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 atomic_t park_pending;
264 /* ctx's that are using this sqd */
265 struct list_head ctx_list;
267 struct task_struct *thread;
268 struct wait_queue_head wait;
270 unsigned sq_thread_idle;
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 xarray io_buffers;
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 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
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++;
1555 if (!percpu_ref_tryget(&ctx->refs))
1558 io_commit_cqring(ctx);
1559 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1562 io_cqring_ev_posted(ctx);
1563 percpu_ref_put(&ctx->refs);
1567 static void io_req_complete_state(struct io_kiocb *req, long res,
1568 unsigned int cflags)
1572 req->compl.cflags = cflags;
1573 req->flags |= REQ_F_COMPLETE_INLINE;
1576 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1577 long res, unsigned cflags)
1579 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1580 io_req_complete_state(req, res, cflags);
1582 io_req_complete_post(req, res, cflags);
1585 static inline void io_req_complete(struct io_kiocb *req, long res)
1587 __io_req_complete(req, 0, res, 0);
1590 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1592 struct io_submit_state *state = &ctx->submit_state;
1593 struct io_comp_state *cs = &state->comp;
1594 struct io_kiocb *req = NULL;
1597 * If we have more than a batch's worth of requests in our IRQ side
1598 * locked cache, grab the lock and move them over to our submission
1601 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1602 spin_lock_irq(&ctx->completion_lock);
1603 list_splice_init(&cs->locked_free_list, &cs->free_list);
1604 cs->locked_free_nr = 0;
1605 spin_unlock_irq(&ctx->completion_lock);
1608 while (!list_empty(&cs->free_list)) {
1609 req = list_first_entry(&cs->free_list, struct io_kiocb,
1611 list_del(&req->compl.list);
1612 state->reqs[state->free_reqs++] = req;
1613 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1620 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1622 struct io_submit_state *state = &ctx->submit_state;
1624 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1626 if (!state->free_reqs) {
1627 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1630 if (io_flush_cached_reqs(ctx))
1633 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1637 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1638 * retry single alloc to be on the safe side.
1640 if (unlikely(ret <= 0)) {
1641 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1642 if (!state->reqs[0])
1646 state->free_reqs = ret;
1650 return state->reqs[state->free_reqs];
1653 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1660 static void io_dismantle_req(struct io_kiocb *req)
1664 if (req->async_data)
1665 kfree(req->async_data);
1667 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1668 if (req->fixed_rsrc_refs)
1669 percpu_ref_put(req->fixed_rsrc_refs);
1670 if (req->work.creds) {
1671 put_cred(req->work.creds);
1672 req->work.creds = NULL;
1675 if (req->flags & REQ_F_INFLIGHT) {
1676 struct io_ring_ctx *ctx = req->ctx;
1677 unsigned long flags;
1679 spin_lock_irqsave(&ctx->inflight_lock, flags);
1680 list_del(&req->inflight_entry);
1681 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1682 req->flags &= ~REQ_F_INFLIGHT;
1686 /* must to be called somewhat shortly after putting a request */
1687 static inline void io_put_task(struct task_struct *task, int nr)
1689 struct io_uring_task *tctx = task->io_uring;
1691 percpu_counter_sub(&tctx->inflight, nr);
1692 if (unlikely(atomic_read(&tctx->in_idle)))
1693 wake_up(&tctx->wait);
1694 put_task_struct_many(task, nr);
1697 static void __io_free_req(struct io_kiocb *req)
1699 struct io_ring_ctx *ctx = req->ctx;
1701 io_dismantle_req(req);
1702 io_put_task(req->task, 1);
1704 kmem_cache_free(req_cachep, req);
1705 percpu_ref_put(&ctx->refs);
1708 static inline void io_remove_next_linked(struct io_kiocb *req)
1710 struct io_kiocb *nxt = req->link;
1712 req->link = nxt->link;
1716 static bool io_kill_linked_timeout(struct io_kiocb *req)
1717 __must_hold(&req->ctx->completion_lock)
1719 struct io_kiocb *link = req->link;
1720 bool cancelled = false;
1723 * Can happen if a linked timeout fired and link had been like
1724 * req -> link t-out -> link t-out [-> ...]
1726 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1727 struct io_timeout_data *io = link->async_data;
1730 io_remove_next_linked(req);
1731 link->timeout.head = NULL;
1732 ret = hrtimer_try_to_cancel(&io->timer);
1734 io_cqring_fill_event(link, -ECANCELED);
1735 io_put_req_deferred(link, 1);
1739 req->flags &= ~REQ_F_LINK_TIMEOUT;
1743 static void io_fail_links(struct io_kiocb *req)
1744 __must_hold(&req->ctx->completion_lock)
1746 struct io_kiocb *nxt, *link = req->link;
1753 trace_io_uring_fail_link(req, link);
1754 io_cqring_fill_event(link, -ECANCELED);
1755 io_put_req_deferred(link, 2);
1760 static bool io_disarm_next(struct io_kiocb *req)
1761 __must_hold(&req->ctx->completion_lock)
1763 bool posted = false;
1765 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1766 posted = io_kill_linked_timeout(req);
1767 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1768 posted |= (req->link != NULL);
1774 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1776 struct io_kiocb *nxt;
1779 * If LINK is set, we have dependent requests in this chain. If we
1780 * didn't fail this request, queue the first one up, moving any other
1781 * dependencies to the next request. In case of failure, fail the rest
1784 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1785 struct io_ring_ctx *ctx = req->ctx;
1786 unsigned long flags;
1789 spin_lock_irqsave(&ctx->completion_lock, flags);
1790 posted = io_disarm_next(req);
1792 io_commit_cqring(req->ctx);
1793 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1795 io_cqring_ev_posted(ctx);
1802 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1804 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1806 return __io_req_find_next(req);
1809 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1813 if (ctx->submit_state.comp.nr) {
1814 mutex_lock(&ctx->uring_lock);
1815 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1816 mutex_unlock(&ctx->uring_lock);
1818 percpu_ref_put(&ctx->refs);
1821 static bool __tctx_task_work(struct io_uring_task *tctx)
1823 struct io_ring_ctx *ctx = NULL;
1824 struct io_wq_work_list list;
1825 struct io_wq_work_node *node;
1827 if (wq_list_empty(&tctx->task_list))
1830 spin_lock_irq(&tctx->task_lock);
1831 list = tctx->task_list;
1832 INIT_WQ_LIST(&tctx->task_list);
1833 spin_unlock_irq(&tctx->task_lock);
1837 struct io_wq_work_node *next = node->next;
1838 struct io_kiocb *req;
1840 req = container_of(node, struct io_kiocb, io_task_work.node);
1841 if (req->ctx != ctx) {
1842 ctx_flush_and_put(ctx);
1844 percpu_ref_get(&ctx->refs);
1847 req->task_work.func(&req->task_work);
1851 ctx_flush_and_put(ctx);
1852 return list.first != NULL;
1855 static void tctx_task_work(struct callback_head *cb)
1857 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1859 clear_bit(0, &tctx->task_state);
1861 while (__tctx_task_work(tctx))
1865 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1866 enum task_work_notify_mode notify)
1868 struct io_uring_task *tctx = tsk->io_uring;
1869 struct io_wq_work_node *node, *prev;
1870 unsigned long flags;
1873 WARN_ON_ONCE(!tctx);
1875 spin_lock_irqsave(&tctx->task_lock, flags);
1876 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1877 spin_unlock_irqrestore(&tctx->task_lock, flags);
1879 /* task_work already pending, we're done */
1880 if (test_bit(0, &tctx->task_state) ||
1881 test_and_set_bit(0, &tctx->task_state))
1884 if (!task_work_add(tsk, &tctx->task_work, notify))
1888 * Slow path - we failed, find and delete work. if the work is not
1889 * in the list, it got run and we're fine.
1892 spin_lock_irqsave(&tctx->task_lock, flags);
1893 wq_list_for_each(node, prev, &tctx->task_list) {
1894 if (&req->io_task_work.node == node) {
1895 wq_list_del(&tctx->task_list, node, prev);
1900 spin_unlock_irqrestore(&tctx->task_lock, flags);
1901 clear_bit(0, &tctx->task_state);
1905 static int io_req_task_work_add(struct io_kiocb *req)
1907 struct task_struct *tsk = req->task;
1908 struct io_ring_ctx *ctx = req->ctx;
1909 enum task_work_notify_mode notify;
1912 if (tsk->flags & PF_EXITING)
1916 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1917 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1918 * processing task_work. There's no reliable way to tell if TWA_RESUME
1922 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1923 notify = TWA_SIGNAL;
1925 ret = io_task_work_add(tsk, req, notify);
1927 wake_up_process(tsk);
1932 static bool io_run_task_work_head(struct callback_head **work_head)
1934 struct callback_head *work, *next;
1935 bool executed = false;
1938 work = xchg(work_head, NULL);
1954 static void io_task_work_add_head(struct callback_head **work_head,
1955 struct callback_head *task_work)
1957 struct callback_head *head;
1960 head = READ_ONCE(*work_head);
1961 task_work->next = head;
1962 } while (cmpxchg(work_head, head, task_work) != head);
1965 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1966 task_work_func_t cb)
1968 init_task_work(&req->task_work, cb);
1969 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1972 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1974 struct io_ring_ctx *ctx = req->ctx;
1976 spin_lock_irq(&ctx->completion_lock);
1977 io_cqring_fill_event(req, error);
1978 io_commit_cqring(ctx);
1979 spin_unlock_irq(&ctx->completion_lock);
1981 io_cqring_ev_posted(ctx);
1982 req_set_fail_links(req);
1983 io_double_put_req(req);
1986 static void io_req_task_cancel(struct callback_head *cb)
1988 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1989 struct io_ring_ctx *ctx = req->ctx;
1991 mutex_lock(&ctx->uring_lock);
1992 __io_req_task_cancel(req, req->result);
1993 mutex_unlock(&ctx->uring_lock);
1994 percpu_ref_put(&ctx->refs);
1997 static void __io_req_task_submit(struct io_kiocb *req)
1999 struct io_ring_ctx *ctx = req->ctx;
2001 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2002 mutex_lock(&ctx->uring_lock);
2003 if (!(current->flags & PF_EXITING) && !current->in_execve)
2004 __io_queue_sqe(req);
2006 __io_req_task_cancel(req, -EFAULT);
2007 mutex_unlock(&ctx->uring_lock);
2010 static void io_req_task_submit(struct callback_head *cb)
2012 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2014 __io_req_task_submit(req);
2017 static void io_req_task_queue(struct io_kiocb *req)
2021 req->task_work.func = io_req_task_submit;
2022 ret = io_req_task_work_add(req);
2023 if (unlikely(ret)) {
2024 req->result = -ECANCELED;
2025 percpu_ref_get(&req->ctx->refs);
2026 io_req_task_work_add_fallback(req, io_req_task_cancel);
2030 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2032 percpu_ref_get(&req->ctx->refs);
2034 req->task_work.func = io_req_task_cancel;
2036 if (unlikely(io_req_task_work_add(req)))
2037 io_req_task_work_add_fallback(req, io_req_task_cancel);
2040 static inline void io_queue_next(struct io_kiocb *req)
2042 struct io_kiocb *nxt = io_req_find_next(req);
2045 io_req_task_queue(nxt);
2048 static void io_free_req(struct io_kiocb *req)
2055 struct task_struct *task;
2060 static inline void io_init_req_batch(struct req_batch *rb)
2067 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2068 struct req_batch *rb)
2071 io_put_task(rb->task, rb->task_refs);
2073 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2076 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2077 struct io_submit_state *state)
2081 if (req->task != rb->task) {
2083 io_put_task(rb->task, rb->task_refs);
2084 rb->task = req->task;
2090 io_dismantle_req(req);
2091 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2092 state->reqs[state->free_reqs++] = req;
2094 list_add(&req->compl.list, &state->comp.free_list);
2097 static void io_submit_flush_completions(struct io_comp_state *cs,
2098 struct io_ring_ctx *ctx)
2101 struct io_kiocb *req;
2102 struct req_batch rb;
2104 io_init_req_batch(&rb);
2105 spin_lock_irq(&ctx->completion_lock);
2106 for (i = 0; i < nr; i++) {
2108 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2110 io_commit_cqring(ctx);
2111 spin_unlock_irq(&ctx->completion_lock);
2113 io_cqring_ev_posted(ctx);
2114 for (i = 0; i < nr; i++) {
2117 /* submission and completion refs */
2118 if (refcount_sub_and_test(2, &req->refs))
2119 io_req_free_batch(&rb, req, &ctx->submit_state);
2122 io_req_free_batch_finish(ctx, &rb);
2127 * Drop reference to request, return next in chain (if there is one) if this
2128 * was the last reference to this request.
2130 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2132 struct io_kiocb *nxt = NULL;
2134 if (refcount_dec_and_test(&req->refs)) {
2135 nxt = io_req_find_next(req);
2141 static void io_put_req(struct io_kiocb *req)
2143 if (refcount_dec_and_test(&req->refs))
2147 static void io_put_req_deferred_cb(struct callback_head *cb)
2149 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2154 static void io_free_req_deferred(struct io_kiocb *req)
2158 req->task_work.func = io_put_req_deferred_cb;
2159 ret = io_req_task_work_add(req);
2161 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2164 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2166 if (refcount_sub_and_test(refs, &req->refs))
2167 io_free_req_deferred(req);
2170 static void io_double_put_req(struct io_kiocb *req)
2172 /* drop both submit and complete references */
2173 if (refcount_sub_and_test(2, &req->refs))
2177 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2179 /* See comment at the top of this file */
2181 return __io_cqring_events(ctx);
2184 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2186 struct io_rings *rings = ctx->rings;
2188 /* make sure SQ entry isn't read before tail */
2189 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2192 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2194 unsigned int cflags;
2196 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2197 cflags |= IORING_CQE_F_BUFFER;
2198 req->flags &= ~REQ_F_BUFFER_SELECTED;
2203 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2205 struct io_buffer *kbuf;
2207 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2208 return io_put_kbuf(req, kbuf);
2211 static inline bool io_run_task_work(void)
2214 * Not safe to run on exiting task, and the task_work handling will
2215 * not add work to such a task.
2217 if (unlikely(current->flags & PF_EXITING))
2219 if (current->task_works) {
2220 __set_current_state(TASK_RUNNING);
2229 * Find and free completed poll iocbs
2231 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2232 struct list_head *done)
2234 struct req_batch rb;
2235 struct io_kiocb *req;
2237 /* order with ->result store in io_complete_rw_iopoll() */
2240 io_init_req_batch(&rb);
2241 while (!list_empty(done)) {
2244 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2245 list_del(&req->inflight_entry);
2247 if (READ_ONCE(req->result) == -EAGAIN) {
2248 req->iopoll_completed = 0;
2249 if (io_rw_reissue(req))
2253 if (req->flags & REQ_F_BUFFER_SELECTED)
2254 cflags = io_put_rw_kbuf(req);
2256 __io_cqring_fill_event(req, req->result, cflags);
2259 if (refcount_dec_and_test(&req->refs))
2260 io_req_free_batch(&rb, req, &ctx->submit_state);
2263 io_commit_cqring(ctx);
2264 io_cqring_ev_posted_iopoll(ctx);
2265 io_req_free_batch_finish(ctx, &rb);
2268 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2271 struct io_kiocb *req, *tmp;
2277 * Only spin for completions if we don't have multiple devices hanging
2278 * off our complete list, and we're under the requested amount.
2280 spin = !ctx->poll_multi_file && *nr_events < min;
2283 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2284 struct kiocb *kiocb = &req->rw.kiocb;
2287 * Move completed and retryable entries to our local lists.
2288 * If we find a request that requires polling, break out
2289 * and complete those lists first, if we have entries there.
2291 if (READ_ONCE(req->iopoll_completed)) {
2292 list_move_tail(&req->inflight_entry, &done);
2295 if (!list_empty(&done))
2298 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2302 /* iopoll may have completed current req */
2303 if (READ_ONCE(req->iopoll_completed))
2304 list_move_tail(&req->inflight_entry, &done);
2311 if (!list_empty(&done))
2312 io_iopoll_complete(ctx, nr_events, &done);
2318 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2319 * non-spinning poll check - we'll still enter the driver poll loop, but only
2320 * as a non-spinning completion check.
2322 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2325 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2328 ret = io_do_iopoll(ctx, nr_events, min);
2331 if (*nr_events >= min)
2339 * We can't just wait for polled events to come to us, we have to actively
2340 * find and complete them.
2342 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2344 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2347 mutex_lock(&ctx->uring_lock);
2348 while (!list_empty(&ctx->iopoll_list)) {
2349 unsigned int nr_events = 0;
2351 io_do_iopoll(ctx, &nr_events, 0);
2353 /* let it sleep and repeat later if can't complete a request */
2357 * Ensure we allow local-to-the-cpu processing to take place,
2358 * in this case we need to ensure that we reap all events.
2359 * Also let task_work, etc. to progress by releasing the mutex
2361 if (need_resched()) {
2362 mutex_unlock(&ctx->uring_lock);
2364 mutex_lock(&ctx->uring_lock);
2367 mutex_unlock(&ctx->uring_lock);
2370 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2372 unsigned int nr_events = 0;
2373 int iters = 0, ret = 0;
2376 * We disallow the app entering submit/complete with polling, but we
2377 * still need to lock the ring to prevent racing with polled issue
2378 * that got punted to a workqueue.
2380 mutex_lock(&ctx->uring_lock);
2383 * Don't enter poll loop if we already have events pending.
2384 * If we do, we can potentially be spinning for commands that
2385 * already triggered a CQE (eg in error).
2387 if (test_bit(0, &ctx->cq_check_overflow))
2388 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2389 if (io_cqring_events(ctx))
2393 * If a submit got punted to a workqueue, we can have the
2394 * application entering polling for a command before it gets
2395 * issued. That app will hold the uring_lock for the duration
2396 * of the poll right here, so we need to take a breather every
2397 * now and then to ensure that the issue has a chance to add
2398 * the poll to the issued list. Otherwise we can spin here
2399 * forever, while the workqueue is stuck trying to acquire the
2402 if (!(++iters & 7)) {
2403 mutex_unlock(&ctx->uring_lock);
2405 mutex_lock(&ctx->uring_lock);
2408 ret = io_iopoll_getevents(ctx, &nr_events, min);
2412 } while (min && !nr_events && !need_resched());
2414 mutex_unlock(&ctx->uring_lock);
2418 static void kiocb_end_write(struct io_kiocb *req)
2421 * Tell lockdep we inherited freeze protection from submission
2424 if (req->flags & REQ_F_ISREG) {
2425 struct inode *inode = file_inode(req->file);
2427 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2429 file_end_write(req->file);
2433 static bool io_resubmit_prep(struct io_kiocb *req)
2435 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2437 struct iov_iter iter;
2439 /* already prepared */
2440 if (req->async_data)
2443 switch (req->opcode) {
2444 case IORING_OP_READV:
2445 case IORING_OP_READ_FIXED:
2446 case IORING_OP_READ:
2449 case IORING_OP_WRITEV:
2450 case IORING_OP_WRITE_FIXED:
2451 case IORING_OP_WRITE:
2455 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2460 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2463 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2466 static bool io_rw_should_reissue(struct io_kiocb *req)
2468 umode_t mode = file_inode(req->file)->i_mode;
2469 struct io_ring_ctx *ctx = req->ctx;
2471 if (!S_ISBLK(mode) && !S_ISREG(mode))
2473 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2474 !(ctx->flags & IORING_SETUP_IOPOLL)))
2477 * If ref is dying, we might be running poll reap from the exit work.
2478 * Don't attempt to reissue from that path, just let it fail with
2481 if (percpu_ref_is_dying(&ctx->refs))
2487 static bool io_rw_reissue(struct io_kiocb *req)
2490 if (!io_rw_should_reissue(req))
2493 lockdep_assert_held(&req->ctx->uring_lock);
2495 if (io_resubmit_prep(req)) {
2496 refcount_inc(&req->refs);
2497 io_queue_async_work(req);
2500 req_set_fail_links(req);
2505 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2506 unsigned int issue_flags)
2510 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2512 if (res != req->result)
2513 req_set_fail_links(req);
2515 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2516 kiocb_end_write(req);
2517 if (req->flags & REQ_F_BUFFER_SELECTED)
2518 cflags = io_put_rw_kbuf(req);
2519 __io_req_complete(req, issue_flags, res, cflags);
2522 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2524 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2526 __io_complete_rw(req, res, res2, 0);
2529 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2531 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2534 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2535 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2536 struct io_async_rw *rw = req->async_data;
2539 iov_iter_revert(&rw->iter,
2540 req->result - iov_iter_count(&rw->iter));
2541 else if (!io_resubmit_prep(req))
2546 if (kiocb->ki_flags & IOCB_WRITE)
2547 kiocb_end_write(req);
2549 if (res != -EAGAIN && res != req->result)
2550 req_set_fail_links(req);
2552 WRITE_ONCE(req->result, res);
2553 /* order with io_poll_complete() checking ->result */
2555 WRITE_ONCE(req->iopoll_completed, 1);
2559 * After the iocb has been issued, it's safe to be found on the poll list.
2560 * Adding the kiocb to the list AFTER submission ensures that we don't
2561 * find it from a io_iopoll_getevents() thread before the issuer is done
2562 * accessing the kiocb cookie.
2564 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2566 struct io_ring_ctx *ctx = req->ctx;
2569 * Track whether we have multiple files in our lists. This will impact
2570 * how we do polling eventually, not spinning if we're on potentially
2571 * different devices.
2573 if (list_empty(&ctx->iopoll_list)) {
2574 ctx->poll_multi_file = false;
2575 } else if (!ctx->poll_multi_file) {
2576 struct io_kiocb *list_req;
2578 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2580 if (list_req->file != req->file)
2581 ctx->poll_multi_file = true;
2585 * For fast devices, IO may have already completed. If it has, add
2586 * it to the front so we find it first.
2588 if (READ_ONCE(req->iopoll_completed))
2589 list_add(&req->inflight_entry, &ctx->iopoll_list);
2591 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2594 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2595 * task context or in io worker task context. If current task context is
2596 * sq thread, we don't need to check whether should wake up sq thread.
2598 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2599 wq_has_sleeper(&ctx->sq_data->wait))
2600 wake_up(&ctx->sq_data->wait);
2603 static inline void io_state_file_put(struct io_submit_state *state)
2605 if (state->file_refs) {
2606 fput_many(state->file, state->file_refs);
2607 state->file_refs = 0;
2612 * Get as many references to a file as we have IOs left in this submission,
2613 * assuming most submissions are for one file, or at least that each file
2614 * has more than one submission.
2616 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2621 if (state->file_refs) {
2622 if (state->fd == fd) {
2626 io_state_file_put(state);
2628 state->file = fget_many(fd, state->ios_left);
2629 if (unlikely(!state->file))
2633 state->file_refs = state->ios_left - 1;
2637 static bool io_bdev_nowait(struct block_device *bdev)
2639 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2643 * If we tracked the file through the SCM inflight mechanism, we could support
2644 * any file. For now, just ensure that anything potentially problematic is done
2647 static bool io_file_supports_async(struct file *file, int rw)
2649 umode_t mode = file_inode(file)->i_mode;
2651 if (S_ISBLK(mode)) {
2652 if (IS_ENABLED(CONFIG_BLOCK) &&
2653 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2657 if (S_ISCHR(mode) || S_ISSOCK(mode))
2659 if (S_ISREG(mode)) {
2660 if (IS_ENABLED(CONFIG_BLOCK) &&
2661 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2662 file->f_op != &io_uring_fops)
2667 /* any ->read/write should understand O_NONBLOCK */
2668 if (file->f_flags & O_NONBLOCK)
2671 if (!(file->f_mode & FMODE_NOWAIT))
2675 return file->f_op->read_iter != NULL;
2677 return file->f_op->write_iter != NULL;
2680 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2682 struct io_ring_ctx *ctx = req->ctx;
2683 struct kiocb *kiocb = &req->rw.kiocb;
2684 struct file *file = req->file;
2688 if (S_ISREG(file_inode(file)->i_mode))
2689 req->flags |= REQ_F_ISREG;
2691 kiocb->ki_pos = READ_ONCE(sqe->off);
2692 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2693 req->flags |= REQ_F_CUR_POS;
2694 kiocb->ki_pos = file->f_pos;
2696 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2697 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2698 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2702 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2703 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2704 req->flags |= REQ_F_NOWAIT;
2706 ioprio = READ_ONCE(sqe->ioprio);
2708 ret = ioprio_check_cap(ioprio);
2712 kiocb->ki_ioprio = ioprio;
2714 kiocb->ki_ioprio = get_current_ioprio();
2716 if (ctx->flags & IORING_SETUP_IOPOLL) {
2717 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2718 !kiocb->ki_filp->f_op->iopoll)
2721 kiocb->ki_flags |= IOCB_HIPRI;
2722 kiocb->ki_complete = io_complete_rw_iopoll;
2723 req->iopoll_completed = 0;
2725 if (kiocb->ki_flags & IOCB_HIPRI)
2727 kiocb->ki_complete = io_complete_rw;
2730 req->rw.addr = READ_ONCE(sqe->addr);
2731 req->rw.len = READ_ONCE(sqe->len);
2732 req->buf_index = READ_ONCE(sqe->buf_index);
2736 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2742 case -ERESTARTNOINTR:
2743 case -ERESTARTNOHAND:
2744 case -ERESTART_RESTARTBLOCK:
2746 * We can't just restart the syscall, since previously
2747 * submitted sqes may already be in progress. Just fail this
2753 kiocb->ki_complete(kiocb, ret, 0);
2757 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2758 unsigned int issue_flags)
2760 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2761 struct io_async_rw *io = req->async_data;
2763 /* add previously done IO, if any */
2764 if (io && io->bytes_done > 0) {
2766 ret = io->bytes_done;
2768 ret += io->bytes_done;
2771 if (req->flags & REQ_F_CUR_POS)
2772 req->file->f_pos = kiocb->ki_pos;
2773 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2774 __io_complete_rw(req, ret, 0, issue_flags);
2776 io_rw_done(kiocb, ret);
2779 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2781 struct io_ring_ctx *ctx = req->ctx;
2782 size_t len = req->rw.len;
2783 struct io_mapped_ubuf *imu;
2784 u16 index, buf_index = req->buf_index;
2788 if (unlikely(buf_index >= ctx->nr_user_bufs))
2790 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2791 imu = &ctx->user_bufs[index];
2792 buf_addr = req->rw.addr;
2795 if (buf_addr + len < buf_addr)
2797 /* not inside the mapped region */
2798 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2802 * May not be a start of buffer, set size appropriately
2803 * and advance us to the beginning.
2805 offset = buf_addr - imu->ubuf;
2806 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2810 * Don't use iov_iter_advance() here, as it's really slow for
2811 * using the latter parts of a big fixed buffer - it iterates
2812 * over each segment manually. We can cheat a bit here, because
2815 * 1) it's a BVEC iter, we set it up
2816 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2817 * first and last bvec
2819 * So just find our index, and adjust the iterator afterwards.
2820 * If the offset is within the first bvec (or the whole first
2821 * bvec, just use iov_iter_advance(). This makes it easier
2822 * since we can just skip the first segment, which may not
2823 * be PAGE_SIZE aligned.
2825 const struct bio_vec *bvec = imu->bvec;
2827 if (offset <= bvec->bv_len) {
2828 iov_iter_advance(iter, offset);
2830 unsigned long seg_skip;
2832 /* skip first vec */
2833 offset -= bvec->bv_len;
2834 seg_skip = 1 + (offset >> PAGE_SHIFT);
2836 iter->bvec = bvec + seg_skip;
2837 iter->nr_segs -= seg_skip;
2838 iter->count -= bvec->bv_len + offset;
2839 iter->iov_offset = offset & ~PAGE_MASK;
2846 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2849 mutex_unlock(&ctx->uring_lock);
2852 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2855 * "Normal" inline submissions always hold the uring_lock, since we
2856 * grab it from the system call. Same is true for the SQPOLL offload.
2857 * The only exception is when we've detached the request and issue it
2858 * from an async worker thread, grab the lock for that case.
2861 mutex_lock(&ctx->uring_lock);
2864 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2865 int bgid, struct io_buffer *kbuf,
2868 struct io_buffer *head;
2870 if (req->flags & REQ_F_BUFFER_SELECTED)
2873 io_ring_submit_lock(req->ctx, needs_lock);
2875 lockdep_assert_held(&req->ctx->uring_lock);
2877 head = xa_load(&req->ctx->io_buffers, bgid);
2879 if (!list_empty(&head->list)) {
2880 kbuf = list_last_entry(&head->list, struct io_buffer,
2882 list_del(&kbuf->list);
2885 xa_erase(&req->ctx->io_buffers, bgid);
2887 if (*len > kbuf->len)
2890 kbuf = ERR_PTR(-ENOBUFS);
2893 io_ring_submit_unlock(req->ctx, needs_lock);
2898 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2901 struct io_buffer *kbuf;
2904 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2905 bgid = req->buf_index;
2906 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2909 req->rw.addr = (u64) (unsigned long) kbuf;
2910 req->flags |= REQ_F_BUFFER_SELECTED;
2911 return u64_to_user_ptr(kbuf->addr);
2914 #ifdef CONFIG_COMPAT
2915 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2918 struct compat_iovec __user *uiov;
2919 compat_ssize_t clen;
2923 uiov = u64_to_user_ptr(req->rw.addr);
2924 if (!access_ok(uiov, sizeof(*uiov)))
2926 if (__get_user(clen, &uiov->iov_len))
2932 buf = io_rw_buffer_select(req, &len, needs_lock);
2934 return PTR_ERR(buf);
2935 iov[0].iov_base = buf;
2936 iov[0].iov_len = (compat_size_t) len;
2941 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2944 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2948 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2951 len = iov[0].iov_len;
2954 buf = io_rw_buffer_select(req, &len, needs_lock);
2956 return PTR_ERR(buf);
2957 iov[0].iov_base = buf;
2958 iov[0].iov_len = len;
2962 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2965 if (req->flags & REQ_F_BUFFER_SELECTED) {
2966 struct io_buffer *kbuf;
2968 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2969 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2970 iov[0].iov_len = kbuf->len;
2973 if (req->rw.len != 1)
2976 #ifdef CONFIG_COMPAT
2977 if (req->ctx->compat)
2978 return io_compat_import(req, iov, needs_lock);
2981 return __io_iov_buffer_select(req, iov, needs_lock);
2984 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2985 struct iov_iter *iter, bool needs_lock)
2987 void __user *buf = u64_to_user_ptr(req->rw.addr);
2988 size_t sqe_len = req->rw.len;
2989 u8 opcode = req->opcode;
2992 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2994 return io_import_fixed(req, rw, iter);
2997 /* buffer index only valid with fixed read/write, or buffer select */
2998 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3001 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3002 if (req->flags & REQ_F_BUFFER_SELECT) {
3003 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3005 return PTR_ERR(buf);
3006 req->rw.len = sqe_len;
3009 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3014 if (req->flags & REQ_F_BUFFER_SELECT) {
3015 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3017 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3022 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3026 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3028 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3032 * For files that don't have ->read_iter() and ->write_iter(), handle them
3033 * by looping over ->read() or ->write() manually.
3035 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3037 struct kiocb *kiocb = &req->rw.kiocb;
3038 struct file *file = req->file;
3042 * Don't support polled IO through this interface, and we can't
3043 * support non-blocking either. For the latter, this just causes
3044 * the kiocb to be handled from an async context.
3046 if (kiocb->ki_flags & IOCB_HIPRI)
3048 if (kiocb->ki_flags & IOCB_NOWAIT)
3051 while (iov_iter_count(iter)) {
3055 if (!iov_iter_is_bvec(iter)) {
3056 iovec = iov_iter_iovec(iter);
3058 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3059 iovec.iov_len = req->rw.len;
3063 nr = file->f_op->read(file, iovec.iov_base,
3064 iovec.iov_len, io_kiocb_ppos(kiocb));
3066 nr = file->f_op->write(file, iovec.iov_base,
3067 iovec.iov_len, io_kiocb_ppos(kiocb));
3076 if (nr != iovec.iov_len)
3080 iov_iter_advance(iter, nr);
3086 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3087 const struct iovec *fast_iov, struct iov_iter *iter)
3089 struct io_async_rw *rw = req->async_data;
3091 memcpy(&rw->iter, iter, sizeof(*iter));
3092 rw->free_iovec = iovec;
3094 /* can only be fixed buffers, no need to do anything */
3095 if (iov_iter_is_bvec(iter))
3098 unsigned iov_off = 0;
3100 rw->iter.iov = rw->fast_iov;
3101 if (iter->iov != fast_iov) {
3102 iov_off = iter->iov - fast_iov;
3103 rw->iter.iov += iov_off;
3105 if (rw->fast_iov != fast_iov)
3106 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3107 sizeof(struct iovec) * iter->nr_segs);
3109 req->flags |= REQ_F_NEED_CLEANUP;
3113 static inline int __io_alloc_async_data(struct io_kiocb *req)
3115 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3116 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3117 return req->async_data == NULL;
3120 static int io_alloc_async_data(struct io_kiocb *req)
3122 if (!io_op_defs[req->opcode].needs_async_data)
3125 return __io_alloc_async_data(req);
3128 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3129 const struct iovec *fast_iov,
3130 struct iov_iter *iter, bool force)
3132 if (!force && !io_op_defs[req->opcode].needs_async_data)
3134 if (!req->async_data) {
3135 if (__io_alloc_async_data(req)) {
3140 io_req_map_rw(req, iovec, fast_iov, iter);
3145 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3147 struct io_async_rw *iorw = req->async_data;
3148 struct iovec *iov = iorw->fast_iov;
3151 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3152 if (unlikely(ret < 0))
3155 iorw->bytes_done = 0;
3156 iorw->free_iovec = iov;
3158 req->flags |= REQ_F_NEED_CLEANUP;
3162 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3164 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3166 return io_prep_rw(req, sqe);
3170 * This is our waitqueue callback handler, registered through lock_page_async()
3171 * when we initially tried to do the IO with the iocb armed our waitqueue.
3172 * This gets called when the page is unlocked, and we generally expect that to
3173 * happen when the page IO is completed and the page is now uptodate. This will
3174 * queue a task_work based retry of the operation, attempting to copy the data
3175 * again. If the latter fails because the page was NOT uptodate, then we will
3176 * do a thread based blocking retry of the operation. That's the unexpected
3179 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3180 int sync, void *arg)
3182 struct wait_page_queue *wpq;
3183 struct io_kiocb *req = wait->private;
3184 struct wait_page_key *key = arg;
3186 wpq = container_of(wait, struct wait_page_queue, wait);
3188 if (!wake_page_match(wpq, key))
3191 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3192 list_del_init(&wait->entry);
3194 /* submit ref gets dropped, acquire a new one */
3195 refcount_inc(&req->refs);
3196 io_req_task_queue(req);
3201 * This controls whether a given IO request should be armed for async page
3202 * based retry. If we return false here, the request is handed to the async
3203 * worker threads for retry. If we're doing buffered reads on a regular file,
3204 * we prepare a private wait_page_queue entry and retry the operation. This
3205 * will either succeed because the page is now uptodate and unlocked, or it
3206 * will register a callback when the page is unlocked at IO completion. Through
3207 * that callback, io_uring uses task_work to setup a retry of the operation.
3208 * That retry will attempt the buffered read again. The retry will generally
3209 * succeed, or in rare cases where it fails, we then fall back to using the
3210 * async worker threads for a blocking retry.
3212 static bool io_rw_should_retry(struct io_kiocb *req)
3214 struct io_async_rw *rw = req->async_data;
3215 struct wait_page_queue *wait = &rw->wpq;
3216 struct kiocb *kiocb = &req->rw.kiocb;
3218 /* never retry for NOWAIT, we just complete with -EAGAIN */
3219 if (req->flags & REQ_F_NOWAIT)
3222 /* Only for buffered IO */
3223 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3227 * just use poll if we can, and don't attempt if the fs doesn't
3228 * support callback based unlocks
3230 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3233 wait->wait.func = io_async_buf_func;
3234 wait->wait.private = req;
3235 wait->wait.flags = 0;
3236 INIT_LIST_HEAD(&wait->wait.entry);
3237 kiocb->ki_flags |= IOCB_WAITQ;
3238 kiocb->ki_flags &= ~IOCB_NOWAIT;
3239 kiocb->ki_waitq = wait;
3243 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3245 if (req->file->f_op->read_iter)
3246 return call_read_iter(req->file, &req->rw.kiocb, iter);
3247 else if (req->file->f_op->read)
3248 return loop_rw_iter(READ, req, iter);
3253 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3255 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3256 struct kiocb *kiocb = &req->rw.kiocb;
3257 struct iov_iter __iter, *iter = &__iter;
3258 struct io_async_rw *rw = req->async_data;
3259 ssize_t io_size, ret, ret2;
3260 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3266 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3270 io_size = iov_iter_count(iter);
3271 req->result = io_size;
3273 /* Ensure we clear previously set non-block flag */
3274 if (!force_nonblock)
3275 kiocb->ki_flags &= ~IOCB_NOWAIT;
3277 kiocb->ki_flags |= IOCB_NOWAIT;
3279 /* If the file doesn't support async, just async punt */
3280 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3281 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3282 return ret ?: -EAGAIN;
3285 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3286 if (unlikely(ret)) {
3291 ret = io_iter_do_read(req, iter);
3293 if (ret == -EIOCBQUEUED) {
3294 if (req->async_data)
3295 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3297 } else if (ret == -EAGAIN) {
3298 /* IOPOLL retry should happen for io-wq threads */
3299 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3301 /* no retry on NONBLOCK nor RWF_NOWAIT */
3302 if (req->flags & REQ_F_NOWAIT)
3304 /* some cases will consume bytes even on error returns */
3305 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3307 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3308 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3309 /* read all, failed, already did sync or don't want to retry */
3313 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3318 rw = req->async_data;
3319 /* now use our persistent iterator, if we aren't already */
3324 rw->bytes_done += ret;
3325 /* if we can retry, do so with the callbacks armed */
3326 if (!io_rw_should_retry(req)) {
3327 kiocb->ki_flags &= ~IOCB_WAITQ;
3332 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3333 * we get -EIOCBQUEUED, then we'll get a notification when the
3334 * desired page gets unlocked. We can also get a partial read
3335 * here, and if we do, then just retry at the new offset.
3337 ret = io_iter_do_read(req, iter);
3338 if (ret == -EIOCBQUEUED)
3340 /* we got some bytes, but not all. retry. */
3341 kiocb->ki_flags &= ~IOCB_WAITQ;
3342 } while (ret > 0 && ret < io_size);
3344 kiocb_done(kiocb, ret, issue_flags);
3346 /* it's faster to check here then delegate to kfree */
3352 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3354 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3356 return io_prep_rw(req, sqe);
3359 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3361 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3362 struct kiocb *kiocb = &req->rw.kiocb;
3363 struct iov_iter __iter, *iter = &__iter;
3364 struct io_async_rw *rw = req->async_data;
3365 ssize_t ret, ret2, io_size;
3366 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3372 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3376 io_size = iov_iter_count(iter);
3377 req->result = io_size;
3379 /* Ensure we clear previously set non-block flag */
3380 if (!force_nonblock)
3381 kiocb->ki_flags &= ~IOCB_NOWAIT;
3383 kiocb->ki_flags |= IOCB_NOWAIT;
3385 /* If the file doesn't support async, just async punt */
3386 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3389 /* file path doesn't support NOWAIT for non-direct_IO */
3390 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3391 (req->flags & REQ_F_ISREG))
3394 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3399 * Open-code file_start_write here to grab freeze protection,
3400 * which will be released by another thread in
3401 * io_complete_rw(). Fool lockdep by telling it the lock got
3402 * released so that it doesn't complain about the held lock when
3403 * we return to userspace.
3405 if (req->flags & REQ_F_ISREG) {
3406 sb_start_write(file_inode(req->file)->i_sb);
3407 __sb_writers_release(file_inode(req->file)->i_sb,
3410 kiocb->ki_flags |= IOCB_WRITE;
3412 if (req->file->f_op->write_iter)
3413 ret2 = call_write_iter(req->file, kiocb, iter);
3414 else if (req->file->f_op->write)
3415 ret2 = loop_rw_iter(WRITE, req, iter);
3420 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3421 * retry them without IOCB_NOWAIT.
3423 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3425 /* no retry on NONBLOCK nor RWF_NOWAIT */
3426 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3428 if (ret2 == -EIOCBQUEUED && req->async_data)
3429 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3430 if (!force_nonblock || ret2 != -EAGAIN) {
3431 /* IOPOLL retry should happen for io-wq threads */
3432 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3435 kiocb_done(kiocb, ret2, issue_flags);
3438 /* some cases will consume bytes even on error returns */
3439 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3440 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3441 return ret ?: -EAGAIN;
3444 /* it's reportedly faster than delegating the null check to kfree() */
3450 static int io_renameat_prep(struct io_kiocb *req,
3451 const struct io_uring_sqe *sqe)
3453 struct io_rename *ren = &req->rename;
3454 const char __user *oldf, *newf;
3456 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3459 ren->old_dfd = READ_ONCE(sqe->fd);
3460 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3461 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3462 ren->new_dfd = READ_ONCE(sqe->len);
3463 ren->flags = READ_ONCE(sqe->rename_flags);
3465 ren->oldpath = getname(oldf);
3466 if (IS_ERR(ren->oldpath))
3467 return PTR_ERR(ren->oldpath);
3469 ren->newpath = getname(newf);
3470 if (IS_ERR(ren->newpath)) {
3471 putname(ren->oldpath);
3472 return PTR_ERR(ren->newpath);
3475 req->flags |= REQ_F_NEED_CLEANUP;
3479 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3481 struct io_rename *ren = &req->rename;
3484 if (issue_flags & IO_URING_F_NONBLOCK)
3487 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3488 ren->newpath, ren->flags);
3490 req->flags &= ~REQ_F_NEED_CLEANUP;
3492 req_set_fail_links(req);
3493 io_req_complete(req, ret);
3497 static int io_unlinkat_prep(struct io_kiocb *req,
3498 const struct io_uring_sqe *sqe)
3500 struct io_unlink *un = &req->unlink;
3501 const char __user *fname;
3503 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3506 un->dfd = READ_ONCE(sqe->fd);
3508 un->flags = READ_ONCE(sqe->unlink_flags);
3509 if (un->flags & ~AT_REMOVEDIR)
3512 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3513 un->filename = getname(fname);
3514 if (IS_ERR(un->filename))
3515 return PTR_ERR(un->filename);
3517 req->flags |= REQ_F_NEED_CLEANUP;
3521 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3523 struct io_unlink *un = &req->unlink;
3526 if (issue_flags & IO_URING_F_NONBLOCK)
3529 if (un->flags & AT_REMOVEDIR)
3530 ret = do_rmdir(un->dfd, un->filename);
3532 ret = do_unlinkat(un->dfd, un->filename);
3534 req->flags &= ~REQ_F_NEED_CLEANUP;
3536 req_set_fail_links(req);
3537 io_req_complete(req, ret);
3541 static int io_shutdown_prep(struct io_kiocb *req,
3542 const struct io_uring_sqe *sqe)
3544 #if defined(CONFIG_NET)
3545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3547 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3551 req->shutdown.how = READ_ONCE(sqe->len);
3558 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3560 #if defined(CONFIG_NET)
3561 struct socket *sock;
3564 if (issue_flags & IO_URING_F_NONBLOCK)
3567 sock = sock_from_file(req->file);
3568 if (unlikely(!sock))
3571 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3573 req_set_fail_links(req);
3574 io_req_complete(req, ret);
3581 static int __io_splice_prep(struct io_kiocb *req,
3582 const struct io_uring_sqe *sqe)
3584 struct io_splice* sp = &req->splice;
3585 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3591 sp->len = READ_ONCE(sqe->len);
3592 sp->flags = READ_ONCE(sqe->splice_flags);
3594 if (unlikely(sp->flags & ~valid_flags))
3597 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3598 (sp->flags & SPLICE_F_FD_IN_FIXED));
3601 req->flags |= REQ_F_NEED_CLEANUP;
3603 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3605 * Splice operation will be punted aync, and here need to
3606 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3608 req->work.flags |= IO_WQ_WORK_UNBOUND;
3614 static int io_tee_prep(struct io_kiocb *req,
3615 const struct io_uring_sqe *sqe)
3617 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3619 return __io_splice_prep(req, sqe);
3622 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3624 struct io_splice *sp = &req->splice;
3625 struct file *in = sp->file_in;
3626 struct file *out = sp->file_out;
3627 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3630 if (issue_flags & IO_URING_F_NONBLOCK)
3633 ret = do_tee(in, out, sp->len, flags);
3635 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3636 req->flags &= ~REQ_F_NEED_CLEANUP;
3639 req_set_fail_links(req);
3640 io_req_complete(req, ret);
3644 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3646 struct io_splice* sp = &req->splice;
3648 sp->off_in = READ_ONCE(sqe->splice_off_in);
3649 sp->off_out = READ_ONCE(sqe->off);
3650 return __io_splice_prep(req, sqe);
3653 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3655 struct io_splice *sp = &req->splice;
3656 struct file *in = sp->file_in;
3657 struct file *out = sp->file_out;
3658 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3659 loff_t *poff_in, *poff_out;
3662 if (issue_flags & IO_URING_F_NONBLOCK)
3665 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3666 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3669 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3671 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3672 req->flags &= ~REQ_F_NEED_CLEANUP;
3675 req_set_fail_links(req);
3676 io_req_complete(req, ret);
3681 * IORING_OP_NOP just posts a completion event, nothing else.
3683 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3685 struct io_ring_ctx *ctx = req->ctx;
3687 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3690 __io_req_complete(req, issue_flags, 0, 0);
3694 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3696 struct io_ring_ctx *ctx = req->ctx;
3701 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3703 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3706 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3707 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3710 req->sync.off = READ_ONCE(sqe->off);
3711 req->sync.len = READ_ONCE(sqe->len);
3715 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3717 loff_t end = req->sync.off + req->sync.len;
3720 /* fsync always requires a blocking context */
3721 if (issue_flags & IO_URING_F_NONBLOCK)
3724 ret = vfs_fsync_range(req->file, req->sync.off,
3725 end > 0 ? end : LLONG_MAX,
3726 req->sync.flags & IORING_FSYNC_DATASYNC);
3728 req_set_fail_links(req);
3729 io_req_complete(req, ret);
3733 static int io_fallocate_prep(struct io_kiocb *req,
3734 const struct io_uring_sqe *sqe)
3736 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3738 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3741 req->sync.off = READ_ONCE(sqe->off);
3742 req->sync.len = READ_ONCE(sqe->addr);
3743 req->sync.mode = READ_ONCE(sqe->len);
3747 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3751 /* fallocate always requiring blocking context */
3752 if (issue_flags & IO_URING_F_NONBLOCK)
3754 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3757 req_set_fail_links(req);
3758 io_req_complete(req, ret);
3762 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3764 const char __user *fname;
3767 if (unlikely(sqe->ioprio || sqe->buf_index))
3769 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3772 /* open.how should be already initialised */
3773 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3774 req->open.how.flags |= O_LARGEFILE;
3776 req->open.dfd = READ_ONCE(sqe->fd);
3777 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3778 req->open.filename = getname(fname);
3779 if (IS_ERR(req->open.filename)) {
3780 ret = PTR_ERR(req->open.filename);
3781 req->open.filename = NULL;
3784 req->open.nofile = rlimit(RLIMIT_NOFILE);
3785 req->flags |= REQ_F_NEED_CLEANUP;
3789 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3795 mode = READ_ONCE(sqe->len);
3796 flags = READ_ONCE(sqe->open_flags);
3797 req->open.how = build_open_how(flags, mode);
3798 return __io_openat_prep(req, sqe);
3801 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3803 struct open_how __user *how;
3807 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3809 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3810 len = READ_ONCE(sqe->len);
3811 if (len < OPEN_HOW_SIZE_VER0)
3814 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3819 return __io_openat_prep(req, sqe);
3822 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3824 struct open_flags op;
3827 bool resolve_nonblock;
3830 ret = build_open_flags(&req->open.how, &op);
3833 nonblock_set = op.open_flag & O_NONBLOCK;
3834 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3835 if (issue_flags & IO_URING_F_NONBLOCK) {
3837 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3838 * it'll always -EAGAIN
3840 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3842 op.lookup_flags |= LOOKUP_CACHED;
3843 op.open_flag |= O_NONBLOCK;
3846 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3850 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3851 /* only retry if RESOLVE_CACHED wasn't already set by application */
3852 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3853 file == ERR_PTR(-EAGAIN)) {
3855 * We could hang on to this 'fd', but seems like marginal
3856 * gain for something that is now known to be a slower path.
3857 * So just put it, and we'll get a new one when we retry.
3865 ret = PTR_ERR(file);
3867 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3868 file->f_flags &= ~O_NONBLOCK;
3869 fsnotify_open(file);
3870 fd_install(ret, file);
3873 putname(req->open.filename);
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3876 req_set_fail_links(req);
3877 io_req_complete(req, ret);
3881 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3883 return io_openat2(req, issue_flags);
3886 static int io_remove_buffers_prep(struct io_kiocb *req,
3887 const struct io_uring_sqe *sqe)
3889 struct io_provide_buf *p = &req->pbuf;
3892 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3895 tmp = READ_ONCE(sqe->fd);
3896 if (!tmp || tmp > USHRT_MAX)
3899 memset(p, 0, sizeof(*p));
3901 p->bgid = READ_ONCE(sqe->buf_group);
3905 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3906 int bgid, unsigned nbufs)
3910 /* shouldn't happen */
3914 /* the head kbuf is the list itself */
3915 while (!list_empty(&buf->list)) {
3916 struct io_buffer *nxt;
3918 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3919 list_del(&nxt->list);
3926 xa_erase(&ctx->io_buffers, bgid);
3931 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3933 struct io_provide_buf *p = &req->pbuf;
3934 struct io_ring_ctx *ctx = req->ctx;
3935 struct io_buffer *head;
3937 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3939 io_ring_submit_lock(ctx, !force_nonblock);
3941 lockdep_assert_held(&ctx->uring_lock);
3944 head = xa_load(&ctx->io_buffers, p->bgid);
3946 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3948 req_set_fail_links(req);
3950 /* need to hold the lock to complete IOPOLL requests */
3951 if (ctx->flags & IORING_SETUP_IOPOLL) {
3952 __io_req_complete(req, issue_flags, ret, 0);
3953 io_ring_submit_unlock(ctx, !force_nonblock);
3955 io_ring_submit_unlock(ctx, !force_nonblock);
3956 __io_req_complete(req, issue_flags, ret, 0);
3961 static int io_provide_buffers_prep(struct io_kiocb *req,
3962 const struct io_uring_sqe *sqe)
3964 struct io_provide_buf *p = &req->pbuf;
3967 if (sqe->ioprio || sqe->rw_flags)
3970 tmp = READ_ONCE(sqe->fd);
3971 if (!tmp || tmp > USHRT_MAX)
3974 p->addr = READ_ONCE(sqe->addr);
3975 p->len = READ_ONCE(sqe->len);
3977 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3980 p->bgid = READ_ONCE(sqe->buf_group);
3981 tmp = READ_ONCE(sqe->off);
3982 if (tmp > USHRT_MAX)
3988 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3990 struct io_buffer *buf;
3991 u64 addr = pbuf->addr;
3992 int i, bid = pbuf->bid;
3994 for (i = 0; i < pbuf->nbufs; i++) {
3995 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4000 buf->len = pbuf->len;
4005 INIT_LIST_HEAD(&buf->list);
4008 list_add_tail(&buf->list, &(*head)->list);
4012 return i ? i : -ENOMEM;
4015 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4017 struct io_provide_buf *p = &req->pbuf;
4018 struct io_ring_ctx *ctx = req->ctx;
4019 struct io_buffer *head, *list;
4021 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4023 io_ring_submit_lock(ctx, !force_nonblock);
4025 lockdep_assert_held(&ctx->uring_lock);
4027 list = head = xa_load(&ctx->io_buffers, p->bgid);
4029 ret = io_add_buffers(p, &head);
4030 if (ret >= 0 && !list) {
4031 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4033 __io_remove_buffers(ctx, head, p->bgid, -1U);
4036 req_set_fail_links(req);
4038 /* need to hold the lock to complete IOPOLL requests */
4039 if (ctx->flags & IORING_SETUP_IOPOLL) {
4040 __io_req_complete(req, issue_flags, ret, 0);
4041 io_ring_submit_unlock(ctx, !force_nonblock);
4043 io_ring_submit_unlock(ctx, !force_nonblock);
4044 __io_req_complete(req, issue_flags, ret, 0);
4049 static int io_epoll_ctl_prep(struct io_kiocb *req,
4050 const struct io_uring_sqe *sqe)
4052 #if defined(CONFIG_EPOLL)
4053 if (sqe->ioprio || sqe->buf_index)
4055 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4058 req->epoll.epfd = READ_ONCE(sqe->fd);
4059 req->epoll.op = READ_ONCE(sqe->len);
4060 req->epoll.fd = READ_ONCE(sqe->off);
4062 if (ep_op_has_event(req->epoll.op)) {
4063 struct epoll_event __user *ev;
4065 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4066 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4076 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4078 #if defined(CONFIG_EPOLL)
4079 struct io_epoll *ie = &req->epoll;
4081 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4083 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4084 if (force_nonblock && ret == -EAGAIN)
4088 req_set_fail_links(req);
4089 __io_req_complete(req, issue_flags, ret, 0);
4096 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4098 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4099 if (sqe->ioprio || sqe->buf_index || sqe->off)
4101 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4104 req->madvise.addr = READ_ONCE(sqe->addr);
4105 req->madvise.len = READ_ONCE(sqe->len);
4106 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4113 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 struct io_madvise *ma = &req->madvise;
4119 if (issue_flags & IO_URING_F_NONBLOCK)
4122 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4124 req_set_fail_links(req);
4125 io_req_complete(req, ret);
4132 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4134 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4136 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4139 req->fadvise.offset = READ_ONCE(sqe->off);
4140 req->fadvise.len = READ_ONCE(sqe->len);
4141 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4145 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4147 struct io_fadvise *fa = &req->fadvise;
4150 if (issue_flags & IO_URING_F_NONBLOCK) {
4151 switch (fa->advice) {
4152 case POSIX_FADV_NORMAL:
4153 case POSIX_FADV_RANDOM:
4154 case POSIX_FADV_SEQUENTIAL:
4161 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4163 req_set_fail_links(req);
4164 io_req_complete(req, ret);
4168 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4170 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4172 if (sqe->ioprio || sqe->buf_index)
4174 if (req->flags & REQ_F_FIXED_FILE)
4177 req->statx.dfd = READ_ONCE(sqe->fd);
4178 req->statx.mask = READ_ONCE(sqe->len);
4179 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4180 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4181 req->statx.flags = READ_ONCE(sqe->statx_flags);
4186 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4188 struct io_statx *ctx = &req->statx;
4191 if (issue_flags & IO_URING_F_NONBLOCK) {
4192 /* only need file table for an actual valid fd */
4193 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4194 req->flags |= REQ_F_NO_FILE_TABLE;
4198 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4202 req_set_fail_links(req);
4203 io_req_complete(req, ret);
4207 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4209 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4211 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4212 sqe->rw_flags || sqe->buf_index)
4214 if (req->flags & REQ_F_FIXED_FILE)
4217 req->close.fd = READ_ONCE(sqe->fd);
4221 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4223 struct files_struct *files = current->files;
4224 struct io_close *close = &req->close;
4225 struct fdtable *fdt;
4231 spin_lock(&files->file_lock);
4232 fdt = files_fdtable(files);
4233 if (close->fd >= fdt->max_fds) {
4234 spin_unlock(&files->file_lock);
4237 file = fdt->fd[close->fd];
4239 spin_unlock(&files->file_lock);
4243 if (file->f_op == &io_uring_fops) {
4244 spin_unlock(&files->file_lock);
4249 /* if the file has a flush method, be safe and punt to async */
4250 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4251 spin_unlock(&files->file_lock);
4255 ret = __close_fd_get_file(close->fd, &file);
4256 spin_unlock(&files->file_lock);
4263 /* No ->flush() or already async, safely close from here */
4264 ret = filp_close(file, current->files);
4267 req_set_fail_links(req);
4270 __io_req_complete(req, issue_flags, ret, 0);
4274 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4276 struct io_ring_ctx *ctx = req->ctx;
4278 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4280 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4283 req->sync.off = READ_ONCE(sqe->off);
4284 req->sync.len = READ_ONCE(sqe->len);
4285 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4289 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4293 /* sync_file_range always requires a blocking context */
4294 if (issue_flags & IO_URING_F_NONBLOCK)
4297 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4300 req_set_fail_links(req);
4301 io_req_complete(req, ret);
4305 #if defined(CONFIG_NET)
4306 static int io_setup_async_msg(struct io_kiocb *req,
4307 struct io_async_msghdr *kmsg)
4309 struct io_async_msghdr *async_msg = req->async_data;
4313 if (io_alloc_async_data(req)) {
4314 kfree(kmsg->free_iov);
4317 async_msg = req->async_data;
4318 req->flags |= REQ_F_NEED_CLEANUP;
4319 memcpy(async_msg, kmsg, sizeof(*kmsg));
4320 async_msg->msg.msg_name = &async_msg->addr;
4321 /* if were using fast_iov, set it to the new one */
4322 if (!async_msg->free_iov)
4323 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4328 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4329 struct io_async_msghdr *iomsg)
4331 iomsg->msg.msg_name = &iomsg->addr;
4332 iomsg->free_iov = iomsg->fast_iov;
4333 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4334 req->sr_msg.msg_flags, &iomsg->free_iov);
4337 static int io_sendmsg_prep_async(struct io_kiocb *req)
4341 if (!io_op_defs[req->opcode].needs_async_data)
4343 ret = io_sendmsg_copy_hdr(req, req->async_data);
4345 req->flags |= REQ_F_NEED_CLEANUP;
4349 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4351 struct io_sr_msg *sr = &req->sr_msg;
4353 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4356 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4357 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4358 sr->len = READ_ONCE(sqe->len);
4360 #ifdef CONFIG_COMPAT
4361 if (req->ctx->compat)
4362 sr->msg_flags |= MSG_CMSG_COMPAT;
4367 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4369 struct io_async_msghdr iomsg, *kmsg;
4370 struct socket *sock;
4374 sock = sock_from_file(req->file);
4375 if (unlikely(!sock))
4378 kmsg = req->async_data;
4380 ret = io_sendmsg_copy_hdr(req, &iomsg);
4386 flags = req->sr_msg.msg_flags;
4387 if (flags & MSG_DONTWAIT)
4388 req->flags |= REQ_F_NOWAIT;
4389 else if (issue_flags & IO_URING_F_NONBLOCK)
4390 flags |= MSG_DONTWAIT;
4392 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4393 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4394 return io_setup_async_msg(req, kmsg);
4395 if (ret == -ERESTARTSYS)
4398 /* fast path, check for non-NULL to avoid function call */
4400 kfree(kmsg->free_iov);
4401 req->flags &= ~REQ_F_NEED_CLEANUP;
4403 req_set_fail_links(req);
4404 __io_req_complete(req, issue_flags, ret, 0);
4408 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4410 struct io_sr_msg *sr = &req->sr_msg;
4413 struct socket *sock;
4417 sock = sock_from_file(req->file);
4418 if (unlikely(!sock))
4421 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4425 msg.msg_name = NULL;
4426 msg.msg_control = NULL;
4427 msg.msg_controllen = 0;
4428 msg.msg_namelen = 0;
4430 flags = req->sr_msg.msg_flags;
4431 if (flags & MSG_DONTWAIT)
4432 req->flags |= REQ_F_NOWAIT;
4433 else if (issue_flags & IO_URING_F_NONBLOCK)
4434 flags |= MSG_DONTWAIT;
4436 msg.msg_flags = flags;
4437 ret = sock_sendmsg(sock, &msg);
4438 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4440 if (ret == -ERESTARTSYS)
4444 req_set_fail_links(req);
4445 __io_req_complete(req, issue_flags, ret, 0);
4449 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4450 struct io_async_msghdr *iomsg)
4452 struct io_sr_msg *sr = &req->sr_msg;
4453 struct iovec __user *uiov;
4457 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4458 &iomsg->uaddr, &uiov, &iov_len);
4462 if (req->flags & REQ_F_BUFFER_SELECT) {
4465 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4467 sr->len = iomsg->fast_iov[0].iov_len;
4468 iomsg->free_iov = NULL;
4470 iomsg->free_iov = iomsg->fast_iov;
4471 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4472 &iomsg->free_iov, &iomsg->msg.msg_iter,
4481 #ifdef CONFIG_COMPAT
4482 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4483 struct io_async_msghdr *iomsg)
4485 struct compat_msghdr __user *msg_compat;
4486 struct io_sr_msg *sr = &req->sr_msg;
4487 struct compat_iovec __user *uiov;
4492 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4493 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4498 uiov = compat_ptr(ptr);
4499 if (req->flags & REQ_F_BUFFER_SELECT) {
4500 compat_ssize_t clen;
4504 if (!access_ok(uiov, sizeof(*uiov)))
4506 if (__get_user(clen, &uiov->iov_len))
4511 iomsg->free_iov = NULL;
4513 iomsg->free_iov = iomsg->fast_iov;
4514 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4515 UIO_FASTIOV, &iomsg->free_iov,
4516 &iomsg->msg.msg_iter, true);
4525 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4526 struct io_async_msghdr *iomsg)
4528 iomsg->msg.msg_name = &iomsg->addr;
4530 #ifdef CONFIG_COMPAT
4531 if (req->ctx->compat)
4532 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4535 return __io_recvmsg_copy_hdr(req, iomsg);
4538 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4541 struct io_sr_msg *sr = &req->sr_msg;
4542 struct io_buffer *kbuf;
4544 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4549 req->flags |= REQ_F_BUFFER_SELECTED;
4553 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4555 return io_put_kbuf(req, req->sr_msg.kbuf);
4558 static int io_recvmsg_prep_async(struct io_kiocb *req)
4562 if (!io_op_defs[req->opcode].needs_async_data)
4564 ret = io_recvmsg_copy_hdr(req, req->async_data);
4566 req->flags |= REQ_F_NEED_CLEANUP;
4570 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4572 struct io_sr_msg *sr = &req->sr_msg;
4574 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4577 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4578 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4579 sr->len = READ_ONCE(sqe->len);
4580 sr->bgid = READ_ONCE(sqe->buf_group);
4582 #ifdef CONFIG_COMPAT
4583 if (req->ctx->compat)
4584 sr->msg_flags |= MSG_CMSG_COMPAT;
4589 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4591 struct io_async_msghdr iomsg, *kmsg;
4592 struct socket *sock;
4593 struct io_buffer *kbuf;
4595 int ret, cflags = 0;
4596 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4598 sock = sock_from_file(req->file);
4599 if (unlikely(!sock))
4602 kmsg = req->async_data;
4604 ret = io_recvmsg_copy_hdr(req, &iomsg);
4610 if (req->flags & REQ_F_BUFFER_SELECT) {
4611 kbuf = io_recv_buffer_select(req, !force_nonblock);
4613 return PTR_ERR(kbuf);
4614 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4615 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4616 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4617 1, req->sr_msg.len);
4620 flags = req->sr_msg.msg_flags;
4621 if (flags & MSG_DONTWAIT)
4622 req->flags |= REQ_F_NOWAIT;
4623 else if (force_nonblock)
4624 flags |= MSG_DONTWAIT;
4626 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4627 kmsg->uaddr, flags);
4628 if (force_nonblock && ret == -EAGAIN)
4629 return io_setup_async_msg(req, kmsg);
4630 if (ret == -ERESTARTSYS)
4633 if (req->flags & REQ_F_BUFFER_SELECTED)
4634 cflags = io_put_recv_kbuf(req);
4635 /* fast path, check for non-NULL to avoid function call */
4637 kfree(kmsg->free_iov);
4638 req->flags &= ~REQ_F_NEED_CLEANUP;
4640 req_set_fail_links(req);
4641 __io_req_complete(req, issue_flags, ret, cflags);
4645 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4647 struct io_buffer *kbuf;
4648 struct io_sr_msg *sr = &req->sr_msg;
4650 void __user *buf = sr->buf;
4651 struct socket *sock;
4654 int ret, cflags = 0;
4655 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4657 sock = sock_from_file(req->file);
4658 if (unlikely(!sock))
4661 if (req->flags & REQ_F_BUFFER_SELECT) {
4662 kbuf = io_recv_buffer_select(req, !force_nonblock);
4664 return PTR_ERR(kbuf);
4665 buf = u64_to_user_ptr(kbuf->addr);
4668 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4672 msg.msg_name = NULL;
4673 msg.msg_control = NULL;
4674 msg.msg_controllen = 0;
4675 msg.msg_namelen = 0;
4676 msg.msg_iocb = NULL;
4679 flags = req->sr_msg.msg_flags;
4680 if (flags & MSG_DONTWAIT)
4681 req->flags |= REQ_F_NOWAIT;
4682 else if (force_nonblock)
4683 flags |= MSG_DONTWAIT;
4685 ret = sock_recvmsg(sock, &msg, flags);
4686 if (force_nonblock && ret == -EAGAIN)
4688 if (ret == -ERESTARTSYS)
4691 if (req->flags & REQ_F_BUFFER_SELECTED)
4692 cflags = io_put_recv_kbuf(req);
4694 req_set_fail_links(req);
4695 __io_req_complete(req, issue_flags, ret, cflags);
4699 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4701 struct io_accept *accept = &req->accept;
4703 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4705 if (sqe->ioprio || sqe->len || sqe->buf_index)
4708 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4709 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4710 accept->flags = READ_ONCE(sqe->accept_flags);
4711 accept->nofile = rlimit(RLIMIT_NOFILE);
4715 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4717 struct io_accept *accept = &req->accept;
4718 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4719 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4722 if (req->file->f_flags & O_NONBLOCK)
4723 req->flags |= REQ_F_NOWAIT;
4725 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4726 accept->addr_len, accept->flags,
4728 if (ret == -EAGAIN && force_nonblock)
4731 if (ret == -ERESTARTSYS)
4733 req_set_fail_links(req);
4735 __io_req_complete(req, issue_flags, ret, 0);
4739 static int io_connect_prep_async(struct io_kiocb *req)
4741 struct io_async_connect *io = req->async_data;
4742 struct io_connect *conn = &req->connect;
4744 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4747 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4749 struct io_connect *conn = &req->connect;
4751 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4753 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4756 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4757 conn->addr_len = READ_ONCE(sqe->addr2);
4761 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4763 struct io_async_connect __io, *io;
4764 unsigned file_flags;
4766 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4768 if (req->async_data) {
4769 io = req->async_data;
4771 ret = move_addr_to_kernel(req->connect.addr,
4772 req->connect.addr_len,
4779 file_flags = force_nonblock ? O_NONBLOCK : 0;
4781 ret = __sys_connect_file(req->file, &io->address,
4782 req->connect.addr_len, file_flags);
4783 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4784 if (req->async_data)
4786 if (io_alloc_async_data(req)) {
4790 io = req->async_data;
4791 memcpy(req->async_data, &__io, sizeof(__io));
4794 if (ret == -ERESTARTSYS)
4798 req_set_fail_links(req);
4799 __io_req_complete(req, issue_flags, ret, 0);
4802 #else /* !CONFIG_NET */
4803 #define IO_NETOP_FN(op) \
4804 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4806 return -EOPNOTSUPP; \
4809 #define IO_NETOP_PREP(op) \
4811 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4813 return -EOPNOTSUPP; \
4816 #define IO_NETOP_PREP_ASYNC(op) \
4818 static int io_##op##_prep_async(struct io_kiocb *req) \
4820 return -EOPNOTSUPP; \
4823 IO_NETOP_PREP_ASYNC(sendmsg);
4824 IO_NETOP_PREP_ASYNC(recvmsg);
4825 IO_NETOP_PREP_ASYNC(connect);
4826 IO_NETOP_PREP(accept);
4829 #endif /* CONFIG_NET */
4831 struct io_poll_table {
4832 struct poll_table_struct pt;
4833 struct io_kiocb *req;
4837 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4838 __poll_t mask, task_work_func_t func)
4842 /* for instances that support it check for an event match first: */
4843 if (mask && !(mask & poll->events))
4846 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4848 list_del_init(&poll->wait.entry);
4851 req->task_work.func = func;
4852 percpu_ref_get(&req->ctx->refs);
4855 * If this fails, then the task is exiting. When a task exits, the
4856 * work gets canceled, so just cancel this request as well instead
4857 * of executing it. We can't safely execute it anyway, as we may not
4858 * have the needed state needed for it anyway.
4860 ret = io_req_task_work_add(req);
4861 if (unlikely(ret)) {
4862 WRITE_ONCE(poll->canceled, true);
4863 io_req_task_work_add_fallback(req, func);
4868 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4869 __acquires(&req->ctx->completion_lock)
4871 struct io_ring_ctx *ctx = req->ctx;
4873 if (!req->result && !READ_ONCE(poll->canceled)) {
4874 struct poll_table_struct pt = { ._key = poll->events };
4876 req->result = vfs_poll(req->file, &pt) & poll->events;
4879 spin_lock_irq(&ctx->completion_lock);
4880 if (!req->result && !READ_ONCE(poll->canceled)) {
4881 add_wait_queue(poll->head, &poll->wait);
4888 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4890 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4891 if (req->opcode == IORING_OP_POLL_ADD)
4892 return req->async_data;
4893 return req->apoll->double_poll;
4896 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4898 if (req->opcode == IORING_OP_POLL_ADD)
4900 return &req->apoll->poll;
4903 static void io_poll_remove_double(struct io_kiocb *req)
4905 struct io_poll_iocb *poll = io_poll_get_double(req);
4907 lockdep_assert_held(&req->ctx->completion_lock);
4909 if (poll && poll->head) {
4910 struct wait_queue_head *head = poll->head;
4912 spin_lock(&head->lock);
4913 list_del_init(&poll->wait.entry);
4914 if (poll->wait.private)
4915 refcount_dec(&req->refs);
4917 spin_unlock(&head->lock);
4921 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4923 struct io_ring_ctx *ctx = req->ctx;
4925 io_poll_remove_double(req);
4926 req->poll.done = true;
4927 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4928 io_commit_cqring(ctx);
4931 static void io_poll_task_func(struct callback_head *cb)
4933 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4934 struct io_ring_ctx *ctx = req->ctx;
4935 struct io_kiocb *nxt;
4937 if (io_poll_rewait(req, &req->poll)) {
4938 spin_unlock_irq(&ctx->completion_lock);
4940 hash_del(&req->hash_node);
4941 io_poll_complete(req, req->result, 0);
4942 spin_unlock_irq(&ctx->completion_lock);
4944 nxt = io_put_req_find_next(req);
4945 io_cqring_ev_posted(ctx);
4947 __io_req_task_submit(nxt);
4950 percpu_ref_put(&ctx->refs);
4953 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4954 int sync, void *key)
4956 struct io_kiocb *req = wait->private;
4957 struct io_poll_iocb *poll = io_poll_get_single(req);
4958 __poll_t mask = key_to_poll(key);
4960 /* for instances that support it check for an event match first: */
4961 if (mask && !(mask & poll->events))
4964 list_del_init(&wait->entry);
4966 if (poll && poll->head) {
4969 spin_lock(&poll->head->lock);
4970 done = list_empty(&poll->wait.entry);
4972 list_del_init(&poll->wait.entry);
4973 /* make sure double remove sees this as being gone */
4974 wait->private = NULL;
4975 spin_unlock(&poll->head->lock);
4977 /* use wait func handler, so it matches the rq type */
4978 poll->wait.func(&poll->wait, mode, sync, key);
4981 refcount_dec(&req->refs);
4985 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4986 wait_queue_func_t wake_func)
4990 poll->canceled = false;
4991 poll->events = events;
4992 INIT_LIST_HEAD(&poll->wait.entry);
4993 init_waitqueue_func_entry(&poll->wait, wake_func);
4996 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4997 struct wait_queue_head *head,
4998 struct io_poll_iocb **poll_ptr)
5000 struct io_kiocb *req = pt->req;
5003 * If poll->head is already set, it's because the file being polled
5004 * uses multiple waitqueues for poll handling (eg one for read, one
5005 * for write). Setup a separate io_poll_iocb if this happens.
5007 if (unlikely(poll->head)) {
5008 struct io_poll_iocb *poll_one = poll;
5010 /* already have a 2nd entry, fail a third attempt */
5012 pt->error = -EINVAL;
5015 /* double add on the same waitqueue head, ignore */
5016 if (poll->head == head)
5018 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5020 pt->error = -ENOMEM;
5023 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5024 refcount_inc(&req->refs);
5025 poll->wait.private = req;
5032 if (poll->events & EPOLLEXCLUSIVE)
5033 add_wait_queue_exclusive(head, &poll->wait);
5035 add_wait_queue(head, &poll->wait);
5038 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5039 struct poll_table_struct *p)
5041 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5042 struct async_poll *apoll = pt->req->apoll;
5044 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5047 static void io_async_task_func(struct callback_head *cb)
5049 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5050 struct async_poll *apoll = req->apoll;
5051 struct io_ring_ctx *ctx = req->ctx;
5053 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5055 if (io_poll_rewait(req, &apoll->poll)) {
5056 spin_unlock_irq(&ctx->completion_lock);
5057 percpu_ref_put(&ctx->refs);
5061 /* If req is still hashed, it cannot have been canceled. Don't check. */
5062 if (hash_hashed(&req->hash_node))
5063 hash_del(&req->hash_node);
5065 io_poll_remove_double(req);
5066 spin_unlock_irq(&ctx->completion_lock);
5068 if (!READ_ONCE(apoll->poll.canceled))
5069 __io_req_task_submit(req);
5071 __io_req_task_cancel(req, -ECANCELED);
5073 percpu_ref_put(&ctx->refs);
5074 kfree(apoll->double_poll);
5078 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5081 struct io_kiocb *req = wait->private;
5082 struct io_poll_iocb *poll = &req->apoll->poll;
5084 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5087 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5090 static void io_poll_req_insert(struct io_kiocb *req)
5092 struct io_ring_ctx *ctx = req->ctx;
5093 struct hlist_head *list;
5095 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5096 hlist_add_head(&req->hash_node, list);
5099 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5100 struct io_poll_iocb *poll,
5101 struct io_poll_table *ipt, __poll_t mask,
5102 wait_queue_func_t wake_func)
5103 __acquires(&ctx->completion_lock)
5105 struct io_ring_ctx *ctx = req->ctx;
5106 bool cancel = false;
5108 INIT_HLIST_NODE(&req->hash_node);
5109 io_init_poll_iocb(poll, mask, wake_func);
5110 poll->file = req->file;
5111 poll->wait.private = req;
5113 ipt->pt._key = mask;
5115 ipt->error = -EINVAL;
5117 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5119 spin_lock_irq(&ctx->completion_lock);
5120 if (likely(poll->head)) {
5121 spin_lock(&poll->head->lock);
5122 if (unlikely(list_empty(&poll->wait.entry))) {
5128 if (mask || ipt->error)
5129 list_del_init(&poll->wait.entry);
5131 WRITE_ONCE(poll->canceled, true);
5132 else if (!poll->done) /* actually waiting for an event */
5133 io_poll_req_insert(req);
5134 spin_unlock(&poll->head->lock);
5140 static bool io_arm_poll_handler(struct io_kiocb *req)
5142 const struct io_op_def *def = &io_op_defs[req->opcode];
5143 struct io_ring_ctx *ctx = req->ctx;
5144 struct async_poll *apoll;
5145 struct io_poll_table ipt;
5149 if (!req->file || !file_can_poll(req->file))
5151 if (req->flags & REQ_F_POLLED)
5155 else if (def->pollout)
5159 /* if we can't nonblock try, then no point in arming a poll handler */
5160 if (!io_file_supports_async(req->file, rw))
5163 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5164 if (unlikely(!apoll))
5166 apoll->double_poll = NULL;
5168 req->flags |= REQ_F_POLLED;
5173 mask |= POLLIN | POLLRDNORM;
5175 mask |= POLLOUT | POLLWRNORM;
5177 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5178 if ((req->opcode == IORING_OP_RECVMSG) &&
5179 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5182 mask |= POLLERR | POLLPRI;
5184 ipt.pt._qproc = io_async_queue_proc;
5186 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5188 if (ret || ipt.error) {
5189 io_poll_remove_double(req);
5190 spin_unlock_irq(&ctx->completion_lock);
5191 kfree(apoll->double_poll);
5195 spin_unlock_irq(&ctx->completion_lock);
5196 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5197 apoll->poll.events);
5201 static bool __io_poll_remove_one(struct io_kiocb *req,
5202 struct io_poll_iocb *poll)
5204 bool do_complete = false;
5206 spin_lock(&poll->head->lock);
5207 WRITE_ONCE(poll->canceled, true);
5208 if (!list_empty(&poll->wait.entry)) {
5209 list_del_init(&poll->wait.entry);
5212 spin_unlock(&poll->head->lock);
5213 hash_del(&req->hash_node);
5217 static bool io_poll_remove_one(struct io_kiocb *req)
5221 io_poll_remove_double(req);
5223 if (req->opcode == IORING_OP_POLL_ADD) {
5224 do_complete = __io_poll_remove_one(req, &req->poll);
5226 struct async_poll *apoll = req->apoll;
5228 /* non-poll requests have submit ref still */
5229 do_complete = __io_poll_remove_one(req, &apoll->poll);
5232 kfree(apoll->double_poll);
5238 io_cqring_fill_event(req, -ECANCELED);
5239 io_commit_cqring(req->ctx);
5240 req_set_fail_links(req);
5241 io_put_req_deferred(req, 1);
5248 * Returns true if we found and killed one or more poll requests
5250 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5251 struct files_struct *files)
5253 struct hlist_node *tmp;
5254 struct io_kiocb *req;
5257 spin_lock_irq(&ctx->completion_lock);
5258 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5259 struct hlist_head *list;
5261 list = &ctx->cancel_hash[i];
5262 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5263 if (io_match_task(req, tsk, files))
5264 posted += io_poll_remove_one(req);
5267 spin_unlock_irq(&ctx->completion_lock);
5270 io_cqring_ev_posted(ctx);
5275 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5277 struct hlist_head *list;
5278 struct io_kiocb *req;
5280 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5281 hlist_for_each_entry(req, list, hash_node) {
5282 if (sqe_addr != req->user_data)
5284 if (io_poll_remove_one(req))
5292 static int io_poll_remove_prep(struct io_kiocb *req,
5293 const struct io_uring_sqe *sqe)
5295 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5297 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5301 req->poll_remove.addr = READ_ONCE(sqe->addr);
5306 * Find a running poll command that matches one specified in sqe->addr,
5307 * and remove it if found.
5309 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5311 struct io_ring_ctx *ctx = req->ctx;
5314 spin_lock_irq(&ctx->completion_lock);
5315 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5316 spin_unlock_irq(&ctx->completion_lock);
5319 req_set_fail_links(req);
5320 io_req_complete(req, ret);
5324 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5327 struct io_kiocb *req = wait->private;
5328 struct io_poll_iocb *poll = &req->poll;
5330 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5333 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5334 struct poll_table_struct *p)
5336 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5338 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5341 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5343 struct io_poll_iocb *poll = &req->poll;
5346 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5348 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5351 events = READ_ONCE(sqe->poll32_events);
5353 events = swahw32(events);
5355 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5356 (events & EPOLLEXCLUSIVE);
5360 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5362 struct io_poll_iocb *poll = &req->poll;
5363 struct io_ring_ctx *ctx = req->ctx;
5364 struct io_poll_table ipt;
5367 ipt.pt._qproc = io_poll_queue_proc;
5369 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5372 if (mask) { /* no async, we'd stolen it */
5374 io_poll_complete(req, mask, 0);
5376 spin_unlock_irq(&ctx->completion_lock);
5379 io_cqring_ev_posted(ctx);
5385 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5387 struct io_timeout_data *data = container_of(timer,
5388 struct io_timeout_data, timer);
5389 struct io_kiocb *req = data->req;
5390 struct io_ring_ctx *ctx = req->ctx;
5391 unsigned long flags;
5393 spin_lock_irqsave(&ctx->completion_lock, flags);
5394 list_del_init(&req->timeout.list);
5395 atomic_set(&req->ctx->cq_timeouts,
5396 atomic_read(&req->ctx->cq_timeouts) + 1);
5398 io_cqring_fill_event(req, -ETIME);
5399 io_commit_cqring(ctx);
5400 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5402 io_cqring_ev_posted(ctx);
5403 req_set_fail_links(req);
5405 return HRTIMER_NORESTART;
5408 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5411 struct io_timeout_data *io;
5412 struct io_kiocb *req;
5415 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5416 if (user_data == req->user_data) {
5423 return ERR_PTR(ret);
5425 io = req->async_data;
5426 ret = hrtimer_try_to_cancel(&io->timer);
5428 return ERR_PTR(-EALREADY);
5429 list_del_init(&req->timeout.list);
5433 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5435 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5438 return PTR_ERR(req);
5440 req_set_fail_links(req);
5441 io_cqring_fill_event(req, -ECANCELED);
5442 io_put_req_deferred(req, 1);
5446 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5447 struct timespec64 *ts, enum hrtimer_mode mode)
5449 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5450 struct io_timeout_data *data;
5453 return PTR_ERR(req);
5455 req->timeout.off = 0; /* noseq */
5456 data = req->async_data;
5457 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5458 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5459 data->timer.function = io_timeout_fn;
5460 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5464 static int io_timeout_remove_prep(struct io_kiocb *req,
5465 const struct io_uring_sqe *sqe)
5467 struct io_timeout_rem *tr = &req->timeout_rem;
5469 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5471 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5473 if (sqe->ioprio || sqe->buf_index || sqe->len)
5476 tr->addr = READ_ONCE(sqe->addr);
5477 tr->flags = READ_ONCE(sqe->timeout_flags);
5478 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5479 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5481 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5483 } else if (tr->flags) {
5484 /* timeout removal doesn't support flags */
5491 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5493 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5498 * Remove or update an existing timeout command
5500 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5502 struct io_timeout_rem *tr = &req->timeout_rem;
5503 struct io_ring_ctx *ctx = req->ctx;
5506 spin_lock_irq(&ctx->completion_lock);
5507 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5508 ret = io_timeout_cancel(ctx, tr->addr);
5510 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5511 io_translate_timeout_mode(tr->flags));
5513 io_cqring_fill_event(req, ret);
5514 io_commit_cqring(ctx);
5515 spin_unlock_irq(&ctx->completion_lock);
5516 io_cqring_ev_posted(ctx);
5518 req_set_fail_links(req);
5523 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5524 bool is_timeout_link)
5526 struct io_timeout_data *data;
5528 u32 off = READ_ONCE(sqe->off);
5530 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5532 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5534 if (off && is_timeout_link)
5536 flags = READ_ONCE(sqe->timeout_flags);
5537 if (flags & ~IORING_TIMEOUT_ABS)
5540 req->timeout.off = off;
5542 if (!req->async_data && io_alloc_async_data(req))
5545 data = req->async_data;
5548 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5551 data->mode = io_translate_timeout_mode(flags);
5552 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5553 io_req_track_inflight(req);
5557 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5559 struct io_ring_ctx *ctx = req->ctx;
5560 struct io_timeout_data *data = req->async_data;
5561 struct list_head *entry;
5562 u32 tail, off = req->timeout.off;
5564 spin_lock_irq(&ctx->completion_lock);
5567 * sqe->off holds how many events that need to occur for this
5568 * timeout event to be satisfied. If it isn't set, then this is
5569 * a pure timeout request, sequence isn't used.
5571 if (io_is_timeout_noseq(req)) {
5572 entry = ctx->timeout_list.prev;
5576 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5577 req->timeout.target_seq = tail + off;
5579 /* Update the last seq here in case io_flush_timeouts() hasn't.
5580 * This is safe because ->completion_lock is held, and submissions
5581 * and completions are never mixed in the same ->completion_lock section.
5583 ctx->cq_last_tm_flush = tail;
5586 * Insertion sort, ensuring the first entry in the list is always
5587 * the one we need first.
5589 list_for_each_prev(entry, &ctx->timeout_list) {
5590 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5593 if (io_is_timeout_noseq(nxt))
5595 /* nxt.seq is behind @tail, otherwise would've been completed */
5596 if (off >= nxt->timeout.target_seq - tail)
5600 list_add(&req->timeout.list, entry);
5601 data->timer.function = io_timeout_fn;
5602 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5603 spin_unlock_irq(&ctx->completion_lock);
5607 struct io_cancel_data {
5608 struct io_ring_ctx *ctx;
5612 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5614 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5615 struct io_cancel_data *cd = data;
5617 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5620 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5621 struct io_ring_ctx *ctx)
5623 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5624 enum io_wq_cancel cancel_ret;
5627 if (!tctx || !tctx->io_wq)
5630 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5631 switch (cancel_ret) {
5632 case IO_WQ_CANCEL_OK:
5635 case IO_WQ_CANCEL_RUNNING:
5638 case IO_WQ_CANCEL_NOTFOUND:
5646 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5647 struct io_kiocb *req, __u64 sqe_addr,
5650 unsigned long flags;
5653 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5654 if (ret != -ENOENT) {
5655 spin_lock_irqsave(&ctx->completion_lock, flags);
5659 spin_lock_irqsave(&ctx->completion_lock, flags);
5660 ret = io_timeout_cancel(ctx, sqe_addr);
5663 ret = io_poll_cancel(ctx, sqe_addr);
5667 io_cqring_fill_event(req, ret);
5668 io_commit_cqring(ctx);
5669 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5670 io_cqring_ev_posted(ctx);
5673 req_set_fail_links(req);
5677 static int io_async_cancel_prep(struct io_kiocb *req,
5678 const struct io_uring_sqe *sqe)
5680 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5682 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5684 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5687 req->cancel.addr = READ_ONCE(sqe->addr);
5691 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5693 struct io_ring_ctx *ctx = req->ctx;
5694 u64 sqe_addr = req->cancel.addr;
5695 struct io_tctx_node *node;
5698 /* tasks should wait for their io-wq threads, so safe w/o sync */
5699 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5700 spin_lock_irq(&ctx->completion_lock);
5703 ret = io_timeout_cancel(ctx, sqe_addr);
5706 ret = io_poll_cancel(ctx, sqe_addr);
5709 spin_unlock_irq(&ctx->completion_lock);
5711 /* slow path, try all io-wq's */
5712 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5714 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5715 struct io_uring_task *tctx = node->task->io_uring;
5717 if (!tctx || !tctx->io_wq)
5719 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5723 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5725 spin_lock_irq(&ctx->completion_lock);
5727 io_cqring_fill_event(req, ret);
5728 io_commit_cqring(ctx);
5729 spin_unlock_irq(&ctx->completion_lock);
5730 io_cqring_ev_posted(ctx);
5733 req_set_fail_links(req);
5738 static int io_rsrc_update_prep(struct io_kiocb *req,
5739 const struct io_uring_sqe *sqe)
5741 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5743 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5745 if (sqe->ioprio || sqe->rw_flags)
5748 req->rsrc_update.offset = READ_ONCE(sqe->off);
5749 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5750 if (!req->rsrc_update.nr_args)
5752 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5756 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5758 struct io_ring_ctx *ctx = req->ctx;
5759 struct io_uring_rsrc_update up;
5762 if (issue_flags & IO_URING_F_NONBLOCK)
5765 up.offset = req->rsrc_update.offset;
5766 up.data = req->rsrc_update.arg;
5768 mutex_lock(&ctx->uring_lock);
5769 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5770 mutex_unlock(&ctx->uring_lock);
5773 req_set_fail_links(req);
5774 __io_req_complete(req, issue_flags, ret, 0);
5778 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5780 switch (req->opcode) {
5783 case IORING_OP_READV:
5784 case IORING_OP_READ_FIXED:
5785 case IORING_OP_READ:
5786 return io_read_prep(req, sqe);
5787 case IORING_OP_WRITEV:
5788 case IORING_OP_WRITE_FIXED:
5789 case IORING_OP_WRITE:
5790 return io_write_prep(req, sqe);
5791 case IORING_OP_POLL_ADD:
5792 return io_poll_add_prep(req, sqe);
5793 case IORING_OP_POLL_REMOVE:
5794 return io_poll_remove_prep(req, sqe);
5795 case IORING_OP_FSYNC:
5796 return io_fsync_prep(req, sqe);
5797 case IORING_OP_SYNC_FILE_RANGE:
5798 return io_sfr_prep(req, sqe);
5799 case IORING_OP_SENDMSG:
5800 case IORING_OP_SEND:
5801 return io_sendmsg_prep(req, sqe);
5802 case IORING_OP_RECVMSG:
5803 case IORING_OP_RECV:
5804 return io_recvmsg_prep(req, sqe);
5805 case IORING_OP_CONNECT:
5806 return io_connect_prep(req, sqe);
5807 case IORING_OP_TIMEOUT:
5808 return io_timeout_prep(req, sqe, false);
5809 case IORING_OP_TIMEOUT_REMOVE:
5810 return io_timeout_remove_prep(req, sqe);
5811 case IORING_OP_ASYNC_CANCEL:
5812 return io_async_cancel_prep(req, sqe);
5813 case IORING_OP_LINK_TIMEOUT:
5814 return io_timeout_prep(req, sqe, true);
5815 case IORING_OP_ACCEPT:
5816 return io_accept_prep(req, sqe);
5817 case IORING_OP_FALLOCATE:
5818 return io_fallocate_prep(req, sqe);
5819 case IORING_OP_OPENAT:
5820 return io_openat_prep(req, sqe);
5821 case IORING_OP_CLOSE:
5822 return io_close_prep(req, sqe);
5823 case IORING_OP_FILES_UPDATE:
5824 return io_rsrc_update_prep(req, sqe);
5825 case IORING_OP_STATX:
5826 return io_statx_prep(req, sqe);
5827 case IORING_OP_FADVISE:
5828 return io_fadvise_prep(req, sqe);
5829 case IORING_OP_MADVISE:
5830 return io_madvise_prep(req, sqe);
5831 case IORING_OP_OPENAT2:
5832 return io_openat2_prep(req, sqe);
5833 case IORING_OP_EPOLL_CTL:
5834 return io_epoll_ctl_prep(req, sqe);
5835 case IORING_OP_SPLICE:
5836 return io_splice_prep(req, sqe);
5837 case IORING_OP_PROVIDE_BUFFERS:
5838 return io_provide_buffers_prep(req, sqe);
5839 case IORING_OP_REMOVE_BUFFERS:
5840 return io_remove_buffers_prep(req, sqe);
5842 return io_tee_prep(req, sqe);
5843 case IORING_OP_SHUTDOWN:
5844 return io_shutdown_prep(req, sqe);
5845 case IORING_OP_RENAMEAT:
5846 return io_renameat_prep(req, sqe);
5847 case IORING_OP_UNLINKAT:
5848 return io_unlinkat_prep(req, sqe);
5851 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5856 static int io_req_prep_async(struct io_kiocb *req)
5858 switch (req->opcode) {
5859 case IORING_OP_READV:
5860 case IORING_OP_READ_FIXED:
5861 case IORING_OP_READ:
5862 return io_rw_prep_async(req, READ);
5863 case IORING_OP_WRITEV:
5864 case IORING_OP_WRITE_FIXED:
5865 case IORING_OP_WRITE:
5866 return io_rw_prep_async(req, WRITE);
5867 case IORING_OP_SENDMSG:
5868 case IORING_OP_SEND:
5869 return io_sendmsg_prep_async(req);
5870 case IORING_OP_RECVMSG:
5871 case IORING_OP_RECV:
5872 return io_recvmsg_prep_async(req);
5873 case IORING_OP_CONNECT:
5874 return io_connect_prep_async(req);
5879 static int io_req_defer_prep(struct io_kiocb *req)
5881 if (!io_op_defs[req->opcode].needs_async_data)
5883 /* some opcodes init it during the inital prep */
5884 if (req->async_data)
5886 if (__io_alloc_async_data(req))
5888 return io_req_prep_async(req);
5891 static u32 io_get_sequence(struct io_kiocb *req)
5893 struct io_kiocb *pos;
5894 struct io_ring_ctx *ctx = req->ctx;
5895 u32 total_submitted, nr_reqs = 0;
5897 io_for_each_link(pos, req)
5900 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5901 return total_submitted - nr_reqs;
5904 static int io_req_defer(struct io_kiocb *req)
5906 struct io_ring_ctx *ctx = req->ctx;
5907 struct io_defer_entry *de;
5911 /* Still need defer if there is pending req in defer list. */
5912 if (likely(list_empty_careful(&ctx->defer_list) &&
5913 !(req->flags & REQ_F_IO_DRAIN)))
5916 seq = io_get_sequence(req);
5917 /* Still a chance to pass the sequence check */
5918 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5921 ret = io_req_defer_prep(req);
5924 io_prep_async_link(req);
5925 de = kmalloc(sizeof(*de), GFP_KERNEL);
5929 spin_lock_irq(&ctx->completion_lock);
5930 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5931 spin_unlock_irq(&ctx->completion_lock);
5933 io_queue_async_work(req);
5934 return -EIOCBQUEUED;
5937 trace_io_uring_defer(ctx, req, req->user_data);
5940 list_add_tail(&de->list, &ctx->defer_list);
5941 spin_unlock_irq(&ctx->completion_lock);
5942 return -EIOCBQUEUED;
5945 static void __io_clean_op(struct io_kiocb *req)
5947 if (req->flags & REQ_F_BUFFER_SELECTED) {
5948 switch (req->opcode) {
5949 case IORING_OP_READV:
5950 case IORING_OP_READ_FIXED:
5951 case IORING_OP_READ:
5952 kfree((void *)(unsigned long)req->rw.addr);
5954 case IORING_OP_RECVMSG:
5955 case IORING_OP_RECV:
5956 kfree(req->sr_msg.kbuf);
5959 req->flags &= ~REQ_F_BUFFER_SELECTED;
5962 if (req->flags & REQ_F_NEED_CLEANUP) {
5963 switch (req->opcode) {
5964 case IORING_OP_READV:
5965 case IORING_OP_READ_FIXED:
5966 case IORING_OP_READ:
5967 case IORING_OP_WRITEV:
5968 case IORING_OP_WRITE_FIXED:
5969 case IORING_OP_WRITE: {
5970 struct io_async_rw *io = req->async_data;
5972 kfree(io->free_iovec);
5975 case IORING_OP_RECVMSG:
5976 case IORING_OP_SENDMSG: {
5977 struct io_async_msghdr *io = req->async_data;
5979 kfree(io->free_iov);
5982 case IORING_OP_SPLICE:
5984 io_put_file(req, req->splice.file_in,
5985 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5987 case IORING_OP_OPENAT:
5988 case IORING_OP_OPENAT2:
5989 if (req->open.filename)
5990 putname(req->open.filename);
5992 case IORING_OP_RENAMEAT:
5993 putname(req->rename.oldpath);
5994 putname(req->rename.newpath);
5996 case IORING_OP_UNLINKAT:
5997 putname(req->unlink.filename);
6000 req->flags &= ~REQ_F_NEED_CLEANUP;
6004 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6006 struct io_ring_ctx *ctx = req->ctx;
6007 const struct cred *creds = NULL;
6010 if (req->work.creds && req->work.creds != current_cred())
6011 creds = override_creds(req->work.creds);
6013 switch (req->opcode) {
6015 ret = io_nop(req, issue_flags);
6017 case IORING_OP_READV:
6018 case IORING_OP_READ_FIXED:
6019 case IORING_OP_READ:
6020 ret = io_read(req, issue_flags);
6022 case IORING_OP_WRITEV:
6023 case IORING_OP_WRITE_FIXED:
6024 case IORING_OP_WRITE:
6025 ret = io_write(req, issue_flags);
6027 case IORING_OP_FSYNC:
6028 ret = io_fsync(req, issue_flags);
6030 case IORING_OP_POLL_ADD:
6031 ret = io_poll_add(req, issue_flags);
6033 case IORING_OP_POLL_REMOVE:
6034 ret = io_poll_remove(req, issue_flags);
6036 case IORING_OP_SYNC_FILE_RANGE:
6037 ret = io_sync_file_range(req, issue_flags);
6039 case IORING_OP_SENDMSG:
6040 ret = io_sendmsg(req, issue_flags);
6042 case IORING_OP_SEND:
6043 ret = io_send(req, issue_flags);
6045 case IORING_OP_RECVMSG:
6046 ret = io_recvmsg(req, issue_flags);
6048 case IORING_OP_RECV:
6049 ret = io_recv(req, issue_flags);
6051 case IORING_OP_TIMEOUT:
6052 ret = io_timeout(req, issue_flags);
6054 case IORING_OP_TIMEOUT_REMOVE:
6055 ret = io_timeout_remove(req, issue_flags);
6057 case IORING_OP_ACCEPT:
6058 ret = io_accept(req, issue_flags);
6060 case IORING_OP_CONNECT:
6061 ret = io_connect(req, issue_flags);
6063 case IORING_OP_ASYNC_CANCEL:
6064 ret = io_async_cancel(req, issue_flags);
6066 case IORING_OP_FALLOCATE:
6067 ret = io_fallocate(req, issue_flags);
6069 case IORING_OP_OPENAT:
6070 ret = io_openat(req, issue_flags);
6072 case IORING_OP_CLOSE:
6073 ret = io_close(req, issue_flags);
6075 case IORING_OP_FILES_UPDATE:
6076 ret = io_files_update(req, issue_flags);
6078 case IORING_OP_STATX:
6079 ret = io_statx(req, issue_flags);
6081 case IORING_OP_FADVISE:
6082 ret = io_fadvise(req, issue_flags);
6084 case IORING_OP_MADVISE:
6085 ret = io_madvise(req, issue_flags);
6087 case IORING_OP_OPENAT2:
6088 ret = io_openat2(req, issue_flags);
6090 case IORING_OP_EPOLL_CTL:
6091 ret = io_epoll_ctl(req, issue_flags);
6093 case IORING_OP_SPLICE:
6094 ret = io_splice(req, issue_flags);
6096 case IORING_OP_PROVIDE_BUFFERS:
6097 ret = io_provide_buffers(req, issue_flags);
6099 case IORING_OP_REMOVE_BUFFERS:
6100 ret = io_remove_buffers(req, issue_flags);
6103 ret = io_tee(req, issue_flags);
6105 case IORING_OP_SHUTDOWN:
6106 ret = io_shutdown(req, issue_flags);
6108 case IORING_OP_RENAMEAT:
6109 ret = io_renameat(req, issue_flags);
6111 case IORING_OP_UNLINKAT:
6112 ret = io_unlinkat(req, issue_flags);
6120 revert_creds(creds);
6125 /* If the op doesn't have a file, we're not polling for it */
6126 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6127 const bool in_async = io_wq_current_is_worker();
6129 /* workqueue context doesn't hold uring_lock, grab it now */
6131 mutex_lock(&ctx->uring_lock);
6133 io_iopoll_req_issued(req, in_async);
6136 mutex_unlock(&ctx->uring_lock);
6142 static void io_wq_submit_work(struct io_wq_work *work)
6144 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6145 struct io_kiocb *timeout;
6148 timeout = io_prep_linked_timeout(req);
6150 io_queue_linked_timeout(timeout);
6152 if (work->flags & IO_WQ_WORK_CANCEL)
6157 ret = io_issue_sqe(req, 0);
6159 * We can get EAGAIN for polled IO even though we're
6160 * forcing a sync submission from here, since we can't
6161 * wait for request slots on the block side.
6169 /* avoid locking problems by failing it from a clean context */
6171 /* io-wq is going to take one down */
6172 refcount_inc(&req->refs);
6173 io_req_task_queue_fail(req, ret);
6177 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6180 struct fixed_rsrc_table *table;
6182 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6183 return table->files[index & IORING_FILE_TABLE_MASK];
6186 static struct file *io_file_get(struct io_submit_state *state,
6187 struct io_kiocb *req, int fd, bool fixed)
6189 struct io_ring_ctx *ctx = req->ctx;
6193 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6195 fd = array_index_nospec(fd, ctx->nr_user_files);
6196 file = io_file_from_index(ctx, fd);
6197 io_set_resource_node(req);
6199 trace_io_uring_file_get(ctx, fd);
6200 file = __io_file_get(state, fd);
6203 if (file && unlikely(file->f_op == &io_uring_fops))
6204 io_req_track_inflight(req);
6208 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6210 struct io_timeout_data *data = container_of(timer,
6211 struct io_timeout_data, timer);
6212 struct io_kiocb *prev, *req = data->req;
6213 struct io_ring_ctx *ctx = req->ctx;
6214 unsigned long flags;
6216 spin_lock_irqsave(&ctx->completion_lock, flags);
6217 prev = req->timeout.head;
6218 req->timeout.head = NULL;
6221 * We don't expect the list to be empty, that will only happen if we
6222 * race with the completion of the linked work.
6224 if (prev && refcount_inc_not_zero(&prev->refs))
6225 io_remove_next_linked(prev);
6228 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6231 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6232 io_put_req_deferred(prev, 1);
6234 io_req_complete_post(req, -ETIME, 0);
6235 io_put_req_deferred(req, 1);
6237 return HRTIMER_NORESTART;
6240 static void __io_queue_linked_timeout(struct io_kiocb *req)
6243 * If the back reference is NULL, then our linked request finished
6244 * before we got a chance to setup the timer
6246 if (req->timeout.head) {
6247 struct io_timeout_data *data = req->async_data;
6249 data->timer.function = io_link_timeout_fn;
6250 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6255 static void io_queue_linked_timeout(struct io_kiocb *req)
6257 struct io_ring_ctx *ctx = req->ctx;
6259 spin_lock_irq(&ctx->completion_lock);
6260 __io_queue_linked_timeout(req);
6261 spin_unlock_irq(&ctx->completion_lock);
6263 /* drop submission reference */
6267 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6269 struct io_kiocb *nxt = req->link;
6271 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6272 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6275 nxt->timeout.head = req;
6276 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6277 req->flags |= REQ_F_LINK_TIMEOUT;
6281 static void __io_queue_sqe(struct io_kiocb *req)
6283 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6286 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6289 * We async punt it if the file wasn't marked NOWAIT, or if the file
6290 * doesn't support non-blocking read/write attempts
6292 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6293 if (!io_arm_poll_handler(req)) {
6295 * Queued up for async execution, worker will release
6296 * submit reference when the iocb is actually submitted.
6298 io_queue_async_work(req);
6300 } else if (likely(!ret)) {
6301 /* drop submission reference */
6302 if (req->flags & REQ_F_COMPLETE_INLINE) {
6303 struct io_ring_ctx *ctx = req->ctx;
6304 struct io_comp_state *cs = &ctx->submit_state.comp;
6306 cs->reqs[cs->nr++] = req;
6307 if (cs->nr == ARRAY_SIZE(cs->reqs))
6308 io_submit_flush_completions(cs, ctx);
6313 req_set_fail_links(req);
6315 io_req_complete(req, ret);
6318 io_queue_linked_timeout(linked_timeout);
6321 static void io_queue_sqe(struct io_kiocb *req)
6325 ret = io_req_defer(req);
6327 if (ret != -EIOCBQUEUED) {
6329 req_set_fail_links(req);
6331 io_req_complete(req, ret);
6333 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6334 ret = io_req_defer_prep(req);
6337 io_queue_async_work(req);
6339 __io_queue_sqe(req);
6344 * Check SQE restrictions (opcode and flags).
6346 * Returns 'true' if SQE is allowed, 'false' otherwise.
6348 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6349 struct io_kiocb *req,
6350 unsigned int sqe_flags)
6352 if (!ctx->restricted)
6355 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6358 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6359 ctx->restrictions.sqe_flags_required)
6362 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6363 ctx->restrictions.sqe_flags_required))
6369 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6370 const struct io_uring_sqe *sqe)
6372 struct io_submit_state *state;
6373 unsigned int sqe_flags;
6374 int personality, ret = 0;
6376 req->opcode = READ_ONCE(sqe->opcode);
6377 /* same numerical values with corresponding REQ_F_*, safe to copy */
6378 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6379 req->user_data = READ_ONCE(sqe->user_data);
6380 req->async_data = NULL;
6384 req->fixed_rsrc_refs = NULL;
6385 /* one is dropped after submission, the other at completion */
6386 refcount_set(&req->refs, 2);
6387 req->task = current;
6389 req->work.list.next = NULL;
6390 req->work.creds = NULL;
6391 req->work.flags = 0;
6393 /* enforce forwards compatibility on users */
6394 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6399 if (unlikely(req->opcode >= IORING_OP_LAST))
6402 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6405 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6406 !io_op_defs[req->opcode].buffer_select)
6409 personality = READ_ONCE(sqe->personality);
6411 req->work.creds = xa_load(&ctx->personalities, personality);
6412 if (!req->work.creds)
6414 get_cred(req->work.creds);
6416 state = &ctx->submit_state;
6419 * Plug now if we have more than 1 IO left after this, and the target
6420 * is potentially a read/write to block based storage.
6422 if (!state->plug_started && state->ios_left > 1 &&
6423 io_op_defs[req->opcode].plug) {
6424 blk_start_plug(&state->plug);
6425 state->plug_started = true;
6428 if (io_op_defs[req->opcode].needs_file) {
6429 bool fixed = req->flags & REQ_F_FIXED_FILE;
6431 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6432 if (unlikely(!req->file))
6440 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6441 const struct io_uring_sqe *sqe)
6443 struct io_submit_link *link = &ctx->submit_state.link;
6446 ret = io_init_req(ctx, req, sqe);
6447 if (unlikely(ret)) {
6450 io_req_complete(req, ret);
6452 /* fail even hard links since we don't submit */
6453 link->head->flags |= REQ_F_FAIL_LINK;
6454 io_put_req(link->head);
6455 io_req_complete(link->head, -ECANCELED);
6460 ret = io_req_prep(req, sqe);
6464 /* don't need @sqe from now on */
6465 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6466 true, ctx->flags & IORING_SETUP_SQPOLL);
6469 * If we already have a head request, queue this one for async
6470 * submittal once the head completes. If we don't have a head but
6471 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6472 * submitted sync once the chain is complete. If none of those
6473 * conditions are true (normal request), then just queue it.
6476 struct io_kiocb *head = link->head;
6479 * Taking sequential execution of a link, draining both sides
6480 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6481 * requests in the link. So, it drains the head and the
6482 * next after the link request. The last one is done via
6483 * drain_next flag to persist the effect across calls.
6485 if (req->flags & REQ_F_IO_DRAIN) {
6486 head->flags |= REQ_F_IO_DRAIN;
6487 ctx->drain_next = 1;
6489 ret = io_req_defer_prep(req);
6492 trace_io_uring_link(ctx, req, head);
6493 link->last->link = req;
6496 /* last request of a link, enqueue the link */
6497 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6502 if (unlikely(ctx->drain_next)) {
6503 req->flags |= REQ_F_IO_DRAIN;
6504 ctx->drain_next = 0;
6506 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6518 * Batched submission is done, ensure local IO is flushed out.
6520 static void io_submit_state_end(struct io_submit_state *state,
6521 struct io_ring_ctx *ctx)
6523 if (state->link.head)
6524 io_queue_sqe(state->link.head);
6526 io_submit_flush_completions(&state->comp, ctx);
6527 if (state->plug_started)
6528 blk_finish_plug(&state->plug);
6529 io_state_file_put(state);
6533 * Start submission side cache.
6535 static void io_submit_state_start(struct io_submit_state *state,
6536 unsigned int max_ios)
6538 state->plug_started = false;
6539 state->ios_left = max_ios;
6540 /* set only head, no need to init link_last in advance */
6541 state->link.head = NULL;
6544 static void io_commit_sqring(struct io_ring_ctx *ctx)
6546 struct io_rings *rings = ctx->rings;
6549 * Ensure any loads from the SQEs are done at this point,
6550 * since once we write the new head, the application could
6551 * write new data to them.
6553 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6557 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6558 * that is mapped by userspace. This means that care needs to be taken to
6559 * ensure that reads are stable, as we cannot rely on userspace always
6560 * being a good citizen. If members of the sqe are validated and then later
6561 * used, it's important that those reads are done through READ_ONCE() to
6562 * prevent a re-load down the line.
6564 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6566 u32 *sq_array = ctx->sq_array;
6570 * The cached sq head (or cq tail) serves two purposes:
6572 * 1) allows us to batch the cost of updating the user visible
6574 * 2) allows the kernel side to track the head on its own, even
6575 * though the application is the one updating it.
6577 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6578 if (likely(head < ctx->sq_entries))
6579 return &ctx->sq_sqes[head];
6581 /* drop invalid entries */
6582 ctx->cached_sq_dropped++;
6583 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6587 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6591 /* if we have a backlog and couldn't flush it all, return BUSY */
6592 if (test_bit(0, &ctx->sq_check_overflow)) {
6593 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6597 /* make sure SQ entry isn't read before tail */
6598 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6600 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6603 percpu_counter_add(¤t->io_uring->inflight, nr);
6604 refcount_add(nr, ¤t->usage);
6605 io_submit_state_start(&ctx->submit_state, nr);
6607 while (submitted < nr) {
6608 const struct io_uring_sqe *sqe;
6609 struct io_kiocb *req;
6611 req = io_alloc_req(ctx);
6612 if (unlikely(!req)) {
6614 submitted = -EAGAIN;
6617 sqe = io_get_sqe(ctx);
6618 if (unlikely(!sqe)) {
6619 kmem_cache_free(req_cachep, req);
6622 /* will complete beyond this point, count as submitted */
6624 if (io_submit_sqe(ctx, req, sqe))
6628 if (unlikely(submitted != nr)) {
6629 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6630 struct io_uring_task *tctx = current->io_uring;
6631 int unused = nr - ref_used;
6633 percpu_ref_put_many(&ctx->refs, unused);
6634 percpu_counter_sub(&tctx->inflight, unused);
6635 put_task_struct_many(current, unused);
6638 io_submit_state_end(&ctx->submit_state, ctx);
6639 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6640 io_commit_sqring(ctx);
6645 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6647 /* Tell userspace we may need a wakeup call */
6648 spin_lock_irq(&ctx->completion_lock);
6649 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6650 spin_unlock_irq(&ctx->completion_lock);
6653 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6655 spin_lock_irq(&ctx->completion_lock);
6656 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6657 spin_unlock_irq(&ctx->completion_lock);
6660 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6662 unsigned int to_submit;
6665 to_submit = io_sqring_entries(ctx);
6666 /* if we're handling multiple rings, cap submit size for fairness */
6667 if (cap_entries && to_submit > 8)
6670 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6671 unsigned nr_events = 0;
6673 mutex_lock(&ctx->uring_lock);
6674 if (!list_empty(&ctx->iopoll_list))
6675 io_do_iopoll(ctx, &nr_events, 0);
6677 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6678 !(ctx->flags & IORING_SETUP_R_DISABLED))
6679 ret = io_submit_sqes(ctx, to_submit);
6680 mutex_unlock(&ctx->uring_lock);
6683 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6684 wake_up(&ctx->sqo_sq_wait);
6689 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6691 struct io_ring_ctx *ctx;
6692 unsigned sq_thread_idle = 0;
6694 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6695 if (sq_thread_idle < ctx->sq_thread_idle)
6696 sq_thread_idle = ctx->sq_thread_idle;
6699 sqd->sq_thread_idle = sq_thread_idle;
6702 static int io_sq_thread(void *data)
6704 struct io_sq_data *sqd = data;
6705 struct io_ring_ctx *ctx;
6706 unsigned long timeout = 0;
6707 char buf[TASK_COMM_LEN];
6710 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6711 set_task_comm(current, buf);
6712 current->pf_io_worker = NULL;
6714 if (sqd->sq_cpu != -1)
6715 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6717 set_cpus_allowed_ptr(current, cpu_online_mask);
6718 current->flags |= PF_NO_SETAFFINITY;
6720 mutex_lock(&sqd->lock);
6721 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6723 bool cap_entries, sqt_spin, needs_sched;
6725 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6726 mutex_unlock(&sqd->lock);
6728 mutex_lock(&sqd->lock);
6730 timeout = jiffies + sqd->sq_thread_idle;
6733 if (fatal_signal_pending(current))
6736 cap_entries = !list_is_singular(&sqd->ctx_list);
6737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6738 const struct cred *creds = NULL;
6740 if (ctx->sq_creds != current_cred())
6741 creds = override_creds(ctx->sq_creds);
6742 ret = __io_sq_thread(ctx, cap_entries);
6744 revert_creds(creds);
6745 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6749 if (sqt_spin || !time_after(jiffies, timeout)) {
6753 timeout = jiffies + sqd->sq_thread_idle;
6758 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6759 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6760 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6761 !list_empty_careful(&ctx->iopoll_list)) {
6762 needs_sched = false;
6765 if (io_sqring_entries(ctx)) {
6766 needs_sched = false;
6771 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6772 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6773 io_ring_set_wakeup_flag(ctx);
6775 mutex_unlock(&sqd->lock);
6778 mutex_lock(&sqd->lock);
6779 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6780 io_ring_clear_wakeup_flag(ctx);
6783 finish_wait(&sqd->wait, &wait);
6784 timeout = jiffies + sqd->sq_thread_idle;
6787 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6788 io_uring_cancel_sqpoll(ctx);
6790 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6791 io_ring_set_wakeup_flag(ctx);
6792 mutex_unlock(&sqd->lock);
6795 complete(&sqd->exited);
6799 struct io_wait_queue {
6800 struct wait_queue_entry wq;
6801 struct io_ring_ctx *ctx;
6803 unsigned nr_timeouts;
6806 static inline bool io_should_wake(struct io_wait_queue *iowq)
6808 struct io_ring_ctx *ctx = iowq->ctx;
6811 * Wake up if we have enough events, or if a timeout occurred since we
6812 * started waiting. For timeouts, we always want to return to userspace,
6813 * regardless of event count.
6815 return io_cqring_events(ctx) >= iowq->to_wait ||
6816 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6819 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6820 int wake_flags, void *key)
6822 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6826 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6827 * the task, and the next invocation will do it.
6829 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6830 return autoremove_wake_function(curr, mode, wake_flags, key);
6834 static int io_run_task_work_sig(void)
6836 if (io_run_task_work())
6838 if (!signal_pending(current))
6840 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6841 return -ERESTARTSYS;
6845 /* when returns >0, the caller should retry */
6846 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6847 struct io_wait_queue *iowq,
6848 signed long *timeout)
6852 /* make sure we run task_work before checking for signals */
6853 ret = io_run_task_work_sig();
6854 if (ret || io_should_wake(iowq))
6856 /* let the caller flush overflows, retry */
6857 if (test_bit(0, &ctx->cq_check_overflow))
6860 *timeout = schedule_timeout(*timeout);
6861 return !*timeout ? -ETIME : 1;
6865 * Wait until events become available, if we don't already have some. The
6866 * application must reap them itself, as they reside on the shared cq ring.
6868 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6869 const sigset_t __user *sig, size_t sigsz,
6870 struct __kernel_timespec __user *uts)
6872 struct io_wait_queue iowq = {
6875 .func = io_wake_function,
6876 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6879 .to_wait = min_events,
6881 struct io_rings *rings = ctx->rings;
6882 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6886 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6887 if (io_cqring_events(ctx) >= min_events)
6889 if (!io_run_task_work())
6894 #ifdef CONFIG_COMPAT
6895 if (in_compat_syscall())
6896 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6900 ret = set_user_sigmask(sig, sigsz);
6907 struct timespec64 ts;
6909 if (get_timespec64(&ts, uts))
6911 timeout = timespec64_to_jiffies(&ts);
6914 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6915 trace_io_uring_cqring_wait(ctx, min_events);
6917 /* if we can't even flush overflow, don't wait for more */
6918 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6922 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6923 TASK_INTERRUPTIBLE);
6924 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6925 finish_wait(&ctx->wait, &iowq.wq);
6929 restore_saved_sigmask_unless(ret == -EINTR);
6931 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6934 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6936 #if defined(CONFIG_UNIX)
6937 if (ctx->ring_sock) {
6938 struct sock *sock = ctx->ring_sock->sk;
6939 struct sk_buff *skb;
6941 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6947 for (i = 0; i < ctx->nr_user_files; i++) {
6950 file = io_file_from_index(ctx, i);
6957 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6959 struct fixed_rsrc_data *data;
6961 data = container_of(ref, struct fixed_rsrc_data, refs);
6962 complete(&data->done);
6965 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6967 spin_lock_bh(&ctx->rsrc_ref_lock);
6970 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6972 spin_unlock_bh(&ctx->rsrc_ref_lock);
6975 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6976 struct fixed_rsrc_data *rsrc_data,
6977 struct fixed_rsrc_ref_node *ref_node)
6979 io_rsrc_ref_lock(ctx);
6980 rsrc_data->node = ref_node;
6981 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6982 io_rsrc_ref_unlock(ctx);
6983 percpu_ref_get(&rsrc_data->refs);
6986 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6988 struct fixed_rsrc_ref_node *ref_node = NULL;
6990 io_rsrc_ref_lock(ctx);
6991 ref_node = data->node;
6993 io_rsrc_ref_unlock(ctx);
6995 percpu_ref_kill(&ref_node->refs);
6998 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6999 struct io_ring_ctx *ctx,
7000 void (*rsrc_put)(struct io_ring_ctx *ctx,
7001 struct io_rsrc_put *prsrc))
7003 struct fixed_rsrc_ref_node *backup_node;
7009 data->quiesce = true;
7012 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7015 backup_node->rsrc_data = data;
7016 backup_node->rsrc_put = rsrc_put;
7018 io_sqe_rsrc_kill_node(ctx, data);
7019 percpu_ref_kill(&data->refs);
7020 flush_delayed_work(&ctx->rsrc_put_work);
7022 ret = wait_for_completion_interruptible(&data->done);
7026 percpu_ref_resurrect(&data->refs);
7027 io_sqe_rsrc_set_node(ctx, data, backup_node);
7029 reinit_completion(&data->done);
7030 mutex_unlock(&ctx->uring_lock);
7031 ret = io_run_task_work_sig();
7032 mutex_lock(&ctx->uring_lock);
7034 data->quiesce = false;
7037 destroy_fixed_rsrc_ref_node(backup_node);
7041 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7043 struct fixed_rsrc_data *data;
7045 data = kzalloc(sizeof(*data), GFP_KERNEL);
7049 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7050 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7055 init_completion(&data->done);
7059 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7061 percpu_ref_exit(&data->refs);
7066 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7068 struct fixed_rsrc_data *data = ctx->file_data;
7069 unsigned nr_tables, i;
7073 * percpu_ref_is_dying() is to stop parallel files unregister
7074 * Since we possibly drop uring lock later in this function to
7077 if (!data || percpu_ref_is_dying(&data->refs))
7079 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7083 __io_sqe_files_unregister(ctx);
7084 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7085 for (i = 0; i < nr_tables; i++)
7086 kfree(data->table[i].files);
7087 free_fixed_rsrc_data(data);
7088 ctx->file_data = NULL;
7089 ctx->nr_user_files = 0;
7093 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7094 __releases(&sqd->lock)
7096 WARN_ON_ONCE(sqd->thread == current);
7099 * Do the dance but not conditional clear_bit() because it'd race with
7100 * other threads incrementing park_pending and setting the bit.
7102 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7103 if (atomic_dec_return(&sqd->park_pending))
7104 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7105 mutex_unlock(&sqd->lock);
7108 static void io_sq_thread_park(struct io_sq_data *sqd)
7109 __acquires(&sqd->lock)
7111 WARN_ON_ONCE(sqd->thread == current);
7113 atomic_inc(&sqd->park_pending);
7114 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7115 mutex_lock(&sqd->lock);
7117 wake_up_process(sqd->thread);
7120 static void io_sq_thread_stop(struct io_sq_data *sqd)
7122 WARN_ON_ONCE(sqd->thread == current);
7124 mutex_lock(&sqd->lock);
7125 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7127 wake_up_process(sqd->thread);
7128 mutex_unlock(&sqd->lock);
7129 wait_for_completion(&sqd->exited);
7132 static void io_put_sq_data(struct io_sq_data *sqd)
7134 if (refcount_dec_and_test(&sqd->refs)) {
7135 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7137 io_sq_thread_stop(sqd);
7142 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7144 struct io_sq_data *sqd = ctx->sq_data;
7147 io_sq_thread_park(sqd);
7148 list_del_init(&ctx->sqd_list);
7149 io_sqd_update_thread_idle(sqd);
7150 io_sq_thread_unpark(sqd);
7152 io_put_sq_data(sqd);
7153 ctx->sq_data = NULL;
7155 put_cred(ctx->sq_creds);
7159 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7161 struct io_ring_ctx *ctx_attach;
7162 struct io_sq_data *sqd;
7165 f = fdget(p->wq_fd);
7167 return ERR_PTR(-ENXIO);
7168 if (f.file->f_op != &io_uring_fops) {
7170 return ERR_PTR(-EINVAL);
7173 ctx_attach = f.file->private_data;
7174 sqd = ctx_attach->sq_data;
7177 return ERR_PTR(-EINVAL);
7179 if (sqd->task_tgid != current->tgid) {
7181 return ERR_PTR(-EPERM);
7184 refcount_inc(&sqd->refs);
7189 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7192 struct io_sq_data *sqd;
7195 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7196 sqd = io_attach_sq_data(p);
7201 /* fall through for EPERM case, setup new sqd/task */
7202 if (PTR_ERR(sqd) != -EPERM)
7206 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7208 return ERR_PTR(-ENOMEM);
7210 atomic_set(&sqd->park_pending, 0);
7211 refcount_set(&sqd->refs, 1);
7212 INIT_LIST_HEAD(&sqd->ctx_list);
7213 mutex_init(&sqd->lock);
7214 init_waitqueue_head(&sqd->wait);
7215 init_completion(&sqd->exited);
7219 #if defined(CONFIG_UNIX)
7221 * Ensure the UNIX gc is aware of our file set, so we are certain that
7222 * the io_uring can be safely unregistered on process exit, even if we have
7223 * loops in the file referencing.
7225 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7227 struct sock *sk = ctx->ring_sock->sk;
7228 struct scm_fp_list *fpl;
7229 struct sk_buff *skb;
7232 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7236 skb = alloc_skb(0, GFP_KERNEL);
7245 fpl->user = get_uid(current_user());
7246 for (i = 0; i < nr; i++) {
7247 struct file *file = io_file_from_index(ctx, i + offset);
7251 fpl->fp[nr_files] = get_file(file);
7252 unix_inflight(fpl->user, fpl->fp[nr_files]);
7257 fpl->max = SCM_MAX_FD;
7258 fpl->count = nr_files;
7259 UNIXCB(skb).fp = fpl;
7260 skb->destructor = unix_destruct_scm;
7261 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7262 skb_queue_head(&sk->sk_receive_queue, skb);
7264 for (i = 0; i < nr_files; i++)
7275 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7276 * causes regular reference counting to break down. We rely on the UNIX
7277 * garbage collection to take care of this problem for us.
7279 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7281 unsigned left, total;
7285 left = ctx->nr_user_files;
7287 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7289 ret = __io_sqe_files_scm(ctx, this_files, total);
7293 total += this_files;
7299 while (total < ctx->nr_user_files) {
7300 struct file *file = io_file_from_index(ctx, total);
7310 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7316 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7317 unsigned nr_tables, unsigned nr_files)
7321 for (i = 0; i < nr_tables; i++) {
7322 struct fixed_rsrc_table *table = &file_data->table[i];
7323 unsigned this_files;
7325 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7326 table->files = kcalloc(this_files, sizeof(struct file *),
7330 nr_files -= this_files;
7336 for (i = 0; i < nr_tables; i++) {
7337 struct fixed_rsrc_table *table = &file_data->table[i];
7338 kfree(table->files);
7343 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7345 struct file *file = prsrc->file;
7346 #if defined(CONFIG_UNIX)
7347 struct sock *sock = ctx->ring_sock->sk;
7348 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7349 struct sk_buff *skb;
7352 __skb_queue_head_init(&list);
7355 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7356 * remove this entry and rearrange the file array.
7358 skb = skb_dequeue(head);
7360 struct scm_fp_list *fp;
7362 fp = UNIXCB(skb).fp;
7363 for (i = 0; i < fp->count; i++) {
7366 if (fp->fp[i] != file)
7369 unix_notinflight(fp->user, fp->fp[i]);
7370 left = fp->count - 1 - i;
7372 memmove(&fp->fp[i], &fp->fp[i + 1],
7373 left * sizeof(struct file *));
7380 __skb_queue_tail(&list, skb);
7390 __skb_queue_tail(&list, skb);
7392 skb = skb_dequeue(head);
7395 if (skb_peek(&list)) {
7396 spin_lock_irq(&head->lock);
7397 while ((skb = __skb_dequeue(&list)) != NULL)
7398 __skb_queue_tail(head, skb);
7399 spin_unlock_irq(&head->lock);
7406 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7408 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7409 struct io_ring_ctx *ctx = rsrc_data->ctx;
7410 struct io_rsrc_put *prsrc, *tmp;
7412 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7413 list_del(&prsrc->list);
7414 ref_node->rsrc_put(ctx, prsrc);
7418 percpu_ref_exit(&ref_node->refs);
7420 percpu_ref_put(&rsrc_data->refs);
7423 static void io_rsrc_put_work(struct work_struct *work)
7425 struct io_ring_ctx *ctx;
7426 struct llist_node *node;
7428 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7429 node = llist_del_all(&ctx->rsrc_put_llist);
7432 struct fixed_rsrc_ref_node *ref_node;
7433 struct llist_node *next = node->next;
7435 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7436 __io_rsrc_put_work(ref_node);
7441 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7444 struct fixed_rsrc_table *table;
7446 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7447 return &table->files[i & IORING_FILE_TABLE_MASK];
7450 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7452 struct fixed_rsrc_ref_node *ref_node;
7453 struct fixed_rsrc_data *data;
7454 struct io_ring_ctx *ctx;
7455 bool first_add = false;
7458 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7459 data = ref_node->rsrc_data;
7462 io_rsrc_ref_lock(ctx);
7463 ref_node->done = true;
7465 while (!list_empty(&ctx->rsrc_ref_list)) {
7466 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7467 struct fixed_rsrc_ref_node, node);
7468 /* recycle ref nodes in order */
7469 if (!ref_node->done)
7471 list_del(&ref_node->node);
7472 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7474 io_rsrc_ref_unlock(ctx);
7476 if (percpu_ref_is_dying(&data->refs))
7480 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7482 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7485 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7486 struct io_ring_ctx *ctx)
7488 struct fixed_rsrc_ref_node *ref_node;
7490 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7494 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7499 INIT_LIST_HEAD(&ref_node->node);
7500 INIT_LIST_HEAD(&ref_node->rsrc_list);
7501 ref_node->done = false;
7505 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7506 struct fixed_rsrc_ref_node *ref_node)
7508 ref_node->rsrc_data = ctx->file_data;
7509 ref_node->rsrc_put = io_ring_file_put;
7512 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7514 percpu_ref_exit(&ref_node->refs);
7519 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7522 __s32 __user *fds = (__s32 __user *) arg;
7523 unsigned nr_tables, i;
7525 int fd, ret = -ENOMEM;
7526 struct fixed_rsrc_ref_node *ref_node;
7527 struct fixed_rsrc_data *file_data;
7533 if (nr_args > IORING_MAX_FIXED_FILES)
7536 file_data = alloc_fixed_rsrc_data(ctx);
7539 ctx->file_data = file_data;
7541 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7542 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7544 if (!file_data->table)
7547 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7550 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7551 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7555 /* allow sparse sets */
7565 * Don't allow io_uring instances to be registered. If UNIX
7566 * isn't enabled, then this causes a reference cycle and this
7567 * instance can never get freed. If UNIX is enabled we'll
7568 * handle it just fine, but there's still no point in allowing
7569 * a ring fd as it doesn't support regular read/write anyway.
7571 if (file->f_op == &io_uring_fops) {
7575 *io_fixed_file_slot(file_data, i) = file;
7578 ret = io_sqe_files_scm(ctx);
7580 io_sqe_files_unregister(ctx);
7584 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7586 io_sqe_files_unregister(ctx);
7589 init_fixed_file_ref_node(ctx, ref_node);
7591 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7594 for (i = 0; i < ctx->nr_user_files; i++) {
7595 file = io_file_from_index(ctx, i);
7599 for (i = 0; i < nr_tables; i++)
7600 kfree(file_data->table[i].files);
7601 ctx->nr_user_files = 0;
7603 free_fixed_rsrc_data(ctx->file_data);
7604 ctx->file_data = NULL;
7608 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7611 #if defined(CONFIG_UNIX)
7612 struct sock *sock = ctx->ring_sock->sk;
7613 struct sk_buff_head *head = &sock->sk_receive_queue;
7614 struct sk_buff *skb;
7617 * See if we can merge this file into an existing skb SCM_RIGHTS
7618 * file set. If there's no room, fall back to allocating a new skb
7619 * and filling it in.
7621 spin_lock_irq(&head->lock);
7622 skb = skb_peek(head);
7624 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7626 if (fpl->count < SCM_MAX_FD) {
7627 __skb_unlink(skb, head);
7628 spin_unlock_irq(&head->lock);
7629 fpl->fp[fpl->count] = get_file(file);
7630 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7632 spin_lock_irq(&head->lock);
7633 __skb_queue_head(head, skb);
7638 spin_unlock_irq(&head->lock);
7645 return __io_sqe_files_scm(ctx, 1, index);
7651 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7653 struct io_rsrc_put *prsrc;
7654 struct fixed_rsrc_ref_node *ref_node = data->node;
7656 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7661 list_add(&prsrc->list, &ref_node->rsrc_list);
7666 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7669 return io_queue_rsrc_removal(data, (void *)file);
7672 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7673 struct io_uring_rsrc_update *up,
7676 struct fixed_rsrc_data *data = ctx->file_data;
7677 struct fixed_rsrc_ref_node *ref_node;
7678 struct file *file, **file_slot;
7682 bool needs_switch = false;
7684 if (check_add_overflow(up->offset, nr_args, &done))
7686 if (done > ctx->nr_user_files)
7689 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7692 init_fixed_file_ref_node(ctx, ref_node);
7694 fds = u64_to_user_ptr(up->data);
7695 for (done = 0; done < nr_args; done++) {
7697 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7701 if (fd == IORING_REGISTER_FILES_SKIP)
7704 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7705 file_slot = io_fixed_file_slot(ctx->file_data, i);
7708 err = io_queue_file_removal(data, *file_slot);
7712 needs_switch = true;
7721 * Don't allow io_uring instances to be registered. If
7722 * UNIX isn't enabled, then this causes a reference
7723 * cycle and this instance can never get freed. If UNIX
7724 * is enabled we'll handle it just fine, but there's
7725 * still no point in allowing a ring fd as it doesn't
7726 * support regular read/write anyway.
7728 if (file->f_op == &io_uring_fops) {
7734 err = io_sqe_file_register(ctx, file, i);
7744 percpu_ref_kill(&data->node->refs);
7745 io_sqe_rsrc_set_node(ctx, data, ref_node);
7747 destroy_fixed_rsrc_ref_node(ref_node);
7749 return done ? done : err;
7752 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7755 struct io_uring_rsrc_update up;
7757 if (!ctx->file_data)
7761 if (copy_from_user(&up, arg, sizeof(up)))
7766 return __io_sqe_files_update(ctx, &up, nr_args);
7769 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7771 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7773 req = io_put_req_find_next(req);
7774 return req ? &req->work : NULL;
7777 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7779 struct io_wq_hash *hash;
7780 struct io_wq_data data;
7781 unsigned int concurrency;
7783 hash = ctx->hash_map;
7785 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7787 return ERR_PTR(-ENOMEM);
7788 refcount_set(&hash->refs, 1);
7789 init_waitqueue_head(&hash->wait);
7790 ctx->hash_map = hash;
7794 data.free_work = io_free_work;
7795 data.do_work = io_wq_submit_work;
7797 /* Do QD, or 4 * CPUS, whatever is smallest */
7798 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7800 return io_wq_create(concurrency, &data);
7803 static int io_uring_alloc_task_context(struct task_struct *task,
7804 struct io_ring_ctx *ctx)
7806 struct io_uring_task *tctx;
7809 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7810 if (unlikely(!tctx))
7813 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7814 if (unlikely(ret)) {
7819 tctx->io_wq = io_init_wq_offload(ctx);
7820 if (IS_ERR(tctx->io_wq)) {
7821 ret = PTR_ERR(tctx->io_wq);
7822 percpu_counter_destroy(&tctx->inflight);
7828 init_waitqueue_head(&tctx->wait);
7830 atomic_set(&tctx->in_idle, 0);
7831 task->io_uring = tctx;
7832 spin_lock_init(&tctx->task_lock);
7833 INIT_WQ_LIST(&tctx->task_list);
7834 tctx->task_state = 0;
7835 init_task_work(&tctx->task_work, tctx_task_work);
7839 void __io_uring_free(struct task_struct *tsk)
7841 struct io_uring_task *tctx = tsk->io_uring;
7843 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7844 WARN_ON_ONCE(tctx->io_wq);
7846 percpu_counter_destroy(&tctx->inflight);
7848 tsk->io_uring = NULL;
7851 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7852 struct io_uring_params *p)
7856 /* Retain compatibility with failing for an invalid attach attempt */
7857 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7858 IORING_SETUP_ATTACH_WQ) {
7861 f = fdget(p->wq_fd);
7864 if (f.file->f_op != &io_uring_fops) {
7870 if (ctx->flags & IORING_SETUP_SQPOLL) {
7871 struct task_struct *tsk;
7872 struct io_sq_data *sqd;
7876 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7879 sqd = io_get_sq_data(p, &attached);
7885 ctx->sq_creds = get_current_cred();
7887 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7888 if (!ctx->sq_thread_idle)
7889 ctx->sq_thread_idle = HZ;
7892 io_sq_thread_park(sqd);
7893 /* don't attach to a dying SQPOLL thread, would be racy */
7894 if (attached && !sqd->thread) {
7897 list_add(&ctx->sqd_list, &sqd->ctx_list);
7898 io_sqd_update_thread_idle(sqd);
7900 io_sq_thread_unpark(sqd);
7903 io_put_sq_data(sqd);
7904 ctx->sq_data = NULL;
7906 } else if (attached) {
7910 if (p->flags & IORING_SETUP_SQ_AFF) {
7911 int cpu = p->sq_thread_cpu;
7914 if (cpu >= nr_cpu_ids)
7916 if (!cpu_online(cpu))
7924 sqd->task_pid = current->pid;
7925 sqd->task_tgid = current->tgid;
7926 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7933 ret = io_uring_alloc_task_context(tsk, ctx);
7934 wake_up_new_task(tsk);
7937 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7938 /* Can't have SQ_AFF without SQPOLL */
7945 io_sq_thread_finish(ctx);
7948 complete(&ctx->sq_data->exited);
7952 static inline void __io_unaccount_mem(struct user_struct *user,
7953 unsigned long nr_pages)
7955 atomic_long_sub(nr_pages, &user->locked_vm);
7958 static inline int __io_account_mem(struct user_struct *user,
7959 unsigned long nr_pages)
7961 unsigned long page_limit, cur_pages, new_pages;
7963 /* Don't allow more pages than we can safely lock */
7964 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7967 cur_pages = atomic_long_read(&user->locked_vm);
7968 new_pages = cur_pages + nr_pages;
7969 if (new_pages > page_limit)
7971 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7972 new_pages) != cur_pages);
7977 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7980 __io_unaccount_mem(ctx->user, nr_pages);
7982 if (ctx->mm_account)
7983 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7986 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7991 ret = __io_account_mem(ctx->user, nr_pages);
7996 if (ctx->mm_account)
7997 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8002 static void io_mem_free(void *ptr)
8009 page = virt_to_head_page(ptr);
8010 if (put_page_testzero(page))
8011 free_compound_page(page);
8014 static void *io_mem_alloc(size_t size)
8016 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8017 __GFP_NORETRY | __GFP_ACCOUNT;
8019 return (void *) __get_free_pages(gfp_flags, get_order(size));
8022 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8025 struct io_rings *rings;
8026 size_t off, sq_array_size;
8028 off = struct_size(rings, cqes, cq_entries);
8029 if (off == SIZE_MAX)
8033 off = ALIGN(off, SMP_CACHE_BYTES);
8041 sq_array_size = array_size(sizeof(u32), sq_entries);
8042 if (sq_array_size == SIZE_MAX)
8045 if (check_add_overflow(off, sq_array_size, &off))
8051 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8055 if (!ctx->user_bufs)
8058 for (i = 0; i < ctx->nr_user_bufs; i++) {
8059 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8061 for (j = 0; j < imu->nr_bvecs; j++)
8062 unpin_user_page(imu->bvec[j].bv_page);
8064 if (imu->acct_pages)
8065 io_unaccount_mem(ctx, imu->acct_pages);
8070 kfree(ctx->user_bufs);
8071 ctx->user_bufs = NULL;
8072 ctx->nr_user_bufs = 0;
8076 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8077 void __user *arg, unsigned index)
8079 struct iovec __user *src;
8081 #ifdef CONFIG_COMPAT
8083 struct compat_iovec __user *ciovs;
8084 struct compat_iovec ciov;
8086 ciovs = (struct compat_iovec __user *) arg;
8087 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8090 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8091 dst->iov_len = ciov.iov_len;
8095 src = (struct iovec __user *) arg;
8096 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8102 * Not super efficient, but this is just a registration time. And we do cache
8103 * the last compound head, so generally we'll only do a full search if we don't
8106 * We check if the given compound head page has already been accounted, to
8107 * avoid double accounting it. This allows us to account the full size of the
8108 * page, not just the constituent pages of a huge page.
8110 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8111 int nr_pages, struct page *hpage)
8115 /* check current page array */
8116 for (i = 0; i < nr_pages; i++) {
8117 if (!PageCompound(pages[i]))
8119 if (compound_head(pages[i]) == hpage)
8123 /* check previously registered pages */
8124 for (i = 0; i < ctx->nr_user_bufs; i++) {
8125 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8127 for (j = 0; j < imu->nr_bvecs; j++) {
8128 if (!PageCompound(imu->bvec[j].bv_page))
8130 if (compound_head(imu->bvec[j].bv_page) == hpage)
8138 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8139 int nr_pages, struct io_mapped_ubuf *imu,
8140 struct page **last_hpage)
8144 for (i = 0; i < nr_pages; i++) {
8145 if (!PageCompound(pages[i])) {
8150 hpage = compound_head(pages[i]);
8151 if (hpage == *last_hpage)
8153 *last_hpage = hpage;
8154 if (headpage_already_acct(ctx, pages, i, hpage))
8156 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8160 if (!imu->acct_pages)
8163 ret = io_account_mem(ctx, imu->acct_pages);
8165 imu->acct_pages = 0;
8169 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8170 struct io_mapped_ubuf *imu,
8171 struct page **last_hpage)
8173 struct vm_area_struct **vmas = NULL;
8174 struct page **pages = NULL;
8175 unsigned long off, start, end, ubuf;
8177 int ret, pret, nr_pages, i;
8179 ubuf = (unsigned long) iov->iov_base;
8180 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8181 start = ubuf >> PAGE_SHIFT;
8182 nr_pages = end - start;
8186 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8190 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8195 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8201 mmap_read_lock(current->mm);
8202 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8204 if (pret == nr_pages) {
8205 /* don't support file backed memory */
8206 for (i = 0; i < nr_pages; i++) {
8207 struct vm_area_struct *vma = vmas[i];
8210 !is_file_hugepages(vma->vm_file)) {
8216 ret = pret < 0 ? pret : -EFAULT;
8218 mmap_read_unlock(current->mm);
8221 * if we did partial map, or found file backed vmas,
8222 * release any pages we did get
8225 unpin_user_pages(pages, pret);
8230 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8232 unpin_user_pages(pages, pret);
8237 off = ubuf & ~PAGE_MASK;
8238 size = iov->iov_len;
8239 for (i = 0; i < nr_pages; i++) {
8242 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8243 imu->bvec[i].bv_page = pages[i];
8244 imu->bvec[i].bv_len = vec_len;
8245 imu->bvec[i].bv_offset = off;
8249 /* store original address for later verification */
8251 imu->len = iov->iov_len;
8252 imu->nr_bvecs = nr_pages;
8260 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8264 if (!nr_args || nr_args > UIO_MAXIOV)
8267 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8269 if (!ctx->user_bufs)
8275 static int io_buffer_validate(struct iovec *iov)
8278 * Don't impose further limits on the size and buffer
8279 * constraints here, we'll -EINVAL later when IO is
8280 * submitted if they are wrong.
8282 if (!iov->iov_base || !iov->iov_len)
8285 /* arbitrary limit, but we need something */
8286 if (iov->iov_len > SZ_1G)
8292 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8293 unsigned int nr_args)
8297 struct page *last_hpage = NULL;
8299 ret = io_buffers_map_alloc(ctx, nr_args);
8303 for (i = 0; i < nr_args; i++) {
8304 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8306 ret = io_copy_iov(ctx, &iov, arg, i);
8310 ret = io_buffer_validate(&iov);
8314 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8318 ctx->nr_user_bufs++;
8322 io_sqe_buffers_unregister(ctx);
8327 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8329 __s32 __user *fds = arg;
8335 if (copy_from_user(&fd, fds, sizeof(*fds)))
8338 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8339 if (IS_ERR(ctx->cq_ev_fd)) {
8340 int ret = PTR_ERR(ctx->cq_ev_fd);
8341 ctx->cq_ev_fd = NULL;
8348 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8350 if (ctx->cq_ev_fd) {
8351 eventfd_ctx_put(ctx->cq_ev_fd);
8352 ctx->cq_ev_fd = NULL;
8359 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8361 struct io_buffer *buf;
8362 unsigned long index;
8364 xa_for_each(&ctx->io_buffers, index, buf)
8365 __io_remove_buffers(ctx, buf, index, -1U);
8368 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8370 struct io_kiocb *req, *nxt;
8372 list_for_each_entry_safe(req, nxt, list, compl.list) {
8373 if (tsk && req->task != tsk)
8375 list_del(&req->compl.list);
8376 kmem_cache_free(req_cachep, req);
8380 static void io_req_caches_free(struct io_ring_ctx *ctx)
8382 struct io_submit_state *submit_state = &ctx->submit_state;
8383 struct io_comp_state *cs = &ctx->submit_state.comp;
8385 mutex_lock(&ctx->uring_lock);
8387 if (submit_state->free_reqs) {
8388 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8389 submit_state->reqs);
8390 submit_state->free_reqs = 0;
8393 spin_lock_irq(&ctx->completion_lock);
8394 list_splice_init(&cs->locked_free_list, &cs->free_list);
8395 cs->locked_free_nr = 0;
8396 spin_unlock_irq(&ctx->completion_lock);
8398 io_req_cache_free(&cs->free_list, NULL);
8400 mutex_unlock(&ctx->uring_lock);
8403 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8406 * Some may use context even when all refs and requests have been put,
8407 * and they are free to do so while still holding uring_lock or
8408 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8410 mutex_lock(&ctx->uring_lock);
8411 mutex_unlock(&ctx->uring_lock);
8412 spin_lock_irq(&ctx->completion_lock);
8413 spin_unlock_irq(&ctx->completion_lock);
8415 io_sq_thread_finish(ctx);
8416 io_sqe_buffers_unregister(ctx);
8418 if (ctx->mm_account) {
8419 mmdrop(ctx->mm_account);
8420 ctx->mm_account = NULL;
8423 mutex_lock(&ctx->uring_lock);
8424 io_sqe_files_unregister(ctx);
8425 mutex_unlock(&ctx->uring_lock);
8426 io_eventfd_unregister(ctx);
8427 io_destroy_buffers(ctx);
8429 #if defined(CONFIG_UNIX)
8430 if (ctx->ring_sock) {
8431 ctx->ring_sock->file = NULL; /* so that iput() is called */
8432 sock_release(ctx->ring_sock);
8436 io_mem_free(ctx->rings);
8437 io_mem_free(ctx->sq_sqes);
8439 percpu_ref_exit(&ctx->refs);
8440 free_uid(ctx->user);
8441 io_req_caches_free(ctx);
8443 io_wq_put_hash(ctx->hash_map);
8444 kfree(ctx->cancel_hash);
8448 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8450 struct io_ring_ctx *ctx = file->private_data;
8453 poll_wait(file, &ctx->cq_wait, wait);
8455 * synchronizes with barrier from wq_has_sleeper call in
8459 if (!io_sqring_full(ctx))
8460 mask |= EPOLLOUT | EPOLLWRNORM;
8463 * Don't flush cqring overflow list here, just do a simple check.
8464 * Otherwise there could possible be ABBA deadlock:
8467 * lock(&ctx->uring_lock);
8469 * lock(&ctx->uring_lock);
8472 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8473 * pushs them to do the flush.
8475 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8476 mask |= EPOLLIN | EPOLLRDNORM;
8481 static int io_uring_fasync(int fd, struct file *file, int on)
8483 struct io_ring_ctx *ctx = file->private_data;
8485 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8488 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8490 const struct cred *creds;
8492 creds = xa_erase(&ctx->personalities, id);
8501 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8503 return io_run_task_work_head(&ctx->exit_task_work);
8506 struct io_tctx_exit {
8507 struct callback_head task_work;
8508 struct completion completion;
8509 struct io_ring_ctx *ctx;
8512 static void io_tctx_exit_cb(struct callback_head *cb)
8514 struct io_uring_task *tctx = current->io_uring;
8515 struct io_tctx_exit *work;
8517 work = container_of(cb, struct io_tctx_exit, task_work);
8519 * When @in_idle, we're in cancellation and it's racy to remove the
8520 * node. It'll be removed by the end of cancellation, just ignore it.
8522 if (!atomic_read(&tctx->in_idle))
8523 io_uring_del_task_file((unsigned long)work->ctx);
8524 complete(&work->completion);
8527 static void io_ring_exit_work(struct work_struct *work)
8529 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8530 unsigned long timeout = jiffies + HZ * 60 * 5;
8531 struct io_tctx_exit exit;
8532 struct io_tctx_node *node;
8536 * If we're doing polled IO and end up having requests being
8537 * submitted async (out-of-line), then completions can come in while
8538 * we're waiting for refs to drop. We need to reap these manually,
8539 * as nobody else will be looking for them.
8542 io_uring_try_cancel_requests(ctx, NULL, NULL);
8544 WARN_ON_ONCE(time_after(jiffies, timeout));
8545 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8547 mutex_lock(&ctx->uring_lock);
8548 while (!list_empty(&ctx->tctx_list)) {
8549 WARN_ON_ONCE(time_after(jiffies, timeout));
8551 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8554 init_completion(&exit.completion);
8555 init_task_work(&exit.task_work, io_tctx_exit_cb);
8556 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8557 if (WARN_ON_ONCE(ret))
8559 wake_up_process(node->task);
8561 mutex_unlock(&ctx->uring_lock);
8562 wait_for_completion(&exit.completion);
8564 mutex_lock(&ctx->uring_lock);
8566 mutex_unlock(&ctx->uring_lock);
8568 io_ring_ctx_free(ctx);
8571 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8573 unsigned long index;
8574 struct creds *creds;
8576 mutex_lock(&ctx->uring_lock);
8577 percpu_ref_kill(&ctx->refs);
8578 /* if force is set, the ring is going away. always drop after that */
8579 ctx->cq_overflow_flushed = 1;
8581 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8582 xa_for_each(&ctx->personalities, index, creds)
8583 io_unregister_personality(ctx, index);
8584 mutex_unlock(&ctx->uring_lock);
8586 /* prevent SQPOLL from submitting new requests */
8588 io_sq_thread_park(ctx->sq_data);
8589 list_del_init(&ctx->sqd_list);
8590 io_sqd_update_thread_idle(ctx->sq_data);
8591 io_sq_thread_unpark(ctx->sq_data);
8594 io_kill_timeouts(ctx, NULL, NULL);
8595 io_poll_remove_all(ctx, NULL, NULL);
8597 /* if we failed setting up the ctx, we might not have any rings */
8598 io_iopoll_try_reap_events(ctx);
8600 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8602 * Use system_unbound_wq to avoid spawning tons of event kworkers
8603 * if we're exiting a ton of rings at the same time. It just adds
8604 * noise and overhead, there's no discernable change in runtime
8605 * over using system_wq.
8607 queue_work(system_unbound_wq, &ctx->exit_work);
8610 static int io_uring_release(struct inode *inode, struct file *file)
8612 struct io_ring_ctx *ctx = file->private_data;
8614 file->private_data = NULL;
8615 io_ring_ctx_wait_and_kill(ctx);
8619 struct io_task_cancel {
8620 struct task_struct *task;
8621 struct files_struct *files;
8624 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8626 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8627 struct io_task_cancel *cancel = data;
8630 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8631 unsigned long flags;
8632 struct io_ring_ctx *ctx = req->ctx;
8634 /* protect against races with linked timeouts */
8635 spin_lock_irqsave(&ctx->completion_lock, flags);
8636 ret = io_match_task(req, cancel->task, cancel->files);
8637 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8639 ret = io_match_task(req, cancel->task, cancel->files);
8644 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8645 struct task_struct *task,
8646 struct files_struct *files)
8648 struct io_defer_entry *de;
8651 spin_lock_irq(&ctx->completion_lock);
8652 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8653 if (io_match_task(de->req, task, files)) {
8654 list_cut_position(&list, &ctx->defer_list, &de->list);
8658 spin_unlock_irq(&ctx->completion_lock);
8659 if (list_empty(&list))
8662 while (!list_empty(&list)) {
8663 de = list_first_entry(&list, struct io_defer_entry, list);
8664 list_del_init(&de->list);
8665 req_set_fail_links(de->req);
8666 io_put_req(de->req);
8667 io_req_complete(de->req, -ECANCELED);
8673 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8675 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8677 return req->ctx == data;
8680 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8682 struct io_tctx_node *node;
8683 enum io_wq_cancel cret;
8686 mutex_lock(&ctx->uring_lock);
8687 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8688 struct io_uring_task *tctx = node->task->io_uring;
8691 * io_wq will stay alive while we hold uring_lock, because it's
8692 * killed after ctx nodes, which requires to take the lock.
8694 if (!tctx || !tctx->io_wq)
8696 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8697 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8699 mutex_unlock(&ctx->uring_lock);
8704 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8705 struct task_struct *task,
8706 struct files_struct *files)
8708 struct io_task_cancel cancel = { .task = task, .files = files, };
8709 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8712 enum io_wq_cancel cret;
8716 ret |= io_uring_try_cancel_iowq(ctx);
8717 } else if (tctx && tctx->io_wq) {
8719 * Cancels requests of all rings, not only @ctx, but
8720 * it's fine as the task is in exit/exec.
8722 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8724 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8727 /* SQPOLL thread does its own polling */
8728 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8729 (ctx->sq_data && ctx->sq_data->thread == current)) {
8730 while (!list_empty_careful(&ctx->iopoll_list)) {
8731 io_iopoll_try_reap_events(ctx);
8736 ret |= io_cancel_defer_files(ctx, task, files);
8737 ret |= io_poll_remove_all(ctx, task, files);
8738 ret |= io_kill_timeouts(ctx, task, files);
8739 ret |= io_run_task_work();
8740 ret |= io_run_ctx_fallback(ctx);
8741 io_cqring_overflow_flush(ctx, true, task, files);
8748 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8749 struct task_struct *task,
8750 struct files_struct *files)
8752 struct io_kiocb *req;
8755 spin_lock_irq(&ctx->inflight_lock);
8756 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8757 cnt += io_match_task(req, task, files);
8758 spin_unlock_irq(&ctx->inflight_lock);
8762 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8763 struct task_struct *task,
8764 struct files_struct *files)
8766 while (!list_empty_careful(&ctx->inflight_list)) {
8770 inflight = io_uring_count_inflight(ctx, task, files);
8774 io_uring_try_cancel_requests(ctx, task, files);
8776 prepare_to_wait(&task->io_uring->wait, &wait,
8777 TASK_UNINTERRUPTIBLE);
8778 if (inflight == io_uring_count_inflight(ctx, task, files))
8780 finish_wait(&task->io_uring->wait, &wait);
8785 * Note that this task has used io_uring. We use it for cancelation purposes.
8787 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8789 struct io_uring_task *tctx = current->io_uring;
8790 struct io_tctx_node *node;
8793 if (unlikely(!tctx)) {
8794 ret = io_uring_alloc_task_context(current, ctx);
8797 tctx = current->io_uring;
8799 if (tctx->last != ctx) {
8800 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8803 node = kmalloc(sizeof(*node), GFP_KERNEL);
8807 node->task = current;
8809 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8816 mutex_lock(&ctx->uring_lock);
8817 list_add(&node->ctx_node, &ctx->tctx_list);
8818 mutex_unlock(&ctx->uring_lock);
8826 * Remove this io_uring_file -> task mapping.
8828 static void io_uring_del_task_file(unsigned long index)
8830 struct io_uring_task *tctx = current->io_uring;
8831 struct io_tctx_node *node;
8835 node = xa_erase(&tctx->xa, index);
8839 WARN_ON_ONCE(current != node->task);
8840 WARN_ON_ONCE(list_empty(&node->ctx_node));
8842 mutex_lock(&node->ctx->uring_lock);
8843 list_del(&node->ctx_node);
8844 mutex_unlock(&node->ctx->uring_lock);
8846 if (tctx->last == node->ctx)
8851 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8853 struct io_tctx_node *node;
8854 unsigned long index;
8856 xa_for_each(&tctx->xa, index, node)
8857 io_uring_del_task_file(index);
8859 io_wq_put_and_exit(tctx->io_wq);
8864 static s64 tctx_inflight(struct io_uring_task *tctx)
8866 return percpu_counter_sum(&tctx->inflight);
8869 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8871 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8872 struct io_ring_ctx *ctx = work->ctx;
8873 struct io_sq_data *sqd = ctx->sq_data;
8876 io_uring_cancel_sqpoll(ctx);
8877 complete(&work->completion);
8880 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8882 struct io_sq_data *sqd = ctx->sq_data;
8883 struct io_tctx_exit work = { .ctx = ctx, };
8884 struct task_struct *task;
8886 io_sq_thread_park(sqd);
8887 list_del_init(&ctx->sqd_list);
8888 io_sqd_update_thread_idle(sqd);
8891 init_completion(&work.completion);
8892 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8893 WARN_ON_ONCE(task_work_add(task, &work.task_work, TWA_SIGNAL));
8894 wake_up_process(task);
8896 io_sq_thread_unpark(sqd);
8899 wait_for_completion(&work.completion);
8902 void __io_uring_files_cancel(struct files_struct *files)
8904 struct io_uring_task *tctx = current->io_uring;
8905 struct io_tctx_node *node;
8906 unsigned long index;
8908 /* make sure overflow events are dropped */
8909 atomic_inc(&tctx->in_idle);
8910 xa_for_each(&tctx->xa, index, node) {
8911 struct io_ring_ctx *ctx = node->ctx;
8914 io_sqpoll_cancel_sync(ctx);
8917 io_uring_cancel_files(ctx, current, files);
8919 io_uring_try_cancel_requests(ctx, current, NULL);
8921 atomic_dec(&tctx->in_idle);
8924 io_uring_clean_tctx(tctx);
8927 /* should only be called by SQPOLL task */
8928 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8930 struct io_sq_data *sqd = ctx->sq_data;
8931 struct io_uring_task *tctx = current->io_uring;
8935 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8937 atomic_inc(&tctx->in_idle);
8939 /* read completions before cancelations */
8940 inflight = tctx_inflight(tctx);
8943 io_uring_try_cancel_requests(ctx, current, NULL);
8945 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8947 * If we've seen completions, retry without waiting. This
8948 * avoids a race where a completion comes in before we did
8949 * prepare_to_wait().
8951 if (inflight == tctx_inflight(tctx))
8953 finish_wait(&tctx->wait, &wait);
8955 atomic_dec(&tctx->in_idle);
8959 * Find any io_uring fd that this task has registered or done IO on, and cancel
8962 void __io_uring_task_cancel(void)
8964 struct io_uring_task *tctx = current->io_uring;
8968 /* make sure overflow events are dropped */
8969 atomic_inc(&tctx->in_idle);
8971 /* read completions before cancelations */
8972 inflight = tctx_inflight(tctx);
8975 __io_uring_files_cancel(NULL);
8977 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8980 * If we've seen completions, retry without waiting. This
8981 * avoids a race where a completion comes in before we did
8982 * prepare_to_wait().
8984 if (inflight == tctx_inflight(tctx))
8986 finish_wait(&tctx->wait, &wait);
8989 atomic_dec(&tctx->in_idle);
8991 io_uring_clean_tctx(tctx);
8992 /* all current's requests should be gone, we can kill tctx */
8993 __io_uring_free(current);
8996 static void *io_uring_validate_mmap_request(struct file *file,
8997 loff_t pgoff, size_t sz)
8999 struct io_ring_ctx *ctx = file->private_data;
9000 loff_t offset = pgoff << PAGE_SHIFT;
9005 case IORING_OFF_SQ_RING:
9006 case IORING_OFF_CQ_RING:
9009 case IORING_OFF_SQES:
9013 return ERR_PTR(-EINVAL);
9016 page = virt_to_head_page(ptr);
9017 if (sz > page_size(page))
9018 return ERR_PTR(-EINVAL);
9025 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9027 size_t sz = vma->vm_end - vma->vm_start;
9031 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9033 return PTR_ERR(ptr);
9035 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9036 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9039 #else /* !CONFIG_MMU */
9041 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9043 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9046 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9048 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9051 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9052 unsigned long addr, unsigned long len,
9053 unsigned long pgoff, unsigned long flags)
9057 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9059 return PTR_ERR(ptr);
9061 return (unsigned long) ptr;
9064 #endif /* !CONFIG_MMU */
9066 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9071 if (!io_sqring_full(ctx))
9073 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9075 if (!io_sqring_full(ctx))
9078 } while (!signal_pending(current));
9080 finish_wait(&ctx->sqo_sq_wait, &wait);
9084 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9085 struct __kernel_timespec __user **ts,
9086 const sigset_t __user **sig)
9088 struct io_uring_getevents_arg arg;
9091 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9092 * is just a pointer to the sigset_t.
9094 if (!(flags & IORING_ENTER_EXT_ARG)) {
9095 *sig = (const sigset_t __user *) argp;
9101 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9102 * timespec and sigset_t pointers if good.
9104 if (*argsz != sizeof(arg))
9106 if (copy_from_user(&arg, argp, sizeof(arg)))
9108 *sig = u64_to_user_ptr(arg.sigmask);
9109 *argsz = arg.sigmask_sz;
9110 *ts = u64_to_user_ptr(arg.ts);
9114 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9115 u32, min_complete, u32, flags, const void __user *, argp,
9118 struct io_ring_ctx *ctx;
9125 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9126 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9134 if (f.file->f_op != &io_uring_fops)
9138 ctx = f.file->private_data;
9139 if (!percpu_ref_tryget(&ctx->refs))
9143 if (ctx->flags & IORING_SETUP_R_DISABLED)
9147 * For SQ polling, the thread will do all submissions and completions.
9148 * Just return the requested submit count, and wake the thread if
9152 if (ctx->flags & IORING_SETUP_SQPOLL) {
9153 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9156 if (unlikely(ctx->sq_data->thread == NULL)) {
9159 if (flags & IORING_ENTER_SQ_WAKEUP)
9160 wake_up(&ctx->sq_data->wait);
9161 if (flags & IORING_ENTER_SQ_WAIT) {
9162 ret = io_sqpoll_wait_sq(ctx);
9166 submitted = to_submit;
9167 } else if (to_submit) {
9168 ret = io_uring_add_task_file(ctx);
9171 mutex_lock(&ctx->uring_lock);
9172 submitted = io_submit_sqes(ctx, to_submit);
9173 mutex_unlock(&ctx->uring_lock);
9175 if (submitted != to_submit)
9178 if (flags & IORING_ENTER_GETEVENTS) {
9179 const sigset_t __user *sig;
9180 struct __kernel_timespec __user *ts;
9182 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9186 min_complete = min(min_complete, ctx->cq_entries);
9189 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9190 * space applications don't need to do io completion events
9191 * polling again, they can rely on io_sq_thread to do polling
9192 * work, which can reduce cpu usage and uring_lock contention.
9194 if (ctx->flags & IORING_SETUP_IOPOLL &&
9195 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9196 ret = io_iopoll_check(ctx, min_complete);
9198 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9203 percpu_ref_put(&ctx->refs);
9206 return submitted ? submitted : ret;
9209 #ifdef CONFIG_PROC_FS
9210 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9211 const struct cred *cred)
9213 struct user_namespace *uns = seq_user_ns(m);
9214 struct group_info *gi;
9219 seq_printf(m, "%5d\n", id);
9220 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9221 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9222 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9223 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9224 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9225 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9226 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9227 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9228 seq_puts(m, "\n\tGroups:\t");
9229 gi = cred->group_info;
9230 for (g = 0; g < gi->ngroups; g++) {
9231 seq_put_decimal_ull(m, g ? " " : "",
9232 from_kgid_munged(uns, gi->gid[g]));
9234 seq_puts(m, "\n\tCapEff:\t");
9235 cap = cred->cap_effective;
9236 CAP_FOR_EACH_U32(__capi)
9237 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9242 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9244 struct io_sq_data *sq = NULL;
9249 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9250 * since fdinfo case grabs it in the opposite direction of normal use
9251 * cases. If we fail to get the lock, we just don't iterate any
9252 * structures that could be going away outside the io_uring mutex.
9254 has_lock = mutex_trylock(&ctx->uring_lock);
9256 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9262 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9263 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9264 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9265 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9266 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9269 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9271 seq_printf(m, "%5u: <none>\n", i);
9273 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9274 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9275 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9277 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9278 (unsigned int) buf->len);
9280 if (has_lock && !xa_empty(&ctx->personalities)) {
9281 unsigned long index;
9282 const struct cred *cred;
9284 seq_printf(m, "Personalities:\n");
9285 xa_for_each(&ctx->personalities, index, cred)
9286 io_uring_show_cred(m, index, cred);
9288 seq_printf(m, "PollList:\n");
9289 spin_lock_irq(&ctx->completion_lock);
9290 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9291 struct hlist_head *list = &ctx->cancel_hash[i];
9292 struct io_kiocb *req;
9294 hlist_for_each_entry(req, list, hash_node)
9295 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9296 req->task->task_works != NULL);
9298 spin_unlock_irq(&ctx->completion_lock);
9300 mutex_unlock(&ctx->uring_lock);
9303 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9305 struct io_ring_ctx *ctx = f->private_data;
9307 if (percpu_ref_tryget(&ctx->refs)) {
9308 __io_uring_show_fdinfo(ctx, m);
9309 percpu_ref_put(&ctx->refs);
9314 static const struct file_operations io_uring_fops = {
9315 .release = io_uring_release,
9316 .mmap = io_uring_mmap,
9318 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9319 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9321 .poll = io_uring_poll,
9322 .fasync = io_uring_fasync,
9323 #ifdef CONFIG_PROC_FS
9324 .show_fdinfo = io_uring_show_fdinfo,
9328 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9329 struct io_uring_params *p)
9331 struct io_rings *rings;
9332 size_t size, sq_array_offset;
9334 /* make sure these are sane, as we already accounted them */
9335 ctx->sq_entries = p->sq_entries;
9336 ctx->cq_entries = p->cq_entries;
9338 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9339 if (size == SIZE_MAX)
9342 rings = io_mem_alloc(size);
9347 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9348 rings->sq_ring_mask = p->sq_entries - 1;
9349 rings->cq_ring_mask = p->cq_entries - 1;
9350 rings->sq_ring_entries = p->sq_entries;
9351 rings->cq_ring_entries = p->cq_entries;
9352 ctx->sq_mask = rings->sq_ring_mask;
9353 ctx->cq_mask = rings->cq_ring_mask;
9355 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9356 if (size == SIZE_MAX) {
9357 io_mem_free(ctx->rings);
9362 ctx->sq_sqes = io_mem_alloc(size);
9363 if (!ctx->sq_sqes) {
9364 io_mem_free(ctx->rings);
9372 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9376 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9380 ret = io_uring_add_task_file(ctx);
9385 fd_install(fd, file);
9390 * Allocate an anonymous fd, this is what constitutes the application
9391 * visible backing of an io_uring instance. The application mmaps this
9392 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9393 * we have to tie this fd to a socket for file garbage collection purposes.
9395 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9398 #if defined(CONFIG_UNIX)
9401 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9404 return ERR_PTR(ret);
9407 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9408 O_RDWR | O_CLOEXEC);
9409 #if defined(CONFIG_UNIX)
9411 sock_release(ctx->ring_sock);
9412 ctx->ring_sock = NULL;
9414 ctx->ring_sock->file = file;
9420 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9421 struct io_uring_params __user *params)
9423 struct io_ring_ctx *ctx;
9429 if (entries > IORING_MAX_ENTRIES) {
9430 if (!(p->flags & IORING_SETUP_CLAMP))
9432 entries = IORING_MAX_ENTRIES;
9436 * Use twice as many entries for the CQ ring. It's possible for the
9437 * application to drive a higher depth than the size of the SQ ring,
9438 * since the sqes are only used at submission time. This allows for
9439 * some flexibility in overcommitting a bit. If the application has
9440 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9441 * of CQ ring entries manually.
9443 p->sq_entries = roundup_pow_of_two(entries);
9444 if (p->flags & IORING_SETUP_CQSIZE) {
9446 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9447 * to a power-of-two, if it isn't already. We do NOT impose
9448 * any cq vs sq ring sizing.
9452 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9453 if (!(p->flags & IORING_SETUP_CLAMP))
9455 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9457 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9458 if (p->cq_entries < p->sq_entries)
9461 p->cq_entries = 2 * p->sq_entries;
9464 ctx = io_ring_ctx_alloc(p);
9467 ctx->compat = in_compat_syscall();
9468 if (!capable(CAP_IPC_LOCK))
9469 ctx->user = get_uid(current_user());
9472 * This is just grabbed for accounting purposes. When a process exits,
9473 * the mm is exited and dropped before the files, hence we need to hang
9474 * on to this mm purely for the purposes of being able to unaccount
9475 * memory (locked/pinned vm). It's not used for anything else.
9477 mmgrab(current->mm);
9478 ctx->mm_account = current->mm;
9480 ret = io_allocate_scq_urings(ctx, p);
9484 ret = io_sq_offload_create(ctx, p);
9488 memset(&p->sq_off, 0, sizeof(p->sq_off));
9489 p->sq_off.head = offsetof(struct io_rings, sq.head);
9490 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9491 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9492 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9493 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9494 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9495 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9497 memset(&p->cq_off, 0, sizeof(p->cq_off));
9498 p->cq_off.head = offsetof(struct io_rings, cq.head);
9499 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9500 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9501 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9502 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9503 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9504 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9506 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9507 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9508 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9509 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9510 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9512 if (copy_to_user(params, p, sizeof(*p))) {
9517 file = io_uring_get_file(ctx);
9519 ret = PTR_ERR(file);
9524 * Install ring fd as the very last thing, so we don't risk someone
9525 * having closed it before we finish setup
9527 ret = io_uring_install_fd(ctx, file);
9529 /* fput will clean it up */
9534 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9537 io_ring_ctx_wait_and_kill(ctx);
9542 * Sets up an aio uring context, and returns the fd. Applications asks for a
9543 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9544 * params structure passed in.
9546 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9548 struct io_uring_params p;
9551 if (copy_from_user(&p, params, sizeof(p)))
9553 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9558 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9559 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9560 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9561 IORING_SETUP_R_DISABLED))
9564 return io_uring_create(entries, &p, params);
9567 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9568 struct io_uring_params __user *, params)
9570 return io_uring_setup(entries, params);
9573 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9575 struct io_uring_probe *p;
9579 size = struct_size(p, ops, nr_args);
9580 if (size == SIZE_MAX)
9582 p = kzalloc(size, GFP_KERNEL);
9587 if (copy_from_user(p, arg, size))
9590 if (memchr_inv(p, 0, size))
9593 p->last_op = IORING_OP_LAST - 1;
9594 if (nr_args > IORING_OP_LAST)
9595 nr_args = IORING_OP_LAST;
9597 for (i = 0; i < nr_args; i++) {
9599 if (!io_op_defs[i].not_supported)
9600 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9605 if (copy_to_user(arg, p, size))
9612 static int io_register_personality(struct io_ring_ctx *ctx)
9614 const struct cred *creds;
9618 creds = get_current_cred();
9620 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9621 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9628 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9629 unsigned int nr_args)
9631 struct io_uring_restriction *res;
9635 /* Restrictions allowed only if rings started disabled */
9636 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9639 /* We allow only a single restrictions registration */
9640 if (ctx->restrictions.registered)
9643 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9646 size = array_size(nr_args, sizeof(*res));
9647 if (size == SIZE_MAX)
9650 res = memdup_user(arg, size);
9652 return PTR_ERR(res);
9656 for (i = 0; i < nr_args; i++) {
9657 switch (res[i].opcode) {
9658 case IORING_RESTRICTION_REGISTER_OP:
9659 if (res[i].register_op >= IORING_REGISTER_LAST) {
9664 __set_bit(res[i].register_op,
9665 ctx->restrictions.register_op);
9667 case IORING_RESTRICTION_SQE_OP:
9668 if (res[i].sqe_op >= IORING_OP_LAST) {
9673 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9675 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9676 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9678 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9679 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9688 /* Reset all restrictions if an error happened */
9690 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9692 ctx->restrictions.registered = true;
9698 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9700 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9703 if (ctx->restrictions.registered)
9704 ctx->restricted = 1;
9706 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9707 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9708 wake_up(&ctx->sq_data->wait);
9712 static bool io_register_op_must_quiesce(int op)
9715 case IORING_UNREGISTER_FILES:
9716 case IORING_REGISTER_FILES_UPDATE:
9717 case IORING_REGISTER_PROBE:
9718 case IORING_REGISTER_PERSONALITY:
9719 case IORING_UNREGISTER_PERSONALITY:
9726 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9727 void __user *arg, unsigned nr_args)
9728 __releases(ctx->uring_lock)
9729 __acquires(ctx->uring_lock)
9734 * We're inside the ring mutex, if the ref is already dying, then
9735 * someone else killed the ctx or is already going through
9736 * io_uring_register().
9738 if (percpu_ref_is_dying(&ctx->refs))
9741 if (io_register_op_must_quiesce(opcode)) {
9742 percpu_ref_kill(&ctx->refs);
9745 * Drop uring mutex before waiting for references to exit. If
9746 * another thread is currently inside io_uring_enter() it might
9747 * need to grab the uring_lock to make progress. If we hold it
9748 * here across the drain wait, then we can deadlock. It's safe
9749 * to drop the mutex here, since no new references will come in
9750 * after we've killed the percpu ref.
9752 mutex_unlock(&ctx->uring_lock);
9754 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9757 ret = io_run_task_work_sig();
9762 mutex_lock(&ctx->uring_lock);
9765 percpu_ref_resurrect(&ctx->refs);
9770 if (ctx->restricted) {
9771 if (opcode >= IORING_REGISTER_LAST) {
9776 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9783 case IORING_REGISTER_BUFFERS:
9784 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9786 case IORING_UNREGISTER_BUFFERS:
9790 ret = io_sqe_buffers_unregister(ctx);
9792 case IORING_REGISTER_FILES:
9793 ret = io_sqe_files_register(ctx, arg, nr_args);
9795 case IORING_UNREGISTER_FILES:
9799 ret = io_sqe_files_unregister(ctx);
9801 case IORING_REGISTER_FILES_UPDATE:
9802 ret = io_sqe_files_update(ctx, arg, nr_args);
9804 case IORING_REGISTER_EVENTFD:
9805 case IORING_REGISTER_EVENTFD_ASYNC:
9809 ret = io_eventfd_register(ctx, arg);
9812 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9813 ctx->eventfd_async = 1;
9815 ctx->eventfd_async = 0;
9817 case IORING_UNREGISTER_EVENTFD:
9821 ret = io_eventfd_unregister(ctx);
9823 case IORING_REGISTER_PROBE:
9825 if (!arg || nr_args > 256)
9827 ret = io_probe(ctx, arg, nr_args);
9829 case IORING_REGISTER_PERSONALITY:
9833 ret = io_register_personality(ctx);
9835 case IORING_UNREGISTER_PERSONALITY:
9839 ret = io_unregister_personality(ctx, nr_args);
9841 case IORING_REGISTER_ENABLE_RINGS:
9845 ret = io_register_enable_rings(ctx);
9847 case IORING_REGISTER_RESTRICTIONS:
9848 ret = io_register_restrictions(ctx, arg, nr_args);
9856 if (io_register_op_must_quiesce(opcode)) {
9857 /* bring the ctx back to life */
9858 percpu_ref_reinit(&ctx->refs);
9860 reinit_completion(&ctx->ref_comp);
9865 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9866 void __user *, arg, unsigned int, nr_args)
9868 struct io_ring_ctx *ctx;
9877 if (f.file->f_op != &io_uring_fops)
9880 ctx = f.file->private_data;
9884 mutex_lock(&ctx->uring_lock);
9885 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9886 mutex_unlock(&ctx->uring_lock);
9887 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9888 ctx->cq_ev_fd != NULL, ret);
9894 static int __init io_uring_init(void)
9896 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9897 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9898 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9901 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9902 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9903 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9904 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9905 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9906 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9907 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9908 BUILD_BUG_SQE_ELEM(8, __u64, off);
9909 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9910 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9911 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9912 BUILD_BUG_SQE_ELEM(24, __u32, len);
9913 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9914 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9915 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9916 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9917 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9918 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9919 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9920 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9921 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9922 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9923 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9924 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9925 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9926 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9927 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9928 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9929 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9930 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9931 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9933 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9934 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9935 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9939 __initcall(io_uring_init);