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:
1230 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1231 req->work.flags |= IO_WQ_WORK_UNBOUND;
1236 static void io_prep_async_link(struct io_kiocb *req)
1238 struct io_kiocb *cur;
1240 io_for_each_link(cur, req)
1241 io_prep_async_work(cur);
1244 static void io_queue_async_work(struct io_kiocb *req)
1246 struct io_ring_ctx *ctx = req->ctx;
1247 struct io_kiocb *link = io_prep_linked_timeout(req);
1248 struct io_uring_task *tctx = req->task->io_uring;
1251 BUG_ON(!tctx->io_wq);
1253 /* init ->work of the whole link before punting */
1254 io_prep_async_link(req);
1255 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1256 &req->work, req->flags);
1257 io_wq_enqueue(tctx->io_wq, &req->work);
1259 io_queue_linked_timeout(link);
1262 static void io_kill_timeout(struct io_kiocb *req, int status)
1264 struct io_timeout_data *io = req->async_data;
1267 ret = hrtimer_try_to_cancel(&io->timer);
1269 atomic_set(&req->ctx->cq_timeouts,
1270 atomic_read(&req->ctx->cq_timeouts) + 1);
1271 list_del_init(&req->timeout.list);
1272 io_cqring_fill_event(req, status);
1273 io_put_req_deferred(req, 1);
1277 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1280 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1281 struct io_defer_entry, list);
1283 if (req_need_defer(de->req, de->seq))
1285 list_del_init(&de->list);
1286 io_req_task_queue(de->req);
1288 } while (!list_empty(&ctx->defer_list));
1291 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1295 if (list_empty(&ctx->timeout_list))
1298 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1301 u32 events_needed, events_got;
1302 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1303 struct io_kiocb, timeout.list);
1305 if (io_is_timeout_noseq(req))
1309 * Since seq can easily wrap around over time, subtract
1310 * the last seq at which timeouts were flushed before comparing.
1311 * Assuming not more than 2^31-1 events have happened since,
1312 * these subtractions won't have wrapped, so we can check if
1313 * target is in [last_seq, current_seq] by comparing the two.
1315 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1316 events_got = seq - ctx->cq_last_tm_flush;
1317 if (events_got < events_needed)
1320 list_del_init(&req->timeout.list);
1321 io_kill_timeout(req, 0);
1322 } while (!list_empty(&ctx->timeout_list));
1324 ctx->cq_last_tm_flush = seq;
1327 static void io_commit_cqring(struct io_ring_ctx *ctx)
1329 io_flush_timeouts(ctx);
1331 /* order cqe stores with ring update */
1332 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1334 if (unlikely(!list_empty(&ctx->defer_list)))
1335 __io_queue_deferred(ctx);
1338 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1340 struct io_rings *r = ctx->rings;
1342 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1345 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1347 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1350 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1352 struct io_rings *rings = ctx->rings;
1356 * writes to the cq entry need to come after reading head; the
1357 * control dependency is enough as we're using WRITE_ONCE to
1360 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1363 tail = ctx->cached_cq_tail++;
1364 return &rings->cqes[tail & ctx->cq_mask];
1367 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1371 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1373 if (!ctx->eventfd_async)
1375 return io_wq_current_is_worker();
1378 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1380 /* see waitqueue_active() comment */
1383 if (waitqueue_active(&ctx->wait))
1384 wake_up(&ctx->wait);
1385 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1386 wake_up(&ctx->sq_data->wait);
1387 if (io_should_trigger_evfd(ctx))
1388 eventfd_signal(ctx->cq_ev_fd, 1);
1389 if (waitqueue_active(&ctx->cq_wait)) {
1390 wake_up_interruptible(&ctx->cq_wait);
1391 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1395 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1397 /* see waitqueue_active() comment */
1400 if (ctx->flags & IORING_SETUP_SQPOLL) {
1401 if (waitqueue_active(&ctx->wait))
1402 wake_up(&ctx->wait);
1404 if (io_should_trigger_evfd(ctx))
1405 eventfd_signal(ctx->cq_ev_fd, 1);
1406 if (waitqueue_active(&ctx->cq_wait)) {
1407 wake_up_interruptible(&ctx->cq_wait);
1408 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1412 /* Returns true if there are no backlogged entries after the flush */
1413 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1415 struct io_rings *rings = ctx->rings;
1416 unsigned long flags;
1417 bool all_flushed, posted;
1419 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1423 spin_lock_irqsave(&ctx->completion_lock, flags);
1424 while (!list_empty(&ctx->cq_overflow_list)) {
1425 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1426 struct io_overflow_cqe *ocqe;
1430 ocqe = list_first_entry(&ctx->cq_overflow_list,
1431 struct io_overflow_cqe, list);
1433 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1435 WRITE_ONCE(ctx->rings->cq_overflow,
1436 ++ctx->cached_cq_overflow);
1438 list_del(&ocqe->list);
1442 all_flushed = list_empty(&ctx->cq_overflow_list);
1444 clear_bit(0, &ctx->sq_check_overflow);
1445 clear_bit(0, &ctx->cq_check_overflow);
1446 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1450 io_commit_cqring(ctx);
1451 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1453 io_cqring_ev_posted(ctx);
1457 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1461 if (test_bit(0, &ctx->cq_check_overflow)) {
1462 /* iopoll syncs against uring_lock, not completion_lock */
1463 if (ctx->flags & IORING_SETUP_IOPOLL)
1464 mutex_lock(&ctx->uring_lock);
1465 ret = __io_cqring_overflow_flush(ctx, force);
1466 if (ctx->flags & IORING_SETUP_IOPOLL)
1467 mutex_unlock(&ctx->uring_lock);
1474 * Shamelessly stolen from the mm implementation of page reference checking,
1475 * see commit f958d7b528b1 for details.
1477 #define req_ref_zero_or_close_to_overflow(req) \
1478 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1480 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1482 return atomic_inc_not_zero(&req->refs);
1485 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1487 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1488 return atomic_sub_and_test(refs, &req->refs);
1491 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1493 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1494 return atomic_dec_and_test(&req->refs);
1497 static inline void req_ref_put(struct io_kiocb *req)
1499 WARN_ON_ONCE(req_ref_put_and_test(req));
1502 static inline void req_ref_get(struct io_kiocb *req)
1504 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1505 atomic_inc(&req->refs);
1508 static bool __io_cqring_fill_event(struct io_kiocb *req, long res,
1509 unsigned int cflags)
1511 struct io_ring_ctx *ctx = req->ctx;
1512 struct io_uring_cqe *cqe;
1514 trace_io_uring_complete(ctx, req->user_data, res, cflags);
1517 * If we can't get a cq entry, userspace overflowed the
1518 * submission (by quite a lot). Increment the overflow count in
1521 cqe = io_get_cqring(ctx);
1523 WRITE_ONCE(cqe->user_data, req->user_data);
1524 WRITE_ONCE(cqe->res, res);
1525 WRITE_ONCE(cqe->flags, cflags);
1528 if (!atomic_read(&req->task->io_uring->in_idle)) {
1529 struct io_overflow_cqe *ocqe;
1531 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1534 if (list_empty(&ctx->cq_overflow_list)) {
1535 set_bit(0, &ctx->sq_check_overflow);
1536 set_bit(0, &ctx->cq_check_overflow);
1537 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1539 ocqe->cqe.user_data = req->user_data;
1540 ocqe->cqe.res = res;
1541 ocqe->cqe.flags = cflags;
1542 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1547 * If we're in ring overflow flush mode, or in task cancel mode,
1548 * or cannot allocate an overflow entry, then we need to drop it
1551 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1555 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1557 __io_cqring_fill_event(req, res, 0);
1560 static void io_req_complete_post(struct io_kiocb *req, long res,
1561 unsigned int cflags)
1563 struct io_ring_ctx *ctx = req->ctx;
1564 unsigned long flags;
1566 spin_lock_irqsave(&ctx->completion_lock, flags);
1567 __io_cqring_fill_event(req, res, cflags);
1569 * If we're the last reference to this request, add to our locked
1572 if (req_ref_put_and_test(req)) {
1573 struct io_comp_state *cs = &ctx->submit_state.comp;
1575 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1576 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1577 io_disarm_next(req);
1579 io_req_task_queue(req->link);
1583 io_dismantle_req(req);
1584 io_put_task(req->task, 1);
1585 list_add(&req->compl.list, &cs->locked_free_list);
1586 cs->locked_free_nr++;
1588 if (!percpu_ref_tryget(&ctx->refs))
1591 io_commit_cqring(ctx);
1592 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1595 io_cqring_ev_posted(ctx);
1596 percpu_ref_put(&ctx->refs);
1600 static void io_req_complete_state(struct io_kiocb *req, long res,
1601 unsigned int cflags)
1603 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1606 req->compl.cflags = cflags;
1607 req->flags |= REQ_F_COMPLETE_INLINE;
1610 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1611 long res, unsigned cflags)
1613 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1614 io_req_complete_state(req, res, cflags);
1616 io_req_complete_post(req, res, cflags);
1619 static inline void io_req_complete(struct io_kiocb *req, long res)
1621 __io_req_complete(req, 0, res, 0);
1624 static void io_req_complete_failed(struct io_kiocb *req, long res)
1626 req_set_fail_links(req);
1628 io_req_complete_post(req, res, 0);
1631 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1632 struct io_comp_state *cs)
1634 spin_lock_irq(&ctx->completion_lock);
1635 list_splice_init(&cs->locked_free_list, &cs->free_list);
1636 cs->locked_free_nr = 0;
1637 spin_unlock_irq(&ctx->completion_lock);
1640 /* Returns true IFF there are requests in the cache */
1641 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1643 struct io_submit_state *state = &ctx->submit_state;
1644 struct io_comp_state *cs = &state->comp;
1648 * If we have more than a batch's worth of requests in our IRQ side
1649 * locked cache, grab the lock and move them over to our submission
1652 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1653 io_flush_cached_locked_reqs(ctx, cs);
1655 nr = state->free_reqs;
1656 while (!list_empty(&cs->free_list)) {
1657 struct io_kiocb *req = list_first_entry(&cs->free_list,
1658 struct io_kiocb, compl.list);
1660 list_del(&req->compl.list);
1661 state->reqs[nr++] = req;
1662 if (nr == ARRAY_SIZE(state->reqs))
1666 state->free_reqs = nr;
1670 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1672 struct io_submit_state *state = &ctx->submit_state;
1674 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1676 if (!state->free_reqs) {
1677 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1680 if (io_flush_cached_reqs(ctx))
1683 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1687 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1688 * retry single alloc to be on the safe side.
1690 if (unlikely(ret <= 0)) {
1691 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1692 if (!state->reqs[0])
1696 state->free_reqs = ret;
1700 return state->reqs[state->free_reqs];
1703 static inline void io_put_file(struct file *file)
1709 static void io_dismantle_req(struct io_kiocb *req)
1711 unsigned int flags = req->flags;
1713 if (!(flags & REQ_F_FIXED_FILE))
1714 io_put_file(req->file);
1715 if (flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1719 if (req->flags & REQ_F_INFLIGHT) {
1720 struct io_ring_ctx *ctx = req->ctx;
1721 unsigned long flags;
1723 spin_lock_irqsave(&ctx->inflight_lock, flags);
1724 list_del(&req->inflight_entry);
1725 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1726 req->flags &= ~REQ_F_INFLIGHT;
1729 if (req->fixed_rsrc_refs)
1730 percpu_ref_put(req->fixed_rsrc_refs);
1731 if (req->async_data)
1732 kfree(req->async_data);
1733 if (req->work.creds) {
1734 put_cred(req->work.creds);
1735 req->work.creds = NULL;
1739 /* must to be called somewhat shortly after putting a request */
1740 static inline void io_put_task(struct task_struct *task, int nr)
1742 struct io_uring_task *tctx = task->io_uring;
1744 percpu_counter_sub(&tctx->inflight, nr);
1745 if (unlikely(atomic_read(&tctx->in_idle)))
1746 wake_up(&tctx->wait);
1747 put_task_struct_many(task, nr);
1750 static void __io_free_req(struct io_kiocb *req)
1752 struct io_ring_ctx *ctx = req->ctx;
1754 io_dismantle_req(req);
1755 io_put_task(req->task, 1);
1757 kmem_cache_free(req_cachep, req);
1758 percpu_ref_put(&ctx->refs);
1761 static inline void io_remove_next_linked(struct io_kiocb *req)
1763 struct io_kiocb *nxt = req->link;
1765 req->link = nxt->link;
1769 static bool io_kill_linked_timeout(struct io_kiocb *req)
1770 __must_hold(&req->ctx->completion_lock)
1772 struct io_kiocb *link = req->link;
1775 * Can happen if a linked timeout fired and link had been like
1776 * req -> link t-out -> link t-out [-> ...]
1778 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1779 struct io_timeout_data *io = link->async_data;
1782 io_remove_next_linked(req);
1783 link->timeout.head = NULL;
1784 ret = hrtimer_try_to_cancel(&io->timer);
1786 io_cqring_fill_event(link, -ECANCELED);
1787 io_put_req_deferred(link, 1);
1794 static void io_fail_links(struct io_kiocb *req)
1795 __must_hold(&req->ctx->completion_lock)
1797 struct io_kiocb *nxt, *link = req->link;
1804 trace_io_uring_fail_link(req, link);
1805 io_cqring_fill_event(link, -ECANCELED);
1806 io_put_req_deferred(link, 2);
1811 static bool io_disarm_next(struct io_kiocb *req)
1812 __must_hold(&req->ctx->completion_lock)
1814 bool posted = false;
1816 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1817 posted = io_kill_linked_timeout(req);
1818 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1819 posted |= (req->link != NULL);
1825 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1827 struct io_kiocb *nxt;
1830 * If LINK is set, we have dependent requests in this chain. If we
1831 * didn't fail this request, queue the first one up, moving any other
1832 * dependencies to the next request. In case of failure, fail the rest
1835 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1836 struct io_ring_ctx *ctx = req->ctx;
1837 unsigned long flags;
1840 spin_lock_irqsave(&ctx->completion_lock, flags);
1841 posted = io_disarm_next(req);
1843 io_commit_cqring(req->ctx);
1844 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1846 io_cqring_ev_posted(ctx);
1853 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1855 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1857 return __io_req_find_next(req);
1860 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1864 if (ctx->submit_state.comp.nr) {
1865 mutex_lock(&ctx->uring_lock);
1866 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1867 mutex_unlock(&ctx->uring_lock);
1869 percpu_ref_put(&ctx->refs);
1872 static bool __tctx_task_work(struct io_uring_task *tctx)
1874 struct io_ring_ctx *ctx = NULL;
1875 struct io_wq_work_list list;
1876 struct io_wq_work_node *node;
1878 if (wq_list_empty(&tctx->task_list))
1881 spin_lock_irq(&tctx->task_lock);
1882 list = tctx->task_list;
1883 INIT_WQ_LIST(&tctx->task_list);
1884 spin_unlock_irq(&tctx->task_lock);
1888 struct io_wq_work_node *next = node->next;
1889 struct io_kiocb *req;
1891 req = container_of(node, struct io_kiocb, io_task_work.node);
1892 if (req->ctx != ctx) {
1893 ctx_flush_and_put(ctx);
1895 percpu_ref_get(&ctx->refs);
1898 req->task_work.func(&req->task_work);
1902 ctx_flush_and_put(ctx);
1903 return list.first != NULL;
1906 static void tctx_task_work(struct callback_head *cb)
1908 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1910 clear_bit(0, &tctx->task_state);
1912 while (__tctx_task_work(tctx))
1916 static int io_req_task_work_add(struct io_kiocb *req)
1918 struct task_struct *tsk = req->task;
1919 struct io_uring_task *tctx = tsk->io_uring;
1920 enum task_work_notify_mode notify;
1921 struct io_wq_work_node *node, *prev;
1922 unsigned long flags;
1925 if (unlikely(tsk->flags & PF_EXITING))
1928 WARN_ON_ONCE(!tctx);
1930 spin_lock_irqsave(&tctx->task_lock, flags);
1931 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1932 spin_unlock_irqrestore(&tctx->task_lock, flags);
1934 /* task_work already pending, we're done */
1935 if (test_bit(0, &tctx->task_state) ||
1936 test_and_set_bit(0, &tctx->task_state))
1940 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1941 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1942 * processing task_work. There's no reliable way to tell if TWA_RESUME
1945 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1947 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1948 wake_up_process(tsk);
1953 * Slow path - we failed, find and delete work. if the work is not
1954 * in the list, it got run and we're fine.
1956 spin_lock_irqsave(&tctx->task_lock, flags);
1957 wq_list_for_each(node, prev, &tctx->task_list) {
1958 if (&req->io_task_work.node == node) {
1959 wq_list_del(&tctx->task_list, node, prev);
1964 spin_unlock_irqrestore(&tctx->task_lock, flags);
1965 clear_bit(0, &tctx->task_state);
1969 static bool io_run_task_work_head(struct callback_head **work_head)
1971 struct callback_head *work, *next;
1972 bool executed = false;
1975 work = xchg(work_head, NULL);
1991 static void io_task_work_add_head(struct callback_head **work_head,
1992 struct callback_head *task_work)
1994 struct callback_head *head;
1997 head = READ_ONCE(*work_head);
1998 task_work->next = head;
1999 } while (cmpxchg(work_head, head, task_work) != head);
2002 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2003 task_work_func_t cb)
2005 init_task_work(&req->task_work, cb);
2006 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2009 static void io_req_task_cancel(struct callback_head *cb)
2011 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2012 struct io_ring_ctx *ctx = req->ctx;
2014 /* ctx is guaranteed to stay alive while we hold uring_lock */
2015 mutex_lock(&ctx->uring_lock);
2016 io_req_complete_failed(req, req->result);
2017 mutex_unlock(&ctx->uring_lock);
2020 static void __io_req_task_submit(struct io_kiocb *req)
2022 struct io_ring_ctx *ctx = req->ctx;
2024 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2025 mutex_lock(&ctx->uring_lock);
2026 if (!(current->flags & PF_EXITING) && !current->in_execve)
2027 __io_queue_sqe(req);
2029 io_req_complete_failed(req, -EFAULT);
2030 mutex_unlock(&ctx->uring_lock);
2033 static void io_req_task_submit(struct callback_head *cb)
2035 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2037 __io_req_task_submit(req);
2040 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2043 req->task_work.func = io_req_task_cancel;
2045 if (unlikely(io_req_task_work_add(req)))
2046 io_req_task_work_add_fallback(req, io_req_task_cancel);
2049 static void io_req_task_queue(struct io_kiocb *req)
2051 req->task_work.func = io_req_task_submit;
2053 if (unlikely(io_req_task_work_add(req)))
2054 io_req_task_queue_fail(req, -ECANCELED);
2057 static inline void io_queue_next(struct io_kiocb *req)
2059 struct io_kiocb *nxt = io_req_find_next(req);
2062 io_req_task_queue(nxt);
2065 static void io_free_req(struct io_kiocb *req)
2072 struct task_struct *task;
2077 static inline void io_init_req_batch(struct req_batch *rb)
2084 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2085 struct req_batch *rb)
2088 io_put_task(rb->task, rb->task_refs);
2090 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2093 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2094 struct io_submit_state *state)
2097 io_dismantle_req(req);
2099 if (req->task != rb->task) {
2101 io_put_task(rb->task, rb->task_refs);
2102 rb->task = req->task;
2108 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2109 state->reqs[state->free_reqs++] = req;
2111 list_add(&req->compl.list, &state->comp.free_list);
2114 static void io_submit_flush_completions(struct io_comp_state *cs,
2115 struct io_ring_ctx *ctx)
2118 struct io_kiocb *req;
2119 struct req_batch rb;
2121 io_init_req_batch(&rb);
2122 spin_lock_irq(&ctx->completion_lock);
2123 for (i = 0; i < nr; i++) {
2125 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2127 io_commit_cqring(ctx);
2128 spin_unlock_irq(&ctx->completion_lock);
2130 io_cqring_ev_posted(ctx);
2131 for (i = 0; i < nr; i++) {
2134 /* submission and completion refs */
2135 if (req_ref_sub_and_test(req, 2))
2136 io_req_free_batch(&rb, req, &ctx->submit_state);
2139 io_req_free_batch_finish(ctx, &rb);
2144 * Drop reference to request, return next in chain (if there is one) if this
2145 * was the last reference to this request.
2147 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2149 struct io_kiocb *nxt = NULL;
2151 if (req_ref_put_and_test(req)) {
2152 nxt = io_req_find_next(req);
2158 static inline void io_put_req(struct io_kiocb *req)
2160 if (req_ref_put_and_test(req))
2164 static void io_put_req_deferred_cb(struct callback_head *cb)
2166 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2171 static void io_free_req_deferred(struct io_kiocb *req)
2173 req->task_work.func = io_put_req_deferred_cb;
2174 if (unlikely(io_req_task_work_add(req)))
2175 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2178 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2180 if (req_ref_sub_and_test(req, refs))
2181 io_free_req_deferred(req);
2184 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2186 /* See comment at the top of this file */
2188 return __io_cqring_events(ctx);
2191 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2193 struct io_rings *rings = ctx->rings;
2195 /* make sure SQ entry isn't read before tail */
2196 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2199 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2201 unsigned int cflags;
2203 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2204 cflags |= IORING_CQE_F_BUFFER;
2205 req->flags &= ~REQ_F_BUFFER_SELECTED;
2210 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2212 struct io_buffer *kbuf;
2214 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2215 return io_put_kbuf(req, kbuf);
2218 static inline bool io_run_task_work(void)
2221 * Not safe to run on exiting task, and the task_work handling will
2222 * not add work to such a task.
2224 if (unlikely(current->flags & PF_EXITING))
2226 if (current->task_works) {
2227 __set_current_state(TASK_RUNNING);
2236 * Find and free completed poll iocbs
2238 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2239 struct list_head *done)
2241 struct req_batch rb;
2242 struct io_kiocb *req;
2244 /* order with ->result store in io_complete_rw_iopoll() */
2247 io_init_req_batch(&rb);
2248 while (!list_empty(done)) {
2251 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2252 list_del(&req->inflight_entry);
2254 if (READ_ONCE(req->result) == -EAGAIN &&
2255 !(req->flags & REQ_F_DONT_REISSUE)) {
2256 req->iopoll_completed = 0;
2258 io_queue_async_work(req);
2262 if (req->flags & REQ_F_BUFFER_SELECTED)
2263 cflags = io_put_rw_kbuf(req);
2265 __io_cqring_fill_event(req, req->result, cflags);
2268 if (req_ref_put_and_test(req))
2269 io_req_free_batch(&rb, req, &ctx->submit_state);
2272 io_commit_cqring(ctx);
2273 io_cqring_ev_posted_iopoll(ctx);
2274 io_req_free_batch_finish(ctx, &rb);
2277 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2280 struct io_kiocb *req, *tmp;
2286 * Only spin for completions if we don't have multiple devices hanging
2287 * off our complete list, and we're under the requested amount.
2289 spin = !ctx->poll_multi_file && *nr_events < min;
2292 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2293 struct kiocb *kiocb = &req->rw.kiocb;
2296 * Move completed and retryable entries to our local lists.
2297 * If we find a request that requires polling, break out
2298 * and complete those lists first, if we have entries there.
2300 if (READ_ONCE(req->iopoll_completed)) {
2301 list_move_tail(&req->inflight_entry, &done);
2304 if (!list_empty(&done))
2307 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2311 /* iopoll may have completed current req */
2312 if (READ_ONCE(req->iopoll_completed))
2313 list_move_tail(&req->inflight_entry, &done);
2320 if (!list_empty(&done))
2321 io_iopoll_complete(ctx, nr_events, &done);
2327 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2328 * non-spinning poll check - we'll still enter the driver poll loop, but only
2329 * as a non-spinning completion check.
2331 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2334 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2337 ret = io_do_iopoll(ctx, nr_events, min);
2340 if (*nr_events >= min)
2348 * We can't just wait for polled events to come to us, we have to actively
2349 * find and complete them.
2351 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2353 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2356 mutex_lock(&ctx->uring_lock);
2357 while (!list_empty(&ctx->iopoll_list)) {
2358 unsigned int nr_events = 0;
2360 io_do_iopoll(ctx, &nr_events, 0);
2362 /* let it sleep and repeat later if can't complete a request */
2366 * Ensure we allow local-to-the-cpu processing to take place,
2367 * in this case we need to ensure that we reap all events.
2368 * Also let task_work, etc. to progress by releasing the mutex
2370 if (need_resched()) {
2371 mutex_unlock(&ctx->uring_lock);
2373 mutex_lock(&ctx->uring_lock);
2376 mutex_unlock(&ctx->uring_lock);
2379 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2381 unsigned int nr_events = 0;
2382 int iters = 0, ret = 0;
2385 * We disallow the app entering submit/complete with polling, but we
2386 * still need to lock the ring to prevent racing with polled issue
2387 * that got punted to a workqueue.
2389 mutex_lock(&ctx->uring_lock);
2392 * Don't enter poll loop if we already have events pending.
2393 * If we do, we can potentially be spinning for commands that
2394 * already triggered a CQE (eg in error).
2396 if (test_bit(0, &ctx->cq_check_overflow))
2397 __io_cqring_overflow_flush(ctx, false);
2398 if (io_cqring_events(ctx))
2402 * If a submit got punted to a workqueue, we can have the
2403 * application entering polling for a command before it gets
2404 * issued. That app will hold the uring_lock for the duration
2405 * of the poll right here, so we need to take a breather every
2406 * now and then to ensure that the issue has a chance to add
2407 * the poll to the issued list. Otherwise we can spin here
2408 * forever, while the workqueue is stuck trying to acquire the
2411 if (!(++iters & 7)) {
2412 mutex_unlock(&ctx->uring_lock);
2414 mutex_lock(&ctx->uring_lock);
2417 ret = io_iopoll_getevents(ctx, &nr_events, min);
2421 } while (min && !nr_events && !need_resched());
2423 mutex_unlock(&ctx->uring_lock);
2427 static void kiocb_end_write(struct io_kiocb *req)
2430 * Tell lockdep we inherited freeze protection from submission
2433 if (req->flags & REQ_F_ISREG) {
2434 struct super_block *sb = file_inode(req->file)->i_sb;
2436 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2442 static bool io_resubmit_prep(struct io_kiocb *req)
2444 struct io_async_rw *rw = req->async_data;
2447 return !io_req_prep_async(req);
2448 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2449 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2453 static bool io_rw_should_reissue(struct io_kiocb *req)
2455 umode_t mode = file_inode(req->file)->i_mode;
2456 struct io_ring_ctx *ctx = req->ctx;
2458 if (!S_ISBLK(mode) && !S_ISREG(mode))
2460 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2461 !(ctx->flags & IORING_SETUP_IOPOLL)))
2464 * If ref is dying, we might be running poll reap from the exit work.
2465 * Don't attempt to reissue from that path, just let it fail with
2468 if (percpu_ref_is_dying(&ctx->refs))
2473 static bool io_rw_should_reissue(struct io_kiocb *req)
2479 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2480 unsigned int issue_flags)
2484 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2485 kiocb_end_write(req);
2486 if (res != req->result) {
2487 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2488 io_rw_should_reissue(req)) {
2489 req->flags |= REQ_F_REISSUE;
2492 req_set_fail_links(req);
2494 if (req->flags & REQ_F_BUFFER_SELECTED)
2495 cflags = io_put_rw_kbuf(req);
2496 __io_req_complete(req, issue_flags, res, cflags);
2499 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 __io_complete_rw(req, res, res2, 0);
2506 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2508 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2510 if (kiocb->ki_flags & IOCB_WRITE)
2511 kiocb_end_write(req);
2512 if (unlikely(res != req->result)) {
2516 if (res == -EAGAIN && io_rw_should_reissue(req) &&
2517 io_resubmit_prep(req))
2521 req_set_fail_links(req);
2522 req->flags |= REQ_F_DONT_REISSUE;
2526 WRITE_ONCE(req->result, res);
2527 /* order with io_iopoll_complete() checking ->result */
2529 WRITE_ONCE(req->iopoll_completed, 1);
2533 * After the iocb has been issued, it's safe to be found on the poll list.
2534 * Adding the kiocb to the list AFTER submission ensures that we don't
2535 * find it from a io_iopoll_getevents() thread before the issuer is done
2536 * accessing the kiocb cookie.
2538 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2540 struct io_ring_ctx *ctx = req->ctx;
2543 * Track whether we have multiple files in our lists. This will impact
2544 * how we do polling eventually, not spinning if we're on potentially
2545 * different devices.
2547 if (list_empty(&ctx->iopoll_list)) {
2548 ctx->poll_multi_file = false;
2549 } else if (!ctx->poll_multi_file) {
2550 struct io_kiocb *list_req;
2552 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2554 if (list_req->file != req->file)
2555 ctx->poll_multi_file = true;
2559 * For fast devices, IO may have already completed. If it has, add
2560 * it to the front so we find it first.
2562 if (READ_ONCE(req->iopoll_completed))
2563 list_add(&req->inflight_entry, &ctx->iopoll_list);
2565 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2568 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2569 * task context or in io worker task context. If current task context is
2570 * sq thread, we don't need to check whether should wake up sq thread.
2572 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2573 wq_has_sleeper(&ctx->sq_data->wait))
2574 wake_up(&ctx->sq_data->wait);
2577 static inline void io_state_file_put(struct io_submit_state *state)
2579 if (state->file_refs) {
2580 fput_many(state->file, state->file_refs);
2581 state->file_refs = 0;
2586 * Get as many references to a file as we have IOs left in this submission,
2587 * assuming most submissions are for one file, or at least that each file
2588 * has more than one submission.
2590 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2595 if (state->file_refs) {
2596 if (state->fd == fd) {
2600 io_state_file_put(state);
2602 state->file = fget_many(fd, state->ios_left);
2603 if (unlikely(!state->file))
2607 state->file_refs = state->ios_left - 1;
2611 static bool io_bdev_nowait(struct block_device *bdev)
2613 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2617 * If we tracked the file through the SCM inflight mechanism, we could support
2618 * any file. For now, just ensure that anything potentially problematic is done
2621 static bool __io_file_supports_async(struct file *file, int rw)
2623 umode_t mode = file_inode(file)->i_mode;
2625 if (S_ISBLK(mode)) {
2626 if (IS_ENABLED(CONFIG_BLOCK) &&
2627 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2631 if (S_ISCHR(mode) || S_ISSOCK(mode))
2633 if (S_ISREG(mode)) {
2634 if (IS_ENABLED(CONFIG_BLOCK) &&
2635 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2636 file->f_op != &io_uring_fops)
2641 /* any ->read/write should understand O_NONBLOCK */
2642 if (file->f_flags & O_NONBLOCK)
2645 if (!(file->f_mode & FMODE_NOWAIT))
2649 return file->f_op->read_iter != NULL;
2651 return file->f_op->write_iter != NULL;
2654 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2656 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2658 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2661 return __io_file_supports_async(req->file, rw);
2664 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2666 struct io_ring_ctx *ctx = req->ctx;
2667 struct kiocb *kiocb = &req->rw.kiocb;
2668 struct file *file = req->file;
2672 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2673 req->flags |= REQ_F_ISREG;
2675 kiocb->ki_pos = READ_ONCE(sqe->off);
2676 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2677 req->flags |= REQ_F_CUR_POS;
2678 kiocb->ki_pos = file->f_pos;
2680 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2681 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2682 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2686 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2687 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2688 req->flags |= REQ_F_NOWAIT;
2690 ioprio = READ_ONCE(sqe->ioprio);
2692 ret = ioprio_check_cap(ioprio);
2696 kiocb->ki_ioprio = ioprio;
2698 kiocb->ki_ioprio = get_current_ioprio();
2700 if (ctx->flags & IORING_SETUP_IOPOLL) {
2701 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2702 !kiocb->ki_filp->f_op->iopoll)
2705 kiocb->ki_flags |= IOCB_HIPRI;
2706 kiocb->ki_complete = io_complete_rw_iopoll;
2707 req->iopoll_completed = 0;
2709 if (kiocb->ki_flags & IOCB_HIPRI)
2711 kiocb->ki_complete = io_complete_rw;
2714 req->rw.addr = READ_ONCE(sqe->addr);
2715 req->rw.len = READ_ONCE(sqe->len);
2716 req->buf_index = READ_ONCE(sqe->buf_index);
2720 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2726 case -ERESTARTNOINTR:
2727 case -ERESTARTNOHAND:
2728 case -ERESTART_RESTARTBLOCK:
2730 * We can't just restart the syscall, since previously
2731 * submitted sqes may already be in progress. Just fail this
2737 kiocb->ki_complete(kiocb, ret, 0);
2741 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2742 unsigned int issue_flags)
2744 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2745 struct io_async_rw *io = req->async_data;
2746 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2748 /* add previously done IO, if any */
2749 if (io && io->bytes_done > 0) {
2751 ret = io->bytes_done;
2753 ret += io->bytes_done;
2756 if (req->flags & REQ_F_CUR_POS)
2757 req->file->f_pos = kiocb->ki_pos;
2758 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2759 __io_complete_rw(req, ret, 0, issue_flags);
2761 io_rw_done(kiocb, ret);
2763 if (check_reissue && req->flags & REQ_F_REISSUE) {
2764 req->flags &= ~REQ_F_REISSUE;
2765 if (!io_resubmit_prep(req)) {
2767 io_queue_async_work(req);
2771 req_set_fail_links(req);
2772 if (req->flags & REQ_F_BUFFER_SELECTED)
2773 cflags = io_put_rw_kbuf(req);
2774 __io_req_complete(req, issue_flags, ret, cflags);
2779 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2781 struct io_ring_ctx *ctx = req->ctx;
2782 size_t len = req->rw.len;
2783 struct io_mapped_ubuf *imu;
2784 u16 index, buf_index = req->buf_index;
2785 u64 buf_end, buf_addr = req->rw.addr;
2788 if (unlikely(buf_index >= ctx->nr_user_bufs))
2790 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2791 imu = &ctx->user_bufs[index];
2792 buf_addr = req->rw.addr;
2794 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2796 /* not inside the mapped region */
2797 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2801 * May not be a start of buffer, set size appropriately
2802 * and advance us to the beginning.
2804 offset = buf_addr - imu->ubuf;
2805 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2809 * Don't use iov_iter_advance() here, as it's really slow for
2810 * using the latter parts of a big fixed buffer - it iterates
2811 * over each segment manually. We can cheat a bit here, because
2814 * 1) it's a BVEC iter, we set it up
2815 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2816 * first and last bvec
2818 * So just find our index, and adjust the iterator afterwards.
2819 * If the offset is within the first bvec (or the whole first
2820 * bvec, just use iov_iter_advance(). This makes it easier
2821 * since we can just skip the first segment, which may not
2822 * be PAGE_SIZE aligned.
2824 const struct bio_vec *bvec = imu->bvec;
2826 if (offset <= bvec->bv_len) {
2827 iov_iter_advance(iter, offset);
2829 unsigned long seg_skip;
2831 /* skip first vec */
2832 offset -= bvec->bv_len;
2833 seg_skip = 1 + (offset >> PAGE_SHIFT);
2835 iter->bvec = bvec + seg_skip;
2836 iter->nr_segs -= seg_skip;
2837 iter->count -= bvec->bv_len + offset;
2838 iter->iov_offset = offset & ~PAGE_MASK;
2845 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2848 mutex_unlock(&ctx->uring_lock);
2851 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2854 * "Normal" inline submissions always hold the uring_lock, since we
2855 * grab it from the system call. Same is true for the SQPOLL offload.
2856 * The only exception is when we've detached the request and issue it
2857 * from an async worker thread, grab the lock for that case.
2860 mutex_lock(&ctx->uring_lock);
2863 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2864 int bgid, struct io_buffer *kbuf,
2867 struct io_buffer *head;
2869 if (req->flags & REQ_F_BUFFER_SELECTED)
2872 io_ring_submit_lock(req->ctx, needs_lock);
2874 lockdep_assert_held(&req->ctx->uring_lock);
2876 head = xa_load(&req->ctx->io_buffers, bgid);
2878 if (!list_empty(&head->list)) {
2879 kbuf = list_last_entry(&head->list, struct io_buffer,
2881 list_del(&kbuf->list);
2884 xa_erase(&req->ctx->io_buffers, bgid);
2886 if (*len > kbuf->len)
2889 kbuf = ERR_PTR(-ENOBUFS);
2892 io_ring_submit_unlock(req->ctx, needs_lock);
2897 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2900 struct io_buffer *kbuf;
2903 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2904 bgid = req->buf_index;
2905 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2908 req->rw.addr = (u64) (unsigned long) kbuf;
2909 req->flags |= REQ_F_BUFFER_SELECTED;
2910 return u64_to_user_ptr(kbuf->addr);
2913 #ifdef CONFIG_COMPAT
2914 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2917 struct compat_iovec __user *uiov;
2918 compat_ssize_t clen;
2922 uiov = u64_to_user_ptr(req->rw.addr);
2923 if (!access_ok(uiov, sizeof(*uiov)))
2925 if (__get_user(clen, &uiov->iov_len))
2931 buf = io_rw_buffer_select(req, &len, needs_lock);
2933 return PTR_ERR(buf);
2934 iov[0].iov_base = buf;
2935 iov[0].iov_len = (compat_size_t) len;
2940 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2943 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2947 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2950 len = iov[0].iov_len;
2953 buf = io_rw_buffer_select(req, &len, needs_lock);
2955 return PTR_ERR(buf);
2956 iov[0].iov_base = buf;
2957 iov[0].iov_len = len;
2961 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 if (req->flags & REQ_F_BUFFER_SELECTED) {
2965 struct io_buffer *kbuf;
2967 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2968 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2969 iov[0].iov_len = kbuf->len;
2972 if (req->rw.len != 1)
2975 #ifdef CONFIG_COMPAT
2976 if (req->ctx->compat)
2977 return io_compat_import(req, iov, needs_lock);
2980 return __io_iov_buffer_select(req, iov, needs_lock);
2983 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2984 struct iov_iter *iter, bool needs_lock)
2986 void __user *buf = u64_to_user_ptr(req->rw.addr);
2987 size_t sqe_len = req->rw.len;
2988 u8 opcode = req->opcode;
2991 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2993 return io_import_fixed(req, rw, iter);
2996 /* buffer index only valid with fixed read/write, or buffer select */
2997 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3000 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3001 if (req->flags & REQ_F_BUFFER_SELECT) {
3002 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3004 return PTR_ERR(buf);
3005 req->rw.len = sqe_len;
3008 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3013 if (req->flags & REQ_F_BUFFER_SELECT) {
3014 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3016 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3021 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3025 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3027 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3031 * For files that don't have ->read_iter() and ->write_iter(), handle them
3032 * by looping over ->read() or ->write() manually.
3034 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3036 struct kiocb *kiocb = &req->rw.kiocb;
3037 struct file *file = req->file;
3041 * Don't support polled IO through this interface, and we can't
3042 * support non-blocking either. For the latter, this just causes
3043 * the kiocb to be handled from an async context.
3045 if (kiocb->ki_flags & IOCB_HIPRI)
3047 if (kiocb->ki_flags & IOCB_NOWAIT)
3050 while (iov_iter_count(iter)) {
3054 if (!iov_iter_is_bvec(iter)) {
3055 iovec = iov_iter_iovec(iter);
3057 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3058 iovec.iov_len = req->rw.len;
3062 nr = file->f_op->read(file, iovec.iov_base,
3063 iovec.iov_len, io_kiocb_ppos(kiocb));
3065 nr = file->f_op->write(file, iovec.iov_base,
3066 iovec.iov_len, io_kiocb_ppos(kiocb));
3075 if (nr != iovec.iov_len)
3079 iov_iter_advance(iter, nr);
3085 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3086 const struct iovec *fast_iov, struct iov_iter *iter)
3088 struct io_async_rw *rw = req->async_data;
3090 memcpy(&rw->iter, iter, sizeof(*iter));
3091 rw->free_iovec = iovec;
3093 /* can only be fixed buffers, no need to do anything */
3094 if (iov_iter_is_bvec(iter))
3097 unsigned iov_off = 0;
3099 rw->iter.iov = rw->fast_iov;
3100 if (iter->iov != fast_iov) {
3101 iov_off = iter->iov - fast_iov;
3102 rw->iter.iov += iov_off;
3104 if (rw->fast_iov != fast_iov)
3105 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3106 sizeof(struct iovec) * iter->nr_segs);
3108 req->flags |= REQ_F_NEED_CLEANUP;
3112 static inline int io_alloc_async_data(struct io_kiocb *req)
3114 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3115 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3116 return req->async_data == NULL;
3119 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3120 const struct iovec *fast_iov,
3121 struct iov_iter *iter, bool force)
3123 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3125 if (!req->async_data) {
3126 if (io_alloc_async_data(req)) {
3131 io_req_map_rw(req, iovec, fast_iov, iter);
3136 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3138 struct io_async_rw *iorw = req->async_data;
3139 struct iovec *iov = iorw->fast_iov;
3142 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3143 if (unlikely(ret < 0))
3146 iorw->bytes_done = 0;
3147 iorw->free_iovec = iov;
3149 req->flags |= REQ_F_NEED_CLEANUP;
3153 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3155 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3157 return io_prep_rw(req, sqe);
3161 * This is our waitqueue callback handler, registered through lock_page_async()
3162 * when we initially tried to do the IO with the iocb armed our waitqueue.
3163 * This gets called when the page is unlocked, and we generally expect that to
3164 * happen when the page IO is completed and the page is now uptodate. This will
3165 * queue a task_work based retry of the operation, attempting to copy the data
3166 * again. If the latter fails because the page was NOT uptodate, then we will
3167 * do a thread based blocking retry of the operation. That's the unexpected
3170 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3171 int sync, void *arg)
3173 struct wait_page_queue *wpq;
3174 struct io_kiocb *req = wait->private;
3175 struct wait_page_key *key = arg;
3177 wpq = container_of(wait, struct wait_page_queue, wait);
3179 if (!wake_page_match(wpq, key))
3182 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3183 list_del_init(&wait->entry);
3185 /* submit ref gets dropped, acquire a new one */
3187 io_req_task_queue(req);
3192 * This controls whether a given IO request should be armed for async page
3193 * based retry. If we return false here, the request is handed to the async
3194 * worker threads for retry. If we're doing buffered reads on a regular file,
3195 * we prepare a private wait_page_queue entry and retry the operation. This
3196 * will either succeed because the page is now uptodate and unlocked, or it
3197 * will register a callback when the page is unlocked at IO completion. Through
3198 * that callback, io_uring uses task_work to setup a retry of the operation.
3199 * That retry will attempt the buffered read again. The retry will generally
3200 * succeed, or in rare cases where it fails, we then fall back to using the
3201 * async worker threads for a blocking retry.
3203 static bool io_rw_should_retry(struct io_kiocb *req)
3205 struct io_async_rw *rw = req->async_data;
3206 struct wait_page_queue *wait = &rw->wpq;
3207 struct kiocb *kiocb = &req->rw.kiocb;
3209 /* never retry for NOWAIT, we just complete with -EAGAIN */
3210 if (req->flags & REQ_F_NOWAIT)
3213 /* Only for buffered IO */
3214 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3218 * just use poll if we can, and don't attempt if the fs doesn't
3219 * support callback based unlocks
3221 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3224 wait->wait.func = io_async_buf_func;
3225 wait->wait.private = req;
3226 wait->wait.flags = 0;
3227 INIT_LIST_HEAD(&wait->wait.entry);
3228 kiocb->ki_flags |= IOCB_WAITQ;
3229 kiocb->ki_flags &= ~IOCB_NOWAIT;
3230 kiocb->ki_waitq = wait;
3234 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3236 if (req->file->f_op->read_iter)
3237 return call_read_iter(req->file, &req->rw.kiocb, iter);
3238 else if (req->file->f_op->read)
3239 return loop_rw_iter(READ, req, iter);
3244 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3246 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3247 struct kiocb *kiocb = &req->rw.kiocb;
3248 struct iov_iter __iter, *iter = &__iter;
3249 struct io_async_rw *rw = req->async_data;
3250 ssize_t io_size, ret, ret2;
3251 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3257 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3261 io_size = iov_iter_count(iter);
3262 req->result = io_size;
3264 /* Ensure we clear previously set non-block flag */
3265 if (!force_nonblock)
3266 kiocb->ki_flags &= ~IOCB_NOWAIT;
3268 kiocb->ki_flags |= IOCB_NOWAIT;
3270 /* If the file doesn't support async, just async punt */
3271 if (force_nonblock && !io_file_supports_async(req, READ)) {
3272 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3273 return ret ?: -EAGAIN;
3276 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3277 if (unlikely(ret)) {
3282 ret = io_iter_do_read(req, iter);
3284 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3285 req->flags &= ~REQ_F_REISSUE;
3286 /* IOPOLL retry should happen for io-wq threads */
3287 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3289 /* no retry on NONBLOCK nor RWF_NOWAIT */
3290 if (req->flags & REQ_F_NOWAIT)
3292 /* some cases will consume bytes even on error returns */
3293 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3295 } else if (ret == -EIOCBQUEUED) {
3297 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3298 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3299 /* read all, failed, already did sync or don't want to retry */
3303 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3308 rw = req->async_data;
3309 /* now use our persistent iterator, if we aren't already */
3314 rw->bytes_done += ret;
3315 /* if we can retry, do so with the callbacks armed */
3316 if (!io_rw_should_retry(req)) {
3317 kiocb->ki_flags &= ~IOCB_WAITQ;
3322 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3323 * we get -EIOCBQUEUED, then we'll get a notification when the
3324 * desired page gets unlocked. We can also get a partial read
3325 * here, and if we do, then just retry at the new offset.
3327 ret = io_iter_do_read(req, iter);
3328 if (ret == -EIOCBQUEUED)
3330 /* we got some bytes, but not all. retry. */
3331 kiocb->ki_flags &= ~IOCB_WAITQ;
3332 } while (ret > 0 && ret < io_size);
3334 kiocb_done(kiocb, ret, issue_flags);
3336 /* it's faster to check here then delegate to kfree */
3342 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3344 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3346 return io_prep_rw(req, sqe);
3349 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3351 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3352 struct kiocb *kiocb = &req->rw.kiocb;
3353 struct iov_iter __iter, *iter = &__iter;
3354 struct io_async_rw *rw = req->async_data;
3355 ssize_t ret, ret2, io_size;
3356 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3362 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3366 io_size = iov_iter_count(iter);
3367 req->result = io_size;
3369 /* Ensure we clear previously set non-block flag */
3370 if (!force_nonblock)
3371 kiocb->ki_flags &= ~IOCB_NOWAIT;
3373 kiocb->ki_flags |= IOCB_NOWAIT;
3375 /* If the file doesn't support async, just async punt */
3376 if (force_nonblock && !io_file_supports_async(req, WRITE))
3379 /* file path doesn't support NOWAIT for non-direct_IO */
3380 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3381 (req->flags & REQ_F_ISREG))
3384 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3389 * Open-code file_start_write here to grab freeze protection,
3390 * which will be released by another thread in
3391 * io_complete_rw(). Fool lockdep by telling it the lock got
3392 * released so that it doesn't complain about the held lock when
3393 * we return to userspace.
3395 if (req->flags & REQ_F_ISREG) {
3396 sb_start_write(file_inode(req->file)->i_sb);
3397 __sb_writers_release(file_inode(req->file)->i_sb,
3400 kiocb->ki_flags |= IOCB_WRITE;
3402 if (req->file->f_op->write_iter)
3403 ret2 = call_write_iter(req->file, kiocb, iter);
3404 else if (req->file->f_op->write)
3405 ret2 = loop_rw_iter(WRITE, req, iter);
3409 if (req->flags & REQ_F_REISSUE) {
3410 req->flags &= ~REQ_F_REISSUE;
3415 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3416 * retry them without IOCB_NOWAIT.
3418 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3420 /* no retry on NONBLOCK nor RWF_NOWAIT */
3421 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3423 if (!force_nonblock || ret2 != -EAGAIN) {
3424 /* IOPOLL retry should happen for io-wq threads */
3425 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3428 kiocb_done(kiocb, ret2, issue_flags);
3431 /* some cases will consume bytes even on error returns */
3432 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3433 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3434 return ret ?: -EAGAIN;
3437 /* it's reportedly faster than delegating the null check to kfree() */
3443 static int io_renameat_prep(struct io_kiocb *req,
3444 const struct io_uring_sqe *sqe)
3446 struct io_rename *ren = &req->rename;
3447 const char __user *oldf, *newf;
3449 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3452 ren->old_dfd = READ_ONCE(sqe->fd);
3453 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3454 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3455 ren->new_dfd = READ_ONCE(sqe->len);
3456 ren->flags = READ_ONCE(sqe->rename_flags);
3458 ren->oldpath = getname(oldf);
3459 if (IS_ERR(ren->oldpath))
3460 return PTR_ERR(ren->oldpath);
3462 ren->newpath = getname(newf);
3463 if (IS_ERR(ren->newpath)) {
3464 putname(ren->oldpath);
3465 return PTR_ERR(ren->newpath);
3468 req->flags |= REQ_F_NEED_CLEANUP;
3472 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3474 struct io_rename *ren = &req->rename;
3477 if (issue_flags & IO_URING_F_NONBLOCK)
3480 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3481 ren->newpath, ren->flags);
3483 req->flags &= ~REQ_F_NEED_CLEANUP;
3485 req_set_fail_links(req);
3486 io_req_complete(req, ret);
3490 static int io_unlinkat_prep(struct io_kiocb *req,
3491 const struct io_uring_sqe *sqe)
3493 struct io_unlink *un = &req->unlink;
3494 const char __user *fname;
3496 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3499 un->dfd = READ_ONCE(sqe->fd);
3501 un->flags = READ_ONCE(sqe->unlink_flags);
3502 if (un->flags & ~AT_REMOVEDIR)
3505 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3506 un->filename = getname(fname);
3507 if (IS_ERR(un->filename))
3508 return PTR_ERR(un->filename);
3510 req->flags |= REQ_F_NEED_CLEANUP;
3514 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3516 struct io_unlink *un = &req->unlink;
3519 if (issue_flags & IO_URING_F_NONBLOCK)
3522 if (un->flags & AT_REMOVEDIR)
3523 ret = do_rmdir(un->dfd, un->filename);
3525 ret = do_unlinkat(un->dfd, un->filename);
3527 req->flags &= ~REQ_F_NEED_CLEANUP;
3529 req_set_fail_links(req);
3530 io_req_complete(req, ret);
3534 static int io_shutdown_prep(struct io_kiocb *req,
3535 const struct io_uring_sqe *sqe)
3537 #if defined(CONFIG_NET)
3538 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3540 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3544 req->shutdown.how = READ_ONCE(sqe->len);
3551 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3553 #if defined(CONFIG_NET)
3554 struct socket *sock;
3557 if (issue_flags & IO_URING_F_NONBLOCK)
3560 sock = sock_from_file(req->file);
3561 if (unlikely(!sock))
3564 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3566 req_set_fail_links(req);
3567 io_req_complete(req, ret);
3574 static int __io_splice_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 struct io_splice* sp = &req->splice;
3578 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3584 sp->len = READ_ONCE(sqe->len);
3585 sp->flags = READ_ONCE(sqe->splice_flags);
3587 if (unlikely(sp->flags & ~valid_flags))
3590 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3591 (sp->flags & SPLICE_F_FD_IN_FIXED));
3594 req->flags |= REQ_F_NEED_CLEANUP;
3598 static int io_tee_prep(struct io_kiocb *req,
3599 const struct io_uring_sqe *sqe)
3601 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3603 return __io_splice_prep(req, sqe);
3606 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3608 struct io_splice *sp = &req->splice;
3609 struct file *in = sp->file_in;
3610 struct file *out = sp->file_out;
3611 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3614 if (issue_flags & IO_URING_F_NONBLOCK)
3617 ret = do_tee(in, out, sp->len, flags);
3619 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3621 req->flags &= ~REQ_F_NEED_CLEANUP;
3624 req_set_fail_links(req);
3625 io_req_complete(req, ret);
3629 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3631 struct io_splice* sp = &req->splice;
3633 sp->off_in = READ_ONCE(sqe->splice_off_in);
3634 sp->off_out = READ_ONCE(sqe->off);
3635 return __io_splice_prep(req, sqe);
3638 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3640 struct io_splice *sp = &req->splice;
3641 struct file *in = sp->file_in;
3642 struct file *out = sp->file_out;
3643 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3644 loff_t *poff_in, *poff_out;
3647 if (issue_flags & IO_URING_F_NONBLOCK)
3650 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3651 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3654 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3656 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3658 req->flags &= ~REQ_F_NEED_CLEANUP;
3661 req_set_fail_links(req);
3662 io_req_complete(req, ret);
3667 * IORING_OP_NOP just posts a completion event, nothing else.
3669 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3671 struct io_ring_ctx *ctx = req->ctx;
3673 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3676 __io_req_complete(req, issue_flags, 0, 0);
3680 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3682 struct io_ring_ctx *ctx = req->ctx;
3687 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3689 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3692 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3693 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3696 req->sync.off = READ_ONCE(sqe->off);
3697 req->sync.len = READ_ONCE(sqe->len);
3701 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3703 loff_t end = req->sync.off + req->sync.len;
3706 /* fsync always requires a blocking context */
3707 if (issue_flags & IO_URING_F_NONBLOCK)
3710 ret = vfs_fsync_range(req->file, req->sync.off,
3711 end > 0 ? end : LLONG_MAX,
3712 req->sync.flags & IORING_FSYNC_DATASYNC);
3714 req_set_fail_links(req);
3715 io_req_complete(req, ret);
3719 static int io_fallocate_prep(struct io_kiocb *req,
3720 const struct io_uring_sqe *sqe)
3722 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3724 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3727 req->sync.off = READ_ONCE(sqe->off);
3728 req->sync.len = READ_ONCE(sqe->addr);
3729 req->sync.mode = READ_ONCE(sqe->len);
3733 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3737 /* fallocate always requiring blocking context */
3738 if (issue_flags & IO_URING_F_NONBLOCK)
3740 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3743 req_set_fail_links(req);
3744 io_req_complete(req, ret);
3748 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3750 const char __user *fname;
3753 if (unlikely(sqe->ioprio || sqe->buf_index))
3755 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3758 /* open.how should be already initialised */
3759 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3760 req->open.how.flags |= O_LARGEFILE;
3762 req->open.dfd = READ_ONCE(sqe->fd);
3763 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3764 req->open.filename = getname(fname);
3765 if (IS_ERR(req->open.filename)) {
3766 ret = PTR_ERR(req->open.filename);
3767 req->open.filename = NULL;
3770 req->open.nofile = rlimit(RLIMIT_NOFILE);
3771 req->flags |= REQ_F_NEED_CLEANUP;
3775 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3781 mode = READ_ONCE(sqe->len);
3782 flags = READ_ONCE(sqe->open_flags);
3783 req->open.how = build_open_how(flags, mode);
3784 return __io_openat_prep(req, sqe);
3787 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3789 struct open_how __user *how;
3793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3795 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3796 len = READ_ONCE(sqe->len);
3797 if (len < OPEN_HOW_SIZE_VER0)
3800 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3805 return __io_openat_prep(req, sqe);
3808 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3810 struct open_flags op;
3813 bool resolve_nonblock;
3816 ret = build_open_flags(&req->open.how, &op);
3819 nonblock_set = op.open_flag & O_NONBLOCK;
3820 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3821 if (issue_flags & IO_URING_F_NONBLOCK) {
3823 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3824 * it'll always -EAGAIN
3826 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3828 op.lookup_flags |= LOOKUP_CACHED;
3829 op.open_flag |= O_NONBLOCK;
3832 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3836 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3837 /* only retry if RESOLVE_CACHED wasn't already set by application */
3838 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3839 file == ERR_PTR(-EAGAIN)) {
3841 * We could hang on to this 'fd', but seems like marginal
3842 * gain for something that is now known to be a slower path.
3843 * So just put it, and we'll get a new one when we retry.
3851 ret = PTR_ERR(file);
3853 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3854 file->f_flags &= ~O_NONBLOCK;
3855 fsnotify_open(file);
3856 fd_install(ret, file);
3859 putname(req->open.filename);
3860 req->flags &= ~REQ_F_NEED_CLEANUP;
3862 req_set_fail_links(req);
3863 io_req_complete(req, ret);
3867 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3869 return io_openat2(req, issue_flags);
3872 static int io_remove_buffers_prep(struct io_kiocb *req,
3873 const struct io_uring_sqe *sqe)
3875 struct io_provide_buf *p = &req->pbuf;
3878 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3881 tmp = READ_ONCE(sqe->fd);
3882 if (!tmp || tmp > USHRT_MAX)
3885 memset(p, 0, sizeof(*p));
3887 p->bgid = READ_ONCE(sqe->buf_group);
3891 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3892 int bgid, unsigned nbufs)
3896 /* shouldn't happen */
3900 /* the head kbuf is the list itself */
3901 while (!list_empty(&buf->list)) {
3902 struct io_buffer *nxt;
3904 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3905 list_del(&nxt->list);
3912 xa_erase(&ctx->io_buffers, bgid);
3917 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3919 struct io_provide_buf *p = &req->pbuf;
3920 struct io_ring_ctx *ctx = req->ctx;
3921 struct io_buffer *head;
3923 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3925 io_ring_submit_lock(ctx, !force_nonblock);
3927 lockdep_assert_held(&ctx->uring_lock);
3930 head = xa_load(&ctx->io_buffers, p->bgid);
3932 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3934 req_set_fail_links(req);
3936 /* complete before unlock, IOPOLL may need the lock */
3937 __io_req_complete(req, issue_flags, ret, 0);
3938 io_ring_submit_unlock(ctx, !force_nonblock);
3942 static int io_provide_buffers_prep(struct io_kiocb *req,
3943 const struct io_uring_sqe *sqe)
3946 struct io_provide_buf *p = &req->pbuf;
3949 if (sqe->ioprio || sqe->rw_flags)
3952 tmp = READ_ONCE(sqe->fd);
3953 if (!tmp || tmp > USHRT_MAX)
3956 p->addr = READ_ONCE(sqe->addr);
3957 p->len = READ_ONCE(sqe->len);
3959 size = (unsigned long)p->len * p->nbufs;
3960 if (!access_ok(u64_to_user_ptr(p->addr), size))
3963 p->bgid = READ_ONCE(sqe->buf_group);
3964 tmp = READ_ONCE(sqe->off);
3965 if (tmp > USHRT_MAX)
3971 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3973 struct io_buffer *buf;
3974 u64 addr = pbuf->addr;
3975 int i, bid = pbuf->bid;
3977 for (i = 0; i < pbuf->nbufs; i++) {
3978 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3983 buf->len = pbuf->len;
3988 INIT_LIST_HEAD(&buf->list);
3991 list_add_tail(&buf->list, &(*head)->list);
3995 return i ? i : -ENOMEM;
3998 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4000 struct io_provide_buf *p = &req->pbuf;
4001 struct io_ring_ctx *ctx = req->ctx;
4002 struct io_buffer *head, *list;
4004 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4006 io_ring_submit_lock(ctx, !force_nonblock);
4008 lockdep_assert_held(&ctx->uring_lock);
4010 list = head = xa_load(&ctx->io_buffers, p->bgid);
4012 ret = io_add_buffers(p, &head);
4013 if (ret >= 0 && !list) {
4014 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4016 __io_remove_buffers(ctx, head, p->bgid, -1U);
4019 req_set_fail_links(req);
4020 /* complete before unlock, IOPOLL may need the lock */
4021 __io_req_complete(req, issue_flags, ret, 0);
4022 io_ring_submit_unlock(ctx, !force_nonblock);
4026 static int io_epoll_ctl_prep(struct io_kiocb *req,
4027 const struct io_uring_sqe *sqe)
4029 #if defined(CONFIG_EPOLL)
4030 if (sqe->ioprio || sqe->buf_index)
4032 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4035 req->epoll.epfd = READ_ONCE(sqe->fd);
4036 req->epoll.op = READ_ONCE(sqe->len);
4037 req->epoll.fd = READ_ONCE(sqe->off);
4039 if (ep_op_has_event(req->epoll.op)) {
4040 struct epoll_event __user *ev;
4042 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4043 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4053 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4055 #if defined(CONFIG_EPOLL)
4056 struct io_epoll *ie = &req->epoll;
4058 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4060 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4061 if (force_nonblock && ret == -EAGAIN)
4065 req_set_fail_links(req);
4066 __io_req_complete(req, issue_flags, ret, 0);
4073 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4075 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4076 if (sqe->ioprio || sqe->buf_index || sqe->off)
4078 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4081 req->madvise.addr = READ_ONCE(sqe->addr);
4082 req->madvise.len = READ_ONCE(sqe->len);
4083 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4090 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4092 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4093 struct io_madvise *ma = &req->madvise;
4096 if (issue_flags & IO_URING_F_NONBLOCK)
4099 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4101 req_set_fail_links(req);
4102 io_req_complete(req, ret);
4109 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4111 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4113 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4116 req->fadvise.offset = READ_ONCE(sqe->off);
4117 req->fadvise.len = READ_ONCE(sqe->len);
4118 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4122 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4124 struct io_fadvise *fa = &req->fadvise;
4127 if (issue_flags & IO_URING_F_NONBLOCK) {
4128 switch (fa->advice) {
4129 case POSIX_FADV_NORMAL:
4130 case POSIX_FADV_RANDOM:
4131 case POSIX_FADV_SEQUENTIAL:
4138 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4140 req_set_fail_links(req);
4141 io_req_complete(req, ret);
4145 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4147 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4149 if (sqe->ioprio || sqe->buf_index)
4151 if (req->flags & REQ_F_FIXED_FILE)
4154 req->statx.dfd = READ_ONCE(sqe->fd);
4155 req->statx.mask = READ_ONCE(sqe->len);
4156 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4157 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4158 req->statx.flags = READ_ONCE(sqe->statx_flags);
4163 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4165 struct io_statx *ctx = &req->statx;
4168 if (issue_flags & IO_URING_F_NONBLOCK)
4171 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4175 req_set_fail_links(req);
4176 io_req_complete(req, ret);
4180 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4182 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4184 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4185 sqe->rw_flags || sqe->buf_index)
4187 if (req->flags & REQ_F_FIXED_FILE)
4190 req->close.fd = READ_ONCE(sqe->fd);
4194 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4196 struct files_struct *files = current->files;
4197 struct io_close *close = &req->close;
4198 struct fdtable *fdt;
4204 spin_lock(&files->file_lock);
4205 fdt = files_fdtable(files);
4206 if (close->fd >= fdt->max_fds) {
4207 spin_unlock(&files->file_lock);
4210 file = fdt->fd[close->fd];
4212 spin_unlock(&files->file_lock);
4216 if (file->f_op == &io_uring_fops) {
4217 spin_unlock(&files->file_lock);
4222 /* if the file has a flush method, be safe and punt to async */
4223 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4224 spin_unlock(&files->file_lock);
4228 ret = __close_fd_get_file(close->fd, &file);
4229 spin_unlock(&files->file_lock);
4236 /* No ->flush() or already async, safely close from here */
4237 ret = filp_close(file, current->files);
4240 req_set_fail_links(req);
4243 __io_req_complete(req, issue_flags, ret, 0);
4247 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4249 struct io_ring_ctx *ctx = req->ctx;
4251 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4253 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4256 req->sync.off = READ_ONCE(sqe->off);
4257 req->sync.len = READ_ONCE(sqe->len);
4258 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4262 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4266 /* sync_file_range always requires a blocking context */
4267 if (issue_flags & IO_URING_F_NONBLOCK)
4270 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4273 req_set_fail_links(req);
4274 io_req_complete(req, ret);
4278 #if defined(CONFIG_NET)
4279 static int io_setup_async_msg(struct io_kiocb *req,
4280 struct io_async_msghdr *kmsg)
4282 struct io_async_msghdr *async_msg = req->async_data;
4286 if (io_alloc_async_data(req)) {
4287 kfree(kmsg->free_iov);
4290 async_msg = req->async_data;
4291 req->flags |= REQ_F_NEED_CLEANUP;
4292 memcpy(async_msg, kmsg, sizeof(*kmsg));
4293 async_msg->msg.msg_name = &async_msg->addr;
4294 /* if were using fast_iov, set it to the new one */
4295 if (!async_msg->free_iov)
4296 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4301 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4302 struct io_async_msghdr *iomsg)
4304 iomsg->msg.msg_name = &iomsg->addr;
4305 iomsg->free_iov = iomsg->fast_iov;
4306 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4307 req->sr_msg.msg_flags, &iomsg->free_iov);
4310 static int io_sendmsg_prep_async(struct io_kiocb *req)
4314 ret = io_sendmsg_copy_hdr(req, req->async_data);
4316 req->flags |= REQ_F_NEED_CLEANUP;
4320 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4322 struct io_sr_msg *sr = &req->sr_msg;
4324 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4327 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4328 sr->len = READ_ONCE(sqe->len);
4329 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4330 if (sr->msg_flags & MSG_DONTWAIT)
4331 req->flags |= REQ_F_NOWAIT;
4333 #ifdef CONFIG_COMPAT
4334 if (req->ctx->compat)
4335 sr->msg_flags |= MSG_CMSG_COMPAT;
4340 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4342 struct io_async_msghdr iomsg, *kmsg;
4343 struct socket *sock;
4348 sock = sock_from_file(req->file);
4349 if (unlikely(!sock))
4352 kmsg = req->async_data;
4354 ret = io_sendmsg_copy_hdr(req, &iomsg);
4360 flags = req->sr_msg.msg_flags;
4361 if (issue_flags & IO_URING_F_NONBLOCK)
4362 flags |= MSG_DONTWAIT;
4363 if (flags & MSG_WAITALL)
4364 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4366 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4367 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4368 return io_setup_async_msg(req, kmsg);
4369 if (ret == -ERESTARTSYS)
4372 /* fast path, check for non-NULL to avoid function call */
4374 kfree(kmsg->free_iov);
4375 req->flags &= ~REQ_F_NEED_CLEANUP;
4377 req_set_fail_links(req);
4378 __io_req_complete(req, issue_flags, ret, 0);
4382 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4384 struct io_sr_msg *sr = &req->sr_msg;
4387 struct socket *sock;
4392 sock = sock_from_file(req->file);
4393 if (unlikely(!sock))
4396 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4400 msg.msg_name = NULL;
4401 msg.msg_control = NULL;
4402 msg.msg_controllen = 0;
4403 msg.msg_namelen = 0;
4405 flags = req->sr_msg.msg_flags;
4406 if (issue_flags & IO_URING_F_NONBLOCK)
4407 flags |= MSG_DONTWAIT;
4408 if (flags & MSG_WAITALL)
4409 min_ret = iov_iter_count(&msg.msg_iter);
4411 msg.msg_flags = flags;
4412 ret = sock_sendmsg(sock, &msg);
4413 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4415 if (ret == -ERESTARTSYS)
4419 req_set_fail_links(req);
4420 __io_req_complete(req, issue_flags, ret, 0);
4424 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4425 struct io_async_msghdr *iomsg)
4427 struct io_sr_msg *sr = &req->sr_msg;
4428 struct iovec __user *uiov;
4432 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4433 &iomsg->uaddr, &uiov, &iov_len);
4437 if (req->flags & REQ_F_BUFFER_SELECT) {
4440 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4442 sr->len = iomsg->fast_iov[0].iov_len;
4443 iomsg->free_iov = NULL;
4445 iomsg->free_iov = iomsg->fast_iov;
4446 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4447 &iomsg->free_iov, &iomsg->msg.msg_iter,
4456 #ifdef CONFIG_COMPAT
4457 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4458 struct io_async_msghdr *iomsg)
4460 struct compat_msghdr __user *msg_compat;
4461 struct io_sr_msg *sr = &req->sr_msg;
4462 struct compat_iovec __user *uiov;
4467 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4468 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4473 uiov = compat_ptr(ptr);
4474 if (req->flags & REQ_F_BUFFER_SELECT) {
4475 compat_ssize_t clen;
4479 if (!access_ok(uiov, sizeof(*uiov)))
4481 if (__get_user(clen, &uiov->iov_len))
4486 iomsg->free_iov = NULL;
4488 iomsg->free_iov = iomsg->fast_iov;
4489 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4490 UIO_FASTIOV, &iomsg->free_iov,
4491 &iomsg->msg.msg_iter, true);
4500 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4501 struct io_async_msghdr *iomsg)
4503 iomsg->msg.msg_name = &iomsg->addr;
4505 #ifdef CONFIG_COMPAT
4506 if (req->ctx->compat)
4507 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4510 return __io_recvmsg_copy_hdr(req, iomsg);
4513 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4516 struct io_sr_msg *sr = &req->sr_msg;
4517 struct io_buffer *kbuf;
4519 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4524 req->flags |= REQ_F_BUFFER_SELECTED;
4528 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4530 return io_put_kbuf(req, req->sr_msg.kbuf);
4533 static int io_recvmsg_prep_async(struct io_kiocb *req)
4537 ret = io_recvmsg_copy_hdr(req, req->async_data);
4539 req->flags |= REQ_F_NEED_CLEANUP;
4543 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4545 struct io_sr_msg *sr = &req->sr_msg;
4547 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4550 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4551 sr->len = READ_ONCE(sqe->len);
4552 sr->bgid = READ_ONCE(sqe->buf_group);
4553 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4554 if (sr->msg_flags & MSG_DONTWAIT)
4555 req->flags |= REQ_F_NOWAIT;
4557 #ifdef CONFIG_COMPAT
4558 if (req->ctx->compat)
4559 sr->msg_flags |= MSG_CMSG_COMPAT;
4564 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4566 struct io_async_msghdr iomsg, *kmsg;
4567 struct socket *sock;
4568 struct io_buffer *kbuf;
4571 int ret, cflags = 0;
4572 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4574 sock = sock_from_file(req->file);
4575 if (unlikely(!sock))
4578 kmsg = req->async_data;
4580 ret = io_recvmsg_copy_hdr(req, &iomsg);
4586 if (req->flags & REQ_F_BUFFER_SELECT) {
4587 kbuf = io_recv_buffer_select(req, !force_nonblock);
4589 return PTR_ERR(kbuf);
4590 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4591 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4592 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4593 1, req->sr_msg.len);
4596 flags = req->sr_msg.msg_flags;
4598 flags |= MSG_DONTWAIT;
4599 if (flags & MSG_WAITALL)
4600 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4602 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4603 kmsg->uaddr, flags);
4604 if (force_nonblock && ret == -EAGAIN)
4605 return io_setup_async_msg(req, kmsg);
4606 if (ret == -ERESTARTSYS)
4609 if (req->flags & REQ_F_BUFFER_SELECTED)
4610 cflags = io_put_recv_kbuf(req);
4611 /* fast path, check for non-NULL to avoid function call */
4613 kfree(kmsg->free_iov);
4614 req->flags &= ~REQ_F_NEED_CLEANUP;
4615 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4616 req_set_fail_links(req);
4617 __io_req_complete(req, issue_flags, ret, cflags);
4621 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4623 struct io_buffer *kbuf;
4624 struct io_sr_msg *sr = &req->sr_msg;
4626 void __user *buf = sr->buf;
4627 struct socket *sock;
4631 int ret, cflags = 0;
4632 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4634 sock = sock_from_file(req->file);
4635 if (unlikely(!sock))
4638 if (req->flags & REQ_F_BUFFER_SELECT) {
4639 kbuf = io_recv_buffer_select(req, !force_nonblock);
4641 return PTR_ERR(kbuf);
4642 buf = u64_to_user_ptr(kbuf->addr);
4645 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4649 msg.msg_name = NULL;
4650 msg.msg_control = NULL;
4651 msg.msg_controllen = 0;
4652 msg.msg_namelen = 0;
4653 msg.msg_iocb = NULL;
4656 flags = req->sr_msg.msg_flags;
4658 flags |= MSG_DONTWAIT;
4659 if (flags & MSG_WAITALL)
4660 min_ret = iov_iter_count(&msg.msg_iter);
4662 ret = sock_recvmsg(sock, &msg, flags);
4663 if (force_nonblock && ret == -EAGAIN)
4665 if (ret == -ERESTARTSYS)
4668 if (req->flags & REQ_F_BUFFER_SELECTED)
4669 cflags = io_put_recv_kbuf(req);
4670 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4671 req_set_fail_links(req);
4672 __io_req_complete(req, issue_flags, ret, cflags);
4676 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4678 struct io_accept *accept = &req->accept;
4680 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4682 if (sqe->ioprio || sqe->len || sqe->buf_index)
4685 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4686 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4687 accept->flags = READ_ONCE(sqe->accept_flags);
4688 accept->nofile = rlimit(RLIMIT_NOFILE);
4692 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4694 struct io_accept *accept = &req->accept;
4695 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4696 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4699 if (req->file->f_flags & O_NONBLOCK)
4700 req->flags |= REQ_F_NOWAIT;
4702 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4703 accept->addr_len, accept->flags,
4705 if (ret == -EAGAIN && force_nonblock)
4708 if (ret == -ERESTARTSYS)
4710 req_set_fail_links(req);
4712 __io_req_complete(req, issue_flags, ret, 0);
4716 static int io_connect_prep_async(struct io_kiocb *req)
4718 struct io_async_connect *io = req->async_data;
4719 struct io_connect *conn = &req->connect;
4721 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4724 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4726 struct io_connect *conn = &req->connect;
4728 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4730 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4733 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4734 conn->addr_len = READ_ONCE(sqe->addr2);
4738 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4740 struct io_async_connect __io, *io;
4741 unsigned file_flags;
4743 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4745 if (req->async_data) {
4746 io = req->async_data;
4748 ret = move_addr_to_kernel(req->connect.addr,
4749 req->connect.addr_len,
4756 file_flags = force_nonblock ? O_NONBLOCK : 0;
4758 ret = __sys_connect_file(req->file, &io->address,
4759 req->connect.addr_len, file_flags);
4760 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4761 if (req->async_data)
4763 if (io_alloc_async_data(req)) {
4767 memcpy(req->async_data, &__io, sizeof(__io));
4770 if (ret == -ERESTARTSYS)
4774 req_set_fail_links(req);
4775 __io_req_complete(req, issue_flags, ret, 0);
4778 #else /* !CONFIG_NET */
4779 #define IO_NETOP_FN(op) \
4780 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4782 return -EOPNOTSUPP; \
4785 #define IO_NETOP_PREP(op) \
4787 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4789 return -EOPNOTSUPP; \
4792 #define IO_NETOP_PREP_ASYNC(op) \
4794 static int io_##op##_prep_async(struct io_kiocb *req) \
4796 return -EOPNOTSUPP; \
4799 IO_NETOP_PREP_ASYNC(sendmsg);
4800 IO_NETOP_PREP_ASYNC(recvmsg);
4801 IO_NETOP_PREP_ASYNC(connect);
4802 IO_NETOP_PREP(accept);
4805 #endif /* CONFIG_NET */
4807 struct io_poll_table {
4808 struct poll_table_struct pt;
4809 struct io_kiocb *req;
4813 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4814 __poll_t mask, task_work_func_t func)
4818 /* for instances that support it check for an event match first: */
4819 if (mask && !(mask & poll->events))
4822 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4824 list_del_init(&poll->wait.entry);
4827 req->task_work.func = func;
4830 * If this fails, then the task is exiting. When a task exits, the
4831 * work gets canceled, so just cancel this request as well instead
4832 * of executing it. We can't safely execute it anyway, as we may not
4833 * have the needed state needed for it anyway.
4835 ret = io_req_task_work_add(req);
4836 if (unlikely(ret)) {
4837 WRITE_ONCE(poll->canceled, true);
4838 io_req_task_work_add_fallback(req, func);
4843 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4844 __acquires(&req->ctx->completion_lock)
4846 struct io_ring_ctx *ctx = req->ctx;
4848 if (!req->result && !READ_ONCE(poll->canceled)) {
4849 struct poll_table_struct pt = { ._key = poll->events };
4851 req->result = vfs_poll(req->file, &pt) & poll->events;
4854 spin_lock_irq(&ctx->completion_lock);
4855 if (!req->result && !READ_ONCE(poll->canceled)) {
4856 add_wait_queue(poll->head, &poll->wait);
4863 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4865 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4866 if (req->opcode == IORING_OP_POLL_ADD)
4867 return req->async_data;
4868 return req->apoll->double_poll;
4871 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4873 if (req->opcode == IORING_OP_POLL_ADD)
4875 return &req->apoll->poll;
4878 static void io_poll_remove_double(struct io_kiocb *req)
4879 __must_hold(&req->ctx->completion_lock)
4881 struct io_poll_iocb *poll = io_poll_get_double(req);
4883 lockdep_assert_held(&req->ctx->completion_lock);
4885 if (poll && poll->head) {
4886 struct wait_queue_head *head = poll->head;
4888 spin_lock(&head->lock);
4889 list_del_init(&poll->wait.entry);
4890 if (poll->wait.private)
4893 spin_unlock(&head->lock);
4897 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4898 __must_hold(&req->ctx->completion_lock)
4900 struct io_ring_ctx *ctx = req->ctx;
4901 unsigned flags = IORING_CQE_F_MORE;
4904 if (READ_ONCE(req->poll.canceled)) {
4906 req->poll.events |= EPOLLONESHOT;
4908 error = mangle_poll(mask);
4910 if (req->poll.events & EPOLLONESHOT)
4912 if (!__io_cqring_fill_event(req, error, flags)) {
4913 io_poll_remove_waitqs(req);
4914 req->poll.done = true;
4917 io_commit_cqring(ctx);
4918 return !(flags & IORING_CQE_F_MORE);
4921 static void io_poll_task_func(struct callback_head *cb)
4923 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4924 struct io_ring_ctx *ctx = req->ctx;
4925 struct io_kiocb *nxt;
4927 if (io_poll_rewait(req, &req->poll)) {
4928 spin_unlock_irq(&ctx->completion_lock);
4932 done = io_poll_complete(req, req->result);
4934 hash_del(&req->hash_node);
4937 add_wait_queue(req->poll.head, &req->poll.wait);
4939 spin_unlock_irq(&ctx->completion_lock);
4940 io_cqring_ev_posted(ctx);
4943 nxt = io_put_req_find_next(req);
4945 __io_req_task_submit(nxt);
4950 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4951 int sync, void *key)
4953 struct io_kiocb *req = wait->private;
4954 struct io_poll_iocb *poll = io_poll_get_single(req);
4955 __poll_t mask = key_to_poll(key);
4957 /* for instances that support it check for an event match first: */
4958 if (mask && !(mask & poll->events))
4960 if (!(poll->events & EPOLLONESHOT))
4961 return poll->wait.func(&poll->wait, mode, sync, key);
4963 list_del_init(&wait->entry);
4965 if (poll && poll->head) {
4968 spin_lock(&poll->head->lock);
4969 done = list_empty(&poll->wait.entry);
4971 list_del_init(&poll->wait.entry);
4972 /* make sure double remove sees this as being gone */
4973 wait->private = NULL;
4974 spin_unlock(&poll->head->lock);
4976 /* use wait func handler, so it matches the rq type */
4977 poll->wait.func(&poll->wait, mode, sync, key);
4984 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4985 wait_queue_func_t wake_func)
4989 poll->canceled = false;
4990 poll->update_events = poll->update_user_data = false;
4991 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4992 /* mask in events that we always want/need */
4993 poll->events = events | IO_POLL_UNMASK;
4994 INIT_LIST_HEAD(&poll->wait.entry);
4995 init_waitqueue_func_entry(&poll->wait, wake_func);
4998 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4999 struct wait_queue_head *head,
5000 struct io_poll_iocb **poll_ptr)
5002 struct io_kiocb *req = pt->req;
5005 * If poll->head is already set, it's because the file being polled
5006 * uses multiple waitqueues for poll handling (eg one for read, one
5007 * for write). Setup a separate io_poll_iocb if this happens.
5009 if (unlikely(poll->head)) {
5010 struct io_poll_iocb *poll_one = poll;
5012 /* already have a 2nd entry, fail a third attempt */
5014 pt->error = -EINVAL;
5017 /* double add on the same waitqueue head, ignore */
5018 if (poll->head == head)
5020 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5022 pt->error = -ENOMEM;
5025 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5027 poll->wait.private = req;
5034 if (poll->events & EPOLLEXCLUSIVE)
5035 add_wait_queue_exclusive(head, &poll->wait);
5037 add_wait_queue(head, &poll->wait);
5040 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5041 struct poll_table_struct *p)
5043 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5044 struct async_poll *apoll = pt->req->apoll;
5046 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5049 static void io_async_task_func(struct callback_head *cb)
5051 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5052 struct async_poll *apoll = req->apoll;
5053 struct io_ring_ctx *ctx = req->ctx;
5055 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5057 if (io_poll_rewait(req, &apoll->poll)) {
5058 spin_unlock_irq(&ctx->completion_lock);
5062 /* If req is still hashed, it cannot have been canceled. Don't check. */
5063 if (hash_hashed(&req->hash_node))
5064 hash_del(&req->hash_node);
5066 io_poll_remove_double(req);
5067 spin_unlock_irq(&ctx->completion_lock);
5069 if (!READ_ONCE(apoll->poll.canceled))
5070 __io_req_task_submit(req);
5072 io_req_complete_failed(req, -ECANCELED);
5074 kfree(apoll->double_poll);
5078 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5081 struct io_kiocb *req = wait->private;
5082 struct io_poll_iocb *poll = &req->apoll->poll;
5084 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5087 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5090 static void io_poll_req_insert(struct io_kiocb *req)
5092 struct io_ring_ctx *ctx = req->ctx;
5093 struct hlist_head *list;
5095 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5096 hlist_add_head(&req->hash_node, list);
5099 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5100 struct io_poll_iocb *poll,
5101 struct io_poll_table *ipt, __poll_t mask,
5102 wait_queue_func_t wake_func)
5103 __acquires(&ctx->completion_lock)
5105 struct io_ring_ctx *ctx = req->ctx;
5106 bool cancel = false;
5108 INIT_HLIST_NODE(&req->hash_node);
5109 io_init_poll_iocb(poll, mask, wake_func);
5110 poll->file = req->file;
5111 poll->wait.private = req;
5113 ipt->pt._key = mask;
5115 ipt->error = -EINVAL;
5117 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5119 spin_lock_irq(&ctx->completion_lock);
5120 if (likely(poll->head)) {
5121 spin_lock(&poll->head->lock);
5122 if (unlikely(list_empty(&poll->wait.entry))) {
5128 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5129 list_del_init(&poll->wait.entry);
5131 WRITE_ONCE(poll->canceled, true);
5132 else if (!poll->done) /* actually waiting for an event */
5133 io_poll_req_insert(req);
5134 spin_unlock(&poll->head->lock);
5140 static bool io_arm_poll_handler(struct io_kiocb *req)
5142 const struct io_op_def *def = &io_op_defs[req->opcode];
5143 struct io_ring_ctx *ctx = req->ctx;
5144 struct async_poll *apoll;
5145 struct io_poll_table ipt;
5149 if (!req->file || !file_can_poll(req->file))
5151 if (req->flags & REQ_F_POLLED)
5155 else if (def->pollout)
5159 /* if we can't nonblock try, then no point in arming a poll handler */
5160 if (!io_file_supports_async(req, rw))
5163 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5164 if (unlikely(!apoll))
5166 apoll->double_poll = NULL;
5168 req->flags |= REQ_F_POLLED;
5171 mask = EPOLLONESHOT;
5173 mask |= POLLIN | POLLRDNORM;
5175 mask |= POLLOUT | POLLWRNORM;
5177 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5178 if ((req->opcode == IORING_OP_RECVMSG) &&
5179 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5182 mask |= POLLERR | POLLPRI;
5184 ipt.pt._qproc = io_async_queue_proc;
5186 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5188 if (ret || ipt.error) {
5189 io_poll_remove_double(req);
5190 spin_unlock_irq(&ctx->completion_lock);
5191 kfree(apoll->double_poll);
5195 spin_unlock_irq(&ctx->completion_lock);
5196 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5197 apoll->poll.events);
5201 static bool __io_poll_remove_one(struct io_kiocb *req,
5202 struct io_poll_iocb *poll, bool do_cancel)
5203 __must_hold(&req->ctx->completion_lock)
5205 bool do_complete = false;
5209 spin_lock(&poll->head->lock);
5211 WRITE_ONCE(poll->canceled, true);
5212 if (!list_empty(&poll->wait.entry)) {
5213 list_del_init(&poll->wait.entry);
5216 spin_unlock(&poll->head->lock);
5217 hash_del(&req->hash_node);
5221 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5222 __must_hold(&req->ctx->completion_lock)
5226 io_poll_remove_double(req);
5228 if (req->opcode == IORING_OP_POLL_ADD) {
5229 do_complete = __io_poll_remove_one(req, &req->poll, true);
5231 struct async_poll *apoll = req->apoll;
5233 /* non-poll requests have submit ref still */
5234 do_complete = __io_poll_remove_one(req, &apoll->poll, true);
5237 kfree(apoll->double_poll);
5245 static bool io_poll_remove_one(struct io_kiocb *req)
5246 __must_hold(&req->ctx->completion_lock)
5250 do_complete = io_poll_remove_waitqs(req);
5252 io_cqring_fill_event(req, -ECANCELED);
5253 io_commit_cqring(req->ctx);
5254 req_set_fail_links(req);
5255 io_put_req_deferred(req, 1);
5262 * Returns true if we found and killed one or more poll requests
5264 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5265 struct files_struct *files)
5267 struct hlist_node *tmp;
5268 struct io_kiocb *req;
5271 spin_lock_irq(&ctx->completion_lock);
5272 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5273 struct hlist_head *list;
5275 list = &ctx->cancel_hash[i];
5276 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5277 if (io_match_task(req, tsk, files))
5278 posted += io_poll_remove_one(req);
5281 spin_unlock_irq(&ctx->completion_lock);
5284 io_cqring_ev_posted(ctx);
5289 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr)
5290 __must_hold(&ctx->completion_lock)
5292 struct hlist_head *list;
5293 struct io_kiocb *req;
5295 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5296 hlist_for_each_entry(req, list, hash_node) {
5297 if (sqe_addr != req->user_data)
5305 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5306 __must_hold(&ctx->completion_lock)
5308 struct io_kiocb *req;
5310 req = io_poll_find(ctx, sqe_addr);
5313 if (io_poll_remove_one(req))
5319 static int io_poll_remove_prep(struct io_kiocb *req,
5320 const struct io_uring_sqe *sqe)
5322 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5324 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5328 req->poll_remove.addr = READ_ONCE(sqe->addr);
5333 * Find a running poll command that matches one specified in sqe->addr,
5334 * and remove it if found.
5336 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5338 struct io_ring_ctx *ctx = req->ctx;
5341 spin_lock_irq(&ctx->completion_lock);
5342 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5343 spin_unlock_irq(&ctx->completion_lock);
5346 req_set_fail_links(req);
5347 io_req_complete(req, ret);
5351 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5354 struct io_kiocb *req = wait->private;
5355 struct io_poll_iocb *poll = &req->poll;
5357 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5360 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5361 struct poll_table_struct *p)
5363 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5365 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5368 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5370 struct io_poll_iocb *poll = &req->poll;
5373 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5375 if (sqe->ioprio || sqe->buf_index)
5377 flags = READ_ONCE(sqe->len);
5378 if (flags & ~(IORING_POLL_ADD_MULTI | IORING_POLL_UPDATE_EVENTS |
5379 IORING_POLL_UPDATE_USER_DATA))
5381 events = READ_ONCE(sqe->poll32_events);
5383 events = swahw32(events);
5385 if (!(flags & IORING_POLL_ADD_MULTI))
5386 events |= EPOLLONESHOT;
5387 poll->update_events = poll->update_user_data = false;
5388 if (flags & IORING_POLL_UPDATE_EVENTS) {
5389 poll->update_events = true;
5390 poll->old_user_data = READ_ONCE(sqe->addr);
5392 if (flags & IORING_POLL_UPDATE_USER_DATA) {
5393 poll->update_user_data = true;
5394 poll->new_user_data = READ_ONCE(sqe->off);
5396 if (!(poll->update_events || poll->update_user_data) &&
5397 (sqe->off || sqe->addr))
5399 poll->events = demangle_poll(events) |
5400 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5404 static int __io_poll_add(struct io_kiocb *req)
5406 struct io_poll_iocb *poll = &req->poll;
5407 struct io_ring_ctx *ctx = req->ctx;
5408 struct io_poll_table ipt;
5411 ipt.pt._qproc = io_poll_queue_proc;
5413 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5416 if (mask) { /* no async, we'd stolen it */
5418 io_poll_complete(req, mask);
5420 spin_unlock_irq(&ctx->completion_lock);
5423 io_cqring_ev_posted(ctx);
5424 if (poll->events & EPOLLONESHOT)
5430 static int io_poll_update(struct io_kiocb *req)
5432 struct io_ring_ctx *ctx = req->ctx;
5433 struct io_kiocb *preq;
5437 spin_lock_irq(&ctx->completion_lock);
5438 preq = io_poll_find(ctx, req->poll.old_user_data);
5442 } else if (preq->opcode != IORING_OP_POLL_ADD) {
5443 /* don't allow internal poll updates */
5449 * Don't allow racy completion with singleshot, as we cannot safely
5450 * update those. For multishot, if we're racing with completion, just
5451 * let completion re-add it.
5453 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5454 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5458 /* we now have a detached poll request. reissue. */
5462 spin_unlock_irq(&ctx->completion_lock);
5463 req_set_fail_links(req);
5464 io_req_complete(req, ret);
5467 /* only mask one event flags, keep behavior flags */
5468 if (req->poll.update_events) {
5469 preq->poll.events &= ~0xffff;
5470 preq->poll.events |= req->poll.events & 0xffff;
5471 preq->poll.events |= IO_POLL_UNMASK;
5473 if (req->poll.update_user_data)
5474 preq->user_data = req->poll.new_user_data;
5476 spin_unlock_irq(&ctx->completion_lock);
5478 /* complete update request, we're done with it */
5479 io_req_complete(req, ret);
5482 ret = __io_poll_add(preq);
5484 req_set_fail_links(preq);
5485 io_req_complete(preq, ret);
5491 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5493 if (!req->poll.update_events && !req->poll.update_user_data)
5494 return __io_poll_add(req);
5495 return io_poll_update(req);
5498 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5500 struct io_timeout_data *data = container_of(timer,
5501 struct io_timeout_data, timer);
5502 struct io_kiocb *req = data->req;
5503 struct io_ring_ctx *ctx = req->ctx;
5504 unsigned long flags;
5506 spin_lock_irqsave(&ctx->completion_lock, flags);
5507 list_del_init(&req->timeout.list);
5508 atomic_set(&req->ctx->cq_timeouts,
5509 atomic_read(&req->ctx->cq_timeouts) + 1);
5511 io_cqring_fill_event(req, -ETIME);
5512 io_commit_cqring(ctx);
5513 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5515 io_cqring_ev_posted(ctx);
5516 req_set_fail_links(req);
5518 return HRTIMER_NORESTART;
5521 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5523 __must_hold(&ctx->completion_lock)
5525 struct io_timeout_data *io;
5526 struct io_kiocb *req;
5529 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5530 if (user_data == req->user_data) {
5537 return ERR_PTR(ret);
5539 io = req->async_data;
5540 ret = hrtimer_try_to_cancel(&io->timer);
5542 return ERR_PTR(-EALREADY);
5543 list_del_init(&req->timeout.list);
5547 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5548 __must_hold(&ctx->completion_lock)
5550 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5553 return PTR_ERR(req);
5555 req_set_fail_links(req);
5556 io_cqring_fill_event(req, -ECANCELED);
5557 io_put_req_deferred(req, 1);
5561 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5562 struct timespec64 *ts, enum hrtimer_mode mode)
5563 __must_hold(&ctx->completion_lock)
5565 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5566 struct io_timeout_data *data;
5569 return PTR_ERR(req);
5571 req->timeout.off = 0; /* noseq */
5572 data = req->async_data;
5573 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5574 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5575 data->timer.function = io_timeout_fn;
5576 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5580 static int io_timeout_remove_prep(struct io_kiocb *req,
5581 const struct io_uring_sqe *sqe)
5583 struct io_timeout_rem *tr = &req->timeout_rem;
5585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5587 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5589 if (sqe->ioprio || sqe->buf_index || sqe->len)
5592 tr->addr = READ_ONCE(sqe->addr);
5593 tr->flags = READ_ONCE(sqe->timeout_flags);
5594 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5595 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5597 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5599 } else if (tr->flags) {
5600 /* timeout removal doesn't support flags */
5607 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5609 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5614 * Remove or update an existing timeout command
5616 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5618 struct io_timeout_rem *tr = &req->timeout_rem;
5619 struct io_ring_ctx *ctx = req->ctx;
5622 spin_lock_irq(&ctx->completion_lock);
5623 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5624 ret = io_timeout_cancel(ctx, tr->addr);
5626 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5627 io_translate_timeout_mode(tr->flags));
5629 io_cqring_fill_event(req, ret);
5630 io_commit_cqring(ctx);
5631 spin_unlock_irq(&ctx->completion_lock);
5632 io_cqring_ev_posted(ctx);
5634 req_set_fail_links(req);
5639 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5640 bool is_timeout_link)
5642 struct io_timeout_data *data;
5644 u32 off = READ_ONCE(sqe->off);
5646 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5648 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5650 if (off && is_timeout_link)
5652 flags = READ_ONCE(sqe->timeout_flags);
5653 if (flags & ~IORING_TIMEOUT_ABS)
5656 req->timeout.off = off;
5658 if (!req->async_data && io_alloc_async_data(req))
5661 data = req->async_data;
5664 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5667 data->mode = io_translate_timeout_mode(flags);
5668 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5669 if (is_timeout_link)
5670 io_req_track_inflight(req);
5674 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5676 struct io_ring_ctx *ctx = req->ctx;
5677 struct io_timeout_data *data = req->async_data;
5678 struct list_head *entry;
5679 u32 tail, off = req->timeout.off;
5681 spin_lock_irq(&ctx->completion_lock);
5684 * sqe->off holds how many events that need to occur for this
5685 * timeout event to be satisfied. If it isn't set, then this is
5686 * a pure timeout request, sequence isn't used.
5688 if (io_is_timeout_noseq(req)) {
5689 entry = ctx->timeout_list.prev;
5693 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5694 req->timeout.target_seq = tail + off;
5696 /* Update the last seq here in case io_flush_timeouts() hasn't.
5697 * This is safe because ->completion_lock is held, and submissions
5698 * and completions are never mixed in the same ->completion_lock section.
5700 ctx->cq_last_tm_flush = tail;
5703 * Insertion sort, ensuring the first entry in the list is always
5704 * the one we need first.
5706 list_for_each_prev(entry, &ctx->timeout_list) {
5707 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5710 if (io_is_timeout_noseq(nxt))
5712 /* nxt.seq is behind @tail, otherwise would've been completed */
5713 if (off >= nxt->timeout.target_seq - tail)
5717 list_add(&req->timeout.list, entry);
5718 data->timer.function = io_timeout_fn;
5719 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5720 spin_unlock_irq(&ctx->completion_lock);
5724 struct io_cancel_data {
5725 struct io_ring_ctx *ctx;
5729 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5731 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5732 struct io_cancel_data *cd = data;
5734 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5737 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5738 struct io_ring_ctx *ctx)
5740 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5741 enum io_wq_cancel cancel_ret;
5744 if (!tctx || !tctx->io_wq)
5747 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5748 switch (cancel_ret) {
5749 case IO_WQ_CANCEL_OK:
5752 case IO_WQ_CANCEL_RUNNING:
5755 case IO_WQ_CANCEL_NOTFOUND:
5763 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5764 struct io_kiocb *req, __u64 sqe_addr,
5767 unsigned long flags;
5770 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5771 spin_lock_irqsave(&ctx->completion_lock, flags);
5774 ret = io_timeout_cancel(ctx, sqe_addr);
5777 ret = io_poll_cancel(ctx, sqe_addr);
5781 io_cqring_fill_event(req, ret);
5782 io_commit_cqring(ctx);
5783 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5784 io_cqring_ev_posted(ctx);
5787 req_set_fail_links(req);
5790 static int io_async_cancel_prep(struct io_kiocb *req,
5791 const struct io_uring_sqe *sqe)
5793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5795 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5797 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5800 req->cancel.addr = READ_ONCE(sqe->addr);
5804 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5806 struct io_ring_ctx *ctx = req->ctx;
5807 u64 sqe_addr = req->cancel.addr;
5808 struct io_tctx_node *node;
5811 /* tasks should wait for their io-wq threads, so safe w/o sync */
5812 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5813 spin_lock_irq(&ctx->completion_lock);
5816 ret = io_timeout_cancel(ctx, sqe_addr);
5819 ret = io_poll_cancel(ctx, sqe_addr);
5822 spin_unlock_irq(&ctx->completion_lock);
5824 /* slow path, try all io-wq's */
5825 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5827 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5828 struct io_uring_task *tctx = node->task->io_uring;
5830 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5834 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5836 spin_lock_irq(&ctx->completion_lock);
5838 io_cqring_fill_event(req, ret);
5839 io_commit_cqring(ctx);
5840 spin_unlock_irq(&ctx->completion_lock);
5841 io_cqring_ev_posted(ctx);
5844 req_set_fail_links(req);
5849 static int io_rsrc_update_prep(struct io_kiocb *req,
5850 const struct io_uring_sqe *sqe)
5852 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5854 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5856 if (sqe->ioprio || sqe->rw_flags)
5859 req->rsrc_update.offset = READ_ONCE(sqe->off);
5860 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5861 if (!req->rsrc_update.nr_args)
5863 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5867 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5869 struct io_ring_ctx *ctx = req->ctx;
5870 struct io_uring_rsrc_update up;
5873 if (issue_flags & IO_URING_F_NONBLOCK)
5876 up.offset = req->rsrc_update.offset;
5877 up.data = req->rsrc_update.arg;
5879 mutex_lock(&ctx->uring_lock);
5880 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5881 mutex_unlock(&ctx->uring_lock);
5884 req_set_fail_links(req);
5885 __io_req_complete(req, issue_flags, ret, 0);
5889 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5891 switch (req->opcode) {
5894 case IORING_OP_READV:
5895 case IORING_OP_READ_FIXED:
5896 case IORING_OP_READ:
5897 return io_read_prep(req, sqe);
5898 case IORING_OP_WRITEV:
5899 case IORING_OP_WRITE_FIXED:
5900 case IORING_OP_WRITE:
5901 return io_write_prep(req, sqe);
5902 case IORING_OP_POLL_ADD:
5903 return io_poll_add_prep(req, sqe);
5904 case IORING_OP_POLL_REMOVE:
5905 return io_poll_remove_prep(req, sqe);
5906 case IORING_OP_FSYNC:
5907 return io_fsync_prep(req, sqe);
5908 case IORING_OP_SYNC_FILE_RANGE:
5909 return io_sfr_prep(req, sqe);
5910 case IORING_OP_SENDMSG:
5911 case IORING_OP_SEND:
5912 return io_sendmsg_prep(req, sqe);
5913 case IORING_OP_RECVMSG:
5914 case IORING_OP_RECV:
5915 return io_recvmsg_prep(req, sqe);
5916 case IORING_OP_CONNECT:
5917 return io_connect_prep(req, sqe);
5918 case IORING_OP_TIMEOUT:
5919 return io_timeout_prep(req, sqe, false);
5920 case IORING_OP_TIMEOUT_REMOVE:
5921 return io_timeout_remove_prep(req, sqe);
5922 case IORING_OP_ASYNC_CANCEL:
5923 return io_async_cancel_prep(req, sqe);
5924 case IORING_OP_LINK_TIMEOUT:
5925 return io_timeout_prep(req, sqe, true);
5926 case IORING_OP_ACCEPT:
5927 return io_accept_prep(req, sqe);
5928 case IORING_OP_FALLOCATE:
5929 return io_fallocate_prep(req, sqe);
5930 case IORING_OP_OPENAT:
5931 return io_openat_prep(req, sqe);
5932 case IORING_OP_CLOSE:
5933 return io_close_prep(req, sqe);
5934 case IORING_OP_FILES_UPDATE:
5935 return io_rsrc_update_prep(req, sqe);
5936 case IORING_OP_STATX:
5937 return io_statx_prep(req, sqe);
5938 case IORING_OP_FADVISE:
5939 return io_fadvise_prep(req, sqe);
5940 case IORING_OP_MADVISE:
5941 return io_madvise_prep(req, sqe);
5942 case IORING_OP_OPENAT2:
5943 return io_openat2_prep(req, sqe);
5944 case IORING_OP_EPOLL_CTL:
5945 return io_epoll_ctl_prep(req, sqe);
5946 case IORING_OP_SPLICE:
5947 return io_splice_prep(req, sqe);
5948 case IORING_OP_PROVIDE_BUFFERS:
5949 return io_provide_buffers_prep(req, sqe);
5950 case IORING_OP_REMOVE_BUFFERS:
5951 return io_remove_buffers_prep(req, sqe);
5953 return io_tee_prep(req, sqe);
5954 case IORING_OP_SHUTDOWN:
5955 return io_shutdown_prep(req, sqe);
5956 case IORING_OP_RENAMEAT:
5957 return io_renameat_prep(req, sqe);
5958 case IORING_OP_UNLINKAT:
5959 return io_unlinkat_prep(req, sqe);
5962 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5967 static int io_req_prep_async(struct io_kiocb *req)
5969 if (!io_op_defs[req->opcode].needs_async_setup)
5971 if (WARN_ON_ONCE(req->async_data))
5973 if (io_alloc_async_data(req))
5976 switch (req->opcode) {
5977 case IORING_OP_READV:
5978 return io_rw_prep_async(req, READ);
5979 case IORING_OP_WRITEV:
5980 return io_rw_prep_async(req, WRITE);
5981 case IORING_OP_SENDMSG:
5982 return io_sendmsg_prep_async(req);
5983 case IORING_OP_RECVMSG:
5984 return io_recvmsg_prep_async(req);
5985 case IORING_OP_CONNECT:
5986 return io_connect_prep_async(req);
5988 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5993 static u32 io_get_sequence(struct io_kiocb *req)
5995 struct io_kiocb *pos;
5996 struct io_ring_ctx *ctx = req->ctx;
5997 u32 total_submitted, nr_reqs = 0;
5999 io_for_each_link(pos, req)
6002 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6003 return total_submitted - nr_reqs;
6006 static int io_req_defer(struct io_kiocb *req)
6008 struct io_ring_ctx *ctx = req->ctx;
6009 struct io_defer_entry *de;
6013 /* Still need defer if there is pending req in defer list. */
6014 if (likely(list_empty_careful(&ctx->defer_list) &&
6015 !(req->flags & REQ_F_IO_DRAIN)))
6018 seq = io_get_sequence(req);
6019 /* Still a chance to pass the sequence check */
6020 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6023 ret = io_req_prep_async(req);
6026 io_prep_async_link(req);
6027 de = kmalloc(sizeof(*de), GFP_KERNEL);
6031 spin_lock_irq(&ctx->completion_lock);
6032 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6033 spin_unlock_irq(&ctx->completion_lock);
6035 io_queue_async_work(req);
6036 return -EIOCBQUEUED;
6039 trace_io_uring_defer(ctx, req, req->user_data);
6042 list_add_tail(&de->list, &ctx->defer_list);
6043 spin_unlock_irq(&ctx->completion_lock);
6044 return -EIOCBQUEUED;
6047 static void io_clean_op(struct io_kiocb *req)
6049 if (req->flags & REQ_F_BUFFER_SELECTED) {
6050 switch (req->opcode) {
6051 case IORING_OP_READV:
6052 case IORING_OP_READ_FIXED:
6053 case IORING_OP_READ:
6054 kfree((void *)(unsigned long)req->rw.addr);
6056 case IORING_OP_RECVMSG:
6057 case IORING_OP_RECV:
6058 kfree(req->sr_msg.kbuf);
6061 req->flags &= ~REQ_F_BUFFER_SELECTED;
6064 if (req->flags & REQ_F_NEED_CLEANUP) {
6065 switch (req->opcode) {
6066 case IORING_OP_READV:
6067 case IORING_OP_READ_FIXED:
6068 case IORING_OP_READ:
6069 case IORING_OP_WRITEV:
6070 case IORING_OP_WRITE_FIXED:
6071 case IORING_OP_WRITE: {
6072 struct io_async_rw *io = req->async_data;
6074 kfree(io->free_iovec);
6077 case IORING_OP_RECVMSG:
6078 case IORING_OP_SENDMSG: {
6079 struct io_async_msghdr *io = req->async_data;
6081 kfree(io->free_iov);
6084 case IORING_OP_SPLICE:
6086 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6087 io_put_file(req->splice.file_in);
6089 case IORING_OP_OPENAT:
6090 case IORING_OP_OPENAT2:
6091 if (req->open.filename)
6092 putname(req->open.filename);
6094 case IORING_OP_RENAMEAT:
6095 putname(req->rename.oldpath);
6096 putname(req->rename.newpath);
6098 case IORING_OP_UNLINKAT:
6099 putname(req->unlink.filename);
6102 req->flags &= ~REQ_F_NEED_CLEANUP;
6106 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6108 struct io_ring_ctx *ctx = req->ctx;
6109 const struct cred *creds = NULL;
6112 if (req->work.creds && req->work.creds != current_cred())
6113 creds = override_creds(req->work.creds);
6115 switch (req->opcode) {
6117 ret = io_nop(req, issue_flags);
6119 case IORING_OP_READV:
6120 case IORING_OP_READ_FIXED:
6121 case IORING_OP_READ:
6122 ret = io_read(req, issue_flags);
6124 case IORING_OP_WRITEV:
6125 case IORING_OP_WRITE_FIXED:
6126 case IORING_OP_WRITE:
6127 ret = io_write(req, issue_flags);
6129 case IORING_OP_FSYNC:
6130 ret = io_fsync(req, issue_flags);
6132 case IORING_OP_POLL_ADD:
6133 ret = io_poll_add(req, issue_flags);
6135 case IORING_OP_POLL_REMOVE:
6136 ret = io_poll_remove(req, issue_flags);
6138 case IORING_OP_SYNC_FILE_RANGE:
6139 ret = io_sync_file_range(req, issue_flags);
6141 case IORING_OP_SENDMSG:
6142 ret = io_sendmsg(req, issue_flags);
6144 case IORING_OP_SEND:
6145 ret = io_send(req, issue_flags);
6147 case IORING_OP_RECVMSG:
6148 ret = io_recvmsg(req, issue_flags);
6150 case IORING_OP_RECV:
6151 ret = io_recv(req, issue_flags);
6153 case IORING_OP_TIMEOUT:
6154 ret = io_timeout(req, issue_flags);
6156 case IORING_OP_TIMEOUT_REMOVE:
6157 ret = io_timeout_remove(req, issue_flags);
6159 case IORING_OP_ACCEPT:
6160 ret = io_accept(req, issue_flags);
6162 case IORING_OP_CONNECT:
6163 ret = io_connect(req, issue_flags);
6165 case IORING_OP_ASYNC_CANCEL:
6166 ret = io_async_cancel(req, issue_flags);
6168 case IORING_OP_FALLOCATE:
6169 ret = io_fallocate(req, issue_flags);
6171 case IORING_OP_OPENAT:
6172 ret = io_openat(req, issue_flags);
6174 case IORING_OP_CLOSE:
6175 ret = io_close(req, issue_flags);
6177 case IORING_OP_FILES_UPDATE:
6178 ret = io_files_update(req, issue_flags);
6180 case IORING_OP_STATX:
6181 ret = io_statx(req, issue_flags);
6183 case IORING_OP_FADVISE:
6184 ret = io_fadvise(req, issue_flags);
6186 case IORING_OP_MADVISE:
6187 ret = io_madvise(req, issue_flags);
6189 case IORING_OP_OPENAT2:
6190 ret = io_openat2(req, issue_flags);
6192 case IORING_OP_EPOLL_CTL:
6193 ret = io_epoll_ctl(req, issue_flags);
6195 case IORING_OP_SPLICE:
6196 ret = io_splice(req, issue_flags);
6198 case IORING_OP_PROVIDE_BUFFERS:
6199 ret = io_provide_buffers(req, issue_flags);
6201 case IORING_OP_REMOVE_BUFFERS:
6202 ret = io_remove_buffers(req, issue_flags);
6205 ret = io_tee(req, issue_flags);
6207 case IORING_OP_SHUTDOWN:
6208 ret = io_shutdown(req, issue_flags);
6210 case IORING_OP_RENAMEAT:
6211 ret = io_renameat(req, issue_flags);
6213 case IORING_OP_UNLINKAT:
6214 ret = io_unlinkat(req, issue_flags);
6222 revert_creds(creds);
6227 /* If the op doesn't have a file, we're not polling for it */
6228 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6229 const bool in_async = io_wq_current_is_worker();
6231 /* workqueue context doesn't hold uring_lock, grab it now */
6233 mutex_lock(&ctx->uring_lock);
6235 io_iopoll_req_issued(req, in_async);
6238 mutex_unlock(&ctx->uring_lock);
6244 static void io_wq_submit_work(struct io_wq_work *work)
6246 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6247 struct io_kiocb *timeout;
6250 timeout = io_prep_linked_timeout(req);
6252 io_queue_linked_timeout(timeout);
6254 if (work->flags & IO_WQ_WORK_CANCEL)
6259 ret = io_issue_sqe(req, 0);
6261 * We can get EAGAIN for polled IO even though we're
6262 * forcing a sync submission from here, since we can't
6263 * wait for request slots on the block side.
6271 /* avoid locking problems by failing it from a clean context */
6273 /* io-wq is going to take one down */
6275 io_req_task_queue_fail(req, ret);
6279 #define FFS_ASYNC_READ 0x1UL
6280 #define FFS_ASYNC_WRITE 0x2UL
6282 #define FFS_ISREG 0x4UL
6284 #define FFS_ISREG 0x0UL
6286 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6288 static inline struct io_fixed_file *io_fixed_file_slot(struct io_rsrc_data *file_data,
6291 struct fixed_rsrc_table *table;
6293 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6294 return &table->files[i & IORING_FILE_TABLE_MASK];
6297 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6300 struct io_fixed_file *slot = io_fixed_file_slot(ctx->file_data, index);
6302 return (struct file *) (slot->file_ptr & FFS_MASK);
6305 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6307 unsigned long file_ptr = (unsigned long) file;
6309 if (__io_file_supports_async(file, READ))
6310 file_ptr |= FFS_ASYNC_READ;
6311 if (__io_file_supports_async(file, WRITE))
6312 file_ptr |= FFS_ASYNC_WRITE;
6313 if (S_ISREG(file_inode(file)->i_mode))
6314 file_ptr |= FFS_ISREG;
6315 file_slot->file_ptr = file_ptr;
6318 static struct file *io_file_get(struct io_submit_state *state,
6319 struct io_kiocb *req, int fd, bool fixed)
6321 struct io_ring_ctx *ctx = req->ctx;
6325 unsigned long file_ptr;
6327 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6329 fd = array_index_nospec(fd, ctx->nr_user_files);
6330 file_ptr = io_fixed_file_slot(ctx->file_data, fd)->file_ptr;
6331 file = (struct file *) (file_ptr & FFS_MASK);
6332 file_ptr &= ~FFS_MASK;
6333 /* mask in overlapping REQ_F and FFS bits */
6334 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6335 io_req_set_rsrc_node(req);
6337 trace_io_uring_file_get(ctx, fd);
6338 file = __io_file_get(state, fd);
6340 /* we don't allow fixed io_uring files */
6341 if (file && unlikely(file->f_op == &io_uring_fops))
6342 io_req_track_inflight(req);
6348 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6350 struct io_timeout_data *data = container_of(timer,
6351 struct io_timeout_data, timer);
6352 struct io_kiocb *prev, *req = data->req;
6353 struct io_ring_ctx *ctx = req->ctx;
6354 unsigned long flags;
6356 spin_lock_irqsave(&ctx->completion_lock, flags);
6357 prev = req->timeout.head;
6358 req->timeout.head = NULL;
6361 * We don't expect the list to be empty, that will only happen if we
6362 * race with the completion of the linked work.
6364 if (prev && req_ref_inc_not_zero(prev))
6365 io_remove_next_linked(prev);
6368 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6371 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6372 io_put_req_deferred(prev, 1);
6374 io_req_complete_post(req, -ETIME, 0);
6376 io_put_req_deferred(req, 1);
6377 return HRTIMER_NORESTART;
6380 static void io_queue_linked_timeout(struct io_kiocb *req)
6382 struct io_ring_ctx *ctx = req->ctx;
6384 spin_lock_irq(&ctx->completion_lock);
6386 * If the back reference is NULL, then our linked request finished
6387 * before we got a chance to setup the timer
6389 if (req->timeout.head) {
6390 struct io_timeout_data *data = req->async_data;
6392 data->timer.function = io_link_timeout_fn;
6393 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6396 spin_unlock_irq(&ctx->completion_lock);
6397 /* drop submission reference */
6401 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6403 struct io_kiocb *nxt = req->link;
6405 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6406 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6409 nxt->timeout.head = req;
6410 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6411 req->flags |= REQ_F_LINK_TIMEOUT;
6415 static void __io_queue_sqe(struct io_kiocb *req)
6417 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6420 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6423 * We async punt it if the file wasn't marked NOWAIT, or if the file
6424 * doesn't support non-blocking read/write attempts
6427 /* drop submission reference */
6428 if (req->flags & REQ_F_COMPLETE_INLINE) {
6429 struct io_ring_ctx *ctx = req->ctx;
6430 struct io_comp_state *cs = &ctx->submit_state.comp;
6432 cs->reqs[cs->nr++] = req;
6433 if (cs->nr == ARRAY_SIZE(cs->reqs))
6434 io_submit_flush_completions(cs, ctx);
6438 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6439 if (!io_arm_poll_handler(req)) {
6441 * Queued up for async execution, worker will release
6442 * submit reference when the iocb is actually submitted.
6444 io_queue_async_work(req);
6447 io_req_complete_failed(req, ret);
6450 io_queue_linked_timeout(linked_timeout);
6453 static void io_queue_sqe(struct io_kiocb *req)
6457 ret = io_req_defer(req);
6459 if (ret != -EIOCBQUEUED) {
6461 io_req_complete_failed(req, ret);
6463 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6464 ret = io_req_prep_async(req);
6467 io_queue_async_work(req);
6469 __io_queue_sqe(req);
6474 * Check SQE restrictions (opcode and flags).
6476 * Returns 'true' if SQE is allowed, 'false' otherwise.
6478 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6479 struct io_kiocb *req,
6480 unsigned int sqe_flags)
6482 if (!ctx->restricted)
6485 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6488 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6489 ctx->restrictions.sqe_flags_required)
6492 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6493 ctx->restrictions.sqe_flags_required))
6499 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6500 const struct io_uring_sqe *sqe)
6502 struct io_submit_state *state;
6503 unsigned int sqe_flags;
6504 int personality, ret = 0;
6506 req->opcode = READ_ONCE(sqe->opcode);
6507 /* same numerical values with corresponding REQ_F_*, safe to copy */
6508 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6509 req->user_data = READ_ONCE(sqe->user_data);
6510 req->async_data = NULL;
6514 req->fixed_rsrc_refs = NULL;
6515 /* one is dropped after submission, the other at completion */
6516 atomic_set(&req->refs, 2);
6517 req->task = current;
6519 req->work.creds = NULL;
6521 /* enforce forwards compatibility on users */
6522 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6527 if (unlikely(req->opcode >= IORING_OP_LAST))
6530 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6533 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6534 !io_op_defs[req->opcode].buffer_select)
6537 personality = READ_ONCE(sqe->personality);
6539 req->work.creds = xa_load(&ctx->personalities, personality);
6540 if (!req->work.creds)
6542 get_cred(req->work.creds);
6544 state = &ctx->submit_state;
6547 * Plug now if we have more than 1 IO left after this, and the target
6548 * is potentially a read/write to block based storage.
6550 if (!state->plug_started && state->ios_left > 1 &&
6551 io_op_defs[req->opcode].plug) {
6552 blk_start_plug(&state->plug);
6553 state->plug_started = true;
6556 if (io_op_defs[req->opcode].needs_file) {
6557 bool fixed = req->flags & REQ_F_FIXED_FILE;
6559 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6560 if (unlikely(!req->file))
6568 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6569 const struct io_uring_sqe *sqe)
6571 struct io_submit_link *link = &ctx->submit_state.link;
6574 ret = io_init_req(ctx, req, sqe);
6575 if (unlikely(ret)) {
6578 /* fail even hard links since we don't submit */
6579 link->head->flags |= REQ_F_FAIL_LINK;
6580 io_req_complete_failed(link->head, -ECANCELED);
6583 io_req_complete_failed(req, ret);
6586 ret = io_req_prep(req, sqe);
6590 /* don't need @sqe from now on */
6591 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6592 true, ctx->flags & IORING_SETUP_SQPOLL);
6595 * If we already have a head request, queue this one for async
6596 * submittal once the head completes. If we don't have a head but
6597 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6598 * submitted sync once the chain is complete. If none of those
6599 * conditions are true (normal request), then just queue it.
6602 struct io_kiocb *head = link->head;
6605 * Taking sequential execution of a link, draining both sides
6606 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6607 * requests in the link. So, it drains the head and the
6608 * next after the link request. The last one is done via
6609 * drain_next flag to persist the effect across calls.
6611 if (req->flags & REQ_F_IO_DRAIN) {
6612 head->flags |= REQ_F_IO_DRAIN;
6613 ctx->drain_next = 1;
6615 ret = io_req_prep_async(req);
6618 trace_io_uring_link(ctx, req, head);
6619 link->last->link = req;
6622 /* last request of a link, enqueue the link */
6623 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6628 if (unlikely(ctx->drain_next)) {
6629 req->flags |= REQ_F_IO_DRAIN;
6630 ctx->drain_next = 0;
6632 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6644 * Batched submission is done, ensure local IO is flushed out.
6646 static void io_submit_state_end(struct io_submit_state *state,
6647 struct io_ring_ctx *ctx)
6649 if (state->link.head)
6650 io_queue_sqe(state->link.head);
6652 io_submit_flush_completions(&state->comp, ctx);
6653 if (state->plug_started)
6654 blk_finish_plug(&state->plug);
6655 io_state_file_put(state);
6659 * Start submission side cache.
6661 static void io_submit_state_start(struct io_submit_state *state,
6662 unsigned int max_ios)
6664 state->plug_started = false;
6665 state->ios_left = max_ios;
6666 /* set only head, no need to init link_last in advance */
6667 state->link.head = NULL;
6670 static void io_commit_sqring(struct io_ring_ctx *ctx)
6672 struct io_rings *rings = ctx->rings;
6675 * Ensure any loads from the SQEs are done at this point,
6676 * since once we write the new head, the application could
6677 * write new data to them.
6679 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6683 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6684 * that is mapped by userspace. This means that care needs to be taken to
6685 * ensure that reads are stable, as we cannot rely on userspace always
6686 * being a good citizen. If members of the sqe are validated and then later
6687 * used, it's important that those reads are done through READ_ONCE() to
6688 * prevent a re-load down the line.
6690 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6692 u32 *sq_array = ctx->sq_array;
6696 * The cached sq head (or cq tail) serves two purposes:
6698 * 1) allows us to batch the cost of updating the user visible
6700 * 2) allows the kernel side to track the head on its own, even
6701 * though the application is the one updating it.
6703 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6704 if (likely(head < ctx->sq_entries))
6705 return &ctx->sq_sqes[head];
6707 /* drop invalid entries */
6708 ctx->cached_sq_dropped++;
6709 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6713 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6717 /* if we have a backlog and couldn't flush it all, return BUSY */
6718 if (test_bit(0, &ctx->sq_check_overflow)) {
6719 if (!__io_cqring_overflow_flush(ctx, false))
6723 /* make sure SQ entry isn't read before tail */
6724 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6726 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6729 percpu_counter_add(¤t->io_uring->inflight, nr);
6730 refcount_add(nr, ¤t->usage);
6731 io_submit_state_start(&ctx->submit_state, nr);
6733 while (submitted < nr) {
6734 const struct io_uring_sqe *sqe;
6735 struct io_kiocb *req;
6737 req = io_alloc_req(ctx);
6738 if (unlikely(!req)) {
6740 submitted = -EAGAIN;
6743 sqe = io_get_sqe(ctx);
6744 if (unlikely(!sqe)) {
6745 kmem_cache_free(req_cachep, req);
6748 /* will complete beyond this point, count as submitted */
6750 if (io_submit_sqe(ctx, req, sqe))
6754 if (unlikely(submitted != nr)) {
6755 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6756 struct io_uring_task *tctx = current->io_uring;
6757 int unused = nr - ref_used;
6759 percpu_ref_put_many(&ctx->refs, unused);
6760 percpu_counter_sub(&tctx->inflight, unused);
6761 put_task_struct_many(current, unused);
6764 io_submit_state_end(&ctx->submit_state, ctx);
6765 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6766 io_commit_sqring(ctx);
6771 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6773 /* Tell userspace we may need a wakeup call */
6774 spin_lock_irq(&ctx->completion_lock);
6775 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6776 spin_unlock_irq(&ctx->completion_lock);
6779 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6781 spin_lock_irq(&ctx->completion_lock);
6782 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6783 spin_unlock_irq(&ctx->completion_lock);
6786 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6788 unsigned int to_submit;
6791 to_submit = io_sqring_entries(ctx);
6792 /* if we're handling multiple rings, cap submit size for fairness */
6793 if (cap_entries && to_submit > 8)
6796 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6797 unsigned nr_events = 0;
6799 mutex_lock(&ctx->uring_lock);
6800 if (!list_empty(&ctx->iopoll_list))
6801 io_do_iopoll(ctx, &nr_events, 0);
6803 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6804 !(ctx->flags & IORING_SETUP_R_DISABLED))
6805 ret = io_submit_sqes(ctx, to_submit);
6806 mutex_unlock(&ctx->uring_lock);
6809 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6810 wake_up(&ctx->sqo_sq_wait);
6815 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6817 struct io_ring_ctx *ctx;
6818 unsigned sq_thread_idle = 0;
6820 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6821 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6822 sqd->sq_thread_idle = sq_thread_idle;
6825 static int io_sq_thread(void *data)
6827 struct io_sq_data *sqd = data;
6828 struct io_ring_ctx *ctx;
6829 unsigned long timeout = 0;
6830 char buf[TASK_COMM_LEN];
6833 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6834 set_task_comm(current, buf);
6835 current->pf_io_worker = NULL;
6837 if (sqd->sq_cpu != -1)
6838 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6840 set_cpus_allowed_ptr(current, cpu_online_mask);
6841 current->flags |= PF_NO_SETAFFINITY;
6843 mutex_lock(&sqd->lock);
6844 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6846 bool cap_entries, sqt_spin, needs_sched;
6848 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6849 signal_pending(current)) {
6850 bool did_sig = false;
6852 mutex_unlock(&sqd->lock);
6853 if (signal_pending(current)) {
6854 struct ksignal ksig;
6856 did_sig = get_signal(&ksig);
6859 mutex_lock(&sqd->lock);
6863 io_run_task_work_head(&sqd->park_task_work);
6864 timeout = jiffies + sqd->sq_thread_idle;
6868 cap_entries = !list_is_singular(&sqd->ctx_list);
6869 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6870 const struct cred *creds = NULL;
6872 if (ctx->sq_creds != current_cred())
6873 creds = override_creds(ctx->sq_creds);
6874 ret = __io_sq_thread(ctx, cap_entries);
6876 revert_creds(creds);
6877 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6881 if (sqt_spin || !time_after(jiffies, timeout)) {
6885 timeout = jiffies + sqd->sq_thread_idle;
6890 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6891 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6892 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6893 !list_empty_careful(&ctx->iopoll_list)) {
6894 needs_sched = false;
6897 if (io_sqring_entries(ctx)) {
6898 needs_sched = false;
6903 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6904 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6905 io_ring_set_wakeup_flag(ctx);
6907 mutex_unlock(&sqd->lock);
6909 mutex_lock(&sqd->lock);
6910 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6911 io_ring_clear_wakeup_flag(ctx);
6914 finish_wait(&sqd->wait, &wait);
6915 io_run_task_work_head(&sqd->park_task_work);
6916 timeout = jiffies + sqd->sq_thread_idle;
6919 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6920 io_uring_cancel_sqpoll(ctx);
6922 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6923 io_ring_set_wakeup_flag(ctx);
6924 mutex_unlock(&sqd->lock);
6927 io_run_task_work_head(&sqd->park_task_work);
6928 complete(&sqd->exited);
6932 struct io_wait_queue {
6933 struct wait_queue_entry wq;
6934 struct io_ring_ctx *ctx;
6936 unsigned nr_timeouts;
6939 static inline bool io_should_wake(struct io_wait_queue *iowq)
6941 struct io_ring_ctx *ctx = iowq->ctx;
6944 * Wake up if we have enough events, or if a timeout occurred since we
6945 * started waiting. For timeouts, we always want to return to userspace,
6946 * regardless of event count.
6948 return io_cqring_events(ctx) >= iowq->to_wait ||
6949 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6952 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6953 int wake_flags, void *key)
6955 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6959 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6960 * the task, and the next invocation will do it.
6962 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6963 return autoremove_wake_function(curr, mode, wake_flags, key);
6967 static int io_run_task_work_sig(void)
6969 if (io_run_task_work())
6971 if (!signal_pending(current))
6973 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6974 return -ERESTARTSYS;
6978 /* when returns >0, the caller should retry */
6979 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6980 struct io_wait_queue *iowq,
6981 signed long *timeout)
6985 /* make sure we run task_work before checking for signals */
6986 ret = io_run_task_work_sig();
6987 if (ret || io_should_wake(iowq))
6989 /* let the caller flush overflows, retry */
6990 if (test_bit(0, &ctx->cq_check_overflow))
6993 *timeout = schedule_timeout(*timeout);
6994 return !*timeout ? -ETIME : 1;
6998 * Wait until events become available, if we don't already have some. The
6999 * application must reap them itself, as they reside on the shared cq ring.
7001 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7002 const sigset_t __user *sig, size_t sigsz,
7003 struct __kernel_timespec __user *uts)
7005 struct io_wait_queue iowq = {
7008 .func = io_wake_function,
7009 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7012 .to_wait = min_events,
7014 struct io_rings *rings = ctx->rings;
7015 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7019 io_cqring_overflow_flush(ctx, false);
7020 if (io_cqring_events(ctx) >= min_events)
7022 if (!io_run_task_work())
7027 #ifdef CONFIG_COMPAT
7028 if (in_compat_syscall())
7029 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7033 ret = set_user_sigmask(sig, sigsz);
7040 struct timespec64 ts;
7042 if (get_timespec64(&ts, uts))
7044 timeout = timespec64_to_jiffies(&ts);
7047 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7048 trace_io_uring_cqring_wait(ctx, min_events);
7050 /* if we can't even flush overflow, don't wait for more */
7051 if (!io_cqring_overflow_flush(ctx, false)) {
7055 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7056 TASK_INTERRUPTIBLE);
7057 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7058 finish_wait(&ctx->wait, &iowq.wq);
7062 restore_saved_sigmask_unless(ret == -EINTR);
7064 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7067 static void io_free_file_tables(struct io_rsrc_data *data, unsigned nr_files)
7069 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7071 for (i = 0; i < nr_tables; i++)
7072 kfree(data->table[i].files);
7077 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7079 #if defined(CONFIG_UNIX)
7080 if (ctx->ring_sock) {
7081 struct sock *sock = ctx->ring_sock->sk;
7082 struct sk_buff *skb;
7084 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7090 for (i = 0; i < ctx->nr_user_files; i++) {
7093 file = io_file_from_index(ctx, i);
7100 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7102 struct io_rsrc_data *data = container_of(ref, struct io_rsrc_data, refs);
7104 complete(&data->done);
7107 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7109 spin_lock_bh(&ctx->rsrc_ref_lock);
7112 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7114 spin_unlock_bh(&ctx->rsrc_ref_lock);
7117 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7119 percpu_ref_exit(&ref_node->refs);
7123 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7124 struct io_rsrc_data *data_to_kill)
7126 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7127 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7130 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7132 rsrc_node->rsrc_data = data_to_kill;
7133 io_rsrc_ref_lock(ctx);
7134 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7135 io_rsrc_ref_unlock(ctx);
7137 percpu_ref_get(&data_to_kill->refs);
7138 percpu_ref_kill(&rsrc_node->refs);
7139 ctx->rsrc_node = NULL;
7142 if (!ctx->rsrc_node) {
7143 ctx->rsrc_node = ctx->rsrc_backup_node;
7144 ctx->rsrc_backup_node = NULL;
7148 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7150 if (ctx->rsrc_backup_node)
7152 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7153 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7156 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7160 /* As we may drop ->uring_lock, other task may have started quiesce */
7164 data->quiesce = true;
7166 ret = io_rsrc_node_switch_start(ctx);
7169 io_rsrc_node_switch(ctx, data);
7171 percpu_ref_kill(&data->refs);
7172 flush_delayed_work(&ctx->rsrc_put_work);
7174 ret = wait_for_completion_interruptible(&data->done);
7178 percpu_ref_resurrect(&data->refs);
7179 reinit_completion(&data->done);
7181 mutex_unlock(&ctx->uring_lock);
7182 ret = io_run_task_work_sig();
7183 mutex_lock(&ctx->uring_lock);
7185 data->quiesce = false;
7190 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7191 rsrc_put_fn *do_put)
7193 struct io_rsrc_data *data;
7195 data = kzalloc(sizeof(*data), GFP_KERNEL);
7199 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7200 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7205 data->do_put = do_put;
7206 init_completion(&data->done);
7210 static void io_rsrc_data_free(struct io_rsrc_data *data)
7212 percpu_ref_exit(&data->refs);
7216 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7218 struct io_rsrc_data *data = ctx->file_data;
7223 ret = io_rsrc_ref_quiesce(data, ctx);
7227 __io_sqe_files_unregister(ctx);
7228 io_free_file_tables(data, ctx->nr_user_files);
7229 io_rsrc_data_free(data);
7230 ctx->file_data = NULL;
7231 ctx->nr_user_files = 0;
7235 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7236 __releases(&sqd->lock)
7238 WARN_ON_ONCE(sqd->thread == current);
7241 * Do the dance but not conditional clear_bit() because it'd race with
7242 * other threads incrementing park_pending and setting the bit.
7244 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7245 if (atomic_dec_return(&sqd->park_pending))
7246 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7247 mutex_unlock(&sqd->lock);
7250 static void io_sq_thread_park(struct io_sq_data *sqd)
7251 __acquires(&sqd->lock)
7253 WARN_ON_ONCE(sqd->thread == current);
7255 atomic_inc(&sqd->park_pending);
7256 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7257 mutex_lock(&sqd->lock);
7259 wake_up_process(sqd->thread);
7262 static void io_sq_thread_stop(struct io_sq_data *sqd)
7264 WARN_ON_ONCE(sqd->thread == current);
7266 mutex_lock(&sqd->lock);
7267 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7269 wake_up_process(sqd->thread);
7270 mutex_unlock(&sqd->lock);
7271 wait_for_completion(&sqd->exited);
7274 static void io_put_sq_data(struct io_sq_data *sqd)
7276 if (refcount_dec_and_test(&sqd->refs)) {
7277 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7279 io_sq_thread_stop(sqd);
7284 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7286 struct io_sq_data *sqd = ctx->sq_data;
7289 io_sq_thread_park(sqd);
7290 list_del_init(&ctx->sqd_list);
7291 io_sqd_update_thread_idle(sqd);
7292 io_sq_thread_unpark(sqd);
7294 io_put_sq_data(sqd);
7295 ctx->sq_data = NULL;
7297 put_cred(ctx->sq_creds);
7301 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7303 struct io_ring_ctx *ctx_attach;
7304 struct io_sq_data *sqd;
7307 f = fdget(p->wq_fd);
7309 return ERR_PTR(-ENXIO);
7310 if (f.file->f_op != &io_uring_fops) {
7312 return ERR_PTR(-EINVAL);
7315 ctx_attach = f.file->private_data;
7316 sqd = ctx_attach->sq_data;
7319 return ERR_PTR(-EINVAL);
7321 if (sqd->task_tgid != current->tgid) {
7323 return ERR_PTR(-EPERM);
7326 refcount_inc(&sqd->refs);
7331 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7334 struct io_sq_data *sqd;
7337 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7338 sqd = io_attach_sq_data(p);
7343 /* fall through for EPERM case, setup new sqd/task */
7344 if (PTR_ERR(sqd) != -EPERM)
7348 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7350 return ERR_PTR(-ENOMEM);
7352 atomic_set(&sqd->park_pending, 0);
7353 refcount_set(&sqd->refs, 1);
7354 INIT_LIST_HEAD(&sqd->ctx_list);
7355 mutex_init(&sqd->lock);
7356 init_waitqueue_head(&sqd->wait);
7357 init_completion(&sqd->exited);
7361 #if defined(CONFIG_UNIX)
7363 * Ensure the UNIX gc is aware of our file set, so we are certain that
7364 * the io_uring can be safely unregistered on process exit, even if we have
7365 * loops in the file referencing.
7367 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7369 struct sock *sk = ctx->ring_sock->sk;
7370 struct scm_fp_list *fpl;
7371 struct sk_buff *skb;
7374 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7378 skb = alloc_skb(0, GFP_KERNEL);
7387 fpl->user = get_uid(current_user());
7388 for (i = 0; i < nr; i++) {
7389 struct file *file = io_file_from_index(ctx, i + offset);
7393 fpl->fp[nr_files] = get_file(file);
7394 unix_inflight(fpl->user, fpl->fp[nr_files]);
7399 fpl->max = SCM_MAX_FD;
7400 fpl->count = nr_files;
7401 UNIXCB(skb).fp = fpl;
7402 skb->destructor = unix_destruct_scm;
7403 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7404 skb_queue_head(&sk->sk_receive_queue, skb);
7406 for (i = 0; i < nr_files; i++)
7417 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7418 * causes regular reference counting to break down. We rely on the UNIX
7419 * garbage collection to take care of this problem for us.
7421 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7423 unsigned left, total;
7427 left = ctx->nr_user_files;
7429 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7431 ret = __io_sqe_files_scm(ctx, this_files, total);
7435 total += this_files;
7441 while (total < ctx->nr_user_files) {
7442 struct file *file = io_file_from_index(ctx, total);
7452 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7458 static bool io_alloc_file_tables(struct io_rsrc_data *file_data,
7461 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7463 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7465 if (!file_data->table)
7468 for (i = 0; i < nr_tables; i++) {
7469 struct fixed_rsrc_table *table = &file_data->table[i];
7470 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7472 table->files = kcalloc(this_files, sizeof(struct file *),
7476 nr_files -= this_files;
7482 io_free_file_tables(file_data, nr_tables * IORING_MAX_FILES_TABLE);
7486 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7488 struct file *file = prsrc->file;
7489 #if defined(CONFIG_UNIX)
7490 struct sock *sock = ctx->ring_sock->sk;
7491 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7492 struct sk_buff *skb;
7495 __skb_queue_head_init(&list);
7498 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7499 * remove this entry and rearrange the file array.
7501 skb = skb_dequeue(head);
7503 struct scm_fp_list *fp;
7505 fp = UNIXCB(skb).fp;
7506 for (i = 0; i < fp->count; i++) {
7509 if (fp->fp[i] != file)
7512 unix_notinflight(fp->user, fp->fp[i]);
7513 left = fp->count - 1 - i;
7515 memmove(&fp->fp[i], &fp->fp[i + 1],
7516 left * sizeof(struct file *));
7523 __skb_queue_tail(&list, skb);
7533 __skb_queue_tail(&list, skb);
7535 skb = skb_dequeue(head);
7538 if (skb_peek(&list)) {
7539 spin_lock_irq(&head->lock);
7540 while ((skb = __skb_dequeue(&list)) != NULL)
7541 __skb_queue_tail(head, skb);
7542 spin_unlock_irq(&head->lock);
7549 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7551 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7552 struct io_ring_ctx *ctx = rsrc_data->ctx;
7553 struct io_rsrc_put *prsrc, *tmp;
7555 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7556 list_del(&prsrc->list);
7557 rsrc_data->do_put(ctx, prsrc);
7561 io_rsrc_node_destroy(ref_node);
7562 percpu_ref_put(&rsrc_data->refs);
7565 static void io_rsrc_put_work(struct work_struct *work)
7567 struct io_ring_ctx *ctx;
7568 struct llist_node *node;
7570 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7571 node = llist_del_all(&ctx->rsrc_put_llist);
7574 struct io_rsrc_node *ref_node;
7575 struct llist_node *next = node->next;
7577 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7578 __io_rsrc_put_work(ref_node);
7583 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7585 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7586 struct io_rsrc_data *data = node->rsrc_data;
7587 struct io_ring_ctx *ctx = data->ctx;
7588 bool first_add = false;
7591 io_rsrc_ref_lock(ctx);
7594 while (!list_empty(&ctx->rsrc_ref_list)) {
7595 node = list_first_entry(&ctx->rsrc_ref_list,
7596 struct io_rsrc_node, node);
7597 /* recycle ref nodes in order */
7600 list_del(&node->node);
7601 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7603 io_rsrc_ref_unlock(ctx);
7605 delay = percpu_ref_is_dying(&data->refs) ? 0 : HZ;
7606 if (first_add || !delay)
7607 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7610 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7612 struct io_rsrc_node *ref_node;
7614 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7618 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7623 INIT_LIST_HEAD(&ref_node->node);
7624 INIT_LIST_HEAD(&ref_node->rsrc_list);
7625 ref_node->done = false;
7629 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7632 __s32 __user *fds = (__s32 __user *) arg;
7636 struct io_rsrc_data *file_data;
7642 if (nr_args > IORING_MAX_FIXED_FILES)
7644 ret = io_rsrc_node_switch_start(ctx);
7648 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put);
7651 ctx->file_data = file_data;
7654 if (!io_alloc_file_tables(file_data, nr_args))
7657 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7658 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7662 /* allow sparse sets */
7672 * Don't allow io_uring instances to be registered. If UNIX
7673 * isn't enabled, then this causes a reference cycle and this
7674 * instance can never get freed. If UNIX is enabled we'll
7675 * handle it just fine, but there's still no point in allowing
7676 * a ring fd as it doesn't support regular read/write anyway.
7678 if (file->f_op == &io_uring_fops) {
7682 io_fixed_file_set(io_fixed_file_slot(file_data, i), file);
7685 ret = io_sqe_files_scm(ctx);
7687 io_sqe_files_unregister(ctx);
7691 io_rsrc_node_switch(ctx, NULL);
7694 for (i = 0; i < ctx->nr_user_files; i++) {
7695 file = io_file_from_index(ctx, i);
7699 io_free_file_tables(file_data, nr_args);
7700 ctx->nr_user_files = 0;
7702 io_rsrc_data_free(ctx->file_data);
7703 ctx->file_data = NULL;
7707 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7710 #if defined(CONFIG_UNIX)
7711 struct sock *sock = ctx->ring_sock->sk;
7712 struct sk_buff_head *head = &sock->sk_receive_queue;
7713 struct sk_buff *skb;
7716 * See if we can merge this file into an existing skb SCM_RIGHTS
7717 * file set. If there's no room, fall back to allocating a new skb
7718 * and filling it in.
7720 spin_lock_irq(&head->lock);
7721 skb = skb_peek(head);
7723 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7725 if (fpl->count < SCM_MAX_FD) {
7726 __skb_unlink(skb, head);
7727 spin_unlock_irq(&head->lock);
7728 fpl->fp[fpl->count] = get_file(file);
7729 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7731 spin_lock_irq(&head->lock);
7732 __skb_queue_head(head, skb);
7737 spin_unlock_irq(&head->lock);
7744 return __io_sqe_files_scm(ctx, 1, index);
7750 static int io_queue_rsrc_removal(struct io_rsrc_data *data,
7751 struct io_rsrc_node *node, void *rsrc)
7753 struct io_rsrc_put *prsrc;
7755 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7760 list_add(&prsrc->list, &node->rsrc_list);
7764 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7765 struct io_uring_rsrc_update *up,
7768 struct io_rsrc_data *data = ctx->file_data;
7769 struct io_fixed_file *file_slot;
7774 bool needs_switch = false;
7776 if (check_add_overflow(up->offset, nr_args, &done))
7778 if (done > ctx->nr_user_files)
7780 err = io_rsrc_node_switch_start(ctx);
7784 fds = u64_to_user_ptr(up->data);
7785 for (done = 0; done < nr_args; done++) {
7787 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7791 if (fd == IORING_REGISTER_FILES_SKIP)
7794 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7795 file_slot = io_fixed_file_slot(ctx->file_data, i);
7797 if (file_slot->file_ptr) {
7798 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7799 err = io_queue_rsrc_removal(data, ctx->rsrc_node, file);
7802 file_slot->file_ptr = 0;
7803 needs_switch = true;
7812 * Don't allow io_uring instances to be registered. If
7813 * UNIX isn't enabled, then this causes a reference
7814 * cycle and this instance can never get freed. If UNIX
7815 * is enabled we'll handle it just fine, but there's
7816 * still no point in allowing a ring fd as it doesn't
7817 * support regular read/write anyway.
7819 if (file->f_op == &io_uring_fops) {
7824 io_fixed_file_set(file_slot, file);
7825 err = io_sqe_file_register(ctx, file, i);
7827 file_slot->file_ptr = 0;
7835 io_rsrc_node_switch(ctx, data);
7836 return done ? done : err;
7839 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7842 struct io_uring_rsrc_update up;
7844 if (!ctx->file_data)
7848 if (copy_from_user(&up, arg, sizeof(up)))
7853 return __io_sqe_files_update(ctx, &up, nr_args);
7856 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7858 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7860 req = io_put_req_find_next(req);
7861 return req ? &req->work : NULL;
7864 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7865 struct task_struct *task)
7867 struct io_wq_hash *hash;
7868 struct io_wq_data data;
7869 unsigned int concurrency;
7871 hash = ctx->hash_map;
7873 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7875 return ERR_PTR(-ENOMEM);
7876 refcount_set(&hash->refs, 1);
7877 init_waitqueue_head(&hash->wait);
7878 ctx->hash_map = hash;
7883 data.free_work = io_free_work;
7884 data.do_work = io_wq_submit_work;
7886 /* Do QD, or 4 * CPUS, whatever is smallest */
7887 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7889 return io_wq_create(concurrency, &data);
7892 static int io_uring_alloc_task_context(struct task_struct *task,
7893 struct io_ring_ctx *ctx)
7895 struct io_uring_task *tctx;
7898 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7899 if (unlikely(!tctx))
7902 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7903 if (unlikely(ret)) {
7908 tctx->io_wq = io_init_wq_offload(ctx, task);
7909 if (IS_ERR(tctx->io_wq)) {
7910 ret = PTR_ERR(tctx->io_wq);
7911 percpu_counter_destroy(&tctx->inflight);
7917 init_waitqueue_head(&tctx->wait);
7919 atomic_set(&tctx->in_idle, 0);
7920 task->io_uring = tctx;
7921 spin_lock_init(&tctx->task_lock);
7922 INIT_WQ_LIST(&tctx->task_list);
7923 tctx->task_state = 0;
7924 init_task_work(&tctx->task_work, tctx_task_work);
7928 void __io_uring_free(struct task_struct *tsk)
7930 struct io_uring_task *tctx = tsk->io_uring;
7932 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7933 WARN_ON_ONCE(tctx->io_wq);
7935 percpu_counter_destroy(&tctx->inflight);
7937 tsk->io_uring = NULL;
7940 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7941 struct io_uring_params *p)
7945 /* Retain compatibility with failing for an invalid attach attempt */
7946 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7947 IORING_SETUP_ATTACH_WQ) {
7950 f = fdget(p->wq_fd);
7953 if (f.file->f_op != &io_uring_fops) {
7959 if (ctx->flags & IORING_SETUP_SQPOLL) {
7960 struct task_struct *tsk;
7961 struct io_sq_data *sqd;
7964 sqd = io_get_sq_data(p, &attached);
7970 ctx->sq_creds = get_current_cred();
7972 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7973 if (!ctx->sq_thread_idle)
7974 ctx->sq_thread_idle = HZ;
7977 io_sq_thread_park(sqd);
7978 list_add(&ctx->sqd_list, &sqd->ctx_list);
7979 io_sqd_update_thread_idle(sqd);
7980 /* don't attach to a dying SQPOLL thread, would be racy */
7981 if (attached && !sqd->thread)
7983 io_sq_thread_unpark(sqd);
7990 if (p->flags & IORING_SETUP_SQ_AFF) {
7991 int cpu = p->sq_thread_cpu;
7994 if (cpu >= nr_cpu_ids)
7996 if (!cpu_online(cpu))
8004 sqd->task_pid = current->pid;
8005 sqd->task_tgid = current->tgid;
8006 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8013 ret = io_uring_alloc_task_context(tsk, ctx);
8014 wake_up_new_task(tsk);
8017 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8018 /* Can't have SQ_AFF without SQPOLL */
8025 io_sq_thread_finish(ctx);
8028 complete(&ctx->sq_data->exited);
8032 static inline void __io_unaccount_mem(struct user_struct *user,
8033 unsigned long nr_pages)
8035 atomic_long_sub(nr_pages, &user->locked_vm);
8038 static inline int __io_account_mem(struct user_struct *user,
8039 unsigned long nr_pages)
8041 unsigned long page_limit, cur_pages, new_pages;
8043 /* Don't allow more pages than we can safely lock */
8044 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8047 cur_pages = atomic_long_read(&user->locked_vm);
8048 new_pages = cur_pages + nr_pages;
8049 if (new_pages > page_limit)
8051 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8052 new_pages) != cur_pages);
8057 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8060 __io_unaccount_mem(ctx->user, nr_pages);
8062 if (ctx->mm_account)
8063 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8066 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8071 ret = __io_account_mem(ctx->user, nr_pages);
8076 if (ctx->mm_account)
8077 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8082 static void io_mem_free(void *ptr)
8089 page = virt_to_head_page(ptr);
8090 if (put_page_testzero(page))
8091 free_compound_page(page);
8094 static void *io_mem_alloc(size_t size)
8096 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8097 __GFP_NORETRY | __GFP_ACCOUNT;
8099 return (void *) __get_free_pages(gfp_flags, get_order(size));
8102 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8105 struct io_rings *rings;
8106 size_t off, sq_array_size;
8108 off = struct_size(rings, cqes, cq_entries);
8109 if (off == SIZE_MAX)
8113 off = ALIGN(off, SMP_CACHE_BYTES);
8121 sq_array_size = array_size(sizeof(u32), sq_entries);
8122 if (sq_array_size == SIZE_MAX)
8125 if (check_add_overflow(off, sq_array_size, &off))
8131 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8135 if (!ctx->user_bufs)
8138 for (i = 0; i < ctx->nr_user_bufs; i++) {
8139 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8141 for (j = 0; j < imu->nr_bvecs; j++)
8142 unpin_user_page(imu->bvec[j].bv_page);
8144 if (imu->acct_pages)
8145 io_unaccount_mem(ctx, imu->acct_pages);
8150 kfree(ctx->user_bufs);
8151 ctx->user_bufs = NULL;
8152 ctx->nr_user_bufs = 0;
8156 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8157 void __user *arg, unsigned index)
8159 struct iovec __user *src;
8161 #ifdef CONFIG_COMPAT
8163 struct compat_iovec __user *ciovs;
8164 struct compat_iovec ciov;
8166 ciovs = (struct compat_iovec __user *) arg;
8167 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8170 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8171 dst->iov_len = ciov.iov_len;
8175 src = (struct iovec __user *) arg;
8176 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8182 * Not super efficient, but this is just a registration time. And we do cache
8183 * the last compound head, so generally we'll only do a full search if we don't
8186 * We check if the given compound head page has already been accounted, to
8187 * avoid double accounting it. This allows us to account the full size of the
8188 * page, not just the constituent pages of a huge page.
8190 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8191 int nr_pages, struct page *hpage)
8195 /* check current page array */
8196 for (i = 0; i < nr_pages; i++) {
8197 if (!PageCompound(pages[i]))
8199 if (compound_head(pages[i]) == hpage)
8203 /* check previously registered pages */
8204 for (i = 0; i < ctx->nr_user_bufs; i++) {
8205 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8207 for (j = 0; j < imu->nr_bvecs; j++) {
8208 if (!PageCompound(imu->bvec[j].bv_page))
8210 if (compound_head(imu->bvec[j].bv_page) == hpage)
8218 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8219 int nr_pages, struct io_mapped_ubuf *imu,
8220 struct page **last_hpage)
8224 for (i = 0; i < nr_pages; i++) {
8225 if (!PageCompound(pages[i])) {
8230 hpage = compound_head(pages[i]);
8231 if (hpage == *last_hpage)
8233 *last_hpage = hpage;
8234 if (headpage_already_acct(ctx, pages, i, hpage))
8236 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8240 if (!imu->acct_pages)
8243 ret = io_account_mem(ctx, imu->acct_pages);
8245 imu->acct_pages = 0;
8249 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8250 struct io_mapped_ubuf *imu,
8251 struct page **last_hpage)
8253 struct vm_area_struct **vmas = NULL;
8254 struct page **pages = NULL;
8255 unsigned long off, start, end, ubuf;
8257 int ret, pret, nr_pages, i;
8259 ubuf = (unsigned long) iov->iov_base;
8260 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8261 start = ubuf >> PAGE_SHIFT;
8262 nr_pages = end - start;
8266 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8270 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8275 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8281 mmap_read_lock(current->mm);
8282 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8284 if (pret == nr_pages) {
8285 /* don't support file backed memory */
8286 for (i = 0; i < nr_pages; i++) {
8287 struct vm_area_struct *vma = vmas[i];
8290 !is_file_hugepages(vma->vm_file)) {
8296 ret = pret < 0 ? pret : -EFAULT;
8298 mmap_read_unlock(current->mm);
8301 * if we did partial map, or found file backed vmas,
8302 * release any pages we did get
8305 unpin_user_pages(pages, pret);
8310 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8312 unpin_user_pages(pages, pret);
8317 off = ubuf & ~PAGE_MASK;
8318 size = iov->iov_len;
8319 for (i = 0; i < nr_pages; i++) {
8322 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8323 imu->bvec[i].bv_page = pages[i];
8324 imu->bvec[i].bv_len = vec_len;
8325 imu->bvec[i].bv_offset = off;
8329 /* store original address for later verification */
8331 imu->ubuf_end = ubuf + iov->iov_len;
8332 imu->nr_bvecs = nr_pages;
8340 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8344 if (!nr_args || nr_args > UIO_MAXIOV)
8347 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8349 if (!ctx->user_bufs)
8355 static int io_buffer_validate(struct iovec *iov)
8357 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8360 * Don't impose further limits on the size and buffer
8361 * constraints here, we'll -EINVAL later when IO is
8362 * submitted if they are wrong.
8364 if (!iov->iov_base || !iov->iov_len)
8367 /* arbitrary limit, but we need something */
8368 if (iov->iov_len > SZ_1G)
8371 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8377 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8378 unsigned int nr_args)
8382 struct page *last_hpage = NULL;
8384 ret = io_buffers_map_alloc(ctx, nr_args);
8388 for (i = 0; i < nr_args; i++) {
8389 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8391 ret = io_copy_iov(ctx, &iov, arg, i);
8395 ret = io_buffer_validate(&iov);
8399 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8403 ctx->nr_user_bufs++;
8407 io_sqe_buffers_unregister(ctx);
8412 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8414 __s32 __user *fds = arg;
8420 if (copy_from_user(&fd, fds, sizeof(*fds)))
8423 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8424 if (IS_ERR(ctx->cq_ev_fd)) {
8425 int ret = PTR_ERR(ctx->cq_ev_fd);
8426 ctx->cq_ev_fd = NULL;
8433 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8435 if (ctx->cq_ev_fd) {
8436 eventfd_ctx_put(ctx->cq_ev_fd);
8437 ctx->cq_ev_fd = NULL;
8444 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8446 struct io_buffer *buf;
8447 unsigned long index;
8449 xa_for_each(&ctx->io_buffers, index, buf)
8450 __io_remove_buffers(ctx, buf, index, -1U);
8453 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8455 struct io_kiocb *req, *nxt;
8457 list_for_each_entry_safe(req, nxt, list, compl.list) {
8458 if (tsk && req->task != tsk)
8460 list_del(&req->compl.list);
8461 kmem_cache_free(req_cachep, req);
8465 static void io_req_caches_free(struct io_ring_ctx *ctx)
8467 struct io_submit_state *submit_state = &ctx->submit_state;
8468 struct io_comp_state *cs = &ctx->submit_state.comp;
8470 mutex_lock(&ctx->uring_lock);
8472 if (submit_state->free_reqs) {
8473 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8474 submit_state->reqs);
8475 submit_state->free_reqs = 0;
8478 io_flush_cached_locked_reqs(ctx, cs);
8479 io_req_cache_free(&cs->free_list, NULL);
8480 mutex_unlock(&ctx->uring_lock);
8483 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8485 io_sq_thread_finish(ctx);
8486 io_sqe_buffers_unregister(ctx);
8488 if (ctx->mm_account) {
8489 mmdrop(ctx->mm_account);
8490 ctx->mm_account = NULL;
8493 mutex_lock(&ctx->uring_lock);
8494 io_sqe_files_unregister(ctx);
8496 __io_cqring_overflow_flush(ctx, true);
8497 mutex_unlock(&ctx->uring_lock);
8498 io_eventfd_unregister(ctx);
8499 io_destroy_buffers(ctx);
8501 /* there are no registered resources left, nobody uses it */
8503 io_rsrc_node_destroy(ctx->rsrc_node);
8504 if (ctx->rsrc_backup_node)
8505 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8506 flush_delayed_work(&ctx->rsrc_put_work);
8508 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8509 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8511 #if defined(CONFIG_UNIX)
8512 if (ctx->ring_sock) {
8513 ctx->ring_sock->file = NULL; /* so that iput() is called */
8514 sock_release(ctx->ring_sock);
8518 io_mem_free(ctx->rings);
8519 io_mem_free(ctx->sq_sqes);
8521 percpu_ref_exit(&ctx->refs);
8522 free_uid(ctx->user);
8523 io_req_caches_free(ctx);
8525 io_wq_put_hash(ctx->hash_map);
8526 kfree(ctx->cancel_hash);
8530 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8532 struct io_ring_ctx *ctx = file->private_data;
8535 poll_wait(file, &ctx->cq_wait, wait);
8537 * synchronizes with barrier from wq_has_sleeper call in
8541 if (!io_sqring_full(ctx))
8542 mask |= EPOLLOUT | EPOLLWRNORM;
8545 * Don't flush cqring overflow list here, just do a simple check.
8546 * Otherwise there could possible be ABBA deadlock:
8549 * lock(&ctx->uring_lock);
8551 * lock(&ctx->uring_lock);
8554 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8555 * pushs them to do the flush.
8557 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8558 mask |= EPOLLIN | EPOLLRDNORM;
8563 static int io_uring_fasync(int fd, struct file *file, int on)
8565 struct io_ring_ctx *ctx = file->private_data;
8567 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8570 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8572 const struct cred *creds;
8574 creds = xa_erase(&ctx->personalities, id);
8583 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8585 return io_run_task_work_head(&ctx->exit_task_work);
8588 struct io_tctx_exit {
8589 struct callback_head task_work;
8590 struct completion completion;
8591 struct io_ring_ctx *ctx;
8594 static void io_tctx_exit_cb(struct callback_head *cb)
8596 struct io_uring_task *tctx = current->io_uring;
8597 struct io_tctx_exit *work;
8599 work = container_of(cb, struct io_tctx_exit, task_work);
8601 * When @in_idle, we're in cancellation and it's racy to remove the
8602 * node. It'll be removed by the end of cancellation, just ignore it.
8604 if (!atomic_read(&tctx->in_idle))
8605 io_uring_del_task_file((unsigned long)work->ctx);
8606 complete(&work->completion);
8609 static void io_ring_exit_work(struct work_struct *work)
8611 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8612 unsigned long timeout = jiffies + HZ * 60 * 5;
8613 struct io_tctx_exit exit;
8614 struct io_tctx_node *node;
8617 /* prevent SQPOLL from submitting new requests */
8619 io_sq_thread_park(ctx->sq_data);
8620 list_del_init(&ctx->sqd_list);
8621 io_sqd_update_thread_idle(ctx->sq_data);
8622 io_sq_thread_unpark(ctx->sq_data);
8626 * If we're doing polled IO and end up having requests being
8627 * submitted async (out-of-line), then completions can come in while
8628 * we're waiting for refs to drop. We need to reap these manually,
8629 * as nobody else will be looking for them.
8632 io_uring_try_cancel_requests(ctx, NULL, NULL);
8634 WARN_ON_ONCE(time_after(jiffies, timeout));
8635 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8638 * Some may use context even when all refs and requests have been put,
8639 * and they are free to do so while still holding uring_lock or
8640 * completion_lock, see __io_req_task_submit(). Apart from other work,
8641 * this lock/unlock section also waits them to finish.
8643 mutex_lock(&ctx->uring_lock);
8644 while (!list_empty(&ctx->tctx_list)) {
8645 WARN_ON_ONCE(time_after(jiffies, timeout));
8647 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8650 init_completion(&exit.completion);
8651 init_task_work(&exit.task_work, io_tctx_exit_cb);
8652 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8653 if (WARN_ON_ONCE(ret))
8655 wake_up_process(node->task);
8657 mutex_unlock(&ctx->uring_lock);
8658 wait_for_completion(&exit.completion);
8660 mutex_lock(&ctx->uring_lock);
8662 mutex_unlock(&ctx->uring_lock);
8663 spin_lock_irq(&ctx->completion_lock);
8664 spin_unlock_irq(&ctx->completion_lock);
8666 io_ring_ctx_free(ctx);
8669 /* Returns true if we found and killed one or more timeouts */
8670 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8671 struct files_struct *files)
8673 struct io_kiocb *req, *tmp;
8676 spin_lock_irq(&ctx->completion_lock);
8677 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8678 if (io_match_task(req, tsk, files)) {
8679 io_kill_timeout(req, -ECANCELED);
8684 io_commit_cqring(ctx);
8685 spin_unlock_irq(&ctx->completion_lock);
8687 io_cqring_ev_posted(ctx);
8688 return canceled != 0;
8691 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8693 unsigned long index;
8694 struct creds *creds;
8696 mutex_lock(&ctx->uring_lock);
8697 percpu_ref_kill(&ctx->refs);
8699 __io_cqring_overflow_flush(ctx, true);
8700 xa_for_each(&ctx->personalities, index, creds)
8701 io_unregister_personality(ctx, index);
8702 mutex_unlock(&ctx->uring_lock);
8704 io_kill_timeouts(ctx, NULL, NULL);
8705 io_poll_remove_all(ctx, NULL, NULL);
8707 /* if we failed setting up the ctx, we might not have any rings */
8708 io_iopoll_try_reap_events(ctx);
8710 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8712 * Use system_unbound_wq to avoid spawning tons of event kworkers
8713 * if we're exiting a ton of rings at the same time. It just adds
8714 * noise and overhead, there's no discernable change in runtime
8715 * over using system_wq.
8717 queue_work(system_unbound_wq, &ctx->exit_work);
8720 static int io_uring_release(struct inode *inode, struct file *file)
8722 struct io_ring_ctx *ctx = file->private_data;
8724 file->private_data = NULL;
8725 io_ring_ctx_wait_and_kill(ctx);
8729 struct io_task_cancel {
8730 struct task_struct *task;
8731 struct files_struct *files;
8734 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8736 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8737 struct io_task_cancel *cancel = data;
8740 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8741 unsigned long flags;
8742 struct io_ring_ctx *ctx = req->ctx;
8744 /* protect against races with linked timeouts */
8745 spin_lock_irqsave(&ctx->completion_lock, flags);
8746 ret = io_match_task(req, cancel->task, cancel->files);
8747 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8749 ret = io_match_task(req, cancel->task, cancel->files);
8754 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8755 struct task_struct *task,
8756 struct files_struct *files)
8758 struct io_defer_entry *de;
8761 spin_lock_irq(&ctx->completion_lock);
8762 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8763 if (io_match_task(de->req, task, files)) {
8764 list_cut_position(&list, &ctx->defer_list, &de->list);
8768 spin_unlock_irq(&ctx->completion_lock);
8769 if (list_empty(&list))
8772 while (!list_empty(&list)) {
8773 de = list_first_entry(&list, struct io_defer_entry, list);
8774 list_del_init(&de->list);
8775 io_req_complete_failed(de->req, -ECANCELED);
8781 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8783 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8785 return req->ctx == data;
8788 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8790 struct io_tctx_node *node;
8791 enum io_wq_cancel cret;
8794 mutex_lock(&ctx->uring_lock);
8795 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8796 struct io_uring_task *tctx = node->task->io_uring;
8799 * io_wq will stay alive while we hold uring_lock, because it's
8800 * killed after ctx nodes, which requires to take the lock.
8802 if (!tctx || !tctx->io_wq)
8804 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8805 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8807 mutex_unlock(&ctx->uring_lock);
8812 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8813 struct task_struct *task,
8814 struct files_struct *files)
8816 struct io_task_cancel cancel = { .task = task, .files = files, };
8817 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8820 enum io_wq_cancel cret;
8824 ret |= io_uring_try_cancel_iowq(ctx);
8825 } else if (tctx && tctx->io_wq) {
8827 * Cancels requests of all rings, not only @ctx, but
8828 * it's fine as the task is in exit/exec.
8830 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8832 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8835 /* SQPOLL thread does its own polling */
8836 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8837 (ctx->sq_data && ctx->sq_data->thread == current)) {
8838 while (!list_empty_careful(&ctx->iopoll_list)) {
8839 io_iopoll_try_reap_events(ctx);
8844 ret |= io_cancel_defer_files(ctx, task, files);
8845 ret |= io_poll_remove_all(ctx, task, files);
8846 ret |= io_kill_timeouts(ctx, task, files);
8847 ret |= io_run_task_work();
8848 ret |= io_run_ctx_fallback(ctx);
8855 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8856 struct task_struct *task,
8857 struct files_struct *files)
8859 struct io_kiocb *req;
8862 spin_lock_irq(&ctx->inflight_lock);
8863 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8864 cnt += io_match_task(req, task, files);
8865 spin_unlock_irq(&ctx->inflight_lock);
8869 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8870 struct task_struct *task,
8871 struct files_struct *files)
8873 while (!list_empty_careful(&ctx->inflight_list)) {
8877 inflight = io_uring_count_inflight(ctx, task, files);
8881 io_uring_try_cancel_requests(ctx, task, files);
8883 prepare_to_wait(&task->io_uring->wait, &wait,
8884 TASK_UNINTERRUPTIBLE);
8885 if (inflight == io_uring_count_inflight(ctx, task, files))
8887 finish_wait(&task->io_uring->wait, &wait);
8891 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8893 struct io_uring_task *tctx = current->io_uring;
8894 struct io_tctx_node *node;
8897 if (unlikely(!tctx)) {
8898 ret = io_uring_alloc_task_context(current, ctx);
8901 tctx = current->io_uring;
8903 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8904 node = kmalloc(sizeof(*node), GFP_KERNEL);
8908 node->task = current;
8910 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8917 mutex_lock(&ctx->uring_lock);
8918 list_add(&node->ctx_node, &ctx->tctx_list);
8919 mutex_unlock(&ctx->uring_lock);
8926 * Note that this task has used io_uring. We use it for cancelation purposes.
8928 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8930 struct io_uring_task *tctx = current->io_uring;
8932 if (likely(tctx && tctx->last == ctx))
8934 return __io_uring_add_task_file(ctx);
8938 * Remove this io_uring_file -> task mapping.
8940 static void io_uring_del_task_file(unsigned long index)
8942 struct io_uring_task *tctx = current->io_uring;
8943 struct io_tctx_node *node;
8947 node = xa_erase(&tctx->xa, index);
8951 WARN_ON_ONCE(current != node->task);
8952 WARN_ON_ONCE(list_empty(&node->ctx_node));
8954 mutex_lock(&node->ctx->uring_lock);
8955 list_del(&node->ctx_node);
8956 mutex_unlock(&node->ctx->uring_lock);
8958 if (tctx->last == node->ctx)
8963 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8965 struct io_tctx_node *node;
8966 unsigned long index;
8968 xa_for_each(&tctx->xa, index, node)
8969 io_uring_del_task_file(index);
8971 io_wq_put_and_exit(tctx->io_wq);
8976 static s64 tctx_inflight(struct io_uring_task *tctx)
8978 return percpu_counter_sum(&tctx->inflight);
8981 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8983 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8984 struct io_ring_ctx *ctx = work->ctx;
8985 struct io_sq_data *sqd = ctx->sq_data;
8988 io_uring_cancel_sqpoll(ctx);
8989 complete(&work->completion);
8992 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8994 struct io_sq_data *sqd = ctx->sq_data;
8995 struct io_tctx_exit work = { .ctx = ctx, };
8996 struct task_struct *task;
8998 io_sq_thread_park(sqd);
8999 list_del_init(&ctx->sqd_list);
9000 io_sqd_update_thread_idle(sqd);
9003 init_completion(&work.completion);
9004 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
9005 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
9006 wake_up_process(task);
9008 io_sq_thread_unpark(sqd);
9011 wait_for_completion(&work.completion);
9014 void __io_uring_files_cancel(struct files_struct *files)
9016 struct io_uring_task *tctx = current->io_uring;
9017 struct io_tctx_node *node;
9018 unsigned long index;
9020 /* make sure overflow events are dropped */
9021 atomic_inc(&tctx->in_idle);
9022 xa_for_each(&tctx->xa, index, node) {
9023 struct io_ring_ctx *ctx = node->ctx;
9026 io_sqpoll_cancel_sync(ctx);
9029 io_uring_cancel_files(ctx, current, files);
9031 io_uring_try_cancel_requests(ctx, current, NULL);
9033 atomic_dec(&tctx->in_idle);
9036 io_uring_clean_tctx(tctx);
9039 /* should only be called by SQPOLL task */
9040 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9042 struct io_sq_data *sqd = ctx->sq_data;
9043 struct io_uring_task *tctx = current->io_uring;
9047 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
9049 atomic_inc(&tctx->in_idle);
9051 /* read completions before cancelations */
9052 inflight = tctx_inflight(tctx);
9055 io_uring_try_cancel_requests(ctx, current, NULL);
9057 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9059 * If we've seen completions, retry without waiting. This
9060 * avoids a race where a completion comes in before we did
9061 * prepare_to_wait().
9063 if (inflight == tctx_inflight(tctx))
9065 finish_wait(&tctx->wait, &wait);
9067 atomic_dec(&tctx->in_idle);
9071 * Find any io_uring fd that this task has registered or done IO on, and cancel
9074 void __io_uring_task_cancel(void)
9076 struct io_uring_task *tctx = current->io_uring;
9080 /* make sure overflow events are dropped */
9081 atomic_inc(&tctx->in_idle);
9082 __io_uring_files_cancel(NULL);
9085 /* read completions before cancelations */
9086 inflight = tctx_inflight(tctx);
9089 __io_uring_files_cancel(NULL);
9091 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9094 * If we've seen completions, retry without waiting. This
9095 * avoids a race where a completion comes in before we did
9096 * prepare_to_wait().
9098 if (inflight == tctx_inflight(tctx))
9100 finish_wait(&tctx->wait, &wait);
9103 atomic_dec(&tctx->in_idle);
9105 io_uring_clean_tctx(tctx);
9106 /* all current's requests should be gone, we can kill tctx */
9107 __io_uring_free(current);
9110 static void *io_uring_validate_mmap_request(struct file *file,
9111 loff_t pgoff, size_t sz)
9113 struct io_ring_ctx *ctx = file->private_data;
9114 loff_t offset = pgoff << PAGE_SHIFT;
9119 case IORING_OFF_SQ_RING:
9120 case IORING_OFF_CQ_RING:
9123 case IORING_OFF_SQES:
9127 return ERR_PTR(-EINVAL);
9130 page = virt_to_head_page(ptr);
9131 if (sz > page_size(page))
9132 return ERR_PTR(-EINVAL);
9139 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9141 size_t sz = vma->vm_end - vma->vm_start;
9145 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9147 return PTR_ERR(ptr);
9149 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9150 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9153 #else /* !CONFIG_MMU */
9155 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9157 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9160 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9162 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9165 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9166 unsigned long addr, unsigned long len,
9167 unsigned long pgoff, unsigned long flags)
9171 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9173 return PTR_ERR(ptr);
9175 return (unsigned long) ptr;
9178 #endif /* !CONFIG_MMU */
9180 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9185 if (!io_sqring_full(ctx))
9187 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9189 if (!io_sqring_full(ctx))
9192 } while (!signal_pending(current));
9194 finish_wait(&ctx->sqo_sq_wait, &wait);
9198 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9199 struct __kernel_timespec __user **ts,
9200 const sigset_t __user **sig)
9202 struct io_uring_getevents_arg arg;
9205 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9206 * is just a pointer to the sigset_t.
9208 if (!(flags & IORING_ENTER_EXT_ARG)) {
9209 *sig = (const sigset_t __user *) argp;
9215 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9216 * timespec and sigset_t pointers if good.
9218 if (*argsz != sizeof(arg))
9220 if (copy_from_user(&arg, argp, sizeof(arg)))
9222 *sig = u64_to_user_ptr(arg.sigmask);
9223 *argsz = arg.sigmask_sz;
9224 *ts = u64_to_user_ptr(arg.ts);
9228 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9229 u32, min_complete, u32, flags, const void __user *, argp,
9232 struct io_ring_ctx *ctx;
9239 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9240 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9244 if (unlikely(!f.file))
9248 if (unlikely(f.file->f_op != &io_uring_fops))
9252 ctx = f.file->private_data;
9253 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9257 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9261 * For SQ polling, the thread will do all submissions and completions.
9262 * Just return the requested submit count, and wake the thread if
9266 if (ctx->flags & IORING_SETUP_SQPOLL) {
9267 io_cqring_overflow_flush(ctx, false);
9270 if (unlikely(ctx->sq_data->thread == NULL)) {
9273 if (flags & IORING_ENTER_SQ_WAKEUP)
9274 wake_up(&ctx->sq_data->wait);
9275 if (flags & IORING_ENTER_SQ_WAIT) {
9276 ret = io_sqpoll_wait_sq(ctx);
9280 submitted = to_submit;
9281 } else if (to_submit) {
9282 ret = io_uring_add_task_file(ctx);
9285 mutex_lock(&ctx->uring_lock);
9286 submitted = io_submit_sqes(ctx, to_submit);
9287 mutex_unlock(&ctx->uring_lock);
9289 if (submitted != to_submit)
9292 if (flags & IORING_ENTER_GETEVENTS) {
9293 const sigset_t __user *sig;
9294 struct __kernel_timespec __user *ts;
9296 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9300 min_complete = min(min_complete, ctx->cq_entries);
9303 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9304 * space applications don't need to do io completion events
9305 * polling again, they can rely on io_sq_thread to do polling
9306 * work, which can reduce cpu usage and uring_lock contention.
9308 if (ctx->flags & IORING_SETUP_IOPOLL &&
9309 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9310 ret = io_iopoll_check(ctx, min_complete);
9312 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9317 percpu_ref_put(&ctx->refs);
9320 return submitted ? submitted : ret;
9323 #ifdef CONFIG_PROC_FS
9324 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9325 const struct cred *cred)
9327 struct user_namespace *uns = seq_user_ns(m);
9328 struct group_info *gi;
9333 seq_printf(m, "%5d\n", id);
9334 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9335 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9336 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9337 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9338 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9339 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9340 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9341 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9342 seq_puts(m, "\n\tGroups:\t");
9343 gi = cred->group_info;
9344 for (g = 0; g < gi->ngroups; g++) {
9345 seq_put_decimal_ull(m, g ? " " : "",
9346 from_kgid_munged(uns, gi->gid[g]));
9348 seq_puts(m, "\n\tCapEff:\t");
9349 cap = cred->cap_effective;
9350 CAP_FOR_EACH_U32(__capi)
9351 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9356 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9358 struct io_sq_data *sq = NULL;
9363 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9364 * since fdinfo case grabs it in the opposite direction of normal use
9365 * cases. If we fail to get the lock, we just don't iterate any
9366 * structures that could be going away outside the io_uring mutex.
9368 has_lock = mutex_trylock(&ctx->uring_lock);
9370 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9376 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9377 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9378 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9379 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9380 struct file *f = io_file_from_index(ctx, i);
9383 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9385 seq_printf(m, "%5u: <none>\n", i);
9387 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9388 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9389 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9390 unsigned int len = buf->ubuf_end - buf->ubuf;
9392 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9394 if (has_lock && !xa_empty(&ctx->personalities)) {
9395 unsigned long index;
9396 const struct cred *cred;
9398 seq_printf(m, "Personalities:\n");
9399 xa_for_each(&ctx->personalities, index, cred)
9400 io_uring_show_cred(m, index, cred);
9402 seq_printf(m, "PollList:\n");
9403 spin_lock_irq(&ctx->completion_lock);
9404 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9405 struct hlist_head *list = &ctx->cancel_hash[i];
9406 struct io_kiocb *req;
9408 hlist_for_each_entry(req, list, hash_node)
9409 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9410 req->task->task_works != NULL);
9412 spin_unlock_irq(&ctx->completion_lock);
9414 mutex_unlock(&ctx->uring_lock);
9417 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9419 struct io_ring_ctx *ctx = f->private_data;
9421 if (percpu_ref_tryget(&ctx->refs)) {
9422 __io_uring_show_fdinfo(ctx, m);
9423 percpu_ref_put(&ctx->refs);
9428 static const struct file_operations io_uring_fops = {
9429 .release = io_uring_release,
9430 .mmap = io_uring_mmap,
9432 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9433 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9435 .poll = io_uring_poll,
9436 .fasync = io_uring_fasync,
9437 #ifdef CONFIG_PROC_FS
9438 .show_fdinfo = io_uring_show_fdinfo,
9442 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9443 struct io_uring_params *p)
9445 struct io_rings *rings;
9446 size_t size, sq_array_offset;
9448 /* make sure these are sane, as we already accounted them */
9449 ctx->sq_entries = p->sq_entries;
9450 ctx->cq_entries = p->cq_entries;
9452 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9453 if (size == SIZE_MAX)
9456 rings = io_mem_alloc(size);
9461 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9462 rings->sq_ring_mask = p->sq_entries - 1;
9463 rings->cq_ring_mask = p->cq_entries - 1;
9464 rings->sq_ring_entries = p->sq_entries;
9465 rings->cq_ring_entries = p->cq_entries;
9466 ctx->sq_mask = rings->sq_ring_mask;
9467 ctx->cq_mask = rings->cq_ring_mask;
9469 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9470 if (size == SIZE_MAX) {
9471 io_mem_free(ctx->rings);
9476 ctx->sq_sqes = io_mem_alloc(size);
9477 if (!ctx->sq_sqes) {
9478 io_mem_free(ctx->rings);
9486 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9490 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9494 ret = io_uring_add_task_file(ctx);
9499 fd_install(fd, file);
9504 * Allocate an anonymous fd, this is what constitutes the application
9505 * visible backing of an io_uring instance. The application mmaps this
9506 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9507 * we have to tie this fd to a socket for file garbage collection purposes.
9509 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9512 #if defined(CONFIG_UNIX)
9515 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9518 return ERR_PTR(ret);
9521 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9522 O_RDWR | O_CLOEXEC);
9523 #if defined(CONFIG_UNIX)
9525 sock_release(ctx->ring_sock);
9526 ctx->ring_sock = NULL;
9528 ctx->ring_sock->file = file;
9534 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9535 struct io_uring_params __user *params)
9537 struct io_ring_ctx *ctx;
9543 if (entries > IORING_MAX_ENTRIES) {
9544 if (!(p->flags & IORING_SETUP_CLAMP))
9546 entries = IORING_MAX_ENTRIES;
9550 * Use twice as many entries for the CQ ring. It's possible for the
9551 * application to drive a higher depth than the size of the SQ ring,
9552 * since the sqes are only used at submission time. This allows for
9553 * some flexibility in overcommitting a bit. If the application has
9554 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9555 * of CQ ring entries manually.
9557 p->sq_entries = roundup_pow_of_two(entries);
9558 if (p->flags & IORING_SETUP_CQSIZE) {
9560 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9561 * to a power-of-two, if it isn't already. We do NOT impose
9562 * any cq vs sq ring sizing.
9566 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9567 if (!(p->flags & IORING_SETUP_CLAMP))
9569 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9571 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9572 if (p->cq_entries < p->sq_entries)
9575 p->cq_entries = 2 * p->sq_entries;
9578 ctx = io_ring_ctx_alloc(p);
9581 ctx->compat = in_compat_syscall();
9582 if (!capable(CAP_IPC_LOCK))
9583 ctx->user = get_uid(current_user());
9586 * This is just grabbed for accounting purposes. When a process exits,
9587 * the mm is exited and dropped before the files, hence we need to hang
9588 * on to this mm purely for the purposes of being able to unaccount
9589 * memory (locked/pinned vm). It's not used for anything else.
9591 mmgrab(current->mm);
9592 ctx->mm_account = current->mm;
9594 ret = io_allocate_scq_urings(ctx, p);
9598 ret = io_sq_offload_create(ctx, p);
9602 memset(&p->sq_off, 0, sizeof(p->sq_off));
9603 p->sq_off.head = offsetof(struct io_rings, sq.head);
9604 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9605 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9606 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9607 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9608 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9609 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9611 memset(&p->cq_off, 0, sizeof(p->cq_off));
9612 p->cq_off.head = offsetof(struct io_rings, cq.head);
9613 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9614 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9615 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9616 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9617 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9618 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9620 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9621 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9622 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9623 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9624 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9626 if (copy_to_user(params, p, sizeof(*p))) {
9631 file = io_uring_get_file(ctx);
9633 ret = PTR_ERR(file);
9638 * Install ring fd as the very last thing, so we don't risk someone
9639 * having closed it before we finish setup
9641 ret = io_uring_install_fd(ctx, file);
9643 /* fput will clean it up */
9648 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9651 io_ring_ctx_wait_and_kill(ctx);
9656 * Sets up an aio uring context, and returns the fd. Applications asks for a
9657 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9658 * params structure passed in.
9660 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9662 struct io_uring_params p;
9665 if (copy_from_user(&p, params, sizeof(p)))
9667 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9672 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9673 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9674 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9675 IORING_SETUP_R_DISABLED))
9678 return io_uring_create(entries, &p, params);
9681 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9682 struct io_uring_params __user *, params)
9684 return io_uring_setup(entries, params);
9687 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9689 struct io_uring_probe *p;
9693 size = struct_size(p, ops, nr_args);
9694 if (size == SIZE_MAX)
9696 p = kzalloc(size, GFP_KERNEL);
9701 if (copy_from_user(p, arg, size))
9704 if (memchr_inv(p, 0, size))
9707 p->last_op = IORING_OP_LAST - 1;
9708 if (nr_args > IORING_OP_LAST)
9709 nr_args = IORING_OP_LAST;
9711 for (i = 0; i < nr_args; i++) {
9713 if (!io_op_defs[i].not_supported)
9714 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9719 if (copy_to_user(arg, p, size))
9726 static int io_register_personality(struct io_ring_ctx *ctx)
9728 const struct cred *creds;
9732 creds = get_current_cred();
9734 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9735 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9742 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9743 unsigned int nr_args)
9745 struct io_uring_restriction *res;
9749 /* Restrictions allowed only if rings started disabled */
9750 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9753 /* We allow only a single restrictions registration */
9754 if (ctx->restrictions.registered)
9757 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9760 size = array_size(nr_args, sizeof(*res));
9761 if (size == SIZE_MAX)
9764 res = memdup_user(arg, size);
9766 return PTR_ERR(res);
9770 for (i = 0; i < nr_args; i++) {
9771 switch (res[i].opcode) {
9772 case IORING_RESTRICTION_REGISTER_OP:
9773 if (res[i].register_op >= IORING_REGISTER_LAST) {
9778 __set_bit(res[i].register_op,
9779 ctx->restrictions.register_op);
9781 case IORING_RESTRICTION_SQE_OP:
9782 if (res[i].sqe_op >= IORING_OP_LAST) {
9787 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9789 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9790 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9792 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9793 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9802 /* Reset all restrictions if an error happened */
9804 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9806 ctx->restrictions.registered = true;
9812 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9814 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9817 if (ctx->restrictions.registered)
9818 ctx->restricted = 1;
9820 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9821 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9822 wake_up(&ctx->sq_data->wait);
9826 static bool io_register_op_must_quiesce(int op)
9829 case IORING_REGISTER_FILES:
9830 case IORING_UNREGISTER_FILES:
9831 case IORING_REGISTER_FILES_UPDATE:
9832 case IORING_REGISTER_PROBE:
9833 case IORING_REGISTER_PERSONALITY:
9834 case IORING_UNREGISTER_PERSONALITY:
9841 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9842 void __user *arg, unsigned nr_args)
9843 __releases(ctx->uring_lock)
9844 __acquires(ctx->uring_lock)
9849 * We're inside the ring mutex, if the ref is already dying, then
9850 * someone else killed the ctx or is already going through
9851 * io_uring_register().
9853 if (percpu_ref_is_dying(&ctx->refs))
9856 if (io_register_op_must_quiesce(opcode)) {
9857 percpu_ref_kill(&ctx->refs);
9860 * Drop uring mutex before waiting for references to exit. If
9861 * another thread is currently inside io_uring_enter() it might
9862 * need to grab the uring_lock to make progress. If we hold it
9863 * here across the drain wait, then we can deadlock. It's safe
9864 * to drop the mutex here, since no new references will come in
9865 * after we've killed the percpu ref.
9867 mutex_unlock(&ctx->uring_lock);
9869 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9872 ret = io_run_task_work_sig();
9877 mutex_lock(&ctx->uring_lock);
9880 percpu_ref_resurrect(&ctx->refs);
9885 if (ctx->restricted) {
9886 if (opcode >= IORING_REGISTER_LAST) {
9891 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9898 case IORING_REGISTER_BUFFERS:
9899 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9901 case IORING_UNREGISTER_BUFFERS:
9905 ret = io_sqe_buffers_unregister(ctx);
9907 case IORING_REGISTER_FILES:
9908 ret = io_sqe_files_register(ctx, arg, nr_args);
9910 case IORING_UNREGISTER_FILES:
9914 ret = io_sqe_files_unregister(ctx);
9916 case IORING_REGISTER_FILES_UPDATE:
9917 ret = io_sqe_files_update(ctx, arg, nr_args);
9919 case IORING_REGISTER_EVENTFD:
9920 case IORING_REGISTER_EVENTFD_ASYNC:
9924 ret = io_eventfd_register(ctx, arg);
9927 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9928 ctx->eventfd_async = 1;
9930 ctx->eventfd_async = 0;
9932 case IORING_UNREGISTER_EVENTFD:
9936 ret = io_eventfd_unregister(ctx);
9938 case IORING_REGISTER_PROBE:
9940 if (!arg || nr_args > 256)
9942 ret = io_probe(ctx, arg, nr_args);
9944 case IORING_REGISTER_PERSONALITY:
9948 ret = io_register_personality(ctx);
9950 case IORING_UNREGISTER_PERSONALITY:
9954 ret = io_unregister_personality(ctx, nr_args);
9956 case IORING_REGISTER_ENABLE_RINGS:
9960 ret = io_register_enable_rings(ctx);
9962 case IORING_REGISTER_RESTRICTIONS:
9963 ret = io_register_restrictions(ctx, arg, nr_args);
9971 if (io_register_op_must_quiesce(opcode)) {
9972 /* bring the ctx back to life */
9973 percpu_ref_reinit(&ctx->refs);
9975 reinit_completion(&ctx->ref_comp);
9980 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9981 void __user *, arg, unsigned int, nr_args)
9983 struct io_ring_ctx *ctx;
9992 if (f.file->f_op != &io_uring_fops)
9995 ctx = f.file->private_data;
9999 mutex_lock(&ctx->uring_lock);
10000 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10001 mutex_unlock(&ctx->uring_lock);
10002 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10003 ctx->cq_ev_fd != NULL, ret);
10009 static int __init io_uring_init(void)
10011 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10012 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10013 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10016 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10017 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10018 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10019 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10020 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10021 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10022 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10023 BUILD_BUG_SQE_ELEM(8, __u64, off);
10024 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10025 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10026 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10027 BUILD_BUG_SQE_ELEM(24, __u32, len);
10028 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10029 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10030 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10031 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10032 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10033 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10034 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10035 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10036 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10037 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10038 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10039 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10040 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10041 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10042 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10043 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10044 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10045 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10046 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10048 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10049 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10050 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10054 __initcall(io_uring_init);