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 struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1311 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1312 req->flags |= REQ_F_LINK_TIMEOUT;
1314 /* linked timeouts should have two refs once prep'ed */
1315 io_req_set_refcount(req);
1316 __io_req_set_refcount(req->link, 2);
1320 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1322 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1324 return __io_prep_linked_timeout(req);
1327 static void io_prep_async_work(struct io_kiocb *req)
1329 const struct io_op_def *def = &io_op_defs[req->opcode];
1330 struct io_ring_ctx *ctx = req->ctx;
1332 if (!(req->flags & REQ_F_CREDS)) {
1333 req->flags |= REQ_F_CREDS;
1334 req->creds = get_current_cred();
1337 req->work.list.next = NULL;
1338 req->work.flags = 0;
1339 if (req->flags & REQ_F_FORCE_ASYNC)
1340 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1342 if (req->flags & REQ_F_ISREG) {
1343 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1344 io_wq_hash_work(&req->work, file_inode(req->file));
1345 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1346 if (def->unbound_nonreg_file)
1347 req->work.flags |= IO_WQ_WORK_UNBOUND;
1350 switch (req->opcode) {
1351 case IORING_OP_SPLICE:
1353 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1354 req->work.flags |= IO_WQ_WORK_UNBOUND;
1359 static void io_prep_async_link(struct io_kiocb *req)
1361 struct io_kiocb *cur;
1363 if (req->flags & REQ_F_LINK_TIMEOUT) {
1364 struct io_ring_ctx *ctx = req->ctx;
1366 spin_lock(&ctx->completion_lock);
1367 io_for_each_link(cur, req)
1368 io_prep_async_work(cur);
1369 spin_unlock(&ctx->completion_lock);
1371 io_for_each_link(cur, req)
1372 io_prep_async_work(cur);
1376 static void io_queue_async_work(struct io_kiocb *req)
1378 struct io_ring_ctx *ctx = req->ctx;
1379 struct io_kiocb *link = io_prep_linked_timeout(req);
1380 struct io_uring_task *tctx = req->task->io_uring;
1383 BUG_ON(!tctx->io_wq);
1385 /* init ->work of the whole link before punting */
1386 io_prep_async_link(req);
1389 * Not expected to happen, but if we do have a bug where this _can_
1390 * happen, catch it here and ensure the request is marked as
1391 * canceled. That will make io-wq go through the usual work cancel
1392 * procedure rather than attempt to run this request (or create a new
1395 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1396 req->work.flags |= IO_WQ_WORK_CANCEL;
1398 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1399 &req->work, req->flags);
1400 io_wq_enqueue(tctx->io_wq, &req->work);
1402 io_queue_linked_timeout(link);
1405 static void io_kill_timeout(struct io_kiocb *req, int status)
1406 __must_hold(&req->ctx->completion_lock)
1407 __must_hold(&req->ctx->timeout_lock)
1409 struct io_timeout_data *io = req->async_data;
1411 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1412 atomic_set(&req->ctx->cq_timeouts,
1413 atomic_read(&req->ctx->cq_timeouts) + 1);
1414 list_del_init(&req->timeout.list);
1415 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1416 io_put_req_deferred(req);
1420 static void io_queue_deferred(struct io_ring_ctx *ctx)
1422 while (!list_empty(&ctx->defer_list)) {
1423 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1424 struct io_defer_entry, list);
1426 if (req_need_defer(de->req, de->seq))
1428 list_del_init(&de->list);
1429 io_req_task_queue(de->req);
1434 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1435 __must_hold(&ctx->completion_lock)
1437 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1439 spin_lock_irq(&ctx->timeout_lock);
1440 while (!list_empty(&ctx->timeout_list)) {
1441 u32 events_needed, events_got;
1442 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1443 struct io_kiocb, timeout.list);
1445 if (io_is_timeout_noseq(req))
1449 * Since seq can easily wrap around over time, subtract
1450 * the last seq at which timeouts were flushed before comparing.
1451 * Assuming not more than 2^31-1 events have happened since,
1452 * these subtractions won't have wrapped, so we can check if
1453 * target is in [last_seq, current_seq] by comparing the two.
1455 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1456 events_got = seq - ctx->cq_last_tm_flush;
1457 if (events_got < events_needed)
1460 list_del_init(&req->timeout.list);
1461 io_kill_timeout(req, 0);
1463 ctx->cq_last_tm_flush = seq;
1464 spin_unlock_irq(&ctx->timeout_lock);
1467 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1469 if (ctx->off_timeout_used)
1470 io_flush_timeouts(ctx);
1471 if (ctx->drain_active)
1472 io_queue_deferred(ctx);
1475 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1477 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1478 __io_commit_cqring_flush(ctx);
1479 /* order cqe stores with ring update */
1480 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1483 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1485 struct io_rings *r = ctx->rings;
1487 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1490 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1492 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1495 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1497 struct io_rings *rings = ctx->rings;
1498 unsigned tail, mask = ctx->cq_entries - 1;
1501 * writes to the cq entry need to come after reading head; the
1502 * control dependency is enough as we're using WRITE_ONCE to
1505 if (__io_cqring_events(ctx) == ctx->cq_entries)
1508 tail = ctx->cached_cq_tail++;
1509 return &rings->cqes[tail & mask];
1512 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1514 if (likely(!ctx->cq_ev_fd))
1516 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1518 return !ctx->eventfd_async || io_wq_current_is_worker();
1521 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1524 * wake_up_all() may seem excessive, but io_wake_function() and
1525 * io_should_wake() handle the termination of the loop and only
1526 * wake as many waiters as we need to.
1528 if (wq_has_sleeper(&ctx->cq_wait))
1529 wake_up_all(&ctx->cq_wait);
1530 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1531 wake_up(&ctx->sq_data->wait);
1532 if (io_should_trigger_evfd(ctx))
1533 eventfd_signal(ctx->cq_ev_fd, 1);
1534 if (waitqueue_active(&ctx->poll_wait)) {
1535 wake_up_interruptible(&ctx->poll_wait);
1536 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1540 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1542 if (ctx->flags & IORING_SETUP_SQPOLL) {
1543 if (wq_has_sleeper(&ctx->cq_wait))
1544 wake_up_all(&ctx->cq_wait);
1546 if (io_should_trigger_evfd(ctx))
1547 eventfd_signal(ctx->cq_ev_fd, 1);
1548 if (waitqueue_active(&ctx->poll_wait)) {
1549 wake_up_interruptible(&ctx->poll_wait);
1550 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1554 /* Returns true if there are no backlogged entries after the flush */
1555 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1557 bool all_flushed, posted;
1559 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1563 spin_lock(&ctx->completion_lock);
1564 while (!list_empty(&ctx->cq_overflow_list)) {
1565 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1566 struct io_overflow_cqe *ocqe;
1570 ocqe = list_first_entry(&ctx->cq_overflow_list,
1571 struct io_overflow_cqe, list);
1573 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1575 io_account_cq_overflow(ctx);
1578 list_del(&ocqe->list);
1582 all_flushed = list_empty(&ctx->cq_overflow_list);
1584 clear_bit(0, &ctx->check_cq_overflow);
1585 WRITE_ONCE(ctx->rings->sq_flags,
1586 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1590 io_commit_cqring(ctx);
1591 spin_unlock(&ctx->completion_lock);
1593 io_cqring_ev_posted(ctx);
1597 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1601 if (test_bit(0, &ctx->check_cq_overflow)) {
1602 /* iopoll syncs against uring_lock, not completion_lock */
1603 if (ctx->flags & IORING_SETUP_IOPOLL)
1604 mutex_lock(&ctx->uring_lock);
1605 ret = __io_cqring_overflow_flush(ctx, false);
1606 if (ctx->flags & IORING_SETUP_IOPOLL)
1607 mutex_unlock(&ctx->uring_lock);
1613 /* must to be called somewhat shortly after putting a request */
1614 static inline void io_put_task(struct task_struct *task, int nr)
1616 struct io_uring_task *tctx = task->io_uring;
1618 percpu_counter_sub(&tctx->inflight, nr);
1619 if (unlikely(atomic_read(&tctx->in_idle)))
1620 wake_up(&tctx->wait);
1621 put_task_struct_many(task, nr);
1624 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1625 long res, unsigned int cflags)
1627 struct io_overflow_cqe *ocqe;
1629 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1632 * If we're in ring overflow flush mode, or in task cancel mode,
1633 * or cannot allocate an overflow entry, then we need to drop it
1636 io_account_cq_overflow(ctx);
1639 if (list_empty(&ctx->cq_overflow_list)) {
1640 set_bit(0, &ctx->check_cq_overflow);
1641 WRITE_ONCE(ctx->rings->sq_flags,
1642 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1645 ocqe->cqe.user_data = user_data;
1646 ocqe->cqe.res = res;
1647 ocqe->cqe.flags = cflags;
1648 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1652 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1653 long res, unsigned int cflags)
1655 struct io_uring_cqe *cqe;
1657 trace_io_uring_complete(ctx, user_data, res, cflags);
1660 * If we can't get a cq entry, userspace overflowed the
1661 * submission (by quite a lot). Increment the overflow count in
1664 cqe = io_get_cqe(ctx);
1666 WRITE_ONCE(cqe->user_data, user_data);
1667 WRITE_ONCE(cqe->res, res);
1668 WRITE_ONCE(cqe->flags, cflags);
1671 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1674 /* not as hot to bloat with inlining */
1675 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1676 long res, unsigned int cflags)
1678 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1681 static void io_req_complete_post(struct io_kiocb *req, long res,
1682 unsigned int cflags)
1684 struct io_ring_ctx *ctx = req->ctx;
1686 spin_lock(&ctx->completion_lock);
1687 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1689 * If we're the last reference to this request, add to our locked
1692 if (req_ref_put_and_test(req)) {
1693 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1694 if (req->flags & IO_DISARM_MASK)
1695 io_disarm_next(req);
1697 io_req_task_queue(req->link);
1701 io_dismantle_req(req);
1702 io_put_task(req->task, 1);
1703 list_add(&req->inflight_entry, &ctx->locked_free_list);
1704 ctx->locked_free_nr++;
1706 if (!percpu_ref_tryget(&ctx->refs))
1709 io_commit_cqring(ctx);
1710 spin_unlock(&ctx->completion_lock);
1713 io_cqring_ev_posted(ctx);
1714 percpu_ref_put(&ctx->refs);
1718 static inline bool io_req_needs_clean(struct io_kiocb *req)
1720 return req->flags & IO_REQ_CLEAN_FLAGS;
1723 static void io_req_complete_state(struct io_kiocb *req, long res,
1724 unsigned int cflags)
1726 if (io_req_needs_clean(req))
1729 req->compl.cflags = cflags;
1730 req->flags |= REQ_F_COMPLETE_INLINE;
1733 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1734 long res, unsigned cflags)
1736 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1737 io_req_complete_state(req, res, cflags);
1739 io_req_complete_post(req, res, cflags);
1742 static inline void io_req_complete(struct io_kiocb *req, long res)
1744 __io_req_complete(req, 0, res, 0);
1747 static void io_req_complete_failed(struct io_kiocb *req, long res)
1750 io_req_complete_post(req, res, 0);
1754 * Don't initialise the fields below on every allocation, but do that in
1755 * advance and keep them valid across allocations.
1757 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1761 req->async_data = NULL;
1762 /* not necessary, but safer to zero */
1766 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1767 struct io_submit_state *state)
1769 spin_lock(&ctx->completion_lock);
1770 list_splice_init(&ctx->locked_free_list, &state->free_list);
1771 ctx->locked_free_nr = 0;
1772 spin_unlock(&ctx->completion_lock);
1775 /* Returns true IFF there are requests in the cache */
1776 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1778 struct io_submit_state *state = &ctx->submit_state;
1782 * If we have more than a batch's worth of requests in our IRQ side
1783 * locked cache, grab the lock and move them over to our submission
1786 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1787 io_flush_cached_locked_reqs(ctx, state);
1789 nr = state->free_reqs;
1790 while (!list_empty(&state->free_list)) {
1791 struct io_kiocb *req = list_first_entry(&state->free_list,
1792 struct io_kiocb, inflight_entry);
1794 list_del(&req->inflight_entry);
1795 state->reqs[nr++] = req;
1796 if (nr == ARRAY_SIZE(state->reqs))
1800 state->free_reqs = nr;
1805 * A request might get retired back into the request caches even before opcode
1806 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1807 * Because of that, io_alloc_req() should be called only under ->uring_lock
1808 * and with extra caution to not get a request that is still worked on.
1810 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1811 __must_hold(&ctx->uring_lock)
1813 struct io_submit_state *state = &ctx->submit_state;
1814 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1817 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1819 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1822 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1826 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1827 * retry single alloc to be on the safe side.
1829 if (unlikely(ret <= 0)) {
1830 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1831 if (!state->reqs[0])
1836 for (i = 0; i < ret; i++)
1837 io_preinit_req(state->reqs[i], ctx);
1838 state->free_reqs = ret;
1841 return state->reqs[state->free_reqs];
1844 static inline void io_put_file(struct file *file)
1850 static void io_dismantle_req(struct io_kiocb *req)
1852 unsigned int flags = req->flags;
1854 if (io_req_needs_clean(req))
1856 if (!(flags & REQ_F_FIXED_FILE))
1857 io_put_file(req->file);
1858 if (req->fixed_rsrc_refs)
1859 percpu_ref_put(req->fixed_rsrc_refs);
1860 if (req->async_data) {
1861 kfree(req->async_data);
1862 req->async_data = NULL;
1866 static void __io_free_req(struct io_kiocb *req)
1868 struct io_ring_ctx *ctx = req->ctx;
1870 io_dismantle_req(req);
1871 io_put_task(req->task, 1);
1873 spin_lock(&ctx->completion_lock);
1874 list_add(&req->inflight_entry, &ctx->locked_free_list);
1875 ctx->locked_free_nr++;
1876 spin_unlock(&ctx->completion_lock);
1878 percpu_ref_put(&ctx->refs);
1881 static inline void io_remove_next_linked(struct io_kiocb *req)
1883 struct io_kiocb *nxt = req->link;
1885 req->link = nxt->link;
1889 static bool io_kill_linked_timeout(struct io_kiocb *req)
1890 __must_hold(&req->ctx->completion_lock)
1891 __must_hold(&req->ctx->timeout_lock)
1893 struct io_kiocb *link = req->link;
1895 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1896 struct io_timeout_data *io = link->async_data;
1898 io_remove_next_linked(req);
1899 link->timeout.head = NULL;
1900 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1901 io_cqring_fill_event(link->ctx, link->user_data,
1903 io_put_req_deferred(link);
1910 static void io_fail_links(struct io_kiocb *req)
1911 __must_hold(&req->ctx->completion_lock)
1913 struct io_kiocb *nxt, *link = req->link;
1920 trace_io_uring_fail_link(req, link);
1921 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1922 io_put_req_deferred(link);
1927 static bool io_disarm_next(struct io_kiocb *req)
1928 __must_hold(&req->ctx->completion_lock)
1930 bool posted = false;
1932 if (req->flags & REQ_F_ARM_LTIMEOUT) {
1933 struct io_kiocb *link = req->link;
1935 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1936 io_remove_next_linked(req);
1937 io_cqring_fill_event(link->ctx, link->user_data,
1939 io_put_req_deferred(link);
1942 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
1943 struct io_ring_ctx *ctx = req->ctx;
1945 spin_lock_irq(&ctx->timeout_lock);
1946 posted = io_kill_linked_timeout(req);
1947 spin_unlock_irq(&ctx->timeout_lock);
1949 if (unlikely((req->flags & REQ_F_FAIL) &&
1950 !(req->flags & REQ_F_HARDLINK))) {
1951 posted |= (req->link != NULL);
1957 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1959 struct io_kiocb *nxt;
1962 * If LINK is set, we have dependent requests in this chain. If we
1963 * didn't fail this request, queue the first one up, moving any other
1964 * dependencies to the next request. In case of failure, fail the rest
1967 if (req->flags & IO_DISARM_MASK) {
1968 struct io_ring_ctx *ctx = req->ctx;
1971 spin_lock(&ctx->completion_lock);
1972 posted = io_disarm_next(req);
1974 io_commit_cqring(req->ctx);
1975 spin_unlock(&ctx->completion_lock);
1977 io_cqring_ev_posted(ctx);
1984 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1986 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1988 return __io_req_find_next(req);
1991 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1995 if (ctx->submit_state.compl_nr) {
1996 mutex_lock(&ctx->uring_lock);
1997 io_submit_flush_completions(ctx);
1998 mutex_unlock(&ctx->uring_lock);
2000 percpu_ref_put(&ctx->refs);
2003 static void tctx_task_work(struct callback_head *cb)
2005 struct io_ring_ctx *ctx = NULL;
2006 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2010 struct io_wq_work_node *node;
2012 spin_lock_irq(&tctx->task_lock);
2013 node = tctx->task_list.first;
2014 INIT_WQ_LIST(&tctx->task_list);
2016 tctx->task_running = false;
2017 spin_unlock_irq(&tctx->task_lock);
2022 struct io_wq_work_node *next = node->next;
2023 struct io_kiocb *req = container_of(node, struct io_kiocb,
2026 if (req->ctx != ctx) {
2027 ctx_flush_and_put(ctx);
2029 percpu_ref_get(&ctx->refs);
2031 req->io_task_work.func(req);
2038 ctx_flush_and_put(ctx);
2041 static void io_req_task_work_add(struct io_kiocb *req)
2043 struct task_struct *tsk = req->task;
2044 struct io_uring_task *tctx = tsk->io_uring;
2045 enum task_work_notify_mode notify;
2046 struct io_wq_work_node *node;
2047 unsigned long flags;
2050 WARN_ON_ONCE(!tctx);
2052 spin_lock_irqsave(&tctx->task_lock, flags);
2053 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2054 running = tctx->task_running;
2056 tctx->task_running = true;
2057 spin_unlock_irqrestore(&tctx->task_lock, flags);
2059 /* task_work already pending, we're done */
2064 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2065 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2066 * processing task_work. There's no reliable way to tell if TWA_RESUME
2069 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2070 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2071 wake_up_process(tsk);
2075 spin_lock_irqsave(&tctx->task_lock, flags);
2076 tctx->task_running = false;
2077 node = tctx->task_list.first;
2078 INIT_WQ_LIST(&tctx->task_list);
2079 spin_unlock_irqrestore(&tctx->task_lock, flags);
2082 req = container_of(node, struct io_kiocb, io_task_work.node);
2084 if (llist_add(&req->io_task_work.fallback_node,
2085 &req->ctx->fallback_llist))
2086 schedule_delayed_work(&req->ctx->fallback_work, 1);
2090 static void io_req_task_cancel(struct io_kiocb *req)
2092 struct io_ring_ctx *ctx = req->ctx;
2094 /* ctx is guaranteed to stay alive while we hold uring_lock */
2095 mutex_lock(&ctx->uring_lock);
2096 io_req_complete_failed(req, req->result);
2097 mutex_unlock(&ctx->uring_lock);
2100 static void io_req_task_submit(struct io_kiocb *req)
2102 struct io_ring_ctx *ctx = req->ctx;
2104 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2105 mutex_lock(&ctx->uring_lock);
2106 if (likely(!(req->task->flags & PF_EXITING)))
2107 __io_queue_sqe(req);
2109 io_req_complete_failed(req, -EFAULT);
2110 mutex_unlock(&ctx->uring_lock);
2113 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2116 req->io_task_work.func = io_req_task_cancel;
2117 io_req_task_work_add(req);
2120 static void io_req_task_queue(struct io_kiocb *req)
2122 req->io_task_work.func = io_req_task_submit;
2123 io_req_task_work_add(req);
2126 static void io_req_task_queue_reissue(struct io_kiocb *req)
2128 req->io_task_work.func = io_queue_async_work;
2129 io_req_task_work_add(req);
2132 static inline void io_queue_next(struct io_kiocb *req)
2134 struct io_kiocb *nxt = io_req_find_next(req);
2137 io_req_task_queue(nxt);
2140 static void io_free_req(struct io_kiocb *req)
2147 struct task_struct *task;
2152 static inline void io_init_req_batch(struct req_batch *rb)
2159 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2160 struct req_batch *rb)
2163 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2164 if (rb->task == current)
2165 current->io_uring->cached_refs += rb->task_refs;
2167 io_put_task(rb->task, rb->task_refs);
2170 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2171 struct io_submit_state *state)
2174 io_dismantle_req(req);
2176 if (req->task != rb->task) {
2178 io_put_task(rb->task, rb->task_refs);
2179 rb->task = req->task;
2185 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2186 state->reqs[state->free_reqs++] = req;
2188 list_add(&req->inflight_entry, &state->free_list);
2191 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2192 __must_hold(&ctx->uring_lock)
2194 struct io_submit_state *state = &ctx->submit_state;
2195 int i, nr = state->compl_nr;
2196 struct req_batch rb;
2198 spin_lock(&ctx->completion_lock);
2199 for (i = 0; i < nr; i++) {
2200 struct io_kiocb *req = state->compl_reqs[i];
2202 __io_cqring_fill_event(ctx, req->user_data, req->result,
2205 io_commit_cqring(ctx);
2206 spin_unlock(&ctx->completion_lock);
2207 io_cqring_ev_posted(ctx);
2209 io_init_req_batch(&rb);
2210 for (i = 0; i < nr; i++) {
2211 struct io_kiocb *req = state->compl_reqs[i];
2213 if (req_ref_put_and_test(req))
2214 io_req_free_batch(&rb, req, &ctx->submit_state);
2217 io_req_free_batch_finish(ctx, &rb);
2218 state->compl_nr = 0;
2222 * Drop reference to request, return next in chain (if there is one) if this
2223 * was the last reference to this request.
2225 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2227 struct io_kiocb *nxt = NULL;
2229 if (req_ref_put_and_test(req)) {
2230 nxt = io_req_find_next(req);
2236 static inline void io_put_req(struct io_kiocb *req)
2238 if (req_ref_put_and_test(req))
2242 static inline void io_put_req_deferred(struct io_kiocb *req)
2244 if (req_ref_put_and_test(req)) {
2245 req->io_task_work.func = io_free_req;
2246 io_req_task_work_add(req);
2250 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2252 /* See comment at the top of this file */
2254 return __io_cqring_events(ctx);
2257 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2259 struct io_rings *rings = ctx->rings;
2261 /* make sure SQ entry isn't read before tail */
2262 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2265 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2267 unsigned int cflags;
2269 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2270 cflags |= IORING_CQE_F_BUFFER;
2271 req->flags &= ~REQ_F_BUFFER_SELECTED;
2276 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2278 struct io_buffer *kbuf;
2280 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2281 return io_put_kbuf(req, kbuf);
2284 static inline bool io_run_task_work(void)
2286 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2287 __set_current_state(TASK_RUNNING);
2288 tracehook_notify_signal();
2296 * Find and free completed poll iocbs
2298 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2299 struct list_head *done)
2301 struct req_batch rb;
2302 struct io_kiocb *req;
2304 /* order with ->result store in io_complete_rw_iopoll() */
2307 io_init_req_batch(&rb);
2308 while (!list_empty(done)) {
2311 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2312 list_del(&req->inflight_entry);
2314 if (READ_ONCE(req->result) == -EAGAIN &&
2315 !(req->flags & REQ_F_DONT_REISSUE)) {
2316 req->iopoll_completed = 0;
2317 io_req_task_queue_reissue(req);
2321 if (req->flags & REQ_F_BUFFER_SELECTED)
2322 cflags = io_put_rw_kbuf(req);
2324 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2327 if (req_ref_put_and_test(req))
2328 io_req_free_batch(&rb, req, &ctx->submit_state);
2331 io_commit_cqring(ctx);
2332 io_cqring_ev_posted_iopoll(ctx);
2333 io_req_free_batch_finish(ctx, &rb);
2336 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2339 struct io_kiocb *req, *tmp;
2344 * Only spin for completions if we don't have multiple devices hanging
2345 * off our complete list, and we're under the requested amount.
2347 spin = !ctx->poll_multi_queue && *nr_events < min;
2349 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2350 struct kiocb *kiocb = &req->rw.kiocb;
2354 * Move completed and retryable entries to our local lists.
2355 * If we find a request that requires polling, break out
2356 * and complete those lists first, if we have entries there.
2358 if (READ_ONCE(req->iopoll_completed)) {
2359 list_move_tail(&req->inflight_entry, &done);
2362 if (!list_empty(&done))
2365 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2366 if (unlikely(ret < 0))
2371 /* iopoll may have completed current req */
2372 if (READ_ONCE(req->iopoll_completed))
2373 list_move_tail(&req->inflight_entry, &done);
2376 if (!list_empty(&done))
2377 io_iopoll_complete(ctx, nr_events, &done);
2383 * We can't just wait for polled events to come to us, we have to actively
2384 * find and complete them.
2386 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2388 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2391 mutex_lock(&ctx->uring_lock);
2392 while (!list_empty(&ctx->iopoll_list)) {
2393 unsigned int nr_events = 0;
2395 io_do_iopoll(ctx, &nr_events, 0);
2397 /* let it sleep and repeat later if can't complete a request */
2401 * Ensure we allow local-to-the-cpu processing to take place,
2402 * in this case we need to ensure that we reap all events.
2403 * Also let task_work, etc. to progress by releasing the mutex
2405 if (need_resched()) {
2406 mutex_unlock(&ctx->uring_lock);
2408 mutex_lock(&ctx->uring_lock);
2411 mutex_unlock(&ctx->uring_lock);
2414 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2416 unsigned int nr_events = 0;
2420 * We disallow the app entering submit/complete with polling, but we
2421 * still need to lock the ring to prevent racing with polled issue
2422 * that got punted to a workqueue.
2424 mutex_lock(&ctx->uring_lock);
2426 * Don't enter poll loop if we already have events pending.
2427 * If we do, we can potentially be spinning for commands that
2428 * already triggered a CQE (eg in error).
2430 if (test_bit(0, &ctx->check_cq_overflow))
2431 __io_cqring_overflow_flush(ctx, false);
2432 if (io_cqring_events(ctx))
2436 * If a submit got punted to a workqueue, we can have the
2437 * application entering polling for a command before it gets
2438 * issued. That app will hold the uring_lock for the duration
2439 * of the poll right here, so we need to take a breather every
2440 * now and then to ensure that the issue has a chance to add
2441 * the poll to the issued list. Otherwise we can spin here
2442 * forever, while the workqueue is stuck trying to acquire the
2445 if (list_empty(&ctx->iopoll_list)) {
2446 u32 tail = ctx->cached_cq_tail;
2448 mutex_unlock(&ctx->uring_lock);
2450 mutex_lock(&ctx->uring_lock);
2452 /* some requests don't go through iopoll_list */
2453 if (tail != ctx->cached_cq_tail ||
2454 list_empty(&ctx->iopoll_list))
2457 ret = io_do_iopoll(ctx, &nr_events, min);
2458 } while (!ret && nr_events < min && !need_resched());
2460 mutex_unlock(&ctx->uring_lock);
2464 static void kiocb_end_write(struct io_kiocb *req)
2467 * Tell lockdep we inherited freeze protection from submission
2470 if (req->flags & REQ_F_ISREG) {
2471 struct super_block *sb = file_inode(req->file)->i_sb;
2473 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2479 static bool io_resubmit_prep(struct io_kiocb *req)
2481 struct io_async_rw *rw = req->async_data;
2484 return !io_req_prep_async(req);
2485 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2486 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2490 static bool io_rw_should_reissue(struct io_kiocb *req)
2492 umode_t mode = file_inode(req->file)->i_mode;
2493 struct io_ring_ctx *ctx = req->ctx;
2495 if (!S_ISBLK(mode) && !S_ISREG(mode))
2497 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2498 !(ctx->flags & IORING_SETUP_IOPOLL)))
2501 * If ref is dying, we might be running poll reap from the exit work.
2502 * Don't attempt to reissue from that path, just let it fail with
2505 if (percpu_ref_is_dying(&ctx->refs))
2508 * Play it safe and assume not safe to re-import and reissue if we're
2509 * not in the original thread group (or in task context).
2511 if (!same_thread_group(req->task, current) || !in_task())
2516 static bool io_resubmit_prep(struct io_kiocb *req)
2520 static bool io_rw_should_reissue(struct io_kiocb *req)
2526 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2528 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2529 kiocb_end_write(req);
2530 if (res != req->result) {
2531 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2532 io_rw_should_reissue(req)) {
2533 req->flags |= REQ_F_REISSUE;
2542 static void io_req_task_complete(struct io_kiocb *req)
2546 if (req->flags & REQ_F_BUFFER_SELECTED)
2547 cflags = io_put_rw_kbuf(req);
2548 __io_req_complete(req, 0, req->result, cflags);
2551 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2552 unsigned int issue_flags)
2554 if (__io_complete_rw_common(req, res))
2556 io_req_task_complete(req);
2559 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2561 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2563 if (__io_complete_rw_common(req, res))
2566 req->io_task_work.func = io_req_task_complete;
2567 io_req_task_work_add(req);
2570 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2572 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2574 if (kiocb->ki_flags & IOCB_WRITE)
2575 kiocb_end_write(req);
2576 if (unlikely(res != req->result)) {
2577 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2578 io_resubmit_prep(req))) {
2580 req->flags |= REQ_F_DONT_REISSUE;
2584 WRITE_ONCE(req->result, res);
2585 /* order with io_iopoll_complete() checking ->result */
2587 WRITE_ONCE(req->iopoll_completed, 1);
2591 * After the iocb has been issued, it's safe to be found on the poll list.
2592 * Adding the kiocb to the list AFTER submission ensures that we don't
2593 * find it from a io_do_iopoll() thread before the issuer is done
2594 * accessing the kiocb cookie.
2596 static void io_iopoll_req_issued(struct io_kiocb *req)
2598 struct io_ring_ctx *ctx = req->ctx;
2599 const bool in_async = io_wq_current_is_worker();
2601 /* workqueue context doesn't hold uring_lock, grab it now */
2602 if (unlikely(in_async))
2603 mutex_lock(&ctx->uring_lock);
2606 * Track whether we have multiple files in our lists. This will impact
2607 * how we do polling eventually, not spinning if we're on potentially
2608 * different devices.
2610 if (list_empty(&ctx->iopoll_list)) {
2611 ctx->poll_multi_queue = false;
2612 } else if (!ctx->poll_multi_queue) {
2613 struct io_kiocb *list_req;
2614 unsigned int queue_num0, queue_num1;
2616 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2619 if (list_req->file != req->file) {
2620 ctx->poll_multi_queue = true;
2622 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2623 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2624 if (queue_num0 != queue_num1)
2625 ctx->poll_multi_queue = true;
2630 * For fast devices, IO may have already completed. If it has, add
2631 * it to the front so we find it first.
2633 if (READ_ONCE(req->iopoll_completed))
2634 list_add(&req->inflight_entry, &ctx->iopoll_list);
2636 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2638 if (unlikely(in_async)) {
2640 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2641 * in sq thread task context or in io worker task context. If
2642 * current task context is sq thread, we don't need to check
2643 * whether should wake up sq thread.
2645 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2646 wq_has_sleeper(&ctx->sq_data->wait))
2647 wake_up(&ctx->sq_data->wait);
2649 mutex_unlock(&ctx->uring_lock);
2653 static bool io_bdev_nowait(struct block_device *bdev)
2655 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2659 * If we tracked the file through the SCM inflight mechanism, we could support
2660 * any file. For now, just ensure that anything potentially problematic is done
2663 static bool __io_file_supports_nowait(struct file *file, int rw)
2665 umode_t mode = file_inode(file)->i_mode;
2667 if (S_ISBLK(mode)) {
2668 if (IS_ENABLED(CONFIG_BLOCK) &&
2669 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2675 if (S_ISREG(mode)) {
2676 if (IS_ENABLED(CONFIG_BLOCK) &&
2677 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2678 file->f_op != &io_uring_fops)
2683 /* any ->read/write should understand O_NONBLOCK */
2684 if (file->f_flags & O_NONBLOCK)
2687 if (!(file->f_mode & FMODE_NOWAIT))
2691 return file->f_op->read_iter != NULL;
2693 return file->f_op->write_iter != NULL;
2696 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2698 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2700 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2703 return __io_file_supports_nowait(req->file, rw);
2706 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2708 struct io_ring_ctx *ctx = req->ctx;
2709 struct kiocb *kiocb = &req->rw.kiocb;
2710 struct file *file = req->file;
2714 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2715 req->flags |= REQ_F_ISREG;
2717 kiocb->ki_pos = READ_ONCE(sqe->off);
2718 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2719 req->flags |= REQ_F_CUR_POS;
2720 kiocb->ki_pos = file->f_pos;
2722 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2723 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2724 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2728 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2729 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2730 req->flags |= REQ_F_NOWAIT;
2732 ioprio = READ_ONCE(sqe->ioprio);
2734 ret = ioprio_check_cap(ioprio);
2738 kiocb->ki_ioprio = ioprio;
2740 kiocb->ki_ioprio = get_current_ioprio();
2742 if (ctx->flags & IORING_SETUP_IOPOLL) {
2743 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2744 !kiocb->ki_filp->f_op->iopoll)
2747 kiocb->ki_flags |= IOCB_HIPRI;
2748 kiocb->ki_complete = io_complete_rw_iopoll;
2749 req->iopoll_completed = 0;
2751 if (kiocb->ki_flags & IOCB_HIPRI)
2753 kiocb->ki_complete = io_complete_rw;
2756 if (req->opcode == IORING_OP_READ_FIXED ||
2757 req->opcode == IORING_OP_WRITE_FIXED) {
2759 io_req_set_rsrc_node(req);
2762 req->rw.addr = READ_ONCE(sqe->addr);
2763 req->rw.len = READ_ONCE(sqe->len);
2764 req->buf_index = READ_ONCE(sqe->buf_index);
2768 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2774 case -ERESTARTNOINTR:
2775 case -ERESTARTNOHAND:
2776 case -ERESTART_RESTARTBLOCK:
2778 * We can't just restart the syscall, since previously
2779 * submitted sqes may already be in progress. Just fail this
2785 kiocb->ki_complete(kiocb, ret, 0);
2789 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2790 unsigned int issue_flags)
2792 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2793 struct io_async_rw *io = req->async_data;
2794 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2796 /* add previously done IO, if any */
2797 if (io && io->bytes_done > 0) {
2799 ret = io->bytes_done;
2801 ret += io->bytes_done;
2804 if (req->flags & REQ_F_CUR_POS)
2805 req->file->f_pos = kiocb->ki_pos;
2806 if (ret >= 0 && check_reissue)
2807 __io_complete_rw(req, ret, 0, issue_flags);
2809 io_rw_done(kiocb, ret);
2811 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2812 req->flags &= ~REQ_F_REISSUE;
2813 if (io_resubmit_prep(req)) {
2814 io_req_task_queue_reissue(req);
2819 if (req->flags & REQ_F_BUFFER_SELECTED)
2820 cflags = io_put_rw_kbuf(req);
2821 __io_req_complete(req, issue_flags, ret, cflags);
2826 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2827 struct io_mapped_ubuf *imu)
2829 size_t len = req->rw.len;
2830 u64 buf_end, buf_addr = req->rw.addr;
2833 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2835 /* not inside the mapped region */
2836 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2840 * May not be a start of buffer, set size appropriately
2841 * and advance us to the beginning.
2843 offset = buf_addr - imu->ubuf;
2844 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2848 * Don't use iov_iter_advance() here, as it's really slow for
2849 * using the latter parts of a big fixed buffer - it iterates
2850 * over each segment manually. We can cheat a bit here, because
2853 * 1) it's a BVEC iter, we set it up
2854 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2855 * first and last bvec
2857 * So just find our index, and adjust the iterator afterwards.
2858 * If the offset is within the first bvec (or the whole first
2859 * bvec, just use iov_iter_advance(). This makes it easier
2860 * since we can just skip the first segment, which may not
2861 * be PAGE_SIZE aligned.
2863 const struct bio_vec *bvec = imu->bvec;
2865 if (offset <= bvec->bv_len) {
2866 iov_iter_advance(iter, offset);
2868 unsigned long seg_skip;
2870 /* skip first vec */
2871 offset -= bvec->bv_len;
2872 seg_skip = 1 + (offset >> PAGE_SHIFT);
2874 iter->bvec = bvec + seg_skip;
2875 iter->nr_segs -= seg_skip;
2876 iter->count -= bvec->bv_len + offset;
2877 iter->iov_offset = offset & ~PAGE_MASK;
2884 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2886 struct io_ring_ctx *ctx = req->ctx;
2887 struct io_mapped_ubuf *imu = req->imu;
2888 u16 index, buf_index = req->buf_index;
2891 if (unlikely(buf_index >= ctx->nr_user_bufs))
2893 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2894 imu = READ_ONCE(ctx->user_bufs[index]);
2897 return __io_import_fixed(req, rw, iter, imu);
2900 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2903 mutex_unlock(&ctx->uring_lock);
2906 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2909 * "Normal" inline submissions always hold the uring_lock, since we
2910 * grab it from the system call. Same is true for the SQPOLL offload.
2911 * The only exception is when we've detached the request and issue it
2912 * from an async worker thread, grab the lock for that case.
2915 mutex_lock(&ctx->uring_lock);
2918 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2919 int bgid, struct io_buffer *kbuf,
2922 struct io_buffer *head;
2924 if (req->flags & REQ_F_BUFFER_SELECTED)
2927 io_ring_submit_lock(req->ctx, needs_lock);
2929 lockdep_assert_held(&req->ctx->uring_lock);
2931 head = xa_load(&req->ctx->io_buffers, bgid);
2933 if (!list_empty(&head->list)) {
2934 kbuf = list_last_entry(&head->list, struct io_buffer,
2936 list_del(&kbuf->list);
2939 xa_erase(&req->ctx->io_buffers, bgid);
2941 if (*len > kbuf->len)
2944 kbuf = ERR_PTR(-ENOBUFS);
2947 io_ring_submit_unlock(req->ctx, needs_lock);
2952 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2955 struct io_buffer *kbuf;
2958 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2959 bgid = req->buf_index;
2960 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2963 req->rw.addr = (u64) (unsigned long) kbuf;
2964 req->flags |= REQ_F_BUFFER_SELECTED;
2965 return u64_to_user_ptr(kbuf->addr);
2968 #ifdef CONFIG_COMPAT
2969 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2972 struct compat_iovec __user *uiov;
2973 compat_ssize_t clen;
2977 uiov = u64_to_user_ptr(req->rw.addr);
2978 if (!access_ok(uiov, sizeof(*uiov)))
2980 if (__get_user(clen, &uiov->iov_len))
2986 buf = io_rw_buffer_select(req, &len, needs_lock);
2988 return PTR_ERR(buf);
2989 iov[0].iov_base = buf;
2990 iov[0].iov_len = (compat_size_t) len;
2995 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2998 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3002 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3005 len = iov[0].iov_len;
3008 buf = io_rw_buffer_select(req, &len, needs_lock);
3010 return PTR_ERR(buf);
3011 iov[0].iov_base = buf;
3012 iov[0].iov_len = len;
3016 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3019 if (req->flags & REQ_F_BUFFER_SELECTED) {
3020 struct io_buffer *kbuf;
3022 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3023 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3024 iov[0].iov_len = kbuf->len;
3027 if (req->rw.len != 1)
3030 #ifdef CONFIG_COMPAT
3031 if (req->ctx->compat)
3032 return io_compat_import(req, iov, needs_lock);
3035 return __io_iov_buffer_select(req, iov, needs_lock);
3038 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3039 struct iov_iter *iter, bool needs_lock)
3041 void __user *buf = u64_to_user_ptr(req->rw.addr);
3042 size_t sqe_len = req->rw.len;
3043 u8 opcode = req->opcode;
3046 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3048 return io_import_fixed(req, rw, iter);
3051 /* buffer index only valid with fixed read/write, or buffer select */
3052 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3055 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3056 if (req->flags & REQ_F_BUFFER_SELECT) {
3057 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3059 return PTR_ERR(buf);
3060 req->rw.len = sqe_len;
3063 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3068 if (req->flags & REQ_F_BUFFER_SELECT) {
3069 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3071 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3076 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3080 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3082 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3086 * For files that don't have ->read_iter() and ->write_iter(), handle them
3087 * by looping over ->read() or ->write() manually.
3089 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3091 struct kiocb *kiocb = &req->rw.kiocb;
3092 struct file *file = req->file;
3096 * Don't support polled IO through this interface, and we can't
3097 * support non-blocking either. For the latter, this just causes
3098 * the kiocb to be handled from an async context.
3100 if (kiocb->ki_flags & IOCB_HIPRI)
3102 if (kiocb->ki_flags & IOCB_NOWAIT)
3105 while (iov_iter_count(iter)) {
3109 if (!iov_iter_is_bvec(iter)) {
3110 iovec = iov_iter_iovec(iter);
3112 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3113 iovec.iov_len = req->rw.len;
3117 nr = file->f_op->read(file, iovec.iov_base,
3118 iovec.iov_len, io_kiocb_ppos(kiocb));
3120 nr = file->f_op->write(file, iovec.iov_base,
3121 iovec.iov_len, io_kiocb_ppos(kiocb));
3130 if (nr != iovec.iov_len)
3134 iov_iter_advance(iter, nr);
3140 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3141 const struct iovec *fast_iov, struct iov_iter *iter)
3143 struct io_async_rw *rw = req->async_data;
3145 memcpy(&rw->iter, iter, sizeof(*iter));
3146 rw->free_iovec = iovec;
3148 /* can only be fixed buffers, no need to do anything */
3149 if (iov_iter_is_bvec(iter))
3152 unsigned iov_off = 0;
3154 rw->iter.iov = rw->fast_iov;
3155 if (iter->iov != fast_iov) {
3156 iov_off = iter->iov - fast_iov;
3157 rw->iter.iov += iov_off;
3159 if (rw->fast_iov != fast_iov)
3160 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3161 sizeof(struct iovec) * iter->nr_segs);
3163 req->flags |= REQ_F_NEED_CLEANUP;
3167 static inline int io_alloc_async_data(struct io_kiocb *req)
3169 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3170 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3171 return req->async_data == NULL;
3174 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3175 const struct iovec *fast_iov,
3176 struct iov_iter *iter, bool force)
3178 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3180 if (!req->async_data) {
3181 if (io_alloc_async_data(req)) {
3186 io_req_map_rw(req, iovec, fast_iov, iter);
3191 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3193 struct io_async_rw *iorw = req->async_data;
3194 struct iovec *iov = iorw->fast_iov;
3197 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3198 if (unlikely(ret < 0))
3201 iorw->bytes_done = 0;
3202 iorw->free_iovec = iov;
3204 req->flags |= REQ_F_NEED_CLEANUP;
3208 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3210 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3212 return io_prep_rw(req, sqe);
3216 * This is our waitqueue callback handler, registered through lock_page_async()
3217 * when we initially tried to do the IO with the iocb armed our waitqueue.
3218 * This gets called when the page is unlocked, and we generally expect that to
3219 * happen when the page IO is completed and the page is now uptodate. This will
3220 * queue a task_work based retry of the operation, attempting to copy the data
3221 * again. If the latter fails because the page was NOT uptodate, then we will
3222 * do a thread based blocking retry of the operation. That's the unexpected
3225 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3226 int sync, void *arg)
3228 struct wait_page_queue *wpq;
3229 struct io_kiocb *req = wait->private;
3230 struct wait_page_key *key = arg;
3232 wpq = container_of(wait, struct wait_page_queue, wait);
3234 if (!wake_page_match(wpq, key))
3237 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3238 list_del_init(&wait->entry);
3239 io_req_task_queue(req);
3244 * This controls whether a given IO request should be armed for async page
3245 * based retry. If we return false here, the request is handed to the async
3246 * worker threads for retry. If we're doing buffered reads on a regular file,
3247 * we prepare a private wait_page_queue entry and retry the operation. This
3248 * will either succeed because the page is now uptodate and unlocked, or it
3249 * will register a callback when the page is unlocked at IO completion. Through
3250 * that callback, io_uring uses task_work to setup a retry of the operation.
3251 * That retry will attempt the buffered read again. The retry will generally
3252 * succeed, or in rare cases where it fails, we then fall back to using the
3253 * async worker threads for a blocking retry.
3255 static bool io_rw_should_retry(struct io_kiocb *req)
3257 struct io_async_rw *rw = req->async_data;
3258 struct wait_page_queue *wait = &rw->wpq;
3259 struct kiocb *kiocb = &req->rw.kiocb;
3261 /* never retry for NOWAIT, we just complete with -EAGAIN */
3262 if (req->flags & REQ_F_NOWAIT)
3265 /* Only for buffered IO */
3266 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3270 * just use poll if we can, and don't attempt if the fs doesn't
3271 * support callback based unlocks
3273 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3276 wait->wait.func = io_async_buf_func;
3277 wait->wait.private = req;
3278 wait->wait.flags = 0;
3279 INIT_LIST_HEAD(&wait->wait.entry);
3280 kiocb->ki_flags |= IOCB_WAITQ;
3281 kiocb->ki_flags &= ~IOCB_NOWAIT;
3282 kiocb->ki_waitq = wait;
3286 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3288 if (req->file->f_op->read_iter)
3289 return call_read_iter(req->file, &req->rw.kiocb, iter);
3290 else if (req->file->f_op->read)
3291 return loop_rw_iter(READ, req, iter);
3296 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3298 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3299 struct kiocb *kiocb = &req->rw.kiocb;
3300 struct iov_iter __iter, *iter = &__iter;
3301 struct io_async_rw *rw = req->async_data;
3302 ssize_t io_size, ret, ret2;
3303 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3309 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3313 io_size = iov_iter_count(iter);
3314 req->result = io_size;
3316 /* Ensure we clear previously set non-block flag */
3317 if (!force_nonblock)
3318 kiocb->ki_flags &= ~IOCB_NOWAIT;
3320 kiocb->ki_flags |= IOCB_NOWAIT;
3322 /* If the file doesn't support async, just async punt */
3323 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3324 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3325 return ret ?: -EAGAIN;
3328 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3329 if (unlikely(ret)) {
3334 ret = io_iter_do_read(req, iter);
3336 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3337 req->flags &= ~REQ_F_REISSUE;
3338 /* IOPOLL retry should happen for io-wq threads */
3339 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3341 /* no retry on NONBLOCK nor RWF_NOWAIT */
3342 if (req->flags & REQ_F_NOWAIT)
3344 /* some cases will consume bytes even on error returns */
3345 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3347 } else if (ret == -EIOCBQUEUED) {
3349 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3350 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3351 /* read all, failed, already did sync or don't want to retry */
3355 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3360 rw = req->async_data;
3361 /* now use our persistent iterator, if we aren't already */
3366 rw->bytes_done += ret;
3367 /* if we can retry, do so with the callbacks armed */
3368 if (!io_rw_should_retry(req)) {
3369 kiocb->ki_flags &= ~IOCB_WAITQ;
3374 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3375 * we get -EIOCBQUEUED, then we'll get a notification when the
3376 * desired page gets unlocked. We can also get a partial read
3377 * here, and if we do, then just retry at the new offset.
3379 ret = io_iter_do_read(req, iter);
3380 if (ret == -EIOCBQUEUED)
3382 /* we got some bytes, but not all. retry. */
3383 kiocb->ki_flags &= ~IOCB_WAITQ;
3384 } while (ret > 0 && ret < io_size);
3386 kiocb_done(kiocb, ret, issue_flags);
3388 /* it's faster to check here then delegate to kfree */
3394 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3396 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3398 return io_prep_rw(req, sqe);
3401 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3403 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3404 struct kiocb *kiocb = &req->rw.kiocb;
3405 struct iov_iter __iter, *iter = &__iter;
3406 struct io_async_rw *rw = req->async_data;
3407 ssize_t ret, ret2, io_size;
3408 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3414 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3418 io_size = iov_iter_count(iter);
3419 req->result = io_size;
3421 /* Ensure we clear previously set non-block flag */
3422 if (!force_nonblock)
3423 kiocb->ki_flags &= ~IOCB_NOWAIT;
3425 kiocb->ki_flags |= IOCB_NOWAIT;
3427 /* If the file doesn't support async, just async punt */
3428 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3431 /* file path doesn't support NOWAIT for non-direct_IO */
3432 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3433 (req->flags & REQ_F_ISREG))
3436 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3441 * Open-code file_start_write here to grab freeze protection,
3442 * which will be released by another thread in
3443 * io_complete_rw(). Fool lockdep by telling it the lock got
3444 * released so that it doesn't complain about the held lock when
3445 * we return to userspace.
3447 if (req->flags & REQ_F_ISREG) {
3448 sb_start_write(file_inode(req->file)->i_sb);
3449 __sb_writers_release(file_inode(req->file)->i_sb,
3452 kiocb->ki_flags |= IOCB_WRITE;
3454 if (req->file->f_op->write_iter)
3455 ret2 = call_write_iter(req->file, kiocb, iter);
3456 else if (req->file->f_op->write)
3457 ret2 = loop_rw_iter(WRITE, req, iter);
3461 if (req->flags & REQ_F_REISSUE) {
3462 req->flags &= ~REQ_F_REISSUE;
3467 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3468 * retry them without IOCB_NOWAIT.
3470 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3472 /* no retry on NONBLOCK nor RWF_NOWAIT */
3473 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3475 if (!force_nonblock || ret2 != -EAGAIN) {
3476 /* IOPOLL retry should happen for io-wq threads */
3477 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3480 kiocb_done(kiocb, ret2, issue_flags);
3483 /* some cases will consume bytes even on error returns */
3484 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3485 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3486 return ret ?: -EAGAIN;
3489 /* it's reportedly faster than delegating the null check to kfree() */
3495 static int io_renameat_prep(struct io_kiocb *req,
3496 const struct io_uring_sqe *sqe)
3498 struct io_rename *ren = &req->rename;
3499 const char __user *oldf, *newf;
3501 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3503 if (sqe->ioprio || sqe->buf_index)
3505 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3508 ren->old_dfd = READ_ONCE(sqe->fd);
3509 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3510 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3511 ren->new_dfd = READ_ONCE(sqe->len);
3512 ren->flags = READ_ONCE(sqe->rename_flags);
3514 ren->oldpath = getname(oldf);
3515 if (IS_ERR(ren->oldpath))
3516 return PTR_ERR(ren->oldpath);
3518 ren->newpath = getname(newf);
3519 if (IS_ERR(ren->newpath)) {
3520 putname(ren->oldpath);
3521 return PTR_ERR(ren->newpath);
3524 req->flags |= REQ_F_NEED_CLEANUP;
3528 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3530 struct io_rename *ren = &req->rename;
3533 if (issue_flags & IO_URING_F_NONBLOCK)
3536 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3537 ren->newpath, ren->flags);
3539 req->flags &= ~REQ_F_NEED_CLEANUP;
3542 io_req_complete(req, ret);
3546 static int io_unlinkat_prep(struct io_kiocb *req,
3547 const struct io_uring_sqe *sqe)
3549 struct io_unlink *un = &req->unlink;
3550 const char __user *fname;
3552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3554 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3556 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3559 un->dfd = READ_ONCE(sqe->fd);
3561 un->flags = READ_ONCE(sqe->unlink_flags);
3562 if (un->flags & ~AT_REMOVEDIR)
3565 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3566 un->filename = getname(fname);
3567 if (IS_ERR(un->filename))
3568 return PTR_ERR(un->filename);
3570 req->flags |= REQ_F_NEED_CLEANUP;
3574 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3576 struct io_unlink *un = &req->unlink;
3579 if (issue_flags & IO_URING_F_NONBLOCK)
3582 if (un->flags & AT_REMOVEDIR)
3583 ret = do_rmdir(un->dfd, un->filename);
3585 ret = do_unlinkat(un->dfd, un->filename);
3587 req->flags &= ~REQ_F_NEED_CLEANUP;
3590 io_req_complete(req, ret);
3594 static int io_shutdown_prep(struct io_kiocb *req,
3595 const struct io_uring_sqe *sqe)
3597 #if defined(CONFIG_NET)
3598 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3600 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3604 req->shutdown.how = READ_ONCE(sqe->len);
3611 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3613 #if defined(CONFIG_NET)
3614 struct socket *sock;
3617 if (issue_flags & IO_URING_F_NONBLOCK)
3620 sock = sock_from_file(req->file);
3621 if (unlikely(!sock))
3624 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3627 io_req_complete(req, ret);
3634 static int __io_splice_prep(struct io_kiocb *req,
3635 const struct io_uring_sqe *sqe)
3637 struct io_splice *sp = &req->splice;
3638 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3640 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3644 sp->len = READ_ONCE(sqe->len);
3645 sp->flags = READ_ONCE(sqe->splice_flags);
3647 if (unlikely(sp->flags & ~valid_flags))
3650 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3651 (sp->flags & SPLICE_F_FD_IN_FIXED));
3654 req->flags |= REQ_F_NEED_CLEANUP;
3658 static int io_tee_prep(struct io_kiocb *req,
3659 const struct io_uring_sqe *sqe)
3661 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3663 return __io_splice_prep(req, sqe);
3666 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3668 struct io_splice *sp = &req->splice;
3669 struct file *in = sp->file_in;
3670 struct file *out = sp->file_out;
3671 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3674 if (issue_flags & IO_URING_F_NONBLOCK)
3677 ret = do_tee(in, out, sp->len, flags);
3679 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3681 req->flags &= ~REQ_F_NEED_CLEANUP;
3685 io_req_complete(req, ret);
3689 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3691 struct io_splice *sp = &req->splice;
3693 sp->off_in = READ_ONCE(sqe->splice_off_in);
3694 sp->off_out = READ_ONCE(sqe->off);
3695 return __io_splice_prep(req, sqe);
3698 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3700 struct io_splice *sp = &req->splice;
3701 struct file *in = sp->file_in;
3702 struct file *out = sp->file_out;
3703 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3704 loff_t *poff_in, *poff_out;
3707 if (issue_flags & IO_URING_F_NONBLOCK)
3710 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3711 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3714 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3716 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3718 req->flags &= ~REQ_F_NEED_CLEANUP;
3722 io_req_complete(req, ret);
3727 * IORING_OP_NOP just posts a completion event, nothing else.
3729 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3731 struct io_ring_ctx *ctx = req->ctx;
3733 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3736 __io_req_complete(req, issue_flags, 0, 0);
3740 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3742 struct io_ring_ctx *ctx = req->ctx;
3747 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3749 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3752 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3753 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3756 req->sync.off = READ_ONCE(sqe->off);
3757 req->sync.len = READ_ONCE(sqe->len);
3761 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3763 loff_t end = req->sync.off + req->sync.len;
3766 /* fsync always requires a blocking context */
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3770 ret = vfs_fsync_range(req->file, req->sync.off,
3771 end > 0 ? end : LLONG_MAX,
3772 req->sync.flags & IORING_FSYNC_DATASYNC);
3775 io_req_complete(req, ret);
3779 static int io_fallocate_prep(struct io_kiocb *req,
3780 const struct io_uring_sqe *sqe)
3782 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3784 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3787 req->sync.off = READ_ONCE(sqe->off);
3788 req->sync.len = READ_ONCE(sqe->addr);
3789 req->sync.mode = READ_ONCE(sqe->len);
3793 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3797 /* fallocate always requiring blocking context */
3798 if (issue_flags & IO_URING_F_NONBLOCK)
3800 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3804 io_req_complete(req, ret);
3808 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3810 const char __user *fname;
3813 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3815 if (unlikely(sqe->ioprio || sqe->buf_index))
3817 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3820 /* open.how should be already initialised */
3821 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3822 req->open.how.flags |= O_LARGEFILE;
3824 req->open.dfd = READ_ONCE(sqe->fd);
3825 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3826 req->open.filename = getname(fname);
3827 if (IS_ERR(req->open.filename)) {
3828 ret = PTR_ERR(req->open.filename);
3829 req->open.filename = NULL;
3832 req->open.nofile = rlimit(RLIMIT_NOFILE);
3833 req->flags |= REQ_F_NEED_CLEANUP;
3837 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3839 u64 mode = READ_ONCE(sqe->len);
3840 u64 flags = READ_ONCE(sqe->open_flags);
3842 req->open.how = build_open_how(flags, mode);
3843 return __io_openat_prep(req, sqe);
3846 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3848 struct open_how __user *how;
3852 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3853 len = READ_ONCE(sqe->len);
3854 if (len < OPEN_HOW_SIZE_VER0)
3857 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3862 return __io_openat_prep(req, sqe);
3865 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3867 struct open_flags op;
3870 bool resolve_nonblock;
3873 ret = build_open_flags(&req->open.how, &op);
3876 nonblock_set = op.open_flag & O_NONBLOCK;
3877 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3878 if (issue_flags & IO_URING_F_NONBLOCK) {
3880 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3881 * it'll always -EAGAIN
3883 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3885 op.lookup_flags |= LOOKUP_CACHED;
3886 op.open_flag |= O_NONBLOCK;
3889 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3893 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3896 * We could hang on to this 'fd' on retrying, but seems like
3897 * marginal gain for something that is now known to be a slower
3898 * path. So just put it, and we'll get a new one when we retry.
3902 ret = PTR_ERR(file);
3903 /* only retry if RESOLVE_CACHED wasn't already set by application */
3904 if (ret == -EAGAIN &&
3905 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3910 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3911 file->f_flags &= ~O_NONBLOCK;
3912 fsnotify_open(file);
3913 fd_install(ret, file);
3915 putname(req->open.filename);
3916 req->flags &= ~REQ_F_NEED_CLEANUP;
3919 __io_req_complete(req, issue_flags, ret, 0);
3923 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3925 return io_openat2(req, issue_flags);
3928 static int io_remove_buffers_prep(struct io_kiocb *req,
3929 const struct io_uring_sqe *sqe)
3931 struct io_provide_buf *p = &req->pbuf;
3934 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3937 tmp = READ_ONCE(sqe->fd);
3938 if (!tmp || tmp > USHRT_MAX)
3941 memset(p, 0, sizeof(*p));
3943 p->bgid = READ_ONCE(sqe->buf_group);
3947 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3948 int bgid, unsigned nbufs)
3952 /* shouldn't happen */
3956 /* the head kbuf is the list itself */
3957 while (!list_empty(&buf->list)) {
3958 struct io_buffer *nxt;
3960 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3961 list_del(&nxt->list);
3968 xa_erase(&ctx->io_buffers, bgid);
3973 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3975 struct io_provide_buf *p = &req->pbuf;
3976 struct io_ring_ctx *ctx = req->ctx;
3977 struct io_buffer *head;
3979 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3981 io_ring_submit_lock(ctx, !force_nonblock);
3983 lockdep_assert_held(&ctx->uring_lock);
3986 head = xa_load(&ctx->io_buffers, p->bgid);
3988 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3992 /* complete before unlock, IOPOLL may need the lock */
3993 __io_req_complete(req, issue_flags, ret, 0);
3994 io_ring_submit_unlock(ctx, !force_nonblock);
3998 static int io_provide_buffers_prep(struct io_kiocb *req,
3999 const struct io_uring_sqe *sqe)
4001 unsigned long size, tmp_check;
4002 struct io_provide_buf *p = &req->pbuf;
4005 if (sqe->ioprio || sqe->rw_flags)
4008 tmp = READ_ONCE(sqe->fd);
4009 if (!tmp || tmp > USHRT_MAX)
4012 p->addr = READ_ONCE(sqe->addr);
4013 p->len = READ_ONCE(sqe->len);
4015 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4018 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4021 size = (unsigned long)p->len * p->nbufs;
4022 if (!access_ok(u64_to_user_ptr(p->addr), size))
4025 p->bgid = READ_ONCE(sqe->buf_group);
4026 tmp = READ_ONCE(sqe->off);
4027 if (tmp > USHRT_MAX)
4033 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4035 struct io_buffer *buf;
4036 u64 addr = pbuf->addr;
4037 int i, bid = pbuf->bid;
4039 for (i = 0; i < pbuf->nbufs; i++) {
4040 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4045 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4050 INIT_LIST_HEAD(&buf->list);
4053 list_add_tail(&buf->list, &(*head)->list);
4057 return i ? i : -ENOMEM;
4060 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4062 struct io_provide_buf *p = &req->pbuf;
4063 struct io_ring_ctx *ctx = req->ctx;
4064 struct io_buffer *head, *list;
4066 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4068 io_ring_submit_lock(ctx, !force_nonblock);
4070 lockdep_assert_held(&ctx->uring_lock);
4072 list = head = xa_load(&ctx->io_buffers, p->bgid);
4074 ret = io_add_buffers(p, &head);
4075 if (ret >= 0 && !list) {
4076 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4078 __io_remove_buffers(ctx, head, p->bgid, -1U);
4082 /* complete before unlock, IOPOLL may need the lock */
4083 __io_req_complete(req, issue_flags, ret, 0);
4084 io_ring_submit_unlock(ctx, !force_nonblock);
4088 static int io_epoll_ctl_prep(struct io_kiocb *req,
4089 const struct io_uring_sqe *sqe)
4091 #if defined(CONFIG_EPOLL)
4092 if (sqe->ioprio || sqe->buf_index)
4094 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4097 req->epoll.epfd = READ_ONCE(sqe->fd);
4098 req->epoll.op = READ_ONCE(sqe->len);
4099 req->epoll.fd = READ_ONCE(sqe->off);
4101 if (ep_op_has_event(req->epoll.op)) {
4102 struct epoll_event __user *ev;
4104 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4105 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4115 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4117 #if defined(CONFIG_EPOLL)
4118 struct io_epoll *ie = &req->epoll;
4120 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4122 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4123 if (force_nonblock && ret == -EAGAIN)
4128 __io_req_complete(req, issue_flags, ret, 0);
4135 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4138 if (sqe->ioprio || sqe->buf_index || sqe->off)
4140 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4143 req->madvise.addr = READ_ONCE(sqe->addr);
4144 req->madvise.len = READ_ONCE(sqe->len);
4145 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4152 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4154 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4155 struct io_madvise *ma = &req->madvise;
4158 if (issue_flags & IO_URING_F_NONBLOCK)
4161 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4164 io_req_complete(req, ret);
4171 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4173 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4175 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4178 req->fadvise.offset = READ_ONCE(sqe->off);
4179 req->fadvise.len = READ_ONCE(sqe->len);
4180 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4184 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4186 struct io_fadvise *fa = &req->fadvise;
4189 if (issue_flags & IO_URING_F_NONBLOCK) {
4190 switch (fa->advice) {
4191 case POSIX_FADV_NORMAL:
4192 case POSIX_FADV_RANDOM:
4193 case POSIX_FADV_SEQUENTIAL:
4200 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4203 __io_req_complete(req, issue_flags, ret, 0);
4207 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4209 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4211 if (sqe->ioprio || sqe->buf_index)
4213 if (req->flags & REQ_F_FIXED_FILE)
4216 req->statx.dfd = READ_ONCE(sqe->fd);
4217 req->statx.mask = READ_ONCE(sqe->len);
4218 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4219 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4220 req->statx.flags = READ_ONCE(sqe->statx_flags);
4225 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4227 struct io_statx *ctx = &req->statx;
4230 if (issue_flags & IO_URING_F_NONBLOCK)
4233 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4238 io_req_complete(req, ret);
4242 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4244 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4246 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4247 sqe->rw_flags || sqe->buf_index)
4249 if (req->flags & REQ_F_FIXED_FILE)
4252 req->close.fd = READ_ONCE(sqe->fd);
4256 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4258 struct files_struct *files = current->files;
4259 struct io_close *close = &req->close;
4260 struct fdtable *fdt;
4261 struct file *file = NULL;
4264 spin_lock(&files->file_lock);
4265 fdt = files_fdtable(files);
4266 if (close->fd >= fdt->max_fds) {
4267 spin_unlock(&files->file_lock);
4270 file = fdt->fd[close->fd];
4271 if (!file || file->f_op == &io_uring_fops) {
4272 spin_unlock(&files->file_lock);
4277 /* if the file has a flush method, be safe and punt to async */
4278 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4279 spin_unlock(&files->file_lock);
4283 ret = __close_fd_get_file(close->fd, &file);
4284 spin_unlock(&files->file_lock);
4291 /* No ->flush() or already async, safely close from here */
4292 ret = filp_close(file, current->files);
4298 __io_req_complete(req, issue_flags, ret, 0);
4302 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4304 struct io_ring_ctx *ctx = req->ctx;
4306 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4308 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4311 req->sync.off = READ_ONCE(sqe->off);
4312 req->sync.len = READ_ONCE(sqe->len);
4313 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4317 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4321 /* sync_file_range always requires a blocking context */
4322 if (issue_flags & IO_URING_F_NONBLOCK)
4325 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4329 io_req_complete(req, ret);
4333 #if defined(CONFIG_NET)
4334 static int io_setup_async_msg(struct io_kiocb *req,
4335 struct io_async_msghdr *kmsg)
4337 struct io_async_msghdr *async_msg = req->async_data;
4341 if (io_alloc_async_data(req)) {
4342 kfree(kmsg->free_iov);
4345 async_msg = req->async_data;
4346 req->flags |= REQ_F_NEED_CLEANUP;
4347 memcpy(async_msg, kmsg, sizeof(*kmsg));
4348 async_msg->msg.msg_name = &async_msg->addr;
4349 /* if were using fast_iov, set it to the new one */
4350 if (!async_msg->free_iov)
4351 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4356 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4357 struct io_async_msghdr *iomsg)
4359 iomsg->msg.msg_name = &iomsg->addr;
4360 iomsg->free_iov = iomsg->fast_iov;
4361 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4362 req->sr_msg.msg_flags, &iomsg->free_iov);
4365 static int io_sendmsg_prep_async(struct io_kiocb *req)
4369 ret = io_sendmsg_copy_hdr(req, req->async_data);
4371 req->flags |= REQ_F_NEED_CLEANUP;
4375 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4377 struct io_sr_msg *sr = &req->sr_msg;
4379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4382 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4383 sr->len = READ_ONCE(sqe->len);
4384 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4385 if (sr->msg_flags & MSG_DONTWAIT)
4386 req->flags |= REQ_F_NOWAIT;
4388 #ifdef CONFIG_COMPAT
4389 if (req->ctx->compat)
4390 sr->msg_flags |= MSG_CMSG_COMPAT;
4395 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4397 struct io_async_msghdr iomsg, *kmsg;
4398 struct socket *sock;
4403 sock = sock_from_file(req->file);
4404 if (unlikely(!sock))
4407 kmsg = req->async_data;
4409 ret = io_sendmsg_copy_hdr(req, &iomsg);
4415 flags = req->sr_msg.msg_flags;
4416 if (issue_flags & IO_URING_F_NONBLOCK)
4417 flags |= MSG_DONTWAIT;
4418 if (flags & MSG_WAITALL)
4419 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4421 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4422 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4423 return io_setup_async_msg(req, kmsg);
4424 if (ret == -ERESTARTSYS)
4427 /* fast path, check for non-NULL to avoid function call */
4429 kfree(kmsg->free_iov);
4430 req->flags &= ~REQ_F_NEED_CLEANUP;
4433 __io_req_complete(req, issue_flags, ret, 0);
4437 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4439 struct io_sr_msg *sr = &req->sr_msg;
4442 struct socket *sock;
4447 sock = sock_from_file(req->file);
4448 if (unlikely(!sock))
4451 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4455 msg.msg_name = NULL;
4456 msg.msg_control = NULL;
4457 msg.msg_controllen = 0;
4458 msg.msg_namelen = 0;
4460 flags = req->sr_msg.msg_flags;
4461 if (issue_flags & IO_URING_F_NONBLOCK)
4462 flags |= MSG_DONTWAIT;
4463 if (flags & MSG_WAITALL)
4464 min_ret = iov_iter_count(&msg.msg_iter);
4466 msg.msg_flags = flags;
4467 ret = sock_sendmsg(sock, &msg);
4468 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4470 if (ret == -ERESTARTSYS)
4475 __io_req_complete(req, issue_flags, ret, 0);
4479 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4480 struct io_async_msghdr *iomsg)
4482 struct io_sr_msg *sr = &req->sr_msg;
4483 struct iovec __user *uiov;
4487 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4488 &iomsg->uaddr, &uiov, &iov_len);
4492 if (req->flags & REQ_F_BUFFER_SELECT) {
4495 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4497 sr->len = iomsg->fast_iov[0].iov_len;
4498 iomsg->free_iov = NULL;
4500 iomsg->free_iov = iomsg->fast_iov;
4501 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4502 &iomsg->free_iov, &iomsg->msg.msg_iter,
4511 #ifdef CONFIG_COMPAT
4512 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4513 struct io_async_msghdr *iomsg)
4515 struct io_sr_msg *sr = &req->sr_msg;
4516 struct compat_iovec __user *uiov;
4521 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4526 uiov = compat_ptr(ptr);
4527 if (req->flags & REQ_F_BUFFER_SELECT) {
4528 compat_ssize_t clen;
4532 if (!access_ok(uiov, sizeof(*uiov)))
4534 if (__get_user(clen, &uiov->iov_len))
4539 iomsg->free_iov = NULL;
4541 iomsg->free_iov = iomsg->fast_iov;
4542 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4543 UIO_FASTIOV, &iomsg->free_iov,
4544 &iomsg->msg.msg_iter, true);
4553 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4554 struct io_async_msghdr *iomsg)
4556 iomsg->msg.msg_name = &iomsg->addr;
4558 #ifdef CONFIG_COMPAT
4559 if (req->ctx->compat)
4560 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4563 return __io_recvmsg_copy_hdr(req, iomsg);
4566 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4569 struct io_sr_msg *sr = &req->sr_msg;
4570 struct io_buffer *kbuf;
4572 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4577 req->flags |= REQ_F_BUFFER_SELECTED;
4581 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4583 return io_put_kbuf(req, req->sr_msg.kbuf);
4586 static int io_recvmsg_prep_async(struct io_kiocb *req)
4590 ret = io_recvmsg_copy_hdr(req, req->async_data);
4592 req->flags |= REQ_F_NEED_CLEANUP;
4596 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4598 struct io_sr_msg *sr = &req->sr_msg;
4600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4603 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4604 sr->len = READ_ONCE(sqe->len);
4605 sr->bgid = READ_ONCE(sqe->buf_group);
4606 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4607 if (sr->msg_flags & MSG_DONTWAIT)
4608 req->flags |= REQ_F_NOWAIT;
4610 #ifdef CONFIG_COMPAT
4611 if (req->ctx->compat)
4612 sr->msg_flags |= MSG_CMSG_COMPAT;
4617 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4619 struct io_async_msghdr iomsg, *kmsg;
4620 struct socket *sock;
4621 struct io_buffer *kbuf;
4624 int ret, cflags = 0;
4625 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4627 sock = sock_from_file(req->file);
4628 if (unlikely(!sock))
4631 kmsg = req->async_data;
4633 ret = io_recvmsg_copy_hdr(req, &iomsg);
4639 if (req->flags & REQ_F_BUFFER_SELECT) {
4640 kbuf = io_recv_buffer_select(req, !force_nonblock);
4642 return PTR_ERR(kbuf);
4643 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4644 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4645 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4646 1, req->sr_msg.len);
4649 flags = req->sr_msg.msg_flags;
4651 flags |= MSG_DONTWAIT;
4652 if (flags & MSG_WAITALL)
4653 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4655 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4656 kmsg->uaddr, flags);
4657 if (force_nonblock && ret == -EAGAIN)
4658 return io_setup_async_msg(req, kmsg);
4659 if (ret == -ERESTARTSYS)
4662 if (req->flags & REQ_F_BUFFER_SELECTED)
4663 cflags = io_put_recv_kbuf(req);
4664 /* fast path, check for non-NULL to avoid function call */
4666 kfree(kmsg->free_iov);
4667 req->flags &= ~REQ_F_NEED_CLEANUP;
4668 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4670 __io_req_complete(req, issue_flags, ret, cflags);
4674 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4676 struct io_buffer *kbuf;
4677 struct io_sr_msg *sr = &req->sr_msg;
4679 void __user *buf = sr->buf;
4680 struct socket *sock;
4684 int ret, cflags = 0;
4685 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4687 sock = sock_from_file(req->file);
4688 if (unlikely(!sock))
4691 if (req->flags & REQ_F_BUFFER_SELECT) {
4692 kbuf = io_recv_buffer_select(req, !force_nonblock);
4694 return PTR_ERR(kbuf);
4695 buf = u64_to_user_ptr(kbuf->addr);
4698 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4702 msg.msg_name = NULL;
4703 msg.msg_control = NULL;
4704 msg.msg_controllen = 0;
4705 msg.msg_namelen = 0;
4706 msg.msg_iocb = NULL;
4709 flags = req->sr_msg.msg_flags;
4711 flags |= MSG_DONTWAIT;
4712 if (flags & MSG_WAITALL)
4713 min_ret = iov_iter_count(&msg.msg_iter);
4715 ret = sock_recvmsg(sock, &msg, flags);
4716 if (force_nonblock && ret == -EAGAIN)
4718 if (ret == -ERESTARTSYS)
4721 if (req->flags & REQ_F_BUFFER_SELECTED)
4722 cflags = io_put_recv_kbuf(req);
4723 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4725 __io_req_complete(req, issue_flags, ret, cflags);
4729 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4731 struct io_accept *accept = &req->accept;
4733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4735 if (sqe->ioprio || sqe->len || sqe->buf_index)
4738 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4739 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4740 accept->flags = READ_ONCE(sqe->accept_flags);
4741 accept->nofile = rlimit(RLIMIT_NOFILE);
4745 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4747 struct io_accept *accept = &req->accept;
4748 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4749 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4752 if (req->file->f_flags & O_NONBLOCK)
4753 req->flags |= REQ_F_NOWAIT;
4755 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4756 accept->addr_len, accept->flags,
4758 if (ret == -EAGAIN && force_nonblock)
4761 if (ret == -ERESTARTSYS)
4765 __io_req_complete(req, issue_flags, ret, 0);
4769 static int io_connect_prep_async(struct io_kiocb *req)
4771 struct io_async_connect *io = req->async_data;
4772 struct io_connect *conn = &req->connect;
4774 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4777 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4779 struct io_connect *conn = &req->connect;
4781 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4783 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4786 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4787 conn->addr_len = READ_ONCE(sqe->addr2);
4791 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4793 struct io_async_connect __io, *io;
4794 unsigned file_flags;
4796 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4798 if (req->async_data) {
4799 io = req->async_data;
4801 ret = move_addr_to_kernel(req->connect.addr,
4802 req->connect.addr_len,
4809 file_flags = force_nonblock ? O_NONBLOCK : 0;
4811 ret = __sys_connect_file(req->file, &io->address,
4812 req->connect.addr_len, file_flags);
4813 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4814 if (req->async_data)
4816 if (io_alloc_async_data(req)) {
4820 memcpy(req->async_data, &__io, sizeof(__io));
4823 if (ret == -ERESTARTSYS)
4828 __io_req_complete(req, issue_flags, ret, 0);
4831 #else /* !CONFIG_NET */
4832 #define IO_NETOP_FN(op) \
4833 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4835 return -EOPNOTSUPP; \
4838 #define IO_NETOP_PREP(op) \
4840 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4842 return -EOPNOTSUPP; \
4845 #define IO_NETOP_PREP_ASYNC(op) \
4847 static int io_##op##_prep_async(struct io_kiocb *req) \
4849 return -EOPNOTSUPP; \
4852 IO_NETOP_PREP_ASYNC(sendmsg);
4853 IO_NETOP_PREP_ASYNC(recvmsg);
4854 IO_NETOP_PREP_ASYNC(connect);
4855 IO_NETOP_PREP(accept);
4858 #endif /* CONFIG_NET */
4860 struct io_poll_table {
4861 struct poll_table_struct pt;
4862 struct io_kiocb *req;
4867 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4868 __poll_t mask, io_req_tw_func_t func)
4870 /* for instances that support it check for an event match first: */
4871 if (mask && !(mask & poll->events))
4874 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4876 list_del_init(&poll->wait.entry);
4879 req->io_task_work.func = func;
4882 * If this fails, then the task is exiting. When a task exits, the
4883 * work gets canceled, so just cancel this request as well instead
4884 * of executing it. We can't safely execute it anyway, as we may not
4885 * have the needed state needed for it anyway.
4887 io_req_task_work_add(req);
4891 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4892 __acquires(&req->ctx->completion_lock)
4894 struct io_ring_ctx *ctx = req->ctx;
4896 if (unlikely(req->task->flags & PF_EXITING))
4897 WRITE_ONCE(poll->canceled, true);
4899 if (!req->result && !READ_ONCE(poll->canceled)) {
4900 struct poll_table_struct pt = { ._key = poll->events };
4902 req->result = vfs_poll(req->file, &pt) & poll->events;
4905 spin_lock(&ctx->completion_lock);
4906 if (!req->result && !READ_ONCE(poll->canceled)) {
4907 add_wait_queue(poll->head, &poll->wait);
4914 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4916 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4917 if (req->opcode == IORING_OP_POLL_ADD)
4918 return req->async_data;
4919 return req->apoll->double_poll;
4922 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4924 if (req->opcode == IORING_OP_POLL_ADD)
4926 return &req->apoll->poll;
4929 static void io_poll_remove_double(struct io_kiocb *req)
4930 __must_hold(&req->ctx->completion_lock)
4932 struct io_poll_iocb *poll = io_poll_get_double(req);
4934 lockdep_assert_held(&req->ctx->completion_lock);
4936 if (poll && poll->head) {
4937 struct wait_queue_head *head = poll->head;
4939 spin_lock_irq(&head->lock);
4940 list_del_init(&poll->wait.entry);
4941 if (poll->wait.private)
4944 spin_unlock_irq(&head->lock);
4948 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4949 __must_hold(&req->ctx->completion_lock)
4951 struct io_ring_ctx *ctx = req->ctx;
4952 unsigned flags = IORING_CQE_F_MORE;
4955 if (READ_ONCE(req->poll.canceled)) {
4957 req->poll.events |= EPOLLONESHOT;
4959 error = mangle_poll(mask);
4961 if (req->poll.events & EPOLLONESHOT)
4963 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4964 req->poll.done = true;
4967 if (flags & IORING_CQE_F_MORE)
4970 io_commit_cqring(ctx);
4971 return !(flags & IORING_CQE_F_MORE);
4974 static void io_poll_task_func(struct io_kiocb *req)
4976 struct io_ring_ctx *ctx = req->ctx;
4977 struct io_kiocb *nxt;
4979 if (io_poll_rewait(req, &req->poll)) {
4980 spin_unlock(&ctx->completion_lock);
4984 done = io_poll_complete(req, req->result);
4986 io_poll_remove_double(req);
4987 hash_del(&req->hash_node);
4990 add_wait_queue(req->poll.head, &req->poll.wait);
4992 spin_unlock(&ctx->completion_lock);
4993 io_cqring_ev_posted(ctx);
4996 nxt = io_put_req_find_next(req);
4998 io_req_task_submit(nxt);
5003 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5004 int sync, void *key)
5006 struct io_kiocb *req = wait->private;
5007 struct io_poll_iocb *poll = io_poll_get_single(req);
5008 __poll_t mask = key_to_poll(key);
5009 unsigned long flags;
5011 /* for instances that support it check for an event match first: */
5012 if (mask && !(mask & poll->events))
5014 if (!(poll->events & EPOLLONESHOT))
5015 return poll->wait.func(&poll->wait, mode, sync, key);
5017 list_del_init(&wait->entry);
5022 spin_lock_irqsave(&poll->head->lock, flags);
5023 done = list_empty(&poll->wait.entry);
5025 list_del_init(&poll->wait.entry);
5026 /* make sure double remove sees this as being gone */
5027 wait->private = NULL;
5028 spin_unlock_irqrestore(&poll->head->lock, flags);
5030 /* use wait func handler, so it matches the rq type */
5031 poll->wait.func(&poll->wait, mode, sync, key);
5038 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5039 wait_queue_func_t wake_func)
5043 poll->canceled = false;
5044 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5045 /* mask in events that we always want/need */
5046 poll->events = events | IO_POLL_UNMASK;
5047 INIT_LIST_HEAD(&poll->wait.entry);
5048 init_waitqueue_func_entry(&poll->wait, wake_func);
5051 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5052 struct wait_queue_head *head,
5053 struct io_poll_iocb **poll_ptr)
5055 struct io_kiocb *req = pt->req;
5058 * The file being polled uses multiple waitqueues for poll handling
5059 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5062 if (unlikely(pt->nr_entries)) {
5063 struct io_poll_iocb *poll_one = poll;
5065 /* already have a 2nd entry, fail a third attempt */
5067 pt->error = -EINVAL;
5071 * Can't handle multishot for double wait for now, turn it
5072 * into one-shot mode.
5074 if (!(poll_one->events & EPOLLONESHOT))
5075 poll_one->events |= EPOLLONESHOT;
5076 /* double add on the same waitqueue head, ignore */
5077 if (poll_one->head == head)
5079 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5081 pt->error = -ENOMEM;
5084 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5086 poll->wait.private = req;
5093 if (poll->events & EPOLLEXCLUSIVE)
5094 add_wait_queue_exclusive(head, &poll->wait);
5096 add_wait_queue(head, &poll->wait);
5099 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5100 struct poll_table_struct *p)
5102 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5103 struct async_poll *apoll = pt->req->apoll;
5105 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5108 static void io_async_task_func(struct io_kiocb *req)
5110 struct async_poll *apoll = req->apoll;
5111 struct io_ring_ctx *ctx = req->ctx;
5113 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5115 if (io_poll_rewait(req, &apoll->poll)) {
5116 spin_unlock(&ctx->completion_lock);
5120 hash_del(&req->hash_node);
5121 io_poll_remove_double(req);
5122 spin_unlock(&ctx->completion_lock);
5124 if (!READ_ONCE(apoll->poll.canceled))
5125 io_req_task_submit(req);
5127 io_req_complete_failed(req, -ECANCELED);
5130 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5133 struct io_kiocb *req = wait->private;
5134 struct io_poll_iocb *poll = &req->apoll->poll;
5136 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5139 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5142 static void io_poll_req_insert(struct io_kiocb *req)
5144 struct io_ring_ctx *ctx = req->ctx;
5145 struct hlist_head *list;
5147 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5148 hlist_add_head(&req->hash_node, list);
5151 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5152 struct io_poll_iocb *poll,
5153 struct io_poll_table *ipt, __poll_t mask,
5154 wait_queue_func_t wake_func)
5155 __acquires(&ctx->completion_lock)
5157 struct io_ring_ctx *ctx = req->ctx;
5158 bool cancel = false;
5160 INIT_HLIST_NODE(&req->hash_node);
5161 io_init_poll_iocb(poll, mask, wake_func);
5162 poll->file = req->file;
5163 poll->wait.private = req;
5165 ipt->pt._key = mask;
5168 ipt->nr_entries = 0;
5170 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5171 if (unlikely(!ipt->nr_entries) && !ipt->error)
5172 ipt->error = -EINVAL;
5174 spin_lock(&ctx->completion_lock);
5175 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5176 io_poll_remove_double(req);
5177 if (likely(poll->head)) {
5178 spin_lock_irq(&poll->head->lock);
5179 if (unlikely(list_empty(&poll->wait.entry))) {
5185 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5186 list_del_init(&poll->wait.entry);
5188 WRITE_ONCE(poll->canceled, true);
5189 else if (!poll->done) /* actually waiting for an event */
5190 io_poll_req_insert(req);
5191 spin_unlock_irq(&poll->head->lock);
5203 static int io_arm_poll_handler(struct io_kiocb *req)
5205 const struct io_op_def *def = &io_op_defs[req->opcode];
5206 struct io_ring_ctx *ctx = req->ctx;
5207 struct async_poll *apoll;
5208 struct io_poll_table ipt;
5209 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5212 if (!req->file || !file_can_poll(req->file))
5213 return IO_APOLL_ABORTED;
5214 if (req->flags & REQ_F_POLLED)
5215 return IO_APOLL_ABORTED;
5216 if (!def->pollin && !def->pollout)
5217 return IO_APOLL_ABORTED;
5221 mask |= POLLIN | POLLRDNORM;
5223 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5224 if ((req->opcode == IORING_OP_RECVMSG) &&
5225 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5229 mask |= POLLOUT | POLLWRNORM;
5232 /* if we can't nonblock try, then no point in arming a poll handler */
5233 if (!io_file_supports_nowait(req, rw))
5234 return IO_APOLL_ABORTED;
5236 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5237 if (unlikely(!apoll))
5238 return IO_APOLL_ABORTED;
5239 apoll->double_poll = NULL;
5241 req->flags |= REQ_F_POLLED;
5242 ipt.pt._qproc = io_async_queue_proc;
5243 io_req_set_refcount(req);
5245 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5247 spin_unlock(&ctx->completion_lock);
5248 if (ret || ipt.error)
5249 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5251 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5252 mask, apoll->poll.events);
5256 static bool __io_poll_remove_one(struct io_kiocb *req,
5257 struct io_poll_iocb *poll, bool do_cancel)
5258 __must_hold(&req->ctx->completion_lock)
5260 bool do_complete = false;
5264 spin_lock_irq(&poll->head->lock);
5266 WRITE_ONCE(poll->canceled, true);
5267 if (!list_empty(&poll->wait.entry)) {
5268 list_del_init(&poll->wait.entry);
5271 spin_unlock_irq(&poll->head->lock);
5272 hash_del(&req->hash_node);
5276 static bool io_poll_remove_one(struct io_kiocb *req)
5277 __must_hold(&req->ctx->completion_lock)
5281 io_poll_remove_double(req);
5282 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5285 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5286 io_commit_cqring(req->ctx);
5288 io_put_req_deferred(req);
5294 * Returns true if we found and killed one or more poll requests
5296 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5299 struct hlist_node *tmp;
5300 struct io_kiocb *req;
5303 spin_lock(&ctx->completion_lock);
5304 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5305 struct hlist_head *list;
5307 list = &ctx->cancel_hash[i];
5308 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5309 if (io_match_task(req, tsk, cancel_all))
5310 posted += io_poll_remove_one(req);
5313 spin_unlock(&ctx->completion_lock);
5316 io_cqring_ev_posted(ctx);
5321 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5323 __must_hold(&ctx->completion_lock)
5325 struct hlist_head *list;
5326 struct io_kiocb *req;
5328 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5329 hlist_for_each_entry(req, list, hash_node) {
5330 if (sqe_addr != req->user_data)
5332 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5339 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5341 __must_hold(&ctx->completion_lock)
5343 struct io_kiocb *req;
5345 req = io_poll_find(ctx, sqe_addr, poll_only);
5348 if (io_poll_remove_one(req))
5354 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5359 events = READ_ONCE(sqe->poll32_events);
5361 events = swahw32(events);
5363 if (!(flags & IORING_POLL_ADD_MULTI))
5364 events |= EPOLLONESHOT;
5365 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5368 static int io_poll_update_prep(struct io_kiocb *req,
5369 const struct io_uring_sqe *sqe)
5371 struct io_poll_update *upd = &req->poll_update;
5374 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5376 if (sqe->ioprio || sqe->buf_index)
5378 flags = READ_ONCE(sqe->len);
5379 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5380 IORING_POLL_ADD_MULTI))
5382 /* meaningless without update */
5383 if (flags == IORING_POLL_ADD_MULTI)
5386 upd->old_user_data = READ_ONCE(sqe->addr);
5387 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5388 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5390 upd->new_user_data = READ_ONCE(sqe->off);
5391 if (!upd->update_user_data && upd->new_user_data)
5393 if (upd->update_events)
5394 upd->events = io_poll_parse_events(sqe, flags);
5395 else if (sqe->poll32_events)
5401 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5404 struct io_kiocb *req = wait->private;
5405 struct io_poll_iocb *poll = &req->poll;
5407 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5410 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5411 struct poll_table_struct *p)
5413 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5415 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5418 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5420 struct io_poll_iocb *poll = &req->poll;
5423 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5425 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5427 flags = READ_ONCE(sqe->len);
5428 if (flags & ~IORING_POLL_ADD_MULTI)
5431 io_req_set_refcount(req);
5432 poll->events = io_poll_parse_events(sqe, flags);
5436 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5438 struct io_poll_iocb *poll = &req->poll;
5439 struct io_ring_ctx *ctx = req->ctx;
5440 struct io_poll_table ipt;
5443 ipt.pt._qproc = io_poll_queue_proc;
5445 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5448 if (mask) { /* no async, we'd stolen it */
5450 io_poll_complete(req, mask);
5452 spin_unlock(&ctx->completion_lock);
5455 io_cqring_ev_posted(ctx);
5456 if (poll->events & EPOLLONESHOT)
5462 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5464 struct io_ring_ctx *ctx = req->ctx;
5465 struct io_kiocb *preq;
5469 spin_lock(&ctx->completion_lock);
5470 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5476 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5478 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5483 * Don't allow racy completion with singleshot, as we cannot safely
5484 * update those. For multishot, if we're racing with completion, just
5485 * let completion re-add it.
5487 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5488 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5492 /* we now have a detached poll request. reissue. */
5496 spin_unlock(&ctx->completion_lock);
5498 io_req_complete(req, ret);
5501 /* only mask one event flags, keep behavior flags */
5502 if (req->poll_update.update_events) {
5503 preq->poll.events &= ~0xffff;
5504 preq->poll.events |= req->poll_update.events & 0xffff;
5505 preq->poll.events |= IO_POLL_UNMASK;
5507 if (req->poll_update.update_user_data)
5508 preq->user_data = req->poll_update.new_user_data;
5509 spin_unlock(&ctx->completion_lock);
5511 /* complete update request, we're done with it */
5512 io_req_complete(req, ret);
5515 ret = io_poll_add(preq, issue_flags);
5518 io_req_complete(preq, ret);
5524 static void io_req_task_timeout(struct io_kiocb *req)
5526 struct io_ring_ctx *ctx = req->ctx;
5528 spin_lock(&ctx->completion_lock);
5529 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5530 io_commit_cqring(ctx);
5531 spin_unlock(&ctx->completion_lock);
5533 io_cqring_ev_posted(ctx);
5538 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5540 struct io_timeout_data *data = container_of(timer,
5541 struct io_timeout_data, timer);
5542 struct io_kiocb *req = data->req;
5543 struct io_ring_ctx *ctx = req->ctx;
5544 unsigned long flags;
5546 spin_lock_irqsave(&ctx->timeout_lock, flags);
5547 list_del_init(&req->timeout.list);
5548 atomic_set(&req->ctx->cq_timeouts,
5549 atomic_read(&req->ctx->cq_timeouts) + 1);
5550 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5552 req->io_task_work.func = io_req_task_timeout;
5553 io_req_task_work_add(req);
5554 return HRTIMER_NORESTART;
5557 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5559 __must_hold(&ctx->timeout_lock)
5561 struct io_timeout_data *io;
5562 struct io_kiocb *req;
5565 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5566 found = user_data == req->user_data;
5571 return ERR_PTR(-ENOENT);
5573 io = req->async_data;
5574 if (hrtimer_try_to_cancel(&io->timer) == -1)
5575 return ERR_PTR(-EALREADY);
5576 list_del_init(&req->timeout.list);
5580 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5581 __must_hold(&ctx->timeout_lock)
5583 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5586 return PTR_ERR(req);
5589 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5590 io_put_req_deferred(req);
5594 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5595 struct timespec64 *ts, enum hrtimer_mode mode)
5596 __must_hold(&ctx->timeout_lock)
5598 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5599 struct io_timeout_data *data;
5602 return PTR_ERR(req);
5604 req->timeout.off = 0; /* noseq */
5605 data = req->async_data;
5606 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5607 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5608 data->timer.function = io_timeout_fn;
5609 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5613 static int io_timeout_remove_prep(struct io_kiocb *req,
5614 const struct io_uring_sqe *sqe)
5616 struct io_timeout_rem *tr = &req->timeout_rem;
5618 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5620 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5622 if (sqe->ioprio || sqe->buf_index || sqe->len)
5625 tr->addr = READ_ONCE(sqe->addr);
5626 tr->flags = READ_ONCE(sqe->timeout_flags);
5627 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5628 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5630 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5632 } else if (tr->flags) {
5633 /* timeout removal doesn't support flags */
5640 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5642 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5647 * Remove or update an existing timeout command
5649 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5651 struct io_timeout_rem *tr = &req->timeout_rem;
5652 struct io_ring_ctx *ctx = req->ctx;
5655 spin_lock_irq(&ctx->timeout_lock);
5656 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5657 ret = io_timeout_cancel(ctx, tr->addr);
5659 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5660 io_translate_timeout_mode(tr->flags));
5661 spin_unlock_irq(&ctx->timeout_lock);
5663 spin_lock(&ctx->completion_lock);
5664 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5665 io_commit_cqring(ctx);
5666 spin_unlock(&ctx->completion_lock);
5667 io_cqring_ev_posted(ctx);
5674 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5675 bool is_timeout_link)
5677 struct io_timeout_data *data;
5679 u32 off = READ_ONCE(sqe->off);
5681 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5683 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5685 if (off && is_timeout_link)
5687 flags = READ_ONCE(sqe->timeout_flags);
5688 if (flags & ~IORING_TIMEOUT_ABS)
5691 req->timeout.off = off;
5692 if (unlikely(off && !req->ctx->off_timeout_used))
5693 req->ctx->off_timeout_used = true;
5695 if (!req->async_data && io_alloc_async_data(req))
5698 data = req->async_data;
5701 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5704 data->mode = io_translate_timeout_mode(flags);
5705 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5707 if (is_timeout_link) {
5708 struct io_submit_link *link = &req->ctx->submit_state.link;
5712 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5714 req->timeout.head = link->last;
5715 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5720 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5722 struct io_ring_ctx *ctx = req->ctx;
5723 struct io_timeout_data *data = req->async_data;
5724 struct list_head *entry;
5725 u32 tail, off = req->timeout.off;
5727 spin_lock_irq(&ctx->timeout_lock);
5730 * sqe->off holds how many events that need to occur for this
5731 * timeout event to be satisfied. If it isn't set, then this is
5732 * a pure timeout request, sequence isn't used.
5734 if (io_is_timeout_noseq(req)) {
5735 entry = ctx->timeout_list.prev;
5739 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5740 req->timeout.target_seq = tail + off;
5742 /* Update the last seq here in case io_flush_timeouts() hasn't.
5743 * This is safe because ->completion_lock is held, and submissions
5744 * and completions are never mixed in the same ->completion_lock section.
5746 ctx->cq_last_tm_flush = tail;
5749 * Insertion sort, ensuring the first entry in the list is always
5750 * the one we need first.
5752 list_for_each_prev(entry, &ctx->timeout_list) {
5753 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5756 if (io_is_timeout_noseq(nxt))
5758 /* nxt.seq is behind @tail, otherwise would've been completed */
5759 if (off >= nxt->timeout.target_seq - tail)
5763 list_add(&req->timeout.list, entry);
5764 data->timer.function = io_timeout_fn;
5765 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5766 spin_unlock_irq(&ctx->timeout_lock);
5770 struct io_cancel_data {
5771 struct io_ring_ctx *ctx;
5775 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5778 struct io_cancel_data *cd = data;
5780 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5783 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5784 struct io_ring_ctx *ctx)
5786 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5787 enum io_wq_cancel cancel_ret;
5790 if (!tctx || !tctx->io_wq)
5793 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5794 switch (cancel_ret) {
5795 case IO_WQ_CANCEL_OK:
5798 case IO_WQ_CANCEL_RUNNING:
5801 case IO_WQ_CANCEL_NOTFOUND:
5809 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5810 __acquires(&req->ctx->completion_lock)
5812 struct io_ring_ctx *ctx = req->ctx;
5815 WARN_ON_ONCE(req->task != current);
5817 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5818 spin_lock(&ctx->completion_lock);
5821 spin_lock_irq(&ctx->timeout_lock);
5822 ret = io_timeout_cancel(ctx, sqe_addr);
5823 spin_unlock_irq(&ctx->timeout_lock);
5826 return io_poll_cancel(ctx, sqe_addr, false);
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);
5853 spin_unlock(&ctx->completion_lock);
5855 /* slow path, try all io-wq's */
5856 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5858 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5859 struct io_uring_task *tctx = node->task->io_uring;
5861 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5865 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5867 spin_lock(&ctx->completion_lock);
5869 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5870 io_commit_cqring(ctx);
5871 spin_unlock(&ctx->completion_lock);
5872 io_cqring_ev_posted(ctx);
5880 static int io_rsrc_update_prep(struct io_kiocb *req,
5881 const struct io_uring_sqe *sqe)
5883 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5885 if (sqe->ioprio || sqe->rw_flags)
5888 req->rsrc_update.offset = READ_ONCE(sqe->off);
5889 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5890 if (!req->rsrc_update.nr_args)
5892 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5896 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5898 struct io_ring_ctx *ctx = req->ctx;
5899 struct io_uring_rsrc_update2 up;
5902 if (issue_flags & IO_URING_F_NONBLOCK)
5905 up.offset = req->rsrc_update.offset;
5906 up.data = req->rsrc_update.arg;
5911 mutex_lock(&ctx->uring_lock);
5912 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5913 &up, req->rsrc_update.nr_args);
5914 mutex_unlock(&ctx->uring_lock);
5918 __io_req_complete(req, issue_flags, ret, 0);
5922 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5924 switch (req->opcode) {
5927 case IORING_OP_READV:
5928 case IORING_OP_READ_FIXED:
5929 case IORING_OP_READ:
5930 return io_read_prep(req, sqe);
5931 case IORING_OP_WRITEV:
5932 case IORING_OP_WRITE_FIXED:
5933 case IORING_OP_WRITE:
5934 return io_write_prep(req, sqe);
5935 case IORING_OP_POLL_ADD:
5936 return io_poll_add_prep(req, sqe);
5937 case IORING_OP_POLL_REMOVE:
5938 return io_poll_update_prep(req, sqe);
5939 case IORING_OP_FSYNC:
5940 return io_fsync_prep(req, sqe);
5941 case IORING_OP_SYNC_FILE_RANGE:
5942 return io_sfr_prep(req, sqe);
5943 case IORING_OP_SENDMSG:
5944 case IORING_OP_SEND:
5945 return io_sendmsg_prep(req, sqe);
5946 case IORING_OP_RECVMSG:
5947 case IORING_OP_RECV:
5948 return io_recvmsg_prep(req, sqe);
5949 case IORING_OP_CONNECT:
5950 return io_connect_prep(req, sqe);
5951 case IORING_OP_TIMEOUT:
5952 return io_timeout_prep(req, sqe, false);
5953 case IORING_OP_TIMEOUT_REMOVE:
5954 return io_timeout_remove_prep(req, sqe);
5955 case IORING_OP_ASYNC_CANCEL:
5956 return io_async_cancel_prep(req, sqe);
5957 case IORING_OP_LINK_TIMEOUT:
5958 return io_timeout_prep(req, sqe, true);
5959 case IORING_OP_ACCEPT:
5960 return io_accept_prep(req, sqe);
5961 case IORING_OP_FALLOCATE:
5962 return io_fallocate_prep(req, sqe);
5963 case IORING_OP_OPENAT:
5964 return io_openat_prep(req, sqe);
5965 case IORING_OP_CLOSE:
5966 return io_close_prep(req, sqe);
5967 case IORING_OP_FILES_UPDATE:
5968 return io_rsrc_update_prep(req, sqe);
5969 case IORING_OP_STATX:
5970 return io_statx_prep(req, sqe);
5971 case IORING_OP_FADVISE:
5972 return io_fadvise_prep(req, sqe);
5973 case IORING_OP_MADVISE:
5974 return io_madvise_prep(req, sqe);
5975 case IORING_OP_OPENAT2:
5976 return io_openat2_prep(req, sqe);
5977 case IORING_OP_EPOLL_CTL:
5978 return io_epoll_ctl_prep(req, sqe);
5979 case IORING_OP_SPLICE:
5980 return io_splice_prep(req, sqe);
5981 case IORING_OP_PROVIDE_BUFFERS:
5982 return io_provide_buffers_prep(req, sqe);
5983 case IORING_OP_REMOVE_BUFFERS:
5984 return io_remove_buffers_prep(req, sqe);
5986 return io_tee_prep(req, sqe);
5987 case IORING_OP_SHUTDOWN:
5988 return io_shutdown_prep(req, sqe);
5989 case IORING_OP_RENAMEAT:
5990 return io_renameat_prep(req, sqe);
5991 case IORING_OP_UNLINKAT:
5992 return io_unlinkat_prep(req, sqe);
5995 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6000 static int io_req_prep_async(struct io_kiocb *req)
6002 if (!io_op_defs[req->opcode].needs_async_setup)
6004 if (WARN_ON_ONCE(req->async_data))
6006 if (io_alloc_async_data(req))
6009 switch (req->opcode) {
6010 case IORING_OP_READV:
6011 return io_rw_prep_async(req, READ);
6012 case IORING_OP_WRITEV:
6013 return io_rw_prep_async(req, WRITE);
6014 case IORING_OP_SENDMSG:
6015 return io_sendmsg_prep_async(req);
6016 case IORING_OP_RECVMSG:
6017 return io_recvmsg_prep_async(req);
6018 case IORING_OP_CONNECT:
6019 return io_connect_prep_async(req);
6021 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6026 static u32 io_get_sequence(struct io_kiocb *req)
6028 u32 seq = req->ctx->cached_sq_head;
6030 /* need original cached_sq_head, but it was increased for each req */
6031 io_for_each_link(req, req)
6036 static bool io_drain_req(struct io_kiocb *req)
6038 struct io_kiocb *pos;
6039 struct io_ring_ctx *ctx = req->ctx;
6040 struct io_defer_entry *de;
6045 * If we need to drain a request in the middle of a link, drain the
6046 * head request and the next request/link after the current link.
6047 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6048 * maintained for every request of our link.
6050 if (ctx->drain_next) {
6051 req->flags |= REQ_F_IO_DRAIN;
6052 ctx->drain_next = false;
6054 /* not interested in head, start from the first linked */
6055 io_for_each_link(pos, req->link) {
6056 if (pos->flags & REQ_F_IO_DRAIN) {
6057 ctx->drain_next = true;
6058 req->flags |= REQ_F_IO_DRAIN;
6063 /* Still need defer if there is pending req in defer list. */
6064 if (likely(list_empty_careful(&ctx->defer_list) &&
6065 !(req->flags & REQ_F_IO_DRAIN))) {
6066 ctx->drain_active = false;
6070 seq = io_get_sequence(req);
6071 /* Still a chance to pass the sequence check */
6072 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6075 ret = io_req_prep_async(req);
6078 io_prep_async_link(req);
6079 de = kmalloc(sizeof(*de), GFP_KERNEL);
6083 io_req_complete_failed(req, ret);
6087 spin_lock(&ctx->completion_lock);
6088 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6089 spin_unlock(&ctx->completion_lock);
6091 io_queue_async_work(req);
6095 trace_io_uring_defer(ctx, req, req->user_data);
6098 list_add_tail(&de->list, &ctx->defer_list);
6099 spin_unlock(&ctx->completion_lock);
6103 static void io_clean_op(struct io_kiocb *req)
6105 if (req->flags & REQ_F_BUFFER_SELECTED) {
6106 switch (req->opcode) {
6107 case IORING_OP_READV:
6108 case IORING_OP_READ_FIXED:
6109 case IORING_OP_READ:
6110 kfree((void *)(unsigned long)req->rw.addr);
6112 case IORING_OP_RECVMSG:
6113 case IORING_OP_RECV:
6114 kfree(req->sr_msg.kbuf);
6119 if (req->flags & REQ_F_NEED_CLEANUP) {
6120 switch (req->opcode) {
6121 case IORING_OP_READV:
6122 case IORING_OP_READ_FIXED:
6123 case IORING_OP_READ:
6124 case IORING_OP_WRITEV:
6125 case IORING_OP_WRITE_FIXED:
6126 case IORING_OP_WRITE: {
6127 struct io_async_rw *io = req->async_data;
6129 kfree(io->free_iovec);
6132 case IORING_OP_RECVMSG:
6133 case IORING_OP_SENDMSG: {
6134 struct io_async_msghdr *io = req->async_data;
6136 kfree(io->free_iov);
6139 case IORING_OP_SPLICE:
6141 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6142 io_put_file(req->splice.file_in);
6144 case IORING_OP_OPENAT:
6145 case IORING_OP_OPENAT2:
6146 if (req->open.filename)
6147 putname(req->open.filename);
6149 case IORING_OP_RENAMEAT:
6150 putname(req->rename.oldpath);
6151 putname(req->rename.newpath);
6153 case IORING_OP_UNLINKAT:
6154 putname(req->unlink.filename);
6158 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6159 kfree(req->apoll->double_poll);
6163 if (req->flags & REQ_F_INFLIGHT) {
6164 struct io_uring_task *tctx = req->task->io_uring;
6166 atomic_dec(&tctx->inflight_tracked);
6168 if (req->flags & REQ_F_CREDS)
6169 put_cred(req->creds);
6171 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6174 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6176 struct io_ring_ctx *ctx = req->ctx;
6177 const struct cred *creds = NULL;
6180 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6181 creds = override_creds(req->creds);
6183 switch (req->opcode) {
6185 ret = io_nop(req, issue_flags);
6187 case IORING_OP_READV:
6188 case IORING_OP_READ_FIXED:
6189 case IORING_OP_READ:
6190 ret = io_read(req, issue_flags);
6192 case IORING_OP_WRITEV:
6193 case IORING_OP_WRITE_FIXED:
6194 case IORING_OP_WRITE:
6195 ret = io_write(req, issue_flags);
6197 case IORING_OP_FSYNC:
6198 ret = io_fsync(req, issue_flags);
6200 case IORING_OP_POLL_ADD:
6201 ret = io_poll_add(req, issue_flags);
6203 case IORING_OP_POLL_REMOVE:
6204 ret = io_poll_update(req, issue_flags);
6206 case IORING_OP_SYNC_FILE_RANGE:
6207 ret = io_sync_file_range(req, issue_flags);
6209 case IORING_OP_SENDMSG:
6210 ret = io_sendmsg(req, issue_flags);
6212 case IORING_OP_SEND:
6213 ret = io_send(req, issue_flags);
6215 case IORING_OP_RECVMSG:
6216 ret = io_recvmsg(req, issue_flags);
6218 case IORING_OP_RECV:
6219 ret = io_recv(req, issue_flags);
6221 case IORING_OP_TIMEOUT:
6222 ret = io_timeout(req, issue_flags);
6224 case IORING_OP_TIMEOUT_REMOVE:
6225 ret = io_timeout_remove(req, issue_flags);
6227 case IORING_OP_ACCEPT:
6228 ret = io_accept(req, issue_flags);
6230 case IORING_OP_CONNECT:
6231 ret = io_connect(req, issue_flags);
6233 case IORING_OP_ASYNC_CANCEL:
6234 ret = io_async_cancel(req, issue_flags);
6236 case IORING_OP_FALLOCATE:
6237 ret = io_fallocate(req, issue_flags);
6239 case IORING_OP_OPENAT:
6240 ret = io_openat(req, issue_flags);
6242 case IORING_OP_CLOSE:
6243 ret = io_close(req, issue_flags);
6245 case IORING_OP_FILES_UPDATE:
6246 ret = io_files_update(req, issue_flags);
6248 case IORING_OP_STATX:
6249 ret = io_statx(req, issue_flags);
6251 case IORING_OP_FADVISE:
6252 ret = io_fadvise(req, issue_flags);
6254 case IORING_OP_MADVISE:
6255 ret = io_madvise(req, issue_flags);
6257 case IORING_OP_OPENAT2:
6258 ret = io_openat2(req, issue_flags);
6260 case IORING_OP_EPOLL_CTL:
6261 ret = io_epoll_ctl(req, issue_flags);
6263 case IORING_OP_SPLICE:
6264 ret = io_splice(req, issue_flags);
6266 case IORING_OP_PROVIDE_BUFFERS:
6267 ret = io_provide_buffers(req, issue_flags);
6269 case IORING_OP_REMOVE_BUFFERS:
6270 ret = io_remove_buffers(req, issue_flags);
6273 ret = io_tee(req, issue_flags);
6275 case IORING_OP_SHUTDOWN:
6276 ret = io_shutdown(req, issue_flags);
6278 case IORING_OP_RENAMEAT:
6279 ret = io_renameat(req, issue_flags);
6281 case IORING_OP_UNLINKAT:
6282 ret = io_unlinkat(req, issue_flags);
6290 revert_creds(creds);
6293 /* If the op doesn't have a file, we're not polling for it */
6294 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6295 io_iopoll_req_issued(req);
6300 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6302 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6304 req = io_put_req_find_next(req);
6305 return req ? &req->work : NULL;
6308 static void io_wq_submit_work(struct io_wq_work *work)
6310 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6311 struct io_kiocb *timeout;
6314 /* one will be dropped by ->io_free_work() after returning to io-wq */
6315 if (!(req->flags & REQ_F_REFCOUNT))
6316 __io_req_set_refcount(req, 2);
6320 timeout = io_prep_linked_timeout(req);
6322 io_queue_linked_timeout(timeout);
6324 if (work->flags & IO_WQ_WORK_CANCEL)
6329 ret = io_issue_sqe(req, 0);
6331 * We can get EAGAIN for polled IO even though we're
6332 * forcing a sync submission from here, since we can't
6333 * wait for request slots on the block side.
6341 /* avoid locking problems by failing it from a clean context */
6343 io_req_task_queue_fail(req, ret);
6346 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6349 return &table->files[i];
6352 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6355 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6357 return (struct file *) (slot->file_ptr & FFS_MASK);
6360 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6362 unsigned long file_ptr = (unsigned long) file;
6364 if (__io_file_supports_nowait(file, READ))
6365 file_ptr |= FFS_ASYNC_READ;
6366 if (__io_file_supports_nowait(file, WRITE))
6367 file_ptr |= FFS_ASYNC_WRITE;
6368 if (S_ISREG(file_inode(file)->i_mode))
6369 file_ptr |= FFS_ISREG;
6370 file_slot->file_ptr = file_ptr;
6373 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6374 struct io_kiocb *req, int fd)
6377 unsigned long file_ptr;
6379 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6381 fd = array_index_nospec(fd, ctx->nr_user_files);
6382 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6383 file = (struct file *) (file_ptr & FFS_MASK);
6384 file_ptr &= ~FFS_MASK;
6385 /* mask in overlapping REQ_F and FFS bits */
6386 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6387 io_req_set_rsrc_node(req);
6391 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6392 struct io_kiocb *req, int fd)
6394 struct file *file = fget(fd);
6396 trace_io_uring_file_get(ctx, fd);
6398 /* we don't allow fixed io_uring files */
6399 if (file && unlikely(file->f_op == &io_uring_fops))
6400 io_req_track_inflight(req);
6404 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6405 struct io_kiocb *req, int fd, bool fixed)
6408 return io_file_get_fixed(ctx, req, fd);
6410 return io_file_get_normal(ctx, req, fd);
6413 static void io_req_task_link_timeout(struct io_kiocb *req)
6415 struct io_kiocb *prev = req->timeout.prev;
6416 struct io_ring_ctx *ctx = req->ctx;
6420 ret = io_try_cancel_userdata(req, prev->user_data);
6423 io_cqring_fill_event(ctx, req->user_data, ret, 0);
6424 io_commit_cqring(ctx);
6425 spin_unlock(&ctx->completion_lock);
6426 io_cqring_ev_posted(ctx);
6431 io_req_complete_post(req, -ETIME, 0);
6435 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6437 struct io_timeout_data *data = container_of(timer,
6438 struct io_timeout_data, timer);
6439 struct io_kiocb *prev, *req = data->req;
6440 struct io_ring_ctx *ctx = req->ctx;
6441 unsigned long flags;
6443 spin_lock_irqsave(&ctx->timeout_lock, flags);
6444 prev = req->timeout.head;
6445 req->timeout.head = NULL;
6448 * We don't expect the list to be empty, that will only happen if we
6449 * race with the completion of the linked work.
6452 io_remove_next_linked(prev);
6453 if (!req_ref_inc_not_zero(prev))
6456 req->timeout.prev = prev;
6457 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6459 req->io_task_work.func = io_req_task_link_timeout;
6460 io_req_task_work_add(req);
6461 return HRTIMER_NORESTART;
6464 static void io_queue_linked_timeout(struct io_kiocb *req)
6466 struct io_ring_ctx *ctx = req->ctx;
6468 spin_lock_irq(&ctx->timeout_lock);
6470 * If the back reference is NULL, then our linked request finished
6471 * before we got a chance to setup the timer
6473 if (req->timeout.head) {
6474 struct io_timeout_data *data = req->async_data;
6476 data->timer.function = io_link_timeout_fn;
6477 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6480 spin_unlock_irq(&ctx->timeout_lock);
6481 /* drop submission reference */
6485 static void __io_queue_sqe(struct io_kiocb *req)
6486 __must_hold(&req->ctx->uring_lock)
6488 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6492 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6495 * We async punt it if the file wasn't marked NOWAIT, or if the file
6496 * doesn't support non-blocking read/write attempts
6499 if (req->flags & REQ_F_COMPLETE_INLINE) {
6500 struct io_ring_ctx *ctx = req->ctx;
6501 struct io_submit_state *state = &ctx->submit_state;
6503 state->compl_reqs[state->compl_nr++] = req;
6504 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6505 io_submit_flush_completions(ctx);
6507 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6508 switch (io_arm_poll_handler(req)) {
6509 case IO_APOLL_READY:
6511 case IO_APOLL_ABORTED:
6513 * Queued up for async execution, worker will release
6514 * submit reference when the iocb is actually submitted.
6516 io_queue_async_work(req);
6520 io_req_complete_failed(req, ret);
6523 io_queue_linked_timeout(linked_timeout);
6526 static inline void io_queue_sqe(struct io_kiocb *req)
6527 __must_hold(&req->ctx->uring_lock)
6529 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6532 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6533 __io_queue_sqe(req);
6535 int ret = io_req_prep_async(req);
6538 io_req_complete_failed(req, ret);
6540 io_queue_async_work(req);
6545 * Check SQE restrictions (opcode and flags).
6547 * Returns 'true' if SQE is allowed, 'false' otherwise.
6549 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6550 struct io_kiocb *req,
6551 unsigned int sqe_flags)
6553 if (likely(!ctx->restricted))
6556 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6559 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6560 ctx->restrictions.sqe_flags_required)
6563 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6564 ctx->restrictions.sqe_flags_required))
6570 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6571 const struct io_uring_sqe *sqe)
6572 __must_hold(&ctx->uring_lock)
6574 struct io_submit_state *state;
6575 unsigned int sqe_flags;
6576 int personality, ret = 0;
6578 /* req is partially pre-initialised, see io_preinit_req() */
6579 req->opcode = READ_ONCE(sqe->opcode);
6580 /* same numerical values with corresponding REQ_F_*, safe to copy */
6581 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6582 req->user_data = READ_ONCE(sqe->user_data);
6584 req->fixed_rsrc_refs = NULL;
6585 req->task = current;
6587 /* enforce forwards compatibility on users */
6588 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6590 if (unlikely(req->opcode >= IORING_OP_LAST))
6592 if (!io_check_restriction(ctx, req, sqe_flags))
6595 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6596 !io_op_defs[req->opcode].buffer_select)
6598 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6599 ctx->drain_active = true;
6601 personality = READ_ONCE(sqe->personality);
6603 req->creds = xa_load(&ctx->personalities, personality);
6606 get_cred(req->creds);
6607 req->flags |= REQ_F_CREDS;
6609 state = &ctx->submit_state;
6612 * Plug now if we have more than 1 IO left after this, and the target
6613 * is potentially a read/write to block based storage.
6615 if (!state->plug_started && state->ios_left > 1 &&
6616 io_op_defs[req->opcode].plug) {
6617 blk_start_plug(&state->plug);
6618 state->plug_started = true;
6621 if (io_op_defs[req->opcode].needs_file) {
6622 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6623 (sqe_flags & IOSQE_FIXED_FILE));
6624 if (unlikely(!req->file))
6632 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6633 const struct io_uring_sqe *sqe)
6634 __must_hold(&ctx->uring_lock)
6636 struct io_submit_link *link = &ctx->submit_state.link;
6639 ret = io_init_req(ctx, req, sqe);
6640 if (unlikely(ret)) {
6643 /* fail even hard links since we don't submit */
6644 req_set_fail(link->head);
6645 io_req_complete_failed(link->head, -ECANCELED);
6648 io_req_complete_failed(req, ret);
6652 ret = io_req_prep(req, sqe);
6656 /* don't need @sqe from now on */
6657 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6659 ctx->flags & IORING_SETUP_SQPOLL);
6662 * If we already have a head request, queue this one for async
6663 * submittal once the head completes. If we don't have a head but
6664 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6665 * submitted sync once the chain is complete. If none of those
6666 * conditions are true (normal request), then just queue it.
6669 struct io_kiocb *head = link->head;
6671 ret = io_req_prep_async(req);
6674 trace_io_uring_link(ctx, req, head);
6675 link->last->link = req;
6678 /* last request of a link, enqueue the link */
6679 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6684 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6696 * Batched submission is done, ensure local IO is flushed out.
6698 static void io_submit_state_end(struct io_submit_state *state,
6699 struct io_ring_ctx *ctx)
6701 if (state->link.head)
6702 io_queue_sqe(state->link.head);
6703 if (state->compl_nr)
6704 io_submit_flush_completions(ctx);
6705 if (state->plug_started)
6706 blk_finish_plug(&state->plug);
6710 * Start submission side cache.
6712 static void io_submit_state_start(struct io_submit_state *state,
6713 unsigned int max_ios)
6715 state->plug_started = false;
6716 state->ios_left = max_ios;
6717 /* set only head, no need to init link_last in advance */
6718 state->link.head = NULL;
6721 static void io_commit_sqring(struct io_ring_ctx *ctx)
6723 struct io_rings *rings = ctx->rings;
6726 * Ensure any loads from the SQEs are done at this point,
6727 * since once we write the new head, the application could
6728 * write new data to them.
6730 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6734 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6735 * that is mapped by userspace. This means that care needs to be taken to
6736 * ensure that reads are stable, as we cannot rely on userspace always
6737 * being a good citizen. If members of the sqe are validated and then later
6738 * used, it's important that those reads are done through READ_ONCE() to
6739 * prevent a re-load down the line.
6741 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6743 unsigned head, mask = ctx->sq_entries - 1;
6744 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6747 * The cached sq head (or cq tail) serves two purposes:
6749 * 1) allows us to batch the cost of updating the user visible
6751 * 2) allows the kernel side to track the head on its own, even
6752 * though the application is the one updating it.
6754 head = READ_ONCE(ctx->sq_array[sq_idx]);
6755 if (likely(head < ctx->sq_entries))
6756 return &ctx->sq_sqes[head];
6758 /* drop invalid entries */
6760 WRITE_ONCE(ctx->rings->sq_dropped,
6761 READ_ONCE(ctx->rings->sq_dropped) + 1);
6765 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6766 __must_hold(&ctx->uring_lock)
6768 struct io_uring_task *tctx;
6771 /* make sure SQ entry isn't read before tail */
6772 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6773 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6776 tctx = current->io_uring;
6777 tctx->cached_refs -= nr;
6778 if (unlikely(tctx->cached_refs < 0)) {
6779 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6781 percpu_counter_add(&tctx->inflight, refill);
6782 refcount_add(refill, ¤t->usage);
6783 tctx->cached_refs += refill;
6785 io_submit_state_start(&ctx->submit_state, nr);
6787 while (submitted < nr) {
6788 const struct io_uring_sqe *sqe;
6789 struct io_kiocb *req;
6791 req = io_alloc_req(ctx);
6792 if (unlikely(!req)) {
6794 submitted = -EAGAIN;
6797 sqe = io_get_sqe(ctx);
6798 if (unlikely(!sqe)) {
6799 kmem_cache_free(req_cachep, req);
6802 /* will complete beyond this point, count as submitted */
6804 if (io_submit_sqe(ctx, req, sqe))
6808 if (unlikely(submitted != nr)) {
6809 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6810 int unused = nr - ref_used;
6812 current->io_uring->cached_refs += unused;
6813 percpu_ref_put_many(&ctx->refs, unused);
6816 io_submit_state_end(&ctx->submit_state, ctx);
6817 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6818 io_commit_sqring(ctx);
6823 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6825 return READ_ONCE(sqd->state);
6828 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6830 /* Tell userspace we may need a wakeup call */
6831 spin_lock(&ctx->completion_lock);
6832 WRITE_ONCE(ctx->rings->sq_flags,
6833 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6834 spin_unlock(&ctx->completion_lock);
6837 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6839 spin_lock(&ctx->completion_lock);
6840 WRITE_ONCE(ctx->rings->sq_flags,
6841 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6842 spin_unlock(&ctx->completion_lock);
6845 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6847 unsigned int to_submit;
6850 to_submit = io_sqring_entries(ctx);
6851 /* if we're handling multiple rings, cap submit size for fairness */
6852 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6853 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6855 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6856 unsigned nr_events = 0;
6857 const struct cred *creds = NULL;
6859 if (ctx->sq_creds != current_cred())
6860 creds = override_creds(ctx->sq_creds);
6862 mutex_lock(&ctx->uring_lock);
6863 if (!list_empty(&ctx->iopoll_list))
6864 io_do_iopoll(ctx, &nr_events, 0);
6867 * Don't submit if refs are dying, good for io_uring_register(),
6868 * but also it is relied upon by io_ring_exit_work()
6870 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6871 !(ctx->flags & IORING_SETUP_R_DISABLED))
6872 ret = io_submit_sqes(ctx, to_submit);
6873 mutex_unlock(&ctx->uring_lock);
6875 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6876 wake_up(&ctx->sqo_sq_wait);
6878 revert_creds(creds);
6884 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6886 struct io_ring_ctx *ctx;
6887 unsigned sq_thread_idle = 0;
6889 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6890 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6891 sqd->sq_thread_idle = sq_thread_idle;
6894 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6896 bool did_sig = false;
6897 struct ksignal ksig;
6899 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6900 signal_pending(current)) {
6901 mutex_unlock(&sqd->lock);
6902 if (signal_pending(current))
6903 did_sig = get_signal(&ksig);
6905 mutex_lock(&sqd->lock);
6907 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6910 static int io_sq_thread(void *data)
6912 struct io_sq_data *sqd = data;
6913 struct io_ring_ctx *ctx;
6914 unsigned long timeout = 0;
6915 char buf[TASK_COMM_LEN];
6918 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6919 set_task_comm(current, buf);
6921 if (sqd->sq_cpu != -1)
6922 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6924 set_cpus_allowed_ptr(current, cpu_online_mask);
6925 current->flags |= PF_NO_SETAFFINITY;
6927 mutex_lock(&sqd->lock);
6929 bool cap_entries, sqt_spin = false;
6931 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6932 if (io_sqd_handle_event(sqd))
6934 timeout = jiffies + sqd->sq_thread_idle;
6937 cap_entries = !list_is_singular(&sqd->ctx_list);
6938 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6939 int ret = __io_sq_thread(ctx, cap_entries);
6941 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6944 if (io_run_task_work())
6947 if (sqt_spin || !time_after(jiffies, timeout)) {
6950 timeout = jiffies + sqd->sq_thread_idle;
6954 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6955 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6956 bool needs_sched = true;
6958 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6959 io_ring_set_wakeup_flag(ctx);
6961 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6962 !list_empty_careful(&ctx->iopoll_list)) {
6963 needs_sched = false;
6966 if (io_sqring_entries(ctx)) {
6967 needs_sched = false;
6973 mutex_unlock(&sqd->lock);
6975 mutex_lock(&sqd->lock);
6977 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6978 io_ring_clear_wakeup_flag(ctx);
6981 finish_wait(&sqd->wait, &wait);
6982 timeout = jiffies + sqd->sq_thread_idle;
6985 io_uring_cancel_generic(true, sqd);
6987 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6988 io_ring_set_wakeup_flag(ctx);
6990 mutex_unlock(&sqd->lock);
6992 complete(&sqd->exited);
6996 struct io_wait_queue {
6997 struct wait_queue_entry wq;
6998 struct io_ring_ctx *ctx;
7000 unsigned nr_timeouts;
7003 static inline bool io_should_wake(struct io_wait_queue *iowq)
7005 struct io_ring_ctx *ctx = iowq->ctx;
7006 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7009 * Wake up if we have enough events, or if a timeout occurred since we
7010 * started waiting. For timeouts, we always want to return to userspace,
7011 * regardless of event count.
7013 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7016 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7017 int wake_flags, void *key)
7019 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7023 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7024 * the task, and the next invocation will do it.
7026 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7027 return autoremove_wake_function(curr, mode, wake_flags, key);
7031 static int io_run_task_work_sig(void)
7033 if (io_run_task_work())
7035 if (!signal_pending(current))
7037 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7038 return -ERESTARTSYS;
7042 /* when returns >0, the caller should retry */
7043 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7044 struct io_wait_queue *iowq,
7045 signed long *timeout)
7049 /* make sure we run task_work before checking for signals */
7050 ret = io_run_task_work_sig();
7051 if (ret || io_should_wake(iowq))
7053 /* let the caller flush overflows, retry */
7054 if (test_bit(0, &ctx->check_cq_overflow))
7057 *timeout = schedule_timeout(*timeout);
7058 return !*timeout ? -ETIME : 1;
7062 * Wait until events become available, if we don't already have some. The
7063 * application must reap them itself, as they reside on the shared cq ring.
7065 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7066 const sigset_t __user *sig, size_t sigsz,
7067 struct __kernel_timespec __user *uts)
7069 struct io_wait_queue iowq;
7070 struct io_rings *rings = ctx->rings;
7071 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7075 io_cqring_overflow_flush(ctx);
7076 if (io_cqring_events(ctx) >= min_events)
7078 if (!io_run_task_work())
7083 #ifdef CONFIG_COMPAT
7084 if (in_compat_syscall())
7085 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7089 ret = set_user_sigmask(sig, sigsz);
7096 struct timespec64 ts;
7098 if (get_timespec64(&ts, uts))
7100 timeout = timespec64_to_jiffies(&ts);
7103 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7104 iowq.wq.private = current;
7105 INIT_LIST_HEAD(&iowq.wq.entry);
7107 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7108 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7110 trace_io_uring_cqring_wait(ctx, min_events);
7112 /* if we can't even flush overflow, don't wait for more */
7113 if (!io_cqring_overflow_flush(ctx)) {
7117 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7118 TASK_INTERRUPTIBLE);
7119 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7120 finish_wait(&ctx->cq_wait, &iowq.wq);
7124 restore_saved_sigmask_unless(ret == -EINTR);
7126 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7129 static void io_free_page_table(void **table, size_t size)
7131 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7133 for (i = 0; i < nr_tables; i++)
7138 static void **io_alloc_page_table(size_t size)
7140 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7141 size_t init_size = size;
7144 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7148 for (i = 0; i < nr_tables; i++) {
7149 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7151 table[i] = kzalloc(this_size, GFP_KERNEL);
7153 io_free_page_table(table, init_size);
7161 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7163 percpu_ref_exit(&ref_node->refs);
7167 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7169 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7170 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7171 unsigned long flags;
7172 bool first_add = false;
7174 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7177 while (!list_empty(&ctx->rsrc_ref_list)) {
7178 node = list_first_entry(&ctx->rsrc_ref_list,
7179 struct io_rsrc_node, node);
7180 /* recycle ref nodes in order */
7183 list_del(&node->node);
7184 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7186 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7189 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7192 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7194 struct io_rsrc_node *ref_node;
7196 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7200 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7205 INIT_LIST_HEAD(&ref_node->node);
7206 INIT_LIST_HEAD(&ref_node->rsrc_list);
7207 ref_node->done = false;
7211 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7212 struct io_rsrc_data *data_to_kill)
7214 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7215 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7218 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7220 rsrc_node->rsrc_data = data_to_kill;
7221 spin_lock_irq(&ctx->rsrc_ref_lock);
7222 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7223 spin_unlock_irq(&ctx->rsrc_ref_lock);
7225 atomic_inc(&data_to_kill->refs);
7226 percpu_ref_kill(&rsrc_node->refs);
7227 ctx->rsrc_node = NULL;
7230 if (!ctx->rsrc_node) {
7231 ctx->rsrc_node = ctx->rsrc_backup_node;
7232 ctx->rsrc_backup_node = NULL;
7236 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7238 if (ctx->rsrc_backup_node)
7240 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7241 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7244 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7248 /* As we may drop ->uring_lock, other task may have started quiesce */
7252 data->quiesce = true;
7254 ret = io_rsrc_node_switch_start(ctx);
7257 io_rsrc_node_switch(ctx, data);
7259 /* kill initial ref, already quiesced if zero */
7260 if (atomic_dec_and_test(&data->refs))
7262 mutex_unlock(&ctx->uring_lock);
7263 flush_delayed_work(&ctx->rsrc_put_work);
7264 ret = wait_for_completion_interruptible(&data->done);
7266 mutex_lock(&ctx->uring_lock);
7270 atomic_inc(&data->refs);
7271 /* wait for all works potentially completing data->done */
7272 flush_delayed_work(&ctx->rsrc_put_work);
7273 reinit_completion(&data->done);
7275 ret = io_run_task_work_sig();
7276 mutex_lock(&ctx->uring_lock);
7278 data->quiesce = false;
7283 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7285 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7286 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7288 return &data->tags[table_idx][off];
7291 static void io_rsrc_data_free(struct io_rsrc_data *data)
7293 size_t size = data->nr * sizeof(data->tags[0][0]);
7296 io_free_page_table((void **)data->tags, size);
7300 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7301 u64 __user *utags, unsigned nr,
7302 struct io_rsrc_data **pdata)
7304 struct io_rsrc_data *data;
7308 data = kzalloc(sizeof(*data), GFP_KERNEL);
7311 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7319 data->do_put = do_put;
7322 for (i = 0; i < nr; i++) {
7323 u64 *tag_slot = io_get_tag_slot(data, i);
7325 if (copy_from_user(tag_slot, &utags[i],
7331 atomic_set(&data->refs, 1);
7332 init_completion(&data->done);
7336 io_rsrc_data_free(data);
7340 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7342 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7343 return !!table->files;
7346 static void io_free_file_tables(struct io_file_table *table)
7348 kvfree(table->files);
7349 table->files = NULL;
7352 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7354 #if defined(CONFIG_UNIX)
7355 if (ctx->ring_sock) {
7356 struct sock *sock = ctx->ring_sock->sk;
7357 struct sk_buff *skb;
7359 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7365 for (i = 0; i < ctx->nr_user_files; i++) {
7368 file = io_file_from_index(ctx, i);
7373 io_free_file_tables(&ctx->file_table);
7374 io_rsrc_data_free(ctx->file_data);
7375 ctx->file_data = NULL;
7376 ctx->nr_user_files = 0;
7379 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7383 if (!ctx->file_data)
7385 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7387 __io_sqe_files_unregister(ctx);
7391 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7392 __releases(&sqd->lock)
7394 WARN_ON_ONCE(sqd->thread == current);
7397 * Do the dance but not conditional clear_bit() because it'd race with
7398 * other threads incrementing park_pending and setting the bit.
7400 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7401 if (atomic_dec_return(&sqd->park_pending))
7402 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7403 mutex_unlock(&sqd->lock);
7406 static void io_sq_thread_park(struct io_sq_data *sqd)
7407 __acquires(&sqd->lock)
7409 WARN_ON_ONCE(sqd->thread == current);
7411 atomic_inc(&sqd->park_pending);
7412 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7413 mutex_lock(&sqd->lock);
7415 wake_up_process(sqd->thread);
7418 static void io_sq_thread_stop(struct io_sq_data *sqd)
7420 WARN_ON_ONCE(sqd->thread == current);
7421 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7423 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7424 mutex_lock(&sqd->lock);
7426 wake_up_process(sqd->thread);
7427 mutex_unlock(&sqd->lock);
7428 wait_for_completion(&sqd->exited);
7431 static void io_put_sq_data(struct io_sq_data *sqd)
7433 if (refcount_dec_and_test(&sqd->refs)) {
7434 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7436 io_sq_thread_stop(sqd);
7441 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7443 struct io_sq_data *sqd = ctx->sq_data;
7446 io_sq_thread_park(sqd);
7447 list_del_init(&ctx->sqd_list);
7448 io_sqd_update_thread_idle(sqd);
7449 io_sq_thread_unpark(sqd);
7451 io_put_sq_data(sqd);
7452 ctx->sq_data = NULL;
7456 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7458 struct io_ring_ctx *ctx_attach;
7459 struct io_sq_data *sqd;
7462 f = fdget(p->wq_fd);
7464 return ERR_PTR(-ENXIO);
7465 if (f.file->f_op != &io_uring_fops) {
7467 return ERR_PTR(-EINVAL);
7470 ctx_attach = f.file->private_data;
7471 sqd = ctx_attach->sq_data;
7474 return ERR_PTR(-EINVAL);
7476 if (sqd->task_tgid != current->tgid) {
7478 return ERR_PTR(-EPERM);
7481 refcount_inc(&sqd->refs);
7486 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7489 struct io_sq_data *sqd;
7492 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7493 sqd = io_attach_sq_data(p);
7498 /* fall through for EPERM case, setup new sqd/task */
7499 if (PTR_ERR(sqd) != -EPERM)
7503 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7505 return ERR_PTR(-ENOMEM);
7507 atomic_set(&sqd->park_pending, 0);
7508 refcount_set(&sqd->refs, 1);
7509 INIT_LIST_HEAD(&sqd->ctx_list);
7510 mutex_init(&sqd->lock);
7511 init_waitqueue_head(&sqd->wait);
7512 init_completion(&sqd->exited);
7516 #if defined(CONFIG_UNIX)
7518 * Ensure the UNIX gc is aware of our file set, so we are certain that
7519 * the io_uring can be safely unregistered on process exit, even if we have
7520 * loops in the file referencing.
7522 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7524 struct sock *sk = ctx->ring_sock->sk;
7525 struct scm_fp_list *fpl;
7526 struct sk_buff *skb;
7529 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7533 skb = alloc_skb(0, GFP_KERNEL);
7542 fpl->user = get_uid(current_user());
7543 for (i = 0; i < nr; i++) {
7544 struct file *file = io_file_from_index(ctx, i + offset);
7548 fpl->fp[nr_files] = get_file(file);
7549 unix_inflight(fpl->user, fpl->fp[nr_files]);
7554 fpl->max = SCM_MAX_FD;
7555 fpl->count = nr_files;
7556 UNIXCB(skb).fp = fpl;
7557 skb->destructor = unix_destruct_scm;
7558 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7559 skb_queue_head(&sk->sk_receive_queue, skb);
7561 for (i = 0; i < nr_files; i++)
7572 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7573 * causes regular reference counting to break down. We rely on the UNIX
7574 * garbage collection to take care of this problem for us.
7576 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7578 unsigned left, total;
7582 left = ctx->nr_user_files;
7584 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7586 ret = __io_sqe_files_scm(ctx, this_files, total);
7590 total += this_files;
7596 while (total < ctx->nr_user_files) {
7597 struct file *file = io_file_from_index(ctx, total);
7607 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7613 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7615 struct file *file = prsrc->file;
7616 #if defined(CONFIG_UNIX)
7617 struct sock *sock = ctx->ring_sock->sk;
7618 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7619 struct sk_buff *skb;
7622 __skb_queue_head_init(&list);
7625 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7626 * remove this entry and rearrange the file array.
7628 skb = skb_dequeue(head);
7630 struct scm_fp_list *fp;
7632 fp = UNIXCB(skb).fp;
7633 for (i = 0; i < fp->count; i++) {
7636 if (fp->fp[i] != file)
7639 unix_notinflight(fp->user, fp->fp[i]);
7640 left = fp->count - 1 - i;
7642 memmove(&fp->fp[i], &fp->fp[i + 1],
7643 left * sizeof(struct file *));
7650 __skb_queue_tail(&list, skb);
7660 __skb_queue_tail(&list, skb);
7662 skb = skb_dequeue(head);
7665 if (skb_peek(&list)) {
7666 spin_lock_irq(&head->lock);
7667 while ((skb = __skb_dequeue(&list)) != NULL)
7668 __skb_queue_tail(head, skb);
7669 spin_unlock_irq(&head->lock);
7676 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7678 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7679 struct io_ring_ctx *ctx = rsrc_data->ctx;
7680 struct io_rsrc_put *prsrc, *tmp;
7682 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7683 list_del(&prsrc->list);
7686 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7688 io_ring_submit_lock(ctx, lock_ring);
7689 spin_lock(&ctx->completion_lock);
7690 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7692 io_commit_cqring(ctx);
7693 spin_unlock(&ctx->completion_lock);
7694 io_cqring_ev_posted(ctx);
7695 io_ring_submit_unlock(ctx, lock_ring);
7698 rsrc_data->do_put(ctx, prsrc);
7702 io_rsrc_node_destroy(ref_node);
7703 if (atomic_dec_and_test(&rsrc_data->refs))
7704 complete(&rsrc_data->done);
7707 static void io_rsrc_put_work(struct work_struct *work)
7709 struct io_ring_ctx *ctx;
7710 struct llist_node *node;
7712 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7713 node = llist_del_all(&ctx->rsrc_put_llist);
7716 struct io_rsrc_node *ref_node;
7717 struct llist_node *next = node->next;
7719 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7720 __io_rsrc_put_work(ref_node);
7725 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7726 unsigned nr_args, u64 __user *tags)
7728 __s32 __user *fds = (__s32 __user *) arg;
7737 if (nr_args > IORING_MAX_FIXED_FILES)
7739 ret = io_rsrc_node_switch_start(ctx);
7742 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7748 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7751 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7752 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7756 /* allow sparse sets */
7759 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7766 if (unlikely(!file))
7770 * Don't allow io_uring instances to be registered. If UNIX
7771 * isn't enabled, then this causes a reference cycle and this
7772 * instance can never get freed. If UNIX is enabled we'll
7773 * handle it just fine, but there's still no point in allowing
7774 * a ring fd as it doesn't support regular read/write anyway.
7776 if (file->f_op == &io_uring_fops) {
7780 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7783 ret = io_sqe_files_scm(ctx);
7785 __io_sqe_files_unregister(ctx);
7789 io_rsrc_node_switch(ctx, NULL);
7792 for (i = 0; i < ctx->nr_user_files; i++) {
7793 file = io_file_from_index(ctx, i);
7797 io_free_file_tables(&ctx->file_table);
7798 ctx->nr_user_files = 0;
7800 io_rsrc_data_free(ctx->file_data);
7801 ctx->file_data = NULL;
7805 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7808 #if defined(CONFIG_UNIX)
7809 struct sock *sock = ctx->ring_sock->sk;
7810 struct sk_buff_head *head = &sock->sk_receive_queue;
7811 struct sk_buff *skb;
7814 * See if we can merge this file into an existing skb SCM_RIGHTS
7815 * file set. If there's no room, fall back to allocating a new skb
7816 * and filling it in.
7818 spin_lock_irq(&head->lock);
7819 skb = skb_peek(head);
7821 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7823 if (fpl->count < SCM_MAX_FD) {
7824 __skb_unlink(skb, head);
7825 spin_unlock_irq(&head->lock);
7826 fpl->fp[fpl->count] = get_file(file);
7827 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7829 spin_lock_irq(&head->lock);
7830 __skb_queue_head(head, skb);
7835 spin_unlock_irq(&head->lock);
7842 return __io_sqe_files_scm(ctx, 1, index);
7848 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7849 struct io_rsrc_node *node, void *rsrc)
7851 struct io_rsrc_put *prsrc;
7853 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7857 prsrc->tag = *io_get_tag_slot(data, idx);
7859 list_add(&prsrc->list, &node->rsrc_list);
7863 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7864 struct io_uring_rsrc_update2 *up,
7867 u64 __user *tags = u64_to_user_ptr(up->tags);
7868 __s32 __user *fds = u64_to_user_ptr(up->data);
7869 struct io_rsrc_data *data = ctx->file_data;
7870 struct io_fixed_file *file_slot;
7874 bool needs_switch = false;
7876 if (!ctx->file_data)
7878 if (up->offset + nr_args > ctx->nr_user_files)
7881 for (done = 0; done < nr_args; done++) {
7884 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7885 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7889 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7893 if (fd == IORING_REGISTER_FILES_SKIP)
7896 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7897 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7899 if (file_slot->file_ptr) {
7900 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7901 err = io_queue_rsrc_removal(data, up->offset + done,
7902 ctx->rsrc_node, file);
7905 file_slot->file_ptr = 0;
7906 needs_switch = true;
7915 * Don't allow io_uring instances to be registered. If
7916 * UNIX isn't enabled, then this causes a reference
7917 * cycle and this instance can never get freed. If UNIX
7918 * is enabled we'll handle it just fine, but there's
7919 * still no point in allowing a ring fd as it doesn't
7920 * support regular read/write anyway.
7922 if (file->f_op == &io_uring_fops) {
7927 *io_get_tag_slot(data, up->offset + done) = tag;
7928 io_fixed_file_set(file_slot, file);
7929 err = io_sqe_file_register(ctx, file, i);
7931 file_slot->file_ptr = 0;
7939 io_rsrc_node_switch(ctx, data);
7940 return done ? done : err;
7943 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7944 struct task_struct *task)
7946 struct io_wq_hash *hash;
7947 struct io_wq_data data;
7948 unsigned int concurrency;
7950 mutex_lock(&ctx->uring_lock);
7951 hash = ctx->hash_map;
7953 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7955 mutex_unlock(&ctx->uring_lock);
7956 return ERR_PTR(-ENOMEM);
7958 refcount_set(&hash->refs, 1);
7959 init_waitqueue_head(&hash->wait);
7960 ctx->hash_map = hash;
7962 mutex_unlock(&ctx->uring_lock);
7966 data.free_work = io_wq_free_work;
7967 data.do_work = io_wq_submit_work;
7969 /* Do QD, or 4 * CPUS, whatever is smallest */
7970 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7972 return io_wq_create(concurrency, &data);
7975 static int io_uring_alloc_task_context(struct task_struct *task,
7976 struct io_ring_ctx *ctx)
7978 struct io_uring_task *tctx;
7981 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7982 if (unlikely(!tctx))
7985 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7986 if (unlikely(ret)) {
7991 tctx->io_wq = io_init_wq_offload(ctx, task);
7992 if (IS_ERR(tctx->io_wq)) {
7993 ret = PTR_ERR(tctx->io_wq);
7994 percpu_counter_destroy(&tctx->inflight);
8000 init_waitqueue_head(&tctx->wait);
8001 atomic_set(&tctx->in_idle, 0);
8002 atomic_set(&tctx->inflight_tracked, 0);
8003 task->io_uring = tctx;
8004 spin_lock_init(&tctx->task_lock);
8005 INIT_WQ_LIST(&tctx->task_list);
8006 init_task_work(&tctx->task_work, tctx_task_work);
8010 void __io_uring_free(struct task_struct *tsk)
8012 struct io_uring_task *tctx = tsk->io_uring;
8014 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8015 WARN_ON_ONCE(tctx->io_wq);
8016 WARN_ON_ONCE(tctx->cached_refs);
8018 percpu_counter_destroy(&tctx->inflight);
8020 tsk->io_uring = NULL;
8023 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8024 struct io_uring_params *p)
8028 /* Retain compatibility with failing for an invalid attach attempt */
8029 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8030 IORING_SETUP_ATTACH_WQ) {
8033 f = fdget(p->wq_fd);
8036 if (f.file->f_op != &io_uring_fops) {
8042 if (ctx->flags & IORING_SETUP_SQPOLL) {
8043 struct task_struct *tsk;
8044 struct io_sq_data *sqd;
8047 sqd = io_get_sq_data(p, &attached);
8053 ctx->sq_creds = get_current_cred();
8055 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8056 if (!ctx->sq_thread_idle)
8057 ctx->sq_thread_idle = HZ;
8059 io_sq_thread_park(sqd);
8060 list_add(&ctx->sqd_list, &sqd->ctx_list);
8061 io_sqd_update_thread_idle(sqd);
8062 /* don't attach to a dying SQPOLL thread, would be racy */
8063 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8064 io_sq_thread_unpark(sqd);
8071 if (p->flags & IORING_SETUP_SQ_AFF) {
8072 int cpu = p->sq_thread_cpu;
8075 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8082 sqd->task_pid = current->pid;
8083 sqd->task_tgid = current->tgid;
8084 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8091 ret = io_uring_alloc_task_context(tsk, ctx);
8092 wake_up_new_task(tsk);
8095 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8096 /* Can't have SQ_AFF without SQPOLL */
8103 complete(&ctx->sq_data->exited);
8105 io_sq_thread_finish(ctx);
8109 static inline void __io_unaccount_mem(struct user_struct *user,
8110 unsigned long nr_pages)
8112 atomic_long_sub(nr_pages, &user->locked_vm);
8115 static inline int __io_account_mem(struct user_struct *user,
8116 unsigned long nr_pages)
8118 unsigned long page_limit, cur_pages, new_pages;
8120 /* Don't allow more pages than we can safely lock */
8121 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8124 cur_pages = atomic_long_read(&user->locked_vm);
8125 new_pages = cur_pages + nr_pages;
8126 if (new_pages > page_limit)
8128 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8129 new_pages) != cur_pages);
8134 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8137 __io_unaccount_mem(ctx->user, nr_pages);
8139 if (ctx->mm_account)
8140 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8143 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8148 ret = __io_account_mem(ctx->user, nr_pages);
8153 if (ctx->mm_account)
8154 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8159 static void io_mem_free(void *ptr)
8166 page = virt_to_head_page(ptr);
8167 if (put_page_testzero(page))
8168 free_compound_page(page);
8171 static void *io_mem_alloc(size_t size)
8173 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8174 __GFP_NORETRY | __GFP_ACCOUNT;
8176 return (void *) __get_free_pages(gfp_flags, get_order(size));
8179 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8182 struct io_rings *rings;
8183 size_t off, sq_array_size;
8185 off = struct_size(rings, cqes, cq_entries);
8186 if (off == SIZE_MAX)
8190 off = ALIGN(off, SMP_CACHE_BYTES);
8198 sq_array_size = array_size(sizeof(u32), sq_entries);
8199 if (sq_array_size == SIZE_MAX)
8202 if (check_add_overflow(off, sq_array_size, &off))
8208 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8210 struct io_mapped_ubuf *imu = *slot;
8213 if (imu != ctx->dummy_ubuf) {
8214 for (i = 0; i < imu->nr_bvecs; i++)
8215 unpin_user_page(imu->bvec[i].bv_page);
8216 if (imu->acct_pages)
8217 io_unaccount_mem(ctx, imu->acct_pages);
8223 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8225 io_buffer_unmap(ctx, &prsrc->buf);
8229 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8233 for (i = 0; i < ctx->nr_user_bufs; i++)
8234 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8235 kfree(ctx->user_bufs);
8236 io_rsrc_data_free(ctx->buf_data);
8237 ctx->user_bufs = NULL;
8238 ctx->buf_data = NULL;
8239 ctx->nr_user_bufs = 0;
8242 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8249 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8251 __io_sqe_buffers_unregister(ctx);
8255 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8256 void __user *arg, unsigned index)
8258 struct iovec __user *src;
8260 #ifdef CONFIG_COMPAT
8262 struct compat_iovec __user *ciovs;
8263 struct compat_iovec ciov;
8265 ciovs = (struct compat_iovec __user *) arg;
8266 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8269 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8270 dst->iov_len = ciov.iov_len;
8274 src = (struct iovec __user *) arg;
8275 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8281 * Not super efficient, but this is just a registration time. And we do cache
8282 * the last compound head, so generally we'll only do a full search if we don't
8285 * We check if the given compound head page has already been accounted, to
8286 * avoid double accounting it. This allows us to account the full size of the
8287 * page, not just the constituent pages of a huge page.
8289 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8290 int nr_pages, struct page *hpage)
8294 /* check current page array */
8295 for (i = 0; i < nr_pages; i++) {
8296 if (!PageCompound(pages[i]))
8298 if (compound_head(pages[i]) == hpage)
8302 /* check previously registered pages */
8303 for (i = 0; i < ctx->nr_user_bufs; i++) {
8304 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8306 for (j = 0; j < imu->nr_bvecs; j++) {
8307 if (!PageCompound(imu->bvec[j].bv_page))
8309 if (compound_head(imu->bvec[j].bv_page) == hpage)
8317 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8318 int nr_pages, struct io_mapped_ubuf *imu,
8319 struct page **last_hpage)
8323 imu->acct_pages = 0;
8324 for (i = 0; i < nr_pages; i++) {
8325 if (!PageCompound(pages[i])) {
8330 hpage = compound_head(pages[i]);
8331 if (hpage == *last_hpage)
8333 *last_hpage = hpage;
8334 if (headpage_already_acct(ctx, pages, i, hpage))
8336 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8340 if (!imu->acct_pages)
8343 ret = io_account_mem(ctx, imu->acct_pages);
8345 imu->acct_pages = 0;
8349 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8350 struct io_mapped_ubuf **pimu,
8351 struct page **last_hpage)
8353 struct io_mapped_ubuf *imu = NULL;
8354 struct vm_area_struct **vmas = NULL;
8355 struct page **pages = NULL;
8356 unsigned long off, start, end, ubuf;
8358 int ret, pret, nr_pages, i;
8360 if (!iov->iov_base) {
8361 *pimu = ctx->dummy_ubuf;
8365 ubuf = (unsigned long) iov->iov_base;
8366 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8367 start = ubuf >> PAGE_SHIFT;
8368 nr_pages = end - start;
8373 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8377 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8382 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8387 mmap_read_lock(current->mm);
8388 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8390 if (pret == nr_pages) {
8391 /* don't support file backed memory */
8392 for (i = 0; i < nr_pages; i++) {
8393 struct vm_area_struct *vma = vmas[i];
8395 if (vma_is_shmem(vma))
8398 !is_file_hugepages(vma->vm_file)) {
8404 ret = pret < 0 ? pret : -EFAULT;
8406 mmap_read_unlock(current->mm);
8409 * if we did partial map, or found file backed vmas,
8410 * release any pages we did get
8413 unpin_user_pages(pages, pret);
8417 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8419 unpin_user_pages(pages, pret);
8423 off = ubuf & ~PAGE_MASK;
8424 size = iov->iov_len;
8425 for (i = 0; i < nr_pages; i++) {
8428 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8429 imu->bvec[i].bv_page = pages[i];
8430 imu->bvec[i].bv_len = vec_len;
8431 imu->bvec[i].bv_offset = off;
8435 /* store original address for later verification */
8437 imu->ubuf_end = ubuf + iov->iov_len;
8438 imu->nr_bvecs = nr_pages;
8449 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8451 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8452 return ctx->user_bufs ? 0 : -ENOMEM;
8455 static int io_buffer_validate(struct iovec *iov)
8457 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8460 * Don't impose further limits on the size and buffer
8461 * constraints here, we'll -EINVAL later when IO is
8462 * submitted if they are wrong.
8465 return iov->iov_len ? -EFAULT : 0;
8469 /* arbitrary limit, but we need something */
8470 if (iov->iov_len > SZ_1G)
8473 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8479 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8480 unsigned int nr_args, u64 __user *tags)
8482 struct page *last_hpage = NULL;
8483 struct io_rsrc_data *data;
8489 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8491 ret = io_rsrc_node_switch_start(ctx);
8494 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8497 ret = io_buffers_map_alloc(ctx, nr_args);
8499 io_rsrc_data_free(data);
8503 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8504 ret = io_copy_iov(ctx, &iov, arg, i);
8507 ret = io_buffer_validate(&iov);
8510 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8515 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8521 WARN_ON_ONCE(ctx->buf_data);
8523 ctx->buf_data = data;
8525 __io_sqe_buffers_unregister(ctx);
8527 io_rsrc_node_switch(ctx, NULL);
8531 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8532 struct io_uring_rsrc_update2 *up,
8533 unsigned int nr_args)
8535 u64 __user *tags = u64_to_user_ptr(up->tags);
8536 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8537 struct page *last_hpage = NULL;
8538 bool needs_switch = false;
8544 if (up->offset + nr_args > ctx->nr_user_bufs)
8547 for (done = 0; done < nr_args; done++) {
8548 struct io_mapped_ubuf *imu;
8549 int offset = up->offset + done;
8552 err = io_copy_iov(ctx, &iov, iovs, done);
8555 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8559 err = io_buffer_validate(&iov);
8562 if (!iov.iov_base && tag) {
8566 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8570 i = array_index_nospec(offset, ctx->nr_user_bufs);
8571 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8572 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8573 ctx->rsrc_node, ctx->user_bufs[i]);
8574 if (unlikely(err)) {
8575 io_buffer_unmap(ctx, &imu);
8578 ctx->user_bufs[i] = NULL;
8579 needs_switch = true;
8582 ctx->user_bufs[i] = imu;
8583 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8587 io_rsrc_node_switch(ctx, ctx->buf_data);
8588 return done ? done : err;
8591 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8593 __s32 __user *fds = arg;
8599 if (copy_from_user(&fd, fds, sizeof(*fds)))
8602 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8603 if (IS_ERR(ctx->cq_ev_fd)) {
8604 int ret = PTR_ERR(ctx->cq_ev_fd);
8606 ctx->cq_ev_fd = NULL;
8613 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8615 if (ctx->cq_ev_fd) {
8616 eventfd_ctx_put(ctx->cq_ev_fd);
8617 ctx->cq_ev_fd = NULL;
8624 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8626 struct io_buffer *buf;
8627 unsigned long index;
8629 xa_for_each(&ctx->io_buffers, index, buf)
8630 __io_remove_buffers(ctx, buf, index, -1U);
8633 static void io_req_cache_free(struct list_head *list)
8635 struct io_kiocb *req, *nxt;
8637 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8638 list_del(&req->inflight_entry);
8639 kmem_cache_free(req_cachep, req);
8643 static void io_req_caches_free(struct io_ring_ctx *ctx)
8645 struct io_submit_state *state = &ctx->submit_state;
8647 mutex_lock(&ctx->uring_lock);
8649 if (state->free_reqs) {
8650 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8651 state->free_reqs = 0;
8654 io_flush_cached_locked_reqs(ctx, state);
8655 io_req_cache_free(&state->free_list);
8656 mutex_unlock(&ctx->uring_lock);
8659 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8661 if (data && !atomic_dec_and_test(&data->refs))
8662 wait_for_completion(&data->done);
8665 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8667 io_sq_thread_finish(ctx);
8669 if (ctx->mm_account) {
8670 mmdrop(ctx->mm_account);
8671 ctx->mm_account = NULL;
8674 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8675 io_wait_rsrc_data(ctx->buf_data);
8676 io_wait_rsrc_data(ctx->file_data);
8678 mutex_lock(&ctx->uring_lock);
8680 __io_sqe_buffers_unregister(ctx);
8682 __io_sqe_files_unregister(ctx);
8684 __io_cqring_overflow_flush(ctx, true);
8685 mutex_unlock(&ctx->uring_lock);
8686 io_eventfd_unregister(ctx);
8687 io_destroy_buffers(ctx);
8689 put_cred(ctx->sq_creds);
8691 /* there are no registered resources left, nobody uses it */
8693 io_rsrc_node_destroy(ctx->rsrc_node);
8694 if (ctx->rsrc_backup_node)
8695 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8696 flush_delayed_work(&ctx->rsrc_put_work);
8698 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8699 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8701 #if defined(CONFIG_UNIX)
8702 if (ctx->ring_sock) {
8703 ctx->ring_sock->file = NULL; /* so that iput() is called */
8704 sock_release(ctx->ring_sock);
8708 io_mem_free(ctx->rings);
8709 io_mem_free(ctx->sq_sqes);
8711 percpu_ref_exit(&ctx->refs);
8712 free_uid(ctx->user);
8713 io_req_caches_free(ctx);
8715 io_wq_put_hash(ctx->hash_map);
8716 kfree(ctx->cancel_hash);
8717 kfree(ctx->dummy_ubuf);
8721 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8723 struct io_ring_ctx *ctx = file->private_data;
8726 poll_wait(file, &ctx->poll_wait, wait);
8728 * synchronizes with barrier from wq_has_sleeper call in
8732 if (!io_sqring_full(ctx))
8733 mask |= EPOLLOUT | EPOLLWRNORM;
8736 * Don't flush cqring overflow list here, just do a simple check.
8737 * Otherwise there could possible be ABBA deadlock:
8740 * lock(&ctx->uring_lock);
8742 * lock(&ctx->uring_lock);
8745 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8746 * pushs them to do the flush.
8748 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8749 mask |= EPOLLIN | EPOLLRDNORM;
8754 static int io_uring_fasync(int fd, struct file *file, int on)
8756 struct io_ring_ctx *ctx = file->private_data;
8758 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8761 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8763 const struct cred *creds;
8765 creds = xa_erase(&ctx->personalities, id);
8774 struct io_tctx_exit {
8775 struct callback_head task_work;
8776 struct completion completion;
8777 struct io_ring_ctx *ctx;
8780 static void io_tctx_exit_cb(struct callback_head *cb)
8782 struct io_uring_task *tctx = current->io_uring;
8783 struct io_tctx_exit *work;
8785 work = container_of(cb, struct io_tctx_exit, task_work);
8787 * When @in_idle, we're in cancellation and it's racy to remove the
8788 * node. It'll be removed by the end of cancellation, just ignore it.
8790 if (!atomic_read(&tctx->in_idle))
8791 io_uring_del_tctx_node((unsigned long)work->ctx);
8792 complete(&work->completion);
8795 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8797 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8799 return req->ctx == data;
8802 static void io_ring_exit_work(struct work_struct *work)
8804 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8805 unsigned long timeout = jiffies + HZ * 60 * 5;
8806 unsigned long interval = HZ / 20;
8807 struct io_tctx_exit exit;
8808 struct io_tctx_node *node;
8812 * If we're doing polled IO and end up having requests being
8813 * submitted async (out-of-line), then completions can come in while
8814 * we're waiting for refs to drop. We need to reap these manually,
8815 * as nobody else will be looking for them.
8818 io_uring_try_cancel_requests(ctx, NULL, true);
8820 struct io_sq_data *sqd = ctx->sq_data;
8821 struct task_struct *tsk;
8823 io_sq_thread_park(sqd);
8825 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8826 io_wq_cancel_cb(tsk->io_uring->io_wq,
8827 io_cancel_ctx_cb, ctx, true);
8828 io_sq_thread_unpark(sqd);
8831 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8832 /* there is little hope left, don't run it too often */
8835 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8837 init_completion(&exit.completion);
8838 init_task_work(&exit.task_work, io_tctx_exit_cb);
8841 * Some may use context even when all refs and requests have been put,
8842 * and they are free to do so while still holding uring_lock or
8843 * completion_lock, see io_req_task_submit(). Apart from other work,
8844 * this lock/unlock section also waits them to finish.
8846 mutex_lock(&ctx->uring_lock);
8847 while (!list_empty(&ctx->tctx_list)) {
8848 WARN_ON_ONCE(time_after(jiffies, timeout));
8850 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8852 /* don't spin on a single task if cancellation failed */
8853 list_rotate_left(&ctx->tctx_list);
8854 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8855 if (WARN_ON_ONCE(ret))
8857 wake_up_process(node->task);
8859 mutex_unlock(&ctx->uring_lock);
8860 wait_for_completion(&exit.completion);
8861 mutex_lock(&ctx->uring_lock);
8863 mutex_unlock(&ctx->uring_lock);
8864 spin_lock(&ctx->completion_lock);
8865 spin_unlock(&ctx->completion_lock);
8867 io_ring_ctx_free(ctx);
8870 /* Returns true if we found and killed one or more timeouts */
8871 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8874 struct io_kiocb *req, *tmp;
8877 spin_lock(&ctx->completion_lock);
8878 spin_lock_irq(&ctx->timeout_lock);
8879 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8880 if (io_match_task(req, tsk, cancel_all)) {
8881 io_kill_timeout(req, -ECANCELED);
8885 spin_unlock_irq(&ctx->timeout_lock);
8887 io_commit_cqring(ctx);
8888 spin_unlock(&ctx->completion_lock);
8890 io_cqring_ev_posted(ctx);
8891 return canceled != 0;
8894 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8896 unsigned long index;
8897 struct creds *creds;
8899 mutex_lock(&ctx->uring_lock);
8900 percpu_ref_kill(&ctx->refs);
8902 __io_cqring_overflow_flush(ctx, true);
8903 xa_for_each(&ctx->personalities, index, creds)
8904 io_unregister_personality(ctx, index);
8905 mutex_unlock(&ctx->uring_lock);
8907 io_kill_timeouts(ctx, NULL, true);
8908 io_poll_remove_all(ctx, NULL, true);
8910 /* if we failed setting up the ctx, we might not have any rings */
8911 io_iopoll_try_reap_events(ctx);
8913 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8915 * Use system_unbound_wq to avoid spawning tons of event kworkers
8916 * if we're exiting a ton of rings at the same time. It just adds
8917 * noise and overhead, there's no discernable change in runtime
8918 * over using system_wq.
8920 queue_work(system_unbound_wq, &ctx->exit_work);
8923 static int io_uring_release(struct inode *inode, struct file *file)
8925 struct io_ring_ctx *ctx = file->private_data;
8927 file->private_data = NULL;
8928 io_ring_ctx_wait_and_kill(ctx);
8932 struct io_task_cancel {
8933 struct task_struct *task;
8937 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8939 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8940 struct io_task_cancel *cancel = data;
8943 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8944 struct io_ring_ctx *ctx = req->ctx;
8946 /* protect against races with linked timeouts */
8947 spin_lock(&ctx->completion_lock);
8948 ret = io_match_task(req, cancel->task, cancel->all);
8949 spin_unlock(&ctx->completion_lock);
8951 ret = io_match_task(req, cancel->task, cancel->all);
8956 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8957 struct task_struct *task, bool cancel_all)
8959 struct io_defer_entry *de;
8962 spin_lock(&ctx->completion_lock);
8963 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8964 if (io_match_task(de->req, task, cancel_all)) {
8965 list_cut_position(&list, &ctx->defer_list, &de->list);
8969 spin_unlock(&ctx->completion_lock);
8970 if (list_empty(&list))
8973 while (!list_empty(&list)) {
8974 de = list_first_entry(&list, struct io_defer_entry, list);
8975 list_del_init(&de->list);
8976 io_req_complete_failed(de->req, -ECANCELED);
8982 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8984 struct io_tctx_node *node;
8985 enum io_wq_cancel cret;
8988 mutex_lock(&ctx->uring_lock);
8989 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8990 struct io_uring_task *tctx = node->task->io_uring;
8993 * io_wq will stay alive while we hold uring_lock, because it's
8994 * killed after ctx nodes, which requires to take the lock.
8996 if (!tctx || !tctx->io_wq)
8998 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8999 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9001 mutex_unlock(&ctx->uring_lock);
9006 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9007 struct task_struct *task,
9010 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9011 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9014 enum io_wq_cancel cret;
9018 ret |= io_uring_try_cancel_iowq(ctx);
9019 } else if (tctx && tctx->io_wq) {
9021 * Cancels requests of all rings, not only @ctx, but
9022 * it's fine as the task is in exit/exec.
9024 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9026 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9029 /* SQPOLL thread does its own polling */
9030 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9031 (ctx->sq_data && ctx->sq_data->thread == current)) {
9032 while (!list_empty_careful(&ctx->iopoll_list)) {
9033 io_iopoll_try_reap_events(ctx);
9038 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9039 ret |= io_poll_remove_all(ctx, task, cancel_all);
9040 ret |= io_kill_timeouts(ctx, task, cancel_all);
9042 ret |= io_run_task_work();
9049 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9051 struct io_uring_task *tctx = current->io_uring;
9052 struct io_tctx_node *node;
9055 if (unlikely(!tctx)) {
9056 ret = io_uring_alloc_task_context(current, ctx);
9059 tctx = current->io_uring;
9061 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9062 node = kmalloc(sizeof(*node), GFP_KERNEL);
9066 node->task = current;
9068 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9075 mutex_lock(&ctx->uring_lock);
9076 list_add(&node->ctx_node, &ctx->tctx_list);
9077 mutex_unlock(&ctx->uring_lock);
9084 * Note that this task has used io_uring. We use it for cancelation purposes.
9086 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9088 struct io_uring_task *tctx = current->io_uring;
9090 if (likely(tctx && tctx->last == ctx))
9092 return __io_uring_add_tctx_node(ctx);
9096 * Remove this io_uring_file -> task mapping.
9098 static void io_uring_del_tctx_node(unsigned long index)
9100 struct io_uring_task *tctx = current->io_uring;
9101 struct io_tctx_node *node;
9105 node = xa_erase(&tctx->xa, index);
9109 WARN_ON_ONCE(current != node->task);
9110 WARN_ON_ONCE(list_empty(&node->ctx_node));
9112 mutex_lock(&node->ctx->uring_lock);
9113 list_del(&node->ctx_node);
9114 mutex_unlock(&node->ctx->uring_lock);
9116 if (tctx->last == node->ctx)
9121 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9123 struct io_wq *wq = tctx->io_wq;
9124 struct io_tctx_node *node;
9125 unsigned long index;
9127 xa_for_each(&tctx->xa, index, node)
9128 io_uring_del_tctx_node(index);
9131 * Must be after io_uring_del_task_file() (removes nodes under
9132 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9135 io_wq_put_and_exit(wq);
9139 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9142 return atomic_read(&tctx->inflight_tracked);
9143 return percpu_counter_sum(&tctx->inflight);
9146 static void io_uring_drop_tctx_refs(struct task_struct *task)
9148 struct io_uring_task *tctx = task->io_uring;
9149 unsigned int refs = tctx->cached_refs;
9152 tctx->cached_refs = 0;
9153 percpu_counter_sub(&tctx->inflight, refs);
9154 put_task_struct_many(task, refs);
9159 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9160 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9162 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9164 struct io_uring_task *tctx = current->io_uring;
9165 struct io_ring_ctx *ctx;
9169 WARN_ON_ONCE(sqd && sqd->thread != current);
9171 if (!current->io_uring)
9174 io_wq_exit_start(tctx->io_wq);
9176 atomic_inc(&tctx->in_idle);
9178 io_uring_drop_tctx_refs(current);
9179 /* read completions before cancelations */
9180 inflight = tctx_inflight(tctx, !cancel_all);
9185 struct io_tctx_node *node;
9186 unsigned long index;
9188 xa_for_each(&tctx->xa, index, node) {
9189 /* sqpoll task will cancel all its requests */
9190 if (node->ctx->sq_data)
9192 io_uring_try_cancel_requests(node->ctx, current,
9196 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9197 io_uring_try_cancel_requests(ctx, current,
9201 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9202 io_uring_drop_tctx_refs(current);
9204 * If we've seen completions, retry without waiting. This
9205 * avoids a race where a completion comes in before we did
9206 * prepare_to_wait().
9208 if (inflight == tctx_inflight(tctx, !cancel_all))
9210 finish_wait(&tctx->wait, &wait);
9212 atomic_dec(&tctx->in_idle);
9214 io_uring_clean_tctx(tctx);
9216 /* for exec all current's requests should be gone, kill tctx */
9217 __io_uring_free(current);
9221 void __io_uring_cancel(bool cancel_all)
9223 io_uring_cancel_generic(cancel_all, NULL);
9226 static void *io_uring_validate_mmap_request(struct file *file,
9227 loff_t pgoff, size_t sz)
9229 struct io_ring_ctx *ctx = file->private_data;
9230 loff_t offset = pgoff << PAGE_SHIFT;
9235 case IORING_OFF_SQ_RING:
9236 case IORING_OFF_CQ_RING:
9239 case IORING_OFF_SQES:
9243 return ERR_PTR(-EINVAL);
9246 page = virt_to_head_page(ptr);
9247 if (sz > page_size(page))
9248 return ERR_PTR(-EINVAL);
9255 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9257 size_t sz = vma->vm_end - vma->vm_start;
9261 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9263 return PTR_ERR(ptr);
9265 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9266 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9269 #else /* !CONFIG_MMU */
9271 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9273 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9276 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9278 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9281 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9282 unsigned long addr, unsigned long len,
9283 unsigned long pgoff, unsigned long flags)
9287 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9289 return PTR_ERR(ptr);
9291 return (unsigned long) ptr;
9294 #endif /* !CONFIG_MMU */
9296 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9301 if (!io_sqring_full(ctx))
9303 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9305 if (!io_sqring_full(ctx))
9308 } while (!signal_pending(current));
9310 finish_wait(&ctx->sqo_sq_wait, &wait);
9314 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9315 struct __kernel_timespec __user **ts,
9316 const sigset_t __user **sig)
9318 struct io_uring_getevents_arg arg;
9321 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9322 * is just a pointer to the sigset_t.
9324 if (!(flags & IORING_ENTER_EXT_ARG)) {
9325 *sig = (const sigset_t __user *) argp;
9331 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9332 * timespec and sigset_t pointers if good.
9334 if (*argsz != sizeof(arg))
9336 if (copy_from_user(&arg, argp, sizeof(arg)))
9338 *sig = u64_to_user_ptr(arg.sigmask);
9339 *argsz = arg.sigmask_sz;
9340 *ts = u64_to_user_ptr(arg.ts);
9344 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9345 u32, min_complete, u32, flags, const void __user *, argp,
9348 struct io_ring_ctx *ctx;
9355 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9356 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9360 if (unlikely(!f.file))
9364 if (unlikely(f.file->f_op != &io_uring_fops))
9368 ctx = f.file->private_data;
9369 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9373 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9377 * For SQ polling, the thread will do all submissions and completions.
9378 * Just return the requested submit count, and wake the thread if
9382 if (ctx->flags & IORING_SETUP_SQPOLL) {
9383 io_cqring_overflow_flush(ctx);
9385 if (unlikely(ctx->sq_data->thread == NULL)) {
9389 if (flags & IORING_ENTER_SQ_WAKEUP)
9390 wake_up(&ctx->sq_data->wait);
9391 if (flags & IORING_ENTER_SQ_WAIT) {
9392 ret = io_sqpoll_wait_sq(ctx);
9396 submitted = to_submit;
9397 } else if (to_submit) {
9398 ret = io_uring_add_tctx_node(ctx);
9401 mutex_lock(&ctx->uring_lock);
9402 submitted = io_submit_sqes(ctx, to_submit);
9403 mutex_unlock(&ctx->uring_lock);
9405 if (submitted != to_submit)
9408 if (flags & IORING_ENTER_GETEVENTS) {
9409 const sigset_t __user *sig;
9410 struct __kernel_timespec __user *ts;
9412 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9416 min_complete = min(min_complete, ctx->cq_entries);
9419 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9420 * space applications don't need to do io completion events
9421 * polling again, they can rely on io_sq_thread to do polling
9422 * work, which can reduce cpu usage and uring_lock contention.
9424 if (ctx->flags & IORING_SETUP_IOPOLL &&
9425 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9426 ret = io_iopoll_check(ctx, min_complete);
9428 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9433 percpu_ref_put(&ctx->refs);
9436 return submitted ? submitted : ret;
9439 #ifdef CONFIG_PROC_FS
9440 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9441 const struct cred *cred)
9443 struct user_namespace *uns = seq_user_ns(m);
9444 struct group_info *gi;
9449 seq_printf(m, "%5d\n", id);
9450 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9451 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9452 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9453 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9454 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9455 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9456 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9457 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9458 seq_puts(m, "\n\tGroups:\t");
9459 gi = cred->group_info;
9460 for (g = 0; g < gi->ngroups; g++) {
9461 seq_put_decimal_ull(m, g ? " " : "",
9462 from_kgid_munged(uns, gi->gid[g]));
9464 seq_puts(m, "\n\tCapEff:\t");
9465 cap = cred->cap_effective;
9466 CAP_FOR_EACH_U32(__capi)
9467 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9472 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9474 struct io_sq_data *sq = NULL;
9479 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9480 * since fdinfo case grabs it in the opposite direction of normal use
9481 * cases. If we fail to get the lock, we just don't iterate any
9482 * structures that could be going away outside the io_uring mutex.
9484 has_lock = mutex_trylock(&ctx->uring_lock);
9486 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9492 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9493 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9494 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9495 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9496 struct file *f = io_file_from_index(ctx, i);
9499 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9501 seq_printf(m, "%5u: <none>\n", i);
9503 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9504 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9505 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9506 unsigned int len = buf->ubuf_end - buf->ubuf;
9508 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9510 if (has_lock && !xa_empty(&ctx->personalities)) {
9511 unsigned long index;
9512 const struct cred *cred;
9514 seq_printf(m, "Personalities:\n");
9515 xa_for_each(&ctx->personalities, index, cred)
9516 io_uring_show_cred(m, index, cred);
9518 seq_printf(m, "PollList:\n");
9519 spin_lock(&ctx->completion_lock);
9520 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9521 struct hlist_head *list = &ctx->cancel_hash[i];
9522 struct io_kiocb *req;
9524 hlist_for_each_entry(req, list, hash_node)
9525 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9526 req->task->task_works != NULL);
9528 spin_unlock(&ctx->completion_lock);
9530 mutex_unlock(&ctx->uring_lock);
9533 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9535 struct io_ring_ctx *ctx = f->private_data;
9537 if (percpu_ref_tryget(&ctx->refs)) {
9538 __io_uring_show_fdinfo(ctx, m);
9539 percpu_ref_put(&ctx->refs);
9544 static const struct file_operations io_uring_fops = {
9545 .release = io_uring_release,
9546 .mmap = io_uring_mmap,
9548 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9549 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9551 .poll = io_uring_poll,
9552 .fasync = io_uring_fasync,
9553 #ifdef CONFIG_PROC_FS
9554 .show_fdinfo = io_uring_show_fdinfo,
9558 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9559 struct io_uring_params *p)
9561 struct io_rings *rings;
9562 size_t size, sq_array_offset;
9564 /* make sure these are sane, as we already accounted them */
9565 ctx->sq_entries = p->sq_entries;
9566 ctx->cq_entries = p->cq_entries;
9568 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9569 if (size == SIZE_MAX)
9572 rings = io_mem_alloc(size);
9577 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9578 rings->sq_ring_mask = p->sq_entries - 1;
9579 rings->cq_ring_mask = p->cq_entries - 1;
9580 rings->sq_ring_entries = p->sq_entries;
9581 rings->cq_ring_entries = p->cq_entries;
9583 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9584 if (size == SIZE_MAX) {
9585 io_mem_free(ctx->rings);
9590 ctx->sq_sqes = io_mem_alloc(size);
9591 if (!ctx->sq_sqes) {
9592 io_mem_free(ctx->rings);
9600 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9604 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9608 ret = io_uring_add_tctx_node(ctx);
9613 fd_install(fd, file);
9618 * Allocate an anonymous fd, this is what constitutes the application
9619 * visible backing of an io_uring instance. The application mmaps this
9620 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9621 * we have to tie this fd to a socket for file garbage collection purposes.
9623 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9626 #if defined(CONFIG_UNIX)
9629 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9632 return ERR_PTR(ret);
9635 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9636 O_RDWR | O_CLOEXEC);
9637 #if defined(CONFIG_UNIX)
9639 sock_release(ctx->ring_sock);
9640 ctx->ring_sock = NULL;
9642 ctx->ring_sock->file = file;
9648 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9649 struct io_uring_params __user *params)
9651 struct io_ring_ctx *ctx;
9657 if (entries > IORING_MAX_ENTRIES) {
9658 if (!(p->flags & IORING_SETUP_CLAMP))
9660 entries = IORING_MAX_ENTRIES;
9664 * Use twice as many entries for the CQ ring. It's possible for the
9665 * application to drive a higher depth than the size of the SQ ring,
9666 * since the sqes are only used at submission time. This allows for
9667 * some flexibility in overcommitting a bit. If the application has
9668 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9669 * of CQ ring entries manually.
9671 p->sq_entries = roundup_pow_of_two(entries);
9672 if (p->flags & IORING_SETUP_CQSIZE) {
9674 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9675 * to a power-of-two, if it isn't already. We do NOT impose
9676 * any cq vs sq ring sizing.
9680 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9681 if (!(p->flags & IORING_SETUP_CLAMP))
9683 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9685 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9686 if (p->cq_entries < p->sq_entries)
9689 p->cq_entries = 2 * p->sq_entries;
9692 ctx = io_ring_ctx_alloc(p);
9695 ctx->compat = in_compat_syscall();
9696 if (!capable(CAP_IPC_LOCK))
9697 ctx->user = get_uid(current_user());
9700 * This is just grabbed for accounting purposes. When a process exits,
9701 * the mm is exited and dropped before the files, hence we need to hang
9702 * on to this mm purely for the purposes of being able to unaccount
9703 * memory (locked/pinned vm). It's not used for anything else.
9705 mmgrab(current->mm);
9706 ctx->mm_account = current->mm;
9708 ret = io_allocate_scq_urings(ctx, p);
9712 ret = io_sq_offload_create(ctx, p);
9715 /* always set a rsrc node */
9716 ret = io_rsrc_node_switch_start(ctx);
9719 io_rsrc_node_switch(ctx, NULL);
9721 memset(&p->sq_off, 0, sizeof(p->sq_off));
9722 p->sq_off.head = offsetof(struct io_rings, sq.head);
9723 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9724 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9725 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9726 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9727 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9728 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9730 memset(&p->cq_off, 0, sizeof(p->cq_off));
9731 p->cq_off.head = offsetof(struct io_rings, cq.head);
9732 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9733 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9734 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9735 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9736 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9737 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9739 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9740 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9741 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9742 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9743 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9744 IORING_FEAT_RSRC_TAGS;
9746 if (copy_to_user(params, p, sizeof(*p))) {
9751 file = io_uring_get_file(ctx);
9753 ret = PTR_ERR(file);
9758 * Install ring fd as the very last thing, so we don't risk someone
9759 * having closed it before we finish setup
9761 ret = io_uring_install_fd(ctx, file);
9763 /* fput will clean it up */
9768 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9771 io_ring_ctx_wait_and_kill(ctx);
9776 * Sets up an aio uring context, and returns the fd. Applications asks for a
9777 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9778 * params structure passed in.
9780 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9782 struct io_uring_params p;
9785 if (copy_from_user(&p, params, sizeof(p)))
9787 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9792 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9793 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9794 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9795 IORING_SETUP_R_DISABLED))
9798 return io_uring_create(entries, &p, params);
9801 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9802 struct io_uring_params __user *, params)
9804 return io_uring_setup(entries, params);
9807 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9809 struct io_uring_probe *p;
9813 size = struct_size(p, ops, nr_args);
9814 if (size == SIZE_MAX)
9816 p = kzalloc(size, GFP_KERNEL);
9821 if (copy_from_user(p, arg, size))
9824 if (memchr_inv(p, 0, size))
9827 p->last_op = IORING_OP_LAST - 1;
9828 if (nr_args > IORING_OP_LAST)
9829 nr_args = IORING_OP_LAST;
9831 for (i = 0; i < nr_args; i++) {
9833 if (!io_op_defs[i].not_supported)
9834 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9839 if (copy_to_user(arg, p, size))
9846 static int io_register_personality(struct io_ring_ctx *ctx)
9848 const struct cred *creds;
9852 creds = get_current_cred();
9854 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9855 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9863 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9864 unsigned int nr_args)
9866 struct io_uring_restriction *res;
9870 /* Restrictions allowed only if rings started disabled */
9871 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9874 /* We allow only a single restrictions registration */
9875 if (ctx->restrictions.registered)
9878 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9881 size = array_size(nr_args, sizeof(*res));
9882 if (size == SIZE_MAX)
9885 res = memdup_user(arg, size);
9887 return PTR_ERR(res);
9891 for (i = 0; i < nr_args; i++) {
9892 switch (res[i].opcode) {
9893 case IORING_RESTRICTION_REGISTER_OP:
9894 if (res[i].register_op >= IORING_REGISTER_LAST) {
9899 __set_bit(res[i].register_op,
9900 ctx->restrictions.register_op);
9902 case IORING_RESTRICTION_SQE_OP:
9903 if (res[i].sqe_op >= IORING_OP_LAST) {
9908 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9910 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9911 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9913 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9914 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9923 /* Reset all restrictions if an error happened */
9925 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9927 ctx->restrictions.registered = true;
9933 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9935 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9938 if (ctx->restrictions.registered)
9939 ctx->restricted = 1;
9941 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9942 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9943 wake_up(&ctx->sq_data->wait);
9947 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9948 struct io_uring_rsrc_update2 *up,
9956 if (check_add_overflow(up->offset, nr_args, &tmp))
9958 err = io_rsrc_node_switch_start(ctx);
9963 case IORING_RSRC_FILE:
9964 return __io_sqe_files_update(ctx, up, nr_args);
9965 case IORING_RSRC_BUFFER:
9966 return __io_sqe_buffers_update(ctx, up, nr_args);
9971 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9974 struct io_uring_rsrc_update2 up;
9978 memset(&up, 0, sizeof(up));
9979 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9981 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9984 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9985 unsigned size, unsigned type)
9987 struct io_uring_rsrc_update2 up;
9989 if (size != sizeof(up))
9991 if (copy_from_user(&up, arg, sizeof(up)))
9993 if (!up.nr || up.resv)
9995 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9998 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9999 unsigned int size, unsigned int type)
10001 struct io_uring_rsrc_register rr;
10003 /* keep it extendible */
10004 if (size != sizeof(rr))
10007 memset(&rr, 0, sizeof(rr));
10008 if (copy_from_user(&rr, arg, size))
10010 if (!rr.nr || rr.resv || rr.resv2)
10014 case IORING_RSRC_FILE:
10015 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10016 rr.nr, u64_to_user_ptr(rr.tags));
10017 case IORING_RSRC_BUFFER:
10018 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10019 rr.nr, u64_to_user_ptr(rr.tags));
10024 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10027 struct io_uring_task *tctx = current->io_uring;
10028 cpumask_var_t new_mask;
10031 if (!tctx || !tctx->io_wq)
10034 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10037 cpumask_clear(new_mask);
10038 if (len > cpumask_size())
10039 len = cpumask_size();
10041 if (copy_from_user(new_mask, arg, len)) {
10042 free_cpumask_var(new_mask);
10046 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10047 free_cpumask_var(new_mask);
10051 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10053 struct io_uring_task *tctx = current->io_uring;
10055 if (!tctx || !tctx->io_wq)
10058 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10061 static bool io_register_op_must_quiesce(int op)
10064 case IORING_REGISTER_BUFFERS:
10065 case IORING_UNREGISTER_BUFFERS:
10066 case IORING_REGISTER_FILES:
10067 case IORING_UNREGISTER_FILES:
10068 case IORING_REGISTER_FILES_UPDATE:
10069 case IORING_REGISTER_PROBE:
10070 case IORING_REGISTER_PERSONALITY:
10071 case IORING_UNREGISTER_PERSONALITY:
10072 case IORING_REGISTER_FILES2:
10073 case IORING_REGISTER_FILES_UPDATE2:
10074 case IORING_REGISTER_BUFFERS2:
10075 case IORING_REGISTER_BUFFERS_UPDATE:
10076 case IORING_REGISTER_IOWQ_AFF:
10077 case IORING_UNREGISTER_IOWQ_AFF:
10084 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10088 percpu_ref_kill(&ctx->refs);
10091 * Drop uring mutex before waiting for references to exit. If another
10092 * thread is currently inside io_uring_enter() it might need to grab the
10093 * uring_lock to make progress. If we hold it here across the drain
10094 * wait, then we can deadlock. It's safe to drop the mutex here, since
10095 * no new references will come in after we've killed the percpu ref.
10097 mutex_unlock(&ctx->uring_lock);
10099 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10102 ret = io_run_task_work_sig();
10103 } while (ret >= 0);
10104 mutex_lock(&ctx->uring_lock);
10107 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10111 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10112 void __user *arg, unsigned nr_args)
10113 __releases(ctx->uring_lock)
10114 __acquires(ctx->uring_lock)
10119 * We're inside the ring mutex, if the ref is already dying, then
10120 * someone else killed the ctx or is already going through
10121 * io_uring_register().
10123 if (percpu_ref_is_dying(&ctx->refs))
10126 if (ctx->restricted) {
10127 if (opcode >= IORING_REGISTER_LAST)
10129 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10130 if (!test_bit(opcode, ctx->restrictions.register_op))
10134 if (io_register_op_must_quiesce(opcode)) {
10135 ret = io_ctx_quiesce(ctx);
10141 case IORING_REGISTER_BUFFERS:
10142 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10144 case IORING_UNREGISTER_BUFFERS:
10146 if (arg || nr_args)
10148 ret = io_sqe_buffers_unregister(ctx);
10150 case IORING_REGISTER_FILES:
10151 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10153 case IORING_UNREGISTER_FILES:
10155 if (arg || nr_args)
10157 ret = io_sqe_files_unregister(ctx);
10159 case IORING_REGISTER_FILES_UPDATE:
10160 ret = io_register_files_update(ctx, arg, nr_args);
10162 case IORING_REGISTER_EVENTFD:
10163 case IORING_REGISTER_EVENTFD_ASYNC:
10167 ret = io_eventfd_register(ctx, arg);
10170 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10171 ctx->eventfd_async = 1;
10173 ctx->eventfd_async = 0;
10175 case IORING_UNREGISTER_EVENTFD:
10177 if (arg || nr_args)
10179 ret = io_eventfd_unregister(ctx);
10181 case IORING_REGISTER_PROBE:
10183 if (!arg || nr_args > 256)
10185 ret = io_probe(ctx, arg, nr_args);
10187 case IORING_REGISTER_PERSONALITY:
10189 if (arg || nr_args)
10191 ret = io_register_personality(ctx);
10193 case IORING_UNREGISTER_PERSONALITY:
10197 ret = io_unregister_personality(ctx, nr_args);
10199 case IORING_REGISTER_ENABLE_RINGS:
10201 if (arg || nr_args)
10203 ret = io_register_enable_rings(ctx);
10205 case IORING_REGISTER_RESTRICTIONS:
10206 ret = io_register_restrictions(ctx, arg, nr_args);
10208 case IORING_REGISTER_FILES2:
10209 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10211 case IORING_REGISTER_FILES_UPDATE2:
10212 ret = io_register_rsrc_update(ctx, arg, nr_args,
10215 case IORING_REGISTER_BUFFERS2:
10216 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10218 case IORING_REGISTER_BUFFERS_UPDATE:
10219 ret = io_register_rsrc_update(ctx, arg, nr_args,
10220 IORING_RSRC_BUFFER);
10222 case IORING_REGISTER_IOWQ_AFF:
10224 if (!arg || !nr_args)
10226 ret = io_register_iowq_aff(ctx, arg, nr_args);
10228 case IORING_UNREGISTER_IOWQ_AFF:
10230 if (arg || nr_args)
10232 ret = io_unregister_iowq_aff(ctx);
10239 if (io_register_op_must_quiesce(opcode)) {
10240 /* bring the ctx back to life */
10241 percpu_ref_reinit(&ctx->refs);
10242 reinit_completion(&ctx->ref_comp);
10247 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10248 void __user *, arg, unsigned int, nr_args)
10250 struct io_ring_ctx *ctx;
10259 if (f.file->f_op != &io_uring_fops)
10262 ctx = f.file->private_data;
10264 io_run_task_work();
10266 mutex_lock(&ctx->uring_lock);
10267 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10268 mutex_unlock(&ctx->uring_lock);
10269 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10270 ctx->cq_ev_fd != NULL, ret);
10276 static int __init io_uring_init(void)
10278 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10279 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10280 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10283 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10284 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10285 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10286 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10287 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10288 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10289 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10290 BUILD_BUG_SQE_ELEM(8, __u64, off);
10291 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10292 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10293 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10294 BUILD_BUG_SQE_ELEM(24, __u32, len);
10295 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10296 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10297 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10299 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10300 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10301 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10302 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10303 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10304 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10305 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10306 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10307 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10308 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10309 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10310 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10311 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10312 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10313 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10314 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10316 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10317 sizeof(struct io_uring_rsrc_update));
10318 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10319 sizeof(struct io_uring_rsrc_update2));
10320 /* should fit into one byte */
10321 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10323 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10324 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10326 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10330 __initcall(io_uring_init);