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,
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
265 struct list_head ctx_new_list;
266 struct mutex ctx_lock;
268 struct task_struct *thread;
269 struct wait_queue_head wait;
271 unsigned sq_thread_idle;
276 struct completion startup;
277 struct completion parked;
278 struct completion exited;
281 #define IO_IOPOLL_BATCH 8
282 #define IO_COMPL_BATCH 32
283 #define IO_REQ_CACHE_SIZE 32
284 #define IO_REQ_ALLOC_BATCH 8
286 struct io_comp_state {
287 struct io_kiocb *reqs[IO_COMPL_BATCH];
289 unsigned int locked_free_nr;
290 /* inline/task_work completion list, under ->uring_lock */
291 struct list_head free_list;
292 /* IRQ completion list, under ->completion_lock */
293 struct list_head locked_free_list;
296 struct io_submit_link {
297 struct io_kiocb *head;
298 struct io_kiocb *last;
301 struct io_submit_state {
302 struct blk_plug plug;
303 struct io_submit_link link;
306 * io_kiocb alloc cache
308 void *reqs[IO_REQ_CACHE_SIZE];
309 unsigned int free_reqs;
314 * Batch completion logic
316 struct io_comp_state comp;
319 * File reference cache
323 unsigned int file_refs;
324 unsigned int ios_left;
329 struct percpu_ref refs;
330 } ____cacheline_aligned_in_smp;
334 unsigned int compat: 1;
335 unsigned int cq_overflow_flushed: 1;
336 unsigned int drain_next: 1;
337 unsigned int eventfd_async: 1;
338 unsigned int restricted: 1;
339 unsigned int sqo_exec: 1;
342 * Ring buffer of indices into array of io_uring_sqe, which is
343 * mmapped by the application using the IORING_OFF_SQES offset.
345 * This indirection could e.g. be used to assign fixed
346 * io_uring_sqe entries to operations and only submit them to
347 * the queue when needed.
349 * The kernel modifies neither the indices array nor the entries
353 unsigned cached_sq_head;
356 unsigned sq_thread_idle;
357 unsigned cached_sq_dropped;
358 unsigned cached_cq_overflow;
359 unsigned long sq_check_overflow;
361 /* hashed buffered write serialization */
362 struct io_wq_hash *hash_map;
364 struct list_head defer_list;
365 struct list_head timeout_list;
366 struct list_head cq_overflow_list;
368 struct io_uring_sqe *sq_sqes;
369 } ____cacheline_aligned_in_smp;
372 struct mutex uring_lock;
373 wait_queue_head_t wait;
374 } ____cacheline_aligned_in_smp;
376 struct io_submit_state submit_state;
378 struct io_rings *rings;
380 /* Only used for accounting purposes */
381 struct mm_struct *mm_account;
383 struct io_sq_data *sq_data; /* if using sq thread polling */
385 struct wait_queue_head sqo_sq_wait;
386 struct list_head sqd_list;
389 * If used, fixed file set. Writers must ensure that ->refs is dead,
390 * readers must ensure that ->refs is alive as long as the file* is
391 * used. Only updated through io_uring_register(2).
393 struct fixed_rsrc_data *file_data;
394 unsigned nr_user_files;
396 /* if used, fixed mapped user buffers */
397 unsigned nr_user_bufs;
398 struct io_mapped_ubuf *user_bufs;
400 struct user_struct *user;
402 struct completion ref_comp;
403 struct completion sq_thread_comp;
405 #if defined(CONFIG_UNIX)
406 struct socket *ring_sock;
409 struct idr io_buffer_idr;
411 struct idr personality_idr;
414 unsigned cached_cq_tail;
417 atomic_t cq_timeouts;
418 unsigned cq_last_tm_flush;
419 unsigned long cq_check_overflow;
420 struct wait_queue_head cq_wait;
421 struct fasync_struct *cq_fasync;
422 struct eventfd_ctx *cq_ev_fd;
423 } ____cacheline_aligned_in_smp;
426 spinlock_t completion_lock;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list;
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
437 bool poll_multi_file;
439 spinlock_t inflight_lock;
440 struct list_head inflight_list;
441 } ____cacheline_aligned_in_smp;
443 struct delayed_work rsrc_put_work;
444 struct llist_head rsrc_put_llist;
445 struct list_head rsrc_ref_list;
446 spinlock_t rsrc_ref_lock;
448 struct io_restriction restrictions;
451 struct callback_head *exit_task_work;
453 struct wait_queue_head hash_wait;
455 /* Keep this last, we don't need it for the fast path */
456 struct work_struct exit_work;
457 struct list_head tctx_list;
461 * First field must be the file pointer in all the
462 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
464 struct io_poll_iocb {
466 struct wait_queue_head *head;
470 struct wait_queue_entry wait;
473 struct io_poll_remove {
483 struct io_timeout_data {
484 struct io_kiocb *req;
485 struct hrtimer timer;
486 struct timespec64 ts;
487 enum hrtimer_mode mode;
492 struct sockaddr __user *addr;
493 int __user *addr_len;
495 unsigned long nofile;
515 struct list_head list;
516 /* head of the link, used by linked timeouts only */
517 struct io_kiocb *head;
520 struct io_timeout_rem {
525 struct timespec64 ts;
530 /* NOTE: kiocb has the file as the first member, so don't do it here */
538 struct sockaddr __user *addr;
545 struct user_msghdr __user *umsg;
551 struct io_buffer *kbuf;
557 struct filename *filename;
559 unsigned long nofile;
562 struct io_rsrc_update {
588 struct epoll_event event;
592 struct file *file_out;
593 struct file *file_in;
600 struct io_provide_buf {
614 const char __user *filename;
615 struct statx __user *buffer;
627 struct filename *oldpath;
628 struct filename *newpath;
636 struct filename *filename;
639 struct io_completion {
641 struct list_head list;
645 struct io_async_connect {
646 struct sockaddr_storage address;
649 struct io_async_msghdr {
650 struct iovec fast_iov[UIO_FASTIOV];
651 /* points to an allocated iov, if NULL we use fast_iov instead */
652 struct iovec *free_iov;
653 struct sockaddr __user *uaddr;
655 struct sockaddr_storage addr;
659 struct iovec fast_iov[UIO_FASTIOV];
660 const struct iovec *free_iovec;
661 struct iov_iter iter;
663 struct wait_page_queue wpq;
667 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
668 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
669 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
670 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
671 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
672 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
678 REQ_F_LINK_TIMEOUT_BIT,
680 REQ_F_NEED_CLEANUP_BIT,
682 REQ_F_BUFFER_SELECTED_BIT,
683 REQ_F_NO_FILE_TABLE_BIT,
684 REQ_F_LTIMEOUT_ACTIVE_BIT,
685 REQ_F_COMPLETE_INLINE_BIT,
687 /* not a real bit, just to check we're not overflowing the space */
693 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
694 /* drain existing IO first */
695 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
697 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
698 /* doesn't sever on completion < 0 */
699 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
701 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
702 /* IOSQE_BUFFER_SELECT */
703 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
705 /* fail rest of links */
706 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
707 /* on inflight list, should be cancelled and waited on exit reliably */
708 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
709 /* read/write uses file position */
710 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
711 /* must not punt to workers */
712 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
713 /* has or had linked timeout */
714 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
716 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
718 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
719 /* already went through poll handler */
720 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
721 /* buffer already selected */
722 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
723 /* doesn't need file table for this request */
724 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
725 /* linked timeout is active, i.e. prepared by link's head */
726 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
727 /* completion is deferred through io_comp_state */
728 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
732 struct io_poll_iocb poll;
733 struct io_poll_iocb *double_poll;
736 struct io_task_work {
737 struct io_wq_work_node node;
738 task_work_func_t func;
742 * NOTE! Each of the iocb union members has the file pointer
743 * as the first entry in their struct definition. So you can
744 * access the file pointer through any of the sub-structs,
745 * or directly as just 'ki_filp' in this struct.
751 struct io_poll_iocb poll;
752 struct io_poll_remove poll_remove;
753 struct io_accept accept;
755 struct io_cancel cancel;
756 struct io_timeout timeout;
757 struct io_timeout_rem timeout_rem;
758 struct io_connect connect;
759 struct io_sr_msg sr_msg;
761 struct io_close close;
762 struct io_rsrc_update rsrc_update;
763 struct io_fadvise fadvise;
764 struct io_madvise madvise;
765 struct io_epoll epoll;
766 struct io_splice splice;
767 struct io_provide_buf pbuf;
768 struct io_statx statx;
769 struct io_shutdown shutdown;
770 struct io_rename rename;
771 struct io_unlink unlink;
772 /* use only after cleaning per-op data, see io_clean_op() */
773 struct io_completion compl;
776 /* opcode allocated if it needs to store data for async defer */
779 /* polled IO has completed */
785 struct io_ring_ctx *ctx;
788 struct task_struct *task;
791 struct io_kiocb *link;
792 struct percpu_ref *fixed_rsrc_refs;
795 * 1. used with ctx->iopoll_list with reads/writes
796 * 2. to track reqs with ->files (see io_op_def::file_table)
798 struct list_head inflight_entry;
800 struct io_task_work io_task_work;
801 struct callback_head task_work;
803 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
804 struct hlist_node hash_node;
805 struct async_poll *apoll;
806 struct io_wq_work work;
809 struct io_tctx_node {
810 struct list_head ctx_node;
811 struct task_struct *task;
813 struct io_ring_ctx *ctx;
816 struct io_defer_entry {
817 struct list_head list;
818 struct io_kiocb *req;
823 /* needs req->file assigned */
824 unsigned needs_file : 1;
825 /* hash wq insertion if file is a regular file */
826 unsigned hash_reg_file : 1;
827 /* unbound wq insertion if file is a non-regular file */
828 unsigned unbound_nonreg_file : 1;
829 /* opcode is not supported by this kernel */
830 unsigned not_supported : 1;
831 /* set if opcode supports polled "wait" */
833 unsigned pollout : 1;
834 /* op supports buffer selection */
835 unsigned buffer_select : 1;
836 /* must always have async data allocated */
837 unsigned needs_async_data : 1;
838 /* should block plug */
840 /* size of async data needed, if any */
841 unsigned short async_size;
844 static const struct io_op_def io_op_defs[] = {
845 [IORING_OP_NOP] = {},
846 [IORING_OP_READV] = {
848 .unbound_nonreg_file = 1,
851 .needs_async_data = 1,
853 .async_size = sizeof(struct io_async_rw),
855 [IORING_OP_WRITEV] = {
858 .unbound_nonreg_file = 1,
860 .needs_async_data = 1,
862 .async_size = sizeof(struct io_async_rw),
864 [IORING_OP_FSYNC] = {
867 [IORING_OP_READ_FIXED] = {
869 .unbound_nonreg_file = 1,
872 .async_size = sizeof(struct io_async_rw),
874 [IORING_OP_WRITE_FIXED] = {
877 .unbound_nonreg_file = 1,
880 .async_size = sizeof(struct io_async_rw),
882 [IORING_OP_POLL_ADD] = {
884 .unbound_nonreg_file = 1,
886 [IORING_OP_POLL_REMOVE] = {},
887 [IORING_OP_SYNC_FILE_RANGE] = {
890 [IORING_OP_SENDMSG] = {
892 .unbound_nonreg_file = 1,
894 .needs_async_data = 1,
895 .async_size = sizeof(struct io_async_msghdr),
897 [IORING_OP_RECVMSG] = {
899 .unbound_nonreg_file = 1,
902 .needs_async_data = 1,
903 .async_size = sizeof(struct io_async_msghdr),
905 [IORING_OP_TIMEOUT] = {
906 .needs_async_data = 1,
907 .async_size = sizeof(struct io_timeout_data),
909 [IORING_OP_TIMEOUT_REMOVE] = {
910 /* used by timeout updates' prep() */
912 [IORING_OP_ACCEPT] = {
914 .unbound_nonreg_file = 1,
917 [IORING_OP_ASYNC_CANCEL] = {},
918 [IORING_OP_LINK_TIMEOUT] = {
919 .needs_async_data = 1,
920 .async_size = sizeof(struct io_timeout_data),
922 [IORING_OP_CONNECT] = {
924 .unbound_nonreg_file = 1,
926 .needs_async_data = 1,
927 .async_size = sizeof(struct io_async_connect),
929 [IORING_OP_FALLOCATE] = {
932 [IORING_OP_OPENAT] = {},
933 [IORING_OP_CLOSE] = {},
934 [IORING_OP_FILES_UPDATE] = {},
935 [IORING_OP_STATX] = {},
938 .unbound_nonreg_file = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_WRITE] = {
946 .unbound_nonreg_file = 1,
949 .async_size = sizeof(struct io_async_rw),
951 [IORING_OP_FADVISE] = {
954 [IORING_OP_MADVISE] = {},
957 .unbound_nonreg_file = 1,
962 .unbound_nonreg_file = 1,
966 [IORING_OP_OPENAT2] = {
968 [IORING_OP_EPOLL_CTL] = {
969 .unbound_nonreg_file = 1,
971 [IORING_OP_SPLICE] = {
974 .unbound_nonreg_file = 1,
976 [IORING_OP_PROVIDE_BUFFERS] = {},
977 [IORING_OP_REMOVE_BUFFERS] = {},
981 .unbound_nonreg_file = 1,
983 [IORING_OP_SHUTDOWN] = {
986 [IORING_OP_RENAMEAT] = {},
987 [IORING_OP_UNLINKAT] = {},
990 static void io_uring_del_task_file(unsigned long index);
991 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
992 struct task_struct *task,
993 struct files_struct *files);
994 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
995 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
996 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
997 struct io_ring_ctx *ctx);
998 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1000 static bool io_rw_reissue(struct io_kiocb *req);
1001 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1002 static void io_put_req(struct io_kiocb *req);
1003 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1004 static void io_double_put_req(struct io_kiocb *req);
1005 static void io_dismantle_req(struct io_kiocb *req);
1006 static void io_put_task(struct task_struct *task, int nr);
1007 static void io_queue_next(struct io_kiocb *req);
1008 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1009 static void __io_queue_linked_timeout(struct io_kiocb *req);
1010 static void io_queue_linked_timeout(struct io_kiocb *req);
1011 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1012 struct io_uring_rsrc_update *ip,
1014 static void __io_clean_op(struct io_kiocb *req);
1015 static struct file *io_file_get(struct io_submit_state *state,
1016 struct io_kiocb *req, int fd, bool fixed);
1017 static void __io_queue_sqe(struct io_kiocb *req);
1018 static void io_rsrc_put_work(struct work_struct *work);
1020 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1021 struct iov_iter *iter, bool needs_lock);
1022 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1023 const struct iovec *fast_iov,
1024 struct iov_iter *iter, bool force);
1025 static void io_req_task_queue(struct io_kiocb *req);
1026 static void io_submit_flush_completions(struct io_comp_state *cs,
1027 struct io_ring_ctx *ctx);
1029 static struct kmem_cache *req_cachep;
1031 static const struct file_operations io_uring_fops;
1033 struct sock *io_uring_get_socket(struct file *file)
1035 #if defined(CONFIG_UNIX)
1036 if (file->f_op == &io_uring_fops) {
1037 struct io_ring_ctx *ctx = file->private_data;
1039 return ctx->ring_sock->sk;
1044 EXPORT_SYMBOL(io_uring_get_socket);
1046 #define io_for_each_link(pos, head) \
1047 for (pos = (head); pos; pos = pos->link)
1049 static inline void io_clean_op(struct io_kiocb *req)
1051 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1055 static inline void io_set_resource_node(struct io_kiocb *req)
1057 struct io_ring_ctx *ctx = req->ctx;
1059 if (!req->fixed_rsrc_refs) {
1060 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1061 percpu_ref_get(req->fixed_rsrc_refs);
1065 static bool io_match_task(struct io_kiocb *head,
1066 struct task_struct *task,
1067 struct files_struct *files)
1069 struct io_kiocb *req;
1071 if (task && head->task != task) {
1072 /* in terms of cancelation, always match if req task is dead */
1073 if (head->task->flags & PF_EXITING)
1080 io_for_each_link(req, head) {
1081 if (req->flags & REQ_F_INFLIGHT)
1083 if (req->task->files == files)
1089 static inline void req_set_fail_links(struct io_kiocb *req)
1091 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1092 req->flags |= REQ_F_FAIL_LINK;
1095 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1097 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1099 complete(&ctx->ref_comp);
1102 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1104 return !req->timeout.off;
1107 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1109 struct io_ring_ctx *ctx;
1112 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1117 * Use 5 bits less than the max cq entries, that should give us around
1118 * 32 entries per hash list if totally full and uniformly spread.
1120 hash_bits = ilog2(p->cq_entries);
1124 ctx->cancel_hash_bits = hash_bits;
1125 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1127 if (!ctx->cancel_hash)
1129 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1131 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1132 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1135 ctx->flags = p->flags;
1136 init_waitqueue_head(&ctx->sqo_sq_wait);
1137 INIT_LIST_HEAD(&ctx->sqd_list);
1138 init_waitqueue_head(&ctx->cq_wait);
1139 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1140 init_completion(&ctx->ref_comp);
1141 init_completion(&ctx->sq_thread_comp);
1142 idr_init(&ctx->io_buffer_idr);
1143 idr_init(&ctx->personality_idr);
1144 mutex_init(&ctx->uring_lock);
1145 init_waitqueue_head(&ctx->wait);
1146 spin_lock_init(&ctx->completion_lock);
1147 INIT_LIST_HEAD(&ctx->iopoll_list);
1148 INIT_LIST_HEAD(&ctx->defer_list);
1149 INIT_LIST_HEAD(&ctx->timeout_list);
1150 spin_lock_init(&ctx->inflight_lock);
1151 INIT_LIST_HEAD(&ctx->inflight_list);
1152 spin_lock_init(&ctx->rsrc_ref_lock);
1153 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1154 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1155 init_llist_head(&ctx->rsrc_put_llist);
1156 INIT_LIST_HEAD(&ctx->tctx_list);
1157 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1158 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1161 kfree(ctx->cancel_hash);
1166 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1168 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1169 struct io_ring_ctx *ctx = req->ctx;
1171 return seq != ctx->cached_cq_tail
1172 + READ_ONCE(ctx->cached_cq_overflow);
1178 static void io_req_track_inflight(struct io_kiocb *req)
1180 struct io_ring_ctx *ctx = req->ctx;
1182 if (!(req->flags & REQ_F_INFLIGHT)) {
1183 req->flags |= REQ_F_INFLIGHT;
1185 spin_lock_irq(&ctx->inflight_lock);
1186 list_add(&req->inflight_entry, &ctx->inflight_list);
1187 spin_unlock_irq(&ctx->inflight_lock);
1191 static void io_prep_async_work(struct io_kiocb *req)
1193 const struct io_op_def *def = &io_op_defs[req->opcode];
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!req->work.creds)
1197 req->work.creds = get_current_cred();
1199 if (req->flags & REQ_F_FORCE_ASYNC)
1200 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1202 if (req->flags & REQ_F_ISREG) {
1203 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1204 io_wq_hash_work(&req->work, file_inode(req->file));
1206 if (def->unbound_nonreg_file)
1207 req->work.flags |= IO_WQ_WORK_UNBOUND;
1211 static void io_prep_async_link(struct io_kiocb *req)
1213 struct io_kiocb *cur;
1215 io_for_each_link(cur, req)
1216 io_prep_async_work(cur);
1219 static void io_queue_async_work(struct io_kiocb *req)
1221 struct io_ring_ctx *ctx = req->ctx;
1222 struct io_kiocb *link = io_prep_linked_timeout(req);
1223 struct io_uring_task *tctx = req->task->io_uring;
1226 BUG_ON(!tctx->io_wq);
1228 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1229 &req->work, req->flags);
1230 /* init ->work of the whole link before punting */
1231 io_prep_async_link(req);
1232 io_wq_enqueue(tctx->io_wq, &req->work);
1234 io_queue_linked_timeout(link);
1237 static void io_kill_timeout(struct io_kiocb *req)
1239 struct io_timeout_data *io = req->async_data;
1242 ret = hrtimer_try_to_cancel(&io->timer);
1244 atomic_set(&req->ctx->cq_timeouts,
1245 atomic_read(&req->ctx->cq_timeouts) + 1);
1246 list_del_init(&req->timeout.list);
1247 io_cqring_fill_event(req, 0);
1248 io_put_req_deferred(req, 1);
1253 * Returns true if we found and killed one or more timeouts
1255 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1256 struct files_struct *files)
1258 struct io_kiocb *req, *tmp;
1261 spin_lock_irq(&ctx->completion_lock);
1262 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1263 if (io_match_task(req, tsk, files)) {
1264 io_kill_timeout(req);
1268 spin_unlock_irq(&ctx->completion_lock);
1269 return canceled != 0;
1272 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1275 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1276 struct io_defer_entry, list);
1278 if (req_need_defer(de->req, de->seq))
1280 list_del_init(&de->list);
1281 io_req_task_queue(de->req);
1283 } while (!list_empty(&ctx->defer_list));
1286 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1290 if (list_empty(&ctx->timeout_list))
1293 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1296 u32 events_needed, events_got;
1297 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1298 struct io_kiocb, timeout.list);
1300 if (io_is_timeout_noseq(req))
1304 * Since seq can easily wrap around over time, subtract
1305 * the last seq at which timeouts were flushed before comparing.
1306 * Assuming not more than 2^31-1 events have happened since,
1307 * these subtractions won't have wrapped, so we can check if
1308 * target is in [last_seq, current_seq] by comparing the two.
1310 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1311 events_got = seq - ctx->cq_last_tm_flush;
1312 if (events_got < events_needed)
1315 list_del_init(&req->timeout.list);
1316 io_kill_timeout(req);
1317 } while (!list_empty(&ctx->timeout_list));
1319 ctx->cq_last_tm_flush = seq;
1322 static void io_commit_cqring(struct io_ring_ctx *ctx)
1324 io_flush_timeouts(ctx);
1326 /* order cqe stores with ring update */
1327 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1329 if (unlikely(!list_empty(&ctx->defer_list)))
1330 __io_queue_deferred(ctx);
1333 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1335 struct io_rings *r = ctx->rings;
1337 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1340 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1342 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1345 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1347 struct io_rings *rings = ctx->rings;
1351 * writes to the cq entry need to come after reading head; the
1352 * control dependency is enough as we're using WRITE_ONCE to
1355 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1358 tail = ctx->cached_cq_tail++;
1359 return &rings->cqes[tail & ctx->cq_mask];
1362 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1366 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1368 if (!ctx->eventfd_async)
1370 return io_wq_current_is_worker();
1373 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1375 /* see waitqueue_active() comment */
1378 if (waitqueue_active(&ctx->wait))
1379 wake_up(&ctx->wait);
1380 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1381 wake_up(&ctx->sq_data->wait);
1382 if (io_should_trigger_evfd(ctx))
1383 eventfd_signal(ctx->cq_ev_fd, 1);
1384 if (waitqueue_active(&ctx->cq_wait)) {
1385 wake_up_interruptible(&ctx->cq_wait);
1386 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1390 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1392 /* see waitqueue_active() comment */
1395 if (ctx->flags & IORING_SETUP_SQPOLL) {
1396 if (waitqueue_active(&ctx->wait))
1397 wake_up(&ctx->wait);
1399 if (io_should_trigger_evfd(ctx))
1400 eventfd_signal(ctx->cq_ev_fd, 1);
1401 if (waitqueue_active(&ctx->cq_wait)) {
1402 wake_up_interruptible(&ctx->cq_wait);
1403 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1407 /* Returns true if there are no backlogged entries after the flush */
1408 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1409 struct task_struct *tsk,
1410 struct files_struct *files)
1412 struct io_rings *rings = ctx->rings;
1413 struct io_kiocb *req, *tmp;
1414 struct io_uring_cqe *cqe;
1415 unsigned long flags;
1416 bool all_flushed, posted;
1419 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1423 spin_lock_irqsave(&ctx->completion_lock, flags);
1424 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1425 if (!io_match_task(req, tsk, files))
1428 cqe = io_get_cqring(ctx);
1432 list_move(&req->compl.list, &list);
1434 WRITE_ONCE(cqe->user_data, req->user_data);
1435 WRITE_ONCE(cqe->res, req->result);
1436 WRITE_ONCE(cqe->flags, req->compl.cflags);
1438 ctx->cached_cq_overflow++;
1439 WRITE_ONCE(ctx->rings->cq_overflow,
1440 ctx->cached_cq_overflow);
1445 all_flushed = list_empty(&ctx->cq_overflow_list);
1447 clear_bit(0, &ctx->sq_check_overflow);
1448 clear_bit(0, &ctx->cq_check_overflow);
1449 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1453 io_commit_cqring(ctx);
1454 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1456 io_cqring_ev_posted(ctx);
1458 while (!list_empty(&list)) {
1459 req = list_first_entry(&list, struct io_kiocb, compl.list);
1460 list_del(&req->compl.list);
1467 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1468 struct task_struct *tsk,
1469 struct files_struct *files)
1473 if (test_bit(0, &ctx->cq_check_overflow)) {
1474 /* iopoll syncs against uring_lock, not completion_lock */
1475 if (ctx->flags & IORING_SETUP_IOPOLL)
1476 mutex_lock(&ctx->uring_lock);
1477 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1478 if (ctx->flags & IORING_SETUP_IOPOLL)
1479 mutex_unlock(&ctx->uring_lock);
1485 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1487 struct io_ring_ctx *ctx = req->ctx;
1488 struct io_uring_cqe *cqe;
1490 trace_io_uring_complete(ctx, req->user_data, res);
1493 * If we can't get a cq entry, userspace overflowed the
1494 * submission (by quite a lot). Increment the overflow count in
1497 cqe = io_get_cqring(ctx);
1499 WRITE_ONCE(cqe->user_data, req->user_data);
1500 WRITE_ONCE(cqe->res, res);
1501 WRITE_ONCE(cqe->flags, cflags);
1502 } else if (ctx->cq_overflow_flushed ||
1503 atomic_read(&req->task->io_uring->in_idle)) {
1505 * If we're in ring overflow flush mode, or in task cancel mode,
1506 * then we cannot store the request for later flushing, we need
1507 * to drop it on the floor.
1509 ctx->cached_cq_overflow++;
1510 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1512 if (list_empty(&ctx->cq_overflow_list)) {
1513 set_bit(0, &ctx->sq_check_overflow);
1514 set_bit(0, &ctx->cq_check_overflow);
1515 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1519 req->compl.cflags = cflags;
1520 refcount_inc(&req->refs);
1521 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1525 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1527 __io_cqring_fill_event(req, res, 0);
1530 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1531 unsigned int cflags)
1533 struct io_ring_ctx *ctx = req->ctx;
1534 unsigned long flags;
1536 spin_lock_irqsave(&ctx->completion_lock, flags);
1537 __io_cqring_fill_event(req, res, cflags);
1538 io_commit_cqring(ctx);
1540 * If we're the last reference to this request, add to our locked
1543 if (refcount_dec_and_test(&req->refs)) {
1544 struct io_comp_state *cs = &ctx->submit_state.comp;
1546 io_dismantle_req(req);
1547 io_put_task(req->task, 1);
1548 list_add(&req->compl.list, &cs->locked_free_list);
1549 cs->locked_free_nr++;
1552 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1554 io_cqring_ev_posted(ctx);
1557 percpu_ref_put(&ctx->refs);
1561 static void io_req_complete_state(struct io_kiocb *req, long res,
1562 unsigned int cflags)
1566 req->compl.cflags = cflags;
1567 req->flags |= REQ_F_COMPLETE_INLINE;
1570 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1571 long res, unsigned cflags)
1573 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1574 io_req_complete_state(req, res, cflags);
1576 io_req_complete_post(req, res, cflags);
1579 static inline void io_req_complete(struct io_kiocb *req, long res)
1581 __io_req_complete(req, 0, res, 0);
1584 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1586 struct io_submit_state *state = &ctx->submit_state;
1587 struct io_comp_state *cs = &state->comp;
1588 struct io_kiocb *req = NULL;
1591 * If we have more than a batch's worth of requests in our IRQ side
1592 * locked cache, grab the lock and move them over to our submission
1595 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1596 spin_lock_irq(&ctx->completion_lock);
1597 list_splice_init(&cs->locked_free_list, &cs->free_list);
1598 cs->locked_free_nr = 0;
1599 spin_unlock_irq(&ctx->completion_lock);
1602 while (!list_empty(&cs->free_list)) {
1603 req = list_first_entry(&cs->free_list, struct io_kiocb,
1605 list_del(&req->compl.list);
1606 state->reqs[state->free_reqs++] = req;
1607 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1614 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1616 struct io_submit_state *state = &ctx->submit_state;
1618 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1620 if (!state->free_reqs) {
1621 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1624 if (io_flush_cached_reqs(ctx))
1627 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1631 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1632 * retry single alloc to be on the safe side.
1634 if (unlikely(ret <= 0)) {
1635 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1636 if (!state->reqs[0])
1640 state->free_reqs = ret;
1644 return state->reqs[state->free_reqs];
1647 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1654 static void io_dismantle_req(struct io_kiocb *req)
1658 if (req->async_data)
1659 kfree(req->async_data);
1661 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1662 if (req->fixed_rsrc_refs)
1663 percpu_ref_put(req->fixed_rsrc_refs);
1664 if (req->work.creds) {
1665 put_cred(req->work.creds);
1666 req->work.creds = NULL;
1669 if (req->flags & REQ_F_INFLIGHT) {
1670 struct io_ring_ctx *ctx = req->ctx;
1671 unsigned long flags;
1673 spin_lock_irqsave(&ctx->inflight_lock, flags);
1674 list_del(&req->inflight_entry);
1675 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1676 req->flags &= ~REQ_F_INFLIGHT;
1680 /* must to be called somewhat shortly after putting a request */
1681 static inline void io_put_task(struct task_struct *task, int nr)
1683 struct io_uring_task *tctx = task->io_uring;
1685 percpu_counter_sub(&tctx->inflight, nr);
1686 if (unlikely(atomic_read(&tctx->in_idle)))
1687 wake_up(&tctx->wait);
1688 put_task_struct_many(task, nr);
1691 static void __io_free_req(struct io_kiocb *req)
1693 struct io_ring_ctx *ctx = req->ctx;
1695 io_dismantle_req(req);
1696 io_put_task(req->task, 1);
1698 kmem_cache_free(req_cachep, req);
1699 percpu_ref_put(&ctx->refs);
1702 static inline void io_remove_next_linked(struct io_kiocb *req)
1704 struct io_kiocb *nxt = req->link;
1706 req->link = nxt->link;
1710 static void io_kill_linked_timeout(struct io_kiocb *req)
1712 struct io_ring_ctx *ctx = req->ctx;
1713 struct io_kiocb *link;
1714 bool cancelled = false;
1715 unsigned long flags;
1717 spin_lock_irqsave(&ctx->completion_lock, flags);
1721 * Can happen if a linked timeout fired and link had been like
1722 * req -> link t-out -> link t-out [-> ...]
1724 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1725 struct io_timeout_data *io = link->async_data;
1728 io_remove_next_linked(req);
1729 link->timeout.head = NULL;
1730 ret = hrtimer_try_to_cancel(&io->timer);
1732 io_cqring_fill_event(link, -ECANCELED);
1733 io_commit_cqring(ctx);
1737 req->flags &= ~REQ_F_LINK_TIMEOUT;
1738 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1741 io_cqring_ev_posted(ctx);
1747 static void io_fail_links(struct io_kiocb *req)
1749 struct io_kiocb *link, *nxt;
1750 struct io_ring_ctx *ctx = req->ctx;
1751 unsigned long flags;
1753 spin_lock_irqsave(&ctx->completion_lock, flags);
1761 trace_io_uring_fail_link(req, link);
1762 io_cqring_fill_event(link, -ECANCELED);
1764 io_put_req_deferred(link, 2);
1767 io_commit_cqring(ctx);
1768 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1770 io_cqring_ev_posted(ctx);
1773 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1775 if (req->flags & REQ_F_LINK_TIMEOUT)
1776 io_kill_linked_timeout(req);
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 (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1785 struct io_kiocb *nxt = req->link;
1794 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1796 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1798 return __io_req_find_next(req);
1801 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1805 if (ctx->submit_state.comp.nr) {
1806 mutex_lock(&ctx->uring_lock);
1807 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1808 mutex_unlock(&ctx->uring_lock);
1810 percpu_ref_put(&ctx->refs);
1813 static bool __tctx_task_work(struct io_uring_task *tctx)
1815 struct io_ring_ctx *ctx = NULL;
1816 struct io_wq_work_list list;
1817 struct io_wq_work_node *node;
1819 if (wq_list_empty(&tctx->task_list))
1822 spin_lock_irq(&tctx->task_lock);
1823 list = tctx->task_list;
1824 INIT_WQ_LIST(&tctx->task_list);
1825 spin_unlock_irq(&tctx->task_lock);
1829 struct io_wq_work_node *next = node->next;
1830 struct io_kiocb *req;
1832 req = container_of(node, struct io_kiocb, io_task_work.node);
1833 if (req->ctx != ctx) {
1834 ctx_flush_and_put(ctx);
1836 percpu_ref_get(&ctx->refs);
1839 req->task_work.func(&req->task_work);
1843 ctx_flush_and_put(ctx);
1844 return list.first != NULL;
1847 static void tctx_task_work(struct callback_head *cb)
1849 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1851 clear_bit(0, &tctx->task_state);
1853 while (__tctx_task_work(tctx))
1857 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1858 enum task_work_notify_mode notify)
1860 struct io_uring_task *tctx = tsk->io_uring;
1861 struct io_wq_work_node *node, *prev;
1862 unsigned long flags;
1865 WARN_ON_ONCE(!tctx);
1867 spin_lock_irqsave(&tctx->task_lock, flags);
1868 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1869 spin_unlock_irqrestore(&tctx->task_lock, flags);
1871 /* task_work already pending, we're done */
1872 if (test_bit(0, &tctx->task_state) ||
1873 test_and_set_bit(0, &tctx->task_state))
1876 if (!task_work_add(tsk, &tctx->task_work, notify))
1880 * Slow path - we failed, find and delete work. if the work is not
1881 * in the list, it got run and we're fine.
1884 spin_lock_irqsave(&tctx->task_lock, flags);
1885 wq_list_for_each(node, prev, &tctx->task_list) {
1886 if (&req->io_task_work.node == node) {
1887 wq_list_del(&tctx->task_list, node, prev);
1892 spin_unlock_irqrestore(&tctx->task_lock, flags);
1893 clear_bit(0, &tctx->task_state);
1897 static int io_req_task_work_add(struct io_kiocb *req)
1899 struct task_struct *tsk = req->task;
1900 struct io_ring_ctx *ctx = req->ctx;
1901 enum task_work_notify_mode notify;
1904 if (tsk->flags & PF_EXITING)
1908 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1909 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1910 * processing task_work. There's no reliable way to tell if TWA_RESUME
1914 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1915 notify = TWA_SIGNAL;
1917 ret = io_task_work_add(tsk, req, notify);
1919 wake_up_process(tsk);
1924 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1925 task_work_func_t cb)
1927 struct io_ring_ctx *ctx = req->ctx;
1928 struct callback_head *head;
1930 init_task_work(&req->task_work, cb);
1932 head = READ_ONCE(ctx->exit_task_work);
1933 req->task_work.next = head;
1934 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1937 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1939 struct io_ring_ctx *ctx = req->ctx;
1941 spin_lock_irq(&ctx->completion_lock);
1942 io_cqring_fill_event(req, error);
1943 io_commit_cqring(ctx);
1944 spin_unlock_irq(&ctx->completion_lock);
1946 io_cqring_ev_posted(ctx);
1947 req_set_fail_links(req);
1948 io_double_put_req(req);
1951 static void io_req_task_cancel(struct callback_head *cb)
1953 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1954 struct io_ring_ctx *ctx = req->ctx;
1956 mutex_lock(&ctx->uring_lock);
1957 __io_req_task_cancel(req, req->result);
1958 mutex_unlock(&ctx->uring_lock);
1959 percpu_ref_put(&ctx->refs);
1962 static void __io_req_task_submit(struct io_kiocb *req)
1964 struct io_ring_ctx *ctx = req->ctx;
1966 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1967 mutex_lock(&ctx->uring_lock);
1968 if (!(current->flags & PF_EXITING) && !current->in_execve)
1969 __io_queue_sqe(req);
1971 __io_req_task_cancel(req, -EFAULT);
1972 mutex_unlock(&ctx->uring_lock);
1975 static void io_req_task_submit(struct callback_head *cb)
1977 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1979 __io_req_task_submit(req);
1982 static void io_req_task_queue(struct io_kiocb *req)
1986 req->task_work.func = io_req_task_submit;
1987 ret = io_req_task_work_add(req);
1988 if (unlikely(ret)) {
1989 req->result = -ECANCELED;
1990 percpu_ref_get(&req->ctx->refs);
1991 io_req_task_work_add_fallback(req, io_req_task_cancel);
1995 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1997 percpu_ref_get(&req->ctx->refs);
1999 req->task_work.func = io_req_task_cancel;
2001 if (unlikely(io_req_task_work_add(req)))
2002 io_req_task_work_add_fallback(req, io_req_task_cancel);
2005 static inline void io_queue_next(struct io_kiocb *req)
2007 struct io_kiocb *nxt = io_req_find_next(req);
2010 io_req_task_queue(nxt);
2013 static void io_free_req(struct io_kiocb *req)
2020 struct task_struct *task;
2025 static inline void io_init_req_batch(struct req_batch *rb)
2032 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2033 struct req_batch *rb)
2036 io_put_task(rb->task, rb->task_refs);
2038 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2041 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2042 struct io_submit_state *state)
2046 if (req->task != rb->task) {
2048 io_put_task(rb->task, rb->task_refs);
2049 rb->task = req->task;
2055 io_dismantle_req(req);
2056 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2057 state->reqs[state->free_reqs++] = req;
2059 list_add(&req->compl.list, &state->comp.free_list);
2062 static void io_submit_flush_completions(struct io_comp_state *cs,
2063 struct io_ring_ctx *ctx)
2066 struct io_kiocb *req;
2067 struct req_batch rb;
2069 io_init_req_batch(&rb);
2070 spin_lock_irq(&ctx->completion_lock);
2071 for (i = 0; i < nr; i++) {
2073 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2075 io_commit_cqring(ctx);
2076 spin_unlock_irq(&ctx->completion_lock);
2078 io_cqring_ev_posted(ctx);
2079 for (i = 0; i < nr; i++) {
2082 /* submission and completion refs */
2083 if (refcount_sub_and_test(2, &req->refs))
2084 io_req_free_batch(&rb, req, &ctx->submit_state);
2087 io_req_free_batch_finish(ctx, &rb);
2092 * Drop reference to request, return next in chain (if there is one) if this
2093 * was the last reference to this request.
2095 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2097 struct io_kiocb *nxt = NULL;
2099 if (refcount_dec_and_test(&req->refs)) {
2100 nxt = io_req_find_next(req);
2106 static void io_put_req(struct io_kiocb *req)
2108 if (refcount_dec_and_test(&req->refs))
2112 static void io_put_req_deferred_cb(struct callback_head *cb)
2114 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2119 static void io_free_req_deferred(struct io_kiocb *req)
2123 req->task_work.func = io_put_req_deferred_cb;
2124 ret = io_req_task_work_add(req);
2126 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2129 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2131 if (refcount_sub_and_test(refs, &req->refs))
2132 io_free_req_deferred(req);
2135 static void io_double_put_req(struct io_kiocb *req)
2137 /* drop both submit and complete references */
2138 if (refcount_sub_and_test(2, &req->refs))
2142 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2144 /* See comment at the top of this file */
2146 return __io_cqring_events(ctx);
2149 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2151 struct io_rings *rings = ctx->rings;
2153 /* make sure SQ entry isn't read before tail */
2154 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2157 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2159 unsigned int cflags;
2161 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2162 cflags |= IORING_CQE_F_BUFFER;
2163 req->flags &= ~REQ_F_BUFFER_SELECTED;
2168 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2170 struct io_buffer *kbuf;
2172 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2173 return io_put_kbuf(req, kbuf);
2176 static inline bool io_run_task_work(void)
2179 * Not safe to run on exiting task, and the task_work handling will
2180 * not add work to such a task.
2182 if (unlikely(current->flags & PF_EXITING))
2184 if (current->task_works) {
2185 __set_current_state(TASK_RUNNING);
2194 * Find and free completed poll iocbs
2196 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2197 struct list_head *done)
2199 struct req_batch rb;
2200 struct io_kiocb *req;
2202 /* order with ->result store in io_complete_rw_iopoll() */
2205 io_init_req_batch(&rb);
2206 while (!list_empty(done)) {
2209 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2210 list_del(&req->inflight_entry);
2212 if (READ_ONCE(req->result) == -EAGAIN) {
2213 req->iopoll_completed = 0;
2214 if (io_rw_reissue(req))
2218 if (req->flags & REQ_F_BUFFER_SELECTED)
2219 cflags = io_put_rw_kbuf(req);
2221 __io_cqring_fill_event(req, req->result, cflags);
2224 if (refcount_dec_and_test(&req->refs))
2225 io_req_free_batch(&rb, req, &ctx->submit_state);
2228 io_commit_cqring(ctx);
2229 io_cqring_ev_posted_iopoll(ctx);
2230 io_req_free_batch_finish(ctx, &rb);
2233 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2236 struct io_kiocb *req, *tmp;
2242 * Only spin for completions if we don't have multiple devices hanging
2243 * off our complete list, and we're under the requested amount.
2245 spin = !ctx->poll_multi_file && *nr_events < min;
2248 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2249 struct kiocb *kiocb = &req->rw.kiocb;
2252 * Move completed and retryable entries to our local lists.
2253 * If we find a request that requires polling, break out
2254 * and complete those lists first, if we have entries there.
2256 if (READ_ONCE(req->iopoll_completed)) {
2257 list_move_tail(&req->inflight_entry, &done);
2260 if (!list_empty(&done))
2263 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2267 /* iopoll may have completed current req */
2268 if (READ_ONCE(req->iopoll_completed))
2269 list_move_tail(&req->inflight_entry, &done);
2276 if (!list_empty(&done))
2277 io_iopoll_complete(ctx, nr_events, &done);
2283 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2284 * non-spinning poll check - we'll still enter the driver poll loop, but only
2285 * as a non-spinning completion check.
2287 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2290 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2293 ret = io_do_iopoll(ctx, nr_events, min);
2296 if (*nr_events >= min)
2304 * We can't just wait for polled events to come to us, we have to actively
2305 * find and complete them.
2307 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2309 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2312 mutex_lock(&ctx->uring_lock);
2313 while (!list_empty(&ctx->iopoll_list)) {
2314 unsigned int nr_events = 0;
2316 io_do_iopoll(ctx, &nr_events, 0);
2318 /* let it sleep and repeat later if can't complete a request */
2322 * Ensure we allow local-to-the-cpu processing to take place,
2323 * in this case we need to ensure that we reap all events.
2324 * Also let task_work, etc. to progress by releasing the mutex
2326 if (need_resched()) {
2327 mutex_unlock(&ctx->uring_lock);
2329 mutex_lock(&ctx->uring_lock);
2332 mutex_unlock(&ctx->uring_lock);
2335 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2337 unsigned int nr_events = 0;
2338 int iters = 0, ret = 0;
2341 * We disallow the app entering submit/complete with polling, but we
2342 * still need to lock the ring to prevent racing with polled issue
2343 * that got punted to a workqueue.
2345 mutex_lock(&ctx->uring_lock);
2348 * Don't enter poll loop if we already have events pending.
2349 * If we do, we can potentially be spinning for commands that
2350 * already triggered a CQE (eg in error).
2352 if (test_bit(0, &ctx->cq_check_overflow))
2353 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2354 if (io_cqring_events(ctx))
2358 * If a submit got punted to a workqueue, we can have the
2359 * application entering polling for a command before it gets
2360 * issued. That app will hold the uring_lock for the duration
2361 * of the poll right here, so we need to take a breather every
2362 * now and then to ensure that the issue has a chance to add
2363 * the poll to the issued list. Otherwise we can spin here
2364 * forever, while the workqueue is stuck trying to acquire the
2367 if (!(++iters & 7)) {
2368 mutex_unlock(&ctx->uring_lock);
2370 mutex_lock(&ctx->uring_lock);
2373 ret = io_iopoll_getevents(ctx, &nr_events, min);
2377 } while (min && !nr_events && !need_resched());
2379 mutex_unlock(&ctx->uring_lock);
2383 static void kiocb_end_write(struct io_kiocb *req)
2386 * Tell lockdep we inherited freeze protection from submission
2389 if (req->flags & REQ_F_ISREG) {
2390 struct inode *inode = file_inode(req->file);
2392 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2394 file_end_write(req->file);
2398 static bool io_resubmit_prep(struct io_kiocb *req)
2400 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2402 struct iov_iter iter;
2404 /* already prepared */
2405 if (req->async_data)
2408 switch (req->opcode) {
2409 case IORING_OP_READV:
2410 case IORING_OP_READ_FIXED:
2411 case IORING_OP_READ:
2414 case IORING_OP_WRITEV:
2415 case IORING_OP_WRITE_FIXED:
2416 case IORING_OP_WRITE:
2420 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2425 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2428 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2431 static bool io_rw_should_reissue(struct io_kiocb *req)
2433 umode_t mode = file_inode(req->file)->i_mode;
2434 struct io_ring_ctx *ctx = req->ctx;
2436 if (!S_ISBLK(mode) && !S_ISREG(mode))
2438 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2439 !(ctx->flags & IORING_SETUP_IOPOLL)))
2442 * If ref is dying, we might be running poll reap from the exit work.
2443 * Don't attempt to reissue from that path, just let it fail with
2446 if (percpu_ref_is_dying(&ctx->refs))
2452 static bool io_rw_reissue(struct io_kiocb *req)
2455 if (!io_rw_should_reissue(req))
2458 lockdep_assert_held(&req->ctx->uring_lock);
2460 if (io_resubmit_prep(req)) {
2461 refcount_inc(&req->refs);
2462 io_queue_async_work(req);
2465 req_set_fail_links(req);
2470 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2471 unsigned int issue_flags)
2475 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2477 if (res != req->result)
2478 req_set_fail_links(req);
2480 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2481 kiocb_end_write(req);
2482 if (req->flags & REQ_F_BUFFER_SELECTED)
2483 cflags = io_put_rw_kbuf(req);
2484 __io_req_complete(req, issue_flags, res, cflags);
2487 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2489 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2491 __io_complete_rw(req, res, res2, 0);
2494 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2496 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2499 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2500 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2501 struct io_async_rw *rw = req->async_data;
2504 iov_iter_revert(&rw->iter,
2505 req->result - iov_iter_count(&rw->iter));
2506 else if (!io_resubmit_prep(req))
2511 if (kiocb->ki_flags & IOCB_WRITE)
2512 kiocb_end_write(req);
2514 if (res != -EAGAIN && res != req->result)
2515 req_set_fail_links(req);
2517 WRITE_ONCE(req->result, res);
2518 /* order with io_poll_complete() checking ->result */
2520 WRITE_ONCE(req->iopoll_completed, 1);
2524 * After the iocb has been issued, it's safe to be found on the poll list.
2525 * Adding the kiocb to the list AFTER submission ensures that we don't
2526 * find it from a io_iopoll_getevents() thread before the issuer is done
2527 * accessing the kiocb cookie.
2529 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2531 struct io_ring_ctx *ctx = req->ctx;
2534 * Track whether we have multiple files in our lists. This will impact
2535 * how we do polling eventually, not spinning if we're on potentially
2536 * different devices.
2538 if (list_empty(&ctx->iopoll_list)) {
2539 ctx->poll_multi_file = false;
2540 } else if (!ctx->poll_multi_file) {
2541 struct io_kiocb *list_req;
2543 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2545 if (list_req->file != req->file)
2546 ctx->poll_multi_file = true;
2550 * For fast devices, IO may have already completed. If it has, add
2551 * it to the front so we find it first.
2553 if (READ_ONCE(req->iopoll_completed))
2554 list_add(&req->inflight_entry, &ctx->iopoll_list);
2556 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2559 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2560 * task context or in io worker task context. If current task context is
2561 * sq thread, we don't need to check whether should wake up sq thread.
2563 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2564 wq_has_sleeper(&ctx->sq_data->wait))
2565 wake_up(&ctx->sq_data->wait);
2568 static inline void io_state_file_put(struct io_submit_state *state)
2570 if (state->file_refs) {
2571 fput_many(state->file, state->file_refs);
2572 state->file_refs = 0;
2577 * Get as many references to a file as we have IOs left in this submission,
2578 * assuming most submissions are for one file, or at least that each file
2579 * has more than one submission.
2581 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2586 if (state->file_refs) {
2587 if (state->fd == fd) {
2591 io_state_file_put(state);
2593 state->file = fget_many(fd, state->ios_left);
2594 if (unlikely(!state->file))
2598 state->file_refs = state->ios_left - 1;
2602 static bool io_bdev_nowait(struct block_device *bdev)
2604 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2608 * If we tracked the file through the SCM inflight mechanism, we could support
2609 * any file. For now, just ensure that anything potentially problematic is done
2612 static bool io_file_supports_async(struct file *file, int rw)
2614 umode_t mode = file_inode(file)->i_mode;
2616 if (S_ISBLK(mode)) {
2617 if (IS_ENABLED(CONFIG_BLOCK) &&
2618 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2622 if (S_ISCHR(mode) || S_ISSOCK(mode))
2624 if (S_ISREG(mode)) {
2625 if (IS_ENABLED(CONFIG_BLOCK) &&
2626 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2627 file->f_op != &io_uring_fops)
2632 /* any ->read/write should understand O_NONBLOCK */
2633 if (file->f_flags & O_NONBLOCK)
2636 if (!(file->f_mode & FMODE_NOWAIT))
2640 return file->f_op->read_iter != NULL;
2642 return file->f_op->write_iter != NULL;
2645 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2647 struct io_ring_ctx *ctx = req->ctx;
2648 struct kiocb *kiocb = &req->rw.kiocb;
2649 struct file *file = req->file;
2653 if (S_ISREG(file_inode(file)->i_mode))
2654 req->flags |= REQ_F_ISREG;
2656 kiocb->ki_pos = READ_ONCE(sqe->off);
2657 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2658 req->flags |= REQ_F_CUR_POS;
2659 kiocb->ki_pos = file->f_pos;
2661 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2662 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2663 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2667 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2668 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2669 req->flags |= REQ_F_NOWAIT;
2671 ioprio = READ_ONCE(sqe->ioprio);
2673 ret = ioprio_check_cap(ioprio);
2677 kiocb->ki_ioprio = ioprio;
2679 kiocb->ki_ioprio = get_current_ioprio();
2681 if (ctx->flags & IORING_SETUP_IOPOLL) {
2682 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2683 !kiocb->ki_filp->f_op->iopoll)
2686 kiocb->ki_flags |= IOCB_HIPRI;
2687 kiocb->ki_complete = io_complete_rw_iopoll;
2688 req->iopoll_completed = 0;
2690 if (kiocb->ki_flags & IOCB_HIPRI)
2692 kiocb->ki_complete = io_complete_rw;
2695 req->rw.addr = READ_ONCE(sqe->addr);
2696 req->rw.len = READ_ONCE(sqe->len);
2697 req->buf_index = READ_ONCE(sqe->buf_index);
2701 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2707 case -ERESTARTNOINTR:
2708 case -ERESTARTNOHAND:
2709 case -ERESTART_RESTARTBLOCK:
2711 * We can't just restart the syscall, since previously
2712 * submitted sqes may already be in progress. Just fail this
2718 kiocb->ki_complete(kiocb, ret, 0);
2722 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2723 unsigned int issue_flags)
2725 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2726 struct io_async_rw *io = req->async_data;
2728 /* add previously done IO, if any */
2729 if (io && io->bytes_done > 0) {
2731 ret = io->bytes_done;
2733 ret += io->bytes_done;
2736 if (req->flags & REQ_F_CUR_POS)
2737 req->file->f_pos = kiocb->ki_pos;
2738 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2739 __io_complete_rw(req, ret, 0, issue_flags);
2741 io_rw_done(kiocb, ret);
2744 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2746 struct io_ring_ctx *ctx = req->ctx;
2747 size_t len = req->rw.len;
2748 struct io_mapped_ubuf *imu;
2749 u16 index, buf_index = req->buf_index;
2753 if (unlikely(buf_index >= ctx->nr_user_bufs))
2755 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2756 imu = &ctx->user_bufs[index];
2757 buf_addr = req->rw.addr;
2760 if (buf_addr + len < buf_addr)
2762 /* not inside the mapped region */
2763 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2767 * May not be a start of buffer, set size appropriately
2768 * and advance us to the beginning.
2770 offset = buf_addr - imu->ubuf;
2771 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2775 * Don't use iov_iter_advance() here, as it's really slow for
2776 * using the latter parts of a big fixed buffer - it iterates
2777 * over each segment manually. We can cheat a bit here, because
2780 * 1) it's a BVEC iter, we set it up
2781 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2782 * first and last bvec
2784 * So just find our index, and adjust the iterator afterwards.
2785 * If the offset is within the first bvec (or the whole first
2786 * bvec, just use iov_iter_advance(). This makes it easier
2787 * since we can just skip the first segment, which may not
2788 * be PAGE_SIZE aligned.
2790 const struct bio_vec *bvec = imu->bvec;
2792 if (offset <= bvec->bv_len) {
2793 iov_iter_advance(iter, offset);
2795 unsigned long seg_skip;
2797 /* skip first vec */
2798 offset -= bvec->bv_len;
2799 seg_skip = 1 + (offset >> PAGE_SHIFT);
2801 iter->bvec = bvec + seg_skip;
2802 iter->nr_segs -= seg_skip;
2803 iter->count -= bvec->bv_len + offset;
2804 iter->iov_offset = offset & ~PAGE_MASK;
2811 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2814 mutex_unlock(&ctx->uring_lock);
2817 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2820 * "Normal" inline submissions always hold the uring_lock, since we
2821 * grab it from the system call. Same is true for the SQPOLL offload.
2822 * The only exception is when we've detached the request and issue it
2823 * from an async worker thread, grab the lock for that case.
2826 mutex_lock(&ctx->uring_lock);
2829 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2830 int bgid, struct io_buffer *kbuf,
2833 struct io_buffer *head;
2835 if (req->flags & REQ_F_BUFFER_SELECTED)
2838 io_ring_submit_lock(req->ctx, needs_lock);
2840 lockdep_assert_held(&req->ctx->uring_lock);
2842 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2844 if (!list_empty(&head->list)) {
2845 kbuf = list_last_entry(&head->list, struct io_buffer,
2847 list_del(&kbuf->list);
2850 idr_remove(&req->ctx->io_buffer_idr, bgid);
2852 if (*len > kbuf->len)
2855 kbuf = ERR_PTR(-ENOBUFS);
2858 io_ring_submit_unlock(req->ctx, needs_lock);
2863 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2866 struct io_buffer *kbuf;
2869 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2870 bgid = req->buf_index;
2871 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2874 req->rw.addr = (u64) (unsigned long) kbuf;
2875 req->flags |= REQ_F_BUFFER_SELECTED;
2876 return u64_to_user_ptr(kbuf->addr);
2879 #ifdef CONFIG_COMPAT
2880 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2883 struct compat_iovec __user *uiov;
2884 compat_ssize_t clen;
2888 uiov = u64_to_user_ptr(req->rw.addr);
2889 if (!access_ok(uiov, sizeof(*uiov)))
2891 if (__get_user(clen, &uiov->iov_len))
2897 buf = io_rw_buffer_select(req, &len, needs_lock);
2899 return PTR_ERR(buf);
2900 iov[0].iov_base = buf;
2901 iov[0].iov_len = (compat_size_t) len;
2906 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2909 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2913 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2916 len = iov[0].iov_len;
2919 buf = io_rw_buffer_select(req, &len, needs_lock);
2921 return PTR_ERR(buf);
2922 iov[0].iov_base = buf;
2923 iov[0].iov_len = len;
2927 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2930 if (req->flags & REQ_F_BUFFER_SELECTED) {
2931 struct io_buffer *kbuf;
2933 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2934 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2935 iov[0].iov_len = kbuf->len;
2938 if (req->rw.len != 1)
2941 #ifdef CONFIG_COMPAT
2942 if (req->ctx->compat)
2943 return io_compat_import(req, iov, needs_lock);
2946 return __io_iov_buffer_select(req, iov, needs_lock);
2949 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2950 struct iov_iter *iter, bool needs_lock)
2952 void __user *buf = u64_to_user_ptr(req->rw.addr);
2953 size_t sqe_len = req->rw.len;
2954 u8 opcode = req->opcode;
2957 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2959 return io_import_fixed(req, rw, iter);
2962 /* buffer index only valid with fixed read/write, or buffer select */
2963 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2966 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2967 if (req->flags & REQ_F_BUFFER_SELECT) {
2968 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2970 return PTR_ERR(buf);
2971 req->rw.len = sqe_len;
2974 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2979 if (req->flags & REQ_F_BUFFER_SELECT) {
2980 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2982 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2987 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2991 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2993 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2997 * For files that don't have ->read_iter() and ->write_iter(), handle them
2998 * by looping over ->read() or ->write() manually.
3000 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3002 struct kiocb *kiocb = &req->rw.kiocb;
3003 struct file *file = req->file;
3007 * Don't support polled IO through this interface, and we can't
3008 * support non-blocking either. For the latter, this just causes
3009 * the kiocb to be handled from an async context.
3011 if (kiocb->ki_flags & IOCB_HIPRI)
3013 if (kiocb->ki_flags & IOCB_NOWAIT)
3016 while (iov_iter_count(iter)) {
3020 if (!iov_iter_is_bvec(iter)) {
3021 iovec = iov_iter_iovec(iter);
3023 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3024 iovec.iov_len = req->rw.len;
3028 nr = file->f_op->read(file, iovec.iov_base,
3029 iovec.iov_len, io_kiocb_ppos(kiocb));
3031 nr = file->f_op->write(file, iovec.iov_base,
3032 iovec.iov_len, io_kiocb_ppos(kiocb));
3041 if (nr != iovec.iov_len)
3045 iov_iter_advance(iter, nr);
3051 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3052 const struct iovec *fast_iov, struct iov_iter *iter)
3054 struct io_async_rw *rw = req->async_data;
3056 memcpy(&rw->iter, iter, sizeof(*iter));
3057 rw->free_iovec = iovec;
3059 /* can only be fixed buffers, no need to do anything */
3060 if (iov_iter_is_bvec(iter))
3063 unsigned iov_off = 0;
3065 rw->iter.iov = rw->fast_iov;
3066 if (iter->iov != fast_iov) {
3067 iov_off = iter->iov - fast_iov;
3068 rw->iter.iov += iov_off;
3070 if (rw->fast_iov != fast_iov)
3071 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3072 sizeof(struct iovec) * iter->nr_segs);
3074 req->flags |= REQ_F_NEED_CLEANUP;
3078 static inline int __io_alloc_async_data(struct io_kiocb *req)
3080 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3081 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3082 return req->async_data == NULL;
3085 static int io_alloc_async_data(struct io_kiocb *req)
3087 if (!io_op_defs[req->opcode].needs_async_data)
3090 return __io_alloc_async_data(req);
3093 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3094 const struct iovec *fast_iov,
3095 struct iov_iter *iter, bool force)
3097 if (!force && !io_op_defs[req->opcode].needs_async_data)
3099 if (!req->async_data) {
3100 if (__io_alloc_async_data(req)) {
3105 io_req_map_rw(req, iovec, fast_iov, iter);
3110 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3112 struct io_async_rw *iorw = req->async_data;
3113 struct iovec *iov = iorw->fast_iov;
3116 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3117 if (unlikely(ret < 0))
3120 iorw->bytes_done = 0;
3121 iorw->free_iovec = iov;
3123 req->flags |= REQ_F_NEED_CLEANUP;
3127 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3129 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3131 return io_prep_rw(req, sqe);
3135 * This is our waitqueue callback handler, registered through lock_page_async()
3136 * when we initially tried to do the IO with the iocb armed our waitqueue.
3137 * This gets called when the page is unlocked, and we generally expect that to
3138 * happen when the page IO is completed and the page is now uptodate. This will
3139 * queue a task_work based retry of the operation, attempting to copy the data
3140 * again. If the latter fails because the page was NOT uptodate, then we will
3141 * do a thread based blocking retry of the operation. That's the unexpected
3144 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3145 int sync, void *arg)
3147 struct wait_page_queue *wpq;
3148 struct io_kiocb *req = wait->private;
3149 struct wait_page_key *key = arg;
3151 wpq = container_of(wait, struct wait_page_queue, wait);
3153 if (!wake_page_match(wpq, key))
3156 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3157 list_del_init(&wait->entry);
3159 /* submit ref gets dropped, acquire a new one */
3160 refcount_inc(&req->refs);
3161 io_req_task_queue(req);
3166 * This controls whether a given IO request should be armed for async page
3167 * based retry. If we return false here, the request is handed to the async
3168 * worker threads for retry. If we're doing buffered reads on a regular file,
3169 * we prepare a private wait_page_queue entry and retry the operation. This
3170 * will either succeed because the page is now uptodate and unlocked, or it
3171 * will register a callback when the page is unlocked at IO completion. Through
3172 * that callback, io_uring uses task_work to setup a retry of the operation.
3173 * That retry will attempt the buffered read again. The retry will generally
3174 * succeed, or in rare cases where it fails, we then fall back to using the
3175 * async worker threads for a blocking retry.
3177 static bool io_rw_should_retry(struct io_kiocb *req)
3179 struct io_async_rw *rw = req->async_data;
3180 struct wait_page_queue *wait = &rw->wpq;
3181 struct kiocb *kiocb = &req->rw.kiocb;
3183 /* never retry for NOWAIT, we just complete with -EAGAIN */
3184 if (req->flags & REQ_F_NOWAIT)
3187 /* Only for buffered IO */
3188 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3192 * just use poll if we can, and don't attempt if the fs doesn't
3193 * support callback based unlocks
3195 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3198 wait->wait.func = io_async_buf_func;
3199 wait->wait.private = req;
3200 wait->wait.flags = 0;
3201 INIT_LIST_HEAD(&wait->wait.entry);
3202 kiocb->ki_flags |= IOCB_WAITQ;
3203 kiocb->ki_flags &= ~IOCB_NOWAIT;
3204 kiocb->ki_waitq = wait;
3208 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3210 if (req->file->f_op->read_iter)
3211 return call_read_iter(req->file, &req->rw.kiocb, iter);
3212 else if (req->file->f_op->read)
3213 return loop_rw_iter(READ, req, iter);
3218 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3220 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3221 struct kiocb *kiocb = &req->rw.kiocb;
3222 struct iov_iter __iter, *iter = &__iter;
3223 struct io_async_rw *rw = req->async_data;
3224 ssize_t io_size, ret, ret2;
3225 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3231 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3235 io_size = iov_iter_count(iter);
3236 req->result = io_size;
3238 /* Ensure we clear previously set non-block flag */
3239 if (!force_nonblock)
3240 kiocb->ki_flags &= ~IOCB_NOWAIT;
3242 kiocb->ki_flags |= IOCB_NOWAIT;
3244 /* If the file doesn't support async, just async punt */
3245 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3246 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3247 return ret ?: -EAGAIN;
3250 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3251 if (unlikely(ret)) {
3256 ret = io_iter_do_read(req, iter);
3258 if (ret == -EIOCBQUEUED) {
3259 if (req->async_data)
3260 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3262 } else if (ret == -EAGAIN) {
3263 /* IOPOLL retry should happen for io-wq threads */
3264 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3266 /* no retry on NONBLOCK nor RWF_NOWAIT */
3267 if (req->flags & REQ_F_NOWAIT)
3269 /* some cases will consume bytes even on error returns */
3270 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3272 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3273 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3274 /* read all, failed, already did sync or don't want to retry */
3278 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3283 rw = req->async_data;
3284 /* now use our persistent iterator, if we aren't already */
3289 rw->bytes_done += ret;
3290 /* if we can retry, do so with the callbacks armed */
3291 if (!io_rw_should_retry(req)) {
3292 kiocb->ki_flags &= ~IOCB_WAITQ;
3297 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3298 * we get -EIOCBQUEUED, then we'll get a notification when the
3299 * desired page gets unlocked. We can also get a partial read
3300 * here, and if we do, then just retry at the new offset.
3302 ret = io_iter_do_read(req, iter);
3303 if (ret == -EIOCBQUEUED)
3305 /* we got some bytes, but not all. retry. */
3306 kiocb->ki_flags &= ~IOCB_WAITQ;
3307 } while (ret > 0 && ret < io_size);
3309 kiocb_done(kiocb, ret, issue_flags);
3311 /* it's faster to check here then delegate to kfree */
3317 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3319 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3321 return io_prep_rw(req, sqe);
3324 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3326 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3327 struct kiocb *kiocb = &req->rw.kiocb;
3328 struct iov_iter __iter, *iter = &__iter;
3329 struct io_async_rw *rw = req->async_data;
3330 ssize_t ret, ret2, io_size;
3331 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3337 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3341 io_size = iov_iter_count(iter);
3342 req->result = io_size;
3344 /* Ensure we clear previously set non-block flag */
3345 if (!force_nonblock)
3346 kiocb->ki_flags &= ~IOCB_NOWAIT;
3348 kiocb->ki_flags |= IOCB_NOWAIT;
3350 /* If the file doesn't support async, just async punt */
3351 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3354 /* file path doesn't support NOWAIT for non-direct_IO */
3355 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3356 (req->flags & REQ_F_ISREG))
3359 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3364 * Open-code file_start_write here to grab freeze protection,
3365 * which will be released by another thread in
3366 * io_complete_rw(). Fool lockdep by telling it the lock got
3367 * released so that it doesn't complain about the held lock when
3368 * we return to userspace.
3370 if (req->flags & REQ_F_ISREG) {
3371 sb_start_write(file_inode(req->file)->i_sb);
3372 __sb_writers_release(file_inode(req->file)->i_sb,
3375 kiocb->ki_flags |= IOCB_WRITE;
3377 if (req->file->f_op->write_iter)
3378 ret2 = call_write_iter(req->file, kiocb, iter);
3379 else if (req->file->f_op->write)
3380 ret2 = loop_rw_iter(WRITE, req, iter);
3385 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3386 * retry them without IOCB_NOWAIT.
3388 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3390 /* no retry on NONBLOCK nor RWF_NOWAIT */
3391 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3393 if (ret2 == -EIOCBQUEUED && req->async_data)
3394 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3395 if (!force_nonblock || ret2 != -EAGAIN) {
3396 /* IOPOLL retry should happen for io-wq threads */
3397 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3400 kiocb_done(kiocb, ret2, issue_flags);
3403 /* some cases will consume bytes even on error returns */
3404 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3405 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3406 return ret ?: -EAGAIN;
3409 /* it's reportedly faster than delegating the null check to kfree() */
3415 static int io_renameat_prep(struct io_kiocb *req,
3416 const struct io_uring_sqe *sqe)
3418 struct io_rename *ren = &req->rename;
3419 const char __user *oldf, *newf;
3421 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3424 ren->old_dfd = READ_ONCE(sqe->fd);
3425 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3426 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3427 ren->new_dfd = READ_ONCE(sqe->len);
3428 ren->flags = READ_ONCE(sqe->rename_flags);
3430 ren->oldpath = getname(oldf);
3431 if (IS_ERR(ren->oldpath))
3432 return PTR_ERR(ren->oldpath);
3434 ren->newpath = getname(newf);
3435 if (IS_ERR(ren->newpath)) {
3436 putname(ren->oldpath);
3437 return PTR_ERR(ren->newpath);
3440 req->flags |= REQ_F_NEED_CLEANUP;
3444 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3446 struct io_rename *ren = &req->rename;
3449 if (issue_flags & IO_URING_F_NONBLOCK)
3452 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3453 ren->newpath, ren->flags);
3455 req->flags &= ~REQ_F_NEED_CLEANUP;
3457 req_set_fail_links(req);
3458 io_req_complete(req, ret);
3462 static int io_unlinkat_prep(struct io_kiocb *req,
3463 const struct io_uring_sqe *sqe)
3465 struct io_unlink *un = &req->unlink;
3466 const char __user *fname;
3468 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3471 un->dfd = READ_ONCE(sqe->fd);
3473 un->flags = READ_ONCE(sqe->unlink_flags);
3474 if (un->flags & ~AT_REMOVEDIR)
3477 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3478 un->filename = getname(fname);
3479 if (IS_ERR(un->filename))
3480 return PTR_ERR(un->filename);
3482 req->flags |= REQ_F_NEED_CLEANUP;
3486 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3488 struct io_unlink *un = &req->unlink;
3491 if (issue_flags & IO_URING_F_NONBLOCK)
3494 if (un->flags & AT_REMOVEDIR)
3495 ret = do_rmdir(un->dfd, un->filename);
3497 ret = do_unlinkat(un->dfd, un->filename);
3499 req->flags &= ~REQ_F_NEED_CLEANUP;
3501 req_set_fail_links(req);
3502 io_req_complete(req, ret);
3506 static int io_shutdown_prep(struct io_kiocb *req,
3507 const struct io_uring_sqe *sqe)
3509 #if defined(CONFIG_NET)
3510 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3512 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3516 req->shutdown.how = READ_ONCE(sqe->len);
3523 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3525 #if defined(CONFIG_NET)
3526 struct socket *sock;
3529 if (issue_flags & IO_URING_F_NONBLOCK)
3532 sock = sock_from_file(req->file);
3533 if (unlikely(!sock))
3536 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3538 req_set_fail_links(req);
3539 io_req_complete(req, ret);
3546 static int __io_splice_prep(struct io_kiocb *req,
3547 const struct io_uring_sqe *sqe)
3549 struct io_splice* sp = &req->splice;
3550 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3556 sp->len = READ_ONCE(sqe->len);
3557 sp->flags = READ_ONCE(sqe->splice_flags);
3559 if (unlikely(sp->flags & ~valid_flags))
3562 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3563 (sp->flags & SPLICE_F_FD_IN_FIXED));
3566 req->flags |= REQ_F_NEED_CLEANUP;
3568 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3570 * Splice operation will be punted aync, and here need to
3571 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3573 req->work.flags |= IO_WQ_WORK_UNBOUND;
3579 static int io_tee_prep(struct io_kiocb *req,
3580 const struct io_uring_sqe *sqe)
3582 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3584 return __io_splice_prep(req, sqe);
3587 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3589 struct io_splice *sp = &req->splice;
3590 struct file *in = sp->file_in;
3591 struct file *out = sp->file_out;
3592 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3595 if (issue_flags & IO_URING_F_NONBLOCK)
3598 ret = do_tee(in, out, sp->len, flags);
3600 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3601 req->flags &= ~REQ_F_NEED_CLEANUP;
3604 req_set_fail_links(req);
3605 io_req_complete(req, ret);
3609 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3611 struct io_splice* sp = &req->splice;
3613 sp->off_in = READ_ONCE(sqe->splice_off_in);
3614 sp->off_out = READ_ONCE(sqe->off);
3615 return __io_splice_prep(req, sqe);
3618 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3620 struct io_splice *sp = &req->splice;
3621 struct file *in = sp->file_in;
3622 struct file *out = sp->file_out;
3623 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3624 loff_t *poff_in, *poff_out;
3627 if (issue_flags & IO_URING_F_NONBLOCK)
3630 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3631 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3634 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3636 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3637 req->flags &= ~REQ_F_NEED_CLEANUP;
3640 req_set_fail_links(req);
3641 io_req_complete(req, ret);
3646 * IORING_OP_NOP just posts a completion event, nothing else.
3648 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3650 struct io_ring_ctx *ctx = req->ctx;
3652 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3655 __io_req_complete(req, issue_flags, 0, 0);
3659 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3661 struct io_ring_ctx *ctx = req->ctx;
3666 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3668 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3671 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3672 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3675 req->sync.off = READ_ONCE(sqe->off);
3676 req->sync.len = READ_ONCE(sqe->len);
3680 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3682 loff_t end = req->sync.off + req->sync.len;
3685 /* fsync always requires a blocking context */
3686 if (issue_flags & IO_URING_F_NONBLOCK)
3689 ret = vfs_fsync_range(req->file, req->sync.off,
3690 end > 0 ? end : LLONG_MAX,
3691 req->sync.flags & IORING_FSYNC_DATASYNC);
3693 req_set_fail_links(req);
3694 io_req_complete(req, ret);
3698 static int io_fallocate_prep(struct io_kiocb *req,
3699 const struct io_uring_sqe *sqe)
3701 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3703 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3706 req->sync.off = READ_ONCE(sqe->off);
3707 req->sync.len = READ_ONCE(sqe->addr);
3708 req->sync.mode = READ_ONCE(sqe->len);
3712 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3716 /* fallocate always requiring blocking context */
3717 if (issue_flags & IO_URING_F_NONBLOCK)
3719 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3722 req_set_fail_links(req);
3723 io_req_complete(req, ret);
3727 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3729 const char __user *fname;
3732 if (unlikely(sqe->ioprio || sqe->buf_index))
3734 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3737 /* open.how should be already initialised */
3738 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3739 req->open.how.flags |= O_LARGEFILE;
3741 req->open.dfd = READ_ONCE(sqe->fd);
3742 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3743 req->open.filename = getname(fname);
3744 if (IS_ERR(req->open.filename)) {
3745 ret = PTR_ERR(req->open.filename);
3746 req->open.filename = NULL;
3749 req->open.nofile = rlimit(RLIMIT_NOFILE);
3750 req->flags |= REQ_F_NEED_CLEANUP;
3754 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3758 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3760 mode = READ_ONCE(sqe->len);
3761 flags = READ_ONCE(sqe->open_flags);
3762 req->open.how = build_open_how(flags, mode);
3763 return __io_openat_prep(req, sqe);
3766 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3768 struct open_how __user *how;
3772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3774 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3775 len = READ_ONCE(sqe->len);
3776 if (len < OPEN_HOW_SIZE_VER0)
3779 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3784 return __io_openat_prep(req, sqe);
3787 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3789 struct open_flags op;
3792 bool resolve_nonblock;
3795 ret = build_open_flags(&req->open.how, &op);
3798 nonblock_set = op.open_flag & O_NONBLOCK;
3799 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3800 if (issue_flags & IO_URING_F_NONBLOCK) {
3802 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3803 * it'll always -EAGAIN
3805 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3807 op.lookup_flags |= LOOKUP_CACHED;
3808 op.open_flag |= O_NONBLOCK;
3811 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3815 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3816 /* only retry if RESOLVE_CACHED wasn't already set by application */
3817 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3818 file == ERR_PTR(-EAGAIN)) {
3820 * We could hang on to this 'fd', but seems like marginal
3821 * gain for something that is now known to be a slower path.
3822 * So just put it, and we'll get a new one when we retry.
3830 ret = PTR_ERR(file);
3832 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3833 file->f_flags &= ~O_NONBLOCK;
3834 fsnotify_open(file);
3835 fd_install(ret, file);
3838 putname(req->open.filename);
3839 req->flags &= ~REQ_F_NEED_CLEANUP;
3841 req_set_fail_links(req);
3842 io_req_complete(req, ret);
3846 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3848 return io_openat2(req, issue_flags);
3851 static int io_remove_buffers_prep(struct io_kiocb *req,
3852 const struct io_uring_sqe *sqe)
3854 struct io_provide_buf *p = &req->pbuf;
3857 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3860 tmp = READ_ONCE(sqe->fd);
3861 if (!tmp || tmp > USHRT_MAX)
3864 memset(p, 0, sizeof(*p));
3866 p->bgid = READ_ONCE(sqe->buf_group);
3870 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3871 int bgid, unsigned nbufs)
3875 /* shouldn't happen */
3879 /* the head kbuf is the list itself */
3880 while (!list_empty(&buf->list)) {
3881 struct io_buffer *nxt;
3883 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3884 list_del(&nxt->list);
3891 idr_remove(&ctx->io_buffer_idr, bgid);
3896 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3898 struct io_provide_buf *p = &req->pbuf;
3899 struct io_ring_ctx *ctx = req->ctx;
3900 struct io_buffer *head;
3902 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3904 io_ring_submit_lock(ctx, !force_nonblock);
3906 lockdep_assert_held(&ctx->uring_lock);
3909 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3911 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3913 req_set_fail_links(req);
3915 /* need to hold the lock to complete IOPOLL requests */
3916 if (ctx->flags & IORING_SETUP_IOPOLL) {
3917 __io_req_complete(req, issue_flags, ret, 0);
3918 io_ring_submit_unlock(ctx, !force_nonblock);
3920 io_ring_submit_unlock(ctx, !force_nonblock);
3921 __io_req_complete(req, issue_flags, ret, 0);
3926 static int io_provide_buffers_prep(struct io_kiocb *req,
3927 const struct io_uring_sqe *sqe)
3929 struct io_provide_buf *p = &req->pbuf;
3932 if (sqe->ioprio || sqe->rw_flags)
3935 tmp = READ_ONCE(sqe->fd);
3936 if (!tmp || tmp > USHRT_MAX)
3939 p->addr = READ_ONCE(sqe->addr);
3940 p->len = READ_ONCE(sqe->len);
3942 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3945 p->bgid = READ_ONCE(sqe->buf_group);
3946 tmp = READ_ONCE(sqe->off);
3947 if (tmp > USHRT_MAX)
3953 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3955 struct io_buffer *buf;
3956 u64 addr = pbuf->addr;
3957 int i, bid = pbuf->bid;
3959 for (i = 0; i < pbuf->nbufs; i++) {
3960 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3965 buf->len = pbuf->len;
3970 INIT_LIST_HEAD(&buf->list);
3973 list_add_tail(&buf->list, &(*head)->list);
3977 return i ? i : -ENOMEM;
3980 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3982 struct io_provide_buf *p = &req->pbuf;
3983 struct io_ring_ctx *ctx = req->ctx;
3984 struct io_buffer *head, *list;
3986 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3988 io_ring_submit_lock(ctx, !force_nonblock);
3990 lockdep_assert_held(&ctx->uring_lock);
3992 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3994 ret = io_add_buffers(p, &head);
3999 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4002 __io_remove_buffers(ctx, head, p->bgid, -1U);
4008 req_set_fail_links(req);
4010 /* need to hold the lock to complete IOPOLL requests */
4011 if (ctx->flags & IORING_SETUP_IOPOLL) {
4012 __io_req_complete(req, issue_flags, ret, 0);
4013 io_ring_submit_unlock(ctx, !force_nonblock);
4015 io_ring_submit_unlock(ctx, !force_nonblock);
4016 __io_req_complete(req, issue_flags, ret, 0);
4021 static int io_epoll_ctl_prep(struct io_kiocb *req,
4022 const struct io_uring_sqe *sqe)
4024 #if defined(CONFIG_EPOLL)
4025 if (sqe->ioprio || sqe->buf_index)
4027 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4030 req->epoll.epfd = READ_ONCE(sqe->fd);
4031 req->epoll.op = READ_ONCE(sqe->len);
4032 req->epoll.fd = READ_ONCE(sqe->off);
4034 if (ep_op_has_event(req->epoll.op)) {
4035 struct epoll_event __user *ev;
4037 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4038 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4048 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4050 #if defined(CONFIG_EPOLL)
4051 struct io_epoll *ie = &req->epoll;
4053 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4055 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4056 if (force_nonblock && ret == -EAGAIN)
4060 req_set_fail_links(req);
4061 __io_req_complete(req, issue_flags, ret, 0);
4068 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4070 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4071 if (sqe->ioprio || sqe->buf_index || sqe->off)
4073 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4076 req->madvise.addr = READ_ONCE(sqe->addr);
4077 req->madvise.len = READ_ONCE(sqe->len);
4078 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4085 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4087 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4088 struct io_madvise *ma = &req->madvise;
4091 if (issue_flags & IO_URING_F_NONBLOCK)
4094 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4096 req_set_fail_links(req);
4097 io_req_complete(req, ret);
4104 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4106 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4108 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4111 req->fadvise.offset = READ_ONCE(sqe->off);
4112 req->fadvise.len = READ_ONCE(sqe->len);
4113 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4117 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4119 struct io_fadvise *fa = &req->fadvise;
4122 if (issue_flags & IO_URING_F_NONBLOCK) {
4123 switch (fa->advice) {
4124 case POSIX_FADV_NORMAL:
4125 case POSIX_FADV_RANDOM:
4126 case POSIX_FADV_SEQUENTIAL:
4133 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4135 req_set_fail_links(req);
4136 io_req_complete(req, ret);
4140 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4142 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4144 if (sqe->ioprio || sqe->buf_index)
4146 if (req->flags & REQ_F_FIXED_FILE)
4149 req->statx.dfd = READ_ONCE(sqe->fd);
4150 req->statx.mask = READ_ONCE(sqe->len);
4151 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4152 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4153 req->statx.flags = READ_ONCE(sqe->statx_flags);
4158 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4160 struct io_statx *ctx = &req->statx;
4163 if (issue_flags & IO_URING_F_NONBLOCK) {
4164 /* only need file table for an actual valid fd */
4165 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4166 req->flags |= REQ_F_NO_FILE_TABLE;
4170 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4174 req_set_fail_links(req);
4175 io_req_complete(req, ret);
4179 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4181 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4183 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4184 sqe->rw_flags || sqe->buf_index)
4186 if (req->flags & REQ_F_FIXED_FILE)
4189 req->close.fd = READ_ONCE(sqe->fd);
4193 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4195 struct files_struct *files = current->files;
4196 struct io_close *close = &req->close;
4197 struct fdtable *fdt;
4203 spin_lock(&files->file_lock);
4204 fdt = files_fdtable(files);
4205 if (close->fd >= fdt->max_fds) {
4206 spin_unlock(&files->file_lock);
4209 file = fdt->fd[close->fd];
4211 spin_unlock(&files->file_lock);
4215 if (file->f_op == &io_uring_fops) {
4216 spin_unlock(&files->file_lock);
4221 /* if the file has a flush method, be safe and punt to async */
4222 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4223 spin_unlock(&files->file_lock);
4227 ret = __close_fd_get_file(close->fd, &file);
4228 spin_unlock(&files->file_lock);
4235 /* No ->flush() or already async, safely close from here */
4236 ret = filp_close(file, current->files);
4239 req_set_fail_links(req);
4242 __io_req_complete(req, issue_flags, ret, 0);
4246 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4248 struct io_ring_ctx *ctx = req->ctx;
4250 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4252 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4255 req->sync.off = READ_ONCE(sqe->off);
4256 req->sync.len = READ_ONCE(sqe->len);
4257 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4261 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4265 /* sync_file_range always requires a blocking context */
4266 if (issue_flags & IO_URING_F_NONBLOCK)
4269 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4272 req_set_fail_links(req);
4273 io_req_complete(req, ret);
4277 #if defined(CONFIG_NET)
4278 static int io_setup_async_msg(struct io_kiocb *req,
4279 struct io_async_msghdr *kmsg)
4281 struct io_async_msghdr *async_msg = req->async_data;
4285 if (io_alloc_async_data(req)) {
4286 kfree(kmsg->free_iov);
4289 async_msg = req->async_data;
4290 req->flags |= REQ_F_NEED_CLEANUP;
4291 memcpy(async_msg, kmsg, sizeof(*kmsg));
4292 async_msg->msg.msg_name = &async_msg->addr;
4293 /* if were using fast_iov, set it to the new one */
4294 if (!async_msg->free_iov)
4295 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4300 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4301 struct io_async_msghdr *iomsg)
4303 iomsg->msg.msg_name = &iomsg->addr;
4304 iomsg->free_iov = iomsg->fast_iov;
4305 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4306 req->sr_msg.msg_flags, &iomsg->free_iov);
4309 static int io_sendmsg_prep_async(struct io_kiocb *req)
4313 if (!io_op_defs[req->opcode].needs_async_data)
4315 ret = io_sendmsg_copy_hdr(req, req->async_data);
4317 req->flags |= REQ_F_NEED_CLEANUP;
4321 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4323 struct io_sr_msg *sr = &req->sr_msg;
4325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4328 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4329 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4330 sr->len = READ_ONCE(sqe->len);
4332 #ifdef CONFIG_COMPAT
4333 if (req->ctx->compat)
4334 sr->msg_flags |= MSG_CMSG_COMPAT;
4339 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4341 struct io_async_msghdr iomsg, *kmsg;
4342 struct socket *sock;
4346 sock = sock_from_file(req->file);
4347 if (unlikely(!sock))
4350 kmsg = req->async_data;
4352 ret = io_sendmsg_copy_hdr(req, &iomsg);
4358 flags = req->sr_msg.msg_flags;
4359 if (flags & MSG_DONTWAIT)
4360 req->flags |= REQ_F_NOWAIT;
4361 else if (issue_flags & IO_URING_F_NONBLOCK)
4362 flags |= MSG_DONTWAIT;
4364 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4365 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4366 return io_setup_async_msg(req, kmsg);
4367 if (ret == -ERESTARTSYS)
4370 /* fast path, check for non-NULL to avoid function call */
4372 kfree(kmsg->free_iov);
4373 req->flags &= ~REQ_F_NEED_CLEANUP;
4375 req_set_fail_links(req);
4376 __io_req_complete(req, issue_flags, ret, 0);
4380 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4382 struct io_sr_msg *sr = &req->sr_msg;
4385 struct socket *sock;
4389 sock = sock_from_file(req->file);
4390 if (unlikely(!sock))
4393 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4397 msg.msg_name = NULL;
4398 msg.msg_control = NULL;
4399 msg.msg_controllen = 0;
4400 msg.msg_namelen = 0;
4402 flags = req->sr_msg.msg_flags;
4403 if (flags & MSG_DONTWAIT)
4404 req->flags |= REQ_F_NOWAIT;
4405 else if (issue_flags & IO_URING_F_NONBLOCK)
4406 flags |= MSG_DONTWAIT;
4408 msg.msg_flags = flags;
4409 ret = sock_sendmsg(sock, &msg);
4410 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4412 if (ret == -ERESTARTSYS)
4416 req_set_fail_links(req);
4417 __io_req_complete(req, issue_flags, ret, 0);
4421 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4422 struct io_async_msghdr *iomsg)
4424 struct io_sr_msg *sr = &req->sr_msg;
4425 struct iovec __user *uiov;
4429 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4430 &iomsg->uaddr, &uiov, &iov_len);
4434 if (req->flags & REQ_F_BUFFER_SELECT) {
4437 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4439 sr->len = iomsg->fast_iov[0].iov_len;
4440 iomsg->free_iov = NULL;
4442 iomsg->free_iov = iomsg->fast_iov;
4443 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4444 &iomsg->free_iov, &iomsg->msg.msg_iter,
4453 #ifdef CONFIG_COMPAT
4454 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4455 struct io_async_msghdr *iomsg)
4457 struct compat_msghdr __user *msg_compat;
4458 struct io_sr_msg *sr = &req->sr_msg;
4459 struct compat_iovec __user *uiov;
4464 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4465 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4470 uiov = compat_ptr(ptr);
4471 if (req->flags & REQ_F_BUFFER_SELECT) {
4472 compat_ssize_t clen;
4476 if (!access_ok(uiov, sizeof(*uiov)))
4478 if (__get_user(clen, &uiov->iov_len))
4483 iomsg->free_iov = NULL;
4485 iomsg->free_iov = iomsg->fast_iov;
4486 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4487 UIO_FASTIOV, &iomsg->free_iov,
4488 &iomsg->msg.msg_iter, true);
4497 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4498 struct io_async_msghdr *iomsg)
4500 iomsg->msg.msg_name = &iomsg->addr;
4502 #ifdef CONFIG_COMPAT
4503 if (req->ctx->compat)
4504 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4507 return __io_recvmsg_copy_hdr(req, iomsg);
4510 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4513 struct io_sr_msg *sr = &req->sr_msg;
4514 struct io_buffer *kbuf;
4516 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4521 req->flags |= REQ_F_BUFFER_SELECTED;
4525 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4527 return io_put_kbuf(req, req->sr_msg.kbuf);
4530 static int io_recvmsg_prep_async(struct io_kiocb *req)
4534 if (!io_op_defs[req->opcode].needs_async_data)
4536 ret = io_recvmsg_copy_hdr(req, req->async_data);
4538 req->flags |= REQ_F_NEED_CLEANUP;
4542 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4544 struct io_sr_msg *sr = &req->sr_msg;
4546 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4549 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4550 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4551 sr->len = READ_ONCE(sqe->len);
4552 sr->bgid = READ_ONCE(sqe->buf_group);
4554 #ifdef CONFIG_COMPAT
4555 if (req->ctx->compat)
4556 sr->msg_flags |= MSG_CMSG_COMPAT;
4561 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4563 struct io_async_msghdr iomsg, *kmsg;
4564 struct socket *sock;
4565 struct io_buffer *kbuf;
4567 int ret, cflags = 0;
4568 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4570 sock = sock_from_file(req->file);
4571 if (unlikely(!sock))
4574 kmsg = req->async_data;
4576 ret = io_recvmsg_copy_hdr(req, &iomsg);
4582 if (req->flags & REQ_F_BUFFER_SELECT) {
4583 kbuf = io_recv_buffer_select(req, !force_nonblock);
4585 return PTR_ERR(kbuf);
4586 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4587 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4588 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4589 1, req->sr_msg.len);
4592 flags = req->sr_msg.msg_flags;
4593 if (flags & MSG_DONTWAIT)
4594 req->flags |= REQ_F_NOWAIT;
4595 else if (force_nonblock)
4596 flags |= MSG_DONTWAIT;
4598 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4599 kmsg->uaddr, flags);
4600 if (force_nonblock && ret == -EAGAIN)
4601 return io_setup_async_msg(req, kmsg);
4602 if (ret == -ERESTARTSYS)
4605 if (req->flags & REQ_F_BUFFER_SELECTED)
4606 cflags = io_put_recv_kbuf(req);
4607 /* fast path, check for non-NULL to avoid function call */
4609 kfree(kmsg->free_iov);
4610 req->flags &= ~REQ_F_NEED_CLEANUP;
4612 req_set_fail_links(req);
4613 __io_req_complete(req, issue_flags, ret, cflags);
4617 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4619 struct io_buffer *kbuf;
4620 struct io_sr_msg *sr = &req->sr_msg;
4622 void __user *buf = sr->buf;
4623 struct socket *sock;
4626 int ret, cflags = 0;
4627 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4629 sock = sock_from_file(req->file);
4630 if (unlikely(!sock))
4633 if (req->flags & REQ_F_BUFFER_SELECT) {
4634 kbuf = io_recv_buffer_select(req, !force_nonblock);
4636 return PTR_ERR(kbuf);
4637 buf = u64_to_user_ptr(kbuf->addr);
4640 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4644 msg.msg_name = NULL;
4645 msg.msg_control = NULL;
4646 msg.msg_controllen = 0;
4647 msg.msg_namelen = 0;
4648 msg.msg_iocb = NULL;
4651 flags = req->sr_msg.msg_flags;
4652 if (flags & MSG_DONTWAIT)
4653 req->flags |= REQ_F_NOWAIT;
4654 else if (force_nonblock)
4655 flags |= MSG_DONTWAIT;
4657 ret = sock_recvmsg(sock, &msg, flags);
4658 if (force_nonblock && ret == -EAGAIN)
4660 if (ret == -ERESTARTSYS)
4663 if (req->flags & REQ_F_BUFFER_SELECTED)
4664 cflags = io_put_recv_kbuf(req);
4666 req_set_fail_links(req);
4667 __io_req_complete(req, issue_flags, ret, cflags);
4671 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4673 struct io_accept *accept = &req->accept;
4675 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4677 if (sqe->ioprio || sqe->len || sqe->buf_index)
4680 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4681 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4682 accept->flags = READ_ONCE(sqe->accept_flags);
4683 accept->nofile = rlimit(RLIMIT_NOFILE);
4687 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4689 struct io_accept *accept = &req->accept;
4690 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4691 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4694 if (req->file->f_flags & O_NONBLOCK)
4695 req->flags |= REQ_F_NOWAIT;
4697 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4698 accept->addr_len, accept->flags,
4700 if (ret == -EAGAIN && force_nonblock)
4703 if (ret == -ERESTARTSYS)
4705 req_set_fail_links(req);
4707 __io_req_complete(req, issue_flags, ret, 0);
4711 static int io_connect_prep_async(struct io_kiocb *req)
4713 struct io_async_connect *io = req->async_data;
4714 struct io_connect *conn = &req->connect;
4716 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4719 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4721 struct io_connect *conn = &req->connect;
4723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4725 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4728 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4729 conn->addr_len = READ_ONCE(sqe->addr2);
4733 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4735 struct io_async_connect __io, *io;
4736 unsigned file_flags;
4738 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4740 if (req->async_data) {
4741 io = req->async_data;
4743 ret = move_addr_to_kernel(req->connect.addr,
4744 req->connect.addr_len,
4751 file_flags = force_nonblock ? O_NONBLOCK : 0;
4753 ret = __sys_connect_file(req->file, &io->address,
4754 req->connect.addr_len, file_flags);
4755 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4756 if (req->async_data)
4758 if (io_alloc_async_data(req)) {
4762 io = req->async_data;
4763 memcpy(req->async_data, &__io, sizeof(__io));
4766 if (ret == -ERESTARTSYS)
4770 req_set_fail_links(req);
4771 __io_req_complete(req, issue_flags, ret, 0);
4774 #else /* !CONFIG_NET */
4775 #define IO_NETOP_FN(op) \
4776 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4778 return -EOPNOTSUPP; \
4781 #define IO_NETOP_PREP(op) \
4783 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4785 return -EOPNOTSUPP; \
4788 #define IO_NETOP_PREP_ASYNC(op) \
4790 static int io_##op##_prep_async(struct io_kiocb *req) \
4792 return -EOPNOTSUPP; \
4795 IO_NETOP_PREP_ASYNC(sendmsg);
4796 IO_NETOP_PREP_ASYNC(recvmsg);
4797 IO_NETOP_PREP_ASYNC(connect);
4798 IO_NETOP_PREP(accept);
4801 #endif /* CONFIG_NET */
4803 struct io_poll_table {
4804 struct poll_table_struct pt;
4805 struct io_kiocb *req;
4809 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4810 __poll_t mask, task_work_func_t func)
4814 /* for instances that support it check for an event match first: */
4815 if (mask && !(mask & poll->events))
4818 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4820 list_del_init(&poll->wait.entry);
4823 req->task_work.func = func;
4824 percpu_ref_get(&req->ctx->refs);
4827 * If this fails, then the task is exiting. When a task exits, the
4828 * work gets canceled, so just cancel this request as well instead
4829 * of executing it. We can't safely execute it anyway, as we may not
4830 * have the needed state needed for it anyway.
4832 ret = io_req_task_work_add(req);
4833 if (unlikely(ret)) {
4834 WRITE_ONCE(poll->canceled, true);
4835 io_req_task_work_add_fallback(req, func);
4840 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4841 __acquires(&req->ctx->completion_lock)
4843 struct io_ring_ctx *ctx = req->ctx;
4845 if (!req->result && !READ_ONCE(poll->canceled)) {
4846 struct poll_table_struct pt = { ._key = poll->events };
4848 req->result = vfs_poll(req->file, &pt) & poll->events;
4851 spin_lock_irq(&ctx->completion_lock);
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 add_wait_queue(poll->head, &poll->wait);
4860 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4862 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4863 if (req->opcode == IORING_OP_POLL_ADD)
4864 return req->async_data;
4865 return req->apoll->double_poll;
4868 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4870 if (req->opcode == IORING_OP_POLL_ADD)
4872 return &req->apoll->poll;
4875 static void io_poll_remove_double(struct io_kiocb *req)
4877 struct io_poll_iocb *poll = io_poll_get_double(req);
4879 lockdep_assert_held(&req->ctx->completion_lock);
4881 if (poll && poll->head) {
4882 struct wait_queue_head *head = poll->head;
4884 spin_lock(&head->lock);
4885 list_del_init(&poll->wait.entry);
4886 if (poll->wait.private)
4887 refcount_dec(&req->refs);
4889 spin_unlock(&head->lock);
4893 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4895 struct io_ring_ctx *ctx = req->ctx;
4897 io_poll_remove_double(req);
4898 req->poll.done = true;
4899 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4900 io_commit_cqring(ctx);
4903 static void io_poll_task_func(struct callback_head *cb)
4905 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4906 struct io_ring_ctx *ctx = req->ctx;
4907 struct io_kiocb *nxt;
4909 if (io_poll_rewait(req, &req->poll)) {
4910 spin_unlock_irq(&ctx->completion_lock);
4912 hash_del(&req->hash_node);
4913 io_poll_complete(req, req->result, 0);
4914 spin_unlock_irq(&ctx->completion_lock);
4916 nxt = io_put_req_find_next(req);
4917 io_cqring_ev_posted(ctx);
4919 __io_req_task_submit(nxt);
4922 percpu_ref_put(&ctx->refs);
4925 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4926 int sync, void *key)
4928 struct io_kiocb *req = wait->private;
4929 struct io_poll_iocb *poll = io_poll_get_single(req);
4930 __poll_t mask = key_to_poll(key);
4932 /* for instances that support it check for an event match first: */
4933 if (mask && !(mask & poll->events))
4936 list_del_init(&wait->entry);
4938 if (poll && poll->head) {
4941 spin_lock(&poll->head->lock);
4942 done = list_empty(&poll->wait.entry);
4944 list_del_init(&poll->wait.entry);
4945 /* make sure double remove sees this as being gone */
4946 wait->private = NULL;
4947 spin_unlock(&poll->head->lock);
4949 /* use wait func handler, so it matches the rq type */
4950 poll->wait.func(&poll->wait, mode, sync, key);
4953 refcount_dec(&req->refs);
4957 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4958 wait_queue_func_t wake_func)
4962 poll->canceled = false;
4963 poll->events = events;
4964 INIT_LIST_HEAD(&poll->wait.entry);
4965 init_waitqueue_func_entry(&poll->wait, wake_func);
4968 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4969 struct wait_queue_head *head,
4970 struct io_poll_iocb **poll_ptr)
4972 struct io_kiocb *req = pt->req;
4975 * If poll->head is already set, it's because the file being polled
4976 * uses multiple waitqueues for poll handling (eg one for read, one
4977 * for write). Setup a separate io_poll_iocb if this happens.
4979 if (unlikely(poll->head)) {
4980 struct io_poll_iocb *poll_one = poll;
4982 /* already have a 2nd entry, fail a third attempt */
4984 pt->error = -EINVAL;
4987 /* double add on the same waitqueue head, ignore */
4988 if (poll->head == head)
4990 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4992 pt->error = -ENOMEM;
4995 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4996 refcount_inc(&req->refs);
4997 poll->wait.private = req;
5004 if (poll->events & EPOLLEXCLUSIVE)
5005 add_wait_queue_exclusive(head, &poll->wait);
5007 add_wait_queue(head, &poll->wait);
5010 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5011 struct poll_table_struct *p)
5013 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5014 struct async_poll *apoll = pt->req->apoll;
5016 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5019 static void io_async_task_func(struct callback_head *cb)
5021 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5022 struct async_poll *apoll = req->apoll;
5023 struct io_ring_ctx *ctx = req->ctx;
5025 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5027 if (io_poll_rewait(req, &apoll->poll)) {
5028 spin_unlock_irq(&ctx->completion_lock);
5029 percpu_ref_put(&ctx->refs);
5033 /* If req is still hashed, it cannot have been canceled. Don't check. */
5034 if (hash_hashed(&req->hash_node))
5035 hash_del(&req->hash_node);
5037 io_poll_remove_double(req);
5038 spin_unlock_irq(&ctx->completion_lock);
5040 if (!READ_ONCE(apoll->poll.canceled))
5041 __io_req_task_submit(req);
5043 __io_req_task_cancel(req, -ECANCELED);
5045 percpu_ref_put(&ctx->refs);
5046 kfree(apoll->double_poll);
5050 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5053 struct io_kiocb *req = wait->private;
5054 struct io_poll_iocb *poll = &req->apoll->poll;
5056 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5059 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5062 static void io_poll_req_insert(struct io_kiocb *req)
5064 struct io_ring_ctx *ctx = req->ctx;
5065 struct hlist_head *list;
5067 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5068 hlist_add_head(&req->hash_node, list);
5071 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5072 struct io_poll_iocb *poll,
5073 struct io_poll_table *ipt, __poll_t mask,
5074 wait_queue_func_t wake_func)
5075 __acquires(&ctx->completion_lock)
5077 struct io_ring_ctx *ctx = req->ctx;
5078 bool cancel = false;
5080 INIT_HLIST_NODE(&req->hash_node);
5081 io_init_poll_iocb(poll, mask, wake_func);
5082 poll->file = req->file;
5083 poll->wait.private = req;
5085 ipt->pt._key = mask;
5087 ipt->error = -EINVAL;
5089 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5091 spin_lock_irq(&ctx->completion_lock);
5092 if (likely(poll->head)) {
5093 spin_lock(&poll->head->lock);
5094 if (unlikely(list_empty(&poll->wait.entry))) {
5100 if (mask || ipt->error)
5101 list_del_init(&poll->wait.entry);
5103 WRITE_ONCE(poll->canceled, true);
5104 else if (!poll->done) /* actually waiting for an event */
5105 io_poll_req_insert(req);
5106 spin_unlock(&poll->head->lock);
5112 static bool io_arm_poll_handler(struct io_kiocb *req)
5114 const struct io_op_def *def = &io_op_defs[req->opcode];
5115 struct io_ring_ctx *ctx = req->ctx;
5116 struct async_poll *apoll;
5117 struct io_poll_table ipt;
5121 if (!req->file || !file_can_poll(req->file))
5123 if (req->flags & REQ_F_POLLED)
5127 else if (def->pollout)
5131 /* if we can't nonblock try, then no point in arming a poll handler */
5132 if (!io_file_supports_async(req->file, rw))
5135 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5136 if (unlikely(!apoll))
5138 apoll->double_poll = NULL;
5140 req->flags |= REQ_F_POLLED;
5145 mask |= POLLIN | POLLRDNORM;
5147 mask |= POLLOUT | POLLWRNORM;
5149 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5150 if ((req->opcode == IORING_OP_RECVMSG) &&
5151 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5154 mask |= POLLERR | POLLPRI;
5156 ipt.pt._qproc = io_async_queue_proc;
5158 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5160 if (ret || ipt.error) {
5161 io_poll_remove_double(req);
5162 spin_unlock_irq(&ctx->completion_lock);
5163 kfree(apoll->double_poll);
5167 spin_unlock_irq(&ctx->completion_lock);
5168 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5169 apoll->poll.events);
5173 static bool __io_poll_remove_one(struct io_kiocb *req,
5174 struct io_poll_iocb *poll)
5176 bool do_complete = false;
5178 spin_lock(&poll->head->lock);
5179 WRITE_ONCE(poll->canceled, true);
5180 if (!list_empty(&poll->wait.entry)) {
5181 list_del_init(&poll->wait.entry);
5184 spin_unlock(&poll->head->lock);
5185 hash_del(&req->hash_node);
5189 static bool io_poll_remove_one(struct io_kiocb *req)
5193 io_poll_remove_double(req);
5195 if (req->opcode == IORING_OP_POLL_ADD) {
5196 do_complete = __io_poll_remove_one(req, &req->poll);
5198 struct async_poll *apoll = req->apoll;
5200 /* non-poll requests have submit ref still */
5201 do_complete = __io_poll_remove_one(req, &apoll->poll);
5204 kfree(apoll->double_poll);
5210 io_cqring_fill_event(req, -ECANCELED);
5211 io_commit_cqring(req->ctx);
5212 req_set_fail_links(req);
5213 io_put_req_deferred(req, 1);
5220 * Returns true if we found and killed one or more poll requests
5222 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5223 struct files_struct *files)
5225 struct hlist_node *tmp;
5226 struct io_kiocb *req;
5229 spin_lock_irq(&ctx->completion_lock);
5230 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5231 struct hlist_head *list;
5233 list = &ctx->cancel_hash[i];
5234 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5235 if (io_match_task(req, tsk, files))
5236 posted += io_poll_remove_one(req);
5239 spin_unlock_irq(&ctx->completion_lock);
5242 io_cqring_ev_posted(ctx);
5247 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5249 struct hlist_head *list;
5250 struct io_kiocb *req;
5252 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5253 hlist_for_each_entry(req, list, hash_node) {
5254 if (sqe_addr != req->user_data)
5256 if (io_poll_remove_one(req))
5264 static int io_poll_remove_prep(struct io_kiocb *req,
5265 const struct io_uring_sqe *sqe)
5267 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5269 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5273 req->poll_remove.addr = READ_ONCE(sqe->addr);
5278 * Find a running poll command that matches one specified in sqe->addr,
5279 * and remove it if found.
5281 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5283 struct io_ring_ctx *ctx = req->ctx;
5286 spin_lock_irq(&ctx->completion_lock);
5287 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5288 spin_unlock_irq(&ctx->completion_lock);
5291 req_set_fail_links(req);
5292 io_req_complete(req, ret);
5296 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5299 struct io_kiocb *req = wait->private;
5300 struct io_poll_iocb *poll = &req->poll;
5302 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5305 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5306 struct poll_table_struct *p)
5308 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5310 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5313 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5315 struct io_poll_iocb *poll = &req->poll;
5318 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5320 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5323 events = READ_ONCE(sqe->poll32_events);
5325 events = swahw32(events);
5327 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5328 (events & EPOLLEXCLUSIVE);
5332 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5334 struct io_poll_iocb *poll = &req->poll;
5335 struct io_ring_ctx *ctx = req->ctx;
5336 struct io_poll_table ipt;
5339 ipt.pt._qproc = io_poll_queue_proc;
5341 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5344 if (mask) { /* no async, we'd stolen it */
5346 io_poll_complete(req, mask, 0);
5348 spin_unlock_irq(&ctx->completion_lock);
5351 io_cqring_ev_posted(ctx);
5357 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5359 struct io_timeout_data *data = container_of(timer,
5360 struct io_timeout_data, timer);
5361 struct io_kiocb *req = data->req;
5362 struct io_ring_ctx *ctx = req->ctx;
5363 unsigned long flags;
5365 spin_lock_irqsave(&ctx->completion_lock, flags);
5366 list_del_init(&req->timeout.list);
5367 atomic_set(&req->ctx->cq_timeouts,
5368 atomic_read(&req->ctx->cq_timeouts) + 1);
5370 io_cqring_fill_event(req, -ETIME);
5371 io_commit_cqring(ctx);
5372 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5374 io_cqring_ev_posted(ctx);
5375 req_set_fail_links(req);
5377 return HRTIMER_NORESTART;
5380 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5383 struct io_timeout_data *io;
5384 struct io_kiocb *req;
5387 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5388 if (user_data == req->user_data) {
5395 return ERR_PTR(ret);
5397 io = req->async_data;
5398 ret = hrtimer_try_to_cancel(&io->timer);
5400 return ERR_PTR(-EALREADY);
5401 list_del_init(&req->timeout.list);
5405 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5407 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5410 return PTR_ERR(req);
5412 req_set_fail_links(req);
5413 io_cqring_fill_event(req, -ECANCELED);
5414 io_put_req_deferred(req, 1);
5418 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5419 struct timespec64 *ts, enum hrtimer_mode mode)
5421 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5422 struct io_timeout_data *data;
5425 return PTR_ERR(req);
5427 req->timeout.off = 0; /* noseq */
5428 data = req->async_data;
5429 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5430 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5431 data->timer.function = io_timeout_fn;
5432 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5436 static int io_timeout_remove_prep(struct io_kiocb *req,
5437 const struct io_uring_sqe *sqe)
5439 struct io_timeout_rem *tr = &req->timeout_rem;
5441 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5443 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5445 if (sqe->ioprio || sqe->buf_index || sqe->len)
5448 tr->addr = READ_ONCE(sqe->addr);
5449 tr->flags = READ_ONCE(sqe->timeout_flags);
5450 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5451 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5453 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5455 } else if (tr->flags) {
5456 /* timeout removal doesn't support flags */
5463 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5465 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5470 * Remove or update an existing timeout command
5472 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5474 struct io_timeout_rem *tr = &req->timeout_rem;
5475 struct io_ring_ctx *ctx = req->ctx;
5478 spin_lock_irq(&ctx->completion_lock);
5479 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5480 ret = io_timeout_cancel(ctx, tr->addr);
5482 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5483 io_translate_timeout_mode(tr->flags));
5485 io_cqring_fill_event(req, ret);
5486 io_commit_cqring(ctx);
5487 spin_unlock_irq(&ctx->completion_lock);
5488 io_cqring_ev_posted(ctx);
5490 req_set_fail_links(req);
5495 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5496 bool is_timeout_link)
5498 struct io_timeout_data *data;
5500 u32 off = READ_ONCE(sqe->off);
5502 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5504 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5506 if (off && is_timeout_link)
5508 flags = READ_ONCE(sqe->timeout_flags);
5509 if (flags & ~IORING_TIMEOUT_ABS)
5512 req->timeout.off = off;
5514 if (!req->async_data && io_alloc_async_data(req))
5517 data = req->async_data;
5520 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5523 data->mode = io_translate_timeout_mode(flags);
5524 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5525 io_req_track_inflight(req);
5529 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5531 struct io_ring_ctx *ctx = req->ctx;
5532 struct io_timeout_data *data = req->async_data;
5533 struct list_head *entry;
5534 u32 tail, off = req->timeout.off;
5536 spin_lock_irq(&ctx->completion_lock);
5539 * sqe->off holds how many events that need to occur for this
5540 * timeout event to be satisfied. If it isn't set, then this is
5541 * a pure timeout request, sequence isn't used.
5543 if (io_is_timeout_noseq(req)) {
5544 entry = ctx->timeout_list.prev;
5548 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5549 req->timeout.target_seq = tail + off;
5551 /* Update the last seq here in case io_flush_timeouts() hasn't.
5552 * This is safe because ->completion_lock is held, and submissions
5553 * and completions are never mixed in the same ->completion_lock section.
5555 ctx->cq_last_tm_flush = tail;
5558 * Insertion sort, ensuring the first entry in the list is always
5559 * the one we need first.
5561 list_for_each_prev(entry, &ctx->timeout_list) {
5562 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5565 if (io_is_timeout_noseq(nxt))
5567 /* nxt.seq is behind @tail, otherwise would've been completed */
5568 if (off >= nxt->timeout.target_seq - tail)
5572 list_add(&req->timeout.list, entry);
5573 data->timer.function = io_timeout_fn;
5574 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5575 spin_unlock_irq(&ctx->completion_lock);
5579 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5581 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5583 return req->user_data == (unsigned long) data;
5586 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5588 enum io_wq_cancel cancel_ret;
5594 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5595 switch (cancel_ret) {
5596 case IO_WQ_CANCEL_OK:
5599 case IO_WQ_CANCEL_RUNNING:
5602 case IO_WQ_CANCEL_NOTFOUND:
5610 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5611 struct io_kiocb *req, __u64 sqe_addr,
5614 unsigned long flags;
5617 ret = io_async_cancel_one(req->task->io_uring,
5618 (void *) (unsigned long) sqe_addr);
5619 if (ret != -ENOENT) {
5620 spin_lock_irqsave(&ctx->completion_lock, flags);
5624 spin_lock_irqsave(&ctx->completion_lock, flags);
5625 ret = io_timeout_cancel(ctx, sqe_addr);
5628 ret = io_poll_cancel(ctx, sqe_addr);
5632 io_cqring_fill_event(req, ret);
5633 io_commit_cqring(ctx);
5634 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5635 io_cqring_ev_posted(ctx);
5638 req_set_fail_links(req);
5642 static int io_async_cancel_prep(struct io_kiocb *req,
5643 const struct io_uring_sqe *sqe)
5645 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5647 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5649 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5652 req->cancel.addr = READ_ONCE(sqe->addr);
5656 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5658 struct io_ring_ctx *ctx = req->ctx;
5660 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5664 static int io_rsrc_update_prep(struct io_kiocb *req,
5665 const struct io_uring_sqe *sqe)
5667 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5669 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5671 if (sqe->ioprio || sqe->rw_flags)
5674 req->rsrc_update.offset = READ_ONCE(sqe->off);
5675 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5676 if (!req->rsrc_update.nr_args)
5678 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5682 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5684 struct io_ring_ctx *ctx = req->ctx;
5685 struct io_uring_rsrc_update up;
5688 if (issue_flags & IO_URING_F_NONBLOCK)
5691 up.offset = req->rsrc_update.offset;
5692 up.data = req->rsrc_update.arg;
5694 mutex_lock(&ctx->uring_lock);
5695 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5696 mutex_unlock(&ctx->uring_lock);
5699 req_set_fail_links(req);
5700 __io_req_complete(req, issue_flags, ret, 0);
5704 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5706 switch (req->opcode) {
5709 case IORING_OP_READV:
5710 case IORING_OP_READ_FIXED:
5711 case IORING_OP_READ:
5712 return io_read_prep(req, sqe);
5713 case IORING_OP_WRITEV:
5714 case IORING_OP_WRITE_FIXED:
5715 case IORING_OP_WRITE:
5716 return io_write_prep(req, sqe);
5717 case IORING_OP_POLL_ADD:
5718 return io_poll_add_prep(req, sqe);
5719 case IORING_OP_POLL_REMOVE:
5720 return io_poll_remove_prep(req, sqe);
5721 case IORING_OP_FSYNC:
5722 return io_fsync_prep(req, sqe);
5723 case IORING_OP_SYNC_FILE_RANGE:
5724 return io_sfr_prep(req, sqe);
5725 case IORING_OP_SENDMSG:
5726 case IORING_OP_SEND:
5727 return io_sendmsg_prep(req, sqe);
5728 case IORING_OP_RECVMSG:
5729 case IORING_OP_RECV:
5730 return io_recvmsg_prep(req, sqe);
5731 case IORING_OP_CONNECT:
5732 return io_connect_prep(req, sqe);
5733 case IORING_OP_TIMEOUT:
5734 return io_timeout_prep(req, sqe, false);
5735 case IORING_OP_TIMEOUT_REMOVE:
5736 return io_timeout_remove_prep(req, sqe);
5737 case IORING_OP_ASYNC_CANCEL:
5738 return io_async_cancel_prep(req, sqe);
5739 case IORING_OP_LINK_TIMEOUT:
5740 return io_timeout_prep(req, sqe, true);
5741 case IORING_OP_ACCEPT:
5742 return io_accept_prep(req, sqe);
5743 case IORING_OP_FALLOCATE:
5744 return io_fallocate_prep(req, sqe);
5745 case IORING_OP_OPENAT:
5746 return io_openat_prep(req, sqe);
5747 case IORING_OP_CLOSE:
5748 return io_close_prep(req, sqe);
5749 case IORING_OP_FILES_UPDATE:
5750 return io_rsrc_update_prep(req, sqe);
5751 case IORING_OP_STATX:
5752 return io_statx_prep(req, sqe);
5753 case IORING_OP_FADVISE:
5754 return io_fadvise_prep(req, sqe);
5755 case IORING_OP_MADVISE:
5756 return io_madvise_prep(req, sqe);
5757 case IORING_OP_OPENAT2:
5758 return io_openat2_prep(req, sqe);
5759 case IORING_OP_EPOLL_CTL:
5760 return io_epoll_ctl_prep(req, sqe);
5761 case IORING_OP_SPLICE:
5762 return io_splice_prep(req, sqe);
5763 case IORING_OP_PROVIDE_BUFFERS:
5764 return io_provide_buffers_prep(req, sqe);
5765 case IORING_OP_REMOVE_BUFFERS:
5766 return io_remove_buffers_prep(req, sqe);
5768 return io_tee_prep(req, sqe);
5769 case IORING_OP_SHUTDOWN:
5770 return io_shutdown_prep(req, sqe);
5771 case IORING_OP_RENAMEAT:
5772 return io_renameat_prep(req, sqe);
5773 case IORING_OP_UNLINKAT:
5774 return io_unlinkat_prep(req, sqe);
5777 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5782 static int io_req_prep_async(struct io_kiocb *req)
5784 switch (req->opcode) {
5785 case IORING_OP_READV:
5786 case IORING_OP_READ_FIXED:
5787 case IORING_OP_READ:
5788 return io_rw_prep_async(req, READ);
5789 case IORING_OP_WRITEV:
5790 case IORING_OP_WRITE_FIXED:
5791 case IORING_OP_WRITE:
5792 return io_rw_prep_async(req, WRITE);
5793 case IORING_OP_SENDMSG:
5794 case IORING_OP_SEND:
5795 return io_sendmsg_prep_async(req);
5796 case IORING_OP_RECVMSG:
5797 case IORING_OP_RECV:
5798 return io_recvmsg_prep_async(req);
5799 case IORING_OP_CONNECT:
5800 return io_connect_prep_async(req);
5805 static int io_req_defer_prep(struct io_kiocb *req)
5807 if (!io_op_defs[req->opcode].needs_async_data)
5809 /* some opcodes init it during the inital prep */
5810 if (req->async_data)
5812 if (__io_alloc_async_data(req))
5814 return io_req_prep_async(req);
5817 static u32 io_get_sequence(struct io_kiocb *req)
5819 struct io_kiocb *pos;
5820 struct io_ring_ctx *ctx = req->ctx;
5821 u32 total_submitted, nr_reqs = 0;
5823 io_for_each_link(pos, req)
5826 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5827 return total_submitted - nr_reqs;
5830 static int io_req_defer(struct io_kiocb *req)
5832 struct io_ring_ctx *ctx = req->ctx;
5833 struct io_defer_entry *de;
5837 /* Still need defer if there is pending req in defer list. */
5838 if (likely(list_empty_careful(&ctx->defer_list) &&
5839 !(req->flags & REQ_F_IO_DRAIN)))
5842 seq = io_get_sequence(req);
5843 /* Still a chance to pass the sequence check */
5844 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5847 ret = io_req_defer_prep(req);
5850 io_prep_async_link(req);
5851 de = kmalloc(sizeof(*de), GFP_KERNEL);
5855 spin_lock_irq(&ctx->completion_lock);
5856 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5857 spin_unlock_irq(&ctx->completion_lock);
5859 io_queue_async_work(req);
5860 return -EIOCBQUEUED;
5863 trace_io_uring_defer(ctx, req, req->user_data);
5866 list_add_tail(&de->list, &ctx->defer_list);
5867 spin_unlock_irq(&ctx->completion_lock);
5868 return -EIOCBQUEUED;
5871 static void __io_clean_op(struct io_kiocb *req)
5873 if (req->flags & REQ_F_BUFFER_SELECTED) {
5874 switch (req->opcode) {
5875 case IORING_OP_READV:
5876 case IORING_OP_READ_FIXED:
5877 case IORING_OP_READ:
5878 kfree((void *)(unsigned long)req->rw.addr);
5880 case IORING_OP_RECVMSG:
5881 case IORING_OP_RECV:
5882 kfree(req->sr_msg.kbuf);
5885 req->flags &= ~REQ_F_BUFFER_SELECTED;
5888 if (req->flags & REQ_F_NEED_CLEANUP) {
5889 switch (req->opcode) {
5890 case IORING_OP_READV:
5891 case IORING_OP_READ_FIXED:
5892 case IORING_OP_READ:
5893 case IORING_OP_WRITEV:
5894 case IORING_OP_WRITE_FIXED:
5895 case IORING_OP_WRITE: {
5896 struct io_async_rw *io = req->async_data;
5898 kfree(io->free_iovec);
5901 case IORING_OP_RECVMSG:
5902 case IORING_OP_SENDMSG: {
5903 struct io_async_msghdr *io = req->async_data;
5905 kfree(io->free_iov);
5908 case IORING_OP_SPLICE:
5910 io_put_file(req, req->splice.file_in,
5911 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5913 case IORING_OP_OPENAT:
5914 case IORING_OP_OPENAT2:
5915 if (req->open.filename)
5916 putname(req->open.filename);
5918 case IORING_OP_RENAMEAT:
5919 putname(req->rename.oldpath);
5920 putname(req->rename.newpath);
5922 case IORING_OP_UNLINKAT:
5923 putname(req->unlink.filename);
5926 req->flags &= ~REQ_F_NEED_CLEANUP;
5930 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5932 struct io_ring_ctx *ctx = req->ctx;
5933 const struct cred *creds = NULL;
5936 if (req->work.creds && req->work.creds != current_cred())
5937 creds = override_creds(req->work.creds);
5939 switch (req->opcode) {
5941 ret = io_nop(req, issue_flags);
5943 case IORING_OP_READV:
5944 case IORING_OP_READ_FIXED:
5945 case IORING_OP_READ:
5946 ret = io_read(req, issue_flags);
5948 case IORING_OP_WRITEV:
5949 case IORING_OP_WRITE_FIXED:
5950 case IORING_OP_WRITE:
5951 ret = io_write(req, issue_flags);
5953 case IORING_OP_FSYNC:
5954 ret = io_fsync(req, issue_flags);
5956 case IORING_OP_POLL_ADD:
5957 ret = io_poll_add(req, issue_flags);
5959 case IORING_OP_POLL_REMOVE:
5960 ret = io_poll_remove(req, issue_flags);
5962 case IORING_OP_SYNC_FILE_RANGE:
5963 ret = io_sync_file_range(req, issue_flags);
5965 case IORING_OP_SENDMSG:
5966 ret = io_sendmsg(req, issue_flags);
5968 case IORING_OP_SEND:
5969 ret = io_send(req, issue_flags);
5971 case IORING_OP_RECVMSG:
5972 ret = io_recvmsg(req, issue_flags);
5974 case IORING_OP_RECV:
5975 ret = io_recv(req, issue_flags);
5977 case IORING_OP_TIMEOUT:
5978 ret = io_timeout(req, issue_flags);
5980 case IORING_OP_TIMEOUT_REMOVE:
5981 ret = io_timeout_remove(req, issue_flags);
5983 case IORING_OP_ACCEPT:
5984 ret = io_accept(req, issue_flags);
5986 case IORING_OP_CONNECT:
5987 ret = io_connect(req, issue_flags);
5989 case IORING_OP_ASYNC_CANCEL:
5990 ret = io_async_cancel(req, issue_flags);
5992 case IORING_OP_FALLOCATE:
5993 ret = io_fallocate(req, issue_flags);
5995 case IORING_OP_OPENAT:
5996 ret = io_openat(req, issue_flags);
5998 case IORING_OP_CLOSE:
5999 ret = io_close(req, issue_flags);
6001 case IORING_OP_FILES_UPDATE:
6002 ret = io_files_update(req, issue_flags);
6004 case IORING_OP_STATX:
6005 ret = io_statx(req, issue_flags);
6007 case IORING_OP_FADVISE:
6008 ret = io_fadvise(req, issue_flags);
6010 case IORING_OP_MADVISE:
6011 ret = io_madvise(req, issue_flags);
6013 case IORING_OP_OPENAT2:
6014 ret = io_openat2(req, issue_flags);
6016 case IORING_OP_EPOLL_CTL:
6017 ret = io_epoll_ctl(req, issue_flags);
6019 case IORING_OP_SPLICE:
6020 ret = io_splice(req, issue_flags);
6022 case IORING_OP_PROVIDE_BUFFERS:
6023 ret = io_provide_buffers(req, issue_flags);
6025 case IORING_OP_REMOVE_BUFFERS:
6026 ret = io_remove_buffers(req, issue_flags);
6029 ret = io_tee(req, issue_flags);
6031 case IORING_OP_SHUTDOWN:
6032 ret = io_shutdown(req, issue_flags);
6034 case IORING_OP_RENAMEAT:
6035 ret = io_renameat(req, issue_flags);
6037 case IORING_OP_UNLINKAT:
6038 ret = io_unlinkat(req, issue_flags);
6046 revert_creds(creds);
6051 /* If the op doesn't have a file, we're not polling for it */
6052 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6053 const bool in_async = io_wq_current_is_worker();
6055 /* workqueue context doesn't hold uring_lock, grab it now */
6057 mutex_lock(&ctx->uring_lock);
6059 io_iopoll_req_issued(req, in_async);
6062 mutex_unlock(&ctx->uring_lock);
6068 static void io_wq_submit_work(struct io_wq_work *work)
6070 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6071 struct io_kiocb *timeout;
6074 timeout = io_prep_linked_timeout(req);
6076 io_queue_linked_timeout(timeout);
6078 if (work->flags & IO_WQ_WORK_CANCEL)
6083 ret = io_issue_sqe(req, 0);
6085 * We can get EAGAIN for polled IO even though we're
6086 * forcing a sync submission from here, since we can't
6087 * wait for request slots on the block side.
6095 /* avoid locking problems by failing it from a clean context */
6097 /* io-wq is going to take one down */
6098 refcount_inc(&req->refs);
6099 io_req_task_queue_fail(req, ret);
6103 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6106 struct fixed_rsrc_table *table;
6108 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6109 return table->files[index & IORING_FILE_TABLE_MASK];
6112 static struct file *io_file_get(struct io_submit_state *state,
6113 struct io_kiocb *req, int fd, bool fixed)
6115 struct io_ring_ctx *ctx = req->ctx;
6119 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6121 fd = array_index_nospec(fd, ctx->nr_user_files);
6122 file = io_file_from_index(ctx, fd);
6123 io_set_resource_node(req);
6125 trace_io_uring_file_get(ctx, fd);
6126 file = __io_file_get(state, fd);
6129 if (file && unlikely(file->f_op == &io_uring_fops))
6130 io_req_track_inflight(req);
6134 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6136 struct io_timeout_data *data = container_of(timer,
6137 struct io_timeout_data, timer);
6138 struct io_kiocb *prev, *req = data->req;
6139 struct io_ring_ctx *ctx = req->ctx;
6140 unsigned long flags;
6142 spin_lock_irqsave(&ctx->completion_lock, flags);
6143 prev = req->timeout.head;
6144 req->timeout.head = NULL;
6147 * We don't expect the list to be empty, that will only happen if we
6148 * race with the completion of the linked work.
6150 if (prev && refcount_inc_not_zero(&prev->refs))
6151 io_remove_next_linked(prev);
6154 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6157 req_set_fail_links(prev);
6158 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6159 io_put_req_deferred(prev, 1);
6161 io_req_complete_post(req, -ETIME, 0);
6162 io_put_req_deferred(req, 1);
6164 return HRTIMER_NORESTART;
6167 static void __io_queue_linked_timeout(struct io_kiocb *req)
6170 * If the back reference is NULL, then our linked request finished
6171 * before we got a chance to setup the timer
6173 if (req->timeout.head) {
6174 struct io_timeout_data *data = req->async_data;
6176 data->timer.function = io_link_timeout_fn;
6177 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6182 static void io_queue_linked_timeout(struct io_kiocb *req)
6184 struct io_ring_ctx *ctx = req->ctx;
6186 spin_lock_irq(&ctx->completion_lock);
6187 __io_queue_linked_timeout(req);
6188 spin_unlock_irq(&ctx->completion_lock);
6190 /* drop submission reference */
6194 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6196 struct io_kiocb *nxt = req->link;
6198 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6199 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6202 nxt->timeout.head = req;
6203 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6204 req->flags |= REQ_F_LINK_TIMEOUT;
6208 static void __io_queue_sqe(struct io_kiocb *req)
6210 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6213 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6216 * We async punt it if the file wasn't marked NOWAIT, or if the file
6217 * doesn't support non-blocking read/write attempts
6219 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6220 if (!io_arm_poll_handler(req)) {
6222 * Queued up for async execution, worker will release
6223 * submit reference when the iocb is actually submitted.
6225 io_queue_async_work(req);
6227 } else if (likely(!ret)) {
6228 /* drop submission reference */
6229 if (req->flags & REQ_F_COMPLETE_INLINE) {
6230 struct io_ring_ctx *ctx = req->ctx;
6231 struct io_comp_state *cs = &ctx->submit_state.comp;
6233 cs->reqs[cs->nr++] = req;
6234 if (cs->nr == ARRAY_SIZE(cs->reqs))
6235 io_submit_flush_completions(cs, ctx);
6240 req_set_fail_links(req);
6242 io_req_complete(req, ret);
6245 io_queue_linked_timeout(linked_timeout);
6248 static void io_queue_sqe(struct io_kiocb *req)
6252 ret = io_req_defer(req);
6254 if (ret != -EIOCBQUEUED) {
6256 req_set_fail_links(req);
6258 io_req_complete(req, ret);
6260 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6261 ret = io_req_defer_prep(req);
6264 io_queue_async_work(req);
6266 __io_queue_sqe(req);
6271 * Check SQE restrictions (opcode and flags).
6273 * Returns 'true' if SQE is allowed, 'false' otherwise.
6275 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6276 struct io_kiocb *req,
6277 unsigned int sqe_flags)
6279 if (!ctx->restricted)
6282 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6285 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6286 ctx->restrictions.sqe_flags_required)
6289 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6290 ctx->restrictions.sqe_flags_required))
6296 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6297 const struct io_uring_sqe *sqe)
6299 struct io_submit_state *state;
6300 unsigned int sqe_flags;
6301 int personality, ret = 0;
6303 req->opcode = READ_ONCE(sqe->opcode);
6304 /* same numerical values with corresponding REQ_F_*, safe to copy */
6305 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6306 req->user_data = READ_ONCE(sqe->user_data);
6307 req->async_data = NULL;
6311 req->fixed_rsrc_refs = NULL;
6312 /* one is dropped after submission, the other at completion */
6313 refcount_set(&req->refs, 2);
6314 req->task = current;
6317 /* enforce forwards compatibility on users */
6318 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6323 if (unlikely(req->opcode >= IORING_OP_LAST))
6326 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6329 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6330 !io_op_defs[req->opcode].buffer_select)
6333 req->work.list.next = NULL;
6334 personality = READ_ONCE(sqe->personality);
6336 req->work.creds = idr_find(&ctx->personality_idr, personality);
6337 if (!req->work.creds)
6339 get_cred(req->work.creds);
6341 req->work.creds = NULL;
6343 req->work.flags = 0;
6344 state = &ctx->submit_state;
6347 * Plug now if we have more than 1 IO left after this, and the target
6348 * is potentially a read/write to block based storage.
6350 if (!state->plug_started && state->ios_left > 1 &&
6351 io_op_defs[req->opcode].plug) {
6352 blk_start_plug(&state->plug);
6353 state->plug_started = true;
6356 if (io_op_defs[req->opcode].needs_file) {
6357 bool fixed = req->flags & REQ_F_FIXED_FILE;
6359 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6360 if (unlikely(!req->file))
6368 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6369 const struct io_uring_sqe *sqe)
6371 struct io_submit_link *link = &ctx->submit_state.link;
6374 ret = io_init_req(ctx, req, sqe);
6375 if (unlikely(ret)) {
6378 io_req_complete(req, ret);
6380 /* fail even hard links since we don't submit */
6381 link->head->flags |= REQ_F_FAIL_LINK;
6382 io_put_req(link->head);
6383 io_req_complete(link->head, -ECANCELED);
6388 ret = io_req_prep(req, sqe);
6392 /* don't need @sqe from now on */
6393 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6394 true, ctx->flags & IORING_SETUP_SQPOLL);
6397 * If we already have a head request, queue this one for async
6398 * submittal once the head completes. If we don't have a head but
6399 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6400 * submitted sync once the chain is complete. If none of those
6401 * conditions are true (normal request), then just queue it.
6404 struct io_kiocb *head = link->head;
6407 * Taking sequential execution of a link, draining both sides
6408 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6409 * requests in the link. So, it drains the head and the
6410 * next after the link request. The last one is done via
6411 * drain_next flag to persist the effect across calls.
6413 if (req->flags & REQ_F_IO_DRAIN) {
6414 head->flags |= REQ_F_IO_DRAIN;
6415 ctx->drain_next = 1;
6417 ret = io_req_defer_prep(req);
6420 trace_io_uring_link(ctx, req, head);
6421 link->last->link = req;
6424 /* last request of a link, enqueue the link */
6425 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6430 if (unlikely(ctx->drain_next)) {
6431 req->flags |= REQ_F_IO_DRAIN;
6432 ctx->drain_next = 0;
6434 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6446 * Batched submission is done, ensure local IO is flushed out.
6448 static void io_submit_state_end(struct io_submit_state *state,
6449 struct io_ring_ctx *ctx)
6451 if (state->link.head)
6452 io_queue_sqe(state->link.head);
6454 io_submit_flush_completions(&state->comp, ctx);
6455 if (state->plug_started)
6456 blk_finish_plug(&state->plug);
6457 io_state_file_put(state);
6461 * Start submission side cache.
6463 static void io_submit_state_start(struct io_submit_state *state,
6464 unsigned int max_ios)
6466 state->plug_started = false;
6467 state->ios_left = max_ios;
6468 /* set only head, no need to init link_last in advance */
6469 state->link.head = NULL;
6472 static void io_commit_sqring(struct io_ring_ctx *ctx)
6474 struct io_rings *rings = ctx->rings;
6477 * Ensure any loads from the SQEs are done at this point,
6478 * since once we write the new head, the application could
6479 * write new data to them.
6481 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6485 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6486 * that is mapped by userspace. This means that care needs to be taken to
6487 * ensure that reads are stable, as we cannot rely on userspace always
6488 * being a good citizen. If members of the sqe are validated and then later
6489 * used, it's important that those reads are done through READ_ONCE() to
6490 * prevent a re-load down the line.
6492 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6494 u32 *sq_array = ctx->sq_array;
6498 * The cached sq head (or cq tail) serves two purposes:
6500 * 1) allows us to batch the cost of updating the user visible
6502 * 2) allows the kernel side to track the head on its own, even
6503 * though the application is the one updating it.
6505 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6506 if (likely(head < ctx->sq_entries))
6507 return &ctx->sq_sqes[head];
6509 /* drop invalid entries */
6510 ctx->cached_sq_dropped++;
6511 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6515 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6519 /* if we have a backlog and couldn't flush it all, return BUSY */
6520 if (test_bit(0, &ctx->sq_check_overflow)) {
6521 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6525 /* make sure SQ entry isn't read before tail */
6526 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6528 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6531 percpu_counter_add(¤t->io_uring->inflight, nr);
6532 refcount_add(nr, ¤t->usage);
6533 io_submit_state_start(&ctx->submit_state, nr);
6535 while (submitted < nr) {
6536 const struct io_uring_sqe *sqe;
6537 struct io_kiocb *req;
6539 req = io_alloc_req(ctx);
6540 if (unlikely(!req)) {
6542 submitted = -EAGAIN;
6545 sqe = io_get_sqe(ctx);
6546 if (unlikely(!sqe)) {
6547 kmem_cache_free(req_cachep, req);
6550 /* will complete beyond this point, count as submitted */
6552 if (io_submit_sqe(ctx, req, sqe))
6556 if (unlikely(submitted != nr)) {
6557 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6558 struct io_uring_task *tctx = current->io_uring;
6559 int unused = nr - ref_used;
6561 percpu_ref_put_many(&ctx->refs, unused);
6562 percpu_counter_sub(&tctx->inflight, unused);
6563 put_task_struct_many(current, unused);
6566 io_submit_state_end(&ctx->submit_state, ctx);
6567 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6568 io_commit_sqring(ctx);
6573 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6575 /* Tell userspace we may need a wakeup call */
6576 spin_lock_irq(&ctx->completion_lock);
6577 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6578 spin_unlock_irq(&ctx->completion_lock);
6581 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6583 spin_lock_irq(&ctx->completion_lock);
6584 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6585 spin_unlock_irq(&ctx->completion_lock);
6588 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6590 unsigned int to_submit;
6593 to_submit = io_sqring_entries(ctx);
6594 /* if we're handling multiple rings, cap submit size for fairness */
6595 if (cap_entries && to_submit > 8)
6598 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6599 unsigned nr_events = 0;
6601 mutex_lock(&ctx->uring_lock);
6602 if (!list_empty(&ctx->iopoll_list))
6603 io_do_iopoll(ctx, &nr_events, 0);
6605 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6606 ret = io_submit_sqes(ctx, to_submit);
6607 mutex_unlock(&ctx->uring_lock);
6610 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6611 wake_up(&ctx->sqo_sq_wait);
6616 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6618 struct io_ring_ctx *ctx;
6619 unsigned sq_thread_idle = 0;
6621 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6622 if (sq_thread_idle < ctx->sq_thread_idle)
6623 sq_thread_idle = ctx->sq_thread_idle;
6626 sqd->sq_thread_idle = sq_thread_idle;
6629 static void io_sqd_init_new(struct io_sq_data *sqd)
6631 struct io_ring_ctx *ctx;
6633 while (!list_empty(&sqd->ctx_new_list)) {
6634 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6635 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6636 complete(&ctx->sq_thread_comp);
6639 io_sqd_update_thread_idle(sqd);
6642 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6644 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6647 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6649 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6652 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6656 * TASK_PARKED is a special state; we must serialize against
6657 * possible pending wakeups to avoid store-store collisions on
6660 * Such a collision might possibly result in the task state
6661 * changin from TASK_PARKED and us failing the
6662 * wait_task_inactive() in kthread_park().
6664 set_special_state(TASK_PARKED);
6665 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6669 * Thread is going to call schedule(), do not preempt it,
6670 * or the caller of kthread_park() may spend more time in
6671 * wait_task_inactive().
6674 complete(&sqd->parked);
6675 schedule_preempt_disabled();
6678 __set_current_state(TASK_RUNNING);
6681 static int io_sq_thread(void *data)
6683 struct io_sq_data *sqd = data;
6684 struct io_ring_ctx *ctx;
6685 unsigned long timeout = 0;
6686 char buf[TASK_COMM_LEN];
6689 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6690 set_task_comm(current, buf);
6691 current->pf_io_worker = NULL;
6693 if (sqd->sq_cpu != -1)
6694 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6696 set_cpus_allowed_ptr(current, cpu_online_mask);
6697 current->flags |= PF_NO_SETAFFINITY;
6699 wait_for_completion(&sqd->startup);
6701 while (!io_sq_thread_should_stop(sqd)) {
6703 bool cap_entries, sqt_spin, needs_sched;
6706 * Any changes to the sqd lists are synchronized through the
6707 * thread parking. This synchronizes the thread vs users,
6708 * the users are synchronized on the sqd->ctx_lock.
6710 if (io_sq_thread_should_park(sqd)) {
6711 io_sq_thread_parkme(sqd);
6714 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6715 io_sqd_init_new(sqd);
6716 timeout = jiffies + sqd->sq_thread_idle;
6718 if (fatal_signal_pending(current))
6721 cap_entries = !list_is_singular(&sqd->ctx_list);
6722 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6723 ret = __io_sq_thread(ctx, cap_entries);
6724 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6728 if (sqt_spin || !time_after(jiffies, timeout)) {
6732 timeout = jiffies + sqd->sq_thread_idle;
6737 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6738 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6739 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6740 !list_empty_careful(&ctx->iopoll_list)) {
6741 needs_sched = false;
6744 if (io_sqring_entries(ctx)) {
6745 needs_sched = false;
6750 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6751 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6752 io_ring_set_wakeup_flag(ctx);
6756 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6757 io_ring_clear_wakeup_flag(ctx);
6760 finish_wait(&sqd->wait, &wait);
6761 timeout = jiffies + sqd->sq_thread_idle;
6764 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6765 io_uring_cancel_sqpoll(ctx);
6770 * Ensure that we park properly if racing with someone trying to park
6771 * while we're exiting. If we fail to grab the lock, check park and
6772 * park if necessary. The ordering with the park bit and the lock
6773 * ensures that we catch this reliably.
6775 if (!mutex_trylock(&sqd->lock)) {
6776 if (io_sq_thread_should_park(sqd))
6777 io_sq_thread_parkme(sqd);
6778 mutex_lock(&sqd->lock);
6782 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6784 io_ring_set_wakeup_flag(ctx);
6787 complete(&sqd->exited);
6788 mutex_unlock(&sqd->lock);
6792 struct io_wait_queue {
6793 struct wait_queue_entry wq;
6794 struct io_ring_ctx *ctx;
6796 unsigned nr_timeouts;
6799 static inline bool io_should_wake(struct io_wait_queue *iowq)
6801 struct io_ring_ctx *ctx = iowq->ctx;
6804 * Wake up if we have enough events, or if a timeout occurred since we
6805 * started waiting. For timeouts, we always want to return to userspace,
6806 * regardless of event count.
6808 return io_cqring_events(ctx) >= iowq->to_wait ||
6809 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6812 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6813 int wake_flags, void *key)
6815 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6819 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6820 * the task, and the next invocation will do it.
6822 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6823 return autoremove_wake_function(curr, mode, wake_flags, key);
6827 static int io_run_task_work_sig(void)
6829 if (io_run_task_work())
6831 if (!signal_pending(current))
6833 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6834 return -ERESTARTSYS;
6838 /* when returns >0, the caller should retry */
6839 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6840 struct io_wait_queue *iowq,
6841 signed long *timeout)
6845 /* make sure we run task_work before checking for signals */
6846 ret = io_run_task_work_sig();
6847 if (ret || io_should_wake(iowq))
6849 /* let the caller flush overflows, retry */
6850 if (test_bit(0, &ctx->cq_check_overflow))
6853 *timeout = schedule_timeout(*timeout);
6854 return !*timeout ? -ETIME : 1;
6858 * Wait until events become available, if we don't already have some. The
6859 * application must reap them itself, as they reside on the shared cq ring.
6861 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6862 const sigset_t __user *sig, size_t sigsz,
6863 struct __kernel_timespec __user *uts)
6865 struct io_wait_queue iowq = {
6868 .func = io_wake_function,
6869 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6872 .to_wait = min_events,
6874 struct io_rings *rings = ctx->rings;
6875 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6879 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6880 if (io_cqring_events(ctx) >= min_events)
6882 if (!io_run_task_work())
6887 #ifdef CONFIG_COMPAT
6888 if (in_compat_syscall())
6889 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6893 ret = set_user_sigmask(sig, sigsz);
6900 struct timespec64 ts;
6902 if (get_timespec64(&ts, uts))
6904 timeout = timespec64_to_jiffies(&ts);
6907 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6908 trace_io_uring_cqring_wait(ctx, min_events);
6910 /* if we can't even flush overflow, don't wait for more */
6911 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6915 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6916 TASK_INTERRUPTIBLE);
6917 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6918 finish_wait(&ctx->wait, &iowq.wq);
6922 restore_saved_sigmask_unless(ret == -EINTR);
6924 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6927 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6929 #if defined(CONFIG_UNIX)
6930 if (ctx->ring_sock) {
6931 struct sock *sock = ctx->ring_sock->sk;
6932 struct sk_buff *skb;
6934 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6940 for (i = 0; i < ctx->nr_user_files; i++) {
6943 file = io_file_from_index(ctx, i);
6950 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6952 struct fixed_rsrc_data *data;
6954 data = container_of(ref, struct fixed_rsrc_data, refs);
6955 complete(&data->done);
6958 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6960 spin_lock_bh(&ctx->rsrc_ref_lock);
6963 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6965 spin_unlock_bh(&ctx->rsrc_ref_lock);
6968 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6969 struct fixed_rsrc_data *rsrc_data,
6970 struct fixed_rsrc_ref_node *ref_node)
6972 io_rsrc_ref_lock(ctx);
6973 rsrc_data->node = ref_node;
6974 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6975 io_rsrc_ref_unlock(ctx);
6976 percpu_ref_get(&rsrc_data->refs);
6979 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6981 struct fixed_rsrc_ref_node *ref_node = NULL;
6983 io_rsrc_ref_lock(ctx);
6984 ref_node = data->node;
6986 io_rsrc_ref_unlock(ctx);
6988 percpu_ref_kill(&ref_node->refs);
6991 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6992 struct io_ring_ctx *ctx,
6993 void (*rsrc_put)(struct io_ring_ctx *ctx,
6994 struct io_rsrc_put *prsrc))
6996 struct fixed_rsrc_ref_node *backup_node;
7002 data->quiesce = true;
7005 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7008 backup_node->rsrc_data = data;
7009 backup_node->rsrc_put = rsrc_put;
7011 io_sqe_rsrc_kill_node(ctx, data);
7012 percpu_ref_kill(&data->refs);
7013 flush_delayed_work(&ctx->rsrc_put_work);
7015 ret = wait_for_completion_interruptible(&data->done);
7019 percpu_ref_resurrect(&data->refs);
7020 io_sqe_rsrc_set_node(ctx, data, backup_node);
7022 reinit_completion(&data->done);
7023 mutex_unlock(&ctx->uring_lock);
7024 ret = io_run_task_work_sig();
7025 mutex_lock(&ctx->uring_lock);
7027 data->quiesce = false;
7030 destroy_fixed_rsrc_ref_node(backup_node);
7034 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7036 struct fixed_rsrc_data *data;
7038 data = kzalloc(sizeof(*data), GFP_KERNEL);
7042 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7043 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7048 init_completion(&data->done);
7052 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7054 percpu_ref_exit(&data->refs);
7059 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7061 struct fixed_rsrc_data *data = ctx->file_data;
7062 unsigned nr_tables, i;
7066 * percpu_ref_is_dying() is to stop parallel files unregister
7067 * Since we possibly drop uring lock later in this function to
7070 if (!data || percpu_ref_is_dying(&data->refs))
7072 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7076 __io_sqe_files_unregister(ctx);
7077 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7078 for (i = 0; i < nr_tables; i++)
7079 kfree(data->table[i].files);
7080 free_fixed_rsrc_data(data);
7081 ctx->file_data = NULL;
7082 ctx->nr_user_files = 0;
7086 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7087 __releases(&sqd->lock)
7089 if (sqd->thread == current)
7091 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7093 wake_up_state(sqd->thread, TASK_PARKED);
7094 mutex_unlock(&sqd->lock);
7097 static void io_sq_thread_park(struct io_sq_data *sqd)
7098 __acquires(&sqd->lock)
7100 if (sqd->thread == current)
7102 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7103 mutex_lock(&sqd->lock);
7105 wake_up_process(sqd->thread);
7106 wait_for_completion(&sqd->parked);
7110 static void io_sq_thread_stop(struct io_sq_data *sqd)
7112 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7114 mutex_lock(&sqd->lock);
7116 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7117 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7118 wake_up_process(sqd->thread);
7119 mutex_unlock(&sqd->lock);
7120 wait_for_completion(&sqd->exited);
7121 WARN_ON_ONCE(sqd->thread);
7123 mutex_unlock(&sqd->lock);
7127 static void io_put_sq_data(struct io_sq_data *sqd)
7129 if (refcount_dec_and_test(&sqd->refs)) {
7130 io_sq_thread_stop(sqd);
7135 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7137 struct io_sq_data *sqd = ctx->sq_data;
7140 complete(&sqd->startup);
7142 wait_for_completion(&ctx->sq_thread_comp);
7143 io_sq_thread_park(sqd);
7146 mutex_lock(&sqd->ctx_lock);
7147 list_del(&ctx->sqd_list);
7148 io_sqd_update_thread_idle(sqd);
7149 mutex_unlock(&sqd->ctx_lock);
7152 io_sq_thread_unpark(sqd);
7154 io_put_sq_data(sqd);
7155 ctx->sq_data = NULL;
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);
7180 refcount_inc(&sqd->refs);
7185 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7187 struct io_sq_data *sqd;
7189 if (p->flags & IORING_SETUP_ATTACH_WQ)
7190 return io_attach_sq_data(p);
7192 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7194 return ERR_PTR(-ENOMEM);
7196 refcount_set(&sqd->refs, 1);
7197 INIT_LIST_HEAD(&sqd->ctx_list);
7198 INIT_LIST_HEAD(&sqd->ctx_new_list);
7199 mutex_init(&sqd->ctx_lock);
7200 mutex_init(&sqd->lock);
7201 init_waitqueue_head(&sqd->wait);
7202 init_completion(&sqd->startup);
7203 init_completion(&sqd->parked);
7204 init_completion(&sqd->exited);
7208 #if defined(CONFIG_UNIX)
7210 * Ensure the UNIX gc is aware of our file set, so we are certain that
7211 * the io_uring can be safely unregistered on process exit, even if we have
7212 * loops in the file referencing.
7214 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7216 struct sock *sk = ctx->ring_sock->sk;
7217 struct scm_fp_list *fpl;
7218 struct sk_buff *skb;
7221 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7225 skb = alloc_skb(0, GFP_KERNEL);
7234 fpl->user = get_uid(current_user());
7235 for (i = 0; i < nr; i++) {
7236 struct file *file = io_file_from_index(ctx, i + offset);
7240 fpl->fp[nr_files] = get_file(file);
7241 unix_inflight(fpl->user, fpl->fp[nr_files]);
7246 fpl->max = SCM_MAX_FD;
7247 fpl->count = nr_files;
7248 UNIXCB(skb).fp = fpl;
7249 skb->destructor = unix_destruct_scm;
7250 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7251 skb_queue_head(&sk->sk_receive_queue, skb);
7253 for (i = 0; i < nr_files; i++)
7264 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7265 * causes regular reference counting to break down. We rely on the UNIX
7266 * garbage collection to take care of this problem for us.
7268 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7270 unsigned left, total;
7274 left = ctx->nr_user_files;
7276 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7278 ret = __io_sqe_files_scm(ctx, this_files, total);
7282 total += this_files;
7288 while (total < ctx->nr_user_files) {
7289 struct file *file = io_file_from_index(ctx, total);
7299 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7305 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7306 unsigned nr_tables, unsigned nr_files)
7310 for (i = 0; i < nr_tables; i++) {
7311 struct fixed_rsrc_table *table = &file_data->table[i];
7312 unsigned this_files;
7314 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7315 table->files = kcalloc(this_files, sizeof(struct file *),
7319 nr_files -= this_files;
7325 for (i = 0; i < nr_tables; i++) {
7326 struct fixed_rsrc_table *table = &file_data->table[i];
7327 kfree(table->files);
7332 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7334 struct file *file = prsrc->file;
7335 #if defined(CONFIG_UNIX)
7336 struct sock *sock = ctx->ring_sock->sk;
7337 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7338 struct sk_buff *skb;
7341 __skb_queue_head_init(&list);
7344 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7345 * remove this entry and rearrange the file array.
7347 skb = skb_dequeue(head);
7349 struct scm_fp_list *fp;
7351 fp = UNIXCB(skb).fp;
7352 for (i = 0; i < fp->count; i++) {
7355 if (fp->fp[i] != file)
7358 unix_notinflight(fp->user, fp->fp[i]);
7359 left = fp->count - 1 - i;
7361 memmove(&fp->fp[i], &fp->fp[i + 1],
7362 left * sizeof(struct file *));
7369 __skb_queue_tail(&list, skb);
7379 __skb_queue_tail(&list, skb);
7381 skb = skb_dequeue(head);
7384 if (skb_peek(&list)) {
7385 spin_lock_irq(&head->lock);
7386 while ((skb = __skb_dequeue(&list)) != NULL)
7387 __skb_queue_tail(head, skb);
7388 spin_unlock_irq(&head->lock);
7395 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7397 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7398 struct io_ring_ctx *ctx = rsrc_data->ctx;
7399 struct io_rsrc_put *prsrc, *tmp;
7401 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7402 list_del(&prsrc->list);
7403 ref_node->rsrc_put(ctx, prsrc);
7407 percpu_ref_exit(&ref_node->refs);
7409 percpu_ref_put(&rsrc_data->refs);
7412 static void io_rsrc_put_work(struct work_struct *work)
7414 struct io_ring_ctx *ctx;
7415 struct llist_node *node;
7417 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7418 node = llist_del_all(&ctx->rsrc_put_llist);
7421 struct fixed_rsrc_ref_node *ref_node;
7422 struct llist_node *next = node->next;
7424 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7425 __io_rsrc_put_work(ref_node);
7430 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7433 struct fixed_rsrc_table *table;
7435 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7436 return &table->files[i & IORING_FILE_TABLE_MASK];
7439 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7441 struct fixed_rsrc_ref_node *ref_node;
7442 struct fixed_rsrc_data *data;
7443 struct io_ring_ctx *ctx;
7444 bool first_add = false;
7447 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7448 data = ref_node->rsrc_data;
7451 io_rsrc_ref_lock(ctx);
7452 ref_node->done = true;
7454 while (!list_empty(&ctx->rsrc_ref_list)) {
7455 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7456 struct fixed_rsrc_ref_node, node);
7457 /* recycle ref nodes in order */
7458 if (!ref_node->done)
7460 list_del(&ref_node->node);
7461 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7463 io_rsrc_ref_unlock(ctx);
7465 if (percpu_ref_is_dying(&data->refs))
7469 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7471 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7474 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7475 struct io_ring_ctx *ctx)
7477 struct fixed_rsrc_ref_node *ref_node;
7479 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7483 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7488 INIT_LIST_HEAD(&ref_node->node);
7489 INIT_LIST_HEAD(&ref_node->rsrc_list);
7490 ref_node->done = false;
7494 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7495 struct fixed_rsrc_ref_node *ref_node)
7497 ref_node->rsrc_data = ctx->file_data;
7498 ref_node->rsrc_put = io_ring_file_put;
7501 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7503 percpu_ref_exit(&ref_node->refs);
7508 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7511 __s32 __user *fds = (__s32 __user *) arg;
7512 unsigned nr_tables, i;
7514 int fd, ret = -ENOMEM;
7515 struct fixed_rsrc_ref_node *ref_node;
7516 struct fixed_rsrc_data *file_data;
7522 if (nr_args > IORING_MAX_FIXED_FILES)
7525 file_data = alloc_fixed_rsrc_data(ctx);
7528 ctx->file_data = file_data;
7530 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7531 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7533 if (!file_data->table)
7536 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7539 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7540 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7544 /* allow sparse sets */
7554 * Don't allow io_uring instances to be registered. If UNIX
7555 * isn't enabled, then this causes a reference cycle and this
7556 * instance can never get freed. If UNIX is enabled we'll
7557 * handle it just fine, but there's still no point in allowing
7558 * a ring fd as it doesn't support regular read/write anyway.
7560 if (file->f_op == &io_uring_fops) {
7564 *io_fixed_file_slot(file_data, i) = file;
7567 ret = io_sqe_files_scm(ctx);
7569 io_sqe_files_unregister(ctx);
7573 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7575 io_sqe_files_unregister(ctx);
7578 init_fixed_file_ref_node(ctx, ref_node);
7580 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7583 for (i = 0; i < ctx->nr_user_files; i++) {
7584 file = io_file_from_index(ctx, i);
7588 for (i = 0; i < nr_tables; i++)
7589 kfree(file_data->table[i].files);
7590 ctx->nr_user_files = 0;
7592 free_fixed_rsrc_data(ctx->file_data);
7593 ctx->file_data = NULL;
7597 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7600 #if defined(CONFIG_UNIX)
7601 struct sock *sock = ctx->ring_sock->sk;
7602 struct sk_buff_head *head = &sock->sk_receive_queue;
7603 struct sk_buff *skb;
7606 * See if we can merge this file into an existing skb SCM_RIGHTS
7607 * file set. If there's no room, fall back to allocating a new skb
7608 * and filling it in.
7610 spin_lock_irq(&head->lock);
7611 skb = skb_peek(head);
7613 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7615 if (fpl->count < SCM_MAX_FD) {
7616 __skb_unlink(skb, head);
7617 spin_unlock_irq(&head->lock);
7618 fpl->fp[fpl->count] = get_file(file);
7619 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7621 spin_lock_irq(&head->lock);
7622 __skb_queue_head(head, skb);
7627 spin_unlock_irq(&head->lock);
7634 return __io_sqe_files_scm(ctx, 1, index);
7640 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7642 struct io_rsrc_put *prsrc;
7643 struct fixed_rsrc_ref_node *ref_node = data->node;
7645 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7650 list_add(&prsrc->list, &ref_node->rsrc_list);
7655 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7658 return io_queue_rsrc_removal(data, (void *)file);
7661 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7662 struct io_uring_rsrc_update *up,
7665 struct fixed_rsrc_data *data = ctx->file_data;
7666 struct fixed_rsrc_ref_node *ref_node;
7667 struct file *file, **file_slot;
7671 bool needs_switch = false;
7673 if (check_add_overflow(up->offset, nr_args, &done))
7675 if (done > ctx->nr_user_files)
7678 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7681 init_fixed_file_ref_node(ctx, ref_node);
7683 fds = u64_to_user_ptr(up->data);
7684 for (done = 0; done < nr_args; done++) {
7686 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7690 if (fd == IORING_REGISTER_FILES_SKIP)
7693 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7694 file_slot = io_fixed_file_slot(ctx->file_data, i);
7697 err = io_queue_file_removal(data, *file_slot);
7701 needs_switch = true;
7710 * Don't allow io_uring instances to be registered. If
7711 * UNIX isn't enabled, then this causes a reference
7712 * cycle and this instance can never get freed. If UNIX
7713 * is enabled we'll handle it just fine, but there's
7714 * still no point in allowing a ring fd as it doesn't
7715 * support regular read/write anyway.
7717 if (file->f_op == &io_uring_fops) {
7723 err = io_sqe_file_register(ctx, file, i);
7733 percpu_ref_kill(&data->node->refs);
7734 io_sqe_rsrc_set_node(ctx, data, ref_node);
7736 destroy_fixed_rsrc_ref_node(ref_node);
7738 return done ? done : err;
7741 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7744 struct io_uring_rsrc_update up;
7746 if (!ctx->file_data)
7750 if (copy_from_user(&up, arg, sizeof(up)))
7755 return __io_sqe_files_update(ctx, &up, nr_args);
7758 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7760 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7762 req = io_put_req_find_next(req);
7763 return req ? &req->work : NULL;
7766 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7768 struct io_wq_hash *hash;
7769 struct io_wq_data data;
7770 unsigned int concurrency;
7772 hash = ctx->hash_map;
7774 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7776 return ERR_PTR(-ENOMEM);
7777 refcount_set(&hash->refs, 1);
7778 init_waitqueue_head(&hash->wait);
7779 ctx->hash_map = hash;
7783 data.free_work = io_free_work;
7784 data.do_work = io_wq_submit_work;
7786 /* Do QD, or 4 * CPUS, whatever is smallest */
7787 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7789 return io_wq_create(concurrency, &data);
7792 static int io_uring_alloc_task_context(struct task_struct *task,
7793 struct io_ring_ctx *ctx)
7795 struct io_uring_task *tctx;
7798 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7799 if (unlikely(!tctx))
7802 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7803 if (unlikely(ret)) {
7808 tctx->io_wq = io_init_wq_offload(ctx);
7809 if (IS_ERR(tctx->io_wq)) {
7810 ret = PTR_ERR(tctx->io_wq);
7811 percpu_counter_destroy(&tctx->inflight);
7817 init_waitqueue_head(&tctx->wait);
7819 atomic_set(&tctx->in_idle, 0);
7820 tctx->sqpoll = false;
7821 task->io_uring = tctx;
7822 spin_lock_init(&tctx->task_lock);
7823 INIT_WQ_LIST(&tctx->task_list);
7824 tctx->task_state = 0;
7825 init_task_work(&tctx->task_work, tctx_task_work);
7829 void __io_uring_free(struct task_struct *tsk)
7831 struct io_uring_task *tctx = tsk->io_uring;
7833 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7834 WARN_ON_ONCE(tctx->io_wq);
7836 percpu_counter_destroy(&tctx->inflight);
7838 tsk->io_uring = NULL;
7841 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7843 struct task_struct *tsk;
7846 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7847 reinit_completion(&sqd->parked);
7849 sqd->task_pid = current->pid;
7850 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7852 return PTR_ERR(tsk);
7853 ret = io_uring_alloc_task_context(tsk, ctx);
7855 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7857 wake_up_new_task(tsk);
7861 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7862 struct io_uring_params *p)
7866 /* Retain compatibility with failing for an invalid attach attempt */
7867 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7868 IORING_SETUP_ATTACH_WQ) {
7871 f = fdget(p->wq_fd);
7874 if (f.file->f_op != &io_uring_fops) {
7880 if (ctx->flags & IORING_SETUP_SQPOLL) {
7881 struct task_struct *tsk;
7882 struct io_sq_data *sqd;
7885 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7888 sqd = io_get_sq_data(p);
7895 io_sq_thread_park(sqd);
7896 mutex_lock(&sqd->ctx_lock);
7897 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7898 mutex_unlock(&sqd->ctx_lock);
7899 io_sq_thread_unpark(sqd);
7901 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7902 if (!ctx->sq_thread_idle)
7903 ctx->sq_thread_idle = HZ;
7908 if (p->flags & IORING_SETUP_SQ_AFF) {
7909 int cpu = p->sq_thread_cpu;
7912 if (cpu >= nr_cpu_ids)
7914 if (!cpu_online(cpu))
7922 sqd->task_pid = current->pid;
7923 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7928 ret = io_uring_alloc_task_context(tsk, ctx);
7930 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7932 wake_up_new_task(tsk);
7935 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7936 /* Can't have SQ_AFF without SQPOLL */
7943 io_sq_thread_finish(ctx);
7947 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7949 struct io_sq_data *sqd = ctx->sq_data;
7951 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7952 if (ctx->flags & IORING_SETUP_SQPOLL)
7953 complete(&sqd->startup);
7956 static inline void __io_unaccount_mem(struct user_struct *user,
7957 unsigned long nr_pages)
7959 atomic_long_sub(nr_pages, &user->locked_vm);
7962 static inline int __io_account_mem(struct user_struct *user,
7963 unsigned long nr_pages)
7965 unsigned long page_limit, cur_pages, new_pages;
7967 /* Don't allow more pages than we can safely lock */
7968 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7971 cur_pages = atomic_long_read(&user->locked_vm);
7972 new_pages = cur_pages + nr_pages;
7973 if (new_pages > page_limit)
7975 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7976 new_pages) != cur_pages);
7981 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7984 __io_unaccount_mem(ctx->user, nr_pages);
7986 if (ctx->mm_account)
7987 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7990 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7995 ret = __io_account_mem(ctx->user, nr_pages);
8000 if (ctx->mm_account)
8001 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8006 static void io_mem_free(void *ptr)
8013 page = virt_to_head_page(ptr);
8014 if (put_page_testzero(page))
8015 free_compound_page(page);
8018 static void *io_mem_alloc(size_t size)
8020 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8021 __GFP_NORETRY | __GFP_ACCOUNT;
8023 return (void *) __get_free_pages(gfp_flags, get_order(size));
8026 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8029 struct io_rings *rings;
8030 size_t off, sq_array_size;
8032 off = struct_size(rings, cqes, cq_entries);
8033 if (off == SIZE_MAX)
8037 off = ALIGN(off, SMP_CACHE_BYTES);
8045 sq_array_size = array_size(sizeof(u32), sq_entries);
8046 if (sq_array_size == SIZE_MAX)
8049 if (check_add_overflow(off, sq_array_size, &off))
8055 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8059 if (!ctx->user_bufs)
8062 for (i = 0; i < ctx->nr_user_bufs; i++) {
8063 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8065 for (j = 0; j < imu->nr_bvecs; j++)
8066 unpin_user_page(imu->bvec[j].bv_page);
8068 if (imu->acct_pages)
8069 io_unaccount_mem(ctx, imu->acct_pages);
8074 kfree(ctx->user_bufs);
8075 ctx->user_bufs = NULL;
8076 ctx->nr_user_bufs = 0;
8080 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8081 void __user *arg, unsigned index)
8083 struct iovec __user *src;
8085 #ifdef CONFIG_COMPAT
8087 struct compat_iovec __user *ciovs;
8088 struct compat_iovec ciov;
8090 ciovs = (struct compat_iovec __user *) arg;
8091 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8094 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8095 dst->iov_len = ciov.iov_len;
8099 src = (struct iovec __user *) arg;
8100 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8106 * Not super efficient, but this is just a registration time. And we do cache
8107 * the last compound head, so generally we'll only do a full search if we don't
8110 * We check if the given compound head page has already been accounted, to
8111 * avoid double accounting it. This allows us to account the full size of the
8112 * page, not just the constituent pages of a huge page.
8114 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8115 int nr_pages, struct page *hpage)
8119 /* check current page array */
8120 for (i = 0; i < nr_pages; i++) {
8121 if (!PageCompound(pages[i]))
8123 if (compound_head(pages[i]) == hpage)
8127 /* check previously registered pages */
8128 for (i = 0; i < ctx->nr_user_bufs; i++) {
8129 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8131 for (j = 0; j < imu->nr_bvecs; j++) {
8132 if (!PageCompound(imu->bvec[j].bv_page))
8134 if (compound_head(imu->bvec[j].bv_page) == hpage)
8142 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8143 int nr_pages, struct io_mapped_ubuf *imu,
8144 struct page **last_hpage)
8148 for (i = 0; i < nr_pages; i++) {
8149 if (!PageCompound(pages[i])) {
8154 hpage = compound_head(pages[i]);
8155 if (hpage == *last_hpage)
8157 *last_hpage = hpage;
8158 if (headpage_already_acct(ctx, pages, i, hpage))
8160 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8164 if (!imu->acct_pages)
8167 ret = io_account_mem(ctx, imu->acct_pages);
8169 imu->acct_pages = 0;
8173 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8174 struct io_mapped_ubuf *imu,
8175 struct page **last_hpage)
8177 struct vm_area_struct **vmas = NULL;
8178 struct page **pages = NULL;
8179 unsigned long off, start, end, ubuf;
8181 int ret, pret, nr_pages, i;
8183 ubuf = (unsigned long) iov->iov_base;
8184 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8185 start = ubuf >> PAGE_SHIFT;
8186 nr_pages = end - start;
8190 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8194 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8199 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8205 mmap_read_lock(current->mm);
8206 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8208 if (pret == nr_pages) {
8209 /* don't support file backed memory */
8210 for (i = 0; i < nr_pages; i++) {
8211 struct vm_area_struct *vma = vmas[i];
8214 !is_file_hugepages(vma->vm_file)) {
8220 ret = pret < 0 ? pret : -EFAULT;
8222 mmap_read_unlock(current->mm);
8225 * if we did partial map, or found file backed vmas,
8226 * release any pages we did get
8229 unpin_user_pages(pages, pret);
8234 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8236 unpin_user_pages(pages, pret);
8241 off = ubuf & ~PAGE_MASK;
8242 size = iov->iov_len;
8243 for (i = 0; i < nr_pages; i++) {
8246 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8247 imu->bvec[i].bv_page = pages[i];
8248 imu->bvec[i].bv_len = vec_len;
8249 imu->bvec[i].bv_offset = off;
8253 /* store original address for later verification */
8255 imu->len = iov->iov_len;
8256 imu->nr_bvecs = nr_pages;
8264 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8268 if (!nr_args || nr_args > UIO_MAXIOV)
8271 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8273 if (!ctx->user_bufs)
8279 static int io_buffer_validate(struct iovec *iov)
8282 * Don't impose further limits on the size and buffer
8283 * constraints here, we'll -EINVAL later when IO is
8284 * submitted if they are wrong.
8286 if (!iov->iov_base || !iov->iov_len)
8289 /* arbitrary limit, but we need something */
8290 if (iov->iov_len > SZ_1G)
8296 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8297 unsigned int nr_args)
8301 struct page *last_hpage = NULL;
8303 ret = io_buffers_map_alloc(ctx, nr_args);
8307 for (i = 0; i < nr_args; i++) {
8308 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8310 ret = io_copy_iov(ctx, &iov, arg, i);
8314 ret = io_buffer_validate(&iov);
8318 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8322 ctx->nr_user_bufs++;
8326 io_sqe_buffers_unregister(ctx);
8331 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8333 __s32 __user *fds = arg;
8339 if (copy_from_user(&fd, fds, sizeof(*fds)))
8342 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8343 if (IS_ERR(ctx->cq_ev_fd)) {
8344 int ret = PTR_ERR(ctx->cq_ev_fd);
8345 ctx->cq_ev_fd = NULL;
8352 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8354 if (ctx->cq_ev_fd) {
8355 eventfd_ctx_put(ctx->cq_ev_fd);
8356 ctx->cq_ev_fd = NULL;
8363 static int __io_destroy_buffers(int id, void *p, void *data)
8365 struct io_ring_ctx *ctx = data;
8366 struct io_buffer *buf = p;
8368 __io_remove_buffers(ctx, buf, id, -1U);
8372 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8374 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8375 idr_destroy(&ctx->io_buffer_idr);
8378 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8380 struct io_kiocb *req, *nxt;
8382 list_for_each_entry_safe(req, nxt, list, compl.list) {
8383 if (tsk && req->task != tsk)
8385 list_del(&req->compl.list);
8386 kmem_cache_free(req_cachep, req);
8390 static void io_req_caches_free(struct io_ring_ctx *ctx)
8392 struct io_submit_state *submit_state = &ctx->submit_state;
8393 struct io_comp_state *cs = &ctx->submit_state.comp;
8395 mutex_lock(&ctx->uring_lock);
8397 if (submit_state->free_reqs) {
8398 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8399 submit_state->reqs);
8400 submit_state->free_reqs = 0;
8403 spin_lock_irq(&ctx->completion_lock);
8404 list_splice_init(&cs->locked_free_list, &cs->free_list);
8405 cs->locked_free_nr = 0;
8406 spin_unlock_irq(&ctx->completion_lock);
8408 io_req_cache_free(&cs->free_list, NULL);
8410 mutex_unlock(&ctx->uring_lock);
8413 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8416 * Some may use context even when all refs and requests have been put,
8417 * and they are free to do so while still holding uring_lock, see
8418 * __io_req_task_submit(). Wait for them to finish.
8420 mutex_lock(&ctx->uring_lock);
8421 mutex_unlock(&ctx->uring_lock);
8423 io_sq_thread_finish(ctx);
8424 io_sqe_buffers_unregister(ctx);
8426 if (ctx->mm_account) {
8427 mmdrop(ctx->mm_account);
8428 ctx->mm_account = NULL;
8431 mutex_lock(&ctx->uring_lock);
8432 io_sqe_files_unregister(ctx);
8433 mutex_unlock(&ctx->uring_lock);
8434 io_eventfd_unregister(ctx);
8435 io_destroy_buffers(ctx);
8436 idr_destroy(&ctx->personality_idr);
8438 #if defined(CONFIG_UNIX)
8439 if (ctx->ring_sock) {
8440 ctx->ring_sock->file = NULL; /* so that iput() is called */
8441 sock_release(ctx->ring_sock);
8445 io_mem_free(ctx->rings);
8446 io_mem_free(ctx->sq_sqes);
8448 percpu_ref_exit(&ctx->refs);
8449 free_uid(ctx->user);
8450 io_req_caches_free(ctx);
8452 io_wq_put_hash(ctx->hash_map);
8453 kfree(ctx->cancel_hash);
8457 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8459 struct io_ring_ctx *ctx = file->private_data;
8462 poll_wait(file, &ctx->cq_wait, wait);
8464 * synchronizes with barrier from wq_has_sleeper call in
8468 if (!io_sqring_full(ctx))
8469 mask |= EPOLLOUT | EPOLLWRNORM;
8472 * Don't flush cqring overflow list here, just do a simple check.
8473 * Otherwise there could possible be ABBA deadlock:
8476 * lock(&ctx->uring_lock);
8478 * lock(&ctx->uring_lock);
8481 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8482 * pushs them to do the flush.
8484 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8485 mask |= EPOLLIN | EPOLLRDNORM;
8490 static int io_uring_fasync(int fd, struct file *file, int on)
8492 struct io_ring_ctx *ctx = file->private_data;
8494 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8497 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8499 const struct cred *creds;
8501 creds = idr_remove(&ctx->personality_idr, id);
8510 static int io_remove_personalities(int id, void *p, void *data)
8512 struct io_ring_ctx *ctx = data;
8514 io_unregister_personality(ctx, id);
8518 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8520 struct callback_head *work, *next;
8521 bool executed = false;
8524 work = xchg(&ctx->exit_task_work, NULL);
8540 struct io_tctx_exit {
8541 struct callback_head task_work;
8542 struct completion completion;
8543 unsigned long index;
8546 static void io_tctx_exit_cb(struct callback_head *cb)
8548 struct io_uring_task *tctx = current->io_uring;
8549 struct io_tctx_exit *work;
8551 work = container_of(cb, struct io_tctx_exit, task_work);
8553 * When @in_idle, we're in cancellation and it's racy to remove the
8554 * node. It'll be removed by the end of cancellation, just ignore it.
8556 if (!atomic_read(&tctx->in_idle))
8557 io_uring_del_task_file(work->index);
8558 complete(&work->completion);
8561 static void io_ring_exit_work(struct work_struct *work)
8563 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8564 struct io_tctx_exit exit;
8565 struct io_tctx_node *node;
8569 * If we're doing polled IO and end up having requests being
8570 * submitted async (out-of-line), then completions can come in while
8571 * we're waiting for refs to drop. We need to reap these manually,
8572 * as nobody else will be looking for them.
8575 io_uring_try_cancel_requests(ctx, NULL, NULL);
8576 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8578 mutex_lock(&ctx->uring_lock);
8579 while (!list_empty(&ctx->tctx_list)) {
8580 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8582 exit.index = (unsigned long)node->file;
8583 init_completion(&exit.completion);
8584 init_task_work(&exit.task_work, io_tctx_exit_cb);
8585 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8586 if (WARN_ON_ONCE(ret))
8588 wake_up_process(node->task);
8590 mutex_unlock(&ctx->uring_lock);
8591 wait_for_completion(&exit.completion);
8593 mutex_lock(&ctx->uring_lock);
8595 mutex_unlock(&ctx->uring_lock);
8597 io_ring_ctx_free(ctx);
8600 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8602 mutex_lock(&ctx->uring_lock);
8603 percpu_ref_kill(&ctx->refs);
8604 /* if force is set, the ring is going away. always drop after that */
8605 ctx->cq_overflow_flushed = 1;
8607 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8608 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8609 mutex_unlock(&ctx->uring_lock);
8611 io_kill_timeouts(ctx, NULL, NULL);
8612 io_poll_remove_all(ctx, NULL, NULL);
8614 /* if we failed setting up the ctx, we might not have any rings */
8615 io_iopoll_try_reap_events(ctx);
8617 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8619 * Use system_unbound_wq to avoid spawning tons of event kworkers
8620 * if we're exiting a ton of rings at the same time. It just adds
8621 * noise and overhead, there's no discernable change in runtime
8622 * over using system_wq.
8624 queue_work(system_unbound_wq, &ctx->exit_work);
8627 static int io_uring_release(struct inode *inode, struct file *file)
8629 struct io_ring_ctx *ctx = file->private_data;
8631 file->private_data = NULL;
8632 io_ring_ctx_wait_and_kill(ctx);
8636 struct io_task_cancel {
8637 struct task_struct *task;
8638 struct files_struct *files;
8641 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8643 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8644 struct io_task_cancel *cancel = data;
8647 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8648 unsigned long flags;
8649 struct io_ring_ctx *ctx = req->ctx;
8651 /* protect against races with linked timeouts */
8652 spin_lock_irqsave(&ctx->completion_lock, flags);
8653 ret = io_match_task(req, cancel->task, cancel->files);
8654 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8656 ret = io_match_task(req, cancel->task, cancel->files);
8661 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8662 struct task_struct *task,
8663 struct files_struct *files)
8665 struct io_defer_entry *de = NULL;
8668 spin_lock_irq(&ctx->completion_lock);
8669 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8670 if (io_match_task(de->req, task, files)) {
8671 list_cut_position(&list, &ctx->defer_list, &de->list);
8675 spin_unlock_irq(&ctx->completion_lock);
8677 while (!list_empty(&list)) {
8678 de = list_first_entry(&list, struct io_defer_entry, list);
8679 list_del_init(&de->list);
8680 req_set_fail_links(de->req);
8681 io_put_req(de->req);
8682 io_req_complete(de->req, -ECANCELED);
8687 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8688 struct task_struct *task,
8689 struct files_struct *files)
8691 struct io_task_cancel cancel = { .task = task, .files = files, };
8692 struct task_struct *tctx_task = task ?: current;
8693 struct io_uring_task *tctx = tctx_task->io_uring;
8696 enum io_wq_cancel cret;
8699 if (tctx && tctx->io_wq) {
8700 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8702 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8705 /* SQPOLL thread does its own polling */
8706 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8707 while (!list_empty_careful(&ctx->iopoll_list)) {
8708 io_iopoll_try_reap_events(ctx);
8713 ret |= io_poll_remove_all(ctx, task, files);
8714 ret |= io_kill_timeouts(ctx, task, files);
8715 ret |= io_run_task_work();
8716 ret |= io_run_ctx_fallback(ctx);
8717 io_cqring_overflow_flush(ctx, true, task, files);
8724 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8725 struct task_struct *task,
8726 struct files_struct *files)
8728 struct io_kiocb *req;
8731 spin_lock_irq(&ctx->inflight_lock);
8732 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8733 cnt += io_match_task(req, task, files);
8734 spin_unlock_irq(&ctx->inflight_lock);
8738 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8739 struct task_struct *task,
8740 struct files_struct *files)
8742 while (!list_empty_careful(&ctx->inflight_list)) {
8746 inflight = io_uring_count_inflight(ctx, task, files);
8750 io_uring_try_cancel_requests(ctx, task, files);
8753 io_sq_thread_unpark(ctx->sq_data);
8754 prepare_to_wait(&task->io_uring->wait, &wait,
8755 TASK_UNINTERRUPTIBLE);
8756 if (inflight == io_uring_count_inflight(ctx, task, files))
8758 finish_wait(&task->io_uring->wait, &wait);
8760 io_sq_thread_park(ctx->sq_data);
8765 * We need to iteratively cancel requests, in case a request has dependent
8766 * hard links. These persist even for failure of cancelations, hence keep
8767 * looping until none are found.
8769 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8770 struct files_struct *files)
8772 struct task_struct *task = current;
8774 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8775 /* never started, nothing to cancel */
8776 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8777 io_sq_offload_start(ctx);
8780 io_sq_thread_park(ctx->sq_data);
8781 task = ctx->sq_data->thread;
8783 atomic_inc(&task->io_uring->in_idle);
8786 io_cancel_defer_files(ctx, task, files);
8788 io_uring_cancel_files(ctx, task, files);
8790 io_uring_try_cancel_requests(ctx, task, NULL);
8793 atomic_dec(&task->io_uring->in_idle);
8795 io_sq_thread_unpark(ctx->sq_data);
8799 * Note that this task has used io_uring. We use it for cancelation purposes.
8801 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8803 struct io_uring_task *tctx = current->io_uring;
8804 struct io_tctx_node *node;
8807 if (unlikely(!tctx)) {
8808 ret = io_uring_alloc_task_context(current, ctx);
8811 tctx = current->io_uring;
8813 if (tctx->last != file) {
8814 void *old = xa_load(&tctx->xa, (unsigned long)file);
8817 node = kmalloc(sizeof(*node), GFP_KERNEL);
8822 node->task = current;
8824 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8831 mutex_lock(&ctx->uring_lock);
8832 list_add(&node->ctx_node, &ctx->tctx_list);
8833 mutex_unlock(&ctx->uring_lock);
8839 * This is race safe in that the task itself is doing this, hence it
8840 * cannot be going through the exit/cancel paths at the same time.
8841 * This cannot be modified while exit/cancel is running.
8843 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8844 tctx->sqpoll = true;
8850 * Remove this io_uring_file -> task mapping.
8852 static void io_uring_del_task_file(unsigned long index)
8854 struct io_uring_task *tctx = current->io_uring;
8855 struct io_tctx_node *node;
8859 node = xa_erase(&tctx->xa, index);
8863 WARN_ON_ONCE(current != node->task);
8864 WARN_ON_ONCE(list_empty(&node->ctx_node));
8866 mutex_lock(&node->ctx->uring_lock);
8867 list_del(&node->ctx_node);
8868 mutex_unlock(&node->ctx->uring_lock);
8870 if (tctx->last == node->file)
8875 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8877 struct io_tctx_node *node;
8878 unsigned long index;
8880 xa_for_each(&tctx->xa, index, node)
8881 io_uring_del_task_file(index);
8883 io_wq_put_and_exit(tctx->io_wq);
8888 void __io_uring_files_cancel(struct files_struct *files)
8890 struct io_uring_task *tctx = current->io_uring;
8891 struct io_tctx_node *node;
8892 unsigned long index;
8894 /* make sure overflow events are dropped */
8895 atomic_inc(&tctx->in_idle);
8896 xa_for_each(&tctx->xa, index, node)
8897 io_uring_cancel_task_requests(node->ctx, files);
8898 atomic_dec(&tctx->in_idle);
8901 io_uring_clean_tctx(tctx);
8904 static s64 tctx_inflight(struct io_uring_task *tctx)
8906 return percpu_counter_sum(&tctx->inflight);
8909 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8911 struct io_sq_data *sqd = ctx->sq_data;
8912 struct io_uring_task *tctx;
8918 io_sq_thread_park(sqd);
8919 if (!sqd->thread || !sqd->thread->io_uring) {
8920 io_sq_thread_unpark(sqd);
8923 tctx = ctx->sq_data->thread->io_uring;
8924 atomic_inc(&tctx->in_idle);
8926 /* read completions before cancelations */
8927 inflight = tctx_inflight(tctx);
8930 io_uring_cancel_task_requests(ctx, NULL);
8932 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8934 * If we've seen completions, retry without waiting. This
8935 * avoids a race where a completion comes in before we did
8936 * prepare_to_wait().
8938 if (inflight == tctx_inflight(tctx))
8940 finish_wait(&tctx->wait, &wait);
8942 atomic_dec(&tctx->in_idle);
8943 io_sq_thread_unpark(sqd);
8947 * Find any io_uring fd that this task has registered or done IO on, and cancel
8950 void __io_uring_task_cancel(void)
8952 struct io_uring_task *tctx = current->io_uring;
8956 /* make sure overflow events are dropped */
8957 atomic_inc(&tctx->in_idle);
8960 struct io_tctx_node *node;
8961 unsigned long index;
8963 xa_for_each(&tctx->xa, index, node)
8964 io_uring_cancel_sqpoll(node->ctx);
8968 /* read completions before cancelations */
8969 inflight = tctx_inflight(tctx);
8972 __io_uring_files_cancel(NULL);
8974 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8977 * If we've seen completions, retry without waiting. This
8978 * avoids a race where a completion comes in before we did
8979 * prepare_to_wait().
8981 if (inflight == tctx_inflight(tctx))
8983 finish_wait(&tctx->wait, &wait);
8986 atomic_dec(&tctx->in_idle);
8988 io_uring_clean_tctx(tctx);
8989 /* all current's requests should be gone, we can kill tctx */
8990 __io_uring_free(current);
8993 static void *io_uring_validate_mmap_request(struct file *file,
8994 loff_t pgoff, size_t sz)
8996 struct io_ring_ctx *ctx = file->private_data;
8997 loff_t offset = pgoff << PAGE_SHIFT;
9002 case IORING_OFF_SQ_RING:
9003 case IORING_OFF_CQ_RING:
9006 case IORING_OFF_SQES:
9010 return ERR_PTR(-EINVAL);
9013 page = virt_to_head_page(ptr);
9014 if (sz > page_size(page))
9015 return ERR_PTR(-EINVAL);
9022 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9024 size_t sz = vma->vm_end - vma->vm_start;
9028 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9030 return PTR_ERR(ptr);
9032 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9033 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9036 #else /* !CONFIG_MMU */
9038 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9040 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9043 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9045 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9048 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9049 unsigned long addr, unsigned long len,
9050 unsigned long pgoff, unsigned long flags)
9054 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9056 return PTR_ERR(ptr);
9058 return (unsigned long) ptr;
9061 #endif /* !CONFIG_MMU */
9063 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9069 if (!io_sqring_full(ctx))
9071 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9073 if (!io_sqring_full(ctx))
9076 } while (!signal_pending(current));
9078 finish_wait(&ctx->sqo_sq_wait, &wait);
9082 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9083 struct __kernel_timespec __user **ts,
9084 const sigset_t __user **sig)
9086 struct io_uring_getevents_arg arg;
9089 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9090 * is just a pointer to the sigset_t.
9092 if (!(flags & IORING_ENTER_EXT_ARG)) {
9093 *sig = (const sigset_t __user *) argp;
9099 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9100 * timespec and sigset_t pointers if good.
9102 if (*argsz != sizeof(arg))
9104 if (copy_from_user(&arg, argp, sizeof(arg)))
9106 *sig = u64_to_user_ptr(arg.sigmask);
9107 *argsz = arg.sigmask_sz;
9108 *ts = u64_to_user_ptr(arg.ts);
9112 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9113 u32, min_complete, u32, flags, const void __user *, argp,
9116 struct io_ring_ctx *ctx;
9123 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9124 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9132 if (f.file->f_op != &io_uring_fops)
9136 ctx = f.file->private_data;
9137 if (!percpu_ref_tryget(&ctx->refs))
9141 if (ctx->flags & IORING_SETUP_R_DISABLED)
9145 * For SQ polling, the thread will do all submissions and completions.
9146 * Just return the requested submit count, and wake the thread if
9150 if (ctx->flags & IORING_SETUP_SQPOLL) {
9151 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9153 if (unlikely(ctx->sqo_exec)) {
9154 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9160 if (flags & IORING_ENTER_SQ_WAKEUP)
9161 wake_up(&ctx->sq_data->wait);
9162 if (flags & IORING_ENTER_SQ_WAIT) {
9163 ret = io_sqpoll_wait_sq(ctx);
9167 submitted = to_submit;
9168 } else if (to_submit) {
9169 ret = io_uring_add_task_file(ctx, f.file);
9172 mutex_lock(&ctx->uring_lock);
9173 submitted = io_submit_sqes(ctx, to_submit);
9174 mutex_unlock(&ctx->uring_lock);
9176 if (submitted != to_submit)
9179 if (flags & IORING_ENTER_GETEVENTS) {
9180 const sigset_t __user *sig;
9181 struct __kernel_timespec __user *ts;
9183 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9187 min_complete = min(min_complete, ctx->cq_entries);
9190 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9191 * space applications don't need to do io completion events
9192 * polling again, they can rely on io_sq_thread to do polling
9193 * work, which can reduce cpu usage and uring_lock contention.
9195 if (ctx->flags & IORING_SETUP_IOPOLL &&
9196 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9197 ret = io_iopoll_check(ctx, min_complete);
9199 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9204 percpu_ref_put(&ctx->refs);
9207 return submitted ? submitted : ret;
9210 #ifdef CONFIG_PROC_FS
9211 static int io_uring_show_cred(int id, void *p, void *data)
9213 const struct cred *cred = p;
9214 struct seq_file *m = data;
9215 struct user_namespace *uns = seq_user_ns(m);
9216 struct group_info *gi;
9221 seq_printf(m, "%5d\n", id);
9222 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9223 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9224 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9225 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9226 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9227 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9228 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9229 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9230 seq_puts(m, "\n\tGroups:\t");
9231 gi = cred->group_info;
9232 for (g = 0; g < gi->ngroups; g++) {
9233 seq_put_decimal_ull(m, g ? " " : "",
9234 from_kgid_munged(uns, gi->gid[g]));
9236 seq_puts(m, "\n\tCapEff:\t");
9237 cap = cred->cap_effective;
9238 CAP_FOR_EACH_U32(__capi)
9239 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9244 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9246 struct io_sq_data *sq = NULL;
9251 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9252 * since fdinfo case grabs it in the opposite direction of normal use
9253 * cases. If we fail to get the lock, we just don't iterate any
9254 * structures that could be going away outside the io_uring mutex.
9256 has_lock = mutex_trylock(&ctx->uring_lock);
9258 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9264 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9265 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9266 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9267 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9268 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9271 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9273 seq_printf(m, "%5u: <none>\n", i);
9275 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9276 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9277 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9279 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9280 (unsigned int) buf->len);
9282 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9283 seq_printf(m, "Personalities:\n");
9284 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9286 seq_printf(m, "PollList:\n");
9287 spin_lock_irq(&ctx->completion_lock);
9288 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9289 struct hlist_head *list = &ctx->cancel_hash[i];
9290 struct io_kiocb *req;
9292 hlist_for_each_entry(req, list, hash_node)
9293 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9294 req->task->task_works != NULL);
9296 spin_unlock_irq(&ctx->completion_lock);
9298 mutex_unlock(&ctx->uring_lock);
9301 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9303 struct io_ring_ctx *ctx = f->private_data;
9305 if (percpu_ref_tryget(&ctx->refs)) {
9306 __io_uring_show_fdinfo(ctx, m);
9307 percpu_ref_put(&ctx->refs);
9312 static const struct file_operations io_uring_fops = {
9313 .release = io_uring_release,
9314 .mmap = io_uring_mmap,
9316 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9317 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9319 .poll = io_uring_poll,
9320 .fasync = io_uring_fasync,
9321 #ifdef CONFIG_PROC_FS
9322 .show_fdinfo = io_uring_show_fdinfo,
9326 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9327 struct io_uring_params *p)
9329 struct io_rings *rings;
9330 size_t size, sq_array_offset;
9332 /* make sure these are sane, as we already accounted them */
9333 ctx->sq_entries = p->sq_entries;
9334 ctx->cq_entries = p->cq_entries;
9336 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9337 if (size == SIZE_MAX)
9340 rings = io_mem_alloc(size);
9345 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9346 rings->sq_ring_mask = p->sq_entries - 1;
9347 rings->cq_ring_mask = p->cq_entries - 1;
9348 rings->sq_ring_entries = p->sq_entries;
9349 rings->cq_ring_entries = p->cq_entries;
9350 ctx->sq_mask = rings->sq_ring_mask;
9351 ctx->cq_mask = rings->cq_ring_mask;
9353 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9354 if (size == SIZE_MAX) {
9355 io_mem_free(ctx->rings);
9360 ctx->sq_sqes = io_mem_alloc(size);
9361 if (!ctx->sq_sqes) {
9362 io_mem_free(ctx->rings);
9370 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9374 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9378 ret = io_uring_add_task_file(ctx, file);
9383 fd_install(fd, file);
9388 * Allocate an anonymous fd, this is what constitutes the application
9389 * visible backing of an io_uring instance. The application mmaps this
9390 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9391 * we have to tie this fd to a socket for file garbage collection purposes.
9393 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9396 #if defined(CONFIG_UNIX)
9399 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9402 return ERR_PTR(ret);
9405 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9406 O_RDWR | O_CLOEXEC);
9407 #if defined(CONFIG_UNIX)
9409 sock_release(ctx->ring_sock);
9410 ctx->ring_sock = NULL;
9412 ctx->ring_sock->file = file;
9418 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9419 struct io_uring_params __user *params)
9421 struct io_ring_ctx *ctx;
9427 if (entries > IORING_MAX_ENTRIES) {
9428 if (!(p->flags & IORING_SETUP_CLAMP))
9430 entries = IORING_MAX_ENTRIES;
9434 * Use twice as many entries for the CQ ring. It's possible for the
9435 * application to drive a higher depth than the size of the SQ ring,
9436 * since the sqes are only used at submission time. This allows for
9437 * some flexibility in overcommitting a bit. If the application has
9438 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9439 * of CQ ring entries manually.
9441 p->sq_entries = roundup_pow_of_two(entries);
9442 if (p->flags & IORING_SETUP_CQSIZE) {
9444 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9445 * to a power-of-two, if it isn't already. We do NOT impose
9446 * any cq vs sq ring sizing.
9450 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9451 if (!(p->flags & IORING_SETUP_CLAMP))
9453 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9455 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9456 if (p->cq_entries < p->sq_entries)
9459 p->cq_entries = 2 * p->sq_entries;
9462 ctx = io_ring_ctx_alloc(p);
9465 ctx->compat = in_compat_syscall();
9466 if (!capable(CAP_IPC_LOCK))
9467 ctx->user = get_uid(current_user());
9470 * This is just grabbed for accounting purposes. When a process exits,
9471 * the mm is exited and dropped before the files, hence we need to hang
9472 * on to this mm purely for the purposes of being able to unaccount
9473 * memory (locked/pinned vm). It's not used for anything else.
9475 mmgrab(current->mm);
9476 ctx->mm_account = current->mm;
9478 ret = io_allocate_scq_urings(ctx, p);
9482 ret = io_sq_offload_create(ctx, p);
9486 if (!(p->flags & IORING_SETUP_R_DISABLED))
9487 io_sq_offload_start(ctx);
9489 memset(&p->sq_off, 0, sizeof(p->sq_off));
9490 p->sq_off.head = offsetof(struct io_rings, sq.head);
9491 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9492 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9493 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9494 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9495 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9496 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9498 memset(&p->cq_off, 0, sizeof(p->cq_off));
9499 p->cq_off.head = offsetof(struct io_rings, cq.head);
9500 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9501 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9502 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9503 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9504 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9505 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9507 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9508 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9509 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9510 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9511 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9513 if (copy_to_user(params, p, sizeof(*p))) {
9518 file = io_uring_get_file(ctx);
9520 ret = PTR_ERR(file);
9525 * Install ring fd as the very last thing, so we don't risk someone
9526 * having closed it before we finish setup
9528 ret = io_uring_install_fd(ctx, file);
9530 /* fput will clean it up */
9535 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9538 io_ring_ctx_wait_and_kill(ctx);
9543 * Sets up an aio uring context, and returns the fd. Applications asks for a
9544 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9545 * params structure passed in.
9547 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9549 struct io_uring_params p;
9552 if (copy_from_user(&p, params, sizeof(p)))
9554 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9559 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9560 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9561 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9562 IORING_SETUP_R_DISABLED))
9565 return io_uring_create(entries, &p, params);
9568 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9569 struct io_uring_params __user *, params)
9571 return io_uring_setup(entries, params);
9574 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9576 struct io_uring_probe *p;
9580 size = struct_size(p, ops, nr_args);
9581 if (size == SIZE_MAX)
9583 p = kzalloc(size, GFP_KERNEL);
9588 if (copy_from_user(p, arg, size))
9591 if (memchr_inv(p, 0, size))
9594 p->last_op = IORING_OP_LAST - 1;
9595 if (nr_args > IORING_OP_LAST)
9596 nr_args = IORING_OP_LAST;
9598 for (i = 0; i < nr_args; i++) {
9600 if (!io_op_defs[i].not_supported)
9601 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9606 if (copy_to_user(arg, p, size))
9613 static int io_register_personality(struct io_ring_ctx *ctx)
9615 const struct cred *creds;
9618 creds = get_current_cred();
9620 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9621 USHRT_MAX, GFP_KERNEL);
9627 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9628 unsigned int nr_args)
9630 struct io_uring_restriction *res;
9634 /* Restrictions allowed only if rings started disabled */
9635 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9638 /* We allow only a single restrictions registration */
9639 if (ctx->restrictions.registered)
9642 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9645 size = array_size(nr_args, sizeof(*res));
9646 if (size == SIZE_MAX)
9649 res = memdup_user(arg, size);
9651 return PTR_ERR(res);
9655 for (i = 0; i < nr_args; i++) {
9656 switch (res[i].opcode) {
9657 case IORING_RESTRICTION_REGISTER_OP:
9658 if (res[i].register_op >= IORING_REGISTER_LAST) {
9663 __set_bit(res[i].register_op,
9664 ctx->restrictions.register_op);
9666 case IORING_RESTRICTION_SQE_OP:
9667 if (res[i].sqe_op >= IORING_OP_LAST) {
9672 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9674 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9675 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9677 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9678 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9687 /* Reset all restrictions if an error happened */
9689 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9691 ctx->restrictions.registered = true;
9697 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9699 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9702 if (ctx->restrictions.registered)
9703 ctx->restricted = 1;
9705 io_sq_offload_start(ctx);
9709 static bool io_register_op_must_quiesce(int op)
9712 case IORING_UNREGISTER_FILES:
9713 case IORING_REGISTER_FILES_UPDATE:
9714 case IORING_REGISTER_PROBE:
9715 case IORING_REGISTER_PERSONALITY:
9716 case IORING_UNREGISTER_PERSONALITY:
9723 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9724 void __user *arg, unsigned nr_args)
9725 __releases(ctx->uring_lock)
9726 __acquires(ctx->uring_lock)
9731 * We're inside the ring mutex, if the ref is already dying, then
9732 * someone else killed the ctx or is already going through
9733 * io_uring_register().
9735 if (percpu_ref_is_dying(&ctx->refs))
9738 if (io_register_op_must_quiesce(opcode)) {
9739 percpu_ref_kill(&ctx->refs);
9742 * Drop uring mutex before waiting for references to exit. If
9743 * another thread is currently inside io_uring_enter() it might
9744 * need to grab the uring_lock to make progress. If we hold it
9745 * here across the drain wait, then we can deadlock. It's safe
9746 * to drop the mutex here, since no new references will come in
9747 * after we've killed the percpu ref.
9749 mutex_unlock(&ctx->uring_lock);
9751 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9754 ret = io_run_task_work_sig();
9759 mutex_lock(&ctx->uring_lock);
9762 percpu_ref_resurrect(&ctx->refs);
9767 if (ctx->restricted) {
9768 if (opcode >= IORING_REGISTER_LAST) {
9773 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9780 case IORING_REGISTER_BUFFERS:
9781 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9783 case IORING_UNREGISTER_BUFFERS:
9787 ret = io_sqe_buffers_unregister(ctx);
9789 case IORING_REGISTER_FILES:
9790 ret = io_sqe_files_register(ctx, arg, nr_args);
9792 case IORING_UNREGISTER_FILES:
9796 ret = io_sqe_files_unregister(ctx);
9798 case IORING_REGISTER_FILES_UPDATE:
9799 ret = io_sqe_files_update(ctx, arg, nr_args);
9801 case IORING_REGISTER_EVENTFD:
9802 case IORING_REGISTER_EVENTFD_ASYNC:
9806 ret = io_eventfd_register(ctx, arg);
9809 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9810 ctx->eventfd_async = 1;
9812 ctx->eventfd_async = 0;
9814 case IORING_UNREGISTER_EVENTFD:
9818 ret = io_eventfd_unregister(ctx);
9820 case IORING_REGISTER_PROBE:
9822 if (!arg || nr_args > 256)
9824 ret = io_probe(ctx, arg, nr_args);
9826 case IORING_REGISTER_PERSONALITY:
9830 ret = io_register_personality(ctx);
9832 case IORING_UNREGISTER_PERSONALITY:
9836 ret = io_unregister_personality(ctx, nr_args);
9838 case IORING_REGISTER_ENABLE_RINGS:
9842 ret = io_register_enable_rings(ctx);
9844 case IORING_REGISTER_RESTRICTIONS:
9845 ret = io_register_restrictions(ctx, arg, nr_args);
9853 if (io_register_op_must_quiesce(opcode)) {
9854 /* bring the ctx back to life */
9855 percpu_ref_reinit(&ctx->refs);
9857 reinit_completion(&ctx->ref_comp);
9862 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9863 void __user *, arg, unsigned int, nr_args)
9865 struct io_ring_ctx *ctx;
9874 if (f.file->f_op != &io_uring_fops)
9877 ctx = f.file->private_data;
9881 mutex_lock(&ctx->uring_lock);
9882 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9883 mutex_unlock(&ctx->uring_lock);
9884 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9885 ctx->cq_ev_fd != NULL, ret);
9891 static int __init io_uring_init(void)
9893 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9894 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9895 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9898 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9899 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9900 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9901 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9902 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9903 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9904 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9905 BUILD_BUG_SQE_ELEM(8, __u64, off);
9906 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9907 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9908 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9909 BUILD_BUG_SQE_ELEM(24, __u32, len);
9910 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9911 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9912 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9913 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9914 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9915 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9916 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9917 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9918 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9919 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9920 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9921 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9922 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9923 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9924 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9925 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9926 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9927 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9928 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9930 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9931 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9932 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9936 __initcall(io_uring_init);