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>
81 #include <linux/freezer.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
129 struct io_uring sq, cq;
131 * Bitmasks to apply to head and tail offsets (constant, equals
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
153 * Written by the kernel, shouldn't be modified by the
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
163 * Written by the application, shouldn't be modified by the
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending than there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
191 enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
193 IO_URING_F_COMPLETE_DEFER = 2,
196 struct io_mapped_ubuf {
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
207 struct list_head list;
214 struct fixed_rsrc_table {
218 struct fixed_rsrc_ref_node {
219 struct percpu_ref refs;
220 struct list_head node;
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
225 struct llist_node llist;
229 struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
231 struct io_ring_ctx *ctx;
233 struct fixed_rsrc_ref_node *node;
234 struct percpu_ref refs;
235 struct completion done;
240 struct list_head list;
246 struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
261 struct rw_semaphore rw_lock;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
265 struct list_head ctx_new_list;
267 struct task_struct *thread;
268 struct wait_queue_head wait;
270 unsigned sq_thread_idle;
275 struct completion startup;
276 struct completion exited;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
401 struct completion sq_thread_comp;
403 #if defined(CONFIG_UNIX)
404 struct socket *ring_sock;
407 struct idr io_buffer_idr;
409 struct idr personality_idr;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
449 struct callback_head *exit_task_work;
451 struct wait_queue_head hash_wait;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work;
455 struct list_head tctx_list;
459 * First field must be the file pointer in all the
460 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
462 struct io_poll_iocb {
464 struct wait_queue_head *head;
468 struct wait_queue_entry wait;
471 struct io_poll_remove {
481 struct io_timeout_data {
482 struct io_kiocb *req;
483 struct hrtimer timer;
484 struct timespec64 ts;
485 enum hrtimer_mode mode;
490 struct sockaddr __user *addr;
491 int __user *addr_len;
493 unsigned long nofile;
513 struct list_head list;
514 /* head of the link, used by linked timeouts only */
515 struct io_kiocb *head;
518 struct io_timeout_rem {
523 struct timespec64 ts;
528 /* NOTE: kiocb has the file as the first member, so don't do it here */
536 struct sockaddr __user *addr;
543 struct user_msghdr __user *umsg;
549 struct io_buffer *kbuf;
555 struct filename *filename;
557 unsigned long nofile;
560 struct io_rsrc_update {
586 struct epoll_event event;
590 struct file *file_out;
591 struct file *file_in;
598 struct io_provide_buf {
612 const char __user *filename;
613 struct statx __user *buffer;
625 struct filename *oldpath;
626 struct filename *newpath;
634 struct filename *filename;
637 struct io_completion {
639 struct list_head list;
643 struct io_async_connect {
644 struct sockaddr_storage address;
647 struct io_async_msghdr {
648 struct iovec fast_iov[UIO_FASTIOV];
649 /* points to an allocated iov, if NULL we use fast_iov instead */
650 struct iovec *free_iov;
651 struct sockaddr __user *uaddr;
653 struct sockaddr_storage addr;
657 struct iovec fast_iov[UIO_FASTIOV];
658 const struct iovec *free_iovec;
659 struct iov_iter iter;
661 struct wait_page_queue wpq;
665 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
666 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
667 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
668 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
669 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
670 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
676 REQ_F_LINK_TIMEOUT_BIT,
678 REQ_F_NEED_CLEANUP_BIT,
680 REQ_F_BUFFER_SELECTED_BIT,
681 REQ_F_NO_FILE_TABLE_BIT,
682 REQ_F_LTIMEOUT_ACTIVE_BIT,
683 REQ_F_COMPLETE_INLINE_BIT,
685 /* not a real bit, just to check we're not overflowing the space */
691 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
692 /* drain existing IO first */
693 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
695 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
696 /* doesn't sever on completion < 0 */
697 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
699 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
700 /* IOSQE_BUFFER_SELECT */
701 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
703 /* fail rest of links */
704 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
705 /* on inflight list, should be cancelled and waited on exit reliably */
706 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
707 /* read/write uses file position */
708 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
709 /* must not punt to workers */
710 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
711 /* has or had linked timeout */
712 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
714 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
716 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
717 /* already went through poll handler */
718 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
719 /* buffer already selected */
720 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
721 /* doesn't need file table for this request */
722 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
723 /* linked timeout is active, i.e. prepared by link's head */
724 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
725 /* completion is deferred through io_comp_state */
726 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
730 struct io_poll_iocb poll;
731 struct io_poll_iocb *double_poll;
734 struct io_task_work {
735 struct io_wq_work_node node;
736 task_work_func_t func;
740 * NOTE! Each of the iocb union members has the file pointer
741 * as the first entry in their struct definition. So you can
742 * access the file pointer through any of the sub-structs,
743 * or directly as just 'ki_filp' in this struct.
749 struct io_poll_iocb poll;
750 struct io_poll_remove poll_remove;
751 struct io_accept accept;
753 struct io_cancel cancel;
754 struct io_timeout timeout;
755 struct io_timeout_rem timeout_rem;
756 struct io_connect connect;
757 struct io_sr_msg sr_msg;
759 struct io_close close;
760 struct io_rsrc_update rsrc_update;
761 struct io_fadvise fadvise;
762 struct io_madvise madvise;
763 struct io_epoll epoll;
764 struct io_splice splice;
765 struct io_provide_buf pbuf;
766 struct io_statx statx;
767 struct io_shutdown shutdown;
768 struct io_rename rename;
769 struct io_unlink unlink;
770 /* use only after cleaning per-op data, see io_clean_op() */
771 struct io_completion compl;
774 /* opcode allocated if it needs to store data for async defer */
777 /* polled IO has completed */
783 struct io_ring_ctx *ctx;
786 struct task_struct *task;
789 struct io_kiocb *link;
790 struct percpu_ref *fixed_rsrc_refs;
793 * 1. used with ctx->iopoll_list with reads/writes
794 * 2. to track reqs with ->files (see io_op_def::file_table)
796 struct list_head inflight_entry;
798 struct io_task_work io_task_work;
799 struct callback_head task_work;
801 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
802 struct hlist_node hash_node;
803 struct async_poll *apoll;
804 struct io_wq_work work;
807 struct io_tctx_node {
808 struct list_head ctx_node;
809 struct task_struct *task;
810 struct io_ring_ctx *ctx;
813 struct io_defer_entry {
814 struct list_head list;
815 struct io_kiocb *req;
820 /* needs req->file assigned */
821 unsigned needs_file : 1;
822 /* hash wq insertion if file is a regular file */
823 unsigned hash_reg_file : 1;
824 /* unbound wq insertion if file is a non-regular file */
825 unsigned unbound_nonreg_file : 1;
826 /* opcode is not supported by this kernel */
827 unsigned not_supported : 1;
828 /* set if opcode supports polled "wait" */
830 unsigned pollout : 1;
831 /* op supports buffer selection */
832 unsigned buffer_select : 1;
833 /* must always have async data allocated */
834 unsigned needs_async_data : 1;
835 /* should block plug */
837 /* size of async data needed, if any */
838 unsigned short async_size;
841 static const struct io_op_def io_op_defs[] = {
842 [IORING_OP_NOP] = {},
843 [IORING_OP_READV] = {
845 .unbound_nonreg_file = 1,
848 .needs_async_data = 1,
850 .async_size = sizeof(struct io_async_rw),
852 [IORING_OP_WRITEV] = {
855 .unbound_nonreg_file = 1,
857 .needs_async_data = 1,
859 .async_size = sizeof(struct io_async_rw),
861 [IORING_OP_FSYNC] = {
864 [IORING_OP_READ_FIXED] = {
866 .unbound_nonreg_file = 1,
869 .async_size = sizeof(struct io_async_rw),
871 [IORING_OP_WRITE_FIXED] = {
874 .unbound_nonreg_file = 1,
877 .async_size = sizeof(struct io_async_rw),
879 [IORING_OP_POLL_ADD] = {
881 .unbound_nonreg_file = 1,
883 [IORING_OP_POLL_REMOVE] = {},
884 [IORING_OP_SYNC_FILE_RANGE] = {
887 [IORING_OP_SENDMSG] = {
889 .unbound_nonreg_file = 1,
891 .needs_async_data = 1,
892 .async_size = sizeof(struct io_async_msghdr),
894 [IORING_OP_RECVMSG] = {
896 .unbound_nonreg_file = 1,
899 .needs_async_data = 1,
900 .async_size = sizeof(struct io_async_msghdr),
902 [IORING_OP_TIMEOUT] = {
903 .needs_async_data = 1,
904 .async_size = sizeof(struct io_timeout_data),
906 [IORING_OP_TIMEOUT_REMOVE] = {
907 /* used by timeout updates' prep() */
909 [IORING_OP_ACCEPT] = {
911 .unbound_nonreg_file = 1,
914 [IORING_OP_ASYNC_CANCEL] = {},
915 [IORING_OP_LINK_TIMEOUT] = {
916 .needs_async_data = 1,
917 .async_size = sizeof(struct io_timeout_data),
919 [IORING_OP_CONNECT] = {
921 .unbound_nonreg_file = 1,
923 .needs_async_data = 1,
924 .async_size = sizeof(struct io_async_connect),
926 [IORING_OP_FALLOCATE] = {
929 [IORING_OP_OPENAT] = {},
930 [IORING_OP_CLOSE] = {},
931 [IORING_OP_FILES_UPDATE] = {},
932 [IORING_OP_STATX] = {},
935 .unbound_nonreg_file = 1,
939 .async_size = sizeof(struct io_async_rw),
941 [IORING_OP_WRITE] = {
943 .unbound_nonreg_file = 1,
946 .async_size = sizeof(struct io_async_rw),
948 [IORING_OP_FADVISE] = {
951 [IORING_OP_MADVISE] = {},
954 .unbound_nonreg_file = 1,
959 .unbound_nonreg_file = 1,
963 [IORING_OP_OPENAT2] = {
965 [IORING_OP_EPOLL_CTL] = {
966 .unbound_nonreg_file = 1,
968 [IORING_OP_SPLICE] = {
971 .unbound_nonreg_file = 1,
973 [IORING_OP_PROVIDE_BUFFERS] = {},
974 [IORING_OP_REMOVE_BUFFERS] = {},
978 .unbound_nonreg_file = 1,
980 [IORING_OP_SHUTDOWN] = {
983 [IORING_OP_RENAMEAT] = {},
984 [IORING_OP_UNLINKAT] = {},
987 static void io_uring_del_task_file(unsigned long index);
988 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
989 struct task_struct *task,
990 struct files_struct *files);
991 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
992 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
993 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
994 struct io_ring_ctx *ctx);
995 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
997 static bool io_rw_reissue(struct io_kiocb *req);
998 static void io_cqring_fill_event(struct io_kiocb *req, long res);
999 static void io_put_req(struct io_kiocb *req);
1000 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1001 static void io_double_put_req(struct io_kiocb *req);
1002 static void io_dismantle_req(struct io_kiocb *req);
1003 static void io_put_task(struct task_struct *task, int nr);
1004 static void io_queue_next(struct io_kiocb *req);
1005 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1006 static void __io_queue_linked_timeout(struct io_kiocb *req);
1007 static void io_queue_linked_timeout(struct io_kiocb *req);
1008 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1009 struct io_uring_rsrc_update *ip,
1011 static void __io_clean_op(struct io_kiocb *req);
1012 static struct file *io_file_get(struct io_submit_state *state,
1013 struct io_kiocb *req, int fd, bool fixed);
1014 static void __io_queue_sqe(struct io_kiocb *req);
1015 static void io_rsrc_put_work(struct work_struct *work);
1017 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1018 struct iov_iter *iter, bool needs_lock);
1019 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1020 const struct iovec *fast_iov,
1021 struct iov_iter *iter, bool force);
1022 static void io_req_task_queue(struct io_kiocb *req);
1023 static void io_submit_flush_completions(struct io_comp_state *cs,
1024 struct io_ring_ctx *ctx);
1026 static struct kmem_cache *req_cachep;
1028 static const struct file_operations io_uring_fops;
1030 struct sock *io_uring_get_socket(struct file *file)
1032 #if defined(CONFIG_UNIX)
1033 if (file->f_op == &io_uring_fops) {
1034 struct io_ring_ctx *ctx = file->private_data;
1036 return ctx->ring_sock->sk;
1041 EXPORT_SYMBOL(io_uring_get_socket);
1043 #define io_for_each_link(pos, head) \
1044 for (pos = (head); pos; pos = pos->link)
1046 static inline void io_clean_op(struct io_kiocb *req)
1048 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1052 static inline void io_set_resource_node(struct io_kiocb *req)
1054 struct io_ring_ctx *ctx = req->ctx;
1056 if (!req->fixed_rsrc_refs) {
1057 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1058 percpu_ref_get(req->fixed_rsrc_refs);
1062 static bool io_match_task(struct io_kiocb *head,
1063 struct task_struct *task,
1064 struct files_struct *files)
1066 struct io_kiocb *req;
1068 if (task && head->task != task) {
1069 /* in terms of cancelation, always match if req task is dead */
1070 if (head->task->flags & PF_EXITING)
1077 io_for_each_link(req, head) {
1078 if (req->flags & REQ_F_INFLIGHT)
1080 if (req->task->files == files)
1086 static inline void req_set_fail_links(struct io_kiocb *req)
1088 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1089 req->flags |= REQ_F_FAIL_LINK;
1092 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1094 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1096 complete(&ctx->ref_comp);
1099 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1101 return !req->timeout.off;
1104 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1106 struct io_ring_ctx *ctx;
1109 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1114 * Use 5 bits less than the max cq entries, that should give us around
1115 * 32 entries per hash list if totally full and uniformly spread.
1117 hash_bits = ilog2(p->cq_entries);
1121 ctx->cancel_hash_bits = hash_bits;
1122 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1124 if (!ctx->cancel_hash)
1126 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1128 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1129 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1132 ctx->flags = p->flags;
1133 init_waitqueue_head(&ctx->sqo_sq_wait);
1134 INIT_LIST_HEAD(&ctx->sqd_list);
1135 init_waitqueue_head(&ctx->cq_wait);
1136 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1137 init_completion(&ctx->ref_comp);
1138 init_completion(&ctx->sq_thread_comp);
1139 idr_init(&ctx->io_buffer_idr);
1140 idr_init(&ctx->personality_idr);
1141 mutex_init(&ctx->uring_lock);
1142 init_waitqueue_head(&ctx->wait);
1143 spin_lock_init(&ctx->completion_lock);
1144 INIT_LIST_HEAD(&ctx->iopoll_list);
1145 INIT_LIST_HEAD(&ctx->defer_list);
1146 INIT_LIST_HEAD(&ctx->timeout_list);
1147 spin_lock_init(&ctx->inflight_lock);
1148 INIT_LIST_HEAD(&ctx->inflight_list);
1149 spin_lock_init(&ctx->rsrc_ref_lock);
1150 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1151 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1152 init_llist_head(&ctx->rsrc_put_llist);
1153 INIT_LIST_HEAD(&ctx->tctx_list);
1154 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1155 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1158 kfree(ctx->cancel_hash);
1163 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1165 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1166 struct io_ring_ctx *ctx = req->ctx;
1168 return seq != ctx->cached_cq_tail
1169 + READ_ONCE(ctx->cached_cq_overflow);
1175 static void io_req_track_inflight(struct io_kiocb *req)
1177 struct io_ring_ctx *ctx = req->ctx;
1179 if (!(req->flags & REQ_F_INFLIGHT)) {
1180 req->flags |= REQ_F_INFLIGHT;
1182 spin_lock_irq(&ctx->inflight_lock);
1183 list_add(&req->inflight_entry, &ctx->inflight_list);
1184 spin_unlock_irq(&ctx->inflight_lock);
1188 static void io_prep_async_work(struct io_kiocb *req)
1190 const struct io_op_def *def = &io_op_defs[req->opcode];
1191 struct io_ring_ctx *ctx = req->ctx;
1193 if (!req->work.creds)
1194 req->work.creds = get_current_cred();
1196 if (req->flags & REQ_F_FORCE_ASYNC)
1197 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1199 if (req->flags & REQ_F_ISREG) {
1200 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1201 io_wq_hash_work(&req->work, file_inode(req->file));
1203 if (def->unbound_nonreg_file)
1204 req->work.flags |= IO_WQ_WORK_UNBOUND;
1208 static void io_prep_async_link(struct io_kiocb *req)
1210 struct io_kiocb *cur;
1212 io_for_each_link(cur, req)
1213 io_prep_async_work(cur);
1216 static void io_queue_async_work(struct io_kiocb *req)
1218 struct io_ring_ctx *ctx = req->ctx;
1219 struct io_kiocb *link = io_prep_linked_timeout(req);
1220 struct io_uring_task *tctx = req->task->io_uring;
1223 BUG_ON(!tctx->io_wq);
1225 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1226 &req->work, req->flags);
1227 /* init ->work of the whole link before punting */
1228 io_prep_async_link(req);
1229 io_wq_enqueue(tctx->io_wq, &req->work);
1231 io_queue_linked_timeout(link);
1234 static void io_kill_timeout(struct io_kiocb *req)
1236 struct io_timeout_data *io = req->async_data;
1239 ret = hrtimer_try_to_cancel(&io->timer);
1241 atomic_set(&req->ctx->cq_timeouts,
1242 atomic_read(&req->ctx->cq_timeouts) + 1);
1243 list_del_init(&req->timeout.list);
1244 io_cqring_fill_event(req, 0);
1245 io_put_req_deferred(req, 1);
1250 * Returns true if we found and killed one or more timeouts
1252 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1253 struct files_struct *files)
1255 struct io_kiocb *req, *tmp;
1258 spin_lock_irq(&ctx->completion_lock);
1259 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1260 if (io_match_task(req, tsk, files)) {
1261 io_kill_timeout(req);
1265 spin_unlock_irq(&ctx->completion_lock);
1266 return canceled != 0;
1269 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1272 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1273 struct io_defer_entry, list);
1275 if (req_need_defer(de->req, de->seq))
1277 list_del_init(&de->list);
1278 io_req_task_queue(de->req);
1280 } while (!list_empty(&ctx->defer_list));
1283 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1287 if (list_empty(&ctx->timeout_list))
1290 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1293 u32 events_needed, events_got;
1294 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1295 struct io_kiocb, timeout.list);
1297 if (io_is_timeout_noseq(req))
1301 * Since seq can easily wrap around over time, subtract
1302 * the last seq at which timeouts were flushed before comparing.
1303 * Assuming not more than 2^31-1 events have happened since,
1304 * these subtractions won't have wrapped, so we can check if
1305 * target is in [last_seq, current_seq] by comparing the two.
1307 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1308 events_got = seq - ctx->cq_last_tm_flush;
1309 if (events_got < events_needed)
1312 list_del_init(&req->timeout.list);
1313 io_kill_timeout(req);
1314 } while (!list_empty(&ctx->timeout_list));
1316 ctx->cq_last_tm_flush = seq;
1319 static void io_commit_cqring(struct io_ring_ctx *ctx)
1321 io_flush_timeouts(ctx);
1323 /* order cqe stores with ring update */
1324 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1326 if (unlikely(!list_empty(&ctx->defer_list)))
1327 __io_queue_deferred(ctx);
1330 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1332 struct io_rings *r = ctx->rings;
1334 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1337 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1339 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1342 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1344 struct io_rings *rings = ctx->rings;
1348 * writes to the cq entry need to come after reading head; the
1349 * control dependency is enough as we're using WRITE_ONCE to
1352 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1355 tail = ctx->cached_cq_tail++;
1356 return &rings->cqes[tail & ctx->cq_mask];
1359 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1363 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1365 if (!ctx->eventfd_async)
1367 return io_wq_current_is_worker();
1370 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1372 /* see waitqueue_active() comment */
1375 if (waitqueue_active(&ctx->wait))
1376 wake_up(&ctx->wait);
1377 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1378 wake_up(&ctx->sq_data->wait);
1379 if (io_should_trigger_evfd(ctx))
1380 eventfd_signal(ctx->cq_ev_fd, 1);
1381 if (waitqueue_active(&ctx->cq_wait)) {
1382 wake_up_interruptible(&ctx->cq_wait);
1383 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1387 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1389 /* see waitqueue_active() comment */
1392 if (ctx->flags & IORING_SETUP_SQPOLL) {
1393 if (waitqueue_active(&ctx->wait))
1394 wake_up(&ctx->wait);
1396 if (io_should_trigger_evfd(ctx))
1397 eventfd_signal(ctx->cq_ev_fd, 1);
1398 if (waitqueue_active(&ctx->cq_wait)) {
1399 wake_up_interruptible(&ctx->cq_wait);
1400 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1404 /* Returns true if there are no backlogged entries after the flush */
1405 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1406 struct task_struct *tsk,
1407 struct files_struct *files)
1409 struct io_rings *rings = ctx->rings;
1410 struct io_kiocb *req, *tmp;
1411 struct io_uring_cqe *cqe;
1412 unsigned long flags;
1413 bool all_flushed, posted;
1416 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1420 spin_lock_irqsave(&ctx->completion_lock, flags);
1421 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1422 if (!io_match_task(req, tsk, files))
1425 cqe = io_get_cqring(ctx);
1429 list_move(&req->compl.list, &list);
1431 WRITE_ONCE(cqe->user_data, req->user_data);
1432 WRITE_ONCE(cqe->res, req->result);
1433 WRITE_ONCE(cqe->flags, req->compl.cflags);
1435 ctx->cached_cq_overflow++;
1436 WRITE_ONCE(ctx->rings->cq_overflow,
1437 ctx->cached_cq_overflow);
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);
1455 while (!list_empty(&list)) {
1456 req = list_first_entry(&list, struct io_kiocb, compl.list);
1457 list_del(&req->compl.list);
1464 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1465 struct task_struct *tsk,
1466 struct files_struct *files)
1470 if (test_bit(0, &ctx->cq_check_overflow)) {
1471 /* iopoll syncs against uring_lock, not completion_lock */
1472 if (ctx->flags & IORING_SETUP_IOPOLL)
1473 mutex_lock(&ctx->uring_lock);
1474 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1475 if (ctx->flags & IORING_SETUP_IOPOLL)
1476 mutex_unlock(&ctx->uring_lock);
1482 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1484 struct io_ring_ctx *ctx = req->ctx;
1485 struct io_uring_cqe *cqe;
1487 trace_io_uring_complete(ctx, req->user_data, res);
1490 * If we can't get a cq entry, userspace overflowed the
1491 * submission (by quite a lot). Increment the overflow count in
1494 cqe = io_get_cqring(ctx);
1496 WRITE_ONCE(cqe->user_data, req->user_data);
1497 WRITE_ONCE(cqe->res, res);
1498 WRITE_ONCE(cqe->flags, cflags);
1499 } else if (ctx->cq_overflow_flushed ||
1500 atomic_read(&req->task->io_uring->in_idle)) {
1502 * If we're in ring overflow flush mode, or in task cancel mode,
1503 * then we cannot store the request for later flushing, we need
1504 * to drop it on the floor.
1506 ctx->cached_cq_overflow++;
1507 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1509 if (list_empty(&ctx->cq_overflow_list)) {
1510 set_bit(0, &ctx->sq_check_overflow);
1511 set_bit(0, &ctx->cq_check_overflow);
1512 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1516 req->compl.cflags = cflags;
1517 refcount_inc(&req->refs);
1518 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1522 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1524 __io_cqring_fill_event(req, res, 0);
1527 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1528 unsigned int cflags)
1530 struct io_ring_ctx *ctx = req->ctx;
1531 unsigned long flags;
1533 spin_lock_irqsave(&ctx->completion_lock, flags);
1534 __io_cqring_fill_event(req, res, cflags);
1535 io_commit_cqring(ctx);
1537 * If we're the last reference to this request, add to our locked
1540 if (refcount_dec_and_test(&req->refs)) {
1541 struct io_comp_state *cs = &ctx->submit_state.comp;
1543 io_dismantle_req(req);
1544 io_put_task(req->task, 1);
1545 list_add(&req->compl.list, &cs->locked_free_list);
1546 cs->locked_free_nr++;
1549 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1551 io_cqring_ev_posted(ctx);
1554 percpu_ref_put(&ctx->refs);
1558 static void io_req_complete_state(struct io_kiocb *req, long res,
1559 unsigned int cflags)
1563 req->compl.cflags = cflags;
1564 req->flags |= REQ_F_COMPLETE_INLINE;
1567 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1568 long res, unsigned cflags)
1570 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1571 io_req_complete_state(req, res, cflags);
1573 io_req_complete_post(req, res, cflags);
1576 static inline void io_req_complete(struct io_kiocb *req, long res)
1578 __io_req_complete(req, 0, res, 0);
1581 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1583 struct io_submit_state *state = &ctx->submit_state;
1584 struct io_comp_state *cs = &state->comp;
1585 struct io_kiocb *req = NULL;
1588 * If we have more than a batch's worth of requests in our IRQ side
1589 * locked cache, grab the lock and move them over to our submission
1592 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1593 spin_lock_irq(&ctx->completion_lock);
1594 list_splice_init(&cs->locked_free_list, &cs->free_list);
1595 cs->locked_free_nr = 0;
1596 spin_unlock_irq(&ctx->completion_lock);
1599 while (!list_empty(&cs->free_list)) {
1600 req = list_first_entry(&cs->free_list, struct io_kiocb,
1602 list_del(&req->compl.list);
1603 state->reqs[state->free_reqs++] = req;
1604 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1611 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1613 struct io_submit_state *state = &ctx->submit_state;
1615 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1617 if (!state->free_reqs) {
1618 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1621 if (io_flush_cached_reqs(ctx))
1624 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1628 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1629 * retry single alloc to be on the safe side.
1631 if (unlikely(ret <= 0)) {
1632 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1633 if (!state->reqs[0])
1637 state->free_reqs = ret;
1641 return state->reqs[state->free_reqs];
1644 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1651 static void io_dismantle_req(struct io_kiocb *req)
1655 if (req->async_data)
1656 kfree(req->async_data);
1658 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1659 if (req->fixed_rsrc_refs)
1660 percpu_ref_put(req->fixed_rsrc_refs);
1661 if (req->work.creds) {
1662 put_cred(req->work.creds);
1663 req->work.creds = NULL;
1666 if (req->flags & REQ_F_INFLIGHT) {
1667 struct io_ring_ctx *ctx = req->ctx;
1668 unsigned long flags;
1670 spin_lock_irqsave(&ctx->inflight_lock, flags);
1671 list_del(&req->inflight_entry);
1672 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1673 req->flags &= ~REQ_F_INFLIGHT;
1677 /* must to be called somewhat shortly after putting a request */
1678 static inline void io_put_task(struct task_struct *task, int nr)
1680 struct io_uring_task *tctx = task->io_uring;
1682 percpu_counter_sub(&tctx->inflight, nr);
1683 if (unlikely(atomic_read(&tctx->in_idle)))
1684 wake_up(&tctx->wait);
1685 put_task_struct_many(task, nr);
1688 static void __io_free_req(struct io_kiocb *req)
1690 struct io_ring_ctx *ctx = req->ctx;
1692 io_dismantle_req(req);
1693 io_put_task(req->task, 1);
1695 kmem_cache_free(req_cachep, req);
1696 percpu_ref_put(&ctx->refs);
1699 static inline void io_remove_next_linked(struct io_kiocb *req)
1701 struct io_kiocb *nxt = req->link;
1703 req->link = nxt->link;
1707 static void io_kill_linked_timeout(struct io_kiocb *req)
1709 struct io_ring_ctx *ctx = req->ctx;
1710 struct io_kiocb *link;
1711 bool cancelled = false;
1712 unsigned long flags;
1714 spin_lock_irqsave(&ctx->completion_lock, flags);
1718 * Can happen if a linked timeout fired and link had been like
1719 * req -> link t-out -> link t-out [-> ...]
1721 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1722 struct io_timeout_data *io = link->async_data;
1725 io_remove_next_linked(req);
1726 link->timeout.head = NULL;
1727 ret = hrtimer_try_to_cancel(&io->timer);
1729 io_cqring_fill_event(link, -ECANCELED);
1730 io_commit_cqring(ctx);
1734 req->flags &= ~REQ_F_LINK_TIMEOUT;
1735 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1738 io_cqring_ev_posted(ctx);
1744 static void io_fail_links(struct io_kiocb *req)
1746 struct io_kiocb *link, *nxt;
1747 struct io_ring_ctx *ctx = req->ctx;
1748 unsigned long flags;
1750 spin_lock_irqsave(&ctx->completion_lock, flags);
1758 trace_io_uring_fail_link(req, link);
1759 io_cqring_fill_event(link, -ECANCELED);
1761 io_put_req_deferred(link, 2);
1764 io_commit_cqring(ctx);
1765 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1767 io_cqring_ev_posted(ctx);
1770 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1772 if (req->flags & REQ_F_LINK_TIMEOUT)
1773 io_kill_linked_timeout(req);
1776 * If LINK is set, we have dependent requests in this chain. If we
1777 * didn't fail this request, queue the first one up, moving any other
1778 * dependencies to the next request. In case of failure, fail the rest
1781 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1782 struct io_kiocb *nxt = req->link;
1791 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1793 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1795 return __io_req_find_next(req);
1798 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1802 if (ctx->submit_state.comp.nr) {
1803 mutex_lock(&ctx->uring_lock);
1804 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1805 mutex_unlock(&ctx->uring_lock);
1807 percpu_ref_put(&ctx->refs);
1810 static bool __tctx_task_work(struct io_uring_task *tctx)
1812 struct io_ring_ctx *ctx = NULL;
1813 struct io_wq_work_list list;
1814 struct io_wq_work_node *node;
1816 if (wq_list_empty(&tctx->task_list))
1819 spin_lock_irq(&tctx->task_lock);
1820 list = tctx->task_list;
1821 INIT_WQ_LIST(&tctx->task_list);
1822 spin_unlock_irq(&tctx->task_lock);
1826 struct io_wq_work_node *next = node->next;
1827 struct io_kiocb *req;
1829 req = container_of(node, struct io_kiocb, io_task_work.node);
1830 if (req->ctx != ctx) {
1831 ctx_flush_and_put(ctx);
1833 percpu_ref_get(&ctx->refs);
1836 req->task_work.func(&req->task_work);
1840 ctx_flush_and_put(ctx);
1841 return list.first != NULL;
1844 static void tctx_task_work(struct callback_head *cb)
1846 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1848 clear_bit(0, &tctx->task_state);
1850 while (__tctx_task_work(tctx))
1854 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1855 enum task_work_notify_mode notify)
1857 struct io_uring_task *tctx = tsk->io_uring;
1858 struct io_wq_work_node *node, *prev;
1859 unsigned long flags;
1862 WARN_ON_ONCE(!tctx);
1864 spin_lock_irqsave(&tctx->task_lock, flags);
1865 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1866 spin_unlock_irqrestore(&tctx->task_lock, flags);
1868 /* task_work already pending, we're done */
1869 if (test_bit(0, &tctx->task_state) ||
1870 test_and_set_bit(0, &tctx->task_state))
1873 if (!task_work_add(tsk, &tctx->task_work, notify))
1877 * Slow path - we failed, find and delete work. if the work is not
1878 * in the list, it got run and we're fine.
1881 spin_lock_irqsave(&tctx->task_lock, flags);
1882 wq_list_for_each(node, prev, &tctx->task_list) {
1883 if (&req->io_task_work.node == node) {
1884 wq_list_del(&tctx->task_list, node, prev);
1889 spin_unlock_irqrestore(&tctx->task_lock, flags);
1890 clear_bit(0, &tctx->task_state);
1894 static int io_req_task_work_add(struct io_kiocb *req)
1896 struct task_struct *tsk = req->task;
1897 struct io_ring_ctx *ctx = req->ctx;
1898 enum task_work_notify_mode notify;
1901 if (tsk->flags & PF_EXITING)
1905 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1906 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1907 * processing task_work. There's no reliable way to tell if TWA_RESUME
1911 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1912 notify = TWA_SIGNAL;
1914 ret = io_task_work_add(tsk, req, notify);
1916 wake_up_process(tsk);
1921 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1922 task_work_func_t cb)
1924 struct io_ring_ctx *ctx = req->ctx;
1925 struct callback_head *head;
1927 init_task_work(&req->task_work, cb);
1929 head = READ_ONCE(ctx->exit_task_work);
1930 req->task_work.next = head;
1931 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1934 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1936 struct io_ring_ctx *ctx = req->ctx;
1938 spin_lock_irq(&ctx->completion_lock);
1939 io_cqring_fill_event(req, error);
1940 io_commit_cqring(ctx);
1941 spin_unlock_irq(&ctx->completion_lock);
1943 io_cqring_ev_posted(ctx);
1944 req_set_fail_links(req);
1945 io_double_put_req(req);
1948 static void io_req_task_cancel(struct callback_head *cb)
1950 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1951 struct io_ring_ctx *ctx = req->ctx;
1953 mutex_lock(&ctx->uring_lock);
1954 __io_req_task_cancel(req, req->result);
1955 mutex_unlock(&ctx->uring_lock);
1956 percpu_ref_put(&ctx->refs);
1959 static void __io_req_task_submit(struct io_kiocb *req)
1961 struct io_ring_ctx *ctx = req->ctx;
1963 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1964 mutex_lock(&ctx->uring_lock);
1965 if (!(current->flags & PF_EXITING) && !current->in_execve)
1966 __io_queue_sqe(req);
1968 __io_req_task_cancel(req, -EFAULT);
1969 mutex_unlock(&ctx->uring_lock);
1972 static void io_req_task_submit(struct callback_head *cb)
1974 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1976 __io_req_task_submit(req);
1979 static void io_req_task_queue(struct io_kiocb *req)
1983 req->task_work.func = io_req_task_submit;
1984 ret = io_req_task_work_add(req);
1985 if (unlikely(ret)) {
1986 req->result = -ECANCELED;
1987 percpu_ref_get(&req->ctx->refs);
1988 io_req_task_work_add_fallback(req, io_req_task_cancel);
1992 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1994 percpu_ref_get(&req->ctx->refs);
1996 req->task_work.func = io_req_task_cancel;
1998 if (unlikely(io_req_task_work_add(req)))
1999 io_req_task_work_add_fallback(req, io_req_task_cancel);
2002 static inline void io_queue_next(struct io_kiocb *req)
2004 struct io_kiocb *nxt = io_req_find_next(req);
2007 io_req_task_queue(nxt);
2010 static void io_free_req(struct io_kiocb *req)
2017 struct task_struct *task;
2022 static inline void io_init_req_batch(struct req_batch *rb)
2029 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2030 struct req_batch *rb)
2033 io_put_task(rb->task, rb->task_refs);
2035 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2038 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2039 struct io_submit_state *state)
2043 if (req->task != rb->task) {
2045 io_put_task(rb->task, rb->task_refs);
2046 rb->task = req->task;
2052 io_dismantle_req(req);
2053 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2054 state->reqs[state->free_reqs++] = req;
2056 list_add(&req->compl.list, &state->comp.free_list);
2059 static void io_submit_flush_completions(struct io_comp_state *cs,
2060 struct io_ring_ctx *ctx)
2063 struct io_kiocb *req;
2064 struct req_batch rb;
2066 io_init_req_batch(&rb);
2067 spin_lock_irq(&ctx->completion_lock);
2068 for (i = 0; i < nr; i++) {
2070 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2072 io_commit_cqring(ctx);
2073 spin_unlock_irq(&ctx->completion_lock);
2075 io_cqring_ev_posted(ctx);
2076 for (i = 0; i < nr; i++) {
2079 /* submission and completion refs */
2080 if (refcount_sub_and_test(2, &req->refs))
2081 io_req_free_batch(&rb, req, &ctx->submit_state);
2084 io_req_free_batch_finish(ctx, &rb);
2089 * Drop reference to request, return next in chain (if there is one) if this
2090 * was the last reference to this request.
2092 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2094 struct io_kiocb *nxt = NULL;
2096 if (refcount_dec_and_test(&req->refs)) {
2097 nxt = io_req_find_next(req);
2103 static void io_put_req(struct io_kiocb *req)
2105 if (refcount_dec_and_test(&req->refs))
2109 static void io_put_req_deferred_cb(struct callback_head *cb)
2111 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2116 static void io_free_req_deferred(struct io_kiocb *req)
2120 req->task_work.func = io_put_req_deferred_cb;
2121 ret = io_req_task_work_add(req);
2123 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2126 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2128 if (refcount_sub_and_test(refs, &req->refs))
2129 io_free_req_deferred(req);
2132 static void io_double_put_req(struct io_kiocb *req)
2134 /* drop both submit and complete references */
2135 if (refcount_sub_and_test(2, &req->refs))
2139 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2141 /* See comment at the top of this file */
2143 return __io_cqring_events(ctx);
2146 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2148 struct io_rings *rings = ctx->rings;
2150 /* make sure SQ entry isn't read before tail */
2151 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2154 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2156 unsigned int cflags;
2158 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2159 cflags |= IORING_CQE_F_BUFFER;
2160 req->flags &= ~REQ_F_BUFFER_SELECTED;
2165 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2167 struct io_buffer *kbuf;
2169 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2170 return io_put_kbuf(req, kbuf);
2173 static inline bool io_run_task_work(void)
2176 * Not safe to run on exiting task, and the task_work handling will
2177 * not add work to such a task.
2179 if (unlikely(current->flags & PF_EXITING))
2181 if (current->task_works) {
2182 __set_current_state(TASK_RUNNING);
2191 * Find and free completed poll iocbs
2193 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2194 struct list_head *done)
2196 struct req_batch rb;
2197 struct io_kiocb *req;
2199 /* order with ->result store in io_complete_rw_iopoll() */
2202 io_init_req_batch(&rb);
2203 while (!list_empty(done)) {
2206 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2207 list_del(&req->inflight_entry);
2209 if (READ_ONCE(req->result) == -EAGAIN) {
2210 req->iopoll_completed = 0;
2211 if (io_rw_reissue(req))
2215 if (req->flags & REQ_F_BUFFER_SELECTED)
2216 cflags = io_put_rw_kbuf(req);
2218 __io_cqring_fill_event(req, req->result, cflags);
2221 if (refcount_dec_and_test(&req->refs))
2222 io_req_free_batch(&rb, req, &ctx->submit_state);
2225 io_commit_cqring(ctx);
2226 io_cqring_ev_posted_iopoll(ctx);
2227 io_req_free_batch_finish(ctx, &rb);
2230 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2233 struct io_kiocb *req, *tmp;
2239 * Only spin for completions if we don't have multiple devices hanging
2240 * off our complete list, and we're under the requested amount.
2242 spin = !ctx->poll_multi_file && *nr_events < min;
2245 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2246 struct kiocb *kiocb = &req->rw.kiocb;
2249 * Move completed and retryable entries to our local lists.
2250 * If we find a request that requires polling, break out
2251 * and complete those lists first, if we have entries there.
2253 if (READ_ONCE(req->iopoll_completed)) {
2254 list_move_tail(&req->inflight_entry, &done);
2257 if (!list_empty(&done))
2260 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2264 /* iopoll may have completed current req */
2265 if (READ_ONCE(req->iopoll_completed))
2266 list_move_tail(&req->inflight_entry, &done);
2273 if (!list_empty(&done))
2274 io_iopoll_complete(ctx, nr_events, &done);
2280 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2281 * non-spinning poll check - we'll still enter the driver poll loop, but only
2282 * as a non-spinning completion check.
2284 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2287 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2290 ret = io_do_iopoll(ctx, nr_events, min);
2293 if (*nr_events >= min)
2301 * We can't just wait for polled events to come to us, we have to actively
2302 * find and complete them.
2304 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2306 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2309 mutex_lock(&ctx->uring_lock);
2310 while (!list_empty(&ctx->iopoll_list)) {
2311 unsigned int nr_events = 0;
2313 io_do_iopoll(ctx, &nr_events, 0);
2315 /* let it sleep and repeat later if can't complete a request */
2319 * Ensure we allow local-to-the-cpu processing to take place,
2320 * in this case we need to ensure that we reap all events.
2321 * Also let task_work, etc. to progress by releasing the mutex
2323 if (need_resched()) {
2324 mutex_unlock(&ctx->uring_lock);
2326 mutex_lock(&ctx->uring_lock);
2329 mutex_unlock(&ctx->uring_lock);
2332 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2334 unsigned int nr_events = 0;
2335 int iters = 0, ret = 0;
2338 * We disallow the app entering submit/complete with polling, but we
2339 * still need to lock the ring to prevent racing with polled issue
2340 * that got punted to a workqueue.
2342 mutex_lock(&ctx->uring_lock);
2345 * Don't enter poll loop if we already have events pending.
2346 * If we do, we can potentially be spinning for commands that
2347 * already triggered a CQE (eg in error).
2349 if (test_bit(0, &ctx->cq_check_overflow))
2350 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2351 if (io_cqring_events(ctx))
2355 * If a submit got punted to a workqueue, we can have the
2356 * application entering polling for a command before it gets
2357 * issued. That app will hold the uring_lock for the duration
2358 * of the poll right here, so we need to take a breather every
2359 * now and then to ensure that the issue has a chance to add
2360 * the poll to the issued list. Otherwise we can spin here
2361 * forever, while the workqueue is stuck trying to acquire the
2364 if (!(++iters & 7)) {
2365 mutex_unlock(&ctx->uring_lock);
2367 mutex_lock(&ctx->uring_lock);
2370 ret = io_iopoll_getevents(ctx, &nr_events, min);
2374 } while (min && !nr_events && !need_resched());
2376 mutex_unlock(&ctx->uring_lock);
2380 static void kiocb_end_write(struct io_kiocb *req)
2383 * Tell lockdep we inherited freeze protection from submission
2386 if (req->flags & REQ_F_ISREG) {
2387 struct inode *inode = file_inode(req->file);
2389 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2391 file_end_write(req->file);
2395 static bool io_resubmit_prep(struct io_kiocb *req)
2397 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2399 struct iov_iter iter;
2401 /* already prepared */
2402 if (req->async_data)
2405 switch (req->opcode) {
2406 case IORING_OP_READV:
2407 case IORING_OP_READ_FIXED:
2408 case IORING_OP_READ:
2411 case IORING_OP_WRITEV:
2412 case IORING_OP_WRITE_FIXED:
2413 case IORING_OP_WRITE:
2417 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2422 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2425 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2428 static bool io_rw_should_reissue(struct io_kiocb *req)
2430 umode_t mode = file_inode(req->file)->i_mode;
2431 struct io_ring_ctx *ctx = req->ctx;
2433 if (!S_ISBLK(mode) && !S_ISREG(mode))
2435 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2436 !(ctx->flags & IORING_SETUP_IOPOLL)))
2439 * If ref is dying, we might be running poll reap from the exit work.
2440 * Don't attempt to reissue from that path, just let it fail with
2443 if (percpu_ref_is_dying(&ctx->refs))
2449 static bool io_rw_reissue(struct io_kiocb *req)
2452 if (!io_rw_should_reissue(req))
2455 lockdep_assert_held(&req->ctx->uring_lock);
2457 if (io_resubmit_prep(req)) {
2458 refcount_inc(&req->refs);
2459 io_queue_async_work(req);
2462 req_set_fail_links(req);
2467 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2468 unsigned int issue_flags)
2472 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2474 if (res != req->result)
2475 req_set_fail_links(req);
2477 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2478 kiocb_end_write(req);
2479 if (req->flags & REQ_F_BUFFER_SELECTED)
2480 cflags = io_put_rw_kbuf(req);
2481 __io_req_complete(req, issue_flags, res, cflags);
2484 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2486 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2488 __io_complete_rw(req, res, res2, 0);
2491 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2493 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2496 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2497 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2498 struct io_async_rw *rw = req->async_data;
2501 iov_iter_revert(&rw->iter,
2502 req->result - iov_iter_count(&rw->iter));
2503 else if (!io_resubmit_prep(req))
2508 if (kiocb->ki_flags & IOCB_WRITE)
2509 kiocb_end_write(req);
2511 if (res != -EAGAIN && res != req->result)
2512 req_set_fail_links(req);
2514 WRITE_ONCE(req->result, res);
2515 /* order with io_poll_complete() checking ->result */
2517 WRITE_ONCE(req->iopoll_completed, 1);
2521 * After the iocb has been issued, it's safe to be found on the poll list.
2522 * Adding the kiocb to the list AFTER submission ensures that we don't
2523 * find it from a io_iopoll_getevents() thread before the issuer is done
2524 * accessing the kiocb cookie.
2526 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2528 struct io_ring_ctx *ctx = req->ctx;
2531 * Track whether we have multiple files in our lists. This will impact
2532 * how we do polling eventually, not spinning if we're on potentially
2533 * different devices.
2535 if (list_empty(&ctx->iopoll_list)) {
2536 ctx->poll_multi_file = false;
2537 } else if (!ctx->poll_multi_file) {
2538 struct io_kiocb *list_req;
2540 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2542 if (list_req->file != req->file)
2543 ctx->poll_multi_file = true;
2547 * For fast devices, IO may have already completed. If it has, add
2548 * it to the front so we find it first.
2550 if (READ_ONCE(req->iopoll_completed))
2551 list_add(&req->inflight_entry, &ctx->iopoll_list);
2553 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2556 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2557 * task context or in io worker task context. If current task context is
2558 * sq thread, we don't need to check whether should wake up sq thread.
2560 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2561 wq_has_sleeper(&ctx->sq_data->wait))
2562 wake_up(&ctx->sq_data->wait);
2565 static inline void io_state_file_put(struct io_submit_state *state)
2567 if (state->file_refs) {
2568 fput_many(state->file, state->file_refs);
2569 state->file_refs = 0;
2574 * Get as many references to a file as we have IOs left in this submission,
2575 * assuming most submissions are for one file, or at least that each file
2576 * has more than one submission.
2578 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2583 if (state->file_refs) {
2584 if (state->fd == fd) {
2588 io_state_file_put(state);
2590 state->file = fget_many(fd, state->ios_left);
2591 if (unlikely(!state->file))
2595 state->file_refs = state->ios_left - 1;
2599 static bool io_bdev_nowait(struct block_device *bdev)
2601 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2605 * If we tracked the file through the SCM inflight mechanism, we could support
2606 * any file. For now, just ensure that anything potentially problematic is done
2609 static bool io_file_supports_async(struct file *file, int rw)
2611 umode_t mode = file_inode(file)->i_mode;
2613 if (S_ISBLK(mode)) {
2614 if (IS_ENABLED(CONFIG_BLOCK) &&
2615 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2619 if (S_ISCHR(mode) || S_ISSOCK(mode))
2621 if (S_ISREG(mode)) {
2622 if (IS_ENABLED(CONFIG_BLOCK) &&
2623 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2624 file->f_op != &io_uring_fops)
2629 /* any ->read/write should understand O_NONBLOCK */
2630 if (file->f_flags & O_NONBLOCK)
2633 if (!(file->f_mode & FMODE_NOWAIT))
2637 return file->f_op->read_iter != NULL;
2639 return file->f_op->write_iter != NULL;
2642 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2644 struct io_ring_ctx *ctx = req->ctx;
2645 struct kiocb *kiocb = &req->rw.kiocb;
2646 struct file *file = req->file;
2650 if (S_ISREG(file_inode(file)->i_mode))
2651 req->flags |= REQ_F_ISREG;
2653 kiocb->ki_pos = READ_ONCE(sqe->off);
2654 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2655 req->flags |= REQ_F_CUR_POS;
2656 kiocb->ki_pos = file->f_pos;
2658 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2659 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2660 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2664 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2665 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2666 req->flags |= REQ_F_NOWAIT;
2668 ioprio = READ_ONCE(sqe->ioprio);
2670 ret = ioprio_check_cap(ioprio);
2674 kiocb->ki_ioprio = ioprio;
2676 kiocb->ki_ioprio = get_current_ioprio();
2678 if (ctx->flags & IORING_SETUP_IOPOLL) {
2679 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2680 !kiocb->ki_filp->f_op->iopoll)
2683 kiocb->ki_flags |= IOCB_HIPRI;
2684 kiocb->ki_complete = io_complete_rw_iopoll;
2685 req->iopoll_completed = 0;
2687 if (kiocb->ki_flags & IOCB_HIPRI)
2689 kiocb->ki_complete = io_complete_rw;
2692 req->rw.addr = READ_ONCE(sqe->addr);
2693 req->rw.len = READ_ONCE(sqe->len);
2694 req->buf_index = READ_ONCE(sqe->buf_index);
2698 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2704 case -ERESTARTNOINTR:
2705 case -ERESTARTNOHAND:
2706 case -ERESTART_RESTARTBLOCK:
2708 * We can't just restart the syscall, since previously
2709 * submitted sqes may already be in progress. Just fail this
2715 kiocb->ki_complete(kiocb, ret, 0);
2719 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2720 unsigned int issue_flags)
2722 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2723 struct io_async_rw *io = req->async_data;
2725 /* add previously done IO, if any */
2726 if (io && io->bytes_done > 0) {
2728 ret = io->bytes_done;
2730 ret += io->bytes_done;
2733 if (req->flags & REQ_F_CUR_POS)
2734 req->file->f_pos = kiocb->ki_pos;
2735 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2736 __io_complete_rw(req, ret, 0, issue_flags);
2738 io_rw_done(kiocb, ret);
2741 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2743 struct io_ring_ctx *ctx = req->ctx;
2744 size_t len = req->rw.len;
2745 struct io_mapped_ubuf *imu;
2746 u16 index, buf_index = req->buf_index;
2750 if (unlikely(buf_index >= ctx->nr_user_bufs))
2752 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2753 imu = &ctx->user_bufs[index];
2754 buf_addr = req->rw.addr;
2757 if (buf_addr + len < buf_addr)
2759 /* not inside the mapped region */
2760 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2764 * May not be a start of buffer, set size appropriately
2765 * and advance us to the beginning.
2767 offset = buf_addr - imu->ubuf;
2768 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2772 * Don't use iov_iter_advance() here, as it's really slow for
2773 * using the latter parts of a big fixed buffer - it iterates
2774 * over each segment manually. We can cheat a bit here, because
2777 * 1) it's a BVEC iter, we set it up
2778 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2779 * first and last bvec
2781 * So just find our index, and adjust the iterator afterwards.
2782 * If the offset is within the first bvec (or the whole first
2783 * bvec, just use iov_iter_advance(). This makes it easier
2784 * since we can just skip the first segment, which may not
2785 * be PAGE_SIZE aligned.
2787 const struct bio_vec *bvec = imu->bvec;
2789 if (offset <= bvec->bv_len) {
2790 iov_iter_advance(iter, offset);
2792 unsigned long seg_skip;
2794 /* skip first vec */
2795 offset -= bvec->bv_len;
2796 seg_skip = 1 + (offset >> PAGE_SHIFT);
2798 iter->bvec = bvec + seg_skip;
2799 iter->nr_segs -= seg_skip;
2800 iter->count -= bvec->bv_len + offset;
2801 iter->iov_offset = offset & ~PAGE_MASK;
2808 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2811 mutex_unlock(&ctx->uring_lock);
2814 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2817 * "Normal" inline submissions always hold the uring_lock, since we
2818 * grab it from the system call. Same is true for the SQPOLL offload.
2819 * The only exception is when we've detached the request and issue it
2820 * from an async worker thread, grab the lock for that case.
2823 mutex_lock(&ctx->uring_lock);
2826 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2827 int bgid, struct io_buffer *kbuf,
2830 struct io_buffer *head;
2832 if (req->flags & REQ_F_BUFFER_SELECTED)
2835 io_ring_submit_lock(req->ctx, needs_lock);
2837 lockdep_assert_held(&req->ctx->uring_lock);
2839 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2841 if (!list_empty(&head->list)) {
2842 kbuf = list_last_entry(&head->list, struct io_buffer,
2844 list_del(&kbuf->list);
2847 idr_remove(&req->ctx->io_buffer_idr, bgid);
2849 if (*len > kbuf->len)
2852 kbuf = ERR_PTR(-ENOBUFS);
2855 io_ring_submit_unlock(req->ctx, needs_lock);
2860 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2863 struct io_buffer *kbuf;
2866 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2867 bgid = req->buf_index;
2868 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2871 req->rw.addr = (u64) (unsigned long) kbuf;
2872 req->flags |= REQ_F_BUFFER_SELECTED;
2873 return u64_to_user_ptr(kbuf->addr);
2876 #ifdef CONFIG_COMPAT
2877 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2880 struct compat_iovec __user *uiov;
2881 compat_ssize_t clen;
2885 uiov = u64_to_user_ptr(req->rw.addr);
2886 if (!access_ok(uiov, sizeof(*uiov)))
2888 if (__get_user(clen, &uiov->iov_len))
2894 buf = io_rw_buffer_select(req, &len, needs_lock);
2896 return PTR_ERR(buf);
2897 iov[0].iov_base = buf;
2898 iov[0].iov_len = (compat_size_t) len;
2903 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2906 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2910 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2913 len = iov[0].iov_len;
2916 buf = io_rw_buffer_select(req, &len, needs_lock);
2918 return PTR_ERR(buf);
2919 iov[0].iov_base = buf;
2920 iov[0].iov_len = len;
2924 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2927 if (req->flags & REQ_F_BUFFER_SELECTED) {
2928 struct io_buffer *kbuf;
2930 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2931 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2932 iov[0].iov_len = kbuf->len;
2935 if (req->rw.len != 1)
2938 #ifdef CONFIG_COMPAT
2939 if (req->ctx->compat)
2940 return io_compat_import(req, iov, needs_lock);
2943 return __io_iov_buffer_select(req, iov, needs_lock);
2946 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2947 struct iov_iter *iter, bool needs_lock)
2949 void __user *buf = u64_to_user_ptr(req->rw.addr);
2950 size_t sqe_len = req->rw.len;
2951 u8 opcode = req->opcode;
2954 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2956 return io_import_fixed(req, rw, iter);
2959 /* buffer index only valid with fixed read/write, or buffer select */
2960 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2963 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2964 if (req->flags & REQ_F_BUFFER_SELECT) {
2965 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2967 return PTR_ERR(buf);
2968 req->rw.len = sqe_len;
2971 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2976 if (req->flags & REQ_F_BUFFER_SELECT) {
2977 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2979 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2984 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2988 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2990 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2994 * For files that don't have ->read_iter() and ->write_iter(), handle them
2995 * by looping over ->read() or ->write() manually.
2997 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2999 struct kiocb *kiocb = &req->rw.kiocb;
3000 struct file *file = req->file;
3004 * Don't support polled IO through this interface, and we can't
3005 * support non-blocking either. For the latter, this just causes
3006 * the kiocb to be handled from an async context.
3008 if (kiocb->ki_flags & IOCB_HIPRI)
3010 if (kiocb->ki_flags & IOCB_NOWAIT)
3013 while (iov_iter_count(iter)) {
3017 if (!iov_iter_is_bvec(iter)) {
3018 iovec = iov_iter_iovec(iter);
3020 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3021 iovec.iov_len = req->rw.len;
3025 nr = file->f_op->read(file, iovec.iov_base,
3026 iovec.iov_len, io_kiocb_ppos(kiocb));
3028 nr = file->f_op->write(file, iovec.iov_base,
3029 iovec.iov_len, io_kiocb_ppos(kiocb));
3038 if (nr != iovec.iov_len)
3042 iov_iter_advance(iter, nr);
3048 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3049 const struct iovec *fast_iov, struct iov_iter *iter)
3051 struct io_async_rw *rw = req->async_data;
3053 memcpy(&rw->iter, iter, sizeof(*iter));
3054 rw->free_iovec = iovec;
3056 /* can only be fixed buffers, no need to do anything */
3057 if (iov_iter_is_bvec(iter))
3060 unsigned iov_off = 0;
3062 rw->iter.iov = rw->fast_iov;
3063 if (iter->iov != fast_iov) {
3064 iov_off = iter->iov - fast_iov;
3065 rw->iter.iov += iov_off;
3067 if (rw->fast_iov != fast_iov)
3068 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3069 sizeof(struct iovec) * iter->nr_segs);
3071 req->flags |= REQ_F_NEED_CLEANUP;
3075 static inline int __io_alloc_async_data(struct io_kiocb *req)
3077 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3078 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3079 return req->async_data == NULL;
3082 static int io_alloc_async_data(struct io_kiocb *req)
3084 if (!io_op_defs[req->opcode].needs_async_data)
3087 return __io_alloc_async_data(req);
3090 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3091 const struct iovec *fast_iov,
3092 struct iov_iter *iter, bool force)
3094 if (!force && !io_op_defs[req->opcode].needs_async_data)
3096 if (!req->async_data) {
3097 if (__io_alloc_async_data(req)) {
3102 io_req_map_rw(req, iovec, fast_iov, iter);
3107 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3109 struct io_async_rw *iorw = req->async_data;
3110 struct iovec *iov = iorw->fast_iov;
3113 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3114 if (unlikely(ret < 0))
3117 iorw->bytes_done = 0;
3118 iorw->free_iovec = iov;
3120 req->flags |= REQ_F_NEED_CLEANUP;
3124 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3126 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3128 return io_prep_rw(req, sqe);
3132 * This is our waitqueue callback handler, registered through lock_page_async()
3133 * when we initially tried to do the IO with the iocb armed our waitqueue.
3134 * This gets called when the page is unlocked, and we generally expect that to
3135 * happen when the page IO is completed and the page is now uptodate. This will
3136 * queue a task_work based retry of the operation, attempting to copy the data
3137 * again. If the latter fails because the page was NOT uptodate, then we will
3138 * do a thread based blocking retry of the operation. That's the unexpected
3141 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3142 int sync, void *arg)
3144 struct wait_page_queue *wpq;
3145 struct io_kiocb *req = wait->private;
3146 struct wait_page_key *key = arg;
3148 wpq = container_of(wait, struct wait_page_queue, wait);
3150 if (!wake_page_match(wpq, key))
3153 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3154 list_del_init(&wait->entry);
3156 /* submit ref gets dropped, acquire a new one */
3157 refcount_inc(&req->refs);
3158 io_req_task_queue(req);
3163 * This controls whether a given IO request should be armed for async page
3164 * based retry. If we return false here, the request is handed to the async
3165 * worker threads for retry. If we're doing buffered reads on a regular file,
3166 * we prepare a private wait_page_queue entry and retry the operation. This
3167 * will either succeed because the page is now uptodate and unlocked, or it
3168 * will register a callback when the page is unlocked at IO completion. Through
3169 * that callback, io_uring uses task_work to setup a retry of the operation.
3170 * That retry will attempt the buffered read again. The retry will generally
3171 * succeed, or in rare cases where it fails, we then fall back to using the
3172 * async worker threads for a blocking retry.
3174 static bool io_rw_should_retry(struct io_kiocb *req)
3176 struct io_async_rw *rw = req->async_data;
3177 struct wait_page_queue *wait = &rw->wpq;
3178 struct kiocb *kiocb = &req->rw.kiocb;
3180 /* never retry for NOWAIT, we just complete with -EAGAIN */
3181 if (req->flags & REQ_F_NOWAIT)
3184 /* Only for buffered IO */
3185 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3189 * just use poll if we can, and don't attempt if the fs doesn't
3190 * support callback based unlocks
3192 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3195 wait->wait.func = io_async_buf_func;
3196 wait->wait.private = req;
3197 wait->wait.flags = 0;
3198 INIT_LIST_HEAD(&wait->wait.entry);
3199 kiocb->ki_flags |= IOCB_WAITQ;
3200 kiocb->ki_flags &= ~IOCB_NOWAIT;
3201 kiocb->ki_waitq = wait;
3205 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3207 if (req->file->f_op->read_iter)
3208 return call_read_iter(req->file, &req->rw.kiocb, iter);
3209 else if (req->file->f_op->read)
3210 return loop_rw_iter(READ, req, iter);
3215 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3217 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3218 struct kiocb *kiocb = &req->rw.kiocb;
3219 struct iov_iter __iter, *iter = &__iter;
3220 struct io_async_rw *rw = req->async_data;
3221 ssize_t io_size, ret, ret2;
3222 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3228 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3232 io_size = iov_iter_count(iter);
3233 req->result = io_size;
3235 /* Ensure we clear previously set non-block flag */
3236 if (!force_nonblock)
3237 kiocb->ki_flags &= ~IOCB_NOWAIT;
3239 kiocb->ki_flags |= IOCB_NOWAIT;
3241 /* If the file doesn't support async, just async punt */
3242 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3243 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3244 return ret ?: -EAGAIN;
3247 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3248 if (unlikely(ret)) {
3253 ret = io_iter_do_read(req, iter);
3255 if (ret == -EIOCBQUEUED) {
3256 if (req->async_data)
3257 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3259 } else if (ret == -EAGAIN) {
3260 /* IOPOLL retry should happen for io-wq threads */
3261 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3263 /* no retry on NONBLOCK nor RWF_NOWAIT */
3264 if (req->flags & REQ_F_NOWAIT)
3266 /* some cases will consume bytes even on error returns */
3267 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3269 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3270 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3271 /* read all, failed, already did sync or don't want to retry */
3275 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3280 rw = req->async_data;
3281 /* now use our persistent iterator, if we aren't already */
3286 rw->bytes_done += ret;
3287 /* if we can retry, do so with the callbacks armed */
3288 if (!io_rw_should_retry(req)) {
3289 kiocb->ki_flags &= ~IOCB_WAITQ;
3294 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3295 * we get -EIOCBQUEUED, then we'll get a notification when the
3296 * desired page gets unlocked. We can also get a partial read
3297 * here, and if we do, then just retry at the new offset.
3299 ret = io_iter_do_read(req, iter);
3300 if (ret == -EIOCBQUEUED)
3302 /* we got some bytes, but not all. retry. */
3303 kiocb->ki_flags &= ~IOCB_WAITQ;
3304 } while (ret > 0 && ret < io_size);
3306 kiocb_done(kiocb, ret, issue_flags);
3308 /* it's faster to check here then delegate to kfree */
3314 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3316 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3318 return io_prep_rw(req, sqe);
3321 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3323 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3324 struct kiocb *kiocb = &req->rw.kiocb;
3325 struct iov_iter __iter, *iter = &__iter;
3326 struct io_async_rw *rw = req->async_data;
3327 ssize_t ret, ret2, io_size;
3328 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3334 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3338 io_size = iov_iter_count(iter);
3339 req->result = io_size;
3341 /* Ensure we clear previously set non-block flag */
3342 if (!force_nonblock)
3343 kiocb->ki_flags &= ~IOCB_NOWAIT;
3345 kiocb->ki_flags |= IOCB_NOWAIT;
3347 /* If the file doesn't support async, just async punt */
3348 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3351 /* file path doesn't support NOWAIT for non-direct_IO */
3352 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3353 (req->flags & REQ_F_ISREG))
3356 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3361 * Open-code file_start_write here to grab freeze protection,
3362 * which will be released by another thread in
3363 * io_complete_rw(). Fool lockdep by telling it the lock got
3364 * released so that it doesn't complain about the held lock when
3365 * we return to userspace.
3367 if (req->flags & REQ_F_ISREG) {
3368 sb_start_write(file_inode(req->file)->i_sb);
3369 __sb_writers_release(file_inode(req->file)->i_sb,
3372 kiocb->ki_flags |= IOCB_WRITE;
3374 if (req->file->f_op->write_iter)
3375 ret2 = call_write_iter(req->file, kiocb, iter);
3376 else if (req->file->f_op->write)
3377 ret2 = loop_rw_iter(WRITE, req, iter);
3382 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3383 * retry them without IOCB_NOWAIT.
3385 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3387 /* no retry on NONBLOCK nor RWF_NOWAIT */
3388 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3390 if (ret2 == -EIOCBQUEUED && req->async_data)
3391 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3392 if (!force_nonblock || ret2 != -EAGAIN) {
3393 /* IOPOLL retry should happen for io-wq threads */
3394 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3397 kiocb_done(kiocb, ret2, issue_flags);
3400 /* some cases will consume bytes even on error returns */
3401 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3402 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3403 return ret ?: -EAGAIN;
3406 /* it's reportedly faster than delegating the null check to kfree() */
3412 static int io_renameat_prep(struct io_kiocb *req,
3413 const struct io_uring_sqe *sqe)
3415 struct io_rename *ren = &req->rename;
3416 const char __user *oldf, *newf;
3418 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3421 ren->old_dfd = READ_ONCE(sqe->fd);
3422 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3423 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3424 ren->new_dfd = READ_ONCE(sqe->len);
3425 ren->flags = READ_ONCE(sqe->rename_flags);
3427 ren->oldpath = getname(oldf);
3428 if (IS_ERR(ren->oldpath))
3429 return PTR_ERR(ren->oldpath);
3431 ren->newpath = getname(newf);
3432 if (IS_ERR(ren->newpath)) {
3433 putname(ren->oldpath);
3434 return PTR_ERR(ren->newpath);
3437 req->flags |= REQ_F_NEED_CLEANUP;
3441 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3443 struct io_rename *ren = &req->rename;
3446 if (issue_flags & IO_URING_F_NONBLOCK)
3449 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3450 ren->newpath, ren->flags);
3452 req->flags &= ~REQ_F_NEED_CLEANUP;
3454 req_set_fail_links(req);
3455 io_req_complete(req, ret);
3459 static int io_unlinkat_prep(struct io_kiocb *req,
3460 const struct io_uring_sqe *sqe)
3462 struct io_unlink *un = &req->unlink;
3463 const char __user *fname;
3465 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3468 un->dfd = READ_ONCE(sqe->fd);
3470 un->flags = READ_ONCE(sqe->unlink_flags);
3471 if (un->flags & ~AT_REMOVEDIR)
3474 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3475 un->filename = getname(fname);
3476 if (IS_ERR(un->filename))
3477 return PTR_ERR(un->filename);
3479 req->flags |= REQ_F_NEED_CLEANUP;
3483 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3485 struct io_unlink *un = &req->unlink;
3488 if (issue_flags & IO_URING_F_NONBLOCK)
3491 if (un->flags & AT_REMOVEDIR)
3492 ret = do_rmdir(un->dfd, un->filename);
3494 ret = do_unlinkat(un->dfd, un->filename);
3496 req->flags &= ~REQ_F_NEED_CLEANUP;
3498 req_set_fail_links(req);
3499 io_req_complete(req, ret);
3503 static int io_shutdown_prep(struct io_kiocb *req,
3504 const struct io_uring_sqe *sqe)
3506 #if defined(CONFIG_NET)
3507 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3509 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3513 req->shutdown.how = READ_ONCE(sqe->len);
3520 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3522 #if defined(CONFIG_NET)
3523 struct socket *sock;
3526 if (issue_flags & IO_URING_F_NONBLOCK)
3529 sock = sock_from_file(req->file);
3530 if (unlikely(!sock))
3533 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3535 req_set_fail_links(req);
3536 io_req_complete(req, ret);
3543 static int __io_splice_prep(struct io_kiocb *req,
3544 const struct io_uring_sqe *sqe)
3546 struct io_splice* sp = &req->splice;
3547 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3549 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3553 sp->len = READ_ONCE(sqe->len);
3554 sp->flags = READ_ONCE(sqe->splice_flags);
3556 if (unlikely(sp->flags & ~valid_flags))
3559 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3560 (sp->flags & SPLICE_F_FD_IN_FIXED));
3563 req->flags |= REQ_F_NEED_CLEANUP;
3565 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3567 * Splice operation will be punted aync, and here need to
3568 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3570 req->work.flags |= IO_WQ_WORK_UNBOUND;
3576 static int io_tee_prep(struct io_kiocb *req,
3577 const struct io_uring_sqe *sqe)
3579 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3581 return __io_splice_prep(req, sqe);
3584 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3586 struct io_splice *sp = &req->splice;
3587 struct file *in = sp->file_in;
3588 struct file *out = sp->file_out;
3589 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3592 if (issue_flags & IO_URING_F_NONBLOCK)
3595 ret = do_tee(in, out, sp->len, flags);
3597 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3598 req->flags &= ~REQ_F_NEED_CLEANUP;
3601 req_set_fail_links(req);
3602 io_req_complete(req, ret);
3606 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3608 struct io_splice* sp = &req->splice;
3610 sp->off_in = READ_ONCE(sqe->splice_off_in);
3611 sp->off_out = READ_ONCE(sqe->off);
3612 return __io_splice_prep(req, sqe);
3615 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3617 struct io_splice *sp = &req->splice;
3618 struct file *in = sp->file_in;
3619 struct file *out = sp->file_out;
3620 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3621 loff_t *poff_in, *poff_out;
3624 if (issue_flags & IO_URING_F_NONBLOCK)
3627 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3628 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3631 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3633 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3634 req->flags &= ~REQ_F_NEED_CLEANUP;
3637 req_set_fail_links(req);
3638 io_req_complete(req, ret);
3643 * IORING_OP_NOP just posts a completion event, nothing else.
3645 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3647 struct io_ring_ctx *ctx = req->ctx;
3649 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3652 __io_req_complete(req, issue_flags, 0, 0);
3656 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3658 struct io_ring_ctx *ctx = req->ctx;
3663 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3665 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3668 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3669 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3672 req->sync.off = READ_ONCE(sqe->off);
3673 req->sync.len = READ_ONCE(sqe->len);
3677 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3679 loff_t end = req->sync.off + req->sync.len;
3682 /* fsync always requires a blocking context */
3683 if (issue_flags & IO_URING_F_NONBLOCK)
3686 ret = vfs_fsync_range(req->file, req->sync.off,
3687 end > 0 ? end : LLONG_MAX,
3688 req->sync.flags & IORING_FSYNC_DATASYNC);
3690 req_set_fail_links(req);
3691 io_req_complete(req, ret);
3695 static int io_fallocate_prep(struct io_kiocb *req,
3696 const struct io_uring_sqe *sqe)
3698 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3703 req->sync.off = READ_ONCE(sqe->off);
3704 req->sync.len = READ_ONCE(sqe->addr);
3705 req->sync.mode = READ_ONCE(sqe->len);
3709 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3713 /* fallocate always requiring blocking context */
3714 if (issue_flags & IO_URING_F_NONBLOCK)
3716 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3719 req_set_fail_links(req);
3720 io_req_complete(req, ret);
3724 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3726 const char __user *fname;
3729 if (unlikely(sqe->ioprio || sqe->buf_index))
3731 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3734 /* open.how should be already initialised */
3735 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3736 req->open.how.flags |= O_LARGEFILE;
3738 req->open.dfd = READ_ONCE(sqe->fd);
3739 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3740 req->open.filename = getname(fname);
3741 if (IS_ERR(req->open.filename)) {
3742 ret = PTR_ERR(req->open.filename);
3743 req->open.filename = NULL;
3746 req->open.nofile = rlimit(RLIMIT_NOFILE);
3747 req->flags |= REQ_F_NEED_CLEANUP;
3751 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3757 mode = READ_ONCE(sqe->len);
3758 flags = READ_ONCE(sqe->open_flags);
3759 req->open.how = build_open_how(flags, mode);
3760 return __io_openat_prep(req, sqe);
3763 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3765 struct open_how __user *how;
3769 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3771 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3772 len = READ_ONCE(sqe->len);
3773 if (len < OPEN_HOW_SIZE_VER0)
3776 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3781 return __io_openat_prep(req, sqe);
3784 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3786 struct open_flags op;
3789 bool resolve_nonblock;
3792 ret = build_open_flags(&req->open.how, &op);
3795 nonblock_set = op.open_flag & O_NONBLOCK;
3796 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3797 if (issue_flags & IO_URING_F_NONBLOCK) {
3799 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3800 * it'll always -EAGAIN
3802 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3804 op.lookup_flags |= LOOKUP_CACHED;
3805 op.open_flag |= O_NONBLOCK;
3808 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3812 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3813 /* only retry if RESOLVE_CACHED wasn't already set by application */
3814 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3815 file == ERR_PTR(-EAGAIN)) {
3817 * We could hang on to this 'fd', but seems like marginal
3818 * gain for something that is now known to be a slower path.
3819 * So just put it, and we'll get a new one when we retry.
3827 ret = PTR_ERR(file);
3829 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3830 file->f_flags &= ~O_NONBLOCK;
3831 fsnotify_open(file);
3832 fd_install(ret, file);
3835 putname(req->open.filename);
3836 req->flags &= ~REQ_F_NEED_CLEANUP;
3838 req_set_fail_links(req);
3839 io_req_complete(req, ret);
3843 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3845 return io_openat2(req, issue_flags);
3848 static int io_remove_buffers_prep(struct io_kiocb *req,
3849 const struct io_uring_sqe *sqe)
3851 struct io_provide_buf *p = &req->pbuf;
3854 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3857 tmp = READ_ONCE(sqe->fd);
3858 if (!tmp || tmp > USHRT_MAX)
3861 memset(p, 0, sizeof(*p));
3863 p->bgid = READ_ONCE(sqe->buf_group);
3867 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3868 int bgid, unsigned nbufs)
3872 /* shouldn't happen */
3876 /* the head kbuf is the list itself */
3877 while (!list_empty(&buf->list)) {
3878 struct io_buffer *nxt;
3880 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3881 list_del(&nxt->list);
3888 idr_remove(&ctx->io_buffer_idr, bgid);
3893 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3895 struct io_provide_buf *p = &req->pbuf;
3896 struct io_ring_ctx *ctx = req->ctx;
3897 struct io_buffer *head;
3899 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3901 io_ring_submit_lock(ctx, !force_nonblock);
3903 lockdep_assert_held(&ctx->uring_lock);
3906 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3908 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3910 req_set_fail_links(req);
3912 /* need to hold the lock to complete IOPOLL requests */
3913 if (ctx->flags & IORING_SETUP_IOPOLL) {
3914 __io_req_complete(req, issue_flags, ret, 0);
3915 io_ring_submit_unlock(ctx, !force_nonblock);
3917 io_ring_submit_unlock(ctx, !force_nonblock);
3918 __io_req_complete(req, issue_flags, ret, 0);
3923 static int io_provide_buffers_prep(struct io_kiocb *req,
3924 const struct io_uring_sqe *sqe)
3926 struct io_provide_buf *p = &req->pbuf;
3929 if (sqe->ioprio || sqe->rw_flags)
3932 tmp = READ_ONCE(sqe->fd);
3933 if (!tmp || tmp > USHRT_MAX)
3936 p->addr = READ_ONCE(sqe->addr);
3937 p->len = READ_ONCE(sqe->len);
3939 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3942 p->bgid = READ_ONCE(sqe->buf_group);
3943 tmp = READ_ONCE(sqe->off);
3944 if (tmp > USHRT_MAX)
3950 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3952 struct io_buffer *buf;
3953 u64 addr = pbuf->addr;
3954 int i, bid = pbuf->bid;
3956 for (i = 0; i < pbuf->nbufs; i++) {
3957 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3962 buf->len = pbuf->len;
3967 INIT_LIST_HEAD(&buf->list);
3970 list_add_tail(&buf->list, &(*head)->list);
3974 return i ? i : -ENOMEM;
3977 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3979 struct io_provide_buf *p = &req->pbuf;
3980 struct io_ring_ctx *ctx = req->ctx;
3981 struct io_buffer *head, *list;
3983 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3985 io_ring_submit_lock(ctx, !force_nonblock);
3987 lockdep_assert_held(&ctx->uring_lock);
3989 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3991 ret = io_add_buffers(p, &head);
3996 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3999 __io_remove_buffers(ctx, head, p->bgid, -1U);
4005 req_set_fail_links(req);
4007 /* need to hold the lock to complete IOPOLL requests */
4008 if (ctx->flags & IORING_SETUP_IOPOLL) {
4009 __io_req_complete(req, issue_flags, ret, 0);
4010 io_ring_submit_unlock(ctx, !force_nonblock);
4012 io_ring_submit_unlock(ctx, !force_nonblock);
4013 __io_req_complete(req, issue_flags, ret, 0);
4018 static int io_epoll_ctl_prep(struct io_kiocb *req,
4019 const struct io_uring_sqe *sqe)
4021 #if defined(CONFIG_EPOLL)
4022 if (sqe->ioprio || sqe->buf_index)
4024 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4027 req->epoll.epfd = READ_ONCE(sqe->fd);
4028 req->epoll.op = READ_ONCE(sqe->len);
4029 req->epoll.fd = READ_ONCE(sqe->off);
4031 if (ep_op_has_event(req->epoll.op)) {
4032 struct epoll_event __user *ev;
4034 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4035 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4045 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4047 #if defined(CONFIG_EPOLL)
4048 struct io_epoll *ie = &req->epoll;
4050 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4052 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4053 if (force_nonblock && ret == -EAGAIN)
4057 req_set_fail_links(req);
4058 __io_req_complete(req, issue_flags, ret, 0);
4065 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4067 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4068 if (sqe->ioprio || sqe->buf_index || sqe->off)
4070 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073 req->madvise.addr = READ_ONCE(sqe->addr);
4074 req->madvise.len = READ_ONCE(sqe->len);
4075 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4082 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4084 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4085 struct io_madvise *ma = &req->madvise;
4088 if (issue_flags & IO_URING_F_NONBLOCK)
4091 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4093 req_set_fail_links(req);
4094 io_req_complete(req, ret);
4101 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4103 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4105 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4108 req->fadvise.offset = READ_ONCE(sqe->off);
4109 req->fadvise.len = READ_ONCE(sqe->len);
4110 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4114 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4116 struct io_fadvise *fa = &req->fadvise;
4119 if (issue_flags & IO_URING_F_NONBLOCK) {
4120 switch (fa->advice) {
4121 case POSIX_FADV_NORMAL:
4122 case POSIX_FADV_RANDOM:
4123 case POSIX_FADV_SEQUENTIAL:
4130 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4132 req_set_fail_links(req);
4133 io_req_complete(req, ret);
4137 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4141 if (sqe->ioprio || sqe->buf_index)
4143 if (req->flags & REQ_F_FIXED_FILE)
4146 req->statx.dfd = READ_ONCE(sqe->fd);
4147 req->statx.mask = READ_ONCE(sqe->len);
4148 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4149 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4150 req->statx.flags = READ_ONCE(sqe->statx_flags);
4155 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4157 struct io_statx *ctx = &req->statx;
4160 if (issue_flags & IO_URING_F_NONBLOCK) {
4161 /* only need file table for an actual valid fd */
4162 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4163 req->flags |= REQ_F_NO_FILE_TABLE;
4167 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4171 req_set_fail_links(req);
4172 io_req_complete(req, ret);
4176 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4178 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4180 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4181 sqe->rw_flags || sqe->buf_index)
4183 if (req->flags & REQ_F_FIXED_FILE)
4186 req->close.fd = READ_ONCE(sqe->fd);
4190 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4192 struct files_struct *files = current->files;
4193 struct io_close *close = &req->close;
4194 struct fdtable *fdt;
4200 spin_lock(&files->file_lock);
4201 fdt = files_fdtable(files);
4202 if (close->fd >= fdt->max_fds) {
4203 spin_unlock(&files->file_lock);
4206 file = fdt->fd[close->fd];
4208 spin_unlock(&files->file_lock);
4212 if (file->f_op == &io_uring_fops) {
4213 spin_unlock(&files->file_lock);
4218 /* if the file has a flush method, be safe and punt to async */
4219 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4220 spin_unlock(&files->file_lock);
4224 ret = __close_fd_get_file(close->fd, &file);
4225 spin_unlock(&files->file_lock);
4232 /* No ->flush() or already async, safely close from here */
4233 ret = filp_close(file, current->files);
4236 req_set_fail_links(req);
4239 __io_req_complete(req, issue_flags, ret, 0);
4243 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4245 struct io_ring_ctx *ctx = req->ctx;
4247 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4249 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4252 req->sync.off = READ_ONCE(sqe->off);
4253 req->sync.len = READ_ONCE(sqe->len);
4254 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4258 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4262 /* sync_file_range always requires a blocking context */
4263 if (issue_flags & IO_URING_F_NONBLOCK)
4266 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4269 req_set_fail_links(req);
4270 io_req_complete(req, ret);
4274 #if defined(CONFIG_NET)
4275 static int io_setup_async_msg(struct io_kiocb *req,
4276 struct io_async_msghdr *kmsg)
4278 struct io_async_msghdr *async_msg = req->async_data;
4282 if (io_alloc_async_data(req)) {
4283 kfree(kmsg->free_iov);
4286 async_msg = req->async_data;
4287 req->flags |= REQ_F_NEED_CLEANUP;
4288 memcpy(async_msg, kmsg, sizeof(*kmsg));
4289 async_msg->msg.msg_name = &async_msg->addr;
4290 /* if were using fast_iov, set it to the new one */
4291 if (!async_msg->free_iov)
4292 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4297 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4298 struct io_async_msghdr *iomsg)
4300 iomsg->msg.msg_name = &iomsg->addr;
4301 iomsg->free_iov = iomsg->fast_iov;
4302 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4303 req->sr_msg.msg_flags, &iomsg->free_iov);
4306 static int io_sendmsg_prep_async(struct io_kiocb *req)
4310 if (!io_op_defs[req->opcode].needs_async_data)
4312 ret = io_sendmsg_copy_hdr(req, req->async_data);
4314 req->flags |= REQ_F_NEED_CLEANUP;
4318 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4320 struct io_sr_msg *sr = &req->sr_msg;
4322 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4325 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4326 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4327 sr->len = READ_ONCE(sqe->len);
4329 #ifdef CONFIG_COMPAT
4330 if (req->ctx->compat)
4331 sr->msg_flags |= MSG_CMSG_COMPAT;
4336 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4338 struct io_async_msghdr iomsg, *kmsg;
4339 struct socket *sock;
4343 sock = sock_from_file(req->file);
4344 if (unlikely(!sock))
4347 kmsg = req->async_data;
4349 ret = io_sendmsg_copy_hdr(req, &iomsg);
4355 flags = req->sr_msg.msg_flags;
4356 if (flags & MSG_DONTWAIT)
4357 req->flags |= REQ_F_NOWAIT;
4358 else if (issue_flags & IO_URING_F_NONBLOCK)
4359 flags |= MSG_DONTWAIT;
4361 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4362 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4363 return io_setup_async_msg(req, kmsg);
4364 if (ret == -ERESTARTSYS)
4367 /* fast path, check for non-NULL to avoid function call */
4369 kfree(kmsg->free_iov);
4370 req->flags &= ~REQ_F_NEED_CLEANUP;
4372 req_set_fail_links(req);
4373 __io_req_complete(req, issue_flags, ret, 0);
4377 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4379 struct io_sr_msg *sr = &req->sr_msg;
4382 struct socket *sock;
4386 sock = sock_from_file(req->file);
4387 if (unlikely(!sock))
4390 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4394 msg.msg_name = NULL;
4395 msg.msg_control = NULL;
4396 msg.msg_controllen = 0;
4397 msg.msg_namelen = 0;
4399 flags = req->sr_msg.msg_flags;
4400 if (flags & MSG_DONTWAIT)
4401 req->flags |= REQ_F_NOWAIT;
4402 else if (issue_flags & IO_URING_F_NONBLOCK)
4403 flags |= MSG_DONTWAIT;
4405 msg.msg_flags = flags;
4406 ret = sock_sendmsg(sock, &msg);
4407 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4409 if (ret == -ERESTARTSYS)
4413 req_set_fail_links(req);
4414 __io_req_complete(req, issue_flags, ret, 0);
4418 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4419 struct io_async_msghdr *iomsg)
4421 struct io_sr_msg *sr = &req->sr_msg;
4422 struct iovec __user *uiov;
4426 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4427 &iomsg->uaddr, &uiov, &iov_len);
4431 if (req->flags & REQ_F_BUFFER_SELECT) {
4434 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4436 sr->len = iomsg->fast_iov[0].iov_len;
4437 iomsg->free_iov = NULL;
4439 iomsg->free_iov = iomsg->fast_iov;
4440 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4441 &iomsg->free_iov, &iomsg->msg.msg_iter,
4450 #ifdef CONFIG_COMPAT
4451 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4452 struct io_async_msghdr *iomsg)
4454 struct compat_msghdr __user *msg_compat;
4455 struct io_sr_msg *sr = &req->sr_msg;
4456 struct compat_iovec __user *uiov;
4461 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4462 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4467 uiov = compat_ptr(ptr);
4468 if (req->flags & REQ_F_BUFFER_SELECT) {
4469 compat_ssize_t clen;
4473 if (!access_ok(uiov, sizeof(*uiov)))
4475 if (__get_user(clen, &uiov->iov_len))
4480 iomsg->free_iov = NULL;
4482 iomsg->free_iov = iomsg->fast_iov;
4483 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4484 UIO_FASTIOV, &iomsg->free_iov,
4485 &iomsg->msg.msg_iter, true);
4494 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4495 struct io_async_msghdr *iomsg)
4497 iomsg->msg.msg_name = &iomsg->addr;
4499 #ifdef CONFIG_COMPAT
4500 if (req->ctx->compat)
4501 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4504 return __io_recvmsg_copy_hdr(req, iomsg);
4507 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4510 struct io_sr_msg *sr = &req->sr_msg;
4511 struct io_buffer *kbuf;
4513 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4518 req->flags |= REQ_F_BUFFER_SELECTED;
4522 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4524 return io_put_kbuf(req, req->sr_msg.kbuf);
4527 static int io_recvmsg_prep_async(struct io_kiocb *req)
4531 if (!io_op_defs[req->opcode].needs_async_data)
4533 ret = io_recvmsg_copy_hdr(req, req->async_data);
4535 req->flags |= REQ_F_NEED_CLEANUP;
4539 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4541 struct io_sr_msg *sr = &req->sr_msg;
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4546 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4547 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4548 sr->len = READ_ONCE(sqe->len);
4549 sr->bgid = READ_ONCE(sqe->buf_group);
4551 #ifdef CONFIG_COMPAT
4552 if (req->ctx->compat)
4553 sr->msg_flags |= MSG_CMSG_COMPAT;
4558 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4560 struct io_async_msghdr iomsg, *kmsg;
4561 struct socket *sock;
4562 struct io_buffer *kbuf;
4564 int ret, cflags = 0;
4565 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4567 sock = sock_from_file(req->file);
4568 if (unlikely(!sock))
4571 kmsg = req->async_data;
4573 ret = io_recvmsg_copy_hdr(req, &iomsg);
4579 if (req->flags & REQ_F_BUFFER_SELECT) {
4580 kbuf = io_recv_buffer_select(req, !force_nonblock);
4582 return PTR_ERR(kbuf);
4583 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4584 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4585 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4586 1, req->sr_msg.len);
4589 flags = req->sr_msg.msg_flags;
4590 if (flags & MSG_DONTWAIT)
4591 req->flags |= REQ_F_NOWAIT;
4592 else if (force_nonblock)
4593 flags |= MSG_DONTWAIT;
4595 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4596 kmsg->uaddr, flags);
4597 if (force_nonblock && ret == -EAGAIN)
4598 return io_setup_async_msg(req, kmsg);
4599 if (ret == -ERESTARTSYS)
4602 if (req->flags & REQ_F_BUFFER_SELECTED)
4603 cflags = io_put_recv_kbuf(req);
4604 /* fast path, check for non-NULL to avoid function call */
4606 kfree(kmsg->free_iov);
4607 req->flags &= ~REQ_F_NEED_CLEANUP;
4609 req_set_fail_links(req);
4610 __io_req_complete(req, issue_flags, ret, cflags);
4614 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4616 struct io_buffer *kbuf;
4617 struct io_sr_msg *sr = &req->sr_msg;
4619 void __user *buf = sr->buf;
4620 struct socket *sock;
4623 int ret, cflags = 0;
4624 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4626 sock = sock_from_file(req->file);
4627 if (unlikely(!sock))
4630 if (req->flags & REQ_F_BUFFER_SELECT) {
4631 kbuf = io_recv_buffer_select(req, !force_nonblock);
4633 return PTR_ERR(kbuf);
4634 buf = u64_to_user_ptr(kbuf->addr);
4637 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4641 msg.msg_name = NULL;
4642 msg.msg_control = NULL;
4643 msg.msg_controllen = 0;
4644 msg.msg_namelen = 0;
4645 msg.msg_iocb = NULL;
4648 flags = req->sr_msg.msg_flags;
4649 if (flags & MSG_DONTWAIT)
4650 req->flags |= REQ_F_NOWAIT;
4651 else if (force_nonblock)
4652 flags |= MSG_DONTWAIT;
4654 ret = sock_recvmsg(sock, &msg, flags);
4655 if (force_nonblock && ret == -EAGAIN)
4657 if (ret == -ERESTARTSYS)
4660 if (req->flags & REQ_F_BUFFER_SELECTED)
4661 cflags = io_put_recv_kbuf(req);
4663 req_set_fail_links(req);
4664 __io_req_complete(req, issue_flags, ret, cflags);
4668 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4670 struct io_accept *accept = &req->accept;
4672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4674 if (sqe->ioprio || sqe->len || sqe->buf_index)
4677 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4678 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4679 accept->flags = READ_ONCE(sqe->accept_flags);
4680 accept->nofile = rlimit(RLIMIT_NOFILE);
4684 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4686 struct io_accept *accept = &req->accept;
4687 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4688 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4691 if (req->file->f_flags & O_NONBLOCK)
4692 req->flags |= REQ_F_NOWAIT;
4694 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4695 accept->addr_len, accept->flags,
4697 if (ret == -EAGAIN && force_nonblock)
4700 if (ret == -ERESTARTSYS)
4702 req_set_fail_links(req);
4704 __io_req_complete(req, issue_flags, ret, 0);
4708 static int io_connect_prep_async(struct io_kiocb *req)
4710 struct io_async_connect *io = req->async_data;
4711 struct io_connect *conn = &req->connect;
4713 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4716 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_connect *conn = &req->connect;
4720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4725 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726 conn->addr_len = READ_ONCE(sqe->addr2);
4730 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4732 struct io_async_connect __io, *io;
4733 unsigned file_flags;
4735 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4737 if (req->async_data) {
4738 io = req->async_data;
4740 ret = move_addr_to_kernel(req->connect.addr,
4741 req->connect.addr_len,
4748 file_flags = force_nonblock ? O_NONBLOCK : 0;
4750 ret = __sys_connect_file(req->file, &io->address,
4751 req->connect.addr_len, file_flags);
4752 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4753 if (req->async_data)
4755 if (io_alloc_async_data(req)) {
4759 io = req->async_data;
4760 memcpy(req->async_data, &__io, sizeof(__io));
4763 if (ret == -ERESTARTSYS)
4767 req_set_fail_links(req);
4768 __io_req_complete(req, issue_flags, ret, 0);
4771 #else /* !CONFIG_NET */
4772 #define IO_NETOP_FN(op) \
4773 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4775 return -EOPNOTSUPP; \
4778 #define IO_NETOP_PREP(op) \
4780 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4782 return -EOPNOTSUPP; \
4785 #define IO_NETOP_PREP_ASYNC(op) \
4787 static int io_##op##_prep_async(struct io_kiocb *req) \
4789 return -EOPNOTSUPP; \
4792 IO_NETOP_PREP_ASYNC(sendmsg);
4793 IO_NETOP_PREP_ASYNC(recvmsg);
4794 IO_NETOP_PREP_ASYNC(connect);
4795 IO_NETOP_PREP(accept);
4798 #endif /* CONFIG_NET */
4800 struct io_poll_table {
4801 struct poll_table_struct pt;
4802 struct io_kiocb *req;
4806 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4807 __poll_t mask, task_work_func_t func)
4811 /* for instances that support it check for an event match first: */
4812 if (mask && !(mask & poll->events))
4815 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4817 list_del_init(&poll->wait.entry);
4820 req->task_work.func = func;
4821 percpu_ref_get(&req->ctx->refs);
4824 * If this fails, then the task is exiting. When a task exits, the
4825 * work gets canceled, so just cancel this request as well instead
4826 * of executing it. We can't safely execute it anyway, as we may not
4827 * have the needed state needed for it anyway.
4829 ret = io_req_task_work_add(req);
4830 if (unlikely(ret)) {
4831 WRITE_ONCE(poll->canceled, true);
4832 io_req_task_work_add_fallback(req, func);
4837 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4838 __acquires(&req->ctx->completion_lock)
4840 struct io_ring_ctx *ctx = req->ctx;
4842 if (!req->result && !READ_ONCE(poll->canceled)) {
4843 struct poll_table_struct pt = { ._key = poll->events };
4845 req->result = vfs_poll(req->file, &pt) & poll->events;
4848 spin_lock_irq(&ctx->completion_lock);
4849 if (!req->result && !READ_ONCE(poll->canceled)) {
4850 add_wait_queue(poll->head, &poll->wait);
4857 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4859 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4860 if (req->opcode == IORING_OP_POLL_ADD)
4861 return req->async_data;
4862 return req->apoll->double_poll;
4865 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4867 if (req->opcode == IORING_OP_POLL_ADD)
4869 return &req->apoll->poll;
4872 static void io_poll_remove_double(struct io_kiocb *req)
4874 struct io_poll_iocb *poll = io_poll_get_double(req);
4876 lockdep_assert_held(&req->ctx->completion_lock);
4878 if (poll && poll->head) {
4879 struct wait_queue_head *head = poll->head;
4881 spin_lock(&head->lock);
4882 list_del_init(&poll->wait.entry);
4883 if (poll->wait.private)
4884 refcount_dec(&req->refs);
4886 spin_unlock(&head->lock);
4890 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4892 struct io_ring_ctx *ctx = req->ctx;
4894 io_poll_remove_double(req);
4895 req->poll.done = true;
4896 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4897 io_commit_cqring(ctx);
4900 static void io_poll_task_func(struct callback_head *cb)
4902 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4903 struct io_ring_ctx *ctx = req->ctx;
4904 struct io_kiocb *nxt;
4906 if (io_poll_rewait(req, &req->poll)) {
4907 spin_unlock_irq(&ctx->completion_lock);
4909 hash_del(&req->hash_node);
4910 io_poll_complete(req, req->result, 0);
4911 spin_unlock_irq(&ctx->completion_lock);
4913 nxt = io_put_req_find_next(req);
4914 io_cqring_ev_posted(ctx);
4916 __io_req_task_submit(nxt);
4919 percpu_ref_put(&ctx->refs);
4922 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4923 int sync, void *key)
4925 struct io_kiocb *req = wait->private;
4926 struct io_poll_iocb *poll = io_poll_get_single(req);
4927 __poll_t mask = key_to_poll(key);
4929 /* for instances that support it check for an event match first: */
4930 if (mask && !(mask & poll->events))
4933 list_del_init(&wait->entry);
4935 if (poll && poll->head) {
4938 spin_lock(&poll->head->lock);
4939 done = list_empty(&poll->wait.entry);
4941 list_del_init(&poll->wait.entry);
4942 /* make sure double remove sees this as being gone */
4943 wait->private = NULL;
4944 spin_unlock(&poll->head->lock);
4946 /* use wait func handler, so it matches the rq type */
4947 poll->wait.func(&poll->wait, mode, sync, key);
4950 refcount_dec(&req->refs);
4954 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4955 wait_queue_func_t wake_func)
4959 poll->canceled = false;
4960 poll->events = events;
4961 INIT_LIST_HEAD(&poll->wait.entry);
4962 init_waitqueue_func_entry(&poll->wait, wake_func);
4965 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4966 struct wait_queue_head *head,
4967 struct io_poll_iocb **poll_ptr)
4969 struct io_kiocb *req = pt->req;
4972 * If poll->head is already set, it's because the file being polled
4973 * uses multiple waitqueues for poll handling (eg one for read, one
4974 * for write). Setup a separate io_poll_iocb if this happens.
4976 if (unlikely(poll->head)) {
4977 struct io_poll_iocb *poll_one = poll;
4979 /* already have a 2nd entry, fail a third attempt */
4981 pt->error = -EINVAL;
4984 /* double add on the same waitqueue head, ignore */
4985 if (poll->head == head)
4987 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4989 pt->error = -ENOMEM;
4992 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4993 refcount_inc(&req->refs);
4994 poll->wait.private = req;
5001 if (poll->events & EPOLLEXCLUSIVE)
5002 add_wait_queue_exclusive(head, &poll->wait);
5004 add_wait_queue(head, &poll->wait);
5007 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5008 struct poll_table_struct *p)
5010 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5011 struct async_poll *apoll = pt->req->apoll;
5013 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5016 static void io_async_task_func(struct callback_head *cb)
5018 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5019 struct async_poll *apoll = req->apoll;
5020 struct io_ring_ctx *ctx = req->ctx;
5022 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5024 if (io_poll_rewait(req, &apoll->poll)) {
5025 spin_unlock_irq(&ctx->completion_lock);
5026 percpu_ref_put(&ctx->refs);
5030 /* If req is still hashed, it cannot have been canceled. Don't check. */
5031 if (hash_hashed(&req->hash_node))
5032 hash_del(&req->hash_node);
5034 io_poll_remove_double(req);
5035 spin_unlock_irq(&ctx->completion_lock);
5037 if (!READ_ONCE(apoll->poll.canceled))
5038 __io_req_task_submit(req);
5040 __io_req_task_cancel(req, -ECANCELED);
5042 percpu_ref_put(&ctx->refs);
5043 kfree(apoll->double_poll);
5047 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5050 struct io_kiocb *req = wait->private;
5051 struct io_poll_iocb *poll = &req->apoll->poll;
5053 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5056 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5059 static void io_poll_req_insert(struct io_kiocb *req)
5061 struct io_ring_ctx *ctx = req->ctx;
5062 struct hlist_head *list;
5064 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5065 hlist_add_head(&req->hash_node, list);
5068 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5069 struct io_poll_iocb *poll,
5070 struct io_poll_table *ipt, __poll_t mask,
5071 wait_queue_func_t wake_func)
5072 __acquires(&ctx->completion_lock)
5074 struct io_ring_ctx *ctx = req->ctx;
5075 bool cancel = false;
5077 INIT_HLIST_NODE(&req->hash_node);
5078 io_init_poll_iocb(poll, mask, wake_func);
5079 poll->file = req->file;
5080 poll->wait.private = req;
5082 ipt->pt._key = mask;
5084 ipt->error = -EINVAL;
5086 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5088 spin_lock_irq(&ctx->completion_lock);
5089 if (likely(poll->head)) {
5090 spin_lock(&poll->head->lock);
5091 if (unlikely(list_empty(&poll->wait.entry))) {
5097 if (mask || ipt->error)
5098 list_del_init(&poll->wait.entry);
5100 WRITE_ONCE(poll->canceled, true);
5101 else if (!poll->done) /* actually waiting for an event */
5102 io_poll_req_insert(req);
5103 spin_unlock(&poll->head->lock);
5109 static bool io_arm_poll_handler(struct io_kiocb *req)
5111 const struct io_op_def *def = &io_op_defs[req->opcode];
5112 struct io_ring_ctx *ctx = req->ctx;
5113 struct async_poll *apoll;
5114 struct io_poll_table ipt;
5118 if (!req->file || !file_can_poll(req->file))
5120 if (req->flags & REQ_F_POLLED)
5124 else if (def->pollout)
5128 /* if we can't nonblock try, then no point in arming a poll handler */
5129 if (!io_file_supports_async(req->file, rw))
5132 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5133 if (unlikely(!apoll))
5135 apoll->double_poll = NULL;
5137 req->flags |= REQ_F_POLLED;
5142 mask |= POLLIN | POLLRDNORM;
5144 mask |= POLLOUT | POLLWRNORM;
5146 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5147 if ((req->opcode == IORING_OP_RECVMSG) &&
5148 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5151 mask |= POLLERR | POLLPRI;
5153 ipt.pt._qproc = io_async_queue_proc;
5155 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5157 if (ret || ipt.error) {
5158 io_poll_remove_double(req);
5159 spin_unlock_irq(&ctx->completion_lock);
5160 kfree(apoll->double_poll);
5164 spin_unlock_irq(&ctx->completion_lock);
5165 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5166 apoll->poll.events);
5170 static bool __io_poll_remove_one(struct io_kiocb *req,
5171 struct io_poll_iocb *poll)
5173 bool do_complete = false;
5175 spin_lock(&poll->head->lock);
5176 WRITE_ONCE(poll->canceled, true);
5177 if (!list_empty(&poll->wait.entry)) {
5178 list_del_init(&poll->wait.entry);
5181 spin_unlock(&poll->head->lock);
5182 hash_del(&req->hash_node);
5186 static bool io_poll_remove_one(struct io_kiocb *req)
5190 io_poll_remove_double(req);
5192 if (req->opcode == IORING_OP_POLL_ADD) {
5193 do_complete = __io_poll_remove_one(req, &req->poll);
5195 struct async_poll *apoll = req->apoll;
5197 /* non-poll requests have submit ref still */
5198 do_complete = __io_poll_remove_one(req, &apoll->poll);
5201 kfree(apoll->double_poll);
5207 io_cqring_fill_event(req, -ECANCELED);
5208 io_commit_cqring(req->ctx);
5209 req_set_fail_links(req);
5210 io_put_req_deferred(req, 1);
5217 * Returns true if we found and killed one or more poll requests
5219 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5220 struct files_struct *files)
5222 struct hlist_node *tmp;
5223 struct io_kiocb *req;
5226 spin_lock_irq(&ctx->completion_lock);
5227 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5228 struct hlist_head *list;
5230 list = &ctx->cancel_hash[i];
5231 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5232 if (io_match_task(req, tsk, files))
5233 posted += io_poll_remove_one(req);
5236 spin_unlock_irq(&ctx->completion_lock);
5239 io_cqring_ev_posted(ctx);
5244 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5246 struct hlist_head *list;
5247 struct io_kiocb *req;
5249 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5250 hlist_for_each_entry(req, list, hash_node) {
5251 if (sqe_addr != req->user_data)
5253 if (io_poll_remove_one(req))
5261 static int io_poll_remove_prep(struct io_kiocb *req,
5262 const struct io_uring_sqe *sqe)
5264 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5266 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5270 req->poll_remove.addr = READ_ONCE(sqe->addr);
5275 * Find a running poll command that matches one specified in sqe->addr,
5276 * and remove it if found.
5278 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5280 struct io_ring_ctx *ctx = req->ctx;
5283 spin_lock_irq(&ctx->completion_lock);
5284 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5285 spin_unlock_irq(&ctx->completion_lock);
5288 req_set_fail_links(req);
5289 io_req_complete(req, ret);
5293 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5296 struct io_kiocb *req = wait->private;
5297 struct io_poll_iocb *poll = &req->poll;
5299 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5302 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5303 struct poll_table_struct *p)
5305 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5307 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5310 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5312 struct io_poll_iocb *poll = &req->poll;
5315 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5317 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5320 events = READ_ONCE(sqe->poll32_events);
5322 events = swahw32(events);
5324 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5325 (events & EPOLLEXCLUSIVE);
5329 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5331 struct io_poll_iocb *poll = &req->poll;
5332 struct io_ring_ctx *ctx = req->ctx;
5333 struct io_poll_table ipt;
5336 ipt.pt._qproc = io_poll_queue_proc;
5338 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5341 if (mask) { /* no async, we'd stolen it */
5343 io_poll_complete(req, mask, 0);
5345 spin_unlock_irq(&ctx->completion_lock);
5348 io_cqring_ev_posted(ctx);
5354 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5356 struct io_timeout_data *data = container_of(timer,
5357 struct io_timeout_data, timer);
5358 struct io_kiocb *req = data->req;
5359 struct io_ring_ctx *ctx = req->ctx;
5360 unsigned long flags;
5362 spin_lock_irqsave(&ctx->completion_lock, flags);
5363 list_del_init(&req->timeout.list);
5364 atomic_set(&req->ctx->cq_timeouts,
5365 atomic_read(&req->ctx->cq_timeouts) + 1);
5367 io_cqring_fill_event(req, -ETIME);
5368 io_commit_cqring(ctx);
5369 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5371 io_cqring_ev_posted(ctx);
5372 req_set_fail_links(req);
5374 return HRTIMER_NORESTART;
5377 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5380 struct io_timeout_data *io;
5381 struct io_kiocb *req;
5384 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5385 if (user_data == req->user_data) {
5392 return ERR_PTR(ret);
5394 io = req->async_data;
5395 ret = hrtimer_try_to_cancel(&io->timer);
5397 return ERR_PTR(-EALREADY);
5398 list_del_init(&req->timeout.list);
5402 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5404 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5407 return PTR_ERR(req);
5409 req_set_fail_links(req);
5410 io_cqring_fill_event(req, -ECANCELED);
5411 io_put_req_deferred(req, 1);
5415 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5416 struct timespec64 *ts, enum hrtimer_mode mode)
5418 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5419 struct io_timeout_data *data;
5422 return PTR_ERR(req);
5424 req->timeout.off = 0; /* noseq */
5425 data = req->async_data;
5426 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5427 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5428 data->timer.function = io_timeout_fn;
5429 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5433 static int io_timeout_remove_prep(struct io_kiocb *req,
5434 const struct io_uring_sqe *sqe)
5436 struct io_timeout_rem *tr = &req->timeout_rem;
5438 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5440 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5442 if (sqe->ioprio || sqe->buf_index || sqe->len)
5445 tr->addr = READ_ONCE(sqe->addr);
5446 tr->flags = READ_ONCE(sqe->timeout_flags);
5447 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5448 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5450 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5452 } else if (tr->flags) {
5453 /* timeout removal doesn't support flags */
5460 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5462 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5467 * Remove or update an existing timeout command
5469 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5471 struct io_timeout_rem *tr = &req->timeout_rem;
5472 struct io_ring_ctx *ctx = req->ctx;
5475 spin_lock_irq(&ctx->completion_lock);
5476 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5477 ret = io_timeout_cancel(ctx, tr->addr);
5479 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5480 io_translate_timeout_mode(tr->flags));
5482 io_cqring_fill_event(req, ret);
5483 io_commit_cqring(ctx);
5484 spin_unlock_irq(&ctx->completion_lock);
5485 io_cqring_ev_posted(ctx);
5487 req_set_fail_links(req);
5492 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5493 bool is_timeout_link)
5495 struct io_timeout_data *data;
5497 u32 off = READ_ONCE(sqe->off);
5499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5501 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5503 if (off && is_timeout_link)
5505 flags = READ_ONCE(sqe->timeout_flags);
5506 if (flags & ~IORING_TIMEOUT_ABS)
5509 req->timeout.off = off;
5511 if (!req->async_data && io_alloc_async_data(req))
5514 data = req->async_data;
5517 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5520 data->mode = io_translate_timeout_mode(flags);
5521 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5522 io_req_track_inflight(req);
5526 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5528 struct io_ring_ctx *ctx = req->ctx;
5529 struct io_timeout_data *data = req->async_data;
5530 struct list_head *entry;
5531 u32 tail, off = req->timeout.off;
5533 spin_lock_irq(&ctx->completion_lock);
5536 * sqe->off holds how many events that need to occur for this
5537 * timeout event to be satisfied. If it isn't set, then this is
5538 * a pure timeout request, sequence isn't used.
5540 if (io_is_timeout_noseq(req)) {
5541 entry = ctx->timeout_list.prev;
5545 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5546 req->timeout.target_seq = tail + off;
5548 /* Update the last seq here in case io_flush_timeouts() hasn't.
5549 * This is safe because ->completion_lock is held, and submissions
5550 * and completions are never mixed in the same ->completion_lock section.
5552 ctx->cq_last_tm_flush = tail;
5555 * Insertion sort, ensuring the first entry in the list is always
5556 * the one we need first.
5558 list_for_each_prev(entry, &ctx->timeout_list) {
5559 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5562 if (io_is_timeout_noseq(nxt))
5564 /* nxt.seq is behind @tail, otherwise would've been completed */
5565 if (off >= nxt->timeout.target_seq - tail)
5569 list_add(&req->timeout.list, entry);
5570 data->timer.function = io_timeout_fn;
5571 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5572 spin_unlock_irq(&ctx->completion_lock);
5576 struct io_cancel_data {
5577 struct io_ring_ctx *ctx;
5581 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5583 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5584 struct io_cancel_data *cd = data;
5586 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5589 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5590 struct io_ring_ctx *ctx)
5592 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5593 enum io_wq_cancel cancel_ret;
5596 if (!tctx || !tctx->io_wq)
5599 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5600 switch (cancel_ret) {
5601 case IO_WQ_CANCEL_OK:
5604 case IO_WQ_CANCEL_RUNNING:
5607 case IO_WQ_CANCEL_NOTFOUND:
5615 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5616 struct io_kiocb *req, __u64 sqe_addr,
5619 unsigned long flags;
5622 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5623 if (ret != -ENOENT) {
5624 spin_lock_irqsave(&ctx->completion_lock, flags);
5628 spin_lock_irqsave(&ctx->completion_lock, flags);
5629 ret = io_timeout_cancel(ctx, sqe_addr);
5632 ret = io_poll_cancel(ctx, sqe_addr);
5636 io_cqring_fill_event(req, ret);
5637 io_commit_cqring(ctx);
5638 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5639 io_cqring_ev_posted(ctx);
5642 req_set_fail_links(req);
5646 static int io_async_cancel_prep(struct io_kiocb *req,
5647 const struct io_uring_sqe *sqe)
5649 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5651 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5653 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5656 req->cancel.addr = READ_ONCE(sqe->addr);
5660 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5662 struct io_ring_ctx *ctx = req->ctx;
5664 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5668 static int io_rsrc_update_prep(struct io_kiocb *req,
5669 const struct io_uring_sqe *sqe)
5671 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5673 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5675 if (sqe->ioprio || sqe->rw_flags)
5678 req->rsrc_update.offset = READ_ONCE(sqe->off);
5679 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5680 if (!req->rsrc_update.nr_args)
5682 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5686 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5688 struct io_ring_ctx *ctx = req->ctx;
5689 struct io_uring_rsrc_update up;
5692 if (issue_flags & IO_URING_F_NONBLOCK)
5695 up.offset = req->rsrc_update.offset;
5696 up.data = req->rsrc_update.arg;
5698 mutex_lock(&ctx->uring_lock);
5699 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5700 mutex_unlock(&ctx->uring_lock);
5703 req_set_fail_links(req);
5704 __io_req_complete(req, issue_flags, ret, 0);
5708 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5710 switch (req->opcode) {
5713 case IORING_OP_READV:
5714 case IORING_OP_READ_FIXED:
5715 case IORING_OP_READ:
5716 return io_read_prep(req, sqe);
5717 case IORING_OP_WRITEV:
5718 case IORING_OP_WRITE_FIXED:
5719 case IORING_OP_WRITE:
5720 return io_write_prep(req, sqe);
5721 case IORING_OP_POLL_ADD:
5722 return io_poll_add_prep(req, sqe);
5723 case IORING_OP_POLL_REMOVE:
5724 return io_poll_remove_prep(req, sqe);
5725 case IORING_OP_FSYNC:
5726 return io_fsync_prep(req, sqe);
5727 case IORING_OP_SYNC_FILE_RANGE:
5728 return io_sfr_prep(req, sqe);
5729 case IORING_OP_SENDMSG:
5730 case IORING_OP_SEND:
5731 return io_sendmsg_prep(req, sqe);
5732 case IORING_OP_RECVMSG:
5733 case IORING_OP_RECV:
5734 return io_recvmsg_prep(req, sqe);
5735 case IORING_OP_CONNECT:
5736 return io_connect_prep(req, sqe);
5737 case IORING_OP_TIMEOUT:
5738 return io_timeout_prep(req, sqe, false);
5739 case IORING_OP_TIMEOUT_REMOVE:
5740 return io_timeout_remove_prep(req, sqe);
5741 case IORING_OP_ASYNC_CANCEL:
5742 return io_async_cancel_prep(req, sqe);
5743 case IORING_OP_LINK_TIMEOUT:
5744 return io_timeout_prep(req, sqe, true);
5745 case IORING_OP_ACCEPT:
5746 return io_accept_prep(req, sqe);
5747 case IORING_OP_FALLOCATE:
5748 return io_fallocate_prep(req, sqe);
5749 case IORING_OP_OPENAT:
5750 return io_openat_prep(req, sqe);
5751 case IORING_OP_CLOSE:
5752 return io_close_prep(req, sqe);
5753 case IORING_OP_FILES_UPDATE:
5754 return io_rsrc_update_prep(req, sqe);
5755 case IORING_OP_STATX:
5756 return io_statx_prep(req, sqe);
5757 case IORING_OP_FADVISE:
5758 return io_fadvise_prep(req, sqe);
5759 case IORING_OP_MADVISE:
5760 return io_madvise_prep(req, sqe);
5761 case IORING_OP_OPENAT2:
5762 return io_openat2_prep(req, sqe);
5763 case IORING_OP_EPOLL_CTL:
5764 return io_epoll_ctl_prep(req, sqe);
5765 case IORING_OP_SPLICE:
5766 return io_splice_prep(req, sqe);
5767 case IORING_OP_PROVIDE_BUFFERS:
5768 return io_provide_buffers_prep(req, sqe);
5769 case IORING_OP_REMOVE_BUFFERS:
5770 return io_remove_buffers_prep(req, sqe);
5772 return io_tee_prep(req, sqe);
5773 case IORING_OP_SHUTDOWN:
5774 return io_shutdown_prep(req, sqe);
5775 case IORING_OP_RENAMEAT:
5776 return io_renameat_prep(req, sqe);
5777 case IORING_OP_UNLINKAT:
5778 return io_unlinkat_prep(req, sqe);
5781 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5786 static int io_req_prep_async(struct io_kiocb *req)
5788 switch (req->opcode) {
5789 case IORING_OP_READV:
5790 case IORING_OP_READ_FIXED:
5791 case IORING_OP_READ:
5792 return io_rw_prep_async(req, READ);
5793 case IORING_OP_WRITEV:
5794 case IORING_OP_WRITE_FIXED:
5795 case IORING_OP_WRITE:
5796 return io_rw_prep_async(req, WRITE);
5797 case IORING_OP_SENDMSG:
5798 case IORING_OP_SEND:
5799 return io_sendmsg_prep_async(req);
5800 case IORING_OP_RECVMSG:
5801 case IORING_OP_RECV:
5802 return io_recvmsg_prep_async(req);
5803 case IORING_OP_CONNECT:
5804 return io_connect_prep_async(req);
5809 static int io_req_defer_prep(struct io_kiocb *req)
5811 if (!io_op_defs[req->opcode].needs_async_data)
5813 /* some opcodes init it during the inital prep */
5814 if (req->async_data)
5816 if (__io_alloc_async_data(req))
5818 return io_req_prep_async(req);
5821 static u32 io_get_sequence(struct io_kiocb *req)
5823 struct io_kiocb *pos;
5824 struct io_ring_ctx *ctx = req->ctx;
5825 u32 total_submitted, nr_reqs = 0;
5827 io_for_each_link(pos, req)
5830 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5831 return total_submitted - nr_reqs;
5834 static int io_req_defer(struct io_kiocb *req)
5836 struct io_ring_ctx *ctx = req->ctx;
5837 struct io_defer_entry *de;
5841 /* Still need defer if there is pending req in defer list. */
5842 if (likely(list_empty_careful(&ctx->defer_list) &&
5843 !(req->flags & REQ_F_IO_DRAIN)))
5846 seq = io_get_sequence(req);
5847 /* Still a chance to pass the sequence check */
5848 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5851 ret = io_req_defer_prep(req);
5854 io_prep_async_link(req);
5855 de = kmalloc(sizeof(*de), GFP_KERNEL);
5859 spin_lock_irq(&ctx->completion_lock);
5860 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5861 spin_unlock_irq(&ctx->completion_lock);
5863 io_queue_async_work(req);
5864 return -EIOCBQUEUED;
5867 trace_io_uring_defer(ctx, req, req->user_data);
5870 list_add_tail(&de->list, &ctx->defer_list);
5871 spin_unlock_irq(&ctx->completion_lock);
5872 return -EIOCBQUEUED;
5875 static void __io_clean_op(struct io_kiocb *req)
5877 if (req->flags & REQ_F_BUFFER_SELECTED) {
5878 switch (req->opcode) {
5879 case IORING_OP_READV:
5880 case IORING_OP_READ_FIXED:
5881 case IORING_OP_READ:
5882 kfree((void *)(unsigned long)req->rw.addr);
5884 case IORING_OP_RECVMSG:
5885 case IORING_OP_RECV:
5886 kfree(req->sr_msg.kbuf);
5889 req->flags &= ~REQ_F_BUFFER_SELECTED;
5892 if (req->flags & REQ_F_NEED_CLEANUP) {
5893 switch (req->opcode) {
5894 case IORING_OP_READV:
5895 case IORING_OP_READ_FIXED:
5896 case IORING_OP_READ:
5897 case IORING_OP_WRITEV:
5898 case IORING_OP_WRITE_FIXED:
5899 case IORING_OP_WRITE: {
5900 struct io_async_rw *io = req->async_data;
5902 kfree(io->free_iovec);
5905 case IORING_OP_RECVMSG:
5906 case IORING_OP_SENDMSG: {
5907 struct io_async_msghdr *io = req->async_data;
5909 kfree(io->free_iov);
5912 case IORING_OP_SPLICE:
5914 io_put_file(req, req->splice.file_in,
5915 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5917 case IORING_OP_OPENAT:
5918 case IORING_OP_OPENAT2:
5919 if (req->open.filename)
5920 putname(req->open.filename);
5922 case IORING_OP_RENAMEAT:
5923 putname(req->rename.oldpath);
5924 putname(req->rename.newpath);
5926 case IORING_OP_UNLINKAT:
5927 putname(req->unlink.filename);
5930 req->flags &= ~REQ_F_NEED_CLEANUP;
5934 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5936 struct io_ring_ctx *ctx = req->ctx;
5937 const struct cred *creds = NULL;
5940 if (req->work.creds && req->work.creds != current_cred())
5941 creds = override_creds(req->work.creds);
5943 switch (req->opcode) {
5945 ret = io_nop(req, issue_flags);
5947 case IORING_OP_READV:
5948 case IORING_OP_READ_FIXED:
5949 case IORING_OP_READ:
5950 ret = io_read(req, issue_flags);
5952 case IORING_OP_WRITEV:
5953 case IORING_OP_WRITE_FIXED:
5954 case IORING_OP_WRITE:
5955 ret = io_write(req, issue_flags);
5957 case IORING_OP_FSYNC:
5958 ret = io_fsync(req, issue_flags);
5960 case IORING_OP_POLL_ADD:
5961 ret = io_poll_add(req, issue_flags);
5963 case IORING_OP_POLL_REMOVE:
5964 ret = io_poll_remove(req, issue_flags);
5966 case IORING_OP_SYNC_FILE_RANGE:
5967 ret = io_sync_file_range(req, issue_flags);
5969 case IORING_OP_SENDMSG:
5970 ret = io_sendmsg(req, issue_flags);
5972 case IORING_OP_SEND:
5973 ret = io_send(req, issue_flags);
5975 case IORING_OP_RECVMSG:
5976 ret = io_recvmsg(req, issue_flags);
5978 case IORING_OP_RECV:
5979 ret = io_recv(req, issue_flags);
5981 case IORING_OP_TIMEOUT:
5982 ret = io_timeout(req, issue_flags);
5984 case IORING_OP_TIMEOUT_REMOVE:
5985 ret = io_timeout_remove(req, issue_flags);
5987 case IORING_OP_ACCEPT:
5988 ret = io_accept(req, issue_flags);
5990 case IORING_OP_CONNECT:
5991 ret = io_connect(req, issue_flags);
5993 case IORING_OP_ASYNC_CANCEL:
5994 ret = io_async_cancel(req, issue_flags);
5996 case IORING_OP_FALLOCATE:
5997 ret = io_fallocate(req, issue_flags);
5999 case IORING_OP_OPENAT:
6000 ret = io_openat(req, issue_flags);
6002 case IORING_OP_CLOSE:
6003 ret = io_close(req, issue_flags);
6005 case IORING_OP_FILES_UPDATE:
6006 ret = io_files_update(req, issue_flags);
6008 case IORING_OP_STATX:
6009 ret = io_statx(req, issue_flags);
6011 case IORING_OP_FADVISE:
6012 ret = io_fadvise(req, issue_flags);
6014 case IORING_OP_MADVISE:
6015 ret = io_madvise(req, issue_flags);
6017 case IORING_OP_OPENAT2:
6018 ret = io_openat2(req, issue_flags);
6020 case IORING_OP_EPOLL_CTL:
6021 ret = io_epoll_ctl(req, issue_flags);
6023 case IORING_OP_SPLICE:
6024 ret = io_splice(req, issue_flags);
6026 case IORING_OP_PROVIDE_BUFFERS:
6027 ret = io_provide_buffers(req, issue_flags);
6029 case IORING_OP_REMOVE_BUFFERS:
6030 ret = io_remove_buffers(req, issue_flags);
6033 ret = io_tee(req, issue_flags);
6035 case IORING_OP_SHUTDOWN:
6036 ret = io_shutdown(req, issue_flags);
6038 case IORING_OP_RENAMEAT:
6039 ret = io_renameat(req, issue_flags);
6041 case IORING_OP_UNLINKAT:
6042 ret = io_unlinkat(req, issue_flags);
6050 revert_creds(creds);
6055 /* If the op doesn't have a file, we're not polling for it */
6056 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6057 const bool in_async = io_wq_current_is_worker();
6059 /* workqueue context doesn't hold uring_lock, grab it now */
6061 mutex_lock(&ctx->uring_lock);
6063 io_iopoll_req_issued(req, in_async);
6066 mutex_unlock(&ctx->uring_lock);
6072 static void io_wq_submit_work(struct io_wq_work *work)
6074 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6075 struct io_kiocb *timeout;
6078 timeout = io_prep_linked_timeout(req);
6080 io_queue_linked_timeout(timeout);
6082 if (work->flags & IO_WQ_WORK_CANCEL)
6087 ret = io_issue_sqe(req, 0);
6089 * We can get EAGAIN for polled IO even though we're
6090 * forcing a sync submission from here, since we can't
6091 * wait for request slots on the block side.
6099 /* avoid locking problems by failing it from a clean context */
6101 /* io-wq is going to take one down */
6102 refcount_inc(&req->refs);
6103 io_req_task_queue_fail(req, ret);
6107 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6110 struct fixed_rsrc_table *table;
6112 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6113 return table->files[index & IORING_FILE_TABLE_MASK];
6116 static struct file *io_file_get(struct io_submit_state *state,
6117 struct io_kiocb *req, int fd, bool fixed)
6119 struct io_ring_ctx *ctx = req->ctx;
6123 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6125 fd = array_index_nospec(fd, ctx->nr_user_files);
6126 file = io_file_from_index(ctx, fd);
6127 io_set_resource_node(req);
6129 trace_io_uring_file_get(ctx, fd);
6130 file = __io_file_get(state, fd);
6133 if (file && unlikely(file->f_op == &io_uring_fops))
6134 io_req_track_inflight(req);
6138 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6140 struct io_timeout_data *data = container_of(timer,
6141 struct io_timeout_data, timer);
6142 struct io_kiocb *prev, *req = data->req;
6143 struct io_ring_ctx *ctx = req->ctx;
6144 unsigned long flags;
6146 spin_lock_irqsave(&ctx->completion_lock, flags);
6147 prev = req->timeout.head;
6148 req->timeout.head = NULL;
6151 * We don't expect the list to be empty, that will only happen if we
6152 * race with the completion of the linked work.
6154 if (prev && refcount_inc_not_zero(&prev->refs))
6155 io_remove_next_linked(prev);
6158 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6161 req_set_fail_links(prev);
6162 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6163 io_put_req_deferred(prev, 1);
6165 io_req_complete_post(req, -ETIME, 0);
6166 io_put_req_deferred(req, 1);
6168 return HRTIMER_NORESTART;
6171 static void __io_queue_linked_timeout(struct io_kiocb *req)
6174 * If the back reference is NULL, then our linked request finished
6175 * before we got a chance to setup the timer
6177 if (req->timeout.head) {
6178 struct io_timeout_data *data = req->async_data;
6180 data->timer.function = io_link_timeout_fn;
6181 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6186 static void io_queue_linked_timeout(struct io_kiocb *req)
6188 struct io_ring_ctx *ctx = req->ctx;
6190 spin_lock_irq(&ctx->completion_lock);
6191 __io_queue_linked_timeout(req);
6192 spin_unlock_irq(&ctx->completion_lock);
6194 /* drop submission reference */
6198 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6200 struct io_kiocb *nxt = req->link;
6202 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6203 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6206 nxt->timeout.head = req;
6207 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6208 req->flags |= REQ_F_LINK_TIMEOUT;
6212 static void __io_queue_sqe(struct io_kiocb *req)
6214 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6217 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6220 * We async punt it if the file wasn't marked NOWAIT, or if the file
6221 * doesn't support non-blocking read/write attempts
6223 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6224 if (!io_arm_poll_handler(req)) {
6226 * Queued up for async execution, worker will release
6227 * submit reference when the iocb is actually submitted.
6229 io_queue_async_work(req);
6231 } else if (likely(!ret)) {
6232 /* drop submission reference */
6233 if (req->flags & REQ_F_COMPLETE_INLINE) {
6234 struct io_ring_ctx *ctx = req->ctx;
6235 struct io_comp_state *cs = &ctx->submit_state.comp;
6237 cs->reqs[cs->nr++] = req;
6238 if (cs->nr == ARRAY_SIZE(cs->reqs))
6239 io_submit_flush_completions(cs, ctx);
6244 req_set_fail_links(req);
6246 io_req_complete(req, ret);
6249 io_queue_linked_timeout(linked_timeout);
6252 static void io_queue_sqe(struct io_kiocb *req)
6256 ret = io_req_defer(req);
6258 if (ret != -EIOCBQUEUED) {
6260 req_set_fail_links(req);
6262 io_req_complete(req, ret);
6264 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6265 ret = io_req_defer_prep(req);
6268 io_queue_async_work(req);
6270 __io_queue_sqe(req);
6275 * Check SQE restrictions (opcode and flags).
6277 * Returns 'true' if SQE is allowed, 'false' otherwise.
6279 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6280 struct io_kiocb *req,
6281 unsigned int sqe_flags)
6283 if (!ctx->restricted)
6286 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6289 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6290 ctx->restrictions.sqe_flags_required)
6293 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6294 ctx->restrictions.sqe_flags_required))
6300 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6301 const struct io_uring_sqe *sqe)
6303 struct io_submit_state *state;
6304 unsigned int sqe_flags;
6305 int personality, ret = 0;
6307 req->opcode = READ_ONCE(sqe->opcode);
6308 /* same numerical values with corresponding REQ_F_*, safe to copy */
6309 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6310 req->user_data = READ_ONCE(sqe->user_data);
6311 req->async_data = NULL;
6315 req->fixed_rsrc_refs = NULL;
6316 /* one is dropped after submission, the other at completion */
6317 refcount_set(&req->refs, 2);
6318 req->task = current;
6321 /* enforce forwards compatibility on users */
6322 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6327 if (unlikely(req->opcode >= IORING_OP_LAST))
6330 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6333 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6334 !io_op_defs[req->opcode].buffer_select)
6337 req->work.list.next = NULL;
6338 personality = READ_ONCE(sqe->personality);
6340 req->work.creds = idr_find(&ctx->personality_idr, personality);
6341 if (!req->work.creds)
6343 get_cred(req->work.creds);
6345 req->work.creds = NULL;
6347 req->work.flags = 0;
6348 state = &ctx->submit_state;
6351 * Plug now if we have more than 1 IO left after this, and the target
6352 * is potentially a read/write to block based storage.
6354 if (!state->plug_started && state->ios_left > 1 &&
6355 io_op_defs[req->opcode].plug) {
6356 blk_start_plug(&state->plug);
6357 state->plug_started = true;
6360 if (io_op_defs[req->opcode].needs_file) {
6361 bool fixed = req->flags & REQ_F_FIXED_FILE;
6363 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6364 if (unlikely(!req->file))
6372 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6373 const struct io_uring_sqe *sqe)
6375 struct io_submit_link *link = &ctx->submit_state.link;
6378 ret = io_init_req(ctx, req, sqe);
6379 if (unlikely(ret)) {
6382 io_req_complete(req, ret);
6384 /* fail even hard links since we don't submit */
6385 link->head->flags |= REQ_F_FAIL_LINK;
6386 io_put_req(link->head);
6387 io_req_complete(link->head, -ECANCELED);
6392 ret = io_req_prep(req, sqe);
6396 /* don't need @sqe from now on */
6397 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6398 true, ctx->flags & IORING_SETUP_SQPOLL);
6401 * If we already have a head request, queue this one for async
6402 * submittal once the head completes. If we don't have a head but
6403 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6404 * submitted sync once the chain is complete. If none of those
6405 * conditions are true (normal request), then just queue it.
6408 struct io_kiocb *head = link->head;
6411 * Taking sequential execution of a link, draining both sides
6412 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6413 * requests in the link. So, it drains the head and the
6414 * next after the link request. The last one is done via
6415 * drain_next flag to persist the effect across calls.
6417 if (req->flags & REQ_F_IO_DRAIN) {
6418 head->flags |= REQ_F_IO_DRAIN;
6419 ctx->drain_next = 1;
6421 ret = io_req_defer_prep(req);
6424 trace_io_uring_link(ctx, req, head);
6425 link->last->link = req;
6428 /* last request of a link, enqueue the link */
6429 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6434 if (unlikely(ctx->drain_next)) {
6435 req->flags |= REQ_F_IO_DRAIN;
6436 ctx->drain_next = 0;
6438 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6450 * Batched submission is done, ensure local IO is flushed out.
6452 static void io_submit_state_end(struct io_submit_state *state,
6453 struct io_ring_ctx *ctx)
6455 if (state->link.head)
6456 io_queue_sqe(state->link.head);
6458 io_submit_flush_completions(&state->comp, ctx);
6459 if (state->plug_started)
6460 blk_finish_plug(&state->plug);
6461 io_state_file_put(state);
6465 * Start submission side cache.
6467 static void io_submit_state_start(struct io_submit_state *state,
6468 unsigned int max_ios)
6470 state->plug_started = false;
6471 state->ios_left = max_ios;
6472 /* set only head, no need to init link_last in advance */
6473 state->link.head = NULL;
6476 static void io_commit_sqring(struct io_ring_ctx *ctx)
6478 struct io_rings *rings = ctx->rings;
6481 * Ensure any loads from the SQEs are done at this point,
6482 * since once we write the new head, the application could
6483 * write new data to them.
6485 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6489 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6490 * that is mapped by userspace. This means that care needs to be taken to
6491 * ensure that reads are stable, as we cannot rely on userspace always
6492 * being a good citizen. If members of the sqe are validated and then later
6493 * used, it's important that those reads are done through READ_ONCE() to
6494 * prevent a re-load down the line.
6496 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6498 u32 *sq_array = ctx->sq_array;
6502 * The cached sq head (or cq tail) serves two purposes:
6504 * 1) allows us to batch the cost of updating the user visible
6506 * 2) allows the kernel side to track the head on its own, even
6507 * though the application is the one updating it.
6509 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6510 if (likely(head < ctx->sq_entries))
6511 return &ctx->sq_sqes[head];
6513 /* drop invalid entries */
6514 ctx->cached_sq_dropped++;
6515 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6519 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6523 /* if we have a backlog and couldn't flush it all, return BUSY */
6524 if (test_bit(0, &ctx->sq_check_overflow)) {
6525 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6529 /* make sure SQ entry isn't read before tail */
6530 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6532 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6535 percpu_counter_add(¤t->io_uring->inflight, nr);
6536 refcount_add(nr, ¤t->usage);
6537 io_submit_state_start(&ctx->submit_state, nr);
6539 while (submitted < nr) {
6540 const struct io_uring_sqe *sqe;
6541 struct io_kiocb *req;
6543 req = io_alloc_req(ctx);
6544 if (unlikely(!req)) {
6546 submitted = -EAGAIN;
6549 sqe = io_get_sqe(ctx);
6550 if (unlikely(!sqe)) {
6551 kmem_cache_free(req_cachep, req);
6554 /* will complete beyond this point, count as submitted */
6556 if (io_submit_sqe(ctx, req, sqe))
6560 if (unlikely(submitted != nr)) {
6561 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6562 struct io_uring_task *tctx = current->io_uring;
6563 int unused = nr - ref_used;
6565 percpu_ref_put_many(&ctx->refs, unused);
6566 percpu_counter_sub(&tctx->inflight, unused);
6567 put_task_struct_many(current, unused);
6570 io_submit_state_end(&ctx->submit_state, ctx);
6571 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6572 io_commit_sqring(ctx);
6577 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6579 /* Tell userspace we may need a wakeup call */
6580 spin_lock_irq(&ctx->completion_lock);
6581 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6582 spin_unlock_irq(&ctx->completion_lock);
6585 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6587 spin_lock_irq(&ctx->completion_lock);
6588 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6589 spin_unlock_irq(&ctx->completion_lock);
6592 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6594 unsigned int to_submit;
6597 to_submit = io_sqring_entries(ctx);
6598 /* if we're handling multiple rings, cap submit size for fairness */
6599 if (cap_entries && to_submit > 8)
6602 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6603 unsigned nr_events = 0;
6605 mutex_lock(&ctx->uring_lock);
6606 if (!list_empty(&ctx->iopoll_list))
6607 io_do_iopoll(ctx, &nr_events, 0);
6609 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6610 !(ctx->flags & IORING_SETUP_R_DISABLED))
6611 ret = io_submit_sqes(ctx, to_submit);
6612 mutex_unlock(&ctx->uring_lock);
6615 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6616 wake_up(&ctx->sqo_sq_wait);
6621 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6623 struct io_ring_ctx *ctx;
6624 unsigned sq_thread_idle = 0;
6626 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6627 if (sq_thread_idle < ctx->sq_thread_idle)
6628 sq_thread_idle = ctx->sq_thread_idle;
6631 sqd->sq_thread_idle = sq_thread_idle;
6634 static void io_sqd_init_new(struct io_sq_data *sqd)
6636 struct io_ring_ctx *ctx;
6638 while (!list_empty(&sqd->ctx_new_list)) {
6639 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6640 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6641 complete(&ctx->sq_thread_comp);
6644 io_sqd_update_thread_idle(sqd);
6647 static int io_sq_thread(void *data)
6649 struct io_sq_data *sqd = data;
6650 struct io_ring_ctx *ctx;
6651 unsigned long timeout = 0;
6652 char buf[TASK_COMM_LEN];
6655 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6656 set_task_comm(current, buf);
6657 current->pf_io_worker = NULL;
6659 if (sqd->sq_cpu != -1)
6660 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6662 set_cpus_allowed_ptr(current, cpu_online_mask);
6663 current->flags |= PF_NO_SETAFFINITY;
6665 wait_for_completion(&sqd->startup);
6667 down_read(&sqd->rw_lock);
6669 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6671 bool cap_entries, sqt_spin, needs_sched;
6673 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6674 up_read(&sqd->rw_lock);
6676 down_read(&sqd->rw_lock);
6679 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6680 io_sqd_init_new(sqd);
6681 timeout = jiffies + sqd->sq_thread_idle;
6683 if (fatal_signal_pending(current))
6686 cap_entries = !list_is_singular(&sqd->ctx_list);
6687 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6688 const struct cred *creds = NULL;
6690 if (ctx->sq_creds != current_cred())
6691 creds = override_creds(ctx->sq_creds);
6692 ret = __io_sq_thread(ctx, cap_entries);
6694 revert_creds(creds);
6695 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6699 if (sqt_spin || !time_after(jiffies, timeout)) {
6703 timeout = jiffies + sqd->sq_thread_idle;
6708 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6709 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6710 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6711 !list_empty_careful(&ctx->iopoll_list)) {
6712 needs_sched = false;
6715 if (io_sqring_entries(ctx)) {
6716 needs_sched = false;
6721 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6722 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6723 io_ring_set_wakeup_flag(ctx);
6725 up_read(&sqd->rw_lock);
6728 down_read(&sqd->rw_lock);
6729 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6730 io_ring_clear_wakeup_flag(ctx);
6733 finish_wait(&sqd->wait, &wait);
6734 timeout = jiffies + sqd->sq_thread_idle;
6737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6738 io_uring_cancel_sqpoll(ctx);
6739 up_read(&sqd->rw_lock);
6743 down_write(&sqd->rw_lock);
6745 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6746 io_ring_set_wakeup_flag(ctx);
6747 up_write(&sqd->rw_lock);
6748 complete(&sqd->exited);
6752 struct io_wait_queue {
6753 struct wait_queue_entry wq;
6754 struct io_ring_ctx *ctx;
6756 unsigned nr_timeouts;
6759 static inline bool io_should_wake(struct io_wait_queue *iowq)
6761 struct io_ring_ctx *ctx = iowq->ctx;
6764 * Wake up if we have enough events, or if a timeout occurred since we
6765 * started waiting. For timeouts, we always want to return to userspace,
6766 * regardless of event count.
6768 return io_cqring_events(ctx) >= iowq->to_wait ||
6769 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6772 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6773 int wake_flags, void *key)
6775 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6779 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6780 * the task, and the next invocation will do it.
6782 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6783 return autoremove_wake_function(curr, mode, wake_flags, key);
6787 static int io_run_task_work_sig(void)
6789 if (io_run_task_work())
6791 if (!signal_pending(current))
6793 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6794 return -ERESTARTSYS;
6798 /* when returns >0, the caller should retry */
6799 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6800 struct io_wait_queue *iowq,
6801 signed long *timeout)
6805 /* make sure we run task_work before checking for signals */
6806 ret = io_run_task_work_sig();
6807 if (ret || io_should_wake(iowq))
6809 /* let the caller flush overflows, retry */
6810 if (test_bit(0, &ctx->cq_check_overflow))
6813 *timeout = schedule_timeout(*timeout);
6814 return !*timeout ? -ETIME : 1;
6818 * Wait until events become available, if we don't already have some. The
6819 * application must reap them itself, as they reside on the shared cq ring.
6821 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6822 const sigset_t __user *sig, size_t sigsz,
6823 struct __kernel_timespec __user *uts)
6825 struct io_wait_queue iowq = {
6828 .func = io_wake_function,
6829 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6832 .to_wait = min_events,
6834 struct io_rings *rings = ctx->rings;
6835 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6839 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6840 if (io_cqring_events(ctx) >= min_events)
6842 if (!io_run_task_work())
6847 #ifdef CONFIG_COMPAT
6848 if (in_compat_syscall())
6849 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6853 ret = set_user_sigmask(sig, sigsz);
6860 struct timespec64 ts;
6862 if (get_timespec64(&ts, uts))
6864 timeout = timespec64_to_jiffies(&ts);
6867 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6868 trace_io_uring_cqring_wait(ctx, min_events);
6870 /* if we can't even flush overflow, don't wait for more */
6871 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6875 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6876 TASK_INTERRUPTIBLE);
6877 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6878 finish_wait(&ctx->wait, &iowq.wq);
6882 restore_saved_sigmask_unless(ret == -EINTR);
6884 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6887 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6889 #if defined(CONFIG_UNIX)
6890 if (ctx->ring_sock) {
6891 struct sock *sock = ctx->ring_sock->sk;
6892 struct sk_buff *skb;
6894 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6900 for (i = 0; i < ctx->nr_user_files; i++) {
6903 file = io_file_from_index(ctx, i);
6910 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6912 struct fixed_rsrc_data *data;
6914 data = container_of(ref, struct fixed_rsrc_data, refs);
6915 complete(&data->done);
6918 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6920 spin_lock_bh(&ctx->rsrc_ref_lock);
6923 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6925 spin_unlock_bh(&ctx->rsrc_ref_lock);
6928 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6929 struct fixed_rsrc_data *rsrc_data,
6930 struct fixed_rsrc_ref_node *ref_node)
6932 io_rsrc_ref_lock(ctx);
6933 rsrc_data->node = ref_node;
6934 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6935 io_rsrc_ref_unlock(ctx);
6936 percpu_ref_get(&rsrc_data->refs);
6939 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6941 struct fixed_rsrc_ref_node *ref_node = NULL;
6943 io_rsrc_ref_lock(ctx);
6944 ref_node = data->node;
6946 io_rsrc_ref_unlock(ctx);
6948 percpu_ref_kill(&ref_node->refs);
6951 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6952 struct io_ring_ctx *ctx,
6953 void (*rsrc_put)(struct io_ring_ctx *ctx,
6954 struct io_rsrc_put *prsrc))
6956 struct fixed_rsrc_ref_node *backup_node;
6962 data->quiesce = true;
6965 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6968 backup_node->rsrc_data = data;
6969 backup_node->rsrc_put = rsrc_put;
6971 io_sqe_rsrc_kill_node(ctx, data);
6972 percpu_ref_kill(&data->refs);
6973 flush_delayed_work(&ctx->rsrc_put_work);
6975 ret = wait_for_completion_interruptible(&data->done);
6979 percpu_ref_resurrect(&data->refs);
6980 io_sqe_rsrc_set_node(ctx, data, backup_node);
6982 reinit_completion(&data->done);
6983 mutex_unlock(&ctx->uring_lock);
6984 ret = io_run_task_work_sig();
6985 mutex_lock(&ctx->uring_lock);
6987 data->quiesce = false;
6990 destroy_fixed_rsrc_ref_node(backup_node);
6994 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
6996 struct fixed_rsrc_data *data;
6998 data = kzalloc(sizeof(*data), GFP_KERNEL);
7002 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7003 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7008 init_completion(&data->done);
7012 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7014 percpu_ref_exit(&data->refs);
7019 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7021 struct fixed_rsrc_data *data = ctx->file_data;
7022 unsigned nr_tables, i;
7026 * percpu_ref_is_dying() is to stop parallel files unregister
7027 * Since we possibly drop uring lock later in this function to
7030 if (!data || percpu_ref_is_dying(&data->refs))
7032 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7036 __io_sqe_files_unregister(ctx);
7037 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7038 for (i = 0; i < nr_tables; i++)
7039 kfree(data->table[i].files);
7040 free_fixed_rsrc_data(data);
7041 ctx->file_data = NULL;
7042 ctx->nr_user_files = 0;
7046 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7047 __releases(&sqd->rw_lock)
7049 if (sqd->thread == current)
7051 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7052 up_write(&sqd->rw_lock);
7055 static void io_sq_thread_park(struct io_sq_data *sqd)
7056 __acquires(&sqd->rw_lock)
7058 if (sqd->thread == current)
7060 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7061 down_write(&sqd->rw_lock);
7062 /* set again for consistency, in case concurrent parks are happening */
7063 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7065 wake_up_process(sqd->thread);
7068 static void io_sq_thread_stop(struct io_sq_data *sqd)
7070 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7072 down_write(&sqd->rw_lock);
7074 up_write(&sqd->rw_lock);
7077 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7078 wake_up_process(sqd->thread);
7079 up_write(&sqd->rw_lock);
7080 wait_for_completion(&sqd->exited);
7083 static void io_put_sq_data(struct io_sq_data *sqd)
7085 if (refcount_dec_and_test(&sqd->refs)) {
7086 io_sq_thread_stop(sqd);
7091 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7093 struct io_sq_data *sqd = ctx->sq_data;
7096 complete(&sqd->startup);
7098 wait_for_completion(&ctx->sq_thread_comp);
7100 io_sq_thread_park(sqd);
7101 list_del(&ctx->sqd_list);
7102 io_sqd_update_thread_idle(sqd);
7103 io_sq_thread_unpark(sqd);
7105 io_put_sq_data(sqd);
7106 ctx->sq_data = NULL;
7108 put_cred(ctx->sq_creds);
7112 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7114 struct io_ring_ctx *ctx_attach;
7115 struct io_sq_data *sqd;
7118 f = fdget(p->wq_fd);
7120 return ERR_PTR(-ENXIO);
7121 if (f.file->f_op != &io_uring_fops) {
7123 return ERR_PTR(-EINVAL);
7126 ctx_attach = f.file->private_data;
7127 sqd = ctx_attach->sq_data;
7130 return ERR_PTR(-EINVAL);
7133 refcount_inc(&sqd->refs);
7138 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7140 struct io_sq_data *sqd;
7142 if (p->flags & IORING_SETUP_ATTACH_WQ)
7143 return io_attach_sq_data(p);
7145 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7147 return ERR_PTR(-ENOMEM);
7149 refcount_set(&sqd->refs, 1);
7150 INIT_LIST_HEAD(&sqd->ctx_list);
7151 INIT_LIST_HEAD(&sqd->ctx_new_list);
7152 init_rwsem(&sqd->rw_lock);
7153 init_waitqueue_head(&sqd->wait);
7154 init_completion(&sqd->startup);
7155 init_completion(&sqd->exited);
7159 #if defined(CONFIG_UNIX)
7161 * Ensure the UNIX gc is aware of our file set, so we are certain that
7162 * the io_uring can be safely unregistered on process exit, even if we have
7163 * loops in the file referencing.
7165 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7167 struct sock *sk = ctx->ring_sock->sk;
7168 struct scm_fp_list *fpl;
7169 struct sk_buff *skb;
7172 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7176 skb = alloc_skb(0, GFP_KERNEL);
7185 fpl->user = get_uid(current_user());
7186 for (i = 0; i < nr; i++) {
7187 struct file *file = io_file_from_index(ctx, i + offset);
7191 fpl->fp[nr_files] = get_file(file);
7192 unix_inflight(fpl->user, fpl->fp[nr_files]);
7197 fpl->max = SCM_MAX_FD;
7198 fpl->count = nr_files;
7199 UNIXCB(skb).fp = fpl;
7200 skb->destructor = unix_destruct_scm;
7201 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7202 skb_queue_head(&sk->sk_receive_queue, skb);
7204 for (i = 0; i < nr_files; i++)
7215 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7216 * causes regular reference counting to break down. We rely on the UNIX
7217 * garbage collection to take care of this problem for us.
7219 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7221 unsigned left, total;
7225 left = ctx->nr_user_files;
7227 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7229 ret = __io_sqe_files_scm(ctx, this_files, total);
7233 total += this_files;
7239 while (total < ctx->nr_user_files) {
7240 struct file *file = io_file_from_index(ctx, total);
7250 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7256 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7257 unsigned nr_tables, unsigned nr_files)
7261 for (i = 0; i < nr_tables; i++) {
7262 struct fixed_rsrc_table *table = &file_data->table[i];
7263 unsigned this_files;
7265 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7266 table->files = kcalloc(this_files, sizeof(struct file *),
7270 nr_files -= this_files;
7276 for (i = 0; i < nr_tables; i++) {
7277 struct fixed_rsrc_table *table = &file_data->table[i];
7278 kfree(table->files);
7283 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7285 struct file *file = prsrc->file;
7286 #if defined(CONFIG_UNIX)
7287 struct sock *sock = ctx->ring_sock->sk;
7288 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7289 struct sk_buff *skb;
7292 __skb_queue_head_init(&list);
7295 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7296 * remove this entry and rearrange the file array.
7298 skb = skb_dequeue(head);
7300 struct scm_fp_list *fp;
7302 fp = UNIXCB(skb).fp;
7303 for (i = 0; i < fp->count; i++) {
7306 if (fp->fp[i] != file)
7309 unix_notinflight(fp->user, fp->fp[i]);
7310 left = fp->count - 1 - i;
7312 memmove(&fp->fp[i], &fp->fp[i + 1],
7313 left * sizeof(struct file *));
7320 __skb_queue_tail(&list, skb);
7330 __skb_queue_tail(&list, skb);
7332 skb = skb_dequeue(head);
7335 if (skb_peek(&list)) {
7336 spin_lock_irq(&head->lock);
7337 while ((skb = __skb_dequeue(&list)) != NULL)
7338 __skb_queue_tail(head, skb);
7339 spin_unlock_irq(&head->lock);
7346 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7348 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7349 struct io_ring_ctx *ctx = rsrc_data->ctx;
7350 struct io_rsrc_put *prsrc, *tmp;
7352 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7353 list_del(&prsrc->list);
7354 ref_node->rsrc_put(ctx, prsrc);
7358 percpu_ref_exit(&ref_node->refs);
7360 percpu_ref_put(&rsrc_data->refs);
7363 static void io_rsrc_put_work(struct work_struct *work)
7365 struct io_ring_ctx *ctx;
7366 struct llist_node *node;
7368 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7369 node = llist_del_all(&ctx->rsrc_put_llist);
7372 struct fixed_rsrc_ref_node *ref_node;
7373 struct llist_node *next = node->next;
7375 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7376 __io_rsrc_put_work(ref_node);
7381 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7384 struct fixed_rsrc_table *table;
7386 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7387 return &table->files[i & IORING_FILE_TABLE_MASK];
7390 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7392 struct fixed_rsrc_ref_node *ref_node;
7393 struct fixed_rsrc_data *data;
7394 struct io_ring_ctx *ctx;
7395 bool first_add = false;
7398 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7399 data = ref_node->rsrc_data;
7402 io_rsrc_ref_lock(ctx);
7403 ref_node->done = true;
7405 while (!list_empty(&ctx->rsrc_ref_list)) {
7406 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7407 struct fixed_rsrc_ref_node, node);
7408 /* recycle ref nodes in order */
7409 if (!ref_node->done)
7411 list_del(&ref_node->node);
7412 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7414 io_rsrc_ref_unlock(ctx);
7416 if (percpu_ref_is_dying(&data->refs))
7420 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7422 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7425 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7426 struct io_ring_ctx *ctx)
7428 struct fixed_rsrc_ref_node *ref_node;
7430 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7434 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7439 INIT_LIST_HEAD(&ref_node->node);
7440 INIT_LIST_HEAD(&ref_node->rsrc_list);
7441 ref_node->done = false;
7445 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7446 struct fixed_rsrc_ref_node *ref_node)
7448 ref_node->rsrc_data = ctx->file_data;
7449 ref_node->rsrc_put = io_ring_file_put;
7452 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7454 percpu_ref_exit(&ref_node->refs);
7459 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7462 __s32 __user *fds = (__s32 __user *) arg;
7463 unsigned nr_tables, i;
7465 int fd, ret = -ENOMEM;
7466 struct fixed_rsrc_ref_node *ref_node;
7467 struct fixed_rsrc_data *file_data;
7473 if (nr_args > IORING_MAX_FIXED_FILES)
7476 file_data = alloc_fixed_rsrc_data(ctx);
7479 ctx->file_data = file_data;
7481 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7482 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7484 if (!file_data->table)
7487 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7490 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7491 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7495 /* allow sparse sets */
7505 * Don't allow io_uring instances to be registered. If UNIX
7506 * isn't enabled, then this causes a reference cycle and this
7507 * instance can never get freed. If UNIX is enabled we'll
7508 * handle it just fine, but there's still no point in allowing
7509 * a ring fd as it doesn't support regular read/write anyway.
7511 if (file->f_op == &io_uring_fops) {
7515 *io_fixed_file_slot(file_data, i) = file;
7518 ret = io_sqe_files_scm(ctx);
7520 io_sqe_files_unregister(ctx);
7524 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7526 io_sqe_files_unregister(ctx);
7529 init_fixed_file_ref_node(ctx, ref_node);
7531 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7534 for (i = 0; i < ctx->nr_user_files; i++) {
7535 file = io_file_from_index(ctx, i);
7539 for (i = 0; i < nr_tables; i++)
7540 kfree(file_data->table[i].files);
7541 ctx->nr_user_files = 0;
7543 free_fixed_rsrc_data(ctx->file_data);
7544 ctx->file_data = NULL;
7548 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7551 #if defined(CONFIG_UNIX)
7552 struct sock *sock = ctx->ring_sock->sk;
7553 struct sk_buff_head *head = &sock->sk_receive_queue;
7554 struct sk_buff *skb;
7557 * See if we can merge this file into an existing skb SCM_RIGHTS
7558 * file set. If there's no room, fall back to allocating a new skb
7559 * and filling it in.
7561 spin_lock_irq(&head->lock);
7562 skb = skb_peek(head);
7564 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7566 if (fpl->count < SCM_MAX_FD) {
7567 __skb_unlink(skb, head);
7568 spin_unlock_irq(&head->lock);
7569 fpl->fp[fpl->count] = get_file(file);
7570 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7572 spin_lock_irq(&head->lock);
7573 __skb_queue_head(head, skb);
7578 spin_unlock_irq(&head->lock);
7585 return __io_sqe_files_scm(ctx, 1, index);
7591 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7593 struct io_rsrc_put *prsrc;
7594 struct fixed_rsrc_ref_node *ref_node = data->node;
7596 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7601 list_add(&prsrc->list, &ref_node->rsrc_list);
7606 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7609 return io_queue_rsrc_removal(data, (void *)file);
7612 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7613 struct io_uring_rsrc_update *up,
7616 struct fixed_rsrc_data *data = ctx->file_data;
7617 struct fixed_rsrc_ref_node *ref_node;
7618 struct file *file, **file_slot;
7622 bool needs_switch = false;
7624 if (check_add_overflow(up->offset, nr_args, &done))
7626 if (done > ctx->nr_user_files)
7629 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7632 init_fixed_file_ref_node(ctx, ref_node);
7634 fds = u64_to_user_ptr(up->data);
7635 for (done = 0; done < nr_args; done++) {
7637 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7641 if (fd == IORING_REGISTER_FILES_SKIP)
7644 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7645 file_slot = io_fixed_file_slot(ctx->file_data, i);
7648 err = io_queue_file_removal(data, *file_slot);
7652 needs_switch = true;
7661 * Don't allow io_uring instances to be registered. If
7662 * UNIX isn't enabled, then this causes a reference
7663 * cycle and this instance can never get freed. If UNIX
7664 * is enabled we'll handle it just fine, but there's
7665 * still no point in allowing a ring fd as it doesn't
7666 * support regular read/write anyway.
7668 if (file->f_op == &io_uring_fops) {
7674 err = io_sqe_file_register(ctx, file, i);
7684 percpu_ref_kill(&data->node->refs);
7685 io_sqe_rsrc_set_node(ctx, data, ref_node);
7687 destroy_fixed_rsrc_ref_node(ref_node);
7689 return done ? done : err;
7692 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7695 struct io_uring_rsrc_update up;
7697 if (!ctx->file_data)
7701 if (copy_from_user(&up, arg, sizeof(up)))
7706 return __io_sqe_files_update(ctx, &up, nr_args);
7709 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7711 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7713 req = io_put_req_find_next(req);
7714 return req ? &req->work : NULL;
7717 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7719 struct io_wq_hash *hash;
7720 struct io_wq_data data;
7721 unsigned int concurrency;
7723 hash = ctx->hash_map;
7725 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7727 return ERR_PTR(-ENOMEM);
7728 refcount_set(&hash->refs, 1);
7729 init_waitqueue_head(&hash->wait);
7730 ctx->hash_map = hash;
7734 data.free_work = io_free_work;
7735 data.do_work = io_wq_submit_work;
7737 /* Do QD, or 4 * CPUS, whatever is smallest */
7738 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7740 return io_wq_create(concurrency, &data);
7743 static int io_uring_alloc_task_context(struct task_struct *task,
7744 struct io_ring_ctx *ctx)
7746 struct io_uring_task *tctx;
7749 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7750 if (unlikely(!tctx))
7753 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7754 if (unlikely(ret)) {
7759 tctx->io_wq = io_init_wq_offload(ctx);
7760 if (IS_ERR(tctx->io_wq)) {
7761 ret = PTR_ERR(tctx->io_wq);
7762 percpu_counter_destroy(&tctx->inflight);
7768 init_waitqueue_head(&tctx->wait);
7770 atomic_set(&tctx->in_idle, 0);
7771 tctx->sqpoll = false;
7772 task->io_uring = tctx;
7773 spin_lock_init(&tctx->task_lock);
7774 INIT_WQ_LIST(&tctx->task_list);
7775 tctx->task_state = 0;
7776 init_task_work(&tctx->task_work, tctx_task_work);
7780 void __io_uring_free(struct task_struct *tsk)
7782 struct io_uring_task *tctx = tsk->io_uring;
7784 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7785 WARN_ON_ONCE(tctx->io_wq);
7787 percpu_counter_destroy(&tctx->inflight);
7789 tsk->io_uring = NULL;
7792 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7793 struct io_uring_params *p)
7797 /* Retain compatibility with failing for an invalid attach attempt */
7798 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7799 IORING_SETUP_ATTACH_WQ) {
7802 f = fdget(p->wq_fd);
7805 if (f.file->f_op != &io_uring_fops) {
7811 if (ctx->flags & IORING_SETUP_SQPOLL) {
7812 struct task_struct *tsk;
7813 struct io_sq_data *sqd;
7816 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7819 sqd = io_get_sq_data(p);
7825 ctx->sq_creds = get_current_cred();
7827 io_sq_thread_park(sqd);
7828 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7829 io_sq_thread_unpark(sqd);
7831 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7832 if (!ctx->sq_thread_idle)
7833 ctx->sq_thread_idle = HZ;
7838 if (p->flags & IORING_SETUP_SQ_AFF) {
7839 int cpu = p->sq_thread_cpu;
7842 if (cpu >= nr_cpu_ids)
7844 if (!cpu_online(cpu))
7852 sqd->task_pid = current->pid;
7853 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7858 ret = io_uring_alloc_task_context(tsk, ctx);
7860 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7862 wake_up_new_task(tsk);
7865 complete(&sqd->startup);
7866 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7867 /* Can't have SQ_AFF without SQPOLL */
7874 io_sq_thread_finish(ctx);
7878 static inline void __io_unaccount_mem(struct user_struct *user,
7879 unsigned long nr_pages)
7881 atomic_long_sub(nr_pages, &user->locked_vm);
7884 static inline int __io_account_mem(struct user_struct *user,
7885 unsigned long nr_pages)
7887 unsigned long page_limit, cur_pages, new_pages;
7889 /* Don't allow more pages than we can safely lock */
7890 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7893 cur_pages = atomic_long_read(&user->locked_vm);
7894 new_pages = cur_pages + nr_pages;
7895 if (new_pages > page_limit)
7897 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7898 new_pages) != cur_pages);
7903 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7906 __io_unaccount_mem(ctx->user, nr_pages);
7908 if (ctx->mm_account)
7909 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7912 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7917 ret = __io_account_mem(ctx->user, nr_pages);
7922 if (ctx->mm_account)
7923 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7928 static void io_mem_free(void *ptr)
7935 page = virt_to_head_page(ptr);
7936 if (put_page_testzero(page))
7937 free_compound_page(page);
7940 static void *io_mem_alloc(size_t size)
7942 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7943 __GFP_NORETRY | __GFP_ACCOUNT;
7945 return (void *) __get_free_pages(gfp_flags, get_order(size));
7948 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7951 struct io_rings *rings;
7952 size_t off, sq_array_size;
7954 off = struct_size(rings, cqes, cq_entries);
7955 if (off == SIZE_MAX)
7959 off = ALIGN(off, SMP_CACHE_BYTES);
7967 sq_array_size = array_size(sizeof(u32), sq_entries);
7968 if (sq_array_size == SIZE_MAX)
7971 if (check_add_overflow(off, sq_array_size, &off))
7977 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
7981 if (!ctx->user_bufs)
7984 for (i = 0; i < ctx->nr_user_bufs; i++) {
7985 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7987 for (j = 0; j < imu->nr_bvecs; j++)
7988 unpin_user_page(imu->bvec[j].bv_page);
7990 if (imu->acct_pages)
7991 io_unaccount_mem(ctx, imu->acct_pages);
7996 kfree(ctx->user_bufs);
7997 ctx->user_bufs = NULL;
7998 ctx->nr_user_bufs = 0;
8002 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8003 void __user *arg, unsigned index)
8005 struct iovec __user *src;
8007 #ifdef CONFIG_COMPAT
8009 struct compat_iovec __user *ciovs;
8010 struct compat_iovec ciov;
8012 ciovs = (struct compat_iovec __user *) arg;
8013 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8016 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8017 dst->iov_len = ciov.iov_len;
8021 src = (struct iovec __user *) arg;
8022 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8028 * Not super efficient, but this is just a registration time. And we do cache
8029 * the last compound head, so generally we'll only do a full search if we don't
8032 * We check if the given compound head page has already been accounted, to
8033 * avoid double accounting it. This allows us to account the full size of the
8034 * page, not just the constituent pages of a huge page.
8036 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8037 int nr_pages, struct page *hpage)
8041 /* check current page array */
8042 for (i = 0; i < nr_pages; i++) {
8043 if (!PageCompound(pages[i]))
8045 if (compound_head(pages[i]) == hpage)
8049 /* check previously registered pages */
8050 for (i = 0; i < ctx->nr_user_bufs; i++) {
8051 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8053 for (j = 0; j < imu->nr_bvecs; j++) {
8054 if (!PageCompound(imu->bvec[j].bv_page))
8056 if (compound_head(imu->bvec[j].bv_page) == hpage)
8064 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8065 int nr_pages, struct io_mapped_ubuf *imu,
8066 struct page **last_hpage)
8070 for (i = 0; i < nr_pages; i++) {
8071 if (!PageCompound(pages[i])) {
8076 hpage = compound_head(pages[i]);
8077 if (hpage == *last_hpage)
8079 *last_hpage = hpage;
8080 if (headpage_already_acct(ctx, pages, i, hpage))
8082 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8086 if (!imu->acct_pages)
8089 ret = io_account_mem(ctx, imu->acct_pages);
8091 imu->acct_pages = 0;
8095 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8096 struct io_mapped_ubuf *imu,
8097 struct page **last_hpage)
8099 struct vm_area_struct **vmas = NULL;
8100 struct page **pages = NULL;
8101 unsigned long off, start, end, ubuf;
8103 int ret, pret, nr_pages, i;
8105 ubuf = (unsigned long) iov->iov_base;
8106 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8107 start = ubuf >> PAGE_SHIFT;
8108 nr_pages = end - start;
8112 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8116 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8121 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8127 mmap_read_lock(current->mm);
8128 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8130 if (pret == nr_pages) {
8131 /* don't support file backed memory */
8132 for (i = 0; i < nr_pages; i++) {
8133 struct vm_area_struct *vma = vmas[i];
8136 !is_file_hugepages(vma->vm_file)) {
8142 ret = pret < 0 ? pret : -EFAULT;
8144 mmap_read_unlock(current->mm);
8147 * if we did partial map, or found file backed vmas,
8148 * release any pages we did get
8151 unpin_user_pages(pages, pret);
8156 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8158 unpin_user_pages(pages, pret);
8163 off = ubuf & ~PAGE_MASK;
8164 size = iov->iov_len;
8165 for (i = 0; i < nr_pages; i++) {
8168 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8169 imu->bvec[i].bv_page = pages[i];
8170 imu->bvec[i].bv_len = vec_len;
8171 imu->bvec[i].bv_offset = off;
8175 /* store original address for later verification */
8177 imu->len = iov->iov_len;
8178 imu->nr_bvecs = nr_pages;
8186 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8190 if (!nr_args || nr_args > UIO_MAXIOV)
8193 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8195 if (!ctx->user_bufs)
8201 static int io_buffer_validate(struct iovec *iov)
8204 * Don't impose further limits on the size and buffer
8205 * constraints here, we'll -EINVAL later when IO is
8206 * submitted if they are wrong.
8208 if (!iov->iov_base || !iov->iov_len)
8211 /* arbitrary limit, but we need something */
8212 if (iov->iov_len > SZ_1G)
8218 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8219 unsigned int nr_args)
8223 struct page *last_hpage = NULL;
8225 ret = io_buffers_map_alloc(ctx, nr_args);
8229 for (i = 0; i < nr_args; i++) {
8230 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8232 ret = io_copy_iov(ctx, &iov, arg, i);
8236 ret = io_buffer_validate(&iov);
8240 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8244 ctx->nr_user_bufs++;
8248 io_sqe_buffers_unregister(ctx);
8253 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8255 __s32 __user *fds = arg;
8261 if (copy_from_user(&fd, fds, sizeof(*fds)))
8264 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8265 if (IS_ERR(ctx->cq_ev_fd)) {
8266 int ret = PTR_ERR(ctx->cq_ev_fd);
8267 ctx->cq_ev_fd = NULL;
8274 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8276 if (ctx->cq_ev_fd) {
8277 eventfd_ctx_put(ctx->cq_ev_fd);
8278 ctx->cq_ev_fd = NULL;
8285 static int __io_destroy_buffers(int id, void *p, void *data)
8287 struct io_ring_ctx *ctx = data;
8288 struct io_buffer *buf = p;
8290 __io_remove_buffers(ctx, buf, id, -1U);
8294 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8296 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8297 idr_destroy(&ctx->io_buffer_idr);
8300 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8302 struct io_kiocb *req, *nxt;
8304 list_for_each_entry_safe(req, nxt, list, compl.list) {
8305 if (tsk && req->task != tsk)
8307 list_del(&req->compl.list);
8308 kmem_cache_free(req_cachep, req);
8312 static void io_req_caches_free(struct io_ring_ctx *ctx)
8314 struct io_submit_state *submit_state = &ctx->submit_state;
8315 struct io_comp_state *cs = &ctx->submit_state.comp;
8317 mutex_lock(&ctx->uring_lock);
8319 if (submit_state->free_reqs) {
8320 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8321 submit_state->reqs);
8322 submit_state->free_reqs = 0;
8325 spin_lock_irq(&ctx->completion_lock);
8326 list_splice_init(&cs->locked_free_list, &cs->free_list);
8327 cs->locked_free_nr = 0;
8328 spin_unlock_irq(&ctx->completion_lock);
8330 io_req_cache_free(&cs->free_list, NULL);
8332 mutex_unlock(&ctx->uring_lock);
8335 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8338 * Some may use context even when all refs and requests have been put,
8339 * and they are free to do so while still holding uring_lock, see
8340 * __io_req_task_submit(). Wait for them to finish.
8342 mutex_lock(&ctx->uring_lock);
8343 mutex_unlock(&ctx->uring_lock);
8345 io_sq_thread_finish(ctx);
8346 io_sqe_buffers_unregister(ctx);
8348 if (ctx->mm_account) {
8349 mmdrop(ctx->mm_account);
8350 ctx->mm_account = NULL;
8353 mutex_lock(&ctx->uring_lock);
8354 io_sqe_files_unregister(ctx);
8355 mutex_unlock(&ctx->uring_lock);
8356 io_eventfd_unregister(ctx);
8357 io_destroy_buffers(ctx);
8358 idr_destroy(&ctx->personality_idr);
8360 #if defined(CONFIG_UNIX)
8361 if (ctx->ring_sock) {
8362 ctx->ring_sock->file = NULL; /* so that iput() is called */
8363 sock_release(ctx->ring_sock);
8367 io_mem_free(ctx->rings);
8368 io_mem_free(ctx->sq_sqes);
8370 percpu_ref_exit(&ctx->refs);
8371 free_uid(ctx->user);
8372 io_req_caches_free(ctx);
8374 io_wq_put_hash(ctx->hash_map);
8375 kfree(ctx->cancel_hash);
8379 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8381 struct io_ring_ctx *ctx = file->private_data;
8384 poll_wait(file, &ctx->cq_wait, wait);
8386 * synchronizes with barrier from wq_has_sleeper call in
8390 if (!io_sqring_full(ctx))
8391 mask |= EPOLLOUT | EPOLLWRNORM;
8394 * Don't flush cqring overflow list here, just do a simple check.
8395 * Otherwise there could possible be ABBA deadlock:
8398 * lock(&ctx->uring_lock);
8400 * lock(&ctx->uring_lock);
8403 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8404 * pushs them to do the flush.
8406 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8407 mask |= EPOLLIN | EPOLLRDNORM;
8412 static int io_uring_fasync(int fd, struct file *file, int on)
8414 struct io_ring_ctx *ctx = file->private_data;
8416 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8419 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8421 const struct cred *creds;
8423 creds = idr_remove(&ctx->personality_idr, id);
8432 static int io_remove_personalities(int id, void *p, void *data)
8434 struct io_ring_ctx *ctx = data;
8436 io_unregister_personality(ctx, id);
8440 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8442 struct callback_head *work, *next;
8443 bool executed = false;
8446 work = xchg(&ctx->exit_task_work, NULL);
8462 struct io_tctx_exit {
8463 struct callback_head task_work;
8464 struct completion completion;
8465 struct io_ring_ctx *ctx;
8468 static void io_tctx_exit_cb(struct callback_head *cb)
8470 struct io_uring_task *tctx = current->io_uring;
8471 struct io_tctx_exit *work;
8473 work = container_of(cb, struct io_tctx_exit, task_work);
8475 * When @in_idle, we're in cancellation and it's racy to remove the
8476 * node. It'll be removed by the end of cancellation, just ignore it.
8478 if (!atomic_read(&tctx->in_idle))
8479 io_uring_del_task_file((unsigned long)work->ctx);
8480 complete(&work->completion);
8483 static void io_ring_exit_work(struct work_struct *work)
8485 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8486 unsigned long timeout = jiffies + HZ * 60 * 5;
8487 struct io_tctx_exit exit;
8488 struct io_tctx_node *node;
8492 * If we're doing polled IO and end up having requests being
8493 * submitted async (out-of-line), then completions can come in while
8494 * we're waiting for refs to drop. We need to reap these manually,
8495 * as nobody else will be looking for them.
8498 io_uring_try_cancel_requests(ctx, NULL, NULL);
8500 WARN_ON_ONCE(time_after(jiffies, timeout));
8501 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8503 mutex_lock(&ctx->uring_lock);
8504 while (!list_empty(&ctx->tctx_list)) {
8505 WARN_ON_ONCE(time_after(jiffies, timeout));
8507 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8510 init_completion(&exit.completion);
8511 init_task_work(&exit.task_work, io_tctx_exit_cb);
8512 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8513 if (WARN_ON_ONCE(ret))
8515 wake_up_process(node->task);
8517 mutex_unlock(&ctx->uring_lock);
8518 wait_for_completion(&exit.completion);
8520 mutex_lock(&ctx->uring_lock);
8522 mutex_unlock(&ctx->uring_lock);
8524 io_ring_ctx_free(ctx);
8527 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8529 mutex_lock(&ctx->uring_lock);
8530 percpu_ref_kill(&ctx->refs);
8531 /* if force is set, the ring is going away. always drop after that */
8532 ctx->cq_overflow_flushed = 1;
8534 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8535 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8536 mutex_unlock(&ctx->uring_lock);
8538 io_kill_timeouts(ctx, NULL, NULL);
8539 io_poll_remove_all(ctx, NULL, NULL);
8541 /* if we failed setting up the ctx, we might not have any rings */
8542 io_iopoll_try_reap_events(ctx);
8544 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8546 * Use system_unbound_wq to avoid spawning tons of event kworkers
8547 * if we're exiting a ton of rings at the same time. It just adds
8548 * noise and overhead, there's no discernable change in runtime
8549 * over using system_wq.
8551 queue_work(system_unbound_wq, &ctx->exit_work);
8554 static int io_uring_release(struct inode *inode, struct file *file)
8556 struct io_ring_ctx *ctx = file->private_data;
8558 file->private_data = NULL;
8559 io_ring_ctx_wait_and_kill(ctx);
8563 struct io_task_cancel {
8564 struct task_struct *task;
8565 struct files_struct *files;
8568 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8570 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8571 struct io_task_cancel *cancel = data;
8574 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8575 unsigned long flags;
8576 struct io_ring_ctx *ctx = req->ctx;
8578 /* protect against races with linked timeouts */
8579 spin_lock_irqsave(&ctx->completion_lock, flags);
8580 ret = io_match_task(req, cancel->task, cancel->files);
8581 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8583 ret = io_match_task(req, cancel->task, cancel->files);
8588 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8589 struct task_struct *task,
8590 struct files_struct *files)
8592 struct io_defer_entry *de = NULL;
8595 spin_lock_irq(&ctx->completion_lock);
8596 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8597 if (io_match_task(de->req, task, files)) {
8598 list_cut_position(&list, &ctx->defer_list, &de->list);
8602 spin_unlock_irq(&ctx->completion_lock);
8604 while (!list_empty(&list)) {
8605 de = list_first_entry(&list, struct io_defer_entry, list);
8606 list_del_init(&de->list);
8607 req_set_fail_links(de->req);
8608 io_put_req(de->req);
8609 io_req_complete(de->req, -ECANCELED);
8614 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8616 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8618 return req->ctx == data;
8621 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8623 struct io_tctx_node *node;
8624 enum io_wq_cancel cret;
8627 mutex_lock(&ctx->uring_lock);
8628 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8629 struct io_uring_task *tctx = node->task->io_uring;
8632 * io_wq will stay alive while we hold uring_lock, because it's
8633 * killed after ctx nodes, which requires to take the lock.
8635 if (!tctx || !tctx->io_wq)
8637 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8638 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8640 mutex_unlock(&ctx->uring_lock);
8645 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8646 struct task_struct *task,
8647 struct files_struct *files)
8649 struct io_task_cancel cancel = { .task = task, .files = files, };
8650 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8653 enum io_wq_cancel cret;
8657 ret |= io_uring_try_cancel_iowq(ctx);
8658 } else if (tctx && tctx->io_wq) {
8660 * Cancels requests of all rings, not only @ctx, but
8661 * it's fine as the task is in exit/exec.
8663 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8665 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8668 /* SQPOLL thread does its own polling */
8669 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8670 while (!list_empty_careful(&ctx->iopoll_list)) {
8671 io_iopoll_try_reap_events(ctx);
8676 ret |= io_poll_remove_all(ctx, task, files);
8677 ret |= io_kill_timeouts(ctx, task, files);
8678 ret |= io_run_task_work();
8679 ret |= io_run_ctx_fallback(ctx);
8680 io_cqring_overflow_flush(ctx, true, task, files);
8687 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8688 struct task_struct *task,
8689 struct files_struct *files)
8691 struct io_kiocb *req;
8694 spin_lock_irq(&ctx->inflight_lock);
8695 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8696 cnt += io_match_task(req, task, files);
8697 spin_unlock_irq(&ctx->inflight_lock);
8701 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8702 struct task_struct *task,
8703 struct files_struct *files)
8705 while (!list_empty_careful(&ctx->inflight_list)) {
8709 inflight = io_uring_count_inflight(ctx, task, files);
8713 io_uring_try_cancel_requests(ctx, task, files);
8716 io_sq_thread_unpark(ctx->sq_data);
8717 prepare_to_wait(&task->io_uring->wait, &wait,
8718 TASK_UNINTERRUPTIBLE);
8719 if (inflight == io_uring_count_inflight(ctx, task, files))
8721 finish_wait(&task->io_uring->wait, &wait);
8723 io_sq_thread_park(ctx->sq_data);
8728 * We need to iteratively cancel requests, in case a request has dependent
8729 * hard links. These persist even for failure of cancelations, hence keep
8730 * looping until none are found.
8732 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8733 struct files_struct *files)
8735 struct task_struct *task = current;
8737 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8738 io_sq_thread_park(ctx->sq_data);
8739 task = ctx->sq_data->thread;
8741 atomic_inc(&task->io_uring->in_idle);
8744 io_cancel_defer_files(ctx, task, files);
8746 io_uring_cancel_files(ctx, task, files);
8748 io_uring_try_cancel_requests(ctx, task, NULL);
8751 atomic_dec(&task->io_uring->in_idle);
8753 io_sq_thread_unpark(ctx->sq_data);
8757 * Note that this task has used io_uring. We use it for cancelation purposes.
8759 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8761 struct io_uring_task *tctx = current->io_uring;
8762 struct io_tctx_node *node;
8765 if (unlikely(!tctx)) {
8766 ret = io_uring_alloc_task_context(current, ctx);
8769 tctx = current->io_uring;
8771 if (tctx->last != ctx) {
8772 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8775 node = kmalloc(sizeof(*node), GFP_KERNEL);
8779 node->task = current;
8781 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8788 mutex_lock(&ctx->uring_lock);
8789 list_add(&node->ctx_node, &ctx->tctx_list);
8790 mutex_unlock(&ctx->uring_lock);
8796 * This is race safe in that the task itself is doing this, hence it
8797 * cannot be going through the exit/cancel paths at the same time.
8798 * This cannot be modified while exit/cancel is running.
8800 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8801 tctx->sqpoll = true;
8807 * Remove this io_uring_file -> task mapping.
8809 static void io_uring_del_task_file(unsigned long index)
8811 struct io_uring_task *tctx = current->io_uring;
8812 struct io_tctx_node *node;
8816 node = xa_erase(&tctx->xa, index);
8820 WARN_ON_ONCE(current != node->task);
8821 WARN_ON_ONCE(list_empty(&node->ctx_node));
8823 mutex_lock(&node->ctx->uring_lock);
8824 list_del(&node->ctx_node);
8825 mutex_unlock(&node->ctx->uring_lock);
8827 if (tctx->last == node->ctx)
8832 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8834 struct io_tctx_node *node;
8835 unsigned long index;
8837 xa_for_each(&tctx->xa, index, node)
8838 io_uring_del_task_file(index);
8840 io_wq_put_and_exit(tctx->io_wq);
8845 void __io_uring_files_cancel(struct files_struct *files)
8847 struct io_uring_task *tctx = current->io_uring;
8848 struct io_tctx_node *node;
8849 unsigned long index;
8851 /* make sure overflow events are dropped */
8852 atomic_inc(&tctx->in_idle);
8853 xa_for_each(&tctx->xa, index, node)
8854 io_uring_cancel_task_requests(node->ctx, files);
8855 atomic_dec(&tctx->in_idle);
8858 io_uring_clean_tctx(tctx);
8861 static s64 tctx_inflight(struct io_uring_task *tctx)
8863 return percpu_counter_sum(&tctx->inflight);
8866 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8868 struct io_sq_data *sqd = ctx->sq_data;
8869 struct io_uring_task *tctx;
8875 io_sq_thread_park(sqd);
8876 if (!sqd->thread || !sqd->thread->io_uring) {
8877 io_sq_thread_unpark(sqd);
8880 tctx = ctx->sq_data->thread->io_uring;
8881 atomic_inc(&tctx->in_idle);
8883 /* read completions before cancelations */
8884 inflight = tctx_inflight(tctx);
8887 io_uring_cancel_task_requests(ctx, NULL);
8889 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8891 * If we've seen completions, retry without waiting. This
8892 * avoids a race where a completion comes in before we did
8893 * prepare_to_wait().
8895 if (inflight == tctx_inflight(tctx))
8897 finish_wait(&tctx->wait, &wait);
8899 atomic_dec(&tctx->in_idle);
8900 io_sq_thread_unpark(sqd);
8904 * Find any io_uring fd that this task has registered or done IO on, and cancel
8907 void __io_uring_task_cancel(void)
8909 struct io_uring_task *tctx = current->io_uring;
8913 /* make sure overflow events are dropped */
8914 atomic_inc(&tctx->in_idle);
8917 struct io_tctx_node *node;
8918 unsigned long index;
8920 xa_for_each(&tctx->xa, index, node)
8921 io_uring_cancel_sqpoll(node->ctx);
8925 /* read completions before cancelations */
8926 inflight = tctx_inflight(tctx);
8929 __io_uring_files_cancel(NULL);
8931 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8934 * If we've seen completions, retry without waiting. This
8935 * avoids a race where a completion comes in before we did
8936 * prepare_to_wait().
8938 if (inflight == tctx_inflight(tctx))
8940 finish_wait(&tctx->wait, &wait);
8943 atomic_dec(&tctx->in_idle);
8945 io_uring_clean_tctx(tctx);
8946 /* all current's requests should be gone, we can kill tctx */
8947 __io_uring_free(current);
8950 static void *io_uring_validate_mmap_request(struct file *file,
8951 loff_t pgoff, size_t sz)
8953 struct io_ring_ctx *ctx = file->private_data;
8954 loff_t offset = pgoff << PAGE_SHIFT;
8959 case IORING_OFF_SQ_RING:
8960 case IORING_OFF_CQ_RING:
8963 case IORING_OFF_SQES:
8967 return ERR_PTR(-EINVAL);
8970 page = virt_to_head_page(ptr);
8971 if (sz > page_size(page))
8972 return ERR_PTR(-EINVAL);
8979 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8981 size_t sz = vma->vm_end - vma->vm_start;
8985 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8987 return PTR_ERR(ptr);
8989 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8990 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8993 #else /* !CONFIG_MMU */
8995 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8997 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9000 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9002 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9005 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9006 unsigned long addr, unsigned long len,
9007 unsigned long pgoff, unsigned long flags)
9011 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9013 return PTR_ERR(ptr);
9015 return (unsigned long) ptr;
9018 #endif /* !CONFIG_MMU */
9020 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9026 if (!io_sqring_full(ctx))
9028 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9030 if (!io_sqring_full(ctx))
9033 } while (!signal_pending(current));
9035 finish_wait(&ctx->sqo_sq_wait, &wait);
9039 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9040 struct __kernel_timespec __user **ts,
9041 const sigset_t __user **sig)
9043 struct io_uring_getevents_arg arg;
9046 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9047 * is just a pointer to the sigset_t.
9049 if (!(flags & IORING_ENTER_EXT_ARG)) {
9050 *sig = (const sigset_t __user *) argp;
9056 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9057 * timespec and sigset_t pointers if good.
9059 if (*argsz != sizeof(arg))
9061 if (copy_from_user(&arg, argp, sizeof(arg)))
9063 *sig = u64_to_user_ptr(arg.sigmask);
9064 *argsz = arg.sigmask_sz;
9065 *ts = u64_to_user_ptr(arg.ts);
9069 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9070 u32, min_complete, u32, flags, const void __user *, argp,
9073 struct io_ring_ctx *ctx;
9080 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9081 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9089 if (f.file->f_op != &io_uring_fops)
9093 ctx = f.file->private_data;
9094 if (!percpu_ref_tryget(&ctx->refs))
9098 if (ctx->flags & IORING_SETUP_R_DISABLED)
9102 * For SQ polling, the thread will do all submissions and completions.
9103 * Just return the requested submit count, and wake the thread if
9107 if (ctx->flags & IORING_SETUP_SQPOLL) {
9108 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9111 if (unlikely(ctx->sq_data->thread == NULL)) {
9114 if (flags & IORING_ENTER_SQ_WAKEUP)
9115 wake_up(&ctx->sq_data->wait);
9116 if (flags & IORING_ENTER_SQ_WAIT) {
9117 ret = io_sqpoll_wait_sq(ctx);
9121 submitted = to_submit;
9122 } else if (to_submit) {
9123 ret = io_uring_add_task_file(ctx);
9126 mutex_lock(&ctx->uring_lock);
9127 submitted = io_submit_sqes(ctx, to_submit);
9128 mutex_unlock(&ctx->uring_lock);
9130 if (submitted != to_submit)
9133 if (flags & IORING_ENTER_GETEVENTS) {
9134 const sigset_t __user *sig;
9135 struct __kernel_timespec __user *ts;
9137 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9141 min_complete = min(min_complete, ctx->cq_entries);
9144 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9145 * space applications don't need to do io completion events
9146 * polling again, they can rely on io_sq_thread to do polling
9147 * work, which can reduce cpu usage and uring_lock contention.
9149 if (ctx->flags & IORING_SETUP_IOPOLL &&
9150 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9151 ret = io_iopoll_check(ctx, min_complete);
9153 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9158 percpu_ref_put(&ctx->refs);
9161 return submitted ? submitted : ret;
9164 #ifdef CONFIG_PROC_FS
9165 static int io_uring_show_cred(int id, void *p, void *data)
9167 const struct cred *cred = p;
9168 struct seq_file *m = data;
9169 struct user_namespace *uns = seq_user_ns(m);
9170 struct group_info *gi;
9175 seq_printf(m, "%5d\n", id);
9176 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9177 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9178 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9179 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9180 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9181 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9182 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9183 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9184 seq_puts(m, "\n\tGroups:\t");
9185 gi = cred->group_info;
9186 for (g = 0; g < gi->ngroups; g++) {
9187 seq_put_decimal_ull(m, g ? " " : "",
9188 from_kgid_munged(uns, gi->gid[g]));
9190 seq_puts(m, "\n\tCapEff:\t");
9191 cap = cred->cap_effective;
9192 CAP_FOR_EACH_U32(__capi)
9193 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9198 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9200 struct io_sq_data *sq = NULL;
9205 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9206 * since fdinfo case grabs it in the opposite direction of normal use
9207 * cases. If we fail to get the lock, we just don't iterate any
9208 * structures that could be going away outside the io_uring mutex.
9210 has_lock = mutex_trylock(&ctx->uring_lock);
9212 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9218 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9219 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9220 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9221 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9222 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9225 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9227 seq_printf(m, "%5u: <none>\n", i);
9229 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9230 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9231 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9233 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9234 (unsigned int) buf->len);
9236 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9237 seq_printf(m, "Personalities:\n");
9238 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9240 seq_printf(m, "PollList:\n");
9241 spin_lock_irq(&ctx->completion_lock);
9242 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9243 struct hlist_head *list = &ctx->cancel_hash[i];
9244 struct io_kiocb *req;
9246 hlist_for_each_entry(req, list, hash_node)
9247 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9248 req->task->task_works != NULL);
9250 spin_unlock_irq(&ctx->completion_lock);
9252 mutex_unlock(&ctx->uring_lock);
9255 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9257 struct io_ring_ctx *ctx = f->private_data;
9259 if (percpu_ref_tryget(&ctx->refs)) {
9260 __io_uring_show_fdinfo(ctx, m);
9261 percpu_ref_put(&ctx->refs);
9266 static const struct file_operations io_uring_fops = {
9267 .release = io_uring_release,
9268 .mmap = io_uring_mmap,
9270 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9271 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9273 .poll = io_uring_poll,
9274 .fasync = io_uring_fasync,
9275 #ifdef CONFIG_PROC_FS
9276 .show_fdinfo = io_uring_show_fdinfo,
9280 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9281 struct io_uring_params *p)
9283 struct io_rings *rings;
9284 size_t size, sq_array_offset;
9286 /* make sure these are sane, as we already accounted them */
9287 ctx->sq_entries = p->sq_entries;
9288 ctx->cq_entries = p->cq_entries;
9290 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9291 if (size == SIZE_MAX)
9294 rings = io_mem_alloc(size);
9299 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9300 rings->sq_ring_mask = p->sq_entries - 1;
9301 rings->cq_ring_mask = p->cq_entries - 1;
9302 rings->sq_ring_entries = p->sq_entries;
9303 rings->cq_ring_entries = p->cq_entries;
9304 ctx->sq_mask = rings->sq_ring_mask;
9305 ctx->cq_mask = rings->cq_ring_mask;
9307 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9308 if (size == SIZE_MAX) {
9309 io_mem_free(ctx->rings);
9314 ctx->sq_sqes = io_mem_alloc(size);
9315 if (!ctx->sq_sqes) {
9316 io_mem_free(ctx->rings);
9324 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9328 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9332 ret = io_uring_add_task_file(ctx);
9337 fd_install(fd, file);
9342 * Allocate an anonymous fd, this is what constitutes the application
9343 * visible backing of an io_uring instance. The application mmaps this
9344 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9345 * we have to tie this fd to a socket for file garbage collection purposes.
9347 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9350 #if defined(CONFIG_UNIX)
9353 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9356 return ERR_PTR(ret);
9359 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9360 O_RDWR | O_CLOEXEC);
9361 #if defined(CONFIG_UNIX)
9363 sock_release(ctx->ring_sock);
9364 ctx->ring_sock = NULL;
9366 ctx->ring_sock->file = file;
9372 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9373 struct io_uring_params __user *params)
9375 struct io_ring_ctx *ctx;
9381 if (entries > IORING_MAX_ENTRIES) {
9382 if (!(p->flags & IORING_SETUP_CLAMP))
9384 entries = IORING_MAX_ENTRIES;
9388 * Use twice as many entries for the CQ ring. It's possible for the
9389 * application to drive a higher depth than the size of the SQ ring,
9390 * since the sqes are only used at submission time. This allows for
9391 * some flexibility in overcommitting a bit. If the application has
9392 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9393 * of CQ ring entries manually.
9395 p->sq_entries = roundup_pow_of_two(entries);
9396 if (p->flags & IORING_SETUP_CQSIZE) {
9398 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9399 * to a power-of-two, if it isn't already. We do NOT impose
9400 * any cq vs sq ring sizing.
9404 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9405 if (!(p->flags & IORING_SETUP_CLAMP))
9407 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9409 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9410 if (p->cq_entries < p->sq_entries)
9413 p->cq_entries = 2 * p->sq_entries;
9416 ctx = io_ring_ctx_alloc(p);
9419 ctx->compat = in_compat_syscall();
9420 if (!capable(CAP_IPC_LOCK))
9421 ctx->user = get_uid(current_user());
9424 * This is just grabbed for accounting purposes. When a process exits,
9425 * the mm is exited and dropped before the files, hence we need to hang
9426 * on to this mm purely for the purposes of being able to unaccount
9427 * memory (locked/pinned vm). It's not used for anything else.
9429 mmgrab(current->mm);
9430 ctx->mm_account = current->mm;
9432 ret = io_allocate_scq_urings(ctx, p);
9436 ret = io_sq_offload_create(ctx, p);
9440 memset(&p->sq_off, 0, sizeof(p->sq_off));
9441 p->sq_off.head = offsetof(struct io_rings, sq.head);
9442 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9443 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9444 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9445 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9446 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9447 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9449 memset(&p->cq_off, 0, sizeof(p->cq_off));
9450 p->cq_off.head = offsetof(struct io_rings, cq.head);
9451 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9452 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9453 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9454 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9455 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9456 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9458 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9459 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9460 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9461 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9462 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9464 if (copy_to_user(params, p, sizeof(*p))) {
9469 file = io_uring_get_file(ctx);
9471 ret = PTR_ERR(file);
9476 * Install ring fd as the very last thing, so we don't risk someone
9477 * having closed it before we finish setup
9479 ret = io_uring_install_fd(ctx, file);
9481 /* fput will clean it up */
9486 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9489 io_ring_ctx_wait_and_kill(ctx);
9494 * Sets up an aio uring context, and returns the fd. Applications asks for a
9495 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9496 * params structure passed in.
9498 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9500 struct io_uring_params p;
9503 if (copy_from_user(&p, params, sizeof(p)))
9505 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9510 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9511 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9512 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9513 IORING_SETUP_R_DISABLED))
9516 return io_uring_create(entries, &p, params);
9519 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9520 struct io_uring_params __user *, params)
9522 return io_uring_setup(entries, params);
9525 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9527 struct io_uring_probe *p;
9531 size = struct_size(p, ops, nr_args);
9532 if (size == SIZE_MAX)
9534 p = kzalloc(size, GFP_KERNEL);
9539 if (copy_from_user(p, arg, size))
9542 if (memchr_inv(p, 0, size))
9545 p->last_op = IORING_OP_LAST - 1;
9546 if (nr_args > IORING_OP_LAST)
9547 nr_args = IORING_OP_LAST;
9549 for (i = 0; i < nr_args; i++) {
9551 if (!io_op_defs[i].not_supported)
9552 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9557 if (copy_to_user(arg, p, size))
9564 static int io_register_personality(struct io_ring_ctx *ctx)
9566 const struct cred *creds;
9569 creds = get_current_cred();
9571 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9572 USHRT_MAX, GFP_KERNEL);
9578 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9579 unsigned int nr_args)
9581 struct io_uring_restriction *res;
9585 /* Restrictions allowed only if rings started disabled */
9586 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9589 /* We allow only a single restrictions registration */
9590 if (ctx->restrictions.registered)
9593 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9596 size = array_size(nr_args, sizeof(*res));
9597 if (size == SIZE_MAX)
9600 res = memdup_user(arg, size);
9602 return PTR_ERR(res);
9606 for (i = 0; i < nr_args; i++) {
9607 switch (res[i].opcode) {
9608 case IORING_RESTRICTION_REGISTER_OP:
9609 if (res[i].register_op >= IORING_REGISTER_LAST) {
9614 __set_bit(res[i].register_op,
9615 ctx->restrictions.register_op);
9617 case IORING_RESTRICTION_SQE_OP:
9618 if (res[i].sqe_op >= IORING_OP_LAST) {
9623 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9625 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9626 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9628 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9629 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9638 /* Reset all restrictions if an error happened */
9640 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9642 ctx->restrictions.registered = true;
9648 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9650 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9653 if (ctx->restrictions.registered)
9654 ctx->restricted = 1;
9656 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9657 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9658 wake_up(&ctx->sq_data->wait);
9662 static bool io_register_op_must_quiesce(int op)
9665 case IORING_UNREGISTER_FILES:
9666 case IORING_REGISTER_FILES_UPDATE:
9667 case IORING_REGISTER_PROBE:
9668 case IORING_REGISTER_PERSONALITY:
9669 case IORING_UNREGISTER_PERSONALITY:
9676 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9677 void __user *arg, unsigned nr_args)
9678 __releases(ctx->uring_lock)
9679 __acquires(ctx->uring_lock)
9684 * We're inside the ring mutex, if the ref is already dying, then
9685 * someone else killed the ctx or is already going through
9686 * io_uring_register().
9688 if (percpu_ref_is_dying(&ctx->refs))
9691 if (io_register_op_must_quiesce(opcode)) {
9692 percpu_ref_kill(&ctx->refs);
9695 * Drop uring mutex before waiting for references to exit. If
9696 * another thread is currently inside io_uring_enter() it might
9697 * need to grab the uring_lock to make progress. If we hold it
9698 * here across the drain wait, then we can deadlock. It's safe
9699 * to drop the mutex here, since no new references will come in
9700 * after we've killed the percpu ref.
9702 mutex_unlock(&ctx->uring_lock);
9704 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9707 ret = io_run_task_work_sig();
9712 mutex_lock(&ctx->uring_lock);
9715 percpu_ref_resurrect(&ctx->refs);
9720 if (ctx->restricted) {
9721 if (opcode >= IORING_REGISTER_LAST) {
9726 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9733 case IORING_REGISTER_BUFFERS:
9734 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9736 case IORING_UNREGISTER_BUFFERS:
9740 ret = io_sqe_buffers_unregister(ctx);
9742 case IORING_REGISTER_FILES:
9743 ret = io_sqe_files_register(ctx, arg, nr_args);
9745 case IORING_UNREGISTER_FILES:
9749 ret = io_sqe_files_unregister(ctx);
9751 case IORING_REGISTER_FILES_UPDATE:
9752 ret = io_sqe_files_update(ctx, arg, nr_args);
9754 case IORING_REGISTER_EVENTFD:
9755 case IORING_REGISTER_EVENTFD_ASYNC:
9759 ret = io_eventfd_register(ctx, arg);
9762 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9763 ctx->eventfd_async = 1;
9765 ctx->eventfd_async = 0;
9767 case IORING_UNREGISTER_EVENTFD:
9771 ret = io_eventfd_unregister(ctx);
9773 case IORING_REGISTER_PROBE:
9775 if (!arg || nr_args > 256)
9777 ret = io_probe(ctx, arg, nr_args);
9779 case IORING_REGISTER_PERSONALITY:
9783 ret = io_register_personality(ctx);
9785 case IORING_UNREGISTER_PERSONALITY:
9789 ret = io_unregister_personality(ctx, nr_args);
9791 case IORING_REGISTER_ENABLE_RINGS:
9795 ret = io_register_enable_rings(ctx);
9797 case IORING_REGISTER_RESTRICTIONS:
9798 ret = io_register_restrictions(ctx, arg, nr_args);
9806 if (io_register_op_must_quiesce(opcode)) {
9807 /* bring the ctx back to life */
9808 percpu_ref_reinit(&ctx->refs);
9810 reinit_completion(&ctx->ref_comp);
9815 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9816 void __user *, arg, unsigned int, nr_args)
9818 struct io_ring_ctx *ctx;
9827 if (f.file->f_op != &io_uring_fops)
9830 ctx = f.file->private_data;
9834 mutex_lock(&ctx->uring_lock);
9835 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9836 mutex_unlock(&ctx->uring_lock);
9837 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9838 ctx->cq_ev_fd != NULL, ret);
9844 static int __init io_uring_init(void)
9846 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9847 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9848 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9851 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9852 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9853 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9854 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9855 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9856 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9857 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9858 BUILD_BUG_SQE_ELEM(8, __u64, off);
9859 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9860 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9861 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9862 BUILD_BUG_SQE_ELEM(24, __u32, len);
9863 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9864 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9865 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9866 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9867 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9868 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9869 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9870 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9871 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9872 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9873 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9874 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9875 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9876 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9877 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9878 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9879 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9880 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9881 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9883 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9884 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9885 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9889 __initcall(io_uring_init);