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;
341 * Ring buffer of indices into array of io_uring_sqe, which is
342 * mmapped by the application using the IORING_OFF_SQES offset.
344 * This indirection could e.g. be used to assign fixed
345 * io_uring_sqe entries to operations and only submit them to
346 * the queue when needed.
348 * The kernel modifies neither the indices array nor the entries
352 unsigned cached_sq_head;
355 unsigned sq_thread_idle;
356 unsigned cached_sq_dropped;
357 unsigned cached_cq_overflow;
358 unsigned long sq_check_overflow;
360 /* hashed buffered write serialization */
361 struct io_wq_hash *hash_map;
363 struct list_head defer_list;
364 struct list_head timeout_list;
365 struct list_head cq_overflow_list;
367 struct io_uring_sqe *sq_sqes;
368 } ____cacheline_aligned_in_smp;
371 struct mutex uring_lock;
372 wait_queue_head_t wait;
373 } ____cacheline_aligned_in_smp;
375 struct io_submit_state submit_state;
377 struct io_rings *rings;
379 /* Only used for accounting purposes */
380 struct mm_struct *mm_account;
382 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
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;
812 struct io_ring_ctx *ctx;
815 struct io_defer_entry {
816 struct list_head list;
817 struct io_kiocb *req;
822 /* needs req->file assigned */
823 unsigned needs_file : 1;
824 /* hash wq insertion if file is a regular file */
825 unsigned hash_reg_file : 1;
826 /* unbound wq insertion if file is a non-regular file */
827 unsigned unbound_nonreg_file : 1;
828 /* opcode is not supported by this kernel */
829 unsigned not_supported : 1;
830 /* set if opcode supports polled "wait" */
832 unsigned pollout : 1;
833 /* op supports buffer selection */
834 unsigned buffer_select : 1;
835 /* must always have async data allocated */
836 unsigned needs_async_data : 1;
837 /* should block plug */
839 /* size of async data needed, if any */
840 unsigned short async_size;
843 static const struct io_op_def io_op_defs[] = {
844 [IORING_OP_NOP] = {},
845 [IORING_OP_READV] = {
847 .unbound_nonreg_file = 1,
850 .needs_async_data = 1,
852 .async_size = sizeof(struct io_async_rw),
854 [IORING_OP_WRITEV] = {
857 .unbound_nonreg_file = 1,
859 .needs_async_data = 1,
861 .async_size = sizeof(struct io_async_rw),
863 [IORING_OP_FSYNC] = {
866 [IORING_OP_READ_FIXED] = {
868 .unbound_nonreg_file = 1,
871 .async_size = sizeof(struct io_async_rw),
873 [IORING_OP_WRITE_FIXED] = {
876 .unbound_nonreg_file = 1,
879 .async_size = sizeof(struct io_async_rw),
881 [IORING_OP_POLL_ADD] = {
883 .unbound_nonreg_file = 1,
885 [IORING_OP_POLL_REMOVE] = {},
886 [IORING_OP_SYNC_FILE_RANGE] = {
889 [IORING_OP_SENDMSG] = {
891 .unbound_nonreg_file = 1,
893 .needs_async_data = 1,
894 .async_size = sizeof(struct io_async_msghdr),
896 [IORING_OP_RECVMSG] = {
898 .unbound_nonreg_file = 1,
901 .needs_async_data = 1,
902 .async_size = sizeof(struct io_async_msghdr),
904 [IORING_OP_TIMEOUT] = {
905 .needs_async_data = 1,
906 .async_size = sizeof(struct io_timeout_data),
908 [IORING_OP_TIMEOUT_REMOVE] = {
909 /* used by timeout updates' prep() */
911 [IORING_OP_ACCEPT] = {
913 .unbound_nonreg_file = 1,
916 [IORING_OP_ASYNC_CANCEL] = {},
917 [IORING_OP_LINK_TIMEOUT] = {
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_timeout_data),
921 [IORING_OP_CONNECT] = {
923 .unbound_nonreg_file = 1,
925 .needs_async_data = 1,
926 .async_size = sizeof(struct io_async_connect),
928 [IORING_OP_FALLOCATE] = {
931 [IORING_OP_OPENAT] = {},
932 [IORING_OP_CLOSE] = {},
933 [IORING_OP_FILES_UPDATE] = {},
934 [IORING_OP_STATX] = {},
937 .unbound_nonreg_file = 1,
941 .async_size = sizeof(struct io_async_rw),
943 [IORING_OP_WRITE] = {
945 .unbound_nonreg_file = 1,
948 .async_size = sizeof(struct io_async_rw),
950 [IORING_OP_FADVISE] = {
953 [IORING_OP_MADVISE] = {},
956 .unbound_nonreg_file = 1,
961 .unbound_nonreg_file = 1,
965 [IORING_OP_OPENAT2] = {
967 [IORING_OP_EPOLL_CTL] = {
968 .unbound_nonreg_file = 1,
970 [IORING_OP_SPLICE] = {
973 .unbound_nonreg_file = 1,
975 [IORING_OP_PROVIDE_BUFFERS] = {},
976 [IORING_OP_REMOVE_BUFFERS] = {},
980 .unbound_nonreg_file = 1,
982 [IORING_OP_SHUTDOWN] = {
985 [IORING_OP_RENAMEAT] = {},
986 [IORING_OP_UNLINKAT] = {},
989 static void io_uring_del_task_file(unsigned long index);
990 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
991 struct task_struct *task,
992 struct files_struct *files);
993 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
994 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
995 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
996 struct io_ring_ctx *ctx);
997 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
999 static bool io_rw_reissue(struct io_kiocb *req);
1000 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1001 static void io_put_req(struct io_kiocb *req);
1002 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1003 static void io_double_put_req(struct io_kiocb *req);
1004 static void io_dismantle_req(struct io_kiocb *req);
1005 static void io_put_task(struct task_struct *task, int nr);
1006 static void io_queue_next(struct io_kiocb *req);
1007 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1008 static void __io_queue_linked_timeout(struct io_kiocb *req);
1009 static void io_queue_linked_timeout(struct io_kiocb *req);
1010 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1011 struct io_uring_rsrc_update *ip,
1013 static void __io_clean_op(struct io_kiocb *req);
1014 static struct file *io_file_get(struct io_submit_state *state,
1015 struct io_kiocb *req, int fd, bool fixed);
1016 static void __io_queue_sqe(struct io_kiocb *req);
1017 static void io_rsrc_put_work(struct work_struct *work);
1019 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1020 struct iov_iter *iter, bool needs_lock);
1021 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1022 const struct iovec *fast_iov,
1023 struct iov_iter *iter, bool force);
1024 static void io_req_task_queue(struct io_kiocb *req);
1025 static void io_submit_flush_completions(struct io_comp_state *cs,
1026 struct io_ring_ctx *ctx);
1028 static struct kmem_cache *req_cachep;
1030 static const struct file_operations io_uring_fops;
1032 struct sock *io_uring_get_socket(struct file *file)
1034 #if defined(CONFIG_UNIX)
1035 if (file->f_op == &io_uring_fops) {
1036 struct io_ring_ctx *ctx = file->private_data;
1038 return ctx->ring_sock->sk;
1043 EXPORT_SYMBOL(io_uring_get_socket);
1045 #define io_for_each_link(pos, head) \
1046 for (pos = (head); pos; pos = pos->link)
1048 static inline void io_clean_op(struct io_kiocb *req)
1050 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1054 static inline void io_set_resource_node(struct io_kiocb *req)
1056 struct io_ring_ctx *ctx = req->ctx;
1058 if (!req->fixed_rsrc_refs) {
1059 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1060 percpu_ref_get(req->fixed_rsrc_refs);
1064 static bool io_match_task(struct io_kiocb *head,
1065 struct task_struct *task,
1066 struct files_struct *files)
1068 struct io_kiocb *req;
1070 if (task && head->task != task) {
1071 /* in terms of cancelation, always match if req task is dead */
1072 if (head->task->flags & PF_EXITING)
1079 io_for_each_link(req, head) {
1080 if (req->flags & REQ_F_INFLIGHT)
1082 if (req->task->files == files)
1088 static inline void req_set_fail_links(struct io_kiocb *req)
1090 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1091 req->flags |= REQ_F_FAIL_LINK;
1094 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1096 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1098 complete(&ctx->ref_comp);
1101 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1103 return !req->timeout.off;
1106 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1108 struct io_ring_ctx *ctx;
1111 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1116 * Use 5 bits less than the max cq entries, that should give us around
1117 * 32 entries per hash list if totally full and uniformly spread.
1119 hash_bits = ilog2(p->cq_entries);
1123 ctx->cancel_hash_bits = hash_bits;
1124 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1126 if (!ctx->cancel_hash)
1128 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1130 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1131 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1134 ctx->flags = p->flags;
1135 init_waitqueue_head(&ctx->sqo_sq_wait);
1136 INIT_LIST_HEAD(&ctx->sqd_list);
1137 init_waitqueue_head(&ctx->cq_wait);
1138 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1139 init_completion(&ctx->ref_comp);
1140 init_completion(&ctx->sq_thread_comp);
1141 idr_init(&ctx->io_buffer_idr);
1142 idr_init(&ctx->personality_idr);
1143 mutex_init(&ctx->uring_lock);
1144 init_waitqueue_head(&ctx->wait);
1145 spin_lock_init(&ctx->completion_lock);
1146 INIT_LIST_HEAD(&ctx->iopoll_list);
1147 INIT_LIST_HEAD(&ctx->defer_list);
1148 INIT_LIST_HEAD(&ctx->timeout_list);
1149 spin_lock_init(&ctx->inflight_lock);
1150 INIT_LIST_HEAD(&ctx->inflight_list);
1151 spin_lock_init(&ctx->rsrc_ref_lock);
1152 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1153 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1154 init_llist_head(&ctx->rsrc_put_llist);
1155 INIT_LIST_HEAD(&ctx->tctx_list);
1156 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1157 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1160 kfree(ctx->cancel_hash);
1165 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1167 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1168 struct io_ring_ctx *ctx = req->ctx;
1170 return seq != ctx->cached_cq_tail
1171 + READ_ONCE(ctx->cached_cq_overflow);
1177 static void io_req_track_inflight(struct io_kiocb *req)
1179 struct io_ring_ctx *ctx = req->ctx;
1181 if (!(req->flags & REQ_F_INFLIGHT)) {
1182 req->flags |= REQ_F_INFLIGHT;
1184 spin_lock_irq(&ctx->inflight_lock);
1185 list_add(&req->inflight_entry, &ctx->inflight_list);
1186 spin_unlock_irq(&ctx->inflight_lock);
1190 static void io_prep_async_work(struct io_kiocb *req)
1192 const struct io_op_def *def = &io_op_defs[req->opcode];
1193 struct io_ring_ctx *ctx = req->ctx;
1195 if (!req->work.creds)
1196 req->work.creds = get_current_cred();
1198 if (req->flags & REQ_F_FORCE_ASYNC)
1199 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1201 if (req->flags & REQ_F_ISREG) {
1202 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1203 io_wq_hash_work(&req->work, file_inode(req->file));
1205 if (def->unbound_nonreg_file)
1206 req->work.flags |= IO_WQ_WORK_UNBOUND;
1210 static void io_prep_async_link(struct io_kiocb *req)
1212 struct io_kiocb *cur;
1214 io_for_each_link(cur, req)
1215 io_prep_async_work(cur);
1218 static void io_queue_async_work(struct io_kiocb *req)
1220 struct io_ring_ctx *ctx = req->ctx;
1221 struct io_kiocb *link = io_prep_linked_timeout(req);
1222 struct io_uring_task *tctx = req->task->io_uring;
1225 BUG_ON(!tctx->io_wq);
1227 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1228 &req->work, req->flags);
1229 /* init ->work of the whole link before punting */
1230 io_prep_async_link(req);
1231 io_wq_enqueue(tctx->io_wq, &req->work);
1233 io_queue_linked_timeout(link);
1236 static void io_kill_timeout(struct io_kiocb *req)
1238 struct io_timeout_data *io = req->async_data;
1241 ret = hrtimer_try_to_cancel(&io->timer);
1243 atomic_set(&req->ctx->cq_timeouts,
1244 atomic_read(&req->ctx->cq_timeouts) + 1);
1245 list_del_init(&req->timeout.list);
1246 io_cqring_fill_event(req, 0);
1247 io_put_req_deferred(req, 1);
1252 * Returns true if we found and killed one or more timeouts
1254 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1255 struct files_struct *files)
1257 struct io_kiocb *req, *tmp;
1260 spin_lock_irq(&ctx->completion_lock);
1261 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1262 if (io_match_task(req, tsk, files)) {
1263 io_kill_timeout(req);
1267 spin_unlock_irq(&ctx->completion_lock);
1268 return canceled != 0;
1271 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1274 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1275 struct io_defer_entry, list);
1277 if (req_need_defer(de->req, de->seq))
1279 list_del_init(&de->list);
1280 io_req_task_queue(de->req);
1282 } while (!list_empty(&ctx->defer_list));
1285 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1289 if (list_empty(&ctx->timeout_list))
1292 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1295 u32 events_needed, events_got;
1296 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1297 struct io_kiocb, timeout.list);
1299 if (io_is_timeout_noseq(req))
1303 * Since seq can easily wrap around over time, subtract
1304 * the last seq at which timeouts were flushed before comparing.
1305 * Assuming not more than 2^31-1 events have happened since,
1306 * these subtractions won't have wrapped, so we can check if
1307 * target is in [last_seq, current_seq] by comparing the two.
1309 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1310 events_got = seq - ctx->cq_last_tm_flush;
1311 if (events_got < events_needed)
1314 list_del_init(&req->timeout.list);
1315 io_kill_timeout(req);
1316 } while (!list_empty(&ctx->timeout_list));
1318 ctx->cq_last_tm_flush = seq;
1321 static void io_commit_cqring(struct io_ring_ctx *ctx)
1323 io_flush_timeouts(ctx);
1325 /* order cqe stores with ring update */
1326 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1328 if (unlikely(!list_empty(&ctx->defer_list)))
1329 __io_queue_deferred(ctx);
1332 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1334 struct io_rings *r = ctx->rings;
1336 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1339 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1341 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1344 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1346 struct io_rings *rings = ctx->rings;
1350 * writes to the cq entry need to come after reading head; the
1351 * control dependency is enough as we're using WRITE_ONCE to
1354 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1357 tail = ctx->cached_cq_tail++;
1358 return &rings->cqes[tail & ctx->cq_mask];
1361 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1365 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1367 if (!ctx->eventfd_async)
1369 return io_wq_current_is_worker();
1372 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1374 /* see waitqueue_active() comment */
1377 if (waitqueue_active(&ctx->wait))
1378 wake_up(&ctx->wait);
1379 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1380 wake_up(&ctx->sq_data->wait);
1381 if (io_should_trigger_evfd(ctx))
1382 eventfd_signal(ctx->cq_ev_fd, 1);
1383 if (waitqueue_active(&ctx->cq_wait)) {
1384 wake_up_interruptible(&ctx->cq_wait);
1385 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1389 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1391 /* see waitqueue_active() comment */
1394 if (ctx->flags & IORING_SETUP_SQPOLL) {
1395 if (waitqueue_active(&ctx->wait))
1396 wake_up(&ctx->wait);
1398 if (io_should_trigger_evfd(ctx))
1399 eventfd_signal(ctx->cq_ev_fd, 1);
1400 if (waitqueue_active(&ctx->cq_wait)) {
1401 wake_up_interruptible(&ctx->cq_wait);
1402 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1406 /* Returns true if there are no backlogged entries after the flush */
1407 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1408 struct task_struct *tsk,
1409 struct files_struct *files)
1411 struct io_rings *rings = ctx->rings;
1412 struct io_kiocb *req, *tmp;
1413 struct io_uring_cqe *cqe;
1414 unsigned long flags;
1415 bool all_flushed, posted;
1418 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1422 spin_lock_irqsave(&ctx->completion_lock, flags);
1423 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1424 if (!io_match_task(req, tsk, files))
1427 cqe = io_get_cqring(ctx);
1431 list_move(&req->compl.list, &list);
1433 WRITE_ONCE(cqe->user_data, req->user_data);
1434 WRITE_ONCE(cqe->res, req->result);
1435 WRITE_ONCE(cqe->flags, req->compl.cflags);
1437 ctx->cached_cq_overflow++;
1438 WRITE_ONCE(ctx->rings->cq_overflow,
1439 ctx->cached_cq_overflow);
1444 all_flushed = list_empty(&ctx->cq_overflow_list);
1446 clear_bit(0, &ctx->sq_check_overflow);
1447 clear_bit(0, &ctx->cq_check_overflow);
1448 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1452 io_commit_cqring(ctx);
1453 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1455 io_cqring_ev_posted(ctx);
1457 while (!list_empty(&list)) {
1458 req = list_first_entry(&list, struct io_kiocb, compl.list);
1459 list_del(&req->compl.list);
1466 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1467 struct task_struct *tsk,
1468 struct files_struct *files)
1472 if (test_bit(0, &ctx->cq_check_overflow)) {
1473 /* iopoll syncs against uring_lock, not completion_lock */
1474 if (ctx->flags & IORING_SETUP_IOPOLL)
1475 mutex_lock(&ctx->uring_lock);
1476 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1477 if (ctx->flags & IORING_SETUP_IOPOLL)
1478 mutex_unlock(&ctx->uring_lock);
1484 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1486 struct io_ring_ctx *ctx = req->ctx;
1487 struct io_uring_cqe *cqe;
1489 trace_io_uring_complete(ctx, req->user_data, res);
1492 * If we can't get a cq entry, userspace overflowed the
1493 * submission (by quite a lot). Increment the overflow count in
1496 cqe = io_get_cqring(ctx);
1498 WRITE_ONCE(cqe->user_data, req->user_data);
1499 WRITE_ONCE(cqe->res, res);
1500 WRITE_ONCE(cqe->flags, cflags);
1501 } else if (ctx->cq_overflow_flushed ||
1502 atomic_read(&req->task->io_uring->in_idle)) {
1504 * If we're in ring overflow flush mode, or in task cancel mode,
1505 * then we cannot store the request for later flushing, we need
1506 * to drop it on the floor.
1508 ctx->cached_cq_overflow++;
1509 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1511 if (list_empty(&ctx->cq_overflow_list)) {
1512 set_bit(0, &ctx->sq_check_overflow);
1513 set_bit(0, &ctx->cq_check_overflow);
1514 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1518 req->compl.cflags = cflags;
1519 refcount_inc(&req->refs);
1520 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1524 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1526 __io_cqring_fill_event(req, res, 0);
1529 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1530 unsigned int cflags)
1532 struct io_ring_ctx *ctx = req->ctx;
1533 unsigned long flags;
1535 spin_lock_irqsave(&ctx->completion_lock, flags);
1536 __io_cqring_fill_event(req, res, cflags);
1537 io_commit_cqring(ctx);
1539 * If we're the last reference to this request, add to our locked
1542 if (refcount_dec_and_test(&req->refs)) {
1543 struct io_comp_state *cs = &ctx->submit_state.comp;
1545 io_dismantle_req(req);
1546 io_put_task(req->task, 1);
1547 list_add(&req->compl.list, &cs->locked_free_list);
1548 cs->locked_free_nr++;
1551 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1553 io_cqring_ev_posted(ctx);
1556 percpu_ref_put(&ctx->refs);
1560 static void io_req_complete_state(struct io_kiocb *req, long res,
1561 unsigned int cflags)
1565 req->compl.cflags = cflags;
1566 req->flags |= REQ_F_COMPLETE_INLINE;
1569 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1570 long res, unsigned cflags)
1572 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1573 io_req_complete_state(req, res, cflags);
1575 io_req_complete_post(req, res, cflags);
1578 static inline void io_req_complete(struct io_kiocb *req, long res)
1580 __io_req_complete(req, 0, res, 0);
1583 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1585 struct io_submit_state *state = &ctx->submit_state;
1586 struct io_comp_state *cs = &state->comp;
1587 struct io_kiocb *req = NULL;
1590 * If we have more than a batch's worth of requests in our IRQ side
1591 * locked cache, grab the lock and move them over to our submission
1594 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1595 spin_lock_irq(&ctx->completion_lock);
1596 list_splice_init(&cs->locked_free_list, &cs->free_list);
1597 cs->locked_free_nr = 0;
1598 spin_unlock_irq(&ctx->completion_lock);
1601 while (!list_empty(&cs->free_list)) {
1602 req = list_first_entry(&cs->free_list, struct io_kiocb,
1604 list_del(&req->compl.list);
1605 state->reqs[state->free_reqs++] = req;
1606 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1613 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1615 struct io_submit_state *state = &ctx->submit_state;
1617 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1619 if (!state->free_reqs) {
1620 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1623 if (io_flush_cached_reqs(ctx))
1626 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1630 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1631 * retry single alloc to be on the safe side.
1633 if (unlikely(ret <= 0)) {
1634 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1635 if (!state->reqs[0])
1639 state->free_reqs = ret;
1643 return state->reqs[state->free_reqs];
1646 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1653 static void io_dismantle_req(struct io_kiocb *req)
1657 if (req->async_data)
1658 kfree(req->async_data);
1660 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1661 if (req->fixed_rsrc_refs)
1662 percpu_ref_put(req->fixed_rsrc_refs);
1663 if (req->work.creds) {
1664 put_cred(req->work.creds);
1665 req->work.creds = NULL;
1668 if (req->flags & REQ_F_INFLIGHT) {
1669 struct io_ring_ctx *ctx = req->ctx;
1670 unsigned long flags;
1672 spin_lock_irqsave(&ctx->inflight_lock, flags);
1673 list_del(&req->inflight_entry);
1674 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1675 req->flags &= ~REQ_F_INFLIGHT;
1679 /* must to be called somewhat shortly after putting a request */
1680 static inline void io_put_task(struct task_struct *task, int nr)
1682 struct io_uring_task *tctx = task->io_uring;
1684 percpu_counter_sub(&tctx->inflight, nr);
1685 if (unlikely(atomic_read(&tctx->in_idle)))
1686 wake_up(&tctx->wait);
1687 put_task_struct_many(task, nr);
1690 static void __io_free_req(struct io_kiocb *req)
1692 struct io_ring_ctx *ctx = req->ctx;
1694 io_dismantle_req(req);
1695 io_put_task(req->task, 1);
1697 kmem_cache_free(req_cachep, req);
1698 percpu_ref_put(&ctx->refs);
1701 static inline void io_remove_next_linked(struct io_kiocb *req)
1703 struct io_kiocb *nxt = req->link;
1705 req->link = nxt->link;
1709 static void io_kill_linked_timeout(struct io_kiocb *req)
1711 struct io_ring_ctx *ctx = req->ctx;
1712 struct io_kiocb *link;
1713 bool cancelled = false;
1714 unsigned long flags;
1716 spin_lock_irqsave(&ctx->completion_lock, flags);
1720 * Can happen if a linked timeout fired and link had been like
1721 * req -> link t-out -> link t-out [-> ...]
1723 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1724 struct io_timeout_data *io = link->async_data;
1727 io_remove_next_linked(req);
1728 link->timeout.head = NULL;
1729 ret = hrtimer_try_to_cancel(&io->timer);
1731 io_cqring_fill_event(link, -ECANCELED);
1732 io_commit_cqring(ctx);
1736 req->flags &= ~REQ_F_LINK_TIMEOUT;
1737 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1740 io_cqring_ev_posted(ctx);
1746 static void io_fail_links(struct io_kiocb *req)
1748 struct io_kiocb *link, *nxt;
1749 struct io_ring_ctx *ctx = req->ctx;
1750 unsigned long flags;
1752 spin_lock_irqsave(&ctx->completion_lock, flags);
1760 trace_io_uring_fail_link(req, link);
1761 io_cqring_fill_event(link, -ECANCELED);
1763 io_put_req_deferred(link, 2);
1766 io_commit_cqring(ctx);
1767 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1769 io_cqring_ev_posted(ctx);
1772 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1774 if (req->flags & REQ_F_LINK_TIMEOUT)
1775 io_kill_linked_timeout(req);
1778 * If LINK is set, we have dependent requests in this chain. If we
1779 * didn't fail this request, queue the first one up, moving any other
1780 * dependencies to the next request. In case of failure, fail the rest
1783 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1784 struct io_kiocb *nxt = req->link;
1793 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1795 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1797 return __io_req_find_next(req);
1800 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1804 if (ctx->submit_state.comp.nr) {
1805 mutex_lock(&ctx->uring_lock);
1806 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1807 mutex_unlock(&ctx->uring_lock);
1809 percpu_ref_put(&ctx->refs);
1812 static bool __tctx_task_work(struct io_uring_task *tctx)
1814 struct io_ring_ctx *ctx = NULL;
1815 struct io_wq_work_list list;
1816 struct io_wq_work_node *node;
1818 if (wq_list_empty(&tctx->task_list))
1821 spin_lock_irq(&tctx->task_lock);
1822 list = tctx->task_list;
1823 INIT_WQ_LIST(&tctx->task_list);
1824 spin_unlock_irq(&tctx->task_lock);
1828 struct io_wq_work_node *next = node->next;
1829 struct io_kiocb *req;
1831 req = container_of(node, struct io_kiocb, io_task_work.node);
1832 if (req->ctx != ctx) {
1833 ctx_flush_and_put(ctx);
1835 percpu_ref_get(&ctx->refs);
1838 req->task_work.func(&req->task_work);
1842 ctx_flush_and_put(ctx);
1843 return list.first != NULL;
1846 static void tctx_task_work(struct callback_head *cb)
1848 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1850 clear_bit(0, &tctx->task_state);
1852 while (__tctx_task_work(tctx))
1856 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1857 enum task_work_notify_mode notify)
1859 struct io_uring_task *tctx = tsk->io_uring;
1860 struct io_wq_work_node *node, *prev;
1861 unsigned long flags;
1864 WARN_ON_ONCE(!tctx);
1866 spin_lock_irqsave(&tctx->task_lock, flags);
1867 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1868 spin_unlock_irqrestore(&tctx->task_lock, flags);
1870 /* task_work already pending, we're done */
1871 if (test_bit(0, &tctx->task_state) ||
1872 test_and_set_bit(0, &tctx->task_state))
1875 if (!task_work_add(tsk, &tctx->task_work, notify))
1879 * Slow path - we failed, find and delete work. if the work is not
1880 * in the list, it got run and we're fine.
1883 spin_lock_irqsave(&tctx->task_lock, flags);
1884 wq_list_for_each(node, prev, &tctx->task_list) {
1885 if (&req->io_task_work.node == node) {
1886 wq_list_del(&tctx->task_list, node, prev);
1891 spin_unlock_irqrestore(&tctx->task_lock, flags);
1892 clear_bit(0, &tctx->task_state);
1896 static int io_req_task_work_add(struct io_kiocb *req)
1898 struct task_struct *tsk = req->task;
1899 struct io_ring_ctx *ctx = req->ctx;
1900 enum task_work_notify_mode notify;
1903 if (tsk->flags & PF_EXITING)
1907 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1908 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1909 * processing task_work. There's no reliable way to tell if TWA_RESUME
1913 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1914 notify = TWA_SIGNAL;
1916 ret = io_task_work_add(tsk, req, notify);
1918 wake_up_process(tsk);
1923 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1924 task_work_func_t cb)
1926 struct io_ring_ctx *ctx = req->ctx;
1927 struct callback_head *head;
1929 init_task_work(&req->task_work, cb);
1931 head = READ_ONCE(ctx->exit_task_work);
1932 req->task_work.next = head;
1933 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1936 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1938 struct io_ring_ctx *ctx = req->ctx;
1940 spin_lock_irq(&ctx->completion_lock);
1941 io_cqring_fill_event(req, error);
1942 io_commit_cqring(ctx);
1943 spin_unlock_irq(&ctx->completion_lock);
1945 io_cqring_ev_posted(ctx);
1946 req_set_fail_links(req);
1947 io_double_put_req(req);
1950 static void io_req_task_cancel(struct callback_head *cb)
1952 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1953 struct io_ring_ctx *ctx = req->ctx;
1955 mutex_lock(&ctx->uring_lock);
1956 __io_req_task_cancel(req, req->result);
1957 mutex_unlock(&ctx->uring_lock);
1958 percpu_ref_put(&ctx->refs);
1961 static void __io_req_task_submit(struct io_kiocb *req)
1963 struct io_ring_ctx *ctx = req->ctx;
1965 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1966 mutex_lock(&ctx->uring_lock);
1967 if (!(current->flags & PF_EXITING) && !current->in_execve)
1968 __io_queue_sqe(req);
1970 __io_req_task_cancel(req, -EFAULT);
1971 mutex_unlock(&ctx->uring_lock);
1974 static void io_req_task_submit(struct callback_head *cb)
1976 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1978 __io_req_task_submit(req);
1981 static void io_req_task_queue(struct io_kiocb *req)
1985 req->task_work.func = io_req_task_submit;
1986 ret = io_req_task_work_add(req);
1987 if (unlikely(ret)) {
1988 req->result = -ECANCELED;
1989 percpu_ref_get(&req->ctx->refs);
1990 io_req_task_work_add_fallback(req, io_req_task_cancel);
1994 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1996 percpu_ref_get(&req->ctx->refs);
1998 req->task_work.func = io_req_task_cancel;
2000 if (unlikely(io_req_task_work_add(req)))
2001 io_req_task_work_add_fallback(req, io_req_task_cancel);
2004 static inline void io_queue_next(struct io_kiocb *req)
2006 struct io_kiocb *nxt = io_req_find_next(req);
2009 io_req_task_queue(nxt);
2012 static void io_free_req(struct io_kiocb *req)
2019 struct task_struct *task;
2024 static inline void io_init_req_batch(struct req_batch *rb)
2031 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2032 struct req_batch *rb)
2035 io_put_task(rb->task, rb->task_refs);
2037 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2040 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2041 struct io_submit_state *state)
2045 if (req->task != rb->task) {
2047 io_put_task(rb->task, rb->task_refs);
2048 rb->task = req->task;
2054 io_dismantle_req(req);
2055 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2056 state->reqs[state->free_reqs++] = req;
2058 list_add(&req->compl.list, &state->comp.free_list);
2061 static void io_submit_flush_completions(struct io_comp_state *cs,
2062 struct io_ring_ctx *ctx)
2065 struct io_kiocb *req;
2066 struct req_batch rb;
2068 io_init_req_batch(&rb);
2069 spin_lock_irq(&ctx->completion_lock);
2070 for (i = 0; i < nr; i++) {
2072 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2074 io_commit_cqring(ctx);
2075 spin_unlock_irq(&ctx->completion_lock);
2077 io_cqring_ev_posted(ctx);
2078 for (i = 0; i < nr; i++) {
2081 /* submission and completion refs */
2082 if (refcount_sub_and_test(2, &req->refs))
2083 io_req_free_batch(&rb, req, &ctx->submit_state);
2086 io_req_free_batch_finish(ctx, &rb);
2091 * Drop reference to request, return next in chain (if there is one) if this
2092 * was the last reference to this request.
2094 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2096 struct io_kiocb *nxt = NULL;
2098 if (refcount_dec_and_test(&req->refs)) {
2099 nxt = io_req_find_next(req);
2105 static void io_put_req(struct io_kiocb *req)
2107 if (refcount_dec_and_test(&req->refs))
2111 static void io_put_req_deferred_cb(struct callback_head *cb)
2113 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2118 static void io_free_req_deferred(struct io_kiocb *req)
2122 req->task_work.func = io_put_req_deferred_cb;
2123 ret = io_req_task_work_add(req);
2125 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2128 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2130 if (refcount_sub_and_test(refs, &req->refs))
2131 io_free_req_deferred(req);
2134 static void io_double_put_req(struct io_kiocb *req)
2136 /* drop both submit and complete references */
2137 if (refcount_sub_and_test(2, &req->refs))
2141 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2143 /* See comment at the top of this file */
2145 return __io_cqring_events(ctx);
2148 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2150 struct io_rings *rings = ctx->rings;
2152 /* make sure SQ entry isn't read before tail */
2153 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2156 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2158 unsigned int cflags;
2160 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2161 cflags |= IORING_CQE_F_BUFFER;
2162 req->flags &= ~REQ_F_BUFFER_SELECTED;
2167 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2169 struct io_buffer *kbuf;
2171 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2172 return io_put_kbuf(req, kbuf);
2175 static inline bool io_run_task_work(void)
2178 * Not safe to run on exiting task, and the task_work handling will
2179 * not add work to such a task.
2181 if (unlikely(current->flags & PF_EXITING))
2183 if (current->task_works) {
2184 __set_current_state(TASK_RUNNING);
2193 * Find and free completed poll iocbs
2195 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2196 struct list_head *done)
2198 struct req_batch rb;
2199 struct io_kiocb *req;
2201 /* order with ->result store in io_complete_rw_iopoll() */
2204 io_init_req_batch(&rb);
2205 while (!list_empty(done)) {
2208 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2209 list_del(&req->inflight_entry);
2211 if (READ_ONCE(req->result) == -EAGAIN) {
2212 req->iopoll_completed = 0;
2213 if (io_rw_reissue(req))
2217 if (req->flags & REQ_F_BUFFER_SELECTED)
2218 cflags = io_put_rw_kbuf(req);
2220 __io_cqring_fill_event(req, req->result, cflags);
2223 if (refcount_dec_and_test(&req->refs))
2224 io_req_free_batch(&rb, req, &ctx->submit_state);
2227 io_commit_cqring(ctx);
2228 io_cqring_ev_posted_iopoll(ctx);
2229 io_req_free_batch_finish(ctx, &rb);
2232 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2235 struct io_kiocb *req, *tmp;
2241 * Only spin for completions if we don't have multiple devices hanging
2242 * off our complete list, and we're under the requested amount.
2244 spin = !ctx->poll_multi_file && *nr_events < min;
2247 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2248 struct kiocb *kiocb = &req->rw.kiocb;
2251 * Move completed and retryable entries to our local lists.
2252 * If we find a request that requires polling, break out
2253 * and complete those lists first, if we have entries there.
2255 if (READ_ONCE(req->iopoll_completed)) {
2256 list_move_tail(&req->inflight_entry, &done);
2259 if (!list_empty(&done))
2262 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2266 /* iopoll may have completed current req */
2267 if (READ_ONCE(req->iopoll_completed))
2268 list_move_tail(&req->inflight_entry, &done);
2275 if (!list_empty(&done))
2276 io_iopoll_complete(ctx, nr_events, &done);
2282 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2283 * non-spinning poll check - we'll still enter the driver poll loop, but only
2284 * as a non-spinning completion check.
2286 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2289 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2292 ret = io_do_iopoll(ctx, nr_events, min);
2295 if (*nr_events >= min)
2303 * We can't just wait for polled events to come to us, we have to actively
2304 * find and complete them.
2306 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2308 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2311 mutex_lock(&ctx->uring_lock);
2312 while (!list_empty(&ctx->iopoll_list)) {
2313 unsigned int nr_events = 0;
2315 io_do_iopoll(ctx, &nr_events, 0);
2317 /* let it sleep and repeat later if can't complete a request */
2321 * Ensure we allow local-to-the-cpu processing to take place,
2322 * in this case we need to ensure that we reap all events.
2323 * Also let task_work, etc. to progress by releasing the mutex
2325 if (need_resched()) {
2326 mutex_unlock(&ctx->uring_lock);
2328 mutex_lock(&ctx->uring_lock);
2331 mutex_unlock(&ctx->uring_lock);
2334 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2336 unsigned int nr_events = 0;
2337 int iters = 0, ret = 0;
2340 * We disallow the app entering submit/complete with polling, but we
2341 * still need to lock the ring to prevent racing with polled issue
2342 * that got punted to a workqueue.
2344 mutex_lock(&ctx->uring_lock);
2347 * Don't enter poll loop if we already have events pending.
2348 * If we do, we can potentially be spinning for commands that
2349 * already triggered a CQE (eg in error).
2351 if (test_bit(0, &ctx->cq_check_overflow))
2352 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2353 if (io_cqring_events(ctx))
2357 * If a submit got punted to a workqueue, we can have the
2358 * application entering polling for a command before it gets
2359 * issued. That app will hold the uring_lock for the duration
2360 * of the poll right here, so we need to take a breather every
2361 * now and then to ensure that the issue has a chance to add
2362 * the poll to the issued list. Otherwise we can spin here
2363 * forever, while the workqueue is stuck trying to acquire the
2366 if (!(++iters & 7)) {
2367 mutex_unlock(&ctx->uring_lock);
2369 mutex_lock(&ctx->uring_lock);
2372 ret = io_iopoll_getevents(ctx, &nr_events, min);
2376 } while (min && !nr_events && !need_resched());
2378 mutex_unlock(&ctx->uring_lock);
2382 static void kiocb_end_write(struct io_kiocb *req)
2385 * Tell lockdep we inherited freeze protection from submission
2388 if (req->flags & REQ_F_ISREG) {
2389 struct inode *inode = file_inode(req->file);
2391 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2393 file_end_write(req->file);
2397 static bool io_resubmit_prep(struct io_kiocb *req)
2399 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2401 struct iov_iter iter;
2403 /* already prepared */
2404 if (req->async_data)
2407 switch (req->opcode) {
2408 case IORING_OP_READV:
2409 case IORING_OP_READ_FIXED:
2410 case IORING_OP_READ:
2413 case IORING_OP_WRITEV:
2414 case IORING_OP_WRITE_FIXED:
2415 case IORING_OP_WRITE:
2419 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2424 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2427 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2430 static bool io_rw_should_reissue(struct io_kiocb *req)
2432 umode_t mode = file_inode(req->file)->i_mode;
2433 struct io_ring_ctx *ctx = req->ctx;
2435 if (!S_ISBLK(mode) && !S_ISREG(mode))
2437 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2438 !(ctx->flags & IORING_SETUP_IOPOLL)))
2441 * If ref is dying, we might be running poll reap from the exit work.
2442 * Don't attempt to reissue from that path, just let it fail with
2445 if (percpu_ref_is_dying(&ctx->refs))
2451 static bool io_rw_reissue(struct io_kiocb *req)
2454 if (!io_rw_should_reissue(req))
2457 lockdep_assert_held(&req->ctx->uring_lock);
2459 if (io_resubmit_prep(req)) {
2460 refcount_inc(&req->refs);
2461 io_queue_async_work(req);
2464 req_set_fail_links(req);
2469 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2470 unsigned int issue_flags)
2474 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2476 if (res != req->result)
2477 req_set_fail_links(req);
2479 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2480 kiocb_end_write(req);
2481 if (req->flags & REQ_F_BUFFER_SELECTED)
2482 cflags = io_put_rw_kbuf(req);
2483 __io_req_complete(req, issue_flags, res, cflags);
2486 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2488 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2490 __io_complete_rw(req, res, res2, 0);
2493 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2495 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2498 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2499 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2500 struct io_async_rw *rw = req->async_data;
2503 iov_iter_revert(&rw->iter,
2504 req->result - iov_iter_count(&rw->iter));
2505 else if (!io_resubmit_prep(req))
2510 if (kiocb->ki_flags & IOCB_WRITE)
2511 kiocb_end_write(req);
2513 if (res != -EAGAIN && res != req->result)
2514 req_set_fail_links(req);
2516 WRITE_ONCE(req->result, res);
2517 /* order with io_poll_complete() checking ->result */
2519 WRITE_ONCE(req->iopoll_completed, 1);
2523 * After the iocb has been issued, it's safe to be found on the poll list.
2524 * Adding the kiocb to the list AFTER submission ensures that we don't
2525 * find it from a io_iopoll_getevents() thread before the issuer is done
2526 * accessing the kiocb cookie.
2528 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2530 struct io_ring_ctx *ctx = req->ctx;
2533 * Track whether we have multiple files in our lists. This will impact
2534 * how we do polling eventually, not spinning if we're on potentially
2535 * different devices.
2537 if (list_empty(&ctx->iopoll_list)) {
2538 ctx->poll_multi_file = false;
2539 } else if (!ctx->poll_multi_file) {
2540 struct io_kiocb *list_req;
2542 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2544 if (list_req->file != req->file)
2545 ctx->poll_multi_file = true;
2549 * For fast devices, IO may have already completed. If it has, add
2550 * it to the front so we find it first.
2552 if (READ_ONCE(req->iopoll_completed))
2553 list_add(&req->inflight_entry, &ctx->iopoll_list);
2555 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2558 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2559 * task context or in io worker task context. If current task context is
2560 * sq thread, we don't need to check whether should wake up sq thread.
2562 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2563 wq_has_sleeper(&ctx->sq_data->wait))
2564 wake_up(&ctx->sq_data->wait);
2567 static inline void io_state_file_put(struct io_submit_state *state)
2569 if (state->file_refs) {
2570 fput_many(state->file, state->file_refs);
2571 state->file_refs = 0;
2576 * Get as many references to a file as we have IOs left in this submission,
2577 * assuming most submissions are for one file, or at least that each file
2578 * has more than one submission.
2580 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2585 if (state->file_refs) {
2586 if (state->fd == fd) {
2590 io_state_file_put(state);
2592 state->file = fget_many(fd, state->ios_left);
2593 if (unlikely(!state->file))
2597 state->file_refs = state->ios_left - 1;
2601 static bool io_bdev_nowait(struct block_device *bdev)
2603 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2607 * If we tracked the file through the SCM inflight mechanism, we could support
2608 * any file. For now, just ensure that anything potentially problematic is done
2611 static bool io_file_supports_async(struct file *file, int rw)
2613 umode_t mode = file_inode(file)->i_mode;
2615 if (S_ISBLK(mode)) {
2616 if (IS_ENABLED(CONFIG_BLOCK) &&
2617 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2621 if (S_ISCHR(mode) || S_ISSOCK(mode))
2623 if (S_ISREG(mode)) {
2624 if (IS_ENABLED(CONFIG_BLOCK) &&
2625 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2626 file->f_op != &io_uring_fops)
2631 /* any ->read/write should understand O_NONBLOCK */
2632 if (file->f_flags & O_NONBLOCK)
2635 if (!(file->f_mode & FMODE_NOWAIT))
2639 return file->f_op->read_iter != NULL;
2641 return file->f_op->write_iter != NULL;
2644 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2646 struct io_ring_ctx *ctx = req->ctx;
2647 struct kiocb *kiocb = &req->rw.kiocb;
2648 struct file *file = req->file;
2652 if (S_ISREG(file_inode(file)->i_mode))
2653 req->flags |= REQ_F_ISREG;
2655 kiocb->ki_pos = READ_ONCE(sqe->off);
2656 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2657 req->flags |= REQ_F_CUR_POS;
2658 kiocb->ki_pos = file->f_pos;
2660 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2661 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2662 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2666 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2667 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2668 req->flags |= REQ_F_NOWAIT;
2670 ioprio = READ_ONCE(sqe->ioprio);
2672 ret = ioprio_check_cap(ioprio);
2676 kiocb->ki_ioprio = ioprio;
2678 kiocb->ki_ioprio = get_current_ioprio();
2680 if (ctx->flags & IORING_SETUP_IOPOLL) {
2681 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2682 !kiocb->ki_filp->f_op->iopoll)
2685 kiocb->ki_flags |= IOCB_HIPRI;
2686 kiocb->ki_complete = io_complete_rw_iopoll;
2687 req->iopoll_completed = 0;
2689 if (kiocb->ki_flags & IOCB_HIPRI)
2691 kiocb->ki_complete = io_complete_rw;
2694 req->rw.addr = READ_ONCE(sqe->addr);
2695 req->rw.len = READ_ONCE(sqe->len);
2696 req->buf_index = READ_ONCE(sqe->buf_index);
2700 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2706 case -ERESTARTNOINTR:
2707 case -ERESTARTNOHAND:
2708 case -ERESTART_RESTARTBLOCK:
2710 * We can't just restart the syscall, since previously
2711 * submitted sqes may already be in progress. Just fail this
2717 kiocb->ki_complete(kiocb, ret, 0);
2721 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2722 unsigned int issue_flags)
2724 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2725 struct io_async_rw *io = req->async_data;
2727 /* add previously done IO, if any */
2728 if (io && io->bytes_done > 0) {
2730 ret = io->bytes_done;
2732 ret += io->bytes_done;
2735 if (req->flags & REQ_F_CUR_POS)
2736 req->file->f_pos = kiocb->ki_pos;
2737 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2738 __io_complete_rw(req, ret, 0, issue_flags);
2740 io_rw_done(kiocb, ret);
2743 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2745 struct io_ring_ctx *ctx = req->ctx;
2746 size_t len = req->rw.len;
2747 struct io_mapped_ubuf *imu;
2748 u16 index, buf_index = req->buf_index;
2752 if (unlikely(buf_index >= ctx->nr_user_bufs))
2754 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2755 imu = &ctx->user_bufs[index];
2756 buf_addr = req->rw.addr;
2759 if (buf_addr + len < buf_addr)
2761 /* not inside the mapped region */
2762 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2766 * May not be a start of buffer, set size appropriately
2767 * and advance us to the beginning.
2769 offset = buf_addr - imu->ubuf;
2770 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2774 * Don't use iov_iter_advance() here, as it's really slow for
2775 * using the latter parts of a big fixed buffer - it iterates
2776 * over each segment manually. We can cheat a bit here, because
2779 * 1) it's a BVEC iter, we set it up
2780 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2781 * first and last bvec
2783 * So just find our index, and adjust the iterator afterwards.
2784 * If the offset is within the first bvec (or the whole first
2785 * bvec, just use iov_iter_advance(). This makes it easier
2786 * since we can just skip the first segment, which may not
2787 * be PAGE_SIZE aligned.
2789 const struct bio_vec *bvec = imu->bvec;
2791 if (offset <= bvec->bv_len) {
2792 iov_iter_advance(iter, offset);
2794 unsigned long seg_skip;
2796 /* skip first vec */
2797 offset -= bvec->bv_len;
2798 seg_skip = 1 + (offset >> PAGE_SHIFT);
2800 iter->bvec = bvec + seg_skip;
2801 iter->nr_segs -= seg_skip;
2802 iter->count -= bvec->bv_len + offset;
2803 iter->iov_offset = offset & ~PAGE_MASK;
2810 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2813 mutex_unlock(&ctx->uring_lock);
2816 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2819 * "Normal" inline submissions always hold the uring_lock, since we
2820 * grab it from the system call. Same is true for the SQPOLL offload.
2821 * The only exception is when we've detached the request and issue it
2822 * from an async worker thread, grab the lock for that case.
2825 mutex_lock(&ctx->uring_lock);
2828 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2829 int bgid, struct io_buffer *kbuf,
2832 struct io_buffer *head;
2834 if (req->flags & REQ_F_BUFFER_SELECTED)
2837 io_ring_submit_lock(req->ctx, needs_lock);
2839 lockdep_assert_held(&req->ctx->uring_lock);
2841 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2843 if (!list_empty(&head->list)) {
2844 kbuf = list_last_entry(&head->list, struct io_buffer,
2846 list_del(&kbuf->list);
2849 idr_remove(&req->ctx->io_buffer_idr, bgid);
2851 if (*len > kbuf->len)
2854 kbuf = ERR_PTR(-ENOBUFS);
2857 io_ring_submit_unlock(req->ctx, needs_lock);
2862 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2865 struct io_buffer *kbuf;
2868 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2869 bgid = req->buf_index;
2870 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2873 req->rw.addr = (u64) (unsigned long) kbuf;
2874 req->flags |= REQ_F_BUFFER_SELECTED;
2875 return u64_to_user_ptr(kbuf->addr);
2878 #ifdef CONFIG_COMPAT
2879 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2882 struct compat_iovec __user *uiov;
2883 compat_ssize_t clen;
2887 uiov = u64_to_user_ptr(req->rw.addr);
2888 if (!access_ok(uiov, sizeof(*uiov)))
2890 if (__get_user(clen, &uiov->iov_len))
2896 buf = io_rw_buffer_select(req, &len, needs_lock);
2898 return PTR_ERR(buf);
2899 iov[0].iov_base = buf;
2900 iov[0].iov_len = (compat_size_t) len;
2905 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2908 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2912 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2915 len = iov[0].iov_len;
2918 buf = io_rw_buffer_select(req, &len, needs_lock);
2920 return PTR_ERR(buf);
2921 iov[0].iov_base = buf;
2922 iov[0].iov_len = len;
2926 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2929 if (req->flags & REQ_F_BUFFER_SELECTED) {
2930 struct io_buffer *kbuf;
2932 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2933 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2934 iov[0].iov_len = kbuf->len;
2937 if (req->rw.len != 1)
2940 #ifdef CONFIG_COMPAT
2941 if (req->ctx->compat)
2942 return io_compat_import(req, iov, needs_lock);
2945 return __io_iov_buffer_select(req, iov, needs_lock);
2948 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2949 struct iov_iter *iter, bool needs_lock)
2951 void __user *buf = u64_to_user_ptr(req->rw.addr);
2952 size_t sqe_len = req->rw.len;
2953 u8 opcode = req->opcode;
2956 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2958 return io_import_fixed(req, rw, iter);
2961 /* buffer index only valid with fixed read/write, or buffer select */
2962 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2965 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2966 if (req->flags & REQ_F_BUFFER_SELECT) {
2967 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2969 return PTR_ERR(buf);
2970 req->rw.len = sqe_len;
2973 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2978 if (req->flags & REQ_F_BUFFER_SELECT) {
2979 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2981 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2986 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2990 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2992 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2996 * For files that don't have ->read_iter() and ->write_iter(), handle them
2997 * by looping over ->read() or ->write() manually.
2999 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3001 struct kiocb *kiocb = &req->rw.kiocb;
3002 struct file *file = req->file;
3006 * Don't support polled IO through this interface, and we can't
3007 * support non-blocking either. For the latter, this just causes
3008 * the kiocb to be handled from an async context.
3010 if (kiocb->ki_flags & IOCB_HIPRI)
3012 if (kiocb->ki_flags & IOCB_NOWAIT)
3015 while (iov_iter_count(iter)) {
3019 if (!iov_iter_is_bvec(iter)) {
3020 iovec = iov_iter_iovec(iter);
3022 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3023 iovec.iov_len = req->rw.len;
3027 nr = file->f_op->read(file, iovec.iov_base,
3028 iovec.iov_len, io_kiocb_ppos(kiocb));
3030 nr = file->f_op->write(file, iovec.iov_base,
3031 iovec.iov_len, io_kiocb_ppos(kiocb));
3040 if (nr != iovec.iov_len)
3044 iov_iter_advance(iter, nr);
3050 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3051 const struct iovec *fast_iov, struct iov_iter *iter)
3053 struct io_async_rw *rw = req->async_data;
3055 memcpy(&rw->iter, iter, sizeof(*iter));
3056 rw->free_iovec = iovec;
3058 /* can only be fixed buffers, no need to do anything */
3059 if (iov_iter_is_bvec(iter))
3062 unsigned iov_off = 0;
3064 rw->iter.iov = rw->fast_iov;
3065 if (iter->iov != fast_iov) {
3066 iov_off = iter->iov - fast_iov;
3067 rw->iter.iov += iov_off;
3069 if (rw->fast_iov != fast_iov)
3070 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3071 sizeof(struct iovec) * iter->nr_segs);
3073 req->flags |= REQ_F_NEED_CLEANUP;
3077 static inline int __io_alloc_async_data(struct io_kiocb *req)
3079 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3080 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3081 return req->async_data == NULL;
3084 static int io_alloc_async_data(struct io_kiocb *req)
3086 if (!io_op_defs[req->opcode].needs_async_data)
3089 return __io_alloc_async_data(req);
3092 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3093 const struct iovec *fast_iov,
3094 struct iov_iter *iter, bool force)
3096 if (!force && !io_op_defs[req->opcode].needs_async_data)
3098 if (!req->async_data) {
3099 if (__io_alloc_async_data(req)) {
3104 io_req_map_rw(req, iovec, fast_iov, iter);
3109 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3111 struct io_async_rw *iorw = req->async_data;
3112 struct iovec *iov = iorw->fast_iov;
3115 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3116 if (unlikely(ret < 0))
3119 iorw->bytes_done = 0;
3120 iorw->free_iovec = iov;
3122 req->flags |= REQ_F_NEED_CLEANUP;
3126 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3128 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3130 return io_prep_rw(req, sqe);
3134 * This is our waitqueue callback handler, registered through lock_page_async()
3135 * when we initially tried to do the IO with the iocb armed our waitqueue.
3136 * This gets called when the page is unlocked, and we generally expect that to
3137 * happen when the page IO is completed and the page is now uptodate. This will
3138 * queue a task_work based retry of the operation, attempting to copy the data
3139 * again. If the latter fails because the page was NOT uptodate, then we will
3140 * do a thread based blocking retry of the operation. That's the unexpected
3143 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3144 int sync, void *arg)
3146 struct wait_page_queue *wpq;
3147 struct io_kiocb *req = wait->private;
3148 struct wait_page_key *key = arg;
3150 wpq = container_of(wait, struct wait_page_queue, wait);
3152 if (!wake_page_match(wpq, key))
3155 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3156 list_del_init(&wait->entry);
3158 /* submit ref gets dropped, acquire a new one */
3159 refcount_inc(&req->refs);
3160 io_req_task_queue(req);
3165 * This controls whether a given IO request should be armed for async page
3166 * based retry. If we return false here, the request is handed to the async
3167 * worker threads for retry. If we're doing buffered reads on a regular file,
3168 * we prepare a private wait_page_queue entry and retry the operation. This
3169 * will either succeed because the page is now uptodate and unlocked, or it
3170 * will register a callback when the page is unlocked at IO completion. Through
3171 * that callback, io_uring uses task_work to setup a retry of the operation.
3172 * That retry will attempt the buffered read again. The retry will generally
3173 * succeed, or in rare cases where it fails, we then fall back to using the
3174 * async worker threads for a blocking retry.
3176 static bool io_rw_should_retry(struct io_kiocb *req)
3178 struct io_async_rw *rw = req->async_data;
3179 struct wait_page_queue *wait = &rw->wpq;
3180 struct kiocb *kiocb = &req->rw.kiocb;
3182 /* never retry for NOWAIT, we just complete with -EAGAIN */
3183 if (req->flags & REQ_F_NOWAIT)
3186 /* Only for buffered IO */
3187 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3191 * just use poll if we can, and don't attempt if the fs doesn't
3192 * support callback based unlocks
3194 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3197 wait->wait.func = io_async_buf_func;
3198 wait->wait.private = req;
3199 wait->wait.flags = 0;
3200 INIT_LIST_HEAD(&wait->wait.entry);
3201 kiocb->ki_flags |= IOCB_WAITQ;
3202 kiocb->ki_flags &= ~IOCB_NOWAIT;
3203 kiocb->ki_waitq = wait;
3207 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3209 if (req->file->f_op->read_iter)
3210 return call_read_iter(req->file, &req->rw.kiocb, iter);
3211 else if (req->file->f_op->read)
3212 return loop_rw_iter(READ, req, iter);
3217 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3219 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3220 struct kiocb *kiocb = &req->rw.kiocb;
3221 struct iov_iter __iter, *iter = &__iter;
3222 struct io_async_rw *rw = req->async_data;
3223 ssize_t io_size, ret, ret2;
3224 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3230 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3234 io_size = iov_iter_count(iter);
3235 req->result = io_size;
3237 /* Ensure we clear previously set non-block flag */
3238 if (!force_nonblock)
3239 kiocb->ki_flags &= ~IOCB_NOWAIT;
3241 kiocb->ki_flags |= IOCB_NOWAIT;
3243 /* If the file doesn't support async, just async punt */
3244 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3245 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3246 return ret ?: -EAGAIN;
3249 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3250 if (unlikely(ret)) {
3255 ret = io_iter_do_read(req, iter);
3257 if (ret == -EIOCBQUEUED) {
3258 if (req->async_data)
3259 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3261 } else if (ret == -EAGAIN) {
3262 /* IOPOLL retry should happen for io-wq threads */
3263 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3265 /* no retry on NONBLOCK nor RWF_NOWAIT */
3266 if (req->flags & REQ_F_NOWAIT)
3268 /* some cases will consume bytes even on error returns */
3269 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3271 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3272 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3273 /* read all, failed, already did sync or don't want to retry */
3277 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3282 rw = req->async_data;
3283 /* now use our persistent iterator, if we aren't already */
3288 rw->bytes_done += ret;
3289 /* if we can retry, do so with the callbacks armed */
3290 if (!io_rw_should_retry(req)) {
3291 kiocb->ki_flags &= ~IOCB_WAITQ;
3296 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3297 * we get -EIOCBQUEUED, then we'll get a notification when the
3298 * desired page gets unlocked. We can also get a partial read
3299 * here, and if we do, then just retry at the new offset.
3301 ret = io_iter_do_read(req, iter);
3302 if (ret == -EIOCBQUEUED)
3304 /* we got some bytes, but not all. retry. */
3305 kiocb->ki_flags &= ~IOCB_WAITQ;
3306 } while (ret > 0 && ret < io_size);
3308 kiocb_done(kiocb, ret, issue_flags);
3310 /* it's faster to check here then delegate to kfree */
3316 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3318 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3320 return io_prep_rw(req, sqe);
3323 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3325 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3326 struct kiocb *kiocb = &req->rw.kiocb;
3327 struct iov_iter __iter, *iter = &__iter;
3328 struct io_async_rw *rw = req->async_data;
3329 ssize_t ret, ret2, io_size;
3330 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3336 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3340 io_size = iov_iter_count(iter);
3341 req->result = io_size;
3343 /* Ensure we clear previously set non-block flag */
3344 if (!force_nonblock)
3345 kiocb->ki_flags &= ~IOCB_NOWAIT;
3347 kiocb->ki_flags |= IOCB_NOWAIT;
3349 /* If the file doesn't support async, just async punt */
3350 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3353 /* file path doesn't support NOWAIT for non-direct_IO */
3354 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3355 (req->flags & REQ_F_ISREG))
3358 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3363 * Open-code file_start_write here to grab freeze protection,
3364 * which will be released by another thread in
3365 * io_complete_rw(). Fool lockdep by telling it the lock got
3366 * released so that it doesn't complain about the held lock when
3367 * we return to userspace.
3369 if (req->flags & REQ_F_ISREG) {
3370 sb_start_write(file_inode(req->file)->i_sb);
3371 __sb_writers_release(file_inode(req->file)->i_sb,
3374 kiocb->ki_flags |= IOCB_WRITE;
3376 if (req->file->f_op->write_iter)
3377 ret2 = call_write_iter(req->file, kiocb, iter);
3378 else if (req->file->f_op->write)
3379 ret2 = loop_rw_iter(WRITE, req, iter);
3384 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3385 * retry them without IOCB_NOWAIT.
3387 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3389 /* no retry on NONBLOCK nor RWF_NOWAIT */
3390 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3392 if (ret2 == -EIOCBQUEUED && req->async_data)
3393 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3394 if (!force_nonblock || ret2 != -EAGAIN) {
3395 /* IOPOLL retry should happen for io-wq threads */
3396 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3399 kiocb_done(kiocb, ret2, issue_flags);
3402 /* some cases will consume bytes even on error returns */
3403 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3404 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3405 return ret ?: -EAGAIN;
3408 /* it's reportedly faster than delegating the null check to kfree() */
3414 static int io_renameat_prep(struct io_kiocb *req,
3415 const struct io_uring_sqe *sqe)
3417 struct io_rename *ren = &req->rename;
3418 const char __user *oldf, *newf;
3420 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3423 ren->old_dfd = READ_ONCE(sqe->fd);
3424 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3425 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3426 ren->new_dfd = READ_ONCE(sqe->len);
3427 ren->flags = READ_ONCE(sqe->rename_flags);
3429 ren->oldpath = getname(oldf);
3430 if (IS_ERR(ren->oldpath))
3431 return PTR_ERR(ren->oldpath);
3433 ren->newpath = getname(newf);
3434 if (IS_ERR(ren->newpath)) {
3435 putname(ren->oldpath);
3436 return PTR_ERR(ren->newpath);
3439 req->flags |= REQ_F_NEED_CLEANUP;
3443 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3445 struct io_rename *ren = &req->rename;
3448 if (issue_flags & IO_URING_F_NONBLOCK)
3451 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3452 ren->newpath, ren->flags);
3454 req->flags &= ~REQ_F_NEED_CLEANUP;
3456 req_set_fail_links(req);
3457 io_req_complete(req, ret);
3461 static int io_unlinkat_prep(struct io_kiocb *req,
3462 const struct io_uring_sqe *sqe)
3464 struct io_unlink *un = &req->unlink;
3465 const char __user *fname;
3467 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3470 un->dfd = READ_ONCE(sqe->fd);
3472 un->flags = READ_ONCE(sqe->unlink_flags);
3473 if (un->flags & ~AT_REMOVEDIR)
3476 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3477 un->filename = getname(fname);
3478 if (IS_ERR(un->filename))
3479 return PTR_ERR(un->filename);
3481 req->flags |= REQ_F_NEED_CLEANUP;
3485 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3487 struct io_unlink *un = &req->unlink;
3490 if (issue_flags & IO_URING_F_NONBLOCK)
3493 if (un->flags & AT_REMOVEDIR)
3494 ret = do_rmdir(un->dfd, un->filename);
3496 ret = do_unlinkat(un->dfd, un->filename);
3498 req->flags &= ~REQ_F_NEED_CLEANUP;
3500 req_set_fail_links(req);
3501 io_req_complete(req, ret);
3505 static int io_shutdown_prep(struct io_kiocb *req,
3506 const struct io_uring_sqe *sqe)
3508 #if defined(CONFIG_NET)
3509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3511 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3515 req->shutdown.how = READ_ONCE(sqe->len);
3522 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3524 #if defined(CONFIG_NET)
3525 struct socket *sock;
3528 if (issue_flags & IO_URING_F_NONBLOCK)
3531 sock = sock_from_file(req->file);
3532 if (unlikely(!sock))
3535 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3537 req_set_fail_links(req);
3538 io_req_complete(req, ret);
3545 static int __io_splice_prep(struct io_kiocb *req,
3546 const struct io_uring_sqe *sqe)
3548 struct io_splice* sp = &req->splice;
3549 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3555 sp->len = READ_ONCE(sqe->len);
3556 sp->flags = READ_ONCE(sqe->splice_flags);
3558 if (unlikely(sp->flags & ~valid_flags))
3561 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3562 (sp->flags & SPLICE_F_FD_IN_FIXED));
3565 req->flags |= REQ_F_NEED_CLEANUP;
3567 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3569 * Splice operation will be punted aync, and here need to
3570 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3572 req->work.flags |= IO_WQ_WORK_UNBOUND;
3578 static int io_tee_prep(struct io_kiocb *req,
3579 const struct io_uring_sqe *sqe)
3581 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3583 return __io_splice_prep(req, sqe);
3586 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3588 struct io_splice *sp = &req->splice;
3589 struct file *in = sp->file_in;
3590 struct file *out = sp->file_out;
3591 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3594 if (issue_flags & IO_URING_F_NONBLOCK)
3597 ret = do_tee(in, out, sp->len, flags);
3599 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3600 req->flags &= ~REQ_F_NEED_CLEANUP;
3603 req_set_fail_links(req);
3604 io_req_complete(req, ret);
3608 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3610 struct io_splice* sp = &req->splice;
3612 sp->off_in = READ_ONCE(sqe->splice_off_in);
3613 sp->off_out = READ_ONCE(sqe->off);
3614 return __io_splice_prep(req, sqe);
3617 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3619 struct io_splice *sp = &req->splice;
3620 struct file *in = sp->file_in;
3621 struct file *out = sp->file_out;
3622 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3623 loff_t *poff_in, *poff_out;
3626 if (issue_flags & IO_URING_F_NONBLOCK)
3629 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3630 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3633 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3635 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3636 req->flags &= ~REQ_F_NEED_CLEANUP;
3639 req_set_fail_links(req);
3640 io_req_complete(req, ret);
3645 * IORING_OP_NOP just posts a completion event, nothing else.
3647 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3649 struct io_ring_ctx *ctx = req->ctx;
3651 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3654 __io_req_complete(req, issue_flags, 0, 0);
3658 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3660 struct io_ring_ctx *ctx = req->ctx;
3665 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3667 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3670 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3671 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3674 req->sync.off = READ_ONCE(sqe->off);
3675 req->sync.len = READ_ONCE(sqe->len);
3679 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3681 loff_t end = req->sync.off + req->sync.len;
3684 /* fsync always requires a blocking context */
3685 if (issue_flags & IO_URING_F_NONBLOCK)
3688 ret = vfs_fsync_range(req->file, req->sync.off,
3689 end > 0 ? end : LLONG_MAX,
3690 req->sync.flags & IORING_FSYNC_DATASYNC);
3692 req_set_fail_links(req);
3693 io_req_complete(req, ret);
3697 static int io_fallocate_prep(struct io_kiocb *req,
3698 const struct io_uring_sqe *sqe)
3700 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3702 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3705 req->sync.off = READ_ONCE(sqe->off);
3706 req->sync.len = READ_ONCE(sqe->addr);
3707 req->sync.mode = READ_ONCE(sqe->len);
3711 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3715 /* fallocate always requiring blocking context */
3716 if (issue_flags & IO_URING_F_NONBLOCK)
3718 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3721 req_set_fail_links(req);
3722 io_req_complete(req, ret);
3726 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3728 const char __user *fname;
3731 if (unlikely(sqe->ioprio || sqe->buf_index))
3733 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3736 /* open.how should be already initialised */
3737 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3738 req->open.how.flags |= O_LARGEFILE;
3740 req->open.dfd = READ_ONCE(sqe->fd);
3741 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3742 req->open.filename = getname(fname);
3743 if (IS_ERR(req->open.filename)) {
3744 ret = PTR_ERR(req->open.filename);
3745 req->open.filename = NULL;
3748 req->open.nofile = rlimit(RLIMIT_NOFILE);
3749 req->flags |= REQ_F_NEED_CLEANUP;
3753 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3757 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3759 mode = READ_ONCE(sqe->len);
3760 flags = READ_ONCE(sqe->open_flags);
3761 req->open.how = build_open_how(flags, mode);
3762 return __io_openat_prep(req, sqe);
3765 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 struct open_how __user *how;
3771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3773 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3774 len = READ_ONCE(sqe->len);
3775 if (len < OPEN_HOW_SIZE_VER0)
3778 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3783 return __io_openat_prep(req, sqe);
3786 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3788 struct open_flags op;
3791 bool resolve_nonblock;
3794 ret = build_open_flags(&req->open.how, &op);
3797 nonblock_set = op.open_flag & O_NONBLOCK;
3798 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3799 if (issue_flags & IO_URING_F_NONBLOCK) {
3801 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3802 * it'll always -EAGAIN
3804 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3806 op.lookup_flags |= LOOKUP_CACHED;
3807 op.open_flag |= O_NONBLOCK;
3810 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3814 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3815 /* only retry if RESOLVE_CACHED wasn't already set by application */
3816 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3817 file == ERR_PTR(-EAGAIN)) {
3819 * We could hang on to this 'fd', but seems like marginal
3820 * gain for something that is now known to be a slower path.
3821 * So just put it, and we'll get a new one when we retry.
3829 ret = PTR_ERR(file);
3831 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3832 file->f_flags &= ~O_NONBLOCK;
3833 fsnotify_open(file);
3834 fd_install(ret, file);
3837 putname(req->open.filename);
3838 req->flags &= ~REQ_F_NEED_CLEANUP;
3840 req_set_fail_links(req);
3841 io_req_complete(req, ret);
3845 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3847 return io_openat2(req, issue_flags);
3850 static int io_remove_buffers_prep(struct io_kiocb *req,
3851 const struct io_uring_sqe *sqe)
3853 struct io_provide_buf *p = &req->pbuf;
3856 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3859 tmp = READ_ONCE(sqe->fd);
3860 if (!tmp || tmp > USHRT_MAX)
3863 memset(p, 0, sizeof(*p));
3865 p->bgid = READ_ONCE(sqe->buf_group);
3869 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3870 int bgid, unsigned nbufs)
3874 /* shouldn't happen */
3878 /* the head kbuf is the list itself */
3879 while (!list_empty(&buf->list)) {
3880 struct io_buffer *nxt;
3882 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3883 list_del(&nxt->list);
3890 idr_remove(&ctx->io_buffer_idr, bgid);
3895 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3897 struct io_provide_buf *p = &req->pbuf;
3898 struct io_ring_ctx *ctx = req->ctx;
3899 struct io_buffer *head;
3901 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3903 io_ring_submit_lock(ctx, !force_nonblock);
3905 lockdep_assert_held(&ctx->uring_lock);
3908 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3910 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3912 req_set_fail_links(req);
3914 /* need to hold the lock to complete IOPOLL requests */
3915 if (ctx->flags & IORING_SETUP_IOPOLL) {
3916 __io_req_complete(req, issue_flags, ret, 0);
3917 io_ring_submit_unlock(ctx, !force_nonblock);
3919 io_ring_submit_unlock(ctx, !force_nonblock);
3920 __io_req_complete(req, issue_flags, ret, 0);
3925 static int io_provide_buffers_prep(struct io_kiocb *req,
3926 const struct io_uring_sqe *sqe)
3928 struct io_provide_buf *p = &req->pbuf;
3931 if (sqe->ioprio || sqe->rw_flags)
3934 tmp = READ_ONCE(sqe->fd);
3935 if (!tmp || tmp > USHRT_MAX)
3938 p->addr = READ_ONCE(sqe->addr);
3939 p->len = READ_ONCE(sqe->len);
3941 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3944 p->bgid = READ_ONCE(sqe->buf_group);
3945 tmp = READ_ONCE(sqe->off);
3946 if (tmp > USHRT_MAX)
3952 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3954 struct io_buffer *buf;
3955 u64 addr = pbuf->addr;
3956 int i, bid = pbuf->bid;
3958 for (i = 0; i < pbuf->nbufs; i++) {
3959 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3964 buf->len = pbuf->len;
3969 INIT_LIST_HEAD(&buf->list);
3972 list_add_tail(&buf->list, &(*head)->list);
3976 return i ? i : -ENOMEM;
3979 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3981 struct io_provide_buf *p = &req->pbuf;
3982 struct io_ring_ctx *ctx = req->ctx;
3983 struct io_buffer *head, *list;
3985 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3987 io_ring_submit_lock(ctx, !force_nonblock);
3989 lockdep_assert_held(&ctx->uring_lock);
3991 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3993 ret = io_add_buffers(p, &head);
3998 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4001 __io_remove_buffers(ctx, head, p->bgid, -1U);
4007 req_set_fail_links(req);
4009 /* need to hold the lock to complete IOPOLL requests */
4010 if (ctx->flags & IORING_SETUP_IOPOLL) {
4011 __io_req_complete(req, issue_flags, ret, 0);
4012 io_ring_submit_unlock(ctx, !force_nonblock);
4014 io_ring_submit_unlock(ctx, !force_nonblock);
4015 __io_req_complete(req, issue_flags, ret, 0);
4020 static int io_epoll_ctl_prep(struct io_kiocb *req,
4021 const struct io_uring_sqe *sqe)
4023 #if defined(CONFIG_EPOLL)
4024 if (sqe->ioprio || sqe->buf_index)
4026 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4029 req->epoll.epfd = READ_ONCE(sqe->fd);
4030 req->epoll.op = READ_ONCE(sqe->len);
4031 req->epoll.fd = READ_ONCE(sqe->off);
4033 if (ep_op_has_event(req->epoll.op)) {
4034 struct epoll_event __user *ev;
4036 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4037 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4047 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4049 #if defined(CONFIG_EPOLL)
4050 struct io_epoll *ie = &req->epoll;
4052 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4054 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4055 if (force_nonblock && ret == -EAGAIN)
4059 req_set_fail_links(req);
4060 __io_req_complete(req, issue_flags, ret, 0);
4067 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4070 if (sqe->ioprio || sqe->buf_index || sqe->off)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->madvise.addr = READ_ONCE(sqe->addr);
4076 req->madvise.len = READ_ONCE(sqe->len);
4077 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4084 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4086 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4087 struct io_madvise *ma = &req->madvise;
4090 if (issue_flags & IO_URING_F_NONBLOCK)
4093 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4095 req_set_fail_links(req);
4096 io_req_complete(req, ret);
4103 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4105 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4107 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4110 req->fadvise.offset = READ_ONCE(sqe->off);
4111 req->fadvise.len = READ_ONCE(sqe->len);
4112 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4116 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4118 struct io_fadvise *fa = &req->fadvise;
4121 if (issue_flags & IO_URING_F_NONBLOCK) {
4122 switch (fa->advice) {
4123 case POSIX_FADV_NORMAL:
4124 case POSIX_FADV_RANDOM:
4125 case POSIX_FADV_SEQUENTIAL:
4132 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4134 req_set_fail_links(req);
4135 io_req_complete(req, ret);
4139 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4141 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4143 if (sqe->ioprio || sqe->buf_index)
4145 if (req->flags & REQ_F_FIXED_FILE)
4148 req->statx.dfd = READ_ONCE(sqe->fd);
4149 req->statx.mask = READ_ONCE(sqe->len);
4150 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4151 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4152 req->statx.flags = READ_ONCE(sqe->statx_flags);
4157 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4159 struct io_statx *ctx = &req->statx;
4162 if (issue_flags & IO_URING_F_NONBLOCK) {
4163 /* only need file table for an actual valid fd */
4164 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4165 req->flags |= REQ_F_NO_FILE_TABLE;
4169 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4173 req_set_fail_links(req);
4174 io_req_complete(req, ret);
4178 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4180 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4182 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4183 sqe->rw_flags || sqe->buf_index)
4185 if (req->flags & REQ_F_FIXED_FILE)
4188 req->close.fd = READ_ONCE(sqe->fd);
4192 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4194 struct files_struct *files = current->files;
4195 struct io_close *close = &req->close;
4196 struct fdtable *fdt;
4202 spin_lock(&files->file_lock);
4203 fdt = files_fdtable(files);
4204 if (close->fd >= fdt->max_fds) {
4205 spin_unlock(&files->file_lock);
4208 file = fdt->fd[close->fd];
4210 spin_unlock(&files->file_lock);
4214 if (file->f_op == &io_uring_fops) {
4215 spin_unlock(&files->file_lock);
4220 /* if the file has a flush method, be safe and punt to async */
4221 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4222 spin_unlock(&files->file_lock);
4226 ret = __close_fd_get_file(close->fd, &file);
4227 spin_unlock(&files->file_lock);
4234 /* No ->flush() or already async, safely close from here */
4235 ret = filp_close(file, current->files);
4238 req_set_fail_links(req);
4241 __io_req_complete(req, issue_flags, ret, 0);
4245 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4247 struct io_ring_ctx *ctx = req->ctx;
4249 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4251 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4254 req->sync.off = READ_ONCE(sqe->off);
4255 req->sync.len = READ_ONCE(sqe->len);
4256 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4260 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4264 /* sync_file_range always requires a blocking context */
4265 if (issue_flags & IO_URING_F_NONBLOCK)
4268 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4271 req_set_fail_links(req);
4272 io_req_complete(req, ret);
4276 #if defined(CONFIG_NET)
4277 static int io_setup_async_msg(struct io_kiocb *req,
4278 struct io_async_msghdr *kmsg)
4280 struct io_async_msghdr *async_msg = req->async_data;
4284 if (io_alloc_async_data(req)) {
4285 kfree(kmsg->free_iov);
4288 async_msg = req->async_data;
4289 req->flags |= REQ_F_NEED_CLEANUP;
4290 memcpy(async_msg, kmsg, sizeof(*kmsg));
4291 async_msg->msg.msg_name = &async_msg->addr;
4292 /* if were using fast_iov, set it to the new one */
4293 if (!async_msg->free_iov)
4294 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4299 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4300 struct io_async_msghdr *iomsg)
4302 iomsg->msg.msg_name = &iomsg->addr;
4303 iomsg->free_iov = iomsg->fast_iov;
4304 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4305 req->sr_msg.msg_flags, &iomsg->free_iov);
4308 static int io_sendmsg_prep_async(struct io_kiocb *req)
4312 if (!io_op_defs[req->opcode].needs_async_data)
4314 ret = io_sendmsg_copy_hdr(req, req->async_data);
4316 req->flags |= REQ_F_NEED_CLEANUP;
4320 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4322 struct io_sr_msg *sr = &req->sr_msg;
4324 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4327 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4328 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4329 sr->len = READ_ONCE(sqe->len);
4331 #ifdef CONFIG_COMPAT
4332 if (req->ctx->compat)
4333 sr->msg_flags |= MSG_CMSG_COMPAT;
4338 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4340 struct io_async_msghdr iomsg, *kmsg;
4341 struct socket *sock;
4345 sock = sock_from_file(req->file);
4346 if (unlikely(!sock))
4349 kmsg = req->async_data;
4351 ret = io_sendmsg_copy_hdr(req, &iomsg);
4357 flags = req->sr_msg.msg_flags;
4358 if (flags & MSG_DONTWAIT)
4359 req->flags |= REQ_F_NOWAIT;
4360 else if (issue_flags & IO_URING_F_NONBLOCK)
4361 flags |= MSG_DONTWAIT;
4363 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4364 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4365 return io_setup_async_msg(req, kmsg);
4366 if (ret == -ERESTARTSYS)
4369 /* fast path, check for non-NULL to avoid function call */
4371 kfree(kmsg->free_iov);
4372 req->flags &= ~REQ_F_NEED_CLEANUP;
4374 req_set_fail_links(req);
4375 __io_req_complete(req, issue_flags, ret, 0);
4379 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4381 struct io_sr_msg *sr = &req->sr_msg;
4384 struct socket *sock;
4388 sock = sock_from_file(req->file);
4389 if (unlikely(!sock))
4392 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4396 msg.msg_name = NULL;
4397 msg.msg_control = NULL;
4398 msg.msg_controllen = 0;
4399 msg.msg_namelen = 0;
4401 flags = req->sr_msg.msg_flags;
4402 if (flags & MSG_DONTWAIT)
4403 req->flags |= REQ_F_NOWAIT;
4404 else if (issue_flags & IO_URING_F_NONBLOCK)
4405 flags |= MSG_DONTWAIT;
4407 msg.msg_flags = flags;
4408 ret = sock_sendmsg(sock, &msg);
4409 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4411 if (ret == -ERESTARTSYS)
4415 req_set_fail_links(req);
4416 __io_req_complete(req, issue_flags, ret, 0);
4420 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4421 struct io_async_msghdr *iomsg)
4423 struct io_sr_msg *sr = &req->sr_msg;
4424 struct iovec __user *uiov;
4428 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4429 &iomsg->uaddr, &uiov, &iov_len);
4433 if (req->flags & REQ_F_BUFFER_SELECT) {
4436 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4438 sr->len = iomsg->fast_iov[0].iov_len;
4439 iomsg->free_iov = NULL;
4441 iomsg->free_iov = iomsg->fast_iov;
4442 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4443 &iomsg->free_iov, &iomsg->msg.msg_iter,
4452 #ifdef CONFIG_COMPAT
4453 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4454 struct io_async_msghdr *iomsg)
4456 struct compat_msghdr __user *msg_compat;
4457 struct io_sr_msg *sr = &req->sr_msg;
4458 struct compat_iovec __user *uiov;
4463 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4464 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4469 uiov = compat_ptr(ptr);
4470 if (req->flags & REQ_F_BUFFER_SELECT) {
4471 compat_ssize_t clen;
4475 if (!access_ok(uiov, sizeof(*uiov)))
4477 if (__get_user(clen, &uiov->iov_len))
4482 iomsg->free_iov = NULL;
4484 iomsg->free_iov = iomsg->fast_iov;
4485 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4486 UIO_FASTIOV, &iomsg->free_iov,
4487 &iomsg->msg.msg_iter, true);
4496 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4497 struct io_async_msghdr *iomsg)
4499 iomsg->msg.msg_name = &iomsg->addr;
4501 #ifdef CONFIG_COMPAT
4502 if (req->ctx->compat)
4503 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4506 return __io_recvmsg_copy_hdr(req, iomsg);
4509 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4512 struct io_sr_msg *sr = &req->sr_msg;
4513 struct io_buffer *kbuf;
4515 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4520 req->flags |= REQ_F_BUFFER_SELECTED;
4524 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4526 return io_put_kbuf(req, req->sr_msg.kbuf);
4529 static int io_recvmsg_prep_async(struct io_kiocb *req)
4533 if (!io_op_defs[req->opcode].needs_async_data)
4535 ret = io_recvmsg_copy_hdr(req, req->async_data);
4537 req->flags |= REQ_F_NEED_CLEANUP;
4541 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4543 struct io_sr_msg *sr = &req->sr_msg;
4545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4548 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4549 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4550 sr->len = READ_ONCE(sqe->len);
4551 sr->bgid = READ_ONCE(sqe->buf_group);
4553 #ifdef CONFIG_COMPAT
4554 if (req->ctx->compat)
4555 sr->msg_flags |= MSG_CMSG_COMPAT;
4560 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4562 struct io_async_msghdr iomsg, *kmsg;
4563 struct socket *sock;
4564 struct io_buffer *kbuf;
4566 int ret, cflags = 0;
4567 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4569 sock = sock_from_file(req->file);
4570 if (unlikely(!sock))
4573 kmsg = req->async_data;
4575 ret = io_recvmsg_copy_hdr(req, &iomsg);
4581 if (req->flags & REQ_F_BUFFER_SELECT) {
4582 kbuf = io_recv_buffer_select(req, !force_nonblock);
4584 return PTR_ERR(kbuf);
4585 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4586 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4587 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4588 1, req->sr_msg.len);
4591 flags = req->sr_msg.msg_flags;
4592 if (flags & MSG_DONTWAIT)
4593 req->flags |= REQ_F_NOWAIT;
4594 else if (force_nonblock)
4595 flags |= MSG_DONTWAIT;
4597 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4598 kmsg->uaddr, flags);
4599 if (force_nonblock && ret == -EAGAIN)
4600 return io_setup_async_msg(req, kmsg);
4601 if (ret == -ERESTARTSYS)
4604 if (req->flags & REQ_F_BUFFER_SELECTED)
4605 cflags = io_put_recv_kbuf(req);
4606 /* fast path, check for non-NULL to avoid function call */
4608 kfree(kmsg->free_iov);
4609 req->flags &= ~REQ_F_NEED_CLEANUP;
4611 req_set_fail_links(req);
4612 __io_req_complete(req, issue_flags, ret, cflags);
4616 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4618 struct io_buffer *kbuf;
4619 struct io_sr_msg *sr = &req->sr_msg;
4621 void __user *buf = sr->buf;
4622 struct socket *sock;
4625 int ret, cflags = 0;
4626 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4628 sock = sock_from_file(req->file);
4629 if (unlikely(!sock))
4632 if (req->flags & REQ_F_BUFFER_SELECT) {
4633 kbuf = io_recv_buffer_select(req, !force_nonblock);
4635 return PTR_ERR(kbuf);
4636 buf = u64_to_user_ptr(kbuf->addr);
4639 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4643 msg.msg_name = NULL;
4644 msg.msg_control = NULL;
4645 msg.msg_controllen = 0;
4646 msg.msg_namelen = 0;
4647 msg.msg_iocb = NULL;
4650 flags = req->sr_msg.msg_flags;
4651 if (flags & MSG_DONTWAIT)
4652 req->flags |= REQ_F_NOWAIT;
4653 else if (force_nonblock)
4654 flags |= MSG_DONTWAIT;
4656 ret = sock_recvmsg(sock, &msg, flags);
4657 if (force_nonblock && ret == -EAGAIN)
4659 if (ret == -ERESTARTSYS)
4662 if (req->flags & REQ_F_BUFFER_SELECTED)
4663 cflags = io_put_recv_kbuf(req);
4665 req_set_fail_links(req);
4666 __io_req_complete(req, issue_flags, ret, cflags);
4670 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4672 struct io_accept *accept = &req->accept;
4674 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4676 if (sqe->ioprio || sqe->len || sqe->buf_index)
4679 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4680 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4681 accept->flags = READ_ONCE(sqe->accept_flags);
4682 accept->nofile = rlimit(RLIMIT_NOFILE);
4686 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4688 struct io_accept *accept = &req->accept;
4689 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4690 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4693 if (req->file->f_flags & O_NONBLOCK)
4694 req->flags |= REQ_F_NOWAIT;
4696 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4697 accept->addr_len, accept->flags,
4699 if (ret == -EAGAIN && force_nonblock)
4702 if (ret == -ERESTARTSYS)
4704 req_set_fail_links(req);
4706 __io_req_complete(req, issue_flags, ret, 0);
4710 static int io_connect_prep_async(struct io_kiocb *req)
4712 struct io_async_connect *io = req->async_data;
4713 struct io_connect *conn = &req->connect;
4715 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4718 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4720 struct io_connect *conn = &req->connect;
4722 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4724 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4727 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4728 conn->addr_len = READ_ONCE(sqe->addr2);
4732 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4734 struct io_async_connect __io, *io;
4735 unsigned file_flags;
4737 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4739 if (req->async_data) {
4740 io = req->async_data;
4742 ret = move_addr_to_kernel(req->connect.addr,
4743 req->connect.addr_len,
4750 file_flags = force_nonblock ? O_NONBLOCK : 0;
4752 ret = __sys_connect_file(req->file, &io->address,
4753 req->connect.addr_len, file_flags);
4754 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4755 if (req->async_data)
4757 if (io_alloc_async_data(req)) {
4761 io = req->async_data;
4762 memcpy(req->async_data, &__io, sizeof(__io));
4765 if (ret == -ERESTARTSYS)
4769 req_set_fail_links(req);
4770 __io_req_complete(req, issue_flags, ret, 0);
4773 #else /* !CONFIG_NET */
4774 #define IO_NETOP_FN(op) \
4775 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4777 return -EOPNOTSUPP; \
4780 #define IO_NETOP_PREP(op) \
4782 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4784 return -EOPNOTSUPP; \
4787 #define IO_NETOP_PREP_ASYNC(op) \
4789 static int io_##op##_prep_async(struct io_kiocb *req) \
4791 return -EOPNOTSUPP; \
4794 IO_NETOP_PREP_ASYNC(sendmsg);
4795 IO_NETOP_PREP_ASYNC(recvmsg);
4796 IO_NETOP_PREP_ASYNC(connect);
4797 IO_NETOP_PREP(accept);
4800 #endif /* CONFIG_NET */
4802 struct io_poll_table {
4803 struct poll_table_struct pt;
4804 struct io_kiocb *req;
4808 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4809 __poll_t mask, task_work_func_t func)
4813 /* for instances that support it check for an event match first: */
4814 if (mask && !(mask & poll->events))
4817 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4819 list_del_init(&poll->wait.entry);
4822 req->task_work.func = func;
4823 percpu_ref_get(&req->ctx->refs);
4826 * If this fails, then the task is exiting. When a task exits, the
4827 * work gets canceled, so just cancel this request as well instead
4828 * of executing it. We can't safely execute it anyway, as we may not
4829 * have the needed state needed for it anyway.
4831 ret = io_req_task_work_add(req);
4832 if (unlikely(ret)) {
4833 WRITE_ONCE(poll->canceled, true);
4834 io_req_task_work_add_fallback(req, func);
4839 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4840 __acquires(&req->ctx->completion_lock)
4842 struct io_ring_ctx *ctx = req->ctx;
4844 if (!req->result && !READ_ONCE(poll->canceled)) {
4845 struct poll_table_struct pt = { ._key = poll->events };
4847 req->result = vfs_poll(req->file, &pt) & poll->events;
4850 spin_lock_irq(&ctx->completion_lock);
4851 if (!req->result && !READ_ONCE(poll->canceled)) {
4852 add_wait_queue(poll->head, &poll->wait);
4859 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4861 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4862 if (req->opcode == IORING_OP_POLL_ADD)
4863 return req->async_data;
4864 return req->apoll->double_poll;
4867 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4869 if (req->opcode == IORING_OP_POLL_ADD)
4871 return &req->apoll->poll;
4874 static void io_poll_remove_double(struct io_kiocb *req)
4876 struct io_poll_iocb *poll = io_poll_get_double(req);
4878 lockdep_assert_held(&req->ctx->completion_lock);
4880 if (poll && poll->head) {
4881 struct wait_queue_head *head = poll->head;
4883 spin_lock(&head->lock);
4884 list_del_init(&poll->wait.entry);
4885 if (poll->wait.private)
4886 refcount_dec(&req->refs);
4888 spin_unlock(&head->lock);
4892 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4894 struct io_ring_ctx *ctx = req->ctx;
4896 io_poll_remove_double(req);
4897 req->poll.done = true;
4898 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4899 io_commit_cqring(ctx);
4902 static void io_poll_task_func(struct callback_head *cb)
4904 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4905 struct io_ring_ctx *ctx = req->ctx;
4906 struct io_kiocb *nxt;
4908 if (io_poll_rewait(req, &req->poll)) {
4909 spin_unlock_irq(&ctx->completion_lock);
4911 hash_del(&req->hash_node);
4912 io_poll_complete(req, req->result, 0);
4913 spin_unlock_irq(&ctx->completion_lock);
4915 nxt = io_put_req_find_next(req);
4916 io_cqring_ev_posted(ctx);
4918 __io_req_task_submit(nxt);
4921 percpu_ref_put(&ctx->refs);
4924 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4925 int sync, void *key)
4927 struct io_kiocb *req = wait->private;
4928 struct io_poll_iocb *poll = io_poll_get_single(req);
4929 __poll_t mask = key_to_poll(key);
4931 /* for instances that support it check for an event match first: */
4932 if (mask && !(mask & poll->events))
4935 list_del_init(&wait->entry);
4937 if (poll && poll->head) {
4940 spin_lock(&poll->head->lock);
4941 done = list_empty(&poll->wait.entry);
4943 list_del_init(&poll->wait.entry);
4944 /* make sure double remove sees this as being gone */
4945 wait->private = NULL;
4946 spin_unlock(&poll->head->lock);
4948 /* use wait func handler, so it matches the rq type */
4949 poll->wait.func(&poll->wait, mode, sync, key);
4952 refcount_dec(&req->refs);
4956 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4957 wait_queue_func_t wake_func)
4961 poll->canceled = false;
4962 poll->events = events;
4963 INIT_LIST_HEAD(&poll->wait.entry);
4964 init_waitqueue_func_entry(&poll->wait, wake_func);
4967 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4968 struct wait_queue_head *head,
4969 struct io_poll_iocb **poll_ptr)
4971 struct io_kiocb *req = pt->req;
4974 * If poll->head is already set, it's because the file being polled
4975 * uses multiple waitqueues for poll handling (eg one for read, one
4976 * for write). Setup a separate io_poll_iocb if this happens.
4978 if (unlikely(poll->head)) {
4979 struct io_poll_iocb *poll_one = poll;
4981 /* already have a 2nd entry, fail a third attempt */
4983 pt->error = -EINVAL;
4986 /* double add on the same waitqueue head, ignore */
4987 if (poll->head == head)
4989 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4991 pt->error = -ENOMEM;
4994 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4995 refcount_inc(&req->refs);
4996 poll->wait.private = req;
5003 if (poll->events & EPOLLEXCLUSIVE)
5004 add_wait_queue_exclusive(head, &poll->wait);
5006 add_wait_queue(head, &poll->wait);
5009 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5010 struct poll_table_struct *p)
5012 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5013 struct async_poll *apoll = pt->req->apoll;
5015 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5018 static void io_async_task_func(struct callback_head *cb)
5020 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5021 struct async_poll *apoll = req->apoll;
5022 struct io_ring_ctx *ctx = req->ctx;
5024 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5026 if (io_poll_rewait(req, &apoll->poll)) {
5027 spin_unlock_irq(&ctx->completion_lock);
5028 percpu_ref_put(&ctx->refs);
5032 /* If req is still hashed, it cannot have been canceled. Don't check. */
5033 if (hash_hashed(&req->hash_node))
5034 hash_del(&req->hash_node);
5036 io_poll_remove_double(req);
5037 spin_unlock_irq(&ctx->completion_lock);
5039 if (!READ_ONCE(apoll->poll.canceled))
5040 __io_req_task_submit(req);
5042 __io_req_task_cancel(req, -ECANCELED);
5044 percpu_ref_put(&ctx->refs);
5045 kfree(apoll->double_poll);
5049 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5052 struct io_kiocb *req = wait->private;
5053 struct io_poll_iocb *poll = &req->apoll->poll;
5055 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5058 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5061 static void io_poll_req_insert(struct io_kiocb *req)
5063 struct io_ring_ctx *ctx = req->ctx;
5064 struct hlist_head *list;
5066 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5067 hlist_add_head(&req->hash_node, list);
5070 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5071 struct io_poll_iocb *poll,
5072 struct io_poll_table *ipt, __poll_t mask,
5073 wait_queue_func_t wake_func)
5074 __acquires(&ctx->completion_lock)
5076 struct io_ring_ctx *ctx = req->ctx;
5077 bool cancel = false;
5079 INIT_HLIST_NODE(&req->hash_node);
5080 io_init_poll_iocb(poll, mask, wake_func);
5081 poll->file = req->file;
5082 poll->wait.private = req;
5084 ipt->pt._key = mask;
5086 ipt->error = -EINVAL;
5088 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5090 spin_lock_irq(&ctx->completion_lock);
5091 if (likely(poll->head)) {
5092 spin_lock(&poll->head->lock);
5093 if (unlikely(list_empty(&poll->wait.entry))) {
5099 if (mask || ipt->error)
5100 list_del_init(&poll->wait.entry);
5102 WRITE_ONCE(poll->canceled, true);
5103 else if (!poll->done) /* actually waiting for an event */
5104 io_poll_req_insert(req);
5105 spin_unlock(&poll->head->lock);
5111 static bool io_arm_poll_handler(struct io_kiocb *req)
5113 const struct io_op_def *def = &io_op_defs[req->opcode];
5114 struct io_ring_ctx *ctx = req->ctx;
5115 struct async_poll *apoll;
5116 struct io_poll_table ipt;
5120 if (!req->file || !file_can_poll(req->file))
5122 if (req->flags & REQ_F_POLLED)
5126 else if (def->pollout)
5130 /* if we can't nonblock try, then no point in arming a poll handler */
5131 if (!io_file_supports_async(req->file, rw))
5134 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5135 if (unlikely(!apoll))
5137 apoll->double_poll = NULL;
5139 req->flags |= REQ_F_POLLED;
5144 mask |= POLLIN | POLLRDNORM;
5146 mask |= POLLOUT | POLLWRNORM;
5148 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5149 if ((req->opcode == IORING_OP_RECVMSG) &&
5150 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5153 mask |= POLLERR | POLLPRI;
5155 ipt.pt._qproc = io_async_queue_proc;
5157 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5159 if (ret || ipt.error) {
5160 io_poll_remove_double(req);
5161 spin_unlock_irq(&ctx->completion_lock);
5162 kfree(apoll->double_poll);
5166 spin_unlock_irq(&ctx->completion_lock);
5167 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5168 apoll->poll.events);
5172 static bool __io_poll_remove_one(struct io_kiocb *req,
5173 struct io_poll_iocb *poll)
5175 bool do_complete = false;
5177 spin_lock(&poll->head->lock);
5178 WRITE_ONCE(poll->canceled, true);
5179 if (!list_empty(&poll->wait.entry)) {
5180 list_del_init(&poll->wait.entry);
5183 spin_unlock(&poll->head->lock);
5184 hash_del(&req->hash_node);
5188 static bool io_poll_remove_one(struct io_kiocb *req)
5192 io_poll_remove_double(req);
5194 if (req->opcode == IORING_OP_POLL_ADD) {
5195 do_complete = __io_poll_remove_one(req, &req->poll);
5197 struct async_poll *apoll = req->apoll;
5199 /* non-poll requests have submit ref still */
5200 do_complete = __io_poll_remove_one(req, &apoll->poll);
5203 kfree(apoll->double_poll);
5209 io_cqring_fill_event(req, -ECANCELED);
5210 io_commit_cqring(req->ctx);
5211 req_set_fail_links(req);
5212 io_put_req_deferred(req, 1);
5219 * Returns true if we found and killed one or more poll requests
5221 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5222 struct files_struct *files)
5224 struct hlist_node *tmp;
5225 struct io_kiocb *req;
5228 spin_lock_irq(&ctx->completion_lock);
5229 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5230 struct hlist_head *list;
5232 list = &ctx->cancel_hash[i];
5233 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5234 if (io_match_task(req, tsk, files))
5235 posted += io_poll_remove_one(req);
5238 spin_unlock_irq(&ctx->completion_lock);
5241 io_cqring_ev_posted(ctx);
5246 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5248 struct hlist_head *list;
5249 struct io_kiocb *req;
5251 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5252 hlist_for_each_entry(req, list, hash_node) {
5253 if (sqe_addr != req->user_data)
5255 if (io_poll_remove_one(req))
5263 static int io_poll_remove_prep(struct io_kiocb *req,
5264 const struct io_uring_sqe *sqe)
5266 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5268 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5272 req->poll_remove.addr = READ_ONCE(sqe->addr);
5277 * Find a running poll command that matches one specified in sqe->addr,
5278 * and remove it if found.
5280 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5282 struct io_ring_ctx *ctx = req->ctx;
5285 spin_lock_irq(&ctx->completion_lock);
5286 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5287 spin_unlock_irq(&ctx->completion_lock);
5290 req_set_fail_links(req);
5291 io_req_complete(req, ret);
5295 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5298 struct io_kiocb *req = wait->private;
5299 struct io_poll_iocb *poll = &req->poll;
5301 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5304 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5305 struct poll_table_struct *p)
5307 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5309 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5312 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5314 struct io_poll_iocb *poll = &req->poll;
5317 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5319 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5322 events = READ_ONCE(sqe->poll32_events);
5324 events = swahw32(events);
5326 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5327 (events & EPOLLEXCLUSIVE);
5331 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5333 struct io_poll_iocb *poll = &req->poll;
5334 struct io_ring_ctx *ctx = req->ctx;
5335 struct io_poll_table ipt;
5338 ipt.pt._qproc = io_poll_queue_proc;
5340 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5343 if (mask) { /* no async, we'd stolen it */
5345 io_poll_complete(req, mask, 0);
5347 spin_unlock_irq(&ctx->completion_lock);
5350 io_cqring_ev_posted(ctx);
5356 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5358 struct io_timeout_data *data = container_of(timer,
5359 struct io_timeout_data, timer);
5360 struct io_kiocb *req = data->req;
5361 struct io_ring_ctx *ctx = req->ctx;
5362 unsigned long flags;
5364 spin_lock_irqsave(&ctx->completion_lock, flags);
5365 list_del_init(&req->timeout.list);
5366 atomic_set(&req->ctx->cq_timeouts,
5367 atomic_read(&req->ctx->cq_timeouts) + 1);
5369 io_cqring_fill_event(req, -ETIME);
5370 io_commit_cqring(ctx);
5371 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5373 io_cqring_ev_posted(ctx);
5374 req_set_fail_links(req);
5376 return HRTIMER_NORESTART;
5379 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5382 struct io_timeout_data *io;
5383 struct io_kiocb *req;
5386 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5387 if (user_data == req->user_data) {
5394 return ERR_PTR(ret);
5396 io = req->async_data;
5397 ret = hrtimer_try_to_cancel(&io->timer);
5399 return ERR_PTR(-EALREADY);
5400 list_del_init(&req->timeout.list);
5404 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5406 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5409 return PTR_ERR(req);
5411 req_set_fail_links(req);
5412 io_cqring_fill_event(req, -ECANCELED);
5413 io_put_req_deferred(req, 1);
5417 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5418 struct timespec64 *ts, enum hrtimer_mode mode)
5420 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5421 struct io_timeout_data *data;
5424 return PTR_ERR(req);
5426 req->timeout.off = 0; /* noseq */
5427 data = req->async_data;
5428 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5429 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5430 data->timer.function = io_timeout_fn;
5431 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5435 static int io_timeout_remove_prep(struct io_kiocb *req,
5436 const struct io_uring_sqe *sqe)
5438 struct io_timeout_rem *tr = &req->timeout_rem;
5440 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5442 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5444 if (sqe->ioprio || sqe->buf_index || sqe->len)
5447 tr->addr = READ_ONCE(sqe->addr);
5448 tr->flags = READ_ONCE(sqe->timeout_flags);
5449 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5450 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5452 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5454 } else if (tr->flags) {
5455 /* timeout removal doesn't support flags */
5462 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5464 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5469 * Remove or update an existing timeout command
5471 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5473 struct io_timeout_rem *tr = &req->timeout_rem;
5474 struct io_ring_ctx *ctx = req->ctx;
5477 spin_lock_irq(&ctx->completion_lock);
5478 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5479 ret = io_timeout_cancel(ctx, tr->addr);
5481 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5482 io_translate_timeout_mode(tr->flags));
5484 io_cqring_fill_event(req, ret);
5485 io_commit_cqring(ctx);
5486 spin_unlock_irq(&ctx->completion_lock);
5487 io_cqring_ev_posted(ctx);
5489 req_set_fail_links(req);
5494 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5495 bool is_timeout_link)
5497 struct io_timeout_data *data;
5499 u32 off = READ_ONCE(sqe->off);
5501 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5503 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5505 if (off && is_timeout_link)
5507 flags = READ_ONCE(sqe->timeout_flags);
5508 if (flags & ~IORING_TIMEOUT_ABS)
5511 req->timeout.off = off;
5513 if (!req->async_data && io_alloc_async_data(req))
5516 data = req->async_data;
5519 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5522 data->mode = io_translate_timeout_mode(flags);
5523 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5524 io_req_track_inflight(req);
5528 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5530 struct io_ring_ctx *ctx = req->ctx;
5531 struct io_timeout_data *data = req->async_data;
5532 struct list_head *entry;
5533 u32 tail, off = req->timeout.off;
5535 spin_lock_irq(&ctx->completion_lock);
5538 * sqe->off holds how many events that need to occur for this
5539 * timeout event to be satisfied. If it isn't set, then this is
5540 * a pure timeout request, sequence isn't used.
5542 if (io_is_timeout_noseq(req)) {
5543 entry = ctx->timeout_list.prev;
5547 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5548 req->timeout.target_seq = tail + off;
5550 /* Update the last seq here in case io_flush_timeouts() hasn't.
5551 * This is safe because ->completion_lock is held, and submissions
5552 * and completions are never mixed in the same ->completion_lock section.
5554 ctx->cq_last_tm_flush = tail;
5557 * Insertion sort, ensuring the first entry in the list is always
5558 * the one we need first.
5560 list_for_each_prev(entry, &ctx->timeout_list) {
5561 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5564 if (io_is_timeout_noseq(nxt))
5566 /* nxt.seq is behind @tail, otherwise would've been completed */
5567 if (off >= nxt->timeout.target_seq - tail)
5571 list_add(&req->timeout.list, entry);
5572 data->timer.function = io_timeout_fn;
5573 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5574 spin_unlock_irq(&ctx->completion_lock);
5578 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5580 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5582 return req->user_data == (unsigned long) data;
5585 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5587 enum io_wq_cancel cancel_ret;
5593 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5594 switch (cancel_ret) {
5595 case IO_WQ_CANCEL_OK:
5598 case IO_WQ_CANCEL_RUNNING:
5601 case IO_WQ_CANCEL_NOTFOUND:
5609 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5610 struct io_kiocb *req, __u64 sqe_addr,
5613 unsigned long flags;
5616 ret = io_async_cancel_one(req->task->io_uring,
5617 (void *) (unsigned long) sqe_addr);
5618 if (ret != -ENOENT) {
5619 spin_lock_irqsave(&ctx->completion_lock, flags);
5623 spin_lock_irqsave(&ctx->completion_lock, flags);
5624 ret = io_timeout_cancel(ctx, sqe_addr);
5627 ret = io_poll_cancel(ctx, sqe_addr);
5631 io_cqring_fill_event(req, ret);
5632 io_commit_cqring(ctx);
5633 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5634 io_cqring_ev_posted(ctx);
5637 req_set_fail_links(req);
5641 static int io_async_cancel_prep(struct io_kiocb *req,
5642 const struct io_uring_sqe *sqe)
5644 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5646 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5648 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5651 req->cancel.addr = READ_ONCE(sqe->addr);
5655 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5657 struct io_ring_ctx *ctx = req->ctx;
5659 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5663 static int io_rsrc_update_prep(struct io_kiocb *req,
5664 const struct io_uring_sqe *sqe)
5666 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5668 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5670 if (sqe->ioprio || sqe->rw_flags)
5673 req->rsrc_update.offset = READ_ONCE(sqe->off);
5674 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5675 if (!req->rsrc_update.nr_args)
5677 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5681 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5683 struct io_ring_ctx *ctx = req->ctx;
5684 struct io_uring_rsrc_update up;
5687 if (issue_flags & IO_URING_F_NONBLOCK)
5690 up.offset = req->rsrc_update.offset;
5691 up.data = req->rsrc_update.arg;
5693 mutex_lock(&ctx->uring_lock);
5694 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5695 mutex_unlock(&ctx->uring_lock);
5698 req_set_fail_links(req);
5699 __io_req_complete(req, issue_flags, ret, 0);
5703 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5705 switch (req->opcode) {
5708 case IORING_OP_READV:
5709 case IORING_OP_READ_FIXED:
5710 case IORING_OP_READ:
5711 return io_read_prep(req, sqe);
5712 case IORING_OP_WRITEV:
5713 case IORING_OP_WRITE_FIXED:
5714 case IORING_OP_WRITE:
5715 return io_write_prep(req, sqe);
5716 case IORING_OP_POLL_ADD:
5717 return io_poll_add_prep(req, sqe);
5718 case IORING_OP_POLL_REMOVE:
5719 return io_poll_remove_prep(req, sqe);
5720 case IORING_OP_FSYNC:
5721 return io_fsync_prep(req, sqe);
5722 case IORING_OP_SYNC_FILE_RANGE:
5723 return io_sfr_prep(req, sqe);
5724 case IORING_OP_SENDMSG:
5725 case IORING_OP_SEND:
5726 return io_sendmsg_prep(req, sqe);
5727 case IORING_OP_RECVMSG:
5728 case IORING_OP_RECV:
5729 return io_recvmsg_prep(req, sqe);
5730 case IORING_OP_CONNECT:
5731 return io_connect_prep(req, sqe);
5732 case IORING_OP_TIMEOUT:
5733 return io_timeout_prep(req, sqe, false);
5734 case IORING_OP_TIMEOUT_REMOVE:
5735 return io_timeout_remove_prep(req, sqe);
5736 case IORING_OP_ASYNC_CANCEL:
5737 return io_async_cancel_prep(req, sqe);
5738 case IORING_OP_LINK_TIMEOUT:
5739 return io_timeout_prep(req, sqe, true);
5740 case IORING_OP_ACCEPT:
5741 return io_accept_prep(req, sqe);
5742 case IORING_OP_FALLOCATE:
5743 return io_fallocate_prep(req, sqe);
5744 case IORING_OP_OPENAT:
5745 return io_openat_prep(req, sqe);
5746 case IORING_OP_CLOSE:
5747 return io_close_prep(req, sqe);
5748 case IORING_OP_FILES_UPDATE:
5749 return io_rsrc_update_prep(req, sqe);
5750 case IORING_OP_STATX:
5751 return io_statx_prep(req, sqe);
5752 case IORING_OP_FADVISE:
5753 return io_fadvise_prep(req, sqe);
5754 case IORING_OP_MADVISE:
5755 return io_madvise_prep(req, sqe);
5756 case IORING_OP_OPENAT2:
5757 return io_openat2_prep(req, sqe);
5758 case IORING_OP_EPOLL_CTL:
5759 return io_epoll_ctl_prep(req, sqe);
5760 case IORING_OP_SPLICE:
5761 return io_splice_prep(req, sqe);
5762 case IORING_OP_PROVIDE_BUFFERS:
5763 return io_provide_buffers_prep(req, sqe);
5764 case IORING_OP_REMOVE_BUFFERS:
5765 return io_remove_buffers_prep(req, sqe);
5767 return io_tee_prep(req, sqe);
5768 case IORING_OP_SHUTDOWN:
5769 return io_shutdown_prep(req, sqe);
5770 case IORING_OP_RENAMEAT:
5771 return io_renameat_prep(req, sqe);
5772 case IORING_OP_UNLINKAT:
5773 return io_unlinkat_prep(req, sqe);
5776 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5781 static int io_req_prep_async(struct io_kiocb *req)
5783 switch (req->opcode) {
5784 case IORING_OP_READV:
5785 case IORING_OP_READ_FIXED:
5786 case IORING_OP_READ:
5787 return io_rw_prep_async(req, READ);
5788 case IORING_OP_WRITEV:
5789 case IORING_OP_WRITE_FIXED:
5790 case IORING_OP_WRITE:
5791 return io_rw_prep_async(req, WRITE);
5792 case IORING_OP_SENDMSG:
5793 case IORING_OP_SEND:
5794 return io_sendmsg_prep_async(req);
5795 case IORING_OP_RECVMSG:
5796 case IORING_OP_RECV:
5797 return io_recvmsg_prep_async(req);
5798 case IORING_OP_CONNECT:
5799 return io_connect_prep_async(req);
5804 static int io_req_defer_prep(struct io_kiocb *req)
5806 if (!io_op_defs[req->opcode].needs_async_data)
5808 /* some opcodes init it during the inital prep */
5809 if (req->async_data)
5811 if (__io_alloc_async_data(req))
5813 return io_req_prep_async(req);
5816 static u32 io_get_sequence(struct io_kiocb *req)
5818 struct io_kiocb *pos;
5819 struct io_ring_ctx *ctx = req->ctx;
5820 u32 total_submitted, nr_reqs = 0;
5822 io_for_each_link(pos, req)
5825 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5826 return total_submitted - nr_reqs;
5829 static int io_req_defer(struct io_kiocb *req)
5831 struct io_ring_ctx *ctx = req->ctx;
5832 struct io_defer_entry *de;
5836 /* Still need defer if there is pending req in defer list. */
5837 if (likely(list_empty_careful(&ctx->defer_list) &&
5838 !(req->flags & REQ_F_IO_DRAIN)))
5841 seq = io_get_sequence(req);
5842 /* Still a chance to pass the sequence check */
5843 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5846 ret = io_req_defer_prep(req);
5849 io_prep_async_link(req);
5850 de = kmalloc(sizeof(*de), GFP_KERNEL);
5854 spin_lock_irq(&ctx->completion_lock);
5855 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5856 spin_unlock_irq(&ctx->completion_lock);
5858 io_queue_async_work(req);
5859 return -EIOCBQUEUED;
5862 trace_io_uring_defer(ctx, req, req->user_data);
5865 list_add_tail(&de->list, &ctx->defer_list);
5866 spin_unlock_irq(&ctx->completion_lock);
5867 return -EIOCBQUEUED;
5870 static void __io_clean_op(struct io_kiocb *req)
5872 if (req->flags & REQ_F_BUFFER_SELECTED) {
5873 switch (req->opcode) {
5874 case IORING_OP_READV:
5875 case IORING_OP_READ_FIXED:
5876 case IORING_OP_READ:
5877 kfree((void *)(unsigned long)req->rw.addr);
5879 case IORING_OP_RECVMSG:
5880 case IORING_OP_RECV:
5881 kfree(req->sr_msg.kbuf);
5884 req->flags &= ~REQ_F_BUFFER_SELECTED;
5887 if (req->flags & REQ_F_NEED_CLEANUP) {
5888 switch (req->opcode) {
5889 case IORING_OP_READV:
5890 case IORING_OP_READ_FIXED:
5891 case IORING_OP_READ:
5892 case IORING_OP_WRITEV:
5893 case IORING_OP_WRITE_FIXED:
5894 case IORING_OP_WRITE: {
5895 struct io_async_rw *io = req->async_data;
5897 kfree(io->free_iovec);
5900 case IORING_OP_RECVMSG:
5901 case IORING_OP_SENDMSG: {
5902 struct io_async_msghdr *io = req->async_data;
5904 kfree(io->free_iov);
5907 case IORING_OP_SPLICE:
5909 io_put_file(req, req->splice.file_in,
5910 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5912 case IORING_OP_OPENAT:
5913 case IORING_OP_OPENAT2:
5914 if (req->open.filename)
5915 putname(req->open.filename);
5917 case IORING_OP_RENAMEAT:
5918 putname(req->rename.oldpath);
5919 putname(req->rename.newpath);
5921 case IORING_OP_UNLINKAT:
5922 putname(req->unlink.filename);
5925 req->flags &= ~REQ_F_NEED_CLEANUP;
5929 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5931 struct io_ring_ctx *ctx = req->ctx;
5932 const struct cred *creds = NULL;
5935 if (req->work.creds && req->work.creds != current_cred())
5936 creds = override_creds(req->work.creds);
5938 switch (req->opcode) {
5940 ret = io_nop(req, issue_flags);
5942 case IORING_OP_READV:
5943 case IORING_OP_READ_FIXED:
5944 case IORING_OP_READ:
5945 ret = io_read(req, issue_flags);
5947 case IORING_OP_WRITEV:
5948 case IORING_OP_WRITE_FIXED:
5949 case IORING_OP_WRITE:
5950 ret = io_write(req, issue_flags);
5952 case IORING_OP_FSYNC:
5953 ret = io_fsync(req, issue_flags);
5955 case IORING_OP_POLL_ADD:
5956 ret = io_poll_add(req, issue_flags);
5958 case IORING_OP_POLL_REMOVE:
5959 ret = io_poll_remove(req, issue_flags);
5961 case IORING_OP_SYNC_FILE_RANGE:
5962 ret = io_sync_file_range(req, issue_flags);
5964 case IORING_OP_SENDMSG:
5965 ret = io_sendmsg(req, issue_flags);
5967 case IORING_OP_SEND:
5968 ret = io_send(req, issue_flags);
5970 case IORING_OP_RECVMSG:
5971 ret = io_recvmsg(req, issue_flags);
5973 case IORING_OP_RECV:
5974 ret = io_recv(req, issue_flags);
5976 case IORING_OP_TIMEOUT:
5977 ret = io_timeout(req, issue_flags);
5979 case IORING_OP_TIMEOUT_REMOVE:
5980 ret = io_timeout_remove(req, issue_flags);
5982 case IORING_OP_ACCEPT:
5983 ret = io_accept(req, issue_flags);
5985 case IORING_OP_CONNECT:
5986 ret = io_connect(req, issue_flags);
5988 case IORING_OP_ASYNC_CANCEL:
5989 ret = io_async_cancel(req, issue_flags);
5991 case IORING_OP_FALLOCATE:
5992 ret = io_fallocate(req, issue_flags);
5994 case IORING_OP_OPENAT:
5995 ret = io_openat(req, issue_flags);
5997 case IORING_OP_CLOSE:
5998 ret = io_close(req, issue_flags);
6000 case IORING_OP_FILES_UPDATE:
6001 ret = io_files_update(req, issue_flags);
6003 case IORING_OP_STATX:
6004 ret = io_statx(req, issue_flags);
6006 case IORING_OP_FADVISE:
6007 ret = io_fadvise(req, issue_flags);
6009 case IORING_OP_MADVISE:
6010 ret = io_madvise(req, issue_flags);
6012 case IORING_OP_OPENAT2:
6013 ret = io_openat2(req, issue_flags);
6015 case IORING_OP_EPOLL_CTL:
6016 ret = io_epoll_ctl(req, issue_flags);
6018 case IORING_OP_SPLICE:
6019 ret = io_splice(req, issue_flags);
6021 case IORING_OP_PROVIDE_BUFFERS:
6022 ret = io_provide_buffers(req, issue_flags);
6024 case IORING_OP_REMOVE_BUFFERS:
6025 ret = io_remove_buffers(req, issue_flags);
6028 ret = io_tee(req, issue_flags);
6030 case IORING_OP_SHUTDOWN:
6031 ret = io_shutdown(req, issue_flags);
6033 case IORING_OP_RENAMEAT:
6034 ret = io_renameat(req, issue_flags);
6036 case IORING_OP_UNLINKAT:
6037 ret = io_unlinkat(req, issue_flags);
6045 revert_creds(creds);
6050 /* If the op doesn't have a file, we're not polling for it */
6051 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6052 const bool in_async = io_wq_current_is_worker();
6054 /* workqueue context doesn't hold uring_lock, grab it now */
6056 mutex_lock(&ctx->uring_lock);
6058 io_iopoll_req_issued(req, in_async);
6061 mutex_unlock(&ctx->uring_lock);
6067 static void io_wq_submit_work(struct io_wq_work *work)
6069 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6070 struct io_kiocb *timeout;
6073 timeout = io_prep_linked_timeout(req);
6075 io_queue_linked_timeout(timeout);
6077 if (work->flags & IO_WQ_WORK_CANCEL)
6082 ret = io_issue_sqe(req, 0);
6084 * We can get EAGAIN for polled IO even though we're
6085 * forcing a sync submission from here, since we can't
6086 * wait for request slots on the block side.
6094 /* avoid locking problems by failing it from a clean context */
6096 /* io-wq is going to take one down */
6097 refcount_inc(&req->refs);
6098 io_req_task_queue_fail(req, ret);
6102 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6105 struct fixed_rsrc_table *table;
6107 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6108 return table->files[index & IORING_FILE_TABLE_MASK];
6111 static struct file *io_file_get(struct io_submit_state *state,
6112 struct io_kiocb *req, int fd, bool fixed)
6114 struct io_ring_ctx *ctx = req->ctx;
6118 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6120 fd = array_index_nospec(fd, ctx->nr_user_files);
6121 file = io_file_from_index(ctx, fd);
6122 io_set_resource_node(req);
6124 trace_io_uring_file_get(ctx, fd);
6125 file = __io_file_get(state, fd);
6128 if (file && unlikely(file->f_op == &io_uring_fops))
6129 io_req_track_inflight(req);
6133 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6135 struct io_timeout_data *data = container_of(timer,
6136 struct io_timeout_data, timer);
6137 struct io_kiocb *prev, *req = data->req;
6138 struct io_ring_ctx *ctx = req->ctx;
6139 unsigned long flags;
6141 spin_lock_irqsave(&ctx->completion_lock, flags);
6142 prev = req->timeout.head;
6143 req->timeout.head = NULL;
6146 * We don't expect the list to be empty, that will only happen if we
6147 * race with the completion of the linked work.
6149 if (prev && refcount_inc_not_zero(&prev->refs))
6150 io_remove_next_linked(prev);
6153 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6156 req_set_fail_links(prev);
6157 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6158 io_put_req_deferred(prev, 1);
6160 io_req_complete_post(req, -ETIME, 0);
6161 io_put_req_deferred(req, 1);
6163 return HRTIMER_NORESTART;
6166 static void __io_queue_linked_timeout(struct io_kiocb *req)
6169 * If the back reference is NULL, then our linked request finished
6170 * before we got a chance to setup the timer
6172 if (req->timeout.head) {
6173 struct io_timeout_data *data = req->async_data;
6175 data->timer.function = io_link_timeout_fn;
6176 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6181 static void io_queue_linked_timeout(struct io_kiocb *req)
6183 struct io_ring_ctx *ctx = req->ctx;
6185 spin_lock_irq(&ctx->completion_lock);
6186 __io_queue_linked_timeout(req);
6187 spin_unlock_irq(&ctx->completion_lock);
6189 /* drop submission reference */
6193 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6195 struct io_kiocb *nxt = req->link;
6197 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6198 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6201 nxt->timeout.head = req;
6202 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6203 req->flags |= REQ_F_LINK_TIMEOUT;
6207 static void __io_queue_sqe(struct io_kiocb *req)
6209 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6212 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6215 * We async punt it if the file wasn't marked NOWAIT, or if the file
6216 * doesn't support non-blocking read/write attempts
6218 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6219 if (!io_arm_poll_handler(req)) {
6221 * Queued up for async execution, worker will release
6222 * submit reference when the iocb is actually submitted.
6224 io_queue_async_work(req);
6226 } else if (likely(!ret)) {
6227 /* drop submission reference */
6228 if (req->flags & REQ_F_COMPLETE_INLINE) {
6229 struct io_ring_ctx *ctx = req->ctx;
6230 struct io_comp_state *cs = &ctx->submit_state.comp;
6232 cs->reqs[cs->nr++] = req;
6233 if (cs->nr == ARRAY_SIZE(cs->reqs))
6234 io_submit_flush_completions(cs, ctx);
6239 req_set_fail_links(req);
6241 io_req_complete(req, ret);
6244 io_queue_linked_timeout(linked_timeout);
6247 static void io_queue_sqe(struct io_kiocb *req)
6251 ret = io_req_defer(req);
6253 if (ret != -EIOCBQUEUED) {
6255 req_set_fail_links(req);
6257 io_req_complete(req, ret);
6259 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6260 ret = io_req_defer_prep(req);
6263 io_queue_async_work(req);
6265 __io_queue_sqe(req);
6270 * Check SQE restrictions (opcode and flags).
6272 * Returns 'true' if SQE is allowed, 'false' otherwise.
6274 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6275 struct io_kiocb *req,
6276 unsigned int sqe_flags)
6278 if (!ctx->restricted)
6281 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6284 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6285 ctx->restrictions.sqe_flags_required)
6288 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6289 ctx->restrictions.sqe_flags_required))
6295 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6296 const struct io_uring_sqe *sqe)
6298 struct io_submit_state *state;
6299 unsigned int sqe_flags;
6300 int personality, ret = 0;
6302 req->opcode = READ_ONCE(sqe->opcode);
6303 /* same numerical values with corresponding REQ_F_*, safe to copy */
6304 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6305 req->user_data = READ_ONCE(sqe->user_data);
6306 req->async_data = NULL;
6310 req->fixed_rsrc_refs = NULL;
6311 /* one is dropped after submission, the other at completion */
6312 refcount_set(&req->refs, 2);
6313 req->task = current;
6316 /* enforce forwards compatibility on users */
6317 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6322 if (unlikely(req->opcode >= IORING_OP_LAST))
6325 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6328 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6329 !io_op_defs[req->opcode].buffer_select)
6332 req->work.list.next = NULL;
6333 personality = READ_ONCE(sqe->personality);
6335 req->work.creds = idr_find(&ctx->personality_idr, personality);
6336 if (!req->work.creds)
6338 get_cred(req->work.creds);
6340 req->work.creds = NULL;
6342 req->work.flags = 0;
6343 state = &ctx->submit_state;
6346 * Plug now if we have more than 1 IO left after this, and the target
6347 * is potentially a read/write to block based storage.
6349 if (!state->plug_started && state->ios_left > 1 &&
6350 io_op_defs[req->opcode].plug) {
6351 blk_start_plug(&state->plug);
6352 state->plug_started = true;
6355 if (io_op_defs[req->opcode].needs_file) {
6356 bool fixed = req->flags & REQ_F_FIXED_FILE;
6358 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6359 if (unlikely(!req->file))
6367 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6368 const struct io_uring_sqe *sqe)
6370 struct io_submit_link *link = &ctx->submit_state.link;
6373 ret = io_init_req(ctx, req, sqe);
6374 if (unlikely(ret)) {
6377 io_req_complete(req, ret);
6379 /* fail even hard links since we don't submit */
6380 link->head->flags |= REQ_F_FAIL_LINK;
6381 io_put_req(link->head);
6382 io_req_complete(link->head, -ECANCELED);
6387 ret = io_req_prep(req, sqe);
6391 /* don't need @sqe from now on */
6392 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6393 true, ctx->flags & IORING_SETUP_SQPOLL);
6396 * If we already have a head request, queue this one for async
6397 * submittal once the head completes. If we don't have a head but
6398 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6399 * submitted sync once the chain is complete. If none of those
6400 * conditions are true (normal request), then just queue it.
6403 struct io_kiocb *head = link->head;
6406 * Taking sequential execution of a link, draining both sides
6407 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6408 * requests in the link. So, it drains the head and the
6409 * next after the link request. The last one is done via
6410 * drain_next flag to persist the effect across calls.
6412 if (req->flags & REQ_F_IO_DRAIN) {
6413 head->flags |= REQ_F_IO_DRAIN;
6414 ctx->drain_next = 1;
6416 ret = io_req_defer_prep(req);
6419 trace_io_uring_link(ctx, req, head);
6420 link->last->link = req;
6423 /* last request of a link, enqueue the link */
6424 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6429 if (unlikely(ctx->drain_next)) {
6430 req->flags |= REQ_F_IO_DRAIN;
6431 ctx->drain_next = 0;
6433 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6445 * Batched submission is done, ensure local IO is flushed out.
6447 static void io_submit_state_end(struct io_submit_state *state,
6448 struct io_ring_ctx *ctx)
6450 if (state->link.head)
6451 io_queue_sqe(state->link.head);
6453 io_submit_flush_completions(&state->comp, ctx);
6454 if (state->plug_started)
6455 blk_finish_plug(&state->plug);
6456 io_state_file_put(state);
6460 * Start submission side cache.
6462 static void io_submit_state_start(struct io_submit_state *state,
6463 unsigned int max_ios)
6465 state->plug_started = false;
6466 state->ios_left = max_ios;
6467 /* set only head, no need to init link_last in advance */
6468 state->link.head = NULL;
6471 static void io_commit_sqring(struct io_ring_ctx *ctx)
6473 struct io_rings *rings = ctx->rings;
6476 * Ensure any loads from the SQEs are done at this point,
6477 * since once we write the new head, the application could
6478 * write new data to them.
6480 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6484 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6485 * that is mapped by userspace. This means that care needs to be taken to
6486 * ensure that reads are stable, as we cannot rely on userspace always
6487 * being a good citizen. If members of the sqe are validated and then later
6488 * used, it's important that those reads are done through READ_ONCE() to
6489 * prevent a re-load down the line.
6491 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6493 u32 *sq_array = ctx->sq_array;
6497 * The cached sq head (or cq tail) serves two purposes:
6499 * 1) allows us to batch the cost of updating the user visible
6501 * 2) allows the kernel side to track the head on its own, even
6502 * though the application is the one updating it.
6504 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6505 if (likely(head < ctx->sq_entries))
6506 return &ctx->sq_sqes[head];
6508 /* drop invalid entries */
6509 ctx->cached_sq_dropped++;
6510 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6514 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6518 /* if we have a backlog and couldn't flush it all, return BUSY */
6519 if (test_bit(0, &ctx->sq_check_overflow)) {
6520 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6524 /* make sure SQ entry isn't read before tail */
6525 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6527 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6530 percpu_counter_add(¤t->io_uring->inflight, nr);
6531 refcount_add(nr, ¤t->usage);
6532 io_submit_state_start(&ctx->submit_state, nr);
6534 while (submitted < nr) {
6535 const struct io_uring_sqe *sqe;
6536 struct io_kiocb *req;
6538 req = io_alloc_req(ctx);
6539 if (unlikely(!req)) {
6541 submitted = -EAGAIN;
6544 sqe = io_get_sqe(ctx);
6545 if (unlikely(!sqe)) {
6546 kmem_cache_free(req_cachep, req);
6549 /* will complete beyond this point, count as submitted */
6551 if (io_submit_sqe(ctx, req, sqe))
6555 if (unlikely(submitted != nr)) {
6556 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6557 struct io_uring_task *tctx = current->io_uring;
6558 int unused = nr - ref_used;
6560 percpu_ref_put_many(&ctx->refs, unused);
6561 percpu_counter_sub(&tctx->inflight, unused);
6562 put_task_struct_many(current, unused);
6565 io_submit_state_end(&ctx->submit_state, ctx);
6566 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6567 io_commit_sqring(ctx);
6572 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6574 /* Tell userspace we may need a wakeup call */
6575 spin_lock_irq(&ctx->completion_lock);
6576 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6577 spin_unlock_irq(&ctx->completion_lock);
6580 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6582 spin_lock_irq(&ctx->completion_lock);
6583 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6584 spin_unlock_irq(&ctx->completion_lock);
6587 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6589 unsigned int to_submit;
6592 to_submit = io_sqring_entries(ctx);
6593 /* if we're handling multiple rings, cap submit size for fairness */
6594 if (cap_entries && to_submit > 8)
6597 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6598 unsigned nr_events = 0;
6600 mutex_lock(&ctx->uring_lock);
6601 if (!list_empty(&ctx->iopoll_list))
6602 io_do_iopoll(ctx, &nr_events, 0);
6604 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6605 ret = io_submit_sqes(ctx, to_submit);
6606 mutex_unlock(&ctx->uring_lock);
6609 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6610 wake_up(&ctx->sqo_sq_wait);
6615 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6617 struct io_ring_ctx *ctx;
6618 unsigned sq_thread_idle = 0;
6620 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6621 if (sq_thread_idle < ctx->sq_thread_idle)
6622 sq_thread_idle = ctx->sq_thread_idle;
6625 sqd->sq_thread_idle = sq_thread_idle;
6628 static void io_sqd_init_new(struct io_sq_data *sqd)
6630 struct io_ring_ctx *ctx;
6632 while (!list_empty(&sqd->ctx_new_list)) {
6633 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6634 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6635 complete(&ctx->sq_thread_comp);
6638 io_sqd_update_thread_idle(sqd);
6641 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6643 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6646 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6648 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6651 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6655 * TASK_PARKED is a special state; we must serialize against
6656 * possible pending wakeups to avoid store-store collisions on
6659 * Such a collision might possibly result in the task state
6660 * changin from TASK_PARKED and us failing the
6661 * wait_task_inactive() in kthread_park().
6663 set_special_state(TASK_PARKED);
6664 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6668 * Thread is going to call schedule(), do not preempt it,
6669 * or the caller of kthread_park() may spend more time in
6670 * wait_task_inactive().
6673 complete(&sqd->parked);
6674 schedule_preempt_disabled();
6677 __set_current_state(TASK_RUNNING);
6680 static int io_sq_thread(void *data)
6682 struct io_sq_data *sqd = data;
6683 struct io_ring_ctx *ctx;
6684 unsigned long timeout = 0;
6685 char buf[TASK_COMM_LEN];
6688 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6689 set_task_comm(current, buf);
6690 current->pf_io_worker = NULL;
6692 if (sqd->sq_cpu != -1)
6693 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6695 set_cpus_allowed_ptr(current, cpu_online_mask);
6696 current->flags |= PF_NO_SETAFFINITY;
6698 wait_for_completion(&sqd->startup);
6700 while (!io_sq_thread_should_stop(sqd)) {
6702 bool cap_entries, sqt_spin, needs_sched;
6705 * Any changes to the sqd lists are synchronized through the
6706 * thread parking. This synchronizes the thread vs users,
6707 * the users are synchronized on the sqd->ctx_lock.
6709 if (io_sq_thread_should_park(sqd)) {
6710 io_sq_thread_parkme(sqd);
6713 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6714 io_sqd_init_new(sqd);
6715 timeout = jiffies + sqd->sq_thread_idle;
6717 if (fatal_signal_pending(current))
6720 cap_entries = !list_is_singular(&sqd->ctx_list);
6721 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6722 const struct cred *creds = NULL;
6724 if (ctx->sq_creds != current_cred())
6725 creds = override_creds(ctx->sq_creds);
6726 ret = __io_sq_thread(ctx, cap_entries);
6728 revert_creds(creds);
6729 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6733 if (sqt_spin || !time_after(jiffies, timeout)) {
6737 timeout = jiffies + sqd->sq_thread_idle;
6742 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6743 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6744 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6745 !list_empty_careful(&ctx->iopoll_list)) {
6746 needs_sched = false;
6749 if (io_sqring_entries(ctx)) {
6750 needs_sched = false;
6755 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6756 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6757 io_ring_set_wakeup_flag(ctx);
6761 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6762 io_ring_clear_wakeup_flag(ctx);
6765 finish_wait(&sqd->wait, &wait);
6766 timeout = jiffies + sqd->sq_thread_idle;
6769 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6770 io_uring_cancel_sqpoll(ctx);
6775 * Ensure that we park properly if racing with someone trying to park
6776 * while we're exiting. If we fail to grab the lock, check park and
6777 * park if necessary. The ordering with the park bit and the lock
6778 * ensures that we catch this reliably.
6780 if (!mutex_trylock(&sqd->lock)) {
6781 if (io_sq_thread_should_park(sqd))
6782 io_sq_thread_parkme(sqd);
6783 mutex_lock(&sqd->lock);
6787 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6788 io_ring_set_wakeup_flag(ctx);
6791 complete(&sqd->exited);
6792 mutex_unlock(&sqd->lock);
6796 struct io_wait_queue {
6797 struct wait_queue_entry wq;
6798 struct io_ring_ctx *ctx;
6800 unsigned nr_timeouts;
6803 static inline bool io_should_wake(struct io_wait_queue *iowq)
6805 struct io_ring_ctx *ctx = iowq->ctx;
6808 * Wake up if we have enough events, or if a timeout occurred since we
6809 * started waiting. For timeouts, we always want to return to userspace,
6810 * regardless of event count.
6812 return io_cqring_events(ctx) >= iowq->to_wait ||
6813 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6816 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6817 int wake_flags, void *key)
6819 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6823 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6824 * the task, and the next invocation will do it.
6826 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6827 return autoremove_wake_function(curr, mode, wake_flags, key);
6831 static int io_run_task_work_sig(void)
6833 if (io_run_task_work())
6835 if (!signal_pending(current))
6837 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6838 return -ERESTARTSYS;
6842 /* when returns >0, the caller should retry */
6843 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6844 struct io_wait_queue *iowq,
6845 signed long *timeout)
6849 /* make sure we run task_work before checking for signals */
6850 ret = io_run_task_work_sig();
6851 if (ret || io_should_wake(iowq))
6853 /* let the caller flush overflows, retry */
6854 if (test_bit(0, &ctx->cq_check_overflow))
6857 *timeout = schedule_timeout(*timeout);
6858 return !*timeout ? -ETIME : 1;
6862 * Wait until events become available, if we don't already have some. The
6863 * application must reap them itself, as they reside on the shared cq ring.
6865 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6866 const sigset_t __user *sig, size_t sigsz,
6867 struct __kernel_timespec __user *uts)
6869 struct io_wait_queue iowq = {
6872 .func = io_wake_function,
6873 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6876 .to_wait = min_events,
6878 struct io_rings *rings = ctx->rings;
6879 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6883 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6884 if (io_cqring_events(ctx) >= min_events)
6886 if (!io_run_task_work())
6891 #ifdef CONFIG_COMPAT
6892 if (in_compat_syscall())
6893 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6897 ret = set_user_sigmask(sig, sigsz);
6904 struct timespec64 ts;
6906 if (get_timespec64(&ts, uts))
6908 timeout = timespec64_to_jiffies(&ts);
6911 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6912 trace_io_uring_cqring_wait(ctx, min_events);
6914 /* if we can't even flush overflow, don't wait for more */
6915 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6919 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6920 TASK_INTERRUPTIBLE);
6921 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6922 finish_wait(&ctx->wait, &iowq.wq);
6926 restore_saved_sigmask_unless(ret == -EINTR);
6928 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6931 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6933 #if defined(CONFIG_UNIX)
6934 if (ctx->ring_sock) {
6935 struct sock *sock = ctx->ring_sock->sk;
6936 struct sk_buff *skb;
6938 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6944 for (i = 0; i < ctx->nr_user_files; i++) {
6947 file = io_file_from_index(ctx, i);
6954 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6956 struct fixed_rsrc_data *data;
6958 data = container_of(ref, struct fixed_rsrc_data, refs);
6959 complete(&data->done);
6962 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6964 spin_lock_bh(&ctx->rsrc_ref_lock);
6967 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6969 spin_unlock_bh(&ctx->rsrc_ref_lock);
6972 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6973 struct fixed_rsrc_data *rsrc_data,
6974 struct fixed_rsrc_ref_node *ref_node)
6976 io_rsrc_ref_lock(ctx);
6977 rsrc_data->node = ref_node;
6978 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6979 io_rsrc_ref_unlock(ctx);
6980 percpu_ref_get(&rsrc_data->refs);
6983 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6985 struct fixed_rsrc_ref_node *ref_node = NULL;
6987 io_rsrc_ref_lock(ctx);
6988 ref_node = data->node;
6990 io_rsrc_ref_unlock(ctx);
6992 percpu_ref_kill(&ref_node->refs);
6995 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6996 struct io_ring_ctx *ctx,
6997 void (*rsrc_put)(struct io_ring_ctx *ctx,
6998 struct io_rsrc_put *prsrc))
7000 struct fixed_rsrc_ref_node *backup_node;
7006 data->quiesce = true;
7009 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7012 backup_node->rsrc_data = data;
7013 backup_node->rsrc_put = rsrc_put;
7015 io_sqe_rsrc_kill_node(ctx, data);
7016 percpu_ref_kill(&data->refs);
7017 flush_delayed_work(&ctx->rsrc_put_work);
7019 ret = wait_for_completion_interruptible(&data->done);
7023 percpu_ref_resurrect(&data->refs);
7024 io_sqe_rsrc_set_node(ctx, data, backup_node);
7026 reinit_completion(&data->done);
7027 mutex_unlock(&ctx->uring_lock);
7028 ret = io_run_task_work_sig();
7029 mutex_lock(&ctx->uring_lock);
7031 data->quiesce = false;
7034 destroy_fixed_rsrc_ref_node(backup_node);
7038 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7040 struct fixed_rsrc_data *data;
7042 data = kzalloc(sizeof(*data), GFP_KERNEL);
7046 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7047 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7052 init_completion(&data->done);
7056 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7058 percpu_ref_exit(&data->refs);
7063 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7065 struct fixed_rsrc_data *data = ctx->file_data;
7066 unsigned nr_tables, i;
7070 * percpu_ref_is_dying() is to stop parallel files unregister
7071 * Since we possibly drop uring lock later in this function to
7074 if (!data || percpu_ref_is_dying(&data->refs))
7076 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7080 __io_sqe_files_unregister(ctx);
7081 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7082 for (i = 0; i < nr_tables; i++)
7083 kfree(data->table[i].files);
7084 free_fixed_rsrc_data(data);
7085 ctx->file_data = NULL;
7086 ctx->nr_user_files = 0;
7090 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7091 __releases(&sqd->lock)
7093 if (sqd->thread == current)
7095 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7097 wake_up_state(sqd->thread, TASK_PARKED);
7098 mutex_unlock(&sqd->lock);
7101 static void io_sq_thread_park(struct io_sq_data *sqd)
7102 __acquires(&sqd->lock)
7104 if (sqd->thread == current)
7106 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7107 mutex_lock(&sqd->lock);
7109 wake_up_process(sqd->thread);
7110 wait_for_completion(&sqd->parked);
7114 static void io_sq_thread_stop(struct io_sq_data *sqd)
7116 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7118 mutex_lock(&sqd->lock);
7120 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7121 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7122 wake_up_process(sqd->thread);
7123 mutex_unlock(&sqd->lock);
7124 wait_for_completion(&sqd->exited);
7125 WARN_ON_ONCE(sqd->thread);
7127 mutex_unlock(&sqd->lock);
7131 static void io_put_sq_data(struct io_sq_data *sqd)
7133 if (refcount_dec_and_test(&sqd->refs)) {
7134 io_sq_thread_stop(sqd);
7139 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7141 struct io_sq_data *sqd = ctx->sq_data;
7144 complete(&sqd->startup);
7146 wait_for_completion(&ctx->sq_thread_comp);
7147 io_sq_thread_park(sqd);
7150 mutex_lock(&sqd->ctx_lock);
7151 list_del(&ctx->sqd_list);
7152 io_sqd_update_thread_idle(sqd);
7153 mutex_unlock(&sqd->ctx_lock);
7156 io_sq_thread_unpark(sqd);
7158 io_put_sq_data(sqd);
7159 ctx->sq_data = NULL;
7161 put_cred(ctx->sq_creds);
7165 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7167 struct io_ring_ctx *ctx_attach;
7168 struct io_sq_data *sqd;
7171 f = fdget(p->wq_fd);
7173 return ERR_PTR(-ENXIO);
7174 if (f.file->f_op != &io_uring_fops) {
7176 return ERR_PTR(-EINVAL);
7179 ctx_attach = f.file->private_data;
7180 sqd = ctx_attach->sq_data;
7183 return ERR_PTR(-EINVAL);
7186 refcount_inc(&sqd->refs);
7191 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7193 struct io_sq_data *sqd;
7195 if (p->flags & IORING_SETUP_ATTACH_WQ)
7196 return io_attach_sq_data(p);
7198 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7200 return ERR_PTR(-ENOMEM);
7202 refcount_set(&sqd->refs, 1);
7203 INIT_LIST_HEAD(&sqd->ctx_list);
7204 INIT_LIST_HEAD(&sqd->ctx_new_list);
7205 mutex_init(&sqd->ctx_lock);
7206 mutex_init(&sqd->lock);
7207 init_waitqueue_head(&sqd->wait);
7208 init_completion(&sqd->startup);
7209 init_completion(&sqd->parked);
7210 init_completion(&sqd->exited);
7214 #if defined(CONFIG_UNIX)
7216 * Ensure the UNIX gc is aware of our file set, so we are certain that
7217 * the io_uring can be safely unregistered on process exit, even if we have
7218 * loops in the file referencing.
7220 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7222 struct sock *sk = ctx->ring_sock->sk;
7223 struct scm_fp_list *fpl;
7224 struct sk_buff *skb;
7227 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7231 skb = alloc_skb(0, GFP_KERNEL);
7240 fpl->user = get_uid(current_user());
7241 for (i = 0; i < nr; i++) {
7242 struct file *file = io_file_from_index(ctx, i + offset);
7246 fpl->fp[nr_files] = get_file(file);
7247 unix_inflight(fpl->user, fpl->fp[nr_files]);
7252 fpl->max = SCM_MAX_FD;
7253 fpl->count = nr_files;
7254 UNIXCB(skb).fp = fpl;
7255 skb->destructor = unix_destruct_scm;
7256 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7257 skb_queue_head(&sk->sk_receive_queue, skb);
7259 for (i = 0; i < nr_files; i++)
7270 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7271 * causes regular reference counting to break down. We rely on the UNIX
7272 * garbage collection to take care of this problem for us.
7274 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7276 unsigned left, total;
7280 left = ctx->nr_user_files;
7282 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7284 ret = __io_sqe_files_scm(ctx, this_files, total);
7288 total += this_files;
7294 while (total < ctx->nr_user_files) {
7295 struct file *file = io_file_from_index(ctx, total);
7305 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7311 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7312 unsigned nr_tables, unsigned nr_files)
7316 for (i = 0; i < nr_tables; i++) {
7317 struct fixed_rsrc_table *table = &file_data->table[i];
7318 unsigned this_files;
7320 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7321 table->files = kcalloc(this_files, sizeof(struct file *),
7325 nr_files -= this_files;
7331 for (i = 0; i < nr_tables; i++) {
7332 struct fixed_rsrc_table *table = &file_data->table[i];
7333 kfree(table->files);
7338 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7340 struct file *file = prsrc->file;
7341 #if defined(CONFIG_UNIX)
7342 struct sock *sock = ctx->ring_sock->sk;
7343 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7344 struct sk_buff *skb;
7347 __skb_queue_head_init(&list);
7350 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7351 * remove this entry and rearrange the file array.
7353 skb = skb_dequeue(head);
7355 struct scm_fp_list *fp;
7357 fp = UNIXCB(skb).fp;
7358 for (i = 0; i < fp->count; i++) {
7361 if (fp->fp[i] != file)
7364 unix_notinflight(fp->user, fp->fp[i]);
7365 left = fp->count - 1 - i;
7367 memmove(&fp->fp[i], &fp->fp[i + 1],
7368 left * sizeof(struct file *));
7375 __skb_queue_tail(&list, skb);
7385 __skb_queue_tail(&list, skb);
7387 skb = skb_dequeue(head);
7390 if (skb_peek(&list)) {
7391 spin_lock_irq(&head->lock);
7392 while ((skb = __skb_dequeue(&list)) != NULL)
7393 __skb_queue_tail(head, skb);
7394 spin_unlock_irq(&head->lock);
7401 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7403 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7404 struct io_ring_ctx *ctx = rsrc_data->ctx;
7405 struct io_rsrc_put *prsrc, *tmp;
7407 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7408 list_del(&prsrc->list);
7409 ref_node->rsrc_put(ctx, prsrc);
7413 percpu_ref_exit(&ref_node->refs);
7415 percpu_ref_put(&rsrc_data->refs);
7418 static void io_rsrc_put_work(struct work_struct *work)
7420 struct io_ring_ctx *ctx;
7421 struct llist_node *node;
7423 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7424 node = llist_del_all(&ctx->rsrc_put_llist);
7427 struct fixed_rsrc_ref_node *ref_node;
7428 struct llist_node *next = node->next;
7430 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7431 __io_rsrc_put_work(ref_node);
7436 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7439 struct fixed_rsrc_table *table;
7441 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7442 return &table->files[i & IORING_FILE_TABLE_MASK];
7445 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7447 struct fixed_rsrc_ref_node *ref_node;
7448 struct fixed_rsrc_data *data;
7449 struct io_ring_ctx *ctx;
7450 bool first_add = false;
7453 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7454 data = ref_node->rsrc_data;
7457 io_rsrc_ref_lock(ctx);
7458 ref_node->done = true;
7460 while (!list_empty(&ctx->rsrc_ref_list)) {
7461 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7462 struct fixed_rsrc_ref_node, node);
7463 /* recycle ref nodes in order */
7464 if (!ref_node->done)
7466 list_del(&ref_node->node);
7467 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7469 io_rsrc_ref_unlock(ctx);
7471 if (percpu_ref_is_dying(&data->refs))
7475 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7477 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7480 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7481 struct io_ring_ctx *ctx)
7483 struct fixed_rsrc_ref_node *ref_node;
7485 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7489 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7494 INIT_LIST_HEAD(&ref_node->node);
7495 INIT_LIST_HEAD(&ref_node->rsrc_list);
7496 ref_node->done = false;
7500 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7501 struct fixed_rsrc_ref_node *ref_node)
7503 ref_node->rsrc_data = ctx->file_data;
7504 ref_node->rsrc_put = io_ring_file_put;
7507 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7509 percpu_ref_exit(&ref_node->refs);
7514 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7517 __s32 __user *fds = (__s32 __user *) arg;
7518 unsigned nr_tables, i;
7520 int fd, ret = -ENOMEM;
7521 struct fixed_rsrc_ref_node *ref_node;
7522 struct fixed_rsrc_data *file_data;
7528 if (nr_args > IORING_MAX_FIXED_FILES)
7531 file_data = alloc_fixed_rsrc_data(ctx);
7534 ctx->file_data = file_data;
7536 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7537 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7539 if (!file_data->table)
7542 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7545 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7546 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7550 /* allow sparse sets */
7560 * Don't allow io_uring instances to be registered. If UNIX
7561 * isn't enabled, then this causes a reference cycle and this
7562 * instance can never get freed. If UNIX is enabled we'll
7563 * handle it just fine, but there's still no point in allowing
7564 * a ring fd as it doesn't support regular read/write anyway.
7566 if (file->f_op == &io_uring_fops) {
7570 *io_fixed_file_slot(file_data, i) = file;
7573 ret = io_sqe_files_scm(ctx);
7575 io_sqe_files_unregister(ctx);
7579 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7581 io_sqe_files_unregister(ctx);
7584 init_fixed_file_ref_node(ctx, ref_node);
7586 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7589 for (i = 0; i < ctx->nr_user_files; i++) {
7590 file = io_file_from_index(ctx, i);
7594 for (i = 0; i < nr_tables; i++)
7595 kfree(file_data->table[i].files);
7596 ctx->nr_user_files = 0;
7598 free_fixed_rsrc_data(ctx->file_data);
7599 ctx->file_data = NULL;
7603 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7606 #if defined(CONFIG_UNIX)
7607 struct sock *sock = ctx->ring_sock->sk;
7608 struct sk_buff_head *head = &sock->sk_receive_queue;
7609 struct sk_buff *skb;
7612 * See if we can merge this file into an existing skb SCM_RIGHTS
7613 * file set. If there's no room, fall back to allocating a new skb
7614 * and filling it in.
7616 spin_lock_irq(&head->lock);
7617 skb = skb_peek(head);
7619 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7621 if (fpl->count < SCM_MAX_FD) {
7622 __skb_unlink(skb, head);
7623 spin_unlock_irq(&head->lock);
7624 fpl->fp[fpl->count] = get_file(file);
7625 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7627 spin_lock_irq(&head->lock);
7628 __skb_queue_head(head, skb);
7633 spin_unlock_irq(&head->lock);
7640 return __io_sqe_files_scm(ctx, 1, index);
7646 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7648 struct io_rsrc_put *prsrc;
7649 struct fixed_rsrc_ref_node *ref_node = data->node;
7651 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7656 list_add(&prsrc->list, &ref_node->rsrc_list);
7661 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7664 return io_queue_rsrc_removal(data, (void *)file);
7667 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7668 struct io_uring_rsrc_update *up,
7671 struct fixed_rsrc_data *data = ctx->file_data;
7672 struct fixed_rsrc_ref_node *ref_node;
7673 struct file *file, **file_slot;
7677 bool needs_switch = false;
7679 if (check_add_overflow(up->offset, nr_args, &done))
7681 if (done > ctx->nr_user_files)
7684 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7687 init_fixed_file_ref_node(ctx, ref_node);
7689 fds = u64_to_user_ptr(up->data);
7690 for (done = 0; done < nr_args; done++) {
7692 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7696 if (fd == IORING_REGISTER_FILES_SKIP)
7699 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7700 file_slot = io_fixed_file_slot(ctx->file_data, i);
7703 err = io_queue_file_removal(data, *file_slot);
7707 needs_switch = true;
7716 * Don't allow io_uring instances to be registered. If
7717 * UNIX isn't enabled, then this causes a reference
7718 * cycle and this instance can never get freed. If UNIX
7719 * is enabled we'll handle it just fine, but there's
7720 * still no point in allowing a ring fd as it doesn't
7721 * support regular read/write anyway.
7723 if (file->f_op == &io_uring_fops) {
7729 err = io_sqe_file_register(ctx, file, i);
7739 percpu_ref_kill(&data->node->refs);
7740 io_sqe_rsrc_set_node(ctx, data, ref_node);
7742 destroy_fixed_rsrc_ref_node(ref_node);
7744 return done ? done : err;
7747 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7750 struct io_uring_rsrc_update up;
7752 if (!ctx->file_data)
7756 if (copy_from_user(&up, arg, sizeof(up)))
7761 return __io_sqe_files_update(ctx, &up, nr_args);
7764 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7766 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7768 req = io_put_req_find_next(req);
7769 return req ? &req->work : NULL;
7772 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7774 struct io_wq_hash *hash;
7775 struct io_wq_data data;
7776 unsigned int concurrency;
7778 hash = ctx->hash_map;
7780 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7782 return ERR_PTR(-ENOMEM);
7783 refcount_set(&hash->refs, 1);
7784 init_waitqueue_head(&hash->wait);
7785 ctx->hash_map = hash;
7789 data.free_work = io_free_work;
7790 data.do_work = io_wq_submit_work;
7792 /* Do QD, or 4 * CPUS, whatever is smallest */
7793 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7795 return io_wq_create(concurrency, &data);
7798 static int io_uring_alloc_task_context(struct task_struct *task,
7799 struct io_ring_ctx *ctx)
7801 struct io_uring_task *tctx;
7804 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7805 if (unlikely(!tctx))
7808 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7809 if (unlikely(ret)) {
7814 tctx->io_wq = io_init_wq_offload(ctx);
7815 if (IS_ERR(tctx->io_wq)) {
7816 ret = PTR_ERR(tctx->io_wq);
7817 percpu_counter_destroy(&tctx->inflight);
7823 init_waitqueue_head(&tctx->wait);
7825 atomic_set(&tctx->in_idle, 0);
7826 tctx->sqpoll = false;
7827 task->io_uring = tctx;
7828 spin_lock_init(&tctx->task_lock);
7829 INIT_WQ_LIST(&tctx->task_list);
7830 tctx->task_state = 0;
7831 init_task_work(&tctx->task_work, tctx_task_work);
7835 void __io_uring_free(struct task_struct *tsk)
7837 struct io_uring_task *tctx = tsk->io_uring;
7839 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7840 WARN_ON_ONCE(tctx->io_wq);
7842 percpu_counter_destroy(&tctx->inflight);
7844 tsk->io_uring = NULL;
7847 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7848 struct io_uring_params *p)
7852 /* Retain compatibility with failing for an invalid attach attempt */
7853 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7854 IORING_SETUP_ATTACH_WQ) {
7857 f = fdget(p->wq_fd);
7860 if (f.file->f_op != &io_uring_fops) {
7866 if (ctx->flags & IORING_SETUP_SQPOLL) {
7867 struct task_struct *tsk;
7868 struct io_sq_data *sqd;
7871 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7874 sqd = io_get_sq_data(p);
7880 ctx->sq_creds = get_current_cred();
7882 io_sq_thread_park(sqd);
7883 mutex_lock(&sqd->ctx_lock);
7884 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7885 mutex_unlock(&sqd->ctx_lock);
7886 io_sq_thread_unpark(sqd);
7888 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7889 if (!ctx->sq_thread_idle)
7890 ctx->sq_thread_idle = HZ;
7895 if (p->flags & IORING_SETUP_SQ_AFF) {
7896 int cpu = p->sq_thread_cpu;
7899 if (cpu >= nr_cpu_ids)
7901 if (!cpu_online(cpu))
7909 sqd->task_pid = current->pid;
7910 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7915 ret = io_uring_alloc_task_context(tsk, ctx);
7917 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7919 wake_up_new_task(tsk);
7922 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7923 /* Can't have SQ_AFF without SQPOLL */
7930 io_sq_thread_finish(ctx);
7934 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7936 struct io_sq_data *sqd = ctx->sq_data;
7938 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7939 if (ctx->flags & IORING_SETUP_SQPOLL)
7940 complete(&sqd->startup);
7943 static inline void __io_unaccount_mem(struct user_struct *user,
7944 unsigned long nr_pages)
7946 atomic_long_sub(nr_pages, &user->locked_vm);
7949 static inline int __io_account_mem(struct user_struct *user,
7950 unsigned long nr_pages)
7952 unsigned long page_limit, cur_pages, new_pages;
7954 /* Don't allow more pages than we can safely lock */
7955 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7958 cur_pages = atomic_long_read(&user->locked_vm);
7959 new_pages = cur_pages + nr_pages;
7960 if (new_pages > page_limit)
7962 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7963 new_pages) != cur_pages);
7968 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7971 __io_unaccount_mem(ctx->user, nr_pages);
7973 if (ctx->mm_account)
7974 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7977 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7982 ret = __io_account_mem(ctx->user, nr_pages);
7987 if (ctx->mm_account)
7988 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7993 static void io_mem_free(void *ptr)
8000 page = virt_to_head_page(ptr);
8001 if (put_page_testzero(page))
8002 free_compound_page(page);
8005 static void *io_mem_alloc(size_t size)
8007 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8008 __GFP_NORETRY | __GFP_ACCOUNT;
8010 return (void *) __get_free_pages(gfp_flags, get_order(size));
8013 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8016 struct io_rings *rings;
8017 size_t off, sq_array_size;
8019 off = struct_size(rings, cqes, cq_entries);
8020 if (off == SIZE_MAX)
8024 off = ALIGN(off, SMP_CACHE_BYTES);
8032 sq_array_size = array_size(sizeof(u32), sq_entries);
8033 if (sq_array_size == SIZE_MAX)
8036 if (check_add_overflow(off, sq_array_size, &off))
8042 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8046 if (!ctx->user_bufs)
8049 for (i = 0; i < ctx->nr_user_bufs; i++) {
8050 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8052 for (j = 0; j < imu->nr_bvecs; j++)
8053 unpin_user_page(imu->bvec[j].bv_page);
8055 if (imu->acct_pages)
8056 io_unaccount_mem(ctx, imu->acct_pages);
8061 kfree(ctx->user_bufs);
8062 ctx->user_bufs = NULL;
8063 ctx->nr_user_bufs = 0;
8067 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8068 void __user *arg, unsigned index)
8070 struct iovec __user *src;
8072 #ifdef CONFIG_COMPAT
8074 struct compat_iovec __user *ciovs;
8075 struct compat_iovec ciov;
8077 ciovs = (struct compat_iovec __user *) arg;
8078 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8081 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8082 dst->iov_len = ciov.iov_len;
8086 src = (struct iovec __user *) arg;
8087 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8093 * Not super efficient, but this is just a registration time. And we do cache
8094 * the last compound head, so generally we'll only do a full search if we don't
8097 * We check if the given compound head page has already been accounted, to
8098 * avoid double accounting it. This allows us to account the full size of the
8099 * page, not just the constituent pages of a huge page.
8101 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8102 int nr_pages, struct page *hpage)
8106 /* check current page array */
8107 for (i = 0; i < nr_pages; i++) {
8108 if (!PageCompound(pages[i]))
8110 if (compound_head(pages[i]) == hpage)
8114 /* check previously registered pages */
8115 for (i = 0; i < ctx->nr_user_bufs; i++) {
8116 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8118 for (j = 0; j < imu->nr_bvecs; j++) {
8119 if (!PageCompound(imu->bvec[j].bv_page))
8121 if (compound_head(imu->bvec[j].bv_page) == hpage)
8129 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8130 int nr_pages, struct io_mapped_ubuf *imu,
8131 struct page **last_hpage)
8135 for (i = 0; i < nr_pages; i++) {
8136 if (!PageCompound(pages[i])) {
8141 hpage = compound_head(pages[i]);
8142 if (hpage == *last_hpage)
8144 *last_hpage = hpage;
8145 if (headpage_already_acct(ctx, pages, i, hpage))
8147 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8151 if (!imu->acct_pages)
8154 ret = io_account_mem(ctx, imu->acct_pages);
8156 imu->acct_pages = 0;
8160 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8161 struct io_mapped_ubuf *imu,
8162 struct page **last_hpage)
8164 struct vm_area_struct **vmas = NULL;
8165 struct page **pages = NULL;
8166 unsigned long off, start, end, ubuf;
8168 int ret, pret, nr_pages, i;
8170 ubuf = (unsigned long) iov->iov_base;
8171 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8172 start = ubuf >> PAGE_SHIFT;
8173 nr_pages = end - start;
8177 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8181 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8186 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8192 mmap_read_lock(current->mm);
8193 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8195 if (pret == nr_pages) {
8196 /* don't support file backed memory */
8197 for (i = 0; i < nr_pages; i++) {
8198 struct vm_area_struct *vma = vmas[i];
8201 !is_file_hugepages(vma->vm_file)) {
8207 ret = pret < 0 ? pret : -EFAULT;
8209 mmap_read_unlock(current->mm);
8212 * if we did partial map, or found file backed vmas,
8213 * release any pages we did get
8216 unpin_user_pages(pages, pret);
8221 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8223 unpin_user_pages(pages, pret);
8228 off = ubuf & ~PAGE_MASK;
8229 size = iov->iov_len;
8230 for (i = 0; i < nr_pages; i++) {
8233 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8234 imu->bvec[i].bv_page = pages[i];
8235 imu->bvec[i].bv_len = vec_len;
8236 imu->bvec[i].bv_offset = off;
8240 /* store original address for later verification */
8242 imu->len = iov->iov_len;
8243 imu->nr_bvecs = nr_pages;
8251 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8255 if (!nr_args || nr_args > UIO_MAXIOV)
8258 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8260 if (!ctx->user_bufs)
8266 static int io_buffer_validate(struct iovec *iov)
8269 * Don't impose further limits on the size and buffer
8270 * constraints here, we'll -EINVAL later when IO is
8271 * submitted if they are wrong.
8273 if (!iov->iov_base || !iov->iov_len)
8276 /* arbitrary limit, but we need something */
8277 if (iov->iov_len > SZ_1G)
8283 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8284 unsigned int nr_args)
8288 struct page *last_hpage = NULL;
8290 ret = io_buffers_map_alloc(ctx, nr_args);
8294 for (i = 0; i < nr_args; i++) {
8295 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8297 ret = io_copy_iov(ctx, &iov, arg, i);
8301 ret = io_buffer_validate(&iov);
8305 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8309 ctx->nr_user_bufs++;
8313 io_sqe_buffers_unregister(ctx);
8318 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8320 __s32 __user *fds = arg;
8326 if (copy_from_user(&fd, fds, sizeof(*fds)))
8329 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8330 if (IS_ERR(ctx->cq_ev_fd)) {
8331 int ret = PTR_ERR(ctx->cq_ev_fd);
8332 ctx->cq_ev_fd = NULL;
8339 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8341 if (ctx->cq_ev_fd) {
8342 eventfd_ctx_put(ctx->cq_ev_fd);
8343 ctx->cq_ev_fd = NULL;
8350 static int __io_destroy_buffers(int id, void *p, void *data)
8352 struct io_ring_ctx *ctx = data;
8353 struct io_buffer *buf = p;
8355 __io_remove_buffers(ctx, buf, id, -1U);
8359 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8361 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8362 idr_destroy(&ctx->io_buffer_idr);
8365 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8367 struct io_kiocb *req, *nxt;
8369 list_for_each_entry_safe(req, nxt, list, compl.list) {
8370 if (tsk && req->task != tsk)
8372 list_del(&req->compl.list);
8373 kmem_cache_free(req_cachep, req);
8377 static void io_req_caches_free(struct io_ring_ctx *ctx)
8379 struct io_submit_state *submit_state = &ctx->submit_state;
8380 struct io_comp_state *cs = &ctx->submit_state.comp;
8382 mutex_lock(&ctx->uring_lock);
8384 if (submit_state->free_reqs) {
8385 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8386 submit_state->reqs);
8387 submit_state->free_reqs = 0;
8390 spin_lock_irq(&ctx->completion_lock);
8391 list_splice_init(&cs->locked_free_list, &cs->free_list);
8392 cs->locked_free_nr = 0;
8393 spin_unlock_irq(&ctx->completion_lock);
8395 io_req_cache_free(&cs->free_list, NULL);
8397 mutex_unlock(&ctx->uring_lock);
8400 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8403 * Some may use context even when all refs and requests have been put,
8404 * and they are free to do so while still holding uring_lock, see
8405 * __io_req_task_submit(). Wait for them to finish.
8407 mutex_lock(&ctx->uring_lock);
8408 mutex_unlock(&ctx->uring_lock);
8410 io_sq_thread_finish(ctx);
8411 io_sqe_buffers_unregister(ctx);
8413 if (ctx->mm_account) {
8414 mmdrop(ctx->mm_account);
8415 ctx->mm_account = NULL;
8418 mutex_lock(&ctx->uring_lock);
8419 io_sqe_files_unregister(ctx);
8420 mutex_unlock(&ctx->uring_lock);
8421 io_eventfd_unregister(ctx);
8422 io_destroy_buffers(ctx);
8423 idr_destroy(&ctx->personality_idr);
8425 #if defined(CONFIG_UNIX)
8426 if (ctx->ring_sock) {
8427 ctx->ring_sock->file = NULL; /* so that iput() is called */
8428 sock_release(ctx->ring_sock);
8432 io_mem_free(ctx->rings);
8433 io_mem_free(ctx->sq_sqes);
8435 percpu_ref_exit(&ctx->refs);
8436 free_uid(ctx->user);
8437 io_req_caches_free(ctx);
8439 io_wq_put_hash(ctx->hash_map);
8440 kfree(ctx->cancel_hash);
8444 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8446 struct io_ring_ctx *ctx = file->private_data;
8449 poll_wait(file, &ctx->cq_wait, wait);
8451 * synchronizes with barrier from wq_has_sleeper call in
8455 if (!io_sqring_full(ctx))
8456 mask |= EPOLLOUT | EPOLLWRNORM;
8459 * Don't flush cqring overflow list here, just do a simple check.
8460 * Otherwise there could possible be ABBA deadlock:
8463 * lock(&ctx->uring_lock);
8465 * lock(&ctx->uring_lock);
8468 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8469 * pushs them to do the flush.
8471 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8472 mask |= EPOLLIN | EPOLLRDNORM;
8477 static int io_uring_fasync(int fd, struct file *file, int on)
8479 struct io_ring_ctx *ctx = file->private_data;
8481 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8484 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8486 const struct cred *creds;
8488 creds = idr_remove(&ctx->personality_idr, id);
8497 static int io_remove_personalities(int id, void *p, void *data)
8499 struct io_ring_ctx *ctx = data;
8501 io_unregister_personality(ctx, id);
8505 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8507 struct callback_head *work, *next;
8508 bool executed = false;
8511 work = xchg(&ctx->exit_task_work, NULL);
8527 struct io_tctx_exit {
8528 struct callback_head task_work;
8529 struct completion completion;
8530 struct io_ring_ctx *ctx;
8533 static void io_tctx_exit_cb(struct callback_head *cb)
8535 struct io_uring_task *tctx = current->io_uring;
8536 struct io_tctx_exit *work;
8538 work = container_of(cb, struct io_tctx_exit, task_work);
8540 * When @in_idle, we're in cancellation and it's racy to remove the
8541 * node. It'll be removed by the end of cancellation, just ignore it.
8543 if (!atomic_read(&tctx->in_idle))
8544 io_uring_del_task_file((unsigned long)work->ctx);
8545 complete(&work->completion);
8548 static void io_ring_exit_work(struct work_struct *work)
8550 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8551 unsigned long timeout = jiffies + HZ * 60 * 5;
8552 struct io_tctx_exit exit;
8553 struct io_tctx_node *node;
8557 * If we're doing polled IO and end up having requests being
8558 * submitted async (out-of-line), then completions can come in while
8559 * we're waiting for refs to drop. We need to reap these manually,
8560 * as nobody else will be looking for them.
8563 io_uring_try_cancel_requests(ctx, NULL, NULL);
8565 WARN_ON_ONCE(time_after(jiffies, timeout));
8566 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8568 mutex_lock(&ctx->uring_lock);
8569 while (!list_empty(&ctx->tctx_list)) {
8570 WARN_ON_ONCE(time_after(jiffies, timeout));
8572 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8575 init_completion(&exit.completion);
8576 init_task_work(&exit.task_work, io_tctx_exit_cb);
8577 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8578 if (WARN_ON_ONCE(ret))
8580 wake_up_process(node->task);
8582 mutex_unlock(&ctx->uring_lock);
8583 wait_for_completion(&exit.completion);
8585 mutex_lock(&ctx->uring_lock);
8587 mutex_unlock(&ctx->uring_lock);
8589 io_ring_ctx_free(ctx);
8592 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8594 mutex_lock(&ctx->uring_lock);
8595 percpu_ref_kill(&ctx->refs);
8596 /* if force is set, the ring is going away. always drop after that */
8597 ctx->cq_overflow_flushed = 1;
8599 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8600 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8601 mutex_unlock(&ctx->uring_lock);
8603 io_kill_timeouts(ctx, NULL, NULL);
8604 io_poll_remove_all(ctx, NULL, NULL);
8606 /* if we failed setting up the ctx, we might not have any rings */
8607 io_iopoll_try_reap_events(ctx);
8609 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8611 * Use system_unbound_wq to avoid spawning tons of event kworkers
8612 * if we're exiting a ton of rings at the same time. It just adds
8613 * noise and overhead, there's no discernable change in runtime
8614 * over using system_wq.
8616 queue_work(system_unbound_wq, &ctx->exit_work);
8619 static int io_uring_release(struct inode *inode, struct file *file)
8621 struct io_ring_ctx *ctx = file->private_data;
8623 file->private_data = NULL;
8624 io_ring_ctx_wait_and_kill(ctx);
8628 struct io_task_cancel {
8629 struct task_struct *task;
8630 struct files_struct *files;
8633 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8635 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8636 struct io_task_cancel *cancel = data;
8639 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8640 unsigned long flags;
8641 struct io_ring_ctx *ctx = req->ctx;
8643 /* protect against races with linked timeouts */
8644 spin_lock_irqsave(&ctx->completion_lock, flags);
8645 ret = io_match_task(req, cancel->task, cancel->files);
8646 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8648 ret = io_match_task(req, cancel->task, cancel->files);
8653 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8654 struct task_struct *task,
8655 struct files_struct *files)
8657 struct io_defer_entry *de = NULL;
8660 spin_lock_irq(&ctx->completion_lock);
8661 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8662 if (io_match_task(de->req, task, files)) {
8663 list_cut_position(&list, &ctx->defer_list, &de->list);
8667 spin_unlock_irq(&ctx->completion_lock);
8669 while (!list_empty(&list)) {
8670 de = list_first_entry(&list, struct io_defer_entry, list);
8671 list_del_init(&de->list);
8672 req_set_fail_links(de->req);
8673 io_put_req(de->req);
8674 io_req_complete(de->req, -ECANCELED);
8679 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8681 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8683 return req->ctx == data;
8686 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8688 struct io_tctx_node *node;
8689 enum io_wq_cancel cret;
8692 mutex_lock(&ctx->uring_lock);
8693 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8694 struct io_uring_task *tctx = node->task->io_uring;
8697 * io_wq will stay alive while we hold uring_lock, because it's
8698 * killed after ctx nodes, which requires to take the lock.
8700 if (!tctx || !tctx->io_wq)
8702 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8703 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8705 mutex_unlock(&ctx->uring_lock);
8710 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8711 struct task_struct *task,
8712 struct files_struct *files)
8714 struct io_task_cancel cancel = { .task = task, .files = files, };
8715 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8718 enum io_wq_cancel cret;
8722 ret |= io_uring_try_cancel_iowq(ctx);
8723 } else if (tctx && tctx->io_wq) {
8725 * Cancels requests of all rings, not only @ctx, but
8726 * it's fine as the task is in exit/exec.
8728 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8730 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8733 /* SQPOLL thread does its own polling */
8734 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8735 while (!list_empty_careful(&ctx->iopoll_list)) {
8736 io_iopoll_try_reap_events(ctx);
8741 ret |= io_poll_remove_all(ctx, task, files);
8742 ret |= io_kill_timeouts(ctx, task, files);
8743 ret |= io_run_task_work();
8744 ret |= io_run_ctx_fallback(ctx);
8745 io_cqring_overflow_flush(ctx, true, task, files);
8752 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8753 struct task_struct *task,
8754 struct files_struct *files)
8756 struct io_kiocb *req;
8759 spin_lock_irq(&ctx->inflight_lock);
8760 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8761 cnt += io_match_task(req, task, files);
8762 spin_unlock_irq(&ctx->inflight_lock);
8766 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8767 struct task_struct *task,
8768 struct files_struct *files)
8770 while (!list_empty_careful(&ctx->inflight_list)) {
8774 inflight = io_uring_count_inflight(ctx, task, files);
8778 io_uring_try_cancel_requests(ctx, task, files);
8781 io_sq_thread_unpark(ctx->sq_data);
8782 prepare_to_wait(&task->io_uring->wait, &wait,
8783 TASK_UNINTERRUPTIBLE);
8784 if (inflight == io_uring_count_inflight(ctx, task, files))
8786 finish_wait(&task->io_uring->wait, &wait);
8788 io_sq_thread_park(ctx->sq_data);
8793 * We need to iteratively cancel requests, in case a request has dependent
8794 * hard links. These persist even for failure of cancelations, hence keep
8795 * looping until none are found.
8797 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8798 struct files_struct *files)
8800 struct task_struct *task = current;
8802 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8803 /* never started, nothing to cancel */
8804 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8805 io_sq_offload_start(ctx);
8808 io_sq_thread_park(ctx->sq_data);
8809 task = ctx->sq_data->thread;
8811 atomic_inc(&task->io_uring->in_idle);
8814 io_cancel_defer_files(ctx, task, files);
8816 io_uring_cancel_files(ctx, task, files);
8818 io_uring_try_cancel_requests(ctx, task, NULL);
8821 atomic_dec(&task->io_uring->in_idle);
8823 io_sq_thread_unpark(ctx->sq_data);
8827 * Note that this task has used io_uring. We use it for cancelation purposes.
8829 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8831 struct io_uring_task *tctx = current->io_uring;
8832 struct io_tctx_node *node;
8835 if (unlikely(!tctx)) {
8836 ret = io_uring_alloc_task_context(current, ctx);
8839 tctx = current->io_uring;
8841 if (tctx->last != ctx) {
8842 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8845 node = kmalloc(sizeof(*node), GFP_KERNEL);
8849 node->task = current;
8851 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8858 mutex_lock(&ctx->uring_lock);
8859 list_add(&node->ctx_node, &ctx->tctx_list);
8860 mutex_unlock(&ctx->uring_lock);
8866 * This is race safe in that the task itself is doing this, hence it
8867 * cannot be going through the exit/cancel paths at the same time.
8868 * This cannot be modified while exit/cancel is running.
8870 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8871 tctx->sqpoll = true;
8877 * Remove this io_uring_file -> task mapping.
8879 static void io_uring_del_task_file(unsigned long index)
8881 struct io_uring_task *tctx = current->io_uring;
8882 struct io_tctx_node *node;
8886 node = xa_erase(&tctx->xa, index);
8890 WARN_ON_ONCE(current != node->task);
8891 WARN_ON_ONCE(list_empty(&node->ctx_node));
8893 mutex_lock(&node->ctx->uring_lock);
8894 list_del(&node->ctx_node);
8895 mutex_unlock(&node->ctx->uring_lock);
8897 if (tctx->last == node->ctx)
8902 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8904 struct io_tctx_node *node;
8905 unsigned long index;
8907 xa_for_each(&tctx->xa, index, node)
8908 io_uring_del_task_file(index);
8910 io_wq_put_and_exit(tctx->io_wq);
8915 void __io_uring_files_cancel(struct files_struct *files)
8917 struct io_uring_task *tctx = current->io_uring;
8918 struct io_tctx_node *node;
8919 unsigned long index;
8921 /* make sure overflow events are dropped */
8922 atomic_inc(&tctx->in_idle);
8923 xa_for_each(&tctx->xa, index, node)
8924 io_uring_cancel_task_requests(node->ctx, files);
8925 atomic_dec(&tctx->in_idle);
8928 io_uring_clean_tctx(tctx);
8931 static s64 tctx_inflight(struct io_uring_task *tctx)
8933 return percpu_counter_sum(&tctx->inflight);
8936 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8938 struct io_sq_data *sqd = ctx->sq_data;
8939 struct io_uring_task *tctx;
8945 io_sq_thread_park(sqd);
8946 if (!sqd->thread || !sqd->thread->io_uring) {
8947 io_sq_thread_unpark(sqd);
8950 tctx = ctx->sq_data->thread->io_uring;
8951 atomic_inc(&tctx->in_idle);
8953 /* read completions before cancelations */
8954 inflight = tctx_inflight(tctx);
8957 io_uring_cancel_task_requests(ctx, NULL);
8959 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8961 * If we've seen completions, retry without waiting. This
8962 * avoids a race where a completion comes in before we did
8963 * prepare_to_wait().
8965 if (inflight == tctx_inflight(tctx))
8967 finish_wait(&tctx->wait, &wait);
8969 atomic_dec(&tctx->in_idle);
8970 io_sq_thread_unpark(sqd);
8974 * Find any io_uring fd that this task has registered or done IO on, and cancel
8977 void __io_uring_task_cancel(void)
8979 struct io_uring_task *tctx = current->io_uring;
8983 /* make sure overflow events are dropped */
8984 atomic_inc(&tctx->in_idle);
8987 struct io_tctx_node *node;
8988 unsigned long index;
8990 xa_for_each(&tctx->xa, index, node)
8991 io_uring_cancel_sqpoll(node->ctx);
8995 /* read completions before cancelations */
8996 inflight = tctx_inflight(tctx);
8999 __io_uring_files_cancel(NULL);
9001 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9004 * If we've seen completions, retry without waiting. This
9005 * avoids a race where a completion comes in before we did
9006 * prepare_to_wait().
9008 if (inflight == tctx_inflight(tctx))
9010 finish_wait(&tctx->wait, &wait);
9013 atomic_dec(&tctx->in_idle);
9015 io_uring_clean_tctx(tctx);
9016 /* all current's requests should be gone, we can kill tctx */
9017 __io_uring_free(current);
9020 static void *io_uring_validate_mmap_request(struct file *file,
9021 loff_t pgoff, size_t sz)
9023 struct io_ring_ctx *ctx = file->private_data;
9024 loff_t offset = pgoff << PAGE_SHIFT;
9029 case IORING_OFF_SQ_RING:
9030 case IORING_OFF_CQ_RING:
9033 case IORING_OFF_SQES:
9037 return ERR_PTR(-EINVAL);
9040 page = virt_to_head_page(ptr);
9041 if (sz > page_size(page))
9042 return ERR_PTR(-EINVAL);
9049 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9051 size_t sz = vma->vm_end - vma->vm_start;
9055 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9057 return PTR_ERR(ptr);
9059 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9060 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9063 #else /* !CONFIG_MMU */
9065 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9067 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9070 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9072 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9075 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9076 unsigned long addr, unsigned long len,
9077 unsigned long pgoff, unsigned long flags)
9081 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9083 return PTR_ERR(ptr);
9085 return (unsigned long) ptr;
9088 #endif /* !CONFIG_MMU */
9090 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9096 if (!io_sqring_full(ctx))
9098 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9100 if (!io_sqring_full(ctx))
9103 } while (!signal_pending(current));
9105 finish_wait(&ctx->sqo_sq_wait, &wait);
9109 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9110 struct __kernel_timespec __user **ts,
9111 const sigset_t __user **sig)
9113 struct io_uring_getevents_arg arg;
9116 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9117 * is just a pointer to the sigset_t.
9119 if (!(flags & IORING_ENTER_EXT_ARG)) {
9120 *sig = (const sigset_t __user *) argp;
9126 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9127 * timespec and sigset_t pointers if good.
9129 if (*argsz != sizeof(arg))
9131 if (copy_from_user(&arg, argp, sizeof(arg)))
9133 *sig = u64_to_user_ptr(arg.sigmask);
9134 *argsz = arg.sigmask_sz;
9135 *ts = u64_to_user_ptr(arg.ts);
9139 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9140 u32, min_complete, u32, flags, const void __user *, argp,
9143 struct io_ring_ctx *ctx;
9150 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9151 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9159 if (f.file->f_op != &io_uring_fops)
9163 ctx = f.file->private_data;
9164 if (!percpu_ref_tryget(&ctx->refs))
9168 if (ctx->flags & IORING_SETUP_R_DISABLED)
9172 * For SQ polling, the thread will do all submissions and completions.
9173 * Just return the requested submit count, and wake the thread if
9177 if (ctx->flags & IORING_SETUP_SQPOLL) {
9178 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9181 if (unlikely(ctx->sq_data->thread == NULL)) {
9184 if (flags & IORING_ENTER_SQ_WAKEUP)
9185 wake_up(&ctx->sq_data->wait);
9186 if (flags & IORING_ENTER_SQ_WAIT) {
9187 ret = io_sqpoll_wait_sq(ctx);
9191 submitted = to_submit;
9192 } else if (to_submit) {
9193 ret = io_uring_add_task_file(ctx);
9196 mutex_lock(&ctx->uring_lock);
9197 submitted = io_submit_sqes(ctx, to_submit);
9198 mutex_unlock(&ctx->uring_lock);
9200 if (submitted != to_submit)
9203 if (flags & IORING_ENTER_GETEVENTS) {
9204 const sigset_t __user *sig;
9205 struct __kernel_timespec __user *ts;
9207 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9211 min_complete = min(min_complete, ctx->cq_entries);
9214 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9215 * space applications don't need to do io completion events
9216 * polling again, they can rely on io_sq_thread to do polling
9217 * work, which can reduce cpu usage and uring_lock contention.
9219 if (ctx->flags & IORING_SETUP_IOPOLL &&
9220 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9221 ret = io_iopoll_check(ctx, min_complete);
9223 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9228 percpu_ref_put(&ctx->refs);
9231 return submitted ? submitted : ret;
9234 #ifdef CONFIG_PROC_FS
9235 static int io_uring_show_cred(int id, void *p, void *data)
9237 const struct cred *cred = p;
9238 struct seq_file *m = data;
9239 struct user_namespace *uns = seq_user_ns(m);
9240 struct group_info *gi;
9245 seq_printf(m, "%5d\n", id);
9246 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9247 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9248 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9249 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9250 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9251 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9252 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9253 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9254 seq_puts(m, "\n\tGroups:\t");
9255 gi = cred->group_info;
9256 for (g = 0; g < gi->ngroups; g++) {
9257 seq_put_decimal_ull(m, g ? " " : "",
9258 from_kgid_munged(uns, gi->gid[g]));
9260 seq_puts(m, "\n\tCapEff:\t");
9261 cap = cred->cap_effective;
9262 CAP_FOR_EACH_U32(__capi)
9263 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9268 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9270 struct io_sq_data *sq = NULL;
9275 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9276 * since fdinfo case grabs it in the opposite direction of normal use
9277 * cases. If we fail to get the lock, we just don't iterate any
9278 * structures that could be going away outside the io_uring mutex.
9280 has_lock = mutex_trylock(&ctx->uring_lock);
9282 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9288 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9289 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9290 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9291 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9292 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9295 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9297 seq_printf(m, "%5u: <none>\n", i);
9299 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9300 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9301 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9303 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9304 (unsigned int) buf->len);
9306 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9307 seq_printf(m, "Personalities:\n");
9308 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9310 seq_printf(m, "PollList:\n");
9311 spin_lock_irq(&ctx->completion_lock);
9312 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9313 struct hlist_head *list = &ctx->cancel_hash[i];
9314 struct io_kiocb *req;
9316 hlist_for_each_entry(req, list, hash_node)
9317 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9318 req->task->task_works != NULL);
9320 spin_unlock_irq(&ctx->completion_lock);
9322 mutex_unlock(&ctx->uring_lock);
9325 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9327 struct io_ring_ctx *ctx = f->private_data;
9329 if (percpu_ref_tryget(&ctx->refs)) {
9330 __io_uring_show_fdinfo(ctx, m);
9331 percpu_ref_put(&ctx->refs);
9336 static const struct file_operations io_uring_fops = {
9337 .release = io_uring_release,
9338 .mmap = io_uring_mmap,
9340 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9341 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9343 .poll = io_uring_poll,
9344 .fasync = io_uring_fasync,
9345 #ifdef CONFIG_PROC_FS
9346 .show_fdinfo = io_uring_show_fdinfo,
9350 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9351 struct io_uring_params *p)
9353 struct io_rings *rings;
9354 size_t size, sq_array_offset;
9356 /* make sure these are sane, as we already accounted them */
9357 ctx->sq_entries = p->sq_entries;
9358 ctx->cq_entries = p->cq_entries;
9360 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9361 if (size == SIZE_MAX)
9364 rings = io_mem_alloc(size);
9369 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9370 rings->sq_ring_mask = p->sq_entries - 1;
9371 rings->cq_ring_mask = p->cq_entries - 1;
9372 rings->sq_ring_entries = p->sq_entries;
9373 rings->cq_ring_entries = p->cq_entries;
9374 ctx->sq_mask = rings->sq_ring_mask;
9375 ctx->cq_mask = rings->cq_ring_mask;
9377 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9378 if (size == SIZE_MAX) {
9379 io_mem_free(ctx->rings);
9384 ctx->sq_sqes = io_mem_alloc(size);
9385 if (!ctx->sq_sqes) {
9386 io_mem_free(ctx->rings);
9394 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9398 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9402 ret = io_uring_add_task_file(ctx);
9407 fd_install(fd, file);
9412 * Allocate an anonymous fd, this is what constitutes the application
9413 * visible backing of an io_uring instance. The application mmaps this
9414 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9415 * we have to tie this fd to a socket for file garbage collection purposes.
9417 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9420 #if defined(CONFIG_UNIX)
9423 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9426 return ERR_PTR(ret);
9429 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9430 O_RDWR | O_CLOEXEC);
9431 #if defined(CONFIG_UNIX)
9433 sock_release(ctx->ring_sock);
9434 ctx->ring_sock = NULL;
9436 ctx->ring_sock->file = file;
9442 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9443 struct io_uring_params __user *params)
9445 struct io_ring_ctx *ctx;
9451 if (entries > IORING_MAX_ENTRIES) {
9452 if (!(p->flags & IORING_SETUP_CLAMP))
9454 entries = IORING_MAX_ENTRIES;
9458 * Use twice as many entries for the CQ ring. It's possible for the
9459 * application to drive a higher depth than the size of the SQ ring,
9460 * since the sqes are only used at submission time. This allows for
9461 * some flexibility in overcommitting a bit. If the application has
9462 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9463 * of CQ ring entries manually.
9465 p->sq_entries = roundup_pow_of_two(entries);
9466 if (p->flags & IORING_SETUP_CQSIZE) {
9468 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9469 * to a power-of-two, if it isn't already. We do NOT impose
9470 * any cq vs sq ring sizing.
9474 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9475 if (!(p->flags & IORING_SETUP_CLAMP))
9477 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9479 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9480 if (p->cq_entries < p->sq_entries)
9483 p->cq_entries = 2 * p->sq_entries;
9486 ctx = io_ring_ctx_alloc(p);
9489 ctx->compat = in_compat_syscall();
9490 if (!capable(CAP_IPC_LOCK))
9491 ctx->user = get_uid(current_user());
9494 * This is just grabbed for accounting purposes. When a process exits,
9495 * the mm is exited and dropped before the files, hence we need to hang
9496 * on to this mm purely for the purposes of being able to unaccount
9497 * memory (locked/pinned vm). It's not used for anything else.
9499 mmgrab(current->mm);
9500 ctx->mm_account = current->mm;
9502 ret = io_allocate_scq_urings(ctx, p);
9506 ret = io_sq_offload_create(ctx, p);
9510 if (!(p->flags & IORING_SETUP_R_DISABLED))
9511 io_sq_offload_start(ctx);
9513 memset(&p->sq_off, 0, sizeof(p->sq_off));
9514 p->sq_off.head = offsetof(struct io_rings, sq.head);
9515 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9516 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9517 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9518 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9519 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9520 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9522 memset(&p->cq_off, 0, sizeof(p->cq_off));
9523 p->cq_off.head = offsetof(struct io_rings, cq.head);
9524 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9525 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9526 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9527 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9528 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9529 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9531 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9532 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9533 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9534 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9535 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9537 if (copy_to_user(params, p, sizeof(*p))) {
9542 file = io_uring_get_file(ctx);
9544 ret = PTR_ERR(file);
9549 * Install ring fd as the very last thing, so we don't risk someone
9550 * having closed it before we finish setup
9552 ret = io_uring_install_fd(ctx, file);
9554 /* fput will clean it up */
9559 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9562 io_ring_ctx_wait_and_kill(ctx);
9567 * Sets up an aio uring context, and returns the fd. Applications asks for a
9568 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9569 * params structure passed in.
9571 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9573 struct io_uring_params p;
9576 if (copy_from_user(&p, params, sizeof(p)))
9578 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9583 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9584 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9585 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9586 IORING_SETUP_R_DISABLED))
9589 return io_uring_create(entries, &p, params);
9592 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9593 struct io_uring_params __user *, params)
9595 return io_uring_setup(entries, params);
9598 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9600 struct io_uring_probe *p;
9604 size = struct_size(p, ops, nr_args);
9605 if (size == SIZE_MAX)
9607 p = kzalloc(size, GFP_KERNEL);
9612 if (copy_from_user(p, arg, size))
9615 if (memchr_inv(p, 0, size))
9618 p->last_op = IORING_OP_LAST - 1;
9619 if (nr_args > IORING_OP_LAST)
9620 nr_args = IORING_OP_LAST;
9622 for (i = 0; i < nr_args; i++) {
9624 if (!io_op_defs[i].not_supported)
9625 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9630 if (copy_to_user(arg, p, size))
9637 static int io_register_personality(struct io_ring_ctx *ctx)
9639 const struct cred *creds;
9642 creds = get_current_cred();
9644 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9645 USHRT_MAX, GFP_KERNEL);
9651 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9652 unsigned int nr_args)
9654 struct io_uring_restriction *res;
9658 /* Restrictions allowed only if rings started disabled */
9659 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9662 /* We allow only a single restrictions registration */
9663 if (ctx->restrictions.registered)
9666 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9669 size = array_size(nr_args, sizeof(*res));
9670 if (size == SIZE_MAX)
9673 res = memdup_user(arg, size);
9675 return PTR_ERR(res);
9679 for (i = 0; i < nr_args; i++) {
9680 switch (res[i].opcode) {
9681 case IORING_RESTRICTION_REGISTER_OP:
9682 if (res[i].register_op >= IORING_REGISTER_LAST) {
9687 __set_bit(res[i].register_op,
9688 ctx->restrictions.register_op);
9690 case IORING_RESTRICTION_SQE_OP:
9691 if (res[i].sqe_op >= IORING_OP_LAST) {
9696 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9698 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9699 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9701 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9702 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9711 /* Reset all restrictions if an error happened */
9713 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9715 ctx->restrictions.registered = true;
9721 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9723 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9726 if (ctx->restrictions.registered)
9727 ctx->restricted = 1;
9729 io_sq_offload_start(ctx);
9733 static bool io_register_op_must_quiesce(int op)
9736 case IORING_UNREGISTER_FILES:
9737 case IORING_REGISTER_FILES_UPDATE:
9738 case IORING_REGISTER_PROBE:
9739 case IORING_REGISTER_PERSONALITY:
9740 case IORING_UNREGISTER_PERSONALITY:
9747 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9748 void __user *arg, unsigned nr_args)
9749 __releases(ctx->uring_lock)
9750 __acquires(ctx->uring_lock)
9755 * We're inside the ring mutex, if the ref is already dying, then
9756 * someone else killed the ctx or is already going through
9757 * io_uring_register().
9759 if (percpu_ref_is_dying(&ctx->refs))
9762 if (io_register_op_must_quiesce(opcode)) {
9763 percpu_ref_kill(&ctx->refs);
9766 * Drop uring mutex before waiting for references to exit. If
9767 * another thread is currently inside io_uring_enter() it might
9768 * need to grab the uring_lock to make progress. If we hold it
9769 * here across the drain wait, then we can deadlock. It's safe
9770 * to drop the mutex here, since no new references will come in
9771 * after we've killed the percpu ref.
9773 mutex_unlock(&ctx->uring_lock);
9775 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9778 ret = io_run_task_work_sig();
9783 mutex_lock(&ctx->uring_lock);
9786 percpu_ref_resurrect(&ctx->refs);
9791 if (ctx->restricted) {
9792 if (opcode >= IORING_REGISTER_LAST) {
9797 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9804 case IORING_REGISTER_BUFFERS:
9805 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9807 case IORING_UNREGISTER_BUFFERS:
9811 ret = io_sqe_buffers_unregister(ctx);
9813 case IORING_REGISTER_FILES:
9814 ret = io_sqe_files_register(ctx, arg, nr_args);
9816 case IORING_UNREGISTER_FILES:
9820 ret = io_sqe_files_unregister(ctx);
9822 case IORING_REGISTER_FILES_UPDATE:
9823 ret = io_sqe_files_update(ctx, arg, nr_args);
9825 case IORING_REGISTER_EVENTFD:
9826 case IORING_REGISTER_EVENTFD_ASYNC:
9830 ret = io_eventfd_register(ctx, arg);
9833 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9834 ctx->eventfd_async = 1;
9836 ctx->eventfd_async = 0;
9838 case IORING_UNREGISTER_EVENTFD:
9842 ret = io_eventfd_unregister(ctx);
9844 case IORING_REGISTER_PROBE:
9846 if (!arg || nr_args > 256)
9848 ret = io_probe(ctx, arg, nr_args);
9850 case IORING_REGISTER_PERSONALITY:
9854 ret = io_register_personality(ctx);
9856 case IORING_UNREGISTER_PERSONALITY:
9860 ret = io_unregister_personality(ctx, nr_args);
9862 case IORING_REGISTER_ENABLE_RINGS:
9866 ret = io_register_enable_rings(ctx);
9868 case IORING_REGISTER_RESTRICTIONS:
9869 ret = io_register_restrictions(ctx, arg, nr_args);
9877 if (io_register_op_must_quiesce(opcode)) {
9878 /* bring the ctx back to life */
9879 percpu_ref_reinit(&ctx->refs);
9881 reinit_completion(&ctx->ref_comp);
9886 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9887 void __user *, arg, unsigned int, nr_args)
9889 struct io_ring_ctx *ctx;
9898 if (f.file->f_op != &io_uring_fops)
9901 ctx = f.file->private_data;
9905 mutex_lock(&ctx->uring_lock);
9906 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9907 mutex_unlock(&ctx->uring_lock);
9908 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9909 ctx->cq_ev_fd != NULL, ret);
9915 static int __init io_uring_init(void)
9917 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9918 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9919 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9922 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9923 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9924 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9925 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9926 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9927 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9928 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9929 BUILD_BUG_SQE_ELEM(8, __u64, off);
9930 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9931 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9932 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9933 BUILD_BUG_SQE_ELEM(24, __u32, len);
9934 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9935 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9936 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9937 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9938 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9939 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9940 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9941 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9942 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9943 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9944 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9945 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9946 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9947 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9948 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9949 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9950 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9951 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9952 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9954 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9955 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9956 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9960 __initcall(io_uring_init);