poll;
5443 struct io_ring_ctx *ctx = req->ctx;
5444 struct io_poll_table ipt;
5447 ipt.pt._qproc = io_poll_queue_proc;
5449 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5452 if (mask) { /* no async, we'd stolen it */
5454 io_poll_complete(req, mask);
5456 spin_unlock(&ctx->completion_lock);
5459 io_cqring_ev_posted(ctx);
5460 if (poll->events & EPOLLONESHOT)
5466 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5468 struct io_ring_ctx *ctx = req->ctx;
5469 struct io_kiocb *preq;
5473 spin_lock(&ctx->completion_lock);
5474 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5480 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5482 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5487 * Don't allow racy completion with singleshot, as we cannot safely
5488 * update those. For multishot, if we're racing with completion, just
5489 * let completion re-add it.
5491 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5492 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5496 /* we now have a detached poll request. reissue. */
5500 spin_unlock(&ctx->completion_lock);
5502 io_req_complete(req, ret);
5505 /* only mask one event flags, keep behavior flags */
5506 if (req->poll_update.update_events) {
5507 preq->poll.events &= ~0xffff;
5508 preq->poll.events |= req->poll_update.events & 0xffff;
5509 preq->poll.events |= IO_POLL_UNMASK;
5511 if (req->poll_update.update_user_data)
5512 preq->user_data = req->poll_update.new_user_data;
5513 spin_unlock(&ctx->completion_lock);
5515 /* complete update request, we're done with it */
5516 io_req_complete(req, ret);
5519 ret = io_poll_add(preq, issue_flags);
5522 io_req_complete(preq, ret);
5528 static void io_req_task_timeout(struct io_kiocb *req)
5531 io_req_complete_post(req, -ETIME, 0);
5534 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5536 struct io_timeout_data *data = container_of(timer,
5537 struct io_timeout_data, timer);
5538 struct io_kiocb *req = data->req;
5539 struct io_ring_ctx *ctx = req->ctx;
5540 unsigned long flags;
5542 spin_lock_irqsave(&ctx->timeout_lock, flags);
5543 list_del_init(&req->timeout.list);
5544 atomic_set(&req->ctx->cq_timeouts,
5545 atomic_read(&req->ctx->cq_timeouts) + 1);
5546 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5548 req->io_task_work.func = io_req_task_timeout;
5549 io_req_task_work_add(req);
5550 return HRTIMER_NORESTART;
5553 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5555 __must_hold(&ctx->timeout_lock)
5557 struct io_timeout_data *io;
5558 struct io_kiocb *req;
5561 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5562 found = user_data == req->user_data;
5567 return ERR_PTR(-ENOENT);
5569 io = req->async_data;
5570 if (hrtimer_try_to_cancel(&io->timer) == -1)
5571 return ERR_PTR(-EALREADY);
5572 list_del_init(&req->timeout.list);
5576 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5577 __must_hold(&ctx->completion_lock)
5578 __must_hold(&ctx->timeout_lock)
5580 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5583 return PTR_ERR(req);
5586 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5587 io_put_req_deferred(req);
5591 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5592 struct timespec64 *ts, enum hrtimer_mode mode)
5593 __must_hold(&ctx->timeout_lock)
5595 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5596 struct io_timeout_data *data;
5599 return PTR_ERR(req);
5601 req->timeout.off = 0; /* noseq */
5602 data = req->async_data;
5603 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5604 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5605 data->timer.function = io_timeout_fn;
5606 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5610 static int io_timeout_remove_prep(struct io_kiocb *req,
5611 const struct io_uring_sqe *sqe)
5613 struct io_timeout_rem *tr = &req->timeout_rem;
5615 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5617 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5619 if (sqe->ioprio || sqe->buf_index || sqe->len)
5622 tr->addr = READ_ONCE(sqe->addr);
5623 tr->flags = READ_ONCE(sqe->timeout_flags);
5624 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5625 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5627 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5629 } else if (tr->flags) {
5630 /* timeout removal doesn't support flags */
5637 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5639 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5644 * Remove or update an existing timeout command
5646 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5648 struct io_timeout_rem *tr = &req->timeout_rem;
5649 struct io_ring_ctx *ctx = req->ctx;
5652 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5653 spin_lock(&ctx->completion_lock);
5654 spin_lock_irq(&ctx->timeout_lock);
5655 ret = io_timeout_cancel(ctx, tr->addr);
5656 spin_unlock_irq(&ctx->timeout_lock);
5657 spin_unlock(&ctx->completion_lock);
5659 spin_lock_irq(&ctx->timeout_lock);
5660 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5661 io_translate_timeout_mode(tr->flags));
5662 spin_unlock_irq(&ctx->timeout_lock);
5667 io_req_complete_post(req, ret, 0);
5671 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5672 bool is_timeout_link)
5674 struct io_timeout_data *data;
5676 u32 off = READ_ONCE(sqe->off);
5678 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5680 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5682 if (off && is_timeout_link)
5684 flags = READ_ONCE(sqe->timeout_flags);
5685 if (flags & ~IORING_TIMEOUT_ABS)
5688 req->timeout.off = off;
5689 if (unlikely(off && !req->ctx->off_timeout_used))
5690 req->ctx->off_timeout_used = true;
5692 if (!req->async_data && io_alloc_async_data(req))
5695 data = req->async_data;
5698 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5701 data->mode = io_translate_timeout_mode(flags);
5702 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5704 if (is_timeout_link) {
5705 struct io_submit_link *link = &req->ctx->submit_state.link;
5709 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5711 req->timeout.head = link->last;
5712 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5717 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5719 struct io_ring_ctx *ctx = req->ctx;
5720 struct io_timeout_data *data = req->async_data;
5721 struct list_head *entry;
5722 u32 tail, off = req->timeout.off;
5724 spin_lock_irq(&ctx->timeout_lock);
5727 * sqe->off holds how many events that need to occur for this
5728 * timeout event to be satisfied. If it isn't set, then this is
5729 * a pure timeout request, sequence isn't used.
5731 if (io_is_timeout_noseq(req)) {
5732 entry = ctx->timeout_list.prev;
5736 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5737 req->timeout.target_seq = tail + off;
5739 /* Update the last seq here in case io_flush_timeouts() hasn't.
5740 * This is safe because ->completion_lock is held, and submissions
5741 * and completions are never mixed in the same ->completion_lock section.
5743 ctx->cq_last_tm_flush = tail;
5746 * Insertion sort, ensuring the first entry in the list is always
5747 * the one we need first.
5749 list_for_each_prev(entry, &ctx->timeout_list) {
5750 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5753 if (io_is_timeout_noseq(nxt))
5755 /* nxt.seq is behind @tail, otherwise would've been completed */
5756 if (off >= nxt->timeout.target_seq - tail)
5760 list_add(&req->timeout.list, entry);
5761 data->timer.function = io_timeout_fn;
5762 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5763 spin_unlock_irq(&ctx->timeout_lock);
5767 struct io_cancel_data {
5768 struct io_ring_ctx *ctx;
5772 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5774 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5775 struct io_cancel_data *cd = data;
5777 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5780 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5781 struct io_ring_ctx *ctx)
5783 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5784 enum io_wq_cancel cancel_ret;
5787 if (!tctx || !tctx->io_wq)
5790 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5791 switch (cancel_ret) {
5792 case IO_WQ_CANCEL_OK:
5795 case IO_WQ_CANCEL_RUNNING:
5798 case IO_WQ_CANCEL_NOTFOUND:
5806 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5808 struct io_ring_ctx *ctx = req->ctx;
5811 WARN_ON_ONCE(req->task != current);
5813 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5817 spin_lock(&ctx->completion_lock);
5818 spin_lock_irq(&ctx->timeout_lock);
5819 ret = io_timeout_cancel(ctx, sqe_addr);
5820 spin_unlock_irq(&ctx->timeout_lock);
5823 ret = io_poll_cancel(ctx, sqe_addr, false);
5825 spin_unlock(&ctx->completion_lock);
5829 static int io_async_cancel_prep(struct io_kiocb *req,
5830 const struct io_uring_sqe *sqe)
5832 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5834 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5836 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5839 req->cancel.addr = READ_ONCE(sqe->addr);
5843 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5845 struct io_ring_ctx *ctx = req->ctx;
5846 u64 sqe_addr = req->cancel.addr;
5847 struct io_tctx_node *node;
5850 ret = io_try_cancel_userdata(req, sqe_addr);
5854 /* slow path, try all io-wq's */
5855 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5857 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5858 struct io_uring_task *tctx = node->task->io_uring;
5860 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5864 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5868 io_req_complete_post(req, ret, 0);
5872 static int io_rsrc_update_prep(struct io_kiocb *req,
5873 const struct io_uring_sqe *sqe)
5875 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5877 if (sqe->ioprio || sqe->rw_flags)
5880 req->rsrc_update.offset = READ_ONCE(sqe->off);
5881 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5882 if (!req->rsrc_update.nr_args)
5884 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5888 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5890 struct io_ring_ctx *ctx = req->ctx;
5891 struct io_uring_rsrc_update2 up;
5894 if (issue_flags & IO_URING_F_NONBLOCK)
5897 up.offset = req->rsrc_update.offset;
5898 up.data = req->rsrc_update.arg;
5903 mutex_lock(&ctx->uring_lock);
5904 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5905 &up, req->rsrc_update.nr_args);
5906 mutex_unlock(&ctx->uring_lock);
5910 __io_req_complete(req, issue_flags, ret, 0);
5914 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5916 switch (req->opcode) {
5919 case IORING_OP_READV:
5920 case IORING_OP_READ_FIXED:
5921 case IORING_OP_READ:
5922 return io_read_prep(req, sqe);
5923 case IORING_OP_WRITEV:
5924 case IORING_OP_WRITE_FIXED:
5925 case IORING_OP_WRITE:
5926 return io_write_prep(req, sqe);
5927 case IORING_OP_POLL_ADD:
5928 return io_poll_add_prep(req, sqe);
5929 case IORING_OP_POLL_REMOVE:
5930 return io_poll_update_prep(req, sqe);
5931 case IORING_OP_FSYNC:
5932 return io_fsync_prep(req, sqe);
5933 case IORING_OP_SYNC_FILE_RANGE:
5934 return io_sfr_prep(req, sqe);
5935 case IORING_OP_SENDMSG:
5936 case IORING_OP_SEND:
5937 return io_sendmsg_prep(req, sqe);
5938 case IORING_OP_RECVMSG:
5939 case IORING_OP_RECV:
5940 return io_recvmsg_prep(req, sqe);
5941 case IORING_OP_CONNECT:
5942 return io_connect_prep(req, sqe);
5943 case IORING_OP_TIMEOUT:
5944 return io_timeout_prep(req, sqe, false);
5945 case IORING_OP_TIMEOUT_REMOVE:
5946 return io_timeout_remove_prep(req, sqe);
5947 case IORING_OP_ASYNC_CANCEL:
5948 return io_async_cancel_prep(req, sqe);
5949 case IORING_OP_LINK_TIMEOUT:
5950 return io_timeout_prep(req, sqe, true);
5951 case IORING_OP_ACCEPT:
5952 return io_accept_prep(req, sqe);
5953 case IORING_OP_FALLOCATE:
5954 return io_fallocate_prep(req, sqe);
5955 case IORING_OP_OPENAT:
5956 return io_openat_prep(req, sqe);
5957 case IORING_OP_CLOSE:
5958 return io_close_prep(req, sqe);
5959 case IORING_OP_FILES_UPDATE:
5960 return io_rsrc_update_prep(req, sqe);
5961 case IORING_OP_STATX:
5962 return io_statx_prep(req, sqe);
5963 case IORING_OP_FADVISE:
5964 return io_fadvise_prep(req, sqe);
5965 case IORING_OP_MADVISE:
5966 return io_madvise_prep(req, sqe);
5967 case IORING_OP_OPENAT2:
5968 return io_openat2_prep(req, sqe);
5969 case IORING_OP_EPOLL_CTL:
5970 return io_epoll_ctl_prep(req, sqe);
5971 case IORING_OP_SPLICE:
5972 return io_splice_prep(req, sqe);
5973 case IORING_OP_PROVIDE_BUFFERS:
5974 return io_provide_buffers_prep(req, sqe);
5975 case IORING_OP_REMOVE_BUFFERS:
5976 return io_remove_buffers_prep(req, sqe);
5978 return io_tee_prep(req, sqe);
5979 case IORING_OP_SHUTDOWN:
5980 return io_shutdown_prep(req, sqe);
5981 case IORING_OP_RENAMEAT:
5982 return io_renameat_prep(req, sqe);
5983 case IORING_OP_UNLINKAT:
5984 return io_unlinkat_prep(req, sqe);
5987 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5992 static int io_req_prep_async(struct io_kiocb *req)
5994 if (!io_op_defs[req->opcode].needs_async_setup)
5996 if (WARN_ON_ONCE(req->async_data))
5998 if (io_alloc_async_data(req))
6001 switch (req->opcode) {
6002 case IORING_OP_READV:
6003 return io_rw_prep_async(req, READ);
6004 case IORING_OP_WRITEV:
6005 return io_rw_prep_async(req, WRITE);
6006 case IORING_OP_SENDMSG:
6007 return io_sendmsg_prep_async(req);
6008 case IORING_OP_RECVMSG:
6009 return io_recvmsg_prep_async(req);
6010 case IORING_OP_CONNECT:
6011 return io_connect_prep_async(req);
6013 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6018 static u32 io_get_sequence(struct io_kiocb *req)
6020 u32 seq = req->ctx->cached_sq_head;
6022 /* need original cached_sq_head, but it was increased for each req */
6023 io_for_each_link(req, req)
6028 static bool io_drain_req(struct io_kiocb *req)
6030 struct io_kiocb *pos;
6031 struct io_ring_ctx *ctx = req->ctx;
6032 struct io_defer_entry *de;
6037 * If we need to drain a request in the middle of a link, drain the
6038 * head request and the next request/link after the current link.
6039 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6040 * maintained for every request of our link.
6042 if (ctx->drain_next) {
6043 req->flags |= REQ_F_IO_DRAIN;
6044 ctx->drain_next = false;
6046 /* not interested in head, start from the first linked */
6047 io_for_each_link(pos, req->link) {
6048 if (pos->flags & REQ_F_IO_DRAIN) {
6049 ctx->drain_next = true;
6050 req->flags |= REQ_F_IO_DRAIN;
6055 /* Still need defer if there is pending req in defer list. */
6056 if (likely(list_empty_careful(&ctx->defer_list) &&
6057 !(req->flags & REQ_F_IO_DRAIN))) {
6058 ctx->drain_active = false;
6062 seq = io_get_sequence(req);
6063 /* Still a chance to pass the sequence check */
6064 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6067 ret = io_req_prep_async(req);
6070 io_prep_async_link(req);
6071 de = kmalloc(sizeof(*de), GFP_KERNEL);
6075 io_req_complete_failed(req, ret);
6079 spin_lock(&ctx->completion_lock);
6080 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6081 spin_unlock(&ctx->completion_lock);
6083 io_queue_async_work(req);
6087 trace_io_uring_defer(ctx, req, req->user_data);
6090 list_add_tail(&de->list, &ctx->defer_list);
6091 spin_unlock(&ctx->completion_lock);
6095 static void io_clean_op(struct io_kiocb *req)
6097 if (req->flags & REQ_F_BUFFER_SELECTED) {
6098 switch (req->opcode) {
6099 case IORING_OP_READV:
6100 case IORING_OP_READ_FIXED:
6101 case IORING_OP_READ:
6102 kfree((void *)(unsigned long)req->rw.addr);
6104 case IORING_OP_RECVMSG:
6105 case IORING_OP_RECV:
6106 kfree(req->sr_msg.kbuf);
6111 if (req->flags & REQ_F_NEED_CLEANUP) {
6112 switch (req->opcode) {
6113 case IORING_OP_READV:
6114 case IORING_OP_READ_FIXED:
6115 case IORING_OP_READ:
6116 case IORING_OP_WRITEV:
6117 case IORING_OP_WRITE_FIXED:
6118 case IORING_OP_WRITE: {
6119 struct io_async_rw *io = req->async_data;
6121 kfree(io->free_iovec);
6124 case IORING_OP_RECVMSG:
6125 case IORING_OP_SENDMSG: {
6126 struct io_async_msghdr *io = req->async_data;
6128 kfree(io->free_iov);
6131 case IORING_OP_SPLICE:
6133 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6134 io_put_file(req->splice.file_in);
6136 case IORING_OP_OPENAT:
6137 case IORING_OP_OPENAT2:
6138 if (req->open.filename)
6139 putname(req->open.filename);
6141 case IORING_OP_RENAMEAT:
6142 putname(req->rename.oldpath);
6143 putname(req->rename.newpath);
6145 case IORING_OP_UNLINKAT:
6146 putname(req->unlink.filename);
6150 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6151 kfree(req->apoll->double_poll);
6155 if (req->flags & REQ_F_INFLIGHT) {
6156 struct io_uring_task *tctx = req->task->io_uring;
6158 atomic_dec(&tctx->inflight_tracked);
6160 if (req->flags & REQ_F_CREDS)
6161 put_cred(req->creds);
6163 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6166 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6168 struct io_ring_ctx *ctx = req->ctx;
6169 const struct cred *creds = NULL;
6172 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6173 creds = override_creds(req->creds);
6175 switch (req->opcode) {
6177 ret = io_nop(req, issue_flags);
6179 case IORING_OP_READV:
6180 case IORING_OP_READ_FIXED:
6181 case IORING_OP_READ:
6182 ret = io_read(req, issue_flags);
6184 case IORING_OP_WRITEV:
6185 case IORING_OP_WRITE_FIXED:
6186 case IORING_OP_WRITE:
6187 ret = io_write(req, issue_flags);
6189 case IORING_OP_FSYNC:
6190 ret = io_fsync(req, issue_flags);
6192 case IORING_OP_POLL_ADD:
6193 ret = io_poll_add(req, issue_flags);
6195 case IORING_OP_POLL_REMOVE:
6196 ret = io_poll_update(req, issue_flags);
6198 case IORING_OP_SYNC_FILE_RANGE:
6199 ret = io_sync_file_range(req, issue_flags);
6201 case IORING_OP_SENDMSG:
6202 ret = io_sendmsg(req, issue_flags);
6204 case IORING_OP_SEND:
6205 ret = io_send(req, issue_flags);
6207 case IORING_OP_RECVMSG:
6208 ret = io_recvmsg(req, issue_flags);
6210 case IORING_OP_RECV:
6211 ret = io_recv(req, issue_flags);
6213 case IORING_OP_TIMEOUT:
6214 ret = io_timeout(req, issue_flags);
6216 case IORING_OP_TIMEOUT_REMOVE:
6217 ret = io_timeout_remove(req, issue_flags);
6219 case IORING_OP_ACCEPT:
6220 ret = io_accept(req, issue_flags);
6222 case IORING_OP_CONNECT:
6223 ret = io_connect(req, issue_flags);
6225 case IORING_OP_ASYNC_CANCEL:
6226 ret = io_async_cancel(req, issue_flags);
6228 case IORING_OP_FALLOCATE:
6229 ret = io_fallocate(req, issue_flags);
6231 case IORING_OP_OPENAT:
6232 ret = io_openat(req, issue_flags);
6234 case IORING_OP_CLOSE:
6235 ret = io_close(req, issue_flags);
6237 case IORING_OP_FILES_UPDATE:
6238 ret = io_files_update(req, issue_flags);
6240 case IORING_OP_STATX:
6241 ret = io_statx(req, issue_flags);
6243 case IORING_OP_FADVISE:
6244 ret = io_fadvise(req, issue_flags);
6246 case IORING_OP_MADVISE:
6247 ret = io_madvise(req, issue_flags);
6249 case IORING_OP_OPENAT2:
6250 ret = io_openat2(req, issue_flags);
6252 case IORING_OP_EPOLL_CTL:
6253 ret = io_epoll_ctl(req, issue_flags);
6255 case IORING_OP_SPLICE:
6256 ret = io_splice(req, issue_flags);
6258 case IORING_OP_PROVIDE_BUFFERS:
6259 ret = io_provide_buffers(req, issue_flags);
6261 case IORING_OP_REMOVE_BUFFERS:
6262 ret = io_remove_buffers(req, issue_flags);
6265 ret = io_tee(req, issue_flags);
6267 case IORING_OP_SHUTDOWN:
6268 ret = io_shutdown(req, issue_flags);
6270 case IORING_OP_RENAMEAT:
6271 ret = io_renameat(req, issue_flags);
6273 case IORING_OP_UNLINKAT:
6274 ret = io_unlinkat(req, issue_flags);
6282 revert_creds(creds);
6285 /* If the op doesn't have a file, we're not polling for it */
6286 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6287 io_iopoll_req_issued(req);
6292 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6294 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6296 req = io_put_req_find_next(req);
6297 return req ? &req->work : NULL;
6300 static void io_wq_submit_work(struct io_wq_work *work)
6302 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6303 struct io_kiocb *timeout;
6306 /* one will be dropped by ->io_free_work() after returning to io-wq */
6307 if (!(req->flags & REQ_F_REFCOUNT))
6308 __io_req_set_refcount(req, 2);
6312 timeout = io_prep_linked_timeout(req);
6314 io_queue_linked_timeout(timeout);
6316 if (work->flags & IO_WQ_WORK_CANCEL)
6321 ret = io_issue_sqe(req, 0);
6323 * We can get EAGAIN for polled IO even though we're
6324 * forcing a sync submission from here, since we can't
6325 * wait for request slots on the block side.
6333 /* avoid locking problems by failing it from a clean context */
6335 io_req_task_queue_fail(req, ret);
6338 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6341 return &table->files[i];
6344 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6347 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6349 return (struct file *) (slot->file_ptr & FFS_MASK);
6352 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6354 unsigned long file_ptr = (unsigned long) file;
6356 if (__io_file_supports_nowait(file, READ))
6357 file_ptr |= FFS_ASYNC_READ;
6358 if (__io_file_supports_nowait(file, WRITE))
6359 file_ptr |= FFS_ASYNC_WRITE;
6360 if (S_ISREG(file_inode(file)->i_mode))
6361 file_ptr |= FFS_ISREG;
6362 file_slot->file_ptr = file_ptr;
6365 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6366 struct io_kiocb *req, int fd)
6369 unsigned long file_ptr;
6371 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6373 fd = array_index_nospec(fd, ctx->nr_user_files);
6374 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6375 file = (struct file *) (file_ptr & FFS_MASK);
6376 file_ptr &= ~FFS_MASK;
6377 /* mask in overlapping REQ_F and FFS bits */
6378 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6379 io_req_set_rsrc_node(req);
6383 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6384 struct io_kiocb *req, int fd)
6386 struct file *file = fget(fd);
6388 trace_io_uring_file_get(ctx, fd);
6390 /* we don't allow fixed io_uring files */
6391 if (file && unlikely(file->f_op == &io_uring_fops))
6392 io_req_track_inflight(req);
6396 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6397 struct io_kiocb *req, int fd, bool fixed)
6400 return io_file_get_fixed(ctx, req, fd);
6402 return io_file_get_normal(ctx, req, fd);
6405 static void io_req_task_link_timeout(struct io_kiocb *req)
6407 struct io_kiocb *prev = req->timeout.prev;
6411 ret = io_try_cancel_userdata(req, prev->user_data);
6412 io_req_complete_post(req, ret ?: -ETIME, 0);
6415 io_req_complete_post(req, -ETIME, 0);
6419 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6421 struct io_timeout_data *data = container_of(timer,
6422 struct io_timeout_data, timer);
6423 struct io_kiocb *prev, *req = data->req;
6424 struct io_ring_ctx *ctx = req->ctx;
6425 unsigned long flags;
6427 spin_lock_irqsave(&ctx->timeout_lock, flags);
6428 prev = req->timeout.head;
6429 req->timeout.head = NULL;
6432 * We don't expect the list to be empty, that will only happen if we
6433 * race with the completion of the linked work.
6436 io_remove_next_linked(prev);
6437 if (!req_ref_inc_not_zero(prev))
6440 req->timeout.prev = prev;
6441 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6443 req->io_task_work.func = io_req_task_link_timeout;
6444 io_req_task_work_add(req);
6445 return HRTIMER_NORESTART;
6448 static void io_queue_linked_timeout(struct io_kiocb *req)
6450 struct io_ring_ctx *ctx = req->ctx;
6452 spin_lock_irq(&ctx->timeout_lock);
6454 * If the back reference is NULL, then our linked request finished
6455 * before we got a chance to setup the timer
6457 if (req->timeout.head) {
6458 struct io_timeout_data *data = req->async_data;
6460 data->timer.function = io_link_timeout_fn;
6461 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6464 spin_unlock_irq(&ctx->timeout_lock);
6465 /* drop submission reference */
6469 static void __io_queue_sqe(struct io_kiocb *req)
6470 __must_hold(&req->ctx->uring_lock)
6472 struct io_kiocb *linked_timeout;
6476 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6479 * We async punt it if the file wasn't marked NOWAIT, or if the file
6480 * doesn't support non-blocking read/write attempts
6483 if (req->flags & REQ_F_COMPLETE_INLINE) {
6484 struct io_ring_ctx *ctx = req->ctx;
6485 struct io_submit_state *state = &ctx->submit_state;
6487 state->compl_reqs[state->compl_nr++] = req;
6488 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6489 io_submit_flush_completions(ctx);
6493 linked_timeout = io_prep_linked_timeout(req);
6495 io_queue_linked_timeout(linked_timeout);
6496 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6497 linked_timeout = io_prep_linked_timeout(req);
6499 switch (io_arm_poll_handler(req)) {
6500 case IO_APOLL_READY:
6502 io_unprep_linked_timeout(req);
6504 case IO_APOLL_ABORTED:
6506 * Queued up for async execution, worker will release
6507 * submit reference when the iocb is actually submitted.
6509 io_queue_async_work(req);
6514 io_queue_linked_timeout(linked_timeout);
6516 io_req_complete_failed(req, ret);
6520 static inline void io_queue_sqe(struct io_kiocb *req)
6521 __must_hold(&req->ctx->uring_lock)
6523 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6526 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6527 __io_queue_sqe(req);
6529 int ret = io_req_prep_async(req);
6532 io_req_complete_failed(req, ret);
6534 io_queue_async_work(req);
6539 * Check SQE restrictions (opcode and flags).
6541 * Returns 'true' if SQE is allowed, 'false' otherwise.
6543 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6544 struct io_kiocb *req,
6545 unsigned int sqe_flags)
6547 if (likely(!ctx->restricted))
6550 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6553 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6554 ctx->restrictions.sqe_flags_required)
6557 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6558 ctx->restrictions.sqe_flags_required))
6564 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6565 const struct io_uring_sqe *sqe)
6566 __must_hold(&ctx->uring_lock)
6568 struct io_submit_state *state;
6569 unsigned int sqe_flags;
6570 int personality, ret = 0;
6572 /* req is partially pre-initialised, see io_preinit_req() */
6573 req->opcode = READ_ONCE(sqe->opcode);
6574 /* same numerical values with corresponding REQ_F_*, safe to copy */
6575 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6576 req->user_data = READ_ONCE(sqe->user_data);
6578 req->fixed_rsrc_refs = NULL;
6579 req->task = current;
6581 /* enforce forwards compatibility on users */
6582 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6584 if (unlikely(req->opcode >= IORING_OP_LAST))
6586 if (!io_check_restriction(ctx, req, sqe_flags))
6589 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6590 !io_op_defs[req->opcode].buffer_select)
6592 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6593 ctx->drain_active = true;
6595 personality = READ_ONCE(sqe->personality);
6597 req->creds = xa_load(&ctx->personalities, personality);
6600 get_cred(req->creds);
6601 req->flags |= REQ_F_CREDS;
6603 state = &ctx->submit_state;
6606 * Plug now if we have more than 1 IO left after this, and the target
6607 * is potentially a read/write to block based storage.
6609 if (!state->plug_started && state->ios_left > 1 &&
6610 io_op_defs[req->opcode].plug) {
6611 blk_start_plug(&state->plug);
6612 state->plug_started = true;
6615 if (io_op_defs[req->opcode].needs_file) {
6616 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6617 (sqe_flags & IOSQE_FIXED_FILE));
6618 if (unlikely(!req->file))
6626 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6627 const struct io_uring_sqe *sqe)
6628 __must_hold(&ctx->uring_lock)
6630 struct io_submit_link *link = &ctx->submit_state.link;
6633 ret = io_init_req(ctx, req, sqe);
6634 if (unlikely(ret)) {
6637 /* fail even hard links since we don't submit */
6638 req_set_fail(link->head);
6639 io_req_complete_failed(link->head, -ECANCELED);
6642 io_req_complete_failed(req, ret);
6646 ret = io_req_prep(req, sqe);
6650 /* don't need @sqe from now on */
6651 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6653 ctx->flags & IORING_SETUP_SQPOLL);
6656 * If we already have a head request, queue this one for async
6657 * submittal once the head completes. If we don't have a head but
6658 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6659 * submitted sync once the chain is complete. If none of those
6660 * conditions are true (normal request), then just queue it.
6663 struct io_kiocb *head = link->head;
6665 ret = io_req_prep_async(req);
6668 trace_io_uring_link(ctx, req, head);
6669 link->last->link = req;
6672 /* last request of a link, enqueue the link */
6673 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6678 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6690 * Batched submission is done, ensure local IO is flushed out.
6692 static void io_submit_state_end(struct io_submit_state *state,
6693 struct io_ring_ctx *ctx)
6695 if (state->link.head)
6696 io_queue_sqe(state->link.head);
6697 if (state->compl_nr)
6698 io_submit_flush_completions(ctx);
6699 if (state->plug_started)
6700 blk_finish_plug(&state->plug);
6704 * Start submission side cache.
6706 static void io_submit_state_start(struct io_submit_state *state,
6707 unsigned int max_ios)
6709 state->plug_started = false;
6710 state->ios_left = max_ios;
6711 /* set only head, no need to init link_last in advance */
6712 state->link.head = NULL;
6715 static void io_commit_sqring(struct io_ring_ctx *ctx)
6717 struct io_rings *rings = ctx->rings;
6720 * Ensure any loads from the SQEs are done at this point,
6721 * since once we write the new head, the application could
6722 * write new data to them.
6724 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6728 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6729 * that is mapped by userspace. This means that care needs to be taken to
6730 * ensure that reads are stable, as we cannot rely on userspace always
6731 * being a good citizen. If members of the sqe are validated and then later
6732 * used, it's important that those reads are done through READ_ONCE() to
6733 * prevent a re-load down the line.
6735 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6737 unsigned head, mask = ctx->sq_entries - 1;
6738 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6741 * The cached sq head (or cq tail) serves two purposes:
6743 * 1) allows us to batch the cost of updating the user visible
6745 * 2) allows the kernel side to track the head on its own, even
6746 * though the application is the one updating it.
6748 head = READ_ONCE(ctx->sq_array[sq_idx]);
6749 if (likely(head < ctx->sq_entries))
6750 return &ctx->sq_sqes[head];
6752 /* drop invalid entries */
6754 WRITE_ONCE(ctx->rings->sq_dropped,
6755 READ_ONCE(ctx->rings->sq_dropped) + 1);
6759 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6760 __must_hold(&ctx->uring_lock)
6762 struct io_uring_task *tctx;
6765 /* make sure SQ entry isn't read before tail */
6766 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6767 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6770 tctx = current->io_uring;
6771 tctx->cached_refs -= nr;
6772 if (unlikely(tctx->cached_refs < 0)) {
6773 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6775 percpu_counter_add(&tctx->inflight, refill);
6776 refcount_add(refill, ¤t->usage);
6777 tctx->cached_refs += refill;
6779 io_submit_state_start(&ctx->submit_state, nr);
6781 while (submitted < nr) {
6782 const struct io_uring_sqe *sqe;
6783 struct io_kiocb *req;
6785 req = io_alloc_req(ctx);
6786 if (unlikely(!req)) {
6788 submitted = -EAGAIN;
6791 sqe = io_get_sqe(ctx);
6792 if (unlikely(!sqe)) {
6793 kmem_cache_free(req_cachep, req);
6796 /* will complete beyond this point, count as submitted */
6798 if (io_submit_sqe(ctx, req, sqe))
6802 if (unlikely(submitted != nr)) {
6803 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6804 int unused = nr - ref_used;
6806 current->io_uring->cached_refs += unused;
6807 percpu_ref_put_many(&ctx->refs, unused);
6810 io_submit_state_end(&ctx->submit_state, ctx);
6811 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6812 io_commit_sqring(ctx);
6817 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6819 return READ_ONCE(sqd->state);
6822 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6824 /* Tell userspace we may need a wakeup call */
6825 spin_lock(&ctx->completion_lock);
6826 WRITE_ONCE(ctx->rings->sq_flags,
6827 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6828 spin_unlock(&ctx->completion_lock);
6831 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6833 spin_lock(&ctx->completion_lock);
6834 WRITE_ONCE(ctx->rings->sq_flags,
6835 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6836 spin_unlock(&ctx->completion_lock);
6839 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6841 unsigned int to_submit;
6844 to_submit = io_sqring_entries(ctx);
6845 /* if we're handling multiple rings, cap submit size for fairness */
6846 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6847 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6849 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6850 unsigned nr_events = 0;
6851 const struct cred *creds = NULL;
6853 if (ctx->sq_creds != current_cred())
6854 creds = override_creds(ctx->sq_creds);
6856 mutex_lock(&ctx->uring_lock);
6857 if (!list_empty(&ctx->iopoll_list))
6858 io_do_iopoll(ctx, &nr_events, 0);
6861 * Don't submit if refs are dying, good for io_uring_register(),
6862 * but also it is relied upon by io_ring_exit_work()
6864 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6865 !(ctx->flags & IORING_SETUP_R_DISABLED))
6866 ret = io_submit_sqes(ctx, to_submit);
6867 mutex_unlock(&ctx->uring_lock);
6869 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6870 wake_up(&ctx->sqo_sq_wait);
6872 revert_creds(creds);
6878 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6880 struct io_ring_ctx *ctx;
6881 unsigned sq_thread_idle = 0;
6883 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6884 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6885 sqd->sq_thread_idle = sq_thread_idle;
6888 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6890 bool did_sig = false;
6891 struct ksignal ksig;
6893 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6894 signal_pending(current)) {
6895 mutex_unlock(&sqd->lock);
6896 if (signal_pending(current))
6897 did_sig = get_signal(&ksig);
6899 mutex_lock(&sqd->lock);
6901 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6904 static int io_sq_thread(void *data)
6906 struct io_sq_data *sqd = data;
6907 struct io_ring_ctx *ctx;
6908 unsigned long timeout = 0;
6909 char buf[TASK_COMM_LEN];
6912 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6913 set_task_comm(current, buf);
6915 if (sqd->sq_cpu != -1)
6916 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6918 set_cpus_allowed_ptr(current, cpu_online_mask);
6919 current->flags |= PF_NO_SETAFFINITY;
6921 mutex_lock(&sqd->lock);
6923 bool cap_entries, sqt_spin = false;
6925 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6926 if (io_sqd_handle_event(sqd))
6928 timeout = jiffies + sqd->sq_thread_idle;
6931 cap_entries = !list_is_singular(&sqd->ctx_list);
6932 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6933 int ret = __io_sq_thread(ctx, cap_entries);
6935 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6938 if (io_run_task_work())
6941 if (sqt_spin || !time_after(jiffies, timeout)) {
6944 timeout = jiffies + sqd->sq_thread_idle;
6948 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6949 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6950 bool needs_sched = true;
6952 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6953 io_ring_set_wakeup_flag(ctx);
6955 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6956 !list_empty_careful(&ctx->iopoll_list)) {
6957 needs_sched = false;
6960 if (io_sqring_entries(ctx)) {
6961 needs_sched = false;
6967 mutex_unlock(&sqd->lock);
6969 mutex_lock(&sqd->lock);
6971 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6972 io_ring_clear_wakeup_flag(ctx);
6975 finish_wait(&sqd->wait, &wait);
6976 timeout = jiffies + sqd->sq_thread_idle;
6979 io_uring_cancel_generic(true, sqd);
6981 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6982 io_ring_set_wakeup_flag(ctx);
6984 mutex_unlock(&sqd->lock);
6986 complete(&sqd->exited);
6990 struct io_wait_queue {
6991 struct wait_queue_entry wq;
6992 struct io_ring_ctx *ctx;
6994 unsigned nr_timeouts;
6997 static inline bool io_should_wake(struct io_wait_queue *iowq)
6999 struct io_ring_ctx *ctx = iowq->ctx;
7000 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7003 * Wake up if we have enough events, or if a timeout occurred since we
7004 * started waiting. For timeouts, we always want to return to userspace,
7005 * regardless of event count.
7007 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7010 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7011 int wake_flags, void *key)
7013 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7017 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7018 * the task, and the next invocation will do it.
7020 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7021 return autoremove_wake_function(curr, mode, wake_flags, key);
7025 static int io_run_task_work_sig(void)
7027 if (io_run_task_work())
7029 if (!signal_pending(current))
7031 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7032 return -ERESTARTSYS;
7036 /* when returns >0, the caller should retry */
7037 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7038 struct io_wait_queue *iowq,
7039 signed long *timeout)
7043 /* make sure we run task_work before checking for signals */
7044 ret = io_run_task_work_sig();
7045 if (ret || io_should_wake(iowq))
7047 /* let the caller flush overflows, retry */
7048 if (test_bit(0, &ctx->check_cq_overflow))
7051 *timeout = schedule_timeout(*timeout);
7052 return !*timeout ? -ETIME : 1;
7056 * Wait until events become available, if we don't already have some. The
7057 * application must reap them itself, as they reside on the shared cq ring.
7059 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7060 const sigset_t __user *sig, size_t sigsz,
7061 struct __kernel_timespec __user *uts)
7063 struct io_wait_queue iowq;
7064 struct io_rings *rings = ctx->rings;
7065 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7069 io_cqring_overflow_flush(ctx);
7070 if (io_cqring_events(ctx) >= min_events)
7072 if (!io_run_task_work())
7077 #ifdef CONFIG_COMPAT
7078 if (in_compat_syscall())
7079 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7083 ret = set_user_sigmask(sig, sigsz);
7090 struct timespec64 ts;
7092 if (get_timespec64(&ts, uts))
7094 timeout = timespec64_to_jiffies(&ts);
7097 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7098 iowq.wq.private = current;
7099 INIT_LIST_HEAD(&iowq.wq.entry);
7101 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7102 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7104 trace_io_uring_cqring_wait(ctx, min_events);
7106 /* if we can't even flush overflow, don't wait for more */
7107 if (!io_cqring_overflow_flush(ctx)) {
7111 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7112 TASK_INTERRUPTIBLE);
7113 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7114 finish_wait(&ctx->cq_wait, &iowq.wq);
7118 restore_saved_sigmask_unless(ret == -EINTR);
7120 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7123 static void io_free_page_table(void **table, size_t size)
7125 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7127 for (i = 0; i < nr_tables; i++)
7132 static void **io_alloc_page_table(size_t size)
7134 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7135 size_t init_size = size;
7138 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7142 for (i = 0; i < nr_tables; i++) {
7143 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7145 table[i] = kzalloc(this_size, GFP_KERNEL);
7147 io_free_page_table(table, init_size);
7155 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7157 percpu_ref_exit(&ref_node->refs);
7161 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7163 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7164 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7165 unsigned long flags;
7166 bool first_add = false;
7168 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7171 while (!list_empty(&ctx->rsrc_ref_list)) {
7172 node = list_first_entry(&ctx->rsrc_ref_list,
7173 struct io_rsrc_node, node);
7174 /* recycle ref nodes in order */
7177 list_del(&node->node);
7178 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7180 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7183 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7186 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7188 struct io_rsrc_node *ref_node;
7190 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7194 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7199 INIT_LIST_HEAD(&ref_node->node);
7200 INIT_LIST_HEAD(&ref_node->rsrc_list);
7201 ref_node->done = false;
7205 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7206 struct io_rsrc_data *data_to_kill)
7208 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7209 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7212 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7214 rsrc_node->rsrc_data = data_to_kill;
7215 spin_lock_irq(&ctx->rsrc_ref_lock);
7216 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7217 spin_unlock_irq(&ctx->rsrc_ref_lock);
7219 atomic_inc(&data_to_kill->refs);
7220 percpu_ref_kill(&rsrc_node->refs);
7221 ctx->rsrc_node = NULL;
7224 if (!ctx->rsrc_node) {
7225 ctx->rsrc_node = ctx->rsrc_backup_node;
7226 ctx->rsrc_backup_node = NULL;
7230 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7232 if (ctx->rsrc_backup_node)
7234 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7235 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7238 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7242 /* As we may drop ->uring_lock, other task may have started quiesce */
7246 data->quiesce = true;
7248 ret = io_rsrc_node_switch_start(ctx);
7251 io_rsrc_node_switch(ctx, data);
7253 /* kill initial ref, already quiesced if zero */
7254 if (atomic_dec_and_test(&data->refs))
7256 mutex_unlock(&ctx->uring_lock);
7257 flush_delayed_work(&ctx->rsrc_put_work);
7258 ret = wait_for_completion_interruptible(&data->done);
7260 mutex_lock(&ctx->uring_lock);
7264 atomic_inc(&data->refs);
7265 /* wait for all works potentially completing data->done */
7266 flush_delayed_work(&ctx->rsrc_put_work);
7267 reinit_completion(&data->done);
7269 ret = io_run_task_work_sig();
7270 mutex_lock(&ctx->uring_lock);
7272 data->quiesce = false;
7277 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7279 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7280 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7282 return &data->tags[table_idx][off];
7285 static void io_rsrc_data_free(struct io_rsrc_data *data)
7287 size_t size = data->nr * sizeof(data->tags[0][0]);
7290 io_free_page_table((void **)data->tags, size);
7294 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7295 u64 __user *utags, unsigned nr,
7296 struct io_rsrc_data **pdata)
7298 struct io_rsrc_data *data;
7302 data = kzalloc(sizeof(*data), GFP_KERNEL);
7305 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7313 data->do_put = do_put;
7316 for (i = 0; i < nr; i++) {
7317 u64 *tag_slot = io_get_tag_slot(data, i);
7319 if (copy_from_user(tag_slot, &utags[i],
7325 atomic_set(&data->refs, 1);
7326 init_completion(&data->done);
7330 io_rsrc_data_free(data);
7334 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7336 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7337 return !!table->files;
7340 static void io_free_file_tables(struct io_file_table *table)
7342 kvfree(table->files);
7343 table->files = NULL;
7346 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7348 #if defined(CONFIG_UNIX)
7349 if (ctx->ring_sock) {
7350 struct sock *sock = ctx->ring_sock->sk;
7351 struct sk_buff *skb;
7353 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7359 for (i = 0; i < ctx->nr_user_files; i++) {
7362 file = io_file_from_index(ctx, i);
7367 io_free_file_tables(&ctx->file_table);
7368 io_rsrc_data_free(ctx->file_data);
7369 ctx->file_data = NULL;
7370 ctx->nr_user_files = 0;
7373 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7377 if (!ctx->file_data)
7379 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7381 __io_sqe_files_unregister(ctx);
7385 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7386 __releases(&sqd->lock)
7388 WARN_ON_ONCE(sqd->thread == current);
7391 * Do the dance but not conditional clear_bit() because it'd race with
7392 * other threads incrementing park_pending and setting the bit.
7394 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7395 if (atomic_dec_return(&sqd->park_pending))
7396 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7397 mutex_unlock(&sqd->lock);
7400 static void io_sq_thread_park(struct io_sq_data *sqd)
7401 __acquires(&sqd->lock)
7403 WARN_ON_ONCE(sqd->thread == current);
7405 atomic_inc(&sqd->park_pending);
7406 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7407 mutex_lock(&sqd->lock);
7409 wake_up_process(sqd->thread);
7412 static void io_sq_thread_stop(struct io_sq_data *sqd)
7414 WARN_ON_ONCE(sqd->thread == current);
7415 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7417 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7418 mutex_lock(&sqd->lock);
7420 wake_up_process(sqd->thread);
7421 mutex_unlock(&sqd->lock);
7422 wait_for_completion(&sqd->exited);
7425 static void io_put_sq_data(struct io_sq_data *sqd)
7427 if (refcount_dec_and_test(&sqd->refs)) {
7428 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7430 io_sq_thread_stop(sqd);
7435 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7437 struct io_sq_data *sqd = ctx->sq_data;
7440 io_sq_thread_park(sqd);
7441 list_del_init(&ctx->sqd_list);
7442 io_sqd_update_thread_idle(sqd);
7443 io_sq_thread_unpark(sqd);
7445 io_put_sq_data(sqd);
7446 ctx->sq_data = NULL;
7450 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7452 struct io_ring_ctx *ctx_attach;
7453 struct io_sq_data *sqd;
7456 f = fdget(p->wq_fd);
7458 return ERR_PTR(-ENXIO);
7459 if (f.file->f_op != &io_uring_fops) {
7461 return ERR_PTR(-EINVAL);
7464 ctx_attach = f.file->private_data;
7465 sqd = ctx_attach->sq_data;
7468 return ERR_PTR(-EINVAL);
7470 if (sqd->task_tgid != current->tgid) {
7472 return ERR_PTR(-EPERM);
7475 refcount_inc(&sqd->refs);
7480 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7483 struct io_sq_data *sqd;
7486 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7487 sqd = io_attach_sq_data(p);
7492 /* fall through for EPERM case, setup new sqd/task */
7493 if (PTR_ERR(sqd) != -EPERM)
7497 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7499 return ERR_PTR(-ENOMEM);
7501 atomic_set(&sqd->park_pending, 0);
7502 refcount_set(&sqd->refs, 1);
7503 INIT_LIST_HEAD(&sqd->ctx_list);
7504 mutex_init(&sqd->lock);
7505 init_waitqueue_head(&sqd->wait);
7506 init_completion(&sqd->exited);
7510 #if defined(CONFIG_UNIX)
7512 * Ensure the UNIX gc is aware of our file set, so we are certain that
7513 * the io_uring can be safely unregistered on process exit, even if we have
7514 * loops in the file referencing.
7516 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7518 struct sock *sk = ctx->ring_sock->sk;
7519 struct scm_fp_list *fpl;
7520 struct sk_buff *skb;
7523 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7527 skb = alloc_skb(0, GFP_KERNEL);
7536 fpl->user = get_uid(current_user());
7537 for (i = 0; i < nr; i++) {
7538 struct file *file = io_file_from_index(ctx, i + offset);
7542 fpl->fp[nr_files] = get_file(file);
7543 unix_inflight(fpl->user, fpl->fp[nr_files]);
7548 fpl->max = SCM_MAX_FD;
7549 fpl->count = nr_files;
7550 UNIXCB(skb).fp = fpl;
7551 skb->destructor = unix_destruct_scm;
7552 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7553 skb_queue_head(&sk->sk_receive_queue, skb);
7555 for (i = 0; i < nr_files; i++)
7566 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7567 * causes regular reference counting to break down. We rely on the UNIX
7568 * garbage collection to take care of this problem for us.
7570 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7572 unsigned left, total;
7576 left = ctx->nr_user_files;
7578 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7580 ret = __io_sqe_files_scm(ctx, this_files, total);
7584 total += this_files;
7590 while (total < ctx->nr_user_files) {
7591 struct file *file = io_file_from_index(ctx, total);
7601 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7607 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7609 struct file *file = prsrc->file;
7610 #if defined(CONFIG_UNIX)
7611 struct sock *sock = ctx->ring_sock->sk;
7612 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7613 struct sk_buff *skb;
7616 __skb_queue_head_init(&list);
7619 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7620 * remove this entry and rearrange the file array.
7622 skb = skb_dequeue(head);
7624 struct scm_fp_list *fp;
7626 fp = UNIXCB(skb).fp;
7627 for (i = 0; i < fp->count; i++) {
7630 if (fp->fp[i] != file)
7633 unix_notinflight(fp->user, fp->fp[i]);
7634 left = fp->count - 1 - i;
7636 memmove(&fp->fp[i], &fp->fp[i + 1],
7637 left * sizeof(struct file *));
7644 __skb_queue_tail(&list, skb);
7654 __skb_queue_tail(&list, skb);
7656 skb = skb_dequeue(head);
7659 if (skb_peek(&list)) {
7660 spin_lock_irq(&head->lock);
7661 while ((skb = __skb_dequeue(&list)) != NULL)
7662 __skb_queue_tail(head, skb);
7663 spin_unlock_irq(&head->lock);
7670 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7672 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7673 struct io_ring_ctx *ctx = rsrc_data->ctx;
7674 struct io_rsrc_put *prsrc, *tmp;
7676 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7677 list_del(&prsrc->list);
7680 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7682 io_ring_submit_lock(ctx, lock_ring);
7683 spin_lock(&ctx->completion_lock);
7684 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7686 io_commit_cqring(ctx);
7687 spin_unlock(&ctx->completion_lock);
7688 io_cqring_ev_posted(ctx);
7689 io_ring_submit_unlock(ctx, lock_ring);
7692 rsrc_data->do_put(ctx, prsrc);
7696 io_rsrc_node_destroy(ref_node);
7697 if (atomic_dec_and_test(&rsrc_data->refs))
7698 complete(&rsrc_data->done);
7701 static void io_rsrc_put_work(struct work_struct *work)
7703 struct io_ring_ctx *ctx;
7704 struct llist_node *node;
7706 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7707 node = llist_del_all(&ctx->rsrc_put_llist);
7710 struct io_rsrc_node *ref_node;
7711 struct llist_node *next = node->next;
7713 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7714 __io_rsrc_put_work(ref_node);
7719 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7720 unsigned nr_args, u64 __user *tags)
7722 __s32 __user *fds = (__s32 __user *) arg;
7731 if (nr_args > IORING_MAX_FIXED_FILES)
7733 ret = io_rsrc_node_switch_start(ctx);
7736 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7742 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7745 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7746 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7750 /* allow sparse sets */
7753 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7760 if (unlikely(!file))
7764 * Don't allow io_uring instances to be registered. If UNIX
7765 * isn't enabled, then this causes a reference cycle and this
7766 * instance can never get freed. If UNIX is enabled we'll
7767 * handle it just fine, but there's still no point in allowing
7768 * a ring fd as it doesn't support regular read/write anyway.
7770 if (file->f_op == &io_uring_fops) {
7774 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7777 ret = io_sqe_files_scm(ctx);
7779 __io_sqe_files_unregister(ctx);
7783 io_rsrc_node_switch(ctx, NULL);
7786 for (i = 0; i < ctx->nr_user_files; i++) {
7787 file = io_file_from_index(ctx, i);
7791 io_free_file_tables(&ctx->file_table);
7792 ctx->nr_user_files = 0;
7794 io_rsrc_data_free(ctx->file_data);
7795 ctx->file_data = NULL;
7799 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7802 #if defined(CONFIG_UNIX)
7803 struct sock *sock = ctx->ring_sock->sk;
7804 struct sk_buff_head *head = &sock->sk_receive_queue;
7805 struct sk_buff *skb;
7808 * See if we can merge this file into an existing skb SCM_RIGHTS
7809 * file set. If there's no room, fall back to allocating a new skb
7810 * and filling it in.
7812 spin_lock_irq(&head->lock);
7813 skb = skb_peek(head);
7815 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7817 if (fpl->count < SCM_MAX_FD) {
7818 __skb_unlink(skb, head);
7819 spin_unlock_irq(&head->lock);
7820 fpl->fp[fpl->count] = get_file(file);
7821 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7823 spin_lock_irq(&head->lock);
7824 __skb_queue_head(head, skb);
7829 spin_unlock_irq(&head->lock);
7836 return __io_sqe_files_scm(ctx, 1, index);
7842 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7843 struct io_rsrc_node *node, void *rsrc)
7845 struct io_rsrc_put *prsrc;
7847 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7851 prsrc->tag = *io_get_tag_slot(data, idx);
7853 list_add(&prsrc->list, &node->rsrc_list);
7857 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7858 struct io_uring_rsrc_update2 *up,
7861 u64 __user *tags = u64_to_user_ptr(up->tags);
7862 __s32 __user *fds = u64_to_user_ptr(up->data);
7863 struct io_rsrc_data *data = ctx->file_data;
7864 struct io_fixed_file *file_slot;
7868 bool needs_switch = false;
7870 if (!ctx->file_data)
7872 if (up->offset + nr_args > ctx->nr_user_files)
7875 for (done = 0; done < nr_args; done++) {
7878 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7879 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7883 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7887 if (fd == IORING_REGISTER_FILES_SKIP)
7890 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7891 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7893 if (file_slot->file_ptr) {
7894 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7895 err = io_queue_rsrc_removal(data, up->offset + done,
7896 ctx->rsrc_node, file);
7899 file_slot->file_ptr = 0;
7900 needs_switch = true;
7909 * Don't allow io_uring instances to be registered. If
7910 * UNIX isn't enabled, then this causes a reference
7911 * cycle and this instance can never get freed. If UNIX
7912 * is enabled we'll handle it just fine, but there's
7913 * still no point in allowing a ring fd as it doesn't
7914 * support regular read/write anyway.
7916 if (file->f_op == &io_uring_fops) {
7921 *io_get_tag_slot(data, up->offset + done) = tag;
7922 io_fixed_file_set(file_slot, file);
7923 err = io_sqe_file_register(ctx, file, i);
7925 file_slot->file_ptr = 0;
7933 io_rsrc_node_switch(ctx, data);
7934 return done ? done : err;
7937 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7938 struct task_struct *task)
7940 struct io_wq_hash *hash;
7941 struct io_wq_data data;
7942 unsigned int concurrency;
7944 mutex_lock(&ctx->uring_lock);
7945 hash = ctx->hash_map;
7947 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7949 mutex_unlock(&ctx->uring_lock);
7950 return ERR_PTR(-ENOMEM);
7952 refcount_set(&hash->refs, 1);
7953 init_waitqueue_head(&hash->wait);
7954 ctx->hash_map = hash;
7956 mutex_unlock(&ctx->uring_lock);
7960 data.free_work = io_wq_free_work;
7961 data.do_work = io_wq_submit_work;
7963 /* Do QD, or 4 * CPUS, whatever is smallest */
7964 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7966 return io_wq_create(concurrency, &data);
7969 static int io_uring_alloc_task_context(struct task_struct *task,
7970 struct io_ring_ctx *ctx)
7972 struct io_uring_task *tctx;
7975 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7976 if (unlikely(!tctx))
7979 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7980 if (unlikely(ret)) {
7985 tctx->io_wq = io_init_wq_offload(ctx, task);
7986 if (IS_ERR(tctx->io_wq)) {
7987 ret = PTR_ERR(tctx->io_wq);
7988 percpu_counter_destroy(&tctx->inflight);
7994 init_waitqueue_head(&tctx->wait);
7995 atomic_set(&tctx->in_idle, 0);
7996 atomic_set(&tctx->inflight_tracked, 0);
7997 task->io_uring = tctx;
7998 spin_lock_init(&tctx->task_lock);
7999 INIT_WQ_LIST(&tctx->task_list);
8000 init_task_work(&tctx->task_work, tctx_task_work);
8004 void __io_uring_free(struct task_struct *tsk)
8006 struct io_uring_task *tctx = tsk->io_uring;
8008 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8009 WARN_ON_ONCE(tctx->io_wq);
8010 WARN_ON_ONCE(tctx->cached_refs);
8012 percpu_counter_destroy(&tctx->inflight);
8014 tsk->io_uring = NULL;
8017 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8018 struct io_uring_params *p)
8022 /* Retain compatibility with failing for an invalid attach attempt */
8023 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8024 IORING_SETUP_ATTACH_WQ) {
8027 f = fdget(p->wq_fd);
8030 if (f.file->f_op != &io_uring_fops) {
8036 if (ctx->flags & IORING_SETUP_SQPOLL) {
8037 struct task_struct *tsk;
8038 struct io_sq_data *sqd;
8041 sqd = io_get_sq_data(p, &attached);
8047 ctx->sq_creds = get_current_cred();
8049 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8050 if (!ctx->sq_thread_idle)
8051 ctx->sq_thread_idle = HZ;
8053 io_sq_thread_park(sqd);
8054 list_add(&ctx->sqd_list, &sqd->ctx_list);
8055 io_sqd_update_thread_idle(sqd);
8056 /* don't attach to a dying SQPOLL thread, would be racy */
8057 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8058 io_sq_thread_unpark(sqd);
8065 if (p->flags & IORING_SETUP_SQ_AFF) {
8066 int cpu = p->sq_thread_cpu;
8069 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8076 sqd->task_pid = current->pid;
8077 sqd->task_tgid = current->tgid;
8078 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8085 ret = io_uring_alloc_task_context(tsk, ctx);
8086 wake_up_new_task(tsk);
8089 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8090 /* Can't have SQ_AFF without SQPOLL */
8097 complete(&ctx->sq_data->exited);
8099 io_sq_thread_finish(ctx);
8103 static inline void __io_unaccount_mem(struct user_struct *user,
8104 unsigned long nr_pages)
8106 atomic_long_sub(nr_pages, &user->locked_vm);
8109 static inline int __io_account_mem(struct user_struct *user,
8110 unsigned long nr_pages)
8112 unsigned long page_limit, cur_pages, new_pages;
8114 /* Don't allow more pages than we can safely lock */
8115 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8118 cur_pages = atomic_long_read(&user->locked_vm);
8119 new_pages = cur_pages + nr_pages;
8120 if (new_pages > page_limit)
8122 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8123 new_pages) != cur_pages);
8128 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8131 __io_unaccount_mem(ctx->user, nr_pages);
8133 if (ctx->mm_account)
8134 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8137 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8142 ret = __io_account_mem(ctx->user, nr_pages);
8147 if (ctx->mm_account)
8148 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8153 static void io_mem_free(void *ptr)
8160 page = virt_to_head_page(ptr);
8161 if (put_page_testzero(page))
8162 free_compound_page(page);
8165 static void *io_mem_alloc(size_t size)
8167 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8168 __GFP_NORETRY | __GFP_ACCOUNT;
8170 return (void *) __get_free_pages(gfp_flags, get_order(size));
8173 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8176 struct io_rings *rings;
8177 size_t off, sq_array_size;
8179 off = struct_size(rings, cqes, cq_entries);
8180 if (off == SIZE_MAX)
8184 off = ALIGN(off, SMP_CACHE_BYTES);
8192 sq_array_size = array_size(sizeof(u32), sq_entries);
8193 if (sq_array_size == SIZE_MAX)
8196 if (check_add_overflow(off, sq_array_size, &off))
8202 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8204 struct io_mapped_ubuf *imu = *slot;
8207 if (imu != ctx->dummy_ubuf) {
8208 for (i = 0; i < imu->nr_bvecs; i++)
8209 unpin_user_page(imu->bvec[i].bv_page);
8210 if (imu->acct_pages)
8211 io_unaccount_mem(ctx, imu->acct_pages);
8217 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8219 io_buffer_unmap(ctx, &prsrc->buf);
8223 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8227 for (i = 0; i < ctx->nr_user_bufs; i++)
8228 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8229 kfree(ctx->user_bufs);
8230 io_rsrc_data_free(ctx->buf_data);
8231 ctx->user_bufs = NULL;
8232 ctx->buf_data = NULL;
8233 ctx->nr_user_bufs = 0;
8236 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8243 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8245 __io_sqe_buffers_unregister(ctx);
8249 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8250 void __user *arg, unsigned index)
8252 struct iovec __user *src;
8254 #ifdef CONFIG_COMPAT
8256 struct compat_iovec __user *ciovs;
8257 struct compat_iovec ciov;
8259 ciovs = (struct compat_iovec __user *) arg;
8260 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8263 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8264 dst->iov_len = ciov.iov_len;
8268 src = (struct iovec __user *) arg;
8269 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8275 * Not super efficient, but this is just a registration time. And we do cache
8276 * the last compound head, so generally we'll only do a full search if we don't
8279 * We check if the given compound head page has already been accounted, to
8280 * avoid double accounting it. This allows us to account the full size of the
8281 * page, not just the constituent pages of a huge page.
8283 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8284 int nr_pages, struct page *hpage)
8288 /* check current page array */
8289 for (i = 0; i < nr_pages; i++) {
8290 if (!PageCompound(pages[i]))
8292 if (compound_head(pages[i]) == hpage)
8296 /* check previously registered pages */
8297 for (i = 0; i < ctx->nr_user_bufs; i++) {
8298 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8300 for (j = 0; j < imu->nr_bvecs; j++) {
8301 if (!PageCompound(imu->bvec[j].bv_page))
8303 if (compound_head(imu->bvec[j].bv_page) == hpage)
8311 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8312 int nr_pages, struct io_mapped_ubuf *imu,
8313 struct page **last_hpage)
8317 imu->acct_pages = 0;
8318 for (i = 0; i < nr_pages; i++) {
8319 if (!PageCompound(pages[i])) {
8324 hpage = compound_head(pages[i]);
8325 if (hpage == *last_hpage)
8327 *last_hpage = hpage;
8328 if (headpage_already_acct(ctx, pages, i, hpage))
8330 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8334 if (!imu->acct_pages)
8337 ret = io_account_mem(ctx, imu->acct_pages);
8339 imu->acct_pages = 0;
8343 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8344 struct io_mapped_ubuf **pimu,
8345 struct page **last_hpage)
8347 struct io_mapped_ubuf *imu = NULL;
8348 struct vm_area_struct **vmas = NULL;
8349 struct page **pages = NULL;
8350 unsigned long off, start, end, ubuf;
8352 int ret, pret, nr_pages, i;
8354 if (!iov->iov_base) {
8355 *pimu = ctx->dummy_ubuf;
8359 ubuf = (unsigned long) iov->iov_base;
8360 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8361 start = ubuf >> PAGE_SHIFT;
8362 nr_pages = end - start;
8367 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8371 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8376 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8381 mmap_read_lock(current->mm);
8382 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8384 if (pret == nr_pages) {
8385 /* don't support file backed memory */
8386 for (i = 0; i < nr_pages; i++) {
8387 struct vm_area_struct *vma = vmas[i];
8389 if (vma_is_shmem(vma))
8392 !is_file_hugepages(vma->vm_file)) {
8398 ret = pret < 0 ? pret : -EFAULT;
8400 mmap_read_unlock(current->mm);
8403 * if we did partial map, or found file backed vmas,
8404 * release any pages we did get
8407 unpin_user_pages(pages, pret);
8411 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8413 unpin_user_pages(pages, pret);
8417 off = ubuf & ~PAGE_MASK;
8418 size = iov->iov_len;
8419 for (i = 0; i < nr_pages; i++) {
8422 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8423 imu->bvec[i].bv_page = pages[i];
8424 imu->bvec[i].bv_len = vec_len;
8425 imu->bvec[i].bv_offset = off;
8429 /* store original address for later verification */
8431 imu->ubuf_end = ubuf + iov->iov_len;
8432 imu->nr_bvecs = nr_pages;
8443 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8445 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8446 return ctx->user_bufs ? 0 : -ENOMEM;
8449 static int io_buffer_validate(struct iovec *iov)
8451 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8454 * Don't impose further limits on the size and buffer
8455 * constraints here, we'll -EINVAL later when IO is
8456 * submitted if they are wrong.
8459 return iov->iov_len ? -EFAULT : 0;
8463 /* arbitrary limit, but we need something */
8464 if (iov->iov_len > SZ_1G)
8467 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8473 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8474 unsigned int nr_args, u64 __user *tags)
8476 struct page *last_hpage = NULL;
8477 struct io_rsrc_data *data;
8483 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8485 ret = io_rsrc_node_switch_start(ctx);
8488 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8491 ret = io_buffers_map_alloc(ctx, nr_args);
8493 io_rsrc_data_free(data);
8497 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8498 ret = io_copy_iov(ctx, &iov, arg, i);
8501 ret = io_buffer_validate(&iov);
8504 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8509 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8515 WARN_ON_ONCE(ctx->buf_data);
8517 ctx->buf_data = data;
8519 __io_sqe_buffers_unregister(ctx);
8521 io_rsrc_node_switch(ctx, NULL);
8525 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8526 struct io_uring_rsrc_update2 *up,
8527 unsigned int nr_args)
8529 u64 __user *tags = u64_to_user_ptr(up->tags);
8530 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8531 struct page *last_hpage = NULL;
8532 bool needs_switch = false;
8538 if (up->offset + nr_args > ctx->nr_user_bufs)
8541 for (done = 0; done < nr_args; done++) {
8542 struct io_mapped_ubuf *imu;
8543 int offset = up->offset + done;
8546 err = io_copy_iov(ctx, &iov, iovs, done);
8549 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8553 err = io_buffer_validate(&iov);
8556 if (!iov.iov_base && tag) {
8560 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8564 i = array_index_nospec(offset, ctx->nr_user_bufs);
8565 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8566 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8567 ctx->rsrc_node, ctx->user_bufs[i]);
8568 if (unlikely(err)) {
8569 io_buffer_unmap(ctx, &imu);
8572 ctx->user_bufs[i] = NULL;
8573 needs_switch = true;
8576 ctx->user_bufs[i] = imu;
8577 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8581 io_rsrc_node_switch(ctx, ctx->buf_data);
8582 return done ? done : err;
8585 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8587 __s32 __user *fds = arg;
8593 if (copy_from_user(&fd, fds, sizeof(*fds)))
8596 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8597 if (IS_ERR(ctx->cq_ev_fd)) {
8598 int ret = PTR_ERR(ctx->cq_ev_fd);
8600 ctx->cq_ev_fd = NULL;
8607 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8609 if (ctx->cq_ev_fd) {
8610 eventfd_ctx_put(ctx->cq_ev_fd);
8611 ctx->cq_ev_fd = NULL;
8618 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8620 struct io_buffer *buf;
8621 unsigned long index;
8623 xa_for_each(&ctx->io_buffers, index, buf)
8624 __io_remove_buffers(ctx, buf, index, -1U);
8627 static void io_req_cache_free(struct list_head *list)
8629 struct io_kiocb *req, *nxt;
8631 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8632 list_del(&req->inflight_entry);
8633 kmem_cache_free(req_cachep, req);
8637 static void io_req_caches_free(struct io_ring_ctx *ctx)
8639 struct io_submit_state *state = &ctx->submit_state;
8641 mutex_lock(&ctx->uring_lock);
8643 if (state->free_reqs) {
8644 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8645 state->free_reqs = 0;
8648 io_flush_cached_locked_reqs(ctx, state);
8649 io_req_cache_free(&state->free_list);
8650 mutex_unlock(&ctx->uring_lock);
8653 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8655 if (data && !atomic_dec_and_test(&data->refs))
8656 wait_for_completion(&data->done);
8659 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8661 io_sq_thread_finish(ctx);
8663 if (ctx->mm_account) {
8664 mmdrop(ctx->mm_account);
8665 ctx->mm_account = NULL;
8668 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8669 io_wait_rsrc_data(ctx->buf_data);
8670 io_wait_rsrc_data(ctx->file_data);
8672 mutex_lock(&ctx->uring_lock);
8674 __io_sqe_buffers_unregister(ctx);
8676 __io_sqe_files_unregister(ctx);
8678 __io_cqring_overflow_flush(ctx, true);
8679 mutex_unlock(&ctx->uring_lock);
8680 io_eventfd_unregister(ctx);
8681 io_destroy_buffers(ctx);
8683 put_cred(ctx->sq_creds);
8685 /* there are no registered resources left, nobody uses it */
8687 io_rsrc_node_destroy(ctx->rsrc_node);
8688 if (ctx->rsrc_backup_node)
8689 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8690 flush_delayed_work(&ctx->rsrc_put_work);
8692 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8693 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8695 #if defined(CONFIG_UNIX)
8696 if (ctx->ring_sock) {
8697 ctx->ring_sock->file = NULL; /* so that iput() is called */
8698 sock_release(ctx->ring_sock);
8702 io_mem_free(ctx->rings);
8703 io_mem_free(ctx->sq_sqes);
8705 percpu_ref_exit(&ctx->refs);
8706 free_uid(ctx->user);
8707 io_req_caches_free(ctx);
8709 io_wq_put_hash(ctx->hash_map);
8710 kfree(ctx->cancel_hash);
8711 kfree(ctx->dummy_ubuf);
8715 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8717 struct io_ring_ctx *ctx = file->private_data;
8720 poll_wait(file, &ctx->poll_wait, wait);
8722 * synchronizes with barrier from wq_has_sleeper call in
8726 if (!io_sqring_full(ctx))
8727 mask |= EPOLLOUT | EPOLLWRNORM;
8730 * Don't flush cqring overflow list here, just do a simple check.
8731 * Otherwise there could possible be ABBA deadlock:
8734 * lock(&ctx->uring_lock);
8736 * lock(&ctx->uring_lock);
8739 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8740 * pushs them to do the flush.
8742 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8743 mask |= EPOLLIN | EPOLLRDNORM;
8748 static int io_uring_fasync(int fd, struct file *file, int on)
8750 struct io_ring_ctx *ctx = file->private_data;
8752 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8755 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8757 const struct cred *creds;
8759 creds = xa_erase(&ctx->personalities, id);
8768 struct io_tctx_exit {
8769 struct callback_head task_work;
8770 struct completion completion;
8771 struct io_ring_ctx *ctx;
8774 static void io_tctx_exit_cb(struct callback_head *cb)
8776 struct io_uring_task *tctx = current->io_uring;
8777 struct io_tctx_exit *work;
8779 work = container_of(cb, struct io_tctx_exit, task_work);
8781 * When @in_idle, we're in cancellation and it's racy to remove the
8782 * node. It'll be removed by the end of cancellation, just ignore it.
8784 if (!atomic_read(&tctx->in_idle))
8785 io_uring_del_tctx_node((unsigned long)work->ctx);
8786 complete(&work->completion);
8789 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8791 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8793 return req->ctx == data;
8796 static void io_ring_exit_work(struct work_struct *work)
8798 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8799 unsigned long timeout = jiffies + HZ * 60 * 5;
8800 unsigned long interval = HZ / 20;
8801 struct io_tctx_exit exit;
8802 struct io_tctx_node *node;
8806 * If we're doing polled IO and end up having requests being
8807 * submitted async (out-of-line), then completions can come in while
8808 * we're waiting for refs to drop. We need to reap these manually,
8809 * as nobody else will be looking for them.
8812 io_uring_try_cancel_requests(ctx, NULL, true);
8814 struct io_sq_data *sqd = ctx->sq_data;
8815 struct task_struct *tsk;
8817 io_sq_thread_park(sqd);
8819 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8820 io_wq_cancel_cb(tsk->io_uring->io_wq,
8821 io_cancel_ctx_cb, ctx, true);
8822 io_sq_thread_unpark(sqd);
8825 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8826 /* there is little hope left, don't run it too often */
8829 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8831 init_completion(&exit.completion);
8832 init_task_work(&exit.task_work, io_tctx_exit_cb);
8835 * Some may use context even when all refs and requests have been put,
8836 * and they are free to do so while still holding uring_lock or
8837 * completion_lock, see io_req_task_submit(). Apart from other work,
8838 * this lock/unlock section also waits them to finish.
8840 mutex_lock(&ctx->uring_lock);
8841 while (!list_empty(&ctx->tctx_list)) {
8842 WARN_ON_ONCE(time_after(jiffies, timeout));
8844 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8846 /* don't spin on a single task if cancellation failed */
8847 list_rotate_left(&ctx->tctx_list);
8848 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8849 if (WARN_ON_ONCE(ret))
8851 wake_up_process(node->task);
8853 mutex_unlock(&ctx->uring_lock);
8854 wait_for_completion(&exit.completion);
8855 mutex_lock(&ctx->uring_lock);
8857 mutex_unlock(&ctx->uring_lock);
8858 spin_lock(&ctx->completion_lock);
8859 spin_unlock(&ctx->completion_lock);
8861 io_ring_ctx_free(ctx);
8864 /* Returns true if we found and killed one or more timeouts */
8865 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8868 struct io_kiocb *req, *tmp;
8871 spin_lock(&ctx->completion_lock);
8872 spin_lock_irq(&ctx->timeout_lock);
8873 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8874 if (io_match_task(req, tsk, cancel_all)) {
8875 io_kill_timeout(req, -ECANCELED);
8879 spin_unlock_irq(&ctx->timeout_lock);
8881 io_commit_cqring(ctx);
8882 spin_unlock(&ctx->completion_lock);
8884 io_cqring_ev_posted(ctx);
8885 return canceled != 0;
8888 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8890 unsigned long index;
8891 struct creds *creds;
8893 mutex_lock(&ctx->uring_lock);
8894 percpu_ref_kill(&ctx->refs);
8896 __io_cqring_overflow_flush(ctx, true);
8897 xa_for_each(&ctx->personalities, index, creds)
8898 io_unregister_personality(ctx, index);
8899 mutex_unlock(&ctx->uring_lock);
8901 io_kill_timeouts(ctx, NULL, true);
8902 io_poll_remove_all(ctx, NULL, true);
8904 /* if we failed setting up the ctx, we might not have any rings */
8905 io_iopoll_try_reap_events(ctx);
8907 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8909 * Use system_unbound_wq to avoid spawning tons of event kworkers
8910 * if we're exiting a ton of rings at the same time. It just adds
8911 * noise and overhead, there's no discernable change in runtime
8912 * over using system_wq.
8914 queue_work(system_unbound_wq, &ctx->exit_work);
8917 static int io_uring_release(struct inode *inode, struct file *file)
8919 struct io_ring_ctx *ctx = file->private_data;
8921 file->private_data = NULL;
8922 io_ring_ctx_wait_and_kill(ctx);
8926 struct io_task_cancel {
8927 struct task_struct *task;
8931 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8933 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8934 struct io_task_cancel *cancel = data;
8937 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8938 struct io_ring_ctx *ctx = req->ctx;
8940 /* protect against races with linked timeouts */
8941 spin_lock(&ctx->completion_lock);
8942 ret = io_match_task(req, cancel->task, cancel->all);
8943 spin_unlock(&ctx->completion_lock);
8945 ret = io_match_task(req, cancel->task, cancel->all);
8950 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8951 struct task_struct *task, bool cancel_all)
8953 struct io_defer_entry *de;
8956 spin_lock(&ctx->completion_lock);
8957 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8958 if (io_match_task(de->req, task, cancel_all)) {
8959 list_cut_position(&list, &ctx->defer_list, &de->list);
8963 spin_unlock(&ctx->completion_lock);
8964 if (list_empty(&list))
8967 while (!list_empty(&list)) {
8968 de = list_first_entry(&list, struct io_defer_entry, list);
8969 list_del_init(&de->list);
8970 io_req_complete_failed(de->req, -ECANCELED);
8976 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8978 struct io_tctx_node *node;
8979 enum io_wq_cancel cret;
8982 mutex_lock(&ctx->uring_lock);
8983 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8984 struct io_uring_task *tctx = node->task->io_uring;
8987 * io_wq will stay alive while we hold uring_lock, because it's
8988 * killed after ctx nodes, which requires to take the lock.
8990 if (!tctx || !tctx->io_wq)
8992 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8993 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8995 mutex_unlock(&ctx->uring_lock);
9000 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9001 struct task_struct *task,
9004 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9005 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9008 enum io_wq_cancel cret;
9012 ret |= io_uring_try_cancel_iowq(ctx);
9013 } else if (tctx && tctx->io_wq) {
9015 * Cancels requests of all rings, not only @ctx, but
9016 * it's fine as the task is in exit/exec.
9018 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9020 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9023 /* SQPOLL thread does its own polling */
9024 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9025 (ctx->sq_data && ctx->sq_data->thread == current)) {
9026 while (!list_empty_careful(&ctx->iopoll_list)) {
9027 io_iopoll_try_reap_events(ctx);
9032 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9033 ret |= io_poll_remove_all(ctx, task, cancel_all);
9034 ret |= io_kill_timeouts(ctx, task, cancel_all);
9036 ret |= io_run_task_work();
9043 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9045 struct io_uring_task *tctx = current->io_uring;
9046 struct io_tctx_node *node;
9049 if (unlikely(!tctx)) {
9050 ret = io_uring_alloc_task_context(current, ctx);
9053 tctx = current->io_uring;
9055 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9056 node = kmalloc(sizeof(*node), GFP_KERNEL);
9060 node->task = current;
9062 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9069 mutex_lock(&ctx->uring_lock);
9070 list_add(&node->ctx_node, &ctx->tctx_list);
9071 mutex_unlock(&ctx->uring_lock);
9078 * Note that this task has used io_uring. We use it for cancelation purposes.
9080 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9082 struct io_uring_task *tctx = current->io_uring;
9084 if (likely(tctx && tctx->last == ctx))
9086 return __io_uring_add_tctx_node(ctx);
9090 * Remove this io_uring_file -> task mapping.
9092 static void io_uring_del_tctx_node(unsigned long index)
9094 struct io_uring_task *tctx = current->io_uring;
9095 struct io_tctx_node *node;
9099 node = xa_erase(&tctx->xa, index);
9103 WARN_ON_ONCE(current != node->task);
9104 WARN_ON_ONCE(list_empty(&node->ctx_node));
9106 mutex_lock(&node->ctx->uring_lock);
9107 list_del(&node->ctx_node);
9108 mutex_unlock(&node->ctx->uring_lock);
9110 if (tctx->last == node->ctx)
9115 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9117 struct io_wq *wq = tctx->io_wq;
9118 struct io_tctx_node *node;
9119 unsigned long index;
9121 xa_for_each(&tctx->xa, index, node)
9122 io_uring_del_tctx_node(index);
9125 * Must be after io_uring_del_task_file() (removes nodes under
9126 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9129 io_wq_put_and_exit(wq);
9133 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9136 return atomic_read(&tctx->inflight_tracked);
9137 return percpu_counter_sum(&tctx->inflight);
9140 static void io_uring_drop_tctx_refs(struct task_struct *task)
9142 struct io_uring_task *tctx = task->io_uring;
9143 unsigned int refs = tctx->cached_refs;
9146 tctx->cached_refs = 0;
9147 percpu_counter_sub(&tctx->inflight, refs);
9148 put_task_struct_many(task, refs);
9153 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9154 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9156 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9158 struct io_uring_task *tctx = current->io_uring;
9159 struct io_ring_ctx *ctx;
9163 WARN_ON_ONCE(sqd && sqd->thread != current);
9165 if (!current->io_uring)
9168 io_wq_exit_start(tctx->io_wq);
9170 atomic_inc(&tctx->in_idle);
9172 io_uring_drop_tctx_refs(current);
9173 /* read completions before cancelations */
9174 inflight = tctx_inflight(tctx, !cancel_all);
9179 struct io_tctx_node *node;
9180 unsigned long index;
9182 xa_for_each(&tctx->xa, index, node) {
9183 /* sqpoll task will cancel all its requests */
9184 if (node->ctx->sq_data)
9186 io_uring_try_cancel_requests(node->ctx, current,
9190 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9191 io_uring_try_cancel_requests(ctx, current,
9195 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9196 io_uring_drop_tctx_refs(current);
9198 * If we've seen completions, retry without waiting. This
9199 * avoids a race where a completion comes in before we did
9200 * prepare_to_wait().
9202 if (inflight == tctx_inflight(tctx, !cancel_all))
9204 finish_wait(&tctx->wait, &wait);
9206 atomic_dec(&tctx->in_idle);
9208 io_uring_clean_tctx(tctx);
9210 /* for exec all current's requests should be gone, kill tctx */
9211 __io_uring_free(current);
9215 void __io_uring_cancel(bool cancel_all)
9217 io_uring_cancel_generic(cancel_all, NULL);
9220 static void *io_uring_validate_mmap_request(struct file *file,
9221 loff_t pgoff, size_t sz)
9223 struct io_ring_ctx *ctx = file->private_data;
9224 loff_t offset = pgoff << PAGE_SHIFT;
9229 case IORING_OFF_SQ_RING:
9230 case IORING_OFF_CQ_RING:
9233 case IORING_OFF_SQES:
9237 return ERR_PTR(-EINVAL);
9240 page = virt_to_head_page(ptr);
9241 if (sz > page_size(page))
9242 return ERR_PTR(-EINVAL);
9249 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9251 size_t sz = vma->vm_end - vma->vm_start;
9255 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9257 return PTR_ERR(ptr);
9259 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9260 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9263 #else /* !CONFIG_MMU */
9265 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9267 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9270 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9272 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9275 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9276 unsigned long addr, unsigned long len,
9277 unsigned long pgoff, unsigned long flags)
9281 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9283 return PTR_ERR(ptr);
9285 return (unsigned long) ptr;
9288 #endif /* !CONFIG_MMU */
9290 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9295 if (!io_sqring_full(ctx))
9297 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9299 if (!io_sqring_full(ctx))
9302 } while (!signal_pending(current));
9304 finish_wait(&ctx->sqo_sq_wait, &wait);
9308 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9309 struct __kernel_timespec __user **ts,
9310 const sigset_t __user **sig)
9312 struct io_uring_getevents_arg arg;
9315 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9316 * is just a pointer to the sigset_t.
9318 if (!(flags & IORING_ENTER_EXT_ARG)) {
9319 *sig = (const sigset_t __user *) argp;
9325 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9326 * timespec and sigset_t pointers if good.
9328 if (*argsz != sizeof(arg))
9330 if (copy_from_user(&arg, argp, sizeof(arg)))
9332 *sig = u64_to_user_ptr(arg.sigmask);
9333 *argsz = arg.sigmask_sz;
9334 *ts = u64_to_user_ptr(arg.ts);
9338 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9339 u32, min_complete, u32, flags, const void __user *, argp,
9342 struct io_ring_ctx *ctx;
9349 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9350 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9354 if (unlikely(!f.file))
9358 if (unlikely(f.file->f_op != &io_uring_fops))
9362 ctx = f.file->private_data;
9363 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9367 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9371 * For SQ polling, the thread will do all submissions and completions.
9372 * Just return the requested submit count, and wake the thread if
9376 if (ctx->flags & IORING_SETUP_SQPOLL) {
9377 io_cqring_overflow_flush(ctx);
9379 if (unlikely(ctx->sq_data->thread == NULL)) {
9383 if (flags & IORING_ENTER_SQ_WAKEUP)
9384 wake_up(&ctx->sq_data->wait);
9385 if (flags & IORING_ENTER_SQ_WAIT) {
9386 ret = io_sqpoll_wait_sq(ctx);
9390 submitted = to_submit;
9391 } else if (to_submit) {
9392 ret = io_uring_add_tctx_node(ctx);
9395 mutex_lock(&ctx->uring_lock);
9396 submitted = io_submit_sqes(ctx, to_submit);
9397 mutex_unlock(&ctx->uring_lock);
9399 if (submitted != to_submit)
9402 if (flags & IORING_ENTER_GETEVENTS) {
9403 const sigset_t __user *sig;
9404 struct __kernel_timespec __user *ts;
9406 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9410 min_complete = min(min_complete, ctx->cq_entries);
9413 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9414 * space applications don't need to do io completion events
9415 * polling again, they can rely on io_sq_thread to do polling
9416 * work, which can reduce cpu usage and uring_lock contention.
9418 if (ctx->flags & IORING_SETUP_IOPOLL &&
9419 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9420 ret = io_iopoll_check(ctx, min_complete);
9422 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9427 percpu_ref_put(&ctx->refs);
9430 return submitted ? submitted : ret;
9433 #ifdef CONFIG_PROC_FS
9434 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9435 const struct cred *cred)
9437 struct user_namespace *uns = seq_user_ns(m);
9438 struct group_info *gi;
9443 seq_printf(m, "%5d\n", id);
9444 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9445 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9446 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9447 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9448 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9449 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9450 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9451 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9452 seq_puts(m, "\n\tGroups:\t");
9453 gi = cred->group_info;
9454 for (g = 0; g < gi->ngroups; g++) {
9455 seq_put_decimal_ull(m, g ? " " : "",
9456 from_kgid_munged(uns, gi->gid[g]));
9458 seq_puts(m, "\n\tCapEff:\t");
9459 cap = cred->cap_effective;
9460 CAP_FOR_EACH_U32(__capi)
9461 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9466 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9468 struct io_sq_data *sq = NULL;
9473 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9474 * since fdinfo case grabs it in the opposite direction of normal use
9475 * cases. If we fail to get the lock, we just don't iterate any
9476 * structures that could be going away outside the io_uring mutex.
9478 has_lock = mutex_trylock(&ctx->uring_lock);
9480 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9486 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9487 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9488 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9489 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9490 struct file *f = io_file_from_index(ctx, i);
9493 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9495 seq_printf(m, "%5u: <none>\n", i);
9497 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9498 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9499 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9500 unsigned int len = buf->ubuf_end - buf->ubuf;
9502 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9504 if (has_lock && !xa_empty(&ctx->personalities)) {
9505 unsigned long index;
9506 const struct cred *cred;
9508 seq_printf(m, "Personalities:\n");
9509 xa_for_each(&ctx->personalities, index, cred)
9510 io_uring_show_cred(m, index, cred);
9512 seq_printf(m, "PollList:\n");
9513 spin_lock(&ctx->completion_lock);
9514 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9515 struct hlist_head *list = &ctx->cancel_hash[i];
9516 struct io_kiocb *req;
9518 hlist_for_each_entry(req, list, hash_node)
9519 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9520 req->task->task_works != NULL);
9522 spin_unlock(&ctx->completion_lock);
9524 mutex_unlock(&ctx->uring_lock);
9527 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9529 struct io_ring_ctx *ctx = f->private_data;
9531 if (percpu_ref_tryget(&ctx->refs)) {
9532 __io_uring_show_fdinfo(ctx, m);
9533 percpu_ref_put(&ctx->refs);
9538 static const struct file_operations io_uring_fops = {
9539 .release = io_uring_release,
9540 .mmap = io_uring_mmap,
9542 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9543 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9545 .poll = io_uring_poll,
9546 .fasync = io_uring_fasync,
9547 #ifdef CONFIG_PROC_FS
9548 .show_fdinfo = io_uring_show_fdinfo,
9552 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9553 struct io_uring_params *p)
9555 struct io_rings *rings;
9556 size_t size, sq_array_offset;
9558 /* make sure these are sane, as we already accounted them */
9559 ctx->sq_entries = p->sq_entries;
9560 ctx->cq_entries = p->cq_entries;
9562 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9563 if (size == SIZE_MAX)
9566 rings = io_mem_alloc(size);
9571 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9572 rings->sq_ring_mask = p->sq_entries - 1;
9573 rings->cq_ring_mask = p->cq_entries - 1;
9574 rings->sq_ring_entries = p->sq_entries;
9575 rings->cq_ring_entries = p->cq_entries;
9577 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9578 if (size == SIZE_MAX) {
9579 io_mem_free(ctx->rings);
9584 ctx->sq_sqes = io_mem_alloc(size);
9585 if (!ctx->sq_sqes) {
9586 io_mem_free(ctx->rings);
9594 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9598 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9602 ret = io_uring_add_tctx_node(ctx);
9607 fd_install(fd, file);
9612 * Allocate an anonymous fd, this is what constitutes the application
9613 * visible backing of an io_uring instance. The application mmaps this
9614 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9615 * we have to tie this fd to a socket for file garbage collection purposes.
9617 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9620 #if defined(CONFIG_UNIX)
9623 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9626 return ERR_PTR(ret);
9629 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9630 O_RDWR | O_CLOEXEC);
9631 #if defined(CONFIG_UNIX)
9633 sock_release(ctx->ring_sock);
9634 ctx->ring_sock = NULL;
9636 ctx->ring_sock->file = file;
9642 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9643 struct io_uring_params __user *params)
9645 struct io_ring_ctx *ctx;
9651 if (entries > IORING_MAX_ENTRIES) {
9652 if (!(p->flags & IORING_SETUP_CLAMP))
9654 entries = IORING_MAX_ENTRIES;
9658 * Use twice as many entries for the CQ ring. It's possible for the
9659 * application to drive a higher depth than the size of the SQ ring,
9660 * since the sqes are only used at submission time. This allows for
9661 * some flexibility in overcommitting a bit. If the application has
9662 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9663 * of CQ ring entries manually.
9665 p->sq_entries = roundup_pow_of_two(entries);
9666 if (p->flags & IORING_SETUP_CQSIZE) {
9668 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9669 * to a power-of-two, if it isn't already. We do NOT impose
9670 * any cq vs sq ring sizing.
9674 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9675 if (!(p->flags & IORING_SETUP_CLAMP))
9677 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9679 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9680 if (p->cq_entries < p->sq_entries)
9683 p->cq_entries = 2 * p->sq_entries;
9686 ctx = io_ring_ctx_alloc(p);
9689 ctx->compat = in_compat_syscall();
9690 if (!capable(CAP_IPC_LOCK))
9691 ctx->user = get_uid(current_user());
9694 * This is just grabbed for accounting purposes. When a process exits,
9695 * the mm is exited and dropped before the files, hence we need to hang
9696 * on to this mm purely for the purposes of being able to unaccount
9697 * memory (locked/pinned vm). It's not used for anything else.
9699 mmgrab(current->mm);
9700 ctx->mm_account = current->mm;
9702 ret = io_allocate_scq_urings(ctx, p);
9706 ret = io_sq_offload_create(ctx, p);
9709 /* always set a rsrc node */
9710 ret = io_rsrc_node_switch_start(ctx);
9713 io_rsrc_node_switch(ctx, NULL);
9715 memset(&p->sq_off, 0, sizeof(p->sq_off));
9716 p->sq_off.head = offsetof(struct io_rings, sq.head);
9717 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9718 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9719 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9720 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9721 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9722 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9724 memset(&p->cq_off, 0, sizeof(p->cq_off));
9725 p->cq_off.head = offsetof(struct io_rings, cq.head);
9726 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9727 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9728 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9729 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9730 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9731 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9733 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9734 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9735 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9736 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9737 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9738 IORING_FEAT_RSRC_TAGS;
9740 if (copy_to_user(params, p, sizeof(*p))) {
9745 file = io_uring_get_file(ctx);
9747 ret = PTR_ERR(file);
9752 * Install ring fd as the very last thing, so we don't risk someone
9753 * having closed it before we finish setup
9755 ret = io_uring_install_fd(ctx, file);
9757 /* fput will clean it up */
9762 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9765 io_ring_ctx_wait_and_kill(ctx);
9770 * Sets up an aio uring context, and returns the fd. Applications asks for a
9771 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9772 * params structure passed in.
9774 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9776 struct io_uring_params p;
9779 if (copy_from_user(&p, params, sizeof(p)))
9781 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9786 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9787 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9788 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9789 IORING_SETUP_R_DISABLED))
9792 return io_uring_create(entries, &p, params);
9795 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9796 struct io_uring_params __user *, params)
9798 return io_uring_setup(entries, params);
9801 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9803 struct io_uring_probe *p;
9807 size = struct_size(p, ops, nr_args);
9808 if (size == SIZE_MAX)
9810 p = kzalloc(size, GFP_KERNEL);
9815 if (copy_from_user(p, arg, size))
9818 if (memchr_inv(p, 0, size))
9821 p->last_op = IORING_OP_LAST - 1;
9822 if (nr_args > IORING_OP_LAST)
9823 nr_args = IORING_OP_LAST;
9825 for (i = 0; i < nr_args; i++) {
9827 if (!io_op_defs[i].not_supported)
9828 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9833 if (copy_to_user(arg, p, size))
9840 static int io_register_personality(struct io_ring_ctx *ctx)
9842 const struct cred *creds;
9846 creds = get_current_cred();
9848 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9849 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9857 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9858 unsigned int nr_args)
9860 struct io_uring_restriction *res;
9864 /* Restrictions allowed only if rings started disabled */
9865 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9868 /* We allow only a single restrictions registration */
9869 if (ctx->restrictions.registered)
9872 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9875 size = array_size(nr_args, sizeof(*res));
9876 if (size == SIZE_MAX)
9879 res = memdup_user(arg, size);
9881 return PTR_ERR(res);
9885 for (i = 0; i < nr_args; i++) {
9886 switch (res[i].opcode) {
9887 case IORING_RESTRICTION_REGISTER_OP:
9888 if (res[i].register_op >= IORING_REGISTER_LAST) {
9893 __set_bit(res[i].register_op,
9894 ctx->restrictions.register_op);
9896 case IORING_RESTRICTION_SQE_OP:
9897 if (res[i].sqe_op >= IORING_OP_LAST) {
9902 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9904 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9905 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9907 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9908 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9917 /* Reset all restrictions if an error happened */
9919 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9921 ctx->restrictions.registered = true;
9927 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9929 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9932 if (ctx->restrictions.registered)
9933 ctx->restricted = 1;
9935 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9936 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9937 wake_up(&ctx->sq_data->wait);
9941 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9942 struct io_uring_rsrc_update2 *up,
9950 if (check_add_overflow(up->offset, nr_args, &tmp))
9952 err = io_rsrc_node_switch_start(ctx);
9957 case IORING_RSRC_FILE:
9958 return __io_sqe_files_update(ctx, up, nr_args);
9959 case IORING_RSRC_BUFFER:
9960 return __io_sqe_buffers_update(ctx, up, nr_args);
9965 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9968 struct io_uring_rsrc_update2 up;
9972 memset(&up, 0, sizeof(up));
9973 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9975 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9978 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9979 unsigned size, unsigned type)
9981 struct io_uring_rsrc_update2 up;
9983 if (size != sizeof(up))
9985 if (copy_from_user(&up, arg, sizeof(up)))
9987 if (!up.nr || up.resv)
9989 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9992 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9993 unsigned int size, unsigned int type)
9995 struct io_uring_rsrc_register rr;
9997 /* keep it extendible */
9998 if (size != sizeof(rr))
10001 memset(&rr, 0, sizeof(rr));
10002 if (copy_from_user(&rr, arg, size))
10004 if (!rr.nr || rr.resv || rr.resv2)
10008 case IORING_RSRC_FILE:
10009 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10010 rr.nr, u64_to_user_ptr(rr.tags));
10011 case IORING_RSRC_BUFFER:
10012 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10013 rr.nr, u64_to_user_ptr(rr.tags));
10018 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10021 struct io_uring_task *tctx = current->io_uring;
10022 cpumask_var_t new_mask;
10025 if (!tctx || !tctx->io_wq)
10028 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10031 cpumask_clear(new_mask);
10032 if (len > cpumask_size())
10033 len = cpumask_size();
10035 if (copy_from_user(new_mask, arg, len)) {
10036 free_cpumask_var(new_mask);
10040 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10041 free_cpumask_var(new_mask);
10045 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10047 struct io_uring_task *tctx = current->io_uring;
10049 if (!tctx || !tctx->io_wq)
10052 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10055 static bool io_register_op_must_quiesce(int op)
10058 case IORING_REGISTER_BUFFERS:
10059 case IORING_UNREGISTER_BUFFERS:
10060 case IORING_REGISTER_FILES:
10061 case IORING_UNREGISTER_FILES:
10062 case IORING_REGISTER_FILES_UPDATE:
10063 case IORING_REGISTER_PROBE:
10064 case IORING_REGISTER_PERSONALITY:
10065 case IORING_UNREGISTER_PERSONALITY:
10066 case IORING_REGISTER_FILES2:
10067 case IORING_REGISTER_FILES_UPDATE2:
10068 case IORING_REGISTER_BUFFERS2:
10069 case IORING_REGISTER_BUFFERS_UPDATE:
10070 case IORING_REGISTER_IOWQ_AFF:
10071 case IORING_UNREGISTER_IOWQ_AFF:
10078 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10082 percpu_ref_kill(&ctx->refs);
10085 * Drop uring mutex before waiting for references to exit. If another
10086 * thread is currently inside io_uring_enter() it might need to grab the
10087 * uring_lock to make progress. If we hold it here across the drain
10088 * wait, then we can deadlock. It's safe to drop the mutex here, since
10089 * no new references will come in after we've killed the percpu ref.
10091 mutex_unlock(&ctx->uring_lock);
10093 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10096 ret = io_run_task_work_sig();
10097 } while (ret >= 0);
10098 mutex_lock(&ctx->uring_lock);
10101 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10105 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10106 void __user *arg, unsigned nr_args)
10107 __releases(ctx->uring_lock)
10108 __acquires(ctx->uring_lock)
10113 * We're inside the ring mutex, if the ref is already dying, then
10114 * someone else killed the ctx or is already going through
10115 * io_uring_register().
10117 if (percpu_ref_is_dying(&ctx->refs))
10120 if (ctx->restricted) {
10121 if (opcode >= IORING_REGISTER_LAST)
10123 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10124 if (!test_bit(opcode, ctx->restrictions.register_op))
10128 if (io_register_op_must_quiesce(opcode)) {
10129 ret = io_ctx_quiesce(ctx);
10135 case IORING_REGISTER_BUFFERS:
10136 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10138 case IORING_UNREGISTER_BUFFERS:
10140 if (arg || nr_args)
10142 ret = io_sqe_buffers_unregister(ctx);
10144 case IORING_REGISTER_FILES:
10145 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10147 case IORING_UNREGISTER_FILES:
10149 if (arg || nr_args)
10151 ret = io_sqe_files_unregister(ctx);
10153 case IORING_REGISTER_FILES_UPDATE:
10154 ret = io_register_files_update(ctx, arg, nr_args);
10156 case IORING_REGISTER_EVENTFD:
10157 case IORING_REGISTER_EVENTFD_ASYNC:
10161 ret = io_eventfd_register(ctx, arg);
10164 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10165 ctx->eventfd_async = 1;
10167 ctx->eventfd_async = 0;
10169 case IORING_UNREGISTER_EVENTFD:
10171 if (arg || nr_args)
10173 ret = io_eventfd_unregister(ctx);
10175 case IORING_REGISTER_PROBE:
10177 if (!arg || nr_args > 256)
10179 ret = io_probe(ctx, arg, nr_args);
10181 case IORING_REGISTER_PERSONALITY:
10183 if (arg || nr_args)
10185 ret = io_register_personality(ctx);
10187 case IORING_UNREGISTER_PERSONALITY:
10191 ret = io_unregister_personality(ctx, nr_args);
10193 case IORING_REGISTER_ENABLE_RINGS:
10195 if (arg || nr_args)
10197 ret = io_register_enable_rings(ctx);
10199 case IORING_REGISTER_RESTRICTIONS:
10200 ret = io_register_restrictions(ctx, arg, nr_args);
10202 case IORING_REGISTER_FILES2:
10203 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10205 case IORING_REGISTER_FILES_UPDATE2:
10206 ret = io_register_rsrc_update(ctx, arg, nr_args,
10209 case IORING_REGISTER_BUFFERS2:
10210 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10212 case IORING_REGISTER_BUFFERS_UPDATE:
10213 ret = io_register_rsrc_update(ctx, arg, nr_args,
10214 IORING_RSRC_BUFFER);
10216 case IORING_REGISTER_IOWQ_AFF:
10218 if (!arg || !nr_args)
10220 ret = io_register_iowq_aff(ctx, arg, nr_args);
10222 case IORING_UNREGISTER_IOWQ_AFF:
10224 if (arg || nr_args)
10226 ret = io_unregister_iowq_aff(ctx);
10233 if (io_register_op_must_quiesce(opcode)) {
10234 /* bring the ctx back to life */
10235 percpu_ref_reinit(&ctx->refs);
10236 reinit_completion(&ctx->ref_comp);
10241 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10242 void __user *, arg, unsigned int, nr_args)
10244 struct io_ring_ctx *ctx;
10253 if (f.file->f_op != &io_uring_fops)
10256 ctx = f.file->private_data;
10258 io_run_task_work();
10260 mutex_lock(&ctx->uring_lock);
10261 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10262 mutex_unlock(&ctx->uring_lock);
10263 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10264 ctx->cq_ev_fd != NULL, ret);
10270 static int __init io_uring_init(void)
10272 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10273 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10274 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10277 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10278 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10279 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10280 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10281 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10282 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10283 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10284 BUILD_BUG_SQE_ELEM(8, __u64, off);
10285 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10286 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10287 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10288 BUILD_BUG_SQE_ELEM(24, __u32, len);
10289 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10290 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10291 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10292 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10293 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10294 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10295 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10296 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10299 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10300 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10301 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10302 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10303 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10304 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10305 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10306 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10307 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10308 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10310 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10311 sizeof(struct io_uring_rsrc_update));
10312 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10313 sizeof(struct io_uring_rsrc_update2));
10314 /* should fit into one byte */
10315 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10317 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10318 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10320 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10324 __initcall(io_uring_init);