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_cqe (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/tracehook.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)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 /* 512 entries per page on 64-bit archs, 64 pages max */
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT 9
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head cq_overflow_list;
379 struct xarray io_buffers;
380 struct xarray personalities;
382 unsigned sq_thread_idle;
383 } ____cacheline_aligned_in_smp;
385 /* IRQ completion list, under ->completion_lock */
386 struct list_head locked_free_list;
387 unsigned int locked_free_nr;
389 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
390 struct io_sq_data *sq_data; /* if using sq thread polling */
392 struct wait_queue_head sqo_sq_wait;
393 struct list_head sqd_list;
395 unsigned long check_cq_overflow;
398 unsigned cached_cq_tail;
400 struct eventfd_ctx *cq_ev_fd;
401 struct wait_queue_head poll_wait;
402 struct wait_queue_head cq_wait;
404 atomic_t cq_timeouts;
405 struct fasync_struct *cq_fasync;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
462 struct io_uring_task {
463 /* submission side */
466 struct wait_queue_head wait;
467 const struct io_ring_ctx *last;
469 struct percpu_counter inflight;
470 atomic_t inflight_tracked;
473 spinlock_t task_lock;
474 struct io_wq_work_list task_list;
475 struct callback_head task_work;
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
483 struct io_poll_iocb {
485 struct wait_queue_head *head;
489 struct wait_queue_entry wait;
492 struct io_poll_update {
498 bool update_user_data;
506 struct io_timeout_data {
507 struct io_kiocb *req;
508 struct hrtimer timer;
509 struct timespec64 ts;
510 enum hrtimer_mode mode;
515 struct sockaddr __user *addr;
516 int __user *addr_len;
518 unsigned long nofile;
538 struct list_head list;
539 /* head of the link, used by linked timeouts only */
540 struct io_kiocb *head;
541 /* for linked completions */
542 struct io_kiocb *prev;
545 struct io_timeout_rem {
550 struct timespec64 ts;
555 /* NOTE: kiocb has the file as the first member, so don't do it here */
563 struct sockaddr __user *addr;
570 struct compat_msghdr __user *umsg_compat;
571 struct user_msghdr __user *umsg;
577 struct io_buffer *kbuf;
583 struct filename *filename;
585 unsigned long nofile;
588 struct io_rsrc_update {
614 struct epoll_event event;
618 struct file *file_out;
619 struct file *file_in;
626 struct io_provide_buf {
640 const char __user *filename;
641 struct statx __user *buffer;
653 struct filename *oldpath;
654 struct filename *newpath;
662 struct filename *filename;
665 struct io_completion {
670 struct io_async_connect {
671 struct sockaddr_storage address;
674 struct io_async_msghdr {
675 struct iovec fast_iov[UIO_FASTIOV];
676 /* points to an allocated iov, if NULL we use fast_iov instead */
677 struct iovec *free_iov;
678 struct sockaddr __user *uaddr;
680 struct sockaddr_storage addr;
684 struct iovec fast_iov[UIO_FASTIOV];
685 const struct iovec *free_iovec;
686 struct iov_iter iter;
688 struct wait_page_queue wpq;
692 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
693 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
694 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
695 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
696 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
697 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
699 /* first byte is taken by user flags, shift it to not overlap */
704 REQ_F_LINK_TIMEOUT_BIT,
705 REQ_F_NEED_CLEANUP_BIT,
707 REQ_F_BUFFER_SELECTED_BIT,
708 REQ_F_LTIMEOUT_ACTIVE_BIT,
709 REQ_F_COMPLETE_INLINE_BIT,
711 REQ_F_DONT_REISSUE_BIT,
714 /* keep async read/write and isreg together and in order */
715 REQ_F_NOWAIT_READ_BIT,
716 REQ_F_NOWAIT_WRITE_BIT,
719 /* not a real bit, just to check we're not overflowing the space */
725 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
726 /* drain existing IO first */
727 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
729 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
730 /* doesn't sever on completion < 0 */
731 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
733 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
734 /* IOSQE_BUFFER_SELECT */
735 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
737 /* fail rest of links */
738 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
739 /* on inflight list, should be cancelled and waited on exit reliably */
740 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
741 /* read/write uses file position */
742 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
743 /* must not punt to workers */
744 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
745 /* has or had linked timeout */
746 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
748 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
749 /* already went through poll handler */
750 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
751 /* buffer already selected */
752 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
753 /* linked timeout is active, i.e. prepared by link's head */
754 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
755 /* completion is deferred through io_comp_state */
756 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
757 /* caller should reissue async */
758 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
759 /* don't attempt request reissue, see io_rw_reissue() */
760 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
761 /* supports async reads */
762 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
763 /* supports async writes */
764 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
766 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
767 /* has creds assigned */
768 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
769 /* skip refcounting if not set */
770 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
774 struct io_poll_iocb poll;
775 struct io_poll_iocb *double_poll;
778 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
780 struct io_task_work {
782 struct io_wq_work_node node;
783 struct llist_node fallback_node;
785 io_req_tw_func_t func;
789 IORING_RSRC_FILE = 0,
790 IORING_RSRC_BUFFER = 1,
794 * NOTE! Each of the iocb union members has the file pointer
795 * as the first entry in their struct definition. So you can
796 * access the file pointer through any of the sub-structs,
797 * or directly as just 'ki_filp' in this struct.
803 struct io_poll_iocb poll;
804 struct io_poll_update poll_update;
805 struct io_accept accept;
807 struct io_cancel cancel;
808 struct io_timeout timeout;
809 struct io_timeout_rem timeout_rem;
810 struct io_connect connect;
811 struct io_sr_msg sr_msg;
813 struct io_close close;
814 struct io_rsrc_update rsrc_update;
815 struct io_fadvise fadvise;
816 struct io_madvise madvise;
817 struct io_epoll epoll;
818 struct io_splice splice;
819 struct io_provide_buf pbuf;
820 struct io_statx statx;
821 struct io_shutdown shutdown;
822 struct io_rename rename;
823 struct io_unlink unlink;
824 /* use only after cleaning per-op data, see io_clean_op() */
825 struct io_completion compl;
828 /* opcode allocated if it needs to store data for async defer */
831 /* polled IO has completed */
837 struct io_ring_ctx *ctx;
840 struct task_struct *task;
843 struct io_kiocb *link;
844 struct percpu_ref *fixed_rsrc_refs;
846 /* used with ctx->iopoll_list with reads/writes */
847 struct list_head inflight_entry;
848 struct io_task_work io_task_work;
849 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
850 struct hlist_node hash_node;
851 struct async_poll *apoll;
852 struct io_wq_work work;
853 const struct cred *creds;
855 /* store used ubuf, so we can prevent reloading */
856 struct io_mapped_ubuf *imu;
859 struct io_tctx_node {
860 struct list_head ctx_node;
861 struct task_struct *task;
862 struct io_ring_ctx *ctx;
865 struct io_defer_entry {
866 struct list_head list;
867 struct io_kiocb *req;
872 /* needs req->file assigned */
873 unsigned needs_file : 1;
874 /* hash wq insertion if file is a regular file */
875 unsigned hash_reg_file : 1;
876 /* unbound wq insertion if file is a non-regular file */
877 unsigned unbound_nonreg_file : 1;
878 /* opcode is not supported by this kernel */
879 unsigned not_supported : 1;
880 /* set if opcode supports polled "wait" */
882 unsigned pollout : 1;
883 /* op supports buffer selection */
884 unsigned buffer_select : 1;
885 /* do prep async if is going to be punted */
886 unsigned needs_async_setup : 1;
887 /* should block plug */
889 /* size of async data needed, if any */
890 unsigned short async_size;
893 static const struct io_op_def io_op_defs[] = {
894 [IORING_OP_NOP] = {},
895 [IORING_OP_READV] = {
897 .unbound_nonreg_file = 1,
900 .needs_async_setup = 1,
902 .async_size = sizeof(struct io_async_rw),
904 [IORING_OP_WRITEV] = {
907 .unbound_nonreg_file = 1,
909 .needs_async_setup = 1,
911 .async_size = sizeof(struct io_async_rw),
913 [IORING_OP_FSYNC] = {
916 [IORING_OP_READ_FIXED] = {
918 .unbound_nonreg_file = 1,
921 .async_size = sizeof(struct io_async_rw),
923 [IORING_OP_WRITE_FIXED] = {
926 .unbound_nonreg_file = 1,
929 .async_size = sizeof(struct io_async_rw),
931 [IORING_OP_POLL_ADD] = {
933 .unbound_nonreg_file = 1,
935 [IORING_OP_POLL_REMOVE] = {},
936 [IORING_OP_SYNC_FILE_RANGE] = {
939 [IORING_OP_SENDMSG] = {
941 .unbound_nonreg_file = 1,
943 .needs_async_setup = 1,
944 .async_size = sizeof(struct io_async_msghdr),
946 [IORING_OP_RECVMSG] = {
948 .unbound_nonreg_file = 1,
951 .needs_async_setup = 1,
952 .async_size = sizeof(struct io_async_msghdr),
954 [IORING_OP_TIMEOUT] = {
955 .async_size = sizeof(struct io_timeout_data),
957 [IORING_OP_TIMEOUT_REMOVE] = {
958 /* used by timeout updates' prep() */
960 [IORING_OP_ACCEPT] = {
962 .unbound_nonreg_file = 1,
965 [IORING_OP_ASYNC_CANCEL] = {},
966 [IORING_OP_LINK_TIMEOUT] = {
967 .async_size = sizeof(struct io_timeout_data),
969 [IORING_OP_CONNECT] = {
971 .unbound_nonreg_file = 1,
973 .needs_async_setup = 1,
974 .async_size = sizeof(struct io_async_connect),
976 [IORING_OP_FALLOCATE] = {
979 [IORING_OP_OPENAT] = {},
980 [IORING_OP_CLOSE] = {},
981 [IORING_OP_FILES_UPDATE] = {},
982 [IORING_OP_STATX] = {},
985 .unbound_nonreg_file = 1,
989 .async_size = sizeof(struct io_async_rw),
991 [IORING_OP_WRITE] = {
993 .unbound_nonreg_file = 1,
996 .async_size = sizeof(struct io_async_rw),
998 [IORING_OP_FADVISE] = {
1001 [IORING_OP_MADVISE] = {},
1002 [IORING_OP_SEND] = {
1004 .unbound_nonreg_file = 1,
1007 [IORING_OP_RECV] = {
1009 .unbound_nonreg_file = 1,
1013 [IORING_OP_OPENAT2] = {
1015 [IORING_OP_EPOLL_CTL] = {
1016 .unbound_nonreg_file = 1,
1018 [IORING_OP_SPLICE] = {
1021 .unbound_nonreg_file = 1,
1023 [IORING_OP_PROVIDE_BUFFERS] = {},
1024 [IORING_OP_REMOVE_BUFFERS] = {},
1028 .unbound_nonreg_file = 1,
1030 [IORING_OP_SHUTDOWN] = {
1033 [IORING_OP_RENAMEAT] = {},
1034 [IORING_OP_UNLINKAT] = {},
1037 static bool io_disarm_next(struct io_kiocb *req);
1038 static void io_uring_del_tctx_node(unsigned long index);
1039 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1040 struct task_struct *task,
1042 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_queue_linked_timeout(struct io_kiocb *req);
1050 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1051 struct io_uring_rsrc_update2 *up,
1053 static void io_clean_op(struct io_kiocb *req);
1054 static struct file *io_file_get(struct io_ring_ctx *ctx,
1055 struct io_kiocb *req, int fd, bool fixed);
1056 static void __io_queue_sqe(struct io_kiocb *req);
1057 static void io_rsrc_put_work(struct work_struct *work);
1059 static void io_req_task_queue(struct io_kiocb *req);
1060 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1061 static int io_req_prep_async(struct io_kiocb *req);
1063 static struct kmem_cache *req_cachep;
1065 static const struct file_operations io_uring_fops;
1067 struct sock *io_uring_get_socket(struct file *file)
1069 #if defined(CONFIG_UNIX)
1070 if (file->f_op == &io_uring_fops) {
1071 struct io_ring_ctx *ctx = file->private_data;
1073 return ctx->ring_sock->sk;
1078 EXPORT_SYMBOL(io_uring_get_socket);
1080 #define io_for_each_link(pos, head) \
1081 for (pos = (head); pos; pos = pos->link)
1084 * Shamelessly stolen from the mm implementation of page reference checking,
1085 * see commit f958d7b528b1 for details.
1087 #define req_ref_zero_or_close_to_overflow(req) \
1088 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1090 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1092 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1093 return atomic_inc_not_zero(&req->refs);
1096 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1098 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1101 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1102 return atomic_dec_and_test(&req->refs);
1105 static inline void req_ref_put(struct io_kiocb *req)
1107 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1108 WARN_ON_ONCE(req_ref_put_and_test(req));
1111 static inline void req_ref_get(struct io_kiocb *req)
1113 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1114 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1115 atomic_inc(&req->refs);
1118 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1120 if (!(req->flags & REQ_F_REFCOUNT)) {
1121 req->flags |= REQ_F_REFCOUNT;
1122 atomic_set(&req->refs, nr);
1126 static inline void io_req_set_refcount(struct io_kiocb *req)
1128 __io_req_set_refcount(req, 1);
1131 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1133 struct io_ring_ctx *ctx = req->ctx;
1135 if (!req->fixed_rsrc_refs) {
1136 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1137 percpu_ref_get(req->fixed_rsrc_refs);
1141 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1143 bool got = percpu_ref_tryget(ref);
1145 /* already at zero, wait for ->release() */
1147 wait_for_completion(compl);
1148 percpu_ref_resurrect(ref);
1150 percpu_ref_put(ref);
1153 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1156 struct io_kiocb *req;
1158 if (task && head->task != task)
1163 io_for_each_link(req, head) {
1164 if (req->flags & REQ_F_INFLIGHT)
1170 static inline void req_set_fail(struct io_kiocb *req)
1172 req->flags |= REQ_F_FAIL;
1175 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1177 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1179 complete(&ctx->ref_comp);
1182 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1184 return !req->timeout.off;
1187 static void io_fallback_req_func(struct work_struct *work)
1189 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1190 fallback_work.work);
1191 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1192 struct io_kiocb *req, *tmp;
1194 percpu_ref_get(&ctx->refs);
1195 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1196 req->io_task_work.func(req);
1197 percpu_ref_put(&ctx->refs);
1200 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1202 struct io_ring_ctx *ctx;
1205 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1210 * Use 5 bits less than the max cq entries, that should give us around
1211 * 32 entries per hash list if totally full and uniformly spread.
1213 hash_bits = ilog2(p->cq_entries);
1217 ctx->cancel_hash_bits = hash_bits;
1218 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1220 if (!ctx->cancel_hash)
1222 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1224 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1225 if (!ctx->dummy_ubuf)
1227 /* set invalid range, so io_import_fixed() fails meeting it */
1228 ctx->dummy_ubuf->ubuf = -1UL;
1230 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1231 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1234 ctx->flags = p->flags;
1235 init_waitqueue_head(&ctx->sqo_sq_wait);
1236 INIT_LIST_HEAD(&ctx->sqd_list);
1237 init_waitqueue_head(&ctx->poll_wait);
1238 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1239 init_completion(&ctx->ref_comp);
1240 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1241 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1242 mutex_init(&ctx->uring_lock);
1243 init_waitqueue_head(&ctx->cq_wait);
1244 spin_lock_init(&ctx->completion_lock);
1245 spin_lock_init(&ctx->timeout_lock);
1246 INIT_LIST_HEAD(&ctx->iopoll_list);
1247 INIT_LIST_HEAD(&ctx->defer_list);
1248 INIT_LIST_HEAD(&ctx->timeout_list);
1249 spin_lock_init(&ctx->rsrc_ref_lock);
1250 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1251 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1252 init_llist_head(&ctx->rsrc_put_llist);
1253 INIT_LIST_HEAD(&ctx->tctx_list);
1254 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1255 INIT_LIST_HEAD(&ctx->locked_free_list);
1256 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1259 kfree(ctx->dummy_ubuf);
1260 kfree(ctx->cancel_hash);
1265 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1267 struct io_rings *r = ctx->rings;
1269 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1273 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1275 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1276 struct io_ring_ctx *ctx = req->ctx;
1278 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1284 #define FFS_ASYNC_READ 0x1UL
1285 #define FFS_ASYNC_WRITE 0x2UL
1287 #define FFS_ISREG 0x4UL
1289 #define FFS_ISREG 0x0UL
1291 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1293 static inline bool io_req_ffs_set(struct io_kiocb *req)
1295 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1298 static void io_req_track_inflight(struct io_kiocb *req)
1300 if (!(req->flags & REQ_F_INFLIGHT)) {
1301 req->flags |= REQ_F_INFLIGHT;
1302 atomic_inc(¤t->io_uring->inflight_tracked);
1306 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1308 struct io_kiocb *nxt = req->link;
1310 if (req->flags & REQ_F_LINK_TIMEOUT)
1313 /* linked timeouts should have two refs once prep'ed */
1314 io_req_set_refcount(req);
1315 io_req_set_refcount(nxt);
1318 nxt->timeout.head = req;
1319 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
1320 req->flags |= REQ_F_LINK_TIMEOUT;
1324 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1326 if (likely(!req->link || req->link->opcode != IORING_OP_LINK_TIMEOUT))
1328 return __io_prep_linked_timeout(req);
1331 static void io_prep_async_work(struct io_kiocb *req)
1333 const struct io_op_def *def = &io_op_defs[req->opcode];
1334 struct io_ring_ctx *ctx = req->ctx;
1336 if (!(req->flags & REQ_F_CREDS)) {
1337 req->flags |= REQ_F_CREDS;
1338 req->creds = get_current_cred();
1341 req->work.list.next = NULL;
1342 req->work.flags = 0;
1343 if (req->flags & REQ_F_FORCE_ASYNC)
1344 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1346 if (req->flags & REQ_F_ISREG) {
1347 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1348 io_wq_hash_work(&req->work, file_inode(req->file));
1349 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1350 if (def->unbound_nonreg_file)
1351 req->work.flags |= IO_WQ_WORK_UNBOUND;
1354 switch (req->opcode) {
1355 case IORING_OP_SPLICE:
1357 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1358 req->work.flags |= IO_WQ_WORK_UNBOUND;
1363 static void io_prep_async_link(struct io_kiocb *req)
1365 struct io_kiocb *cur;
1367 if (req->flags & REQ_F_LINK_TIMEOUT) {
1368 struct io_ring_ctx *ctx = req->ctx;
1370 spin_lock(&ctx->completion_lock);
1371 io_for_each_link(cur, req)
1372 io_prep_async_work(cur);
1373 spin_unlock(&ctx->completion_lock);
1375 io_for_each_link(cur, req)
1376 io_prep_async_work(cur);
1380 static void io_queue_async_work(struct io_kiocb *req)
1382 struct io_ring_ctx *ctx = req->ctx;
1383 struct io_kiocb *link = io_prep_linked_timeout(req);
1384 struct io_uring_task *tctx = req->task->io_uring;
1387 BUG_ON(!tctx->io_wq);
1389 /* init ->work of the whole link before punting */
1390 io_prep_async_link(req);
1393 * Not expected to happen, but if we do have a bug where this _can_
1394 * happen, catch it here and ensure the request is marked as
1395 * canceled. That will make io-wq go through the usual work cancel
1396 * procedure rather than attempt to run this request (or create a new
1399 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1400 req->work.flags |= IO_WQ_WORK_CANCEL;
1402 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1403 &req->work, req->flags);
1404 io_wq_enqueue(tctx->io_wq, &req->work);
1406 io_queue_linked_timeout(link);
1409 static void io_kill_timeout(struct io_kiocb *req, int status)
1410 __must_hold(&req->ctx->completion_lock)
1411 __must_hold(&req->ctx->timeout_lock)
1413 struct io_timeout_data *io = req->async_data;
1415 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1416 atomic_set(&req->ctx->cq_timeouts,
1417 atomic_read(&req->ctx->cq_timeouts) + 1);
1418 list_del_init(&req->timeout.list);
1419 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1420 io_put_req_deferred(req);
1424 static void io_queue_deferred(struct io_ring_ctx *ctx)
1426 while (!list_empty(&ctx->defer_list)) {
1427 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1428 struct io_defer_entry, list);
1430 if (req_need_defer(de->req, de->seq))
1432 list_del_init(&de->list);
1433 io_req_task_queue(de->req);
1438 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1439 __must_hold(&ctx->completion_lock)
1441 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1443 spin_lock_irq(&ctx->timeout_lock);
1444 while (!list_empty(&ctx->timeout_list)) {
1445 u32 events_needed, events_got;
1446 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1447 struct io_kiocb, timeout.list);
1449 if (io_is_timeout_noseq(req))
1453 * Since seq can easily wrap around over time, subtract
1454 * the last seq at which timeouts were flushed before comparing.
1455 * Assuming not more than 2^31-1 events have happened since,
1456 * these subtractions won't have wrapped, so we can check if
1457 * target is in [last_seq, current_seq] by comparing the two.
1459 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1460 events_got = seq - ctx->cq_last_tm_flush;
1461 if (events_got < events_needed)
1464 list_del_init(&req->timeout.list);
1465 io_kill_timeout(req, 0);
1467 ctx->cq_last_tm_flush = seq;
1468 spin_unlock_irq(&ctx->timeout_lock);
1471 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1473 if (ctx->off_timeout_used)
1474 io_flush_timeouts(ctx);
1475 if (ctx->drain_active)
1476 io_queue_deferred(ctx);
1479 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1481 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1482 __io_commit_cqring_flush(ctx);
1483 /* order cqe stores with ring update */
1484 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1487 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1489 struct io_rings *r = ctx->rings;
1491 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1494 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1496 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1499 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1501 struct io_rings *rings = ctx->rings;
1502 unsigned tail, mask = ctx->cq_entries - 1;
1505 * writes to the cq entry need to come after reading head; the
1506 * control dependency is enough as we're using WRITE_ONCE to
1509 if (__io_cqring_events(ctx) == ctx->cq_entries)
1512 tail = ctx->cached_cq_tail++;
1513 return &rings->cqes[tail & mask];
1516 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1518 if (likely(!ctx->cq_ev_fd))
1520 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1522 return !ctx->eventfd_async || io_wq_current_is_worker();
1525 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1528 * wake_up_all() may seem excessive, but io_wake_function() and
1529 * io_should_wake() handle the termination of the loop and only
1530 * wake as many waiters as we need to.
1532 if (wq_has_sleeper(&ctx->cq_wait))
1533 wake_up_all(&ctx->cq_wait);
1534 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1535 wake_up(&ctx->sq_data->wait);
1536 if (io_should_trigger_evfd(ctx))
1537 eventfd_signal(ctx->cq_ev_fd, 1);
1538 if (waitqueue_active(&ctx->poll_wait)) {
1539 wake_up_interruptible(&ctx->poll_wait);
1540 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1544 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1546 if (ctx->flags & IORING_SETUP_SQPOLL) {
1547 if (wq_has_sleeper(&ctx->cq_wait))
1548 wake_up_all(&ctx->cq_wait);
1550 if (io_should_trigger_evfd(ctx))
1551 eventfd_signal(ctx->cq_ev_fd, 1);
1552 if (waitqueue_active(&ctx->poll_wait)) {
1553 wake_up_interruptible(&ctx->poll_wait);
1554 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1558 /* Returns true if there are no backlogged entries after the flush */
1559 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1561 bool all_flushed, posted;
1563 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1567 spin_lock(&ctx->completion_lock);
1568 while (!list_empty(&ctx->cq_overflow_list)) {
1569 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1570 struct io_overflow_cqe *ocqe;
1574 ocqe = list_first_entry(&ctx->cq_overflow_list,
1575 struct io_overflow_cqe, list);
1577 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1579 io_account_cq_overflow(ctx);
1582 list_del(&ocqe->list);
1586 all_flushed = list_empty(&ctx->cq_overflow_list);
1588 clear_bit(0, &ctx->check_cq_overflow);
1589 WRITE_ONCE(ctx->rings->sq_flags,
1590 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1594 io_commit_cqring(ctx);
1595 spin_unlock(&ctx->completion_lock);
1597 io_cqring_ev_posted(ctx);
1601 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1605 if (test_bit(0, &ctx->check_cq_overflow)) {
1606 /* iopoll syncs against uring_lock, not completion_lock */
1607 if (ctx->flags & IORING_SETUP_IOPOLL)
1608 mutex_lock(&ctx->uring_lock);
1609 ret = __io_cqring_overflow_flush(ctx, false);
1610 if (ctx->flags & IORING_SETUP_IOPOLL)
1611 mutex_unlock(&ctx->uring_lock);
1617 /* must to be called somewhat shortly after putting a request */
1618 static inline void io_put_task(struct task_struct *task, int nr)
1620 struct io_uring_task *tctx = task->io_uring;
1622 percpu_counter_sub(&tctx->inflight, nr);
1623 if (unlikely(atomic_read(&tctx->in_idle)))
1624 wake_up(&tctx->wait);
1625 put_task_struct_many(task, nr);
1628 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1629 long res, unsigned int cflags)
1631 struct io_overflow_cqe *ocqe;
1633 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1636 * If we're in ring overflow flush mode, or in task cancel mode,
1637 * or cannot allocate an overflow entry, then we need to drop it
1640 io_account_cq_overflow(ctx);
1643 if (list_empty(&ctx->cq_overflow_list)) {
1644 set_bit(0, &ctx->check_cq_overflow);
1645 WRITE_ONCE(ctx->rings->sq_flags,
1646 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1649 ocqe->cqe.user_data = user_data;
1650 ocqe->cqe.res = res;
1651 ocqe->cqe.flags = cflags;
1652 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1656 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1657 long res, unsigned int cflags)
1659 struct io_uring_cqe *cqe;
1661 trace_io_uring_complete(ctx, user_data, res, cflags);
1664 * If we can't get a cq entry, userspace overflowed the
1665 * submission (by quite a lot). Increment the overflow count in
1668 cqe = io_get_cqe(ctx);
1670 WRITE_ONCE(cqe->user_data, user_data);
1671 WRITE_ONCE(cqe->res, res);
1672 WRITE_ONCE(cqe->flags, cflags);
1675 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1678 /* not as hot to bloat with inlining */
1679 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1680 long res, unsigned int cflags)
1682 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1685 static void io_req_complete_post(struct io_kiocb *req, long res,
1686 unsigned int cflags)
1688 struct io_ring_ctx *ctx = req->ctx;
1690 spin_lock(&ctx->completion_lock);
1691 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1693 * If we're the last reference to this request, add to our locked
1696 if (req_ref_put_and_test(req)) {
1697 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1698 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1699 io_disarm_next(req);
1701 io_req_task_queue(req->link);
1705 io_dismantle_req(req);
1706 io_put_task(req->task, 1);
1707 list_add(&req->inflight_entry, &ctx->locked_free_list);
1708 ctx->locked_free_nr++;
1710 if (!percpu_ref_tryget(&ctx->refs))
1713 io_commit_cqring(ctx);
1714 spin_unlock(&ctx->completion_lock);
1717 io_cqring_ev_posted(ctx);
1718 percpu_ref_put(&ctx->refs);
1722 static inline bool io_req_needs_clean(struct io_kiocb *req)
1724 return req->flags & IO_REQ_CLEAN_FLAGS;
1727 static void io_req_complete_state(struct io_kiocb *req, long res,
1728 unsigned int cflags)
1730 if (io_req_needs_clean(req))
1733 req->compl.cflags = cflags;
1734 req->flags |= REQ_F_COMPLETE_INLINE;
1737 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1738 long res, unsigned cflags)
1740 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1741 io_req_complete_state(req, res, cflags);
1743 io_req_complete_post(req, res, cflags);
1746 static inline void io_req_complete(struct io_kiocb *req, long res)
1748 __io_req_complete(req, 0, res, 0);
1751 static void io_req_complete_failed(struct io_kiocb *req, long res)
1754 io_req_complete_post(req, res, 0);
1758 * Don't initialise the fields below on every allocation, but do that in
1759 * advance and keep them valid across allocations.
1761 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1765 req->async_data = NULL;
1766 /* not necessary, but safer to zero */
1770 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1771 struct io_submit_state *state)
1773 spin_lock(&ctx->completion_lock);
1774 list_splice_init(&ctx->locked_free_list, &state->free_list);
1775 ctx->locked_free_nr = 0;
1776 spin_unlock(&ctx->completion_lock);
1779 /* Returns true IFF there are requests in the cache */
1780 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1782 struct io_submit_state *state = &ctx->submit_state;
1786 * If we have more than a batch's worth of requests in our IRQ side
1787 * locked cache, grab the lock and move them over to our submission
1790 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1791 io_flush_cached_locked_reqs(ctx, state);
1793 nr = state->free_reqs;
1794 while (!list_empty(&state->free_list)) {
1795 struct io_kiocb *req = list_first_entry(&state->free_list,
1796 struct io_kiocb, inflight_entry);
1798 list_del(&req->inflight_entry);
1799 state->reqs[nr++] = req;
1800 if (nr == ARRAY_SIZE(state->reqs))
1804 state->free_reqs = nr;
1809 * A request might get retired back into the request caches even before opcode
1810 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1811 * Because of that, io_alloc_req() should be called only under ->uring_lock
1812 * and with extra caution to not get a request that is still worked on.
1814 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1815 __must_hold(&ctx->uring_lock)
1817 struct io_submit_state *state = &ctx->submit_state;
1818 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1821 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1823 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1826 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1830 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1831 * retry single alloc to be on the safe side.
1833 if (unlikely(ret <= 0)) {
1834 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1835 if (!state->reqs[0])
1840 for (i = 0; i < ret; i++)
1841 io_preinit_req(state->reqs[i], ctx);
1842 state->free_reqs = ret;
1845 return state->reqs[state->free_reqs];
1848 static inline void io_put_file(struct file *file)
1854 static void io_dismantle_req(struct io_kiocb *req)
1856 unsigned int flags = req->flags;
1858 if (io_req_needs_clean(req))
1860 if (!(flags & REQ_F_FIXED_FILE))
1861 io_put_file(req->file);
1862 if (req->fixed_rsrc_refs)
1863 percpu_ref_put(req->fixed_rsrc_refs);
1864 if (req->async_data) {
1865 kfree(req->async_data);
1866 req->async_data = NULL;
1870 static void __io_free_req(struct io_kiocb *req)
1872 struct io_ring_ctx *ctx = req->ctx;
1874 io_dismantle_req(req);
1875 io_put_task(req->task, 1);
1877 spin_lock(&ctx->completion_lock);
1878 list_add(&req->inflight_entry, &ctx->locked_free_list);
1879 ctx->locked_free_nr++;
1880 spin_unlock(&ctx->completion_lock);
1882 percpu_ref_put(&ctx->refs);
1885 static inline void io_remove_next_linked(struct io_kiocb *req)
1887 struct io_kiocb *nxt = req->link;
1889 req->link = nxt->link;
1893 static bool io_kill_linked_timeout(struct io_kiocb *req)
1894 __must_hold(&req->ctx->completion_lock)
1895 __must_hold(&req->ctx->timeout_lock)
1897 struct io_kiocb *link = req->link;
1900 * Can happen if a linked timeout fired and link had been like
1901 * req -> link t-out -> link t-out [-> ...]
1903 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1904 struct io_timeout_data *io = link->async_data;
1906 io_remove_next_linked(req);
1907 link->timeout.head = NULL;
1908 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1909 io_cqring_fill_event(link->ctx, link->user_data,
1911 io_put_req_deferred(link);
1918 static void io_fail_links(struct io_kiocb *req)
1919 __must_hold(&req->ctx->completion_lock)
1921 struct io_kiocb *nxt, *link = req->link;
1928 trace_io_uring_fail_link(req, link);
1929 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1930 io_put_req_deferred(link);
1935 static bool io_disarm_next(struct io_kiocb *req)
1936 __must_hold(&req->ctx->completion_lock)
1938 bool posted = false;
1940 if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1941 struct io_ring_ctx *ctx = req->ctx;
1943 spin_lock_irq(&ctx->timeout_lock);
1944 posted = io_kill_linked_timeout(req);
1945 spin_unlock_irq(&ctx->timeout_lock);
1947 if (unlikely((req->flags & REQ_F_FAIL) &&
1948 !(req->flags & REQ_F_HARDLINK))) {
1949 posted |= (req->link != NULL);
1955 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1957 struct io_kiocb *nxt;
1960 * If LINK is set, we have dependent requests in this chain. If we
1961 * didn't fail this request, queue the first one up, moving any other
1962 * dependencies to the next request. In case of failure, fail the rest
1965 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1966 struct io_ring_ctx *ctx = req->ctx;
1969 spin_lock(&ctx->completion_lock);
1970 posted = io_disarm_next(req);
1972 io_commit_cqring(req->ctx);
1973 spin_unlock(&ctx->completion_lock);
1975 io_cqring_ev_posted(ctx);
1982 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1984 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1986 return __io_req_find_next(req);
1989 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1993 if (ctx->submit_state.compl_nr) {
1994 mutex_lock(&ctx->uring_lock);
1995 io_submit_flush_completions(ctx);
1996 mutex_unlock(&ctx->uring_lock);
1998 percpu_ref_put(&ctx->refs);
2001 static void tctx_task_work(struct callback_head *cb)
2003 struct io_ring_ctx *ctx = NULL;
2004 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2008 struct io_wq_work_node *node;
2010 spin_lock_irq(&tctx->task_lock);
2011 node = tctx->task_list.first;
2012 INIT_WQ_LIST(&tctx->task_list);
2014 tctx->task_running = false;
2015 spin_unlock_irq(&tctx->task_lock);
2020 struct io_wq_work_node *next = node->next;
2021 struct io_kiocb *req = container_of(node, struct io_kiocb,
2024 if (req->ctx != ctx) {
2025 ctx_flush_and_put(ctx);
2027 percpu_ref_get(&ctx->refs);
2029 req->io_task_work.func(req);
2036 ctx_flush_and_put(ctx);
2039 static void io_req_task_work_add(struct io_kiocb *req)
2041 struct task_struct *tsk = req->task;
2042 struct io_uring_task *tctx = tsk->io_uring;
2043 enum task_work_notify_mode notify;
2044 struct io_wq_work_node *node;
2045 unsigned long flags;
2048 WARN_ON_ONCE(!tctx);
2050 spin_lock_irqsave(&tctx->task_lock, flags);
2051 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2052 running = tctx->task_running;
2054 tctx->task_running = true;
2055 spin_unlock_irqrestore(&tctx->task_lock, flags);
2057 /* task_work already pending, we're done */
2062 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2063 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2064 * processing task_work. There's no reliable way to tell if TWA_RESUME
2067 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2068 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2069 wake_up_process(tsk);
2073 spin_lock_irqsave(&tctx->task_lock, flags);
2074 tctx->task_running = false;
2075 node = tctx->task_list.first;
2076 INIT_WQ_LIST(&tctx->task_list);
2077 spin_unlock_irqrestore(&tctx->task_lock, flags);
2080 req = container_of(node, struct io_kiocb, io_task_work.node);
2082 if (llist_add(&req->io_task_work.fallback_node,
2083 &req->ctx->fallback_llist))
2084 schedule_delayed_work(&req->ctx->fallback_work, 1);
2088 static void io_req_task_cancel(struct io_kiocb *req)
2090 struct io_ring_ctx *ctx = req->ctx;
2092 /* ctx is guaranteed to stay alive while we hold uring_lock */
2093 mutex_lock(&ctx->uring_lock);
2094 io_req_complete_failed(req, req->result);
2095 mutex_unlock(&ctx->uring_lock);
2098 static void io_req_task_submit(struct io_kiocb *req)
2100 struct io_ring_ctx *ctx = req->ctx;
2102 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2103 mutex_lock(&ctx->uring_lock);
2104 if (likely(!(req->task->flags & PF_EXITING)))
2105 __io_queue_sqe(req);
2107 io_req_complete_failed(req, -EFAULT);
2108 mutex_unlock(&ctx->uring_lock);
2111 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2114 req->io_task_work.func = io_req_task_cancel;
2115 io_req_task_work_add(req);
2118 static void io_req_task_queue(struct io_kiocb *req)
2120 req->io_task_work.func = io_req_task_submit;
2121 io_req_task_work_add(req);
2124 static void io_req_task_queue_reissue(struct io_kiocb *req)
2126 req->io_task_work.func = io_queue_async_work;
2127 io_req_task_work_add(req);
2130 static inline void io_queue_next(struct io_kiocb *req)
2132 struct io_kiocb *nxt = io_req_find_next(req);
2135 io_req_task_queue(nxt);
2138 static void io_free_req(struct io_kiocb *req)
2145 struct task_struct *task;
2150 static inline void io_init_req_batch(struct req_batch *rb)
2157 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2158 struct req_batch *rb)
2161 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2162 if (rb->task == current)
2163 current->io_uring->cached_refs += rb->task_refs;
2165 io_put_task(rb->task, rb->task_refs);
2168 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2169 struct io_submit_state *state)
2172 io_dismantle_req(req);
2174 if (req->task != rb->task) {
2176 io_put_task(rb->task, rb->task_refs);
2177 rb->task = req->task;
2183 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2184 state->reqs[state->free_reqs++] = req;
2186 list_add(&req->inflight_entry, &state->free_list);
2189 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2190 __must_hold(&ctx->uring_lock)
2192 struct io_submit_state *state = &ctx->submit_state;
2193 int i, nr = state->compl_nr;
2194 struct req_batch rb;
2196 spin_lock(&ctx->completion_lock);
2197 for (i = 0; i < nr; i++) {
2198 struct io_kiocb *req = state->compl_reqs[i];
2200 __io_cqring_fill_event(ctx, req->user_data, req->result,
2203 io_commit_cqring(ctx);
2204 spin_unlock(&ctx->completion_lock);
2205 io_cqring_ev_posted(ctx);
2207 io_init_req_batch(&rb);
2208 for (i = 0; i < nr; i++) {
2209 struct io_kiocb *req = state->compl_reqs[i];
2211 if (req_ref_put_and_test(req))
2212 io_req_free_batch(&rb, req, &ctx->submit_state);
2215 io_req_free_batch_finish(ctx, &rb);
2216 state->compl_nr = 0;
2220 * Drop reference to request, return next in chain (if there is one) if this
2221 * was the last reference to this request.
2223 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2225 struct io_kiocb *nxt = NULL;
2227 if (req_ref_put_and_test(req)) {
2228 nxt = io_req_find_next(req);
2234 static inline void io_put_req(struct io_kiocb *req)
2236 if (req_ref_put_and_test(req))
2240 static inline void io_put_req_deferred(struct io_kiocb *req)
2242 if (req_ref_put_and_test(req)) {
2243 req->io_task_work.func = io_free_req;
2244 io_req_task_work_add(req);
2248 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2250 /* See comment at the top of this file */
2252 return __io_cqring_events(ctx);
2255 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2257 struct io_rings *rings = ctx->rings;
2259 /* make sure SQ entry isn't read before tail */
2260 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2263 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2265 unsigned int cflags;
2267 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2268 cflags |= IORING_CQE_F_BUFFER;
2269 req->flags &= ~REQ_F_BUFFER_SELECTED;
2274 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2276 struct io_buffer *kbuf;
2278 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2279 return io_put_kbuf(req, kbuf);
2282 static inline bool io_run_task_work(void)
2284 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2285 __set_current_state(TASK_RUNNING);
2286 tracehook_notify_signal();
2294 * Find and free completed poll iocbs
2296 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2297 struct list_head *done, bool resubmit)
2299 struct req_batch rb;
2300 struct io_kiocb *req;
2302 /* order with ->result store in io_complete_rw_iopoll() */
2305 io_init_req_batch(&rb);
2306 while (!list_empty(done)) {
2309 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2310 list_del(&req->inflight_entry);
2312 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2313 !(req->flags & REQ_F_DONT_REISSUE)) {
2314 req->iopoll_completed = 0;
2315 io_req_task_queue_reissue(req);
2319 if (req->flags & REQ_F_BUFFER_SELECTED)
2320 cflags = io_put_rw_kbuf(req);
2322 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2325 if (req_ref_put_and_test(req))
2326 io_req_free_batch(&rb, req, &ctx->submit_state);
2329 io_commit_cqring(ctx);
2330 io_cqring_ev_posted_iopoll(ctx);
2331 io_req_free_batch_finish(ctx, &rb);
2334 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2335 long min, bool resubmit)
2337 struct io_kiocb *req, *tmp;
2342 * Only spin for completions if we don't have multiple devices hanging
2343 * off our complete list, and we're under the requested amount.
2345 spin = !ctx->poll_multi_queue && *nr_events < min;
2347 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2348 struct kiocb *kiocb = &req->rw.kiocb;
2352 * Move completed and retryable entries to our local lists.
2353 * If we find a request that requires polling, break out
2354 * and complete those lists first, if we have entries there.
2356 if (READ_ONCE(req->iopoll_completed)) {
2357 list_move_tail(&req->inflight_entry, &done);
2360 if (!list_empty(&done))
2363 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2364 if (unlikely(ret < 0))
2369 /* iopoll may have completed current req */
2370 if (READ_ONCE(req->iopoll_completed))
2371 list_move_tail(&req->inflight_entry, &done);
2374 if (!list_empty(&done))
2375 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2381 * We can't just wait for polled events to come to us, we have to actively
2382 * find and complete them.
2384 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2386 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2389 mutex_lock(&ctx->uring_lock);
2390 while (!list_empty(&ctx->iopoll_list)) {
2391 unsigned int nr_events = 0;
2393 io_do_iopoll(ctx, &nr_events, 0, false);
2395 /* let it sleep and repeat later if can't complete a request */
2399 * Ensure we allow local-to-the-cpu processing to take place,
2400 * in this case we need to ensure that we reap all events.
2401 * Also let task_work, etc. to progress by releasing the mutex
2403 if (need_resched()) {
2404 mutex_unlock(&ctx->uring_lock);
2406 mutex_lock(&ctx->uring_lock);
2409 mutex_unlock(&ctx->uring_lock);
2412 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2414 unsigned int nr_events = 0;
2418 * We disallow the app entering submit/complete with polling, but we
2419 * still need to lock the ring to prevent racing with polled issue
2420 * that got punted to a workqueue.
2422 mutex_lock(&ctx->uring_lock);
2424 * Don't enter poll loop if we already have events pending.
2425 * If we do, we can potentially be spinning for commands that
2426 * already triggered a CQE (eg in error).
2428 if (test_bit(0, &ctx->check_cq_overflow))
2429 __io_cqring_overflow_flush(ctx, false);
2430 if (io_cqring_events(ctx))
2434 * If a submit got punted to a workqueue, we can have the
2435 * application entering polling for a command before it gets
2436 * issued. That app will hold the uring_lock for the duration
2437 * of the poll right here, so we need to take a breather every
2438 * now and then to ensure that the issue has a chance to add
2439 * the poll to the issued list. Otherwise we can spin here
2440 * forever, while the workqueue is stuck trying to acquire the
2443 if (list_empty(&ctx->iopoll_list)) {
2444 u32 tail = ctx->cached_cq_tail;
2446 mutex_unlock(&ctx->uring_lock);
2448 mutex_lock(&ctx->uring_lock);
2450 /* some requests don't go through iopoll_list */
2451 if (tail != ctx->cached_cq_tail ||
2452 list_empty(&ctx->iopoll_list))
2455 ret = io_do_iopoll(ctx, &nr_events, min, true);
2456 } while (!ret && nr_events < min && !need_resched());
2458 mutex_unlock(&ctx->uring_lock);
2462 static void kiocb_end_write(struct io_kiocb *req)
2465 * Tell lockdep we inherited freeze protection from submission
2468 if (req->flags & REQ_F_ISREG) {
2469 struct super_block *sb = file_inode(req->file)->i_sb;
2471 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2477 static bool io_resubmit_prep(struct io_kiocb *req)
2479 struct io_async_rw *rw = req->async_data;
2482 return !io_req_prep_async(req);
2483 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2484 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2488 static bool io_rw_should_reissue(struct io_kiocb *req)
2490 umode_t mode = file_inode(req->file)->i_mode;
2491 struct io_ring_ctx *ctx = req->ctx;
2493 if (!S_ISBLK(mode) && !S_ISREG(mode))
2495 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2496 !(ctx->flags & IORING_SETUP_IOPOLL)))
2499 * If ref is dying, we might be running poll reap from the exit work.
2500 * Don't attempt to reissue from that path, just let it fail with
2503 if (percpu_ref_is_dying(&ctx->refs))
2506 * Play it safe and assume not safe to re-import and reissue if we're
2507 * not in the original thread group (or in task context).
2509 if (!same_thread_group(req->task, current) || !in_task())
2514 static bool io_resubmit_prep(struct io_kiocb *req)
2518 static bool io_rw_should_reissue(struct io_kiocb *req)
2524 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2526 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2527 kiocb_end_write(req);
2528 if (res != req->result) {
2529 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2530 io_rw_should_reissue(req)) {
2531 req->flags |= REQ_F_REISSUE;
2540 static void io_req_task_complete(struct io_kiocb *req)
2544 if (req->flags & REQ_F_BUFFER_SELECTED)
2545 cflags = io_put_rw_kbuf(req);
2546 __io_req_complete(req, 0, req->result, cflags);
2549 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2550 unsigned int issue_flags)
2552 if (__io_complete_rw_common(req, res))
2554 io_req_task_complete(req);
2557 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2559 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2561 if (__io_complete_rw_common(req, res))
2564 req->io_task_work.func = io_req_task_complete;
2565 io_req_task_work_add(req);
2568 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2570 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2572 if (kiocb->ki_flags & IOCB_WRITE)
2573 kiocb_end_write(req);
2574 if (unlikely(res != req->result)) {
2575 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2576 io_resubmit_prep(req))) {
2578 req->flags |= REQ_F_DONT_REISSUE;
2582 WRITE_ONCE(req->result, res);
2583 /* order with io_iopoll_complete() checking ->result */
2585 WRITE_ONCE(req->iopoll_completed, 1);
2589 * After the iocb has been issued, it's safe to be found on the poll list.
2590 * Adding the kiocb to the list AFTER submission ensures that we don't
2591 * find it from a io_do_iopoll() thread before the issuer is done
2592 * accessing the kiocb cookie.
2594 static void io_iopoll_req_issued(struct io_kiocb *req)
2596 struct io_ring_ctx *ctx = req->ctx;
2597 const bool in_async = io_wq_current_is_worker();
2599 /* workqueue context doesn't hold uring_lock, grab it now */
2600 if (unlikely(in_async))
2601 mutex_lock(&ctx->uring_lock);
2604 * Track whether we have multiple files in our lists. This will impact
2605 * how we do polling eventually, not spinning if we're on potentially
2606 * different devices.
2608 if (list_empty(&ctx->iopoll_list)) {
2609 ctx->poll_multi_queue = false;
2610 } else if (!ctx->poll_multi_queue) {
2611 struct io_kiocb *list_req;
2612 unsigned int queue_num0, queue_num1;
2614 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2617 if (list_req->file != req->file) {
2618 ctx->poll_multi_queue = true;
2620 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2621 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2622 if (queue_num0 != queue_num1)
2623 ctx->poll_multi_queue = true;
2628 * For fast devices, IO may have already completed. If it has, add
2629 * it to the front so we find it first.
2631 if (READ_ONCE(req->iopoll_completed))
2632 list_add(&req->inflight_entry, &ctx->iopoll_list);
2634 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2636 if (unlikely(in_async)) {
2638 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2639 * in sq thread task context or in io worker task context. If
2640 * current task context is sq thread, we don't need to check
2641 * whether should wake up sq thread.
2643 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2644 wq_has_sleeper(&ctx->sq_data->wait))
2645 wake_up(&ctx->sq_data->wait);
2647 mutex_unlock(&ctx->uring_lock);
2651 static bool io_bdev_nowait(struct block_device *bdev)
2653 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2657 * If we tracked the file through the SCM inflight mechanism, we could support
2658 * any file. For now, just ensure that anything potentially problematic is done
2661 static bool __io_file_supports_nowait(struct file *file, int rw)
2663 umode_t mode = file_inode(file)->i_mode;
2665 if (S_ISBLK(mode)) {
2666 if (IS_ENABLED(CONFIG_BLOCK) &&
2667 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2673 if (S_ISREG(mode)) {
2674 if (IS_ENABLED(CONFIG_BLOCK) &&
2675 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2676 file->f_op != &io_uring_fops)
2681 /* any ->read/write should understand O_NONBLOCK */
2682 if (file->f_flags & O_NONBLOCK)
2685 if (!(file->f_mode & FMODE_NOWAIT))
2689 return file->f_op->read_iter != NULL;
2691 return file->f_op->write_iter != NULL;
2694 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2696 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2698 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2701 return __io_file_supports_nowait(req->file, rw);
2704 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2706 struct io_ring_ctx *ctx = req->ctx;
2707 struct kiocb *kiocb = &req->rw.kiocb;
2708 struct file *file = req->file;
2712 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2713 req->flags |= REQ_F_ISREG;
2715 kiocb->ki_pos = READ_ONCE(sqe->off);
2716 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2717 req->flags |= REQ_F_CUR_POS;
2718 kiocb->ki_pos = file->f_pos;
2720 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2721 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2722 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2726 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2727 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2728 req->flags |= REQ_F_NOWAIT;
2730 ioprio = READ_ONCE(sqe->ioprio);
2732 ret = ioprio_check_cap(ioprio);
2736 kiocb->ki_ioprio = ioprio;
2738 kiocb->ki_ioprio = get_current_ioprio();
2740 if (ctx->flags & IORING_SETUP_IOPOLL) {
2741 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2742 !kiocb->ki_filp->f_op->iopoll)
2745 kiocb->ki_flags |= IOCB_HIPRI;
2746 kiocb->ki_complete = io_complete_rw_iopoll;
2747 req->iopoll_completed = 0;
2749 if (kiocb->ki_flags & IOCB_HIPRI)
2751 kiocb->ki_complete = io_complete_rw;
2754 if (req->opcode == IORING_OP_READ_FIXED ||
2755 req->opcode == IORING_OP_WRITE_FIXED) {
2757 io_req_set_rsrc_node(req);
2760 req->rw.addr = READ_ONCE(sqe->addr);
2761 req->rw.len = READ_ONCE(sqe->len);
2762 req->buf_index = READ_ONCE(sqe->buf_index);
2766 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2772 case -ERESTARTNOINTR:
2773 case -ERESTARTNOHAND:
2774 case -ERESTART_RESTARTBLOCK:
2776 * We can't just restart the syscall, since previously
2777 * submitted sqes may already be in progress. Just fail this
2783 kiocb->ki_complete(kiocb, ret, 0);
2787 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2788 unsigned int issue_flags)
2790 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2791 struct io_async_rw *io = req->async_data;
2792 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2794 /* add previously done IO, if any */
2795 if (io && io->bytes_done > 0) {
2797 ret = io->bytes_done;
2799 ret += io->bytes_done;
2802 if (req->flags & REQ_F_CUR_POS)
2803 req->file->f_pos = kiocb->ki_pos;
2804 if (ret >= 0 && check_reissue)
2805 __io_complete_rw(req, ret, 0, issue_flags);
2807 io_rw_done(kiocb, ret);
2809 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2810 req->flags &= ~REQ_F_REISSUE;
2811 if (io_resubmit_prep(req)) {
2812 io_req_task_queue_reissue(req);
2817 if (req->flags & REQ_F_BUFFER_SELECTED)
2818 cflags = io_put_rw_kbuf(req);
2819 __io_req_complete(req, issue_flags, ret, cflags);
2824 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2825 struct io_mapped_ubuf *imu)
2827 size_t len = req->rw.len;
2828 u64 buf_end, buf_addr = req->rw.addr;
2831 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2833 /* not inside the mapped region */
2834 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2838 * May not be a start of buffer, set size appropriately
2839 * and advance us to the beginning.
2841 offset = buf_addr - imu->ubuf;
2842 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2846 * Don't use iov_iter_advance() here, as it's really slow for
2847 * using the latter parts of a big fixed buffer - it iterates
2848 * over each segment manually. We can cheat a bit here, because
2851 * 1) it's a BVEC iter, we set it up
2852 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2853 * first and last bvec
2855 * So just find our index, and adjust the iterator afterwards.
2856 * If the offset is within the first bvec (or the whole first
2857 * bvec, just use iov_iter_advance(). This makes it easier
2858 * since we can just skip the first segment, which may not
2859 * be PAGE_SIZE aligned.
2861 const struct bio_vec *bvec = imu->bvec;
2863 if (offset <= bvec->bv_len) {
2864 iov_iter_advance(iter, offset);
2866 unsigned long seg_skip;
2868 /* skip first vec */
2869 offset -= bvec->bv_len;
2870 seg_skip = 1 + (offset >> PAGE_SHIFT);
2872 iter->bvec = bvec + seg_skip;
2873 iter->nr_segs -= seg_skip;
2874 iter->count -= bvec->bv_len + offset;
2875 iter->iov_offset = offset & ~PAGE_MASK;
2882 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2884 struct io_ring_ctx *ctx = req->ctx;
2885 struct io_mapped_ubuf *imu = req->imu;
2886 u16 index, buf_index = req->buf_index;
2889 if (unlikely(buf_index >= ctx->nr_user_bufs))
2891 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2892 imu = READ_ONCE(ctx->user_bufs[index]);
2895 return __io_import_fixed(req, rw, iter, imu);
2898 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2901 mutex_unlock(&ctx->uring_lock);
2904 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2907 * "Normal" inline submissions always hold the uring_lock, since we
2908 * grab it from the system call. Same is true for the SQPOLL offload.
2909 * The only exception is when we've detached the request and issue it
2910 * from an async worker thread, grab the lock for that case.
2913 mutex_lock(&ctx->uring_lock);
2916 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2917 int bgid, struct io_buffer *kbuf,
2920 struct io_buffer *head;
2922 if (req->flags & REQ_F_BUFFER_SELECTED)
2925 io_ring_submit_lock(req->ctx, needs_lock);
2927 lockdep_assert_held(&req->ctx->uring_lock);
2929 head = xa_load(&req->ctx->io_buffers, bgid);
2931 if (!list_empty(&head->list)) {
2932 kbuf = list_last_entry(&head->list, struct io_buffer,
2934 list_del(&kbuf->list);
2937 xa_erase(&req->ctx->io_buffers, bgid);
2939 if (*len > kbuf->len)
2942 kbuf = ERR_PTR(-ENOBUFS);
2945 io_ring_submit_unlock(req->ctx, needs_lock);
2950 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2953 struct io_buffer *kbuf;
2956 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2957 bgid = req->buf_index;
2958 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2961 req->rw.addr = (u64) (unsigned long) kbuf;
2962 req->flags |= REQ_F_BUFFER_SELECTED;
2963 return u64_to_user_ptr(kbuf->addr);
2966 #ifdef CONFIG_COMPAT
2967 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2970 struct compat_iovec __user *uiov;
2971 compat_ssize_t clen;
2975 uiov = u64_to_user_ptr(req->rw.addr);
2976 if (!access_ok(uiov, sizeof(*uiov)))
2978 if (__get_user(clen, &uiov->iov_len))
2984 buf = io_rw_buffer_select(req, &len, needs_lock);
2986 return PTR_ERR(buf);
2987 iov[0].iov_base = buf;
2988 iov[0].iov_len = (compat_size_t) len;
2993 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2996 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3000 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3003 len = iov[0].iov_len;
3006 buf = io_rw_buffer_select(req, &len, needs_lock);
3008 return PTR_ERR(buf);
3009 iov[0].iov_base = buf;
3010 iov[0].iov_len = len;
3014 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3017 if (req->flags & REQ_F_BUFFER_SELECTED) {
3018 struct io_buffer *kbuf;
3020 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3021 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3022 iov[0].iov_len = kbuf->len;
3025 if (req->rw.len != 1)
3028 #ifdef CONFIG_COMPAT
3029 if (req->ctx->compat)
3030 return io_compat_import(req, iov, needs_lock);
3033 return __io_iov_buffer_select(req, iov, needs_lock);
3036 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3037 struct iov_iter *iter, bool needs_lock)
3039 void __user *buf = u64_to_user_ptr(req->rw.addr);
3040 size_t sqe_len = req->rw.len;
3041 u8 opcode = req->opcode;
3044 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3046 return io_import_fixed(req, rw, iter);
3049 /* buffer index only valid with fixed read/write, or buffer select */
3050 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3053 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3054 if (req->flags & REQ_F_BUFFER_SELECT) {
3055 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3057 return PTR_ERR(buf);
3058 req->rw.len = sqe_len;
3061 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3066 if (req->flags & REQ_F_BUFFER_SELECT) {
3067 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3069 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3074 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3078 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3080 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3084 * For files that don't have ->read_iter() and ->write_iter(), handle them
3085 * by looping over ->read() or ->write() manually.
3087 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3089 struct kiocb *kiocb = &req->rw.kiocb;
3090 struct file *file = req->file;
3094 * Don't support polled IO through this interface, and we can't
3095 * support non-blocking either. For the latter, this just causes
3096 * the kiocb to be handled from an async context.
3098 if (kiocb->ki_flags & IOCB_HIPRI)
3100 if (kiocb->ki_flags & IOCB_NOWAIT)
3103 while (iov_iter_count(iter)) {
3107 if (!iov_iter_is_bvec(iter)) {
3108 iovec = iov_iter_iovec(iter);
3110 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3111 iovec.iov_len = req->rw.len;
3115 nr = file->f_op->read(file, iovec.iov_base,
3116 iovec.iov_len, io_kiocb_ppos(kiocb));
3118 nr = file->f_op->write(file, iovec.iov_base,
3119 iovec.iov_len, io_kiocb_ppos(kiocb));
3128 if (nr != iovec.iov_len)
3132 iov_iter_advance(iter, nr);
3138 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3139 const struct iovec *fast_iov, struct iov_iter *iter)
3141 struct io_async_rw *rw = req->async_data;
3143 memcpy(&rw->iter, iter, sizeof(*iter));
3144 rw->free_iovec = iovec;
3146 /* can only be fixed buffers, no need to do anything */
3147 if (iov_iter_is_bvec(iter))
3150 unsigned iov_off = 0;
3152 rw->iter.iov = rw->fast_iov;
3153 if (iter->iov != fast_iov) {
3154 iov_off = iter->iov - fast_iov;
3155 rw->iter.iov += iov_off;
3157 if (rw->fast_iov != fast_iov)
3158 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3159 sizeof(struct iovec) * iter->nr_segs);
3161 req->flags |= REQ_F_NEED_CLEANUP;
3165 static inline int io_alloc_async_data(struct io_kiocb *req)
3167 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3168 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3169 return req->async_data == NULL;
3172 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3173 const struct iovec *fast_iov,
3174 struct iov_iter *iter, bool force)
3176 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3178 if (!req->async_data) {
3179 if (io_alloc_async_data(req)) {
3184 io_req_map_rw(req, iovec, fast_iov, iter);
3189 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3191 struct io_async_rw *iorw = req->async_data;
3192 struct iovec *iov = iorw->fast_iov;
3195 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3196 if (unlikely(ret < 0))
3199 iorw->bytes_done = 0;
3200 iorw->free_iovec = iov;
3202 req->flags |= REQ_F_NEED_CLEANUP;
3206 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3208 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3210 return io_prep_rw(req, sqe);
3214 * This is our waitqueue callback handler, registered through lock_page_async()
3215 * when we initially tried to do the IO with the iocb armed our waitqueue.
3216 * This gets called when the page is unlocked, and we generally expect that to
3217 * happen when the page IO is completed and the page is now uptodate. This will
3218 * queue a task_work based retry of the operation, attempting to copy the data
3219 * again. If the latter fails because the page was NOT uptodate, then we will
3220 * do a thread based blocking retry of the operation. That's the unexpected
3223 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3224 int sync, void *arg)
3226 struct wait_page_queue *wpq;
3227 struct io_kiocb *req = wait->private;
3228 struct wait_page_key *key = arg;
3230 wpq = container_of(wait, struct wait_page_queue, wait);
3232 if (!wake_page_match(wpq, key))
3235 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3236 list_del_init(&wait->entry);
3237 io_req_task_queue(req);
3242 * This controls whether a given IO request should be armed for async page
3243 * based retry. If we return false here, the request is handed to the async
3244 * worker threads for retry. If we're doing buffered reads on a regular file,
3245 * we prepare a private wait_page_queue entry and retry the operation. This
3246 * will either succeed because the page is now uptodate and unlocked, or it
3247 * will register a callback when the page is unlocked at IO completion. Through
3248 * that callback, io_uring uses task_work to setup a retry of the operation.
3249 * That retry will attempt the buffered read again. The retry will generally
3250 * succeed, or in rare cases where it fails, we then fall back to using the
3251 * async worker threads for a blocking retry.
3253 static bool io_rw_should_retry(struct io_kiocb *req)
3255 struct io_async_rw *rw = req->async_data;
3256 struct wait_page_queue *wait = &rw->wpq;
3257 struct kiocb *kiocb = &req->rw.kiocb;
3259 /* never retry for NOWAIT, we just complete with -EAGAIN */
3260 if (req->flags & REQ_F_NOWAIT)
3263 /* Only for buffered IO */
3264 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3268 * just use poll if we can, and don't attempt if the fs doesn't
3269 * support callback based unlocks
3271 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3274 wait->wait.func = io_async_buf_func;
3275 wait->wait.private = req;
3276 wait->wait.flags = 0;
3277 INIT_LIST_HEAD(&wait->wait.entry);
3278 kiocb->ki_flags |= IOCB_WAITQ;
3279 kiocb->ki_flags &= ~IOCB_NOWAIT;
3280 kiocb->ki_waitq = wait;
3284 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3286 if (req->file->f_op->read_iter)
3287 return call_read_iter(req->file, &req->rw.kiocb, iter);
3288 else if (req->file->f_op->read)
3289 return loop_rw_iter(READ, req, iter);
3294 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3296 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3297 struct kiocb *kiocb = &req->rw.kiocb;
3298 struct iov_iter __iter, *iter = &__iter;
3299 struct io_async_rw *rw = req->async_data;
3300 ssize_t io_size, ret, ret2;
3301 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3307 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3311 io_size = iov_iter_count(iter);
3312 req->result = io_size;
3314 /* Ensure we clear previously set non-block flag */
3315 if (!force_nonblock)
3316 kiocb->ki_flags &= ~IOCB_NOWAIT;
3318 kiocb->ki_flags |= IOCB_NOWAIT;
3320 /* If the file doesn't support async, just async punt */
3321 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3322 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3323 return ret ?: -EAGAIN;
3326 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3327 if (unlikely(ret)) {
3332 ret = io_iter_do_read(req, iter);
3334 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3335 req->flags &= ~REQ_F_REISSUE;
3336 /* IOPOLL retry should happen for io-wq threads */
3337 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3339 /* no retry on NONBLOCK nor RWF_NOWAIT */
3340 if (req->flags & REQ_F_NOWAIT)
3342 /* some cases will consume bytes even on error returns */
3343 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3345 } else if (ret == -EIOCBQUEUED) {
3347 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3348 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3349 /* read all, failed, already did sync or don't want to retry */
3353 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3358 rw = req->async_data;
3359 /* now use our persistent iterator, if we aren't already */
3364 rw->bytes_done += ret;
3365 /* if we can retry, do so with the callbacks armed */
3366 if (!io_rw_should_retry(req)) {
3367 kiocb->ki_flags &= ~IOCB_WAITQ;
3372 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3373 * we get -EIOCBQUEUED, then we'll get a notification when the
3374 * desired page gets unlocked. We can also get a partial read
3375 * here, and if we do, then just retry at the new offset.
3377 ret = io_iter_do_read(req, iter);
3378 if (ret == -EIOCBQUEUED)
3380 /* we got some bytes, but not all. retry. */
3381 kiocb->ki_flags &= ~IOCB_WAITQ;
3382 } while (ret > 0 && ret < io_size);
3384 kiocb_done(kiocb, ret, issue_flags);
3386 /* it's faster to check here then delegate to kfree */
3392 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3394 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3396 return io_prep_rw(req, sqe);
3399 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3401 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3402 struct kiocb *kiocb = &req->rw.kiocb;
3403 struct iov_iter __iter, *iter = &__iter;
3404 struct io_async_rw *rw = req->async_data;
3405 ssize_t ret, ret2, io_size;
3406 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3412 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3416 io_size = iov_iter_count(iter);
3417 req->result = io_size;
3419 /* Ensure we clear previously set non-block flag */
3420 if (!force_nonblock)
3421 kiocb->ki_flags &= ~IOCB_NOWAIT;
3423 kiocb->ki_flags |= IOCB_NOWAIT;
3425 /* If the file doesn't support async, just async punt */
3426 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3429 /* file path doesn't support NOWAIT for non-direct_IO */
3430 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3431 (req->flags & REQ_F_ISREG))
3434 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3439 * Open-code file_start_write here to grab freeze protection,
3440 * which will be released by another thread in
3441 * io_complete_rw(). Fool lockdep by telling it the lock got
3442 * released so that it doesn't complain about the held lock when
3443 * we return to userspace.
3445 if (req->flags & REQ_F_ISREG) {
3446 sb_start_write(file_inode(req->file)->i_sb);
3447 __sb_writers_release(file_inode(req->file)->i_sb,
3450 kiocb->ki_flags |= IOCB_WRITE;
3452 if (req->file->f_op->write_iter)
3453 ret2 = call_write_iter(req->file, kiocb, iter);
3454 else if (req->file->f_op->write)
3455 ret2 = loop_rw_iter(WRITE, req, iter);
3459 if (req->flags & REQ_F_REISSUE) {
3460 req->flags &= ~REQ_F_REISSUE;
3465 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3466 * retry them without IOCB_NOWAIT.
3468 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3470 /* no retry on NONBLOCK nor RWF_NOWAIT */
3471 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3473 if (!force_nonblock || ret2 != -EAGAIN) {
3474 /* IOPOLL retry should happen for io-wq threads */
3475 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3478 kiocb_done(kiocb, ret2, issue_flags);
3481 /* some cases will consume bytes even on error returns */
3482 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3483 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3484 return ret ?: -EAGAIN;
3487 /* it's reportedly faster than delegating the null check to kfree() */
3493 static int io_renameat_prep(struct io_kiocb *req,
3494 const struct io_uring_sqe *sqe)
3496 struct io_rename *ren = &req->rename;
3497 const char __user *oldf, *newf;
3499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3501 if (sqe->ioprio || sqe->buf_index)
3503 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3506 ren->old_dfd = READ_ONCE(sqe->fd);
3507 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3508 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3509 ren->new_dfd = READ_ONCE(sqe->len);
3510 ren->flags = READ_ONCE(sqe->rename_flags);
3512 ren->oldpath = getname(oldf);
3513 if (IS_ERR(ren->oldpath))
3514 return PTR_ERR(ren->oldpath);
3516 ren->newpath = getname(newf);
3517 if (IS_ERR(ren->newpath)) {
3518 putname(ren->oldpath);
3519 return PTR_ERR(ren->newpath);
3522 req->flags |= REQ_F_NEED_CLEANUP;
3526 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3528 struct io_rename *ren = &req->rename;
3531 if (issue_flags & IO_URING_F_NONBLOCK)
3534 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3535 ren->newpath, ren->flags);
3537 req->flags &= ~REQ_F_NEED_CLEANUP;
3540 io_req_complete(req, ret);
3544 static int io_unlinkat_prep(struct io_kiocb *req,
3545 const struct io_uring_sqe *sqe)
3547 struct io_unlink *un = &req->unlink;
3548 const char __user *fname;
3550 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3552 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3554 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3557 un->dfd = READ_ONCE(sqe->fd);
3559 un->flags = READ_ONCE(sqe->unlink_flags);
3560 if (un->flags & ~AT_REMOVEDIR)
3563 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3564 un->filename = getname(fname);
3565 if (IS_ERR(un->filename))
3566 return PTR_ERR(un->filename);
3568 req->flags |= REQ_F_NEED_CLEANUP;
3572 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3574 struct io_unlink *un = &req->unlink;
3577 if (issue_flags & IO_URING_F_NONBLOCK)
3580 if (un->flags & AT_REMOVEDIR)
3581 ret = do_rmdir(un->dfd, un->filename);
3583 ret = do_unlinkat(un->dfd, un->filename);
3585 req->flags &= ~REQ_F_NEED_CLEANUP;
3588 io_req_complete(req, ret);
3592 static int io_shutdown_prep(struct io_kiocb *req,
3593 const struct io_uring_sqe *sqe)
3595 #if defined(CONFIG_NET)
3596 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3598 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3602 req->shutdown.how = READ_ONCE(sqe->len);
3609 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3611 #if defined(CONFIG_NET)
3612 struct socket *sock;
3615 if (issue_flags & IO_URING_F_NONBLOCK)
3618 sock = sock_from_file(req->file);
3619 if (unlikely(!sock))
3622 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3625 io_req_complete(req, ret);
3632 static int __io_splice_prep(struct io_kiocb *req,
3633 const struct io_uring_sqe *sqe)
3635 struct io_splice *sp = &req->splice;
3636 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3642 sp->len = READ_ONCE(sqe->len);
3643 sp->flags = READ_ONCE(sqe->splice_flags);
3645 if (unlikely(sp->flags & ~valid_flags))
3648 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3649 (sp->flags & SPLICE_F_FD_IN_FIXED));
3652 req->flags |= REQ_F_NEED_CLEANUP;
3656 static int io_tee_prep(struct io_kiocb *req,
3657 const struct io_uring_sqe *sqe)
3659 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3661 return __io_splice_prep(req, sqe);
3664 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3666 struct io_splice *sp = &req->splice;
3667 struct file *in = sp->file_in;
3668 struct file *out = sp->file_out;
3669 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3672 if (issue_flags & IO_URING_F_NONBLOCK)
3675 ret = do_tee(in, out, sp->len, flags);
3677 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3679 req->flags &= ~REQ_F_NEED_CLEANUP;
3683 io_req_complete(req, ret);
3687 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3689 struct io_splice *sp = &req->splice;
3691 sp->off_in = READ_ONCE(sqe->splice_off_in);
3692 sp->off_out = READ_ONCE(sqe->off);
3693 return __io_splice_prep(req, sqe);
3696 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3698 struct io_splice *sp = &req->splice;
3699 struct file *in = sp->file_in;
3700 struct file *out = sp->file_out;
3701 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3702 loff_t *poff_in, *poff_out;
3705 if (issue_flags & IO_URING_F_NONBLOCK)
3708 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3709 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3712 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3714 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3716 req->flags &= ~REQ_F_NEED_CLEANUP;
3720 io_req_complete(req, ret);
3725 * IORING_OP_NOP just posts a completion event, nothing else.
3727 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3729 struct io_ring_ctx *ctx = req->ctx;
3731 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3734 __io_req_complete(req, issue_flags, 0, 0);
3738 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3740 struct io_ring_ctx *ctx = req->ctx;
3745 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3747 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3750 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3751 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3754 req->sync.off = READ_ONCE(sqe->off);
3755 req->sync.len = READ_ONCE(sqe->len);
3759 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3761 loff_t end = req->sync.off + req->sync.len;
3764 /* fsync always requires a blocking context */
3765 if (issue_flags & IO_URING_F_NONBLOCK)
3768 ret = vfs_fsync_range(req->file, req->sync.off,
3769 end > 0 ? end : LLONG_MAX,
3770 req->sync.flags & IORING_FSYNC_DATASYNC);
3773 io_req_complete(req, ret);
3777 static int io_fallocate_prep(struct io_kiocb *req,
3778 const struct io_uring_sqe *sqe)
3780 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3785 req->sync.off = READ_ONCE(sqe->off);
3786 req->sync.len = READ_ONCE(sqe->addr);
3787 req->sync.mode = READ_ONCE(sqe->len);
3791 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3795 /* fallocate always requiring blocking context */
3796 if (issue_flags & IO_URING_F_NONBLOCK)
3798 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3802 io_req_complete(req, ret);
3806 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 const char __user *fname;
3811 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3813 if (unlikely(sqe->ioprio || sqe->buf_index))
3815 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3818 /* open.how should be already initialised */
3819 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3820 req->open.how.flags |= O_LARGEFILE;
3822 req->open.dfd = READ_ONCE(sqe->fd);
3823 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3824 req->open.filename = getname(fname);
3825 if (IS_ERR(req->open.filename)) {
3826 ret = PTR_ERR(req->open.filename);
3827 req->open.filename = NULL;
3830 req->open.nofile = rlimit(RLIMIT_NOFILE);
3831 req->flags |= REQ_F_NEED_CLEANUP;
3835 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3837 u64 mode = READ_ONCE(sqe->len);
3838 u64 flags = READ_ONCE(sqe->open_flags);
3840 req->open.how = build_open_how(flags, mode);
3841 return __io_openat_prep(req, sqe);
3844 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3846 struct open_how __user *how;
3850 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3851 len = READ_ONCE(sqe->len);
3852 if (len < OPEN_HOW_SIZE_VER0)
3855 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3860 return __io_openat_prep(req, sqe);
3863 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3865 struct open_flags op;
3868 bool resolve_nonblock;
3871 ret = build_open_flags(&req->open.how, &op);
3874 nonblock_set = op.open_flag & O_NONBLOCK;
3875 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3876 if (issue_flags & IO_URING_F_NONBLOCK) {
3878 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3879 * it'll always -EAGAIN
3881 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3883 op.lookup_flags |= LOOKUP_CACHED;
3884 op.open_flag |= O_NONBLOCK;
3887 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3891 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3894 * We could hang on to this 'fd' on retrying, but seems like
3895 * marginal gain for something that is now known to be a slower
3896 * path. So just put it, and we'll get a new one when we retry.
3900 ret = PTR_ERR(file);
3901 /* only retry if RESOLVE_CACHED wasn't already set by application */
3902 if (ret == -EAGAIN &&
3903 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3908 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3909 file->f_flags &= ~O_NONBLOCK;
3910 fsnotify_open(file);
3911 fd_install(ret, file);
3913 putname(req->open.filename);
3914 req->flags &= ~REQ_F_NEED_CLEANUP;
3917 __io_req_complete(req, issue_flags, ret, 0);
3921 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3923 return io_openat2(req, issue_flags);
3926 static int io_remove_buffers_prep(struct io_kiocb *req,
3927 const struct io_uring_sqe *sqe)
3929 struct io_provide_buf *p = &req->pbuf;
3932 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3935 tmp = READ_ONCE(sqe->fd);
3936 if (!tmp || tmp > USHRT_MAX)
3939 memset(p, 0, sizeof(*p));
3941 p->bgid = READ_ONCE(sqe->buf_group);
3945 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3946 int bgid, unsigned nbufs)
3950 /* shouldn't happen */
3954 /* the head kbuf is the list itself */
3955 while (!list_empty(&buf->list)) {
3956 struct io_buffer *nxt;
3958 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3959 list_del(&nxt->list);
3966 xa_erase(&ctx->io_buffers, bgid);
3971 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3973 struct io_provide_buf *p = &req->pbuf;
3974 struct io_ring_ctx *ctx = req->ctx;
3975 struct io_buffer *head;
3977 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3979 io_ring_submit_lock(ctx, !force_nonblock);
3981 lockdep_assert_held(&ctx->uring_lock);
3984 head = xa_load(&ctx->io_buffers, p->bgid);
3986 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3990 /* complete before unlock, IOPOLL may need the lock */
3991 __io_req_complete(req, issue_flags, ret, 0);
3992 io_ring_submit_unlock(ctx, !force_nonblock);
3996 static int io_provide_buffers_prep(struct io_kiocb *req,
3997 const struct io_uring_sqe *sqe)
3999 unsigned long size, tmp_check;
4000 struct io_provide_buf *p = &req->pbuf;
4003 if (sqe->ioprio || sqe->rw_flags)
4006 tmp = READ_ONCE(sqe->fd);
4007 if (!tmp || tmp > USHRT_MAX)
4010 p->addr = READ_ONCE(sqe->addr);
4011 p->len = READ_ONCE(sqe->len);
4013 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4016 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4019 size = (unsigned long)p->len * p->nbufs;
4020 if (!access_ok(u64_to_user_ptr(p->addr), size))
4023 p->bgid = READ_ONCE(sqe->buf_group);
4024 tmp = READ_ONCE(sqe->off);
4025 if (tmp > USHRT_MAX)
4031 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4033 struct io_buffer *buf;
4034 u64 addr = pbuf->addr;
4035 int i, bid = pbuf->bid;
4037 for (i = 0; i < pbuf->nbufs; i++) {
4038 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4043 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4048 INIT_LIST_HEAD(&buf->list);
4051 list_add_tail(&buf->list, &(*head)->list);
4055 return i ? i : -ENOMEM;
4058 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4060 struct io_provide_buf *p = &req->pbuf;
4061 struct io_ring_ctx *ctx = req->ctx;
4062 struct io_buffer *head, *list;
4064 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4066 io_ring_submit_lock(ctx, !force_nonblock);
4068 lockdep_assert_held(&ctx->uring_lock);
4070 list = head = xa_load(&ctx->io_buffers, p->bgid);
4072 ret = io_add_buffers(p, &head);
4073 if (ret >= 0 && !list) {
4074 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4076 __io_remove_buffers(ctx, head, p->bgid, -1U);
4080 /* complete before unlock, IOPOLL may need the lock */
4081 __io_req_complete(req, issue_flags, ret, 0);
4082 io_ring_submit_unlock(ctx, !force_nonblock);
4086 static int io_epoll_ctl_prep(struct io_kiocb *req,
4087 const struct io_uring_sqe *sqe)
4089 #if defined(CONFIG_EPOLL)
4090 if (sqe->ioprio || sqe->buf_index)
4092 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4095 req->epoll.epfd = READ_ONCE(sqe->fd);
4096 req->epoll.op = READ_ONCE(sqe->len);
4097 req->epoll.fd = READ_ONCE(sqe->off);
4099 if (ep_op_has_event(req->epoll.op)) {
4100 struct epoll_event __user *ev;
4102 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4103 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4113 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4115 #if defined(CONFIG_EPOLL)
4116 struct io_epoll *ie = &req->epoll;
4118 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4120 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4121 if (force_nonblock && ret == -EAGAIN)
4126 __io_req_complete(req, issue_flags, ret, 0);
4133 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4135 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4136 if (sqe->ioprio || sqe->buf_index || sqe->off)
4138 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4141 req->madvise.addr = READ_ONCE(sqe->addr);
4142 req->madvise.len = READ_ONCE(sqe->len);
4143 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4150 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4152 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4153 struct io_madvise *ma = &req->madvise;
4156 if (issue_flags & IO_URING_F_NONBLOCK)
4159 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4162 io_req_complete(req, ret);
4169 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4171 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4173 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4176 req->fadvise.offset = READ_ONCE(sqe->off);
4177 req->fadvise.len = READ_ONCE(sqe->len);
4178 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4182 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4184 struct io_fadvise *fa = &req->fadvise;
4187 if (issue_flags & IO_URING_F_NONBLOCK) {
4188 switch (fa->advice) {
4189 case POSIX_FADV_NORMAL:
4190 case POSIX_FADV_RANDOM:
4191 case POSIX_FADV_SEQUENTIAL:
4198 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4201 __io_req_complete(req, issue_flags, ret, 0);
4205 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4207 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4209 if (sqe->ioprio || sqe->buf_index)
4211 if (req->flags & REQ_F_FIXED_FILE)
4214 req->statx.dfd = READ_ONCE(sqe->fd);
4215 req->statx.mask = READ_ONCE(sqe->len);
4216 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4217 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4218 req->statx.flags = READ_ONCE(sqe->statx_flags);
4223 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4225 struct io_statx *ctx = &req->statx;
4228 if (issue_flags & IO_URING_F_NONBLOCK)
4231 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4236 io_req_complete(req, ret);
4240 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4242 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4244 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4245 sqe->rw_flags || sqe->buf_index)
4247 if (req->flags & REQ_F_FIXED_FILE)
4250 req->close.fd = READ_ONCE(sqe->fd);
4254 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4256 struct files_struct *files = current->files;
4257 struct io_close *close = &req->close;
4258 struct fdtable *fdt;
4259 struct file *file = NULL;
4262 spin_lock(&files->file_lock);
4263 fdt = files_fdtable(files);
4264 if (close->fd >= fdt->max_fds) {
4265 spin_unlock(&files->file_lock);
4268 file = fdt->fd[close->fd];
4269 if (!file || file->f_op == &io_uring_fops) {
4270 spin_unlock(&files->file_lock);
4275 /* if the file has a flush method, be safe and punt to async */
4276 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4277 spin_unlock(&files->file_lock);
4281 ret = __close_fd_get_file(close->fd, &file);
4282 spin_unlock(&files->file_lock);
4289 /* No ->flush() or already async, safely close from here */
4290 ret = filp_close(file, current->files);
4296 __io_req_complete(req, issue_flags, ret, 0);
4300 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4302 struct io_ring_ctx *ctx = req->ctx;
4304 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4306 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4309 req->sync.off = READ_ONCE(sqe->off);
4310 req->sync.len = READ_ONCE(sqe->len);
4311 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4315 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4319 /* sync_file_range always requires a blocking context */
4320 if (issue_flags & IO_URING_F_NONBLOCK)
4323 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4327 io_req_complete(req, ret);
4331 #if defined(CONFIG_NET)
4332 static int io_setup_async_msg(struct io_kiocb *req,
4333 struct io_async_msghdr *kmsg)
4335 struct io_async_msghdr *async_msg = req->async_data;
4339 if (io_alloc_async_data(req)) {
4340 kfree(kmsg->free_iov);
4343 async_msg = req->async_data;
4344 req->flags |= REQ_F_NEED_CLEANUP;
4345 memcpy(async_msg, kmsg, sizeof(*kmsg));
4346 async_msg->msg.msg_name = &async_msg->addr;
4347 /* if were using fast_iov, set it to the new one */
4348 if (!async_msg->free_iov)
4349 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4354 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4355 struct io_async_msghdr *iomsg)
4357 iomsg->msg.msg_name = &iomsg->addr;
4358 iomsg->free_iov = iomsg->fast_iov;
4359 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4360 req->sr_msg.msg_flags, &iomsg->free_iov);
4363 static int io_sendmsg_prep_async(struct io_kiocb *req)
4367 ret = io_sendmsg_copy_hdr(req, req->async_data);
4369 req->flags |= REQ_F_NEED_CLEANUP;
4373 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4375 struct io_sr_msg *sr = &req->sr_msg;
4377 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4380 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4381 sr->len = READ_ONCE(sqe->len);
4382 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4383 if (sr->msg_flags & MSG_DONTWAIT)
4384 req->flags |= REQ_F_NOWAIT;
4386 #ifdef CONFIG_COMPAT
4387 if (req->ctx->compat)
4388 sr->msg_flags |= MSG_CMSG_COMPAT;
4393 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4395 struct io_async_msghdr iomsg, *kmsg;
4396 struct socket *sock;
4401 sock = sock_from_file(req->file);
4402 if (unlikely(!sock))
4405 kmsg = req->async_data;
4407 ret = io_sendmsg_copy_hdr(req, &iomsg);
4413 flags = req->sr_msg.msg_flags;
4414 if (issue_flags & IO_URING_F_NONBLOCK)
4415 flags |= MSG_DONTWAIT;
4416 if (flags & MSG_WAITALL)
4417 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4419 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4420 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4421 return io_setup_async_msg(req, kmsg);
4422 if (ret == -ERESTARTSYS)
4425 /* fast path, check for non-NULL to avoid function call */
4427 kfree(kmsg->free_iov);
4428 req->flags &= ~REQ_F_NEED_CLEANUP;
4431 __io_req_complete(req, issue_flags, ret, 0);
4435 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4437 struct io_sr_msg *sr = &req->sr_msg;
4440 struct socket *sock;
4445 sock = sock_from_file(req->file);
4446 if (unlikely(!sock))
4449 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4453 msg.msg_name = NULL;
4454 msg.msg_control = NULL;
4455 msg.msg_controllen = 0;
4456 msg.msg_namelen = 0;
4458 flags = req->sr_msg.msg_flags;
4459 if (issue_flags & IO_URING_F_NONBLOCK)
4460 flags |= MSG_DONTWAIT;
4461 if (flags & MSG_WAITALL)
4462 min_ret = iov_iter_count(&msg.msg_iter);
4464 msg.msg_flags = flags;
4465 ret = sock_sendmsg(sock, &msg);
4466 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4468 if (ret == -ERESTARTSYS)
4473 __io_req_complete(req, issue_flags, ret, 0);
4477 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4478 struct io_async_msghdr *iomsg)
4480 struct io_sr_msg *sr = &req->sr_msg;
4481 struct iovec __user *uiov;
4485 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4486 &iomsg->uaddr, &uiov, &iov_len);
4490 if (req->flags & REQ_F_BUFFER_SELECT) {
4493 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4495 sr->len = iomsg->fast_iov[0].iov_len;
4496 iomsg->free_iov = NULL;
4498 iomsg->free_iov = iomsg->fast_iov;
4499 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4500 &iomsg->free_iov, &iomsg->msg.msg_iter,
4509 #ifdef CONFIG_COMPAT
4510 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4511 struct io_async_msghdr *iomsg)
4513 struct io_sr_msg *sr = &req->sr_msg;
4514 struct compat_iovec __user *uiov;
4519 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4524 uiov = compat_ptr(ptr);
4525 if (req->flags & REQ_F_BUFFER_SELECT) {
4526 compat_ssize_t clen;
4530 if (!access_ok(uiov, sizeof(*uiov)))
4532 if (__get_user(clen, &uiov->iov_len))
4537 iomsg->free_iov = NULL;
4539 iomsg->free_iov = iomsg->fast_iov;
4540 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4541 UIO_FASTIOV, &iomsg->free_iov,
4542 &iomsg->msg.msg_iter, true);
4551 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4552 struct io_async_msghdr *iomsg)
4554 iomsg->msg.msg_name = &iomsg->addr;
4556 #ifdef CONFIG_COMPAT
4557 if (req->ctx->compat)
4558 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4561 return __io_recvmsg_copy_hdr(req, iomsg);
4564 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4567 struct io_sr_msg *sr = &req->sr_msg;
4568 struct io_buffer *kbuf;
4570 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4575 req->flags |= REQ_F_BUFFER_SELECTED;
4579 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4581 return io_put_kbuf(req, req->sr_msg.kbuf);
4584 static int io_recvmsg_prep_async(struct io_kiocb *req)
4588 ret = io_recvmsg_copy_hdr(req, req->async_data);
4590 req->flags |= REQ_F_NEED_CLEANUP;
4594 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4596 struct io_sr_msg *sr = &req->sr_msg;
4598 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4601 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4602 sr->len = READ_ONCE(sqe->len);
4603 sr->bgid = READ_ONCE(sqe->buf_group);
4604 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4605 if (sr->msg_flags & MSG_DONTWAIT)
4606 req->flags |= REQ_F_NOWAIT;
4608 #ifdef CONFIG_COMPAT
4609 if (req->ctx->compat)
4610 sr->msg_flags |= MSG_CMSG_COMPAT;
4615 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4617 struct io_async_msghdr iomsg, *kmsg;
4618 struct socket *sock;
4619 struct io_buffer *kbuf;
4622 int ret, cflags = 0;
4623 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4625 sock = sock_from_file(req->file);
4626 if (unlikely(!sock))
4629 kmsg = req->async_data;
4631 ret = io_recvmsg_copy_hdr(req, &iomsg);
4637 if (req->flags & REQ_F_BUFFER_SELECT) {
4638 kbuf = io_recv_buffer_select(req, !force_nonblock);
4640 return PTR_ERR(kbuf);
4641 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4642 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4643 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4644 1, req->sr_msg.len);
4647 flags = req->sr_msg.msg_flags;
4649 flags |= MSG_DONTWAIT;
4650 if (flags & MSG_WAITALL)
4651 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4653 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4654 kmsg->uaddr, flags);
4655 if (force_nonblock && ret == -EAGAIN)
4656 return io_setup_async_msg(req, kmsg);
4657 if (ret == -ERESTARTSYS)
4660 if (req->flags & REQ_F_BUFFER_SELECTED)
4661 cflags = io_put_recv_kbuf(req);
4662 /* fast path, check for non-NULL to avoid function call */
4664 kfree(kmsg->free_iov);
4665 req->flags &= ~REQ_F_NEED_CLEANUP;
4666 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4668 __io_req_complete(req, issue_flags, ret, cflags);
4672 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4674 struct io_buffer *kbuf;
4675 struct io_sr_msg *sr = &req->sr_msg;
4677 void __user *buf = sr->buf;
4678 struct socket *sock;
4682 int ret, cflags = 0;
4683 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4685 sock = sock_from_file(req->file);
4686 if (unlikely(!sock))
4689 if (req->flags & REQ_F_BUFFER_SELECT) {
4690 kbuf = io_recv_buffer_select(req, !force_nonblock);
4692 return PTR_ERR(kbuf);
4693 buf = u64_to_user_ptr(kbuf->addr);
4696 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4700 msg.msg_name = NULL;
4701 msg.msg_control = NULL;
4702 msg.msg_controllen = 0;
4703 msg.msg_namelen = 0;
4704 msg.msg_iocb = NULL;
4707 flags = req->sr_msg.msg_flags;
4709 flags |= MSG_DONTWAIT;
4710 if (flags & MSG_WAITALL)
4711 min_ret = iov_iter_count(&msg.msg_iter);
4713 ret = sock_recvmsg(sock, &msg, flags);
4714 if (force_nonblock && ret == -EAGAIN)
4716 if (ret == -ERESTARTSYS)
4719 if (req->flags & REQ_F_BUFFER_SELECTED)
4720 cflags = io_put_recv_kbuf(req);
4721 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4723 __io_req_complete(req, issue_flags, ret, cflags);
4727 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4729 struct io_accept *accept = &req->accept;
4731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4733 if (sqe->ioprio || sqe->len || sqe->buf_index)
4736 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4737 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4738 accept->flags = READ_ONCE(sqe->accept_flags);
4739 accept->nofile = rlimit(RLIMIT_NOFILE);
4743 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4745 struct io_accept *accept = &req->accept;
4746 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4747 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4750 if (req->file->f_flags & O_NONBLOCK)
4751 req->flags |= REQ_F_NOWAIT;
4753 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4754 accept->addr_len, accept->flags,
4756 if (ret == -EAGAIN && force_nonblock)
4759 if (ret == -ERESTARTSYS)
4763 __io_req_complete(req, issue_flags, ret, 0);
4767 static int io_connect_prep_async(struct io_kiocb *req)
4769 struct io_async_connect *io = req->async_data;
4770 struct io_connect *conn = &req->connect;
4772 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4775 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4777 struct io_connect *conn = &req->connect;
4779 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4781 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4784 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4785 conn->addr_len = READ_ONCE(sqe->addr2);
4789 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4791 struct io_async_connect __io, *io;
4792 unsigned file_flags;
4794 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4796 if (req->async_data) {
4797 io = req->async_data;
4799 ret = move_addr_to_kernel(req->connect.addr,
4800 req->connect.addr_len,
4807 file_flags = force_nonblock ? O_NONBLOCK : 0;
4809 ret = __sys_connect_file(req->file, &io->address,
4810 req->connect.addr_len, file_flags);
4811 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4812 if (req->async_data)
4814 if (io_alloc_async_data(req)) {
4818 memcpy(req->async_data, &__io, sizeof(__io));
4821 if (ret == -ERESTARTSYS)
4826 __io_req_complete(req, issue_flags, ret, 0);
4829 #else /* !CONFIG_NET */
4830 #define IO_NETOP_FN(op) \
4831 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4833 return -EOPNOTSUPP; \
4836 #define IO_NETOP_PREP(op) \
4838 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4840 return -EOPNOTSUPP; \
4843 #define IO_NETOP_PREP_ASYNC(op) \
4845 static int io_##op##_prep_async(struct io_kiocb *req) \
4847 return -EOPNOTSUPP; \
4850 IO_NETOP_PREP_ASYNC(sendmsg);
4851 IO_NETOP_PREP_ASYNC(recvmsg);
4852 IO_NETOP_PREP_ASYNC(connect);
4853 IO_NETOP_PREP(accept);
4856 #endif /* CONFIG_NET */
4858 struct io_poll_table {
4859 struct poll_table_struct pt;
4860 struct io_kiocb *req;
4865 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4866 __poll_t mask, io_req_tw_func_t func)
4868 /* for instances that support it check for an event match first: */
4869 if (mask && !(mask & poll->events))
4872 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4874 list_del_init(&poll->wait.entry);
4877 req->io_task_work.func = func;
4880 * If this fails, then the task is exiting. When a task exits, the
4881 * work gets canceled, so just cancel this request as well instead
4882 * of executing it. We can't safely execute it anyway, as we may not
4883 * have the needed state needed for it anyway.
4885 io_req_task_work_add(req);
4889 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4890 __acquires(&req->ctx->completion_lock)
4892 struct io_ring_ctx *ctx = req->ctx;
4894 if (unlikely(req->task->flags & PF_EXITING))
4895 WRITE_ONCE(poll->canceled, true);
4897 if (!req->result && !READ_ONCE(poll->canceled)) {
4898 struct poll_table_struct pt = { ._key = poll->events };
4900 req->result = vfs_poll(req->file, &pt) & poll->events;
4903 spin_lock(&ctx->completion_lock);
4904 if (!req->result && !READ_ONCE(poll->canceled)) {
4905 add_wait_queue(poll->head, &poll->wait);
4912 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4914 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4915 if (req->opcode == IORING_OP_POLL_ADD)
4916 return req->async_data;
4917 return req->apoll->double_poll;
4920 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4922 if (req->opcode == IORING_OP_POLL_ADD)
4924 return &req->apoll->poll;
4927 static void io_poll_remove_double(struct io_kiocb *req)
4928 __must_hold(&req->ctx->completion_lock)
4930 struct io_poll_iocb *poll = io_poll_get_double(req);
4932 lockdep_assert_held(&req->ctx->completion_lock);
4934 if (poll && poll->head) {
4935 struct wait_queue_head *head = poll->head;
4937 spin_lock_irq(&head->lock);
4938 list_del_init(&poll->wait.entry);
4939 if (poll->wait.private)
4942 spin_unlock_irq(&head->lock);
4946 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4947 __must_hold(&req->ctx->completion_lock)
4949 struct io_ring_ctx *ctx = req->ctx;
4950 unsigned flags = IORING_CQE_F_MORE;
4953 if (READ_ONCE(req->poll.canceled)) {
4955 req->poll.events |= EPOLLONESHOT;
4957 error = mangle_poll(mask);
4959 if (req->poll.events & EPOLLONESHOT)
4961 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4962 req->poll.done = true;
4965 if (flags & IORING_CQE_F_MORE)
4968 io_commit_cqring(ctx);
4969 return !(flags & IORING_CQE_F_MORE);
4972 static void io_poll_task_func(struct io_kiocb *req)
4974 struct io_ring_ctx *ctx = req->ctx;
4975 struct io_kiocb *nxt;
4977 if (io_poll_rewait(req, &req->poll)) {
4978 spin_unlock(&ctx->completion_lock);
4982 done = io_poll_complete(req, req->result);
4984 io_poll_remove_double(req);
4985 hash_del(&req->hash_node);
4988 add_wait_queue(req->poll.head, &req->poll.wait);
4990 spin_unlock(&ctx->completion_lock);
4991 io_cqring_ev_posted(ctx);
4994 nxt = io_put_req_find_next(req);
4996 io_req_task_submit(nxt);
5001 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5002 int sync, void *key)
5004 struct io_kiocb *req = wait->private;
5005 struct io_poll_iocb *poll = io_poll_get_single(req);
5006 __poll_t mask = key_to_poll(key);
5007 unsigned long flags;
5009 /* for instances that support it check for an event match first: */
5010 if (mask && !(mask & poll->events))
5012 if (!(poll->events & EPOLLONESHOT))
5013 return poll->wait.func(&poll->wait, mode, sync, key);
5015 list_del_init(&wait->entry);
5020 spin_lock_irqsave(&poll->head->lock, flags);
5021 done = list_empty(&poll->wait.entry);
5023 list_del_init(&poll->wait.entry);
5024 /* make sure double remove sees this as being gone */
5025 wait->private = NULL;
5026 spin_unlock_irqrestore(&poll->head->lock, flags);
5028 /* use wait func handler, so it matches the rq type */
5029 poll->wait.func(&poll->wait, mode, sync, key);
5036 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5037 wait_queue_func_t wake_func)
5041 poll->canceled = false;
5042 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5043 /* mask in events that we always want/need */
5044 poll->events = events | IO_POLL_UNMASK;
5045 INIT_LIST_HEAD(&poll->wait.entry);
5046 init_waitqueue_func_entry(&poll->wait, wake_func);
5049 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5050 struct wait_queue_head *head,
5051 struct io_poll_iocb **poll_ptr)
5053 struct io_kiocb *req = pt->req;
5056 * The file being polled uses multiple waitqueues for poll handling
5057 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5060 if (unlikely(pt->nr_entries)) {
5061 struct io_poll_iocb *poll_one = poll;
5063 /* already have a 2nd entry, fail a third attempt */
5065 pt->error = -EINVAL;
5069 * Can't handle multishot for double wait for now, turn it
5070 * into one-shot mode.
5072 if (!(poll_one->events & EPOLLONESHOT))
5073 poll_one->events |= EPOLLONESHOT;
5074 /* double add on the same waitqueue head, ignore */
5075 if (poll_one->head == head)
5077 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5079 pt->error = -ENOMEM;
5082 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5084 poll->wait.private = req;
5091 if (poll->events & EPOLLEXCLUSIVE)
5092 add_wait_queue_exclusive(head, &poll->wait);
5094 add_wait_queue(head, &poll->wait);
5097 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5098 struct poll_table_struct *p)
5100 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5101 struct async_poll *apoll = pt->req->apoll;
5103 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5106 static void io_async_task_func(struct io_kiocb *req)
5108 struct async_poll *apoll = req->apoll;
5109 struct io_ring_ctx *ctx = req->ctx;
5111 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5113 if (io_poll_rewait(req, &apoll->poll)) {
5114 spin_unlock(&ctx->completion_lock);
5118 hash_del(&req->hash_node);
5119 io_poll_remove_double(req);
5120 spin_unlock(&ctx->completion_lock);
5122 if (!READ_ONCE(apoll->poll.canceled))
5123 io_req_task_submit(req);
5125 io_req_complete_failed(req, -ECANCELED);
5128 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5131 struct io_kiocb *req = wait->private;
5132 struct io_poll_iocb *poll = &req->apoll->poll;
5134 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5137 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5140 static void io_poll_req_insert(struct io_kiocb *req)
5142 struct io_ring_ctx *ctx = req->ctx;
5143 struct hlist_head *list;
5145 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5146 hlist_add_head(&req->hash_node, list);
5149 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5150 struct io_poll_iocb *poll,
5151 struct io_poll_table *ipt, __poll_t mask,
5152 wait_queue_func_t wake_func)
5153 __acquires(&ctx->completion_lock)
5155 struct io_ring_ctx *ctx = req->ctx;
5156 bool cancel = false;
5158 INIT_HLIST_NODE(&req->hash_node);
5159 io_init_poll_iocb(poll, mask, wake_func);
5160 poll->file = req->file;
5161 poll->wait.private = req;
5163 ipt->pt._key = mask;
5166 ipt->nr_entries = 0;
5168 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5169 if (unlikely(!ipt->nr_entries) && !ipt->error)
5170 ipt->error = -EINVAL;
5172 spin_lock(&ctx->completion_lock);
5173 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5174 io_poll_remove_double(req);
5175 if (likely(poll->head)) {
5176 spin_lock_irq(&poll->head->lock);
5177 if (unlikely(list_empty(&poll->wait.entry))) {
5183 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5184 list_del_init(&poll->wait.entry);
5186 WRITE_ONCE(poll->canceled, true);
5187 else if (!poll->done) /* actually waiting for an event */
5188 io_poll_req_insert(req);
5189 spin_unlock_irq(&poll->head->lock);
5201 static int io_arm_poll_handler(struct io_kiocb *req)
5203 const struct io_op_def *def = &io_op_defs[req->opcode];
5204 struct io_ring_ctx *ctx = req->ctx;
5205 struct async_poll *apoll;
5206 struct io_poll_table ipt;
5207 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5210 if (!req->file || !file_can_poll(req->file))
5211 return IO_APOLL_ABORTED;
5212 if (req->flags & REQ_F_POLLED)
5213 return IO_APOLL_ABORTED;
5214 if (!def->pollin && !def->pollout)
5215 return IO_APOLL_ABORTED;
5219 mask |= POLLIN | POLLRDNORM;
5221 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5222 if ((req->opcode == IORING_OP_RECVMSG) &&
5223 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5227 mask |= POLLOUT | POLLWRNORM;
5230 /* if we can't nonblock try, then no point in arming a poll handler */
5231 if (!io_file_supports_nowait(req, rw))
5232 return IO_APOLL_ABORTED;
5234 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5235 if (unlikely(!apoll))
5236 return IO_APOLL_ABORTED;
5237 apoll->double_poll = NULL;
5239 req->flags |= REQ_F_POLLED;
5240 ipt.pt._qproc = io_async_queue_proc;
5241 io_req_set_refcount(req);
5243 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5245 spin_unlock(&ctx->completion_lock);
5246 if (ret || ipt.error)
5247 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5249 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5250 mask, apoll->poll.events);
5254 static bool __io_poll_remove_one(struct io_kiocb *req,
5255 struct io_poll_iocb *poll, bool do_cancel)
5256 __must_hold(&req->ctx->completion_lock)
5258 bool do_complete = false;
5262 spin_lock_irq(&poll->head->lock);
5264 WRITE_ONCE(poll->canceled, true);
5265 if (!list_empty(&poll->wait.entry)) {
5266 list_del_init(&poll->wait.entry);
5269 spin_unlock_irq(&poll->head->lock);
5270 hash_del(&req->hash_node);
5274 static bool io_poll_remove_one(struct io_kiocb *req)
5275 __must_hold(&req->ctx->completion_lock)
5279 io_poll_remove_double(req);
5280 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5283 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5284 io_commit_cqring(req->ctx);
5286 io_put_req_deferred(req);
5292 * Returns true if we found and killed one or more poll requests
5294 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5297 struct hlist_node *tmp;
5298 struct io_kiocb *req;
5301 spin_lock(&ctx->completion_lock);
5302 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5303 struct hlist_head *list;
5305 list = &ctx->cancel_hash[i];
5306 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5307 if (io_match_task(req, tsk, cancel_all))
5308 posted += io_poll_remove_one(req);
5311 spin_unlock(&ctx->completion_lock);
5314 io_cqring_ev_posted(ctx);
5319 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5321 __must_hold(&ctx->completion_lock)
5323 struct hlist_head *list;
5324 struct io_kiocb *req;
5326 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5327 hlist_for_each_entry(req, list, hash_node) {
5328 if (sqe_addr != req->user_data)
5330 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5337 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5339 __must_hold(&ctx->completion_lock)
5341 struct io_kiocb *req;
5343 req = io_poll_find(ctx, sqe_addr, poll_only);
5346 if (io_poll_remove_one(req))
5352 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5357 events = READ_ONCE(sqe->poll32_events);
5359 events = swahw32(events);
5361 if (!(flags & IORING_POLL_ADD_MULTI))
5362 events |= EPOLLONESHOT;
5363 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5366 static int io_poll_update_prep(struct io_kiocb *req,
5367 const struct io_uring_sqe *sqe)
5369 struct io_poll_update *upd = &req->poll_update;
5372 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5374 if (sqe->ioprio || sqe->buf_index)
5376 flags = READ_ONCE(sqe->len);
5377 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5378 IORING_POLL_ADD_MULTI))
5380 /* meaningless without update */
5381 if (flags == IORING_POLL_ADD_MULTI)
5384 upd->old_user_data = READ_ONCE(sqe->addr);
5385 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5386 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5388 upd->new_user_data = READ_ONCE(sqe->off);
5389 if (!upd->update_user_data && upd->new_user_data)
5391 if (upd->update_events)
5392 upd->events = io_poll_parse_events(sqe, flags);
5393 else if (sqe->poll32_events)
5399 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5402 struct io_kiocb *req = wait->private;
5403 struct io_poll_iocb *poll = &req->poll;
5405 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5408 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5409 struct poll_table_struct *p)
5411 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5413 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5416 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5418 struct io_poll_iocb *poll = &req->poll;
5421 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5423 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5425 flags = READ_ONCE(sqe->len);
5426 if (flags & ~IORING_POLL_ADD_MULTI)
5429 io_req_set_refcount(req);
5430 poll->events = io_poll_parse_events(sqe, flags);
5434 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5436 struct io_poll_iocb *poll = &req->poll;
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct io_poll_table ipt;
5441 ipt.pt._qproc = io_poll_queue_proc;
5443 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5446 if (mask) { /* no async, we'd stolen it */
5448 io_poll_complete(req, mask);
5450 spin_unlock(&ctx->completion_lock);
5453 io_cqring_ev_posted(ctx);
5454 if (poll->events & EPOLLONESHOT)
5460 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5462 struct io_ring_ctx *ctx = req->ctx;
5463 struct io_kiocb *preq;
5467 spin_lock(&ctx->completion_lock);
5468 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5474 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5476 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5481 * Don't allow racy completion with singleshot, as we cannot safely
5482 * update those. For multishot, if we're racing with completion, just
5483 * let completion re-add it.
5485 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5486 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5490 /* we now have a detached poll request. reissue. */
5494 spin_unlock(&ctx->completion_lock);
5496 io_req_complete(req, ret);
5499 /* only mask one event flags, keep behavior flags */
5500 if (req->poll_update.update_events) {
5501 preq->poll.events &= ~0xffff;
5502 preq->poll.events |= req->poll_update.events & 0xffff;
5503 preq->poll.events |= IO_POLL_UNMASK;
5505 if (req->poll_update.update_user_data)
5506 preq->user_data = req->poll_update.new_user_data;
5507 spin_unlock(&ctx->completion_lock);
5509 /* complete update request, we're done with it */
5510 io_req_complete(req, ret);
5513 ret = io_poll_add(preq, issue_flags);
5516 io_req_complete(preq, ret);
5522 static void io_req_task_timeout(struct io_kiocb *req)
5524 struct io_ring_ctx *ctx = req->ctx;
5526 spin_lock(&ctx->completion_lock);
5527 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5528 io_commit_cqring(ctx);
5529 spin_unlock(&ctx->completion_lock);
5531 io_cqring_ev_posted(ctx);
5536 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5538 struct io_timeout_data *data = container_of(timer,
5539 struct io_timeout_data, timer);
5540 struct io_kiocb *req = data->req;
5541 struct io_ring_ctx *ctx = req->ctx;
5542 unsigned long flags;
5544 spin_lock_irqsave(&ctx->timeout_lock, flags);
5545 list_del_init(&req->timeout.list);
5546 atomic_set(&req->ctx->cq_timeouts,
5547 atomic_read(&req->ctx->cq_timeouts) + 1);
5548 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5550 req->io_task_work.func = io_req_task_timeout;
5551 io_req_task_work_add(req);
5552 return HRTIMER_NORESTART;
5555 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5557 __must_hold(&ctx->timeout_lock)
5559 struct io_timeout_data *io;
5560 struct io_kiocb *req;
5563 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5564 found = user_data == req->user_data;
5569 return ERR_PTR(-ENOENT);
5571 io = req->async_data;
5572 if (hrtimer_try_to_cancel(&io->timer) == -1)
5573 return ERR_PTR(-EALREADY);
5574 list_del_init(&req->timeout.list);
5578 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5579 __must_hold(&ctx->timeout_lock)
5581 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5584 return PTR_ERR(req);
5587 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5588 io_put_req_deferred(req);
5592 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5593 struct timespec64 *ts, enum hrtimer_mode mode)
5594 __must_hold(&ctx->timeout_lock)
5596 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5597 struct io_timeout_data *data;
5600 return PTR_ERR(req);
5602 req->timeout.off = 0; /* noseq */
5603 data = req->async_data;
5604 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5605 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5606 data->timer.function = io_timeout_fn;
5607 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5611 static int io_timeout_remove_prep(struct io_kiocb *req,
5612 const struct io_uring_sqe *sqe)
5614 struct io_timeout_rem *tr = &req->timeout_rem;
5616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5618 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5620 if (sqe->ioprio || sqe->buf_index || sqe->len)
5623 tr->addr = READ_ONCE(sqe->addr);
5624 tr->flags = READ_ONCE(sqe->timeout_flags);
5625 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5626 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5628 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5630 } else if (tr->flags) {
5631 /* timeout removal doesn't support flags */
5638 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5640 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5645 * Remove or update an existing timeout command
5647 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5649 struct io_timeout_rem *tr = &req->timeout_rem;
5650 struct io_ring_ctx *ctx = req->ctx;
5653 spin_lock_irq(&ctx->timeout_lock);
5654 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5655 ret = io_timeout_cancel(ctx, tr->addr);
5657 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5658 io_translate_timeout_mode(tr->flags));
5659 spin_unlock_irq(&ctx->timeout_lock);
5661 spin_lock(&ctx->completion_lock);
5662 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5663 io_commit_cqring(ctx);
5664 spin_unlock(&ctx->completion_lock);
5665 io_cqring_ev_posted(ctx);
5672 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5673 bool is_timeout_link)
5675 struct io_timeout_data *data;
5677 u32 off = READ_ONCE(sqe->off);
5679 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5681 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5683 if (off && is_timeout_link)
5685 flags = READ_ONCE(sqe->timeout_flags);
5686 if (flags & ~IORING_TIMEOUT_ABS)
5689 req->timeout.off = off;
5690 if (unlikely(off && !req->ctx->off_timeout_used))
5691 req->ctx->off_timeout_used = true;
5693 if (!req->async_data && io_alloc_async_data(req))
5696 data = req->async_data;
5699 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5702 data->mode = io_translate_timeout_mode(flags);
5703 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5704 if (is_timeout_link)
5705 io_req_track_inflight(req);
5709 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5711 struct io_ring_ctx *ctx = req->ctx;
5712 struct io_timeout_data *data = req->async_data;
5713 struct list_head *entry;
5714 u32 tail, off = req->timeout.off;
5716 spin_lock_irq(&ctx->timeout_lock);
5719 * sqe->off holds how many events that need to occur for this
5720 * timeout event to be satisfied. If it isn't set, then this is
5721 * a pure timeout request, sequence isn't used.
5723 if (io_is_timeout_noseq(req)) {
5724 entry = ctx->timeout_list.prev;
5728 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5729 req->timeout.target_seq = tail + off;
5731 /* Update the last seq here in case io_flush_timeouts() hasn't.
5732 * This is safe because ->completion_lock is held, and submissions
5733 * and completions are never mixed in the same ->completion_lock section.
5735 ctx->cq_last_tm_flush = tail;
5738 * Insertion sort, ensuring the first entry in the list is always
5739 * the one we need first.
5741 list_for_each_prev(entry, &ctx->timeout_list) {
5742 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5745 if (io_is_timeout_noseq(nxt))
5747 /* nxt.seq is behind @tail, otherwise would've been completed */
5748 if (off >= nxt->timeout.target_seq - tail)
5752 list_add(&req->timeout.list, entry);
5753 data->timer.function = io_timeout_fn;
5754 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5755 spin_unlock_irq(&ctx->timeout_lock);
5759 struct io_cancel_data {
5760 struct io_ring_ctx *ctx;
5764 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5766 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5767 struct io_cancel_data *cd = data;
5769 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5772 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5773 struct io_ring_ctx *ctx)
5775 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5776 enum io_wq_cancel cancel_ret;
5779 if (!tctx || !tctx->io_wq)
5782 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5783 switch (cancel_ret) {
5784 case IO_WQ_CANCEL_OK:
5787 case IO_WQ_CANCEL_RUNNING:
5790 case IO_WQ_CANCEL_NOTFOUND:
5798 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5799 struct io_kiocb *req, __u64 sqe_addr,
5804 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5805 spin_lock(&ctx->completion_lock);
5808 spin_lock_irq(&ctx->timeout_lock);
5809 ret = io_timeout_cancel(ctx, sqe_addr);
5810 spin_unlock_irq(&ctx->timeout_lock);
5813 ret = io_poll_cancel(ctx, sqe_addr, false);
5817 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5818 io_commit_cqring(ctx);
5819 spin_unlock(&ctx->completion_lock);
5820 io_cqring_ev_posted(ctx);
5826 static int io_async_cancel_prep(struct io_kiocb *req,
5827 const struct io_uring_sqe *sqe)
5829 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5831 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5833 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5836 req->cancel.addr = READ_ONCE(sqe->addr);
5840 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5842 struct io_ring_ctx *ctx = req->ctx;
5843 u64 sqe_addr = req->cancel.addr;
5844 struct io_tctx_node *node;
5847 /* tasks should wait for their io-wq threads, so safe w/o sync */
5848 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5849 spin_lock(&ctx->completion_lock);
5852 spin_lock_irq(&ctx->timeout_lock);
5853 ret = io_timeout_cancel(ctx, sqe_addr);
5854 spin_unlock_irq(&ctx->timeout_lock);
5857 ret = io_poll_cancel(ctx, sqe_addr, false);
5860 spin_unlock(&ctx->completion_lock);
5862 /* slow path, try all io-wq's */
5863 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5865 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5866 struct io_uring_task *tctx = node->task->io_uring;
5868 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5872 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5874 spin_lock(&ctx->completion_lock);
5876 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5877 io_commit_cqring(ctx);
5878 spin_unlock(&ctx->completion_lock);
5879 io_cqring_ev_posted(ctx);
5887 static int io_rsrc_update_prep(struct io_kiocb *req,
5888 const struct io_uring_sqe *sqe)
5890 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5892 if (sqe->ioprio || sqe->rw_flags)
5895 req->rsrc_update.offset = READ_ONCE(sqe->off);
5896 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5897 if (!req->rsrc_update.nr_args)
5899 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5903 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5905 struct io_ring_ctx *ctx = req->ctx;
5906 struct io_uring_rsrc_update2 up;
5909 if (issue_flags & IO_URING_F_NONBLOCK)
5912 up.offset = req->rsrc_update.offset;
5913 up.data = req->rsrc_update.arg;
5918 mutex_lock(&ctx->uring_lock);
5919 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5920 &up, req->rsrc_update.nr_args);
5921 mutex_unlock(&ctx->uring_lock);
5925 __io_req_complete(req, issue_flags, ret, 0);
5929 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5931 switch (req->opcode) {
5934 case IORING_OP_READV:
5935 case IORING_OP_READ_FIXED:
5936 case IORING_OP_READ:
5937 return io_read_prep(req, sqe);
5938 case IORING_OP_WRITEV:
5939 case IORING_OP_WRITE_FIXED:
5940 case IORING_OP_WRITE:
5941 return io_write_prep(req, sqe);
5942 case IORING_OP_POLL_ADD:
5943 return io_poll_add_prep(req, sqe);
5944 case IORING_OP_POLL_REMOVE:
5945 return io_poll_update_prep(req, sqe);
5946 case IORING_OP_FSYNC:
5947 return io_fsync_prep(req, sqe);
5948 case IORING_OP_SYNC_FILE_RANGE:
5949 return io_sfr_prep(req, sqe);
5950 case IORING_OP_SENDMSG:
5951 case IORING_OP_SEND:
5952 return io_sendmsg_prep(req, sqe);
5953 case IORING_OP_RECVMSG:
5954 case IORING_OP_RECV:
5955 return io_recvmsg_prep(req, sqe);
5956 case IORING_OP_CONNECT:
5957 return io_connect_prep(req, sqe);
5958 case IORING_OP_TIMEOUT:
5959 return io_timeout_prep(req, sqe, false);
5960 case IORING_OP_TIMEOUT_REMOVE:
5961 return io_timeout_remove_prep(req, sqe);
5962 case IORING_OP_ASYNC_CANCEL:
5963 return io_async_cancel_prep(req, sqe);
5964 case IORING_OP_LINK_TIMEOUT:
5965 return io_timeout_prep(req, sqe, true);
5966 case IORING_OP_ACCEPT:
5967 return io_accept_prep(req, sqe);
5968 case IORING_OP_FALLOCATE:
5969 return io_fallocate_prep(req, sqe);
5970 case IORING_OP_OPENAT:
5971 return io_openat_prep(req, sqe);
5972 case IORING_OP_CLOSE:
5973 return io_close_prep(req, sqe);
5974 case IORING_OP_FILES_UPDATE:
5975 return io_rsrc_update_prep(req, sqe);
5976 case IORING_OP_STATX:
5977 return io_statx_prep(req, sqe);
5978 case IORING_OP_FADVISE:
5979 return io_fadvise_prep(req, sqe);
5980 case IORING_OP_MADVISE:
5981 return io_madvise_prep(req, sqe);
5982 case IORING_OP_OPENAT2:
5983 return io_openat2_prep(req, sqe);
5984 case IORING_OP_EPOLL_CTL:
5985 return io_epoll_ctl_prep(req, sqe);
5986 case IORING_OP_SPLICE:
5987 return io_splice_prep(req, sqe);
5988 case IORING_OP_PROVIDE_BUFFERS:
5989 return io_provide_buffers_prep(req, sqe);
5990 case IORING_OP_REMOVE_BUFFERS:
5991 return io_remove_buffers_prep(req, sqe);
5993 return io_tee_prep(req, sqe);
5994 case IORING_OP_SHUTDOWN:
5995 return io_shutdown_prep(req, sqe);
5996 case IORING_OP_RENAMEAT:
5997 return io_renameat_prep(req, sqe);
5998 case IORING_OP_UNLINKAT:
5999 return io_unlinkat_prep(req, sqe);
6002 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6007 static int io_req_prep_async(struct io_kiocb *req)
6009 if (!io_op_defs[req->opcode].needs_async_setup)
6011 if (WARN_ON_ONCE(req->async_data))
6013 if (io_alloc_async_data(req))
6016 switch (req->opcode) {
6017 case IORING_OP_READV:
6018 return io_rw_prep_async(req, READ);
6019 case IORING_OP_WRITEV:
6020 return io_rw_prep_async(req, WRITE);
6021 case IORING_OP_SENDMSG:
6022 return io_sendmsg_prep_async(req);
6023 case IORING_OP_RECVMSG:
6024 return io_recvmsg_prep_async(req);
6025 case IORING_OP_CONNECT:
6026 return io_connect_prep_async(req);
6028 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6033 static u32 io_get_sequence(struct io_kiocb *req)
6035 u32 seq = req->ctx->cached_sq_head;
6037 /* need original cached_sq_head, but it was increased for each req */
6038 io_for_each_link(req, req)
6043 static bool io_drain_req(struct io_kiocb *req)
6045 struct io_kiocb *pos;
6046 struct io_ring_ctx *ctx = req->ctx;
6047 struct io_defer_entry *de;
6052 * If we need to drain a request in the middle of a link, drain the
6053 * head request and the next request/link after the current link.
6054 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6055 * maintained for every request of our link.
6057 if (ctx->drain_next) {
6058 req->flags |= REQ_F_IO_DRAIN;
6059 ctx->drain_next = false;
6061 /* not interested in head, start from the first linked */
6062 io_for_each_link(pos, req->link) {
6063 if (pos->flags & REQ_F_IO_DRAIN) {
6064 ctx->drain_next = true;
6065 req->flags |= REQ_F_IO_DRAIN;
6070 /* Still need defer if there is pending req in defer list. */
6071 if (likely(list_empty_careful(&ctx->defer_list) &&
6072 !(req->flags & REQ_F_IO_DRAIN))) {
6073 ctx->drain_active = false;
6077 seq = io_get_sequence(req);
6078 /* Still a chance to pass the sequence check */
6079 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6082 ret = io_req_prep_async(req);
6085 io_prep_async_link(req);
6086 de = kmalloc(sizeof(*de), GFP_KERNEL);
6090 io_req_complete_failed(req, ret);
6094 spin_lock(&ctx->completion_lock);
6095 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6096 spin_unlock(&ctx->completion_lock);
6098 io_queue_async_work(req);
6102 trace_io_uring_defer(ctx, req, req->user_data);
6105 list_add_tail(&de->list, &ctx->defer_list);
6106 spin_unlock(&ctx->completion_lock);
6110 static void io_clean_op(struct io_kiocb *req)
6112 if (req->flags & REQ_F_BUFFER_SELECTED) {
6113 switch (req->opcode) {
6114 case IORING_OP_READV:
6115 case IORING_OP_READ_FIXED:
6116 case IORING_OP_READ:
6117 kfree((void *)(unsigned long)req->rw.addr);
6119 case IORING_OP_RECVMSG:
6120 case IORING_OP_RECV:
6121 kfree(req->sr_msg.kbuf);
6126 if (req->flags & REQ_F_NEED_CLEANUP) {
6127 switch (req->opcode) {
6128 case IORING_OP_READV:
6129 case IORING_OP_READ_FIXED:
6130 case IORING_OP_READ:
6131 case IORING_OP_WRITEV:
6132 case IORING_OP_WRITE_FIXED:
6133 case IORING_OP_WRITE: {
6134 struct io_async_rw *io = req->async_data;
6136 kfree(io->free_iovec);
6139 case IORING_OP_RECVMSG:
6140 case IORING_OP_SENDMSG: {
6141 struct io_async_msghdr *io = req->async_data;
6143 kfree(io->free_iov);
6146 case IORING_OP_SPLICE:
6148 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6149 io_put_file(req->splice.file_in);
6151 case IORING_OP_OPENAT:
6152 case IORING_OP_OPENAT2:
6153 if (req->open.filename)
6154 putname(req->open.filename);
6156 case IORING_OP_RENAMEAT:
6157 putname(req->rename.oldpath);
6158 putname(req->rename.newpath);
6160 case IORING_OP_UNLINKAT:
6161 putname(req->unlink.filename);
6165 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6166 kfree(req->apoll->double_poll);
6170 if (req->flags & REQ_F_INFLIGHT) {
6171 struct io_uring_task *tctx = req->task->io_uring;
6173 atomic_dec(&tctx->inflight_tracked);
6175 if (req->flags & REQ_F_CREDS)
6176 put_cred(req->creds);
6178 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6181 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6183 struct io_ring_ctx *ctx = req->ctx;
6184 const struct cred *creds = NULL;
6187 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6188 creds = override_creds(req->creds);
6190 switch (req->opcode) {
6192 ret = io_nop(req, issue_flags);
6194 case IORING_OP_READV:
6195 case IORING_OP_READ_FIXED:
6196 case IORING_OP_READ:
6197 ret = io_read(req, issue_flags);
6199 case IORING_OP_WRITEV:
6200 case IORING_OP_WRITE_FIXED:
6201 case IORING_OP_WRITE:
6202 ret = io_write(req, issue_flags);
6204 case IORING_OP_FSYNC:
6205 ret = io_fsync(req, issue_flags);
6207 case IORING_OP_POLL_ADD:
6208 ret = io_poll_add(req, issue_flags);
6210 case IORING_OP_POLL_REMOVE:
6211 ret = io_poll_update(req, issue_flags);
6213 case IORING_OP_SYNC_FILE_RANGE:
6214 ret = io_sync_file_range(req, issue_flags);
6216 case IORING_OP_SENDMSG:
6217 ret = io_sendmsg(req, issue_flags);
6219 case IORING_OP_SEND:
6220 ret = io_send(req, issue_flags);
6222 case IORING_OP_RECVMSG:
6223 ret = io_recvmsg(req, issue_flags);
6225 case IORING_OP_RECV:
6226 ret = io_recv(req, issue_flags);
6228 case IORING_OP_TIMEOUT:
6229 ret = io_timeout(req, issue_flags);
6231 case IORING_OP_TIMEOUT_REMOVE:
6232 ret = io_timeout_remove(req, issue_flags);
6234 case IORING_OP_ACCEPT:
6235 ret = io_accept(req, issue_flags);
6237 case IORING_OP_CONNECT:
6238 ret = io_connect(req, issue_flags);
6240 case IORING_OP_ASYNC_CANCEL:
6241 ret = io_async_cancel(req, issue_flags);
6243 case IORING_OP_FALLOCATE:
6244 ret = io_fallocate(req, issue_flags);
6246 case IORING_OP_OPENAT:
6247 ret = io_openat(req, issue_flags);
6249 case IORING_OP_CLOSE:
6250 ret = io_close(req, issue_flags);
6252 case IORING_OP_FILES_UPDATE:
6253 ret = io_files_update(req, issue_flags);
6255 case IORING_OP_STATX:
6256 ret = io_statx(req, issue_flags);
6258 case IORING_OP_FADVISE:
6259 ret = io_fadvise(req, issue_flags);
6261 case IORING_OP_MADVISE:
6262 ret = io_madvise(req, issue_flags);
6264 case IORING_OP_OPENAT2:
6265 ret = io_openat2(req, issue_flags);
6267 case IORING_OP_EPOLL_CTL:
6268 ret = io_epoll_ctl(req, issue_flags);
6270 case IORING_OP_SPLICE:
6271 ret = io_splice(req, issue_flags);
6273 case IORING_OP_PROVIDE_BUFFERS:
6274 ret = io_provide_buffers(req, issue_flags);
6276 case IORING_OP_REMOVE_BUFFERS:
6277 ret = io_remove_buffers(req, issue_flags);
6280 ret = io_tee(req, issue_flags);
6282 case IORING_OP_SHUTDOWN:
6283 ret = io_shutdown(req, issue_flags);
6285 case IORING_OP_RENAMEAT:
6286 ret = io_renameat(req, issue_flags);
6288 case IORING_OP_UNLINKAT:
6289 ret = io_unlinkat(req, issue_flags);
6297 revert_creds(creds);
6300 /* If the op doesn't have a file, we're not polling for it */
6301 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6302 io_iopoll_req_issued(req);
6307 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6309 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6311 req = io_put_req_find_next(req);
6312 return req ? &req->work : NULL;
6315 static void io_wq_submit_work(struct io_wq_work *work)
6317 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6318 struct io_kiocb *timeout;
6321 /* one will be dropped by ->io_free_work() after returning to io-wq */
6322 if (!(req->flags & REQ_F_REFCOUNT))
6323 __io_req_set_refcount(req, 2);
6327 timeout = io_prep_linked_timeout(req);
6329 io_queue_linked_timeout(timeout);
6331 if (work->flags & IO_WQ_WORK_CANCEL)
6336 ret = io_issue_sqe(req, 0);
6338 * We can get EAGAIN for polled IO even though we're
6339 * forcing a sync submission from here, since we can't
6340 * wait for request slots on the block side.
6348 /* avoid locking problems by failing it from a clean context */
6350 io_req_task_queue_fail(req, ret);
6353 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6356 return &table->files[i];
6359 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6362 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6364 return (struct file *) (slot->file_ptr & FFS_MASK);
6367 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6369 unsigned long file_ptr = (unsigned long) file;
6371 if (__io_file_supports_nowait(file, READ))
6372 file_ptr |= FFS_ASYNC_READ;
6373 if (__io_file_supports_nowait(file, WRITE))
6374 file_ptr |= FFS_ASYNC_WRITE;
6375 if (S_ISREG(file_inode(file)->i_mode))
6376 file_ptr |= FFS_ISREG;
6377 file_slot->file_ptr = file_ptr;
6380 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6381 struct io_kiocb *req, int fd)
6384 unsigned long file_ptr;
6386 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6388 fd = array_index_nospec(fd, ctx->nr_user_files);
6389 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6390 file = (struct file *) (file_ptr & FFS_MASK);
6391 file_ptr &= ~FFS_MASK;
6392 /* mask in overlapping REQ_F and FFS bits */
6393 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6394 io_req_set_rsrc_node(req);
6398 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6399 struct io_kiocb *req, int fd)
6401 struct file *file = fget(fd);
6403 trace_io_uring_file_get(ctx, fd);
6405 /* we don't allow fixed io_uring files */
6406 if (file && unlikely(file->f_op == &io_uring_fops))
6407 io_req_track_inflight(req);
6411 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6412 struct io_kiocb *req, int fd, bool fixed)
6415 return io_file_get_fixed(ctx, req, fd);
6417 return io_file_get_normal(ctx, req, fd);
6420 static void io_req_task_link_timeout(struct io_kiocb *req)
6422 struct io_kiocb *prev = req->timeout.prev;
6423 struct io_ring_ctx *ctx = req->ctx;
6426 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6430 io_req_complete_post(req, -ETIME, 0);
6434 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6436 struct io_timeout_data *data = container_of(timer,
6437 struct io_timeout_data, timer);
6438 struct io_kiocb *prev, *req = data->req;
6439 struct io_ring_ctx *ctx = req->ctx;
6440 unsigned long flags;
6442 spin_lock_irqsave(&ctx->timeout_lock, flags);
6443 prev = req->timeout.head;
6444 req->timeout.head = NULL;
6447 * We don't expect the list to be empty, that will only happen if we
6448 * race with the completion of the linked work.
6451 io_remove_next_linked(prev);
6452 if (!req_ref_inc_not_zero(prev))
6455 req->timeout.prev = prev;
6456 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6458 req->io_task_work.func = io_req_task_link_timeout;
6459 io_req_task_work_add(req);
6460 return HRTIMER_NORESTART;
6463 static void io_queue_linked_timeout(struct io_kiocb *req)
6465 struct io_ring_ctx *ctx = req->ctx;
6467 spin_lock_irq(&ctx->timeout_lock);
6469 * If the back reference is NULL, then our linked request finished
6470 * before we got a chance to setup the timer
6472 if (req->timeout.head) {
6473 struct io_timeout_data *data = req->async_data;
6475 data->timer.function = io_link_timeout_fn;
6476 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6479 spin_unlock_irq(&ctx->timeout_lock);
6480 /* drop submission reference */
6484 static void __io_queue_sqe(struct io_kiocb *req)
6485 __must_hold(&req->ctx->uring_lock)
6487 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6491 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6494 * We async punt it if the file wasn't marked NOWAIT, or if the file
6495 * doesn't support non-blocking read/write attempts
6498 if (req->flags & REQ_F_COMPLETE_INLINE) {
6499 struct io_ring_ctx *ctx = req->ctx;
6500 struct io_submit_state *state = &ctx->submit_state;
6502 state->compl_reqs[state->compl_nr++] = req;
6503 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6504 io_submit_flush_completions(ctx);
6506 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6507 switch (io_arm_poll_handler(req)) {
6508 case IO_APOLL_READY:
6510 case IO_APOLL_ABORTED:
6512 * Queued up for async execution, worker will release
6513 * submit reference when the iocb is actually submitted.
6515 io_queue_async_work(req);
6519 io_req_complete_failed(req, ret);
6522 io_queue_linked_timeout(linked_timeout);
6525 static inline void io_queue_sqe(struct io_kiocb *req)
6526 __must_hold(&req->ctx->uring_lock)
6528 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6531 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6532 __io_queue_sqe(req);
6534 int ret = io_req_prep_async(req);
6537 io_req_complete_failed(req, ret);
6539 io_queue_async_work(req);
6544 * Check SQE restrictions (opcode and flags).
6546 * Returns 'true' if SQE is allowed, 'false' otherwise.
6548 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6549 struct io_kiocb *req,
6550 unsigned int sqe_flags)
6552 if (likely(!ctx->restricted))
6555 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6558 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6559 ctx->restrictions.sqe_flags_required)
6562 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6563 ctx->restrictions.sqe_flags_required))
6569 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6570 const struct io_uring_sqe *sqe)
6571 __must_hold(&ctx->uring_lock)
6573 struct io_submit_state *state;
6574 unsigned int sqe_flags;
6575 int personality, ret = 0;
6577 /* req is partially pre-initialised, see io_preinit_req() */
6578 req->opcode = READ_ONCE(sqe->opcode);
6579 /* same numerical values with corresponding REQ_F_*, safe to copy */
6580 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6581 req->user_data = READ_ONCE(sqe->user_data);
6583 req->fixed_rsrc_refs = NULL;
6584 req->task = current;
6586 /* enforce forwards compatibility on users */
6587 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6589 if (unlikely(req->opcode >= IORING_OP_LAST))
6591 if (!io_check_restriction(ctx, req, sqe_flags))
6594 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6595 !io_op_defs[req->opcode].buffer_select)
6597 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6598 ctx->drain_active = true;
6600 personality = READ_ONCE(sqe->personality);
6602 req->creds = xa_load(&ctx->personalities, personality);
6605 get_cred(req->creds);
6606 req->flags |= REQ_F_CREDS;
6608 state = &ctx->submit_state;
6611 * Plug now if we have more than 1 IO left after this, and the target
6612 * is potentially a read/write to block based storage.
6614 if (!state->plug_started && state->ios_left > 1 &&
6615 io_op_defs[req->opcode].plug) {
6616 blk_start_plug(&state->plug);
6617 state->plug_started = true;
6620 if (io_op_defs[req->opcode].needs_file) {
6621 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6622 (sqe_flags & IOSQE_FIXED_FILE));
6623 if (unlikely(!req->file))
6631 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6632 const struct io_uring_sqe *sqe)
6633 __must_hold(&ctx->uring_lock)
6635 struct io_submit_link *link = &ctx->submit_state.link;
6638 ret = io_init_req(ctx, req, sqe);
6639 if (unlikely(ret)) {
6642 /* fail even hard links since we don't submit */
6643 req_set_fail(link->head);
6644 io_req_complete_failed(link->head, -ECANCELED);
6647 io_req_complete_failed(req, ret);
6651 ret = io_req_prep(req, sqe);
6655 /* don't need @sqe from now on */
6656 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6658 ctx->flags & IORING_SETUP_SQPOLL);
6661 * If we already have a head request, queue this one for async
6662 * submittal once the head completes. If we don't have a head but
6663 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6664 * submitted sync once the chain is complete. If none of those
6665 * conditions are true (normal request), then just queue it.
6668 struct io_kiocb *head = link->head;
6670 ret = io_req_prep_async(req);
6673 trace_io_uring_link(ctx, req, head);
6674 link->last->link = req;
6677 /* last request of a link, enqueue the link */
6678 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6683 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6695 * Batched submission is done, ensure local IO is flushed out.
6697 static void io_submit_state_end(struct io_submit_state *state,
6698 struct io_ring_ctx *ctx)
6700 if (state->link.head)
6701 io_queue_sqe(state->link.head);
6702 if (state->compl_nr)
6703 io_submit_flush_completions(ctx);
6704 if (state->plug_started)
6705 blk_finish_plug(&state->plug);
6709 * Start submission side cache.
6711 static void io_submit_state_start(struct io_submit_state *state,
6712 unsigned int max_ios)
6714 state->plug_started = false;
6715 state->ios_left = max_ios;
6716 /* set only head, no need to init link_last in advance */
6717 state->link.head = NULL;
6720 static void io_commit_sqring(struct io_ring_ctx *ctx)
6722 struct io_rings *rings = ctx->rings;
6725 * Ensure any loads from the SQEs are done at this point,
6726 * since once we write the new head, the application could
6727 * write new data to them.
6729 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6733 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6734 * that is mapped by userspace. This means that care needs to be taken to
6735 * ensure that reads are stable, as we cannot rely on userspace always
6736 * being a good citizen. If members of the sqe are validated and then later
6737 * used, it's important that those reads are done through READ_ONCE() to
6738 * prevent a re-load down the line.
6740 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6742 unsigned head, mask = ctx->sq_entries - 1;
6743 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6746 * The cached sq head (or cq tail) serves two purposes:
6748 * 1) allows us to batch the cost of updating the user visible
6750 * 2) allows the kernel side to track the head on its own, even
6751 * though the application is the one updating it.
6753 head = READ_ONCE(ctx->sq_array[sq_idx]);
6754 if (likely(head < ctx->sq_entries))
6755 return &ctx->sq_sqes[head];
6757 /* drop invalid entries */
6759 WRITE_ONCE(ctx->rings->sq_dropped,
6760 READ_ONCE(ctx->rings->sq_dropped) + 1);
6764 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6765 __must_hold(&ctx->uring_lock)
6767 struct io_uring_task *tctx;
6770 /* make sure SQ entry isn't read before tail */
6771 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6772 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6775 tctx = current->io_uring;
6776 tctx->cached_refs -= nr;
6777 if (unlikely(tctx->cached_refs < 0)) {
6778 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6780 percpu_counter_add(&tctx->inflight, refill);
6781 refcount_add(refill, ¤t->usage);
6782 tctx->cached_refs += refill;
6784 io_submit_state_start(&ctx->submit_state, nr);
6786 while (submitted < nr) {
6787 const struct io_uring_sqe *sqe;
6788 struct io_kiocb *req;
6790 req = io_alloc_req(ctx);
6791 if (unlikely(!req)) {
6793 submitted = -EAGAIN;
6796 sqe = io_get_sqe(ctx);
6797 if (unlikely(!sqe)) {
6798 kmem_cache_free(req_cachep, req);
6801 /* will complete beyond this point, count as submitted */
6803 if (io_submit_sqe(ctx, req, sqe))
6807 if (unlikely(submitted != nr)) {
6808 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6809 int unused = nr - ref_used;
6811 current->io_uring->cached_refs += unused;
6812 percpu_ref_put_many(&ctx->refs, unused);
6815 io_submit_state_end(&ctx->submit_state, ctx);
6816 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6817 io_commit_sqring(ctx);
6822 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6824 return READ_ONCE(sqd->state);
6827 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6829 /* Tell userspace we may need a wakeup call */
6830 spin_lock(&ctx->completion_lock);
6831 WRITE_ONCE(ctx->rings->sq_flags,
6832 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6833 spin_unlock(&ctx->completion_lock);
6836 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6838 spin_lock(&ctx->completion_lock);
6839 WRITE_ONCE(ctx->rings->sq_flags,
6840 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6841 spin_unlock(&ctx->completion_lock);
6844 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6846 unsigned int to_submit;
6849 to_submit = io_sqring_entries(ctx);
6850 /* if we're handling multiple rings, cap submit size for fairness */
6851 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6852 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6854 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6855 unsigned nr_events = 0;
6856 const struct cred *creds = NULL;
6858 if (ctx->sq_creds != current_cred())
6859 creds = override_creds(ctx->sq_creds);
6861 mutex_lock(&ctx->uring_lock);
6862 if (!list_empty(&ctx->iopoll_list))
6863 io_do_iopoll(ctx, &nr_events, 0, true);
6866 * Don't submit if refs are dying, good for io_uring_register(),
6867 * but also it is relied upon by io_ring_exit_work()
6869 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6870 !(ctx->flags & IORING_SETUP_R_DISABLED))
6871 ret = io_submit_sqes(ctx, to_submit);
6872 mutex_unlock(&ctx->uring_lock);
6874 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6875 wake_up(&ctx->sqo_sq_wait);
6877 revert_creds(creds);
6883 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6885 struct io_ring_ctx *ctx;
6886 unsigned sq_thread_idle = 0;
6888 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6889 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6890 sqd->sq_thread_idle = sq_thread_idle;
6893 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6895 bool did_sig = false;
6896 struct ksignal ksig;
6898 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6899 signal_pending(current)) {
6900 mutex_unlock(&sqd->lock);
6901 if (signal_pending(current))
6902 did_sig = get_signal(&ksig);
6904 mutex_lock(&sqd->lock);
6906 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6909 static int io_sq_thread(void *data)
6911 struct io_sq_data *sqd = data;
6912 struct io_ring_ctx *ctx;
6913 unsigned long timeout = 0;
6914 char buf[TASK_COMM_LEN];
6917 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6918 set_task_comm(current, buf);
6920 if (sqd->sq_cpu != -1)
6921 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6923 set_cpus_allowed_ptr(current, cpu_online_mask);
6924 current->flags |= PF_NO_SETAFFINITY;
6926 mutex_lock(&sqd->lock);
6928 bool cap_entries, sqt_spin = false;
6930 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6931 if (io_sqd_handle_event(sqd))
6933 timeout = jiffies + sqd->sq_thread_idle;
6936 cap_entries = !list_is_singular(&sqd->ctx_list);
6937 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6938 int ret = __io_sq_thread(ctx, cap_entries);
6940 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6943 if (io_run_task_work())
6946 if (sqt_spin || !time_after(jiffies, timeout)) {
6949 timeout = jiffies + sqd->sq_thread_idle;
6953 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6954 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6955 bool needs_sched = true;
6957 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6958 io_ring_set_wakeup_flag(ctx);
6960 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6961 !list_empty_careful(&ctx->iopoll_list)) {
6962 needs_sched = false;
6965 if (io_sqring_entries(ctx)) {
6966 needs_sched = false;
6972 mutex_unlock(&sqd->lock);
6974 mutex_lock(&sqd->lock);
6976 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6977 io_ring_clear_wakeup_flag(ctx);
6980 finish_wait(&sqd->wait, &wait);
6981 timeout = jiffies + sqd->sq_thread_idle;
6984 io_uring_cancel_generic(true, sqd);
6986 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6987 io_ring_set_wakeup_flag(ctx);
6989 mutex_unlock(&sqd->lock);
6991 complete(&sqd->exited);
6995 struct io_wait_queue {
6996 struct wait_queue_entry wq;
6997 struct io_ring_ctx *ctx;
6999 unsigned nr_timeouts;
7002 static inline bool io_should_wake(struct io_wait_queue *iowq)
7004 struct io_ring_ctx *ctx = iowq->ctx;
7005 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7008 * Wake up if we have enough events, or if a timeout occurred since we
7009 * started waiting. For timeouts, we always want to return to userspace,
7010 * regardless of event count.
7012 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7015 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7016 int wake_flags, void *key)
7018 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7022 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7023 * the task, and the next invocation will do it.
7025 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7026 return autoremove_wake_function(curr, mode, wake_flags, key);
7030 static int io_run_task_work_sig(void)
7032 if (io_run_task_work())
7034 if (!signal_pending(current))
7036 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7037 return -ERESTARTSYS;
7041 /* when returns >0, the caller should retry */
7042 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7043 struct io_wait_queue *iowq,
7044 signed long *timeout)
7048 /* make sure we run task_work before checking for signals */
7049 ret = io_run_task_work_sig();
7050 if (ret || io_should_wake(iowq))
7052 /* let the caller flush overflows, retry */
7053 if (test_bit(0, &ctx->check_cq_overflow))
7056 *timeout = schedule_timeout(*timeout);
7057 return !*timeout ? -ETIME : 1;
7061 * Wait until events become available, if we don't already have some. The
7062 * application must reap them itself, as they reside on the shared cq ring.
7064 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7065 const sigset_t __user *sig, size_t sigsz,
7066 struct __kernel_timespec __user *uts)
7068 struct io_wait_queue iowq;
7069 struct io_rings *rings = ctx->rings;
7070 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7074 io_cqring_overflow_flush(ctx);
7075 if (io_cqring_events(ctx) >= min_events)
7077 if (!io_run_task_work())
7082 #ifdef CONFIG_COMPAT
7083 if (in_compat_syscall())
7084 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7088 ret = set_user_sigmask(sig, sigsz);
7095 struct timespec64 ts;
7097 if (get_timespec64(&ts, uts))
7099 timeout = timespec64_to_jiffies(&ts);
7102 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7103 iowq.wq.private = current;
7104 INIT_LIST_HEAD(&iowq.wq.entry);
7106 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7107 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7109 trace_io_uring_cqring_wait(ctx, min_events);
7111 /* if we can't even flush overflow, don't wait for more */
7112 if (!io_cqring_overflow_flush(ctx)) {
7116 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7117 TASK_INTERRUPTIBLE);
7118 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7119 finish_wait(&ctx->cq_wait, &iowq.wq);
7123 restore_saved_sigmask_unless(ret == -EINTR);
7125 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7128 static void io_free_page_table(void **table, size_t size)
7130 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7132 for (i = 0; i < nr_tables; i++)
7137 static void **io_alloc_page_table(size_t size)
7139 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7140 size_t init_size = size;
7143 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7147 for (i = 0; i < nr_tables; i++) {
7148 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7150 table[i] = kzalloc(this_size, GFP_KERNEL);
7152 io_free_page_table(table, init_size);
7160 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7162 percpu_ref_exit(&ref_node->refs);
7166 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7168 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7169 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7170 unsigned long flags;
7171 bool first_add = false;
7173 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7176 while (!list_empty(&ctx->rsrc_ref_list)) {
7177 node = list_first_entry(&ctx->rsrc_ref_list,
7178 struct io_rsrc_node, node);
7179 /* recycle ref nodes in order */
7182 list_del(&node->node);
7183 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7185 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7188 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7191 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7193 struct io_rsrc_node *ref_node;
7195 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7199 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7204 INIT_LIST_HEAD(&ref_node->node);
7205 INIT_LIST_HEAD(&ref_node->rsrc_list);
7206 ref_node->done = false;
7210 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7211 struct io_rsrc_data *data_to_kill)
7213 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7214 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7217 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7219 rsrc_node->rsrc_data = data_to_kill;
7220 spin_lock_irq(&ctx->rsrc_ref_lock);
7221 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7222 spin_unlock_irq(&ctx->rsrc_ref_lock);
7224 atomic_inc(&data_to_kill->refs);
7225 percpu_ref_kill(&rsrc_node->refs);
7226 ctx->rsrc_node = NULL;
7229 if (!ctx->rsrc_node) {
7230 ctx->rsrc_node = ctx->rsrc_backup_node;
7231 ctx->rsrc_backup_node = NULL;
7235 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7237 if (ctx->rsrc_backup_node)
7239 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7240 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7243 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7247 /* As we may drop ->uring_lock, other task may have started quiesce */
7251 data->quiesce = true;
7253 ret = io_rsrc_node_switch_start(ctx);
7256 io_rsrc_node_switch(ctx, data);
7258 /* kill initial ref, already quiesced if zero */
7259 if (atomic_dec_and_test(&data->refs))
7261 mutex_unlock(&ctx->uring_lock);
7262 flush_delayed_work(&ctx->rsrc_put_work);
7263 ret = wait_for_completion_interruptible(&data->done);
7265 mutex_lock(&ctx->uring_lock);
7269 atomic_inc(&data->refs);
7270 /* wait for all works potentially completing data->done */
7271 flush_delayed_work(&ctx->rsrc_put_work);
7272 reinit_completion(&data->done);
7274 ret = io_run_task_work_sig();
7275 mutex_lock(&ctx->uring_lock);
7277 data->quiesce = false;
7282 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7284 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7285 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7287 return &data->tags[table_idx][off];
7290 static void io_rsrc_data_free(struct io_rsrc_data *data)
7292 size_t size = data->nr * sizeof(data->tags[0][0]);
7295 io_free_page_table((void **)data->tags, size);
7299 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7300 u64 __user *utags, unsigned nr,
7301 struct io_rsrc_data **pdata)
7303 struct io_rsrc_data *data;
7307 data = kzalloc(sizeof(*data), GFP_KERNEL);
7310 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7318 data->do_put = do_put;
7321 for (i = 0; i < nr; i++) {
7322 u64 *tag_slot = io_get_tag_slot(data, i);
7324 if (copy_from_user(tag_slot, &utags[i],
7330 atomic_set(&data->refs, 1);
7331 init_completion(&data->done);
7335 io_rsrc_data_free(data);
7339 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7341 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7342 return !!table->files;
7345 static void io_free_file_tables(struct io_file_table *table)
7347 kvfree(table->files);
7348 table->files = NULL;
7351 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7353 #if defined(CONFIG_UNIX)
7354 if (ctx->ring_sock) {
7355 struct sock *sock = ctx->ring_sock->sk;
7356 struct sk_buff *skb;
7358 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7364 for (i = 0; i < ctx->nr_user_files; i++) {
7367 file = io_file_from_index(ctx, i);
7372 io_free_file_tables(&ctx->file_table);
7373 io_rsrc_data_free(ctx->file_data);
7374 ctx->file_data = NULL;
7375 ctx->nr_user_files = 0;
7378 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7382 if (!ctx->file_data)
7384 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7386 __io_sqe_files_unregister(ctx);
7390 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7391 __releases(&sqd->lock)
7393 WARN_ON_ONCE(sqd->thread == current);
7396 * Do the dance but not conditional clear_bit() because it'd race with
7397 * other threads incrementing park_pending and setting the bit.
7399 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7400 if (atomic_dec_return(&sqd->park_pending))
7401 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7402 mutex_unlock(&sqd->lock);
7405 static void io_sq_thread_park(struct io_sq_data *sqd)
7406 __acquires(&sqd->lock)
7408 WARN_ON_ONCE(sqd->thread == current);
7410 atomic_inc(&sqd->park_pending);
7411 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7412 mutex_lock(&sqd->lock);
7414 wake_up_process(sqd->thread);
7417 static void io_sq_thread_stop(struct io_sq_data *sqd)
7419 WARN_ON_ONCE(sqd->thread == current);
7420 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7422 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7423 mutex_lock(&sqd->lock);
7425 wake_up_process(sqd->thread);
7426 mutex_unlock(&sqd->lock);
7427 wait_for_completion(&sqd->exited);
7430 static void io_put_sq_data(struct io_sq_data *sqd)
7432 if (refcount_dec_and_test(&sqd->refs)) {
7433 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7435 io_sq_thread_stop(sqd);
7440 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7442 struct io_sq_data *sqd = ctx->sq_data;
7445 io_sq_thread_park(sqd);
7446 list_del_init(&ctx->sqd_list);
7447 io_sqd_update_thread_idle(sqd);
7448 io_sq_thread_unpark(sqd);
7450 io_put_sq_data(sqd);
7451 ctx->sq_data = NULL;
7455 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7457 struct io_ring_ctx *ctx_attach;
7458 struct io_sq_data *sqd;
7461 f = fdget(p->wq_fd);
7463 return ERR_PTR(-ENXIO);
7464 if (f.file->f_op != &io_uring_fops) {
7466 return ERR_PTR(-EINVAL);
7469 ctx_attach = f.file->private_data;
7470 sqd = ctx_attach->sq_data;
7473 return ERR_PTR(-EINVAL);
7475 if (sqd->task_tgid != current->tgid) {
7477 return ERR_PTR(-EPERM);
7480 refcount_inc(&sqd->refs);
7485 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7488 struct io_sq_data *sqd;
7491 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7492 sqd = io_attach_sq_data(p);
7497 /* fall through for EPERM case, setup new sqd/task */
7498 if (PTR_ERR(sqd) != -EPERM)
7502 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7504 return ERR_PTR(-ENOMEM);
7506 atomic_set(&sqd->park_pending, 0);
7507 refcount_set(&sqd->refs, 1);
7508 INIT_LIST_HEAD(&sqd->ctx_list);
7509 mutex_init(&sqd->lock);
7510 init_waitqueue_head(&sqd->wait);
7511 init_completion(&sqd->exited);
7515 #if defined(CONFIG_UNIX)
7517 * Ensure the UNIX gc is aware of our file set, so we are certain that
7518 * the io_uring can be safely unregistered on process exit, even if we have
7519 * loops in the file referencing.
7521 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7523 struct sock *sk = ctx->ring_sock->sk;
7524 struct scm_fp_list *fpl;
7525 struct sk_buff *skb;
7528 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7532 skb = alloc_skb(0, GFP_KERNEL);
7541 fpl->user = get_uid(current_user());
7542 for (i = 0; i < nr; i++) {
7543 struct file *file = io_file_from_index(ctx, i + offset);
7547 fpl->fp[nr_files] = get_file(file);
7548 unix_inflight(fpl->user, fpl->fp[nr_files]);
7553 fpl->max = SCM_MAX_FD;
7554 fpl->count = nr_files;
7555 UNIXCB(skb).fp = fpl;
7556 skb->destructor = unix_destruct_scm;
7557 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7558 skb_queue_head(&sk->sk_receive_queue, skb);
7560 for (i = 0; i < nr_files; i++)
7571 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7572 * causes regular reference counting to break down. We rely on the UNIX
7573 * garbage collection to take care of this problem for us.
7575 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7577 unsigned left, total;
7581 left = ctx->nr_user_files;
7583 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7585 ret = __io_sqe_files_scm(ctx, this_files, total);
7589 total += this_files;
7595 while (total < ctx->nr_user_files) {
7596 struct file *file = io_file_from_index(ctx, total);
7606 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7612 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7614 struct file *file = prsrc->file;
7615 #if defined(CONFIG_UNIX)
7616 struct sock *sock = ctx->ring_sock->sk;
7617 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7618 struct sk_buff *skb;
7621 __skb_queue_head_init(&list);
7624 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7625 * remove this entry and rearrange the file array.
7627 skb = skb_dequeue(head);
7629 struct scm_fp_list *fp;
7631 fp = UNIXCB(skb).fp;
7632 for (i = 0; i < fp->count; i++) {
7635 if (fp->fp[i] != file)
7638 unix_notinflight(fp->user, fp->fp[i]);
7639 left = fp->count - 1 - i;
7641 memmove(&fp->fp[i], &fp->fp[i + 1],
7642 left * sizeof(struct file *));
7649 __skb_queue_tail(&list, skb);
7659 __skb_queue_tail(&list, skb);
7661 skb = skb_dequeue(head);
7664 if (skb_peek(&list)) {
7665 spin_lock_irq(&head->lock);
7666 while ((skb = __skb_dequeue(&list)) != NULL)
7667 __skb_queue_tail(head, skb);
7668 spin_unlock_irq(&head->lock);
7675 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7677 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7678 struct io_ring_ctx *ctx = rsrc_data->ctx;
7679 struct io_rsrc_put *prsrc, *tmp;
7681 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7682 list_del(&prsrc->list);
7685 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7687 io_ring_submit_lock(ctx, lock_ring);
7688 spin_lock(&ctx->completion_lock);
7689 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7691 io_commit_cqring(ctx);
7692 spin_unlock(&ctx->completion_lock);
7693 io_cqring_ev_posted(ctx);
7694 io_ring_submit_unlock(ctx, lock_ring);
7697 rsrc_data->do_put(ctx, prsrc);
7701 io_rsrc_node_destroy(ref_node);
7702 if (atomic_dec_and_test(&rsrc_data->refs))
7703 complete(&rsrc_data->done);
7706 static void io_rsrc_put_work(struct work_struct *work)
7708 struct io_ring_ctx *ctx;
7709 struct llist_node *node;
7711 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7712 node = llist_del_all(&ctx->rsrc_put_llist);
7715 struct io_rsrc_node *ref_node;
7716 struct llist_node *next = node->next;
7718 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7719 __io_rsrc_put_work(ref_node);
7724 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7725 unsigned nr_args, u64 __user *tags)
7727 __s32 __user *fds = (__s32 __user *) arg;
7736 if (nr_args > IORING_MAX_FIXED_FILES)
7738 ret = io_rsrc_node_switch_start(ctx);
7741 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7747 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7750 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7751 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7755 /* allow sparse sets */
7758 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7765 if (unlikely(!file))
7769 * Don't allow io_uring instances to be registered. If UNIX
7770 * isn't enabled, then this causes a reference cycle and this
7771 * instance can never get freed. If UNIX is enabled we'll
7772 * handle it just fine, but there's still no point in allowing
7773 * a ring fd as it doesn't support regular read/write anyway.
7775 if (file->f_op == &io_uring_fops) {
7779 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7782 ret = io_sqe_files_scm(ctx);
7784 __io_sqe_files_unregister(ctx);
7788 io_rsrc_node_switch(ctx, NULL);
7791 for (i = 0; i < ctx->nr_user_files; i++) {
7792 file = io_file_from_index(ctx, i);
7796 io_free_file_tables(&ctx->file_table);
7797 ctx->nr_user_files = 0;
7799 io_rsrc_data_free(ctx->file_data);
7800 ctx->file_data = NULL;
7804 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7807 #if defined(CONFIG_UNIX)
7808 struct sock *sock = ctx->ring_sock->sk;
7809 struct sk_buff_head *head = &sock->sk_receive_queue;
7810 struct sk_buff *skb;
7813 * See if we can merge this file into an existing skb SCM_RIGHTS
7814 * file set. If there's no room, fall back to allocating a new skb
7815 * and filling it in.
7817 spin_lock_irq(&head->lock);
7818 skb = skb_peek(head);
7820 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7822 if (fpl->count < SCM_MAX_FD) {
7823 __skb_unlink(skb, head);
7824 spin_unlock_irq(&head->lock);
7825 fpl->fp[fpl->count] = get_file(file);
7826 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7828 spin_lock_irq(&head->lock);
7829 __skb_queue_head(head, skb);
7834 spin_unlock_irq(&head->lock);
7841 return __io_sqe_files_scm(ctx, 1, index);
7847 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7848 struct io_rsrc_node *node, void *rsrc)
7850 struct io_rsrc_put *prsrc;
7852 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7856 prsrc->tag = *io_get_tag_slot(data, idx);
7858 list_add(&prsrc->list, &node->rsrc_list);
7862 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7863 struct io_uring_rsrc_update2 *up,
7866 u64 __user *tags = u64_to_user_ptr(up->tags);
7867 __s32 __user *fds = u64_to_user_ptr(up->data);
7868 struct io_rsrc_data *data = ctx->file_data;
7869 struct io_fixed_file *file_slot;
7873 bool needs_switch = false;
7875 if (!ctx->file_data)
7877 if (up->offset + nr_args > ctx->nr_user_files)
7880 for (done = 0; done < nr_args; done++) {
7883 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7884 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7888 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7892 if (fd == IORING_REGISTER_FILES_SKIP)
7895 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7896 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7898 if (file_slot->file_ptr) {
7899 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7900 err = io_queue_rsrc_removal(data, up->offset + done,
7901 ctx->rsrc_node, file);
7904 file_slot->file_ptr = 0;
7905 needs_switch = true;
7914 * Don't allow io_uring instances to be registered. If
7915 * UNIX isn't enabled, then this causes a reference
7916 * cycle and this instance can never get freed. If UNIX
7917 * is enabled we'll handle it just fine, but there's
7918 * still no point in allowing a ring fd as it doesn't
7919 * support regular read/write anyway.
7921 if (file->f_op == &io_uring_fops) {
7926 *io_get_tag_slot(data, up->offset + done) = tag;
7927 io_fixed_file_set(file_slot, file);
7928 err = io_sqe_file_register(ctx, file, i);
7930 file_slot->file_ptr = 0;
7938 io_rsrc_node_switch(ctx, data);
7939 return done ? done : err;
7942 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7943 struct task_struct *task)
7945 struct io_wq_hash *hash;
7946 struct io_wq_data data;
7947 unsigned int concurrency;
7949 mutex_lock(&ctx->uring_lock);
7950 hash = ctx->hash_map;
7952 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7954 mutex_unlock(&ctx->uring_lock);
7955 return ERR_PTR(-ENOMEM);
7957 refcount_set(&hash->refs, 1);
7958 init_waitqueue_head(&hash->wait);
7959 ctx->hash_map = hash;
7961 mutex_unlock(&ctx->uring_lock);
7965 data.free_work = io_wq_free_work;
7966 data.do_work = io_wq_submit_work;
7968 /* Do QD, or 4 * CPUS, whatever is smallest */
7969 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7971 return io_wq_create(concurrency, &data);
7974 static int io_uring_alloc_task_context(struct task_struct *task,
7975 struct io_ring_ctx *ctx)
7977 struct io_uring_task *tctx;
7980 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7981 if (unlikely(!tctx))
7984 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7985 if (unlikely(ret)) {
7990 tctx->io_wq = io_init_wq_offload(ctx, task);
7991 if (IS_ERR(tctx->io_wq)) {
7992 ret = PTR_ERR(tctx->io_wq);
7993 percpu_counter_destroy(&tctx->inflight);
7999 init_waitqueue_head(&tctx->wait);
8000 atomic_set(&tctx->in_idle, 0);
8001 atomic_set(&tctx->inflight_tracked, 0);
8002 task->io_uring = tctx;
8003 spin_lock_init(&tctx->task_lock);
8004 INIT_WQ_LIST(&tctx->task_list);
8005 init_task_work(&tctx->task_work, tctx_task_work);
8009 void __io_uring_free(struct task_struct *tsk)
8011 struct io_uring_task *tctx = tsk->io_uring;
8013 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8014 WARN_ON_ONCE(tctx->io_wq);
8015 WARN_ON_ONCE(tctx->cached_refs);
8017 percpu_counter_destroy(&tctx->inflight);
8019 tsk->io_uring = NULL;
8022 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8023 struct io_uring_params *p)
8027 /* Retain compatibility with failing for an invalid attach attempt */
8028 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8029 IORING_SETUP_ATTACH_WQ) {
8032 f = fdget(p->wq_fd);
8035 if (f.file->f_op != &io_uring_fops) {
8041 if (ctx->flags & IORING_SETUP_SQPOLL) {
8042 struct task_struct *tsk;
8043 struct io_sq_data *sqd;
8046 sqd = io_get_sq_data(p, &attached);
8052 ctx->sq_creds = get_current_cred();
8054 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8055 if (!ctx->sq_thread_idle)
8056 ctx->sq_thread_idle = HZ;
8058 io_sq_thread_park(sqd);
8059 list_add(&ctx->sqd_list, &sqd->ctx_list);
8060 io_sqd_update_thread_idle(sqd);
8061 /* don't attach to a dying SQPOLL thread, would be racy */
8062 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8063 io_sq_thread_unpark(sqd);
8070 if (p->flags & IORING_SETUP_SQ_AFF) {
8071 int cpu = p->sq_thread_cpu;
8074 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8081 sqd->task_pid = current->pid;
8082 sqd->task_tgid = current->tgid;
8083 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8090 ret = io_uring_alloc_task_context(tsk, ctx);
8091 wake_up_new_task(tsk);
8094 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8095 /* Can't have SQ_AFF without SQPOLL */
8102 complete(&ctx->sq_data->exited);
8104 io_sq_thread_finish(ctx);
8108 static inline void __io_unaccount_mem(struct user_struct *user,
8109 unsigned long nr_pages)
8111 atomic_long_sub(nr_pages, &user->locked_vm);
8114 static inline int __io_account_mem(struct user_struct *user,
8115 unsigned long nr_pages)
8117 unsigned long page_limit, cur_pages, new_pages;
8119 /* Don't allow more pages than we can safely lock */
8120 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8123 cur_pages = atomic_long_read(&user->locked_vm);
8124 new_pages = cur_pages + nr_pages;
8125 if (new_pages > page_limit)
8127 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8128 new_pages) != cur_pages);
8133 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8136 __io_unaccount_mem(ctx->user, nr_pages);
8138 if (ctx->mm_account)
8139 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8142 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8147 ret = __io_account_mem(ctx->user, nr_pages);
8152 if (ctx->mm_account)
8153 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8158 static void io_mem_free(void *ptr)
8165 page = virt_to_head_page(ptr);
8166 if (put_page_testzero(page))
8167 free_compound_page(page);
8170 static void *io_mem_alloc(size_t size)
8172 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8173 __GFP_NORETRY | __GFP_ACCOUNT;
8175 return (void *) __get_free_pages(gfp_flags, get_order(size));
8178 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8181 struct io_rings *rings;
8182 size_t off, sq_array_size;
8184 off = struct_size(rings, cqes, cq_entries);
8185 if (off == SIZE_MAX)
8189 off = ALIGN(off, SMP_CACHE_BYTES);
8197 sq_array_size = array_size(sizeof(u32), sq_entries);
8198 if (sq_array_size == SIZE_MAX)
8201 if (check_add_overflow(off, sq_array_size, &off))
8207 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8209 struct io_mapped_ubuf *imu = *slot;
8212 if (imu != ctx->dummy_ubuf) {
8213 for (i = 0; i < imu->nr_bvecs; i++)
8214 unpin_user_page(imu->bvec[i].bv_page);
8215 if (imu->acct_pages)
8216 io_unaccount_mem(ctx, imu->acct_pages);
8222 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8224 io_buffer_unmap(ctx, &prsrc->buf);
8228 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8232 for (i = 0; i < ctx->nr_user_bufs; i++)
8233 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8234 kfree(ctx->user_bufs);
8235 io_rsrc_data_free(ctx->buf_data);
8236 ctx->user_bufs = NULL;
8237 ctx->buf_data = NULL;
8238 ctx->nr_user_bufs = 0;
8241 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8248 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8250 __io_sqe_buffers_unregister(ctx);
8254 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8255 void __user *arg, unsigned index)
8257 struct iovec __user *src;
8259 #ifdef CONFIG_COMPAT
8261 struct compat_iovec __user *ciovs;
8262 struct compat_iovec ciov;
8264 ciovs = (struct compat_iovec __user *) arg;
8265 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8268 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8269 dst->iov_len = ciov.iov_len;
8273 src = (struct iovec __user *) arg;
8274 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8280 * Not super efficient, but this is just a registration time. And we do cache
8281 * the last compound head, so generally we'll only do a full search if we don't
8284 * We check if the given compound head page has already been accounted, to
8285 * avoid double accounting it. This allows us to account the full size of the
8286 * page, not just the constituent pages of a huge page.
8288 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8289 int nr_pages, struct page *hpage)
8293 /* check current page array */
8294 for (i = 0; i < nr_pages; i++) {
8295 if (!PageCompound(pages[i]))
8297 if (compound_head(pages[i]) == hpage)
8301 /* check previously registered pages */
8302 for (i = 0; i < ctx->nr_user_bufs; i++) {
8303 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8305 for (j = 0; j < imu->nr_bvecs; j++) {
8306 if (!PageCompound(imu->bvec[j].bv_page))
8308 if (compound_head(imu->bvec[j].bv_page) == hpage)
8316 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8317 int nr_pages, struct io_mapped_ubuf *imu,
8318 struct page **last_hpage)
8322 imu->acct_pages = 0;
8323 for (i = 0; i < nr_pages; i++) {
8324 if (!PageCompound(pages[i])) {
8329 hpage = compound_head(pages[i]);
8330 if (hpage == *last_hpage)
8332 *last_hpage = hpage;
8333 if (headpage_already_acct(ctx, pages, i, hpage))
8335 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8339 if (!imu->acct_pages)
8342 ret = io_account_mem(ctx, imu->acct_pages);
8344 imu->acct_pages = 0;
8348 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8349 struct io_mapped_ubuf **pimu,
8350 struct page **last_hpage)
8352 struct io_mapped_ubuf *imu = NULL;
8353 struct vm_area_struct **vmas = NULL;
8354 struct page **pages = NULL;
8355 unsigned long off, start, end, ubuf;
8357 int ret, pret, nr_pages, i;
8359 if (!iov->iov_base) {
8360 *pimu = ctx->dummy_ubuf;
8364 ubuf = (unsigned long) iov->iov_base;
8365 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8366 start = ubuf >> PAGE_SHIFT;
8367 nr_pages = end - start;
8372 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8376 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8381 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8386 mmap_read_lock(current->mm);
8387 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8389 if (pret == nr_pages) {
8390 /* don't support file backed memory */
8391 for (i = 0; i < nr_pages; i++) {
8392 struct vm_area_struct *vma = vmas[i];
8394 if (vma_is_shmem(vma))
8397 !is_file_hugepages(vma->vm_file)) {
8403 ret = pret < 0 ? pret : -EFAULT;
8405 mmap_read_unlock(current->mm);
8408 * if we did partial map, or found file backed vmas,
8409 * release any pages we did get
8412 unpin_user_pages(pages, pret);
8416 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8418 unpin_user_pages(pages, pret);
8422 off = ubuf & ~PAGE_MASK;
8423 size = iov->iov_len;
8424 for (i = 0; i < nr_pages; i++) {
8427 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8428 imu->bvec[i].bv_page = pages[i];
8429 imu->bvec[i].bv_len = vec_len;
8430 imu->bvec[i].bv_offset = off;
8434 /* store original address for later verification */
8436 imu->ubuf_end = ubuf + iov->iov_len;
8437 imu->nr_bvecs = nr_pages;
8448 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8450 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8451 return ctx->user_bufs ? 0 : -ENOMEM;
8454 static int io_buffer_validate(struct iovec *iov)
8456 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8459 * Don't impose further limits on the size and buffer
8460 * constraints here, we'll -EINVAL later when IO is
8461 * submitted if they are wrong.
8464 return iov->iov_len ? -EFAULT : 0;
8468 /* arbitrary limit, but we need something */
8469 if (iov->iov_len > SZ_1G)
8472 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8478 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8479 unsigned int nr_args, u64 __user *tags)
8481 struct page *last_hpage = NULL;
8482 struct io_rsrc_data *data;
8488 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8490 ret = io_rsrc_node_switch_start(ctx);
8493 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8496 ret = io_buffers_map_alloc(ctx, nr_args);
8498 io_rsrc_data_free(data);
8502 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8503 ret = io_copy_iov(ctx, &iov, arg, i);
8506 ret = io_buffer_validate(&iov);
8509 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8514 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8520 WARN_ON_ONCE(ctx->buf_data);
8522 ctx->buf_data = data;
8524 __io_sqe_buffers_unregister(ctx);
8526 io_rsrc_node_switch(ctx, NULL);
8530 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8531 struct io_uring_rsrc_update2 *up,
8532 unsigned int nr_args)
8534 u64 __user *tags = u64_to_user_ptr(up->tags);
8535 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8536 struct page *last_hpage = NULL;
8537 bool needs_switch = false;
8543 if (up->offset + nr_args > ctx->nr_user_bufs)
8546 for (done = 0; done < nr_args; done++) {
8547 struct io_mapped_ubuf *imu;
8548 int offset = up->offset + done;
8551 err = io_copy_iov(ctx, &iov, iovs, done);
8554 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8558 err = io_buffer_validate(&iov);
8561 if (!iov.iov_base && tag) {
8565 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8569 i = array_index_nospec(offset, ctx->nr_user_bufs);
8570 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8571 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8572 ctx->rsrc_node, ctx->user_bufs[i]);
8573 if (unlikely(err)) {
8574 io_buffer_unmap(ctx, &imu);
8577 ctx->user_bufs[i] = NULL;
8578 needs_switch = true;
8581 ctx->user_bufs[i] = imu;
8582 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8586 io_rsrc_node_switch(ctx, ctx->buf_data);
8587 return done ? done : err;
8590 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8592 __s32 __user *fds = arg;
8598 if (copy_from_user(&fd, fds, sizeof(*fds)))
8601 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8602 if (IS_ERR(ctx->cq_ev_fd)) {
8603 int ret = PTR_ERR(ctx->cq_ev_fd);
8605 ctx->cq_ev_fd = NULL;
8612 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8614 if (ctx->cq_ev_fd) {
8615 eventfd_ctx_put(ctx->cq_ev_fd);
8616 ctx->cq_ev_fd = NULL;
8623 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8625 struct io_buffer *buf;
8626 unsigned long index;
8628 xa_for_each(&ctx->io_buffers, index, buf)
8629 __io_remove_buffers(ctx, buf, index, -1U);
8632 static void io_req_cache_free(struct list_head *list)
8634 struct io_kiocb *req, *nxt;
8636 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8637 list_del(&req->inflight_entry);
8638 kmem_cache_free(req_cachep, req);
8642 static void io_req_caches_free(struct io_ring_ctx *ctx)
8644 struct io_submit_state *state = &ctx->submit_state;
8646 mutex_lock(&ctx->uring_lock);
8648 if (state->free_reqs) {
8649 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8650 state->free_reqs = 0;
8653 io_flush_cached_locked_reqs(ctx, state);
8654 io_req_cache_free(&state->free_list);
8655 mutex_unlock(&ctx->uring_lock);
8658 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8660 if (data && !atomic_dec_and_test(&data->refs))
8661 wait_for_completion(&data->done);
8664 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8666 io_sq_thread_finish(ctx);
8668 if (ctx->mm_account) {
8669 mmdrop(ctx->mm_account);
8670 ctx->mm_account = NULL;
8673 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8674 io_wait_rsrc_data(ctx->buf_data);
8675 io_wait_rsrc_data(ctx->file_data);
8677 mutex_lock(&ctx->uring_lock);
8679 __io_sqe_buffers_unregister(ctx);
8681 __io_sqe_files_unregister(ctx);
8683 __io_cqring_overflow_flush(ctx, true);
8684 mutex_unlock(&ctx->uring_lock);
8685 io_eventfd_unregister(ctx);
8686 io_destroy_buffers(ctx);
8688 put_cred(ctx->sq_creds);
8690 /* there are no registered resources left, nobody uses it */
8692 io_rsrc_node_destroy(ctx->rsrc_node);
8693 if (ctx->rsrc_backup_node)
8694 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8695 flush_delayed_work(&ctx->rsrc_put_work);
8697 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8698 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8700 #if defined(CONFIG_UNIX)
8701 if (ctx->ring_sock) {
8702 ctx->ring_sock->file = NULL; /* so that iput() is called */
8703 sock_release(ctx->ring_sock);
8707 io_mem_free(ctx->rings);
8708 io_mem_free(ctx->sq_sqes);
8710 percpu_ref_exit(&ctx->refs);
8711 free_uid(ctx->user);
8712 io_req_caches_free(ctx);
8714 io_wq_put_hash(ctx->hash_map);
8715 kfree(ctx->cancel_hash);
8716 kfree(ctx->dummy_ubuf);
8720 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8722 struct io_ring_ctx *ctx = file->private_data;
8725 poll_wait(file, &ctx->poll_wait, wait);
8727 * synchronizes with barrier from wq_has_sleeper call in
8731 if (!io_sqring_full(ctx))
8732 mask |= EPOLLOUT | EPOLLWRNORM;
8735 * Don't flush cqring overflow list here, just do a simple check.
8736 * Otherwise there could possible be ABBA deadlock:
8739 * lock(&ctx->uring_lock);
8741 * lock(&ctx->uring_lock);
8744 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8745 * pushs them to do the flush.
8747 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8748 mask |= EPOLLIN | EPOLLRDNORM;
8753 static int io_uring_fasync(int fd, struct file *file, int on)
8755 struct io_ring_ctx *ctx = file->private_data;
8757 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8760 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8762 const struct cred *creds;
8764 creds = xa_erase(&ctx->personalities, id);
8773 struct io_tctx_exit {
8774 struct callback_head task_work;
8775 struct completion completion;
8776 struct io_ring_ctx *ctx;
8779 static void io_tctx_exit_cb(struct callback_head *cb)
8781 struct io_uring_task *tctx = current->io_uring;
8782 struct io_tctx_exit *work;
8784 work = container_of(cb, struct io_tctx_exit, task_work);
8786 * When @in_idle, we're in cancellation and it's racy to remove the
8787 * node. It'll be removed by the end of cancellation, just ignore it.
8789 if (!atomic_read(&tctx->in_idle))
8790 io_uring_del_tctx_node((unsigned long)work->ctx);
8791 complete(&work->completion);
8794 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8796 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8798 return req->ctx == data;
8801 static void io_ring_exit_work(struct work_struct *work)
8803 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8804 unsigned long timeout = jiffies + HZ * 60 * 5;
8805 unsigned long interval = HZ / 20;
8806 struct io_tctx_exit exit;
8807 struct io_tctx_node *node;
8811 * If we're doing polled IO and end up having requests being
8812 * submitted async (out-of-line), then completions can come in while
8813 * we're waiting for refs to drop. We need to reap these manually,
8814 * as nobody else will be looking for them.
8817 io_uring_try_cancel_requests(ctx, NULL, true);
8819 struct io_sq_data *sqd = ctx->sq_data;
8820 struct task_struct *tsk;
8822 io_sq_thread_park(sqd);
8824 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8825 io_wq_cancel_cb(tsk->io_uring->io_wq,
8826 io_cancel_ctx_cb, ctx, true);
8827 io_sq_thread_unpark(sqd);
8830 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8831 /* there is little hope left, don't run it too often */
8834 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8836 init_completion(&exit.completion);
8837 init_task_work(&exit.task_work, io_tctx_exit_cb);
8840 * Some may use context even when all refs and requests have been put,
8841 * and they are free to do so while still holding uring_lock or
8842 * completion_lock, see io_req_task_submit(). Apart from other work,
8843 * this lock/unlock section also waits them to finish.
8845 mutex_lock(&ctx->uring_lock);
8846 while (!list_empty(&ctx->tctx_list)) {
8847 WARN_ON_ONCE(time_after(jiffies, timeout));
8849 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8851 /* don't spin on a single task if cancellation failed */
8852 list_rotate_left(&ctx->tctx_list);
8853 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8854 if (WARN_ON_ONCE(ret))
8856 wake_up_process(node->task);
8858 mutex_unlock(&ctx->uring_lock);
8859 wait_for_completion(&exit.completion);
8860 mutex_lock(&ctx->uring_lock);
8862 mutex_unlock(&ctx->uring_lock);
8863 spin_lock(&ctx->completion_lock);
8864 spin_unlock(&ctx->completion_lock);
8866 io_ring_ctx_free(ctx);
8869 /* Returns true if we found and killed one or more timeouts */
8870 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8873 struct io_kiocb *req, *tmp;
8876 spin_lock(&ctx->completion_lock);
8877 spin_lock_irq(&ctx->timeout_lock);
8878 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8879 if (io_match_task(req, tsk, cancel_all)) {
8880 io_kill_timeout(req, -ECANCELED);
8884 spin_unlock_irq(&ctx->timeout_lock);
8886 io_commit_cqring(ctx);
8887 spin_unlock(&ctx->completion_lock);
8889 io_cqring_ev_posted(ctx);
8890 return canceled != 0;
8893 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8895 unsigned long index;
8896 struct creds *creds;
8898 mutex_lock(&ctx->uring_lock);
8899 percpu_ref_kill(&ctx->refs);
8901 __io_cqring_overflow_flush(ctx, true);
8902 xa_for_each(&ctx->personalities, index, creds)
8903 io_unregister_personality(ctx, index);
8904 mutex_unlock(&ctx->uring_lock);
8906 io_kill_timeouts(ctx, NULL, true);
8907 io_poll_remove_all(ctx, NULL, true);
8909 /* if we failed setting up the ctx, we might not have any rings */
8910 io_iopoll_try_reap_events(ctx);
8912 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8914 * Use system_unbound_wq to avoid spawning tons of event kworkers
8915 * if we're exiting a ton of rings at the same time. It just adds
8916 * noise and overhead, there's no discernable change in runtime
8917 * over using system_wq.
8919 queue_work(system_unbound_wq, &ctx->exit_work);
8922 static int io_uring_release(struct inode *inode, struct file *file)
8924 struct io_ring_ctx *ctx = file->private_data;
8926 file->private_data = NULL;
8927 io_ring_ctx_wait_and_kill(ctx);
8931 struct io_task_cancel {
8932 struct task_struct *task;
8936 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8938 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8939 struct io_task_cancel *cancel = data;
8942 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8943 struct io_ring_ctx *ctx = req->ctx;
8945 /* protect against races with linked timeouts */
8946 spin_lock(&ctx->completion_lock);
8947 ret = io_match_task(req, cancel->task, cancel->all);
8948 spin_unlock(&ctx->completion_lock);
8950 ret = io_match_task(req, cancel->task, cancel->all);
8955 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8956 struct task_struct *task, bool cancel_all)
8958 struct io_defer_entry *de;
8961 spin_lock(&ctx->completion_lock);
8962 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8963 if (io_match_task(de->req, task, cancel_all)) {
8964 list_cut_position(&list, &ctx->defer_list, &de->list);
8968 spin_unlock(&ctx->completion_lock);
8969 if (list_empty(&list))
8972 while (!list_empty(&list)) {
8973 de = list_first_entry(&list, struct io_defer_entry, list);
8974 list_del_init(&de->list);
8975 io_req_complete_failed(de->req, -ECANCELED);
8981 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8983 struct io_tctx_node *node;
8984 enum io_wq_cancel cret;
8987 mutex_lock(&ctx->uring_lock);
8988 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8989 struct io_uring_task *tctx = node->task->io_uring;
8992 * io_wq will stay alive while we hold uring_lock, because it's
8993 * killed after ctx nodes, which requires to take the lock.
8995 if (!tctx || !tctx->io_wq)
8997 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8998 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9000 mutex_unlock(&ctx->uring_lock);
9005 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9006 struct task_struct *task,
9009 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9010 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9013 enum io_wq_cancel cret;
9017 ret |= io_uring_try_cancel_iowq(ctx);
9018 } else if (tctx && tctx->io_wq) {
9020 * Cancels requests of all rings, not only @ctx, but
9021 * it's fine as the task is in exit/exec.
9023 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9025 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9028 /* SQPOLL thread does its own polling */
9029 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9030 (ctx->sq_data && ctx->sq_data->thread == current)) {
9031 while (!list_empty_careful(&ctx->iopoll_list)) {
9032 io_iopoll_try_reap_events(ctx);
9037 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9038 ret |= io_poll_remove_all(ctx, task, cancel_all);
9039 ret |= io_kill_timeouts(ctx, task, cancel_all);
9041 ret |= io_run_task_work();
9048 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9050 struct io_uring_task *tctx = current->io_uring;
9051 struct io_tctx_node *node;
9054 if (unlikely(!tctx)) {
9055 ret = io_uring_alloc_task_context(current, ctx);
9058 tctx = current->io_uring;
9060 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9061 node = kmalloc(sizeof(*node), GFP_KERNEL);
9065 node->task = current;
9067 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9074 mutex_lock(&ctx->uring_lock);
9075 list_add(&node->ctx_node, &ctx->tctx_list);
9076 mutex_unlock(&ctx->uring_lock);
9083 * Note that this task has used io_uring. We use it for cancelation purposes.
9085 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9087 struct io_uring_task *tctx = current->io_uring;
9089 if (likely(tctx && tctx->last == ctx))
9091 return __io_uring_add_tctx_node(ctx);
9095 * Remove this io_uring_file -> task mapping.
9097 static void io_uring_del_tctx_node(unsigned long index)
9099 struct io_uring_task *tctx = current->io_uring;
9100 struct io_tctx_node *node;
9104 node = xa_erase(&tctx->xa, index);
9108 WARN_ON_ONCE(current != node->task);
9109 WARN_ON_ONCE(list_empty(&node->ctx_node));
9111 mutex_lock(&node->ctx->uring_lock);
9112 list_del(&node->ctx_node);
9113 mutex_unlock(&node->ctx->uring_lock);
9115 if (tctx->last == node->ctx)
9120 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9122 struct io_wq *wq = tctx->io_wq;
9123 struct io_tctx_node *node;
9124 unsigned long index;
9126 xa_for_each(&tctx->xa, index, node)
9127 io_uring_del_tctx_node(index);
9130 * Must be after io_uring_del_task_file() (removes nodes under
9131 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9134 io_wq_put_and_exit(wq);
9138 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9141 return atomic_read(&tctx->inflight_tracked);
9142 return percpu_counter_sum(&tctx->inflight);
9145 static void io_uring_drop_tctx_refs(struct task_struct *task)
9147 struct io_uring_task *tctx = task->io_uring;
9148 unsigned int refs = tctx->cached_refs;
9151 tctx->cached_refs = 0;
9152 percpu_counter_sub(&tctx->inflight, refs);
9153 put_task_struct_many(task, refs);
9158 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9159 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9161 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9163 struct io_uring_task *tctx = current->io_uring;
9164 struct io_ring_ctx *ctx;
9168 WARN_ON_ONCE(sqd && sqd->thread != current);
9170 if (!current->io_uring)
9173 io_wq_exit_start(tctx->io_wq);
9175 atomic_inc(&tctx->in_idle);
9177 io_uring_drop_tctx_refs(current);
9178 /* read completions before cancelations */
9179 inflight = tctx_inflight(tctx, !cancel_all);
9184 struct io_tctx_node *node;
9185 unsigned long index;
9187 xa_for_each(&tctx->xa, index, node) {
9188 /* sqpoll task will cancel all its requests */
9189 if (node->ctx->sq_data)
9191 io_uring_try_cancel_requests(node->ctx, current,
9195 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9196 io_uring_try_cancel_requests(ctx, current,
9200 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9201 io_uring_drop_tctx_refs(current);
9203 * If we've seen completions, retry without waiting. This
9204 * avoids a race where a completion comes in before we did
9205 * prepare_to_wait().
9207 if (inflight == tctx_inflight(tctx, !cancel_all))
9209 finish_wait(&tctx->wait, &wait);
9211 atomic_dec(&tctx->in_idle);
9213 io_uring_clean_tctx(tctx);
9215 /* for exec all current's requests should be gone, kill tctx */
9216 __io_uring_free(current);
9220 void __io_uring_cancel(bool cancel_all)
9222 io_uring_cancel_generic(cancel_all, NULL);
9225 static void *io_uring_validate_mmap_request(struct file *file,
9226 loff_t pgoff, size_t sz)
9228 struct io_ring_ctx *ctx = file->private_data;
9229 loff_t offset = pgoff << PAGE_SHIFT;
9234 case IORING_OFF_SQ_RING:
9235 case IORING_OFF_CQ_RING:
9238 case IORING_OFF_SQES:
9242 return ERR_PTR(-EINVAL);
9245 page = virt_to_head_page(ptr);
9246 if (sz > page_size(page))
9247 return ERR_PTR(-EINVAL);
9254 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9256 size_t sz = vma->vm_end - vma->vm_start;
9260 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9262 return PTR_ERR(ptr);
9264 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9265 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9268 #else /* !CONFIG_MMU */
9270 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9272 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9275 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9277 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9280 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9281 unsigned long addr, unsigned long len,
9282 unsigned long pgoff, unsigned long flags)
9286 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9288 return PTR_ERR(ptr);
9290 return (unsigned long) ptr;
9293 #endif /* !CONFIG_MMU */
9295 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9300 if (!io_sqring_full(ctx))
9302 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9304 if (!io_sqring_full(ctx))
9307 } while (!signal_pending(current));
9309 finish_wait(&ctx->sqo_sq_wait, &wait);
9313 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9314 struct __kernel_timespec __user **ts,
9315 const sigset_t __user **sig)
9317 struct io_uring_getevents_arg arg;
9320 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9321 * is just a pointer to the sigset_t.
9323 if (!(flags & IORING_ENTER_EXT_ARG)) {
9324 *sig = (const sigset_t __user *) argp;
9330 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9331 * timespec and sigset_t pointers if good.
9333 if (*argsz != sizeof(arg))
9335 if (copy_from_user(&arg, argp, sizeof(arg)))
9337 *sig = u64_to_user_ptr(arg.sigmask);
9338 *argsz = arg.sigmask_sz;
9339 *ts = u64_to_user_ptr(arg.ts);
9343 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9344 u32, min_complete, u32, flags, const void __user *, argp,
9347 struct io_ring_ctx *ctx;
9354 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9355 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9359 if (unlikely(!f.file))
9363 if (unlikely(f.file->f_op != &io_uring_fops))
9367 ctx = f.file->private_data;
9368 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9372 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9376 * For SQ polling, the thread will do all submissions and completions.
9377 * Just return the requested submit count, and wake the thread if
9381 if (ctx->flags & IORING_SETUP_SQPOLL) {
9382 io_cqring_overflow_flush(ctx);
9384 if (unlikely(ctx->sq_data->thread == NULL)) {
9388 if (flags & IORING_ENTER_SQ_WAKEUP)
9389 wake_up(&ctx->sq_data->wait);
9390 if (flags & IORING_ENTER_SQ_WAIT) {
9391 ret = io_sqpoll_wait_sq(ctx);
9395 submitted = to_submit;
9396 } else if (to_submit) {
9397 ret = io_uring_add_tctx_node(ctx);
9400 mutex_lock(&ctx->uring_lock);
9401 submitted = io_submit_sqes(ctx, to_submit);
9402 mutex_unlock(&ctx->uring_lock);
9404 if (submitted != to_submit)
9407 if (flags & IORING_ENTER_GETEVENTS) {
9408 const sigset_t __user *sig;
9409 struct __kernel_timespec __user *ts;
9411 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9415 min_complete = min(min_complete, ctx->cq_entries);
9418 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9419 * space applications don't need to do io completion events
9420 * polling again, they can rely on io_sq_thread to do polling
9421 * work, which can reduce cpu usage and uring_lock contention.
9423 if (ctx->flags & IORING_SETUP_IOPOLL &&
9424 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9425 ret = io_iopoll_check(ctx, min_complete);
9427 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9432 percpu_ref_put(&ctx->refs);
9435 return submitted ? submitted : ret;
9438 #ifdef CONFIG_PROC_FS
9439 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9440 const struct cred *cred)
9442 struct user_namespace *uns = seq_user_ns(m);
9443 struct group_info *gi;
9448 seq_printf(m, "%5d\n", id);
9449 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9450 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9451 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9452 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9453 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9454 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9455 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9456 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9457 seq_puts(m, "\n\tGroups:\t");
9458 gi = cred->group_info;
9459 for (g = 0; g < gi->ngroups; g++) {
9460 seq_put_decimal_ull(m, g ? " " : "",
9461 from_kgid_munged(uns, gi->gid[g]));
9463 seq_puts(m, "\n\tCapEff:\t");
9464 cap = cred->cap_effective;
9465 CAP_FOR_EACH_U32(__capi)
9466 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9471 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9473 struct io_sq_data *sq = NULL;
9478 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9479 * since fdinfo case grabs it in the opposite direction of normal use
9480 * cases. If we fail to get the lock, we just don't iterate any
9481 * structures that could be going away outside the io_uring mutex.
9483 has_lock = mutex_trylock(&ctx->uring_lock);
9485 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9491 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9492 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9493 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9494 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9495 struct file *f = io_file_from_index(ctx, i);
9498 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9500 seq_printf(m, "%5u: <none>\n", i);
9502 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9503 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9504 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9505 unsigned int len = buf->ubuf_end - buf->ubuf;
9507 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9509 if (has_lock && !xa_empty(&ctx->personalities)) {
9510 unsigned long index;
9511 const struct cred *cred;
9513 seq_printf(m, "Personalities:\n");
9514 xa_for_each(&ctx->personalities, index, cred)
9515 io_uring_show_cred(m, index, cred);
9517 seq_printf(m, "PollList:\n");
9518 spin_lock(&ctx->completion_lock);
9519 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9520 struct hlist_head *list = &ctx->cancel_hash[i];
9521 struct io_kiocb *req;
9523 hlist_for_each_entry(req, list, hash_node)
9524 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9525 req->task->task_works != NULL);
9527 spin_unlock(&ctx->completion_lock);
9529 mutex_unlock(&ctx->uring_lock);
9532 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9534 struct io_ring_ctx *ctx = f->private_data;
9536 if (percpu_ref_tryget(&ctx->refs)) {
9537 __io_uring_show_fdinfo(ctx, m);
9538 percpu_ref_put(&ctx->refs);
9543 static const struct file_operations io_uring_fops = {
9544 .release = io_uring_release,
9545 .mmap = io_uring_mmap,
9547 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9548 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9550 .poll = io_uring_poll,
9551 .fasync = io_uring_fasync,
9552 #ifdef CONFIG_PROC_FS
9553 .show_fdinfo = io_uring_show_fdinfo,
9557 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9558 struct io_uring_params *p)
9560 struct io_rings *rings;
9561 size_t size, sq_array_offset;
9563 /* make sure these are sane, as we already accounted them */
9564 ctx->sq_entries = p->sq_entries;
9565 ctx->cq_entries = p->cq_entries;
9567 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9568 if (size == SIZE_MAX)
9571 rings = io_mem_alloc(size);
9576 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9577 rings->sq_ring_mask = p->sq_entries - 1;
9578 rings->cq_ring_mask = p->cq_entries - 1;
9579 rings->sq_ring_entries = p->sq_entries;
9580 rings->cq_ring_entries = p->cq_entries;
9582 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9583 if (size == SIZE_MAX) {
9584 io_mem_free(ctx->rings);
9589 ctx->sq_sqes = io_mem_alloc(size);
9590 if (!ctx->sq_sqes) {
9591 io_mem_free(ctx->rings);
9599 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9603 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9607 ret = io_uring_add_tctx_node(ctx);
9612 fd_install(fd, file);
9617 * Allocate an anonymous fd, this is what constitutes the application
9618 * visible backing of an io_uring instance. The application mmaps this
9619 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9620 * we have to tie this fd to a socket for file garbage collection purposes.
9622 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9625 #if defined(CONFIG_UNIX)
9628 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9631 return ERR_PTR(ret);
9634 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9635 O_RDWR | O_CLOEXEC);
9636 #if defined(CONFIG_UNIX)
9638 sock_release(ctx->ring_sock);
9639 ctx->ring_sock = NULL;
9641 ctx->ring_sock->file = file;
9647 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9648 struct io_uring_params __user *params)
9650 struct io_ring_ctx *ctx;
9656 if (entries > IORING_MAX_ENTRIES) {
9657 if (!(p->flags & IORING_SETUP_CLAMP))
9659 entries = IORING_MAX_ENTRIES;
9663 * Use twice as many entries for the CQ ring. It's possible for the
9664 * application to drive a higher depth than the size of the SQ ring,
9665 * since the sqes are only used at submission time. This allows for
9666 * some flexibility in overcommitting a bit. If the application has
9667 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9668 * of CQ ring entries manually.
9670 p->sq_entries = roundup_pow_of_two(entries);
9671 if (p->flags & IORING_SETUP_CQSIZE) {
9673 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9674 * to a power-of-two, if it isn't already. We do NOT impose
9675 * any cq vs sq ring sizing.
9679 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9680 if (!(p->flags & IORING_SETUP_CLAMP))
9682 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9684 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9685 if (p->cq_entries < p->sq_entries)
9688 p->cq_entries = 2 * p->sq_entries;
9691 ctx = io_ring_ctx_alloc(p);
9694 ctx->compat = in_compat_syscall();
9695 if (!capable(CAP_IPC_LOCK))
9696 ctx->user = get_uid(current_user());
9699 * This is just grabbed for accounting purposes. When a process exits,
9700 * the mm is exited and dropped before the files, hence we need to hang
9701 * on to this mm purely for the purposes of being able to unaccount
9702 * memory (locked/pinned vm). It's not used for anything else.
9704 mmgrab(current->mm);
9705 ctx->mm_account = current->mm;
9707 ret = io_allocate_scq_urings(ctx, p);
9711 ret = io_sq_offload_create(ctx, p);
9714 /* always set a rsrc node */
9715 ret = io_rsrc_node_switch_start(ctx);
9718 io_rsrc_node_switch(ctx, NULL);
9720 memset(&p->sq_off, 0, sizeof(p->sq_off));
9721 p->sq_off.head = offsetof(struct io_rings, sq.head);
9722 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9723 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9724 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9725 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9726 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9727 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9729 memset(&p->cq_off, 0, sizeof(p->cq_off));
9730 p->cq_off.head = offsetof(struct io_rings, cq.head);
9731 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9732 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9733 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9734 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9735 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9736 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9738 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9739 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9740 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9741 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9742 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9743 IORING_FEAT_RSRC_TAGS;
9745 if (copy_to_user(params, p, sizeof(*p))) {
9750 file = io_uring_get_file(ctx);
9752 ret = PTR_ERR(file);
9757 * Install ring fd as the very last thing, so we don't risk someone
9758 * having closed it before we finish setup
9760 ret = io_uring_install_fd(ctx, file);
9762 /* fput will clean it up */
9767 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9770 io_ring_ctx_wait_and_kill(ctx);
9775 * Sets up an aio uring context, and returns the fd. Applications asks for a
9776 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9777 * params structure passed in.
9779 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9781 struct io_uring_params p;
9784 if (copy_from_user(&p, params, sizeof(p)))
9786 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9791 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9792 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9793 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9794 IORING_SETUP_R_DISABLED))
9797 return io_uring_create(entries, &p, params);
9800 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9801 struct io_uring_params __user *, params)
9803 return io_uring_setup(entries, params);
9806 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9808 struct io_uring_probe *p;
9812 size = struct_size(p, ops, nr_args);
9813 if (size == SIZE_MAX)
9815 p = kzalloc(size, GFP_KERNEL);
9820 if (copy_from_user(p, arg, size))
9823 if (memchr_inv(p, 0, size))
9826 p->last_op = IORING_OP_LAST - 1;
9827 if (nr_args > IORING_OP_LAST)
9828 nr_args = IORING_OP_LAST;
9830 for (i = 0; i < nr_args; i++) {
9832 if (!io_op_defs[i].not_supported)
9833 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9838 if (copy_to_user(arg, p, size))
9845 static int io_register_personality(struct io_ring_ctx *ctx)
9847 const struct cred *creds;
9851 creds = get_current_cred();
9853 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9854 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9862 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9863 unsigned int nr_args)
9865 struct io_uring_restriction *res;
9869 /* Restrictions allowed only if rings started disabled */
9870 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9873 /* We allow only a single restrictions registration */
9874 if (ctx->restrictions.registered)
9877 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9880 size = array_size(nr_args, sizeof(*res));
9881 if (size == SIZE_MAX)
9884 res = memdup_user(arg, size);
9886 return PTR_ERR(res);
9890 for (i = 0; i < nr_args; i++) {
9891 switch (res[i].opcode) {
9892 case IORING_RESTRICTION_REGISTER_OP:
9893 if (res[i].register_op >= IORING_REGISTER_LAST) {
9898 __set_bit(res[i].register_op,
9899 ctx->restrictions.register_op);
9901 case IORING_RESTRICTION_SQE_OP:
9902 if (res[i].sqe_op >= IORING_OP_LAST) {
9907 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9909 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9910 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9912 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9913 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9922 /* Reset all restrictions if an error happened */
9924 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9926 ctx->restrictions.registered = true;
9932 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9934 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9937 if (ctx->restrictions.registered)
9938 ctx->restricted = 1;
9940 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9941 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9942 wake_up(&ctx->sq_data->wait);
9946 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9947 struct io_uring_rsrc_update2 *up,
9955 if (check_add_overflow(up->offset, nr_args, &tmp))
9957 err = io_rsrc_node_switch_start(ctx);
9962 case IORING_RSRC_FILE:
9963 return __io_sqe_files_update(ctx, up, nr_args);
9964 case IORING_RSRC_BUFFER:
9965 return __io_sqe_buffers_update(ctx, up, nr_args);
9970 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9973 struct io_uring_rsrc_update2 up;
9977 memset(&up, 0, sizeof(up));
9978 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9980 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9983 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9984 unsigned size, unsigned type)
9986 struct io_uring_rsrc_update2 up;
9988 if (size != sizeof(up))
9990 if (copy_from_user(&up, arg, sizeof(up)))
9992 if (!up.nr || up.resv)
9994 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9997 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9998 unsigned int size, unsigned int type)
10000 struct io_uring_rsrc_register rr;
10002 /* keep it extendible */
10003 if (size != sizeof(rr))
10006 memset(&rr, 0, sizeof(rr));
10007 if (copy_from_user(&rr, arg, size))
10009 if (!rr.nr || rr.resv || rr.resv2)
10013 case IORING_RSRC_FILE:
10014 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10015 rr.nr, u64_to_user_ptr(rr.tags));
10016 case IORING_RSRC_BUFFER:
10017 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10018 rr.nr, u64_to_user_ptr(rr.tags));
10023 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10026 struct io_uring_task *tctx = current->io_uring;
10027 cpumask_var_t new_mask;
10030 if (!tctx || !tctx->io_wq)
10033 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10036 cpumask_clear(new_mask);
10037 if (len > cpumask_size())
10038 len = cpumask_size();
10040 if (copy_from_user(new_mask, arg, len)) {
10041 free_cpumask_var(new_mask);
10045 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10046 free_cpumask_var(new_mask);
10050 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10052 struct io_uring_task *tctx = current->io_uring;
10054 if (!tctx || !tctx->io_wq)
10057 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10060 static bool io_register_op_must_quiesce(int op)
10063 case IORING_REGISTER_BUFFERS:
10064 case IORING_UNREGISTER_BUFFERS:
10065 case IORING_REGISTER_FILES:
10066 case IORING_UNREGISTER_FILES:
10067 case IORING_REGISTER_FILES_UPDATE:
10068 case IORING_REGISTER_PROBE:
10069 case IORING_REGISTER_PERSONALITY:
10070 case IORING_UNREGISTER_PERSONALITY:
10071 case IORING_REGISTER_FILES2:
10072 case IORING_REGISTER_FILES_UPDATE2:
10073 case IORING_REGISTER_BUFFERS2:
10074 case IORING_REGISTER_BUFFERS_UPDATE:
10075 case IORING_REGISTER_IOWQ_AFF:
10076 case IORING_UNREGISTER_IOWQ_AFF:
10083 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10087 percpu_ref_kill(&ctx->refs);
10090 * Drop uring mutex before waiting for references to exit. If another
10091 * thread is currently inside io_uring_enter() it might need to grab the
10092 * uring_lock to make progress. If we hold it here across the drain
10093 * wait, then we can deadlock. It's safe to drop the mutex here, since
10094 * no new references will come in after we've killed the percpu ref.
10096 mutex_unlock(&ctx->uring_lock);
10098 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10101 ret = io_run_task_work_sig();
10102 } while (ret >= 0);
10103 mutex_lock(&ctx->uring_lock);
10106 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10110 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10111 void __user *arg, unsigned nr_args)
10112 __releases(ctx->uring_lock)
10113 __acquires(ctx->uring_lock)
10118 * We're inside the ring mutex, if the ref is already dying, then
10119 * someone else killed the ctx or is already going through
10120 * io_uring_register().
10122 if (percpu_ref_is_dying(&ctx->refs))
10125 if (ctx->restricted) {
10126 if (opcode >= IORING_REGISTER_LAST)
10128 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10129 if (!test_bit(opcode, ctx->restrictions.register_op))
10133 if (io_register_op_must_quiesce(opcode)) {
10134 ret = io_ctx_quiesce(ctx);
10140 case IORING_REGISTER_BUFFERS:
10141 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10143 case IORING_UNREGISTER_BUFFERS:
10145 if (arg || nr_args)
10147 ret = io_sqe_buffers_unregister(ctx);
10149 case IORING_REGISTER_FILES:
10150 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10152 case IORING_UNREGISTER_FILES:
10154 if (arg || nr_args)
10156 ret = io_sqe_files_unregister(ctx);
10158 case IORING_REGISTER_FILES_UPDATE:
10159 ret = io_register_files_update(ctx, arg, nr_args);
10161 case IORING_REGISTER_EVENTFD:
10162 case IORING_REGISTER_EVENTFD_ASYNC:
10166 ret = io_eventfd_register(ctx, arg);
10169 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10170 ctx->eventfd_async = 1;
10172 ctx->eventfd_async = 0;
10174 case IORING_UNREGISTER_EVENTFD:
10176 if (arg || nr_args)
10178 ret = io_eventfd_unregister(ctx);
10180 case IORING_REGISTER_PROBE:
10182 if (!arg || nr_args > 256)
10184 ret = io_probe(ctx, arg, nr_args);
10186 case IORING_REGISTER_PERSONALITY:
10188 if (arg || nr_args)
10190 ret = io_register_personality(ctx);
10192 case IORING_UNREGISTER_PERSONALITY:
10196 ret = io_unregister_personality(ctx, nr_args);
10198 case IORING_REGISTER_ENABLE_RINGS:
10200 if (arg || nr_args)
10202 ret = io_register_enable_rings(ctx);
10204 case IORING_REGISTER_RESTRICTIONS:
10205 ret = io_register_restrictions(ctx, arg, nr_args);
10207 case IORING_REGISTER_FILES2:
10208 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10210 case IORING_REGISTER_FILES_UPDATE2:
10211 ret = io_register_rsrc_update(ctx, arg, nr_args,
10214 case IORING_REGISTER_BUFFERS2:
10215 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10217 case IORING_REGISTER_BUFFERS_UPDATE:
10218 ret = io_register_rsrc_update(ctx, arg, nr_args,
10219 IORING_RSRC_BUFFER);
10221 case IORING_REGISTER_IOWQ_AFF:
10223 if (!arg || !nr_args)
10225 ret = io_register_iowq_aff(ctx, arg, nr_args);
10227 case IORING_UNREGISTER_IOWQ_AFF:
10229 if (arg || nr_args)
10231 ret = io_unregister_iowq_aff(ctx);
10238 if (io_register_op_must_quiesce(opcode)) {
10239 /* bring the ctx back to life */
10240 percpu_ref_reinit(&ctx->refs);
10241 reinit_completion(&ctx->ref_comp);
10246 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10247 void __user *, arg, unsigned int, nr_args)
10249 struct io_ring_ctx *ctx;
10258 if (f.file->f_op != &io_uring_fops)
10261 ctx = f.file->private_data;
10263 io_run_task_work();
10265 mutex_lock(&ctx->uring_lock);
10266 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10267 mutex_unlock(&ctx->uring_lock);
10268 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10269 ctx->cq_ev_fd != NULL, ret);
10275 static int __init io_uring_init(void)
10277 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10278 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10279 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10282 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10283 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10284 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10285 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10286 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10287 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10288 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10289 BUILD_BUG_SQE_ELEM(8, __u64, off);
10290 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10291 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10292 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10293 BUILD_BUG_SQE_ELEM(24, __u32, len);
10294 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10295 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10296 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10298 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10299 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10300 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10301 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10302 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10303 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10304 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10305 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10306 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10307 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10308 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10309 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10310 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10311 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10312 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10313 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10315 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10316 sizeof(struct io_uring_rsrc_update));
10317 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10318 sizeof(struct io_uring_rsrc_update2));
10319 /* should fit into one byte */
10320 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10322 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10323 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10325 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10329 __initcall(io_uring_init);