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
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 (PAGE_SHIFT - 3)
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;
519 unsigned long nofile;
539 struct list_head list;
540 /* head of the link, used by linked timeouts only */
541 struct io_kiocb *head;
542 /* for linked completions */
543 struct io_kiocb *prev;
546 struct io_timeout_rem {
551 struct timespec64 ts;
556 /* NOTE: kiocb has the file as the first member, so don't do it here */
564 struct sockaddr __user *addr;
571 struct compat_msghdr __user *umsg_compat;
572 struct user_msghdr __user *umsg;
578 struct io_buffer *kbuf;
585 struct filename *filename;
587 unsigned long nofile;
590 struct io_rsrc_update {
616 struct epoll_event event;
620 struct file *file_out;
621 struct file *file_in;
628 struct io_provide_buf {
642 const char __user *filename;
643 struct statx __user *buffer;
655 struct filename *oldpath;
656 struct filename *newpath;
664 struct filename *filename;
667 struct io_completion {
672 struct io_async_connect {
673 struct sockaddr_storage address;
676 struct io_async_msghdr {
677 struct iovec fast_iov[UIO_FASTIOV];
678 /* points to an allocated iov, if NULL we use fast_iov instead */
679 struct iovec *free_iov;
680 struct sockaddr __user *uaddr;
682 struct sockaddr_storage addr;
686 struct iovec fast_iov[UIO_FASTIOV];
687 const struct iovec *free_iovec;
688 struct iov_iter iter;
690 struct wait_page_queue wpq;
694 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
695 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
696 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
697 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
698 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
699 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
701 /* first byte is taken by user flags, shift it to not overlap */
706 REQ_F_LINK_TIMEOUT_BIT,
707 REQ_F_NEED_CLEANUP_BIT,
709 REQ_F_BUFFER_SELECTED_BIT,
710 REQ_F_COMPLETE_INLINE_BIT,
712 REQ_F_DONT_REISSUE_BIT,
715 REQ_F_ARM_LTIMEOUT_BIT,
716 /* keep async read/write and isreg together and in order */
717 REQ_F_NOWAIT_READ_BIT,
718 REQ_F_NOWAIT_WRITE_BIT,
721 /* not a real bit, just to check we're not overflowing the space */
727 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
728 /* drain existing IO first */
729 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
731 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
732 /* doesn't sever on completion < 0 */
733 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
735 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
736 /* IOSQE_BUFFER_SELECT */
737 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
739 /* fail rest of links */
740 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
741 /* on inflight list, should be cancelled and waited on exit reliably */
742 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
743 /* read/write uses file position */
744 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
745 /* must not punt to workers */
746 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
747 /* has or had linked timeout */
748 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
750 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
751 /* already went through poll handler */
752 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
753 /* buffer already selected */
754 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
755 /* completion is deferred through io_comp_state */
756 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
757 /* caller should reissue async */
758 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
759 /* don't attempt request reissue, see io_rw_reissue() */
760 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
761 /* supports async reads */
762 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
763 /* supports async writes */
764 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
766 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
767 /* has creds assigned */
768 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
769 /* skip refcounting if not set */
770 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
771 /* there is a linked timeout that has to be armed */
772 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
776 struct io_poll_iocb poll;
777 struct io_poll_iocb *double_poll;
780 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
782 struct io_task_work {
784 struct io_wq_work_node node;
785 struct llist_node fallback_node;
787 io_req_tw_func_t func;
791 IORING_RSRC_FILE = 0,
792 IORING_RSRC_BUFFER = 1,
796 * NOTE! Each of the iocb union members has the file pointer
797 * as the first entry in their struct definition. So you can
798 * access the file pointer through any of the sub-structs,
799 * or directly as just 'ki_filp' in this struct.
805 struct io_poll_iocb poll;
806 struct io_poll_update poll_update;
807 struct io_accept accept;
809 struct io_cancel cancel;
810 struct io_timeout timeout;
811 struct io_timeout_rem timeout_rem;
812 struct io_connect connect;
813 struct io_sr_msg sr_msg;
815 struct io_close close;
816 struct io_rsrc_update rsrc_update;
817 struct io_fadvise fadvise;
818 struct io_madvise madvise;
819 struct io_epoll epoll;
820 struct io_splice splice;
821 struct io_provide_buf pbuf;
822 struct io_statx statx;
823 struct io_shutdown shutdown;
824 struct io_rename rename;
825 struct io_unlink unlink;
826 /* use only after cleaning per-op data, see io_clean_op() */
827 struct io_completion compl;
830 /* opcode allocated if it needs to store data for async defer */
833 /* polled IO has completed */
839 struct io_ring_ctx *ctx;
842 struct task_struct *task;
845 struct io_kiocb *link;
846 struct percpu_ref *fixed_rsrc_refs;
848 /* used with ctx->iopoll_list with reads/writes */
849 struct list_head inflight_entry;
850 struct io_task_work io_task_work;
851 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
852 struct hlist_node hash_node;
853 struct async_poll *apoll;
854 struct io_wq_work work;
855 const struct cred *creds;
857 /* store used ubuf, so we can prevent reloading */
858 struct io_mapped_ubuf *imu;
861 struct io_tctx_node {
862 struct list_head ctx_node;
863 struct task_struct *task;
864 struct io_ring_ctx *ctx;
867 struct io_defer_entry {
868 struct list_head list;
869 struct io_kiocb *req;
874 /* needs req->file assigned */
875 unsigned needs_file : 1;
876 /* hash wq insertion if file is a regular file */
877 unsigned hash_reg_file : 1;
878 /* unbound wq insertion if file is a non-regular file */
879 unsigned unbound_nonreg_file : 1;
880 /* opcode is not supported by this kernel */
881 unsigned not_supported : 1;
882 /* set if opcode supports polled "wait" */
884 unsigned pollout : 1;
885 /* op supports buffer selection */
886 unsigned buffer_select : 1;
887 /* do prep async if is going to be punted */
888 unsigned needs_async_setup : 1;
889 /* should block plug */
891 /* size of async data needed, if any */
892 unsigned short async_size;
895 static const struct io_op_def io_op_defs[] = {
896 [IORING_OP_NOP] = {},
897 [IORING_OP_READV] = {
899 .unbound_nonreg_file = 1,
902 .needs_async_setup = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_WRITEV] = {
909 .unbound_nonreg_file = 1,
911 .needs_async_setup = 1,
913 .async_size = sizeof(struct io_async_rw),
915 [IORING_OP_FSYNC] = {
918 [IORING_OP_READ_FIXED] = {
920 .unbound_nonreg_file = 1,
923 .async_size = sizeof(struct io_async_rw),
925 [IORING_OP_WRITE_FIXED] = {
928 .unbound_nonreg_file = 1,
931 .async_size = sizeof(struct io_async_rw),
933 [IORING_OP_POLL_ADD] = {
935 .unbound_nonreg_file = 1,
937 [IORING_OP_POLL_REMOVE] = {},
938 [IORING_OP_SYNC_FILE_RANGE] = {
941 [IORING_OP_SENDMSG] = {
943 .unbound_nonreg_file = 1,
945 .needs_async_setup = 1,
946 .async_size = sizeof(struct io_async_msghdr),
948 [IORING_OP_RECVMSG] = {
950 .unbound_nonreg_file = 1,
953 .needs_async_setup = 1,
954 .async_size = sizeof(struct io_async_msghdr),
956 [IORING_OP_TIMEOUT] = {
957 .async_size = sizeof(struct io_timeout_data),
959 [IORING_OP_TIMEOUT_REMOVE] = {
960 /* used by timeout updates' prep() */
962 [IORING_OP_ACCEPT] = {
964 .unbound_nonreg_file = 1,
967 [IORING_OP_ASYNC_CANCEL] = {},
968 [IORING_OP_LINK_TIMEOUT] = {
969 .async_size = sizeof(struct io_timeout_data),
971 [IORING_OP_CONNECT] = {
973 .unbound_nonreg_file = 1,
975 .needs_async_setup = 1,
976 .async_size = sizeof(struct io_async_connect),
978 [IORING_OP_FALLOCATE] = {
981 [IORING_OP_OPENAT] = {},
982 [IORING_OP_CLOSE] = {},
983 [IORING_OP_FILES_UPDATE] = {},
984 [IORING_OP_STATX] = {},
987 .unbound_nonreg_file = 1,
991 .async_size = sizeof(struct io_async_rw),
993 [IORING_OP_WRITE] = {
995 .unbound_nonreg_file = 1,
998 .async_size = sizeof(struct io_async_rw),
1000 [IORING_OP_FADVISE] = {
1003 [IORING_OP_MADVISE] = {},
1004 [IORING_OP_SEND] = {
1006 .unbound_nonreg_file = 1,
1009 [IORING_OP_RECV] = {
1011 .unbound_nonreg_file = 1,
1015 [IORING_OP_OPENAT2] = {
1017 [IORING_OP_EPOLL_CTL] = {
1018 .unbound_nonreg_file = 1,
1020 [IORING_OP_SPLICE] = {
1023 .unbound_nonreg_file = 1,
1025 [IORING_OP_PROVIDE_BUFFERS] = {},
1026 [IORING_OP_REMOVE_BUFFERS] = {},
1030 .unbound_nonreg_file = 1,
1032 [IORING_OP_SHUTDOWN] = {
1035 [IORING_OP_RENAMEAT] = {},
1036 [IORING_OP_UNLINKAT] = {},
1039 /* requests with any of those set should undergo io_disarm_next() */
1040 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1042 static bool io_disarm_next(struct io_kiocb *req);
1043 static void io_uring_del_tctx_node(unsigned long index);
1044 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1045 struct task_struct *task,
1047 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1049 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1050 long res, unsigned int cflags);
1051 static void io_put_req(struct io_kiocb *req);
1052 static void io_put_req_deferred(struct io_kiocb *req);
1053 static void io_dismantle_req(struct io_kiocb *req);
1054 static void io_queue_linked_timeout(struct io_kiocb *req);
1055 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1056 struct io_uring_rsrc_update2 *up,
1058 static void io_clean_op(struct io_kiocb *req);
1059 static struct file *io_file_get(struct io_ring_ctx *ctx,
1060 struct io_kiocb *req, int fd, bool fixed);
1061 static void __io_queue_sqe(struct io_kiocb *req);
1062 static void io_rsrc_put_work(struct work_struct *work);
1064 static void io_req_task_queue(struct io_kiocb *req);
1065 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1066 static int io_req_prep_async(struct io_kiocb *req);
1068 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1069 unsigned int issue_flags, u32 slot_index);
1071 static struct kmem_cache *req_cachep;
1073 static const struct file_operations io_uring_fops;
1075 struct sock *io_uring_get_socket(struct file *file)
1077 #if defined(CONFIG_UNIX)
1078 if (file->f_op == &io_uring_fops) {
1079 struct io_ring_ctx *ctx = file->private_data;
1081 return ctx->ring_sock->sk;
1086 EXPORT_SYMBOL(io_uring_get_socket);
1088 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1091 mutex_lock(&ctx->uring_lock);
1096 #define io_for_each_link(pos, head) \
1097 for (pos = (head); pos; pos = pos->link)
1100 * Shamelessly stolen from the mm implementation of page reference checking,
1101 * see commit f958d7b528b1 for details.
1103 #define req_ref_zero_or_close_to_overflow(req) \
1104 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1106 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1108 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1109 return atomic_inc_not_zero(&req->refs);
1112 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1114 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1117 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1118 return atomic_dec_and_test(&req->refs);
1121 static inline void req_ref_put(struct io_kiocb *req)
1123 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1124 WARN_ON_ONCE(req_ref_put_and_test(req));
1127 static inline void req_ref_get(struct io_kiocb *req)
1129 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1130 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1131 atomic_inc(&req->refs);
1134 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1136 if (!(req->flags & REQ_F_REFCOUNT)) {
1137 req->flags |= REQ_F_REFCOUNT;
1138 atomic_set(&req->refs, nr);
1142 static inline void io_req_set_refcount(struct io_kiocb *req)
1144 __io_req_set_refcount(req, 1);
1147 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1149 struct io_ring_ctx *ctx = req->ctx;
1151 if (!req->fixed_rsrc_refs) {
1152 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1153 percpu_ref_get(req->fixed_rsrc_refs);
1157 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1159 bool got = percpu_ref_tryget(ref);
1161 /* already at zero, wait for ->release() */
1163 wait_for_completion(compl);
1164 percpu_ref_resurrect(ref);
1166 percpu_ref_put(ref);
1169 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1172 struct io_kiocb *req;
1174 if (task && head->task != task)
1179 io_for_each_link(req, head) {
1180 if (req->flags & REQ_F_INFLIGHT)
1186 static inline void req_set_fail(struct io_kiocb *req)
1188 req->flags |= REQ_F_FAIL;
1191 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1197 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1199 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1201 complete(&ctx->ref_comp);
1204 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1206 return !req->timeout.off;
1209 static void io_fallback_req_func(struct work_struct *work)
1211 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1212 fallback_work.work);
1213 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1214 struct io_kiocb *req, *tmp;
1215 bool locked = false;
1217 percpu_ref_get(&ctx->refs);
1218 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1219 req->io_task_work.func(req, &locked);
1222 if (ctx->submit_state.compl_nr)
1223 io_submit_flush_completions(ctx);
1224 mutex_unlock(&ctx->uring_lock);
1226 percpu_ref_put(&ctx->refs);
1230 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1232 struct io_ring_ctx *ctx;
1235 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1240 * Use 5 bits less than the max cq entries, that should give us around
1241 * 32 entries per hash list if totally full and uniformly spread.
1243 hash_bits = ilog2(p->cq_entries);
1247 ctx->cancel_hash_bits = hash_bits;
1248 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1250 if (!ctx->cancel_hash)
1252 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1254 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1255 if (!ctx->dummy_ubuf)
1257 /* set invalid range, so io_import_fixed() fails meeting it */
1258 ctx->dummy_ubuf->ubuf = -1UL;
1260 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1261 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1264 ctx->flags = p->flags;
1265 init_waitqueue_head(&ctx->sqo_sq_wait);
1266 INIT_LIST_HEAD(&ctx->sqd_list);
1267 init_waitqueue_head(&ctx->poll_wait);
1268 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1269 init_completion(&ctx->ref_comp);
1270 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1271 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1272 mutex_init(&ctx->uring_lock);
1273 init_waitqueue_head(&ctx->cq_wait);
1274 spin_lock_init(&ctx->completion_lock);
1275 spin_lock_init(&ctx->timeout_lock);
1276 INIT_LIST_HEAD(&ctx->iopoll_list);
1277 INIT_LIST_HEAD(&ctx->defer_list);
1278 INIT_LIST_HEAD(&ctx->timeout_list);
1279 spin_lock_init(&ctx->rsrc_ref_lock);
1280 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1281 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1282 init_llist_head(&ctx->rsrc_put_llist);
1283 INIT_LIST_HEAD(&ctx->tctx_list);
1284 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1285 INIT_LIST_HEAD(&ctx->locked_free_list);
1286 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1289 kfree(ctx->dummy_ubuf);
1290 kfree(ctx->cancel_hash);
1295 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1297 struct io_rings *r = ctx->rings;
1299 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1303 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1305 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1306 struct io_ring_ctx *ctx = req->ctx;
1308 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1314 #define FFS_ASYNC_READ 0x1UL
1315 #define FFS_ASYNC_WRITE 0x2UL
1317 #define FFS_ISREG 0x4UL
1319 #define FFS_ISREG 0x0UL
1321 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1323 static inline bool io_req_ffs_set(struct io_kiocb *req)
1325 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1328 static void io_req_track_inflight(struct io_kiocb *req)
1330 if (!(req->flags & REQ_F_INFLIGHT)) {
1331 req->flags |= REQ_F_INFLIGHT;
1332 atomic_inc(¤t->io_uring->inflight_tracked);
1336 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1338 req->flags &= ~REQ_F_LINK_TIMEOUT;
1341 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1343 if (WARN_ON_ONCE(!req->link))
1346 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1347 req->flags |= REQ_F_LINK_TIMEOUT;
1349 /* linked timeouts should have two refs once prep'ed */
1350 io_req_set_refcount(req);
1351 __io_req_set_refcount(req->link, 2);
1355 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1357 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1359 return __io_prep_linked_timeout(req);
1362 static void io_prep_async_work(struct io_kiocb *req)
1364 const struct io_op_def *def = &io_op_defs[req->opcode];
1365 struct io_ring_ctx *ctx = req->ctx;
1367 if (!(req->flags & REQ_F_CREDS)) {
1368 req->flags |= REQ_F_CREDS;
1369 req->creds = get_current_cred();
1372 req->work.list.next = NULL;
1373 req->work.flags = 0;
1374 if (req->flags & REQ_F_FORCE_ASYNC)
1375 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1377 if (req->flags & REQ_F_ISREG) {
1378 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1379 io_wq_hash_work(&req->work, file_inode(req->file));
1380 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1381 if (def->unbound_nonreg_file)
1382 req->work.flags |= IO_WQ_WORK_UNBOUND;
1385 switch (req->opcode) {
1386 case IORING_OP_SPLICE:
1388 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1389 req->work.flags |= IO_WQ_WORK_UNBOUND;
1394 static void io_prep_async_link(struct io_kiocb *req)
1396 struct io_kiocb *cur;
1398 if (req->flags & REQ_F_LINK_TIMEOUT) {
1399 struct io_ring_ctx *ctx = req->ctx;
1401 spin_lock(&ctx->completion_lock);
1402 io_for_each_link(cur, req)
1403 io_prep_async_work(cur);
1404 spin_unlock(&ctx->completion_lock);
1406 io_for_each_link(cur, req)
1407 io_prep_async_work(cur);
1411 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1413 struct io_ring_ctx *ctx = req->ctx;
1414 struct io_kiocb *link = io_prep_linked_timeout(req);
1415 struct io_uring_task *tctx = req->task->io_uring;
1417 /* must not take the lock, NULL it as a precaution */
1421 BUG_ON(!tctx->io_wq);
1423 /* init ->work of the whole link before punting */
1424 io_prep_async_link(req);
1427 * Not expected to happen, but if we do have a bug where this _can_
1428 * happen, catch it here and ensure the request is marked as
1429 * canceled. That will make io-wq go through the usual work cancel
1430 * procedure rather than attempt to run this request (or create a new
1433 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1434 req->work.flags |= IO_WQ_WORK_CANCEL;
1436 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1437 &req->work, req->flags);
1438 io_wq_enqueue(tctx->io_wq, &req->work);
1440 io_queue_linked_timeout(link);
1443 static void io_kill_timeout(struct io_kiocb *req, int status)
1444 __must_hold(&req->ctx->completion_lock)
1445 __must_hold(&req->ctx->timeout_lock)
1447 struct io_timeout_data *io = req->async_data;
1449 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1450 atomic_set(&req->ctx->cq_timeouts,
1451 atomic_read(&req->ctx->cq_timeouts) + 1);
1452 list_del_init(&req->timeout.list);
1453 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1454 io_put_req_deferred(req);
1458 static void io_queue_deferred(struct io_ring_ctx *ctx)
1460 while (!list_empty(&ctx->defer_list)) {
1461 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1462 struct io_defer_entry, list);
1464 if (req_need_defer(de->req, de->seq))
1466 list_del_init(&de->list);
1467 io_req_task_queue(de->req);
1472 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1473 __must_hold(&ctx->completion_lock)
1475 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1477 spin_lock_irq(&ctx->timeout_lock);
1478 while (!list_empty(&ctx->timeout_list)) {
1479 u32 events_needed, events_got;
1480 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1481 struct io_kiocb, timeout.list);
1483 if (io_is_timeout_noseq(req))
1487 * Since seq can easily wrap around over time, subtract
1488 * the last seq at which timeouts were flushed before comparing.
1489 * Assuming not more than 2^31-1 events have happened since,
1490 * these subtractions won't have wrapped, so we can check if
1491 * target is in [last_seq, current_seq] by comparing the two.
1493 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1494 events_got = seq - ctx->cq_last_tm_flush;
1495 if (events_got < events_needed)
1498 list_del_init(&req->timeout.list);
1499 io_kill_timeout(req, 0);
1501 ctx->cq_last_tm_flush = seq;
1502 spin_unlock_irq(&ctx->timeout_lock);
1505 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1507 if (ctx->off_timeout_used)
1508 io_flush_timeouts(ctx);
1509 if (ctx->drain_active)
1510 io_queue_deferred(ctx);
1513 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1515 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1516 __io_commit_cqring_flush(ctx);
1517 /* order cqe stores with ring update */
1518 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1521 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1523 struct io_rings *r = ctx->rings;
1525 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1528 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1530 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1533 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1535 struct io_rings *rings = ctx->rings;
1536 unsigned tail, mask = ctx->cq_entries - 1;
1539 * writes to the cq entry need to come after reading head; the
1540 * control dependency is enough as we're using WRITE_ONCE to
1543 if (__io_cqring_events(ctx) == ctx->cq_entries)
1546 tail = ctx->cached_cq_tail++;
1547 return &rings->cqes[tail & mask];
1550 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1552 if (likely(!ctx->cq_ev_fd))
1554 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1556 return !ctx->eventfd_async || io_wq_current_is_worker();
1560 * This should only get called when at least one event has been posted.
1561 * Some applications rely on the eventfd notification count only changing
1562 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1563 * 1:1 relationship between how many times this function is called (and
1564 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1566 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1569 * wake_up_all() may seem excessive, but io_wake_function() and
1570 * io_should_wake() handle the termination of the loop and only
1571 * wake as many waiters as we need to.
1573 if (wq_has_sleeper(&ctx->cq_wait))
1574 wake_up_all(&ctx->cq_wait);
1575 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1576 wake_up(&ctx->sq_data->wait);
1577 if (io_should_trigger_evfd(ctx))
1578 eventfd_signal(ctx->cq_ev_fd, 1);
1579 if (waitqueue_active(&ctx->poll_wait)) {
1580 wake_up_interruptible(&ctx->poll_wait);
1581 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1585 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1587 if (ctx->flags & IORING_SETUP_SQPOLL) {
1588 if (wq_has_sleeper(&ctx->cq_wait))
1589 wake_up_all(&ctx->cq_wait);
1591 if (io_should_trigger_evfd(ctx))
1592 eventfd_signal(ctx->cq_ev_fd, 1);
1593 if (waitqueue_active(&ctx->poll_wait)) {
1594 wake_up_interruptible(&ctx->poll_wait);
1595 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1599 /* Returns true if there are no backlogged entries after the flush */
1600 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1602 bool all_flushed, posted;
1604 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1608 spin_lock(&ctx->completion_lock);
1609 while (!list_empty(&ctx->cq_overflow_list)) {
1610 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1611 struct io_overflow_cqe *ocqe;
1615 ocqe = list_first_entry(&ctx->cq_overflow_list,
1616 struct io_overflow_cqe, list);
1618 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1620 io_account_cq_overflow(ctx);
1623 list_del(&ocqe->list);
1627 all_flushed = list_empty(&ctx->cq_overflow_list);
1629 clear_bit(0, &ctx->check_cq_overflow);
1630 WRITE_ONCE(ctx->rings->sq_flags,
1631 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1635 io_commit_cqring(ctx);
1636 spin_unlock(&ctx->completion_lock);
1638 io_cqring_ev_posted(ctx);
1642 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1646 if (test_bit(0, &ctx->check_cq_overflow)) {
1647 /* iopoll syncs against uring_lock, not completion_lock */
1648 if (ctx->flags & IORING_SETUP_IOPOLL)
1649 mutex_lock(&ctx->uring_lock);
1650 ret = __io_cqring_overflow_flush(ctx, false);
1651 if (ctx->flags & IORING_SETUP_IOPOLL)
1652 mutex_unlock(&ctx->uring_lock);
1658 /* must to be called somewhat shortly after putting a request */
1659 static inline void io_put_task(struct task_struct *task, int nr)
1661 struct io_uring_task *tctx = task->io_uring;
1663 if (likely(task == current)) {
1664 tctx->cached_refs += nr;
1666 percpu_counter_sub(&tctx->inflight, nr);
1667 if (unlikely(atomic_read(&tctx->in_idle)))
1668 wake_up(&tctx->wait);
1669 put_task_struct_many(task, nr);
1673 static void io_task_refs_refill(struct io_uring_task *tctx)
1675 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1677 percpu_counter_add(&tctx->inflight, refill);
1678 refcount_add(refill, ¤t->usage);
1679 tctx->cached_refs += refill;
1682 static inline void io_get_task_refs(int nr)
1684 struct io_uring_task *tctx = current->io_uring;
1686 tctx->cached_refs -= nr;
1687 if (unlikely(tctx->cached_refs < 0))
1688 io_task_refs_refill(tctx);
1691 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1692 long res, unsigned int cflags)
1694 struct io_overflow_cqe *ocqe;
1696 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1699 * If we're in ring overflow flush mode, or in task cancel mode,
1700 * or cannot allocate an overflow entry, then we need to drop it
1703 io_account_cq_overflow(ctx);
1706 if (list_empty(&ctx->cq_overflow_list)) {
1707 set_bit(0, &ctx->check_cq_overflow);
1708 WRITE_ONCE(ctx->rings->sq_flags,
1709 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1712 ocqe->cqe.user_data = user_data;
1713 ocqe->cqe.res = res;
1714 ocqe->cqe.flags = cflags;
1715 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1719 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1720 long res, unsigned int cflags)
1722 struct io_uring_cqe *cqe;
1724 trace_io_uring_complete(ctx, user_data, res, cflags);
1727 * If we can't get a cq entry, userspace overflowed the
1728 * submission (by quite a lot). Increment the overflow count in
1731 cqe = io_get_cqe(ctx);
1733 WRITE_ONCE(cqe->user_data, user_data);
1734 WRITE_ONCE(cqe->res, res);
1735 WRITE_ONCE(cqe->flags, cflags);
1738 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1741 /* not as hot to bloat with inlining */
1742 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1743 long res, unsigned int cflags)
1745 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1748 static void io_req_complete_post(struct io_kiocb *req, long res,
1749 unsigned int cflags)
1751 struct io_ring_ctx *ctx = req->ctx;
1753 spin_lock(&ctx->completion_lock);
1754 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1756 * If we're the last reference to this request, add to our locked
1759 if (req_ref_put_and_test(req)) {
1760 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1761 if (req->flags & IO_DISARM_MASK)
1762 io_disarm_next(req);
1764 io_req_task_queue(req->link);
1768 io_dismantle_req(req);
1769 io_put_task(req->task, 1);
1770 list_add(&req->inflight_entry, &ctx->locked_free_list);
1771 ctx->locked_free_nr++;
1773 if (!percpu_ref_tryget(&ctx->refs))
1776 io_commit_cqring(ctx);
1777 spin_unlock(&ctx->completion_lock);
1780 io_cqring_ev_posted(ctx);
1781 percpu_ref_put(&ctx->refs);
1785 static inline bool io_req_needs_clean(struct io_kiocb *req)
1787 return req->flags & IO_REQ_CLEAN_FLAGS;
1790 static void io_req_complete_state(struct io_kiocb *req, long res,
1791 unsigned int cflags)
1793 if (io_req_needs_clean(req))
1796 req->compl.cflags = cflags;
1797 req->flags |= REQ_F_COMPLETE_INLINE;
1800 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1801 long res, unsigned cflags)
1803 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1804 io_req_complete_state(req, res, cflags);
1806 io_req_complete_post(req, res, cflags);
1809 static inline void io_req_complete(struct io_kiocb *req, long res)
1811 __io_req_complete(req, 0, res, 0);
1814 static void io_req_complete_failed(struct io_kiocb *req, long res)
1817 io_req_complete_post(req, res, 0);
1821 * Don't initialise the fields below on every allocation, but do that in
1822 * advance and keep them valid across allocations.
1824 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1828 req->async_data = NULL;
1829 /* not necessary, but safer to zero */
1833 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1834 struct io_submit_state *state)
1836 spin_lock(&ctx->completion_lock);
1837 list_splice_init(&ctx->locked_free_list, &state->free_list);
1838 ctx->locked_free_nr = 0;
1839 spin_unlock(&ctx->completion_lock);
1842 /* Returns true IFF there are requests in the cache */
1843 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1845 struct io_submit_state *state = &ctx->submit_state;
1849 * If we have more than a batch's worth of requests in our IRQ side
1850 * locked cache, grab the lock and move them over to our submission
1853 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1854 io_flush_cached_locked_reqs(ctx, state);
1856 nr = state->free_reqs;
1857 while (!list_empty(&state->free_list)) {
1858 struct io_kiocb *req = list_first_entry(&state->free_list,
1859 struct io_kiocb, inflight_entry);
1861 list_del(&req->inflight_entry);
1862 state->reqs[nr++] = req;
1863 if (nr == ARRAY_SIZE(state->reqs))
1867 state->free_reqs = nr;
1872 * A request might get retired back into the request caches even before opcode
1873 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1874 * Because of that, io_alloc_req() should be called only under ->uring_lock
1875 * and with extra caution to not get a request that is still worked on.
1877 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1878 __must_hold(&ctx->uring_lock)
1880 struct io_submit_state *state = &ctx->submit_state;
1881 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1884 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1886 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1889 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1893 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1894 * retry single alloc to be on the safe side.
1896 if (unlikely(ret <= 0)) {
1897 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1898 if (!state->reqs[0])
1903 for (i = 0; i < ret; i++)
1904 io_preinit_req(state->reqs[i], ctx);
1905 state->free_reqs = ret;
1908 return state->reqs[state->free_reqs];
1911 static inline void io_put_file(struct file *file)
1917 static void io_dismantle_req(struct io_kiocb *req)
1919 unsigned int flags = req->flags;
1921 if (io_req_needs_clean(req))
1923 if (!(flags & REQ_F_FIXED_FILE))
1924 io_put_file(req->file);
1925 if (req->fixed_rsrc_refs)
1926 percpu_ref_put(req->fixed_rsrc_refs);
1927 if (req->async_data) {
1928 kfree(req->async_data);
1929 req->async_data = NULL;
1933 static void __io_free_req(struct io_kiocb *req)
1935 struct io_ring_ctx *ctx = req->ctx;
1937 io_dismantle_req(req);
1938 io_put_task(req->task, 1);
1940 spin_lock(&ctx->completion_lock);
1941 list_add(&req->inflight_entry, &ctx->locked_free_list);
1942 ctx->locked_free_nr++;
1943 spin_unlock(&ctx->completion_lock);
1945 percpu_ref_put(&ctx->refs);
1948 static inline void io_remove_next_linked(struct io_kiocb *req)
1950 struct io_kiocb *nxt = req->link;
1952 req->link = nxt->link;
1956 static bool io_kill_linked_timeout(struct io_kiocb *req)
1957 __must_hold(&req->ctx->completion_lock)
1958 __must_hold(&req->ctx->timeout_lock)
1960 struct io_kiocb *link = req->link;
1962 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1963 struct io_timeout_data *io = link->async_data;
1965 io_remove_next_linked(req);
1966 link->timeout.head = NULL;
1967 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1968 io_cqring_fill_event(link->ctx, link->user_data,
1970 io_put_req_deferred(link);
1977 static void io_fail_links(struct io_kiocb *req)
1978 __must_hold(&req->ctx->completion_lock)
1980 struct io_kiocb *nxt, *link = req->link;
1984 long res = -ECANCELED;
1986 if (link->flags & REQ_F_FAIL)
1992 trace_io_uring_fail_link(req, link);
1993 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
1994 io_put_req_deferred(link);
1999 static bool io_disarm_next(struct io_kiocb *req)
2000 __must_hold(&req->ctx->completion_lock)
2002 bool posted = false;
2004 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2005 struct io_kiocb *link = req->link;
2007 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2008 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2009 io_remove_next_linked(req);
2010 io_cqring_fill_event(link->ctx, link->user_data,
2012 io_put_req_deferred(link);
2015 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2016 struct io_ring_ctx *ctx = req->ctx;
2018 spin_lock_irq(&ctx->timeout_lock);
2019 posted = io_kill_linked_timeout(req);
2020 spin_unlock_irq(&ctx->timeout_lock);
2022 if (unlikely((req->flags & REQ_F_FAIL) &&
2023 !(req->flags & REQ_F_HARDLINK))) {
2024 posted |= (req->link != NULL);
2030 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2032 struct io_kiocb *nxt;
2035 * If LINK is set, we have dependent requests in this chain. If we
2036 * didn't fail this request, queue the first one up, moving any other
2037 * dependencies to the next request. In case of failure, fail the rest
2040 if (req->flags & IO_DISARM_MASK) {
2041 struct io_ring_ctx *ctx = req->ctx;
2044 spin_lock(&ctx->completion_lock);
2045 posted = io_disarm_next(req);
2047 io_commit_cqring(req->ctx);
2048 spin_unlock(&ctx->completion_lock);
2050 io_cqring_ev_posted(ctx);
2057 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2059 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2061 return __io_req_find_next(req);
2064 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2069 if (ctx->submit_state.compl_nr)
2070 io_submit_flush_completions(ctx);
2071 mutex_unlock(&ctx->uring_lock);
2074 percpu_ref_put(&ctx->refs);
2077 static void tctx_task_work(struct callback_head *cb)
2079 bool locked = false;
2080 struct io_ring_ctx *ctx = NULL;
2081 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2085 struct io_wq_work_node *node;
2087 spin_lock_irq(&tctx->task_lock);
2088 node = tctx->task_list.first;
2089 INIT_WQ_LIST(&tctx->task_list);
2091 tctx->task_running = false;
2092 spin_unlock_irq(&tctx->task_lock);
2097 struct io_wq_work_node *next = node->next;
2098 struct io_kiocb *req = container_of(node, struct io_kiocb,
2101 if (req->ctx != ctx) {
2102 ctx_flush_and_put(ctx, &locked);
2104 /* if not contended, grab and improve batching */
2105 locked = mutex_trylock(&ctx->uring_lock);
2106 percpu_ref_get(&ctx->refs);
2108 req->io_task_work.func(req, &locked);
2115 ctx_flush_and_put(ctx, &locked);
2118 static void io_req_task_work_add(struct io_kiocb *req)
2120 struct task_struct *tsk = req->task;
2121 struct io_uring_task *tctx = tsk->io_uring;
2122 enum task_work_notify_mode notify;
2123 struct io_wq_work_node *node;
2124 unsigned long flags;
2127 WARN_ON_ONCE(!tctx);
2129 spin_lock_irqsave(&tctx->task_lock, flags);
2130 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2131 running = tctx->task_running;
2133 tctx->task_running = true;
2134 spin_unlock_irqrestore(&tctx->task_lock, flags);
2136 /* task_work already pending, we're done */
2141 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2142 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2143 * processing task_work. There's no reliable way to tell if TWA_RESUME
2146 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2147 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2148 wake_up_process(tsk);
2152 spin_lock_irqsave(&tctx->task_lock, flags);
2153 tctx->task_running = false;
2154 node = tctx->task_list.first;
2155 INIT_WQ_LIST(&tctx->task_list);
2156 spin_unlock_irqrestore(&tctx->task_lock, flags);
2159 req = container_of(node, struct io_kiocb, io_task_work.node);
2161 if (llist_add(&req->io_task_work.fallback_node,
2162 &req->ctx->fallback_llist))
2163 schedule_delayed_work(&req->ctx->fallback_work, 1);
2167 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2169 struct io_ring_ctx *ctx = req->ctx;
2171 /* not needed for normal modes, but SQPOLL depends on it */
2172 io_tw_lock(ctx, locked);
2173 io_req_complete_failed(req, req->result);
2176 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2178 struct io_ring_ctx *ctx = req->ctx;
2180 io_tw_lock(ctx, locked);
2181 /* req->task == current here, checking PF_EXITING is safe */
2182 if (likely(!(req->task->flags & PF_EXITING)))
2183 __io_queue_sqe(req);
2185 io_req_complete_failed(req, -EFAULT);
2188 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2191 req->io_task_work.func = io_req_task_cancel;
2192 io_req_task_work_add(req);
2195 static void io_req_task_queue(struct io_kiocb *req)
2197 req->io_task_work.func = io_req_task_submit;
2198 io_req_task_work_add(req);
2201 static void io_req_task_queue_reissue(struct io_kiocb *req)
2203 req->io_task_work.func = io_queue_async_work;
2204 io_req_task_work_add(req);
2207 static inline void io_queue_next(struct io_kiocb *req)
2209 struct io_kiocb *nxt = io_req_find_next(req);
2212 io_req_task_queue(nxt);
2215 static void io_free_req(struct io_kiocb *req)
2221 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2227 struct task_struct *task;
2232 static inline void io_init_req_batch(struct req_batch *rb)
2239 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2240 struct req_batch *rb)
2243 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2245 io_put_task(rb->task, rb->task_refs);
2248 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2249 struct io_submit_state *state)
2252 io_dismantle_req(req);
2254 if (req->task != rb->task) {
2256 io_put_task(rb->task, rb->task_refs);
2257 rb->task = req->task;
2263 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2264 state->reqs[state->free_reqs++] = req;
2266 list_add(&req->inflight_entry, &state->free_list);
2269 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2270 __must_hold(&ctx->uring_lock)
2272 struct io_submit_state *state = &ctx->submit_state;
2273 int i, nr = state->compl_nr;
2274 struct req_batch rb;
2276 spin_lock(&ctx->completion_lock);
2277 for (i = 0; i < nr; i++) {
2278 struct io_kiocb *req = state->compl_reqs[i];
2280 __io_cqring_fill_event(ctx, req->user_data, req->result,
2283 io_commit_cqring(ctx);
2284 spin_unlock(&ctx->completion_lock);
2285 io_cqring_ev_posted(ctx);
2287 io_init_req_batch(&rb);
2288 for (i = 0; i < nr; i++) {
2289 struct io_kiocb *req = state->compl_reqs[i];
2291 if (req_ref_put_and_test(req))
2292 io_req_free_batch(&rb, req, &ctx->submit_state);
2295 io_req_free_batch_finish(ctx, &rb);
2296 state->compl_nr = 0;
2300 * Drop reference to request, return next in chain (if there is one) if this
2301 * was the last reference to this request.
2303 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2305 struct io_kiocb *nxt = NULL;
2307 if (req_ref_put_and_test(req)) {
2308 nxt = io_req_find_next(req);
2314 static inline void io_put_req(struct io_kiocb *req)
2316 if (req_ref_put_and_test(req))
2320 static inline void io_put_req_deferred(struct io_kiocb *req)
2322 if (req_ref_put_and_test(req)) {
2323 req->io_task_work.func = io_free_req_work;
2324 io_req_task_work_add(req);
2328 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2330 /* See comment at the top of this file */
2332 return __io_cqring_events(ctx);
2335 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2337 struct io_rings *rings = ctx->rings;
2339 /* make sure SQ entry isn't read before tail */
2340 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2343 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2345 unsigned int cflags;
2347 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2348 cflags |= IORING_CQE_F_BUFFER;
2349 req->flags &= ~REQ_F_BUFFER_SELECTED;
2354 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2356 struct io_buffer *kbuf;
2358 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2360 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2361 return io_put_kbuf(req, kbuf);
2364 static inline bool io_run_task_work(void)
2366 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2367 __set_current_state(TASK_RUNNING);
2368 tracehook_notify_signal();
2376 * Find and free completed poll iocbs
2378 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2379 struct list_head *done)
2381 struct req_batch rb;
2382 struct io_kiocb *req;
2384 /* order with ->result store in io_complete_rw_iopoll() */
2387 io_init_req_batch(&rb);
2388 while (!list_empty(done)) {
2389 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2390 list_del(&req->inflight_entry);
2392 if (READ_ONCE(req->result) == -EAGAIN &&
2393 !(req->flags & REQ_F_DONT_REISSUE)) {
2394 req->iopoll_completed = 0;
2395 io_req_task_queue_reissue(req);
2399 __io_cqring_fill_event(ctx, req->user_data, req->result,
2400 io_put_rw_kbuf(req));
2403 if (req_ref_put_and_test(req))
2404 io_req_free_batch(&rb, req, &ctx->submit_state);
2407 io_commit_cqring(ctx);
2408 io_cqring_ev_posted_iopoll(ctx);
2409 io_req_free_batch_finish(ctx, &rb);
2412 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2415 struct io_kiocb *req, *tmp;
2420 * Only spin for completions if we don't have multiple devices hanging
2421 * off our complete list, and we're under the requested amount.
2423 spin = !ctx->poll_multi_queue && *nr_events < min;
2425 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2426 struct kiocb *kiocb = &req->rw.kiocb;
2430 * Move completed and retryable entries to our local lists.
2431 * If we find a request that requires polling, break out
2432 * and complete those lists first, if we have entries there.
2434 if (READ_ONCE(req->iopoll_completed)) {
2435 list_move_tail(&req->inflight_entry, &done);
2438 if (!list_empty(&done))
2441 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2442 if (unlikely(ret < 0))
2447 /* iopoll may have completed current req */
2448 if (READ_ONCE(req->iopoll_completed))
2449 list_move_tail(&req->inflight_entry, &done);
2452 if (!list_empty(&done))
2453 io_iopoll_complete(ctx, nr_events, &done);
2459 * We can't just wait for polled events to come to us, we have to actively
2460 * find and complete them.
2462 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2464 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2467 mutex_lock(&ctx->uring_lock);
2468 while (!list_empty(&ctx->iopoll_list)) {
2469 unsigned int nr_events = 0;
2471 io_do_iopoll(ctx, &nr_events, 0);
2473 /* let it sleep and repeat later if can't complete a request */
2477 * Ensure we allow local-to-the-cpu processing to take place,
2478 * in this case we need to ensure that we reap all events.
2479 * Also let task_work, etc. to progress by releasing the mutex
2481 if (need_resched()) {
2482 mutex_unlock(&ctx->uring_lock);
2484 mutex_lock(&ctx->uring_lock);
2487 mutex_unlock(&ctx->uring_lock);
2490 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2492 unsigned int nr_events = 0;
2496 * We disallow the app entering submit/complete with polling, but we
2497 * still need to lock the ring to prevent racing with polled issue
2498 * that got punted to a workqueue.
2500 mutex_lock(&ctx->uring_lock);
2502 * Don't enter poll loop if we already have events pending.
2503 * If we do, we can potentially be spinning for commands that
2504 * already triggered a CQE (eg in error).
2506 if (test_bit(0, &ctx->check_cq_overflow))
2507 __io_cqring_overflow_flush(ctx, false);
2508 if (io_cqring_events(ctx))
2512 * If a submit got punted to a workqueue, we can have the
2513 * application entering polling for a command before it gets
2514 * issued. That app will hold the uring_lock for the duration
2515 * of the poll right here, so we need to take a breather every
2516 * now and then to ensure that the issue has a chance to add
2517 * the poll to the issued list. Otherwise we can spin here
2518 * forever, while the workqueue is stuck trying to acquire the
2521 if (list_empty(&ctx->iopoll_list)) {
2522 u32 tail = ctx->cached_cq_tail;
2524 mutex_unlock(&ctx->uring_lock);
2526 mutex_lock(&ctx->uring_lock);
2528 /* some requests don't go through iopoll_list */
2529 if (tail != ctx->cached_cq_tail ||
2530 list_empty(&ctx->iopoll_list))
2533 ret = io_do_iopoll(ctx, &nr_events, min);
2534 } while (!ret && nr_events < min && !need_resched());
2536 mutex_unlock(&ctx->uring_lock);
2540 static void kiocb_end_write(struct io_kiocb *req)
2543 * Tell lockdep we inherited freeze protection from submission
2546 if (req->flags & REQ_F_ISREG) {
2547 struct super_block *sb = file_inode(req->file)->i_sb;
2549 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2555 static bool io_resubmit_prep(struct io_kiocb *req)
2557 struct io_async_rw *rw = req->async_data;
2560 return !io_req_prep_async(req);
2561 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2562 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2566 static bool io_rw_should_reissue(struct io_kiocb *req)
2568 umode_t mode = file_inode(req->file)->i_mode;
2569 struct io_ring_ctx *ctx = req->ctx;
2571 if (!S_ISBLK(mode) && !S_ISREG(mode))
2573 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2574 !(ctx->flags & IORING_SETUP_IOPOLL)))
2577 * If ref is dying, we might be running poll reap from the exit work.
2578 * Don't attempt to reissue from that path, just let it fail with
2581 if (percpu_ref_is_dying(&ctx->refs))
2584 * Play it safe and assume not safe to re-import and reissue if we're
2585 * not in the original thread group (or in task context).
2587 if (!same_thread_group(req->task, current) || !in_task())
2592 static bool io_resubmit_prep(struct io_kiocb *req)
2596 static bool io_rw_should_reissue(struct io_kiocb *req)
2602 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2604 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2605 kiocb_end_write(req);
2606 if (res != req->result) {
2607 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2608 io_rw_should_reissue(req)) {
2609 req->flags |= REQ_F_REISSUE;
2618 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2620 unsigned int cflags = io_put_rw_kbuf(req);
2621 long res = req->result;
2624 struct io_ring_ctx *ctx = req->ctx;
2625 struct io_submit_state *state = &ctx->submit_state;
2627 io_req_complete_state(req, res, cflags);
2628 state->compl_reqs[state->compl_nr++] = req;
2629 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2630 io_submit_flush_completions(ctx);
2632 io_req_complete_post(req, res, cflags);
2636 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2637 unsigned int issue_flags)
2639 if (__io_complete_rw_common(req, res))
2641 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2644 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2646 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2648 if (__io_complete_rw_common(req, res))
2651 req->io_task_work.func = io_req_task_complete;
2652 io_req_task_work_add(req);
2655 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2657 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2659 if (kiocb->ki_flags & IOCB_WRITE)
2660 kiocb_end_write(req);
2661 if (unlikely(res != req->result)) {
2662 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2663 io_resubmit_prep(req))) {
2665 req->flags |= REQ_F_DONT_REISSUE;
2669 WRITE_ONCE(req->result, res);
2670 /* order with io_iopoll_complete() checking ->result */
2672 WRITE_ONCE(req->iopoll_completed, 1);
2676 * After the iocb has been issued, it's safe to be found on the poll list.
2677 * Adding the kiocb to the list AFTER submission ensures that we don't
2678 * find it from a io_do_iopoll() thread before the issuer is done
2679 * accessing the kiocb cookie.
2681 static void io_iopoll_req_issued(struct io_kiocb *req)
2683 struct io_ring_ctx *ctx = req->ctx;
2684 const bool in_async = io_wq_current_is_worker();
2686 /* workqueue context doesn't hold uring_lock, grab it now */
2687 if (unlikely(in_async))
2688 mutex_lock(&ctx->uring_lock);
2691 * Track whether we have multiple files in our lists. This will impact
2692 * how we do polling eventually, not spinning if we're on potentially
2693 * different devices.
2695 if (list_empty(&ctx->iopoll_list)) {
2696 ctx->poll_multi_queue = false;
2697 } else if (!ctx->poll_multi_queue) {
2698 struct io_kiocb *list_req;
2699 unsigned int queue_num0, queue_num1;
2701 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2704 if (list_req->file != req->file) {
2705 ctx->poll_multi_queue = true;
2707 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2708 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2709 if (queue_num0 != queue_num1)
2710 ctx->poll_multi_queue = true;
2715 * For fast devices, IO may have already completed. If it has, add
2716 * it to the front so we find it first.
2718 if (READ_ONCE(req->iopoll_completed))
2719 list_add(&req->inflight_entry, &ctx->iopoll_list);
2721 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2723 if (unlikely(in_async)) {
2725 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2726 * in sq thread task context or in io worker task context. If
2727 * current task context is sq thread, we don't need to check
2728 * whether should wake up sq thread.
2730 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2731 wq_has_sleeper(&ctx->sq_data->wait))
2732 wake_up(&ctx->sq_data->wait);
2734 mutex_unlock(&ctx->uring_lock);
2738 static bool io_bdev_nowait(struct block_device *bdev)
2740 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2744 * If we tracked the file through the SCM inflight mechanism, we could support
2745 * any file. For now, just ensure that anything potentially problematic is done
2748 static bool __io_file_supports_nowait(struct file *file, int rw)
2750 umode_t mode = file_inode(file)->i_mode;
2752 if (S_ISBLK(mode)) {
2753 if (IS_ENABLED(CONFIG_BLOCK) &&
2754 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2760 if (S_ISREG(mode)) {
2761 if (IS_ENABLED(CONFIG_BLOCK) &&
2762 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2763 file->f_op != &io_uring_fops)
2768 /* any ->read/write should understand O_NONBLOCK */
2769 if (file->f_flags & O_NONBLOCK)
2772 if (!(file->f_mode & FMODE_NOWAIT))
2776 return file->f_op->read_iter != NULL;
2778 return file->f_op->write_iter != NULL;
2781 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2783 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2785 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2788 return __io_file_supports_nowait(req->file, rw);
2791 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2793 struct io_ring_ctx *ctx = req->ctx;
2794 struct kiocb *kiocb = &req->rw.kiocb;
2795 struct file *file = req->file;
2799 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2800 req->flags |= REQ_F_ISREG;
2802 kiocb->ki_pos = READ_ONCE(sqe->off);
2803 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2804 req->flags |= REQ_F_CUR_POS;
2805 kiocb->ki_pos = file->f_pos;
2807 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2808 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2809 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2813 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2814 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2815 req->flags |= REQ_F_NOWAIT;
2817 ioprio = READ_ONCE(sqe->ioprio);
2819 ret = ioprio_check_cap(ioprio);
2823 kiocb->ki_ioprio = ioprio;
2825 kiocb->ki_ioprio = get_current_ioprio();
2827 if (ctx->flags & IORING_SETUP_IOPOLL) {
2828 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2829 !kiocb->ki_filp->f_op->iopoll)
2832 kiocb->ki_flags |= IOCB_HIPRI;
2833 kiocb->ki_complete = io_complete_rw_iopoll;
2834 req->iopoll_completed = 0;
2836 if (kiocb->ki_flags & IOCB_HIPRI)
2838 kiocb->ki_complete = io_complete_rw;
2841 if (req->opcode == IORING_OP_READ_FIXED ||
2842 req->opcode == IORING_OP_WRITE_FIXED) {
2844 io_req_set_rsrc_node(req);
2847 req->rw.addr = READ_ONCE(sqe->addr);
2848 req->rw.len = READ_ONCE(sqe->len);
2849 req->buf_index = READ_ONCE(sqe->buf_index);
2853 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2859 case -ERESTARTNOINTR:
2860 case -ERESTARTNOHAND:
2861 case -ERESTART_RESTARTBLOCK:
2863 * We can't just restart the syscall, since previously
2864 * submitted sqes may already be in progress. Just fail this
2870 kiocb->ki_complete(kiocb, ret, 0);
2874 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2875 unsigned int issue_flags)
2877 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2878 struct io_async_rw *io = req->async_data;
2879 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2881 /* add previously done IO, if any */
2882 if (io && io->bytes_done > 0) {
2884 ret = io->bytes_done;
2886 ret += io->bytes_done;
2889 if (req->flags & REQ_F_CUR_POS)
2890 req->file->f_pos = kiocb->ki_pos;
2891 if (ret >= 0 && check_reissue)
2892 __io_complete_rw(req, ret, 0, issue_flags);
2894 io_rw_done(kiocb, ret);
2896 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2897 req->flags &= ~REQ_F_REISSUE;
2898 if (io_resubmit_prep(req)) {
2899 io_req_task_queue_reissue(req);
2902 __io_req_complete(req, issue_flags, ret,
2903 io_put_rw_kbuf(req));
2908 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2909 struct io_mapped_ubuf *imu)
2911 size_t len = req->rw.len;
2912 u64 buf_end, buf_addr = req->rw.addr;
2915 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2917 /* not inside the mapped region */
2918 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2922 * May not be a start of buffer, set size appropriately
2923 * and advance us to the beginning.
2925 offset = buf_addr - imu->ubuf;
2926 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2930 * Don't use iov_iter_advance() here, as it's really slow for
2931 * using the latter parts of a big fixed buffer - it iterates
2932 * over each segment manually. We can cheat a bit here, because
2935 * 1) it's a BVEC iter, we set it up
2936 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2937 * first and last bvec
2939 * So just find our index, and adjust the iterator afterwards.
2940 * If the offset is within the first bvec (or the whole first
2941 * bvec, just use iov_iter_advance(). This makes it easier
2942 * since we can just skip the first segment, which may not
2943 * be PAGE_SIZE aligned.
2945 const struct bio_vec *bvec = imu->bvec;
2947 if (offset <= bvec->bv_len) {
2948 iov_iter_advance(iter, offset);
2950 unsigned long seg_skip;
2952 /* skip first vec */
2953 offset -= bvec->bv_len;
2954 seg_skip = 1 + (offset >> PAGE_SHIFT);
2956 iter->bvec = bvec + seg_skip;
2957 iter->nr_segs -= seg_skip;
2958 iter->count -= bvec->bv_len + offset;
2959 iter->iov_offset = offset & ~PAGE_MASK;
2966 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2968 struct io_ring_ctx *ctx = req->ctx;
2969 struct io_mapped_ubuf *imu = req->imu;
2970 u16 index, buf_index = req->buf_index;
2973 if (unlikely(buf_index >= ctx->nr_user_bufs))
2975 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2976 imu = READ_ONCE(ctx->user_bufs[index]);
2979 return __io_import_fixed(req, rw, iter, imu);
2982 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2985 mutex_unlock(&ctx->uring_lock);
2988 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2991 * "Normal" inline submissions always hold the uring_lock, since we
2992 * grab it from the system call. Same is true for the SQPOLL offload.
2993 * The only exception is when we've detached the request and issue it
2994 * from an async worker thread, grab the lock for that case.
2997 mutex_lock(&ctx->uring_lock);
3000 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3001 int bgid, struct io_buffer *kbuf,
3004 struct io_buffer *head;
3006 if (req->flags & REQ_F_BUFFER_SELECTED)
3009 io_ring_submit_lock(req->ctx, needs_lock);
3011 lockdep_assert_held(&req->ctx->uring_lock);
3013 head = xa_load(&req->ctx->io_buffers, bgid);
3015 if (!list_empty(&head->list)) {
3016 kbuf = list_last_entry(&head->list, struct io_buffer,
3018 list_del(&kbuf->list);
3021 xa_erase(&req->ctx->io_buffers, bgid);
3023 if (*len > kbuf->len)
3026 kbuf = ERR_PTR(-ENOBUFS);
3029 io_ring_submit_unlock(req->ctx, needs_lock);
3034 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3037 struct io_buffer *kbuf;
3040 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3041 bgid = req->buf_index;
3042 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3045 req->rw.addr = (u64) (unsigned long) kbuf;
3046 req->flags |= REQ_F_BUFFER_SELECTED;
3047 return u64_to_user_ptr(kbuf->addr);
3050 #ifdef CONFIG_COMPAT
3051 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3054 struct compat_iovec __user *uiov;
3055 compat_ssize_t clen;
3059 uiov = u64_to_user_ptr(req->rw.addr);
3060 if (!access_ok(uiov, sizeof(*uiov)))
3062 if (__get_user(clen, &uiov->iov_len))
3068 buf = io_rw_buffer_select(req, &len, needs_lock);
3070 return PTR_ERR(buf);
3071 iov[0].iov_base = buf;
3072 iov[0].iov_len = (compat_size_t) len;
3077 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3080 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3084 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3087 len = iov[0].iov_len;
3090 buf = io_rw_buffer_select(req, &len, needs_lock);
3092 return PTR_ERR(buf);
3093 iov[0].iov_base = buf;
3094 iov[0].iov_len = len;
3098 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3101 if (req->flags & REQ_F_BUFFER_SELECTED) {
3102 struct io_buffer *kbuf;
3104 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3105 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3106 iov[0].iov_len = kbuf->len;
3109 if (req->rw.len != 1)
3112 #ifdef CONFIG_COMPAT
3113 if (req->ctx->compat)
3114 return io_compat_import(req, iov, needs_lock);
3117 return __io_iov_buffer_select(req, iov, needs_lock);
3120 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3121 struct iov_iter *iter, bool needs_lock)
3123 void __user *buf = u64_to_user_ptr(req->rw.addr);
3124 size_t sqe_len = req->rw.len;
3125 u8 opcode = req->opcode;
3128 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3130 return io_import_fixed(req, rw, iter);
3133 /* buffer index only valid with fixed read/write, or buffer select */
3134 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3137 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3138 if (req->flags & REQ_F_BUFFER_SELECT) {
3139 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3141 return PTR_ERR(buf);
3142 req->rw.len = sqe_len;
3145 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3150 if (req->flags & REQ_F_BUFFER_SELECT) {
3151 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3153 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3158 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3162 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3164 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3168 * For files that don't have ->read_iter() and ->write_iter(), handle them
3169 * by looping over ->read() or ->write() manually.
3171 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3173 struct kiocb *kiocb = &req->rw.kiocb;
3174 struct file *file = req->file;
3178 * Don't support polled IO through this interface, and we can't
3179 * support non-blocking either. For the latter, this just causes
3180 * the kiocb to be handled from an async context.
3182 if (kiocb->ki_flags & IOCB_HIPRI)
3184 if (kiocb->ki_flags & IOCB_NOWAIT)
3187 while (iov_iter_count(iter)) {
3191 if (!iov_iter_is_bvec(iter)) {
3192 iovec = iov_iter_iovec(iter);
3194 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3195 iovec.iov_len = req->rw.len;
3199 nr = file->f_op->read(file, iovec.iov_base,
3200 iovec.iov_len, io_kiocb_ppos(kiocb));
3202 nr = file->f_op->write(file, iovec.iov_base,
3203 iovec.iov_len, io_kiocb_ppos(kiocb));
3212 if (nr != iovec.iov_len)
3216 iov_iter_advance(iter, nr);
3222 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3223 const struct iovec *fast_iov, struct iov_iter *iter)
3225 struct io_async_rw *rw = req->async_data;
3227 memcpy(&rw->iter, iter, sizeof(*iter));
3228 rw->free_iovec = iovec;
3230 /* can only be fixed buffers, no need to do anything */
3231 if (iov_iter_is_bvec(iter))
3234 unsigned iov_off = 0;
3236 rw->iter.iov = rw->fast_iov;
3237 if (iter->iov != fast_iov) {
3238 iov_off = iter->iov - fast_iov;
3239 rw->iter.iov += iov_off;
3241 if (rw->fast_iov != fast_iov)
3242 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3243 sizeof(struct iovec) * iter->nr_segs);
3245 req->flags |= REQ_F_NEED_CLEANUP;
3249 static inline int io_alloc_async_data(struct io_kiocb *req)
3251 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3252 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3253 return req->async_data == NULL;
3256 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3257 const struct iovec *fast_iov,
3258 struct iov_iter *iter, bool force)
3260 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3262 if (!req->async_data) {
3263 if (io_alloc_async_data(req)) {
3268 io_req_map_rw(req, iovec, fast_iov, iter);
3273 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3275 struct io_async_rw *iorw = req->async_data;
3276 struct iovec *iov = iorw->fast_iov;
3279 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3280 if (unlikely(ret < 0))
3283 iorw->bytes_done = 0;
3284 iorw->free_iovec = iov;
3286 req->flags |= REQ_F_NEED_CLEANUP;
3290 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3292 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3294 return io_prep_rw(req, sqe);
3298 * This is our waitqueue callback handler, registered through lock_page_async()
3299 * when we initially tried to do the IO with the iocb armed our waitqueue.
3300 * This gets called when the page is unlocked, and we generally expect that to
3301 * happen when the page IO is completed and the page is now uptodate. This will
3302 * queue a task_work based retry of the operation, attempting to copy the data
3303 * again. If the latter fails because the page was NOT uptodate, then we will
3304 * do a thread based blocking retry of the operation. That's the unexpected
3307 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3308 int sync, void *arg)
3310 struct wait_page_queue *wpq;
3311 struct io_kiocb *req = wait->private;
3312 struct wait_page_key *key = arg;
3314 wpq = container_of(wait, struct wait_page_queue, wait);
3316 if (!wake_page_match(wpq, key))
3319 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3320 list_del_init(&wait->entry);
3321 io_req_task_queue(req);
3326 * This controls whether a given IO request should be armed for async page
3327 * based retry. If we return false here, the request is handed to the async
3328 * worker threads for retry. If we're doing buffered reads on a regular file,
3329 * we prepare a private wait_page_queue entry and retry the operation. This
3330 * will either succeed because the page is now uptodate and unlocked, or it
3331 * will register a callback when the page is unlocked at IO completion. Through
3332 * that callback, io_uring uses task_work to setup a retry of the operation.
3333 * That retry will attempt the buffered read again. The retry will generally
3334 * succeed, or in rare cases where it fails, we then fall back to using the
3335 * async worker threads for a blocking retry.
3337 static bool io_rw_should_retry(struct io_kiocb *req)
3339 struct io_async_rw *rw = req->async_data;
3340 struct wait_page_queue *wait = &rw->wpq;
3341 struct kiocb *kiocb = &req->rw.kiocb;
3343 /* never retry for NOWAIT, we just complete with -EAGAIN */
3344 if (req->flags & REQ_F_NOWAIT)
3347 /* Only for buffered IO */
3348 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3352 * just use poll if we can, and don't attempt if the fs doesn't
3353 * support callback based unlocks
3355 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3358 wait->wait.func = io_async_buf_func;
3359 wait->wait.private = req;
3360 wait->wait.flags = 0;
3361 INIT_LIST_HEAD(&wait->wait.entry);
3362 kiocb->ki_flags |= IOCB_WAITQ;
3363 kiocb->ki_flags &= ~IOCB_NOWAIT;
3364 kiocb->ki_waitq = wait;
3368 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3370 if (req->file->f_op->read_iter)
3371 return call_read_iter(req->file, &req->rw.kiocb, iter);
3372 else if (req->file->f_op->read)
3373 return loop_rw_iter(READ, req, iter);
3378 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3380 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3381 struct kiocb *kiocb = &req->rw.kiocb;
3382 struct iov_iter __iter, *iter = &__iter;
3383 struct io_async_rw *rw = req->async_data;
3384 ssize_t io_size, ret, ret2;
3385 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3391 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3395 io_size = iov_iter_count(iter);
3396 req->result = io_size;
3398 /* Ensure we clear previously set non-block flag */
3399 if (!force_nonblock)
3400 kiocb->ki_flags &= ~IOCB_NOWAIT;
3402 kiocb->ki_flags |= IOCB_NOWAIT;
3404 /* If the file doesn't support async, just async punt */
3405 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3406 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3407 return ret ?: -EAGAIN;
3410 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3411 if (unlikely(ret)) {
3416 ret = io_iter_do_read(req, iter);
3418 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3419 req->flags &= ~REQ_F_REISSUE;
3420 /* IOPOLL retry should happen for io-wq threads */
3421 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3423 /* no retry on NONBLOCK nor RWF_NOWAIT */
3424 if (req->flags & REQ_F_NOWAIT)
3426 /* some cases will consume bytes even on error returns */
3427 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3429 } else if (ret == -EIOCBQUEUED) {
3431 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3432 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3433 /* read all, failed, already did sync or don't want to retry */
3437 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3442 rw = req->async_data;
3443 /* now use our persistent iterator, if we aren't already */
3448 rw->bytes_done += ret;
3449 /* if we can retry, do so with the callbacks armed */
3450 if (!io_rw_should_retry(req)) {
3451 kiocb->ki_flags &= ~IOCB_WAITQ;
3456 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3457 * we get -EIOCBQUEUED, then we'll get a notification when the
3458 * desired page gets unlocked. We can also get a partial read
3459 * here, and if we do, then just retry at the new offset.
3461 ret = io_iter_do_read(req, iter);
3462 if (ret == -EIOCBQUEUED)
3464 /* we got some bytes, but not all. retry. */
3465 kiocb->ki_flags &= ~IOCB_WAITQ;
3466 } while (ret > 0 && ret < io_size);
3468 kiocb_done(kiocb, ret, issue_flags);
3470 /* it's faster to check here then delegate to kfree */
3476 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3478 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3480 return io_prep_rw(req, sqe);
3483 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3485 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3486 struct kiocb *kiocb = &req->rw.kiocb;
3487 struct iov_iter __iter, *iter = &__iter;
3488 struct io_async_rw *rw = req->async_data;
3489 ssize_t ret, ret2, io_size;
3490 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3496 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3500 io_size = iov_iter_count(iter);
3501 req->result = io_size;
3503 /* Ensure we clear previously set non-block flag */
3504 if (!force_nonblock)
3505 kiocb->ki_flags &= ~IOCB_NOWAIT;
3507 kiocb->ki_flags |= IOCB_NOWAIT;
3509 /* If the file doesn't support async, just async punt */
3510 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3513 /* file path doesn't support NOWAIT for non-direct_IO */
3514 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3515 (req->flags & REQ_F_ISREG))
3518 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3523 * Open-code file_start_write here to grab freeze protection,
3524 * which will be released by another thread in
3525 * io_complete_rw(). Fool lockdep by telling it the lock got
3526 * released so that it doesn't complain about the held lock when
3527 * we return to userspace.
3529 if (req->flags & REQ_F_ISREG) {
3530 sb_start_write(file_inode(req->file)->i_sb);
3531 __sb_writers_release(file_inode(req->file)->i_sb,
3534 kiocb->ki_flags |= IOCB_WRITE;
3536 if (req->file->f_op->write_iter)
3537 ret2 = call_write_iter(req->file, kiocb, iter);
3538 else if (req->file->f_op->write)
3539 ret2 = loop_rw_iter(WRITE, req, iter);
3543 if (req->flags & REQ_F_REISSUE) {
3544 req->flags &= ~REQ_F_REISSUE;
3549 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3550 * retry them without IOCB_NOWAIT.
3552 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3554 /* no retry on NONBLOCK nor RWF_NOWAIT */
3555 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3557 if (!force_nonblock || ret2 != -EAGAIN) {
3558 /* IOPOLL retry should happen for io-wq threads */
3559 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3562 kiocb_done(kiocb, ret2, issue_flags);
3565 /* some cases will consume bytes even on error returns */
3566 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3567 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3568 return ret ?: -EAGAIN;
3571 /* it's reportedly faster than delegating the null check to kfree() */
3577 static int io_renameat_prep(struct io_kiocb *req,
3578 const struct io_uring_sqe *sqe)
3580 struct io_rename *ren = &req->rename;
3581 const char __user *oldf, *newf;
3583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3585 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3587 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3590 ren->old_dfd = READ_ONCE(sqe->fd);
3591 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3592 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3593 ren->new_dfd = READ_ONCE(sqe->len);
3594 ren->flags = READ_ONCE(sqe->rename_flags);
3596 ren->oldpath = getname(oldf);
3597 if (IS_ERR(ren->oldpath))
3598 return PTR_ERR(ren->oldpath);
3600 ren->newpath = getname(newf);
3601 if (IS_ERR(ren->newpath)) {
3602 putname(ren->oldpath);
3603 return PTR_ERR(ren->newpath);
3606 req->flags |= REQ_F_NEED_CLEANUP;
3610 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3612 struct io_rename *ren = &req->rename;
3615 if (issue_flags & IO_URING_F_NONBLOCK)
3618 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3619 ren->newpath, ren->flags);
3621 req->flags &= ~REQ_F_NEED_CLEANUP;
3624 io_req_complete(req, ret);
3628 static int io_unlinkat_prep(struct io_kiocb *req,
3629 const struct io_uring_sqe *sqe)
3631 struct io_unlink *un = &req->unlink;
3632 const char __user *fname;
3634 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3636 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3639 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3642 un->dfd = READ_ONCE(sqe->fd);
3644 un->flags = READ_ONCE(sqe->unlink_flags);
3645 if (un->flags & ~AT_REMOVEDIR)
3648 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3649 un->filename = getname(fname);
3650 if (IS_ERR(un->filename))
3651 return PTR_ERR(un->filename);
3653 req->flags |= REQ_F_NEED_CLEANUP;
3657 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3659 struct io_unlink *un = &req->unlink;
3662 if (issue_flags & IO_URING_F_NONBLOCK)
3665 if (un->flags & AT_REMOVEDIR)
3666 ret = do_rmdir(un->dfd, un->filename);
3668 ret = do_unlinkat(un->dfd, un->filename);
3670 req->flags &= ~REQ_F_NEED_CLEANUP;
3673 io_req_complete(req, ret);
3677 static int io_shutdown_prep(struct io_kiocb *req,
3678 const struct io_uring_sqe *sqe)
3680 #if defined(CONFIG_NET)
3681 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3683 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3684 sqe->buf_index || sqe->splice_fd_in))
3687 req->shutdown.how = READ_ONCE(sqe->len);
3694 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3696 #if defined(CONFIG_NET)
3697 struct socket *sock;
3700 if (issue_flags & IO_URING_F_NONBLOCK)
3703 sock = sock_from_file(req->file);
3704 if (unlikely(!sock))
3707 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3710 io_req_complete(req, ret);
3717 static int __io_splice_prep(struct io_kiocb *req,
3718 const struct io_uring_sqe *sqe)
3720 struct io_splice *sp = &req->splice;
3721 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3727 sp->len = READ_ONCE(sqe->len);
3728 sp->flags = READ_ONCE(sqe->splice_flags);
3730 if (unlikely(sp->flags & ~valid_flags))
3733 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3734 (sp->flags & SPLICE_F_FD_IN_FIXED));
3737 req->flags |= REQ_F_NEED_CLEANUP;
3741 static int io_tee_prep(struct io_kiocb *req,
3742 const struct io_uring_sqe *sqe)
3744 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3746 return __io_splice_prep(req, sqe);
3749 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3751 struct io_splice *sp = &req->splice;
3752 struct file *in = sp->file_in;
3753 struct file *out = sp->file_out;
3754 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3757 if (issue_flags & IO_URING_F_NONBLOCK)
3760 ret = do_tee(in, out, sp->len, flags);
3762 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3764 req->flags &= ~REQ_F_NEED_CLEANUP;
3768 io_req_complete(req, ret);
3772 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3774 struct io_splice *sp = &req->splice;
3776 sp->off_in = READ_ONCE(sqe->splice_off_in);
3777 sp->off_out = READ_ONCE(sqe->off);
3778 return __io_splice_prep(req, sqe);
3781 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3783 struct io_splice *sp = &req->splice;
3784 struct file *in = sp->file_in;
3785 struct file *out = sp->file_out;
3786 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3787 loff_t *poff_in, *poff_out;
3790 if (issue_flags & IO_URING_F_NONBLOCK)
3793 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3794 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3797 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3799 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3801 req->flags &= ~REQ_F_NEED_CLEANUP;
3805 io_req_complete(req, ret);
3810 * IORING_OP_NOP just posts a completion event, nothing else.
3812 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3814 struct io_ring_ctx *ctx = req->ctx;
3816 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3819 __io_req_complete(req, issue_flags, 0, 0);
3823 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3825 struct io_ring_ctx *ctx = req->ctx;
3830 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3832 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3836 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3837 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3840 req->sync.off = READ_ONCE(sqe->off);
3841 req->sync.len = READ_ONCE(sqe->len);
3845 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3847 loff_t end = req->sync.off + req->sync.len;
3850 /* fsync always requires a blocking context */
3851 if (issue_flags & IO_URING_F_NONBLOCK)
3854 ret = vfs_fsync_range(req->file, req->sync.off,
3855 end > 0 ? end : LLONG_MAX,
3856 req->sync.flags & IORING_FSYNC_DATASYNC);
3859 io_req_complete(req, ret);
3863 static int io_fallocate_prep(struct io_kiocb *req,
3864 const struct io_uring_sqe *sqe)
3866 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3869 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3872 req->sync.off = READ_ONCE(sqe->off);
3873 req->sync.len = READ_ONCE(sqe->addr);
3874 req->sync.mode = READ_ONCE(sqe->len);
3878 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3882 /* fallocate always requiring blocking context */
3883 if (issue_flags & IO_URING_F_NONBLOCK)
3885 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3889 io_req_complete(req, ret);
3893 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3895 const char __user *fname;
3898 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3900 if (unlikely(sqe->ioprio || sqe->buf_index))
3902 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3905 /* open.how should be already initialised */
3906 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3907 req->open.how.flags |= O_LARGEFILE;
3909 req->open.dfd = READ_ONCE(sqe->fd);
3910 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3911 req->open.filename = getname(fname);
3912 if (IS_ERR(req->open.filename)) {
3913 ret = PTR_ERR(req->open.filename);
3914 req->open.filename = NULL;
3918 req->open.file_slot = READ_ONCE(sqe->file_index);
3919 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
3922 req->open.nofile = rlimit(RLIMIT_NOFILE);
3923 req->flags |= REQ_F_NEED_CLEANUP;
3927 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3929 u64 mode = READ_ONCE(sqe->len);
3930 u64 flags = READ_ONCE(sqe->open_flags);
3932 req->open.how = build_open_how(flags, mode);
3933 return __io_openat_prep(req, sqe);
3936 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3938 struct open_how __user *how;
3942 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3943 len = READ_ONCE(sqe->len);
3944 if (len < OPEN_HOW_SIZE_VER0)
3947 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3952 return __io_openat_prep(req, sqe);
3955 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3957 struct open_flags op;
3959 bool resolve_nonblock, nonblock_set;
3960 bool fixed = !!req->open.file_slot;
3963 ret = build_open_flags(&req->open.how, &op);
3966 nonblock_set = op.open_flag & O_NONBLOCK;
3967 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3968 if (issue_flags & IO_URING_F_NONBLOCK) {
3970 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3971 * it'll always -EAGAIN
3973 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3975 op.lookup_flags |= LOOKUP_CACHED;
3976 op.open_flag |= O_NONBLOCK;
3980 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3985 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3988 * We could hang on to this 'fd' on retrying, but seems like
3989 * marginal gain for something that is now known to be a slower
3990 * path. So just put it, and we'll get a new one when we retry.
3995 ret = PTR_ERR(file);
3996 /* only retry if RESOLVE_CACHED wasn't already set by application */
3997 if (ret == -EAGAIN &&
3998 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4003 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4004 file->f_flags &= ~O_NONBLOCK;
4005 fsnotify_open(file);
4008 fd_install(ret, file);
4010 ret = io_install_fixed_file(req, file, issue_flags,
4011 req->open.file_slot - 1);
4013 putname(req->open.filename);
4014 req->flags &= ~REQ_F_NEED_CLEANUP;
4017 __io_req_complete(req, issue_flags, ret, 0);
4021 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4023 return io_openat2(req, issue_flags);
4026 static int io_remove_buffers_prep(struct io_kiocb *req,
4027 const struct io_uring_sqe *sqe)
4029 struct io_provide_buf *p = &req->pbuf;
4032 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4036 tmp = READ_ONCE(sqe->fd);
4037 if (!tmp || tmp > USHRT_MAX)
4040 memset(p, 0, sizeof(*p));
4042 p->bgid = READ_ONCE(sqe->buf_group);
4046 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4047 int bgid, unsigned nbufs)
4051 /* shouldn't happen */
4055 /* the head kbuf is the list itself */
4056 while (!list_empty(&buf->list)) {
4057 struct io_buffer *nxt;
4059 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4060 list_del(&nxt->list);
4067 xa_erase(&ctx->io_buffers, bgid);
4072 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4074 struct io_provide_buf *p = &req->pbuf;
4075 struct io_ring_ctx *ctx = req->ctx;
4076 struct io_buffer *head;
4078 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4080 io_ring_submit_lock(ctx, !force_nonblock);
4082 lockdep_assert_held(&ctx->uring_lock);
4085 head = xa_load(&ctx->io_buffers, p->bgid);
4087 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4091 /* complete before unlock, IOPOLL may need the lock */
4092 __io_req_complete(req, issue_flags, ret, 0);
4093 io_ring_submit_unlock(ctx, !force_nonblock);
4097 static int io_provide_buffers_prep(struct io_kiocb *req,
4098 const struct io_uring_sqe *sqe)
4100 unsigned long size, tmp_check;
4101 struct io_provide_buf *p = &req->pbuf;
4104 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4107 tmp = READ_ONCE(sqe->fd);
4108 if (!tmp || tmp > USHRT_MAX)
4111 p->addr = READ_ONCE(sqe->addr);
4112 p->len = READ_ONCE(sqe->len);
4114 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4117 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4120 size = (unsigned long)p->len * p->nbufs;
4121 if (!access_ok(u64_to_user_ptr(p->addr), size))
4124 p->bgid = READ_ONCE(sqe->buf_group);
4125 tmp = READ_ONCE(sqe->off);
4126 if (tmp > USHRT_MAX)
4132 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4134 struct io_buffer *buf;
4135 u64 addr = pbuf->addr;
4136 int i, bid = pbuf->bid;
4138 for (i = 0; i < pbuf->nbufs; i++) {
4139 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4144 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4149 INIT_LIST_HEAD(&buf->list);
4152 list_add_tail(&buf->list, &(*head)->list);
4156 return i ? i : -ENOMEM;
4159 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4161 struct io_provide_buf *p = &req->pbuf;
4162 struct io_ring_ctx *ctx = req->ctx;
4163 struct io_buffer *head, *list;
4165 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4167 io_ring_submit_lock(ctx, !force_nonblock);
4169 lockdep_assert_held(&ctx->uring_lock);
4171 list = head = xa_load(&ctx->io_buffers, p->bgid);
4173 ret = io_add_buffers(p, &head);
4174 if (ret >= 0 && !list) {
4175 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4177 __io_remove_buffers(ctx, head, p->bgid, -1U);
4181 /* complete before unlock, IOPOLL may need the lock */
4182 __io_req_complete(req, issue_flags, ret, 0);
4183 io_ring_submit_unlock(ctx, !force_nonblock);
4187 static int io_epoll_ctl_prep(struct io_kiocb *req,
4188 const struct io_uring_sqe *sqe)
4190 #if defined(CONFIG_EPOLL)
4191 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4193 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4196 req->epoll.epfd = READ_ONCE(sqe->fd);
4197 req->epoll.op = READ_ONCE(sqe->len);
4198 req->epoll.fd = READ_ONCE(sqe->off);
4200 if (ep_op_has_event(req->epoll.op)) {
4201 struct epoll_event __user *ev;
4203 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4204 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4214 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4216 #if defined(CONFIG_EPOLL)
4217 struct io_epoll *ie = &req->epoll;
4219 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4221 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4222 if (force_nonblock && ret == -EAGAIN)
4227 __io_req_complete(req, issue_flags, ret, 0);
4234 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4236 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4237 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4239 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4242 req->madvise.addr = READ_ONCE(sqe->addr);
4243 req->madvise.len = READ_ONCE(sqe->len);
4244 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4251 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4253 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4254 struct io_madvise *ma = &req->madvise;
4257 if (issue_flags & IO_URING_F_NONBLOCK)
4260 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4263 io_req_complete(req, ret);
4270 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4272 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4274 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4277 req->fadvise.offset = READ_ONCE(sqe->off);
4278 req->fadvise.len = READ_ONCE(sqe->len);
4279 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4283 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4285 struct io_fadvise *fa = &req->fadvise;
4288 if (issue_flags & IO_URING_F_NONBLOCK) {
4289 switch (fa->advice) {
4290 case POSIX_FADV_NORMAL:
4291 case POSIX_FADV_RANDOM:
4292 case POSIX_FADV_SEQUENTIAL:
4299 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4302 __io_req_complete(req, issue_flags, ret, 0);
4306 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4308 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4310 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4312 if (req->flags & REQ_F_FIXED_FILE)
4315 req->statx.dfd = READ_ONCE(sqe->fd);
4316 req->statx.mask = READ_ONCE(sqe->len);
4317 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4318 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4319 req->statx.flags = READ_ONCE(sqe->statx_flags);
4324 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4326 struct io_statx *ctx = &req->statx;
4329 if (issue_flags & IO_URING_F_NONBLOCK)
4332 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4337 io_req_complete(req, ret);
4341 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4343 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4345 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4346 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4348 if (req->flags & REQ_F_FIXED_FILE)
4351 req->close.fd = READ_ONCE(sqe->fd);
4355 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4357 struct files_struct *files = current->files;
4358 struct io_close *close = &req->close;
4359 struct fdtable *fdt;
4360 struct file *file = NULL;
4363 spin_lock(&files->file_lock);
4364 fdt = files_fdtable(files);
4365 if (close->fd >= fdt->max_fds) {
4366 spin_unlock(&files->file_lock);
4369 file = fdt->fd[close->fd];
4370 if (!file || file->f_op == &io_uring_fops) {
4371 spin_unlock(&files->file_lock);
4376 /* if the file has a flush method, be safe and punt to async */
4377 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4378 spin_unlock(&files->file_lock);
4382 ret = __close_fd_get_file(close->fd, &file);
4383 spin_unlock(&files->file_lock);
4390 /* No ->flush() or already async, safely close from here */
4391 ret = filp_close(file, current->files);
4397 __io_req_complete(req, issue_flags, ret, 0);
4401 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4403 struct io_ring_ctx *ctx = req->ctx;
4405 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4407 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4411 req->sync.off = READ_ONCE(sqe->off);
4412 req->sync.len = READ_ONCE(sqe->len);
4413 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4417 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4421 /* sync_file_range always requires a blocking context */
4422 if (issue_flags & IO_URING_F_NONBLOCK)
4425 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4429 io_req_complete(req, ret);
4433 #if defined(CONFIG_NET)
4434 static int io_setup_async_msg(struct io_kiocb *req,
4435 struct io_async_msghdr *kmsg)
4437 struct io_async_msghdr *async_msg = req->async_data;
4441 if (io_alloc_async_data(req)) {
4442 kfree(kmsg->free_iov);
4445 async_msg = req->async_data;
4446 req->flags |= REQ_F_NEED_CLEANUP;
4447 memcpy(async_msg, kmsg, sizeof(*kmsg));
4448 async_msg->msg.msg_name = &async_msg->addr;
4449 /* if were using fast_iov, set it to the new one */
4450 if (!async_msg->free_iov)
4451 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4456 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4457 struct io_async_msghdr *iomsg)
4459 iomsg->msg.msg_name = &iomsg->addr;
4460 iomsg->free_iov = iomsg->fast_iov;
4461 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4462 req->sr_msg.msg_flags, &iomsg->free_iov);
4465 static int io_sendmsg_prep_async(struct io_kiocb *req)
4469 ret = io_sendmsg_copy_hdr(req, req->async_data);
4471 req->flags |= REQ_F_NEED_CLEANUP;
4475 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4477 struct io_sr_msg *sr = &req->sr_msg;
4479 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4482 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4483 sr->len = READ_ONCE(sqe->len);
4484 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4485 if (sr->msg_flags & MSG_DONTWAIT)
4486 req->flags |= REQ_F_NOWAIT;
4488 #ifdef CONFIG_COMPAT
4489 if (req->ctx->compat)
4490 sr->msg_flags |= MSG_CMSG_COMPAT;
4495 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4497 struct io_async_msghdr iomsg, *kmsg;
4498 struct socket *sock;
4503 sock = sock_from_file(req->file);
4504 if (unlikely(!sock))
4507 kmsg = req->async_data;
4509 ret = io_sendmsg_copy_hdr(req, &iomsg);
4515 flags = req->sr_msg.msg_flags;
4516 if (issue_flags & IO_URING_F_NONBLOCK)
4517 flags |= MSG_DONTWAIT;
4518 if (flags & MSG_WAITALL)
4519 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4521 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4522 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4523 return io_setup_async_msg(req, kmsg);
4524 if (ret == -ERESTARTSYS)
4527 /* fast path, check for non-NULL to avoid function call */
4529 kfree(kmsg->free_iov);
4530 req->flags &= ~REQ_F_NEED_CLEANUP;
4533 __io_req_complete(req, issue_flags, ret, 0);
4537 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4539 struct io_sr_msg *sr = &req->sr_msg;
4542 struct socket *sock;
4547 sock = sock_from_file(req->file);
4548 if (unlikely(!sock))
4551 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4555 msg.msg_name = NULL;
4556 msg.msg_control = NULL;
4557 msg.msg_controllen = 0;
4558 msg.msg_namelen = 0;
4560 flags = req->sr_msg.msg_flags;
4561 if (issue_flags & IO_URING_F_NONBLOCK)
4562 flags |= MSG_DONTWAIT;
4563 if (flags & MSG_WAITALL)
4564 min_ret = iov_iter_count(&msg.msg_iter);
4566 msg.msg_flags = flags;
4567 ret = sock_sendmsg(sock, &msg);
4568 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4570 if (ret == -ERESTARTSYS)
4575 __io_req_complete(req, issue_flags, ret, 0);
4579 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4580 struct io_async_msghdr *iomsg)
4582 struct io_sr_msg *sr = &req->sr_msg;
4583 struct iovec __user *uiov;
4587 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4588 &iomsg->uaddr, &uiov, &iov_len);
4592 if (req->flags & REQ_F_BUFFER_SELECT) {
4595 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4597 sr->len = iomsg->fast_iov[0].iov_len;
4598 iomsg->free_iov = NULL;
4600 iomsg->free_iov = iomsg->fast_iov;
4601 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4602 &iomsg->free_iov, &iomsg->msg.msg_iter,
4611 #ifdef CONFIG_COMPAT
4612 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4613 struct io_async_msghdr *iomsg)
4615 struct io_sr_msg *sr = &req->sr_msg;
4616 struct compat_iovec __user *uiov;
4621 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4626 uiov = compat_ptr(ptr);
4627 if (req->flags & REQ_F_BUFFER_SELECT) {
4628 compat_ssize_t clen;
4632 if (!access_ok(uiov, sizeof(*uiov)))
4634 if (__get_user(clen, &uiov->iov_len))
4639 iomsg->free_iov = NULL;
4641 iomsg->free_iov = iomsg->fast_iov;
4642 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4643 UIO_FASTIOV, &iomsg->free_iov,
4644 &iomsg->msg.msg_iter, true);
4653 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4654 struct io_async_msghdr *iomsg)
4656 iomsg->msg.msg_name = &iomsg->addr;
4658 #ifdef CONFIG_COMPAT
4659 if (req->ctx->compat)
4660 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4663 return __io_recvmsg_copy_hdr(req, iomsg);
4666 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4669 struct io_sr_msg *sr = &req->sr_msg;
4670 struct io_buffer *kbuf;
4672 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4677 req->flags |= REQ_F_BUFFER_SELECTED;
4681 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4683 return io_put_kbuf(req, req->sr_msg.kbuf);
4686 static int io_recvmsg_prep_async(struct io_kiocb *req)
4690 ret = io_recvmsg_copy_hdr(req, req->async_data);
4692 req->flags |= REQ_F_NEED_CLEANUP;
4696 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4698 struct io_sr_msg *sr = &req->sr_msg;
4700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4703 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4704 sr->len = READ_ONCE(sqe->len);
4705 sr->bgid = READ_ONCE(sqe->buf_group);
4706 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4707 if (sr->msg_flags & MSG_DONTWAIT)
4708 req->flags |= REQ_F_NOWAIT;
4710 #ifdef CONFIG_COMPAT
4711 if (req->ctx->compat)
4712 sr->msg_flags |= MSG_CMSG_COMPAT;
4717 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4719 struct io_async_msghdr iomsg, *kmsg;
4720 struct socket *sock;
4721 struct io_buffer *kbuf;
4724 int ret, cflags = 0;
4725 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4727 sock = sock_from_file(req->file);
4728 if (unlikely(!sock))
4731 kmsg = req->async_data;
4733 ret = io_recvmsg_copy_hdr(req, &iomsg);
4739 if (req->flags & REQ_F_BUFFER_SELECT) {
4740 kbuf = io_recv_buffer_select(req, !force_nonblock);
4742 return PTR_ERR(kbuf);
4743 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4744 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4745 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4746 1, req->sr_msg.len);
4749 flags = req->sr_msg.msg_flags;
4751 flags |= MSG_DONTWAIT;
4752 if (flags & MSG_WAITALL)
4753 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4755 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4756 kmsg->uaddr, flags);
4757 if (force_nonblock && ret == -EAGAIN)
4758 return io_setup_async_msg(req, kmsg);
4759 if (ret == -ERESTARTSYS)
4762 if (req->flags & REQ_F_BUFFER_SELECTED)
4763 cflags = io_put_recv_kbuf(req);
4764 /* fast path, check for non-NULL to avoid function call */
4766 kfree(kmsg->free_iov);
4767 req->flags &= ~REQ_F_NEED_CLEANUP;
4768 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4770 __io_req_complete(req, issue_flags, ret, cflags);
4774 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4776 struct io_buffer *kbuf;
4777 struct io_sr_msg *sr = &req->sr_msg;
4779 void __user *buf = sr->buf;
4780 struct socket *sock;
4784 int ret, cflags = 0;
4785 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4787 sock = sock_from_file(req->file);
4788 if (unlikely(!sock))
4791 if (req->flags & REQ_F_BUFFER_SELECT) {
4792 kbuf = io_recv_buffer_select(req, !force_nonblock);
4794 return PTR_ERR(kbuf);
4795 buf = u64_to_user_ptr(kbuf->addr);
4798 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4802 msg.msg_name = NULL;
4803 msg.msg_control = NULL;
4804 msg.msg_controllen = 0;
4805 msg.msg_namelen = 0;
4806 msg.msg_iocb = NULL;
4809 flags = req->sr_msg.msg_flags;
4811 flags |= MSG_DONTWAIT;
4812 if (flags & MSG_WAITALL)
4813 min_ret = iov_iter_count(&msg.msg_iter);
4815 ret = sock_recvmsg(sock, &msg, flags);
4816 if (force_nonblock && ret == -EAGAIN)
4818 if (ret == -ERESTARTSYS)
4821 if (req->flags & REQ_F_BUFFER_SELECTED)
4822 cflags = io_put_recv_kbuf(req);
4823 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4825 __io_req_complete(req, issue_flags, ret, cflags);
4829 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4831 struct io_accept *accept = &req->accept;
4833 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4835 if (sqe->ioprio || sqe->len || sqe->buf_index)
4838 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4839 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4840 accept->flags = READ_ONCE(sqe->accept_flags);
4841 accept->nofile = rlimit(RLIMIT_NOFILE);
4843 accept->file_slot = READ_ONCE(sqe->file_index);
4844 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
4845 (accept->flags & SOCK_CLOEXEC)))
4847 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
4849 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
4850 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
4854 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4856 struct io_accept *accept = &req->accept;
4857 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4858 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4859 bool fixed = !!accept->file_slot;
4863 if (req->file->f_flags & O_NONBLOCK)
4864 req->flags |= REQ_F_NOWAIT;
4867 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
4868 if (unlikely(fd < 0))
4871 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
4876 ret = PTR_ERR(file);
4877 if (ret == -EAGAIN && force_nonblock)
4879 if (ret == -ERESTARTSYS)
4882 } else if (!fixed) {
4883 fd_install(fd, file);
4886 ret = io_install_fixed_file(req, file, issue_flags,
4887 accept->file_slot - 1);
4889 __io_req_complete(req, issue_flags, ret, 0);
4893 static int io_connect_prep_async(struct io_kiocb *req)
4895 struct io_async_connect *io = req->async_data;
4896 struct io_connect *conn = &req->connect;
4898 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4901 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4903 struct io_connect *conn = &req->connect;
4905 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4907 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4911 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4912 conn->addr_len = READ_ONCE(sqe->addr2);
4916 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4918 struct io_async_connect __io, *io;
4919 unsigned file_flags;
4921 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4923 if (req->async_data) {
4924 io = req->async_data;
4926 ret = move_addr_to_kernel(req->connect.addr,
4927 req->connect.addr_len,
4934 file_flags = force_nonblock ? O_NONBLOCK : 0;
4936 ret = __sys_connect_file(req->file, &io->address,
4937 req->connect.addr_len, file_flags);
4938 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4939 if (req->async_data)
4941 if (io_alloc_async_data(req)) {
4945 memcpy(req->async_data, &__io, sizeof(__io));
4948 if (ret == -ERESTARTSYS)
4953 __io_req_complete(req, issue_flags, ret, 0);
4956 #else /* !CONFIG_NET */
4957 #define IO_NETOP_FN(op) \
4958 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4960 return -EOPNOTSUPP; \
4963 #define IO_NETOP_PREP(op) \
4965 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4967 return -EOPNOTSUPP; \
4970 #define IO_NETOP_PREP_ASYNC(op) \
4972 static int io_##op##_prep_async(struct io_kiocb *req) \
4974 return -EOPNOTSUPP; \
4977 IO_NETOP_PREP_ASYNC(sendmsg);
4978 IO_NETOP_PREP_ASYNC(recvmsg);
4979 IO_NETOP_PREP_ASYNC(connect);
4980 IO_NETOP_PREP(accept);
4983 #endif /* CONFIG_NET */
4985 struct io_poll_table {
4986 struct poll_table_struct pt;
4987 struct io_kiocb *req;
4992 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4993 __poll_t mask, io_req_tw_func_t func)
4995 /* for instances that support it check for an event match first: */
4996 if (mask && !(mask & poll->events))
4999 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5001 list_del_init(&poll->wait.entry);
5004 req->io_task_work.func = func;
5007 * If this fails, then the task is exiting. When a task exits, the
5008 * work gets canceled, so just cancel this request as well instead
5009 * of executing it. We can't safely execute it anyway, as we may not
5010 * have the needed state needed for it anyway.
5012 io_req_task_work_add(req);
5016 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5017 __acquires(&req->ctx->completion_lock)
5019 struct io_ring_ctx *ctx = req->ctx;
5021 /* req->task == current here, checking PF_EXITING is safe */
5022 if (unlikely(req->task->flags & PF_EXITING))
5023 WRITE_ONCE(poll->canceled, true);
5025 if (!req->result && !READ_ONCE(poll->canceled)) {
5026 struct poll_table_struct pt = { ._key = poll->events };
5028 req->result = vfs_poll(req->file, &pt) & poll->events;
5031 spin_lock(&ctx->completion_lock);
5032 if (!req->result && !READ_ONCE(poll->canceled)) {
5033 add_wait_queue(poll->head, &poll->wait);
5040 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5042 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5043 if (req->opcode == IORING_OP_POLL_ADD)
5044 return req->async_data;
5045 return req->apoll->double_poll;
5048 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5050 if (req->opcode == IORING_OP_POLL_ADD)
5052 return &req->apoll->poll;
5055 static void io_poll_remove_double(struct io_kiocb *req)
5056 __must_hold(&req->ctx->completion_lock)
5058 struct io_poll_iocb *poll = io_poll_get_double(req);
5060 lockdep_assert_held(&req->ctx->completion_lock);
5062 if (poll && poll->head) {
5063 struct wait_queue_head *head = poll->head;
5065 spin_lock_irq(&head->lock);
5066 list_del_init(&poll->wait.entry);
5067 if (poll->wait.private)
5070 spin_unlock_irq(&head->lock);
5074 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5075 __must_hold(&req->ctx->completion_lock)
5077 struct io_ring_ctx *ctx = req->ctx;
5078 unsigned flags = IORING_CQE_F_MORE;
5081 if (READ_ONCE(req->poll.canceled)) {
5083 req->poll.events |= EPOLLONESHOT;
5085 error = mangle_poll(mask);
5087 if (req->poll.events & EPOLLONESHOT)
5089 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5090 req->poll.done = true;
5093 if (flags & IORING_CQE_F_MORE)
5096 io_commit_cqring(ctx);
5097 return !(flags & IORING_CQE_F_MORE);
5100 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5102 struct io_ring_ctx *ctx = req->ctx;
5103 struct io_kiocb *nxt;
5105 if (io_poll_rewait(req, &req->poll)) {
5106 spin_unlock(&ctx->completion_lock);
5110 done = io_poll_complete(req, req->result);
5112 io_poll_remove_double(req);
5113 hash_del(&req->hash_node);
5116 add_wait_queue(req->poll.head, &req->poll.wait);
5118 spin_unlock(&ctx->completion_lock);
5119 io_cqring_ev_posted(ctx);
5122 nxt = io_put_req_find_next(req);
5124 io_req_task_submit(nxt, locked);
5129 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5130 int sync, void *key)
5132 struct io_kiocb *req = wait->private;
5133 struct io_poll_iocb *poll = io_poll_get_single(req);
5134 __poll_t mask = key_to_poll(key);
5135 unsigned long flags;
5137 /* for instances that support it check for an event match first: */
5138 if (mask && !(mask & poll->events))
5140 if (!(poll->events & EPOLLONESHOT))
5141 return poll->wait.func(&poll->wait, mode, sync, key);
5143 list_del_init(&wait->entry);
5148 spin_lock_irqsave(&poll->head->lock, flags);
5149 done = list_empty(&poll->wait.entry);
5151 list_del_init(&poll->wait.entry);
5152 /* make sure double remove sees this as being gone */
5153 wait->private = NULL;
5154 spin_unlock_irqrestore(&poll->head->lock, flags);
5156 /* use wait func handler, so it matches the rq type */
5157 poll->wait.func(&poll->wait, mode, sync, key);
5164 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5165 wait_queue_func_t wake_func)
5169 poll->canceled = false;
5170 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5171 /* mask in events that we always want/need */
5172 poll->events = events | IO_POLL_UNMASK;
5173 INIT_LIST_HEAD(&poll->wait.entry);
5174 init_waitqueue_func_entry(&poll->wait, wake_func);
5177 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5178 struct wait_queue_head *head,
5179 struct io_poll_iocb **poll_ptr)
5181 struct io_kiocb *req = pt->req;
5184 * The file being polled uses multiple waitqueues for poll handling
5185 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5188 if (unlikely(pt->nr_entries)) {
5189 struct io_poll_iocb *poll_one = poll;
5191 /* double add on the same waitqueue head, ignore */
5192 if (poll_one->head == head)
5194 /* already have a 2nd entry, fail a third attempt */
5196 if ((*poll_ptr)->head == head)
5198 pt->error = -EINVAL;
5202 * Can't handle multishot for double wait for now, turn it
5203 * into one-shot mode.
5205 if (!(poll_one->events & EPOLLONESHOT))
5206 poll_one->events |= EPOLLONESHOT;
5207 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5209 pt->error = -ENOMEM;
5212 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5214 poll->wait.private = req;
5221 if (poll->events & EPOLLEXCLUSIVE)
5222 add_wait_queue_exclusive(head, &poll->wait);
5224 add_wait_queue(head, &poll->wait);
5227 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5228 struct poll_table_struct *p)
5230 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5231 struct async_poll *apoll = pt->req->apoll;
5233 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5236 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5238 struct async_poll *apoll = req->apoll;
5239 struct io_ring_ctx *ctx = req->ctx;
5241 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5243 if (io_poll_rewait(req, &apoll->poll)) {
5244 spin_unlock(&ctx->completion_lock);
5248 hash_del(&req->hash_node);
5249 io_poll_remove_double(req);
5250 spin_unlock(&ctx->completion_lock);
5252 if (!READ_ONCE(apoll->poll.canceled))
5253 io_req_task_submit(req, locked);
5255 io_req_complete_failed(req, -ECANCELED);
5258 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5261 struct io_kiocb *req = wait->private;
5262 struct io_poll_iocb *poll = &req->apoll->poll;
5264 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5267 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5270 static void io_poll_req_insert(struct io_kiocb *req)
5272 struct io_ring_ctx *ctx = req->ctx;
5273 struct hlist_head *list;
5275 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5276 hlist_add_head(&req->hash_node, list);
5279 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5280 struct io_poll_iocb *poll,
5281 struct io_poll_table *ipt, __poll_t mask,
5282 wait_queue_func_t wake_func)
5283 __acquires(&ctx->completion_lock)
5285 struct io_ring_ctx *ctx = req->ctx;
5286 bool cancel = false;
5288 INIT_HLIST_NODE(&req->hash_node);
5289 io_init_poll_iocb(poll, mask, wake_func);
5290 poll->file = req->file;
5291 poll->wait.private = req;
5293 ipt->pt._key = mask;
5296 ipt->nr_entries = 0;
5298 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5299 if (unlikely(!ipt->nr_entries) && !ipt->error)
5300 ipt->error = -EINVAL;
5302 spin_lock(&ctx->completion_lock);
5303 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5304 io_poll_remove_double(req);
5305 if (likely(poll->head)) {
5306 spin_lock_irq(&poll->head->lock);
5307 if (unlikely(list_empty(&poll->wait.entry))) {
5313 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5314 list_del_init(&poll->wait.entry);
5316 WRITE_ONCE(poll->canceled, true);
5317 else if (!poll->done) /* actually waiting for an event */
5318 io_poll_req_insert(req);
5319 spin_unlock_irq(&poll->head->lock);
5331 static int io_arm_poll_handler(struct io_kiocb *req)
5333 const struct io_op_def *def = &io_op_defs[req->opcode];
5334 struct io_ring_ctx *ctx = req->ctx;
5335 struct async_poll *apoll;
5336 struct io_poll_table ipt;
5337 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5340 if (!req->file || !file_can_poll(req->file))
5341 return IO_APOLL_ABORTED;
5342 if (req->flags & REQ_F_POLLED)
5343 return IO_APOLL_ABORTED;
5344 if (!def->pollin && !def->pollout)
5345 return IO_APOLL_ABORTED;
5349 mask |= POLLIN | POLLRDNORM;
5351 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5352 if ((req->opcode == IORING_OP_RECVMSG) &&
5353 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5357 mask |= POLLOUT | POLLWRNORM;
5360 /* if we can't nonblock try, then no point in arming a poll handler */
5361 if (!io_file_supports_nowait(req, rw))
5362 return IO_APOLL_ABORTED;
5364 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5365 if (unlikely(!apoll))
5366 return IO_APOLL_ABORTED;
5367 apoll->double_poll = NULL;
5369 req->flags |= REQ_F_POLLED;
5370 ipt.pt._qproc = io_async_queue_proc;
5371 io_req_set_refcount(req);
5373 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5375 spin_unlock(&ctx->completion_lock);
5376 if (ret || ipt.error)
5377 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5379 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5380 mask, apoll->poll.events);
5384 static bool __io_poll_remove_one(struct io_kiocb *req,
5385 struct io_poll_iocb *poll, bool do_cancel)
5386 __must_hold(&req->ctx->completion_lock)
5388 bool do_complete = false;
5392 spin_lock_irq(&poll->head->lock);
5394 WRITE_ONCE(poll->canceled, true);
5395 if (!list_empty(&poll->wait.entry)) {
5396 list_del_init(&poll->wait.entry);
5399 spin_unlock_irq(&poll->head->lock);
5400 hash_del(&req->hash_node);
5404 static bool io_poll_remove_one(struct io_kiocb *req)
5405 __must_hold(&req->ctx->completion_lock)
5409 io_poll_remove_double(req);
5410 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5413 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5414 io_commit_cqring(req->ctx);
5416 io_put_req_deferred(req);
5422 * Returns true if we found and killed one or more poll requests
5424 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5427 struct hlist_node *tmp;
5428 struct io_kiocb *req;
5431 spin_lock(&ctx->completion_lock);
5432 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5433 struct hlist_head *list;
5435 list = &ctx->cancel_hash[i];
5436 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5437 if (io_match_task(req, tsk, cancel_all))
5438 posted += io_poll_remove_one(req);
5441 spin_unlock(&ctx->completion_lock);
5444 io_cqring_ev_posted(ctx);
5449 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5451 __must_hold(&ctx->completion_lock)
5453 struct hlist_head *list;
5454 struct io_kiocb *req;
5456 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5457 hlist_for_each_entry(req, list, hash_node) {
5458 if (sqe_addr != req->user_data)
5460 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5467 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5469 __must_hold(&ctx->completion_lock)
5471 struct io_kiocb *req;
5473 req = io_poll_find(ctx, sqe_addr, poll_only);
5476 if (io_poll_remove_one(req))
5482 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5487 events = READ_ONCE(sqe->poll32_events);
5489 events = swahw32(events);
5491 if (!(flags & IORING_POLL_ADD_MULTI))
5492 events |= EPOLLONESHOT;
5493 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5496 static int io_poll_update_prep(struct io_kiocb *req,
5497 const struct io_uring_sqe *sqe)
5499 struct io_poll_update *upd = &req->poll_update;
5502 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5504 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5506 flags = READ_ONCE(sqe->len);
5507 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5508 IORING_POLL_ADD_MULTI))
5510 /* meaningless without update */
5511 if (flags == IORING_POLL_ADD_MULTI)
5514 upd->old_user_data = READ_ONCE(sqe->addr);
5515 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5516 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5518 upd->new_user_data = READ_ONCE(sqe->off);
5519 if (!upd->update_user_data && upd->new_user_data)
5521 if (upd->update_events)
5522 upd->events = io_poll_parse_events(sqe, flags);
5523 else if (sqe->poll32_events)
5529 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5532 struct io_kiocb *req = wait->private;
5533 struct io_poll_iocb *poll = &req->poll;
5535 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5538 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5539 struct poll_table_struct *p)
5541 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5543 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5546 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5548 struct io_poll_iocb *poll = &req->poll;
5551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5553 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5555 flags = READ_ONCE(sqe->len);
5556 if (flags & ~IORING_POLL_ADD_MULTI)
5559 io_req_set_refcount(req);
5560 poll->events = io_poll_parse_events(sqe, flags);
5564 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5566 struct io_poll_iocb *poll = &req->poll;
5567 struct io_ring_ctx *ctx = req->ctx;
5568 struct io_poll_table ipt;
5571 ipt.pt._qproc = io_poll_queue_proc;
5573 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5576 if (mask) { /* no async, we'd stolen it */
5578 io_poll_complete(req, mask);
5580 spin_unlock(&ctx->completion_lock);
5583 io_cqring_ev_posted(ctx);
5584 if (poll->events & EPOLLONESHOT)
5590 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5592 struct io_ring_ctx *ctx = req->ctx;
5593 struct io_kiocb *preq;
5597 spin_lock(&ctx->completion_lock);
5598 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5604 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5606 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5611 * Don't allow racy completion with singleshot, as we cannot safely
5612 * update those. For multishot, if we're racing with completion, just
5613 * let completion re-add it.
5615 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5616 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5620 /* we now have a detached poll request. reissue. */
5624 spin_unlock(&ctx->completion_lock);
5626 io_req_complete(req, ret);
5629 /* only mask one event flags, keep behavior flags */
5630 if (req->poll_update.update_events) {
5631 preq->poll.events &= ~0xffff;
5632 preq->poll.events |= req->poll_update.events & 0xffff;
5633 preq->poll.events |= IO_POLL_UNMASK;
5635 if (req->poll_update.update_user_data)
5636 preq->user_data = req->poll_update.new_user_data;
5637 spin_unlock(&ctx->completion_lock);
5639 /* complete update request, we're done with it */
5640 io_req_complete(req, ret);
5643 ret = io_poll_add(preq, issue_flags);
5646 io_req_complete(preq, ret);
5652 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5655 io_req_complete_post(req, -ETIME, 0);
5658 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5660 struct io_timeout_data *data = container_of(timer,
5661 struct io_timeout_data, timer);
5662 struct io_kiocb *req = data->req;
5663 struct io_ring_ctx *ctx = req->ctx;
5664 unsigned long flags;
5666 spin_lock_irqsave(&ctx->timeout_lock, flags);
5667 list_del_init(&req->timeout.list);
5668 atomic_set(&req->ctx->cq_timeouts,
5669 atomic_read(&req->ctx->cq_timeouts) + 1);
5670 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5672 req->io_task_work.func = io_req_task_timeout;
5673 io_req_task_work_add(req);
5674 return HRTIMER_NORESTART;
5677 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5679 __must_hold(&ctx->timeout_lock)
5681 struct io_timeout_data *io;
5682 struct io_kiocb *req;
5685 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5686 found = user_data == req->user_data;
5691 return ERR_PTR(-ENOENT);
5693 io = req->async_data;
5694 if (hrtimer_try_to_cancel(&io->timer) == -1)
5695 return ERR_PTR(-EALREADY);
5696 list_del_init(&req->timeout.list);
5700 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5701 __must_hold(&ctx->completion_lock)
5702 __must_hold(&ctx->timeout_lock)
5704 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5707 return PTR_ERR(req);
5710 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5711 io_put_req_deferred(req);
5715 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5716 struct timespec64 *ts, enum hrtimer_mode mode)
5717 __must_hold(&ctx->timeout_lock)
5719 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5720 struct io_timeout_data *data;
5723 return PTR_ERR(req);
5725 req->timeout.off = 0; /* noseq */
5726 data = req->async_data;
5727 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5728 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5729 data->timer.function = io_timeout_fn;
5730 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5734 static int io_timeout_remove_prep(struct io_kiocb *req,
5735 const struct io_uring_sqe *sqe)
5737 struct io_timeout_rem *tr = &req->timeout_rem;
5739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5741 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5743 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5746 tr->addr = READ_ONCE(sqe->addr);
5747 tr->flags = READ_ONCE(sqe->timeout_flags);
5748 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5749 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5751 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5753 } else if (tr->flags) {
5754 /* timeout removal doesn't support flags */
5761 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5763 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5768 * Remove or update an existing timeout command
5770 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5772 struct io_timeout_rem *tr = &req->timeout_rem;
5773 struct io_ring_ctx *ctx = req->ctx;
5776 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5777 spin_lock(&ctx->completion_lock);
5778 spin_lock_irq(&ctx->timeout_lock);
5779 ret = io_timeout_cancel(ctx, tr->addr);
5780 spin_unlock_irq(&ctx->timeout_lock);
5781 spin_unlock(&ctx->completion_lock);
5783 spin_lock_irq(&ctx->timeout_lock);
5784 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5785 io_translate_timeout_mode(tr->flags));
5786 spin_unlock_irq(&ctx->timeout_lock);
5791 io_req_complete_post(req, ret, 0);
5795 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5796 bool is_timeout_link)
5798 struct io_timeout_data *data;
5800 u32 off = READ_ONCE(sqe->off);
5802 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5804 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5807 if (off && is_timeout_link)
5809 flags = READ_ONCE(sqe->timeout_flags);
5810 if (flags & ~IORING_TIMEOUT_ABS)
5813 req->timeout.off = off;
5814 if (unlikely(off && !req->ctx->off_timeout_used))
5815 req->ctx->off_timeout_used = true;
5817 if (!req->async_data && io_alloc_async_data(req))
5820 data = req->async_data;
5823 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5826 data->mode = io_translate_timeout_mode(flags);
5827 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5829 if (is_timeout_link) {
5830 struct io_submit_link *link = &req->ctx->submit_state.link;
5834 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5836 req->timeout.head = link->last;
5837 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5842 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5844 struct io_ring_ctx *ctx = req->ctx;
5845 struct io_timeout_data *data = req->async_data;
5846 struct list_head *entry;
5847 u32 tail, off = req->timeout.off;
5849 spin_lock_irq(&ctx->timeout_lock);
5852 * sqe->off holds how many events that need to occur for this
5853 * timeout event to be satisfied. If it isn't set, then this is
5854 * a pure timeout request, sequence isn't used.
5856 if (io_is_timeout_noseq(req)) {
5857 entry = ctx->timeout_list.prev;
5861 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5862 req->timeout.target_seq = tail + off;
5864 /* Update the last seq here in case io_flush_timeouts() hasn't.
5865 * This is safe because ->completion_lock is held, and submissions
5866 * and completions are never mixed in the same ->completion_lock section.
5868 ctx->cq_last_tm_flush = tail;
5871 * Insertion sort, ensuring the first entry in the list is always
5872 * the one we need first.
5874 list_for_each_prev(entry, &ctx->timeout_list) {
5875 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5878 if (io_is_timeout_noseq(nxt))
5880 /* nxt.seq is behind @tail, otherwise would've been completed */
5881 if (off >= nxt->timeout.target_seq - tail)
5885 list_add(&req->timeout.list, entry);
5886 data->timer.function = io_timeout_fn;
5887 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5888 spin_unlock_irq(&ctx->timeout_lock);
5892 struct io_cancel_data {
5893 struct io_ring_ctx *ctx;
5897 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5899 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5900 struct io_cancel_data *cd = data;
5902 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5905 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5906 struct io_ring_ctx *ctx)
5908 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5909 enum io_wq_cancel cancel_ret;
5912 if (!tctx || !tctx->io_wq)
5915 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5916 switch (cancel_ret) {
5917 case IO_WQ_CANCEL_OK:
5920 case IO_WQ_CANCEL_RUNNING:
5923 case IO_WQ_CANCEL_NOTFOUND:
5931 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5933 struct io_ring_ctx *ctx = req->ctx;
5936 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
5938 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5942 spin_lock(&ctx->completion_lock);
5943 spin_lock_irq(&ctx->timeout_lock);
5944 ret = io_timeout_cancel(ctx, sqe_addr);
5945 spin_unlock_irq(&ctx->timeout_lock);
5948 ret = io_poll_cancel(ctx, sqe_addr, false);
5950 spin_unlock(&ctx->completion_lock);
5954 static int io_async_cancel_prep(struct io_kiocb *req,
5955 const struct io_uring_sqe *sqe)
5957 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5959 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5961 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5965 req->cancel.addr = READ_ONCE(sqe->addr);
5969 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5971 struct io_ring_ctx *ctx = req->ctx;
5972 u64 sqe_addr = req->cancel.addr;
5973 struct io_tctx_node *node;
5976 ret = io_try_cancel_userdata(req, sqe_addr);
5980 /* slow path, try all io-wq's */
5981 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5983 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5984 struct io_uring_task *tctx = node->task->io_uring;
5986 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5990 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5994 io_req_complete_post(req, ret, 0);
5998 static int io_rsrc_update_prep(struct io_kiocb *req,
5999 const struct io_uring_sqe *sqe)
6001 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6003 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6006 req->rsrc_update.offset = READ_ONCE(sqe->off);
6007 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6008 if (!req->rsrc_update.nr_args)
6010 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6014 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6016 struct io_ring_ctx *ctx = req->ctx;
6017 struct io_uring_rsrc_update2 up;
6020 if (issue_flags & IO_URING_F_NONBLOCK)
6023 up.offset = req->rsrc_update.offset;
6024 up.data = req->rsrc_update.arg;
6029 mutex_lock(&ctx->uring_lock);
6030 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6031 &up, req->rsrc_update.nr_args);
6032 mutex_unlock(&ctx->uring_lock);
6036 __io_req_complete(req, issue_flags, ret, 0);
6040 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6042 switch (req->opcode) {
6045 case IORING_OP_READV:
6046 case IORING_OP_READ_FIXED:
6047 case IORING_OP_READ:
6048 return io_read_prep(req, sqe);
6049 case IORING_OP_WRITEV:
6050 case IORING_OP_WRITE_FIXED:
6051 case IORING_OP_WRITE:
6052 return io_write_prep(req, sqe);
6053 case IORING_OP_POLL_ADD:
6054 return io_poll_add_prep(req, sqe);
6055 case IORING_OP_POLL_REMOVE:
6056 return io_poll_update_prep(req, sqe);
6057 case IORING_OP_FSYNC:
6058 return io_fsync_prep(req, sqe);
6059 case IORING_OP_SYNC_FILE_RANGE:
6060 return io_sfr_prep(req, sqe);
6061 case IORING_OP_SENDMSG:
6062 case IORING_OP_SEND:
6063 return io_sendmsg_prep(req, sqe);
6064 case IORING_OP_RECVMSG:
6065 case IORING_OP_RECV:
6066 return io_recvmsg_prep(req, sqe);
6067 case IORING_OP_CONNECT:
6068 return io_connect_prep(req, sqe);
6069 case IORING_OP_TIMEOUT:
6070 return io_timeout_prep(req, sqe, false);
6071 case IORING_OP_TIMEOUT_REMOVE:
6072 return io_timeout_remove_prep(req, sqe);
6073 case IORING_OP_ASYNC_CANCEL:
6074 return io_async_cancel_prep(req, sqe);
6075 case IORING_OP_LINK_TIMEOUT:
6076 return io_timeout_prep(req, sqe, true);
6077 case IORING_OP_ACCEPT:
6078 return io_accept_prep(req, sqe);
6079 case IORING_OP_FALLOCATE:
6080 return io_fallocate_prep(req, sqe);
6081 case IORING_OP_OPENAT:
6082 return io_openat_prep(req, sqe);
6083 case IORING_OP_CLOSE:
6084 return io_close_prep(req, sqe);
6085 case IORING_OP_FILES_UPDATE:
6086 return io_rsrc_update_prep(req, sqe);
6087 case IORING_OP_STATX:
6088 return io_statx_prep(req, sqe);
6089 case IORING_OP_FADVISE:
6090 return io_fadvise_prep(req, sqe);
6091 case IORING_OP_MADVISE:
6092 return io_madvise_prep(req, sqe);
6093 case IORING_OP_OPENAT2:
6094 return io_openat2_prep(req, sqe);
6095 case IORING_OP_EPOLL_CTL:
6096 return io_epoll_ctl_prep(req, sqe);
6097 case IORING_OP_SPLICE:
6098 return io_splice_prep(req, sqe);
6099 case IORING_OP_PROVIDE_BUFFERS:
6100 return io_provide_buffers_prep(req, sqe);
6101 case IORING_OP_REMOVE_BUFFERS:
6102 return io_remove_buffers_prep(req, sqe);
6104 return io_tee_prep(req, sqe);
6105 case IORING_OP_SHUTDOWN:
6106 return io_shutdown_prep(req, sqe);
6107 case IORING_OP_RENAMEAT:
6108 return io_renameat_prep(req, sqe);
6109 case IORING_OP_UNLINKAT:
6110 return io_unlinkat_prep(req, sqe);
6113 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6118 static int io_req_prep_async(struct io_kiocb *req)
6120 if (!io_op_defs[req->opcode].needs_async_setup)
6122 if (WARN_ON_ONCE(req->async_data))
6124 if (io_alloc_async_data(req))
6127 switch (req->opcode) {
6128 case IORING_OP_READV:
6129 return io_rw_prep_async(req, READ);
6130 case IORING_OP_WRITEV:
6131 return io_rw_prep_async(req, WRITE);
6132 case IORING_OP_SENDMSG:
6133 return io_sendmsg_prep_async(req);
6134 case IORING_OP_RECVMSG:
6135 return io_recvmsg_prep_async(req);
6136 case IORING_OP_CONNECT:
6137 return io_connect_prep_async(req);
6139 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6144 static u32 io_get_sequence(struct io_kiocb *req)
6146 u32 seq = req->ctx->cached_sq_head;
6148 /* need original cached_sq_head, but it was increased for each req */
6149 io_for_each_link(req, req)
6154 static bool io_drain_req(struct io_kiocb *req)
6156 struct io_kiocb *pos;
6157 struct io_ring_ctx *ctx = req->ctx;
6158 struct io_defer_entry *de;
6163 * If we need to drain a request in the middle of a link, drain the
6164 * head request and the next request/link after the current link.
6165 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6166 * maintained for every request of our link.
6168 if (ctx->drain_next) {
6169 req->flags |= REQ_F_IO_DRAIN;
6170 ctx->drain_next = false;
6172 /* not interested in head, start from the first linked */
6173 io_for_each_link(pos, req->link) {
6174 if (pos->flags & REQ_F_IO_DRAIN) {
6175 ctx->drain_next = true;
6176 req->flags |= REQ_F_IO_DRAIN;
6181 /* Still need defer if there is pending req in defer list. */
6182 if (likely(list_empty_careful(&ctx->defer_list) &&
6183 !(req->flags & REQ_F_IO_DRAIN))) {
6184 ctx->drain_active = false;
6188 seq = io_get_sequence(req);
6189 /* Still a chance to pass the sequence check */
6190 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6193 ret = io_req_prep_async(req);
6196 io_prep_async_link(req);
6197 de = kmalloc(sizeof(*de), GFP_KERNEL);
6201 io_req_complete_failed(req, ret);
6205 spin_lock(&ctx->completion_lock);
6206 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6207 spin_unlock(&ctx->completion_lock);
6209 io_queue_async_work(req, NULL);
6213 trace_io_uring_defer(ctx, req, req->user_data);
6216 list_add_tail(&de->list, &ctx->defer_list);
6217 spin_unlock(&ctx->completion_lock);
6221 static void io_clean_op(struct io_kiocb *req)
6223 if (req->flags & REQ_F_BUFFER_SELECTED) {
6224 switch (req->opcode) {
6225 case IORING_OP_READV:
6226 case IORING_OP_READ_FIXED:
6227 case IORING_OP_READ:
6228 kfree((void *)(unsigned long)req->rw.addr);
6230 case IORING_OP_RECVMSG:
6231 case IORING_OP_RECV:
6232 kfree(req->sr_msg.kbuf);
6237 if (req->flags & REQ_F_NEED_CLEANUP) {
6238 switch (req->opcode) {
6239 case IORING_OP_READV:
6240 case IORING_OP_READ_FIXED:
6241 case IORING_OP_READ:
6242 case IORING_OP_WRITEV:
6243 case IORING_OP_WRITE_FIXED:
6244 case IORING_OP_WRITE: {
6245 struct io_async_rw *io = req->async_data;
6247 kfree(io->free_iovec);
6250 case IORING_OP_RECVMSG:
6251 case IORING_OP_SENDMSG: {
6252 struct io_async_msghdr *io = req->async_data;
6254 kfree(io->free_iov);
6257 case IORING_OP_SPLICE:
6259 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6260 io_put_file(req->splice.file_in);
6262 case IORING_OP_OPENAT:
6263 case IORING_OP_OPENAT2:
6264 if (req->open.filename)
6265 putname(req->open.filename);
6267 case IORING_OP_RENAMEAT:
6268 putname(req->rename.oldpath);
6269 putname(req->rename.newpath);
6271 case IORING_OP_UNLINKAT:
6272 putname(req->unlink.filename);
6276 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6277 kfree(req->apoll->double_poll);
6281 if (req->flags & REQ_F_INFLIGHT) {
6282 struct io_uring_task *tctx = req->task->io_uring;
6284 atomic_dec(&tctx->inflight_tracked);
6286 if (req->flags & REQ_F_CREDS)
6287 put_cred(req->creds);
6289 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6292 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6294 struct io_ring_ctx *ctx = req->ctx;
6295 const struct cred *creds = NULL;
6298 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6299 creds = override_creds(req->creds);
6301 switch (req->opcode) {
6303 ret = io_nop(req, issue_flags);
6305 case IORING_OP_READV:
6306 case IORING_OP_READ_FIXED:
6307 case IORING_OP_READ:
6308 ret = io_read(req, issue_flags);
6310 case IORING_OP_WRITEV:
6311 case IORING_OP_WRITE_FIXED:
6312 case IORING_OP_WRITE:
6313 ret = io_write(req, issue_flags);
6315 case IORING_OP_FSYNC:
6316 ret = io_fsync(req, issue_flags);
6318 case IORING_OP_POLL_ADD:
6319 ret = io_poll_add(req, issue_flags);
6321 case IORING_OP_POLL_REMOVE:
6322 ret = io_poll_update(req, issue_flags);
6324 case IORING_OP_SYNC_FILE_RANGE:
6325 ret = io_sync_file_range(req, issue_flags);
6327 case IORING_OP_SENDMSG:
6328 ret = io_sendmsg(req, issue_flags);
6330 case IORING_OP_SEND:
6331 ret = io_send(req, issue_flags);
6333 case IORING_OP_RECVMSG:
6334 ret = io_recvmsg(req, issue_flags);
6336 case IORING_OP_RECV:
6337 ret = io_recv(req, issue_flags);
6339 case IORING_OP_TIMEOUT:
6340 ret = io_timeout(req, issue_flags);
6342 case IORING_OP_TIMEOUT_REMOVE:
6343 ret = io_timeout_remove(req, issue_flags);
6345 case IORING_OP_ACCEPT:
6346 ret = io_accept(req, issue_flags);
6348 case IORING_OP_CONNECT:
6349 ret = io_connect(req, issue_flags);
6351 case IORING_OP_ASYNC_CANCEL:
6352 ret = io_async_cancel(req, issue_flags);
6354 case IORING_OP_FALLOCATE:
6355 ret = io_fallocate(req, issue_flags);
6357 case IORING_OP_OPENAT:
6358 ret = io_openat(req, issue_flags);
6360 case IORING_OP_CLOSE:
6361 ret = io_close(req, issue_flags);
6363 case IORING_OP_FILES_UPDATE:
6364 ret = io_files_update(req, issue_flags);
6366 case IORING_OP_STATX:
6367 ret = io_statx(req, issue_flags);
6369 case IORING_OP_FADVISE:
6370 ret = io_fadvise(req, issue_flags);
6372 case IORING_OP_MADVISE:
6373 ret = io_madvise(req, issue_flags);
6375 case IORING_OP_OPENAT2:
6376 ret = io_openat2(req, issue_flags);
6378 case IORING_OP_EPOLL_CTL:
6379 ret = io_epoll_ctl(req, issue_flags);
6381 case IORING_OP_SPLICE:
6382 ret = io_splice(req, issue_flags);
6384 case IORING_OP_PROVIDE_BUFFERS:
6385 ret = io_provide_buffers(req, issue_flags);
6387 case IORING_OP_REMOVE_BUFFERS:
6388 ret = io_remove_buffers(req, issue_flags);
6391 ret = io_tee(req, issue_flags);
6393 case IORING_OP_SHUTDOWN:
6394 ret = io_shutdown(req, issue_flags);
6396 case IORING_OP_RENAMEAT:
6397 ret = io_renameat(req, issue_flags);
6399 case IORING_OP_UNLINKAT:
6400 ret = io_unlinkat(req, issue_flags);
6408 revert_creds(creds);
6411 /* If the op doesn't have a file, we're not polling for it */
6412 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6413 io_iopoll_req_issued(req);
6418 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6420 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6422 req = io_put_req_find_next(req);
6423 return req ? &req->work : NULL;
6426 static void io_wq_submit_work(struct io_wq_work *work)
6428 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6429 struct io_kiocb *timeout;
6432 /* one will be dropped by ->io_free_work() after returning to io-wq */
6433 if (!(req->flags & REQ_F_REFCOUNT))
6434 __io_req_set_refcount(req, 2);
6438 timeout = io_prep_linked_timeout(req);
6440 io_queue_linked_timeout(timeout);
6442 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6443 if (work->flags & IO_WQ_WORK_CANCEL)
6448 ret = io_issue_sqe(req, 0);
6450 * We can get EAGAIN for polled IO even though we're
6451 * forcing a sync submission from here, since we can't
6452 * wait for request slots on the block side.
6460 /* avoid locking problems by failing it from a clean context */
6462 io_req_task_queue_fail(req, ret);
6465 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6468 return &table->files[i];
6471 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6474 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6476 return (struct file *) (slot->file_ptr & FFS_MASK);
6479 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6481 unsigned long file_ptr = (unsigned long) file;
6483 if (__io_file_supports_nowait(file, READ))
6484 file_ptr |= FFS_ASYNC_READ;
6485 if (__io_file_supports_nowait(file, WRITE))
6486 file_ptr |= FFS_ASYNC_WRITE;
6487 if (S_ISREG(file_inode(file)->i_mode))
6488 file_ptr |= FFS_ISREG;
6489 file_slot->file_ptr = file_ptr;
6492 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6493 struct io_kiocb *req, int fd)
6496 unsigned long file_ptr;
6498 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6500 fd = array_index_nospec(fd, ctx->nr_user_files);
6501 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6502 file = (struct file *) (file_ptr & FFS_MASK);
6503 file_ptr &= ~FFS_MASK;
6504 /* mask in overlapping REQ_F and FFS bits */
6505 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6506 io_req_set_rsrc_node(req);
6510 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6511 struct io_kiocb *req, int fd)
6513 struct file *file = fget(fd);
6515 trace_io_uring_file_get(ctx, fd);
6517 /* we don't allow fixed io_uring files */
6518 if (file && unlikely(file->f_op == &io_uring_fops))
6519 io_req_track_inflight(req);
6523 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6524 struct io_kiocb *req, int fd, bool fixed)
6527 return io_file_get_fixed(ctx, req, fd);
6529 return io_file_get_normal(ctx, req, fd);
6532 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6534 struct io_kiocb *prev = req->timeout.prev;
6538 ret = io_try_cancel_userdata(req, prev->user_data);
6539 io_req_complete_post(req, ret ?: -ETIME, 0);
6542 io_req_complete_post(req, -ETIME, 0);
6546 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6548 struct io_timeout_data *data = container_of(timer,
6549 struct io_timeout_data, timer);
6550 struct io_kiocb *prev, *req = data->req;
6551 struct io_ring_ctx *ctx = req->ctx;
6552 unsigned long flags;
6554 spin_lock_irqsave(&ctx->timeout_lock, flags);
6555 prev = req->timeout.head;
6556 req->timeout.head = NULL;
6559 * We don't expect the list to be empty, that will only happen if we
6560 * race with the completion of the linked work.
6563 io_remove_next_linked(prev);
6564 if (!req_ref_inc_not_zero(prev))
6567 req->timeout.prev = prev;
6568 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6570 req->io_task_work.func = io_req_task_link_timeout;
6571 io_req_task_work_add(req);
6572 return HRTIMER_NORESTART;
6575 static void io_queue_linked_timeout(struct io_kiocb *req)
6577 struct io_ring_ctx *ctx = req->ctx;
6579 spin_lock_irq(&ctx->timeout_lock);
6581 * If the back reference is NULL, then our linked request finished
6582 * before we got a chance to setup the timer
6584 if (req->timeout.head) {
6585 struct io_timeout_data *data = req->async_data;
6587 data->timer.function = io_link_timeout_fn;
6588 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6591 spin_unlock_irq(&ctx->timeout_lock);
6592 /* drop submission reference */
6596 static void __io_queue_sqe(struct io_kiocb *req)
6597 __must_hold(&req->ctx->uring_lock)
6599 struct io_kiocb *linked_timeout;
6603 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6606 * We async punt it if the file wasn't marked NOWAIT, or if the file
6607 * doesn't support non-blocking read/write attempts
6610 if (req->flags & REQ_F_COMPLETE_INLINE) {
6611 struct io_ring_ctx *ctx = req->ctx;
6612 struct io_submit_state *state = &ctx->submit_state;
6614 state->compl_reqs[state->compl_nr++] = req;
6615 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6616 io_submit_flush_completions(ctx);
6620 linked_timeout = io_prep_linked_timeout(req);
6622 io_queue_linked_timeout(linked_timeout);
6623 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6624 linked_timeout = io_prep_linked_timeout(req);
6626 switch (io_arm_poll_handler(req)) {
6627 case IO_APOLL_READY:
6629 io_unprep_linked_timeout(req);
6631 case IO_APOLL_ABORTED:
6633 * Queued up for async execution, worker will release
6634 * submit reference when the iocb is actually submitted.
6636 io_queue_async_work(req, NULL);
6641 io_queue_linked_timeout(linked_timeout);
6643 io_req_complete_failed(req, ret);
6647 static inline void io_queue_sqe(struct io_kiocb *req)
6648 __must_hold(&req->ctx->uring_lock)
6650 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6653 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6654 __io_queue_sqe(req);
6655 } else if (req->flags & REQ_F_FAIL) {
6656 io_req_complete_failed(req, req->result);
6658 int ret = io_req_prep_async(req);
6661 io_req_complete_failed(req, ret);
6663 io_queue_async_work(req, NULL);
6668 * Check SQE restrictions (opcode and flags).
6670 * Returns 'true' if SQE is allowed, 'false' otherwise.
6672 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6673 struct io_kiocb *req,
6674 unsigned int sqe_flags)
6676 if (likely(!ctx->restricted))
6679 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6682 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6683 ctx->restrictions.sqe_flags_required)
6686 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6687 ctx->restrictions.sqe_flags_required))
6693 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6694 const struct io_uring_sqe *sqe)
6695 __must_hold(&ctx->uring_lock)
6697 struct io_submit_state *state;
6698 unsigned int sqe_flags;
6699 int personality, ret = 0;
6701 /* req is partially pre-initialised, see io_preinit_req() */
6702 req->opcode = READ_ONCE(sqe->opcode);
6703 /* same numerical values with corresponding REQ_F_*, safe to copy */
6704 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6705 req->user_data = READ_ONCE(sqe->user_data);
6707 req->fixed_rsrc_refs = NULL;
6708 req->task = current;
6710 /* enforce forwards compatibility on users */
6711 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6713 if (unlikely(req->opcode >= IORING_OP_LAST))
6715 if (!io_check_restriction(ctx, req, sqe_flags))
6718 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6719 !io_op_defs[req->opcode].buffer_select)
6721 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6722 ctx->drain_active = true;
6724 personality = READ_ONCE(sqe->personality);
6726 req->creds = xa_load(&ctx->personalities, personality);
6729 get_cred(req->creds);
6730 req->flags |= REQ_F_CREDS;
6732 state = &ctx->submit_state;
6735 * Plug now if we have more than 1 IO left after this, and the target
6736 * is potentially a read/write to block based storage.
6738 if (!state->plug_started && state->ios_left > 1 &&
6739 io_op_defs[req->opcode].plug) {
6740 blk_start_plug(&state->plug);
6741 state->plug_started = true;
6744 if (io_op_defs[req->opcode].needs_file) {
6745 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6746 (sqe_flags & IOSQE_FIXED_FILE));
6747 if (unlikely(!req->file))
6755 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6756 const struct io_uring_sqe *sqe)
6757 __must_hold(&ctx->uring_lock)
6759 struct io_submit_link *link = &ctx->submit_state.link;
6762 ret = io_init_req(ctx, req, sqe);
6763 if (unlikely(ret)) {
6765 /* fail even hard links since we don't submit */
6768 * we can judge a link req is failed or cancelled by if
6769 * REQ_F_FAIL is set, but the head is an exception since
6770 * it may be set REQ_F_FAIL because of other req's failure
6771 * so let's leverage req->result to distinguish if a head
6772 * is set REQ_F_FAIL because of its failure or other req's
6773 * failure so that we can set the correct ret code for it.
6774 * init result here to avoid affecting the normal path.
6776 if (!(link->head->flags & REQ_F_FAIL))
6777 req_fail_link_node(link->head, -ECANCELED);
6778 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6780 * the current req is a normal req, we should return
6781 * error and thus break the submittion loop.
6783 io_req_complete_failed(req, ret);
6786 req_fail_link_node(req, ret);
6788 ret = io_req_prep(req, sqe);
6793 /* don't need @sqe from now on */
6794 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6796 ctx->flags & IORING_SETUP_SQPOLL);
6799 * If we already have a head request, queue this one for async
6800 * submittal once the head completes. If we don't have a head but
6801 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6802 * submitted sync once the chain is complete. If none of those
6803 * conditions are true (normal request), then just queue it.
6806 struct io_kiocb *head = link->head;
6808 if (!(req->flags & REQ_F_FAIL)) {
6809 ret = io_req_prep_async(req);
6810 if (unlikely(ret)) {
6811 req_fail_link_node(req, ret);
6812 if (!(head->flags & REQ_F_FAIL))
6813 req_fail_link_node(head, -ECANCELED);
6816 trace_io_uring_link(ctx, req, head);
6817 link->last->link = req;
6820 /* last request of a link, enqueue the link */
6821 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6826 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6838 * Batched submission is done, ensure local IO is flushed out.
6840 static void io_submit_state_end(struct io_submit_state *state,
6841 struct io_ring_ctx *ctx)
6843 if (state->link.head)
6844 io_queue_sqe(state->link.head);
6845 if (state->compl_nr)
6846 io_submit_flush_completions(ctx);
6847 if (state->plug_started)
6848 blk_finish_plug(&state->plug);
6852 * Start submission side cache.
6854 static void io_submit_state_start(struct io_submit_state *state,
6855 unsigned int max_ios)
6857 state->plug_started = false;
6858 state->ios_left = max_ios;
6859 /* set only head, no need to init link_last in advance */
6860 state->link.head = NULL;
6863 static void io_commit_sqring(struct io_ring_ctx *ctx)
6865 struct io_rings *rings = ctx->rings;
6868 * Ensure any loads from the SQEs are done at this point,
6869 * since once we write the new head, the application could
6870 * write new data to them.
6872 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6876 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6877 * that is mapped by userspace. This means that care needs to be taken to
6878 * ensure that reads are stable, as we cannot rely on userspace always
6879 * being a good citizen. If members of the sqe are validated and then later
6880 * used, it's important that those reads are done through READ_ONCE() to
6881 * prevent a re-load down the line.
6883 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6885 unsigned head, mask = ctx->sq_entries - 1;
6886 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6889 * The cached sq head (or cq tail) serves two purposes:
6891 * 1) allows us to batch the cost of updating the user visible
6893 * 2) allows the kernel side to track the head on its own, even
6894 * though the application is the one updating it.
6896 head = READ_ONCE(ctx->sq_array[sq_idx]);
6897 if (likely(head < ctx->sq_entries))
6898 return &ctx->sq_sqes[head];
6900 /* drop invalid entries */
6902 WRITE_ONCE(ctx->rings->sq_dropped,
6903 READ_ONCE(ctx->rings->sq_dropped) + 1);
6907 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6908 __must_hold(&ctx->uring_lock)
6912 /* make sure SQ entry isn't read before tail */
6913 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6914 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6916 io_get_task_refs(nr);
6918 io_submit_state_start(&ctx->submit_state, nr);
6919 while (submitted < nr) {
6920 const struct io_uring_sqe *sqe;
6921 struct io_kiocb *req;
6923 req = io_alloc_req(ctx);
6924 if (unlikely(!req)) {
6926 submitted = -EAGAIN;
6929 sqe = io_get_sqe(ctx);
6930 if (unlikely(!sqe)) {
6931 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
6934 /* will complete beyond this point, count as submitted */
6936 if (io_submit_sqe(ctx, req, sqe))
6940 if (unlikely(submitted != nr)) {
6941 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6942 int unused = nr - ref_used;
6944 current->io_uring->cached_refs += unused;
6945 percpu_ref_put_many(&ctx->refs, unused);
6948 io_submit_state_end(&ctx->submit_state, ctx);
6949 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6950 io_commit_sqring(ctx);
6955 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6957 return READ_ONCE(sqd->state);
6960 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6962 /* Tell userspace we may need a wakeup call */
6963 spin_lock(&ctx->completion_lock);
6964 WRITE_ONCE(ctx->rings->sq_flags,
6965 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6966 spin_unlock(&ctx->completion_lock);
6969 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6971 spin_lock(&ctx->completion_lock);
6972 WRITE_ONCE(ctx->rings->sq_flags,
6973 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6974 spin_unlock(&ctx->completion_lock);
6977 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6979 unsigned int to_submit;
6982 to_submit = io_sqring_entries(ctx);
6983 /* if we're handling multiple rings, cap submit size for fairness */
6984 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6985 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6987 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6988 unsigned nr_events = 0;
6989 const struct cred *creds = NULL;
6991 if (ctx->sq_creds != current_cred())
6992 creds = override_creds(ctx->sq_creds);
6994 mutex_lock(&ctx->uring_lock);
6995 if (!list_empty(&ctx->iopoll_list))
6996 io_do_iopoll(ctx, &nr_events, 0);
6999 * Don't submit if refs are dying, good for io_uring_register(),
7000 * but also it is relied upon by io_ring_exit_work()
7002 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7003 !(ctx->flags & IORING_SETUP_R_DISABLED))
7004 ret = io_submit_sqes(ctx, to_submit);
7005 mutex_unlock(&ctx->uring_lock);
7007 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7008 wake_up(&ctx->sqo_sq_wait);
7010 revert_creds(creds);
7016 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7018 struct io_ring_ctx *ctx;
7019 unsigned sq_thread_idle = 0;
7021 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7022 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7023 sqd->sq_thread_idle = sq_thread_idle;
7026 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7028 bool did_sig = false;
7029 struct ksignal ksig;
7031 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7032 signal_pending(current)) {
7033 mutex_unlock(&sqd->lock);
7034 if (signal_pending(current))
7035 did_sig = get_signal(&ksig);
7037 mutex_lock(&sqd->lock);
7039 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7042 static int io_sq_thread(void *data)
7044 struct io_sq_data *sqd = data;
7045 struct io_ring_ctx *ctx;
7046 unsigned long timeout = 0;
7047 char buf[TASK_COMM_LEN];
7050 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7051 set_task_comm(current, buf);
7053 if (sqd->sq_cpu != -1)
7054 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7056 set_cpus_allowed_ptr(current, cpu_online_mask);
7057 current->flags |= PF_NO_SETAFFINITY;
7059 mutex_lock(&sqd->lock);
7061 bool cap_entries, sqt_spin = false;
7063 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7064 if (io_sqd_handle_event(sqd))
7066 timeout = jiffies + sqd->sq_thread_idle;
7069 cap_entries = !list_is_singular(&sqd->ctx_list);
7070 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7071 int ret = __io_sq_thread(ctx, cap_entries);
7073 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7076 if (io_run_task_work())
7079 if (sqt_spin || !time_after(jiffies, timeout)) {
7082 timeout = jiffies + sqd->sq_thread_idle;
7086 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7087 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7088 bool needs_sched = true;
7090 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7091 io_ring_set_wakeup_flag(ctx);
7093 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7094 !list_empty_careful(&ctx->iopoll_list)) {
7095 needs_sched = false;
7098 if (io_sqring_entries(ctx)) {
7099 needs_sched = false;
7105 mutex_unlock(&sqd->lock);
7107 mutex_lock(&sqd->lock);
7109 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7110 io_ring_clear_wakeup_flag(ctx);
7113 finish_wait(&sqd->wait, &wait);
7114 timeout = jiffies + sqd->sq_thread_idle;
7117 io_uring_cancel_generic(true, sqd);
7119 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7120 io_ring_set_wakeup_flag(ctx);
7122 mutex_unlock(&sqd->lock);
7124 complete(&sqd->exited);
7128 struct io_wait_queue {
7129 struct wait_queue_entry wq;
7130 struct io_ring_ctx *ctx;
7132 unsigned nr_timeouts;
7135 static inline bool io_should_wake(struct io_wait_queue *iowq)
7137 struct io_ring_ctx *ctx = iowq->ctx;
7138 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7141 * Wake up if we have enough events, or if a timeout occurred since we
7142 * started waiting. For timeouts, we always want to return to userspace,
7143 * regardless of event count.
7145 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7148 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7149 int wake_flags, void *key)
7151 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7155 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7156 * the task, and the next invocation will do it.
7158 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7159 return autoremove_wake_function(curr, mode, wake_flags, key);
7163 static int io_run_task_work_sig(void)
7165 if (io_run_task_work())
7167 if (!signal_pending(current))
7169 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7170 return -ERESTARTSYS;
7174 /* when returns >0, the caller should retry */
7175 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7176 struct io_wait_queue *iowq,
7177 signed long *timeout)
7181 /* make sure we run task_work before checking for signals */
7182 ret = io_run_task_work_sig();
7183 if (ret || io_should_wake(iowq))
7185 /* let the caller flush overflows, retry */
7186 if (test_bit(0, &ctx->check_cq_overflow))
7189 *timeout = schedule_timeout(*timeout);
7190 return !*timeout ? -ETIME : 1;
7194 * Wait until events become available, if we don't already have some. The
7195 * application must reap them itself, as they reside on the shared cq ring.
7197 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7198 const sigset_t __user *sig, size_t sigsz,
7199 struct __kernel_timespec __user *uts)
7201 struct io_wait_queue iowq;
7202 struct io_rings *rings = ctx->rings;
7203 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7207 io_cqring_overflow_flush(ctx);
7208 if (io_cqring_events(ctx) >= min_events)
7210 if (!io_run_task_work())
7215 #ifdef CONFIG_COMPAT
7216 if (in_compat_syscall())
7217 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7221 ret = set_user_sigmask(sig, sigsz);
7228 struct timespec64 ts;
7230 if (get_timespec64(&ts, uts))
7232 timeout = timespec64_to_jiffies(&ts);
7235 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7236 iowq.wq.private = current;
7237 INIT_LIST_HEAD(&iowq.wq.entry);
7239 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7240 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7242 trace_io_uring_cqring_wait(ctx, min_events);
7244 /* if we can't even flush overflow, don't wait for more */
7245 if (!io_cqring_overflow_flush(ctx)) {
7249 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7250 TASK_INTERRUPTIBLE);
7251 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7252 finish_wait(&ctx->cq_wait, &iowq.wq);
7256 restore_saved_sigmask_unless(ret == -EINTR);
7258 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7261 static void io_free_page_table(void **table, size_t size)
7263 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7265 for (i = 0; i < nr_tables; i++)
7270 static void **io_alloc_page_table(size_t size)
7272 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7273 size_t init_size = size;
7276 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7280 for (i = 0; i < nr_tables; i++) {
7281 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7283 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7285 io_free_page_table(table, init_size);
7293 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7295 percpu_ref_exit(&ref_node->refs);
7299 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7301 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7302 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7303 unsigned long flags;
7304 bool first_add = false;
7306 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7309 while (!list_empty(&ctx->rsrc_ref_list)) {
7310 node = list_first_entry(&ctx->rsrc_ref_list,
7311 struct io_rsrc_node, node);
7312 /* recycle ref nodes in order */
7315 list_del(&node->node);
7316 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7318 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7321 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7324 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7326 struct io_rsrc_node *ref_node;
7328 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7332 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7337 INIT_LIST_HEAD(&ref_node->node);
7338 INIT_LIST_HEAD(&ref_node->rsrc_list);
7339 ref_node->done = false;
7343 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7344 struct io_rsrc_data *data_to_kill)
7346 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7347 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7350 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7352 rsrc_node->rsrc_data = data_to_kill;
7353 spin_lock_irq(&ctx->rsrc_ref_lock);
7354 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7355 spin_unlock_irq(&ctx->rsrc_ref_lock);
7357 atomic_inc(&data_to_kill->refs);
7358 percpu_ref_kill(&rsrc_node->refs);
7359 ctx->rsrc_node = NULL;
7362 if (!ctx->rsrc_node) {
7363 ctx->rsrc_node = ctx->rsrc_backup_node;
7364 ctx->rsrc_backup_node = NULL;
7368 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7370 if (ctx->rsrc_backup_node)
7372 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7373 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7376 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7380 /* As we may drop ->uring_lock, other task may have started quiesce */
7384 data->quiesce = true;
7386 ret = io_rsrc_node_switch_start(ctx);
7389 io_rsrc_node_switch(ctx, data);
7391 /* kill initial ref, already quiesced if zero */
7392 if (atomic_dec_and_test(&data->refs))
7394 mutex_unlock(&ctx->uring_lock);
7395 flush_delayed_work(&ctx->rsrc_put_work);
7396 ret = wait_for_completion_interruptible(&data->done);
7398 mutex_lock(&ctx->uring_lock);
7402 atomic_inc(&data->refs);
7403 /* wait for all works potentially completing data->done */
7404 flush_delayed_work(&ctx->rsrc_put_work);
7405 reinit_completion(&data->done);
7407 ret = io_run_task_work_sig();
7408 mutex_lock(&ctx->uring_lock);
7410 data->quiesce = false;
7415 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7417 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7418 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7420 return &data->tags[table_idx][off];
7423 static void io_rsrc_data_free(struct io_rsrc_data *data)
7425 size_t size = data->nr * sizeof(data->tags[0][0]);
7428 io_free_page_table((void **)data->tags, size);
7432 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7433 u64 __user *utags, unsigned nr,
7434 struct io_rsrc_data **pdata)
7436 struct io_rsrc_data *data;
7440 data = kzalloc(sizeof(*data), GFP_KERNEL);
7443 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7451 data->do_put = do_put;
7454 for (i = 0; i < nr; i++) {
7455 u64 *tag_slot = io_get_tag_slot(data, i);
7457 if (copy_from_user(tag_slot, &utags[i],
7463 atomic_set(&data->refs, 1);
7464 init_completion(&data->done);
7468 io_rsrc_data_free(data);
7472 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7474 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7475 GFP_KERNEL_ACCOUNT);
7476 return !!table->files;
7479 static void io_free_file_tables(struct io_file_table *table)
7481 kvfree(table->files);
7482 table->files = NULL;
7485 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7487 #if defined(CONFIG_UNIX)
7488 if (ctx->ring_sock) {
7489 struct sock *sock = ctx->ring_sock->sk;
7490 struct sk_buff *skb;
7492 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7498 for (i = 0; i < ctx->nr_user_files; i++) {
7501 file = io_file_from_index(ctx, i);
7506 io_free_file_tables(&ctx->file_table);
7507 io_rsrc_data_free(ctx->file_data);
7508 ctx->file_data = NULL;
7509 ctx->nr_user_files = 0;
7512 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7516 if (!ctx->file_data)
7518 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7520 __io_sqe_files_unregister(ctx);
7524 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7525 __releases(&sqd->lock)
7527 WARN_ON_ONCE(sqd->thread == current);
7530 * Do the dance but not conditional clear_bit() because it'd race with
7531 * other threads incrementing park_pending and setting the bit.
7533 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7534 if (atomic_dec_return(&sqd->park_pending))
7535 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7536 mutex_unlock(&sqd->lock);
7539 static void io_sq_thread_park(struct io_sq_data *sqd)
7540 __acquires(&sqd->lock)
7542 WARN_ON_ONCE(sqd->thread == current);
7544 atomic_inc(&sqd->park_pending);
7545 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7546 mutex_lock(&sqd->lock);
7548 wake_up_process(sqd->thread);
7551 static void io_sq_thread_stop(struct io_sq_data *sqd)
7553 WARN_ON_ONCE(sqd->thread == current);
7554 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7556 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7557 mutex_lock(&sqd->lock);
7559 wake_up_process(sqd->thread);
7560 mutex_unlock(&sqd->lock);
7561 wait_for_completion(&sqd->exited);
7564 static void io_put_sq_data(struct io_sq_data *sqd)
7566 if (refcount_dec_and_test(&sqd->refs)) {
7567 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7569 io_sq_thread_stop(sqd);
7574 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7576 struct io_sq_data *sqd = ctx->sq_data;
7579 io_sq_thread_park(sqd);
7580 list_del_init(&ctx->sqd_list);
7581 io_sqd_update_thread_idle(sqd);
7582 io_sq_thread_unpark(sqd);
7584 io_put_sq_data(sqd);
7585 ctx->sq_data = NULL;
7589 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7591 struct io_ring_ctx *ctx_attach;
7592 struct io_sq_data *sqd;
7595 f = fdget(p->wq_fd);
7597 return ERR_PTR(-ENXIO);
7598 if (f.file->f_op != &io_uring_fops) {
7600 return ERR_PTR(-EINVAL);
7603 ctx_attach = f.file->private_data;
7604 sqd = ctx_attach->sq_data;
7607 return ERR_PTR(-EINVAL);
7609 if (sqd->task_tgid != current->tgid) {
7611 return ERR_PTR(-EPERM);
7614 refcount_inc(&sqd->refs);
7619 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7622 struct io_sq_data *sqd;
7625 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7626 sqd = io_attach_sq_data(p);
7631 /* fall through for EPERM case, setup new sqd/task */
7632 if (PTR_ERR(sqd) != -EPERM)
7636 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7638 return ERR_PTR(-ENOMEM);
7640 atomic_set(&sqd->park_pending, 0);
7641 refcount_set(&sqd->refs, 1);
7642 INIT_LIST_HEAD(&sqd->ctx_list);
7643 mutex_init(&sqd->lock);
7644 init_waitqueue_head(&sqd->wait);
7645 init_completion(&sqd->exited);
7649 #if defined(CONFIG_UNIX)
7651 * Ensure the UNIX gc is aware of our file set, so we are certain that
7652 * the io_uring can be safely unregistered on process exit, even if we have
7653 * loops in the file referencing.
7655 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7657 struct sock *sk = ctx->ring_sock->sk;
7658 struct scm_fp_list *fpl;
7659 struct sk_buff *skb;
7662 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7666 skb = alloc_skb(0, GFP_KERNEL);
7675 fpl->user = get_uid(current_user());
7676 for (i = 0; i < nr; i++) {
7677 struct file *file = io_file_from_index(ctx, i + offset);
7681 fpl->fp[nr_files] = get_file(file);
7682 unix_inflight(fpl->user, fpl->fp[nr_files]);
7687 fpl->max = SCM_MAX_FD;
7688 fpl->count = nr_files;
7689 UNIXCB(skb).fp = fpl;
7690 skb->destructor = unix_destruct_scm;
7691 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7692 skb_queue_head(&sk->sk_receive_queue, skb);
7694 for (i = 0; i < nr_files; i++)
7705 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7706 * causes regular reference counting to break down. We rely on the UNIX
7707 * garbage collection to take care of this problem for us.
7709 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7711 unsigned left, total;
7715 left = ctx->nr_user_files;
7717 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7719 ret = __io_sqe_files_scm(ctx, this_files, total);
7723 total += this_files;
7729 while (total < ctx->nr_user_files) {
7730 struct file *file = io_file_from_index(ctx, total);
7740 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7746 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7748 struct file *file = prsrc->file;
7749 #if defined(CONFIG_UNIX)
7750 struct sock *sock = ctx->ring_sock->sk;
7751 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7752 struct sk_buff *skb;
7755 __skb_queue_head_init(&list);
7758 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7759 * remove this entry and rearrange the file array.
7761 skb = skb_dequeue(head);
7763 struct scm_fp_list *fp;
7765 fp = UNIXCB(skb).fp;
7766 for (i = 0; i < fp->count; i++) {
7769 if (fp->fp[i] != file)
7772 unix_notinflight(fp->user, fp->fp[i]);
7773 left = fp->count - 1 - i;
7775 memmove(&fp->fp[i], &fp->fp[i + 1],
7776 left * sizeof(struct file *));
7783 __skb_queue_tail(&list, skb);
7793 __skb_queue_tail(&list, skb);
7795 skb = skb_dequeue(head);
7798 if (skb_peek(&list)) {
7799 spin_lock_irq(&head->lock);
7800 while ((skb = __skb_dequeue(&list)) != NULL)
7801 __skb_queue_tail(head, skb);
7802 spin_unlock_irq(&head->lock);
7809 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7811 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7812 struct io_ring_ctx *ctx = rsrc_data->ctx;
7813 struct io_rsrc_put *prsrc, *tmp;
7815 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7816 list_del(&prsrc->list);
7819 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7821 io_ring_submit_lock(ctx, lock_ring);
7822 spin_lock(&ctx->completion_lock);
7823 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7825 io_commit_cqring(ctx);
7826 spin_unlock(&ctx->completion_lock);
7827 io_cqring_ev_posted(ctx);
7828 io_ring_submit_unlock(ctx, lock_ring);
7831 rsrc_data->do_put(ctx, prsrc);
7835 io_rsrc_node_destroy(ref_node);
7836 if (atomic_dec_and_test(&rsrc_data->refs))
7837 complete(&rsrc_data->done);
7840 static void io_rsrc_put_work(struct work_struct *work)
7842 struct io_ring_ctx *ctx;
7843 struct llist_node *node;
7845 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7846 node = llist_del_all(&ctx->rsrc_put_llist);
7849 struct io_rsrc_node *ref_node;
7850 struct llist_node *next = node->next;
7852 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7853 __io_rsrc_put_work(ref_node);
7858 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7859 unsigned nr_args, u64 __user *tags)
7861 __s32 __user *fds = (__s32 __user *) arg;
7870 if (nr_args > IORING_MAX_FIXED_FILES)
7872 if (nr_args > rlimit(RLIMIT_NOFILE))
7874 ret = io_rsrc_node_switch_start(ctx);
7877 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7883 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7886 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7887 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7891 /* allow sparse sets */
7894 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7901 if (unlikely(!file))
7905 * Don't allow io_uring instances to be registered. If UNIX
7906 * isn't enabled, then this causes a reference cycle and this
7907 * instance can never get freed. If UNIX is enabled we'll
7908 * handle it just fine, but there's still no point in allowing
7909 * a ring fd as it doesn't support regular read/write anyway.
7911 if (file->f_op == &io_uring_fops) {
7915 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7918 ret = io_sqe_files_scm(ctx);
7920 __io_sqe_files_unregister(ctx);
7924 io_rsrc_node_switch(ctx, NULL);
7927 for (i = 0; i < ctx->nr_user_files; i++) {
7928 file = io_file_from_index(ctx, i);
7932 io_free_file_tables(&ctx->file_table);
7933 ctx->nr_user_files = 0;
7935 io_rsrc_data_free(ctx->file_data);
7936 ctx->file_data = NULL;
7940 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7943 #if defined(CONFIG_UNIX)
7944 struct sock *sock = ctx->ring_sock->sk;
7945 struct sk_buff_head *head = &sock->sk_receive_queue;
7946 struct sk_buff *skb;
7949 * See if we can merge this file into an existing skb SCM_RIGHTS
7950 * file set. If there's no room, fall back to allocating a new skb
7951 * and filling it in.
7953 spin_lock_irq(&head->lock);
7954 skb = skb_peek(head);
7956 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7958 if (fpl->count < SCM_MAX_FD) {
7959 __skb_unlink(skb, head);
7960 spin_unlock_irq(&head->lock);
7961 fpl->fp[fpl->count] = get_file(file);
7962 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7964 spin_lock_irq(&head->lock);
7965 __skb_queue_head(head, skb);
7970 spin_unlock_irq(&head->lock);
7977 return __io_sqe_files_scm(ctx, 1, index);
7983 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
7984 unsigned int issue_flags, u32 slot_index)
7986 struct io_ring_ctx *ctx = req->ctx;
7987 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
7988 struct io_fixed_file *file_slot;
7991 io_ring_submit_lock(ctx, !force_nonblock);
7992 if (file->f_op == &io_uring_fops)
7995 if (!ctx->file_data)
7998 if (slot_index >= ctx->nr_user_files)
8001 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8002 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8004 if (file_slot->file_ptr)
8007 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8008 io_fixed_file_set(file_slot, file);
8009 ret = io_sqe_file_register(ctx, file, slot_index);
8011 file_slot->file_ptr = 0;
8017 io_ring_submit_unlock(ctx, !force_nonblock);
8023 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8024 struct io_rsrc_node *node, void *rsrc)
8026 struct io_rsrc_put *prsrc;
8028 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8032 prsrc->tag = *io_get_tag_slot(data, idx);
8034 list_add(&prsrc->list, &node->rsrc_list);
8038 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8039 struct io_uring_rsrc_update2 *up,
8042 u64 __user *tags = u64_to_user_ptr(up->tags);
8043 __s32 __user *fds = u64_to_user_ptr(up->data);
8044 struct io_rsrc_data *data = ctx->file_data;
8045 struct io_fixed_file *file_slot;
8049 bool needs_switch = false;
8051 if (!ctx->file_data)
8053 if (up->offset + nr_args > ctx->nr_user_files)
8056 for (done = 0; done < nr_args; done++) {
8059 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8060 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8064 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8068 if (fd == IORING_REGISTER_FILES_SKIP)
8071 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8072 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8074 if (file_slot->file_ptr) {
8075 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8076 err = io_queue_rsrc_removal(data, up->offset + done,
8077 ctx->rsrc_node, file);
8080 file_slot->file_ptr = 0;
8081 needs_switch = true;
8090 * Don't allow io_uring instances to be registered. If
8091 * UNIX isn't enabled, then this causes a reference
8092 * cycle and this instance can never get freed. If UNIX
8093 * is enabled we'll handle it just fine, but there's
8094 * still no point in allowing a ring fd as it doesn't
8095 * support regular read/write anyway.
8097 if (file->f_op == &io_uring_fops) {
8102 *io_get_tag_slot(data, up->offset + done) = tag;
8103 io_fixed_file_set(file_slot, file);
8104 err = io_sqe_file_register(ctx, file, i);
8106 file_slot->file_ptr = 0;
8114 io_rsrc_node_switch(ctx, data);
8115 return done ? done : err;
8118 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8119 struct task_struct *task)
8121 struct io_wq_hash *hash;
8122 struct io_wq_data data;
8123 unsigned int concurrency;
8125 mutex_lock(&ctx->uring_lock);
8126 hash = ctx->hash_map;
8128 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8130 mutex_unlock(&ctx->uring_lock);
8131 return ERR_PTR(-ENOMEM);
8133 refcount_set(&hash->refs, 1);
8134 init_waitqueue_head(&hash->wait);
8135 ctx->hash_map = hash;
8137 mutex_unlock(&ctx->uring_lock);
8141 data.free_work = io_wq_free_work;
8142 data.do_work = io_wq_submit_work;
8144 /* Do QD, or 4 * CPUS, whatever is smallest */
8145 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8147 return io_wq_create(concurrency, &data);
8150 static int io_uring_alloc_task_context(struct task_struct *task,
8151 struct io_ring_ctx *ctx)
8153 struct io_uring_task *tctx;
8156 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8157 if (unlikely(!tctx))
8160 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8161 if (unlikely(ret)) {
8166 tctx->io_wq = io_init_wq_offload(ctx, task);
8167 if (IS_ERR(tctx->io_wq)) {
8168 ret = PTR_ERR(tctx->io_wq);
8169 percpu_counter_destroy(&tctx->inflight);
8175 init_waitqueue_head(&tctx->wait);
8176 atomic_set(&tctx->in_idle, 0);
8177 atomic_set(&tctx->inflight_tracked, 0);
8178 task->io_uring = tctx;
8179 spin_lock_init(&tctx->task_lock);
8180 INIT_WQ_LIST(&tctx->task_list);
8181 init_task_work(&tctx->task_work, tctx_task_work);
8185 void __io_uring_free(struct task_struct *tsk)
8187 struct io_uring_task *tctx = tsk->io_uring;
8189 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8190 WARN_ON_ONCE(tctx->io_wq);
8191 WARN_ON_ONCE(tctx->cached_refs);
8193 percpu_counter_destroy(&tctx->inflight);
8195 tsk->io_uring = NULL;
8198 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8199 struct io_uring_params *p)
8203 /* Retain compatibility with failing for an invalid attach attempt */
8204 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8205 IORING_SETUP_ATTACH_WQ) {
8208 f = fdget(p->wq_fd);
8211 if (f.file->f_op != &io_uring_fops) {
8217 if (ctx->flags & IORING_SETUP_SQPOLL) {
8218 struct task_struct *tsk;
8219 struct io_sq_data *sqd;
8222 sqd = io_get_sq_data(p, &attached);
8228 ctx->sq_creds = get_current_cred();
8230 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8231 if (!ctx->sq_thread_idle)
8232 ctx->sq_thread_idle = HZ;
8234 io_sq_thread_park(sqd);
8235 list_add(&ctx->sqd_list, &sqd->ctx_list);
8236 io_sqd_update_thread_idle(sqd);
8237 /* don't attach to a dying SQPOLL thread, would be racy */
8238 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8239 io_sq_thread_unpark(sqd);
8246 if (p->flags & IORING_SETUP_SQ_AFF) {
8247 int cpu = p->sq_thread_cpu;
8250 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8257 sqd->task_pid = current->pid;
8258 sqd->task_tgid = current->tgid;
8259 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8266 ret = io_uring_alloc_task_context(tsk, ctx);
8267 wake_up_new_task(tsk);
8270 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8271 /* Can't have SQ_AFF without SQPOLL */
8278 complete(&ctx->sq_data->exited);
8280 io_sq_thread_finish(ctx);
8284 static inline void __io_unaccount_mem(struct user_struct *user,
8285 unsigned long nr_pages)
8287 atomic_long_sub(nr_pages, &user->locked_vm);
8290 static inline int __io_account_mem(struct user_struct *user,
8291 unsigned long nr_pages)
8293 unsigned long page_limit, cur_pages, new_pages;
8295 /* Don't allow more pages than we can safely lock */
8296 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8299 cur_pages = atomic_long_read(&user->locked_vm);
8300 new_pages = cur_pages + nr_pages;
8301 if (new_pages > page_limit)
8303 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8304 new_pages) != cur_pages);
8309 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8312 __io_unaccount_mem(ctx->user, nr_pages);
8314 if (ctx->mm_account)
8315 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8318 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8323 ret = __io_account_mem(ctx->user, nr_pages);
8328 if (ctx->mm_account)
8329 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8334 static void io_mem_free(void *ptr)
8341 page = virt_to_head_page(ptr);
8342 if (put_page_testzero(page))
8343 free_compound_page(page);
8346 static void *io_mem_alloc(size_t size)
8348 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8349 __GFP_NORETRY | __GFP_ACCOUNT;
8351 return (void *) __get_free_pages(gfp_flags, get_order(size));
8354 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8357 struct io_rings *rings;
8358 size_t off, sq_array_size;
8360 off = struct_size(rings, cqes, cq_entries);
8361 if (off == SIZE_MAX)
8365 off = ALIGN(off, SMP_CACHE_BYTES);
8373 sq_array_size = array_size(sizeof(u32), sq_entries);
8374 if (sq_array_size == SIZE_MAX)
8377 if (check_add_overflow(off, sq_array_size, &off))
8383 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8385 struct io_mapped_ubuf *imu = *slot;
8388 if (imu != ctx->dummy_ubuf) {
8389 for (i = 0; i < imu->nr_bvecs; i++)
8390 unpin_user_page(imu->bvec[i].bv_page);
8391 if (imu->acct_pages)
8392 io_unaccount_mem(ctx, imu->acct_pages);
8398 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8400 io_buffer_unmap(ctx, &prsrc->buf);
8404 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8408 for (i = 0; i < ctx->nr_user_bufs; i++)
8409 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8410 kfree(ctx->user_bufs);
8411 io_rsrc_data_free(ctx->buf_data);
8412 ctx->user_bufs = NULL;
8413 ctx->buf_data = NULL;
8414 ctx->nr_user_bufs = 0;
8417 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8424 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8426 __io_sqe_buffers_unregister(ctx);
8430 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8431 void __user *arg, unsigned index)
8433 struct iovec __user *src;
8435 #ifdef CONFIG_COMPAT
8437 struct compat_iovec __user *ciovs;
8438 struct compat_iovec ciov;
8440 ciovs = (struct compat_iovec __user *) arg;
8441 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8444 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8445 dst->iov_len = ciov.iov_len;
8449 src = (struct iovec __user *) arg;
8450 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8456 * Not super efficient, but this is just a registration time. And we do cache
8457 * the last compound head, so generally we'll only do a full search if we don't
8460 * We check if the given compound head page has already been accounted, to
8461 * avoid double accounting it. This allows us to account the full size of the
8462 * page, not just the constituent pages of a huge page.
8464 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8465 int nr_pages, struct page *hpage)
8469 /* check current page array */
8470 for (i = 0; i < nr_pages; i++) {
8471 if (!PageCompound(pages[i]))
8473 if (compound_head(pages[i]) == hpage)
8477 /* check previously registered pages */
8478 for (i = 0; i < ctx->nr_user_bufs; i++) {
8479 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8481 for (j = 0; j < imu->nr_bvecs; j++) {
8482 if (!PageCompound(imu->bvec[j].bv_page))
8484 if (compound_head(imu->bvec[j].bv_page) == hpage)
8492 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8493 int nr_pages, struct io_mapped_ubuf *imu,
8494 struct page **last_hpage)
8498 imu->acct_pages = 0;
8499 for (i = 0; i < nr_pages; i++) {
8500 if (!PageCompound(pages[i])) {
8505 hpage = compound_head(pages[i]);
8506 if (hpage == *last_hpage)
8508 *last_hpage = hpage;
8509 if (headpage_already_acct(ctx, pages, i, hpage))
8511 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8515 if (!imu->acct_pages)
8518 ret = io_account_mem(ctx, imu->acct_pages);
8520 imu->acct_pages = 0;
8524 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8525 struct io_mapped_ubuf **pimu,
8526 struct page **last_hpage)
8528 struct io_mapped_ubuf *imu = NULL;
8529 struct vm_area_struct **vmas = NULL;
8530 struct page **pages = NULL;
8531 unsigned long off, start, end, ubuf;
8533 int ret, pret, nr_pages, i;
8535 if (!iov->iov_base) {
8536 *pimu = ctx->dummy_ubuf;
8540 ubuf = (unsigned long) iov->iov_base;
8541 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8542 start = ubuf >> PAGE_SHIFT;
8543 nr_pages = end - start;
8548 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8552 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8557 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8562 mmap_read_lock(current->mm);
8563 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8565 if (pret == nr_pages) {
8566 /* don't support file backed memory */
8567 for (i = 0; i < nr_pages; i++) {
8568 struct vm_area_struct *vma = vmas[i];
8570 if (vma_is_shmem(vma))
8573 !is_file_hugepages(vma->vm_file)) {
8579 ret = pret < 0 ? pret : -EFAULT;
8581 mmap_read_unlock(current->mm);
8584 * if we did partial map, or found file backed vmas,
8585 * release any pages we did get
8588 unpin_user_pages(pages, pret);
8592 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8594 unpin_user_pages(pages, pret);
8598 off = ubuf & ~PAGE_MASK;
8599 size = iov->iov_len;
8600 for (i = 0; i < nr_pages; i++) {
8603 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8604 imu->bvec[i].bv_page = pages[i];
8605 imu->bvec[i].bv_len = vec_len;
8606 imu->bvec[i].bv_offset = off;
8610 /* store original address for later verification */
8612 imu->ubuf_end = ubuf + iov->iov_len;
8613 imu->nr_bvecs = nr_pages;
8624 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8626 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8627 return ctx->user_bufs ? 0 : -ENOMEM;
8630 static int io_buffer_validate(struct iovec *iov)
8632 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8635 * Don't impose further limits on the size and buffer
8636 * constraints here, we'll -EINVAL later when IO is
8637 * submitted if they are wrong.
8640 return iov->iov_len ? -EFAULT : 0;
8644 /* arbitrary limit, but we need something */
8645 if (iov->iov_len > SZ_1G)
8648 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8654 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8655 unsigned int nr_args, u64 __user *tags)
8657 struct page *last_hpage = NULL;
8658 struct io_rsrc_data *data;
8664 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8666 ret = io_rsrc_node_switch_start(ctx);
8669 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8672 ret = io_buffers_map_alloc(ctx, nr_args);
8674 io_rsrc_data_free(data);
8678 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8679 ret = io_copy_iov(ctx, &iov, arg, i);
8682 ret = io_buffer_validate(&iov);
8685 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8690 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8696 WARN_ON_ONCE(ctx->buf_data);
8698 ctx->buf_data = data;
8700 __io_sqe_buffers_unregister(ctx);
8702 io_rsrc_node_switch(ctx, NULL);
8706 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8707 struct io_uring_rsrc_update2 *up,
8708 unsigned int nr_args)
8710 u64 __user *tags = u64_to_user_ptr(up->tags);
8711 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8712 struct page *last_hpage = NULL;
8713 bool needs_switch = false;
8719 if (up->offset + nr_args > ctx->nr_user_bufs)
8722 for (done = 0; done < nr_args; done++) {
8723 struct io_mapped_ubuf *imu;
8724 int offset = up->offset + done;
8727 err = io_copy_iov(ctx, &iov, iovs, done);
8730 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8734 err = io_buffer_validate(&iov);
8737 if (!iov.iov_base && tag) {
8741 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8745 i = array_index_nospec(offset, ctx->nr_user_bufs);
8746 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8747 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8748 ctx->rsrc_node, ctx->user_bufs[i]);
8749 if (unlikely(err)) {
8750 io_buffer_unmap(ctx, &imu);
8753 ctx->user_bufs[i] = NULL;
8754 needs_switch = true;
8757 ctx->user_bufs[i] = imu;
8758 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8762 io_rsrc_node_switch(ctx, ctx->buf_data);
8763 return done ? done : err;
8766 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8768 __s32 __user *fds = arg;
8774 if (copy_from_user(&fd, fds, sizeof(*fds)))
8777 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8778 if (IS_ERR(ctx->cq_ev_fd)) {
8779 int ret = PTR_ERR(ctx->cq_ev_fd);
8781 ctx->cq_ev_fd = NULL;
8788 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8790 if (ctx->cq_ev_fd) {
8791 eventfd_ctx_put(ctx->cq_ev_fd);
8792 ctx->cq_ev_fd = NULL;
8799 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8801 struct io_buffer *buf;
8802 unsigned long index;
8804 xa_for_each(&ctx->io_buffers, index, buf)
8805 __io_remove_buffers(ctx, buf, index, -1U);
8808 static void io_req_cache_free(struct list_head *list)
8810 struct io_kiocb *req, *nxt;
8812 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8813 list_del(&req->inflight_entry);
8814 kmem_cache_free(req_cachep, req);
8818 static void io_req_caches_free(struct io_ring_ctx *ctx)
8820 struct io_submit_state *state = &ctx->submit_state;
8822 mutex_lock(&ctx->uring_lock);
8824 if (state->free_reqs) {
8825 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8826 state->free_reqs = 0;
8829 io_flush_cached_locked_reqs(ctx, state);
8830 io_req_cache_free(&state->free_list);
8831 mutex_unlock(&ctx->uring_lock);
8834 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8836 if (data && !atomic_dec_and_test(&data->refs))
8837 wait_for_completion(&data->done);
8840 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8842 io_sq_thread_finish(ctx);
8844 if (ctx->mm_account) {
8845 mmdrop(ctx->mm_account);
8846 ctx->mm_account = NULL;
8849 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8850 io_wait_rsrc_data(ctx->buf_data);
8851 io_wait_rsrc_data(ctx->file_data);
8853 mutex_lock(&ctx->uring_lock);
8855 __io_sqe_buffers_unregister(ctx);
8857 __io_sqe_files_unregister(ctx);
8859 __io_cqring_overflow_flush(ctx, true);
8860 mutex_unlock(&ctx->uring_lock);
8861 io_eventfd_unregister(ctx);
8862 io_destroy_buffers(ctx);
8864 put_cred(ctx->sq_creds);
8866 /* there are no registered resources left, nobody uses it */
8868 io_rsrc_node_destroy(ctx->rsrc_node);
8869 if (ctx->rsrc_backup_node)
8870 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8871 flush_delayed_work(&ctx->rsrc_put_work);
8873 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8874 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8876 #if defined(CONFIG_UNIX)
8877 if (ctx->ring_sock) {
8878 ctx->ring_sock->file = NULL; /* so that iput() is called */
8879 sock_release(ctx->ring_sock);
8883 io_mem_free(ctx->rings);
8884 io_mem_free(ctx->sq_sqes);
8886 percpu_ref_exit(&ctx->refs);
8887 free_uid(ctx->user);
8888 io_req_caches_free(ctx);
8890 io_wq_put_hash(ctx->hash_map);
8891 kfree(ctx->cancel_hash);
8892 kfree(ctx->dummy_ubuf);
8896 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8898 struct io_ring_ctx *ctx = file->private_data;
8901 poll_wait(file, &ctx->poll_wait, wait);
8903 * synchronizes with barrier from wq_has_sleeper call in
8907 if (!io_sqring_full(ctx))
8908 mask |= EPOLLOUT | EPOLLWRNORM;
8911 * Don't flush cqring overflow list here, just do a simple check.
8912 * Otherwise there could possible be ABBA deadlock:
8915 * lock(&ctx->uring_lock);
8917 * lock(&ctx->uring_lock);
8920 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8921 * pushs them to do the flush.
8923 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8924 mask |= EPOLLIN | EPOLLRDNORM;
8929 static int io_uring_fasync(int fd, struct file *file, int on)
8931 struct io_ring_ctx *ctx = file->private_data;
8933 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8936 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8938 const struct cred *creds;
8940 creds = xa_erase(&ctx->personalities, id);
8949 struct io_tctx_exit {
8950 struct callback_head task_work;
8951 struct completion completion;
8952 struct io_ring_ctx *ctx;
8955 static void io_tctx_exit_cb(struct callback_head *cb)
8957 struct io_uring_task *tctx = current->io_uring;
8958 struct io_tctx_exit *work;
8960 work = container_of(cb, struct io_tctx_exit, task_work);
8962 * When @in_idle, we're in cancellation and it's racy to remove the
8963 * node. It'll be removed by the end of cancellation, just ignore it.
8965 if (!atomic_read(&tctx->in_idle))
8966 io_uring_del_tctx_node((unsigned long)work->ctx);
8967 complete(&work->completion);
8970 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8972 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8974 return req->ctx == data;
8977 static void io_ring_exit_work(struct work_struct *work)
8979 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8980 unsigned long timeout = jiffies + HZ * 60 * 5;
8981 unsigned long interval = HZ / 20;
8982 struct io_tctx_exit exit;
8983 struct io_tctx_node *node;
8987 * If we're doing polled IO and end up having requests being
8988 * submitted async (out-of-line), then completions can come in while
8989 * we're waiting for refs to drop. We need to reap these manually,
8990 * as nobody else will be looking for them.
8993 io_uring_try_cancel_requests(ctx, NULL, true);
8995 struct io_sq_data *sqd = ctx->sq_data;
8996 struct task_struct *tsk;
8998 io_sq_thread_park(sqd);
9000 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9001 io_wq_cancel_cb(tsk->io_uring->io_wq,
9002 io_cancel_ctx_cb, ctx, true);
9003 io_sq_thread_unpark(sqd);
9006 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9007 /* there is little hope left, don't run it too often */
9010 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9012 init_completion(&exit.completion);
9013 init_task_work(&exit.task_work, io_tctx_exit_cb);
9016 * Some may use context even when all refs and requests have been put,
9017 * and they are free to do so while still holding uring_lock or
9018 * completion_lock, see io_req_task_submit(). Apart from other work,
9019 * this lock/unlock section also waits them to finish.
9021 mutex_lock(&ctx->uring_lock);
9022 while (!list_empty(&ctx->tctx_list)) {
9023 WARN_ON_ONCE(time_after(jiffies, timeout));
9025 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9027 /* don't spin on a single task if cancellation failed */
9028 list_rotate_left(&ctx->tctx_list);
9029 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9030 if (WARN_ON_ONCE(ret))
9032 wake_up_process(node->task);
9034 mutex_unlock(&ctx->uring_lock);
9035 wait_for_completion(&exit.completion);
9036 mutex_lock(&ctx->uring_lock);
9038 mutex_unlock(&ctx->uring_lock);
9039 spin_lock(&ctx->completion_lock);
9040 spin_unlock(&ctx->completion_lock);
9042 io_ring_ctx_free(ctx);
9045 /* Returns true if we found and killed one or more timeouts */
9046 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9049 struct io_kiocb *req, *tmp;
9052 spin_lock(&ctx->completion_lock);
9053 spin_lock_irq(&ctx->timeout_lock);
9054 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9055 if (io_match_task(req, tsk, cancel_all)) {
9056 io_kill_timeout(req, -ECANCELED);
9060 spin_unlock_irq(&ctx->timeout_lock);
9062 io_commit_cqring(ctx);
9063 spin_unlock(&ctx->completion_lock);
9065 io_cqring_ev_posted(ctx);
9066 return canceled != 0;
9069 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9071 unsigned long index;
9072 struct creds *creds;
9074 mutex_lock(&ctx->uring_lock);
9075 percpu_ref_kill(&ctx->refs);
9077 __io_cqring_overflow_flush(ctx, true);
9078 xa_for_each(&ctx->personalities, index, creds)
9079 io_unregister_personality(ctx, index);
9080 mutex_unlock(&ctx->uring_lock);
9082 io_kill_timeouts(ctx, NULL, true);
9083 io_poll_remove_all(ctx, NULL, true);
9085 /* if we failed setting up the ctx, we might not have any rings */
9086 io_iopoll_try_reap_events(ctx);
9088 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9090 * Use system_unbound_wq to avoid spawning tons of event kworkers
9091 * if we're exiting a ton of rings at the same time. It just adds
9092 * noise and overhead, there's no discernable change in runtime
9093 * over using system_wq.
9095 queue_work(system_unbound_wq, &ctx->exit_work);
9098 static int io_uring_release(struct inode *inode, struct file *file)
9100 struct io_ring_ctx *ctx = file->private_data;
9102 file->private_data = NULL;
9103 io_ring_ctx_wait_and_kill(ctx);
9107 struct io_task_cancel {
9108 struct task_struct *task;
9112 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9114 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9115 struct io_task_cancel *cancel = data;
9118 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9119 struct io_ring_ctx *ctx = req->ctx;
9121 /* protect against races with linked timeouts */
9122 spin_lock(&ctx->completion_lock);
9123 ret = io_match_task(req, cancel->task, cancel->all);
9124 spin_unlock(&ctx->completion_lock);
9126 ret = io_match_task(req, cancel->task, cancel->all);
9131 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9132 struct task_struct *task, bool cancel_all)
9134 struct io_defer_entry *de;
9137 spin_lock(&ctx->completion_lock);
9138 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9139 if (io_match_task(de->req, task, cancel_all)) {
9140 list_cut_position(&list, &ctx->defer_list, &de->list);
9144 spin_unlock(&ctx->completion_lock);
9145 if (list_empty(&list))
9148 while (!list_empty(&list)) {
9149 de = list_first_entry(&list, struct io_defer_entry, list);
9150 list_del_init(&de->list);
9151 io_req_complete_failed(de->req, -ECANCELED);
9157 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9159 struct io_tctx_node *node;
9160 enum io_wq_cancel cret;
9163 mutex_lock(&ctx->uring_lock);
9164 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9165 struct io_uring_task *tctx = node->task->io_uring;
9168 * io_wq will stay alive while we hold uring_lock, because it's
9169 * killed after ctx nodes, which requires to take the lock.
9171 if (!tctx || !tctx->io_wq)
9173 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9174 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9176 mutex_unlock(&ctx->uring_lock);
9181 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9182 struct task_struct *task,
9185 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9186 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9189 enum io_wq_cancel cret;
9193 ret |= io_uring_try_cancel_iowq(ctx);
9194 } else if (tctx && tctx->io_wq) {
9196 * Cancels requests of all rings, not only @ctx, but
9197 * it's fine as the task is in exit/exec.
9199 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9201 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9204 /* SQPOLL thread does its own polling */
9205 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9206 (ctx->sq_data && ctx->sq_data->thread == current)) {
9207 while (!list_empty_careful(&ctx->iopoll_list)) {
9208 io_iopoll_try_reap_events(ctx);
9213 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9214 ret |= io_poll_remove_all(ctx, task, cancel_all);
9215 ret |= io_kill_timeouts(ctx, task, cancel_all);
9217 ret |= io_run_task_work();
9224 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9226 struct io_uring_task *tctx = current->io_uring;
9227 struct io_tctx_node *node;
9230 if (unlikely(!tctx)) {
9231 ret = io_uring_alloc_task_context(current, ctx);
9234 tctx = current->io_uring;
9236 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9237 node = kmalloc(sizeof(*node), GFP_KERNEL);
9241 node->task = current;
9243 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9250 mutex_lock(&ctx->uring_lock);
9251 list_add(&node->ctx_node, &ctx->tctx_list);
9252 mutex_unlock(&ctx->uring_lock);
9259 * Note that this task has used io_uring. We use it for cancelation purposes.
9261 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9263 struct io_uring_task *tctx = current->io_uring;
9265 if (likely(tctx && tctx->last == ctx))
9267 return __io_uring_add_tctx_node(ctx);
9271 * Remove this io_uring_file -> task mapping.
9273 static void io_uring_del_tctx_node(unsigned long index)
9275 struct io_uring_task *tctx = current->io_uring;
9276 struct io_tctx_node *node;
9280 node = xa_erase(&tctx->xa, index);
9284 WARN_ON_ONCE(current != node->task);
9285 WARN_ON_ONCE(list_empty(&node->ctx_node));
9287 mutex_lock(&node->ctx->uring_lock);
9288 list_del(&node->ctx_node);
9289 mutex_unlock(&node->ctx->uring_lock);
9291 if (tctx->last == node->ctx)
9296 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9298 struct io_wq *wq = tctx->io_wq;
9299 struct io_tctx_node *node;
9300 unsigned long index;
9302 xa_for_each(&tctx->xa, index, node)
9303 io_uring_del_tctx_node(index);
9306 * Must be after io_uring_del_task_file() (removes nodes under
9307 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9309 io_wq_put_and_exit(wq);
9314 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9317 return atomic_read(&tctx->inflight_tracked);
9318 return percpu_counter_sum(&tctx->inflight);
9321 static void io_uring_drop_tctx_refs(struct task_struct *task)
9323 struct io_uring_task *tctx = task->io_uring;
9324 unsigned int refs = tctx->cached_refs;
9327 tctx->cached_refs = 0;
9328 percpu_counter_sub(&tctx->inflight, refs);
9329 put_task_struct_many(task, refs);
9334 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9335 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9337 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9339 struct io_uring_task *tctx = current->io_uring;
9340 struct io_ring_ctx *ctx;
9344 WARN_ON_ONCE(sqd && sqd->thread != current);
9346 if (!current->io_uring)
9349 io_wq_exit_start(tctx->io_wq);
9351 atomic_inc(&tctx->in_idle);
9353 io_uring_drop_tctx_refs(current);
9354 /* read completions before cancelations */
9355 inflight = tctx_inflight(tctx, !cancel_all);
9360 struct io_tctx_node *node;
9361 unsigned long index;
9363 xa_for_each(&tctx->xa, index, node) {
9364 /* sqpoll task will cancel all its requests */
9365 if (node->ctx->sq_data)
9367 io_uring_try_cancel_requests(node->ctx, current,
9371 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9372 io_uring_try_cancel_requests(ctx, current,
9376 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9377 io_uring_drop_tctx_refs(current);
9379 * If we've seen completions, retry without waiting. This
9380 * avoids a race where a completion comes in before we did
9381 * prepare_to_wait().
9383 if (inflight == tctx_inflight(tctx, !cancel_all))
9385 finish_wait(&tctx->wait, &wait);
9387 atomic_dec(&tctx->in_idle);
9389 io_uring_clean_tctx(tctx);
9391 /* for exec all current's requests should be gone, kill tctx */
9392 __io_uring_free(current);
9396 void __io_uring_cancel(bool cancel_all)
9398 io_uring_cancel_generic(cancel_all, NULL);
9401 static void *io_uring_validate_mmap_request(struct file *file,
9402 loff_t pgoff, size_t sz)
9404 struct io_ring_ctx *ctx = file->private_data;
9405 loff_t offset = pgoff << PAGE_SHIFT;
9410 case IORING_OFF_SQ_RING:
9411 case IORING_OFF_CQ_RING:
9414 case IORING_OFF_SQES:
9418 return ERR_PTR(-EINVAL);
9421 page = virt_to_head_page(ptr);
9422 if (sz > page_size(page))
9423 return ERR_PTR(-EINVAL);
9430 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9432 size_t sz = vma->vm_end - vma->vm_start;
9436 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9438 return PTR_ERR(ptr);
9440 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9441 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9444 #else /* !CONFIG_MMU */
9446 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9448 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9451 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9453 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9456 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9457 unsigned long addr, unsigned long len,
9458 unsigned long pgoff, unsigned long flags)
9462 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9464 return PTR_ERR(ptr);
9466 return (unsigned long) ptr;
9469 #endif /* !CONFIG_MMU */
9471 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9476 if (!io_sqring_full(ctx))
9478 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9480 if (!io_sqring_full(ctx))
9483 } while (!signal_pending(current));
9485 finish_wait(&ctx->sqo_sq_wait, &wait);
9489 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9490 struct __kernel_timespec __user **ts,
9491 const sigset_t __user **sig)
9493 struct io_uring_getevents_arg arg;
9496 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9497 * is just a pointer to the sigset_t.
9499 if (!(flags & IORING_ENTER_EXT_ARG)) {
9500 *sig = (const sigset_t __user *) argp;
9506 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9507 * timespec and sigset_t pointers if good.
9509 if (*argsz != sizeof(arg))
9511 if (copy_from_user(&arg, argp, sizeof(arg)))
9513 *sig = u64_to_user_ptr(arg.sigmask);
9514 *argsz = arg.sigmask_sz;
9515 *ts = u64_to_user_ptr(arg.ts);
9519 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9520 u32, min_complete, u32, flags, const void __user *, argp,
9523 struct io_ring_ctx *ctx;
9530 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9531 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9535 if (unlikely(!f.file))
9539 if (unlikely(f.file->f_op != &io_uring_fops))
9543 ctx = f.file->private_data;
9544 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9548 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9552 * For SQ polling, the thread will do all submissions and completions.
9553 * Just return the requested submit count, and wake the thread if
9557 if (ctx->flags & IORING_SETUP_SQPOLL) {
9558 io_cqring_overflow_flush(ctx);
9560 if (unlikely(ctx->sq_data->thread == NULL)) {
9564 if (flags & IORING_ENTER_SQ_WAKEUP)
9565 wake_up(&ctx->sq_data->wait);
9566 if (flags & IORING_ENTER_SQ_WAIT) {
9567 ret = io_sqpoll_wait_sq(ctx);
9571 submitted = to_submit;
9572 } else if (to_submit) {
9573 ret = io_uring_add_tctx_node(ctx);
9576 mutex_lock(&ctx->uring_lock);
9577 submitted = io_submit_sqes(ctx, to_submit);
9578 mutex_unlock(&ctx->uring_lock);
9580 if (submitted != to_submit)
9583 if (flags & IORING_ENTER_GETEVENTS) {
9584 const sigset_t __user *sig;
9585 struct __kernel_timespec __user *ts;
9587 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9591 min_complete = min(min_complete, ctx->cq_entries);
9594 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9595 * space applications don't need to do io completion events
9596 * polling again, they can rely on io_sq_thread to do polling
9597 * work, which can reduce cpu usage and uring_lock contention.
9599 if (ctx->flags & IORING_SETUP_IOPOLL &&
9600 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9601 ret = io_iopoll_check(ctx, min_complete);
9603 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9608 percpu_ref_put(&ctx->refs);
9611 return submitted ? submitted : ret;
9614 #ifdef CONFIG_PROC_FS
9615 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9616 const struct cred *cred)
9618 struct user_namespace *uns = seq_user_ns(m);
9619 struct group_info *gi;
9624 seq_printf(m, "%5d\n", id);
9625 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9626 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9627 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9628 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9629 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9630 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9631 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9632 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9633 seq_puts(m, "\n\tGroups:\t");
9634 gi = cred->group_info;
9635 for (g = 0; g < gi->ngroups; g++) {
9636 seq_put_decimal_ull(m, g ? " " : "",
9637 from_kgid_munged(uns, gi->gid[g]));
9639 seq_puts(m, "\n\tCapEff:\t");
9640 cap = cred->cap_effective;
9641 CAP_FOR_EACH_U32(__capi)
9642 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9647 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9649 struct io_sq_data *sq = NULL;
9654 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9655 * since fdinfo case grabs it in the opposite direction of normal use
9656 * cases. If we fail to get the lock, we just don't iterate any
9657 * structures that could be going away outside the io_uring mutex.
9659 has_lock = mutex_trylock(&ctx->uring_lock);
9661 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9667 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9668 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9669 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9670 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9671 struct file *f = io_file_from_index(ctx, i);
9674 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9676 seq_printf(m, "%5u: <none>\n", i);
9678 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9679 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9680 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9681 unsigned int len = buf->ubuf_end - buf->ubuf;
9683 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9685 if (has_lock && !xa_empty(&ctx->personalities)) {
9686 unsigned long index;
9687 const struct cred *cred;
9689 seq_printf(m, "Personalities:\n");
9690 xa_for_each(&ctx->personalities, index, cred)
9691 io_uring_show_cred(m, index, cred);
9693 seq_printf(m, "PollList:\n");
9694 spin_lock(&ctx->completion_lock);
9695 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9696 struct hlist_head *list = &ctx->cancel_hash[i];
9697 struct io_kiocb *req;
9699 hlist_for_each_entry(req, list, hash_node)
9700 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9701 req->task->task_works != NULL);
9703 spin_unlock(&ctx->completion_lock);
9705 mutex_unlock(&ctx->uring_lock);
9708 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9710 struct io_ring_ctx *ctx = f->private_data;
9712 if (percpu_ref_tryget(&ctx->refs)) {
9713 __io_uring_show_fdinfo(ctx, m);
9714 percpu_ref_put(&ctx->refs);
9719 static const struct file_operations io_uring_fops = {
9720 .release = io_uring_release,
9721 .mmap = io_uring_mmap,
9723 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9724 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9726 .poll = io_uring_poll,
9727 .fasync = io_uring_fasync,
9728 #ifdef CONFIG_PROC_FS
9729 .show_fdinfo = io_uring_show_fdinfo,
9733 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9734 struct io_uring_params *p)
9736 struct io_rings *rings;
9737 size_t size, sq_array_offset;
9739 /* make sure these are sane, as we already accounted them */
9740 ctx->sq_entries = p->sq_entries;
9741 ctx->cq_entries = p->cq_entries;
9743 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9744 if (size == SIZE_MAX)
9747 rings = io_mem_alloc(size);
9752 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9753 rings->sq_ring_mask = p->sq_entries - 1;
9754 rings->cq_ring_mask = p->cq_entries - 1;
9755 rings->sq_ring_entries = p->sq_entries;
9756 rings->cq_ring_entries = p->cq_entries;
9758 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9759 if (size == SIZE_MAX) {
9760 io_mem_free(ctx->rings);
9765 ctx->sq_sqes = io_mem_alloc(size);
9766 if (!ctx->sq_sqes) {
9767 io_mem_free(ctx->rings);
9775 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9779 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9783 ret = io_uring_add_tctx_node(ctx);
9788 fd_install(fd, file);
9793 * Allocate an anonymous fd, this is what constitutes the application
9794 * visible backing of an io_uring instance. The application mmaps this
9795 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9796 * we have to tie this fd to a socket for file garbage collection purposes.
9798 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9801 #if defined(CONFIG_UNIX)
9804 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9807 return ERR_PTR(ret);
9810 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9811 O_RDWR | O_CLOEXEC);
9812 #if defined(CONFIG_UNIX)
9814 sock_release(ctx->ring_sock);
9815 ctx->ring_sock = NULL;
9817 ctx->ring_sock->file = file;
9823 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9824 struct io_uring_params __user *params)
9826 struct io_ring_ctx *ctx;
9832 if (entries > IORING_MAX_ENTRIES) {
9833 if (!(p->flags & IORING_SETUP_CLAMP))
9835 entries = IORING_MAX_ENTRIES;
9839 * Use twice as many entries for the CQ ring. It's possible for the
9840 * application to drive a higher depth than the size of the SQ ring,
9841 * since the sqes are only used at submission time. This allows for
9842 * some flexibility in overcommitting a bit. If the application has
9843 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9844 * of CQ ring entries manually.
9846 p->sq_entries = roundup_pow_of_two(entries);
9847 if (p->flags & IORING_SETUP_CQSIZE) {
9849 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9850 * to a power-of-two, if it isn't already. We do NOT impose
9851 * any cq vs sq ring sizing.
9855 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9856 if (!(p->flags & IORING_SETUP_CLAMP))
9858 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9860 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9861 if (p->cq_entries < p->sq_entries)
9864 p->cq_entries = 2 * p->sq_entries;
9867 ctx = io_ring_ctx_alloc(p);
9870 ctx->compat = in_compat_syscall();
9871 if (!capable(CAP_IPC_LOCK))
9872 ctx->user = get_uid(current_user());
9875 * This is just grabbed for accounting purposes. When a process exits,
9876 * the mm is exited and dropped before the files, hence we need to hang
9877 * on to this mm purely for the purposes of being able to unaccount
9878 * memory (locked/pinned vm). It's not used for anything else.
9880 mmgrab(current->mm);
9881 ctx->mm_account = current->mm;
9883 ret = io_allocate_scq_urings(ctx, p);
9887 ret = io_sq_offload_create(ctx, p);
9890 /* always set a rsrc node */
9891 ret = io_rsrc_node_switch_start(ctx);
9894 io_rsrc_node_switch(ctx, NULL);
9896 memset(&p->sq_off, 0, sizeof(p->sq_off));
9897 p->sq_off.head = offsetof(struct io_rings, sq.head);
9898 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9899 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9900 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9901 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9902 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9903 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9905 memset(&p->cq_off, 0, sizeof(p->cq_off));
9906 p->cq_off.head = offsetof(struct io_rings, cq.head);
9907 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9908 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9909 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9910 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9911 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9912 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9914 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9915 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9916 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9917 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9918 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9919 IORING_FEAT_RSRC_TAGS;
9921 if (copy_to_user(params, p, sizeof(*p))) {
9926 file = io_uring_get_file(ctx);
9928 ret = PTR_ERR(file);
9933 * Install ring fd as the very last thing, so we don't risk someone
9934 * having closed it before we finish setup
9936 ret = io_uring_install_fd(ctx, file);
9938 /* fput will clean it up */
9943 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9946 io_ring_ctx_wait_and_kill(ctx);
9951 * Sets up an aio uring context, and returns the fd. Applications asks for a
9952 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9953 * params structure passed in.
9955 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9957 struct io_uring_params p;
9960 if (copy_from_user(&p, params, sizeof(p)))
9962 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9967 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9968 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9969 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9970 IORING_SETUP_R_DISABLED))
9973 return io_uring_create(entries, &p, params);
9976 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9977 struct io_uring_params __user *, params)
9979 return io_uring_setup(entries, params);
9982 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9984 struct io_uring_probe *p;
9988 size = struct_size(p, ops, nr_args);
9989 if (size == SIZE_MAX)
9991 p = kzalloc(size, GFP_KERNEL);
9996 if (copy_from_user(p, arg, size))
9999 if (memchr_inv(p, 0, size))
10002 p->last_op = IORING_OP_LAST - 1;
10003 if (nr_args > IORING_OP_LAST)
10004 nr_args = IORING_OP_LAST;
10006 for (i = 0; i < nr_args; i++) {
10008 if (!io_op_defs[i].not_supported)
10009 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10014 if (copy_to_user(arg, p, size))
10021 static int io_register_personality(struct io_ring_ctx *ctx)
10023 const struct cred *creds;
10027 creds = get_current_cred();
10029 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10030 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10038 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10039 unsigned int nr_args)
10041 struct io_uring_restriction *res;
10045 /* Restrictions allowed only if rings started disabled */
10046 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10049 /* We allow only a single restrictions registration */
10050 if (ctx->restrictions.registered)
10053 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10056 size = array_size(nr_args, sizeof(*res));
10057 if (size == SIZE_MAX)
10060 res = memdup_user(arg, size);
10062 return PTR_ERR(res);
10066 for (i = 0; i < nr_args; i++) {
10067 switch (res[i].opcode) {
10068 case IORING_RESTRICTION_REGISTER_OP:
10069 if (res[i].register_op >= IORING_REGISTER_LAST) {
10074 __set_bit(res[i].register_op,
10075 ctx->restrictions.register_op);
10077 case IORING_RESTRICTION_SQE_OP:
10078 if (res[i].sqe_op >= IORING_OP_LAST) {
10083 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10085 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10086 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10088 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10089 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10098 /* Reset all restrictions if an error happened */
10100 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10102 ctx->restrictions.registered = true;
10108 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10110 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10113 if (ctx->restrictions.registered)
10114 ctx->restricted = 1;
10116 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10117 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10118 wake_up(&ctx->sq_data->wait);
10122 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10123 struct io_uring_rsrc_update2 *up,
10131 if (check_add_overflow(up->offset, nr_args, &tmp))
10133 err = io_rsrc_node_switch_start(ctx);
10138 case IORING_RSRC_FILE:
10139 return __io_sqe_files_update(ctx, up, nr_args);
10140 case IORING_RSRC_BUFFER:
10141 return __io_sqe_buffers_update(ctx, up, nr_args);
10146 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10149 struct io_uring_rsrc_update2 up;
10153 memset(&up, 0, sizeof(up));
10154 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10156 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10159 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10160 unsigned size, unsigned type)
10162 struct io_uring_rsrc_update2 up;
10164 if (size != sizeof(up))
10166 if (copy_from_user(&up, arg, sizeof(up)))
10168 if (!up.nr || up.resv)
10170 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10173 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10174 unsigned int size, unsigned int type)
10176 struct io_uring_rsrc_register rr;
10178 /* keep it extendible */
10179 if (size != sizeof(rr))
10182 memset(&rr, 0, sizeof(rr));
10183 if (copy_from_user(&rr, arg, size))
10185 if (!rr.nr || rr.resv || rr.resv2)
10189 case IORING_RSRC_FILE:
10190 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10191 rr.nr, u64_to_user_ptr(rr.tags));
10192 case IORING_RSRC_BUFFER:
10193 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10194 rr.nr, u64_to_user_ptr(rr.tags));
10199 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10202 struct io_uring_task *tctx = current->io_uring;
10203 cpumask_var_t new_mask;
10206 if (!tctx || !tctx->io_wq)
10209 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10212 cpumask_clear(new_mask);
10213 if (len > cpumask_size())
10214 len = cpumask_size();
10216 if (copy_from_user(new_mask, arg, len)) {
10217 free_cpumask_var(new_mask);
10221 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10222 free_cpumask_var(new_mask);
10226 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10228 struct io_uring_task *tctx = current->io_uring;
10230 if (!tctx || !tctx->io_wq)
10233 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10236 static bool io_register_op_must_quiesce(int op)
10239 case IORING_REGISTER_BUFFERS:
10240 case IORING_UNREGISTER_BUFFERS:
10241 case IORING_REGISTER_FILES:
10242 case IORING_UNREGISTER_FILES:
10243 case IORING_REGISTER_FILES_UPDATE:
10244 case IORING_REGISTER_PROBE:
10245 case IORING_REGISTER_PERSONALITY:
10246 case IORING_UNREGISTER_PERSONALITY:
10247 case IORING_REGISTER_FILES2:
10248 case IORING_REGISTER_FILES_UPDATE2:
10249 case IORING_REGISTER_BUFFERS2:
10250 case IORING_REGISTER_BUFFERS_UPDATE:
10251 case IORING_REGISTER_IOWQ_AFF:
10252 case IORING_UNREGISTER_IOWQ_AFF:
10259 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10263 percpu_ref_kill(&ctx->refs);
10266 * Drop uring mutex before waiting for references to exit. If another
10267 * thread is currently inside io_uring_enter() it might need to grab the
10268 * uring_lock to make progress. If we hold it here across the drain
10269 * wait, then we can deadlock. It's safe to drop the mutex here, since
10270 * no new references will come in after we've killed the percpu ref.
10272 mutex_unlock(&ctx->uring_lock);
10274 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10277 ret = io_run_task_work_sig();
10278 } while (ret >= 0);
10279 mutex_lock(&ctx->uring_lock);
10282 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10286 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10287 void __user *arg, unsigned nr_args)
10288 __releases(ctx->uring_lock)
10289 __acquires(ctx->uring_lock)
10294 * We're inside the ring mutex, if the ref is already dying, then
10295 * someone else killed the ctx or is already going through
10296 * io_uring_register().
10298 if (percpu_ref_is_dying(&ctx->refs))
10301 if (ctx->restricted) {
10302 if (opcode >= IORING_REGISTER_LAST)
10304 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10305 if (!test_bit(opcode, ctx->restrictions.register_op))
10309 if (io_register_op_must_quiesce(opcode)) {
10310 ret = io_ctx_quiesce(ctx);
10316 case IORING_REGISTER_BUFFERS:
10317 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10319 case IORING_UNREGISTER_BUFFERS:
10321 if (arg || nr_args)
10323 ret = io_sqe_buffers_unregister(ctx);
10325 case IORING_REGISTER_FILES:
10326 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10328 case IORING_UNREGISTER_FILES:
10330 if (arg || nr_args)
10332 ret = io_sqe_files_unregister(ctx);
10334 case IORING_REGISTER_FILES_UPDATE:
10335 ret = io_register_files_update(ctx, arg, nr_args);
10337 case IORING_REGISTER_EVENTFD:
10338 case IORING_REGISTER_EVENTFD_ASYNC:
10342 ret = io_eventfd_register(ctx, arg);
10345 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10346 ctx->eventfd_async = 1;
10348 ctx->eventfd_async = 0;
10350 case IORING_UNREGISTER_EVENTFD:
10352 if (arg || nr_args)
10354 ret = io_eventfd_unregister(ctx);
10356 case IORING_REGISTER_PROBE:
10358 if (!arg || nr_args > 256)
10360 ret = io_probe(ctx, arg, nr_args);
10362 case IORING_REGISTER_PERSONALITY:
10364 if (arg || nr_args)
10366 ret = io_register_personality(ctx);
10368 case IORING_UNREGISTER_PERSONALITY:
10372 ret = io_unregister_personality(ctx, nr_args);
10374 case IORING_REGISTER_ENABLE_RINGS:
10376 if (arg || nr_args)
10378 ret = io_register_enable_rings(ctx);
10380 case IORING_REGISTER_RESTRICTIONS:
10381 ret = io_register_restrictions(ctx, arg, nr_args);
10383 case IORING_REGISTER_FILES2:
10384 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10386 case IORING_REGISTER_FILES_UPDATE2:
10387 ret = io_register_rsrc_update(ctx, arg, nr_args,
10390 case IORING_REGISTER_BUFFERS2:
10391 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10393 case IORING_REGISTER_BUFFERS_UPDATE:
10394 ret = io_register_rsrc_update(ctx, arg, nr_args,
10395 IORING_RSRC_BUFFER);
10397 case IORING_REGISTER_IOWQ_AFF:
10399 if (!arg || !nr_args)
10401 ret = io_register_iowq_aff(ctx, arg, nr_args);
10403 case IORING_UNREGISTER_IOWQ_AFF:
10405 if (arg || nr_args)
10407 ret = io_unregister_iowq_aff(ctx);
10414 if (io_register_op_must_quiesce(opcode)) {
10415 /* bring the ctx back to life */
10416 percpu_ref_reinit(&ctx->refs);
10417 reinit_completion(&ctx->ref_comp);
10422 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10423 void __user *, arg, unsigned int, nr_args)
10425 struct io_ring_ctx *ctx;
10434 if (f.file->f_op != &io_uring_fops)
10437 ctx = f.file->private_data;
10439 io_run_task_work();
10441 mutex_lock(&ctx->uring_lock);
10442 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10443 mutex_unlock(&ctx->uring_lock);
10444 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10445 ctx->cq_ev_fd != NULL, ret);
10451 static int __init io_uring_init(void)
10453 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10454 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10455 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10458 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10459 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10460 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10461 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10462 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10463 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10464 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10465 BUILD_BUG_SQE_ELEM(8, __u64, off);
10466 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10467 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10468 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10469 BUILD_BUG_SQE_ELEM(24, __u32, len);
10470 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10471 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10472 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10473 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10474 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10475 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10476 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10477 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10478 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10479 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10480 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10481 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10482 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10483 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10484 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10485 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10486 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10487 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10488 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10489 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10490 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10492 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10493 sizeof(struct io_uring_rsrc_update));
10494 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10495 sizeof(struct io_uring_rsrc_update2));
10497 /* ->buf_index is u16 */
10498 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10500 /* should fit into one byte */
10501 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10503 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10504 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10506 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10510 __initcall(io_uring_init);