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;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 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;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct idr io_buffer_idr;
408 struct xarray personalities;
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 bool io_disarm_next(struct io_kiocb *req);
988 static void io_uring_del_task_file(unsigned long index);
989 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
990 struct task_struct *task,
991 struct files_struct *files);
992 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
993 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
994 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
995 struct io_ring_ctx *ctx);
996 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
998 static bool io_rw_reissue(struct io_kiocb *req);
999 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1000 static void io_put_req(struct io_kiocb *req);
1001 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1002 static void io_double_put_req(struct io_kiocb *req);
1003 static void io_dismantle_req(struct io_kiocb *req);
1004 static void io_put_task(struct task_struct *task, int nr);
1005 static void io_queue_next(struct io_kiocb *req);
1006 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1007 static void __io_queue_linked_timeout(struct io_kiocb *req);
1008 static void io_queue_linked_timeout(struct io_kiocb *req);
1009 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1010 struct io_uring_rsrc_update *ip,
1012 static void __io_clean_op(struct io_kiocb *req);
1013 static struct file *io_file_get(struct io_submit_state *state,
1014 struct io_kiocb *req, int fd, bool fixed);
1015 static void __io_queue_sqe(struct io_kiocb *req);
1016 static void io_rsrc_put_work(struct work_struct *work);
1018 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1019 struct iov_iter *iter, bool needs_lock);
1020 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1021 const struct iovec *fast_iov,
1022 struct iov_iter *iter, bool force);
1023 static void io_req_task_queue(struct io_kiocb *req);
1024 static void io_submit_flush_completions(struct io_comp_state *cs,
1025 struct io_ring_ctx *ctx);
1027 static struct kmem_cache *req_cachep;
1029 static const struct file_operations io_uring_fops;
1031 struct sock *io_uring_get_socket(struct file *file)
1033 #if defined(CONFIG_UNIX)
1034 if (file->f_op == &io_uring_fops) {
1035 struct io_ring_ctx *ctx = file->private_data;
1037 return ctx->ring_sock->sk;
1042 EXPORT_SYMBOL(io_uring_get_socket);
1044 #define io_for_each_link(pos, head) \
1045 for (pos = (head); pos; pos = pos->link)
1047 static inline void io_clean_op(struct io_kiocb *req)
1049 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1053 static inline void io_set_resource_node(struct io_kiocb *req)
1055 struct io_ring_ctx *ctx = req->ctx;
1057 if (!req->fixed_rsrc_refs) {
1058 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1059 percpu_ref_get(req->fixed_rsrc_refs);
1063 static bool io_match_task(struct io_kiocb *head,
1064 struct task_struct *task,
1065 struct files_struct *files)
1067 struct io_kiocb *req;
1069 if (task && head->task != task) {
1070 /* in terms of cancelation, always match if req task is dead */
1071 if (head->task->flags & PF_EXITING)
1078 io_for_each_link(req, head) {
1079 if (req->flags & REQ_F_INFLIGHT)
1081 if (req->task->files == files)
1087 static inline void req_set_fail_links(struct io_kiocb *req)
1089 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1090 req->flags |= REQ_F_FAIL_LINK;
1093 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1095 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1097 complete(&ctx->ref_comp);
1100 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1102 return !req->timeout.off;
1105 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1107 struct io_ring_ctx *ctx;
1110 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1115 * Use 5 bits less than the max cq entries, that should give us around
1116 * 32 entries per hash list if totally full and uniformly spread.
1118 hash_bits = ilog2(p->cq_entries);
1122 ctx->cancel_hash_bits = hash_bits;
1123 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1125 if (!ctx->cancel_hash)
1127 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1129 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1130 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1133 ctx->flags = p->flags;
1134 init_waitqueue_head(&ctx->sqo_sq_wait);
1135 INIT_LIST_HEAD(&ctx->sqd_list);
1136 init_waitqueue_head(&ctx->cq_wait);
1137 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1138 init_completion(&ctx->ref_comp);
1139 idr_init(&ctx->io_buffer_idr);
1140 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
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 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);
1536 * If we're the last reference to this request, add to our locked
1539 if (refcount_dec_and_test(&req->refs)) {
1540 struct io_comp_state *cs = &ctx->submit_state.comp;
1542 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1543 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1544 io_disarm_next(req);
1546 io_req_task_queue(req->link);
1550 io_dismantle_req(req);
1551 io_put_task(req->task, 1);
1552 list_add(&req->compl.list, &cs->locked_free_list);
1553 cs->locked_free_nr++;
1556 io_commit_cqring(ctx);
1557 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1558 io_cqring_ev_posted(ctx);
1561 percpu_ref_put(&ctx->refs);
1564 static void io_req_complete_state(struct io_kiocb *req, long res,
1565 unsigned int cflags)
1569 req->compl.cflags = cflags;
1570 req->flags |= REQ_F_COMPLETE_INLINE;
1573 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1574 long res, unsigned cflags)
1576 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1577 io_req_complete_state(req, res, cflags);
1579 io_req_complete_post(req, res, cflags);
1582 static inline void io_req_complete(struct io_kiocb *req, long res)
1584 __io_req_complete(req, 0, res, 0);
1587 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1589 struct io_submit_state *state = &ctx->submit_state;
1590 struct io_comp_state *cs = &state->comp;
1591 struct io_kiocb *req = NULL;
1594 * If we have more than a batch's worth of requests in our IRQ side
1595 * locked cache, grab the lock and move them over to our submission
1598 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1599 spin_lock_irq(&ctx->completion_lock);
1600 list_splice_init(&cs->locked_free_list, &cs->free_list);
1601 cs->locked_free_nr = 0;
1602 spin_unlock_irq(&ctx->completion_lock);
1605 while (!list_empty(&cs->free_list)) {
1606 req = list_first_entry(&cs->free_list, struct io_kiocb,
1608 list_del(&req->compl.list);
1609 state->reqs[state->free_reqs++] = req;
1610 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1617 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1619 struct io_submit_state *state = &ctx->submit_state;
1621 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1623 if (!state->free_reqs) {
1624 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1627 if (io_flush_cached_reqs(ctx))
1630 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1634 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1635 * retry single alloc to be on the safe side.
1637 if (unlikely(ret <= 0)) {
1638 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1639 if (!state->reqs[0])
1643 state->free_reqs = ret;
1647 return state->reqs[state->free_reqs];
1650 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1657 static void io_dismantle_req(struct io_kiocb *req)
1661 if (req->async_data)
1662 kfree(req->async_data);
1664 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1665 if (req->fixed_rsrc_refs)
1666 percpu_ref_put(req->fixed_rsrc_refs);
1667 if (req->work.creds) {
1668 put_cred(req->work.creds);
1669 req->work.creds = NULL;
1672 if (req->flags & REQ_F_INFLIGHT) {
1673 struct io_ring_ctx *ctx = req->ctx;
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->inflight_lock, flags);
1677 list_del(&req->inflight_entry);
1678 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1679 req->flags &= ~REQ_F_INFLIGHT;
1683 /* must to be called somewhat shortly after putting a request */
1684 static inline void io_put_task(struct task_struct *task, int nr)
1686 struct io_uring_task *tctx = task->io_uring;
1688 percpu_counter_sub(&tctx->inflight, nr);
1689 if (unlikely(atomic_read(&tctx->in_idle)))
1690 wake_up(&tctx->wait);
1691 put_task_struct_many(task, nr);
1694 static void __io_free_req(struct io_kiocb *req)
1696 struct io_ring_ctx *ctx = req->ctx;
1698 io_dismantle_req(req);
1699 io_put_task(req->task, 1);
1701 kmem_cache_free(req_cachep, req);
1702 percpu_ref_put(&ctx->refs);
1705 static inline void io_remove_next_linked(struct io_kiocb *req)
1707 struct io_kiocb *nxt = req->link;
1709 req->link = nxt->link;
1713 static bool io_kill_linked_timeout(struct io_kiocb *req)
1714 __must_hold(&req->ctx->completion_lock)
1716 struct io_kiocb *link = req->link;
1717 bool cancelled = false;
1720 * Can happen if a linked timeout fired and link had been like
1721 * req -> link t-out -> link t-out [-> ...]
1723 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1724 struct io_timeout_data *io = link->async_data;
1727 io_remove_next_linked(req);
1728 link->timeout.head = NULL;
1729 ret = hrtimer_try_to_cancel(&io->timer);
1731 io_cqring_fill_event(link, -ECANCELED);
1732 io_put_req_deferred(link, 1);
1736 req->flags &= ~REQ_F_LINK_TIMEOUT;
1740 static void io_fail_links(struct io_kiocb *req)
1741 __must_hold(&req->ctx->completion_lock)
1743 struct io_kiocb *nxt, *link = req->link;
1750 trace_io_uring_fail_link(req, link);
1751 io_cqring_fill_event(link, -ECANCELED);
1752 io_put_req_deferred(link, 2);
1757 static bool io_disarm_next(struct io_kiocb *req)
1758 __must_hold(&req->ctx->completion_lock)
1760 bool posted = false;
1762 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1763 posted = io_kill_linked_timeout(req);
1764 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1765 posted |= (req->link != NULL);
1771 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1773 struct io_kiocb *nxt;
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 (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1782 struct io_ring_ctx *ctx = req->ctx;
1783 unsigned long flags;
1786 spin_lock_irqsave(&ctx->completion_lock, flags);
1787 posted = io_disarm_next(req);
1789 io_commit_cqring(req->ctx);
1790 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1792 io_cqring_ev_posted(ctx);
1799 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1801 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1803 return __io_req_find_next(req);
1806 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1810 if (ctx->submit_state.comp.nr) {
1811 mutex_lock(&ctx->uring_lock);
1812 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1813 mutex_unlock(&ctx->uring_lock);
1815 percpu_ref_put(&ctx->refs);
1818 static bool __tctx_task_work(struct io_uring_task *tctx)
1820 struct io_ring_ctx *ctx = NULL;
1821 struct io_wq_work_list list;
1822 struct io_wq_work_node *node;
1824 if (wq_list_empty(&tctx->task_list))
1827 spin_lock_irq(&tctx->task_lock);
1828 list = tctx->task_list;
1829 INIT_WQ_LIST(&tctx->task_list);
1830 spin_unlock_irq(&tctx->task_lock);
1834 struct io_wq_work_node *next = node->next;
1835 struct io_kiocb *req;
1837 req = container_of(node, struct io_kiocb, io_task_work.node);
1838 if (req->ctx != ctx) {
1839 ctx_flush_and_put(ctx);
1841 percpu_ref_get(&ctx->refs);
1844 req->task_work.func(&req->task_work);
1848 ctx_flush_and_put(ctx);
1849 return list.first != NULL;
1852 static void tctx_task_work(struct callback_head *cb)
1854 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1856 clear_bit(0, &tctx->task_state);
1858 while (__tctx_task_work(tctx))
1862 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1863 enum task_work_notify_mode notify)
1865 struct io_uring_task *tctx = tsk->io_uring;
1866 struct io_wq_work_node *node, *prev;
1867 unsigned long flags;
1870 WARN_ON_ONCE(!tctx);
1872 spin_lock_irqsave(&tctx->task_lock, flags);
1873 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1874 spin_unlock_irqrestore(&tctx->task_lock, flags);
1876 /* task_work already pending, we're done */
1877 if (test_bit(0, &tctx->task_state) ||
1878 test_and_set_bit(0, &tctx->task_state))
1881 if (!task_work_add(tsk, &tctx->task_work, notify))
1885 * Slow path - we failed, find and delete work. if the work is not
1886 * in the list, it got run and we're fine.
1889 spin_lock_irqsave(&tctx->task_lock, flags);
1890 wq_list_for_each(node, prev, &tctx->task_list) {
1891 if (&req->io_task_work.node == node) {
1892 wq_list_del(&tctx->task_list, node, prev);
1897 spin_unlock_irqrestore(&tctx->task_lock, flags);
1898 clear_bit(0, &tctx->task_state);
1902 static int io_req_task_work_add(struct io_kiocb *req)
1904 struct task_struct *tsk = req->task;
1905 struct io_ring_ctx *ctx = req->ctx;
1906 enum task_work_notify_mode notify;
1909 if (tsk->flags & PF_EXITING)
1913 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1914 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1915 * processing task_work. There's no reliable way to tell if TWA_RESUME
1919 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1920 notify = TWA_SIGNAL;
1922 ret = io_task_work_add(tsk, req, notify);
1924 wake_up_process(tsk);
1929 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1930 task_work_func_t cb)
1932 struct io_ring_ctx *ctx = req->ctx;
1933 struct callback_head *head;
1935 init_task_work(&req->task_work, cb);
1937 head = READ_ONCE(ctx->exit_task_work);
1938 req->task_work.next = head;
1939 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1942 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1944 struct io_ring_ctx *ctx = req->ctx;
1946 spin_lock_irq(&ctx->completion_lock);
1947 io_cqring_fill_event(req, error);
1948 io_commit_cqring(ctx);
1949 spin_unlock_irq(&ctx->completion_lock);
1951 io_cqring_ev_posted(ctx);
1952 req_set_fail_links(req);
1953 io_double_put_req(req);
1956 static void io_req_task_cancel(struct callback_head *cb)
1958 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1959 struct io_ring_ctx *ctx = req->ctx;
1961 mutex_lock(&ctx->uring_lock);
1962 __io_req_task_cancel(req, req->result);
1963 mutex_unlock(&ctx->uring_lock);
1964 percpu_ref_put(&ctx->refs);
1967 static void __io_req_task_submit(struct io_kiocb *req)
1969 struct io_ring_ctx *ctx = req->ctx;
1971 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1972 mutex_lock(&ctx->uring_lock);
1973 if (!(current->flags & PF_EXITING) && !current->in_execve)
1974 __io_queue_sqe(req);
1976 __io_req_task_cancel(req, -EFAULT);
1977 mutex_unlock(&ctx->uring_lock);
1980 static void io_req_task_submit(struct callback_head *cb)
1982 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1984 __io_req_task_submit(req);
1987 static void io_req_task_queue(struct io_kiocb *req)
1991 req->task_work.func = io_req_task_submit;
1992 ret = io_req_task_work_add(req);
1993 if (unlikely(ret)) {
1994 req->result = -ECANCELED;
1995 percpu_ref_get(&req->ctx->refs);
1996 io_req_task_work_add_fallback(req, io_req_task_cancel);
2000 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2002 percpu_ref_get(&req->ctx->refs);
2004 req->task_work.func = io_req_task_cancel;
2006 if (unlikely(io_req_task_work_add(req)))
2007 io_req_task_work_add_fallback(req, io_req_task_cancel);
2010 static inline void io_queue_next(struct io_kiocb *req)
2012 struct io_kiocb *nxt = io_req_find_next(req);
2015 io_req_task_queue(nxt);
2018 static void io_free_req(struct io_kiocb *req)
2025 struct task_struct *task;
2030 static inline void io_init_req_batch(struct req_batch *rb)
2037 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2038 struct req_batch *rb)
2041 io_put_task(rb->task, rb->task_refs);
2043 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2046 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2047 struct io_submit_state *state)
2051 if (req->task != rb->task) {
2053 io_put_task(rb->task, rb->task_refs);
2054 rb->task = req->task;
2060 io_dismantle_req(req);
2061 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2062 state->reqs[state->free_reqs++] = req;
2064 list_add(&req->compl.list, &state->comp.free_list);
2067 static void io_submit_flush_completions(struct io_comp_state *cs,
2068 struct io_ring_ctx *ctx)
2071 struct io_kiocb *req;
2072 struct req_batch rb;
2074 io_init_req_batch(&rb);
2075 spin_lock_irq(&ctx->completion_lock);
2076 for (i = 0; i < nr; i++) {
2078 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2080 io_commit_cqring(ctx);
2081 spin_unlock_irq(&ctx->completion_lock);
2083 io_cqring_ev_posted(ctx);
2084 for (i = 0; i < nr; i++) {
2087 /* submission and completion refs */
2088 if (refcount_sub_and_test(2, &req->refs))
2089 io_req_free_batch(&rb, req, &ctx->submit_state);
2092 io_req_free_batch_finish(ctx, &rb);
2097 * Drop reference to request, return next in chain (if there is one) if this
2098 * was the last reference to this request.
2100 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2102 struct io_kiocb *nxt = NULL;
2104 if (refcount_dec_and_test(&req->refs)) {
2105 nxt = io_req_find_next(req);
2111 static void io_put_req(struct io_kiocb *req)
2113 if (refcount_dec_and_test(&req->refs))
2117 static void io_put_req_deferred_cb(struct callback_head *cb)
2119 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2124 static void io_free_req_deferred(struct io_kiocb *req)
2128 req->task_work.func = io_put_req_deferred_cb;
2129 ret = io_req_task_work_add(req);
2131 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2134 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2136 if (refcount_sub_and_test(refs, &req->refs))
2137 io_free_req_deferred(req);
2140 static void io_double_put_req(struct io_kiocb *req)
2142 /* drop both submit and complete references */
2143 if (refcount_sub_and_test(2, &req->refs))
2147 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2149 /* See comment at the top of this file */
2151 return __io_cqring_events(ctx);
2154 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2156 struct io_rings *rings = ctx->rings;
2158 /* make sure SQ entry isn't read before tail */
2159 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2162 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2164 unsigned int cflags;
2166 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2167 cflags |= IORING_CQE_F_BUFFER;
2168 req->flags &= ~REQ_F_BUFFER_SELECTED;
2173 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2175 struct io_buffer *kbuf;
2177 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2178 return io_put_kbuf(req, kbuf);
2181 static inline bool io_run_task_work(void)
2184 * Not safe to run on exiting task, and the task_work handling will
2185 * not add work to such a task.
2187 if (unlikely(current->flags & PF_EXITING))
2189 if (current->task_works) {
2190 __set_current_state(TASK_RUNNING);
2199 * Find and free completed poll iocbs
2201 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2202 struct list_head *done)
2204 struct req_batch rb;
2205 struct io_kiocb *req;
2207 /* order with ->result store in io_complete_rw_iopoll() */
2210 io_init_req_batch(&rb);
2211 while (!list_empty(done)) {
2214 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2215 list_del(&req->inflight_entry);
2217 if (READ_ONCE(req->result) == -EAGAIN) {
2218 req->iopoll_completed = 0;
2219 if (io_rw_reissue(req))
2223 if (req->flags & REQ_F_BUFFER_SELECTED)
2224 cflags = io_put_rw_kbuf(req);
2226 __io_cqring_fill_event(req, req->result, cflags);
2229 if (refcount_dec_and_test(&req->refs))
2230 io_req_free_batch(&rb, req, &ctx->submit_state);
2233 io_commit_cqring(ctx);
2234 io_cqring_ev_posted_iopoll(ctx);
2235 io_req_free_batch_finish(ctx, &rb);
2238 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2241 struct io_kiocb *req, *tmp;
2247 * Only spin for completions if we don't have multiple devices hanging
2248 * off our complete list, and we're under the requested amount.
2250 spin = !ctx->poll_multi_file && *nr_events < min;
2253 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2254 struct kiocb *kiocb = &req->rw.kiocb;
2257 * Move completed and retryable entries to our local lists.
2258 * If we find a request that requires polling, break out
2259 * and complete those lists first, if we have entries there.
2261 if (READ_ONCE(req->iopoll_completed)) {
2262 list_move_tail(&req->inflight_entry, &done);
2265 if (!list_empty(&done))
2268 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2272 /* iopoll may have completed current req */
2273 if (READ_ONCE(req->iopoll_completed))
2274 list_move_tail(&req->inflight_entry, &done);
2281 if (!list_empty(&done))
2282 io_iopoll_complete(ctx, nr_events, &done);
2288 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2289 * non-spinning poll check - we'll still enter the driver poll loop, but only
2290 * as a non-spinning completion check.
2292 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2295 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2298 ret = io_do_iopoll(ctx, nr_events, min);
2301 if (*nr_events >= min)
2309 * We can't just wait for polled events to come to us, we have to actively
2310 * find and complete them.
2312 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2314 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2317 mutex_lock(&ctx->uring_lock);
2318 while (!list_empty(&ctx->iopoll_list)) {
2319 unsigned int nr_events = 0;
2321 io_do_iopoll(ctx, &nr_events, 0);
2323 /* let it sleep and repeat later if can't complete a request */
2327 * Ensure we allow local-to-the-cpu processing to take place,
2328 * in this case we need to ensure that we reap all events.
2329 * Also let task_work, etc. to progress by releasing the mutex
2331 if (need_resched()) {
2332 mutex_unlock(&ctx->uring_lock);
2334 mutex_lock(&ctx->uring_lock);
2337 mutex_unlock(&ctx->uring_lock);
2340 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2342 unsigned int nr_events = 0;
2343 int iters = 0, ret = 0;
2346 * We disallow the app entering submit/complete with polling, but we
2347 * still need to lock the ring to prevent racing with polled issue
2348 * that got punted to a workqueue.
2350 mutex_lock(&ctx->uring_lock);
2353 * Don't enter poll loop if we already have events pending.
2354 * If we do, we can potentially be spinning for commands that
2355 * already triggered a CQE (eg in error).
2357 if (test_bit(0, &ctx->cq_check_overflow))
2358 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2359 if (io_cqring_events(ctx))
2363 * If a submit got punted to a workqueue, we can have the
2364 * application entering polling for a command before it gets
2365 * issued. That app will hold the uring_lock for the duration
2366 * of the poll right here, so we need to take a breather every
2367 * now and then to ensure that the issue has a chance to add
2368 * the poll to the issued list. Otherwise we can spin here
2369 * forever, while the workqueue is stuck trying to acquire the
2372 if (!(++iters & 7)) {
2373 mutex_unlock(&ctx->uring_lock);
2375 mutex_lock(&ctx->uring_lock);
2378 ret = io_iopoll_getevents(ctx, &nr_events, min);
2382 } while (min && !nr_events && !need_resched());
2384 mutex_unlock(&ctx->uring_lock);
2388 static void kiocb_end_write(struct io_kiocb *req)
2391 * Tell lockdep we inherited freeze protection from submission
2394 if (req->flags & REQ_F_ISREG) {
2395 struct inode *inode = file_inode(req->file);
2397 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2399 file_end_write(req->file);
2403 static bool io_resubmit_prep(struct io_kiocb *req)
2405 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2407 struct iov_iter iter;
2409 /* already prepared */
2410 if (req->async_data)
2413 switch (req->opcode) {
2414 case IORING_OP_READV:
2415 case IORING_OP_READ_FIXED:
2416 case IORING_OP_READ:
2419 case IORING_OP_WRITEV:
2420 case IORING_OP_WRITE_FIXED:
2421 case IORING_OP_WRITE:
2425 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2430 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2433 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2436 static bool io_rw_should_reissue(struct io_kiocb *req)
2438 umode_t mode = file_inode(req->file)->i_mode;
2439 struct io_ring_ctx *ctx = req->ctx;
2441 if (!S_ISBLK(mode) && !S_ISREG(mode))
2443 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2444 !(ctx->flags & IORING_SETUP_IOPOLL)))
2447 * If ref is dying, we might be running poll reap from the exit work.
2448 * Don't attempt to reissue from that path, just let it fail with
2451 if (percpu_ref_is_dying(&ctx->refs))
2457 static bool io_rw_reissue(struct io_kiocb *req)
2460 if (!io_rw_should_reissue(req))
2463 lockdep_assert_held(&req->ctx->uring_lock);
2465 if (io_resubmit_prep(req)) {
2466 refcount_inc(&req->refs);
2467 io_queue_async_work(req);
2470 req_set_fail_links(req);
2475 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2476 unsigned int issue_flags)
2480 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2482 if (res != req->result)
2483 req_set_fail_links(req);
2485 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2486 kiocb_end_write(req);
2487 if (req->flags & REQ_F_BUFFER_SELECTED)
2488 cflags = io_put_rw_kbuf(req);
2489 __io_req_complete(req, issue_flags, res, cflags);
2492 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2494 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2496 __io_complete_rw(req, res, res2, 0);
2499 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2504 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2505 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2506 struct io_async_rw *rw = req->async_data;
2509 iov_iter_revert(&rw->iter,
2510 req->result - iov_iter_count(&rw->iter));
2511 else if (!io_resubmit_prep(req))
2516 if (kiocb->ki_flags & IOCB_WRITE)
2517 kiocb_end_write(req);
2519 if (res != -EAGAIN && res != req->result)
2520 req_set_fail_links(req);
2522 WRITE_ONCE(req->result, res);
2523 /* order with io_poll_complete() checking ->result */
2525 WRITE_ONCE(req->iopoll_completed, 1);
2529 * After the iocb has been issued, it's safe to be found on the poll list.
2530 * Adding the kiocb to the list AFTER submission ensures that we don't
2531 * find it from a io_iopoll_getevents() thread before the issuer is done
2532 * accessing the kiocb cookie.
2534 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2536 struct io_ring_ctx *ctx = req->ctx;
2539 * Track whether we have multiple files in our lists. This will impact
2540 * how we do polling eventually, not spinning if we're on potentially
2541 * different devices.
2543 if (list_empty(&ctx->iopoll_list)) {
2544 ctx->poll_multi_file = false;
2545 } else if (!ctx->poll_multi_file) {
2546 struct io_kiocb *list_req;
2548 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2550 if (list_req->file != req->file)
2551 ctx->poll_multi_file = true;
2555 * For fast devices, IO may have already completed. If it has, add
2556 * it to the front so we find it first.
2558 if (READ_ONCE(req->iopoll_completed))
2559 list_add(&req->inflight_entry, &ctx->iopoll_list);
2561 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2564 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2565 * task context or in io worker task context. If current task context is
2566 * sq thread, we don't need to check whether should wake up sq thread.
2568 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2569 wq_has_sleeper(&ctx->sq_data->wait))
2570 wake_up(&ctx->sq_data->wait);
2573 static inline void io_state_file_put(struct io_submit_state *state)
2575 if (state->file_refs) {
2576 fput_many(state->file, state->file_refs);
2577 state->file_refs = 0;
2582 * Get as many references to a file as we have IOs left in this submission,
2583 * assuming most submissions are for one file, or at least that each file
2584 * has more than one submission.
2586 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2591 if (state->file_refs) {
2592 if (state->fd == fd) {
2596 io_state_file_put(state);
2598 state->file = fget_many(fd, state->ios_left);
2599 if (unlikely(!state->file))
2603 state->file_refs = state->ios_left - 1;
2607 static bool io_bdev_nowait(struct block_device *bdev)
2609 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2613 * If we tracked the file through the SCM inflight mechanism, we could support
2614 * any file. For now, just ensure that anything potentially problematic is done
2617 static bool io_file_supports_async(struct file *file, int rw)
2619 umode_t mode = file_inode(file)->i_mode;
2621 if (S_ISBLK(mode)) {
2622 if (IS_ENABLED(CONFIG_BLOCK) &&
2623 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2627 if (S_ISCHR(mode) || S_ISSOCK(mode))
2629 if (S_ISREG(mode)) {
2630 if (IS_ENABLED(CONFIG_BLOCK) &&
2631 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2632 file->f_op != &io_uring_fops)
2637 /* any ->read/write should understand O_NONBLOCK */
2638 if (file->f_flags & O_NONBLOCK)
2641 if (!(file->f_mode & FMODE_NOWAIT))
2645 return file->f_op->read_iter != NULL;
2647 return file->f_op->write_iter != NULL;
2650 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2652 struct io_ring_ctx *ctx = req->ctx;
2653 struct kiocb *kiocb = &req->rw.kiocb;
2654 struct file *file = req->file;
2658 if (S_ISREG(file_inode(file)->i_mode))
2659 req->flags |= REQ_F_ISREG;
2661 kiocb->ki_pos = READ_ONCE(sqe->off);
2662 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2663 req->flags |= REQ_F_CUR_POS;
2664 kiocb->ki_pos = file->f_pos;
2666 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2667 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2668 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2672 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2673 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2674 req->flags |= REQ_F_NOWAIT;
2676 ioprio = READ_ONCE(sqe->ioprio);
2678 ret = ioprio_check_cap(ioprio);
2682 kiocb->ki_ioprio = ioprio;
2684 kiocb->ki_ioprio = get_current_ioprio();
2686 if (ctx->flags & IORING_SETUP_IOPOLL) {
2687 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2688 !kiocb->ki_filp->f_op->iopoll)
2691 kiocb->ki_flags |= IOCB_HIPRI;
2692 kiocb->ki_complete = io_complete_rw_iopoll;
2693 req->iopoll_completed = 0;
2695 if (kiocb->ki_flags & IOCB_HIPRI)
2697 kiocb->ki_complete = io_complete_rw;
2700 req->rw.addr = READ_ONCE(sqe->addr);
2701 req->rw.len = READ_ONCE(sqe->len);
2702 req->buf_index = READ_ONCE(sqe->buf_index);
2706 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2712 case -ERESTARTNOINTR:
2713 case -ERESTARTNOHAND:
2714 case -ERESTART_RESTARTBLOCK:
2716 * We can't just restart the syscall, since previously
2717 * submitted sqes may already be in progress. Just fail this
2723 kiocb->ki_complete(kiocb, ret, 0);
2727 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2728 unsigned int issue_flags)
2730 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2731 struct io_async_rw *io = req->async_data;
2733 /* add previously done IO, if any */
2734 if (io && io->bytes_done > 0) {
2736 ret = io->bytes_done;
2738 ret += io->bytes_done;
2741 if (req->flags & REQ_F_CUR_POS)
2742 req->file->f_pos = kiocb->ki_pos;
2743 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2744 __io_complete_rw(req, ret, 0, issue_flags);
2746 io_rw_done(kiocb, ret);
2749 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2751 struct io_ring_ctx *ctx = req->ctx;
2752 size_t len = req->rw.len;
2753 struct io_mapped_ubuf *imu;
2754 u16 index, buf_index = req->buf_index;
2758 if (unlikely(buf_index >= ctx->nr_user_bufs))
2760 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2761 imu = &ctx->user_bufs[index];
2762 buf_addr = req->rw.addr;
2765 if (buf_addr + len < buf_addr)
2767 /* not inside the mapped region */
2768 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2772 * May not be a start of buffer, set size appropriately
2773 * and advance us to the beginning.
2775 offset = buf_addr - imu->ubuf;
2776 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2780 * Don't use iov_iter_advance() here, as it's really slow for
2781 * using the latter parts of a big fixed buffer - it iterates
2782 * over each segment manually. We can cheat a bit here, because
2785 * 1) it's a BVEC iter, we set it up
2786 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2787 * first and last bvec
2789 * So just find our index, and adjust the iterator afterwards.
2790 * If the offset is within the first bvec (or the whole first
2791 * bvec, just use iov_iter_advance(). This makes it easier
2792 * since we can just skip the first segment, which may not
2793 * be PAGE_SIZE aligned.
2795 const struct bio_vec *bvec = imu->bvec;
2797 if (offset <= bvec->bv_len) {
2798 iov_iter_advance(iter, offset);
2800 unsigned long seg_skip;
2802 /* skip first vec */
2803 offset -= bvec->bv_len;
2804 seg_skip = 1 + (offset >> PAGE_SHIFT);
2806 iter->bvec = bvec + seg_skip;
2807 iter->nr_segs -= seg_skip;
2808 iter->count -= bvec->bv_len + offset;
2809 iter->iov_offset = offset & ~PAGE_MASK;
2816 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2819 mutex_unlock(&ctx->uring_lock);
2822 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2825 * "Normal" inline submissions always hold the uring_lock, since we
2826 * grab it from the system call. Same is true for the SQPOLL offload.
2827 * The only exception is when we've detached the request and issue it
2828 * from an async worker thread, grab the lock for that case.
2831 mutex_lock(&ctx->uring_lock);
2834 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2835 int bgid, struct io_buffer *kbuf,
2838 struct io_buffer *head;
2840 if (req->flags & REQ_F_BUFFER_SELECTED)
2843 io_ring_submit_lock(req->ctx, needs_lock);
2845 lockdep_assert_held(&req->ctx->uring_lock);
2847 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2849 if (!list_empty(&head->list)) {
2850 kbuf = list_last_entry(&head->list, struct io_buffer,
2852 list_del(&kbuf->list);
2855 idr_remove(&req->ctx->io_buffer_idr, bgid);
2857 if (*len > kbuf->len)
2860 kbuf = ERR_PTR(-ENOBUFS);
2863 io_ring_submit_unlock(req->ctx, needs_lock);
2868 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2871 struct io_buffer *kbuf;
2874 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2875 bgid = req->buf_index;
2876 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2879 req->rw.addr = (u64) (unsigned long) kbuf;
2880 req->flags |= REQ_F_BUFFER_SELECTED;
2881 return u64_to_user_ptr(kbuf->addr);
2884 #ifdef CONFIG_COMPAT
2885 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2888 struct compat_iovec __user *uiov;
2889 compat_ssize_t clen;
2893 uiov = u64_to_user_ptr(req->rw.addr);
2894 if (!access_ok(uiov, sizeof(*uiov)))
2896 if (__get_user(clen, &uiov->iov_len))
2902 buf = io_rw_buffer_select(req, &len, needs_lock);
2904 return PTR_ERR(buf);
2905 iov[0].iov_base = buf;
2906 iov[0].iov_len = (compat_size_t) len;
2911 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2914 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2918 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2921 len = iov[0].iov_len;
2924 buf = io_rw_buffer_select(req, &len, needs_lock);
2926 return PTR_ERR(buf);
2927 iov[0].iov_base = buf;
2928 iov[0].iov_len = len;
2932 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2935 if (req->flags & REQ_F_BUFFER_SELECTED) {
2936 struct io_buffer *kbuf;
2938 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2939 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2940 iov[0].iov_len = kbuf->len;
2943 if (req->rw.len != 1)
2946 #ifdef CONFIG_COMPAT
2947 if (req->ctx->compat)
2948 return io_compat_import(req, iov, needs_lock);
2951 return __io_iov_buffer_select(req, iov, needs_lock);
2954 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2955 struct iov_iter *iter, bool needs_lock)
2957 void __user *buf = u64_to_user_ptr(req->rw.addr);
2958 size_t sqe_len = req->rw.len;
2959 u8 opcode = req->opcode;
2962 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2964 return io_import_fixed(req, rw, iter);
2967 /* buffer index only valid with fixed read/write, or buffer select */
2968 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2971 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2972 if (req->flags & REQ_F_BUFFER_SELECT) {
2973 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2975 return PTR_ERR(buf);
2976 req->rw.len = sqe_len;
2979 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2984 if (req->flags & REQ_F_BUFFER_SELECT) {
2985 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2987 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2992 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2996 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2998 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3002 * For files that don't have ->read_iter() and ->write_iter(), handle them
3003 * by looping over ->read() or ->write() manually.
3005 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3007 struct kiocb *kiocb = &req->rw.kiocb;
3008 struct file *file = req->file;
3012 * Don't support polled IO through this interface, and we can't
3013 * support non-blocking either. For the latter, this just causes
3014 * the kiocb to be handled from an async context.
3016 if (kiocb->ki_flags & IOCB_HIPRI)
3018 if (kiocb->ki_flags & IOCB_NOWAIT)
3021 while (iov_iter_count(iter)) {
3025 if (!iov_iter_is_bvec(iter)) {
3026 iovec = iov_iter_iovec(iter);
3028 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3029 iovec.iov_len = req->rw.len;
3033 nr = file->f_op->read(file, iovec.iov_base,
3034 iovec.iov_len, io_kiocb_ppos(kiocb));
3036 nr = file->f_op->write(file, iovec.iov_base,
3037 iovec.iov_len, io_kiocb_ppos(kiocb));
3046 if (nr != iovec.iov_len)
3050 iov_iter_advance(iter, nr);
3056 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3057 const struct iovec *fast_iov, struct iov_iter *iter)
3059 struct io_async_rw *rw = req->async_data;
3061 memcpy(&rw->iter, iter, sizeof(*iter));
3062 rw->free_iovec = iovec;
3064 /* can only be fixed buffers, no need to do anything */
3065 if (iov_iter_is_bvec(iter))
3068 unsigned iov_off = 0;
3070 rw->iter.iov = rw->fast_iov;
3071 if (iter->iov != fast_iov) {
3072 iov_off = iter->iov - fast_iov;
3073 rw->iter.iov += iov_off;
3075 if (rw->fast_iov != fast_iov)
3076 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3077 sizeof(struct iovec) * iter->nr_segs);
3079 req->flags |= REQ_F_NEED_CLEANUP;
3083 static inline int __io_alloc_async_data(struct io_kiocb *req)
3085 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3086 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3087 return req->async_data == NULL;
3090 static int io_alloc_async_data(struct io_kiocb *req)
3092 if (!io_op_defs[req->opcode].needs_async_data)
3095 return __io_alloc_async_data(req);
3098 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3099 const struct iovec *fast_iov,
3100 struct iov_iter *iter, bool force)
3102 if (!force && !io_op_defs[req->opcode].needs_async_data)
3104 if (!req->async_data) {
3105 if (__io_alloc_async_data(req)) {
3110 io_req_map_rw(req, iovec, fast_iov, iter);
3115 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3117 struct io_async_rw *iorw = req->async_data;
3118 struct iovec *iov = iorw->fast_iov;
3121 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3122 if (unlikely(ret < 0))
3125 iorw->bytes_done = 0;
3126 iorw->free_iovec = iov;
3128 req->flags |= REQ_F_NEED_CLEANUP;
3132 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3134 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3136 return io_prep_rw(req, sqe);
3140 * This is our waitqueue callback handler, registered through lock_page_async()
3141 * when we initially tried to do the IO with the iocb armed our waitqueue.
3142 * This gets called when the page is unlocked, and we generally expect that to
3143 * happen when the page IO is completed and the page is now uptodate. This will
3144 * queue a task_work based retry of the operation, attempting to copy the data
3145 * again. If the latter fails because the page was NOT uptodate, then we will
3146 * do a thread based blocking retry of the operation. That's the unexpected
3149 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3150 int sync, void *arg)
3152 struct wait_page_queue *wpq;
3153 struct io_kiocb *req = wait->private;
3154 struct wait_page_key *key = arg;
3156 wpq = container_of(wait, struct wait_page_queue, wait);
3158 if (!wake_page_match(wpq, key))
3161 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3162 list_del_init(&wait->entry);
3164 /* submit ref gets dropped, acquire a new one */
3165 refcount_inc(&req->refs);
3166 io_req_task_queue(req);
3171 * This controls whether a given IO request should be armed for async page
3172 * based retry. If we return false here, the request is handed to the async
3173 * worker threads for retry. If we're doing buffered reads on a regular file,
3174 * we prepare a private wait_page_queue entry and retry the operation. This
3175 * will either succeed because the page is now uptodate and unlocked, or it
3176 * will register a callback when the page is unlocked at IO completion. Through
3177 * that callback, io_uring uses task_work to setup a retry of the operation.
3178 * That retry will attempt the buffered read again. The retry will generally
3179 * succeed, or in rare cases where it fails, we then fall back to using the
3180 * async worker threads for a blocking retry.
3182 static bool io_rw_should_retry(struct io_kiocb *req)
3184 struct io_async_rw *rw = req->async_data;
3185 struct wait_page_queue *wait = &rw->wpq;
3186 struct kiocb *kiocb = &req->rw.kiocb;
3188 /* never retry for NOWAIT, we just complete with -EAGAIN */
3189 if (req->flags & REQ_F_NOWAIT)
3192 /* Only for buffered IO */
3193 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3197 * just use poll if we can, and don't attempt if the fs doesn't
3198 * support callback based unlocks
3200 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3203 wait->wait.func = io_async_buf_func;
3204 wait->wait.private = req;
3205 wait->wait.flags = 0;
3206 INIT_LIST_HEAD(&wait->wait.entry);
3207 kiocb->ki_flags |= IOCB_WAITQ;
3208 kiocb->ki_flags &= ~IOCB_NOWAIT;
3209 kiocb->ki_waitq = wait;
3213 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3215 if (req->file->f_op->read_iter)
3216 return call_read_iter(req->file, &req->rw.kiocb, iter);
3217 else if (req->file->f_op->read)
3218 return loop_rw_iter(READ, req, iter);
3223 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3225 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3226 struct kiocb *kiocb = &req->rw.kiocb;
3227 struct iov_iter __iter, *iter = &__iter;
3228 struct io_async_rw *rw = req->async_data;
3229 ssize_t io_size, ret, ret2;
3230 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3236 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3240 io_size = iov_iter_count(iter);
3241 req->result = io_size;
3243 /* Ensure we clear previously set non-block flag */
3244 if (!force_nonblock)
3245 kiocb->ki_flags &= ~IOCB_NOWAIT;
3247 kiocb->ki_flags |= IOCB_NOWAIT;
3249 /* If the file doesn't support async, just async punt */
3250 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3251 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3252 return ret ?: -EAGAIN;
3255 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3256 if (unlikely(ret)) {
3261 ret = io_iter_do_read(req, iter);
3263 if (ret == -EIOCBQUEUED) {
3264 if (req->async_data)
3265 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3267 } else if (ret == -EAGAIN) {
3268 /* IOPOLL retry should happen for io-wq threads */
3269 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3271 /* no retry on NONBLOCK nor RWF_NOWAIT */
3272 if (req->flags & REQ_F_NOWAIT)
3274 /* some cases will consume bytes even on error returns */
3275 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3277 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3278 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3279 /* read all, failed, already did sync or don't want to retry */
3283 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3288 rw = req->async_data;
3289 /* now use our persistent iterator, if we aren't already */
3294 rw->bytes_done += ret;
3295 /* if we can retry, do so with the callbacks armed */
3296 if (!io_rw_should_retry(req)) {
3297 kiocb->ki_flags &= ~IOCB_WAITQ;
3302 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3303 * we get -EIOCBQUEUED, then we'll get a notification when the
3304 * desired page gets unlocked. We can also get a partial read
3305 * here, and if we do, then just retry at the new offset.
3307 ret = io_iter_do_read(req, iter);
3308 if (ret == -EIOCBQUEUED)
3310 /* we got some bytes, but not all. retry. */
3311 kiocb->ki_flags &= ~IOCB_WAITQ;
3312 } while (ret > 0 && ret < io_size);
3314 kiocb_done(kiocb, ret, issue_flags);
3316 /* it's faster to check here then delegate to kfree */
3322 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3324 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3326 return io_prep_rw(req, sqe);
3329 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3331 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3332 struct kiocb *kiocb = &req->rw.kiocb;
3333 struct iov_iter __iter, *iter = &__iter;
3334 struct io_async_rw *rw = req->async_data;
3335 ssize_t ret, ret2, io_size;
3336 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3342 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3346 io_size = iov_iter_count(iter);
3347 req->result = io_size;
3349 /* Ensure we clear previously set non-block flag */
3350 if (!force_nonblock)
3351 kiocb->ki_flags &= ~IOCB_NOWAIT;
3353 kiocb->ki_flags |= IOCB_NOWAIT;
3355 /* If the file doesn't support async, just async punt */
3356 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3359 /* file path doesn't support NOWAIT for non-direct_IO */
3360 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3361 (req->flags & REQ_F_ISREG))
3364 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3369 * Open-code file_start_write here to grab freeze protection,
3370 * which will be released by another thread in
3371 * io_complete_rw(). Fool lockdep by telling it the lock got
3372 * released so that it doesn't complain about the held lock when
3373 * we return to userspace.
3375 if (req->flags & REQ_F_ISREG) {
3376 sb_start_write(file_inode(req->file)->i_sb);
3377 __sb_writers_release(file_inode(req->file)->i_sb,
3380 kiocb->ki_flags |= IOCB_WRITE;
3382 if (req->file->f_op->write_iter)
3383 ret2 = call_write_iter(req->file, kiocb, iter);
3384 else if (req->file->f_op->write)
3385 ret2 = loop_rw_iter(WRITE, req, iter);
3390 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3391 * retry them without IOCB_NOWAIT.
3393 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3395 /* no retry on NONBLOCK nor RWF_NOWAIT */
3396 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3398 if (ret2 == -EIOCBQUEUED && req->async_data)
3399 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3400 if (!force_nonblock || ret2 != -EAGAIN) {
3401 /* IOPOLL retry should happen for io-wq threads */
3402 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3405 kiocb_done(kiocb, ret2, issue_flags);
3408 /* some cases will consume bytes even on error returns */
3409 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3410 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3411 return ret ?: -EAGAIN;
3414 /* it's reportedly faster than delegating the null check to kfree() */
3420 static int io_renameat_prep(struct io_kiocb *req,
3421 const struct io_uring_sqe *sqe)
3423 struct io_rename *ren = &req->rename;
3424 const char __user *oldf, *newf;
3426 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3429 ren->old_dfd = READ_ONCE(sqe->fd);
3430 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3431 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3432 ren->new_dfd = READ_ONCE(sqe->len);
3433 ren->flags = READ_ONCE(sqe->rename_flags);
3435 ren->oldpath = getname(oldf);
3436 if (IS_ERR(ren->oldpath))
3437 return PTR_ERR(ren->oldpath);
3439 ren->newpath = getname(newf);
3440 if (IS_ERR(ren->newpath)) {
3441 putname(ren->oldpath);
3442 return PTR_ERR(ren->newpath);
3445 req->flags |= REQ_F_NEED_CLEANUP;
3449 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3451 struct io_rename *ren = &req->rename;
3454 if (issue_flags & IO_URING_F_NONBLOCK)
3457 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3458 ren->newpath, ren->flags);
3460 req->flags &= ~REQ_F_NEED_CLEANUP;
3462 req_set_fail_links(req);
3463 io_req_complete(req, ret);
3467 static int io_unlinkat_prep(struct io_kiocb *req,
3468 const struct io_uring_sqe *sqe)
3470 struct io_unlink *un = &req->unlink;
3471 const char __user *fname;
3473 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3476 un->dfd = READ_ONCE(sqe->fd);
3478 un->flags = READ_ONCE(sqe->unlink_flags);
3479 if (un->flags & ~AT_REMOVEDIR)
3482 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3483 un->filename = getname(fname);
3484 if (IS_ERR(un->filename))
3485 return PTR_ERR(un->filename);
3487 req->flags |= REQ_F_NEED_CLEANUP;
3491 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3493 struct io_unlink *un = &req->unlink;
3496 if (issue_flags & IO_URING_F_NONBLOCK)
3499 if (un->flags & AT_REMOVEDIR)
3500 ret = do_rmdir(un->dfd, un->filename);
3502 ret = do_unlinkat(un->dfd, un->filename);
3504 req->flags &= ~REQ_F_NEED_CLEANUP;
3506 req_set_fail_links(req);
3507 io_req_complete(req, ret);
3511 static int io_shutdown_prep(struct io_kiocb *req,
3512 const struct io_uring_sqe *sqe)
3514 #if defined(CONFIG_NET)
3515 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3517 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3521 req->shutdown.how = READ_ONCE(sqe->len);
3528 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3530 #if defined(CONFIG_NET)
3531 struct socket *sock;
3534 if (issue_flags & IO_URING_F_NONBLOCK)
3537 sock = sock_from_file(req->file);
3538 if (unlikely(!sock))
3541 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3543 req_set_fail_links(req);
3544 io_req_complete(req, ret);
3551 static int __io_splice_prep(struct io_kiocb *req,
3552 const struct io_uring_sqe *sqe)
3554 struct io_splice* sp = &req->splice;
3555 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3557 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3561 sp->len = READ_ONCE(sqe->len);
3562 sp->flags = READ_ONCE(sqe->splice_flags);
3564 if (unlikely(sp->flags & ~valid_flags))
3567 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3568 (sp->flags & SPLICE_F_FD_IN_FIXED));
3571 req->flags |= REQ_F_NEED_CLEANUP;
3573 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3575 * Splice operation will be punted aync, and here need to
3576 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3578 req->work.flags |= IO_WQ_WORK_UNBOUND;
3584 static int io_tee_prep(struct io_kiocb *req,
3585 const struct io_uring_sqe *sqe)
3587 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3589 return __io_splice_prep(req, sqe);
3592 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3594 struct io_splice *sp = &req->splice;
3595 struct file *in = sp->file_in;
3596 struct file *out = sp->file_out;
3597 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3600 if (issue_flags & IO_URING_F_NONBLOCK)
3603 ret = do_tee(in, out, sp->len, flags);
3605 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3606 req->flags &= ~REQ_F_NEED_CLEANUP;
3609 req_set_fail_links(req);
3610 io_req_complete(req, ret);
3614 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3616 struct io_splice* sp = &req->splice;
3618 sp->off_in = READ_ONCE(sqe->splice_off_in);
3619 sp->off_out = READ_ONCE(sqe->off);
3620 return __io_splice_prep(req, sqe);
3623 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3625 struct io_splice *sp = &req->splice;
3626 struct file *in = sp->file_in;
3627 struct file *out = sp->file_out;
3628 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3629 loff_t *poff_in, *poff_out;
3632 if (issue_flags & IO_URING_F_NONBLOCK)
3635 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3636 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3639 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3641 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3642 req->flags &= ~REQ_F_NEED_CLEANUP;
3645 req_set_fail_links(req);
3646 io_req_complete(req, ret);
3651 * IORING_OP_NOP just posts a completion event, nothing else.
3653 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3655 struct io_ring_ctx *ctx = req->ctx;
3657 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3660 __io_req_complete(req, issue_flags, 0, 0);
3664 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3666 struct io_ring_ctx *ctx = req->ctx;
3671 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3673 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3676 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3677 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3680 req->sync.off = READ_ONCE(sqe->off);
3681 req->sync.len = READ_ONCE(sqe->len);
3685 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3687 loff_t end = req->sync.off + req->sync.len;
3690 /* fsync always requires a blocking context */
3691 if (issue_flags & IO_URING_F_NONBLOCK)
3694 ret = vfs_fsync_range(req->file, req->sync.off,
3695 end > 0 ? end : LLONG_MAX,
3696 req->sync.flags & IORING_FSYNC_DATASYNC);
3698 req_set_fail_links(req);
3699 io_req_complete(req, ret);
3703 static int io_fallocate_prep(struct io_kiocb *req,
3704 const struct io_uring_sqe *sqe)
3706 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3708 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3711 req->sync.off = READ_ONCE(sqe->off);
3712 req->sync.len = READ_ONCE(sqe->addr);
3713 req->sync.mode = READ_ONCE(sqe->len);
3717 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3721 /* fallocate always requiring blocking context */
3722 if (issue_flags & IO_URING_F_NONBLOCK)
3724 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3727 req_set_fail_links(req);
3728 io_req_complete(req, ret);
3732 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3734 const char __user *fname;
3737 if (unlikely(sqe->ioprio || sqe->buf_index))
3739 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3742 /* open.how should be already initialised */
3743 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3744 req->open.how.flags |= O_LARGEFILE;
3746 req->open.dfd = READ_ONCE(sqe->fd);
3747 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3748 req->open.filename = getname(fname);
3749 if (IS_ERR(req->open.filename)) {
3750 ret = PTR_ERR(req->open.filename);
3751 req->open.filename = NULL;
3754 req->open.nofile = rlimit(RLIMIT_NOFILE);
3755 req->flags |= REQ_F_NEED_CLEANUP;
3759 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3765 mode = READ_ONCE(sqe->len);
3766 flags = READ_ONCE(sqe->open_flags);
3767 req->open.how = build_open_how(flags, mode);
3768 return __io_openat_prep(req, sqe);
3771 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3773 struct open_how __user *how;
3777 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3779 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3780 len = READ_ONCE(sqe->len);
3781 if (len < OPEN_HOW_SIZE_VER0)
3784 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3789 return __io_openat_prep(req, sqe);
3792 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3794 struct open_flags op;
3797 bool resolve_nonblock;
3800 ret = build_open_flags(&req->open.how, &op);
3803 nonblock_set = op.open_flag & O_NONBLOCK;
3804 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3805 if (issue_flags & IO_URING_F_NONBLOCK) {
3807 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3808 * it'll always -EAGAIN
3810 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3812 op.lookup_flags |= LOOKUP_CACHED;
3813 op.open_flag |= O_NONBLOCK;
3816 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3820 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3821 /* only retry if RESOLVE_CACHED wasn't already set by application */
3822 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3823 file == ERR_PTR(-EAGAIN)) {
3825 * We could hang on to this 'fd', but seems like marginal
3826 * gain for something that is now known to be a slower path.
3827 * So just put it, and we'll get a new one when we retry.
3835 ret = PTR_ERR(file);
3837 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3838 file->f_flags &= ~O_NONBLOCK;
3839 fsnotify_open(file);
3840 fd_install(ret, file);
3843 putname(req->open.filename);
3844 req->flags &= ~REQ_F_NEED_CLEANUP;
3846 req_set_fail_links(req);
3847 io_req_complete(req, ret);
3851 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3853 return io_openat2(req, issue_flags);
3856 static int io_remove_buffers_prep(struct io_kiocb *req,
3857 const struct io_uring_sqe *sqe)
3859 struct io_provide_buf *p = &req->pbuf;
3862 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3865 tmp = READ_ONCE(sqe->fd);
3866 if (!tmp || tmp > USHRT_MAX)
3869 memset(p, 0, sizeof(*p));
3871 p->bgid = READ_ONCE(sqe->buf_group);
3875 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3876 int bgid, unsigned nbufs)
3880 /* shouldn't happen */
3884 /* the head kbuf is the list itself */
3885 while (!list_empty(&buf->list)) {
3886 struct io_buffer *nxt;
3888 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3889 list_del(&nxt->list);
3896 idr_remove(&ctx->io_buffer_idr, bgid);
3901 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3903 struct io_provide_buf *p = &req->pbuf;
3904 struct io_ring_ctx *ctx = req->ctx;
3905 struct io_buffer *head;
3907 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3909 io_ring_submit_lock(ctx, !force_nonblock);
3911 lockdep_assert_held(&ctx->uring_lock);
3914 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3916 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3918 req_set_fail_links(req);
3920 /* need to hold the lock to complete IOPOLL requests */
3921 if (ctx->flags & IORING_SETUP_IOPOLL) {
3922 __io_req_complete(req, issue_flags, ret, 0);
3923 io_ring_submit_unlock(ctx, !force_nonblock);
3925 io_ring_submit_unlock(ctx, !force_nonblock);
3926 __io_req_complete(req, issue_flags, ret, 0);
3931 static int io_provide_buffers_prep(struct io_kiocb *req,
3932 const struct io_uring_sqe *sqe)
3934 struct io_provide_buf *p = &req->pbuf;
3937 if (sqe->ioprio || sqe->rw_flags)
3940 tmp = READ_ONCE(sqe->fd);
3941 if (!tmp || tmp > USHRT_MAX)
3944 p->addr = READ_ONCE(sqe->addr);
3945 p->len = READ_ONCE(sqe->len);
3947 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3950 p->bgid = READ_ONCE(sqe->buf_group);
3951 tmp = READ_ONCE(sqe->off);
3952 if (tmp > USHRT_MAX)
3958 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3960 struct io_buffer *buf;
3961 u64 addr = pbuf->addr;
3962 int i, bid = pbuf->bid;
3964 for (i = 0; i < pbuf->nbufs; i++) {
3965 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3970 buf->len = pbuf->len;
3975 INIT_LIST_HEAD(&buf->list);
3978 list_add_tail(&buf->list, &(*head)->list);
3982 return i ? i : -ENOMEM;
3985 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3987 struct io_provide_buf *p = &req->pbuf;
3988 struct io_ring_ctx *ctx = req->ctx;
3989 struct io_buffer *head, *list;
3991 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3993 io_ring_submit_lock(ctx, !force_nonblock);
3995 lockdep_assert_held(&ctx->uring_lock);
3997 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3999 ret = io_add_buffers(p, &head);
4004 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4007 __io_remove_buffers(ctx, head, p->bgid, -1U);
4013 req_set_fail_links(req);
4015 /* need to hold the lock to complete IOPOLL requests */
4016 if (ctx->flags & IORING_SETUP_IOPOLL) {
4017 __io_req_complete(req, issue_flags, ret, 0);
4018 io_ring_submit_unlock(ctx, !force_nonblock);
4020 io_ring_submit_unlock(ctx, !force_nonblock);
4021 __io_req_complete(req, issue_flags, ret, 0);
4026 static int io_epoll_ctl_prep(struct io_kiocb *req,
4027 const struct io_uring_sqe *sqe)
4029 #if defined(CONFIG_EPOLL)
4030 if (sqe->ioprio || sqe->buf_index)
4032 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4035 req->epoll.epfd = READ_ONCE(sqe->fd);
4036 req->epoll.op = READ_ONCE(sqe->len);
4037 req->epoll.fd = READ_ONCE(sqe->off);
4039 if (ep_op_has_event(req->epoll.op)) {
4040 struct epoll_event __user *ev;
4042 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4043 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4053 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4055 #if defined(CONFIG_EPOLL)
4056 struct io_epoll *ie = &req->epoll;
4058 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4060 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4061 if (force_nonblock && ret == -EAGAIN)
4065 req_set_fail_links(req);
4066 __io_req_complete(req, issue_flags, ret, 0);
4073 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4075 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4076 if (sqe->ioprio || sqe->buf_index || sqe->off)
4078 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4081 req->madvise.addr = READ_ONCE(sqe->addr);
4082 req->madvise.len = READ_ONCE(sqe->len);
4083 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4090 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4092 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4093 struct io_madvise *ma = &req->madvise;
4096 if (issue_flags & IO_URING_F_NONBLOCK)
4099 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4101 req_set_fail_links(req);
4102 io_req_complete(req, ret);
4109 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4111 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4113 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4116 req->fadvise.offset = READ_ONCE(sqe->off);
4117 req->fadvise.len = READ_ONCE(sqe->len);
4118 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4122 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4124 struct io_fadvise *fa = &req->fadvise;
4127 if (issue_flags & IO_URING_F_NONBLOCK) {
4128 switch (fa->advice) {
4129 case POSIX_FADV_NORMAL:
4130 case POSIX_FADV_RANDOM:
4131 case POSIX_FADV_SEQUENTIAL:
4138 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4140 req_set_fail_links(req);
4141 io_req_complete(req, ret);
4145 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4147 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4149 if (sqe->ioprio || sqe->buf_index)
4151 if (req->flags & REQ_F_FIXED_FILE)
4154 req->statx.dfd = READ_ONCE(sqe->fd);
4155 req->statx.mask = READ_ONCE(sqe->len);
4156 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4157 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4158 req->statx.flags = READ_ONCE(sqe->statx_flags);
4163 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4165 struct io_statx *ctx = &req->statx;
4168 if (issue_flags & IO_URING_F_NONBLOCK) {
4169 /* only need file table for an actual valid fd */
4170 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4171 req->flags |= REQ_F_NO_FILE_TABLE;
4175 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4179 req_set_fail_links(req);
4180 io_req_complete(req, ret);
4184 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4186 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4188 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4189 sqe->rw_flags || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->close.fd = READ_ONCE(sqe->fd);
4198 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4200 struct files_struct *files = current->files;
4201 struct io_close *close = &req->close;
4202 struct fdtable *fdt;
4208 spin_lock(&files->file_lock);
4209 fdt = files_fdtable(files);
4210 if (close->fd >= fdt->max_fds) {
4211 spin_unlock(&files->file_lock);
4214 file = fdt->fd[close->fd];
4216 spin_unlock(&files->file_lock);
4220 if (file->f_op == &io_uring_fops) {
4221 spin_unlock(&files->file_lock);
4226 /* if the file has a flush method, be safe and punt to async */
4227 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4228 spin_unlock(&files->file_lock);
4232 ret = __close_fd_get_file(close->fd, &file);
4233 spin_unlock(&files->file_lock);
4240 /* No ->flush() or already async, safely close from here */
4241 ret = filp_close(file, current->files);
4244 req_set_fail_links(req);
4247 __io_req_complete(req, issue_flags, ret, 0);
4251 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4253 struct io_ring_ctx *ctx = req->ctx;
4255 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4257 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4260 req->sync.off = READ_ONCE(sqe->off);
4261 req->sync.len = READ_ONCE(sqe->len);
4262 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4266 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4270 /* sync_file_range always requires a blocking context */
4271 if (issue_flags & IO_URING_F_NONBLOCK)
4274 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4277 req_set_fail_links(req);
4278 io_req_complete(req, ret);
4282 #if defined(CONFIG_NET)
4283 static int io_setup_async_msg(struct io_kiocb *req,
4284 struct io_async_msghdr *kmsg)
4286 struct io_async_msghdr *async_msg = req->async_data;
4290 if (io_alloc_async_data(req)) {
4291 kfree(kmsg->free_iov);
4294 async_msg = req->async_data;
4295 req->flags |= REQ_F_NEED_CLEANUP;
4296 memcpy(async_msg, kmsg, sizeof(*kmsg));
4297 async_msg->msg.msg_name = &async_msg->addr;
4298 /* if were using fast_iov, set it to the new one */
4299 if (!async_msg->free_iov)
4300 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4305 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4306 struct io_async_msghdr *iomsg)
4308 iomsg->msg.msg_name = &iomsg->addr;
4309 iomsg->free_iov = iomsg->fast_iov;
4310 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4311 req->sr_msg.msg_flags, &iomsg->free_iov);
4314 static int io_sendmsg_prep_async(struct io_kiocb *req)
4318 if (!io_op_defs[req->opcode].needs_async_data)
4320 ret = io_sendmsg_copy_hdr(req, req->async_data);
4322 req->flags |= REQ_F_NEED_CLEANUP;
4326 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4328 struct io_sr_msg *sr = &req->sr_msg;
4330 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4333 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4334 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4335 sr->len = READ_ONCE(sqe->len);
4337 #ifdef CONFIG_COMPAT
4338 if (req->ctx->compat)
4339 sr->msg_flags |= MSG_CMSG_COMPAT;
4344 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4346 struct io_async_msghdr iomsg, *kmsg;
4347 struct socket *sock;
4351 sock = sock_from_file(req->file);
4352 if (unlikely(!sock))
4355 kmsg = req->async_data;
4357 ret = io_sendmsg_copy_hdr(req, &iomsg);
4363 flags = req->sr_msg.msg_flags;
4364 if (flags & MSG_DONTWAIT)
4365 req->flags |= REQ_F_NOWAIT;
4366 else if (issue_flags & IO_URING_F_NONBLOCK)
4367 flags |= MSG_DONTWAIT;
4369 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4370 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4371 return io_setup_async_msg(req, kmsg);
4372 if (ret == -ERESTARTSYS)
4375 /* fast path, check for non-NULL to avoid function call */
4377 kfree(kmsg->free_iov);
4378 req->flags &= ~REQ_F_NEED_CLEANUP;
4380 req_set_fail_links(req);
4381 __io_req_complete(req, issue_flags, ret, 0);
4385 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4387 struct io_sr_msg *sr = &req->sr_msg;
4390 struct socket *sock;
4394 sock = sock_from_file(req->file);
4395 if (unlikely(!sock))
4398 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4402 msg.msg_name = NULL;
4403 msg.msg_control = NULL;
4404 msg.msg_controllen = 0;
4405 msg.msg_namelen = 0;
4407 flags = req->sr_msg.msg_flags;
4408 if (flags & MSG_DONTWAIT)
4409 req->flags |= REQ_F_NOWAIT;
4410 else if (issue_flags & IO_URING_F_NONBLOCK)
4411 flags |= MSG_DONTWAIT;
4413 msg.msg_flags = flags;
4414 ret = sock_sendmsg(sock, &msg);
4415 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4417 if (ret == -ERESTARTSYS)
4421 req_set_fail_links(req);
4422 __io_req_complete(req, issue_flags, ret, 0);
4426 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4427 struct io_async_msghdr *iomsg)
4429 struct io_sr_msg *sr = &req->sr_msg;
4430 struct iovec __user *uiov;
4434 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4435 &iomsg->uaddr, &uiov, &iov_len);
4439 if (req->flags & REQ_F_BUFFER_SELECT) {
4442 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4444 sr->len = iomsg->fast_iov[0].iov_len;
4445 iomsg->free_iov = NULL;
4447 iomsg->free_iov = iomsg->fast_iov;
4448 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4449 &iomsg->free_iov, &iomsg->msg.msg_iter,
4458 #ifdef CONFIG_COMPAT
4459 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4460 struct io_async_msghdr *iomsg)
4462 struct compat_msghdr __user *msg_compat;
4463 struct io_sr_msg *sr = &req->sr_msg;
4464 struct compat_iovec __user *uiov;
4469 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4470 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4475 uiov = compat_ptr(ptr);
4476 if (req->flags & REQ_F_BUFFER_SELECT) {
4477 compat_ssize_t clen;
4481 if (!access_ok(uiov, sizeof(*uiov)))
4483 if (__get_user(clen, &uiov->iov_len))
4488 iomsg->free_iov = NULL;
4490 iomsg->free_iov = iomsg->fast_iov;
4491 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4492 UIO_FASTIOV, &iomsg->free_iov,
4493 &iomsg->msg.msg_iter, true);
4502 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4503 struct io_async_msghdr *iomsg)
4505 iomsg->msg.msg_name = &iomsg->addr;
4507 #ifdef CONFIG_COMPAT
4508 if (req->ctx->compat)
4509 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4512 return __io_recvmsg_copy_hdr(req, iomsg);
4515 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4518 struct io_sr_msg *sr = &req->sr_msg;
4519 struct io_buffer *kbuf;
4521 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4526 req->flags |= REQ_F_BUFFER_SELECTED;
4530 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4532 return io_put_kbuf(req, req->sr_msg.kbuf);
4535 static int io_recvmsg_prep_async(struct io_kiocb *req)
4539 if (!io_op_defs[req->opcode].needs_async_data)
4541 ret = io_recvmsg_copy_hdr(req, req->async_data);
4543 req->flags |= REQ_F_NEED_CLEANUP;
4547 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 struct io_sr_msg *sr = &req->sr_msg;
4551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4554 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4555 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4556 sr->len = READ_ONCE(sqe->len);
4557 sr->bgid = READ_ONCE(sqe->buf_group);
4559 #ifdef CONFIG_COMPAT
4560 if (req->ctx->compat)
4561 sr->msg_flags |= MSG_CMSG_COMPAT;
4566 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4568 struct io_async_msghdr iomsg, *kmsg;
4569 struct socket *sock;
4570 struct io_buffer *kbuf;
4572 int ret, cflags = 0;
4573 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4575 sock = sock_from_file(req->file);
4576 if (unlikely(!sock))
4579 kmsg = req->async_data;
4581 ret = io_recvmsg_copy_hdr(req, &iomsg);
4587 if (req->flags & REQ_F_BUFFER_SELECT) {
4588 kbuf = io_recv_buffer_select(req, !force_nonblock);
4590 return PTR_ERR(kbuf);
4591 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4592 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4593 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4594 1, req->sr_msg.len);
4597 flags = req->sr_msg.msg_flags;
4598 if (flags & MSG_DONTWAIT)
4599 req->flags |= REQ_F_NOWAIT;
4600 else if (force_nonblock)
4601 flags |= MSG_DONTWAIT;
4603 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4604 kmsg->uaddr, flags);
4605 if (force_nonblock && ret == -EAGAIN)
4606 return io_setup_async_msg(req, kmsg);
4607 if (ret == -ERESTARTSYS)
4610 if (req->flags & REQ_F_BUFFER_SELECTED)
4611 cflags = io_put_recv_kbuf(req);
4612 /* fast path, check for non-NULL to avoid function call */
4614 kfree(kmsg->free_iov);
4615 req->flags &= ~REQ_F_NEED_CLEANUP;
4617 req_set_fail_links(req);
4618 __io_req_complete(req, issue_flags, ret, cflags);
4622 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4624 struct io_buffer *kbuf;
4625 struct io_sr_msg *sr = &req->sr_msg;
4627 void __user *buf = sr->buf;
4628 struct socket *sock;
4631 int ret, cflags = 0;
4632 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4634 sock = sock_from_file(req->file);
4635 if (unlikely(!sock))
4638 if (req->flags & REQ_F_BUFFER_SELECT) {
4639 kbuf = io_recv_buffer_select(req, !force_nonblock);
4641 return PTR_ERR(kbuf);
4642 buf = u64_to_user_ptr(kbuf->addr);
4645 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4649 msg.msg_name = NULL;
4650 msg.msg_control = NULL;
4651 msg.msg_controllen = 0;
4652 msg.msg_namelen = 0;
4653 msg.msg_iocb = NULL;
4656 flags = req->sr_msg.msg_flags;
4657 if (flags & MSG_DONTWAIT)
4658 req->flags |= REQ_F_NOWAIT;
4659 else if (force_nonblock)
4660 flags |= MSG_DONTWAIT;
4662 ret = sock_recvmsg(sock, &msg, flags);
4663 if (force_nonblock && ret == -EAGAIN)
4665 if (ret == -ERESTARTSYS)
4668 if (req->flags & REQ_F_BUFFER_SELECTED)
4669 cflags = io_put_recv_kbuf(req);
4671 req_set_fail_links(req);
4672 __io_req_complete(req, issue_flags, ret, cflags);
4676 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4678 struct io_accept *accept = &req->accept;
4680 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4682 if (sqe->ioprio || sqe->len || sqe->buf_index)
4685 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4686 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4687 accept->flags = READ_ONCE(sqe->accept_flags);
4688 accept->nofile = rlimit(RLIMIT_NOFILE);
4692 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4694 struct io_accept *accept = &req->accept;
4695 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4696 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4699 if (req->file->f_flags & O_NONBLOCK)
4700 req->flags |= REQ_F_NOWAIT;
4702 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4703 accept->addr_len, accept->flags,
4705 if (ret == -EAGAIN && force_nonblock)
4708 if (ret == -ERESTARTSYS)
4710 req_set_fail_links(req);
4712 __io_req_complete(req, issue_flags, ret, 0);
4716 static int io_connect_prep_async(struct io_kiocb *req)
4718 struct io_async_connect *io = req->async_data;
4719 struct io_connect *conn = &req->connect;
4721 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4724 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4726 struct io_connect *conn = &req->connect;
4728 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4730 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4733 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4734 conn->addr_len = READ_ONCE(sqe->addr2);
4738 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4740 struct io_async_connect __io, *io;
4741 unsigned file_flags;
4743 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4745 if (req->async_data) {
4746 io = req->async_data;
4748 ret = move_addr_to_kernel(req->connect.addr,
4749 req->connect.addr_len,
4756 file_flags = force_nonblock ? O_NONBLOCK : 0;
4758 ret = __sys_connect_file(req->file, &io->address,
4759 req->connect.addr_len, file_flags);
4760 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4761 if (req->async_data)
4763 if (io_alloc_async_data(req)) {
4767 io = req->async_data;
4768 memcpy(req->async_data, &__io, sizeof(__io));
4771 if (ret == -ERESTARTSYS)
4775 req_set_fail_links(req);
4776 __io_req_complete(req, issue_flags, ret, 0);
4779 #else /* !CONFIG_NET */
4780 #define IO_NETOP_FN(op) \
4781 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4783 return -EOPNOTSUPP; \
4786 #define IO_NETOP_PREP(op) \
4788 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4790 return -EOPNOTSUPP; \
4793 #define IO_NETOP_PREP_ASYNC(op) \
4795 static int io_##op##_prep_async(struct io_kiocb *req) \
4797 return -EOPNOTSUPP; \
4800 IO_NETOP_PREP_ASYNC(sendmsg);
4801 IO_NETOP_PREP_ASYNC(recvmsg);
4802 IO_NETOP_PREP_ASYNC(connect);
4803 IO_NETOP_PREP(accept);
4806 #endif /* CONFIG_NET */
4808 struct io_poll_table {
4809 struct poll_table_struct pt;
4810 struct io_kiocb *req;
4814 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4815 __poll_t mask, task_work_func_t func)
4819 /* for instances that support it check for an event match first: */
4820 if (mask && !(mask & poll->events))
4823 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4825 list_del_init(&poll->wait.entry);
4828 req->task_work.func = func;
4829 percpu_ref_get(&req->ctx->refs);
4832 * If this fails, then the task is exiting. When a task exits, the
4833 * work gets canceled, so just cancel this request as well instead
4834 * of executing it. We can't safely execute it anyway, as we may not
4835 * have the needed state needed for it anyway.
4837 ret = io_req_task_work_add(req);
4838 if (unlikely(ret)) {
4839 WRITE_ONCE(poll->canceled, true);
4840 io_req_task_work_add_fallback(req, func);
4845 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4846 __acquires(&req->ctx->completion_lock)
4848 struct io_ring_ctx *ctx = req->ctx;
4850 if (!req->result && !READ_ONCE(poll->canceled)) {
4851 struct poll_table_struct pt = { ._key = poll->events };
4853 req->result = vfs_poll(req->file, &pt) & poll->events;
4856 spin_lock_irq(&ctx->completion_lock);
4857 if (!req->result && !READ_ONCE(poll->canceled)) {
4858 add_wait_queue(poll->head, &poll->wait);
4865 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4867 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4868 if (req->opcode == IORING_OP_POLL_ADD)
4869 return req->async_data;
4870 return req->apoll->double_poll;
4873 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4875 if (req->opcode == IORING_OP_POLL_ADD)
4877 return &req->apoll->poll;
4880 static void io_poll_remove_double(struct io_kiocb *req)
4882 struct io_poll_iocb *poll = io_poll_get_double(req);
4884 lockdep_assert_held(&req->ctx->completion_lock);
4886 if (poll && poll->head) {
4887 struct wait_queue_head *head = poll->head;
4889 spin_lock(&head->lock);
4890 list_del_init(&poll->wait.entry);
4891 if (poll->wait.private)
4892 refcount_dec(&req->refs);
4894 spin_unlock(&head->lock);
4898 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4900 struct io_ring_ctx *ctx = req->ctx;
4902 io_poll_remove_double(req);
4903 req->poll.done = true;
4904 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4905 io_commit_cqring(ctx);
4908 static void io_poll_task_func(struct callback_head *cb)
4910 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4911 struct io_ring_ctx *ctx = req->ctx;
4912 struct io_kiocb *nxt;
4914 if (io_poll_rewait(req, &req->poll)) {
4915 spin_unlock_irq(&ctx->completion_lock);
4917 hash_del(&req->hash_node);
4918 io_poll_complete(req, req->result, 0);
4919 spin_unlock_irq(&ctx->completion_lock);
4921 nxt = io_put_req_find_next(req);
4922 io_cqring_ev_posted(ctx);
4924 __io_req_task_submit(nxt);
4927 percpu_ref_put(&ctx->refs);
4930 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4931 int sync, void *key)
4933 struct io_kiocb *req = wait->private;
4934 struct io_poll_iocb *poll = io_poll_get_single(req);
4935 __poll_t mask = key_to_poll(key);
4937 /* for instances that support it check for an event match first: */
4938 if (mask && !(mask & poll->events))
4941 list_del_init(&wait->entry);
4943 if (poll && poll->head) {
4946 spin_lock(&poll->head->lock);
4947 done = list_empty(&poll->wait.entry);
4949 list_del_init(&poll->wait.entry);
4950 /* make sure double remove sees this as being gone */
4951 wait->private = NULL;
4952 spin_unlock(&poll->head->lock);
4954 /* use wait func handler, so it matches the rq type */
4955 poll->wait.func(&poll->wait, mode, sync, key);
4958 refcount_dec(&req->refs);
4962 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4963 wait_queue_func_t wake_func)
4967 poll->canceled = false;
4968 poll->events = events;
4969 INIT_LIST_HEAD(&poll->wait.entry);
4970 init_waitqueue_func_entry(&poll->wait, wake_func);
4973 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4974 struct wait_queue_head *head,
4975 struct io_poll_iocb **poll_ptr)
4977 struct io_kiocb *req = pt->req;
4980 * If poll->head is already set, it's because the file being polled
4981 * uses multiple waitqueues for poll handling (eg one for read, one
4982 * for write). Setup a separate io_poll_iocb if this happens.
4984 if (unlikely(poll->head)) {
4985 struct io_poll_iocb *poll_one = poll;
4987 /* already have a 2nd entry, fail a third attempt */
4989 pt->error = -EINVAL;
4992 /* double add on the same waitqueue head, ignore */
4993 if (poll->head == head)
4995 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4997 pt->error = -ENOMEM;
5000 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5001 refcount_inc(&req->refs);
5002 poll->wait.private = req;
5009 if (poll->events & EPOLLEXCLUSIVE)
5010 add_wait_queue_exclusive(head, &poll->wait);
5012 add_wait_queue(head, &poll->wait);
5015 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5016 struct poll_table_struct *p)
5018 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5019 struct async_poll *apoll = pt->req->apoll;
5021 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5024 static void io_async_task_func(struct callback_head *cb)
5026 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5027 struct async_poll *apoll = req->apoll;
5028 struct io_ring_ctx *ctx = req->ctx;
5030 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5032 if (io_poll_rewait(req, &apoll->poll)) {
5033 spin_unlock_irq(&ctx->completion_lock);
5034 percpu_ref_put(&ctx->refs);
5038 /* If req is still hashed, it cannot have been canceled. Don't check. */
5039 if (hash_hashed(&req->hash_node))
5040 hash_del(&req->hash_node);
5042 io_poll_remove_double(req);
5043 spin_unlock_irq(&ctx->completion_lock);
5045 if (!READ_ONCE(apoll->poll.canceled))
5046 __io_req_task_submit(req);
5048 __io_req_task_cancel(req, -ECANCELED);
5050 percpu_ref_put(&ctx->refs);
5051 kfree(apoll->double_poll);
5055 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5058 struct io_kiocb *req = wait->private;
5059 struct io_poll_iocb *poll = &req->apoll->poll;
5061 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5064 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5067 static void io_poll_req_insert(struct io_kiocb *req)
5069 struct io_ring_ctx *ctx = req->ctx;
5070 struct hlist_head *list;
5072 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5073 hlist_add_head(&req->hash_node, list);
5076 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5077 struct io_poll_iocb *poll,
5078 struct io_poll_table *ipt, __poll_t mask,
5079 wait_queue_func_t wake_func)
5080 __acquires(&ctx->completion_lock)
5082 struct io_ring_ctx *ctx = req->ctx;
5083 bool cancel = false;
5085 INIT_HLIST_NODE(&req->hash_node);
5086 io_init_poll_iocb(poll, mask, wake_func);
5087 poll->file = req->file;
5088 poll->wait.private = req;
5090 ipt->pt._key = mask;
5092 ipt->error = -EINVAL;
5094 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5096 spin_lock_irq(&ctx->completion_lock);
5097 if (likely(poll->head)) {
5098 spin_lock(&poll->head->lock);
5099 if (unlikely(list_empty(&poll->wait.entry))) {
5105 if (mask || ipt->error)
5106 list_del_init(&poll->wait.entry);
5108 WRITE_ONCE(poll->canceled, true);
5109 else if (!poll->done) /* actually waiting for an event */
5110 io_poll_req_insert(req);
5111 spin_unlock(&poll->head->lock);
5117 static bool io_arm_poll_handler(struct io_kiocb *req)
5119 const struct io_op_def *def = &io_op_defs[req->opcode];
5120 struct io_ring_ctx *ctx = req->ctx;
5121 struct async_poll *apoll;
5122 struct io_poll_table ipt;
5126 if (!req->file || !file_can_poll(req->file))
5128 if (req->flags & REQ_F_POLLED)
5132 else if (def->pollout)
5136 /* if we can't nonblock try, then no point in arming a poll handler */
5137 if (!io_file_supports_async(req->file, rw))
5140 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5141 if (unlikely(!apoll))
5143 apoll->double_poll = NULL;
5145 req->flags |= REQ_F_POLLED;
5150 mask |= POLLIN | POLLRDNORM;
5152 mask |= POLLOUT | POLLWRNORM;
5154 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5155 if ((req->opcode == IORING_OP_RECVMSG) &&
5156 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5159 mask |= POLLERR | POLLPRI;
5161 ipt.pt._qproc = io_async_queue_proc;
5163 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5165 if (ret || ipt.error) {
5166 io_poll_remove_double(req);
5167 spin_unlock_irq(&ctx->completion_lock);
5168 kfree(apoll->double_poll);
5172 spin_unlock_irq(&ctx->completion_lock);
5173 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5174 apoll->poll.events);
5178 static bool __io_poll_remove_one(struct io_kiocb *req,
5179 struct io_poll_iocb *poll)
5181 bool do_complete = false;
5183 spin_lock(&poll->head->lock);
5184 WRITE_ONCE(poll->canceled, true);
5185 if (!list_empty(&poll->wait.entry)) {
5186 list_del_init(&poll->wait.entry);
5189 spin_unlock(&poll->head->lock);
5190 hash_del(&req->hash_node);
5194 static bool io_poll_remove_one(struct io_kiocb *req)
5198 io_poll_remove_double(req);
5200 if (req->opcode == IORING_OP_POLL_ADD) {
5201 do_complete = __io_poll_remove_one(req, &req->poll);
5203 struct async_poll *apoll = req->apoll;
5205 /* non-poll requests have submit ref still */
5206 do_complete = __io_poll_remove_one(req, &apoll->poll);
5209 kfree(apoll->double_poll);
5215 io_cqring_fill_event(req, -ECANCELED);
5216 io_commit_cqring(req->ctx);
5217 req_set_fail_links(req);
5218 io_put_req_deferred(req, 1);
5225 * Returns true if we found and killed one or more poll requests
5227 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5228 struct files_struct *files)
5230 struct hlist_node *tmp;
5231 struct io_kiocb *req;
5234 spin_lock_irq(&ctx->completion_lock);
5235 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5236 struct hlist_head *list;
5238 list = &ctx->cancel_hash[i];
5239 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5240 if (io_match_task(req, tsk, files))
5241 posted += io_poll_remove_one(req);
5244 spin_unlock_irq(&ctx->completion_lock);
5247 io_cqring_ev_posted(ctx);
5252 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5254 struct hlist_head *list;
5255 struct io_kiocb *req;
5257 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5258 hlist_for_each_entry(req, list, hash_node) {
5259 if (sqe_addr != req->user_data)
5261 if (io_poll_remove_one(req))
5269 static int io_poll_remove_prep(struct io_kiocb *req,
5270 const struct io_uring_sqe *sqe)
5272 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5274 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5278 req->poll_remove.addr = READ_ONCE(sqe->addr);
5283 * Find a running poll command that matches one specified in sqe->addr,
5284 * and remove it if found.
5286 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5288 struct io_ring_ctx *ctx = req->ctx;
5291 spin_lock_irq(&ctx->completion_lock);
5292 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5293 spin_unlock_irq(&ctx->completion_lock);
5296 req_set_fail_links(req);
5297 io_req_complete(req, ret);
5301 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5304 struct io_kiocb *req = wait->private;
5305 struct io_poll_iocb *poll = &req->poll;
5307 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5310 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5311 struct poll_table_struct *p)
5313 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5315 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5318 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5320 struct io_poll_iocb *poll = &req->poll;
5323 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5325 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5328 events = READ_ONCE(sqe->poll32_events);
5330 events = swahw32(events);
5332 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5333 (events & EPOLLEXCLUSIVE);
5337 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5339 struct io_poll_iocb *poll = &req->poll;
5340 struct io_ring_ctx *ctx = req->ctx;
5341 struct io_poll_table ipt;
5344 ipt.pt._qproc = io_poll_queue_proc;
5346 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5349 if (mask) { /* no async, we'd stolen it */
5351 io_poll_complete(req, mask, 0);
5353 spin_unlock_irq(&ctx->completion_lock);
5356 io_cqring_ev_posted(ctx);
5362 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5364 struct io_timeout_data *data = container_of(timer,
5365 struct io_timeout_data, timer);
5366 struct io_kiocb *req = data->req;
5367 struct io_ring_ctx *ctx = req->ctx;
5368 unsigned long flags;
5370 spin_lock_irqsave(&ctx->completion_lock, flags);
5371 list_del_init(&req->timeout.list);
5372 atomic_set(&req->ctx->cq_timeouts,
5373 atomic_read(&req->ctx->cq_timeouts) + 1);
5375 io_cqring_fill_event(req, -ETIME);
5376 io_commit_cqring(ctx);
5377 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5379 io_cqring_ev_posted(ctx);
5380 req_set_fail_links(req);
5382 return HRTIMER_NORESTART;
5385 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5388 struct io_timeout_data *io;
5389 struct io_kiocb *req;
5392 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5393 if (user_data == req->user_data) {
5400 return ERR_PTR(ret);
5402 io = req->async_data;
5403 ret = hrtimer_try_to_cancel(&io->timer);
5405 return ERR_PTR(-EALREADY);
5406 list_del_init(&req->timeout.list);
5410 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5412 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5415 return PTR_ERR(req);
5417 req_set_fail_links(req);
5418 io_cqring_fill_event(req, -ECANCELED);
5419 io_put_req_deferred(req, 1);
5423 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5424 struct timespec64 *ts, enum hrtimer_mode mode)
5426 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5427 struct io_timeout_data *data;
5430 return PTR_ERR(req);
5432 req->timeout.off = 0; /* noseq */
5433 data = req->async_data;
5434 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5435 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5436 data->timer.function = io_timeout_fn;
5437 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5441 static int io_timeout_remove_prep(struct io_kiocb *req,
5442 const struct io_uring_sqe *sqe)
5444 struct io_timeout_rem *tr = &req->timeout_rem;
5446 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5448 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5450 if (sqe->ioprio || sqe->buf_index || sqe->len)
5453 tr->addr = READ_ONCE(sqe->addr);
5454 tr->flags = READ_ONCE(sqe->timeout_flags);
5455 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5456 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5458 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5460 } else if (tr->flags) {
5461 /* timeout removal doesn't support flags */
5468 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5470 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5475 * Remove or update an existing timeout command
5477 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5479 struct io_timeout_rem *tr = &req->timeout_rem;
5480 struct io_ring_ctx *ctx = req->ctx;
5483 spin_lock_irq(&ctx->completion_lock);
5484 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5485 ret = io_timeout_cancel(ctx, tr->addr);
5487 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5488 io_translate_timeout_mode(tr->flags));
5490 io_cqring_fill_event(req, ret);
5491 io_commit_cqring(ctx);
5492 spin_unlock_irq(&ctx->completion_lock);
5493 io_cqring_ev_posted(ctx);
5495 req_set_fail_links(req);
5500 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5501 bool is_timeout_link)
5503 struct io_timeout_data *data;
5505 u32 off = READ_ONCE(sqe->off);
5507 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5509 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5511 if (off && is_timeout_link)
5513 flags = READ_ONCE(sqe->timeout_flags);
5514 if (flags & ~IORING_TIMEOUT_ABS)
5517 req->timeout.off = off;
5519 if (!req->async_data && io_alloc_async_data(req))
5522 data = req->async_data;
5525 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5528 data->mode = io_translate_timeout_mode(flags);
5529 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5530 io_req_track_inflight(req);
5534 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5536 struct io_ring_ctx *ctx = req->ctx;
5537 struct io_timeout_data *data = req->async_data;
5538 struct list_head *entry;
5539 u32 tail, off = req->timeout.off;
5541 spin_lock_irq(&ctx->completion_lock);
5544 * sqe->off holds how many events that need to occur for this
5545 * timeout event to be satisfied. If it isn't set, then this is
5546 * a pure timeout request, sequence isn't used.
5548 if (io_is_timeout_noseq(req)) {
5549 entry = ctx->timeout_list.prev;
5553 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5554 req->timeout.target_seq = tail + off;
5556 /* Update the last seq here in case io_flush_timeouts() hasn't.
5557 * This is safe because ->completion_lock is held, and submissions
5558 * and completions are never mixed in the same ->completion_lock section.
5560 ctx->cq_last_tm_flush = tail;
5563 * Insertion sort, ensuring the first entry in the list is always
5564 * the one we need first.
5566 list_for_each_prev(entry, &ctx->timeout_list) {
5567 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5570 if (io_is_timeout_noseq(nxt))
5572 /* nxt.seq is behind @tail, otherwise would've been completed */
5573 if (off >= nxt->timeout.target_seq - tail)
5577 list_add(&req->timeout.list, entry);
5578 data->timer.function = io_timeout_fn;
5579 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5580 spin_unlock_irq(&ctx->completion_lock);
5584 struct io_cancel_data {
5585 struct io_ring_ctx *ctx;
5589 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5591 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5592 struct io_cancel_data *cd = data;
5594 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5597 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5598 struct io_ring_ctx *ctx)
5600 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5601 enum io_wq_cancel cancel_ret;
5604 if (!tctx || !tctx->io_wq)
5607 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5608 switch (cancel_ret) {
5609 case IO_WQ_CANCEL_OK:
5612 case IO_WQ_CANCEL_RUNNING:
5615 case IO_WQ_CANCEL_NOTFOUND:
5623 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5624 struct io_kiocb *req, __u64 sqe_addr,
5627 unsigned long flags;
5630 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5631 if (ret != -ENOENT) {
5632 spin_lock_irqsave(&ctx->completion_lock, flags);
5636 spin_lock_irqsave(&ctx->completion_lock, flags);
5637 ret = io_timeout_cancel(ctx, sqe_addr);
5640 ret = io_poll_cancel(ctx, sqe_addr);
5644 io_cqring_fill_event(req, ret);
5645 io_commit_cqring(ctx);
5646 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5647 io_cqring_ev_posted(ctx);
5650 req_set_fail_links(req);
5654 static int io_async_cancel_prep(struct io_kiocb *req,
5655 const struct io_uring_sqe *sqe)
5657 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5659 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5661 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5664 req->cancel.addr = READ_ONCE(sqe->addr);
5668 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5670 struct io_ring_ctx *ctx = req->ctx;
5672 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5676 static int io_rsrc_update_prep(struct io_kiocb *req,
5677 const struct io_uring_sqe *sqe)
5679 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5681 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5683 if (sqe->ioprio || sqe->rw_flags)
5686 req->rsrc_update.offset = READ_ONCE(sqe->off);
5687 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5688 if (!req->rsrc_update.nr_args)
5690 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5694 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5696 struct io_ring_ctx *ctx = req->ctx;
5697 struct io_uring_rsrc_update up;
5700 if (issue_flags & IO_URING_F_NONBLOCK)
5703 up.offset = req->rsrc_update.offset;
5704 up.data = req->rsrc_update.arg;
5706 mutex_lock(&ctx->uring_lock);
5707 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5708 mutex_unlock(&ctx->uring_lock);
5711 req_set_fail_links(req);
5712 __io_req_complete(req, issue_flags, ret, 0);
5716 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5718 switch (req->opcode) {
5721 case IORING_OP_READV:
5722 case IORING_OP_READ_FIXED:
5723 case IORING_OP_READ:
5724 return io_read_prep(req, sqe);
5725 case IORING_OP_WRITEV:
5726 case IORING_OP_WRITE_FIXED:
5727 case IORING_OP_WRITE:
5728 return io_write_prep(req, sqe);
5729 case IORING_OP_POLL_ADD:
5730 return io_poll_add_prep(req, sqe);
5731 case IORING_OP_POLL_REMOVE:
5732 return io_poll_remove_prep(req, sqe);
5733 case IORING_OP_FSYNC:
5734 return io_fsync_prep(req, sqe);
5735 case IORING_OP_SYNC_FILE_RANGE:
5736 return io_sfr_prep(req, sqe);
5737 case IORING_OP_SENDMSG:
5738 case IORING_OP_SEND:
5739 return io_sendmsg_prep(req, sqe);
5740 case IORING_OP_RECVMSG:
5741 case IORING_OP_RECV:
5742 return io_recvmsg_prep(req, sqe);
5743 case IORING_OP_CONNECT:
5744 return io_connect_prep(req, sqe);
5745 case IORING_OP_TIMEOUT:
5746 return io_timeout_prep(req, sqe, false);
5747 case IORING_OP_TIMEOUT_REMOVE:
5748 return io_timeout_remove_prep(req, sqe);
5749 case IORING_OP_ASYNC_CANCEL:
5750 return io_async_cancel_prep(req, sqe);
5751 case IORING_OP_LINK_TIMEOUT:
5752 return io_timeout_prep(req, sqe, true);
5753 case IORING_OP_ACCEPT:
5754 return io_accept_prep(req, sqe);
5755 case IORING_OP_FALLOCATE:
5756 return io_fallocate_prep(req, sqe);
5757 case IORING_OP_OPENAT:
5758 return io_openat_prep(req, sqe);
5759 case IORING_OP_CLOSE:
5760 return io_close_prep(req, sqe);
5761 case IORING_OP_FILES_UPDATE:
5762 return io_rsrc_update_prep(req, sqe);
5763 case IORING_OP_STATX:
5764 return io_statx_prep(req, sqe);
5765 case IORING_OP_FADVISE:
5766 return io_fadvise_prep(req, sqe);
5767 case IORING_OP_MADVISE:
5768 return io_madvise_prep(req, sqe);
5769 case IORING_OP_OPENAT2:
5770 return io_openat2_prep(req, sqe);
5771 case IORING_OP_EPOLL_CTL:
5772 return io_epoll_ctl_prep(req, sqe);
5773 case IORING_OP_SPLICE:
5774 return io_splice_prep(req, sqe);
5775 case IORING_OP_PROVIDE_BUFFERS:
5776 return io_provide_buffers_prep(req, sqe);
5777 case IORING_OP_REMOVE_BUFFERS:
5778 return io_remove_buffers_prep(req, sqe);
5780 return io_tee_prep(req, sqe);
5781 case IORING_OP_SHUTDOWN:
5782 return io_shutdown_prep(req, sqe);
5783 case IORING_OP_RENAMEAT:
5784 return io_renameat_prep(req, sqe);
5785 case IORING_OP_UNLINKAT:
5786 return io_unlinkat_prep(req, sqe);
5789 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5794 static int io_req_prep_async(struct io_kiocb *req)
5796 switch (req->opcode) {
5797 case IORING_OP_READV:
5798 case IORING_OP_READ_FIXED:
5799 case IORING_OP_READ:
5800 return io_rw_prep_async(req, READ);
5801 case IORING_OP_WRITEV:
5802 case IORING_OP_WRITE_FIXED:
5803 case IORING_OP_WRITE:
5804 return io_rw_prep_async(req, WRITE);
5805 case IORING_OP_SENDMSG:
5806 case IORING_OP_SEND:
5807 return io_sendmsg_prep_async(req);
5808 case IORING_OP_RECVMSG:
5809 case IORING_OP_RECV:
5810 return io_recvmsg_prep_async(req);
5811 case IORING_OP_CONNECT:
5812 return io_connect_prep_async(req);
5817 static int io_req_defer_prep(struct io_kiocb *req)
5819 if (!io_op_defs[req->opcode].needs_async_data)
5821 /* some opcodes init it during the inital prep */
5822 if (req->async_data)
5824 if (__io_alloc_async_data(req))
5826 return io_req_prep_async(req);
5829 static u32 io_get_sequence(struct io_kiocb *req)
5831 struct io_kiocb *pos;
5832 struct io_ring_ctx *ctx = req->ctx;
5833 u32 total_submitted, nr_reqs = 0;
5835 io_for_each_link(pos, req)
5838 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5839 return total_submitted - nr_reqs;
5842 static int io_req_defer(struct io_kiocb *req)
5844 struct io_ring_ctx *ctx = req->ctx;
5845 struct io_defer_entry *de;
5849 /* Still need defer if there is pending req in defer list. */
5850 if (likely(list_empty_careful(&ctx->defer_list) &&
5851 !(req->flags & REQ_F_IO_DRAIN)))
5854 seq = io_get_sequence(req);
5855 /* Still a chance to pass the sequence check */
5856 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5859 ret = io_req_defer_prep(req);
5862 io_prep_async_link(req);
5863 de = kmalloc(sizeof(*de), GFP_KERNEL);
5867 spin_lock_irq(&ctx->completion_lock);
5868 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5869 spin_unlock_irq(&ctx->completion_lock);
5871 io_queue_async_work(req);
5872 return -EIOCBQUEUED;
5875 trace_io_uring_defer(ctx, req, req->user_data);
5878 list_add_tail(&de->list, &ctx->defer_list);
5879 spin_unlock_irq(&ctx->completion_lock);
5880 return -EIOCBQUEUED;
5883 static void __io_clean_op(struct io_kiocb *req)
5885 if (req->flags & REQ_F_BUFFER_SELECTED) {
5886 switch (req->opcode) {
5887 case IORING_OP_READV:
5888 case IORING_OP_READ_FIXED:
5889 case IORING_OP_READ:
5890 kfree((void *)(unsigned long)req->rw.addr);
5892 case IORING_OP_RECVMSG:
5893 case IORING_OP_RECV:
5894 kfree(req->sr_msg.kbuf);
5897 req->flags &= ~REQ_F_BUFFER_SELECTED;
5900 if (req->flags & REQ_F_NEED_CLEANUP) {
5901 switch (req->opcode) {
5902 case IORING_OP_READV:
5903 case IORING_OP_READ_FIXED:
5904 case IORING_OP_READ:
5905 case IORING_OP_WRITEV:
5906 case IORING_OP_WRITE_FIXED:
5907 case IORING_OP_WRITE: {
5908 struct io_async_rw *io = req->async_data;
5910 kfree(io->free_iovec);
5913 case IORING_OP_RECVMSG:
5914 case IORING_OP_SENDMSG: {
5915 struct io_async_msghdr *io = req->async_data;
5917 kfree(io->free_iov);
5920 case IORING_OP_SPLICE:
5922 io_put_file(req, req->splice.file_in,
5923 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5925 case IORING_OP_OPENAT:
5926 case IORING_OP_OPENAT2:
5927 if (req->open.filename)
5928 putname(req->open.filename);
5930 case IORING_OP_RENAMEAT:
5931 putname(req->rename.oldpath);
5932 putname(req->rename.newpath);
5934 case IORING_OP_UNLINKAT:
5935 putname(req->unlink.filename);
5938 req->flags &= ~REQ_F_NEED_CLEANUP;
5942 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5944 struct io_ring_ctx *ctx = req->ctx;
5945 const struct cred *creds = NULL;
5948 if (req->work.creds && req->work.creds != current_cred())
5949 creds = override_creds(req->work.creds);
5951 switch (req->opcode) {
5953 ret = io_nop(req, issue_flags);
5955 case IORING_OP_READV:
5956 case IORING_OP_READ_FIXED:
5957 case IORING_OP_READ:
5958 ret = io_read(req, issue_flags);
5960 case IORING_OP_WRITEV:
5961 case IORING_OP_WRITE_FIXED:
5962 case IORING_OP_WRITE:
5963 ret = io_write(req, issue_flags);
5965 case IORING_OP_FSYNC:
5966 ret = io_fsync(req, issue_flags);
5968 case IORING_OP_POLL_ADD:
5969 ret = io_poll_add(req, issue_flags);
5971 case IORING_OP_POLL_REMOVE:
5972 ret = io_poll_remove(req, issue_flags);
5974 case IORING_OP_SYNC_FILE_RANGE:
5975 ret = io_sync_file_range(req, issue_flags);
5977 case IORING_OP_SENDMSG:
5978 ret = io_sendmsg(req, issue_flags);
5980 case IORING_OP_SEND:
5981 ret = io_send(req, issue_flags);
5983 case IORING_OP_RECVMSG:
5984 ret = io_recvmsg(req, issue_flags);
5986 case IORING_OP_RECV:
5987 ret = io_recv(req, issue_flags);
5989 case IORING_OP_TIMEOUT:
5990 ret = io_timeout(req, issue_flags);
5992 case IORING_OP_TIMEOUT_REMOVE:
5993 ret = io_timeout_remove(req, issue_flags);
5995 case IORING_OP_ACCEPT:
5996 ret = io_accept(req, issue_flags);
5998 case IORING_OP_CONNECT:
5999 ret = io_connect(req, issue_flags);
6001 case IORING_OP_ASYNC_CANCEL:
6002 ret = io_async_cancel(req, issue_flags);
6004 case IORING_OP_FALLOCATE:
6005 ret = io_fallocate(req, issue_flags);
6007 case IORING_OP_OPENAT:
6008 ret = io_openat(req, issue_flags);
6010 case IORING_OP_CLOSE:
6011 ret = io_close(req, issue_flags);
6013 case IORING_OP_FILES_UPDATE:
6014 ret = io_files_update(req, issue_flags);
6016 case IORING_OP_STATX:
6017 ret = io_statx(req, issue_flags);
6019 case IORING_OP_FADVISE:
6020 ret = io_fadvise(req, issue_flags);
6022 case IORING_OP_MADVISE:
6023 ret = io_madvise(req, issue_flags);
6025 case IORING_OP_OPENAT2:
6026 ret = io_openat2(req, issue_flags);
6028 case IORING_OP_EPOLL_CTL:
6029 ret = io_epoll_ctl(req, issue_flags);
6031 case IORING_OP_SPLICE:
6032 ret = io_splice(req, issue_flags);
6034 case IORING_OP_PROVIDE_BUFFERS:
6035 ret = io_provide_buffers(req, issue_flags);
6037 case IORING_OP_REMOVE_BUFFERS:
6038 ret = io_remove_buffers(req, issue_flags);
6041 ret = io_tee(req, issue_flags);
6043 case IORING_OP_SHUTDOWN:
6044 ret = io_shutdown(req, issue_flags);
6046 case IORING_OP_RENAMEAT:
6047 ret = io_renameat(req, issue_flags);
6049 case IORING_OP_UNLINKAT:
6050 ret = io_unlinkat(req, issue_flags);
6058 revert_creds(creds);
6063 /* If the op doesn't have a file, we're not polling for it */
6064 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6065 const bool in_async = io_wq_current_is_worker();
6067 /* workqueue context doesn't hold uring_lock, grab it now */
6069 mutex_lock(&ctx->uring_lock);
6071 io_iopoll_req_issued(req, in_async);
6074 mutex_unlock(&ctx->uring_lock);
6080 static void io_wq_submit_work(struct io_wq_work *work)
6082 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6083 struct io_kiocb *timeout;
6086 timeout = io_prep_linked_timeout(req);
6088 io_queue_linked_timeout(timeout);
6090 if (work->flags & IO_WQ_WORK_CANCEL)
6095 ret = io_issue_sqe(req, 0);
6097 * We can get EAGAIN for polled IO even though we're
6098 * forcing a sync submission from here, since we can't
6099 * wait for request slots on the block side.
6107 /* avoid locking problems by failing it from a clean context */
6109 /* io-wq is going to take one down */
6110 refcount_inc(&req->refs);
6111 io_req_task_queue_fail(req, ret);
6115 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6118 struct fixed_rsrc_table *table;
6120 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6121 return table->files[index & IORING_FILE_TABLE_MASK];
6124 static struct file *io_file_get(struct io_submit_state *state,
6125 struct io_kiocb *req, int fd, bool fixed)
6127 struct io_ring_ctx *ctx = req->ctx;
6131 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6133 fd = array_index_nospec(fd, ctx->nr_user_files);
6134 file = io_file_from_index(ctx, fd);
6135 io_set_resource_node(req);
6137 trace_io_uring_file_get(ctx, fd);
6138 file = __io_file_get(state, fd);
6141 if (file && unlikely(file->f_op == &io_uring_fops))
6142 io_req_track_inflight(req);
6146 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6148 struct io_timeout_data *data = container_of(timer,
6149 struct io_timeout_data, timer);
6150 struct io_kiocb *prev, *req = data->req;
6151 struct io_ring_ctx *ctx = req->ctx;
6152 unsigned long flags;
6154 spin_lock_irqsave(&ctx->completion_lock, flags);
6155 prev = req->timeout.head;
6156 req->timeout.head = NULL;
6159 * We don't expect the list to be empty, that will only happen if we
6160 * race with the completion of the linked work.
6162 if (prev && refcount_inc_not_zero(&prev->refs))
6163 io_remove_next_linked(prev);
6166 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6169 req_set_fail_links(prev);
6170 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6171 io_put_req_deferred(prev, 1);
6173 io_req_complete_post(req, -ETIME, 0);
6174 io_put_req_deferred(req, 1);
6176 return HRTIMER_NORESTART;
6179 static void __io_queue_linked_timeout(struct io_kiocb *req)
6182 * If the back reference is NULL, then our linked request finished
6183 * before we got a chance to setup the timer
6185 if (req->timeout.head) {
6186 struct io_timeout_data *data = req->async_data;
6188 data->timer.function = io_link_timeout_fn;
6189 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6194 static void io_queue_linked_timeout(struct io_kiocb *req)
6196 struct io_ring_ctx *ctx = req->ctx;
6198 spin_lock_irq(&ctx->completion_lock);
6199 __io_queue_linked_timeout(req);
6200 spin_unlock_irq(&ctx->completion_lock);
6202 /* drop submission reference */
6206 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6208 struct io_kiocb *nxt = req->link;
6210 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6211 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6214 nxt->timeout.head = req;
6215 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6216 req->flags |= REQ_F_LINK_TIMEOUT;
6220 static void __io_queue_sqe(struct io_kiocb *req)
6222 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6225 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6228 * We async punt it if the file wasn't marked NOWAIT, or if the file
6229 * doesn't support non-blocking read/write attempts
6231 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6232 if (!io_arm_poll_handler(req)) {
6234 * Queued up for async execution, worker will release
6235 * submit reference when the iocb is actually submitted.
6237 io_queue_async_work(req);
6239 } else if (likely(!ret)) {
6240 /* drop submission reference */
6241 if (req->flags & REQ_F_COMPLETE_INLINE) {
6242 struct io_ring_ctx *ctx = req->ctx;
6243 struct io_comp_state *cs = &ctx->submit_state.comp;
6245 cs->reqs[cs->nr++] = req;
6246 if (cs->nr == ARRAY_SIZE(cs->reqs))
6247 io_submit_flush_completions(cs, ctx);
6252 req_set_fail_links(req);
6254 io_req_complete(req, ret);
6257 io_queue_linked_timeout(linked_timeout);
6260 static void io_queue_sqe(struct io_kiocb *req)
6264 ret = io_req_defer(req);
6266 if (ret != -EIOCBQUEUED) {
6268 req_set_fail_links(req);
6270 io_req_complete(req, ret);
6272 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6273 ret = io_req_defer_prep(req);
6276 io_queue_async_work(req);
6278 __io_queue_sqe(req);
6283 * Check SQE restrictions (opcode and flags).
6285 * Returns 'true' if SQE is allowed, 'false' otherwise.
6287 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6288 struct io_kiocb *req,
6289 unsigned int sqe_flags)
6291 if (!ctx->restricted)
6294 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6297 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6298 ctx->restrictions.sqe_flags_required)
6301 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6302 ctx->restrictions.sqe_flags_required))
6308 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6309 const struct io_uring_sqe *sqe)
6311 struct io_submit_state *state;
6312 unsigned int sqe_flags;
6313 int personality, ret = 0;
6315 req->opcode = READ_ONCE(sqe->opcode);
6316 /* same numerical values with corresponding REQ_F_*, safe to copy */
6317 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6318 req->user_data = READ_ONCE(sqe->user_data);
6319 req->async_data = NULL;
6323 req->fixed_rsrc_refs = NULL;
6324 /* one is dropped after submission, the other at completion */
6325 refcount_set(&req->refs, 2);
6326 req->task = current;
6328 req->work.list.next = NULL;
6329 req->work.creds = NULL;
6330 req->work.flags = 0;
6332 /* enforce forwards compatibility on users */
6333 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6338 if (unlikely(req->opcode >= IORING_OP_LAST))
6341 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6344 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6345 !io_op_defs[req->opcode].buffer_select)
6348 personality = READ_ONCE(sqe->personality);
6350 req->work.creds = xa_load(&ctx->personalities, personality);
6351 if (!req->work.creds)
6353 get_cred(req->work.creds);
6355 state = &ctx->submit_state;
6358 * Plug now if we have more than 1 IO left after this, and the target
6359 * is potentially a read/write to block based storage.
6361 if (!state->plug_started && state->ios_left > 1 &&
6362 io_op_defs[req->opcode].plug) {
6363 blk_start_plug(&state->plug);
6364 state->plug_started = true;
6367 if (io_op_defs[req->opcode].needs_file) {
6368 bool fixed = req->flags & REQ_F_FIXED_FILE;
6370 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6371 if (unlikely(!req->file))
6379 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6380 const struct io_uring_sqe *sqe)
6382 struct io_submit_link *link = &ctx->submit_state.link;
6385 ret = io_init_req(ctx, req, sqe);
6386 if (unlikely(ret)) {
6389 io_req_complete(req, ret);
6391 /* fail even hard links since we don't submit */
6392 link->head->flags |= REQ_F_FAIL_LINK;
6393 io_put_req(link->head);
6394 io_req_complete(link->head, -ECANCELED);
6399 ret = io_req_prep(req, sqe);
6403 /* don't need @sqe from now on */
6404 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6405 true, ctx->flags & IORING_SETUP_SQPOLL);
6408 * If we already have a head request, queue this one for async
6409 * submittal once the head completes. If we don't have a head but
6410 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6411 * submitted sync once the chain is complete. If none of those
6412 * conditions are true (normal request), then just queue it.
6415 struct io_kiocb *head = link->head;
6418 * Taking sequential execution of a link, draining both sides
6419 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6420 * requests in the link. So, it drains the head and the
6421 * next after the link request. The last one is done via
6422 * drain_next flag to persist the effect across calls.
6424 if (req->flags & REQ_F_IO_DRAIN) {
6425 head->flags |= REQ_F_IO_DRAIN;
6426 ctx->drain_next = 1;
6428 ret = io_req_defer_prep(req);
6431 trace_io_uring_link(ctx, req, head);
6432 link->last->link = req;
6435 /* last request of a link, enqueue the link */
6436 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6441 if (unlikely(ctx->drain_next)) {
6442 req->flags |= REQ_F_IO_DRAIN;
6443 ctx->drain_next = 0;
6445 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6457 * Batched submission is done, ensure local IO is flushed out.
6459 static void io_submit_state_end(struct io_submit_state *state,
6460 struct io_ring_ctx *ctx)
6462 if (state->link.head)
6463 io_queue_sqe(state->link.head);
6465 io_submit_flush_completions(&state->comp, ctx);
6466 if (state->plug_started)
6467 blk_finish_plug(&state->plug);
6468 io_state_file_put(state);
6472 * Start submission side cache.
6474 static void io_submit_state_start(struct io_submit_state *state,
6475 unsigned int max_ios)
6477 state->plug_started = false;
6478 state->ios_left = max_ios;
6479 /* set only head, no need to init link_last in advance */
6480 state->link.head = NULL;
6483 static void io_commit_sqring(struct io_ring_ctx *ctx)
6485 struct io_rings *rings = ctx->rings;
6488 * Ensure any loads from the SQEs are done at this point,
6489 * since once we write the new head, the application could
6490 * write new data to them.
6492 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6496 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6497 * that is mapped by userspace. This means that care needs to be taken to
6498 * ensure that reads are stable, as we cannot rely on userspace always
6499 * being a good citizen. If members of the sqe are validated and then later
6500 * used, it's important that those reads are done through READ_ONCE() to
6501 * prevent a re-load down the line.
6503 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6505 u32 *sq_array = ctx->sq_array;
6509 * The cached sq head (or cq tail) serves two purposes:
6511 * 1) allows us to batch the cost of updating the user visible
6513 * 2) allows the kernel side to track the head on its own, even
6514 * though the application is the one updating it.
6516 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6517 if (likely(head < ctx->sq_entries))
6518 return &ctx->sq_sqes[head];
6520 /* drop invalid entries */
6521 ctx->cached_sq_dropped++;
6522 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6526 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6530 /* if we have a backlog and couldn't flush it all, return BUSY */
6531 if (test_bit(0, &ctx->sq_check_overflow)) {
6532 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6536 /* make sure SQ entry isn't read before tail */
6537 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6539 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6542 percpu_counter_add(¤t->io_uring->inflight, nr);
6543 refcount_add(nr, ¤t->usage);
6544 io_submit_state_start(&ctx->submit_state, nr);
6546 while (submitted < nr) {
6547 const struct io_uring_sqe *sqe;
6548 struct io_kiocb *req;
6550 req = io_alloc_req(ctx);
6551 if (unlikely(!req)) {
6553 submitted = -EAGAIN;
6556 sqe = io_get_sqe(ctx);
6557 if (unlikely(!sqe)) {
6558 kmem_cache_free(req_cachep, req);
6561 /* will complete beyond this point, count as submitted */
6563 if (io_submit_sqe(ctx, req, sqe))
6567 if (unlikely(submitted != nr)) {
6568 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6569 struct io_uring_task *tctx = current->io_uring;
6570 int unused = nr - ref_used;
6572 percpu_ref_put_many(&ctx->refs, unused);
6573 percpu_counter_sub(&tctx->inflight, unused);
6574 put_task_struct_many(current, unused);
6577 io_submit_state_end(&ctx->submit_state, ctx);
6578 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6579 io_commit_sqring(ctx);
6584 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6586 /* Tell userspace we may need a wakeup call */
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 inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6594 spin_lock_irq(&ctx->completion_lock);
6595 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6596 spin_unlock_irq(&ctx->completion_lock);
6599 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6601 unsigned int to_submit;
6604 to_submit = io_sqring_entries(ctx);
6605 /* if we're handling multiple rings, cap submit size for fairness */
6606 if (cap_entries && to_submit > 8)
6609 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6610 unsigned nr_events = 0;
6612 mutex_lock(&ctx->uring_lock);
6613 if (!list_empty(&ctx->iopoll_list))
6614 io_do_iopoll(ctx, &nr_events, 0);
6616 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6617 !(ctx->flags & IORING_SETUP_R_DISABLED))
6618 ret = io_submit_sqes(ctx, to_submit);
6619 mutex_unlock(&ctx->uring_lock);
6622 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6623 wake_up(&ctx->sqo_sq_wait);
6628 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6630 struct io_ring_ctx *ctx;
6631 unsigned sq_thread_idle = 0;
6633 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6634 if (sq_thread_idle < ctx->sq_thread_idle)
6635 sq_thread_idle = ctx->sq_thread_idle;
6638 sqd->sq_thread_idle = sq_thread_idle;
6641 static int io_sq_thread(void *data)
6643 struct io_sq_data *sqd = data;
6644 struct io_ring_ctx *ctx;
6645 unsigned long timeout = 0;
6646 char buf[TASK_COMM_LEN];
6649 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6650 set_task_comm(current, buf);
6651 current->pf_io_worker = NULL;
6653 if (sqd->sq_cpu != -1)
6654 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6656 set_cpus_allowed_ptr(current, cpu_online_mask);
6657 current->flags |= PF_NO_SETAFFINITY;
6659 wait_for_completion(&sqd->startup);
6661 down_read(&sqd->rw_lock);
6663 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6665 bool cap_entries, sqt_spin, needs_sched;
6667 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6668 up_read(&sqd->rw_lock);
6670 down_read(&sqd->rw_lock);
6671 timeout = jiffies + sqd->sq_thread_idle;
6674 if (fatal_signal_pending(current))
6677 cap_entries = !list_is_singular(&sqd->ctx_list);
6678 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6679 const struct cred *creds = NULL;
6681 if (ctx->sq_creds != current_cred())
6682 creds = override_creds(ctx->sq_creds);
6683 ret = __io_sq_thread(ctx, cap_entries);
6685 revert_creds(creds);
6686 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6690 if (sqt_spin || !time_after(jiffies, timeout)) {
6694 timeout = jiffies + sqd->sq_thread_idle;
6699 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6700 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6701 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6702 !list_empty_careful(&ctx->iopoll_list)) {
6703 needs_sched = false;
6706 if (io_sqring_entries(ctx)) {
6707 needs_sched = false;
6712 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6713 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6714 io_ring_set_wakeup_flag(ctx);
6716 up_read(&sqd->rw_lock);
6718 down_read(&sqd->rw_lock);
6719 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6720 io_ring_clear_wakeup_flag(ctx);
6723 finish_wait(&sqd->wait, &wait);
6724 timeout = jiffies + sqd->sq_thread_idle;
6727 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6728 io_uring_cancel_sqpoll(ctx);
6729 up_read(&sqd->rw_lock);
6733 down_write(&sqd->rw_lock);
6735 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6736 io_ring_set_wakeup_flag(ctx);
6737 up_write(&sqd->rw_lock);
6738 complete(&sqd->exited);
6742 struct io_wait_queue {
6743 struct wait_queue_entry wq;
6744 struct io_ring_ctx *ctx;
6746 unsigned nr_timeouts;
6749 static inline bool io_should_wake(struct io_wait_queue *iowq)
6751 struct io_ring_ctx *ctx = iowq->ctx;
6754 * Wake up if we have enough events, or if a timeout occurred since we
6755 * started waiting. For timeouts, we always want to return to userspace,
6756 * regardless of event count.
6758 return io_cqring_events(ctx) >= iowq->to_wait ||
6759 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6762 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6763 int wake_flags, void *key)
6765 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6769 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6770 * the task, and the next invocation will do it.
6772 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6773 return autoremove_wake_function(curr, mode, wake_flags, key);
6777 static int io_run_task_work_sig(void)
6779 if (io_run_task_work())
6781 if (!signal_pending(current))
6783 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6784 return -ERESTARTSYS;
6788 /* when returns >0, the caller should retry */
6789 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6790 struct io_wait_queue *iowq,
6791 signed long *timeout)
6795 /* make sure we run task_work before checking for signals */
6796 ret = io_run_task_work_sig();
6797 if (ret || io_should_wake(iowq))
6799 /* let the caller flush overflows, retry */
6800 if (test_bit(0, &ctx->cq_check_overflow))
6803 *timeout = schedule_timeout(*timeout);
6804 return !*timeout ? -ETIME : 1;
6808 * Wait until events become available, if we don't already have some. The
6809 * application must reap them itself, as they reside on the shared cq ring.
6811 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6812 const sigset_t __user *sig, size_t sigsz,
6813 struct __kernel_timespec __user *uts)
6815 struct io_wait_queue iowq = {
6818 .func = io_wake_function,
6819 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6822 .to_wait = min_events,
6824 struct io_rings *rings = ctx->rings;
6825 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6829 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6830 if (io_cqring_events(ctx) >= min_events)
6832 if (!io_run_task_work())
6837 #ifdef CONFIG_COMPAT
6838 if (in_compat_syscall())
6839 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6843 ret = set_user_sigmask(sig, sigsz);
6850 struct timespec64 ts;
6852 if (get_timespec64(&ts, uts))
6854 timeout = timespec64_to_jiffies(&ts);
6857 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6858 trace_io_uring_cqring_wait(ctx, min_events);
6860 /* if we can't even flush overflow, don't wait for more */
6861 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6865 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6866 TASK_INTERRUPTIBLE);
6867 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6868 finish_wait(&ctx->wait, &iowq.wq);
6872 restore_saved_sigmask_unless(ret == -EINTR);
6874 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6877 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6879 #if defined(CONFIG_UNIX)
6880 if (ctx->ring_sock) {
6881 struct sock *sock = ctx->ring_sock->sk;
6882 struct sk_buff *skb;
6884 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6890 for (i = 0; i < ctx->nr_user_files; i++) {
6893 file = io_file_from_index(ctx, i);
6900 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6902 struct fixed_rsrc_data *data;
6904 data = container_of(ref, struct fixed_rsrc_data, refs);
6905 complete(&data->done);
6908 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6910 spin_lock_bh(&ctx->rsrc_ref_lock);
6913 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6915 spin_unlock_bh(&ctx->rsrc_ref_lock);
6918 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6919 struct fixed_rsrc_data *rsrc_data,
6920 struct fixed_rsrc_ref_node *ref_node)
6922 io_rsrc_ref_lock(ctx);
6923 rsrc_data->node = ref_node;
6924 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6925 io_rsrc_ref_unlock(ctx);
6926 percpu_ref_get(&rsrc_data->refs);
6929 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6931 struct fixed_rsrc_ref_node *ref_node = NULL;
6933 io_rsrc_ref_lock(ctx);
6934 ref_node = data->node;
6936 io_rsrc_ref_unlock(ctx);
6938 percpu_ref_kill(&ref_node->refs);
6941 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6942 struct io_ring_ctx *ctx,
6943 void (*rsrc_put)(struct io_ring_ctx *ctx,
6944 struct io_rsrc_put *prsrc))
6946 struct fixed_rsrc_ref_node *backup_node;
6952 data->quiesce = true;
6955 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6958 backup_node->rsrc_data = data;
6959 backup_node->rsrc_put = rsrc_put;
6961 io_sqe_rsrc_kill_node(ctx, data);
6962 percpu_ref_kill(&data->refs);
6963 flush_delayed_work(&ctx->rsrc_put_work);
6965 ret = wait_for_completion_interruptible(&data->done);
6969 percpu_ref_resurrect(&data->refs);
6970 io_sqe_rsrc_set_node(ctx, data, backup_node);
6972 reinit_completion(&data->done);
6973 mutex_unlock(&ctx->uring_lock);
6974 ret = io_run_task_work_sig();
6975 mutex_lock(&ctx->uring_lock);
6977 data->quiesce = false;
6980 destroy_fixed_rsrc_ref_node(backup_node);
6984 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
6986 struct fixed_rsrc_data *data;
6988 data = kzalloc(sizeof(*data), GFP_KERNEL);
6992 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
6993 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6998 init_completion(&data->done);
7002 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7004 percpu_ref_exit(&data->refs);
7009 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7011 struct fixed_rsrc_data *data = ctx->file_data;
7012 unsigned nr_tables, i;
7016 * percpu_ref_is_dying() is to stop parallel files unregister
7017 * Since we possibly drop uring lock later in this function to
7020 if (!data || percpu_ref_is_dying(&data->refs))
7022 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7026 __io_sqe_files_unregister(ctx);
7027 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7028 for (i = 0; i < nr_tables; i++)
7029 kfree(data->table[i].files);
7030 free_fixed_rsrc_data(data);
7031 ctx->file_data = NULL;
7032 ctx->nr_user_files = 0;
7036 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7037 __releases(&sqd->rw_lock)
7039 if (sqd->thread == current)
7041 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7042 up_write(&sqd->rw_lock);
7045 static void io_sq_thread_park(struct io_sq_data *sqd)
7046 __acquires(&sqd->rw_lock)
7048 if (sqd->thread == current)
7050 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7051 down_write(&sqd->rw_lock);
7052 /* set again for consistency, in case concurrent parks are happening */
7053 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7055 wake_up_process(sqd->thread);
7058 static void io_sq_thread_stop(struct io_sq_data *sqd)
7060 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7062 down_write(&sqd->rw_lock);
7063 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7065 wake_up_process(sqd->thread);
7066 up_write(&sqd->rw_lock);
7067 wait_for_completion(&sqd->exited);
7070 static void io_put_sq_data(struct io_sq_data *sqd)
7072 if (refcount_dec_and_test(&sqd->refs)) {
7073 io_sq_thread_stop(sqd);
7078 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7080 struct io_sq_data *sqd = ctx->sq_data;
7083 complete(&sqd->startup);
7084 io_sq_thread_park(sqd);
7085 list_del(&ctx->sqd_list);
7086 io_sqd_update_thread_idle(sqd);
7087 io_sq_thread_unpark(sqd);
7089 io_put_sq_data(sqd);
7090 ctx->sq_data = NULL;
7092 put_cred(ctx->sq_creds);
7096 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7098 struct io_ring_ctx *ctx_attach;
7099 struct io_sq_data *sqd;
7102 f = fdget(p->wq_fd);
7104 return ERR_PTR(-ENXIO);
7105 if (f.file->f_op != &io_uring_fops) {
7107 return ERR_PTR(-EINVAL);
7110 ctx_attach = f.file->private_data;
7111 sqd = ctx_attach->sq_data;
7114 return ERR_PTR(-EINVAL);
7116 if (sqd->task_tgid != current->tgid) {
7118 return ERR_PTR(-EPERM);
7121 refcount_inc(&sqd->refs);
7126 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7128 struct io_sq_data *sqd;
7130 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7131 sqd = io_attach_sq_data(p);
7134 /* fall through for EPERM case, setup new sqd/task */
7135 if (PTR_ERR(sqd) != -EPERM)
7139 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7141 return ERR_PTR(-ENOMEM);
7143 refcount_set(&sqd->refs, 1);
7144 INIT_LIST_HEAD(&sqd->ctx_list);
7145 init_rwsem(&sqd->rw_lock);
7146 init_waitqueue_head(&sqd->wait);
7147 init_completion(&sqd->startup);
7148 init_completion(&sqd->exited);
7152 #if defined(CONFIG_UNIX)
7154 * Ensure the UNIX gc is aware of our file set, so we are certain that
7155 * the io_uring can be safely unregistered on process exit, even if we have
7156 * loops in the file referencing.
7158 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7160 struct sock *sk = ctx->ring_sock->sk;
7161 struct scm_fp_list *fpl;
7162 struct sk_buff *skb;
7165 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7169 skb = alloc_skb(0, GFP_KERNEL);
7178 fpl->user = get_uid(current_user());
7179 for (i = 0; i < nr; i++) {
7180 struct file *file = io_file_from_index(ctx, i + offset);
7184 fpl->fp[nr_files] = get_file(file);
7185 unix_inflight(fpl->user, fpl->fp[nr_files]);
7190 fpl->max = SCM_MAX_FD;
7191 fpl->count = nr_files;
7192 UNIXCB(skb).fp = fpl;
7193 skb->destructor = unix_destruct_scm;
7194 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7195 skb_queue_head(&sk->sk_receive_queue, skb);
7197 for (i = 0; i < nr_files; i++)
7208 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7209 * causes regular reference counting to break down. We rely on the UNIX
7210 * garbage collection to take care of this problem for us.
7212 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7214 unsigned left, total;
7218 left = ctx->nr_user_files;
7220 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7222 ret = __io_sqe_files_scm(ctx, this_files, total);
7226 total += this_files;
7232 while (total < ctx->nr_user_files) {
7233 struct file *file = io_file_from_index(ctx, total);
7243 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7249 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7250 unsigned nr_tables, unsigned nr_files)
7254 for (i = 0; i < nr_tables; i++) {
7255 struct fixed_rsrc_table *table = &file_data->table[i];
7256 unsigned this_files;
7258 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7259 table->files = kcalloc(this_files, sizeof(struct file *),
7263 nr_files -= this_files;
7269 for (i = 0; i < nr_tables; i++) {
7270 struct fixed_rsrc_table *table = &file_data->table[i];
7271 kfree(table->files);
7276 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7278 struct file *file = prsrc->file;
7279 #if defined(CONFIG_UNIX)
7280 struct sock *sock = ctx->ring_sock->sk;
7281 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7282 struct sk_buff *skb;
7285 __skb_queue_head_init(&list);
7288 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7289 * remove this entry and rearrange the file array.
7291 skb = skb_dequeue(head);
7293 struct scm_fp_list *fp;
7295 fp = UNIXCB(skb).fp;
7296 for (i = 0; i < fp->count; i++) {
7299 if (fp->fp[i] != file)
7302 unix_notinflight(fp->user, fp->fp[i]);
7303 left = fp->count - 1 - i;
7305 memmove(&fp->fp[i], &fp->fp[i + 1],
7306 left * sizeof(struct file *));
7313 __skb_queue_tail(&list, skb);
7323 __skb_queue_tail(&list, skb);
7325 skb = skb_dequeue(head);
7328 if (skb_peek(&list)) {
7329 spin_lock_irq(&head->lock);
7330 while ((skb = __skb_dequeue(&list)) != NULL)
7331 __skb_queue_tail(head, skb);
7332 spin_unlock_irq(&head->lock);
7339 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7341 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7342 struct io_ring_ctx *ctx = rsrc_data->ctx;
7343 struct io_rsrc_put *prsrc, *tmp;
7345 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7346 list_del(&prsrc->list);
7347 ref_node->rsrc_put(ctx, prsrc);
7351 percpu_ref_exit(&ref_node->refs);
7353 percpu_ref_put(&rsrc_data->refs);
7356 static void io_rsrc_put_work(struct work_struct *work)
7358 struct io_ring_ctx *ctx;
7359 struct llist_node *node;
7361 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7362 node = llist_del_all(&ctx->rsrc_put_llist);
7365 struct fixed_rsrc_ref_node *ref_node;
7366 struct llist_node *next = node->next;
7368 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7369 __io_rsrc_put_work(ref_node);
7374 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7377 struct fixed_rsrc_table *table;
7379 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7380 return &table->files[i & IORING_FILE_TABLE_MASK];
7383 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7385 struct fixed_rsrc_ref_node *ref_node;
7386 struct fixed_rsrc_data *data;
7387 struct io_ring_ctx *ctx;
7388 bool first_add = false;
7391 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7392 data = ref_node->rsrc_data;
7395 io_rsrc_ref_lock(ctx);
7396 ref_node->done = true;
7398 while (!list_empty(&ctx->rsrc_ref_list)) {
7399 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7400 struct fixed_rsrc_ref_node, node);
7401 /* recycle ref nodes in order */
7402 if (!ref_node->done)
7404 list_del(&ref_node->node);
7405 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7407 io_rsrc_ref_unlock(ctx);
7409 if (percpu_ref_is_dying(&data->refs))
7413 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7415 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7418 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7419 struct io_ring_ctx *ctx)
7421 struct fixed_rsrc_ref_node *ref_node;
7423 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7427 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7432 INIT_LIST_HEAD(&ref_node->node);
7433 INIT_LIST_HEAD(&ref_node->rsrc_list);
7434 ref_node->done = false;
7438 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7439 struct fixed_rsrc_ref_node *ref_node)
7441 ref_node->rsrc_data = ctx->file_data;
7442 ref_node->rsrc_put = io_ring_file_put;
7445 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7447 percpu_ref_exit(&ref_node->refs);
7452 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7455 __s32 __user *fds = (__s32 __user *) arg;
7456 unsigned nr_tables, i;
7458 int fd, ret = -ENOMEM;
7459 struct fixed_rsrc_ref_node *ref_node;
7460 struct fixed_rsrc_data *file_data;
7466 if (nr_args > IORING_MAX_FIXED_FILES)
7469 file_data = alloc_fixed_rsrc_data(ctx);
7472 ctx->file_data = file_data;
7474 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7475 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7477 if (!file_data->table)
7480 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7483 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7484 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7488 /* allow sparse sets */
7498 * Don't allow io_uring instances to be registered. If UNIX
7499 * isn't enabled, then this causes a reference cycle and this
7500 * instance can never get freed. If UNIX is enabled we'll
7501 * handle it just fine, but there's still no point in allowing
7502 * a ring fd as it doesn't support regular read/write anyway.
7504 if (file->f_op == &io_uring_fops) {
7508 *io_fixed_file_slot(file_data, i) = file;
7511 ret = io_sqe_files_scm(ctx);
7513 io_sqe_files_unregister(ctx);
7517 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7519 io_sqe_files_unregister(ctx);
7522 init_fixed_file_ref_node(ctx, ref_node);
7524 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7527 for (i = 0; i < ctx->nr_user_files; i++) {
7528 file = io_file_from_index(ctx, i);
7532 for (i = 0; i < nr_tables; i++)
7533 kfree(file_data->table[i].files);
7534 ctx->nr_user_files = 0;
7536 free_fixed_rsrc_data(ctx->file_data);
7537 ctx->file_data = NULL;
7541 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7544 #if defined(CONFIG_UNIX)
7545 struct sock *sock = ctx->ring_sock->sk;
7546 struct sk_buff_head *head = &sock->sk_receive_queue;
7547 struct sk_buff *skb;
7550 * See if we can merge this file into an existing skb SCM_RIGHTS
7551 * file set. If there's no room, fall back to allocating a new skb
7552 * and filling it in.
7554 spin_lock_irq(&head->lock);
7555 skb = skb_peek(head);
7557 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7559 if (fpl->count < SCM_MAX_FD) {
7560 __skb_unlink(skb, head);
7561 spin_unlock_irq(&head->lock);
7562 fpl->fp[fpl->count] = get_file(file);
7563 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7565 spin_lock_irq(&head->lock);
7566 __skb_queue_head(head, skb);
7571 spin_unlock_irq(&head->lock);
7578 return __io_sqe_files_scm(ctx, 1, index);
7584 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7586 struct io_rsrc_put *prsrc;
7587 struct fixed_rsrc_ref_node *ref_node = data->node;
7589 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7594 list_add(&prsrc->list, &ref_node->rsrc_list);
7599 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7602 return io_queue_rsrc_removal(data, (void *)file);
7605 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7606 struct io_uring_rsrc_update *up,
7609 struct fixed_rsrc_data *data = ctx->file_data;
7610 struct fixed_rsrc_ref_node *ref_node;
7611 struct file *file, **file_slot;
7615 bool needs_switch = false;
7617 if (check_add_overflow(up->offset, nr_args, &done))
7619 if (done > ctx->nr_user_files)
7622 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7625 init_fixed_file_ref_node(ctx, ref_node);
7627 fds = u64_to_user_ptr(up->data);
7628 for (done = 0; done < nr_args; done++) {
7630 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7634 if (fd == IORING_REGISTER_FILES_SKIP)
7637 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7638 file_slot = io_fixed_file_slot(ctx->file_data, i);
7641 err = io_queue_file_removal(data, *file_slot);
7645 needs_switch = true;
7654 * Don't allow io_uring instances to be registered. If
7655 * UNIX isn't enabled, then this causes a reference
7656 * cycle and this instance can never get freed. If UNIX
7657 * is enabled we'll handle it just fine, but there's
7658 * still no point in allowing a ring fd as it doesn't
7659 * support regular read/write anyway.
7661 if (file->f_op == &io_uring_fops) {
7667 err = io_sqe_file_register(ctx, file, i);
7677 percpu_ref_kill(&data->node->refs);
7678 io_sqe_rsrc_set_node(ctx, data, ref_node);
7680 destroy_fixed_rsrc_ref_node(ref_node);
7682 return done ? done : err;
7685 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7688 struct io_uring_rsrc_update up;
7690 if (!ctx->file_data)
7694 if (copy_from_user(&up, arg, sizeof(up)))
7699 return __io_sqe_files_update(ctx, &up, nr_args);
7702 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7704 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7706 req = io_put_req_find_next(req);
7707 return req ? &req->work : NULL;
7710 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7712 struct io_wq_hash *hash;
7713 struct io_wq_data data;
7714 unsigned int concurrency;
7716 hash = ctx->hash_map;
7718 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7720 return ERR_PTR(-ENOMEM);
7721 refcount_set(&hash->refs, 1);
7722 init_waitqueue_head(&hash->wait);
7723 ctx->hash_map = hash;
7727 data.free_work = io_free_work;
7728 data.do_work = io_wq_submit_work;
7730 /* Do QD, or 4 * CPUS, whatever is smallest */
7731 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7733 return io_wq_create(concurrency, &data);
7736 static int io_uring_alloc_task_context(struct task_struct *task,
7737 struct io_ring_ctx *ctx)
7739 struct io_uring_task *tctx;
7742 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7743 if (unlikely(!tctx))
7746 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7747 if (unlikely(ret)) {
7752 tctx->io_wq = io_init_wq_offload(ctx);
7753 if (IS_ERR(tctx->io_wq)) {
7754 ret = PTR_ERR(tctx->io_wq);
7755 percpu_counter_destroy(&tctx->inflight);
7761 init_waitqueue_head(&tctx->wait);
7763 atomic_set(&tctx->in_idle, 0);
7764 tctx->sqpoll = false;
7765 task->io_uring = tctx;
7766 spin_lock_init(&tctx->task_lock);
7767 INIT_WQ_LIST(&tctx->task_list);
7768 tctx->task_state = 0;
7769 init_task_work(&tctx->task_work, tctx_task_work);
7773 void __io_uring_free(struct task_struct *tsk)
7775 struct io_uring_task *tctx = tsk->io_uring;
7777 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7778 WARN_ON_ONCE(tctx->io_wq);
7780 percpu_counter_destroy(&tctx->inflight);
7782 tsk->io_uring = NULL;
7785 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7786 struct io_uring_params *p)
7790 /* Retain compatibility with failing for an invalid attach attempt */
7791 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7792 IORING_SETUP_ATTACH_WQ) {
7795 f = fdget(p->wq_fd);
7798 if (f.file->f_op != &io_uring_fops) {
7804 if (ctx->flags & IORING_SETUP_SQPOLL) {
7805 struct task_struct *tsk;
7806 struct io_sq_data *sqd;
7809 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7812 sqd = io_get_sq_data(p);
7818 ctx->sq_creds = get_current_cred();
7820 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7821 if (!ctx->sq_thread_idle)
7822 ctx->sq_thread_idle = HZ;
7824 io_sq_thread_park(sqd);
7825 list_add(&ctx->sqd_list, &sqd->ctx_list);
7826 io_sqd_update_thread_idle(sqd);
7827 io_sq_thread_unpark(sqd);
7832 if (p->flags & IORING_SETUP_SQ_AFF) {
7833 int cpu = p->sq_thread_cpu;
7836 if (cpu >= nr_cpu_ids)
7838 if (!cpu_online(cpu))
7846 sqd->task_pid = current->pid;
7847 sqd->task_tgid = current->tgid;
7848 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7855 ret = io_uring_alloc_task_context(tsk, ctx);
7856 wake_up_new_task(tsk);
7859 complete(&sqd->startup);
7860 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7861 /* Can't have SQ_AFF without SQPOLL */
7868 io_sq_thread_finish(ctx);
7871 complete(&ctx->sq_data->exited);
7875 static inline void __io_unaccount_mem(struct user_struct *user,
7876 unsigned long nr_pages)
7878 atomic_long_sub(nr_pages, &user->locked_vm);
7881 static inline int __io_account_mem(struct user_struct *user,
7882 unsigned long nr_pages)
7884 unsigned long page_limit, cur_pages, new_pages;
7886 /* Don't allow more pages than we can safely lock */
7887 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7890 cur_pages = atomic_long_read(&user->locked_vm);
7891 new_pages = cur_pages + nr_pages;
7892 if (new_pages > page_limit)
7894 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7895 new_pages) != cur_pages);
7900 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7903 __io_unaccount_mem(ctx->user, nr_pages);
7905 if (ctx->mm_account)
7906 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7909 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7914 ret = __io_account_mem(ctx->user, nr_pages);
7919 if (ctx->mm_account)
7920 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7925 static void io_mem_free(void *ptr)
7932 page = virt_to_head_page(ptr);
7933 if (put_page_testzero(page))
7934 free_compound_page(page);
7937 static void *io_mem_alloc(size_t size)
7939 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7940 __GFP_NORETRY | __GFP_ACCOUNT;
7942 return (void *) __get_free_pages(gfp_flags, get_order(size));
7945 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7948 struct io_rings *rings;
7949 size_t off, sq_array_size;
7951 off = struct_size(rings, cqes, cq_entries);
7952 if (off == SIZE_MAX)
7956 off = ALIGN(off, SMP_CACHE_BYTES);
7964 sq_array_size = array_size(sizeof(u32), sq_entries);
7965 if (sq_array_size == SIZE_MAX)
7968 if (check_add_overflow(off, sq_array_size, &off))
7974 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
7978 if (!ctx->user_bufs)
7981 for (i = 0; i < ctx->nr_user_bufs; i++) {
7982 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7984 for (j = 0; j < imu->nr_bvecs; j++)
7985 unpin_user_page(imu->bvec[j].bv_page);
7987 if (imu->acct_pages)
7988 io_unaccount_mem(ctx, imu->acct_pages);
7993 kfree(ctx->user_bufs);
7994 ctx->user_bufs = NULL;
7995 ctx->nr_user_bufs = 0;
7999 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8000 void __user *arg, unsigned index)
8002 struct iovec __user *src;
8004 #ifdef CONFIG_COMPAT
8006 struct compat_iovec __user *ciovs;
8007 struct compat_iovec ciov;
8009 ciovs = (struct compat_iovec __user *) arg;
8010 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8013 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8014 dst->iov_len = ciov.iov_len;
8018 src = (struct iovec __user *) arg;
8019 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8025 * Not super efficient, but this is just a registration time. And we do cache
8026 * the last compound head, so generally we'll only do a full search if we don't
8029 * We check if the given compound head page has already been accounted, to
8030 * avoid double accounting it. This allows us to account the full size of the
8031 * page, not just the constituent pages of a huge page.
8033 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8034 int nr_pages, struct page *hpage)
8038 /* check current page array */
8039 for (i = 0; i < nr_pages; i++) {
8040 if (!PageCompound(pages[i]))
8042 if (compound_head(pages[i]) == hpage)
8046 /* check previously registered pages */
8047 for (i = 0; i < ctx->nr_user_bufs; i++) {
8048 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8050 for (j = 0; j < imu->nr_bvecs; j++) {
8051 if (!PageCompound(imu->bvec[j].bv_page))
8053 if (compound_head(imu->bvec[j].bv_page) == hpage)
8061 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8062 int nr_pages, struct io_mapped_ubuf *imu,
8063 struct page **last_hpage)
8067 for (i = 0; i < nr_pages; i++) {
8068 if (!PageCompound(pages[i])) {
8073 hpage = compound_head(pages[i]);
8074 if (hpage == *last_hpage)
8076 *last_hpage = hpage;
8077 if (headpage_already_acct(ctx, pages, i, hpage))
8079 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8083 if (!imu->acct_pages)
8086 ret = io_account_mem(ctx, imu->acct_pages);
8088 imu->acct_pages = 0;
8092 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8093 struct io_mapped_ubuf *imu,
8094 struct page **last_hpage)
8096 struct vm_area_struct **vmas = NULL;
8097 struct page **pages = NULL;
8098 unsigned long off, start, end, ubuf;
8100 int ret, pret, nr_pages, i;
8102 ubuf = (unsigned long) iov->iov_base;
8103 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8104 start = ubuf >> PAGE_SHIFT;
8105 nr_pages = end - start;
8109 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8113 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8118 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8124 mmap_read_lock(current->mm);
8125 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8127 if (pret == nr_pages) {
8128 /* don't support file backed memory */
8129 for (i = 0; i < nr_pages; i++) {
8130 struct vm_area_struct *vma = vmas[i];
8133 !is_file_hugepages(vma->vm_file)) {
8139 ret = pret < 0 ? pret : -EFAULT;
8141 mmap_read_unlock(current->mm);
8144 * if we did partial map, or found file backed vmas,
8145 * release any pages we did get
8148 unpin_user_pages(pages, pret);
8153 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8155 unpin_user_pages(pages, pret);
8160 off = ubuf & ~PAGE_MASK;
8161 size = iov->iov_len;
8162 for (i = 0; i < nr_pages; i++) {
8165 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8166 imu->bvec[i].bv_page = pages[i];
8167 imu->bvec[i].bv_len = vec_len;
8168 imu->bvec[i].bv_offset = off;
8172 /* store original address for later verification */
8174 imu->len = iov->iov_len;
8175 imu->nr_bvecs = nr_pages;
8183 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8187 if (!nr_args || nr_args > UIO_MAXIOV)
8190 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8192 if (!ctx->user_bufs)
8198 static int io_buffer_validate(struct iovec *iov)
8201 * Don't impose further limits on the size and buffer
8202 * constraints here, we'll -EINVAL later when IO is
8203 * submitted if they are wrong.
8205 if (!iov->iov_base || !iov->iov_len)
8208 /* arbitrary limit, but we need something */
8209 if (iov->iov_len > SZ_1G)
8215 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8216 unsigned int nr_args)
8220 struct page *last_hpage = NULL;
8222 ret = io_buffers_map_alloc(ctx, nr_args);
8226 for (i = 0; i < nr_args; i++) {
8227 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8229 ret = io_copy_iov(ctx, &iov, arg, i);
8233 ret = io_buffer_validate(&iov);
8237 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8241 ctx->nr_user_bufs++;
8245 io_sqe_buffers_unregister(ctx);
8250 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8252 __s32 __user *fds = arg;
8258 if (copy_from_user(&fd, fds, sizeof(*fds)))
8261 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8262 if (IS_ERR(ctx->cq_ev_fd)) {
8263 int ret = PTR_ERR(ctx->cq_ev_fd);
8264 ctx->cq_ev_fd = NULL;
8271 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8273 if (ctx->cq_ev_fd) {
8274 eventfd_ctx_put(ctx->cq_ev_fd);
8275 ctx->cq_ev_fd = NULL;
8282 static int __io_destroy_buffers(int id, void *p, void *data)
8284 struct io_ring_ctx *ctx = data;
8285 struct io_buffer *buf = p;
8287 __io_remove_buffers(ctx, buf, id, -1U);
8291 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8293 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8294 idr_destroy(&ctx->io_buffer_idr);
8297 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8299 struct io_kiocb *req, *nxt;
8301 list_for_each_entry_safe(req, nxt, list, compl.list) {
8302 if (tsk && req->task != tsk)
8304 list_del(&req->compl.list);
8305 kmem_cache_free(req_cachep, req);
8309 static void io_req_caches_free(struct io_ring_ctx *ctx)
8311 struct io_submit_state *submit_state = &ctx->submit_state;
8312 struct io_comp_state *cs = &ctx->submit_state.comp;
8314 mutex_lock(&ctx->uring_lock);
8316 if (submit_state->free_reqs) {
8317 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8318 submit_state->reqs);
8319 submit_state->free_reqs = 0;
8322 spin_lock_irq(&ctx->completion_lock);
8323 list_splice_init(&cs->locked_free_list, &cs->free_list);
8324 cs->locked_free_nr = 0;
8325 spin_unlock_irq(&ctx->completion_lock);
8327 io_req_cache_free(&cs->free_list, NULL);
8329 mutex_unlock(&ctx->uring_lock);
8332 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8335 * Some may use context even when all refs and requests have been put,
8336 * and they are free to do so while still holding uring_lock, see
8337 * __io_req_task_submit(). Wait for them to finish.
8339 mutex_lock(&ctx->uring_lock);
8340 mutex_unlock(&ctx->uring_lock);
8342 io_sq_thread_finish(ctx);
8343 io_sqe_buffers_unregister(ctx);
8345 if (ctx->mm_account) {
8346 mmdrop(ctx->mm_account);
8347 ctx->mm_account = NULL;
8350 mutex_lock(&ctx->uring_lock);
8351 io_sqe_files_unregister(ctx);
8352 mutex_unlock(&ctx->uring_lock);
8353 io_eventfd_unregister(ctx);
8354 io_destroy_buffers(ctx);
8356 #if defined(CONFIG_UNIX)
8357 if (ctx->ring_sock) {
8358 ctx->ring_sock->file = NULL; /* so that iput() is called */
8359 sock_release(ctx->ring_sock);
8363 io_mem_free(ctx->rings);
8364 io_mem_free(ctx->sq_sqes);
8366 percpu_ref_exit(&ctx->refs);
8367 free_uid(ctx->user);
8368 io_req_caches_free(ctx);
8370 io_wq_put_hash(ctx->hash_map);
8371 kfree(ctx->cancel_hash);
8375 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8377 struct io_ring_ctx *ctx = file->private_data;
8380 poll_wait(file, &ctx->cq_wait, wait);
8382 * synchronizes with barrier from wq_has_sleeper call in
8386 if (!io_sqring_full(ctx))
8387 mask |= EPOLLOUT | EPOLLWRNORM;
8390 * Don't flush cqring overflow list here, just do a simple check.
8391 * Otherwise there could possible be ABBA deadlock:
8394 * lock(&ctx->uring_lock);
8396 * lock(&ctx->uring_lock);
8399 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8400 * pushs them to do the flush.
8402 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8403 mask |= EPOLLIN | EPOLLRDNORM;
8408 static int io_uring_fasync(int fd, struct file *file, int on)
8410 struct io_ring_ctx *ctx = file->private_data;
8412 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8415 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8417 const struct cred *creds;
8419 creds = xa_erase(&ctx->personalities, id);
8428 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8430 struct callback_head *work, *next;
8431 bool executed = false;
8434 work = xchg(&ctx->exit_task_work, NULL);
8450 struct io_tctx_exit {
8451 struct callback_head task_work;
8452 struct completion completion;
8453 struct io_ring_ctx *ctx;
8456 static void io_tctx_exit_cb(struct callback_head *cb)
8458 struct io_uring_task *tctx = current->io_uring;
8459 struct io_tctx_exit *work;
8461 work = container_of(cb, struct io_tctx_exit, task_work);
8463 * When @in_idle, we're in cancellation and it's racy to remove the
8464 * node. It'll be removed by the end of cancellation, just ignore it.
8466 if (!atomic_read(&tctx->in_idle))
8467 io_uring_del_task_file((unsigned long)work->ctx);
8468 complete(&work->completion);
8471 static void io_ring_exit_work(struct work_struct *work)
8473 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8474 unsigned long timeout = jiffies + HZ * 60 * 5;
8475 struct io_tctx_exit exit;
8476 struct io_tctx_node *node;
8480 * If we're doing polled IO and end up having requests being
8481 * submitted async (out-of-line), then completions can come in while
8482 * we're waiting for refs to drop. We need to reap these manually,
8483 * as nobody else will be looking for them.
8486 io_uring_try_cancel_requests(ctx, NULL, NULL);
8488 WARN_ON_ONCE(time_after(jiffies, timeout));
8489 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8491 mutex_lock(&ctx->uring_lock);
8492 while (!list_empty(&ctx->tctx_list)) {
8493 WARN_ON_ONCE(time_after(jiffies, timeout));
8495 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8498 init_completion(&exit.completion);
8499 init_task_work(&exit.task_work, io_tctx_exit_cb);
8500 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8501 if (WARN_ON_ONCE(ret))
8503 wake_up_process(node->task);
8505 mutex_unlock(&ctx->uring_lock);
8506 wait_for_completion(&exit.completion);
8508 mutex_lock(&ctx->uring_lock);
8510 mutex_unlock(&ctx->uring_lock);
8512 io_ring_ctx_free(ctx);
8515 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8517 unsigned long index;
8518 struct creds *creds;
8520 mutex_lock(&ctx->uring_lock);
8521 percpu_ref_kill(&ctx->refs);
8522 /* if force is set, the ring is going away. always drop after that */
8523 ctx->cq_overflow_flushed = 1;
8525 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8526 xa_for_each(&ctx->personalities, index, creds)
8527 io_unregister_personality(ctx, index);
8528 mutex_unlock(&ctx->uring_lock);
8530 io_kill_timeouts(ctx, NULL, NULL);
8531 io_poll_remove_all(ctx, NULL, NULL);
8533 /* if we failed setting up the ctx, we might not have any rings */
8534 io_iopoll_try_reap_events(ctx);
8536 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8538 * Use system_unbound_wq to avoid spawning tons of event kworkers
8539 * if we're exiting a ton of rings at the same time. It just adds
8540 * noise and overhead, there's no discernable change in runtime
8541 * over using system_wq.
8543 queue_work(system_unbound_wq, &ctx->exit_work);
8546 static int io_uring_release(struct inode *inode, struct file *file)
8548 struct io_ring_ctx *ctx = file->private_data;
8550 file->private_data = NULL;
8551 io_ring_ctx_wait_and_kill(ctx);
8555 struct io_task_cancel {
8556 struct task_struct *task;
8557 struct files_struct *files;
8560 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8562 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8563 struct io_task_cancel *cancel = data;
8566 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8567 unsigned long flags;
8568 struct io_ring_ctx *ctx = req->ctx;
8570 /* protect against races with linked timeouts */
8571 spin_lock_irqsave(&ctx->completion_lock, flags);
8572 ret = io_match_task(req, cancel->task, cancel->files);
8573 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8575 ret = io_match_task(req, cancel->task, cancel->files);
8580 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8581 struct task_struct *task,
8582 struct files_struct *files)
8584 struct io_defer_entry *de = NULL;
8587 spin_lock_irq(&ctx->completion_lock);
8588 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8589 if (io_match_task(de->req, task, files)) {
8590 list_cut_position(&list, &ctx->defer_list, &de->list);
8594 spin_unlock_irq(&ctx->completion_lock);
8596 while (!list_empty(&list)) {
8597 de = list_first_entry(&list, struct io_defer_entry, list);
8598 list_del_init(&de->list);
8599 req_set_fail_links(de->req);
8600 io_put_req(de->req);
8601 io_req_complete(de->req, -ECANCELED);
8606 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8608 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8610 return req->ctx == data;
8613 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8615 struct io_tctx_node *node;
8616 enum io_wq_cancel cret;
8619 mutex_lock(&ctx->uring_lock);
8620 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8621 struct io_uring_task *tctx = node->task->io_uring;
8624 * io_wq will stay alive while we hold uring_lock, because it's
8625 * killed after ctx nodes, which requires to take the lock.
8627 if (!tctx || !tctx->io_wq)
8629 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8630 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8632 mutex_unlock(&ctx->uring_lock);
8637 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8638 struct task_struct *task,
8639 struct files_struct *files)
8641 struct io_task_cancel cancel = { .task = task, .files = files, };
8642 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8645 enum io_wq_cancel cret;
8649 ret |= io_uring_try_cancel_iowq(ctx);
8650 } else if (tctx && tctx->io_wq) {
8652 * Cancels requests of all rings, not only @ctx, but
8653 * it's fine as the task is in exit/exec.
8655 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8657 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8660 /* SQPOLL thread does its own polling */
8661 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8662 (ctx->sq_data && ctx->sq_data->thread == current)) {
8663 while (!list_empty_careful(&ctx->iopoll_list)) {
8664 io_iopoll_try_reap_events(ctx);
8669 ret |= io_poll_remove_all(ctx, task, files);
8670 ret |= io_kill_timeouts(ctx, task, files);
8671 ret |= io_run_task_work();
8672 ret |= io_run_ctx_fallback(ctx);
8673 io_cqring_overflow_flush(ctx, true, task, files);
8680 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8681 struct task_struct *task,
8682 struct files_struct *files)
8684 struct io_kiocb *req;
8687 spin_lock_irq(&ctx->inflight_lock);
8688 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8689 cnt += io_match_task(req, task, files);
8690 spin_unlock_irq(&ctx->inflight_lock);
8694 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8695 struct task_struct *task,
8696 struct files_struct *files)
8698 while (!list_empty_careful(&ctx->inflight_list)) {
8702 inflight = io_uring_count_inflight(ctx, task, files);
8706 io_uring_try_cancel_requests(ctx, task, files);
8709 io_sq_thread_unpark(ctx->sq_data);
8710 prepare_to_wait(&task->io_uring->wait, &wait,
8711 TASK_UNINTERRUPTIBLE);
8712 if (inflight == io_uring_count_inflight(ctx, task, files))
8714 finish_wait(&task->io_uring->wait, &wait);
8716 io_sq_thread_park(ctx->sq_data);
8721 * We need to iteratively cancel requests, in case a request has dependent
8722 * hard links. These persist even for failure of cancelations, hence keep
8723 * looping until none are found.
8725 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8726 struct files_struct *files)
8728 struct task_struct *task = current;
8730 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8731 io_sq_thread_park(ctx->sq_data);
8732 task = ctx->sq_data->thread;
8734 atomic_inc(&task->io_uring->in_idle);
8737 io_cancel_defer_files(ctx, task, files);
8739 io_uring_cancel_files(ctx, task, files);
8741 io_uring_try_cancel_requests(ctx, task, NULL);
8744 atomic_dec(&task->io_uring->in_idle);
8746 io_sq_thread_unpark(ctx->sq_data);
8750 * Note that this task has used io_uring. We use it for cancelation purposes.
8752 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8754 struct io_uring_task *tctx = current->io_uring;
8755 struct io_tctx_node *node;
8758 if (unlikely(!tctx)) {
8759 ret = io_uring_alloc_task_context(current, ctx);
8762 tctx = current->io_uring;
8764 if (tctx->last != ctx) {
8765 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8768 node = kmalloc(sizeof(*node), GFP_KERNEL);
8772 node->task = current;
8774 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8781 mutex_lock(&ctx->uring_lock);
8782 list_add(&node->ctx_node, &ctx->tctx_list);
8783 mutex_unlock(&ctx->uring_lock);
8789 * This is race safe in that the task itself is doing this, hence it
8790 * cannot be going through the exit/cancel paths at the same time.
8791 * This cannot be modified while exit/cancel is running.
8793 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8794 tctx->sqpoll = true;
8800 * Remove this io_uring_file -> task mapping.
8802 static void io_uring_del_task_file(unsigned long index)
8804 struct io_uring_task *tctx = current->io_uring;
8805 struct io_tctx_node *node;
8809 node = xa_erase(&tctx->xa, index);
8813 WARN_ON_ONCE(current != node->task);
8814 WARN_ON_ONCE(list_empty(&node->ctx_node));
8816 mutex_lock(&node->ctx->uring_lock);
8817 list_del(&node->ctx_node);
8818 mutex_unlock(&node->ctx->uring_lock);
8820 if (tctx->last == node->ctx)
8825 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8827 struct io_tctx_node *node;
8828 unsigned long index;
8830 xa_for_each(&tctx->xa, index, node)
8831 io_uring_del_task_file(index);
8833 io_wq_put_and_exit(tctx->io_wq);
8838 void __io_uring_files_cancel(struct files_struct *files)
8840 struct io_uring_task *tctx = current->io_uring;
8841 struct io_tctx_node *node;
8842 unsigned long index;
8844 /* make sure overflow events are dropped */
8845 atomic_inc(&tctx->in_idle);
8846 xa_for_each(&tctx->xa, index, node)
8847 io_uring_cancel_task_requests(node->ctx, files);
8848 atomic_dec(&tctx->in_idle);
8851 io_uring_clean_tctx(tctx);
8854 static s64 tctx_inflight(struct io_uring_task *tctx)
8856 return percpu_counter_sum(&tctx->inflight);
8859 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8861 struct io_sq_data *sqd = ctx->sq_data;
8862 struct io_uring_task *tctx;
8868 io_sq_thread_park(sqd);
8869 if (!sqd->thread || !sqd->thread->io_uring) {
8870 io_sq_thread_unpark(sqd);
8873 tctx = ctx->sq_data->thread->io_uring;
8874 atomic_inc(&tctx->in_idle);
8876 /* read completions before cancelations */
8877 inflight = tctx_inflight(tctx);
8880 io_uring_cancel_task_requests(ctx, NULL);
8882 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8884 * If we've seen completions, retry without waiting. This
8885 * avoids a race where a completion comes in before we did
8886 * prepare_to_wait().
8888 if (inflight == tctx_inflight(tctx))
8890 finish_wait(&tctx->wait, &wait);
8892 atomic_dec(&tctx->in_idle);
8893 io_sq_thread_unpark(sqd);
8897 * Find any io_uring fd that this task has registered or done IO on, and cancel
8900 void __io_uring_task_cancel(void)
8902 struct io_uring_task *tctx = current->io_uring;
8906 /* make sure overflow events are dropped */
8907 atomic_inc(&tctx->in_idle);
8910 struct io_tctx_node *node;
8911 unsigned long index;
8913 xa_for_each(&tctx->xa, index, node)
8914 io_uring_cancel_sqpoll(node->ctx);
8918 /* read completions before cancelations */
8919 inflight = tctx_inflight(tctx);
8922 __io_uring_files_cancel(NULL);
8924 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8927 * If we've seen completions, retry without waiting. This
8928 * avoids a race where a completion comes in before we did
8929 * prepare_to_wait().
8931 if (inflight == tctx_inflight(tctx))
8933 finish_wait(&tctx->wait, &wait);
8936 atomic_dec(&tctx->in_idle);
8938 io_uring_clean_tctx(tctx);
8939 /* all current's requests should be gone, we can kill tctx */
8940 __io_uring_free(current);
8943 static void *io_uring_validate_mmap_request(struct file *file,
8944 loff_t pgoff, size_t sz)
8946 struct io_ring_ctx *ctx = file->private_data;
8947 loff_t offset = pgoff << PAGE_SHIFT;
8952 case IORING_OFF_SQ_RING:
8953 case IORING_OFF_CQ_RING:
8956 case IORING_OFF_SQES:
8960 return ERR_PTR(-EINVAL);
8963 page = virt_to_head_page(ptr);
8964 if (sz > page_size(page))
8965 return ERR_PTR(-EINVAL);
8972 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8974 size_t sz = vma->vm_end - vma->vm_start;
8978 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8980 return PTR_ERR(ptr);
8982 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8983 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8986 #else /* !CONFIG_MMU */
8988 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8990 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8993 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8995 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8998 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8999 unsigned long addr, unsigned long len,
9000 unsigned long pgoff, unsigned long flags)
9004 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9006 return PTR_ERR(ptr);
9008 return (unsigned long) ptr;
9011 #endif /* !CONFIG_MMU */
9013 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9018 if (!io_sqring_full(ctx))
9020 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9022 if (!io_sqring_full(ctx))
9025 } while (!signal_pending(current));
9027 finish_wait(&ctx->sqo_sq_wait, &wait);
9031 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9032 struct __kernel_timespec __user **ts,
9033 const sigset_t __user **sig)
9035 struct io_uring_getevents_arg arg;
9038 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9039 * is just a pointer to the sigset_t.
9041 if (!(flags & IORING_ENTER_EXT_ARG)) {
9042 *sig = (const sigset_t __user *) argp;
9048 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9049 * timespec and sigset_t pointers if good.
9051 if (*argsz != sizeof(arg))
9053 if (copy_from_user(&arg, argp, sizeof(arg)))
9055 *sig = u64_to_user_ptr(arg.sigmask);
9056 *argsz = arg.sigmask_sz;
9057 *ts = u64_to_user_ptr(arg.ts);
9061 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9062 u32, min_complete, u32, flags, const void __user *, argp,
9065 struct io_ring_ctx *ctx;
9072 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9073 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9081 if (f.file->f_op != &io_uring_fops)
9085 ctx = f.file->private_data;
9086 if (!percpu_ref_tryget(&ctx->refs))
9090 if (ctx->flags & IORING_SETUP_R_DISABLED)
9094 * For SQ polling, the thread will do all submissions and completions.
9095 * Just return the requested submit count, and wake the thread if
9099 if (ctx->flags & IORING_SETUP_SQPOLL) {
9100 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9103 if (unlikely(ctx->sq_data->thread == NULL)) {
9106 if (flags & IORING_ENTER_SQ_WAKEUP)
9107 wake_up(&ctx->sq_data->wait);
9108 if (flags & IORING_ENTER_SQ_WAIT) {
9109 ret = io_sqpoll_wait_sq(ctx);
9113 submitted = to_submit;
9114 } else if (to_submit) {
9115 ret = io_uring_add_task_file(ctx);
9118 mutex_lock(&ctx->uring_lock);
9119 submitted = io_submit_sqes(ctx, to_submit);
9120 mutex_unlock(&ctx->uring_lock);
9122 if (submitted != to_submit)
9125 if (flags & IORING_ENTER_GETEVENTS) {
9126 const sigset_t __user *sig;
9127 struct __kernel_timespec __user *ts;
9129 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9133 min_complete = min(min_complete, ctx->cq_entries);
9136 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9137 * space applications don't need to do io completion events
9138 * polling again, they can rely on io_sq_thread to do polling
9139 * work, which can reduce cpu usage and uring_lock contention.
9141 if (ctx->flags & IORING_SETUP_IOPOLL &&
9142 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9143 ret = io_iopoll_check(ctx, min_complete);
9145 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9150 percpu_ref_put(&ctx->refs);
9153 return submitted ? submitted : ret;
9156 #ifdef CONFIG_PROC_FS
9157 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9158 const struct cred *cred)
9160 struct user_namespace *uns = seq_user_ns(m);
9161 struct group_info *gi;
9166 seq_printf(m, "%5d\n", id);
9167 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9168 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9169 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9170 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9171 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9172 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9173 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9174 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9175 seq_puts(m, "\n\tGroups:\t");
9176 gi = cred->group_info;
9177 for (g = 0; g < gi->ngroups; g++) {
9178 seq_put_decimal_ull(m, g ? " " : "",
9179 from_kgid_munged(uns, gi->gid[g]));
9181 seq_puts(m, "\n\tCapEff:\t");
9182 cap = cred->cap_effective;
9183 CAP_FOR_EACH_U32(__capi)
9184 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9189 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9191 struct io_sq_data *sq = NULL;
9196 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9197 * since fdinfo case grabs it in the opposite direction of normal use
9198 * cases. If we fail to get the lock, we just don't iterate any
9199 * structures that could be going away outside the io_uring mutex.
9201 has_lock = mutex_trylock(&ctx->uring_lock);
9203 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9209 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9210 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9211 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9212 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9213 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9216 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9218 seq_printf(m, "%5u: <none>\n", i);
9220 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9221 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9222 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9224 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9225 (unsigned int) buf->len);
9227 if (has_lock && !xa_empty(&ctx->personalities)) {
9228 unsigned long index;
9229 const struct cred *cred;
9231 seq_printf(m, "Personalities:\n");
9232 xa_for_each(&ctx->personalities, index, cred)
9233 io_uring_show_cred(m, index, cred);
9235 seq_printf(m, "PollList:\n");
9236 spin_lock_irq(&ctx->completion_lock);
9237 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9238 struct hlist_head *list = &ctx->cancel_hash[i];
9239 struct io_kiocb *req;
9241 hlist_for_each_entry(req, list, hash_node)
9242 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9243 req->task->task_works != NULL);
9245 spin_unlock_irq(&ctx->completion_lock);
9247 mutex_unlock(&ctx->uring_lock);
9250 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9252 struct io_ring_ctx *ctx = f->private_data;
9254 if (percpu_ref_tryget(&ctx->refs)) {
9255 __io_uring_show_fdinfo(ctx, m);
9256 percpu_ref_put(&ctx->refs);
9261 static const struct file_operations io_uring_fops = {
9262 .release = io_uring_release,
9263 .mmap = io_uring_mmap,
9265 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9266 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9268 .poll = io_uring_poll,
9269 .fasync = io_uring_fasync,
9270 #ifdef CONFIG_PROC_FS
9271 .show_fdinfo = io_uring_show_fdinfo,
9275 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9276 struct io_uring_params *p)
9278 struct io_rings *rings;
9279 size_t size, sq_array_offset;
9281 /* make sure these are sane, as we already accounted them */
9282 ctx->sq_entries = p->sq_entries;
9283 ctx->cq_entries = p->cq_entries;
9285 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9286 if (size == SIZE_MAX)
9289 rings = io_mem_alloc(size);
9294 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9295 rings->sq_ring_mask = p->sq_entries - 1;
9296 rings->cq_ring_mask = p->cq_entries - 1;
9297 rings->sq_ring_entries = p->sq_entries;
9298 rings->cq_ring_entries = p->cq_entries;
9299 ctx->sq_mask = rings->sq_ring_mask;
9300 ctx->cq_mask = rings->cq_ring_mask;
9302 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9303 if (size == SIZE_MAX) {
9304 io_mem_free(ctx->rings);
9309 ctx->sq_sqes = io_mem_alloc(size);
9310 if (!ctx->sq_sqes) {
9311 io_mem_free(ctx->rings);
9319 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9323 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9327 ret = io_uring_add_task_file(ctx);
9332 fd_install(fd, file);
9337 * Allocate an anonymous fd, this is what constitutes the application
9338 * visible backing of an io_uring instance. The application mmaps this
9339 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9340 * we have to tie this fd to a socket for file garbage collection purposes.
9342 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9345 #if defined(CONFIG_UNIX)
9348 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9351 return ERR_PTR(ret);
9354 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9355 O_RDWR | O_CLOEXEC);
9356 #if defined(CONFIG_UNIX)
9358 sock_release(ctx->ring_sock);
9359 ctx->ring_sock = NULL;
9361 ctx->ring_sock->file = file;
9367 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9368 struct io_uring_params __user *params)
9370 struct io_ring_ctx *ctx;
9376 if (entries > IORING_MAX_ENTRIES) {
9377 if (!(p->flags & IORING_SETUP_CLAMP))
9379 entries = IORING_MAX_ENTRIES;
9383 * Use twice as many entries for the CQ ring. It's possible for the
9384 * application to drive a higher depth than the size of the SQ ring,
9385 * since the sqes are only used at submission time. This allows for
9386 * some flexibility in overcommitting a bit. If the application has
9387 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9388 * of CQ ring entries manually.
9390 p->sq_entries = roundup_pow_of_two(entries);
9391 if (p->flags & IORING_SETUP_CQSIZE) {
9393 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9394 * to a power-of-two, if it isn't already. We do NOT impose
9395 * any cq vs sq ring sizing.
9399 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9400 if (!(p->flags & IORING_SETUP_CLAMP))
9402 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9404 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9405 if (p->cq_entries < p->sq_entries)
9408 p->cq_entries = 2 * p->sq_entries;
9411 ctx = io_ring_ctx_alloc(p);
9414 ctx->compat = in_compat_syscall();
9415 if (!capable(CAP_IPC_LOCK))
9416 ctx->user = get_uid(current_user());
9419 * This is just grabbed for accounting purposes. When a process exits,
9420 * the mm is exited and dropped before the files, hence we need to hang
9421 * on to this mm purely for the purposes of being able to unaccount
9422 * memory (locked/pinned vm). It's not used for anything else.
9424 mmgrab(current->mm);
9425 ctx->mm_account = current->mm;
9427 ret = io_allocate_scq_urings(ctx, p);
9431 ret = io_sq_offload_create(ctx, p);
9435 memset(&p->sq_off, 0, sizeof(p->sq_off));
9436 p->sq_off.head = offsetof(struct io_rings, sq.head);
9437 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9438 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9439 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9440 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9441 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9442 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9444 memset(&p->cq_off, 0, sizeof(p->cq_off));
9445 p->cq_off.head = offsetof(struct io_rings, cq.head);
9446 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9447 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9448 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9449 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9450 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9451 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9453 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9454 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9455 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9456 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9457 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9459 if (copy_to_user(params, p, sizeof(*p))) {
9464 file = io_uring_get_file(ctx);
9466 ret = PTR_ERR(file);
9471 * Install ring fd as the very last thing, so we don't risk someone
9472 * having closed it before we finish setup
9474 ret = io_uring_install_fd(ctx, file);
9476 /* fput will clean it up */
9481 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9484 io_ring_ctx_wait_and_kill(ctx);
9489 * Sets up an aio uring context, and returns the fd. Applications asks for a
9490 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9491 * params structure passed in.
9493 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9495 struct io_uring_params p;
9498 if (copy_from_user(&p, params, sizeof(p)))
9500 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9505 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9506 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9507 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9508 IORING_SETUP_R_DISABLED))
9511 return io_uring_create(entries, &p, params);
9514 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9515 struct io_uring_params __user *, params)
9517 return io_uring_setup(entries, params);
9520 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9522 struct io_uring_probe *p;
9526 size = struct_size(p, ops, nr_args);
9527 if (size == SIZE_MAX)
9529 p = kzalloc(size, GFP_KERNEL);
9534 if (copy_from_user(p, arg, size))
9537 if (memchr_inv(p, 0, size))
9540 p->last_op = IORING_OP_LAST - 1;
9541 if (nr_args > IORING_OP_LAST)
9542 nr_args = IORING_OP_LAST;
9544 for (i = 0; i < nr_args; i++) {
9546 if (!io_op_defs[i].not_supported)
9547 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9552 if (copy_to_user(arg, p, size))
9559 static int io_register_personality(struct io_ring_ctx *ctx)
9561 const struct cred *creds;
9565 creds = get_current_cred();
9567 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9568 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9575 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9576 unsigned int nr_args)
9578 struct io_uring_restriction *res;
9582 /* Restrictions allowed only if rings started disabled */
9583 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9586 /* We allow only a single restrictions registration */
9587 if (ctx->restrictions.registered)
9590 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9593 size = array_size(nr_args, sizeof(*res));
9594 if (size == SIZE_MAX)
9597 res = memdup_user(arg, size);
9599 return PTR_ERR(res);
9603 for (i = 0; i < nr_args; i++) {
9604 switch (res[i].opcode) {
9605 case IORING_RESTRICTION_REGISTER_OP:
9606 if (res[i].register_op >= IORING_REGISTER_LAST) {
9611 __set_bit(res[i].register_op,
9612 ctx->restrictions.register_op);
9614 case IORING_RESTRICTION_SQE_OP:
9615 if (res[i].sqe_op >= IORING_OP_LAST) {
9620 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9622 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9623 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9625 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9626 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9635 /* Reset all restrictions if an error happened */
9637 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9639 ctx->restrictions.registered = true;
9645 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9647 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9650 if (ctx->restrictions.registered)
9651 ctx->restricted = 1;
9653 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9654 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9655 wake_up(&ctx->sq_data->wait);
9659 static bool io_register_op_must_quiesce(int op)
9662 case IORING_UNREGISTER_FILES:
9663 case IORING_REGISTER_FILES_UPDATE:
9664 case IORING_REGISTER_PROBE:
9665 case IORING_REGISTER_PERSONALITY:
9666 case IORING_UNREGISTER_PERSONALITY:
9673 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9674 void __user *arg, unsigned nr_args)
9675 __releases(ctx->uring_lock)
9676 __acquires(ctx->uring_lock)
9681 * We're inside the ring mutex, if the ref is already dying, then
9682 * someone else killed the ctx or is already going through
9683 * io_uring_register().
9685 if (percpu_ref_is_dying(&ctx->refs))
9688 if (io_register_op_must_quiesce(opcode)) {
9689 percpu_ref_kill(&ctx->refs);
9692 * Drop uring mutex before waiting for references to exit. If
9693 * another thread is currently inside io_uring_enter() it might
9694 * need to grab the uring_lock to make progress. If we hold it
9695 * here across the drain wait, then we can deadlock. It's safe
9696 * to drop the mutex here, since no new references will come in
9697 * after we've killed the percpu ref.
9699 mutex_unlock(&ctx->uring_lock);
9701 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9704 ret = io_run_task_work_sig();
9709 mutex_lock(&ctx->uring_lock);
9712 percpu_ref_resurrect(&ctx->refs);
9717 if (ctx->restricted) {
9718 if (opcode >= IORING_REGISTER_LAST) {
9723 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9730 case IORING_REGISTER_BUFFERS:
9731 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9733 case IORING_UNREGISTER_BUFFERS:
9737 ret = io_sqe_buffers_unregister(ctx);
9739 case IORING_REGISTER_FILES:
9740 ret = io_sqe_files_register(ctx, arg, nr_args);
9742 case IORING_UNREGISTER_FILES:
9746 ret = io_sqe_files_unregister(ctx);
9748 case IORING_REGISTER_FILES_UPDATE:
9749 ret = io_sqe_files_update(ctx, arg, nr_args);
9751 case IORING_REGISTER_EVENTFD:
9752 case IORING_REGISTER_EVENTFD_ASYNC:
9756 ret = io_eventfd_register(ctx, arg);
9759 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9760 ctx->eventfd_async = 1;
9762 ctx->eventfd_async = 0;
9764 case IORING_UNREGISTER_EVENTFD:
9768 ret = io_eventfd_unregister(ctx);
9770 case IORING_REGISTER_PROBE:
9772 if (!arg || nr_args > 256)
9774 ret = io_probe(ctx, arg, nr_args);
9776 case IORING_REGISTER_PERSONALITY:
9780 ret = io_register_personality(ctx);
9782 case IORING_UNREGISTER_PERSONALITY:
9786 ret = io_unregister_personality(ctx, nr_args);
9788 case IORING_REGISTER_ENABLE_RINGS:
9792 ret = io_register_enable_rings(ctx);
9794 case IORING_REGISTER_RESTRICTIONS:
9795 ret = io_register_restrictions(ctx, arg, nr_args);
9803 if (io_register_op_must_quiesce(opcode)) {
9804 /* bring the ctx back to life */
9805 percpu_ref_reinit(&ctx->refs);
9807 reinit_completion(&ctx->ref_comp);
9812 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9813 void __user *, arg, unsigned int, nr_args)
9815 struct io_ring_ctx *ctx;
9824 if (f.file->f_op != &io_uring_fops)
9827 ctx = f.file->private_data;
9831 mutex_lock(&ctx->uring_lock);
9832 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9833 mutex_unlock(&ctx->uring_lock);
9834 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9835 ctx->cq_ev_fd != NULL, ret);
9841 static int __init io_uring_init(void)
9843 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9844 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9845 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9848 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9849 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9850 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9851 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9852 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9853 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9854 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9855 BUILD_BUG_SQE_ELEM(8, __u64, off);
9856 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9857 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9858 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9859 BUILD_BUG_SQE_ELEM(24, __u32, len);
9860 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9861 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9862 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9863 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9864 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9865 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9866 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9867 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9868 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9869 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9870 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9871 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9872 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9873 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9874 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9875 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9876 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9877 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9878 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9880 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9881 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9882 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9886 __initcall(io_uring_init);