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_cqring (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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
205 struct io_overflow_cqe {
206 struct io_uring_cqe cqe;
207 struct list_head list;
210 struct io_fixed_file {
211 /* file * with additional FFS_* flags */
212 unsigned long file_ptr;
216 struct list_head list;
223 struct fixed_rsrc_table {
224 struct io_fixed_file *files;
227 struct io_rsrc_node {
228 struct percpu_ref refs;
229 struct list_head node;
230 struct list_head rsrc_list;
231 struct io_rsrc_data *rsrc_data;
232 struct llist_node llist;
236 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
238 struct io_rsrc_data {
239 struct fixed_rsrc_table *table;
240 struct io_ring_ctx *ctx;
243 struct percpu_ref refs;
244 struct completion done;
249 struct list_head list;
255 struct io_restriction {
256 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
257 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
258 u8 sqe_flags_allowed;
259 u8 sqe_flags_required;
264 IO_SQ_THREAD_SHOULD_STOP = 0,
265 IO_SQ_THREAD_SHOULD_PARK,
270 atomic_t park_pending;
273 /* ctx's that are using this sqd */
274 struct list_head ctx_list;
276 struct task_struct *thread;
277 struct wait_queue_head wait;
279 unsigned sq_thread_idle;
285 struct completion exited;
286 struct callback_head *park_task_work;
289 #define IO_IOPOLL_BATCH 8
290 #define IO_COMPL_BATCH 32
291 #define IO_REQ_CACHE_SIZE 32
292 #define IO_REQ_ALLOC_BATCH 8
294 struct io_comp_state {
295 struct io_kiocb *reqs[IO_COMPL_BATCH];
297 unsigned int locked_free_nr;
298 /* inline/task_work completion list, under ->uring_lock */
299 struct list_head free_list;
300 /* IRQ completion list, under ->completion_lock */
301 struct list_head locked_free_list;
304 struct io_submit_link {
305 struct io_kiocb *head;
306 struct io_kiocb *last;
309 struct io_submit_state {
310 struct blk_plug plug;
311 struct io_submit_link link;
314 * io_kiocb alloc cache
316 void *reqs[IO_REQ_CACHE_SIZE];
317 unsigned int free_reqs;
322 * Batch completion logic
324 struct io_comp_state comp;
327 * File reference cache
331 unsigned int file_refs;
332 unsigned int ios_left;
337 struct percpu_ref refs;
338 } ____cacheline_aligned_in_smp;
342 unsigned int compat: 1;
343 unsigned int drain_next: 1;
344 unsigned int eventfd_async: 1;
345 unsigned int restricted: 1;
348 * Ring buffer of indices into array of io_uring_sqe, which is
349 * mmapped by the application using the IORING_OFF_SQES offset.
351 * This indirection could e.g. be used to assign fixed
352 * io_uring_sqe entries to operations and only submit them to
353 * the queue when needed.
355 * The kernel modifies neither the indices array nor the entries
359 unsigned cached_sq_head;
362 unsigned sq_thread_idle;
363 unsigned cached_sq_dropped;
364 unsigned cached_cq_overflow;
365 unsigned long sq_check_overflow;
367 /* hashed buffered write serialization */
368 struct io_wq_hash *hash_map;
370 struct list_head defer_list;
371 struct list_head timeout_list;
372 struct list_head cq_overflow_list;
374 struct io_uring_sqe *sq_sqes;
375 } ____cacheline_aligned_in_smp;
378 struct mutex uring_lock;
379 wait_queue_head_t wait;
380 } ____cacheline_aligned_in_smp;
382 struct io_submit_state submit_state;
384 struct io_rings *rings;
386 /* Only used for accounting purposes */
387 struct mm_struct *mm_account;
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;
396 * If used, fixed file set. Writers must ensure that ->refs is dead,
397 * readers must ensure that ->refs is alive as long as the file* is
398 * used. Only updated through io_uring_register(2).
400 struct io_rsrc_data *file_data;
401 unsigned nr_user_files;
403 /* if used, fixed mapped user buffers */
404 unsigned nr_user_bufs;
405 struct io_mapped_ubuf *user_bufs;
407 struct user_struct *user;
409 struct completion ref_comp;
411 #if defined(CONFIG_UNIX)
412 struct socket *ring_sock;
415 struct xarray io_buffers;
417 struct xarray personalities;
421 unsigned cached_cq_tail;
424 atomic_t cq_timeouts;
425 unsigned cq_last_tm_flush;
426 unsigned long cq_check_overflow;
427 struct wait_queue_head cq_wait;
428 struct fasync_struct *cq_fasync;
429 struct eventfd_ctx *cq_ev_fd;
430 } ____cacheline_aligned_in_smp;
433 spinlock_t completion_lock;
436 * ->iopoll_list is protected by the ctx->uring_lock for
437 * io_uring instances that don't use IORING_SETUP_SQPOLL.
438 * For SQPOLL, only the single threaded io_sq_thread() will
439 * manipulate the list, hence no extra locking is needed there.
441 struct list_head iopoll_list;
442 struct hlist_head *cancel_hash;
443 unsigned cancel_hash_bits;
444 bool poll_multi_file;
446 spinlock_t inflight_lock;
447 struct list_head inflight_list;
448 } ____cacheline_aligned_in_smp;
450 struct delayed_work rsrc_put_work;
451 struct llist_head rsrc_put_llist;
452 struct list_head rsrc_ref_list;
453 spinlock_t rsrc_ref_lock;
454 struct io_rsrc_node *rsrc_node;
455 struct io_rsrc_node *rsrc_backup_node;
457 struct io_restriction restrictions;
460 struct callback_head *exit_task_work;
462 /* Keep this last, we don't need it for the fast path */
463 struct work_struct exit_work;
464 struct list_head tctx_list;
467 struct io_uring_task {
468 /* submission side */
470 struct wait_queue_head wait;
471 const struct io_ring_ctx *last;
473 struct percpu_counter inflight;
476 spinlock_t task_lock;
477 struct io_wq_work_list task_list;
478 unsigned long task_state;
479 struct callback_head task_work;
483 * First field must be the file pointer in all the
484 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
486 struct io_poll_iocb {
488 struct wait_queue_head *head;
493 bool update_user_data;
495 struct wait_queue_entry wait;
503 struct io_poll_remove {
513 struct io_timeout_data {
514 struct io_kiocb *req;
515 struct hrtimer timer;
516 struct timespec64 ts;
517 enum hrtimer_mode mode;
522 struct sockaddr __user *addr;
523 int __user *addr_len;
525 unsigned long nofile;
545 struct list_head list;
546 /* head of the link, used by linked timeouts only */
547 struct io_kiocb *head;
550 struct io_timeout_rem {
555 struct timespec64 ts;
560 /* NOTE: kiocb has the file as the first member, so don't do it here */
568 struct sockaddr __user *addr;
575 struct user_msghdr __user *umsg;
581 struct io_buffer *kbuf;
587 struct filename *filename;
589 unsigned long nofile;
592 struct io_rsrc_update {
618 struct epoll_event event;
622 struct file *file_out;
623 struct file *file_in;
630 struct io_provide_buf {
644 const char __user *filename;
645 struct statx __user *buffer;
657 struct filename *oldpath;
658 struct filename *newpath;
666 struct filename *filename;
669 struct io_completion {
671 struct list_head list;
675 struct io_async_connect {
676 struct sockaddr_storage address;
679 struct io_async_msghdr {
680 struct iovec fast_iov[UIO_FASTIOV];
681 /* points to an allocated iov, if NULL we use fast_iov instead */
682 struct iovec *free_iov;
683 struct sockaddr __user *uaddr;
685 struct sockaddr_storage addr;
689 struct iovec fast_iov[UIO_FASTIOV];
690 const struct iovec *free_iovec;
691 struct iov_iter iter;
693 struct wait_page_queue wpq;
697 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
698 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
699 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
700 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
701 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
702 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 REQ_F_LINK_TIMEOUT_BIT,
709 REQ_F_NEED_CLEANUP_BIT,
711 REQ_F_BUFFER_SELECTED_BIT,
712 REQ_F_LTIMEOUT_ACTIVE_BIT,
713 REQ_F_COMPLETE_INLINE_BIT,
715 REQ_F_DONT_REISSUE_BIT,
716 /* keep async read/write and isreg together and in order */
717 REQ_F_ASYNC_READ_BIT,
718 REQ_F_ASYNC_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_LINK = BIT(REQ_F_FAIL_LINK_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 /* linked timeout is active, i.e. prepared by link's head */
756 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
757 /* completion is deferred through io_comp_state */
758 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
759 /* caller should reissue async */
760 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
761 /* don't attempt request reissue, see io_rw_reissue() */
762 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
763 /* supports async reads */
764 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
765 /* supports async writes */
766 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
768 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
772 struct io_poll_iocb poll;
773 struct io_poll_iocb *double_poll;
776 struct io_task_work {
777 struct io_wq_work_node node;
778 task_work_func_t func;
782 * NOTE! Each of the iocb union members has the file pointer
783 * as the first entry in their struct definition. So you can
784 * access the file pointer through any of the sub-structs,
785 * or directly as just 'ki_filp' in this struct.
791 struct io_poll_iocb poll;
792 struct io_poll_remove poll_remove;
793 struct io_accept accept;
795 struct io_cancel cancel;
796 struct io_timeout timeout;
797 struct io_timeout_rem timeout_rem;
798 struct io_connect connect;
799 struct io_sr_msg sr_msg;
801 struct io_close close;
802 struct io_rsrc_update rsrc_update;
803 struct io_fadvise fadvise;
804 struct io_madvise madvise;
805 struct io_epoll epoll;
806 struct io_splice splice;
807 struct io_provide_buf pbuf;
808 struct io_statx statx;
809 struct io_shutdown shutdown;
810 struct io_rename rename;
811 struct io_unlink unlink;
812 /* use only after cleaning per-op data, see io_clean_op() */
813 struct io_completion compl;
816 /* opcode allocated if it needs to store data for async defer */
819 /* polled IO has completed */
825 struct io_ring_ctx *ctx;
828 struct task_struct *task;
831 struct io_kiocb *link;
832 struct percpu_ref *fixed_rsrc_refs;
835 * 1. used with ctx->iopoll_list with reads/writes
836 * 2. to track reqs with ->files (see io_op_def::file_table)
838 struct list_head inflight_entry;
840 struct io_task_work io_task_work;
841 struct callback_head task_work;
843 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
844 struct hlist_node hash_node;
845 struct async_poll *apoll;
846 struct io_wq_work work;
849 struct io_tctx_node {
850 struct list_head ctx_node;
851 struct task_struct *task;
852 struct io_ring_ctx *ctx;
855 struct io_defer_entry {
856 struct list_head list;
857 struct io_kiocb *req;
862 /* needs req->file assigned */
863 unsigned needs_file : 1;
864 /* hash wq insertion if file is a regular file */
865 unsigned hash_reg_file : 1;
866 /* unbound wq insertion if file is a non-regular file */
867 unsigned unbound_nonreg_file : 1;
868 /* opcode is not supported by this kernel */
869 unsigned not_supported : 1;
870 /* set if opcode supports polled "wait" */
872 unsigned pollout : 1;
873 /* op supports buffer selection */
874 unsigned buffer_select : 1;
875 /* do prep async if is going to be punted */
876 unsigned needs_async_setup : 1;
877 /* should block plug */
879 /* size of async data needed, if any */
880 unsigned short async_size;
883 static const struct io_op_def io_op_defs[] = {
884 [IORING_OP_NOP] = {},
885 [IORING_OP_READV] = {
887 .unbound_nonreg_file = 1,
890 .needs_async_setup = 1,
892 .async_size = sizeof(struct io_async_rw),
894 [IORING_OP_WRITEV] = {
897 .unbound_nonreg_file = 1,
899 .needs_async_setup = 1,
901 .async_size = sizeof(struct io_async_rw),
903 [IORING_OP_FSYNC] = {
906 [IORING_OP_READ_FIXED] = {
908 .unbound_nonreg_file = 1,
911 .async_size = sizeof(struct io_async_rw),
913 [IORING_OP_WRITE_FIXED] = {
916 .unbound_nonreg_file = 1,
919 .async_size = sizeof(struct io_async_rw),
921 [IORING_OP_POLL_ADD] = {
923 .unbound_nonreg_file = 1,
925 [IORING_OP_POLL_REMOVE] = {},
926 [IORING_OP_SYNC_FILE_RANGE] = {
929 [IORING_OP_SENDMSG] = {
931 .unbound_nonreg_file = 1,
933 .needs_async_setup = 1,
934 .async_size = sizeof(struct io_async_msghdr),
936 [IORING_OP_RECVMSG] = {
938 .unbound_nonreg_file = 1,
941 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_msghdr),
944 [IORING_OP_TIMEOUT] = {
945 .async_size = sizeof(struct io_timeout_data),
947 [IORING_OP_TIMEOUT_REMOVE] = {
948 /* used by timeout updates' prep() */
950 [IORING_OP_ACCEPT] = {
952 .unbound_nonreg_file = 1,
955 [IORING_OP_ASYNC_CANCEL] = {},
956 [IORING_OP_LINK_TIMEOUT] = {
957 .async_size = sizeof(struct io_timeout_data),
959 [IORING_OP_CONNECT] = {
961 .unbound_nonreg_file = 1,
963 .needs_async_setup = 1,
964 .async_size = sizeof(struct io_async_connect),
966 [IORING_OP_FALLOCATE] = {
969 [IORING_OP_OPENAT] = {},
970 [IORING_OP_CLOSE] = {},
971 [IORING_OP_FILES_UPDATE] = {},
972 [IORING_OP_STATX] = {},
975 .unbound_nonreg_file = 1,
979 .async_size = sizeof(struct io_async_rw),
981 [IORING_OP_WRITE] = {
983 .unbound_nonreg_file = 1,
986 .async_size = sizeof(struct io_async_rw),
988 [IORING_OP_FADVISE] = {
991 [IORING_OP_MADVISE] = {},
994 .unbound_nonreg_file = 1,
999 .unbound_nonreg_file = 1,
1003 [IORING_OP_OPENAT2] = {
1005 [IORING_OP_EPOLL_CTL] = {
1006 .unbound_nonreg_file = 1,
1008 [IORING_OP_SPLICE] = {
1011 .unbound_nonreg_file = 1,
1013 [IORING_OP_PROVIDE_BUFFERS] = {},
1014 [IORING_OP_REMOVE_BUFFERS] = {},
1018 .unbound_nonreg_file = 1,
1020 [IORING_OP_SHUTDOWN] = {
1023 [IORING_OP_RENAMEAT] = {},
1024 [IORING_OP_UNLINKAT] = {},
1027 static bool io_disarm_next(struct io_kiocb *req);
1028 static void io_uring_del_task_file(unsigned long index);
1029 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1030 struct task_struct *task,
1031 struct files_struct *files);
1032 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1033 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1035 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1036 static void io_put_req(struct io_kiocb *req);
1037 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1038 static void io_dismantle_req(struct io_kiocb *req);
1039 static void io_put_task(struct task_struct *task, int nr);
1040 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1041 static void io_queue_linked_timeout(struct io_kiocb *req);
1042 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1043 struct io_uring_rsrc_update *ip,
1045 static void io_clean_op(struct io_kiocb *req);
1046 static struct file *io_file_get(struct io_submit_state *state,
1047 struct io_kiocb *req, int fd, bool fixed);
1048 static void __io_queue_sqe(struct io_kiocb *req);
1049 static void io_rsrc_put_work(struct work_struct *work);
1051 static void io_req_task_queue(struct io_kiocb *req);
1052 static void io_submit_flush_completions(struct io_comp_state *cs,
1053 struct io_ring_ctx *ctx);
1054 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1055 static int io_req_prep_async(struct io_kiocb *req);
1057 static struct kmem_cache *req_cachep;
1059 static const struct file_operations io_uring_fops;
1061 struct sock *io_uring_get_socket(struct file *file)
1063 #if defined(CONFIG_UNIX)
1064 if (file->f_op == &io_uring_fops) {
1065 struct io_ring_ctx *ctx = file->private_data;
1067 return ctx->ring_sock->sk;
1072 EXPORT_SYMBOL(io_uring_get_socket);
1074 #define io_for_each_link(pos, head) \
1075 for (pos = (head); pos; pos = pos->link)
1077 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1079 struct io_ring_ctx *ctx = req->ctx;
1081 if (!req->fixed_rsrc_refs) {
1082 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1083 percpu_ref_get(req->fixed_rsrc_refs);
1087 static bool io_match_task(struct io_kiocb *head,
1088 struct task_struct *task,
1089 struct files_struct *files)
1091 struct io_kiocb *req;
1093 if (task && head->task != task)
1098 io_for_each_link(req, head) {
1099 if (req->flags & REQ_F_INFLIGHT)
1105 static inline void req_set_fail_links(struct io_kiocb *req)
1107 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1108 req->flags |= REQ_F_FAIL_LINK;
1111 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1113 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1115 complete(&ctx->ref_comp);
1118 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1120 return !req->timeout.off;
1123 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1125 struct io_ring_ctx *ctx;
1128 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1133 * Use 5 bits less than the max cq entries, that should give us around
1134 * 32 entries per hash list if totally full and uniformly spread.
1136 hash_bits = ilog2(p->cq_entries);
1140 ctx->cancel_hash_bits = hash_bits;
1141 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1143 if (!ctx->cancel_hash)
1145 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1147 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1148 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1151 ctx->flags = p->flags;
1152 init_waitqueue_head(&ctx->sqo_sq_wait);
1153 INIT_LIST_HEAD(&ctx->sqd_list);
1154 init_waitqueue_head(&ctx->cq_wait);
1155 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1156 init_completion(&ctx->ref_comp);
1157 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1158 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1159 mutex_init(&ctx->uring_lock);
1160 init_waitqueue_head(&ctx->wait);
1161 spin_lock_init(&ctx->completion_lock);
1162 INIT_LIST_HEAD(&ctx->iopoll_list);
1163 INIT_LIST_HEAD(&ctx->defer_list);
1164 INIT_LIST_HEAD(&ctx->timeout_list);
1165 spin_lock_init(&ctx->inflight_lock);
1166 INIT_LIST_HEAD(&ctx->inflight_list);
1167 spin_lock_init(&ctx->rsrc_ref_lock);
1168 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1169 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1170 init_llist_head(&ctx->rsrc_put_llist);
1171 INIT_LIST_HEAD(&ctx->tctx_list);
1172 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1173 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1176 kfree(ctx->cancel_hash);
1181 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1183 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1184 struct io_ring_ctx *ctx = req->ctx;
1186 return seq != ctx->cached_cq_tail
1187 + READ_ONCE(ctx->cached_cq_overflow);
1193 static void io_req_track_inflight(struct io_kiocb *req)
1195 struct io_ring_ctx *ctx = req->ctx;
1197 if (!(req->flags & REQ_F_INFLIGHT)) {
1198 req->flags |= REQ_F_INFLIGHT;
1200 spin_lock_irq(&ctx->inflight_lock);
1201 list_add(&req->inflight_entry, &ctx->inflight_list);
1202 spin_unlock_irq(&ctx->inflight_lock);
1206 static void io_prep_async_work(struct io_kiocb *req)
1208 const struct io_op_def *def = &io_op_defs[req->opcode];
1209 struct io_ring_ctx *ctx = req->ctx;
1211 if (!req->work.creds)
1212 req->work.creds = get_current_cred();
1214 req->work.list.next = NULL;
1215 req->work.flags = 0;
1216 if (req->flags & REQ_F_FORCE_ASYNC)
1217 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1219 if (req->flags & REQ_F_ISREG) {
1220 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1221 io_wq_hash_work(&req->work, file_inode(req->file));
1222 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1223 if (def->unbound_nonreg_file)
1224 req->work.flags |= IO_WQ_WORK_UNBOUND;
1227 switch (req->opcode) {
1228 case IORING_OP_SPLICE:
1231 * Splice operation will be punted aync, and here need to
1232 * modify io_wq_work.flags, so initialize io_wq_work firstly.
1234 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1235 req->work.flags |= IO_WQ_WORK_UNBOUND;
1240 static void io_prep_async_link(struct io_kiocb *req)
1242 struct io_kiocb *cur;
1244 io_for_each_link(cur, req)
1245 io_prep_async_work(cur);
1248 static void io_queue_async_work(struct io_kiocb *req)
1250 struct io_ring_ctx *ctx = req->ctx;
1251 struct io_kiocb *link = io_prep_linked_timeout(req);
1252 struct io_uring_task *tctx = req->task->io_uring;
1255 BUG_ON(!tctx->io_wq);
1257 /* init ->work of the whole link before punting */
1258 io_prep_async_link(req);
1259 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1260 &req->work, req->flags);
1261 io_wq_enqueue(tctx->io_wq, &req->work);
1263 io_queue_linked_timeout(link);
1266 static void io_kill_timeout(struct io_kiocb *req, int status)
1268 struct io_timeout_data *io = req->async_data;
1271 ret = hrtimer_try_to_cancel(&io->timer);
1273 atomic_set(&req->ctx->cq_timeouts,
1274 atomic_read(&req->ctx->cq_timeouts) + 1);
1275 list_del_init(&req->timeout.list);
1276 io_cqring_fill_event(req, status);
1277 io_put_req_deferred(req, 1);
1281 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1284 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1285 struct io_defer_entry, list);
1287 if (req_need_defer(de->req, de->seq))
1289 list_del_init(&de->list);
1290 io_req_task_queue(de->req);
1292 } while (!list_empty(&ctx->defer_list));
1295 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1299 if (list_empty(&ctx->timeout_list))
1302 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1305 u32 events_needed, events_got;
1306 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1307 struct io_kiocb, timeout.list);
1309 if (io_is_timeout_noseq(req))
1313 * Since seq can easily wrap around over time, subtract
1314 * the last seq at which timeouts were flushed before comparing.
1315 * Assuming not more than 2^31-1 events have happened since,
1316 * these subtractions won't have wrapped, so we can check if
1317 * target is in [last_seq, current_seq] by comparing the two.
1319 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1320 events_got = seq - ctx->cq_last_tm_flush;
1321 if (events_got < events_needed)
1324 list_del_init(&req->timeout.list);
1325 io_kill_timeout(req, 0);
1326 } while (!list_empty(&ctx->timeout_list));
1328 ctx->cq_last_tm_flush = seq;
1331 static void io_commit_cqring(struct io_ring_ctx *ctx)
1333 io_flush_timeouts(ctx);
1335 /* order cqe stores with ring update */
1336 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1338 if (unlikely(!list_empty(&ctx->defer_list)))
1339 __io_queue_deferred(ctx);
1342 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1344 struct io_rings *r = ctx->rings;
1346 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1349 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1351 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1354 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1356 struct io_rings *rings = ctx->rings;
1360 * writes to the cq entry need to come after reading head; the
1361 * control dependency is enough as we're using WRITE_ONCE to
1364 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1367 tail = ctx->cached_cq_tail++;
1368 return &rings->cqes[tail & ctx->cq_mask];
1371 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1375 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1377 if (!ctx->eventfd_async)
1379 return io_wq_current_is_worker();
1382 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1384 /* see waitqueue_active() comment */
1387 if (waitqueue_active(&ctx->wait))
1388 wake_up(&ctx->wait);
1389 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1390 wake_up(&ctx->sq_data->wait);
1391 if (io_should_trigger_evfd(ctx))
1392 eventfd_signal(ctx->cq_ev_fd, 1);
1393 if (waitqueue_active(&ctx->cq_wait)) {
1394 wake_up_interruptible(&ctx->cq_wait);
1395 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1399 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1401 /* see waitqueue_active() comment */
1404 if (ctx->flags & IORING_SETUP_SQPOLL) {
1405 if (waitqueue_active(&ctx->wait))
1406 wake_up(&ctx->wait);
1408 if (io_should_trigger_evfd(ctx))
1409 eventfd_signal(ctx->cq_ev_fd, 1);
1410 if (waitqueue_active(&ctx->cq_wait)) {
1411 wake_up_interruptible(&ctx->cq_wait);
1412 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1416 /* Returns true if there are no backlogged entries after the flush */
1417 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1419 struct io_rings *rings = ctx->rings;
1420 unsigned long flags;
1421 bool all_flushed, posted;
1423 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1427 spin_lock_irqsave(&ctx->completion_lock, flags);
1428 while (!list_empty(&ctx->cq_overflow_list)) {
1429 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1430 struct io_overflow_cqe *ocqe;
1434 ocqe = list_first_entry(&ctx->cq_overflow_list,
1435 struct io_overflow_cqe, list);
1437 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1439 WRITE_ONCE(ctx->rings->cq_overflow,
1440 ++ctx->cached_cq_overflow);
1442 list_del(&ocqe->list);
1446 all_flushed = list_empty(&ctx->cq_overflow_list);
1448 clear_bit(0, &ctx->sq_check_overflow);
1449 clear_bit(0, &ctx->cq_check_overflow);
1450 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1454 io_commit_cqring(ctx);
1455 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1457 io_cqring_ev_posted(ctx);
1461 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1465 if (test_bit(0, &ctx->cq_check_overflow)) {
1466 /* iopoll syncs against uring_lock, not completion_lock */
1467 if (ctx->flags & IORING_SETUP_IOPOLL)
1468 mutex_lock(&ctx->uring_lock);
1469 ret = __io_cqring_overflow_flush(ctx, force);
1470 if (ctx->flags & IORING_SETUP_IOPOLL)
1471 mutex_unlock(&ctx->uring_lock);
1478 * Shamelessly stolen from the mm implementation of page reference checking,
1479 * see commit f958d7b528b1 for details.
1481 #define req_ref_zero_or_close_to_overflow(req) \
1482 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1484 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1486 return atomic_inc_not_zero(&req->refs);
1489 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1491 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1492 return atomic_sub_and_test(refs, &req->refs);
1495 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1497 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1498 return atomic_dec_and_test(&req->refs);
1501 static inline void req_ref_put(struct io_kiocb *req)
1503 WARN_ON_ONCE(req_ref_put_and_test(req));
1506 static inline void req_ref_get(struct io_kiocb *req)
1508 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1509 atomic_inc(&req->refs);
1512 static bool __io_cqring_fill_event(struct io_kiocb *req, long res,
1513 unsigned int cflags)
1515 struct io_ring_ctx *ctx = req->ctx;
1516 struct io_uring_cqe *cqe;
1518 trace_io_uring_complete(ctx, req->user_data, res, cflags);
1521 * If we can't get a cq entry, userspace overflowed the
1522 * submission (by quite a lot). Increment the overflow count in
1525 cqe = io_get_cqring(ctx);
1527 WRITE_ONCE(cqe->user_data, req->user_data);
1528 WRITE_ONCE(cqe->res, res);
1529 WRITE_ONCE(cqe->flags, cflags);
1532 if (!atomic_read(&req->task->io_uring->in_idle)) {
1533 struct io_overflow_cqe *ocqe;
1535 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1538 if (list_empty(&ctx->cq_overflow_list)) {
1539 set_bit(0, &ctx->sq_check_overflow);
1540 set_bit(0, &ctx->cq_check_overflow);
1541 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1543 ocqe->cqe.user_data = req->user_data;
1544 ocqe->cqe.res = res;
1545 ocqe->cqe.flags = cflags;
1546 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1551 * If we're in ring overflow flush mode, or in task cancel mode,
1552 * or cannot allocate an overflow entry, then we need to drop it
1555 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1559 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1561 __io_cqring_fill_event(req, res, 0);
1564 static void io_req_complete_post(struct io_kiocb *req, long res,
1565 unsigned int cflags)
1567 struct io_ring_ctx *ctx = req->ctx;
1568 unsigned long flags;
1570 spin_lock_irqsave(&ctx->completion_lock, flags);
1571 __io_cqring_fill_event(req, res, cflags);
1573 * If we're the last reference to this request, add to our locked
1576 if (req_ref_put_and_test(req)) {
1577 struct io_comp_state *cs = &ctx->submit_state.comp;
1579 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1580 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1581 io_disarm_next(req);
1583 io_req_task_queue(req->link);
1587 io_dismantle_req(req);
1588 io_put_task(req->task, 1);
1589 list_add(&req->compl.list, &cs->locked_free_list);
1590 cs->locked_free_nr++;
1592 if (!percpu_ref_tryget(&ctx->refs))
1595 io_commit_cqring(ctx);
1596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1599 io_cqring_ev_posted(ctx);
1600 percpu_ref_put(&ctx->refs);
1604 static void io_req_complete_state(struct io_kiocb *req, long res,
1605 unsigned int cflags)
1607 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1610 req->compl.cflags = cflags;
1611 req->flags |= REQ_F_COMPLETE_INLINE;
1614 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1615 long res, unsigned cflags)
1617 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1618 io_req_complete_state(req, res, cflags);
1620 io_req_complete_post(req, res, cflags);
1623 static inline void io_req_complete(struct io_kiocb *req, long res)
1625 __io_req_complete(req, 0, res, 0);
1628 static void io_req_complete_failed(struct io_kiocb *req, long res)
1630 req_set_fail_links(req);
1632 io_req_complete_post(req, res, 0);
1635 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1636 struct io_comp_state *cs)
1638 spin_lock_irq(&ctx->completion_lock);
1639 list_splice_init(&cs->locked_free_list, &cs->free_list);
1640 cs->locked_free_nr = 0;
1641 spin_unlock_irq(&ctx->completion_lock);
1644 /* Returns true IFF there are requests in the cache */
1645 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1647 struct io_submit_state *state = &ctx->submit_state;
1648 struct io_comp_state *cs = &state->comp;
1652 * If we have more than a batch's worth of requests in our IRQ side
1653 * locked cache, grab the lock and move them over to our submission
1656 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1657 io_flush_cached_locked_reqs(ctx, cs);
1659 nr = state->free_reqs;
1660 while (!list_empty(&cs->free_list)) {
1661 struct io_kiocb *req = list_first_entry(&cs->free_list,
1662 struct io_kiocb, compl.list);
1664 list_del(&req->compl.list);
1665 state->reqs[nr++] = req;
1666 if (nr == ARRAY_SIZE(state->reqs))
1670 state->free_reqs = nr;
1674 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1676 struct io_submit_state *state = &ctx->submit_state;
1678 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1680 if (!state->free_reqs) {
1681 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1684 if (io_flush_cached_reqs(ctx))
1687 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1691 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1692 * retry single alloc to be on the safe side.
1694 if (unlikely(ret <= 0)) {
1695 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1696 if (!state->reqs[0])
1700 state->free_reqs = ret;
1704 return state->reqs[state->free_reqs];
1707 static inline void io_put_file(struct file *file)
1713 static void io_dismantle_req(struct io_kiocb *req)
1715 unsigned int flags = req->flags;
1717 if (!(flags & REQ_F_FIXED_FILE))
1718 io_put_file(req->file);
1719 if (flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1723 if (req->flags & REQ_F_INFLIGHT) {
1724 struct io_ring_ctx *ctx = req->ctx;
1725 unsigned long flags;
1727 spin_lock_irqsave(&ctx->inflight_lock, flags);
1728 list_del(&req->inflight_entry);
1729 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1730 req->flags &= ~REQ_F_INFLIGHT;
1733 if (req->fixed_rsrc_refs)
1734 percpu_ref_put(req->fixed_rsrc_refs);
1735 if (req->async_data)
1736 kfree(req->async_data);
1737 if (req->work.creds) {
1738 put_cred(req->work.creds);
1739 req->work.creds = NULL;
1743 /* must to be called somewhat shortly after putting a request */
1744 static inline void io_put_task(struct task_struct *task, int nr)
1746 struct io_uring_task *tctx = task->io_uring;
1748 percpu_counter_sub(&tctx->inflight, nr);
1749 if (unlikely(atomic_read(&tctx->in_idle)))
1750 wake_up(&tctx->wait);
1751 put_task_struct_many(task, nr);
1754 static void __io_free_req(struct io_kiocb *req)
1756 struct io_ring_ctx *ctx = req->ctx;
1758 io_dismantle_req(req);
1759 io_put_task(req->task, 1);
1761 kmem_cache_free(req_cachep, req);
1762 percpu_ref_put(&ctx->refs);
1765 static inline void io_remove_next_linked(struct io_kiocb *req)
1767 struct io_kiocb *nxt = req->link;
1769 req->link = nxt->link;
1773 static bool io_kill_linked_timeout(struct io_kiocb *req)
1774 __must_hold(&req->ctx->completion_lock)
1776 struct io_kiocb *link = req->link;
1779 * Can happen if a linked timeout fired and link had been like
1780 * req -> link t-out -> link t-out [-> ...]
1782 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1783 struct io_timeout_data *io = link->async_data;
1786 io_remove_next_linked(req);
1787 link->timeout.head = NULL;
1788 ret = hrtimer_try_to_cancel(&io->timer);
1790 io_cqring_fill_event(link, -ECANCELED);
1791 io_put_req_deferred(link, 1);
1798 static void io_fail_links(struct io_kiocb *req)
1799 __must_hold(&req->ctx->completion_lock)
1801 struct io_kiocb *nxt, *link = req->link;
1808 trace_io_uring_fail_link(req, link);
1809 io_cqring_fill_event(link, -ECANCELED);
1810 io_put_req_deferred(link, 2);
1815 static bool io_disarm_next(struct io_kiocb *req)
1816 __must_hold(&req->ctx->completion_lock)
1818 bool posted = false;
1820 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1821 posted = io_kill_linked_timeout(req);
1822 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1823 posted |= (req->link != NULL);
1829 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1831 struct io_kiocb *nxt;
1834 * If LINK is set, we have dependent requests in this chain. If we
1835 * didn't fail this request, queue the first one up, moving any other
1836 * dependencies to the next request. In case of failure, fail the rest
1839 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1840 struct io_ring_ctx *ctx = req->ctx;
1841 unsigned long flags;
1844 spin_lock_irqsave(&ctx->completion_lock, flags);
1845 posted = io_disarm_next(req);
1847 io_commit_cqring(req->ctx);
1848 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1850 io_cqring_ev_posted(ctx);
1857 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1859 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1861 return __io_req_find_next(req);
1864 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1868 if (ctx->submit_state.comp.nr) {
1869 mutex_lock(&ctx->uring_lock);
1870 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1871 mutex_unlock(&ctx->uring_lock);
1873 percpu_ref_put(&ctx->refs);
1876 static bool __tctx_task_work(struct io_uring_task *tctx)
1878 struct io_ring_ctx *ctx = NULL;
1879 struct io_wq_work_list list;
1880 struct io_wq_work_node *node;
1882 if (wq_list_empty(&tctx->task_list))
1885 spin_lock_irq(&tctx->task_lock);
1886 list = tctx->task_list;
1887 INIT_WQ_LIST(&tctx->task_list);
1888 spin_unlock_irq(&tctx->task_lock);
1892 struct io_wq_work_node *next = node->next;
1893 struct io_kiocb *req;
1895 req = container_of(node, struct io_kiocb, io_task_work.node);
1896 if (req->ctx != ctx) {
1897 ctx_flush_and_put(ctx);
1899 percpu_ref_get(&ctx->refs);
1902 req->task_work.func(&req->task_work);
1906 ctx_flush_and_put(ctx);
1907 return list.first != NULL;
1910 static void tctx_task_work(struct callback_head *cb)
1912 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1914 clear_bit(0, &tctx->task_state);
1916 while (__tctx_task_work(tctx))
1920 static int io_req_task_work_add(struct io_kiocb *req)
1922 struct task_struct *tsk = req->task;
1923 struct io_uring_task *tctx = tsk->io_uring;
1924 enum task_work_notify_mode notify;
1925 struct io_wq_work_node *node, *prev;
1926 unsigned long flags;
1929 if (unlikely(tsk->flags & PF_EXITING))
1932 WARN_ON_ONCE(!tctx);
1934 spin_lock_irqsave(&tctx->task_lock, flags);
1935 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1936 spin_unlock_irqrestore(&tctx->task_lock, flags);
1938 /* task_work already pending, we're done */
1939 if (test_bit(0, &tctx->task_state) ||
1940 test_and_set_bit(0, &tctx->task_state))
1944 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1945 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1946 * processing task_work. There's no reliable way to tell if TWA_RESUME
1949 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1951 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1952 wake_up_process(tsk);
1957 * Slow path - we failed, find and delete work. if the work is not
1958 * in the list, it got run and we're fine.
1960 spin_lock_irqsave(&tctx->task_lock, flags);
1961 wq_list_for_each(node, prev, &tctx->task_list) {
1962 if (&req->io_task_work.node == node) {
1963 wq_list_del(&tctx->task_list, node, prev);
1968 spin_unlock_irqrestore(&tctx->task_lock, flags);
1969 clear_bit(0, &tctx->task_state);
1973 static bool io_run_task_work_head(struct callback_head **work_head)
1975 struct callback_head *work, *next;
1976 bool executed = false;
1979 work = xchg(work_head, NULL);
1995 static void io_task_work_add_head(struct callback_head **work_head,
1996 struct callback_head *task_work)
1998 struct callback_head *head;
2001 head = READ_ONCE(*work_head);
2002 task_work->next = head;
2003 } while (cmpxchg(work_head, head, task_work) != head);
2006 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2007 task_work_func_t cb)
2009 init_task_work(&req->task_work, cb);
2010 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2013 static void io_req_task_cancel(struct callback_head *cb)
2015 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2016 struct io_ring_ctx *ctx = req->ctx;
2018 /* ctx is guaranteed to stay alive while we hold uring_lock */
2019 mutex_lock(&ctx->uring_lock);
2020 io_req_complete_failed(req, req->result);
2021 mutex_unlock(&ctx->uring_lock);
2024 static void __io_req_task_submit(struct io_kiocb *req)
2026 struct io_ring_ctx *ctx = req->ctx;
2028 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2029 mutex_lock(&ctx->uring_lock);
2030 if (!(current->flags & PF_EXITING) && !current->in_execve)
2031 __io_queue_sqe(req);
2033 io_req_complete_failed(req, -EFAULT);
2034 mutex_unlock(&ctx->uring_lock);
2037 static void io_req_task_submit(struct callback_head *cb)
2039 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2041 __io_req_task_submit(req);
2044 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2047 req->task_work.func = io_req_task_cancel;
2049 if (unlikely(io_req_task_work_add(req)))
2050 io_req_task_work_add_fallback(req, io_req_task_cancel);
2053 static void io_req_task_queue(struct io_kiocb *req)
2055 req->task_work.func = io_req_task_submit;
2057 if (unlikely(io_req_task_work_add(req)))
2058 io_req_task_queue_fail(req, -ECANCELED);
2061 static inline void io_queue_next(struct io_kiocb *req)
2063 struct io_kiocb *nxt = io_req_find_next(req);
2066 io_req_task_queue(nxt);
2069 static void io_free_req(struct io_kiocb *req)
2076 struct task_struct *task;
2081 static inline void io_init_req_batch(struct req_batch *rb)
2088 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2089 struct req_batch *rb)
2092 io_put_task(rb->task, rb->task_refs);
2094 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2097 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2098 struct io_submit_state *state)
2101 io_dismantle_req(req);
2103 if (req->task != rb->task) {
2105 io_put_task(rb->task, rb->task_refs);
2106 rb->task = req->task;
2112 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2113 state->reqs[state->free_reqs++] = req;
2115 list_add(&req->compl.list, &state->comp.free_list);
2118 static void io_submit_flush_completions(struct io_comp_state *cs,
2119 struct io_ring_ctx *ctx)
2122 struct io_kiocb *req;
2123 struct req_batch rb;
2125 io_init_req_batch(&rb);
2126 spin_lock_irq(&ctx->completion_lock);
2127 for (i = 0; i < nr; i++) {
2129 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2131 io_commit_cqring(ctx);
2132 spin_unlock_irq(&ctx->completion_lock);
2134 io_cqring_ev_posted(ctx);
2135 for (i = 0; i < nr; i++) {
2138 /* submission and completion refs */
2139 if (req_ref_sub_and_test(req, 2))
2140 io_req_free_batch(&rb, req, &ctx->submit_state);
2143 io_req_free_batch_finish(ctx, &rb);
2148 * Drop reference to request, return next in chain (if there is one) if this
2149 * was the last reference to this request.
2151 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2153 struct io_kiocb *nxt = NULL;
2155 if (req_ref_put_and_test(req)) {
2156 nxt = io_req_find_next(req);
2162 static inline void io_put_req(struct io_kiocb *req)
2164 if (req_ref_put_and_test(req))
2168 static void io_put_req_deferred_cb(struct callback_head *cb)
2170 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2175 static void io_free_req_deferred(struct io_kiocb *req)
2177 req->task_work.func = io_put_req_deferred_cb;
2178 if (unlikely(io_req_task_work_add(req)))
2179 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2182 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2184 if (req_ref_sub_and_test(req, refs))
2185 io_free_req_deferred(req);
2188 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2190 /* See comment at the top of this file */
2192 return __io_cqring_events(ctx);
2195 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2197 struct io_rings *rings = ctx->rings;
2199 /* make sure SQ entry isn't read before tail */
2200 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2203 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2205 unsigned int cflags;
2207 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2208 cflags |= IORING_CQE_F_BUFFER;
2209 req->flags &= ~REQ_F_BUFFER_SELECTED;
2214 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2216 struct io_buffer *kbuf;
2218 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2219 return io_put_kbuf(req, kbuf);
2222 static inline bool io_run_task_work(void)
2225 * Not safe to run on exiting task, and the task_work handling will
2226 * not add work to such a task.
2228 if (unlikely(current->flags & PF_EXITING))
2230 if (current->task_works) {
2231 __set_current_state(TASK_RUNNING);
2240 * Find and free completed poll iocbs
2242 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2243 struct list_head *done)
2245 struct req_batch rb;
2246 struct io_kiocb *req;
2248 /* order with ->result store in io_complete_rw_iopoll() */
2251 io_init_req_batch(&rb);
2252 while (!list_empty(done)) {
2255 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2256 list_del(&req->inflight_entry);
2258 if (READ_ONCE(req->result) == -EAGAIN &&
2259 !(req->flags & REQ_F_DONT_REISSUE)) {
2260 req->iopoll_completed = 0;
2262 io_queue_async_work(req);
2266 if (req->flags & REQ_F_BUFFER_SELECTED)
2267 cflags = io_put_rw_kbuf(req);
2269 __io_cqring_fill_event(req, req->result, cflags);
2272 if (req_ref_put_and_test(req))
2273 io_req_free_batch(&rb, req, &ctx->submit_state);
2276 io_commit_cqring(ctx);
2277 io_cqring_ev_posted_iopoll(ctx);
2278 io_req_free_batch_finish(ctx, &rb);
2281 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2284 struct io_kiocb *req, *tmp;
2290 * Only spin for completions if we don't have multiple devices hanging
2291 * off our complete list, and we're under the requested amount.
2293 spin = !ctx->poll_multi_file && *nr_events < min;
2296 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2297 struct kiocb *kiocb = &req->rw.kiocb;
2300 * Move completed and retryable entries to our local lists.
2301 * If we find a request that requires polling, break out
2302 * and complete those lists first, if we have entries there.
2304 if (READ_ONCE(req->iopoll_completed)) {
2305 list_move_tail(&req->inflight_entry, &done);
2308 if (!list_empty(&done))
2311 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2315 /* iopoll may have completed current req */
2316 if (READ_ONCE(req->iopoll_completed))
2317 list_move_tail(&req->inflight_entry, &done);
2324 if (!list_empty(&done))
2325 io_iopoll_complete(ctx, nr_events, &done);
2331 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2332 * non-spinning poll check - we'll still enter the driver poll loop, but only
2333 * as a non-spinning completion check.
2335 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2338 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2341 ret = io_do_iopoll(ctx, nr_events, min);
2344 if (*nr_events >= min)
2352 * We can't just wait for polled events to come to us, we have to actively
2353 * find and complete them.
2355 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2357 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2360 mutex_lock(&ctx->uring_lock);
2361 while (!list_empty(&ctx->iopoll_list)) {
2362 unsigned int nr_events = 0;
2364 io_do_iopoll(ctx, &nr_events, 0);
2366 /* let it sleep and repeat later if can't complete a request */
2370 * Ensure we allow local-to-the-cpu processing to take place,
2371 * in this case we need to ensure that we reap all events.
2372 * Also let task_work, etc. to progress by releasing the mutex
2374 if (need_resched()) {
2375 mutex_unlock(&ctx->uring_lock);
2377 mutex_lock(&ctx->uring_lock);
2380 mutex_unlock(&ctx->uring_lock);
2383 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2385 unsigned int nr_events = 0;
2386 int iters = 0, ret = 0;
2389 * We disallow the app entering submit/complete with polling, but we
2390 * still need to lock the ring to prevent racing with polled issue
2391 * that got punted to a workqueue.
2393 mutex_lock(&ctx->uring_lock);
2396 * Don't enter poll loop if we already have events pending.
2397 * If we do, we can potentially be spinning for commands that
2398 * already triggered a CQE (eg in error).
2400 if (test_bit(0, &ctx->cq_check_overflow))
2401 __io_cqring_overflow_flush(ctx, false);
2402 if (io_cqring_events(ctx))
2406 * If a submit got punted to a workqueue, we can have the
2407 * application entering polling for a command before it gets
2408 * issued. That app will hold the uring_lock for the duration
2409 * of the poll right here, so we need to take a breather every
2410 * now and then to ensure that the issue has a chance to add
2411 * the poll to the issued list. Otherwise we can spin here
2412 * forever, while the workqueue is stuck trying to acquire the
2415 if (!(++iters & 7)) {
2416 mutex_unlock(&ctx->uring_lock);
2418 mutex_lock(&ctx->uring_lock);
2421 ret = io_iopoll_getevents(ctx, &nr_events, min);
2425 } while (min && !nr_events && !need_resched());
2427 mutex_unlock(&ctx->uring_lock);
2431 static void kiocb_end_write(struct io_kiocb *req)
2434 * Tell lockdep we inherited freeze protection from submission
2437 if (req->flags & REQ_F_ISREG) {
2438 struct super_block *sb = file_inode(req->file)->i_sb;
2440 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2446 static bool io_resubmit_prep(struct io_kiocb *req)
2448 struct io_async_rw *rw = req->async_data;
2451 return !io_req_prep_async(req);
2452 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2453 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2457 static bool io_rw_should_reissue(struct io_kiocb *req)
2459 umode_t mode = file_inode(req->file)->i_mode;
2460 struct io_ring_ctx *ctx = req->ctx;
2462 if (!S_ISBLK(mode) && !S_ISREG(mode))
2464 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2465 !(ctx->flags & IORING_SETUP_IOPOLL)))
2468 * If ref is dying, we might be running poll reap from the exit work.
2469 * Don't attempt to reissue from that path, just let it fail with
2472 if (percpu_ref_is_dying(&ctx->refs))
2477 static bool io_rw_should_reissue(struct io_kiocb *req)
2483 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2484 unsigned int issue_flags)
2488 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2489 kiocb_end_write(req);
2490 if (res != req->result) {
2491 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2492 io_rw_should_reissue(req)) {
2493 req->flags |= REQ_F_REISSUE;
2496 req_set_fail_links(req);
2498 if (req->flags & REQ_F_BUFFER_SELECTED)
2499 cflags = io_put_rw_kbuf(req);
2500 __io_req_complete(req, issue_flags, res, cflags);
2503 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2505 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2507 __io_complete_rw(req, res, res2, 0);
2510 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2512 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2514 if (kiocb->ki_flags & IOCB_WRITE)
2515 kiocb_end_write(req);
2516 if (unlikely(res != req->result)) {
2520 if (res == -EAGAIN && io_rw_should_reissue(req) &&
2521 io_resubmit_prep(req))
2525 req_set_fail_links(req);
2526 req->flags |= REQ_F_DONT_REISSUE;
2530 WRITE_ONCE(req->result, res);
2531 /* order with io_iopoll_complete() checking ->result */
2533 WRITE_ONCE(req->iopoll_completed, 1);
2537 * After the iocb has been issued, it's safe to be found on the poll list.
2538 * Adding the kiocb to the list AFTER submission ensures that we don't
2539 * find it from a io_iopoll_getevents() thread before the issuer is done
2540 * accessing the kiocb cookie.
2542 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2544 struct io_ring_ctx *ctx = req->ctx;
2547 * Track whether we have multiple files in our lists. This will impact
2548 * how we do polling eventually, not spinning if we're on potentially
2549 * different devices.
2551 if (list_empty(&ctx->iopoll_list)) {
2552 ctx->poll_multi_file = false;
2553 } else if (!ctx->poll_multi_file) {
2554 struct io_kiocb *list_req;
2556 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2558 if (list_req->file != req->file)
2559 ctx->poll_multi_file = true;
2563 * For fast devices, IO may have already completed. If it has, add
2564 * it to the front so we find it first.
2566 if (READ_ONCE(req->iopoll_completed))
2567 list_add(&req->inflight_entry, &ctx->iopoll_list);
2569 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2572 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2573 * task context or in io worker task context. If current task context is
2574 * sq thread, we don't need to check whether should wake up sq thread.
2576 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2577 wq_has_sleeper(&ctx->sq_data->wait))
2578 wake_up(&ctx->sq_data->wait);
2581 static inline void io_state_file_put(struct io_submit_state *state)
2583 if (state->file_refs) {
2584 fput_many(state->file, state->file_refs);
2585 state->file_refs = 0;
2590 * Get as many references to a file as we have IOs left in this submission,
2591 * assuming most submissions are for one file, or at least that each file
2592 * has more than one submission.
2594 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2599 if (state->file_refs) {
2600 if (state->fd == fd) {
2604 io_state_file_put(state);
2606 state->file = fget_many(fd, state->ios_left);
2607 if (unlikely(!state->file))
2611 state->file_refs = state->ios_left - 1;
2615 static bool io_bdev_nowait(struct block_device *bdev)
2617 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2621 * If we tracked the file through the SCM inflight mechanism, we could support
2622 * any file. For now, just ensure that anything potentially problematic is done
2625 static bool __io_file_supports_async(struct file *file, int rw)
2627 umode_t mode = file_inode(file)->i_mode;
2629 if (S_ISBLK(mode)) {
2630 if (IS_ENABLED(CONFIG_BLOCK) &&
2631 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2635 if (S_ISCHR(mode) || S_ISSOCK(mode))
2637 if (S_ISREG(mode)) {
2638 if (IS_ENABLED(CONFIG_BLOCK) &&
2639 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2640 file->f_op != &io_uring_fops)
2645 /* any ->read/write should understand O_NONBLOCK */
2646 if (file->f_flags & O_NONBLOCK)
2649 if (!(file->f_mode & FMODE_NOWAIT))
2653 return file->f_op->read_iter != NULL;
2655 return file->f_op->write_iter != NULL;
2658 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2660 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2662 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2665 return __io_file_supports_async(req->file, rw);
2668 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2670 struct io_ring_ctx *ctx = req->ctx;
2671 struct kiocb *kiocb = &req->rw.kiocb;
2672 struct file *file = req->file;
2676 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2677 req->flags |= REQ_F_ISREG;
2679 kiocb->ki_pos = READ_ONCE(sqe->off);
2680 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2681 req->flags |= REQ_F_CUR_POS;
2682 kiocb->ki_pos = file->f_pos;
2684 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2685 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2686 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2690 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2691 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2692 req->flags |= REQ_F_NOWAIT;
2694 ioprio = READ_ONCE(sqe->ioprio);
2696 ret = ioprio_check_cap(ioprio);
2700 kiocb->ki_ioprio = ioprio;
2702 kiocb->ki_ioprio = get_current_ioprio();
2704 if (ctx->flags & IORING_SETUP_IOPOLL) {
2705 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2706 !kiocb->ki_filp->f_op->iopoll)
2709 kiocb->ki_flags |= IOCB_HIPRI;
2710 kiocb->ki_complete = io_complete_rw_iopoll;
2711 req->iopoll_completed = 0;
2713 if (kiocb->ki_flags & IOCB_HIPRI)
2715 kiocb->ki_complete = io_complete_rw;
2718 req->rw.addr = READ_ONCE(sqe->addr);
2719 req->rw.len = READ_ONCE(sqe->len);
2720 req->buf_index = READ_ONCE(sqe->buf_index);
2724 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2730 case -ERESTARTNOINTR:
2731 case -ERESTARTNOHAND:
2732 case -ERESTART_RESTARTBLOCK:
2734 * We can't just restart the syscall, since previously
2735 * submitted sqes may already be in progress. Just fail this
2741 kiocb->ki_complete(kiocb, ret, 0);
2745 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2746 unsigned int issue_flags)
2748 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2749 struct io_async_rw *io = req->async_data;
2750 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2752 /* add previously done IO, if any */
2753 if (io && io->bytes_done > 0) {
2755 ret = io->bytes_done;
2757 ret += io->bytes_done;
2760 if (req->flags & REQ_F_CUR_POS)
2761 req->file->f_pos = kiocb->ki_pos;
2762 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2763 __io_complete_rw(req, ret, 0, issue_flags);
2765 io_rw_done(kiocb, ret);
2767 if (check_reissue && req->flags & REQ_F_REISSUE) {
2768 req->flags &= ~REQ_F_REISSUE;
2769 if (!io_resubmit_prep(req)) {
2771 io_queue_async_work(req);
2775 req_set_fail_links(req);
2776 if (req->flags & REQ_F_BUFFER_SELECTED)
2777 cflags = io_put_rw_kbuf(req);
2778 __io_req_complete(req, issue_flags, ret, cflags);
2783 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2785 struct io_ring_ctx *ctx = req->ctx;
2786 size_t len = req->rw.len;
2787 struct io_mapped_ubuf *imu;
2788 u16 index, buf_index = req->buf_index;
2789 u64 buf_end, buf_addr = req->rw.addr;
2792 if (unlikely(buf_index >= ctx->nr_user_bufs))
2794 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2795 imu = &ctx->user_bufs[index];
2796 buf_addr = req->rw.addr;
2798 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2800 /* not inside the mapped region */
2801 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2805 * May not be a start of buffer, set size appropriately
2806 * and advance us to the beginning.
2808 offset = buf_addr - imu->ubuf;
2809 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2813 * Don't use iov_iter_advance() here, as it's really slow for
2814 * using the latter parts of a big fixed buffer - it iterates
2815 * over each segment manually. We can cheat a bit here, because
2818 * 1) it's a BVEC iter, we set it up
2819 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2820 * first and last bvec
2822 * So just find our index, and adjust the iterator afterwards.
2823 * If the offset is within the first bvec (or the whole first
2824 * bvec, just use iov_iter_advance(). This makes it easier
2825 * since we can just skip the first segment, which may not
2826 * be PAGE_SIZE aligned.
2828 const struct bio_vec *bvec = imu->bvec;
2830 if (offset <= bvec->bv_len) {
2831 iov_iter_advance(iter, offset);
2833 unsigned long seg_skip;
2835 /* skip first vec */
2836 offset -= bvec->bv_len;
2837 seg_skip = 1 + (offset >> PAGE_SHIFT);
2839 iter->bvec = bvec + seg_skip;
2840 iter->nr_segs -= seg_skip;
2841 iter->count -= bvec->bv_len + offset;
2842 iter->iov_offset = offset & ~PAGE_MASK;
2849 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2852 mutex_unlock(&ctx->uring_lock);
2855 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2858 * "Normal" inline submissions always hold the uring_lock, since we
2859 * grab it from the system call. Same is true for the SQPOLL offload.
2860 * The only exception is when we've detached the request and issue it
2861 * from an async worker thread, grab the lock for that case.
2864 mutex_lock(&ctx->uring_lock);
2867 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2868 int bgid, struct io_buffer *kbuf,
2871 struct io_buffer *head;
2873 if (req->flags & REQ_F_BUFFER_SELECTED)
2876 io_ring_submit_lock(req->ctx, needs_lock);
2878 lockdep_assert_held(&req->ctx->uring_lock);
2880 head = xa_load(&req->ctx->io_buffers, bgid);
2882 if (!list_empty(&head->list)) {
2883 kbuf = list_last_entry(&head->list, struct io_buffer,
2885 list_del(&kbuf->list);
2888 xa_erase(&req->ctx->io_buffers, bgid);
2890 if (*len > kbuf->len)
2893 kbuf = ERR_PTR(-ENOBUFS);
2896 io_ring_submit_unlock(req->ctx, needs_lock);
2901 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2904 struct io_buffer *kbuf;
2907 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2908 bgid = req->buf_index;
2909 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2912 req->rw.addr = (u64) (unsigned long) kbuf;
2913 req->flags |= REQ_F_BUFFER_SELECTED;
2914 return u64_to_user_ptr(kbuf->addr);
2917 #ifdef CONFIG_COMPAT
2918 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2921 struct compat_iovec __user *uiov;
2922 compat_ssize_t clen;
2926 uiov = u64_to_user_ptr(req->rw.addr);
2927 if (!access_ok(uiov, sizeof(*uiov)))
2929 if (__get_user(clen, &uiov->iov_len))
2935 buf = io_rw_buffer_select(req, &len, needs_lock);
2937 return PTR_ERR(buf);
2938 iov[0].iov_base = buf;
2939 iov[0].iov_len = (compat_size_t) len;
2944 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2947 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2951 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2954 len = iov[0].iov_len;
2957 buf = io_rw_buffer_select(req, &len, needs_lock);
2959 return PTR_ERR(buf);
2960 iov[0].iov_base = buf;
2961 iov[0].iov_len = len;
2965 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2968 if (req->flags & REQ_F_BUFFER_SELECTED) {
2969 struct io_buffer *kbuf;
2971 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2972 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2973 iov[0].iov_len = kbuf->len;
2976 if (req->rw.len != 1)
2979 #ifdef CONFIG_COMPAT
2980 if (req->ctx->compat)
2981 return io_compat_import(req, iov, needs_lock);
2984 return __io_iov_buffer_select(req, iov, needs_lock);
2987 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2988 struct iov_iter *iter, bool needs_lock)
2990 void __user *buf = u64_to_user_ptr(req->rw.addr);
2991 size_t sqe_len = req->rw.len;
2992 u8 opcode = req->opcode;
2995 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2997 return io_import_fixed(req, rw, iter);
3000 /* buffer index only valid with fixed read/write, or buffer select */
3001 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3004 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3005 if (req->flags & REQ_F_BUFFER_SELECT) {
3006 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3008 return PTR_ERR(buf);
3009 req->rw.len = sqe_len;
3012 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3017 if (req->flags & REQ_F_BUFFER_SELECT) {
3018 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3020 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3025 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3029 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3031 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3035 * For files that don't have ->read_iter() and ->write_iter(), handle them
3036 * by looping over ->read() or ->write() manually.
3038 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3040 struct kiocb *kiocb = &req->rw.kiocb;
3041 struct file *file = req->file;
3045 * Don't support polled IO through this interface, and we can't
3046 * support non-blocking either. For the latter, this just causes
3047 * the kiocb to be handled from an async context.
3049 if (kiocb->ki_flags & IOCB_HIPRI)
3051 if (kiocb->ki_flags & IOCB_NOWAIT)
3054 while (iov_iter_count(iter)) {
3058 if (!iov_iter_is_bvec(iter)) {
3059 iovec = iov_iter_iovec(iter);
3061 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3062 iovec.iov_len = req->rw.len;
3066 nr = file->f_op->read(file, iovec.iov_base,
3067 iovec.iov_len, io_kiocb_ppos(kiocb));
3069 nr = file->f_op->write(file, iovec.iov_base,
3070 iovec.iov_len, io_kiocb_ppos(kiocb));
3079 if (nr != iovec.iov_len)
3083 iov_iter_advance(iter, nr);
3089 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3090 const struct iovec *fast_iov, struct iov_iter *iter)
3092 struct io_async_rw *rw = req->async_data;
3094 memcpy(&rw->iter, iter, sizeof(*iter));
3095 rw->free_iovec = iovec;
3097 /* can only be fixed buffers, no need to do anything */
3098 if (iov_iter_is_bvec(iter))
3101 unsigned iov_off = 0;
3103 rw->iter.iov = rw->fast_iov;
3104 if (iter->iov != fast_iov) {
3105 iov_off = iter->iov - fast_iov;
3106 rw->iter.iov += iov_off;
3108 if (rw->fast_iov != fast_iov)
3109 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3110 sizeof(struct iovec) * iter->nr_segs);
3112 req->flags |= REQ_F_NEED_CLEANUP;
3116 static inline int io_alloc_async_data(struct io_kiocb *req)
3118 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3119 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3120 return req->async_data == NULL;
3123 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3124 const struct iovec *fast_iov,
3125 struct iov_iter *iter, bool force)
3127 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3129 if (!req->async_data) {
3130 if (io_alloc_async_data(req)) {
3135 io_req_map_rw(req, iovec, fast_iov, iter);
3140 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3142 struct io_async_rw *iorw = req->async_data;
3143 struct iovec *iov = iorw->fast_iov;
3146 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3147 if (unlikely(ret < 0))
3150 iorw->bytes_done = 0;
3151 iorw->free_iovec = iov;
3153 req->flags |= REQ_F_NEED_CLEANUP;
3157 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3159 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3161 return io_prep_rw(req, sqe);
3165 * This is our waitqueue callback handler, registered through lock_page_async()
3166 * when we initially tried to do the IO with the iocb armed our waitqueue.
3167 * This gets called when the page is unlocked, and we generally expect that to
3168 * happen when the page IO is completed and the page is now uptodate. This will
3169 * queue a task_work based retry of the operation, attempting to copy the data
3170 * again. If the latter fails because the page was NOT uptodate, then we will
3171 * do a thread based blocking retry of the operation. That's the unexpected
3174 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3175 int sync, void *arg)
3177 struct wait_page_queue *wpq;
3178 struct io_kiocb *req = wait->private;
3179 struct wait_page_key *key = arg;
3181 wpq = container_of(wait, struct wait_page_queue, wait);
3183 if (!wake_page_match(wpq, key))
3186 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3187 list_del_init(&wait->entry);
3189 /* submit ref gets dropped, acquire a new one */
3191 io_req_task_queue(req);
3196 * This controls whether a given IO request should be armed for async page
3197 * based retry. If we return false here, the request is handed to the async
3198 * worker threads for retry. If we're doing buffered reads on a regular file,
3199 * we prepare a private wait_page_queue entry and retry the operation. This
3200 * will either succeed because the page is now uptodate and unlocked, or it
3201 * will register a callback when the page is unlocked at IO completion. Through
3202 * that callback, io_uring uses task_work to setup a retry of the operation.
3203 * That retry will attempt the buffered read again. The retry will generally
3204 * succeed, or in rare cases where it fails, we then fall back to using the
3205 * async worker threads for a blocking retry.
3207 static bool io_rw_should_retry(struct io_kiocb *req)
3209 struct io_async_rw *rw = req->async_data;
3210 struct wait_page_queue *wait = &rw->wpq;
3211 struct kiocb *kiocb = &req->rw.kiocb;
3213 /* never retry for NOWAIT, we just complete with -EAGAIN */
3214 if (req->flags & REQ_F_NOWAIT)
3217 /* Only for buffered IO */
3218 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3222 * just use poll if we can, and don't attempt if the fs doesn't
3223 * support callback based unlocks
3225 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3228 wait->wait.func = io_async_buf_func;
3229 wait->wait.private = req;
3230 wait->wait.flags = 0;
3231 INIT_LIST_HEAD(&wait->wait.entry);
3232 kiocb->ki_flags |= IOCB_WAITQ;
3233 kiocb->ki_flags &= ~IOCB_NOWAIT;
3234 kiocb->ki_waitq = wait;
3238 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3240 if (req->file->f_op->read_iter)
3241 return call_read_iter(req->file, &req->rw.kiocb, iter);
3242 else if (req->file->f_op->read)
3243 return loop_rw_iter(READ, req, iter);
3248 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3250 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3251 struct kiocb *kiocb = &req->rw.kiocb;
3252 struct iov_iter __iter, *iter = &__iter;
3253 struct io_async_rw *rw = req->async_data;
3254 ssize_t io_size, ret, ret2;
3255 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3261 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3265 io_size = iov_iter_count(iter);
3266 req->result = io_size;
3268 /* Ensure we clear previously set non-block flag */
3269 if (!force_nonblock)
3270 kiocb->ki_flags &= ~IOCB_NOWAIT;
3272 kiocb->ki_flags |= IOCB_NOWAIT;
3274 /* If the file doesn't support async, just async punt */
3275 if (force_nonblock && !io_file_supports_async(req, READ)) {
3276 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3277 return ret ?: -EAGAIN;
3280 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3281 if (unlikely(ret)) {
3286 ret = io_iter_do_read(req, iter);
3288 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3289 req->flags &= ~REQ_F_REISSUE;
3290 /* IOPOLL retry should happen for io-wq threads */
3291 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3293 /* no retry on NONBLOCK nor RWF_NOWAIT */
3294 if (req->flags & REQ_F_NOWAIT)
3296 /* some cases will consume bytes even on error returns */
3297 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3299 } else if (ret == -EIOCBQUEUED) {
3301 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3302 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3303 /* read all, failed, already did sync or don't want to retry */
3307 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3312 rw = req->async_data;
3313 /* now use our persistent iterator, if we aren't already */
3318 rw->bytes_done += ret;
3319 /* if we can retry, do so with the callbacks armed */
3320 if (!io_rw_should_retry(req)) {
3321 kiocb->ki_flags &= ~IOCB_WAITQ;
3326 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3327 * we get -EIOCBQUEUED, then we'll get a notification when the
3328 * desired page gets unlocked. We can also get a partial read
3329 * here, and if we do, then just retry at the new offset.
3331 ret = io_iter_do_read(req, iter);
3332 if (ret == -EIOCBQUEUED)
3334 /* we got some bytes, but not all. retry. */
3335 kiocb->ki_flags &= ~IOCB_WAITQ;
3336 } while (ret > 0 && ret < io_size);
3338 kiocb_done(kiocb, ret, issue_flags);
3340 /* it's faster to check here then delegate to kfree */
3346 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3348 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3350 return io_prep_rw(req, sqe);
3353 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3355 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3356 struct kiocb *kiocb = &req->rw.kiocb;
3357 struct iov_iter __iter, *iter = &__iter;
3358 struct io_async_rw *rw = req->async_data;
3359 ssize_t ret, ret2, io_size;
3360 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3366 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3370 io_size = iov_iter_count(iter);
3371 req->result = io_size;
3373 /* Ensure we clear previously set non-block flag */
3374 if (!force_nonblock)
3375 kiocb->ki_flags &= ~IOCB_NOWAIT;
3377 kiocb->ki_flags |= IOCB_NOWAIT;
3379 /* If the file doesn't support async, just async punt */
3380 if (force_nonblock && !io_file_supports_async(req, WRITE))
3383 /* file path doesn't support NOWAIT for non-direct_IO */
3384 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3385 (req->flags & REQ_F_ISREG))
3388 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3393 * Open-code file_start_write here to grab freeze protection,
3394 * which will be released by another thread in
3395 * io_complete_rw(). Fool lockdep by telling it the lock got
3396 * released so that it doesn't complain about the held lock when
3397 * we return to userspace.
3399 if (req->flags & REQ_F_ISREG) {
3400 sb_start_write(file_inode(req->file)->i_sb);
3401 __sb_writers_release(file_inode(req->file)->i_sb,
3404 kiocb->ki_flags |= IOCB_WRITE;
3406 if (req->file->f_op->write_iter)
3407 ret2 = call_write_iter(req->file, kiocb, iter);
3408 else if (req->file->f_op->write)
3409 ret2 = loop_rw_iter(WRITE, req, iter);
3413 if (req->flags & REQ_F_REISSUE) {
3414 req->flags &= ~REQ_F_REISSUE;
3419 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3420 * retry them without IOCB_NOWAIT.
3422 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3424 /* no retry on NONBLOCK nor RWF_NOWAIT */
3425 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3427 if (!force_nonblock || ret2 != -EAGAIN) {
3428 /* IOPOLL retry should happen for io-wq threads */
3429 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3432 kiocb_done(kiocb, ret2, issue_flags);
3435 /* some cases will consume bytes even on error returns */
3436 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3437 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3438 return ret ?: -EAGAIN;
3441 /* it's reportedly faster than delegating the null check to kfree() */
3447 static int io_renameat_prep(struct io_kiocb *req,
3448 const struct io_uring_sqe *sqe)
3450 struct io_rename *ren = &req->rename;
3451 const char __user *oldf, *newf;
3453 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3456 ren->old_dfd = READ_ONCE(sqe->fd);
3457 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3458 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3459 ren->new_dfd = READ_ONCE(sqe->len);
3460 ren->flags = READ_ONCE(sqe->rename_flags);
3462 ren->oldpath = getname(oldf);
3463 if (IS_ERR(ren->oldpath))
3464 return PTR_ERR(ren->oldpath);
3466 ren->newpath = getname(newf);
3467 if (IS_ERR(ren->newpath)) {
3468 putname(ren->oldpath);
3469 return PTR_ERR(ren->newpath);
3472 req->flags |= REQ_F_NEED_CLEANUP;
3476 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3478 struct io_rename *ren = &req->rename;
3481 if (issue_flags & IO_URING_F_NONBLOCK)
3484 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3485 ren->newpath, ren->flags);
3487 req->flags &= ~REQ_F_NEED_CLEANUP;
3489 req_set_fail_links(req);
3490 io_req_complete(req, ret);
3494 static int io_unlinkat_prep(struct io_kiocb *req,
3495 const struct io_uring_sqe *sqe)
3497 struct io_unlink *un = &req->unlink;
3498 const char __user *fname;
3500 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3503 un->dfd = READ_ONCE(sqe->fd);
3505 un->flags = READ_ONCE(sqe->unlink_flags);
3506 if (un->flags & ~AT_REMOVEDIR)
3509 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3510 un->filename = getname(fname);
3511 if (IS_ERR(un->filename))
3512 return PTR_ERR(un->filename);
3514 req->flags |= REQ_F_NEED_CLEANUP;
3518 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3520 struct io_unlink *un = &req->unlink;
3523 if (issue_flags & IO_URING_F_NONBLOCK)
3526 if (un->flags & AT_REMOVEDIR)
3527 ret = do_rmdir(un->dfd, un->filename);
3529 ret = do_unlinkat(un->dfd, un->filename);
3531 req->flags &= ~REQ_F_NEED_CLEANUP;
3533 req_set_fail_links(req);
3534 io_req_complete(req, ret);
3538 static int io_shutdown_prep(struct io_kiocb *req,
3539 const struct io_uring_sqe *sqe)
3541 #if defined(CONFIG_NET)
3542 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3544 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3548 req->shutdown.how = READ_ONCE(sqe->len);
3555 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3557 #if defined(CONFIG_NET)
3558 struct socket *sock;
3561 if (issue_flags & IO_URING_F_NONBLOCK)
3564 sock = sock_from_file(req->file);
3565 if (unlikely(!sock))
3568 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3570 req_set_fail_links(req);
3571 io_req_complete(req, ret);
3578 static int __io_splice_prep(struct io_kiocb *req,
3579 const struct io_uring_sqe *sqe)
3581 struct io_splice* sp = &req->splice;
3582 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3584 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3588 sp->len = READ_ONCE(sqe->len);
3589 sp->flags = READ_ONCE(sqe->splice_flags);
3591 if (unlikely(sp->flags & ~valid_flags))
3594 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3595 (sp->flags & SPLICE_F_FD_IN_FIXED));
3598 req->flags |= REQ_F_NEED_CLEANUP;
3602 static int io_tee_prep(struct io_kiocb *req,
3603 const struct io_uring_sqe *sqe)
3605 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3607 return __io_splice_prep(req, sqe);
3610 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3612 struct io_splice *sp = &req->splice;
3613 struct file *in = sp->file_in;
3614 struct file *out = sp->file_out;
3615 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3618 if (issue_flags & IO_URING_F_NONBLOCK)
3621 ret = do_tee(in, out, sp->len, flags);
3623 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3625 req->flags &= ~REQ_F_NEED_CLEANUP;
3628 req_set_fail_links(req);
3629 io_req_complete(req, ret);
3633 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3635 struct io_splice* sp = &req->splice;
3637 sp->off_in = READ_ONCE(sqe->splice_off_in);
3638 sp->off_out = READ_ONCE(sqe->off);
3639 return __io_splice_prep(req, sqe);
3642 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3644 struct io_splice *sp = &req->splice;
3645 struct file *in = sp->file_in;
3646 struct file *out = sp->file_out;
3647 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3648 loff_t *poff_in, *poff_out;
3651 if (issue_flags & IO_URING_F_NONBLOCK)
3654 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3655 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3658 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3660 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3662 req->flags &= ~REQ_F_NEED_CLEANUP;
3665 req_set_fail_links(req);
3666 io_req_complete(req, ret);
3671 * IORING_OP_NOP just posts a completion event, nothing else.
3673 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3675 struct io_ring_ctx *ctx = req->ctx;
3677 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3680 __io_req_complete(req, issue_flags, 0, 0);
3684 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3686 struct io_ring_ctx *ctx = req->ctx;
3691 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3693 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3696 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3697 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3700 req->sync.off = READ_ONCE(sqe->off);
3701 req->sync.len = READ_ONCE(sqe->len);
3705 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3707 loff_t end = req->sync.off + req->sync.len;
3710 /* fsync always requires a blocking context */
3711 if (issue_flags & IO_URING_F_NONBLOCK)
3714 ret = vfs_fsync_range(req->file, req->sync.off,
3715 end > 0 ? end : LLONG_MAX,
3716 req->sync.flags & IORING_FSYNC_DATASYNC);
3718 req_set_fail_links(req);
3719 io_req_complete(req, ret);
3723 static int io_fallocate_prep(struct io_kiocb *req,
3724 const struct io_uring_sqe *sqe)
3726 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3728 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3731 req->sync.off = READ_ONCE(sqe->off);
3732 req->sync.len = READ_ONCE(sqe->addr);
3733 req->sync.mode = READ_ONCE(sqe->len);
3737 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3741 /* fallocate always requiring blocking context */
3742 if (issue_flags & IO_URING_F_NONBLOCK)
3744 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3747 req_set_fail_links(req);
3748 io_req_complete(req, ret);
3752 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3754 const char __user *fname;
3757 if (unlikely(sqe->ioprio || sqe->buf_index))
3759 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3762 /* open.how should be already initialised */
3763 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3764 req->open.how.flags |= O_LARGEFILE;
3766 req->open.dfd = READ_ONCE(sqe->fd);
3767 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3768 req->open.filename = getname(fname);
3769 if (IS_ERR(req->open.filename)) {
3770 ret = PTR_ERR(req->open.filename);
3771 req->open.filename = NULL;
3774 req->open.nofile = rlimit(RLIMIT_NOFILE);
3775 req->flags |= REQ_F_NEED_CLEANUP;
3779 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3783 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3785 mode = READ_ONCE(sqe->len);
3786 flags = READ_ONCE(sqe->open_flags);
3787 req->open.how = build_open_how(flags, mode);
3788 return __io_openat_prep(req, sqe);
3791 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3793 struct open_how __user *how;
3797 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3799 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3800 len = READ_ONCE(sqe->len);
3801 if (len < OPEN_HOW_SIZE_VER0)
3804 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3809 return __io_openat_prep(req, sqe);
3812 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3814 struct open_flags op;
3817 bool resolve_nonblock;
3820 ret = build_open_flags(&req->open.how, &op);
3823 nonblock_set = op.open_flag & O_NONBLOCK;
3824 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3825 if (issue_flags & IO_URING_F_NONBLOCK) {
3827 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3828 * it'll always -EAGAIN
3830 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3832 op.lookup_flags |= LOOKUP_CACHED;
3833 op.open_flag |= O_NONBLOCK;
3836 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3840 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3841 /* only retry if RESOLVE_CACHED wasn't already set by application */
3842 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3843 file == ERR_PTR(-EAGAIN)) {
3845 * We could hang on to this 'fd', but seems like marginal
3846 * gain for something that is now known to be a slower path.
3847 * So just put it, and we'll get a new one when we retry.
3855 ret = PTR_ERR(file);
3857 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3858 file->f_flags &= ~O_NONBLOCK;
3859 fsnotify_open(file);
3860 fd_install(ret, file);
3863 putname(req->open.filename);
3864 req->flags &= ~REQ_F_NEED_CLEANUP;
3866 req_set_fail_links(req);
3867 io_req_complete(req, ret);
3871 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3873 return io_openat2(req, issue_flags);
3876 static int io_remove_buffers_prep(struct io_kiocb *req,
3877 const struct io_uring_sqe *sqe)
3879 struct io_provide_buf *p = &req->pbuf;
3882 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3885 tmp = READ_ONCE(sqe->fd);
3886 if (!tmp || tmp > USHRT_MAX)
3889 memset(p, 0, sizeof(*p));
3891 p->bgid = READ_ONCE(sqe->buf_group);
3895 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3896 int bgid, unsigned nbufs)
3900 /* shouldn't happen */
3904 /* the head kbuf is the list itself */
3905 while (!list_empty(&buf->list)) {
3906 struct io_buffer *nxt;
3908 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3909 list_del(&nxt->list);
3916 xa_erase(&ctx->io_buffers, bgid);
3921 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3923 struct io_provide_buf *p = &req->pbuf;
3924 struct io_ring_ctx *ctx = req->ctx;
3925 struct io_buffer *head;
3927 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3929 io_ring_submit_lock(ctx, !force_nonblock);
3931 lockdep_assert_held(&ctx->uring_lock);
3934 head = xa_load(&ctx->io_buffers, p->bgid);
3936 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3938 req_set_fail_links(req);
3940 /* complete before unlock, IOPOLL may need the lock */
3941 __io_req_complete(req, issue_flags, ret, 0);
3942 io_ring_submit_unlock(ctx, !force_nonblock);
3946 static int io_provide_buffers_prep(struct io_kiocb *req,
3947 const struct io_uring_sqe *sqe)
3950 struct io_provide_buf *p = &req->pbuf;
3953 if (sqe->ioprio || sqe->rw_flags)
3956 tmp = READ_ONCE(sqe->fd);
3957 if (!tmp || tmp > USHRT_MAX)
3960 p->addr = READ_ONCE(sqe->addr);
3961 p->len = READ_ONCE(sqe->len);
3963 size = (unsigned long)p->len * p->nbufs;
3964 if (!access_ok(u64_to_user_ptr(p->addr), size))
3967 p->bgid = READ_ONCE(sqe->buf_group);
3968 tmp = READ_ONCE(sqe->off);
3969 if (tmp > USHRT_MAX)
3975 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3977 struct io_buffer *buf;
3978 u64 addr = pbuf->addr;
3979 int i, bid = pbuf->bid;
3981 for (i = 0; i < pbuf->nbufs; i++) {
3982 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3987 buf->len = pbuf->len;
3992 INIT_LIST_HEAD(&buf->list);
3995 list_add_tail(&buf->list, &(*head)->list);
3999 return i ? i : -ENOMEM;
4002 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4004 struct io_provide_buf *p = &req->pbuf;
4005 struct io_ring_ctx *ctx = req->ctx;
4006 struct io_buffer *head, *list;
4008 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4010 io_ring_submit_lock(ctx, !force_nonblock);
4012 lockdep_assert_held(&ctx->uring_lock);
4014 list = head = xa_load(&ctx->io_buffers, p->bgid);
4016 ret = io_add_buffers(p, &head);
4017 if (ret >= 0 && !list) {
4018 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4020 __io_remove_buffers(ctx, head, p->bgid, -1U);
4023 req_set_fail_links(req);
4024 /* complete before unlock, IOPOLL may need the lock */
4025 __io_req_complete(req, issue_flags, ret, 0);
4026 io_ring_submit_unlock(ctx, !force_nonblock);
4030 static int io_epoll_ctl_prep(struct io_kiocb *req,
4031 const struct io_uring_sqe *sqe)
4033 #if defined(CONFIG_EPOLL)
4034 if (sqe->ioprio || sqe->buf_index)
4036 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4039 req->epoll.epfd = READ_ONCE(sqe->fd);
4040 req->epoll.op = READ_ONCE(sqe->len);
4041 req->epoll.fd = READ_ONCE(sqe->off);
4043 if (ep_op_has_event(req->epoll.op)) {
4044 struct epoll_event __user *ev;
4046 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4047 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4057 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4059 #if defined(CONFIG_EPOLL)
4060 struct io_epoll *ie = &req->epoll;
4062 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4064 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4065 if (force_nonblock && ret == -EAGAIN)
4069 req_set_fail_links(req);
4070 __io_req_complete(req, issue_flags, ret, 0);
4077 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4079 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4080 if (sqe->ioprio || sqe->buf_index || sqe->off)
4082 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4085 req->madvise.addr = READ_ONCE(sqe->addr);
4086 req->madvise.len = READ_ONCE(sqe->len);
4087 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4094 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4096 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4097 struct io_madvise *ma = &req->madvise;
4100 if (issue_flags & IO_URING_F_NONBLOCK)
4103 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4105 req_set_fail_links(req);
4106 io_req_complete(req, ret);
4113 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4117 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4120 req->fadvise.offset = READ_ONCE(sqe->off);
4121 req->fadvise.len = READ_ONCE(sqe->len);
4122 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4126 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4128 struct io_fadvise *fa = &req->fadvise;
4131 if (issue_flags & IO_URING_F_NONBLOCK) {
4132 switch (fa->advice) {
4133 case POSIX_FADV_NORMAL:
4134 case POSIX_FADV_RANDOM:
4135 case POSIX_FADV_SEQUENTIAL:
4142 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4144 req_set_fail_links(req);
4145 io_req_complete(req, ret);
4149 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4153 if (sqe->ioprio || sqe->buf_index)
4155 if (req->flags & REQ_F_FIXED_FILE)
4158 req->statx.dfd = READ_ONCE(sqe->fd);
4159 req->statx.mask = READ_ONCE(sqe->len);
4160 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4161 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4162 req->statx.flags = READ_ONCE(sqe->statx_flags);
4167 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4169 struct io_statx *ctx = &req->statx;
4172 if (issue_flags & IO_URING_F_NONBLOCK)
4175 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4179 req_set_fail_links(req);
4180 io_req_complete(req, ret);
4184 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4186 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4188 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4189 sqe->rw_flags || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->close.fd = READ_ONCE(sqe->fd);
4198 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4200 struct files_struct *files = current->files;
4201 struct io_close *close = &req->close;
4202 struct fdtable *fdt;
4208 spin_lock(&files->file_lock);
4209 fdt = files_fdtable(files);
4210 if (close->fd >= fdt->max_fds) {
4211 spin_unlock(&files->file_lock);
4214 file = fdt->fd[close->fd];
4216 spin_unlock(&files->file_lock);
4220 if (file->f_op == &io_uring_fops) {
4221 spin_unlock(&files->file_lock);
4226 /* if the file has a flush method, be safe and punt to async */
4227 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4228 spin_unlock(&files->file_lock);
4232 ret = __close_fd_get_file(close->fd, &file);
4233 spin_unlock(&files->file_lock);
4240 /* No ->flush() or already async, safely close from here */
4241 ret = filp_close(file, current->files);
4244 req_set_fail_links(req);
4247 __io_req_complete(req, issue_flags, ret, 0);
4251 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4253 struct io_ring_ctx *ctx = req->ctx;
4255 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4257 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4260 req->sync.off = READ_ONCE(sqe->off);
4261 req->sync.len = READ_ONCE(sqe->len);
4262 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4266 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4270 /* sync_file_range always requires a blocking context */
4271 if (issue_flags & IO_URING_F_NONBLOCK)
4274 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4277 req_set_fail_links(req);
4278 io_req_complete(req, ret);
4282 #if defined(CONFIG_NET)
4283 static int io_setup_async_msg(struct io_kiocb *req,
4284 struct io_async_msghdr *kmsg)
4286 struct io_async_msghdr *async_msg = req->async_data;
4290 if (io_alloc_async_data(req)) {
4291 kfree(kmsg->free_iov);
4294 async_msg = req->async_data;
4295 req->flags |= REQ_F_NEED_CLEANUP;
4296 memcpy(async_msg, kmsg, sizeof(*kmsg));
4297 async_msg->msg.msg_name = &async_msg->addr;
4298 /* if were using fast_iov, set it to the new one */
4299 if (!async_msg->free_iov)
4300 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4305 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4306 struct io_async_msghdr *iomsg)
4308 iomsg->msg.msg_name = &iomsg->addr;
4309 iomsg->free_iov = iomsg->fast_iov;
4310 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4311 req->sr_msg.msg_flags, &iomsg->free_iov);
4314 static int io_sendmsg_prep_async(struct io_kiocb *req)
4318 ret = io_sendmsg_copy_hdr(req, req->async_data);
4320 req->flags |= REQ_F_NEED_CLEANUP;
4324 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4326 struct io_sr_msg *sr = &req->sr_msg;
4328 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4331 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4332 sr->len = READ_ONCE(sqe->len);
4333 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4334 if (sr->msg_flags & MSG_DONTWAIT)
4335 req->flags |= REQ_F_NOWAIT;
4337 #ifdef CONFIG_COMPAT
4338 if (req->ctx->compat)
4339 sr->msg_flags |= MSG_CMSG_COMPAT;
4344 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4346 struct io_async_msghdr iomsg, *kmsg;
4347 struct socket *sock;
4352 sock = sock_from_file(req->file);
4353 if (unlikely(!sock))
4356 kmsg = req->async_data;
4358 ret = io_sendmsg_copy_hdr(req, &iomsg);
4364 flags = req->sr_msg.msg_flags;
4365 if (issue_flags & IO_URING_F_NONBLOCK)
4366 flags |= MSG_DONTWAIT;
4367 if (flags & MSG_WAITALL)
4368 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4370 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4371 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4372 return io_setup_async_msg(req, kmsg);
4373 if (ret == -ERESTARTSYS)
4376 /* fast path, check for non-NULL to avoid function call */
4378 kfree(kmsg->free_iov);
4379 req->flags &= ~REQ_F_NEED_CLEANUP;
4381 req_set_fail_links(req);
4382 __io_req_complete(req, issue_flags, ret, 0);
4386 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4388 struct io_sr_msg *sr = &req->sr_msg;
4391 struct socket *sock;
4396 sock = sock_from_file(req->file);
4397 if (unlikely(!sock))
4400 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4404 msg.msg_name = NULL;
4405 msg.msg_control = NULL;
4406 msg.msg_controllen = 0;
4407 msg.msg_namelen = 0;
4409 flags = req->sr_msg.msg_flags;
4410 if (issue_flags & IO_URING_F_NONBLOCK)
4411 flags |= MSG_DONTWAIT;
4412 if (flags & MSG_WAITALL)
4413 min_ret = iov_iter_count(&msg.msg_iter);
4415 msg.msg_flags = flags;
4416 ret = sock_sendmsg(sock, &msg);
4417 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4419 if (ret == -ERESTARTSYS)
4423 req_set_fail_links(req);
4424 __io_req_complete(req, issue_flags, ret, 0);
4428 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4429 struct io_async_msghdr *iomsg)
4431 struct io_sr_msg *sr = &req->sr_msg;
4432 struct iovec __user *uiov;
4436 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4437 &iomsg->uaddr, &uiov, &iov_len);
4441 if (req->flags & REQ_F_BUFFER_SELECT) {
4444 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4446 sr->len = iomsg->fast_iov[0].iov_len;
4447 iomsg->free_iov = NULL;
4449 iomsg->free_iov = iomsg->fast_iov;
4450 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4451 &iomsg->free_iov, &iomsg->msg.msg_iter,
4460 #ifdef CONFIG_COMPAT
4461 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4462 struct io_async_msghdr *iomsg)
4464 struct compat_msghdr __user *msg_compat;
4465 struct io_sr_msg *sr = &req->sr_msg;
4466 struct compat_iovec __user *uiov;
4471 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4472 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4477 uiov = compat_ptr(ptr);
4478 if (req->flags & REQ_F_BUFFER_SELECT) {
4479 compat_ssize_t clen;
4483 if (!access_ok(uiov, sizeof(*uiov)))
4485 if (__get_user(clen, &uiov->iov_len))
4490 iomsg->free_iov = NULL;
4492 iomsg->free_iov = iomsg->fast_iov;
4493 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4494 UIO_FASTIOV, &iomsg->free_iov,
4495 &iomsg->msg.msg_iter, true);
4504 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4505 struct io_async_msghdr *iomsg)
4507 iomsg->msg.msg_name = &iomsg->addr;
4509 #ifdef CONFIG_COMPAT
4510 if (req->ctx->compat)
4511 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4514 return __io_recvmsg_copy_hdr(req, iomsg);
4517 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4520 struct io_sr_msg *sr = &req->sr_msg;
4521 struct io_buffer *kbuf;
4523 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4528 req->flags |= REQ_F_BUFFER_SELECTED;
4532 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4534 return io_put_kbuf(req, req->sr_msg.kbuf);
4537 static int io_recvmsg_prep_async(struct io_kiocb *req)
4541 ret = io_recvmsg_copy_hdr(req, req->async_data);
4543 req->flags |= REQ_F_NEED_CLEANUP;
4547 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 struct io_sr_msg *sr = &req->sr_msg;
4551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4554 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4555 sr->len = READ_ONCE(sqe->len);
4556 sr->bgid = READ_ONCE(sqe->buf_group);
4557 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4558 if (sr->msg_flags & MSG_DONTWAIT)
4559 req->flags |= REQ_F_NOWAIT;
4561 #ifdef CONFIG_COMPAT
4562 if (req->ctx->compat)
4563 sr->msg_flags |= MSG_CMSG_COMPAT;
4568 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4570 struct io_async_msghdr iomsg, *kmsg;
4571 struct socket *sock;
4572 struct io_buffer *kbuf;
4575 int ret, cflags = 0;
4576 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4578 sock = sock_from_file(req->file);
4579 if (unlikely(!sock))
4582 kmsg = req->async_data;
4584 ret = io_recvmsg_copy_hdr(req, &iomsg);
4590 if (req->flags & REQ_F_BUFFER_SELECT) {
4591 kbuf = io_recv_buffer_select(req, !force_nonblock);
4593 return PTR_ERR(kbuf);
4594 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4595 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4596 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4597 1, req->sr_msg.len);
4600 flags = req->sr_msg.msg_flags;
4602 flags |= MSG_DONTWAIT;
4603 if (flags & MSG_WAITALL)
4604 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4606 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4607 kmsg->uaddr, flags);
4608 if (force_nonblock && ret == -EAGAIN)
4609 return io_setup_async_msg(req, kmsg);
4610 if (ret == -ERESTARTSYS)
4613 if (req->flags & REQ_F_BUFFER_SELECTED)
4614 cflags = io_put_recv_kbuf(req);
4615 /* fast path, check for non-NULL to avoid function call */
4617 kfree(kmsg->free_iov);
4618 req->flags &= ~REQ_F_NEED_CLEANUP;
4619 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4620 req_set_fail_links(req);
4621 __io_req_complete(req, issue_flags, ret, cflags);
4625 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4627 struct io_buffer *kbuf;
4628 struct io_sr_msg *sr = &req->sr_msg;
4630 void __user *buf = sr->buf;
4631 struct socket *sock;
4635 int ret, cflags = 0;
4636 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4638 sock = sock_from_file(req->file);
4639 if (unlikely(!sock))
4642 if (req->flags & REQ_F_BUFFER_SELECT) {
4643 kbuf = io_recv_buffer_select(req, !force_nonblock);
4645 return PTR_ERR(kbuf);
4646 buf = u64_to_user_ptr(kbuf->addr);
4649 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4653 msg.msg_name = NULL;
4654 msg.msg_control = NULL;
4655 msg.msg_controllen = 0;
4656 msg.msg_namelen = 0;
4657 msg.msg_iocb = NULL;
4660 flags = req->sr_msg.msg_flags;
4662 flags |= MSG_DONTWAIT;
4663 if (flags & MSG_WAITALL)
4664 min_ret = iov_iter_count(&msg.msg_iter);
4666 ret = sock_recvmsg(sock, &msg, flags);
4667 if (force_nonblock && ret == -EAGAIN)
4669 if (ret == -ERESTARTSYS)
4672 if (req->flags & REQ_F_BUFFER_SELECTED)
4673 cflags = io_put_recv_kbuf(req);
4674 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4675 req_set_fail_links(req);
4676 __io_req_complete(req, issue_flags, ret, cflags);
4680 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4682 struct io_accept *accept = &req->accept;
4684 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4686 if (sqe->ioprio || sqe->len || sqe->buf_index)
4689 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4690 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4691 accept->flags = READ_ONCE(sqe->accept_flags);
4692 accept->nofile = rlimit(RLIMIT_NOFILE);
4696 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4698 struct io_accept *accept = &req->accept;
4699 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4700 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4703 if (req->file->f_flags & O_NONBLOCK)
4704 req->flags |= REQ_F_NOWAIT;
4706 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4707 accept->addr_len, accept->flags,
4709 if (ret == -EAGAIN && force_nonblock)
4712 if (ret == -ERESTARTSYS)
4714 req_set_fail_links(req);
4716 __io_req_complete(req, issue_flags, ret, 0);
4720 static int io_connect_prep_async(struct io_kiocb *req)
4722 struct io_async_connect *io = req->async_data;
4723 struct io_connect *conn = &req->connect;
4725 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4728 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4730 struct io_connect *conn = &req->connect;
4732 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4734 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4737 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4738 conn->addr_len = READ_ONCE(sqe->addr2);
4742 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4744 struct io_async_connect __io, *io;
4745 unsigned file_flags;
4747 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4749 if (req->async_data) {
4750 io = req->async_data;
4752 ret = move_addr_to_kernel(req->connect.addr,
4753 req->connect.addr_len,
4760 file_flags = force_nonblock ? O_NONBLOCK : 0;
4762 ret = __sys_connect_file(req->file, &io->address,
4763 req->connect.addr_len, file_flags);
4764 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4765 if (req->async_data)
4767 if (io_alloc_async_data(req)) {
4771 memcpy(req->async_data, &__io, sizeof(__io));
4774 if (ret == -ERESTARTSYS)
4778 req_set_fail_links(req);
4779 __io_req_complete(req, issue_flags, ret, 0);
4782 #else /* !CONFIG_NET */
4783 #define IO_NETOP_FN(op) \
4784 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4786 return -EOPNOTSUPP; \
4789 #define IO_NETOP_PREP(op) \
4791 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4793 return -EOPNOTSUPP; \
4796 #define IO_NETOP_PREP_ASYNC(op) \
4798 static int io_##op##_prep_async(struct io_kiocb *req) \
4800 return -EOPNOTSUPP; \
4803 IO_NETOP_PREP_ASYNC(sendmsg);
4804 IO_NETOP_PREP_ASYNC(recvmsg);
4805 IO_NETOP_PREP_ASYNC(connect);
4806 IO_NETOP_PREP(accept);
4809 #endif /* CONFIG_NET */
4811 struct io_poll_table {
4812 struct poll_table_struct pt;
4813 struct io_kiocb *req;
4817 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4818 __poll_t mask, task_work_func_t func)
4822 /* for instances that support it check for an event match first: */
4823 if (mask && !(mask & poll->events))
4826 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4828 list_del_init(&poll->wait.entry);
4831 req->task_work.func = func;
4834 * If this fails, then the task is exiting. When a task exits, the
4835 * work gets canceled, so just cancel this request as well instead
4836 * of executing it. We can't safely execute it anyway, as we may not
4837 * have the needed state needed for it anyway.
4839 ret = io_req_task_work_add(req);
4840 if (unlikely(ret)) {
4841 WRITE_ONCE(poll->canceled, true);
4842 io_req_task_work_add_fallback(req, func);
4847 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4848 __acquires(&req->ctx->completion_lock)
4850 struct io_ring_ctx *ctx = req->ctx;
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 struct poll_table_struct pt = { ._key = poll->events };
4855 req->result = vfs_poll(req->file, &pt) & poll->events;
4858 spin_lock_irq(&ctx->completion_lock);
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 add_wait_queue(poll->head, &poll->wait);
4867 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4869 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4870 if (req->opcode == IORING_OP_POLL_ADD)
4871 return req->async_data;
4872 return req->apoll->double_poll;
4875 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4877 if (req->opcode == IORING_OP_POLL_ADD)
4879 return &req->apoll->poll;
4882 static void io_poll_remove_double(struct io_kiocb *req)
4883 __must_hold(&req->ctx->completion_lock)
4885 struct io_poll_iocb *poll = io_poll_get_double(req);
4887 lockdep_assert_held(&req->ctx->completion_lock);
4889 if (poll && poll->head) {
4890 struct wait_queue_head *head = poll->head;
4892 spin_lock(&head->lock);
4893 list_del_init(&poll->wait.entry);
4894 if (poll->wait.private)
4897 spin_unlock(&head->lock);
4901 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4902 __must_hold(&req->ctx->completion_lock)
4904 struct io_ring_ctx *ctx = req->ctx;
4905 unsigned flags = IORING_CQE_F_MORE;
4907 if (!error && req->poll.canceled) {
4909 req->poll.events |= EPOLLONESHOT;
4912 error = mangle_poll(mask);
4913 if (req->poll.events & EPOLLONESHOT)
4915 if (!__io_cqring_fill_event(req, error, flags)) {
4916 io_poll_remove_waitqs(req);
4917 req->poll.done = true;
4920 io_commit_cqring(ctx);
4921 return !(flags & IORING_CQE_F_MORE);
4924 static void io_poll_task_func(struct callback_head *cb)
4926 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4927 struct io_ring_ctx *ctx = req->ctx;
4928 struct io_kiocb *nxt;
4930 if (io_poll_rewait(req, &req->poll)) {
4931 spin_unlock_irq(&ctx->completion_lock);
4935 post_ev = done = io_poll_complete(req, req->result, 0);
4937 hash_del(&req->hash_node);
4938 } else if (!(req->poll.events & EPOLLONESHOT)) {
4941 add_wait_queue(req->poll.head, &req->poll.wait);
4943 spin_unlock_irq(&ctx->completion_lock);
4946 io_cqring_ev_posted(ctx);
4948 nxt = io_put_req_find_next(req);
4950 __io_req_task_submit(nxt);
4955 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4956 int sync, void *key)
4958 struct io_kiocb *req = wait->private;
4959 struct io_poll_iocb *poll = io_poll_get_single(req);
4960 __poll_t mask = key_to_poll(key);
4962 /* for instances that support it check for an event match first: */
4963 if (mask && !(mask & poll->events))
4965 if (!(poll->events & EPOLLONESHOT))
4966 return poll->wait.func(&poll->wait, mode, sync, key);
4968 list_del_init(&wait->entry);
4970 if (poll && poll->head) {
4973 spin_lock(&poll->head->lock);
4974 done = list_empty(&poll->wait.entry);
4976 list_del_init(&poll->wait.entry);
4977 /* make sure double remove sees this as being gone */
4978 wait->private = NULL;
4979 spin_unlock(&poll->head->lock);
4981 /* use wait func handler, so it matches the rq type */
4982 poll->wait.func(&poll->wait, mode, sync, key);
4989 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4990 wait_queue_func_t wake_func)
4994 poll->canceled = false;
4995 poll->update_events = poll->update_user_data = false;
4996 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4997 /* mask in events that we always want/need */
4998 poll->events = events | IO_POLL_UNMASK;
4999 INIT_LIST_HEAD(&poll->wait.entry);
5000 init_waitqueue_func_entry(&poll->wait, wake_func);
5003 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5004 struct wait_queue_head *head,
5005 struct io_poll_iocb **poll_ptr)
5007 struct io_kiocb *req = pt->req;
5010 * If poll->head is already set, it's because the file being polled
5011 * uses multiple waitqueues for poll handling (eg one for read, one
5012 * for write). Setup a separate io_poll_iocb if this happens.
5014 if (unlikely(poll->head)) {
5015 struct io_poll_iocb *poll_one = poll;
5017 /* already have a 2nd entry, fail a third attempt */
5019 pt->error = -EINVAL;
5022 /* double add on the same waitqueue head, ignore */
5023 if (poll->head == head)
5025 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5027 pt->error = -ENOMEM;
5030 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5032 poll->wait.private = req;
5039 if (poll->events & EPOLLEXCLUSIVE)
5040 add_wait_queue_exclusive(head, &poll->wait);
5042 add_wait_queue(head, &poll->wait);
5045 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5046 struct poll_table_struct *p)
5048 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5049 struct async_poll *apoll = pt->req->apoll;
5051 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5054 static void io_async_task_func(struct callback_head *cb)
5056 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5057 struct async_poll *apoll = req->apoll;
5058 struct io_ring_ctx *ctx = req->ctx;
5060 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5062 if (io_poll_rewait(req, &apoll->poll)) {
5063 spin_unlock_irq(&ctx->completion_lock);
5067 /* If req is still hashed, it cannot have been canceled. Don't check. */
5068 if (hash_hashed(&req->hash_node))
5069 hash_del(&req->hash_node);
5071 io_poll_remove_double(req);
5072 spin_unlock_irq(&ctx->completion_lock);
5074 if (!READ_ONCE(apoll->poll.canceled))
5075 __io_req_task_submit(req);
5077 io_req_complete_failed(req, -ECANCELED);
5079 kfree(apoll->double_poll);
5083 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5086 struct io_kiocb *req = wait->private;
5087 struct io_poll_iocb *poll = &req->apoll->poll;
5089 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5092 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5095 static void io_poll_req_insert(struct io_kiocb *req)
5097 struct io_ring_ctx *ctx = req->ctx;
5098 struct hlist_head *list;
5100 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5101 hlist_add_head(&req->hash_node, list);
5104 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5105 struct io_poll_iocb *poll,
5106 struct io_poll_table *ipt, __poll_t mask,
5107 wait_queue_func_t wake_func)
5108 __acquires(&ctx->completion_lock)
5110 struct io_ring_ctx *ctx = req->ctx;
5111 bool cancel = false;
5113 INIT_HLIST_NODE(&req->hash_node);
5114 io_init_poll_iocb(poll, mask, wake_func);
5115 poll->file = req->file;
5116 poll->wait.private = req;
5118 ipt->pt._key = mask;
5120 ipt->error = -EINVAL;
5122 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5124 spin_lock_irq(&ctx->completion_lock);
5125 if (likely(poll->head)) {
5126 spin_lock(&poll->head->lock);
5127 if (unlikely(list_empty(&poll->wait.entry))) {
5133 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5134 list_del_init(&poll->wait.entry);
5136 WRITE_ONCE(poll->canceled, true);
5137 else if (!poll->done) /* actually waiting for an event */
5138 io_poll_req_insert(req);
5139 spin_unlock(&poll->head->lock);
5145 static bool io_arm_poll_handler(struct io_kiocb *req)
5147 const struct io_op_def *def = &io_op_defs[req->opcode];
5148 struct io_ring_ctx *ctx = req->ctx;
5149 struct async_poll *apoll;
5150 struct io_poll_table ipt;
5154 if (!req->file || !file_can_poll(req->file))
5156 if (req->flags & REQ_F_POLLED)
5160 else if (def->pollout)
5164 /* if we can't nonblock try, then no point in arming a poll handler */
5165 if (!io_file_supports_async(req, rw))
5168 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5169 if (unlikely(!apoll))
5171 apoll->double_poll = NULL;
5173 req->flags |= REQ_F_POLLED;
5176 mask = EPOLLONESHOT;
5178 mask |= POLLIN | POLLRDNORM;
5180 mask |= POLLOUT | POLLWRNORM;
5182 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5183 if ((req->opcode == IORING_OP_RECVMSG) &&
5184 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5187 mask |= POLLERR | POLLPRI;
5189 ipt.pt._qproc = io_async_queue_proc;
5191 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5193 if (ret || ipt.error) {
5194 io_poll_remove_double(req);
5195 spin_unlock_irq(&ctx->completion_lock);
5196 kfree(apoll->double_poll);
5200 spin_unlock_irq(&ctx->completion_lock);
5201 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5202 apoll->poll.events);
5206 static bool __io_poll_remove_one(struct io_kiocb *req,
5207 struct io_poll_iocb *poll, bool do_cancel)
5208 __must_hold(&req->ctx->completion_lock)
5210 bool do_complete = false;
5214 spin_lock(&poll->head->lock);
5216 WRITE_ONCE(poll->canceled, true);
5217 if (!list_empty(&poll->wait.entry)) {
5218 list_del_init(&poll->wait.entry);
5221 spin_unlock(&poll->head->lock);
5222 hash_del(&req->hash_node);
5226 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5227 __must_hold(&req->ctx->completion_lock)
5231 io_poll_remove_double(req);
5233 if (req->opcode == IORING_OP_POLL_ADD) {
5234 do_complete = __io_poll_remove_one(req, &req->poll, true);
5236 struct async_poll *apoll = req->apoll;
5238 /* non-poll requests have submit ref still */
5239 do_complete = __io_poll_remove_one(req, &apoll->poll, true);
5242 kfree(apoll->double_poll);
5250 static bool io_poll_remove_one(struct io_kiocb *req)
5251 __must_hold(&req->ctx->completion_lock)
5255 do_complete = io_poll_remove_waitqs(req);
5257 io_cqring_fill_event(req, -ECANCELED);
5258 io_commit_cqring(req->ctx);
5259 req_set_fail_links(req);
5260 io_put_req_deferred(req, 1);
5267 * Returns true if we found and killed one or more poll requests
5269 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5270 struct files_struct *files)
5272 struct hlist_node *tmp;
5273 struct io_kiocb *req;
5276 spin_lock_irq(&ctx->completion_lock);
5277 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5278 struct hlist_head *list;
5280 list = &ctx->cancel_hash[i];
5281 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5282 if (io_match_task(req, tsk, files))
5283 posted += io_poll_remove_one(req);
5286 spin_unlock_irq(&ctx->completion_lock);
5289 io_cqring_ev_posted(ctx);
5294 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr)
5295 __must_hold(&ctx->completion_lock)
5297 struct hlist_head *list;
5298 struct io_kiocb *req;
5300 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5301 hlist_for_each_entry(req, list, hash_node) {
5302 if (sqe_addr != req->user_data)
5310 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5311 __must_hold(&ctx->completion_lock)
5313 struct io_kiocb *req;
5315 req = io_poll_find(ctx, sqe_addr);
5318 if (io_poll_remove_one(req))
5324 static int io_poll_remove_prep(struct io_kiocb *req,
5325 const struct io_uring_sqe *sqe)
5327 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5329 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5333 req->poll_remove.addr = READ_ONCE(sqe->addr);
5338 * Find a running poll command that matches one specified in sqe->addr,
5339 * and remove it if found.
5341 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5343 struct io_ring_ctx *ctx = req->ctx;
5346 spin_lock_irq(&ctx->completion_lock);
5347 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5348 spin_unlock_irq(&ctx->completion_lock);
5351 req_set_fail_links(req);
5352 io_req_complete(req, ret);
5356 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5359 struct io_kiocb *req = wait->private;
5360 struct io_poll_iocb *poll = &req->poll;
5362 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5365 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5366 struct poll_table_struct *p)
5368 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5370 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5373 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5375 struct io_poll_iocb *poll = &req->poll;
5378 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5380 if (sqe->ioprio || sqe->buf_index)
5382 flags = READ_ONCE(sqe->len);
5383 if (flags & ~(IORING_POLL_ADD_MULTI | IORING_POLL_UPDATE_EVENTS |
5384 IORING_POLL_UPDATE_USER_DATA))
5386 events = READ_ONCE(sqe->poll32_events);
5388 events = swahw32(events);
5390 if (!(flags & IORING_POLL_ADD_MULTI))
5391 events |= EPOLLONESHOT;
5392 poll->update_events = poll->update_user_data = false;
5393 if (flags & IORING_POLL_UPDATE_EVENTS) {
5394 poll->update_events = true;
5395 poll->old_user_data = READ_ONCE(sqe->addr);
5397 if (flags & IORING_POLL_UPDATE_USER_DATA) {
5398 poll->update_user_data = true;
5399 poll->new_user_data = READ_ONCE(sqe->off);
5401 if (!(poll->update_events || poll->update_user_data) &&
5402 (sqe->off || sqe->addr))
5404 poll->events = demangle_poll(events) |
5405 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5409 static int __io_poll_add(struct io_kiocb *req)
5411 struct io_poll_iocb *poll = &req->poll;
5412 struct io_ring_ctx *ctx = req->ctx;
5413 struct io_poll_table ipt;
5416 ipt.pt._qproc = io_poll_queue_proc;
5418 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5421 if (mask) { /* no async, we'd stolen it */
5423 io_poll_complete(req, mask, 0);
5425 spin_unlock_irq(&ctx->completion_lock);
5428 io_cqring_ev_posted(ctx);
5429 if (poll->events & EPOLLONESHOT)
5435 static int io_poll_update(struct io_kiocb *req)
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct io_kiocb *preq;
5441 spin_lock_irq(&ctx->completion_lock);
5442 preq = io_poll_find(ctx, req->poll.old_user_data);
5446 } else if (preq->opcode != IORING_OP_POLL_ADD) {
5447 /* don't allow internal poll updates */
5451 if (!__io_poll_remove_one(preq, &preq->poll, false)) {
5452 if (preq->poll.events & EPOLLONESHOT) {
5457 /* we now have a detached poll request. reissue. */
5460 spin_unlock_irq(&ctx->completion_lock);
5462 req_set_fail_links(req);
5463 io_req_complete(req, ret);
5466 /* only mask one event flags, keep behavior flags */
5467 if (req->poll.update_events) {
5468 preq->poll.events &= ~0xffff;
5469 preq->poll.events |= req->poll.events & 0xffff;
5470 preq->poll.events |= IO_POLL_UNMASK;
5472 if (req->poll.update_user_data)
5473 preq->user_data = req->poll.new_user_data;
5475 /* complete update request, we're done with it */
5476 io_req_complete(req, ret);
5478 ret = __io_poll_add(preq);
5480 req_set_fail_links(preq);
5481 io_req_complete(preq, ret);
5486 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5488 if (!req->poll.update_events && !req->poll.update_user_data)
5489 return __io_poll_add(req);
5490 return io_poll_update(req);
5493 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5495 struct io_timeout_data *data = container_of(timer,
5496 struct io_timeout_data, timer);
5497 struct io_kiocb *req = data->req;
5498 struct io_ring_ctx *ctx = req->ctx;
5499 unsigned long flags;
5501 spin_lock_irqsave(&ctx->completion_lock, flags);
5502 list_del_init(&req->timeout.list);
5503 atomic_set(&req->ctx->cq_timeouts,
5504 atomic_read(&req->ctx->cq_timeouts) + 1);
5506 io_cqring_fill_event(req, -ETIME);
5507 io_commit_cqring(ctx);
5508 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5510 io_cqring_ev_posted(ctx);
5511 req_set_fail_links(req);
5513 return HRTIMER_NORESTART;
5516 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5518 __must_hold(&ctx->completion_lock)
5520 struct io_timeout_data *io;
5521 struct io_kiocb *req;
5524 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5525 if (user_data == req->user_data) {
5532 return ERR_PTR(ret);
5534 io = req->async_data;
5535 ret = hrtimer_try_to_cancel(&io->timer);
5537 return ERR_PTR(-EALREADY);
5538 list_del_init(&req->timeout.list);
5542 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5543 __must_hold(&ctx->completion_lock)
5545 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5548 return PTR_ERR(req);
5550 req_set_fail_links(req);
5551 io_cqring_fill_event(req, -ECANCELED);
5552 io_put_req_deferred(req, 1);
5556 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5557 struct timespec64 *ts, enum hrtimer_mode mode)
5558 __must_hold(&ctx->completion_lock)
5560 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5561 struct io_timeout_data *data;
5564 return PTR_ERR(req);
5566 req->timeout.off = 0; /* noseq */
5567 data = req->async_data;
5568 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5569 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5570 data->timer.function = io_timeout_fn;
5571 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5575 static int io_timeout_remove_prep(struct io_kiocb *req,
5576 const struct io_uring_sqe *sqe)
5578 struct io_timeout_rem *tr = &req->timeout_rem;
5580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5582 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5584 if (sqe->ioprio || sqe->buf_index || sqe->len)
5587 tr->addr = READ_ONCE(sqe->addr);
5588 tr->flags = READ_ONCE(sqe->timeout_flags);
5589 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5590 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5592 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5594 } else if (tr->flags) {
5595 /* timeout removal doesn't support flags */
5602 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5604 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5609 * Remove or update an existing timeout command
5611 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5613 struct io_timeout_rem *tr = &req->timeout_rem;
5614 struct io_ring_ctx *ctx = req->ctx;
5617 spin_lock_irq(&ctx->completion_lock);
5618 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5619 ret = io_timeout_cancel(ctx, tr->addr);
5621 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5622 io_translate_timeout_mode(tr->flags));
5624 io_cqring_fill_event(req, ret);
5625 io_commit_cqring(ctx);
5626 spin_unlock_irq(&ctx->completion_lock);
5627 io_cqring_ev_posted(ctx);
5629 req_set_fail_links(req);
5634 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5635 bool is_timeout_link)
5637 struct io_timeout_data *data;
5639 u32 off = READ_ONCE(sqe->off);
5641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5643 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5645 if (off && is_timeout_link)
5647 flags = READ_ONCE(sqe->timeout_flags);
5648 if (flags & ~IORING_TIMEOUT_ABS)
5651 req->timeout.off = off;
5653 if (!req->async_data && io_alloc_async_data(req))
5656 data = req->async_data;
5659 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5662 data->mode = io_translate_timeout_mode(flags);
5663 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5664 if (is_timeout_link)
5665 io_req_track_inflight(req);
5669 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5671 struct io_ring_ctx *ctx = req->ctx;
5672 struct io_timeout_data *data = req->async_data;
5673 struct list_head *entry;
5674 u32 tail, off = req->timeout.off;
5676 spin_lock_irq(&ctx->completion_lock);
5679 * sqe->off holds how many events that need to occur for this
5680 * timeout event to be satisfied. If it isn't set, then this is
5681 * a pure timeout request, sequence isn't used.
5683 if (io_is_timeout_noseq(req)) {
5684 entry = ctx->timeout_list.prev;
5688 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5689 req->timeout.target_seq = tail + off;
5691 /* Update the last seq here in case io_flush_timeouts() hasn't.
5692 * This is safe because ->completion_lock is held, and submissions
5693 * and completions are never mixed in the same ->completion_lock section.
5695 ctx->cq_last_tm_flush = tail;
5698 * Insertion sort, ensuring the first entry in the list is always
5699 * the one we need first.
5701 list_for_each_prev(entry, &ctx->timeout_list) {
5702 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5705 if (io_is_timeout_noseq(nxt))
5707 /* nxt.seq is behind @tail, otherwise would've been completed */
5708 if (off >= nxt->timeout.target_seq - tail)
5712 list_add(&req->timeout.list, entry);
5713 data->timer.function = io_timeout_fn;
5714 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5715 spin_unlock_irq(&ctx->completion_lock);
5719 struct io_cancel_data {
5720 struct io_ring_ctx *ctx;
5724 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5726 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5727 struct io_cancel_data *cd = data;
5729 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5732 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5733 struct io_ring_ctx *ctx)
5735 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5736 enum io_wq_cancel cancel_ret;
5739 if (!tctx || !tctx->io_wq)
5742 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5743 switch (cancel_ret) {
5744 case IO_WQ_CANCEL_OK:
5747 case IO_WQ_CANCEL_RUNNING:
5750 case IO_WQ_CANCEL_NOTFOUND:
5758 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5759 struct io_kiocb *req, __u64 sqe_addr,
5762 unsigned long flags;
5765 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5766 spin_lock_irqsave(&ctx->completion_lock, flags);
5769 ret = io_timeout_cancel(ctx, sqe_addr);
5772 ret = io_poll_cancel(ctx, sqe_addr);
5776 io_cqring_fill_event(req, ret);
5777 io_commit_cqring(ctx);
5778 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5779 io_cqring_ev_posted(ctx);
5782 req_set_fail_links(req);
5785 static int io_async_cancel_prep(struct io_kiocb *req,
5786 const struct io_uring_sqe *sqe)
5788 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5790 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5792 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5795 req->cancel.addr = READ_ONCE(sqe->addr);
5799 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5801 struct io_ring_ctx *ctx = req->ctx;
5802 u64 sqe_addr = req->cancel.addr;
5803 struct io_tctx_node *node;
5806 /* tasks should wait for their io-wq threads, so safe w/o sync */
5807 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5808 spin_lock_irq(&ctx->completion_lock);
5811 ret = io_timeout_cancel(ctx, sqe_addr);
5814 ret = io_poll_cancel(ctx, sqe_addr);
5817 spin_unlock_irq(&ctx->completion_lock);
5819 /* slow path, try all io-wq's */
5820 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5822 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5823 struct io_uring_task *tctx = node->task->io_uring;
5825 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5829 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5831 spin_lock_irq(&ctx->completion_lock);
5833 io_cqring_fill_event(req, ret);
5834 io_commit_cqring(ctx);
5835 spin_unlock_irq(&ctx->completion_lock);
5836 io_cqring_ev_posted(ctx);
5839 req_set_fail_links(req);
5844 static int io_rsrc_update_prep(struct io_kiocb *req,
5845 const struct io_uring_sqe *sqe)
5847 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5849 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5851 if (sqe->ioprio || sqe->rw_flags)
5854 req->rsrc_update.offset = READ_ONCE(sqe->off);
5855 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5856 if (!req->rsrc_update.nr_args)
5858 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5862 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5864 struct io_ring_ctx *ctx = req->ctx;
5865 struct io_uring_rsrc_update up;
5868 if (issue_flags & IO_URING_F_NONBLOCK)
5871 up.offset = req->rsrc_update.offset;
5872 up.data = req->rsrc_update.arg;
5874 mutex_lock(&ctx->uring_lock);
5875 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5876 mutex_unlock(&ctx->uring_lock);
5879 req_set_fail_links(req);
5880 __io_req_complete(req, issue_flags, ret, 0);
5884 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5886 switch (req->opcode) {
5889 case IORING_OP_READV:
5890 case IORING_OP_READ_FIXED:
5891 case IORING_OP_READ:
5892 return io_read_prep(req, sqe);
5893 case IORING_OP_WRITEV:
5894 case IORING_OP_WRITE_FIXED:
5895 case IORING_OP_WRITE:
5896 return io_write_prep(req, sqe);
5897 case IORING_OP_POLL_ADD:
5898 return io_poll_add_prep(req, sqe);
5899 case IORING_OP_POLL_REMOVE:
5900 return io_poll_remove_prep(req, sqe);
5901 case IORING_OP_FSYNC:
5902 return io_fsync_prep(req, sqe);
5903 case IORING_OP_SYNC_FILE_RANGE:
5904 return io_sfr_prep(req, sqe);
5905 case IORING_OP_SENDMSG:
5906 case IORING_OP_SEND:
5907 return io_sendmsg_prep(req, sqe);
5908 case IORING_OP_RECVMSG:
5909 case IORING_OP_RECV:
5910 return io_recvmsg_prep(req, sqe);
5911 case IORING_OP_CONNECT:
5912 return io_connect_prep(req, sqe);
5913 case IORING_OP_TIMEOUT:
5914 return io_timeout_prep(req, sqe, false);
5915 case IORING_OP_TIMEOUT_REMOVE:
5916 return io_timeout_remove_prep(req, sqe);
5917 case IORING_OP_ASYNC_CANCEL:
5918 return io_async_cancel_prep(req, sqe);
5919 case IORING_OP_LINK_TIMEOUT:
5920 return io_timeout_prep(req, sqe, true);
5921 case IORING_OP_ACCEPT:
5922 return io_accept_prep(req, sqe);
5923 case IORING_OP_FALLOCATE:
5924 return io_fallocate_prep(req, sqe);
5925 case IORING_OP_OPENAT:
5926 return io_openat_prep(req, sqe);
5927 case IORING_OP_CLOSE:
5928 return io_close_prep(req, sqe);
5929 case IORING_OP_FILES_UPDATE:
5930 return io_rsrc_update_prep(req, sqe);
5931 case IORING_OP_STATX:
5932 return io_statx_prep(req, sqe);
5933 case IORING_OP_FADVISE:
5934 return io_fadvise_prep(req, sqe);
5935 case IORING_OP_MADVISE:
5936 return io_madvise_prep(req, sqe);
5937 case IORING_OP_OPENAT2:
5938 return io_openat2_prep(req, sqe);
5939 case IORING_OP_EPOLL_CTL:
5940 return io_epoll_ctl_prep(req, sqe);
5941 case IORING_OP_SPLICE:
5942 return io_splice_prep(req, sqe);
5943 case IORING_OP_PROVIDE_BUFFERS:
5944 return io_provide_buffers_prep(req, sqe);
5945 case IORING_OP_REMOVE_BUFFERS:
5946 return io_remove_buffers_prep(req, sqe);
5948 return io_tee_prep(req, sqe);
5949 case IORING_OP_SHUTDOWN:
5950 return io_shutdown_prep(req, sqe);
5951 case IORING_OP_RENAMEAT:
5952 return io_renameat_prep(req, sqe);
5953 case IORING_OP_UNLINKAT:
5954 return io_unlinkat_prep(req, sqe);
5957 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5962 static int io_req_prep_async(struct io_kiocb *req)
5964 if (!io_op_defs[req->opcode].needs_async_setup)
5966 if (WARN_ON_ONCE(req->async_data))
5968 if (io_alloc_async_data(req))
5971 switch (req->opcode) {
5972 case IORING_OP_READV:
5973 return io_rw_prep_async(req, READ);
5974 case IORING_OP_WRITEV:
5975 return io_rw_prep_async(req, WRITE);
5976 case IORING_OP_SENDMSG:
5977 return io_sendmsg_prep_async(req);
5978 case IORING_OP_RECVMSG:
5979 return io_recvmsg_prep_async(req);
5980 case IORING_OP_CONNECT:
5981 return io_connect_prep_async(req);
5983 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5988 static u32 io_get_sequence(struct io_kiocb *req)
5990 struct io_kiocb *pos;
5991 struct io_ring_ctx *ctx = req->ctx;
5992 u32 total_submitted, nr_reqs = 0;
5994 io_for_each_link(pos, req)
5997 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5998 return total_submitted - nr_reqs;
6001 static int io_req_defer(struct io_kiocb *req)
6003 struct io_ring_ctx *ctx = req->ctx;
6004 struct io_defer_entry *de;
6008 /* Still need defer if there is pending req in defer list. */
6009 if (likely(list_empty_careful(&ctx->defer_list) &&
6010 !(req->flags & REQ_F_IO_DRAIN)))
6013 seq = io_get_sequence(req);
6014 /* Still a chance to pass the sequence check */
6015 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6018 ret = io_req_prep_async(req);
6021 io_prep_async_link(req);
6022 de = kmalloc(sizeof(*de), GFP_KERNEL);
6026 spin_lock_irq(&ctx->completion_lock);
6027 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6028 spin_unlock_irq(&ctx->completion_lock);
6030 io_queue_async_work(req);
6031 return -EIOCBQUEUED;
6034 trace_io_uring_defer(ctx, req, req->user_data);
6037 list_add_tail(&de->list, &ctx->defer_list);
6038 spin_unlock_irq(&ctx->completion_lock);
6039 return -EIOCBQUEUED;
6042 static void io_clean_op(struct io_kiocb *req)
6044 if (req->flags & REQ_F_BUFFER_SELECTED) {
6045 switch (req->opcode) {
6046 case IORING_OP_READV:
6047 case IORING_OP_READ_FIXED:
6048 case IORING_OP_READ:
6049 kfree((void *)(unsigned long)req->rw.addr);
6051 case IORING_OP_RECVMSG:
6052 case IORING_OP_RECV:
6053 kfree(req->sr_msg.kbuf);
6056 req->flags &= ~REQ_F_BUFFER_SELECTED;
6059 if (req->flags & REQ_F_NEED_CLEANUP) {
6060 switch (req->opcode) {
6061 case IORING_OP_READV:
6062 case IORING_OP_READ_FIXED:
6063 case IORING_OP_READ:
6064 case IORING_OP_WRITEV:
6065 case IORING_OP_WRITE_FIXED:
6066 case IORING_OP_WRITE: {
6067 struct io_async_rw *io = req->async_data;
6069 kfree(io->free_iovec);
6072 case IORING_OP_RECVMSG:
6073 case IORING_OP_SENDMSG: {
6074 struct io_async_msghdr *io = req->async_data;
6076 kfree(io->free_iov);
6079 case IORING_OP_SPLICE:
6081 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6082 io_put_file(req->splice.file_in);
6084 case IORING_OP_OPENAT:
6085 case IORING_OP_OPENAT2:
6086 if (req->open.filename)
6087 putname(req->open.filename);
6089 case IORING_OP_RENAMEAT:
6090 putname(req->rename.oldpath);
6091 putname(req->rename.newpath);
6093 case IORING_OP_UNLINKAT:
6094 putname(req->unlink.filename);
6097 req->flags &= ~REQ_F_NEED_CLEANUP;
6101 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6103 struct io_ring_ctx *ctx = req->ctx;
6104 const struct cred *creds = NULL;
6107 if (req->work.creds && req->work.creds != current_cred())
6108 creds = override_creds(req->work.creds);
6110 switch (req->opcode) {
6112 ret = io_nop(req, issue_flags);
6114 case IORING_OP_READV:
6115 case IORING_OP_READ_FIXED:
6116 case IORING_OP_READ:
6117 ret = io_read(req, issue_flags);
6119 case IORING_OP_WRITEV:
6120 case IORING_OP_WRITE_FIXED:
6121 case IORING_OP_WRITE:
6122 ret = io_write(req, issue_flags);
6124 case IORING_OP_FSYNC:
6125 ret = io_fsync(req, issue_flags);
6127 case IORING_OP_POLL_ADD:
6128 ret = io_poll_add(req, issue_flags);
6130 case IORING_OP_POLL_REMOVE:
6131 ret = io_poll_remove(req, issue_flags);
6133 case IORING_OP_SYNC_FILE_RANGE:
6134 ret = io_sync_file_range(req, issue_flags);
6136 case IORING_OP_SENDMSG:
6137 ret = io_sendmsg(req, issue_flags);
6139 case IORING_OP_SEND:
6140 ret = io_send(req, issue_flags);
6142 case IORING_OP_RECVMSG:
6143 ret = io_recvmsg(req, issue_flags);
6145 case IORING_OP_RECV:
6146 ret = io_recv(req, issue_flags);
6148 case IORING_OP_TIMEOUT:
6149 ret = io_timeout(req, issue_flags);
6151 case IORING_OP_TIMEOUT_REMOVE:
6152 ret = io_timeout_remove(req, issue_flags);
6154 case IORING_OP_ACCEPT:
6155 ret = io_accept(req, issue_flags);
6157 case IORING_OP_CONNECT:
6158 ret = io_connect(req, issue_flags);
6160 case IORING_OP_ASYNC_CANCEL:
6161 ret = io_async_cancel(req, issue_flags);
6163 case IORING_OP_FALLOCATE:
6164 ret = io_fallocate(req, issue_flags);
6166 case IORING_OP_OPENAT:
6167 ret = io_openat(req, issue_flags);
6169 case IORING_OP_CLOSE:
6170 ret = io_close(req, issue_flags);
6172 case IORING_OP_FILES_UPDATE:
6173 ret = io_files_update(req, issue_flags);
6175 case IORING_OP_STATX:
6176 ret = io_statx(req, issue_flags);
6178 case IORING_OP_FADVISE:
6179 ret = io_fadvise(req, issue_flags);
6181 case IORING_OP_MADVISE:
6182 ret = io_madvise(req, issue_flags);
6184 case IORING_OP_OPENAT2:
6185 ret = io_openat2(req, issue_flags);
6187 case IORING_OP_EPOLL_CTL:
6188 ret = io_epoll_ctl(req, issue_flags);
6190 case IORING_OP_SPLICE:
6191 ret = io_splice(req, issue_flags);
6193 case IORING_OP_PROVIDE_BUFFERS:
6194 ret = io_provide_buffers(req, issue_flags);
6196 case IORING_OP_REMOVE_BUFFERS:
6197 ret = io_remove_buffers(req, issue_flags);
6200 ret = io_tee(req, issue_flags);
6202 case IORING_OP_SHUTDOWN:
6203 ret = io_shutdown(req, issue_flags);
6205 case IORING_OP_RENAMEAT:
6206 ret = io_renameat(req, issue_flags);
6208 case IORING_OP_UNLINKAT:
6209 ret = io_unlinkat(req, issue_flags);
6217 revert_creds(creds);
6222 /* If the op doesn't have a file, we're not polling for it */
6223 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6224 const bool in_async = io_wq_current_is_worker();
6226 /* workqueue context doesn't hold uring_lock, grab it now */
6228 mutex_lock(&ctx->uring_lock);
6230 io_iopoll_req_issued(req, in_async);
6233 mutex_unlock(&ctx->uring_lock);
6239 static void io_wq_submit_work(struct io_wq_work *work)
6241 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6242 struct io_kiocb *timeout;
6245 timeout = io_prep_linked_timeout(req);
6247 io_queue_linked_timeout(timeout);
6249 if (work->flags & IO_WQ_WORK_CANCEL)
6254 ret = io_issue_sqe(req, 0);
6256 * We can get EAGAIN for polled IO even though we're
6257 * forcing a sync submission from here, since we can't
6258 * wait for request slots on the block side.
6266 /* avoid locking problems by failing it from a clean context */
6268 /* io-wq is going to take one down */
6270 io_req_task_queue_fail(req, ret);
6274 #define FFS_ASYNC_READ 0x1UL
6275 #define FFS_ASYNC_WRITE 0x2UL
6277 #define FFS_ISREG 0x4UL
6279 #define FFS_ISREG 0x0UL
6281 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6283 static inline struct io_fixed_file *io_fixed_file_slot(struct io_rsrc_data *file_data,
6286 struct fixed_rsrc_table *table;
6288 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6289 return &table->files[i & IORING_FILE_TABLE_MASK];
6292 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6295 struct io_fixed_file *slot = io_fixed_file_slot(ctx->file_data, index);
6297 return (struct file *) (slot->file_ptr & FFS_MASK);
6300 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6302 unsigned long file_ptr = (unsigned long) file;
6304 if (__io_file_supports_async(file, READ))
6305 file_ptr |= FFS_ASYNC_READ;
6306 if (__io_file_supports_async(file, WRITE))
6307 file_ptr |= FFS_ASYNC_WRITE;
6308 if (S_ISREG(file_inode(file)->i_mode))
6309 file_ptr |= FFS_ISREG;
6310 file_slot->file_ptr = file_ptr;
6313 static struct file *io_file_get(struct io_submit_state *state,
6314 struct io_kiocb *req, int fd, bool fixed)
6316 struct io_ring_ctx *ctx = req->ctx;
6320 unsigned long file_ptr;
6322 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6324 fd = array_index_nospec(fd, ctx->nr_user_files);
6325 file_ptr = io_fixed_file_slot(ctx->file_data, fd)->file_ptr;
6326 file = (struct file *) (file_ptr & FFS_MASK);
6327 file_ptr &= ~FFS_MASK;
6328 /* mask in overlapping REQ_F and FFS bits */
6329 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6330 io_req_set_rsrc_node(req);
6332 trace_io_uring_file_get(ctx, fd);
6333 file = __io_file_get(state, fd);
6335 /* we don't allow fixed io_uring files */
6336 if (file && unlikely(file->f_op == &io_uring_fops))
6337 io_req_track_inflight(req);
6343 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6345 struct io_timeout_data *data = container_of(timer,
6346 struct io_timeout_data, timer);
6347 struct io_kiocb *prev, *req = data->req;
6348 struct io_ring_ctx *ctx = req->ctx;
6349 unsigned long flags;
6351 spin_lock_irqsave(&ctx->completion_lock, flags);
6352 prev = req->timeout.head;
6353 req->timeout.head = NULL;
6356 * We don't expect the list to be empty, that will only happen if we
6357 * race with the completion of the linked work.
6359 if (prev && req_ref_inc_not_zero(prev))
6360 io_remove_next_linked(prev);
6363 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6366 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6367 io_put_req_deferred(prev, 1);
6369 io_req_complete_post(req, -ETIME, 0);
6371 io_put_req_deferred(req, 1);
6372 return HRTIMER_NORESTART;
6375 static void io_queue_linked_timeout(struct io_kiocb *req)
6377 struct io_ring_ctx *ctx = req->ctx;
6379 spin_lock_irq(&ctx->completion_lock);
6381 * If the back reference is NULL, then our linked request finished
6382 * before we got a chance to setup the timer
6384 if (req->timeout.head) {
6385 struct io_timeout_data *data = req->async_data;
6387 data->timer.function = io_link_timeout_fn;
6388 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6391 spin_unlock_irq(&ctx->completion_lock);
6392 /* drop submission reference */
6396 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6398 struct io_kiocb *nxt = req->link;
6400 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6401 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6404 nxt->timeout.head = req;
6405 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6406 req->flags |= REQ_F_LINK_TIMEOUT;
6410 static void __io_queue_sqe(struct io_kiocb *req)
6412 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6415 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6418 * We async punt it if the file wasn't marked NOWAIT, or if the file
6419 * doesn't support non-blocking read/write attempts
6422 /* drop submission reference */
6423 if (req->flags & REQ_F_COMPLETE_INLINE) {
6424 struct io_ring_ctx *ctx = req->ctx;
6425 struct io_comp_state *cs = &ctx->submit_state.comp;
6427 cs->reqs[cs->nr++] = req;
6428 if (cs->nr == ARRAY_SIZE(cs->reqs))
6429 io_submit_flush_completions(cs, ctx);
6433 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6434 if (!io_arm_poll_handler(req)) {
6436 * Queued up for async execution, worker will release
6437 * submit reference when the iocb is actually submitted.
6439 io_queue_async_work(req);
6442 io_req_complete_failed(req, ret);
6445 io_queue_linked_timeout(linked_timeout);
6448 static void io_queue_sqe(struct io_kiocb *req)
6452 ret = io_req_defer(req);
6454 if (ret != -EIOCBQUEUED) {
6456 io_req_complete_failed(req, ret);
6458 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6459 ret = io_req_prep_async(req);
6462 io_queue_async_work(req);
6464 __io_queue_sqe(req);
6469 * Check SQE restrictions (opcode and flags).
6471 * Returns 'true' if SQE is allowed, 'false' otherwise.
6473 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6474 struct io_kiocb *req,
6475 unsigned int sqe_flags)
6477 if (!ctx->restricted)
6480 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6483 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6484 ctx->restrictions.sqe_flags_required)
6487 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6488 ctx->restrictions.sqe_flags_required))
6494 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6495 const struct io_uring_sqe *sqe)
6497 struct io_submit_state *state;
6498 unsigned int sqe_flags;
6499 int personality, ret = 0;
6501 req->opcode = READ_ONCE(sqe->opcode);
6502 /* same numerical values with corresponding REQ_F_*, safe to copy */
6503 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6504 req->user_data = READ_ONCE(sqe->user_data);
6505 req->async_data = NULL;
6509 req->fixed_rsrc_refs = NULL;
6510 /* one is dropped after submission, the other at completion */
6511 atomic_set(&req->refs, 2);
6512 req->task = current;
6514 req->work.creds = NULL;
6516 /* enforce forwards compatibility on users */
6517 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6522 if (unlikely(req->opcode >= IORING_OP_LAST))
6525 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6528 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6529 !io_op_defs[req->opcode].buffer_select)
6532 personality = READ_ONCE(sqe->personality);
6534 req->work.creds = xa_load(&ctx->personalities, personality);
6535 if (!req->work.creds)
6537 get_cred(req->work.creds);
6539 state = &ctx->submit_state;
6542 * Plug now if we have more than 1 IO left after this, and the target
6543 * is potentially a read/write to block based storage.
6545 if (!state->plug_started && state->ios_left > 1 &&
6546 io_op_defs[req->opcode].plug) {
6547 blk_start_plug(&state->plug);
6548 state->plug_started = true;
6551 if (io_op_defs[req->opcode].needs_file) {
6552 bool fixed = req->flags & REQ_F_FIXED_FILE;
6554 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6555 if (unlikely(!req->file))
6563 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6564 const struct io_uring_sqe *sqe)
6566 struct io_submit_link *link = &ctx->submit_state.link;
6569 ret = io_init_req(ctx, req, sqe);
6570 if (unlikely(ret)) {
6573 /* fail even hard links since we don't submit */
6574 link->head->flags |= REQ_F_FAIL_LINK;
6575 io_req_complete_failed(link->head, -ECANCELED);
6578 io_req_complete_failed(req, ret);
6581 ret = io_req_prep(req, sqe);
6585 /* don't need @sqe from now on */
6586 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6587 true, ctx->flags & IORING_SETUP_SQPOLL);
6590 * If we already have a head request, queue this one for async
6591 * submittal once the head completes. If we don't have a head but
6592 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6593 * submitted sync once the chain is complete. If none of those
6594 * conditions are true (normal request), then just queue it.
6597 struct io_kiocb *head = link->head;
6600 * Taking sequential execution of a link, draining both sides
6601 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6602 * requests in the link. So, it drains the head and the
6603 * next after the link request. The last one is done via
6604 * drain_next flag to persist the effect across calls.
6606 if (req->flags & REQ_F_IO_DRAIN) {
6607 head->flags |= REQ_F_IO_DRAIN;
6608 ctx->drain_next = 1;
6610 ret = io_req_prep_async(req);
6613 trace_io_uring_link(ctx, req, head);
6614 link->last->link = req;
6617 /* last request of a link, enqueue the link */
6618 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6623 if (unlikely(ctx->drain_next)) {
6624 req->flags |= REQ_F_IO_DRAIN;
6625 ctx->drain_next = 0;
6627 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6639 * Batched submission is done, ensure local IO is flushed out.
6641 static void io_submit_state_end(struct io_submit_state *state,
6642 struct io_ring_ctx *ctx)
6644 if (state->link.head)
6645 io_queue_sqe(state->link.head);
6647 io_submit_flush_completions(&state->comp, ctx);
6648 if (state->plug_started)
6649 blk_finish_plug(&state->plug);
6650 io_state_file_put(state);
6654 * Start submission side cache.
6656 static void io_submit_state_start(struct io_submit_state *state,
6657 unsigned int max_ios)
6659 state->plug_started = false;
6660 state->ios_left = max_ios;
6661 /* set only head, no need to init link_last in advance */
6662 state->link.head = NULL;
6665 static void io_commit_sqring(struct io_ring_ctx *ctx)
6667 struct io_rings *rings = ctx->rings;
6670 * Ensure any loads from the SQEs are done at this point,
6671 * since once we write the new head, the application could
6672 * write new data to them.
6674 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6678 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6679 * that is mapped by userspace. This means that care needs to be taken to
6680 * ensure that reads are stable, as we cannot rely on userspace always
6681 * being a good citizen. If members of the sqe are validated and then later
6682 * used, it's important that those reads are done through READ_ONCE() to
6683 * prevent a re-load down the line.
6685 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6687 u32 *sq_array = ctx->sq_array;
6691 * The cached sq head (or cq tail) serves two purposes:
6693 * 1) allows us to batch the cost of updating the user visible
6695 * 2) allows the kernel side to track the head on its own, even
6696 * though the application is the one updating it.
6698 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6699 if (likely(head < ctx->sq_entries))
6700 return &ctx->sq_sqes[head];
6702 /* drop invalid entries */
6703 ctx->cached_sq_dropped++;
6704 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6708 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6712 /* if we have a backlog and couldn't flush it all, return BUSY */
6713 if (test_bit(0, &ctx->sq_check_overflow)) {
6714 if (!__io_cqring_overflow_flush(ctx, false))
6718 /* make sure SQ entry isn't read before tail */
6719 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6721 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6724 percpu_counter_add(¤t->io_uring->inflight, nr);
6725 refcount_add(nr, ¤t->usage);
6726 io_submit_state_start(&ctx->submit_state, nr);
6728 while (submitted < nr) {
6729 const struct io_uring_sqe *sqe;
6730 struct io_kiocb *req;
6732 req = io_alloc_req(ctx);
6733 if (unlikely(!req)) {
6735 submitted = -EAGAIN;
6738 sqe = io_get_sqe(ctx);
6739 if (unlikely(!sqe)) {
6740 kmem_cache_free(req_cachep, req);
6743 /* will complete beyond this point, count as submitted */
6745 if (io_submit_sqe(ctx, req, sqe))
6749 if (unlikely(submitted != nr)) {
6750 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6751 struct io_uring_task *tctx = current->io_uring;
6752 int unused = nr - ref_used;
6754 percpu_ref_put_many(&ctx->refs, unused);
6755 percpu_counter_sub(&tctx->inflight, unused);
6756 put_task_struct_many(current, unused);
6759 io_submit_state_end(&ctx->submit_state, ctx);
6760 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6761 io_commit_sqring(ctx);
6766 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6768 /* Tell userspace we may need a wakeup call */
6769 spin_lock_irq(&ctx->completion_lock);
6770 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6771 spin_unlock_irq(&ctx->completion_lock);
6774 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6776 spin_lock_irq(&ctx->completion_lock);
6777 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6778 spin_unlock_irq(&ctx->completion_lock);
6781 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6783 unsigned int to_submit;
6786 to_submit = io_sqring_entries(ctx);
6787 /* if we're handling multiple rings, cap submit size for fairness */
6788 if (cap_entries && to_submit > 8)
6791 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6792 unsigned nr_events = 0;
6794 mutex_lock(&ctx->uring_lock);
6795 if (!list_empty(&ctx->iopoll_list))
6796 io_do_iopoll(ctx, &nr_events, 0);
6798 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6799 !(ctx->flags & IORING_SETUP_R_DISABLED))
6800 ret = io_submit_sqes(ctx, to_submit);
6801 mutex_unlock(&ctx->uring_lock);
6804 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6805 wake_up(&ctx->sqo_sq_wait);
6810 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6812 struct io_ring_ctx *ctx;
6813 unsigned sq_thread_idle = 0;
6815 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6816 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6817 sqd->sq_thread_idle = sq_thread_idle;
6820 static int io_sq_thread(void *data)
6822 struct io_sq_data *sqd = data;
6823 struct io_ring_ctx *ctx;
6824 unsigned long timeout = 0;
6825 char buf[TASK_COMM_LEN];
6828 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6829 set_task_comm(current, buf);
6830 current->pf_io_worker = NULL;
6832 if (sqd->sq_cpu != -1)
6833 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6835 set_cpus_allowed_ptr(current, cpu_online_mask);
6836 current->flags |= PF_NO_SETAFFINITY;
6838 mutex_lock(&sqd->lock);
6839 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6841 bool cap_entries, sqt_spin, needs_sched;
6843 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6844 signal_pending(current)) {
6845 bool did_sig = false;
6847 mutex_unlock(&sqd->lock);
6848 if (signal_pending(current)) {
6849 struct ksignal ksig;
6851 did_sig = get_signal(&ksig);
6854 mutex_lock(&sqd->lock);
6858 io_run_task_work_head(&sqd->park_task_work);
6859 timeout = jiffies + sqd->sq_thread_idle;
6863 cap_entries = !list_is_singular(&sqd->ctx_list);
6864 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6865 const struct cred *creds = NULL;
6867 if (ctx->sq_creds != current_cred())
6868 creds = override_creds(ctx->sq_creds);
6869 ret = __io_sq_thread(ctx, cap_entries);
6871 revert_creds(creds);
6872 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6876 if (sqt_spin || !time_after(jiffies, timeout)) {
6880 timeout = jiffies + sqd->sq_thread_idle;
6885 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6886 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6887 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6888 !list_empty_careful(&ctx->iopoll_list)) {
6889 needs_sched = false;
6892 if (io_sqring_entries(ctx)) {
6893 needs_sched = false;
6898 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6899 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6900 io_ring_set_wakeup_flag(ctx);
6902 mutex_unlock(&sqd->lock);
6904 mutex_lock(&sqd->lock);
6905 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6906 io_ring_clear_wakeup_flag(ctx);
6909 finish_wait(&sqd->wait, &wait);
6910 io_run_task_work_head(&sqd->park_task_work);
6911 timeout = jiffies + sqd->sq_thread_idle;
6914 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6915 io_uring_cancel_sqpoll(ctx);
6917 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6918 io_ring_set_wakeup_flag(ctx);
6919 mutex_unlock(&sqd->lock);
6922 io_run_task_work_head(&sqd->park_task_work);
6923 complete(&sqd->exited);
6927 struct io_wait_queue {
6928 struct wait_queue_entry wq;
6929 struct io_ring_ctx *ctx;
6931 unsigned nr_timeouts;
6934 static inline bool io_should_wake(struct io_wait_queue *iowq)
6936 struct io_ring_ctx *ctx = iowq->ctx;
6939 * Wake up if we have enough events, or if a timeout occurred since we
6940 * started waiting. For timeouts, we always want to return to userspace,
6941 * regardless of event count.
6943 return io_cqring_events(ctx) >= iowq->to_wait ||
6944 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6947 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6948 int wake_flags, void *key)
6950 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6954 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6955 * the task, and the next invocation will do it.
6957 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6958 return autoremove_wake_function(curr, mode, wake_flags, key);
6962 static int io_run_task_work_sig(void)
6964 if (io_run_task_work())
6966 if (!signal_pending(current))
6968 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6969 return -ERESTARTSYS;
6973 /* when returns >0, the caller should retry */
6974 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6975 struct io_wait_queue *iowq,
6976 signed long *timeout)
6980 /* make sure we run task_work before checking for signals */
6981 ret = io_run_task_work_sig();
6982 if (ret || io_should_wake(iowq))
6984 /* let the caller flush overflows, retry */
6985 if (test_bit(0, &ctx->cq_check_overflow))
6988 *timeout = schedule_timeout(*timeout);
6989 return !*timeout ? -ETIME : 1;
6993 * Wait until events become available, if we don't already have some. The
6994 * application must reap them itself, as they reside on the shared cq ring.
6996 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6997 const sigset_t __user *sig, size_t sigsz,
6998 struct __kernel_timespec __user *uts)
7000 struct io_wait_queue iowq = {
7003 .func = io_wake_function,
7004 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7007 .to_wait = min_events,
7009 struct io_rings *rings = ctx->rings;
7010 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7014 io_cqring_overflow_flush(ctx, false);
7015 if (io_cqring_events(ctx) >= min_events)
7017 if (!io_run_task_work())
7022 #ifdef CONFIG_COMPAT
7023 if (in_compat_syscall())
7024 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7028 ret = set_user_sigmask(sig, sigsz);
7035 struct timespec64 ts;
7037 if (get_timespec64(&ts, uts))
7039 timeout = timespec64_to_jiffies(&ts);
7042 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7043 trace_io_uring_cqring_wait(ctx, min_events);
7045 /* if we can't even flush overflow, don't wait for more */
7046 if (!io_cqring_overflow_flush(ctx, false)) {
7050 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7051 TASK_INTERRUPTIBLE);
7052 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7053 finish_wait(&ctx->wait, &iowq.wq);
7057 restore_saved_sigmask_unless(ret == -EINTR);
7059 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7062 static void io_free_file_tables(struct io_rsrc_data *data, unsigned nr_files)
7064 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7066 for (i = 0; i < nr_tables; i++)
7067 kfree(data->table[i].files);
7072 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7074 #if defined(CONFIG_UNIX)
7075 if (ctx->ring_sock) {
7076 struct sock *sock = ctx->ring_sock->sk;
7077 struct sk_buff *skb;
7079 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7085 for (i = 0; i < ctx->nr_user_files; i++) {
7088 file = io_file_from_index(ctx, i);
7095 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7097 struct io_rsrc_data *data = container_of(ref, struct io_rsrc_data, refs);
7099 complete(&data->done);
7102 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7104 spin_lock_bh(&ctx->rsrc_ref_lock);
7107 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7109 spin_unlock_bh(&ctx->rsrc_ref_lock);
7112 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7114 percpu_ref_exit(&ref_node->refs);
7118 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7119 struct io_rsrc_data *data_to_kill)
7121 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7122 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7125 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7127 rsrc_node->rsrc_data = data_to_kill;
7128 io_rsrc_ref_lock(ctx);
7129 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7130 io_rsrc_ref_unlock(ctx);
7132 percpu_ref_get(&data_to_kill->refs);
7133 percpu_ref_kill(&rsrc_node->refs);
7134 ctx->rsrc_node = NULL;
7137 if (!ctx->rsrc_node) {
7138 ctx->rsrc_node = ctx->rsrc_backup_node;
7139 ctx->rsrc_backup_node = NULL;
7143 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7145 if (ctx->rsrc_backup_node)
7147 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7148 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7151 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7155 /* As we may drop ->uring_lock, other task may have started quiesce */
7159 data->quiesce = true;
7161 ret = io_rsrc_node_switch_start(ctx);
7164 io_rsrc_node_switch(ctx, data);
7166 percpu_ref_kill(&data->refs);
7167 flush_delayed_work(&ctx->rsrc_put_work);
7169 ret = wait_for_completion_interruptible(&data->done);
7173 percpu_ref_resurrect(&data->refs);
7174 reinit_completion(&data->done);
7176 mutex_unlock(&ctx->uring_lock);
7177 ret = io_run_task_work_sig();
7178 mutex_lock(&ctx->uring_lock);
7180 data->quiesce = false;
7185 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7186 rsrc_put_fn *do_put)
7188 struct io_rsrc_data *data;
7190 data = kzalloc(sizeof(*data), GFP_KERNEL);
7194 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7195 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7200 data->do_put = do_put;
7201 init_completion(&data->done);
7205 static void io_rsrc_data_free(struct io_rsrc_data *data)
7207 percpu_ref_exit(&data->refs);
7211 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7213 struct io_rsrc_data *data = ctx->file_data;
7218 ret = io_rsrc_ref_quiesce(data, ctx);
7222 __io_sqe_files_unregister(ctx);
7223 io_free_file_tables(data, ctx->nr_user_files);
7224 io_rsrc_data_free(data);
7225 ctx->file_data = NULL;
7226 ctx->nr_user_files = 0;
7230 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7231 __releases(&sqd->lock)
7233 WARN_ON_ONCE(sqd->thread == current);
7236 * Do the dance but not conditional clear_bit() because it'd race with
7237 * other threads incrementing park_pending and setting the bit.
7239 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7240 if (atomic_dec_return(&sqd->park_pending))
7241 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7242 mutex_unlock(&sqd->lock);
7245 static void io_sq_thread_park(struct io_sq_data *sqd)
7246 __acquires(&sqd->lock)
7248 WARN_ON_ONCE(sqd->thread == current);
7250 atomic_inc(&sqd->park_pending);
7251 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7252 mutex_lock(&sqd->lock);
7254 wake_up_process(sqd->thread);
7257 static void io_sq_thread_stop(struct io_sq_data *sqd)
7259 WARN_ON_ONCE(sqd->thread == current);
7261 mutex_lock(&sqd->lock);
7262 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7264 wake_up_process(sqd->thread);
7265 mutex_unlock(&sqd->lock);
7266 wait_for_completion(&sqd->exited);
7269 static void io_put_sq_data(struct io_sq_data *sqd)
7271 if (refcount_dec_and_test(&sqd->refs)) {
7272 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7274 io_sq_thread_stop(sqd);
7279 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7281 struct io_sq_data *sqd = ctx->sq_data;
7284 io_sq_thread_park(sqd);
7285 list_del_init(&ctx->sqd_list);
7286 io_sqd_update_thread_idle(sqd);
7287 io_sq_thread_unpark(sqd);
7289 io_put_sq_data(sqd);
7290 ctx->sq_data = NULL;
7292 put_cred(ctx->sq_creds);
7296 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7298 struct io_ring_ctx *ctx_attach;
7299 struct io_sq_data *sqd;
7302 f = fdget(p->wq_fd);
7304 return ERR_PTR(-ENXIO);
7305 if (f.file->f_op != &io_uring_fops) {
7307 return ERR_PTR(-EINVAL);
7310 ctx_attach = f.file->private_data;
7311 sqd = ctx_attach->sq_data;
7314 return ERR_PTR(-EINVAL);
7316 if (sqd->task_tgid != current->tgid) {
7318 return ERR_PTR(-EPERM);
7321 refcount_inc(&sqd->refs);
7326 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7329 struct io_sq_data *sqd;
7332 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7333 sqd = io_attach_sq_data(p);
7338 /* fall through for EPERM case, setup new sqd/task */
7339 if (PTR_ERR(sqd) != -EPERM)
7343 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7345 return ERR_PTR(-ENOMEM);
7347 atomic_set(&sqd->park_pending, 0);
7348 refcount_set(&sqd->refs, 1);
7349 INIT_LIST_HEAD(&sqd->ctx_list);
7350 mutex_init(&sqd->lock);
7351 init_waitqueue_head(&sqd->wait);
7352 init_completion(&sqd->exited);
7356 #if defined(CONFIG_UNIX)
7358 * Ensure the UNIX gc is aware of our file set, so we are certain that
7359 * the io_uring can be safely unregistered on process exit, even if we have
7360 * loops in the file referencing.
7362 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7364 struct sock *sk = ctx->ring_sock->sk;
7365 struct scm_fp_list *fpl;
7366 struct sk_buff *skb;
7369 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7373 skb = alloc_skb(0, GFP_KERNEL);
7382 fpl->user = get_uid(current_user());
7383 for (i = 0; i < nr; i++) {
7384 struct file *file = io_file_from_index(ctx, i + offset);
7388 fpl->fp[nr_files] = get_file(file);
7389 unix_inflight(fpl->user, fpl->fp[nr_files]);
7394 fpl->max = SCM_MAX_FD;
7395 fpl->count = nr_files;
7396 UNIXCB(skb).fp = fpl;
7397 skb->destructor = unix_destruct_scm;
7398 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7399 skb_queue_head(&sk->sk_receive_queue, skb);
7401 for (i = 0; i < nr_files; i++)
7412 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7413 * causes regular reference counting to break down. We rely on the UNIX
7414 * garbage collection to take care of this problem for us.
7416 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7418 unsigned left, total;
7422 left = ctx->nr_user_files;
7424 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7426 ret = __io_sqe_files_scm(ctx, this_files, total);
7430 total += this_files;
7436 while (total < ctx->nr_user_files) {
7437 struct file *file = io_file_from_index(ctx, total);
7447 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7453 static bool io_alloc_file_tables(struct io_rsrc_data *file_data,
7456 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7458 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7460 if (!file_data->table)
7463 for (i = 0; i < nr_tables; i++) {
7464 struct fixed_rsrc_table *table = &file_data->table[i];
7465 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7467 table->files = kcalloc(this_files, sizeof(struct file *),
7471 nr_files -= this_files;
7477 io_free_file_tables(file_data, nr_tables * IORING_MAX_FILES_TABLE);
7481 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7483 struct file *file = prsrc->file;
7484 #if defined(CONFIG_UNIX)
7485 struct sock *sock = ctx->ring_sock->sk;
7486 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7487 struct sk_buff *skb;
7490 __skb_queue_head_init(&list);
7493 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7494 * remove this entry and rearrange the file array.
7496 skb = skb_dequeue(head);
7498 struct scm_fp_list *fp;
7500 fp = UNIXCB(skb).fp;
7501 for (i = 0; i < fp->count; i++) {
7504 if (fp->fp[i] != file)
7507 unix_notinflight(fp->user, fp->fp[i]);
7508 left = fp->count - 1 - i;
7510 memmove(&fp->fp[i], &fp->fp[i + 1],
7511 left * sizeof(struct file *));
7518 __skb_queue_tail(&list, skb);
7528 __skb_queue_tail(&list, skb);
7530 skb = skb_dequeue(head);
7533 if (skb_peek(&list)) {
7534 spin_lock_irq(&head->lock);
7535 while ((skb = __skb_dequeue(&list)) != NULL)
7536 __skb_queue_tail(head, skb);
7537 spin_unlock_irq(&head->lock);
7544 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7546 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7547 struct io_ring_ctx *ctx = rsrc_data->ctx;
7548 struct io_rsrc_put *prsrc, *tmp;
7550 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7551 list_del(&prsrc->list);
7552 rsrc_data->do_put(ctx, prsrc);
7556 io_rsrc_node_destroy(ref_node);
7557 percpu_ref_put(&rsrc_data->refs);
7560 static void io_rsrc_put_work(struct work_struct *work)
7562 struct io_ring_ctx *ctx;
7563 struct llist_node *node;
7565 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7566 node = llist_del_all(&ctx->rsrc_put_llist);
7569 struct io_rsrc_node *ref_node;
7570 struct llist_node *next = node->next;
7572 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7573 __io_rsrc_put_work(ref_node);
7578 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7580 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7581 struct io_rsrc_data *data = node->rsrc_data;
7582 struct io_ring_ctx *ctx = data->ctx;
7583 bool first_add = false;
7586 io_rsrc_ref_lock(ctx);
7589 while (!list_empty(&ctx->rsrc_ref_list)) {
7590 node = list_first_entry(&ctx->rsrc_ref_list,
7591 struct io_rsrc_node, node);
7592 /* recycle ref nodes in order */
7595 list_del(&node->node);
7596 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7598 io_rsrc_ref_unlock(ctx);
7600 delay = percpu_ref_is_dying(&data->refs) ? 0 : HZ;
7601 if (first_add || !delay)
7602 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7605 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7607 struct io_rsrc_node *ref_node;
7609 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7613 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7618 INIT_LIST_HEAD(&ref_node->node);
7619 INIT_LIST_HEAD(&ref_node->rsrc_list);
7620 ref_node->done = false;
7624 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7627 __s32 __user *fds = (__s32 __user *) arg;
7631 struct io_rsrc_data *file_data;
7637 if (nr_args > IORING_MAX_FIXED_FILES)
7639 ret = io_rsrc_node_switch_start(ctx);
7643 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put);
7646 ctx->file_data = file_data;
7649 if (!io_alloc_file_tables(file_data, nr_args))
7652 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7653 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7657 /* allow sparse sets */
7667 * Don't allow io_uring instances to be registered. If UNIX
7668 * isn't enabled, then this causes a reference cycle and this
7669 * instance can never get freed. If UNIX is enabled we'll
7670 * handle it just fine, but there's still no point in allowing
7671 * a ring fd as it doesn't support regular read/write anyway.
7673 if (file->f_op == &io_uring_fops) {
7677 io_fixed_file_set(io_fixed_file_slot(file_data, i), file);
7680 ret = io_sqe_files_scm(ctx);
7682 io_sqe_files_unregister(ctx);
7686 io_rsrc_node_switch(ctx, NULL);
7689 for (i = 0; i < ctx->nr_user_files; i++) {
7690 file = io_file_from_index(ctx, i);
7694 io_free_file_tables(file_data, nr_args);
7695 ctx->nr_user_files = 0;
7697 io_rsrc_data_free(ctx->file_data);
7698 ctx->file_data = NULL;
7702 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7705 #if defined(CONFIG_UNIX)
7706 struct sock *sock = ctx->ring_sock->sk;
7707 struct sk_buff_head *head = &sock->sk_receive_queue;
7708 struct sk_buff *skb;
7711 * See if we can merge this file into an existing skb SCM_RIGHTS
7712 * file set. If there's no room, fall back to allocating a new skb
7713 * and filling it in.
7715 spin_lock_irq(&head->lock);
7716 skb = skb_peek(head);
7718 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7720 if (fpl->count < SCM_MAX_FD) {
7721 __skb_unlink(skb, head);
7722 spin_unlock_irq(&head->lock);
7723 fpl->fp[fpl->count] = get_file(file);
7724 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7726 spin_lock_irq(&head->lock);
7727 __skb_queue_head(head, skb);
7732 spin_unlock_irq(&head->lock);
7739 return __io_sqe_files_scm(ctx, 1, index);
7745 static int io_queue_rsrc_removal(struct io_rsrc_data *data,
7746 struct io_rsrc_node *node, void *rsrc)
7748 struct io_rsrc_put *prsrc;
7750 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7755 list_add(&prsrc->list, &node->rsrc_list);
7759 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7760 struct io_uring_rsrc_update *up,
7763 struct io_rsrc_data *data = ctx->file_data;
7764 struct io_fixed_file *file_slot;
7769 bool needs_switch = false;
7771 if (check_add_overflow(up->offset, nr_args, &done))
7773 if (done > ctx->nr_user_files)
7775 err = io_rsrc_node_switch_start(ctx);
7779 fds = u64_to_user_ptr(up->data);
7780 for (done = 0; done < nr_args; done++) {
7782 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7786 if (fd == IORING_REGISTER_FILES_SKIP)
7789 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7790 file_slot = io_fixed_file_slot(ctx->file_data, i);
7792 if (file_slot->file_ptr) {
7793 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7794 err = io_queue_rsrc_removal(data, ctx->rsrc_node, file);
7797 file_slot->file_ptr = 0;
7798 needs_switch = true;
7807 * Don't allow io_uring instances to be registered. If
7808 * UNIX isn't enabled, then this causes a reference
7809 * cycle and this instance can never get freed. If UNIX
7810 * is enabled we'll handle it just fine, but there's
7811 * still no point in allowing a ring fd as it doesn't
7812 * support regular read/write anyway.
7814 if (file->f_op == &io_uring_fops) {
7819 io_fixed_file_set(file_slot, file);
7820 err = io_sqe_file_register(ctx, file, i);
7822 file_slot->file_ptr = 0;
7830 io_rsrc_node_switch(ctx, data);
7831 return done ? done : err;
7834 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7837 struct io_uring_rsrc_update up;
7839 if (!ctx->file_data)
7843 if (copy_from_user(&up, arg, sizeof(up)))
7848 return __io_sqe_files_update(ctx, &up, nr_args);
7851 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7853 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7855 req = io_put_req_find_next(req);
7856 return req ? &req->work : NULL;
7859 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7860 struct task_struct *task)
7862 struct io_wq_hash *hash;
7863 struct io_wq_data data;
7864 unsigned int concurrency;
7866 hash = ctx->hash_map;
7868 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7870 return ERR_PTR(-ENOMEM);
7871 refcount_set(&hash->refs, 1);
7872 init_waitqueue_head(&hash->wait);
7873 ctx->hash_map = hash;
7878 data.free_work = io_free_work;
7879 data.do_work = io_wq_submit_work;
7881 /* Do QD, or 4 * CPUS, whatever is smallest */
7882 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7884 return io_wq_create(concurrency, &data);
7887 static int io_uring_alloc_task_context(struct task_struct *task,
7888 struct io_ring_ctx *ctx)
7890 struct io_uring_task *tctx;
7893 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7894 if (unlikely(!tctx))
7897 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7898 if (unlikely(ret)) {
7903 tctx->io_wq = io_init_wq_offload(ctx, task);
7904 if (IS_ERR(tctx->io_wq)) {
7905 ret = PTR_ERR(tctx->io_wq);
7906 percpu_counter_destroy(&tctx->inflight);
7912 init_waitqueue_head(&tctx->wait);
7914 atomic_set(&tctx->in_idle, 0);
7915 task->io_uring = tctx;
7916 spin_lock_init(&tctx->task_lock);
7917 INIT_WQ_LIST(&tctx->task_list);
7918 tctx->task_state = 0;
7919 init_task_work(&tctx->task_work, tctx_task_work);
7923 void __io_uring_free(struct task_struct *tsk)
7925 struct io_uring_task *tctx = tsk->io_uring;
7927 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7928 WARN_ON_ONCE(tctx->io_wq);
7930 percpu_counter_destroy(&tctx->inflight);
7932 tsk->io_uring = NULL;
7935 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7936 struct io_uring_params *p)
7940 /* Retain compatibility with failing for an invalid attach attempt */
7941 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7942 IORING_SETUP_ATTACH_WQ) {
7945 f = fdget(p->wq_fd);
7948 if (f.file->f_op != &io_uring_fops) {
7954 if (ctx->flags & IORING_SETUP_SQPOLL) {
7955 struct task_struct *tsk;
7956 struct io_sq_data *sqd;
7959 sqd = io_get_sq_data(p, &attached);
7965 ctx->sq_creds = get_current_cred();
7967 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7968 if (!ctx->sq_thread_idle)
7969 ctx->sq_thread_idle = HZ;
7972 io_sq_thread_park(sqd);
7973 list_add(&ctx->sqd_list, &sqd->ctx_list);
7974 io_sqd_update_thread_idle(sqd);
7975 /* don't attach to a dying SQPOLL thread, would be racy */
7976 if (attached && !sqd->thread)
7978 io_sq_thread_unpark(sqd);
7985 if (p->flags & IORING_SETUP_SQ_AFF) {
7986 int cpu = p->sq_thread_cpu;
7989 if (cpu >= nr_cpu_ids)
7991 if (!cpu_online(cpu))
7999 sqd->task_pid = current->pid;
8000 sqd->task_tgid = current->tgid;
8001 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8008 ret = io_uring_alloc_task_context(tsk, ctx);
8009 wake_up_new_task(tsk);
8012 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8013 /* Can't have SQ_AFF without SQPOLL */
8020 io_sq_thread_finish(ctx);
8023 complete(&ctx->sq_data->exited);
8027 static inline void __io_unaccount_mem(struct user_struct *user,
8028 unsigned long nr_pages)
8030 atomic_long_sub(nr_pages, &user->locked_vm);
8033 static inline int __io_account_mem(struct user_struct *user,
8034 unsigned long nr_pages)
8036 unsigned long page_limit, cur_pages, new_pages;
8038 /* Don't allow more pages than we can safely lock */
8039 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8042 cur_pages = atomic_long_read(&user->locked_vm);
8043 new_pages = cur_pages + nr_pages;
8044 if (new_pages > page_limit)
8046 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8047 new_pages) != cur_pages);
8052 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8055 __io_unaccount_mem(ctx->user, nr_pages);
8057 if (ctx->mm_account)
8058 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8061 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8066 ret = __io_account_mem(ctx->user, nr_pages);
8071 if (ctx->mm_account)
8072 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8077 static void io_mem_free(void *ptr)
8084 page = virt_to_head_page(ptr);
8085 if (put_page_testzero(page))
8086 free_compound_page(page);
8089 static void *io_mem_alloc(size_t size)
8091 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8092 __GFP_NORETRY | __GFP_ACCOUNT;
8094 return (void *) __get_free_pages(gfp_flags, get_order(size));
8097 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8100 struct io_rings *rings;
8101 size_t off, sq_array_size;
8103 off = struct_size(rings, cqes, cq_entries);
8104 if (off == SIZE_MAX)
8108 off = ALIGN(off, SMP_CACHE_BYTES);
8116 sq_array_size = array_size(sizeof(u32), sq_entries);
8117 if (sq_array_size == SIZE_MAX)
8120 if (check_add_overflow(off, sq_array_size, &off))
8126 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8130 if (!ctx->user_bufs)
8133 for (i = 0; i < ctx->nr_user_bufs; i++) {
8134 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8136 for (j = 0; j < imu->nr_bvecs; j++)
8137 unpin_user_page(imu->bvec[j].bv_page);
8139 if (imu->acct_pages)
8140 io_unaccount_mem(ctx, imu->acct_pages);
8145 kfree(ctx->user_bufs);
8146 ctx->user_bufs = NULL;
8147 ctx->nr_user_bufs = 0;
8151 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8152 void __user *arg, unsigned index)
8154 struct iovec __user *src;
8156 #ifdef CONFIG_COMPAT
8158 struct compat_iovec __user *ciovs;
8159 struct compat_iovec ciov;
8161 ciovs = (struct compat_iovec __user *) arg;
8162 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8165 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8166 dst->iov_len = ciov.iov_len;
8170 src = (struct iovec __user *) arg;
8171 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8177 * Not super efficient, but this is just a registration time. And we do cache
8178 * the last compound head, so generally we'll only do a full search if we don't
8181 * We check if the given compound head page has already been accounted, to
8182 * avoid double accounting it. This allows us to account the full size of the
8183 * page, not just the constituent pages of a huge page.
8185 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8186 int nr_pages, struct page *hpage)
8190 /* check current page array */
8191 for (i = 0; i < nr_pages; i++) {
8192 if (!PageCompound(pages[i]))
8194 if (compound_head(pages[i]) == hpage)
8198 /* check previously registered pages */
8199 for (i = 0; i < ctx->nr_user_bufs; i++) {
8200 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8202 for (j = 0; j < imu->nr_bvecs; j++) {
8203 if (!PageCompound(imu->bvec[j].bv_page))
8205 if (compound_head(imu->bvec[j].bv_page) == hpage)
8213 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8214 int nr_pages, struct io_mapped_ubuf *imu,
8215 struct page **last_hpage)
8219 for (i = 0; i < nr_pages; i++) {
8220 if (!PageCompound(pages[i])) {
8225 hpage = compound_head(pages[i]);
8226 if (hpage == *last_hpage)
8228 *last_hpage = hpage;
8229 if (headpage_already_acct(ctx, pages, i, hpage))
8231 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8235 if (!imu->acct_pages)
8238 ret = io_account_mem(ctx, imu->acct_pages);
8240 imu->acct_pages = 0;
8244 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8245 struct io_mapped_ubuf *imu,
8246 struct page **last_hpage)
8248 struct vm_area_struct **vmas = NULL;
8249 struct page **pages = NULL;
8250 unsigned long off, start, end, ubuf;
8252 int ret, pret, nr_pages, i;
8254 ubuf = (unsigned long) iov->iov_base;
8255 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8256 start = ubuf >> PAGE_SHIFT;
8257 nr_pages = end - start;
8261 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8265 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8270 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8276 mmap_read_lock(current->mm);
8277 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8279 if (pret == nr_pages) {
8280 /* don't support file backed memory */
8281 for (i = 0; i < nr_pages; i++) {
8282 struct vm_area_struct *vma = vmas[i];
8285 !is_file_hugepages(vma->vm_file)) {
8291 ret = pret < 0 ? pret : -EFAULT;
8293 mmap_read_unlock(current->mm);
8296 * if we did partial map, or found file backed vmas,
8297 * release any pages we did get
8300 unpin_user_pages(pages, pret);
8305 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8307 unpin_user_pages(pages, pret);
8312 off = ubuf & ~PAGE_MASK;
8313 size = iov->iov_len;
8314 for (i = 0; i < nr_pages; i++) {
8317 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8318 imu->bvec[i].bv_page = pages[i];
8319 imu->bvec[i].bv_len = vec_len;
8320 imu->bvec[i].bv_offset = off;
8324 /* store original address for later verification */
8326 imu->ubuf_end = ubuf + iov->iov_len;
8327 imu->nr_bvecs = nr_pages;
8335 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8339 if (!nr_args || nr_args > UIO_MAXIOV)
8342 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8344 if (!ctx->user_bufs)
8350 static int io_buffer_validate(struct iovec *iov)
8352 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8355 * Don't impose further limits on the size and buffer
8356 * constraints here, we'll -EINVAL later when IO is
8357 * submitted if they are wrong.
8359 if (!iov->iov_base || !iov->iov_len)
8362 /* arbitrary limit, but we need something */
8363 if (iov->iov_len > SZ_1G)
8366 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8372 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8373 unsigned int nr_args)
8377 struct page *last_hpage = NULL;
8379 ret = io_buffers_map_alloc(ctx, nr_args);
8383 for (i = 0; i < nr_args; i++) {
8384 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8386 ret = io_copy_iov(ctx, &iov, arg, i);
8390 ret = io_buffer_validate(&iov);
8394 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8398 ctx->nr_user_bufs++;
8402 io_sqe_buffers_unregister(ctx);
8407 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8409 __s32 __user *fds = arg;
8415 if (copy_from_user(&fd, fds, sizeof(*fds)))
8418 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8419 if (IS_ERR(ctx->cq_ev_fd)) {
8420 int ret = PTR_ERR(ctx->cq_ev_fd);
8421 ctx->cq_ev_fd = NULL;
8428 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8430 if (ctx->cq_ev_fd) {
8431 eventfd_ctx_put(ctx->cq_ev_fd);
8432 ctx->cq_ev_fd = NULL;
8439 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8441 struct io_buffer *buf;
8442 unsigned long index;
8444 xa_for_each(&ctx->io_buffers, index, buf)
8445 __io_remove_buffers(ctx, buf, index, -1U);
8448 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8450 struct io_kiocb *req, *nxt;
8452 list_for_each_entry_safe(req, nxt, list, compl.list) {
8453 if (tsk && req->task != tsk)
8455 list_del(&req->compl.list);
8456 kmem_cache_free(req_cachep, req);
8460 static void io_req_caches_free(struct io_ring_ctx *ctx)
8462 struct io_submit_state *submit_state = &ctx->submit_state;
8463 struct io_comp_state *cs = &ctx->submit_state.comp;
8465 mutex_lock(&ctx->uring_lock);
8467 if (submit_state->free_reqs) {
8468 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8469 submit_state->reqs);
8470 submit_state->free_reqs = 0;
8473 io_flush_cached_locked_reqs(ctx, cs);
8474 io_req_cache_free(&cs->free_list, NULL);
8475 mutex_unlock(&ctx->uring_lock);
8478 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8480 io_sq_thread_finish(ctx);
8481 io_sqe_buffers_unregister(ctx);
8483 if (ctx->mm_account) {
8484 mmdrop(ctx->mm_account);
8485 ctx->mm_account = NULL;
8488 mutex_lock(&ctx->uring_lock);
8489 io_sqe_files_unregister(ctx);
8491 __io_cqring_overflow_flush(ctx, true);
8492 mutex_unlock(&ctx->uring_lock);
8493 io_eventfd_unregister(ctx);
8494 io_destroy_buffers(ctx);
8496 /* there are no registered resources left, nobody uses it */
8498 io_rsrc_node_destroy(ctx->rsrc_node);
8499 if (ctx->rsrc_backup_node)
8500 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8501 flush_delayed_work(&ctx->rsrc_put_work);
8503 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8504 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8506 #if defined(CONFIG_UNIX)
8507 if (ctx->ring_sock) {
8508 ctx->ring_sock->file = NULL; /* so that iput() is called */
8509 sock_release(ctx->ring_sock);
8513 io_mem_free(ctx->rings);
8514 io_mem_free(ctx->sq_sqes);
8516 percpu_ref_exit(&ctx->refs);
8517 free_uid(ctx->user);
8518 io_req_caches_free(ctx);
8520 io_wq_put_hash(ctx->hash_map);
8521 kfree(ctx->cancel_hash);
8525 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8527 struct io_ring_ctx *ctx = file->private_data;
8530 poll_wait(file, &ctx->cq_wait, wait);
8532 * synchronizes with barrier from wq_has_sleeper call in
8536 if (!io_sqring_full(ctx))
8537 mask |= EPOLLOUT | EPOLLWRNORM;
8540 * Don't flush cqring overflow list here, just do a simple check.
8541 * Otherwise there could possible be ABBA deadlock:
8544 * lock(&ctx->uring_lock);
8546 * lock(&ctx->uring_lock);
8549 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8550 * pushs them to do the flush.
8552 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8553 mask |= EPOLLIN | EPOLLRDNORM;
8558 static int io_uring_fasync(int fd, struct file *file, int on)
8560 struct io_ring_ctx *ctx = file->private_data;
8562 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8565 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8567 const struct cred *creds;
8569 creds = xa_erase(&ctx->personalities, id);
8578 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8580 return io_run_task_work_head(&ctx->exit_task_work);
8583 struct io_tctx_exit {
8584 struct callback_head task_work;
8585 struct completion completion;
8586 struct io_ring_ctx *ctx;
8589 static void io_tctx_exit_cb(struct callback_head *cb)
8591 struct io_uring_task *tctx = current->io_uring;
8592 struct io_tctx_exit *work;
8594 work = container_of(cb, struct io_tctx_exit, task_work);
8596 * When @in_idle, we're in cancellation and it's racy to remove the
8597 * node. It'll be removed by the end of cancellation, just ignore it.
8599 if (!atomic_read(&tctx->in_idle))
8600 io_uring_del_task_file((unsigned long)work->ctx);
8601 complete(&work->completion);
8604 static void io_ring_exit_work(struct work_struct *work)
8606 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8607 unsigned long timeout = jiffies + HZ * 60 * 5;
8608 struct io_tctx_exit exit;
8609 struct io_tctx_node *node;
8612 /* prevent SQPOLL from submitting new requests */
8614 io_sq_thread_park(ctx->sq_data);
8615 list_del_init(&ctx->sqd_list);
8616 io_sqd_update_thread_idle(ctx->sq_data);
8617 io_sq_thread_unpark(ctx->sq_data);
8621 * If we're doing polled IO and end up having requests being
8622 * submitted async (out-of-line), then completions can come in while
8623 * we're waiting for refs to drop. We need to reap these manually,
8624 * as nobody else will be looking for them.
8627 io_uring_try_cancel_requests(ctx, NULL, NULL);
8629 WARN_ON_ONCE(time_after(jiffies, timeout));
8630 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8633 * Some may use context even when all refs and requests have been put,
8634 * and they are free to do so while still holding uring_lock or
8635 * completion_lock, see __io_req_task_submit(). Apart from other work,
8636 * this lock/unlock section also waits them to finish.
8638 mutex_lock(&ctx->uring_lock);
8639 while (!list_empty(&ctx->tctx_list)) {
8640 WARN_ON_ONCE(time_after(jiffies, timeout));
8642 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8645 init_completion(&exit.completion);
8646 init_task_work(&exit.task_work, io_tctx_exit_cb);
8647 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8648 if (WARN_ON_ONCE(ret))
8650 wake_up_process(node->task);
8652 mutex_unlock(&ctx->uring_lock);
8653 wait_for_completion(&exit.completion);
8655 mutex_lock(&ctx->uring_lock);
8657 mutex_unlock(&ctx->uring_lock);
8658 spin_lock_irq(&ctx->completion_lock);
8659 spin_unlock_irq(&ctx->completion_lock);
8661 io_ring_ctx_free(ctx);
8664 /* Returns true if we found and killed one or more timeouts */
8665 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8666 struct files_struct *files)
8668 struct io_kiocb *req, *tmp;
8671 spin_lock_irq(&ctx->completion_lock);
8672 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8673 if (io_match_task(req, tsk, files)) {
8674 io_kill_timeout(req, -ECANCELED);
8679 io_commit_cqring(ctx);
8680 spin_unlock_irq(&ctx->completion_lock);
8682 io_cqring_ev_posted(ctx);
8683 return canceled != 0;
8686 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8688 unsigned long index;
8689 struct creds *creds;
8691 mutex_lock(&ctx->uring_lock);
8692 percpu_ref_kill(&ctx->refs);
8694 __io_cqring_overflow_flush(ctx, true);
8695 xa_for_each(&ctx->personalities, index, creds)
8696 io_unregister_personality(ctx, index);
8697 mutex_unlock(&ctx->uring_lock);
8699 io_kill_timeouts(ctx, NULL, NULL);
8700 io_poll_remove_all(ctx, NULL, NULL);
8702 /* if we failed setting up the ctx, we might not have any rings */
8703 io_iopoll_try_reap_events(ctx);
8705 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8707 * Use system_unbound_wq to avoid spawning tons of event kworkers
8708 * if we're exiting a ton of rings at the same time. It just adds
8709 * noise and overhead, there's no discernable change in runtime
8710 * over using system_wq.
8712 queue_work(system_unbound_wq, &ctx->exit_work);
8715 static int io_uring_release(struct inode *inode, struct file *file)
8717 struct io_ring_ctx *ctx = file->private_data;
8719 file->private_data = NULL;
8720 io_ring_ctx_wait_and_kill(ctx);
8724 struct io_task_cancel {
8725 struct task_struct *task;
8726 struct files_struct *files;
8729 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8731 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8732 struct io_task_cancel *cancel = data;
8735 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8736 unsigned long flags;
8737 struct io_ring_ctx *ctx = req->ctx;
8739 /* protect against races with linked timeouts */
8740 spin_lock_irqsave(&ctx->completion_lock, flags);
8741 ret = io_match_task(req, cancel->task, cancel->files);
8742 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8744 ret = io_match_task(req, cancel->task, cancel->files);
8749 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8750 struct task_struct *task,
8751 struct files_struct *files)
8753 struct io_defer_entry *de;
8756 spin_lock_irq(&ctx->completion_lock);
8757 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8758 if (io_match_task(de->req, task, files)) {
8759 list_cut_position(&list, &ctx->defer_list, &de->list);
8763 spin_unlock_irq(&ctx->completion_lock);
8764 if (list_empty(&list))
8767 while (!list_empty(&list)) {
8768 de = list_first_entry(&list, struct io_defer_entry, list);
8769 list_del_init(&de->list);
8770 io_req_complete_failed(de->req, -ECANCELED);
8776 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8778 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8780 return req->ctx == data;
8783 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8785 struct io_tctx_node *node;
8786 enum io_wq_cancel cret;
8789 mutex_lock(&ctx->uring_lock);
8790 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8791 struct io_uring_task *tctx = node->task->io_uring;
8794 * io_wq will stay alive while we hold uring_lock, because it's
8795 * killed after ctx nodes, which requires to take the lock.
8797 if (!tctx || !tctx->io_wq)
8799 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8800 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8802 mutex_unlock(&ctx->uring_lock);
8807 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8808 struct task_struct *task,
8809 struct files_struct *files)
8811 struct io_task_cancel cancel = { .task = task, .files = files, };
8812 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8815 enum io_wq_cancel cret;
8819 ret |= io_uring_try_cancel_iowq(ctx);
8820 } else if (tctx && tctx->io_wq) {
8822 * Cancels requests of all rings, not only @ctx, but
8823 * it's fine as the task is in exit/exec.
8825 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8827 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8830 /* SQPOLL thread does its own polling */
8831 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8832 (ctx->sq_data && ctx->sq_data->thread == current)) {
8833 while (!list_empty_careful(&ctx->iopoll_list)) {
8834 io_iopoll_try_reap_events(ctx);
8839 ret |= io_cancel_defer_files(ctx, task, files);
8840 ret |= io_poll_remove_all(ctx, task, files);
8841 ret |= io_kill_timeouts(ctx, task, files);
8842 ret |= io_run_task_work();
8843 ret |= io_run_ctx_fallback(ctx);
8850 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8851 struct task_struct *task,
8852 struct files_struct *files)
8854 struct io_kiocb *req;
8857 spin_lock_irq(&ctx->inflight_lock);
8858 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8859 cnt += io_match_task(req, task, files);
8860 spin_unlock_irq(&ctx->inflight_lock);
8864 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8865 struct task_struct *task,
8866 struct files_struct *files)
8868 while (!list_empty_careful(&ctx->inflight_list)) {
8872 inflight = io_uring_count_inflight(ctx, task, files);
8876 io_uring_try_cancel_requests(ctx, task, files);
8878 prepare_to_wait(&task->io_uring->wait, &wait,
8879 TASK_UNINTERRUPTIBLE);
8880 if (inflight == io_uring_count_inflight(ctx, task, files))
8882 finish_wait(&task->io_uring->wait, &wait);
8886 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8888 struct io_uring_task *tctx = current->io_uring;
8889 struct io_tctx_node *node;
8892 if (unlikely(!tctx)) {
8893 ret = io_uring_alloc_task_context(current, ctx);
8896 tctx = current->io_uring;
8898 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8899 node = kmalloc(sizeof(*node), GFP_KERNEL);
8903 node->task = current;
8905 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8912 mutex_lock(&ctx->uring_lock);
8913 list_add(&node->ctx_node, &ctx->tctx_list);
8914 mutex_unlock(&ctx->uring_lock);
8921 * Note that this task has used io_uring. We use it for cancelation purposes.
8923 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8925 struct io_uring_task *tctx = current->io_uring;
8927 if (likely(tctx && tctx->last == ctx))
8929 return __io_uring_add_task_file(ctx);
8933 * Remove this io_uring_file -> task mapping.
8935 static void io_uring_del_task_file(unsigned long index)
8937 struct io_uring_task *tctx = current->io_uring;
8938 struct io_tctx_node *node;
8942 node = xa_erase(&tctx->xa, index);
8946 WARN_ON_ONCE(current != node->task);
8947 WARN_ON_ONCE(list_empty(&node->ctx_node));
8949 mutex_lock(&node->ctx->uring_lock);
8950 list_del(&node->ctx_node);
8951 mutex_unlock(&node->ctx->uring_lock);
8953 if (tctx->last == node->ctx)
8958 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8960 struct io_tctx_node *node;
8961 unsigned long index;
8963 xa_for_each(&tctx->xa, index, node)
8964 io_uring_del_task_file(index);
8966 io_wq_put_and_exit(tctx->io_wq);
8971 static s64 tctx_inflight(struct io_uring_task *tctx)
8973 return percpu_counter_sum(&tctx->inflight);
8976 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8978 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8979 struct io_ring_ctx *ctx = work->ctx;
8980 struct io_sq_data *sqd = ctx->sq_data;
8983 io_uring_cancel_sqpoll(ctx);
8984 complete(&work->completion);
8987 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8989 struct io_sq_data *sqd = ctx->sq_data;
8990 struct io_tctx_exit work = { .ctx = ctx, };
8991 struct task_struct *task;
8993 io_sq_thread_park(sqd);
8994 list_del_init(&ctx->sqd_list);
8995 io_sqd_update_thread_idle(sqd);
8998 init_completion(&work.completion);
8999 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
9000 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
9001 wake_up_process(task);
9003 io_sq_thread_unpark(sqd);
9006 wait_for_completion(&work.completion);
9009 void __io_uring_files_cancel(struct files_struct *files)
9011 struct io_uring_task *tctx = current->io_uring;
9012 struct io_tctx_node *node;
9013 unsigned long index;
9015 /* make sure overflow events are dropped */
9016 atomic_inc(&tctx->in_idle);
9017 xa_for_each(&tctx->xa, index, node) {
9018 struct io_ring_ctx *ctx = node->ctx;
9021 io_sqpoll_cancel_sync(ctx);
9024 io_uring_cancel_files(ctx, current, files);
9026 io_uring_try_cancel_requests(ctx, current, NULL);
9028 atomic_dec(&tctx->in_idle);
9031 io_uring_clean_tctx(tctx);
9034 /* should only be called by SQPOLL task */
9035 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9037 struct io_sq_data *sqd = ctx->sq_data;
9038 struct io_uring_task *tctx = current->io_uring;
9042 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
9044 atomic_inc(&tctx->in_idle);
9046 /* read completions before cancelations */
9047 inflight = tctx_inflight(tctx);
9050 io_uring_try_cancel_requests(ctx, current, NULL);
9052 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9054 * If we've seen completions, retry without waiting. This
9055 * avoids a race where a completion comes in before we did
9056 * prepare_to_wait().
9058 if (inflight == tctx_inflight(tctx))
9060 finish_wait(&tctx->wait, &wait);
9062 atomic_dec(&tctx->in_idle);
9066 * Find any io_uring fd that this task has registered or done IO on, and cancel
9069 void __io_uring_task_cancel(void)
9071 struct io_uring_task *tctx = current->io_uring;
9075 /* make sure overflow events are dropped */
9076 atomic_inc(&tctx->in_idle);
9077 __io_uring_files_cancel(NULL);
9080 /* read completions before cancelations */
9081 inflight = tctx_inflight(tctx);
9084 __io_uring_files_cancel(NULL);
9086 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9089 * If we've seen completions, retry without waiting. This
9090 * avoids a race where a completion comes in before we did
9091 * prepare_to_wait().
9093 if (inflight == tctx_inflight(tctx))
9095 finish_wait(&tctx->wait, &wait);
9098 atomic_dec(&tctx->in_idle);
9100 io_uring_clean_tctx(tctx);
9101 /* all current's requests should be gone, we can kill tctx */
9102 __io_uring_free(current);
9105 static void *io_uring_validate_mmap_request(struct file *file,
9106 loff_t pgoff, size_t sz)
9108 struct io_ring_ctx *ctx = file->private_data;
9109 loff_t offset = pgoff << PAGE_SHIFT;
9114 case IORING_OFF_SQ_RING:
9115 case IORING_OFF_CQ_RING:
9118 case IORING_OFF_SQES:
9122 return ERR_PTR(-EINVAL);
9125 page = virt_to_head_page(ptr);
9126 if (sz > page_size(page))
9127 return ERR_PTR(-EINVAL);
9134 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9136 size_t sz = vma->vm_end - vma->vm_start;
9140 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9142 return PTR_ERR(ptr);
9144 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9145 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9148 #else /* !CONFIG_MMU */
9150 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9152 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9155 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9157 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9160 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9161 unsigned long addr, unsigned long len,
9162 unsigned long pgoff, unsigned long flags)
9166 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9168 return PTR_ERR(ptr);
9170 return (unsigned long) ptr;
9173 #endif /* !CONFIG_MMU */
9175 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9180 if (!io_sqring_full(ctx))
9182 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9184 if (!io_sqring_full(ctx))
9187 } while (!signal_pending(current));
9189 finish_wait(&ctx->sqo_sq_wait, &wait);
9193 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9194 struct __kernel_timespec __user **ts,
9195 const sigset_t __user **sig)
9197 struct io_uring_getevents_arg arg;
9200 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9201 * is just a pointer to the sigset_t.
9203 if (!(flags & IORING_ENTER_EXT_ARG)) {
9204 *sig = (const sigset_t __user *) argp;
9210 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9211 * timespec and sigset_t pointers if good.
9213 if (*argsz != sizeof(arg))
9215 if (copy_from_user(&arg, argp, sizeof(arg)))
9217 *sig = u64_to_user_ptr(arg.sigmask);
9218 *argsz = arg.sigmask_sz;
9219 *ts = u64_to_user_ptr(arg.ts);
9223 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9224 u32, min_complete, u32, flags, const void __user *, argp,
9227 struct io_ring_ctx *ctx;
9234 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9235 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9239 if (unlikely(!f.file))
9243 if (unlikely(f.file->f_op != &io_uring_fops))
9247 ctx = f.file->private_data;
9248 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9252 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9256 * For SQ polling, the thread will do all submissions and completions.
9257 * Just return the requested submit count, and wake the thread if
9261 if (ctx->flags & IORING_SETUP_SQPOLL) {
9262 io_cqring_overflow_flush(ctx, false);
9265 if (unlikely(ctx->sq_data->thread == NULL)) {
9268 if (flags & IORING_ENTER_SQ_WAKEUP)
9269 wake_up(&ctx->sq_data->wait);
9270 if (flags & IORING_ENTER_SQ_WAIT) {
9271 ret = io_sqpoll_wait_sq(ctx);
9275 submitted = to_submit;
9276 } else if (to_submit) {
9277 ret = io_uring_add_task_file(ctx);
9280 mutex_lock(&ctx->uring_lock);
9281 submitted = io_submit_sqes(ctx, to_submit);
9282 mutex_unlock(&ctx->uring_lock);
9284 if (submitted != to_submit)
9287 if (flags & IORING_ENTER_GETEVENTS) {
9288 const sigset_t __user *sig;
9289 struct __kernel_timespec __user *ts;
9291 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9295 min_complete = min(min_complete, ctx->cq_entries);
9298 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9299 * space applications don't need to do io completion events
9300 * polling again, they can rely on io_sq_thread to do polling
9301 * work, which can reduce cpu usage and uring_lock contention.
9303 if (ctx->flags & IORING_SETUP_IOPOLL &&
9304 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9305 ret = io_iopoll_check(ctx, min_complete);
9307 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9312 percpu_ref_put(&ctx->refs);
9315 return submitted ? submitted : ret;
9318 #ifdef CONFIG_PROC_FS
9319 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9320 const struct cred *cred)
9322 struct user_namespace *uns = seq_user_ns(m);
9323 struct group_info *gi;
9328 seq_printf(m, "%5d\n", id);
9329 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9330 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9331 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9332 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9333 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9334 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9335 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9336 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9337 seq_puts(m, "\n\tGroups:\t");
9338 gi = cred->group_info;
9339 for (g = 0; g < gi->ngroups; g++) {
9340 seq_put_decimal_ull(m, g ? " " : "",
9341 from_kgid_munged(uns, gi->gid[g]));
9343 seq_puts(m, "\n\tCapEff:\t");
9344 cap = cred->cap_effective;
9345 CAP_FOR_EACH_U32(__capi)
9346 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9351 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9353 struct io_sq_data *sq = NULL;
9358 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9359 * since fdinfo case grabs it in the opposite direction of normal use
9360 * cases. If we fail to get the lock, we just don't iterate any
9361 * structures that could be going away outside the io_uring mutex.
9363 has_lock = mutex_trylock(&ctx->uring_lock);
9365 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9371 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9372 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9373 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9374 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9375 struct file *f = io_file_from_index(ctx, i);
9378 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9380 seq_printf(m, "%5u: <none>\n", i);
9382 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9383 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9384 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9385 unsigned int len = buf->ubuf_end - buf->ubuf;
9387 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9389 if (has_lock && !xa_empty(&ctx->personalities)) {
9390 unsigned long index;
9391 const struct cred *cred;
9393 seq_printf(m, "Personalities:\n");
9394 xa_for_each(&ctx->personalities, index, cred)
9395 io_uring_show_cred(m, index, cred);
9397 seq_printf(m, "PollList:\n");
9398 spin_lock_irq(&ctx->completion_lock);
9399 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9400 struct hlist_head *list = &ctx->cancel_hash[i];
9401 struct io_kiocb *req;
9403 hlist_for_each_entry(req, list, hash_node)
9404 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9405 req->task->task_works != NULL);
9407 spin_unlock_irq(&ctx->completion_lock);
9409 mutex_unlock(&ctx->uring_lock);
9412 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9414 struct io_ring_ctx *ctx = f->private_data;
9416 if (percpu_ref_tryget(&ctx->refs)) {
9417 __io_uring_show_fdinfo(ctx, m);
9418 percpu_ref_put(&ctx->refs);
9423 static const struct file_operations io_uring_fops = {
9424 .release = io_uring_release,
9425 .mmap = io_uring_mmap,
9427 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9428 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9430 .poll = io_uring_poll,
9431 .fasync = io_uring_fasync,
9432 #ifdef CONFIG_PROC_FS
9433 .show_fdinfo = io_uring_show_fdinfo,
9437 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9438 struct io_uring_params *p)
9440 struct io_rings *rings;
9441 size_t size, sq_array_offset;
9443 /* make sure these are sane, as we already accounted them */
9444 ctx->sq_entries = p->sq_entries;
9445 ctx->cq_entries = p->cq_entries;
9447 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9448 if (size == SIZE_MAX)
9451 rings = io_mem_alloc(size);
9456 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9457 rings->sq_ring_mask = p->sq_entries - 1;
9458 rings->cq_ring_mask = p->cq_entries - 1;
9459 rings->sq_ring_entries = p->sq_entries;
9460 rings->cq_ring_entries = p->cq_entries;
9461 ctx->sq_mask = rings->sq_ring_mask;
9462 ctx->cq_mask = rings->cq_ring_mask;
9464 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9465 if (size == SIZE_MAX) {
9466 io_mem_free(ctx->rings);
9471 ctx->sq_sqes = io_mem_alloc(size);
9472 if (!ctx->sq_sqes) {
9473 io_mem_free(ctx->rings);
9481 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9485 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9489 ret = io_uring_add_task_file(ctx);
9494 fd_install(fd, file);
9499 * Allocate an anonymous fd, this is what constitutes the application
9500 * visible backing of an io_uring instance. The application mmaps this
9501 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9502 * we have to tie this fd to a socket for file garbage collection purposes.
9504 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9507 #if defined(CONFIG_UNIX)
9510 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9513 return ERR_PTR(ret);
9516 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9517 O_RDWR | O_CLOEXEC);
9518 #if defined(CONFIG_UNIX)
9520 sock_release(ctx->ring_sock);
9521 ctx->ring_sock = NULL;
9523 ctx->ring_sock->file = file;
9529 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9530 struct io_uring_params __user *params)
9532 struct io_ring_ctx *ctx;
9538 if (entries > IORING_MAX_ENTRIES) {
9539 if (!(p->flags & IORING_SETUP_CLAMP))
9541 entries = IORING_MAX_ENTRIES;
9545 * Use twice as many entries for the CQ ring. It's possible for the
9546 * application to drive a higher depth than the size of the SQ ring,
9547 * since the sqes are only used at submission time. This allows for
9548 * some flexibility in overcommitting a bit. If the application has
9549 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9550 * of CQ ring entries manually.
9552 p->sq_entries = roundup_pow_of_two(entries);
9553 if (p->flags & IORING_SETUP_CQSIZE) {
9555 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9556 * to a power-of-two, if it isn't already. We do NOT impose
9557 * any cq vs sq ring sizing.
9561 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9562 if (!(p->flags & IORING_SETUP_CLAMP))
9564 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9566 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9567 if (p->cq_entries < p->sq_entries)
9570 p->cq_entries = 2 * p->sq_entries;
9573 ctx = io_ring_ctx_alloc(p);
9576 ctx->compat = in_compat_syscall();
9577 if (!capable(CAP_IPC_LOCK))
9578 ctx->user = get_uid(current_user());
9581 * This is just grabbed for accounting purposes. When a process exits,
9582 * the mm is exited and dropped before the files, hence we need to hang
9583 * on to this mm purely for the purposes of being able to unaccount
9584 * memory (locked/pinned vm). It's not used for anything else.
9586 mmgrab(current->mm);
9587 ctx->mm_account = current->mm;
9589 ret = io_allocate_scq_urings(ctx, p);
9593 ret = io_sq_offload_create(ctx, p);
9597 memset(&p->sq_off, 0, sizeof(p->sq_off));
9598 p->sq_off.head = offsetof(struct io_rings, sq.head);
9599 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9600 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9601 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9602 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9603 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9604 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9606 memset(&p->cq_off, 0, sizeof(p->cq_off));
9607 p->cq_off.head = offsetof(struct io_rings, cq.head);
9608 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9609 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9610 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9611 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9612 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9613 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9615 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9616 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9617 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9618 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9619 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9621 if (copy_to_user(params, p, sizeof(*p))) {
9626 file = io_uring_get_file(ctx);
9628 ret = PTR_ERR(file);
9633 * Install ring fd as the very last thing, so we don't risk someone
9634 * having closed it before we finish setup
9636 ret = io_uring_install_fd(ctx, file);
9638 /* fput will clean it up */
9643 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9646 io_ring_ctx_wait_and_kill(ctx);
9651 * Sets up an aio uring context, and returns the fd. Applications asks for a
9652 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9653 * params structure passed in.
9655 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9657 struct io_uring_params p;
9660 if (copy_from_user(&p, params, sizeof(p)))
9662 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9667 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9668 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9669 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9670 IORING_SETUP_R_DISABLED))
9673 return io_uring_create(entries, &p, params);
9676 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9677 struct io_uring_params __user *, params)
9679 return io_uring_setup(entries, params);
9682 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9684 struct io_uring_probe *p;
9688 size = struct_size(p, ops, nr_args);
9689 if (size == SIZE_MAX)
9691 p = kzalloc(size, GFP_KERNEL);
9696 if (copy_from_user(p, arg, size))
9699 if (memchr_inv(p, 0, size))
9702 p->last_op = IORING_OP_LAST - 1;
9703 if (nr_args > IORING_OP_LAST)
9704 nr_args = IORING_OP_LAST;
9706 for (i = 0; i < nr_args; i++) {
9708 if (!io_op_defs[i].not_supported)
9709 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9714 if (copy_to_user(arg, p, size))
9721 static int io_register_personality(struct io_ring_ctx *ctx)
9723 const struct cred *creds;
9727 creds = get_current_cred();
9729 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9730 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9737 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9738 unsigned int nr_args)
9740 struct io_uring_restriction *res;
9744 /* Restrictions allowed only if rings started disabled */
9745 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9748 /* We allow only a single restrictions registration */
9749 if (ctx->restrictions.registered)
9752 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9755 size = array_size(nr_args, sizeof(*res));
9756 if (size == SIZE_MAX)
9759 res = memdup_user(arg, size);
9761 return PTR_ERR(res);
9765 for (i = 0; i < nr_args; i++) {
9766 switch (res[i].opcode) {
9767 case IORING_RESTRICTION_REGISTER_OP:
9768 if (res[i].register_op >= IORING_REGISTER_LAST) {
9773 __set_bit(res[i].register_op,
9774 ctx->restrictions.register_op);
9776 case IORING_RESTRICTION_SQE_OP:
9777 if (res[i].sqe_op >= IORING_OP_LAST) {
9782 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9784 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9785 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9787 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9788 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9797 /* Reset all restrictions if an error happened */
9799 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9801 ctx->restrictions.registered = true;
9807 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9809 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9812 if (ctx->restrictions.registered)
9813 ctx->restricted = 1;
9815 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9816 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9817 wake_up(&ctx->sq_data->wait);
9821 static bool io_register_op_must_quiesce(int op)
9824 case IORING_REGISTER_FILES:
9825 case IORING_UNREGISTER_FILES:
9826 case IORING_REGISTER_FILES_UPDATE:
9827 case IORING_REGISTER_PROBE:
9828 case IORING_REGISTER_PERSONALITY:
9829 case IORING_UNREGISTER_PERSONALITY:
9836 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9837 void __user *arg, unsigned nr_args)
9838 __releases(ctx->uring_lock)
9839 __acquires(ctx->uring_lock)
9844 * We're inside the ring mutex, if the ref is already dying, then
9845 * someone else killed the ctx or is already going through
9846 * io_uring_register().
9848 if (percpu_ref_is_dying(&ctx->refs))
9851 if (io_register_op_must_quiesce(opcode)) {
9852 percpu_ref_kill(&ctx->refs);
9855 * Drop uring mutex before waiting for references to exit. If
9856 * another thread is currently inside io_uring_enter() it might
9857 * need to grab the uring_lock to make progress. If we hold it
9858 * here across the drain wait, then we can deadlock. It's safe
9859 * to drop the mutex here, since no new references will come in
9860 * after we've killed the percpu ref.
9862 mutex_unlock(&ctx->uring_lock);
9864 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9867 ret = io_run_task_work_sig();
9872 mutex_lock(&ctx->uring_lock);
9875 percpu_ref_resurrect(&ctx->refs);
9880 if (ctx->restricted) {
9881 if (opcode >= IORING_REGISTER_LAST) {
9886 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9893 case IORING_REGISTER_BUFFERS:
9894 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9896 case IORING_UNREGISTER_BUFFERS:
9900 ret = io_sqe_buffers_unregister(ctx);
9902 case IORING_REGISTER_FILES:
9903 ret = io_sqe_files_register(ctx, arg, nr_args);
9905 case IORING_UNREGISTER_FILES:
9909 ret = io_sqe_files_unregister(ctx);
9911 case IORING_REGISTER_FILES_UPDATE:
9912 ret = io_sqe_files_update(ctx, arg, nr_args);
9914 case IORING_REGISTER_EVENTFD:
9915 case IORING_REGISTER_EVENTFD_ASYNC:
9919 ret = io_eventfd_register(ctx, arg);
9922 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9923 ctx->eventfd_async = 1;
9925 ctx->eventfd_async = 0;
9927 case IORING_UNREGISTER_EVENTFD:
9931 ret = io_eventfd_unregister(ctx);
9933 case IORING_REGISTER_PROBE:
9935 if (!arg || nr_args > 256)
9937 ret = io_probe(ctx, arg, nr_args);
9939 case IORING_REGISTER_PERSONALITY:
9943 ret = io_register_personality(ctx);
9945 case IORING_UNREGISTER_PERSONALITY:
9949 ret = io_unregister_personality(ctx, nr_args);
9951 case IORING_REGISTER_ENABLE_RINGS:
9955 ret = io_register_enable_rings(ctx);
9957 case IORING_REGISTER_RESTRICTIONS:
9958 ret = io_register_restrictions(ctx, arg, nr_args);
9966 if (io_register_op_must_quiesce(opcode)) {
9967 /* bring the ctx back to life */
9968 percpu_ref_reinit(&ctx->refs);
9970 reinit_completion(&ctx->ref_comp);
9975 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9976 void __user *, arg, unsigned int, nr_args)
9978 struct io_ring_ctx *ctx;
9987 if (f.file->f_op != &io_uring_fops)
9990 ctx = f.file->private_data;
9994 mutex_lock(&ctx->uring_lock);
9995 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9996 mutex_unlock(&ctx->uring_lock);
9997 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9998 ctx->cq_ev_fd != NULL, ret);
10004 static int __init io_uring_init(void)
10006 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10007 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10008 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10011 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10012 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10013 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10014 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10015 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10016 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10017 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10018 BUILD_BUG_SQE_ELEM(8, __u64, off);
10019 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10020 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10021 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10022 BUILD_BUG_SQE_ELEM(24, __u32, len);
10023 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10024 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10025 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10026 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10027 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10028 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10029 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10030 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10031 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10032 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10033 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10034 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10035 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10036 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10037 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10038 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10039 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10040 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10041 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10043 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10044 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10045 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10049 __initcall(io_uring_init);