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_refcount(struct io_kiocb *req)
1120 if (!(req->flags & REQ_F_REFCOUNT)) {
1121 req->flags |= REQ_F_REFCOUNT;
1122 atomic_set(&req->refs, 1);
1126 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1128 struct io_ring_ctx *ctx = req->ctx;
1130 if (!req->fixed_rsrc_refs) {
1131 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1132 percpu_ref_get(req->fixed_rsrc_refs);
1136 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1138 bool got = percpu_ref_tryget(ref);
1140 /* already at zero, wait for ->release() */
1142 wait_for_completion(compl);
1143 percpu_ref_resurrect(ref);
1145 percpu_ref_put(ref);
1148 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1151 struct io_kiocb *req;
1153 if (task && head->task != task)
1158 io_for_each_link(req, head) {
1159 if (req->flags & REQ_F_INFLIGHT)
1165 static inline void req_set_fail(struct io_kiocb *req)
1167 req->flags |= REQ_F_FAIL;
1170 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1172 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1174 complete(&ctx->ref_comp);
1177 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1179 return !req->timeout.off;
1182 static void io_fallback_req_func(struct work_struct *work)
1184 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1185 fallback_work.work);
1186 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1187 struct io_kiocb *req, *tmp;
1189 percpu_ref_get(&ctx->refs);
1190 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1191 req->io_task_work.func(req);
1192 percpu_ref_put(&ctx->refs);
1195 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1197 struct io_ring_ctx *ctx;
1200 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1205 * Use 5 bits less than the max cq entries, that should give us around
1206 * 32 entries per hash list if totally full and uniformly spread.
1208 hash_bits = ilog2(p->cq_entries);
1212 ctx->cancel_hash_bits = hash_bits;
1213 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1215 if (!ctx->cancel_hash)
1217 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1219 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1220 if (!ctx->dummy_ubuf)
1222 /* set invalid range, so io_import_fixed() fails meeting it */
1223 ctx->dummy_ubuf->ubuf = -1UL;
1225 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1226 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1229 ctx->flags = p->flags;
1230 init_waitqueue_head(&ctx->sqo_sq_wait);
1231 INIT_LIST_HEAD(&ctx->sqd_list);
1232 init_waitqueue_head(&ctx->poll_wait);
1233 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1234 init_completion(&ctx->ref_comp);
1235 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1236 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1237 mutex_init(&ctx->uring_lock);
1238 init_waitqueue_head(&ctx->cq_wait);
1239 spin_lock_init(&ctx->completion_lock);
1240 spin_lock_init(&ctx->timeout_lock);
1241 INIT_LIST_HEAD(&ctx->iopoll_list);
1242 INIT_LIST_HEAD(&ctx->defer_list);
1243 INIT_LIST_HEAD(&ctx->timeout_list);
1244 spin_lock_init(&ctx->rsrc_ref_lock);
1245 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1246 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1247 init_llist_head(&ctx->rsrc_put_llist);
1248 INIT_LIST_HEAD(&ctx->tctx_list);
1249 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1250 INIT_LIST_HEAD(&ctx->locked_free_list);
1251 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1254 kfree(ctx->dummy_ubuf);
1255 kfree(ctx->cancel_hash);
1260 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1262 struct io_rings *r = ctx->rings;
1264 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1268 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1270 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1271 struct io_ring_ctx *ctx = req->ctx;
1273 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1279 #define FFS_ASYNC_READ 0x1UL
1280 #define FFS_ASYNC_WRITE 0x2UL
1282 #define FFS_ISREG 0x4UL
1284 #define FFS_ISREG 0x0UL
1286 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1288 static inline bool io_req_ffs_set(struct io_kiocb *req)
1290 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1293 static void io_req_track_inflight(struct io_kiocb *req)
1295 if (!(req->flags & REQ_F_INFLIGHT)) {
1296 req->flags |= REQ_F_INFLIGHT;
1297 atomic_inc(¤t->io_uring->inflight_tracked);
1301 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1303 struct io_kiocb *nxt = req->link;
1305 if (req->flags & REQ_F_LINK_TIMEOUT)
1308 /* linked timeouts should have two refs once prep'ed */
1309 io_req_refcount(req);
1310 io_req_refcount(nxt);
1313 nxt->timeout.head = req;
1314 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
1315 req->flags |= REQ_F_LINK_TIMEOUT;
1319 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1321 if (likely(!req->link || req->link->opcode != IORING_OP_LINK_TIMEOUT))
1323 return __io_prep_linked_timeout(req);
1326 static void io_prep_async_work(struct io_kiocb *req)
1328 const struct io_op_def *def = &io_op_defs[req->opcode];
1329 struct io_ring_ctx *ctx = req->ctx;
1331 if (!(req->flags & REQ_F_CREDS)) {
1332 req->flags |= REQ_F_CREDS;
1333 req->creds = get_current_cred();
1336 req->work.list.next = NULL;
1337 req->work.flags = 0;
1338 if (req->flags & REQ_F_FORCE_ASYNC)
1339 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1341 if (req->flags & REQ_F_ISREG) {
1342 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1343 io_wq_hash_work(&req->work, file_inode(req->file));
1344 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1345 if (def->unbound_nonreg_file)
1346 req->work.flags |= IO_WQ_WORK_UNBOUND;
1349 switch (req->opcode) {
1350 case IORING_OP_SPLICE:
1352 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1353 req->work.flags |= IO_WQ_WORK_UNBOUND;
1358 static void io_prep_async_link(struct io_kiocb *req)
1360 struct io_kiocb *cur;
1362 if (req->flags & REQ_F_LINK_TIMEOUT) {
1363 struct io_ring_ctx *ctx = req->ctx;
1365 spin_lock(&ctx->completion_lock);
1366 io_for_each_link(cur, req)
1367 io_prep_async_work(cur);
1368 spin_unlock(&ctx->completion_lock);
1370 io_for_each_link(cur, req)
1371 io_prep_async_work(cur);
1375 static void io_queue_async_work(struct io_kiocb *req)
1377 struct io_ring_ctx *ctx = req->ctx;
1378 struct io_kiocb *link = io_prep_linked_timeout(req);
1379 struct io_uring_task *tctx = req->task->io_uring;
1382 BUG_ON(!tctx->io_wq);
1384 /* init ->work of the whole link before punting */
1385 io_prep_async_link(req);
1388 * Not expected to happen, but if we do have a bug where this _can_
1389 * happen, catch it here and ensure the request is marked as
1390 * canceled. That will make io-wq go through the usual work cancel
1391 * procedure rather than attempt to run this request (or create a new
1394 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1395 req->work.flags |= IO_WQ_WORK_CANCEL;
1397 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1398 &req->work, req->flags);
1399 io_wq_enqueue(tctx->io_wq, &req->work);
1401 io_queue_linked_timeout(link);
1404 static void io_kill_timeout(struct io_kiocb *req, int status)
1405 __must_hold(&req->ctx->completion_lock)
1406 __must_hold(&req->ctx->timeout_lock)
1408 struct io_timeout_data *io = req->async_data;
1410 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1411 atomic_set(&req->ctx->cq_timeouts,
1412 atomic_read(&req->ctx->cq_timeouts) + 1);
1413 list_del_init(&req->timeout.list);
1414 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1415 io_put_req_deferred(req);
1419 static void io_queue_deferred(struct io_ring_ctx *ctx)
1421 while (!list_empty(&ctx->defer_list)) {
1422 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1423 struct io_defer_entry, list);
1425 if (req_need_defer(de->req, de->seq))
1427 list_del_init(&de->list);
1428 io_req_task_queue(de->req);
1433 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1434 __must_hold(&ctx->completion_lock)
1436 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1438 spin_lock_irq(&ctx->timeout_lock);
1439 while (!list_empty(&ctx->timeout_list)) {
1440 u32 events_needed, events_got;
1441 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1442 struct io_kiocb, timeout.list);
1444 if (io_is_timeout_noseq(req))
1448 * Since seq can easily wrap around over time, subtract
1449 * the last seq at which timeouts were flushed before comparing.
1450 * Assuming not more than 2^31-1 events have happened since,
1451 * these subtractions won't have wrapped, so we can check if
1452 * target is in [last_seq, current_seq] by comparing the two.
1454 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1455 events_got = seq - ctx->cq_last_tm_flush;
1456 if (events_got < events_needed)
1459 list_del_init(&req->timeout.list);
1460 io_kill_timeout(req, 0);
1462 ctx->cq_last_tm_flush = seq;
1463 spin_unlock_irq(&ctx->timeout_lock);
1466 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1468 if (ctx->off_timeout_used)
1469 io_flush_timeouts(ctx);
1470 if (ctx->drain_active)
1471 io_queue_deferred(ctx);
1474 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1476 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1477 __io_commit_cqring_flush(ctx);
1478 /* order cqe stores with ring update */
1479 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1482 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1484 struct io_rings *r = ctx->rings;
1486 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1489 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1491 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1494 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1496 struct io_rings *rings = ctx->rings;
1497 unsigned tail, mask = ctx->cq_entries - 1;
1500 * writes to the cq entry need to come after reading head; the
1501 * control dependency is enough as we're using WRITE_ONCE to
1504 if (__io_cqring_events(ctx) == ctx->cq_entries)
1507 tail = ctx->cached_cq_tail++;
1508 return &rings->cqes[tail & mask];
1511 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1513 if (likely(!ctx->cq_ev_fd))
1515 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1517 return !ctx->eventfd_async || io_wq_current_is_worker();
1520 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1523 * wake_up_all() may seem excessive, but io_wake_function() and
1524 * io_should_wake() handle the termination of the loop and only
1525 * wake as many waiters as we need to.
1527 if (wq_has_sleeper(&ctx->cq_wait))
1528 wake_up_all(&ctx->cq_wait);
1529 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1530 wake_up(&ctx->sq_data->wait);
1531 if (io_should_trigger_evfd(ctx))
1532 eventfd_signal(ctx->cq_ev_fd, 1);
1533 if (waitqueue_active(&ctx->poll_wait)) {
1534 wake_up_interruptible(&ctx->poll_wait);
1535 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1539 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1541 if (ctx->flags & IORING_SETUP_SQPOLL) {
1542 if (wq_has_sleeper(&ctx->cq_wait))
1543 wake_up_all(&ctx->cq_wait);
1545 if (io_should_trigger_evfd(ctx))
1546 eventfd_signal(ctx->cq_ev_fd, 1);
1547 if (waitqueue_active(&ctx->poll_wait)) {
1548 wake_up_interruptible(&ctx->poll_wait);
1549 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1553 /* Returns true if there are no backlogged entries after the flush */
1554 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1556 bool all_flushed, posted;
1558 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1562 spin_lock(&ctx->completion_lock);
1563 while (!list_empty(&ctx->cq_overflow_list)) {
1564 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1565 struct io_overflow_cqe *ocqe;
1569 ocqe = list_first_entry(&ctx->cq_overflow_list,
1570 struct io_overflow_cqe, list);
1572 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1574 io_account_cq_overflow(ctx);
1577 list_del(&ocqe->list);
1581 all_flushed = list_empty(&ctx->cq_overflow_list);
1583 clear_bit(0, &ctx->check_cq_overflow);
1584 WRITE_ONCE(ctx->rings->sq_flags,
1585 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1589 io_commit_cqring(ctx);
1590 spin_unlock(&ctx->completion_lock);
1592 io_cqring_ev_posted(ctx);
1596 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1600 if (test_bit(0, &ctx->check_cq_overflow)) {
1601 /* iopoll syncs against uring_lock, not completion_lock */
1602 if (ctx->flags & IORING_SETUP_IOPOLL)
1603 mutex_lock(&ctx->uring_lock);
1604 ret = __io_cqring_overflow_flush(ctx, false);
1605 if (ctx->flags & IORING_SETUP_IOPOLL)
1606 mutex_unlock(&ctx->uring_lock);
1612 /* must to be called somewhat shortly after putting a request */
1613 static inline void io_put_task(struct task_struct *task, int nr)
1615 struct io_uring_task *tctx = task->io_uring;
1617 percpu_counter_sub(&tctx->inflight, nr);
1618 if (unlikely(atomic_read(&tctx->in_idle)))
1619 wake_up(&tctx->wait);
1620 put_task_struct_many(task, nr);
1623 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1624 long res, unsigned int cflags)
1626 struct io_overflow_cqe *ocqe;
1628 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1631 * If we're in ring overflow flush mode, or in task cancel mode,
1632 * or cannot allocate an overflow entry, then we need to drop it
1635 io_account_cq_overflow(ctx);
1638 if (list_empty(&ctx->cq_overflow_list)) {
1639 set_bit(0, &ctx->check_cq_overflow);
1640 WRITE_ONCE(ctx->rings->sq_flags,
1641 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1644 ocqe->cqe.user_data = user_data;
1645 ocqe->cqe.res = res;
1646 ocqe->cqe.flags = cflags;
1647 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1651 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1652 long res, unsigned int cflags)
1654 struct io_uring_cqe *cqe;
1656 trace_io_uring_complete(ctx, user_data, res, cflags);
1659 * If we can't get a cq entry, userspace overflowed the
1660 * submission (by quite a lot). Increment the overflow count in
1663 cqe = io_get_cqe(ctx);
1665 WRITE_ONCE(cqe->user_data, user_data);
1666 WRITE_ONCE(cqe->res, res);
1667 WRITE_ONCE(cqe->flags, cflags);
1670 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1673 /* not as hot to bloat with inlining */
1674 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1675 long res, unsigned int cflags)
1677 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1680 static void io_req_complete_post(struct io_kiocb *req, long res,
1681 unsigned int cflags)
1683 struct io_ring_ctx *ctx = req->ctx;
1685 spin_lock(&ctx->completion_lock);
1686 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1688 * If we're the last reference to this request, add to our locked
1691 if (req_ref_put_and_test(req)) {
1692 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1693 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1694 io_disarm_next(req);
1696 io_req_task_queue(req->link);
1700 io_dismantle_req(req);
1701 io_put_task(req->task, 1);
1702 list_add(&req->inflight_entry, &ctx->locked_free_list);
1703 ctx->locked_free_nr++;
1705 if (!percpu_ref_tryget(&ctx->refs))
1708 io_commit_cqring(ctx);
1709 spin_unlock(&ctx->completion_lock);
1712 io_cqring_ev_posted(ctx);
1713 percpu_ref_put(&ctx->refs);
1717 static inline bool io_req_needs_clean(struct io_kiocb *req)
1719 return req->flags & IO_REQ_CLEAN_FLAGS;
1722 static void io_req_complete_state(struct io_kiocb *req, long res,
1723 unsigned int cflags)
1725 if (io_req_needs_clean(req))
1728 req->compl.cflags = cflags;
1729 req->flags |= REQ_F_COMPLETE_INLINE;
1732 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1733 long res, unsigned cflags)
1735 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1736 io_req_complete_state(req, res, cflags);
1738 io_req_complete_post(req, res, cflags);
1741 static inline void io_req_complete(struct io_kiocb *req, long res)
1743 __io_req_complete(req, 0, res, 0);
1746 static void io_req_complete_failed(struct io_kiocb *req, long res)
1749 io_req_complete_post(req, res, 0);
1753 * Don't initialise the fields below on every allocation, but do that in
1754 * advance and keep them valid across allocations.
1756 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1760 req->async_data = NULL;
1761 /* not necessary, but safer to zero */
1765 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1766 struct io_submit_state *state)
1768 spin_lock(&ctx->completion_lock);
1769 list_splice_init(&ctx->locked_free_list, &state->free_list);
1770 ctx->locked_free_nr = 0;
1771 spin_unlock(&ctx->completion_lock);
1774 /* Returns true IFF there are requests in the cache */
1775 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1777 struct io_submit_state *state = &ctx->submit_state;
1781 * If we have more than a batch's worth of requests in our IRQ side
1782 * locked cache, grab the lock and move them over to our submission
1785 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1786 io_flush_cached_locked_reqs(ctx, state);
1788 nr = state->free_reqs;
1789 while (!list_empty(&state->free_list)) {
1790 struct io_kiocb *req = list_first_entry(&state->free_list,
1791 struct io_kiocb, inflight_entry);
1793 list_del(&req->inflight_entry);
1794 state->reqs[nr++] = req;
1795 if (nr == ARRAY_SIZE(state->reqs))
1799 state->free_reqs = nr;
1804 * A request might get retired back into the request caches even before opcode
1805 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1806 * Because of that, io_alloc_req() should be called only under ->uring_lock
1807 * and with extra caution to not get a request that is still worked on.
1809 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1810 __must_hold(&ctx->uring_lock)
1812 struct io_submit_state *state = &ctx->submit_state;
1813 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1816 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1818 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1821 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1825 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1826 * retry single alloc to be on the safe side.
1828 if (unlikely(ret <= 0)) {
1829 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1830 if (!state->reqs[0])
1835 for (i = 0; i < ret; i++)
1836 io_preinit_req(state->reqs[i], ctx);
1837 state->free_reqs = ret;
1840 return state->reqs[state->free_reqs];
1843 static inline void io_put_file(struct file *file)
1849 static void io_dismantle_req(struct io_kiocb *req)
1851 unsigned int flags = req->flags;
1853 if (io_req_needs_clean(req))
1855 if (!(flags & REQ_F_FIXED_FILE))
1856 io_put_file(req->file);
1857 if (req->fixed_rsrc_refs)
1858 percpu_ref_put(req->fixed_rsrc_refs);
1859 if (req->async_data) {
1860 kfree(req->async_data);
1861 req->async_data = NULL;
1865 static void __io_free_req(struct io_kiocb *req)
1867 struct io_ring_ctx *ctx = req->ctx;
1869 io_dismantle_req(req);
1870 io_put_task(req->task, 1);
1872 spin_lock(&ctx->completion_lock);
1873 list_add(&req->inflight_entry, &ctx->locked_free_list);
1874 ctx->locked_free_nr++;
1875 spin_unlock(&ctx->completion_lock);
1877 percpu_ref_put(&ctx->refs);
1880 static inline void io_remove_next_linked(struct io_kiocb *req)
1882 struct io_kiocb *nxt = req->link;
1884 req->link = nxt->link;
1888 static bool io_kill_linked_timeout(struct io_kiocb *req)
1889 __must_hold(&req->ctx->completion_lock)
1890 __must_hold(&req->ctx->timeout_lock)
1892 struct io_kiocb *link = req->link;
1895 * Can happen if a linked timeout fired and link had been like
1896 * req -> link t-out -> link t-out [-> ...]
1898 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1899 struct io_timeout_data *io = link->async_data;
1901 io_remove_next_linked(req);
1902 link->timeout.head = NULL;
1903 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1904 io_cqring_fill_event(link->ctx, link->user_data,
1906 io_put_req_deferred(link);
1913 static void io_fail_links(struct io_kiocb *req)
1914 __must_hold(&req->ctx->completion_lock)
1916 struct io_kiocb *nxt, *link = req->link;
1923 trace_io_uring_fail_link(req, link);
1924 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1925 io_put_req_deferred(link);
1930 static bool io_disarm_next(struct io_kiocb *req)
1931 __must_hold(&req->ctx->completion_lock)
1933 bool posted = false;
1935 if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1936 struct io_ring_ctx *ctx = req->ctx;
1938 spin_lock_irq(&ctx->timeout_lock);
1939 posted = io_kill_linked_timeout(req);
1940 spin_unlock_irq(&ctx->timeout_lock);
1942 if (unlikely((req->flags & REQ_F_FAIL) &&
1943 !(req->flags & REQ_F_HARDLINK))) {
1944 posted |= (req->link != NULL);
1950 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1952 struct io_kiocb *nxt;
1955 * If LINK is set, we have dependent requests in this chain. If we
1956 * didn't fail this request, queue the first one up, moving any other
1957 * dependencies to the next request. In case of failure, fail the rest
1960 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1961 struct io_ring_ctx *ctx = req->ctx;
1964 spin_lock(&ctx->completion_lock);
1965 posted = io_disarm_next(req);
1967 io_commit_cqring(req->ctx);
1968 spin_unlock(&ctx->completion_lock);
1970 io_cqring_ev_posted(ctx);
1977 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1979 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1981 return __io_req_find_next(req);
1984 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1988 if (ctx->submit_state.compl_nr) {
1989 mutex_lock(&ctx->uring_lock);
1990 io_submit_flush_completions(ctx);
1991 mutex_unlock(&ctx->uring_lock);
1993 percpu_ref_put(&ctx->refs);
1996 static void tctx_task_work(struct callback_head *cb)
1998 struct io_ring_ctx *ctx = NULL;
1999 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2003 struct io_wq_work_node *node;
2005 spin_lock_irq(&tctx->task_lock);
2006 node = tctx->task_list.first;
2007 INIT_WQ_LIST(&tctx->task_list);
2009 tctx->task_running = false;
2010 spin_unlock_irq(&tctx->task_lock);
2015 struct io_wq_work_node *next = node->next;
2016 struct io_kiocb *req = container_of(node, struct io_kiocb,
2019 if (req->ctx != ctx) {
2020 ctx_flush_and_put(ctx);
2022 percpu_ref_get(&ctx->refs);
2024 req->io_task_work.func(req);
2031 ctx_flush_and_put(ctx);
2034 static void io_req_task_work_add(struct io_kiocb *req)
2036 struct task_struct *tsk = req->task;
2037 struct io_uring_task *tctx = tsk->io_uring;
2038 enum task_work_notify_mode notify;
2039 struct io_wq_work_node *node;
2040 unsigned long flags;
2043 WARN_ON_ONCE(!tctx);
2045 spin_lock_irqsave(&tctx->task_lock, flags);
2046 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2047 running = tctx->task_running;
2049 tctx->task_running = true;
2050 spin_unlock_irqrestore(&tctx->task_lock, flags);
2052 /* task_work already pending, we're done */
2057 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2058 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2059 * processing task_work. There's no reliable way to tell if TWA_RESUME
2062 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2063 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2064 wake_up_process(tsk);
2068 spin_lock_irqsave(&tctx->task_lock, flags);
2069 tctx->task_running = false;
2070 node = tctx->task_list.first;
2071 INIT_WQ_LIST(&tctx->task_list);
2072 spin_unlock_irqrestore(&tctx->task_lock, flags);
2075 req = container_of(node, struct io_kiocb, io_task_work.node);
2077 if (llist_add(&req->io_task_work.fallback_node,
2078 &req->ctx->fallback_llist))
2079 schedule_delayed_work(&req->ctx->fallback_work, 1);
2083 static void io_req_task_cancel(struct io_kiocb *req)
2085 struct io_ring_ctx *ctx = req->ctx;
2087 /* ctx is guaranteed to stay alive while we hold uring_lock */
2088 mutex_lock(&ctx->uring_lock);
2089 io_req_complete_failed(req, req->result);
2090 mutex_unlock(&ctx->uring_lock);
2093 static void io_req_task_submit(struct io_kiocb *req)
2095 struct io_ring_ctx *ctx = req->ctx;
2097 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2098 mutex_lock(&ctx->uring_lock);
2099 if (likely(!(req->task->flags & PF_EXITING)))
2100 __io_queue_sqe(req);
2102 io_req_complete_failed(req, -EFAULT);
2103 mutex_unlock(&ctx->uring_lock);
2106 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2109 req->io_task_work.func = io_req_task_cancel;
2110 io_req_task_work_add(req);
2113 static void io_req_task_queue(struct io_kiocb *req)
2115 req->io_task_work.func = io_req_task_submit;
2116 io_req_task_work_add(req);
2119 static void io_req_task_queue_reissue(struct io_kiocb *req)
2121 req->io_task_work.func = io_queue_async_work;
2122 io_req_task_work_add(req);
2125 static inline void io_queue_next(struct io_kiocb *req)
2127 struct io_kiocb *nxt = io_req_find_next(req);
2130 io_req_task_queue(nxt);
2133 static void io_free_req(struct io_kiocb *req)
2140 struct task_struct *task;
2145 static inline void io_init_req_batch(struct req_batch *rb)
2152 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2153 struct req_batch *rb)
2156 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2157 if (rb->task == current)
2158 current->io_uring->cached_refs += rb->task_refs;
2160 io_put_task(rb->task, rb->task_refs);
2163 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2164 struct io_submit_state *state)
2167 io_dismantle_req(req);
2169 if (req->task != rb->task) {
2171 io_put_task(rb->task, rb->task_refs);
2172 rb->task = req->task;
2178 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2179 state->reqs[state->free_reqs++] = req;
2181 list_add(&req->inflight_entry, &state->free_list);
2184 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2185 __must_hold(&ctx->uring_lock)
2187 struct io_submit_state *state = &ctx->submit_state;
2188 int i, nr = state->compl_nr;
2189 struct req_batch rb;
2191 spin_lock(&ctx->completion_lock);
2192 for (i = 0; i < nr; i++) {
2193 struct io_kiocb *req = state->compl_reqs[i];
2195 __io_cqring_fill_event(ctx, req->user_data, req->result,
2198 io_commit_cqring(ctx);
2199 spin_unlock(&ctx->completion_lock);
2200 io_cqring_ev_posted(ctx);
2202 io_init_req_batch(&rb);
2203 for (i = 0; i < nr; i++) {
2204 struct io_kiocb *req = state->compl_reqs[i];
2206 if (req_ref_put_and_test(req))
2207 io_req_free_batch(&rb, req, &ctx->submit_state);
2210 io_req_free_batch_finish(ctx, &rb);
2211 state->compl_nr = 0;
2215 * Drop reference to request, return next in chain (if there is one) if this
2216 * was the last reference to this request.
2218 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2220 struct io_kiocb *nxt = NULL;
2222 if (req_ref_put_and_test(req)) {
2223 nxt = io_req_find_next(req);
2229 static inline void io_put_req(struct io_kiocb *req)
2231 if (req_ref_put_and_test(req))
2235 static inline void io_put_req_deferred(struct io_kiocb *req)
2237 if (req_ref_put_and_test(req)) {
2238 req->io_task_work.func = io_free_req;
2239 io_req_task_work_add(req);
2243 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2245 /* See comment at the top of this file */
2247 return __io_cqring_events(ctx);
2250 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2252 struct io_rings *rings = ctx->rings;
2254 /* make sure SQ entry isn't read before tail */
2255 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2258 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2260 unsigned int cflags;
2262 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2263 cflags |= IORING_CQE_F_BUFFER;
2264 req->flags &= ~REQ_F_BUFFER_SELECTED;
2269 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2271 struct io_buffer *kbuf;
2273 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2274 return io_put_kbuf(req, kbuf);
2277 static inline bool io_run_task_work(void)
2279 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2280 __set_current_state(TASK_RUNNING);
2281 tracehook_notify_signal();
2289 * Find and free completed poll iocbs
2291 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2292 struct list_head *done, bool resubmit)
2294 struct req_batch rb;
2295 struct io_kiocb *req;
2297 /* order with ->result store in io_complete_rw_iopoll() */
2300 io_init_req_batch(&rb);
2301 while (!list_empty(done)) {
2304 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2305 list_del(&req->inflight_entry);
2307 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2308 !(req->flags & REQ_F_DONT_REISSUE)) {
2309 req->iopoll_completed = 0;
2310 io_req_task_queue_reissue(req);
2314 if (req->flags & REQ_F_BUFFER_SELECTED)
2315 cflags = io_put_rw_kbuf(req);
2317 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2320 if (req_ref_put_and_test(req))
2321 io_req_free_batch(&rb, req, &ctx->submit_state);
2324 io_commit_cqring(ctx);
2325 io_cqring_ev_posted_iopoll(ctx);
2326 io_req_free_batch_finish(ctx, &rb);
2329 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2330 long min, bool resubmit)
2332 struct io_kiocb *req, *tmp;
2337 * Only spin for completions if we don't have multiple devices hanging
2338 * off our complete list, and we're under the requested amount.
2340 spin = !ctx->poll_multi_queue && *nr_events < min;
2342 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2343 struct kiocb *kiocb = &req->rw.kiocb;
2347 * Move completed and retryable entries to our local lists.
2348 * If we find a request that requires polling, break out
2349 * and complete those lists first, if we have entries there.
2351 if (READ_ONCE(req->iopoll_completed)) {
2352 list_move_tail(&req->inflight_entry, &done);
2355 if (!list_empty(&done))
2358 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2359 if (unlikely(ret < 0))
2364 /* iopoll may have completed current req */
2365 if (READ_ONCE(req->iopoll_completed))
2366 list_move_tail(&req->inflight_entry, &done);
2369 if (!list_empty(&done))
2370 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2376 * We can't just wait for polled events to come to us, we have to actively
2377 * find and complete them.
2379 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2381 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2384 mutex_lock(&ctx->uring_lock);
2385 while (!list_empty(&ctx->iopoll_list)) {
2386 unsigned int nr_events = 0;
2388 io_do_iopoll(ctx, &nr_events, 0, false);
2390 /* let it sleep and repeat later if can't complete a request */
2394 * Ensure we allow local-to-the-cpu processing to take place,
2395 * in this case we need to ensure that we reap all events.
2396 * Also let task_work, etc. to progress by releasing the mutex
2398 if (need_resched()) {
2399 mutex_unlock(&ctx->uring_lock);
2401 mutex_lock(&ctx->uring_lock);
2404 mutex_unlock(&ctx->uring_lock);
2407 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2409 unsigned int nr_events = 0;
2413 * We disallow the app entering submit/complete with polling, but we
2414 * still need to lock the ring to prevent racing with polled issue
2415 * that got punted to a workqueue.
2417 mutex_lock(&ctx->uring_lock);
2419 * Don't enter poll loop if we already have events pending.
2420 * If we do, we can potentially be spinning for commands that
2421 * already triggered a CQE (eg in error).
2423 if (test_bit(0, &ctx->check_cq_overflow))
2424 __io_cqring_overflow_flush(ctx, false);
2425 if (io_cqring_events(ctx))
2429 * If a submit got punted to a workqueue, we can have the
2430 * application entering polling for a command before it gets
2431 * issued. That app will hold the uring_lock for the duration
2432 * of the poll right here, so we need to take a breather every
2433 * now and then to ensure that the issue has a chance to add
2434 * the poll to the issued list. Otherwise we can spin here
2435 * forever, while the workqueue is stuck trying to acquire the
2438 if (list_empty(&ctx->iopoll_list)) {
2439 u32 tail = ctx->cached_cq_tail;
2441 mutex_unlock(&ctx->uring_lock);
2443 mutex_lock(&ctx->uring_lock);
2445 /* some requests don't go through iopoll_list */
2446 if (tail != ctx->cached_cq_tail ||
2447 list_empty(&ctx->iopoll_list))
2450 ret = io_do_iopoll(ctx, &nr_events, min, true);
2451 } while (!ret && nr_events < min && !need_resched());
2453 mutex_unlock(&ctx->uring_lock);
2457 static void kiocb_end_write(struct io_kiocb *req)
2460 * Tell lockdep we inherited freeze protection from submission
2463 if (req->flags & REQ_F_ISREG) {
2464 struct super_block *sb = file_inode(req->file)->i_sb;
2466 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2472 static bool io_resubmit_prep(struct io_kiocb *req)
2474 struct io_async_rw *rw = req->async_data;
2477 return !io_req_prep_async(req);
2478 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2479 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2483 static bool io_rw_should_reissue(struct io_kiocb *req)
2485 umode_t mode = file_inode(req->file)->i_mode;
2486 struct io_ring_ctx *ctx = req->ctx;
2488 if (!S_ISBLK(mode) && !S_ISREG(mode))
2490 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2491 !(ctx->flags & IORING_SETUP_IOPOLL)))
2494 * If ref is dying, we might be running poll reap from the exit work.
2495 * Don't attempt to reissue from that path, just let it fail with
2498 if (percpu_ref_is_dying(&ctx->refs))
2501 * Play it safe and assume not safe to re-import and reissue if we're
2502 * not in the original thread group (or in task context).
2504 if (!same_thread_group(req->task, current) || !in_task())
2509 static bool io_resubmit_prep(struct io_kiocb *req)
2513 static bool io_rw_should_reissue(struct io_kiocb *req)
2519 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2521 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2522 kiocb_end_write(req);
2523 if (res != req->result) {
2524 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2525 io_rw_should_reissue(req)) {
2526 req->flags |= REQ_F_REISSUE;
2535 static void io_req_task_complete(struct io_kiocb *req)
2539 if (req->flags & REQ_F_BUFFER_SELECTED)
2540 cflags = io_put_rw_kbuf(req);
2541 __io_req_complete(req, 0, req->result, cflags);
2544 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2545 unsigned int issue_flags)
2547 if (__io_complete_rw_common(req, res))
2549 io_req_task_complete(req);
2552 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2554 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2556 if (__io_complete_rw_common(req, res))
2559 req->io_task_work.func = io_req_task_complete;
2560 io_req_task_work_add(req);
2563 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2565 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2567 if (kiocb->ki_flags & IOCB_WRITE)
2568 kiocb_end_write(req);
2569 if (unlikely(res != req->result)) {
2570 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2571 io_resubmit_prep(req))) {
2573 req->flags |= REQ_F_DONT_REISSUE;
2577 WRITE_ONCE(req->result, res);
2578 /* order with io_iopoll_complete() checking ->result */
2580 WRITE_ONCE(req->iopoll_completed, 1);
2584 * After the iocb has been issued, it's safe to be found on the poll list.
2585 * Adding the kiocb to the list AFTER submission ensures that we don't
2586 * find it from a io_do_iopoll() thread before the issuer is done
2587 * accessing the kiocb cookie.
2589 static void io_iopoll_req_issued(struct io_kiocb *req)
2591 struct io_ring_ctx *ctx = req->ctx;
2592 const bool in_async = io_wq_current_is_worker();
2594 /* workqueue context doesn't hold uring_lock, grab it now */
2595 if (unlikely(in_async))
2596 mutex_lock(&ctx->uring_lock);
2599 * Track whether we have multiple files in our lists. This will impact
2600 * how we do polling eventually, not spinning if we're on potentially
2601 * different devices.
2603 if (list_empty(&ctx->iopoll_list)) {
2604 ctx->poll_multi_queue = false;
2605 } else if (!ctx->poll_multi_queue) {
2606 struct io_kiocb *list_req;
2607 unsigned int queue_num0, queue_num1;
2609 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2612 if (list_req->file != req->file) {
2613 ctx->poll_multi_queue = true;
2615 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2616 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2617 if (queue_num0 != queue_num1)
2618 ctx->poll_multi_queue = true;
2623 * For fast devices, IO may have already completed. If it has, add
2624 * it to the front so we find it first.
2626 if (READ_ONCE(req->iopoll_completed))
2627 list_add(&req->inflight_entry, &ctx->iopoll_list);
2629 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2631 if (unlikely(in_async)) {
2633 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2634 * in sq thread task context or in io worker task context. If
2635 * current task context is sq thread, we don't need to check
2636 * whether should wake up sq thread.
2638 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2639 wq_has_sleeper(&ctx->sq_data->wait))
2640 wake_up(&ctx->sq_data->wait);
2642 mutex_unlock(&ctx->uring_lock);
2646 static bool io_bdev_nowait(struct block_device *bdev)
2648 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2652 * If we tracked the file through the SCM inflight mechanism, we could support
2653 * any file. For now, just ensure that anything potentially problematic is done
2656 static bool __io_file_supports_nowait(struct file *file, int rw)
2658 umode_t mode = file_inode(file)->i_mode;
2660 if (S_ISBLK(mode)) {
2661 if (IS_ENABLED(CONFIG_BLOCK) &&
2662 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2668 if (S_ISREG(mode)) {
2669 if (IS_ENABLED(CONFIG_BLOCK) &&
2670 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2671 file->f_op != &io_uring_fops)
2676 /* any ->read/write should understand O_NONBLOCK */
2677 if (file->f_flags & O_NONBLOCK)
2680 if (!(file->f_mode & FMODE_NOWAIT))
2684 return file->f_op->read_iter != NULL;
2686 return file->f_op->write_iter != NULL;
2689 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2691 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2693 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2696 return __io_file_supports_nowait(req->file, rw);
2699 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2701 struct io_ring_ctx *ctx = req->ctx;
2702 struct kiocb *kiocb = &req->rw.kiocb;
2703 struct file *file = req->file;
2707 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2708 req->flags |= REQ_F_ISREG;
2710 kiocb->ki_pos = READ_ONCE(sqe->off);
2711 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2712 req->flags |= REQ_F_CUR_POS;
2713 kiocb->ki_pos = file->f_pos;
2715 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2716 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2717 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2721 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2722 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2723 req->flags |= REQ_F_NOWAIT;
2725 ioprio = READ_ONCE(sqe->ioprio);
2727 ret = ioprio_check_cap(ioprio);
2731 kiocb->ki_ioprio = ioprio;
2733 kiocb->ki_ioprio = get_current_ioprio();
2735 if (ctx->flags & IORING_SETUP_IOPOLL) {
2736 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2737 !kiocb->ki_filp->f_op->iopoll)
2740 kiocb->ki_flags |= IOCB_HIPRI;
2741 kiocb->ki_complete = io_complete_rw_iopoll;
2742 req->iopoll_completed = 0;
2744 if (kiocb->ki_flags & IOCB_HIPRI)
2746 kiocb->ki_complete = io_complete_rw;
2749 if (req->opcode == IORING_OP_READ_FIXED ||
2750 req->opcode == IORING_OP_WRITE_FIXED) {
2752 io_req_set_rsrc_node(req);
2755 req->rw.addr = READ_ONCE(sqe->addr);
2756 req->rw.len = READ_ONCE(sqe->len);
2757 req->buf_index = READ_ONCE(sqe->buf_index);
2761 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2767 case -ERESTARTNOINTR:
2768 case -ERESTARTNOHAND:
2769 case -ERESTART_RESTARTBLOCK:
2771 * We can't just restart the syscall, since previously
2772 * submitted sqes may already be in progress. Just fail this
2778 kiocb->ki_complete(kiocb, ret, 0);
2782 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2783 unsigned int issue_flags)
2785 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2786 struct io_async_rw *io = req->async_data;
2787 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2789 /* add previously done IO, if any */
2790 if (io && io->bytes_done > 0) {
2792 ret = io->bytes_done;
2794 ret += io->bytes_done;
2797 if (req->flags & REQ_F_CUR_POS)
2798 req->file->f_pos = kiocb->ki_pos;
2799 if (ret >= 0 && check_reissue)
2800 __io_complete_rw(req, ret, 0, issue_flags);
2802 io_rw_done(kiocb, ret);
2804 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2805 req->flags &= ~REQ_F_REISSUE;
2806 if (io_resubmit_prep(req)) {
2807 io_req_task_queue_reissue(req);
2812 if (req->flags & REQ_F_BUFFER_SELECTED)
2813 cflags = io_put_rw_kbuf(req);
2814 __io_req_complete(req, issue_flags, ret, cflags);
2819 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2820 struct io_mapped_ubuf *imu)
2822 size_t len = req->rw.len;
2823 u64 buf_end, buf_addr = req->rw.addr;
2826 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2828 /* not inside the mapped region */
2829 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2833 * May not be a start of buffer, set size appropriately
2834 * and advance us to the beginning.
2836 offset = buf_addr - imu->ubuf;
2837 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2841 * Don't use iov_iter_advance() here, as it's really slow for
2842 * using the latter parts of a big fixed buffer - it iterates
2843 * over each segment manually. We can cheat a bit here, because
2846 * 1) it's a BVEC iter, we set it up
2847 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2848 * first and last bvec
2850 * So just find our index, and adjust the iterator afterwards.
2851 * If the offset is within the first bvec (or the whole first
2852 * bvec, just use iov_iter_advance(). This makes it easier
2853 * since we can just skip the first segment, which may not
2854 * be PAGE_SIZE aligned.
2856 const struct bio_vec *bvec = imu->bvec;
2858 if (offset <= bvec->bv_len) {
2859 iov_iter_advance(iter, offset);
2861 unsigned long seg_skip;
2863 /* skip first vec */
2864 offset -= bvec->bv_len;
2865 seg_skip = 1 + (offset >> PAGE_SHIFT);
2867 iter->bvec = bvec + seg_skip;
2868 iter->nr_segs -= seg_skip;
2869 iter->count -= bvec->bv_len + offset;
2870 iter->iov_offset = offset & ~PAGE_MASK;
2877 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2879 struct io_ring_ctx *ctx = req->ctx;
2880 struct io_mapped_ubuf *imu = req->imu;
2881 u16 index, buf_index = req->buf_index;
2884 if (unlikely(buf_index >= ctx->nr_user_bufs))
2886 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2887 imu = READ_ONCE(ctx->user_bufs[index]);
2890 return __io_import_fixed(req, rw, iter, imu);
2893 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2896 mutex_unlock(&ctx->uring_lock);
2899 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2902 * "Normal" inline submissions always hold the uring_lock, since we
2903 * grab it from the system call. Same is true for the SQPOLL offload.
2904 * The only exception is when we've detached the request and issue it
2905 * from an async worker thread, grab the lock for that case.
2908 mutex_lock(&ctx->uring_lock);
2911 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2912 int bgid, struct io_buffer *kbuf,
2915 struct io_buffer *head;
2917 if (req->flags & REQ_F_BUFFER_SELECTED)
2920 io_ring_submit_lock(req->ctx, needs_lock);
2922 lockdep_assert_held(&req->ctx->uring_lock);
2924 head = xa_load(&req->ctx->io_buffers, bgid);
2926 if (!list_empty(&head->list)) {
2927 kbuf = list_last_entry(&head->list, struct io_buffer,
2929 list_del(&kbuf->list);
2932 xa_erase(&req->ctx->io_buffers, bgid);
2934 if (*len > kbuf->len)
2937 kbuf = ERR_PTR(-ENOBUFS);
2940 io_ring_submit_unlock(req->ctx, needs_lock);
2945 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2948 struct io_buffer *kbuf;
2951 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2952 bgid = req->buf_index;
2953 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2956 req->rw.addr = (u64) (unsigned long) kbuf;
2957 req->flags |= REQ_F_BUFFER_SELECTED;
2958 return u64_to_user_ptr(kbuf->addr);
2961 #ifdef CONFIG_COMPAT
2962 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2965 struct compat_iovec __user *uiov;
2966 compat_ssize_t clen;
2970 uiov = u64_to_user_ptr(req->rw.addr);
2971 if (!access_ok(uiov, sizeof(*uiov)))
2973 if (__get_user(clen, &uiov->iov_len))
2979 buf = io_rw_buffer_select(req, &len, needs_lock);
2981 return PTR_ERR(buf);
2982 iov[0].iov_base = buf;
2983 iov[0].iov_len = (compat_size_t) len;
2988 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2991 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2995 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2998 len = iov[0].iov_len;
3001 buf = io_rw_buffer_select(req, &len, needs_lock);
3003 return PTR_ERR(buf);
3004 iov[0].iov_base = buf;
3005 iov[0].iov_len = len;
3009 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3012 if (req->flags & REQ_F_BUFFER_SELECTED) {
3013 struct io_buffer *kbuf;
3015 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3016 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3017 iov[0].iov_len = kbuf->len;
3020 if (req->rw.len != 1)
3023 #ifdef CONFIG_COMPAT
3024 if (req->ctx->compat)
3025 return io_compat_import(req, iov, needs_lock);
3028 return __io_iov_buffer_select(req, iov, needs_lock);
3031 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3032 struct iov_iter *iter, bool needs_lock)
3034 void __user *buf = u64_to_user_ptr(req->rw.addr);
3035 size_t sqe_len = req->rw.len;
3036 u8 opcode = req->opcode;
3039 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3041 return io_import_fixed(req, rw, iter);
3044 /* buffer index only valid with fixed read/write, or buffer select */
3045 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3048 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3049 if (req->flags & REQ_F_BUFFER_SELECT) {
3050 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3052 return PTR_ERR(buf);
3053 req->rw.len = sqe_len;
3056 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3061 if (req->flags & REQ_F_BUFFER_SELECT) {
3062 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3064 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3069 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3073 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3075 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3079 * For files that don't have ->read_iter() and ->write_iter(), handle them
3080 * by looping over ->read() or ->write() manually.
3082 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3084 struct kiocb *kiocb = &req->rw.kiocb;
3085 struct file *file = req->file;
3089 * Don't support polled IO through this interface, and we can't
3090 * support non-blocking either. For the latter, this just causes
3091 * the kiocb to be handled from an async context.
3093 if (kiocb->ki_flags & IOCB_HIPRI)
3095 if (kiocb->ki_flags & IOCB_NOWAIT)
3098 while (iov_iter_count(iter)) {
3102 if (!iov_iter_is_bvec(iter)) {
3103 iovec = iov_iter_iovec(iter);
3105 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3106 iovec.iov_len = req->rw.len;
3110 nr = file->f_op->read(file, iovec.iov_base,
3111 iovec.iov_len, io_kiocb_ppos(kiocb));
3113 nr = file->f_op->write(file, iovec.iov_base,
3114 iovec.iov_len, io_kiocb_ppos(kiocb));
3123 if (nr != iovec.iov_len)
3127 iov_iter_advance(iter, nr);
3133 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3134 const struct iovec *fast_iov, struct iov_iter *iter)
3136 struct io_async_rw *rw = req->async_data;
3138 memcpy(&rw->iter, iter, sizeof(*iter));
3139 rw->free_iovec = iovec;
3141 /* can only be fixed buffers, no need to do anything */
3142 if (iov_iter_is_bvec(iter))
3145 unsigned iov_off = 0;
3147 rw->iter.iov = rw->fast_iov;
3148 if (iter->iov != fast_iov) {
3149 iov_off = iter->iov - fast_iov;
3150 rw->iter.iov += iov_off;
3152 if (rw->fast_iov != fast_iov)
3153 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3154 sizeof(struct iovec) * iter->nr_segs);
3156 req->flags |= REQ_F_NEED_CLEANUP;
3160 static inline int io_alloc_async_data(struct io_kiocb *req)
3162 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3163 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3164 return req->async_data == NULL;
3167 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3168 const struct iovec *fast_iov,
3169 struct iov_iter *iter, bool force)
3171 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3173 if (!req->async_data) {
3174 if (io_alloc_async_data(req)) {
3179 io_req_map_rw(req, iovec, fast_iov, iter);
3184 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3186 struct io_async_rw *iorw = req->async_data;
3187 struct iovec *iov = iorw->fast_iov;
3190 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3191 if (unlikely(ret < 0))
3194 iorw->bytes_done = 0;
3195 iorw->free_iovec = iov;
3197 req->flags |= REQ_F_NEED_CLEANUP;
3201 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3203 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3205 return io_prep_rw(req, sqe);
3209 * This is our waitqueue callback handler, registered through lock_page_async()
3210 * when we initially tried to do the IO with the iocb armed our waitqueue.
3211 * This gets called when the page is unlocked, and we generally expect that to
3212 * happen when the page IO is completed and the page is now uptodate. This will
3213 * queue a task_work based retry of the operation, attempting to copy the data
3214 * again. If the latter fails because the page was NOT uptodate, then we will
3215 * do a thread based blocking retry of the operation. That's the unexpected
3218 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3219 int sync, void *arg)
3221 struct wait_page_queue *wpq;
3222 struct io_kiocb *req = wait->private;
3223 struct wait_page_key *key = arg;
3225 wpq = container_of(wait, struct wait_page_queue, wait);
3227 if (!wake_page_match(wpq, key))
3230 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3231 list_del_init(&wait->entry);
3232 io_req_task_queue(req);
3237 * This controls whether a given IO request should be armed for async page
3238 * based retry. If we return false here, the request is handed to the async
3239 * worker threads for retry. If we're doing buffered reads on a regular file,
3240 * we prepare a private wait_page_queue entry and retry the operation. This
3241 * will either succeed because the page is now uptodate and unlocked, or it
3242 * will register a callback when the page is unlocked at IO completion. Through
3243 * that callback, io_uring uses task_work to setup a retry of the operation.
3244 * That retry will attempt the buffered read again. The retry will generally
3245 * succeed, or in rare cases where it fails, we then fall back to using the
3246 * async worker threads for a blocking retry.
3248 static bool io_rw_should_retry(struct io_kiocb *req)
3250 struct io_async_rw *rw = req->async_data;
3251 struct wait_page_queue *wait = &rw->wpq;
3252 struct kiocb *kiocb = &req->rw.kiocb;
3254 /* never retry for NOWAIT, we just complete with -EAGAIN */
3255 if (req->flags & REQ_F_NOWAIT)
3258 /* Only for buffered IO */
3259 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3263 * just use poll if we can, and don't attempt if the fs doesn't
3264 * support callback based unlocks
3266 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3269 wait->wait.func = io_async_buf_func;
3270 wait->wait.private = req;
3271 wait->wait.flags = 0;
3272 INIT_LIST_HEAD(&wait->wait.entry);
3273 kiocb->ki_flags |= IOCB_WAITQ;
3274 kiocb->ki_flags &= ~IOCB_NOWAIT;
3275 kiocb->ki_waitq = wait;
3279 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3281 if (req->file->f_op->read_iter)
3282 return call_read_iter(req->file, &req->rw.kiocb, iter);
3283 else if (req->file->f_op->read)
3284 return loop_rw_iter(READ, req, iter);
3289 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3291 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3292 struct kiocb *kiocb = &req->rw.kiocb;
3293 struct iov_iter __iter, *iter = &__iter;
3294 struct io_async_rw *rw = req->async_data;
3295 ssize_t io_size, ret, ret2;
3296 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3302 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3306 io_size = iov_iter_count(iter);
3307 req->result = io_size;
3309 /* Ensure we clear previously set non-block flag */
3310 if (!force_nonblock)
3311 kiocb->ki_flags &= ~IOCB_NOWAIT;
3313 kiocb->ki_flags |= IOCB_NOWAIT;
3315 /* If the file doesn't support async, just async punt */
3316 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3317 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3318 return ret ?: -EAGAIN;
3321 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3322 if (unlikely(ret)) {
3327 ret = io_iter_do_read(req, iter);
3329 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3330 req->flags &= ~REQ_F_REISSUE;
3331 /* IOPOLL retry should happen for io-wq threads */
3332 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3334 /* no retry on NONBLOCK nor RWF_NOWAIT */
3335 if (req->flags & REQ_F_NOWAIT)
3337 /* some cases will consume bytes even on error returns */
3338 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3340 } else if (ret == -EIOCBQUEUED) {
3342 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3343 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3344 /* read all, failed, already did sync or don't want to retry */
3348 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3353 rw = req->async_data;
3354 /* now use our persistent iterator, if we aren't already */
3359 rw->bytes_done += ret;
3360 /* if we can retry, do so with the callbacks armed */
3361 if (!io_rw_should_retry(req)) {
3362 kiocb->ki_flags &= ~IOCB_WAITQ;
3367 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3368 * we get -EIOCBQUEUED, then we'll get a notification when the
3369 * desired page gets unlocked. We can also get a partial read
3370 * here, and if we do, then just retry at the new offset.
3372 ret = io_iter_do_read(req, iter);
3373 if (ret == -EIOCBQUEUED)
3375 /* we got some bytes, but not all. retry. */
3376 kiocb->ki_flags &= ~IOCB_WAITQ;
3377 } while (ret > 0 && ret < io_size);
3379 kiocb_done(kiocb, ret, issue_flags);
3381 /* it's faster to check here then delegate to kfree */
3387 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3389 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3391 return io_prep_rw(req, sqe);
3394 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3396 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3397 struct kiocb *kiocb = &req->rw.kiocb;
3398 struct iov_iter __iter, *iter = &__iter;
3399 struct io_async_rw *rw = req->async_data;
3400 ssize_t ret, ret2, io_size;
3401 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3407 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3411 io_size = iov_iter_count(iter);
3412 req->result = io_size;
3414 /* Ensure we clear previously set non-block flag */
3415 if (!force_nonblock)
3416 kiocb->ki_flags &= ~IOCB_NOWAIT;
3418 kiocb->ki_flags |= IOCB_NOWAIT;
3420 /* If the file doesn't support async, just async punt */
3421 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3424 /* file path doesn't support NOWAIT for non-direct_IO */
3425 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3426 (req->flags & REQ_F_ISREG))
3429 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3434 * Open-code file_start_write here to grab freeze protection,
3435 * which will be released by another thread in
3436 * io_complete_rw(). Fool lockdep by telling it the lock got
3437 * released so that it doesn't complain about the held lock when
3438 * we return to userspace.
3440 if (req->flags & REQ_F_ISREG) {
3441 sb_start_write(file_inode(req->file)->i_sb);
3442 __sb_writers_release(file_inode(req->file)->i_sb,
3445 kiocb->ki_flags |= IOCB_WRITE;
3447 if (req->file->f_op->write_iter)
3448 ret2 = call_write_iter(req->file, kiocb, iter);
3449 else if (req->file->f_op->write)
3450 ret2 = loop_rw_iter(WRITE, req, iter);
3454 if (req->flags & REQ_F_REISSUE) {
3455 req->flags &= ~REQ_F_REISSUE;
3460 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3461 * retry them without IOCB_NOWAIT.
3463 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3465 /* no retry on NONBLOCK nor RWF_NOWAIT */
3466 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3468 if (!force_nonblock || ret2 != -EAGAIN) {
3469 /* IOPOLL retry should happen for io-wq threads */
3470 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3473 kiocb_done(kiocb, ret2, issue_flags);
3476 /* some cases will consume bytes even on error returns */
3477 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3478 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3479 return ret ?: -EAGAIN;
3482 /* it's reportedly faster than delegating the null check to kfree() */
3488 static int io_renameat_prep(struct io_kiocb *req,
3489 const struct io_uring_sqe *sqe)
3491 struct io_rename *ren = &req->rename;
3492 const char __user *oldf, *newf;
3494 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3496 if (sqe->ioprio || sqe->buf_index)
3498 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3501 ren->old_dfd = READ_ONCE(sqe->fd);
3502 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3503 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3504 ren->new_dfd = READ_ONCE(sqe->len);
3505 ren->flags = READ_ONCE(sqe->rename_flags);
3507 ren->oldpath = getname(oldf);
3508 if (IS_ERR(ren->oldpath))
3509 return PTR_ERR(ren->oldpath);
3511 ren->newpath = getname(newf);
3512 if (IS_ERR(ren->newpath)) {
3513 putname(ren->oldpath);
3514 return PTR_ERR(ren->newpath);
3517 req->flags |= REQ_F_NEED_CLEANUP;
3521 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3523 struct io_rename *ren = &req->rename;
3526 if (issue_flags & IO_URING_F_NONBLOCK)
3529 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3530 ren->newpath, ren->flags);
3532 req->flags &= ~REQ_F_NEED_CLEANUP;
3535 io_req_complete(req, ret);
3539 static int io_unlinkat_prep(struct io_kiocb *req,
3540 const struct io_uring_sqe *sqe)
3542 struct io_unlink *un = &req->unlink;
3543 const char __user *fname;
3545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3547 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3549 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3552 un->dfd = READ_ONCE(sqe->fd);
3554 un->flags = READ_ONCE(sqe->unlink_flags);
3555 if (un->flags & ~AT_REMOVEDIR)
3558 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3559 un->filename = getname(fname);
3560 if (IS_ERR(un->filename))
3561 return PTR_ERR(un->filename);
3563 req->flags |= REQ_F_NEED_CLEANUP;
3567 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3569 struct io_unlink *un = &req->unlink;
3572 if (issue_flags & IO_URING_F_NONBLOCK)
3575 if (un->flags & AT_REMOVEDIR)
3576 ret = do_rmdir(un->dfd, un->filename);
3578 ret = do_unlinkat(un->dfd, un->filename);
3580 req->flags &= ~REQ_F_NEED_CLEANUP;
3583 io_req_complete(req, ret);
3587 static int io_shutdown_prep(struct io_kiocb *req,
3588 const struct io_uring_sqe *sqe)
3590 #if defined(CONFIG_NET)
3591 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3593 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3597 req->shutdown.how = READ_ONCE(sqe->len);
3604 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3606 #if defined(CONFIG_NET)
3607 struct socket *sock;
3610 if (issue_flags & IO_URING_F_NONBLOCK)
3613 sock = sock_from_file(req->file);
3614 if (unlikely(!sock))
3617 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3620 io_req_complete(req, ret);
3627 static int __io_splice_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 struct io_splice *sp = &req->splice;
3631 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3633 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3637 sp->len = READ_ONCE(sqe->len);
3638 sp->flags = READ_ONCE(sqe->splice_flags);
3640 if (unlikely(sp->flags & ~valid_flags))
3643 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3644 (sp->flags & SPLICE_F_FD_IN_FIXED));
3647 req->flags |= REQ_F_NEED_CLEANUP;
3651 static int io_tee_prep(struct io_kiocb *req,
3652 const struct io_uring_sqe *sqe)
3654 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3656 return __io_splice_prep(req, sqe);
3659 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3661 struct io_splice *sp = &req->splice;
3662 struct file *in = sp->file_in;
3663 struct file *out = sp->file_out;
3664 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3667 if (issue_flags & IO_URING_F_NONBLOCK)
3670 ret = do_tee(in, out, sp->len, flags);
3672 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3674 req->flags &= ~REQ_F_NEED_CLEANUP;
3678 io_req_complete(req, ret);
3682 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3684 struct io_splice *sp = &req->splice;
3686 sp->off_in = READ_ONCE(sqe->splice_off_in);
3687 sp->off_out = READ_ONCE(sqe->off);
3688 return __io_splice_prep(req, sqe);
3691 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3693 struct io_splice *sp = &req->splice;
3694 struct file *in = sp->file_in;
3695 struct file *out = sp->file_out;
3696 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3697 loff_t *poff_in, *poff_out;
3700 if (issue_flags & IO_URING_F_NONBLOCK)
3703 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3704 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3707 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3709 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3711 req->flags &= ~REQ_F_NEED_CLEANUP;
3715 io_req_complete(req, ret);
3720 * IORING_OP_NOP just posts a completion event, nothing else.
3722 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3724 struct io_ring_ctx *ctx = req->ctx;
3726 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3729 __io_req_complete(req, issue_flags, 0, 0);
3733 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3735 struct io_ring_ctx *ctx = req->ctx;
3740 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3742 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3745 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3746 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3749 req->sync.off = READ_ONCE(sqe->off);
3750 req->sync.len = READ_ONCE(sqe->len);
3754 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3756 loff_t end = req->sync.off + req->sync.len;
3759 /* fsync always requires a blocking context */
3760 if (issue_flags & IO_URING_F_NONBLOCK)
3763 ret = vfs_fsync_range(req->file, req->sync.off,
3764 end > 0 ? end : LLONG_MAX,
3765 req->sync.flags & IORING_FSYNC_DATASYNC);
3768 io_req_complete(req, ret);
3772 static int io_fallocate_prep(struct io_kiocb *req,
3773 const struct io_uring_sqe *sqe)
3775 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3777 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3780 req->sync.off = READ_ONCE(sqe->off);
3781 req->sync.len = READ_ONCE(sqe->addr);
3782 req->sync.mode = READ_ONCE(sqe->len);
3786 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3790 /* fallocate always requiring blocking context */
3791 if (issue_flags & IO_URING_F_NONBLOCK)
3793 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3797 io_req_complete(req, ret);
3801 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3803 const char __user *fname;
3806 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3808 if (unlikely(sqe->ioprio || sqe->buf_index))
3810 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3813 /* open.how should be already initialised */
3814 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3815 req->open.how.flags |= O_LARGEFILE;
3817 req->open.dfd = READ_ONCE(sqe->fd);
3818 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3819 req->open.filename = getname(fname);
3820 if (IS_ERR(req->open.filename)) {
3821 ret = PTR_ERR(req->open.filename);
3822 req->open.filename = NULL;
3825 req->open.nofile = rlimit(RLIMIT_NOFILE);
3826 req->flags |= REQ_F_NEED_CLEANUP;
3830 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3832 u64 mode = READ_ONCE(sqe->len);
3833 u64 flags = READ_ONCE(sqe->open_flags);
3835 req->open.how = build_open_how(flags, mode);
3836 return __io_openat_prep(req, sqe);
3839 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3841 struct open_how __user *how;
3845 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3846 len = READ_ONCE(sqe->len);
3847 if (len < OPEN_HOW_SIZE_VER0)
3850 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3855 return __io_openat_prep(req, sqe);
3858 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3860 struct open_flags op;
3863 bool resolve_nonblock;
3866 ret = build_open_flags(&req->open.how, &op);
3869 nonblock_set = op.open_flag & O_NONBLOCK;
3870 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3871 if (issue_flags & IO_URING_F_NONBLOCK) {
3873 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3874 * it'll always -EAGAIN
3876 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3878 op.lookup_flags |= LOOKUP_CACHED;
3879 op.open_flag |= O_NONBLOCK;
3882 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3886 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3889 * We could hang on to this 'fd' on retrying, but seems like
3890 * marginal gain for something that is now known to be a slower
3891 * path. So just put it, and we'll get a new one when we retry.
3895 ret = PTR_ERR(file);
3896 /* only retry if RESOLVE_CACHED wasn't already set by application */
3897 if (ret == -EAGAIN &&
3898 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3903 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3904 file->f_flags &= ~O_NONBLOCK;
3905 fsnotify_open(file);
3906 fd_install(ret, file);
3908 putname(req->open.filename);
3909 req->flags &= ~REQ_F_NEED_CLEANUP;
3912 __io_req_complete(req, issue_flags, ret, 0);
3916 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3918 return io_openat2(req, issue_flags);
3921 static int io_remove_buffers_prep(struct io_kiocb *req,
3922 const struct io_uring_sqe *sqe)
3924 struct io_provide_buf *p = &req->pbuf;
3927 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3930 tmp = READ_ONCE(sqe->fd);
3931 if (!tmp || tmp > USHRT_MAX)
3934 memset(p, 0, sizeof(*p));
3936 p->bgid = READ_ONCE(sqe->buf_group);
3940 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3941 int bgid, unsigned nbufs)
3945 /* shouldn't happen */
3949 /* the head kbuf is the list itself */
3950 while (!list_empty(&buf->list)) {
3951 struct io_buffer *nxt;
3953 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3954 list_del(&nxt->list);
3961 xa_erase(&ctx->io_buffers, bgid);
3966 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3968 struct io_provide_buf *p = &req->pbuf;
3969 struct io_ring_ctx *ctx = req->ctx;
3970 struct io_buffer *head;
3972 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3974 io_ring_submit_lock(ctx, !force_nonblock);
3976 lockdep_assert_held(&ctx->uring_lock);
3979 head = xa_load(&ctx->io_buffers, p->bgid);
3981 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3985 /* complete before unlock, IOPOLL may need the lock */
3986 __io_req_complete(req, issue_flags, ret, 0);
3987 io_ring_submit_unlock(ctx, !force_nonblock);
3991 static int io_provide_buffers_prep(struct io_kiocb *req,
3992 const struct io_uring_sqe *sqe)
3994 unsigned long size, tmp_check;
3995 struct io_provide_buf *p = &req->pbuf;
3998 if (sqe->ioprio || sqe->rw_flags)
4001 tmp = READ_ONCE(sqe->fd);
4002 if (!tmp || tmp > USHRT_MAX)
4005 p->addr = READ_ONCE(sqe->addr);
4006 p->len = READ_ONCE(sqe->len);
4008 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4011 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4014 size = (unsigned long)p->len * p->nbufs;
4015 if (!access_ok(u64_to_user_ptr(p->addr), size))
4018 p->bgid = READ_ONCE(sqe->buf_group);
4019 tmp = READ_ONCE(sqe->off);
4020 if (tmp > USHRT_MAX)
4026 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4028 struct io_buffer *buf;
4029 u64 addr = pbuf->addr;
4030 int i, bid = pbuf->bid;
4032 for (i = 0; i < pbuf->nbufs; i++) {
4033 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4038 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4043 INIT_LIST_HEAD(&buf->list);
4046 list_add_tail(&buf->list, &(*head)->list);
4050 return i ? i : -ENOMEM;
4053 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4055 struct io_provide_buf *p = &req->pbuf;
4056 struct io_ring_ctx *ctx = req->ctx;
4057 struct io_buffer *head, *list;
4059 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4061 io_ring_submit_lock(ctx, !force_nonblock);
4063 lockdep_assert_held(&ctx->uring_lock);
4065 list = head = xa_load(&ctx->io_buffers, p->bgid);
4067 ret = io_add_buffers(p, &head);
4068 if (ret >= 0 && !list) {
4069 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4071 __io_remove_buffers(ctx, head, p->bgid, -1U);
4075 /* complete before unlock, IOPOLL may need the lock */
4076 __io_req_complete(req, issue_flags, ret, 0);
4077 io_ring_submit_unlock(ctx, !force_nonblock);
4081 static int io_epoll_ctl_prep(struct io_kiocb *req,
4082 const struct io_uring_sqe *sqe)
4084 #if defined(CONFIG_EPOLL)
4085 if (sqe->ioprio || sqe->buf_index)
4087 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4090 req->epoll.epfd = READ_ONCE(sqe->fd);
4091 req->epoll.op = READ_ONCE(sqe->len);
4092 req->epoll.fd = READ_ONCE(sqe->off);
4094 if (ep_op_has_event(req->epoll.op)) {
4095 struct epoll_event __user *ev;
4097 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4098 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4108 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4110 #if defined(CONFIG_EPOLL)
4111 struct io_epoll *ie = &req->epoll;
4113 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4115 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4116 if (force_nonblock && ret == -EAGAIN)
4121 __io_req_complete(req, issue_flags, ret, 0);
4128 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4130 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4131 if (sqe->ioprio || sqe->buf_index || sqe->off)
4133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4136 req->madvise.addr = READ_ONCE(sqe->addr);
4137 req->madvise.len = READ_ONCE(sqe->len);
4138 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4145 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4147 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4148 struct io_madvise *ma = &req->madvise;
4151 if (issue_flags & IO_URING_F_NONBLOCK)
4154 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4157 io_req_complete(req, ret);
4164 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4166 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4168 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4171 req->fadvise.offset = READ_ONCE(sqe->off);
4172 req->fadvise.len = READ_ONCE(sqe->len);
4173 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4177 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4179 struct io_fadvise *fa = &req->fadvise;
4182 if (issue_flags & IO_URING_F_NONBLOCK) {
4183 switch (fa->advice) {
4184 case POSIX_FADV_NORMAL:
4185 case POSIX_FADV_RANDOM:
4186 case POSIX_FADV_SEQUENTIAL:
4193 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4196 __io_req_complete(req, issue_flags, ret, 0);
4200 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4204 if (sqe->ioprio || sqe->buf_index)
4206 if (req->flags & REQ_F_FIXED_FILE)
4209 req->statx.dfd = READ_ONCE(sqe->fd);
4210 req->statx.mask = READ_ONCE(sqe->len);
4211 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4212 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4213 req->statx.flags = READ_ONCE(sqe->statx_flags);
4218 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4220 struct io_statx *ctx = &req->statx;
4223 if (issue_flags & IO_URING_F_NONBLOCK)
4226 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4231 io_req_complete(req, ret);
4235 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4237 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4239 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4240 sqe->rw_flags || sqe->buf_index)
4242 if (req->flags & REQ_F_FIXED_FILE)
4245 req->close.fd = READ_ONCE(sqe->fd);
4249 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4251 struct files_struct *files = current->files;
4252 struct io_close *close = &req->close;
4253 struct fdtable *fdt;
4254 struct file *file = NULL;
4257 spin_lock(&files->file_lock);
4258 fdt = files_fdtable(files);
4259 if (close->fd >= fdt->max_fds) {
4260 spin_unlock(&files->file_lock);
4263 file = fdt->fd[close->fd];
4264 if (!file || file->f_op == &io_uring_fops) {
4265 spin_unlock(&files->file_lock);
4270 /* if the file has a flush method, be safe and punt to async */
4271 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4272 spin_unlock(&files->file_lock);
4276 ret = __close_fd_get_file(close->fd, &file);
4277 spin_unlock(&files->file_lock);
4284 /* No ->flush() or already async, safely close from here */
4285 ret = filp_close(file, current->files);
4291 __io_req_complete(req, issue_flags, ret, 0);
4295 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4297 struct io_ring_ctx *ctx = req->ctx;
4299 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4301 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4304 req->sync.off = READ_ONCE(sqe->off);
4305 req->sync.len = READ_ONCE(sqe->len);
4306 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4310 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4314 /* sync_file_range always requires a blocking context */
4315 if (issue_flags & IO_URING_F_NONBLOCK)
4318 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4322 io_req_complete(req, ret);
4326 #if defined(CONFIG_NET)
4327 static int io_setup_async_msg(struct io_kiocb *req,
4328 struct io_async_msghdr *kmsg)
4330 struct io_async_msghdr *async_msg = req->async_data;
4334 if (io_alloc_async_data(req)) {
4335 kfree(kmsg->free_iov);
4338 async_msg = req->async_data;
4339 req->flags |= REQ_F_NEED_CLEANUP;
4340 memcpy(async_msg, kmsg, sizeof(*kmsg));
4341 async_msg->msg.msg_name = &async_msg->addr;
4342 /* if were using fast_iov, set it to the new one */
4343 if (!async_msg->free_iov)
4344 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4349 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4350 struct io_async_msghdr *iomsg)
4352 iomsg->msg.msg_name = &iomsg->addr;
4353 iomsg->free_iov = iomsg->fast_iov;
4354 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4355 req->sr_msg.msg_flags, &iomsg->free_iov);
4358 static int io_sendmsg_prep_async(struct io_kiocb *req)
4362 ret = io_sendmsg_copy_hdr(req, req->async_data);
4364 req->flags |= REQ_F_NEED_CLEANUP;
4368 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4370 struct io_sr_msg *sr = &req->sr_msg;
4372 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4375 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4376 sr->len = READ_ONCE(sqe->len);
4377 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4378 if (sr->msg_flags & MSG_DONTWAIT)
4379 req->flags |= REQ_F_NOWAIT;
4381 #ifdef CONFIG_COMPAT
4382 if (req->ctx->compat)
4383 sr->msg_flags |= MSG_CMSG_COMPAT;
4388 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4390 struct io_async_msghdr iomsg, *kmsg;
4391 struct socket *sock;
4396 sock = sock_from_file(req->file);
4397 if (unlikely(!sock))
4400 kmsg = req->async_data;
4402 ret = io_sendmsg_copy_hdr(req, &iomsg);
4408 flags = req->sr_msg.msg_flags;
4409 if (issue_flags & IO_URING_F_NONBLOCK)
4410 flags |= MSG_DONTWAIT;
4411 if (flags & MSG_WAITALL)
4412 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4414 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4415 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4416 return io_setup_async_msg(req, kmsg);
4417 if (ret == -ERESTARTSYS)
4420 /* fast path, check for non-NULL to avoid function call */
4422 kfree(kmsg->free_iov);
4423 req->flags &= ~REQ_F_NEED_CLEANUP;
4426 __io_req_complete(req, issue_flags, ret, 0);
4430 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4432 struct io_sr_msg *sr = &req->sr_msg;
4435 struct socket *sock;
4440 sock = sock_from_file(req->file);
4441 if (unlikely(!sock))
4444 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4448 msg.msg_name = NULL;
4449 msg.msg_control = NULL;
4450 msg.msg_controllen = 0;
4451 msg.msg_namelen = 0;
4453 flags = req->sr_msg.msg_flags;
4454 if (issue_flags & IO_URING_F_NONBLOCK)
4455 flags |= MSG_DONTWAIT;
4456 if (flags & MSG_WAITALL)
4457 min_ret = iov_iter_count(&msg.msg_iter);
4459 msg.msg_flags = flags;
4460 ret = sock_sendmsg(sock, &msg);
4461 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4463 if (ret == -ERESTARTSYS)
4468 __io_req_complete(req, issue_flags, ret, 0);
4472 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4473 struct io_async_msghdr *iomsg)
4475 struct io_sr_msg *sr = &req->sr_msg;
4476 struct iovec __user *uiov;
4480 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4481 &iomsg->uaddr, &uiov, &iov_len);
4485 if (req->flags & REQ_F_BUFFER_SELECT) {
4488 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4490 sr->len = iomsg->fast_iov[0].iov_len;
4491 iomsg->free_iov = NULL;
4493 iomsg->free_iov = iomsg->fast_iov;
4494 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4495 &iomsg->free_iov, &iomsg->msg.msg_iter,
4504 #ifdef CONFIG_COMPAT
4505 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4506 struct io_async_msghdr *iomsg)
4508 struct io_sr_msg *sr = &req->sr_msg;
4509 struct compat_iovec __user *uiov;
4514 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4519 uiov = compat_ptr(ptr);
4520 if (req->flags & REQ_F_BUFFER_SELECT) {
4521 compat_ssize_t clen;
4525 if (!access_ok(uiov, sizeof(*uiov)))
4527 if (__get_user(clen, &uiov->iov_len))
4532 iomsg->free_iov = NULL;
4534 iomsg->free_iov = iomsg->fast_iov;
4535 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4536 UIO_FASTIOV, &iomsg->free_iov,
4537 &iomsg->msg.msg_iter, true);
4546 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4547 struct io_async_msghdr *iomsg)
4549 iomsg->msg.msg_name = &iomsg->addr;
4551 #ifdef CONFIG_COMPAT
4552 if (req->ctx->compat)
4553 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4556 return __io_recvmsg_copy_hdr(req, iomsg);
4559 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4562 struct io_sr_msg *sr = &req->sr_msg;
4563 struct io_buffer *kbuf;
4565 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4570 req->flags |= REQ_F_BUFFER_SELECTED;
4574 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4576 return io_put_kbuf(req, req->sr_msg.kbuf);
4579 static int io_recvmsg_prep_async(struct io_kiocb *req)
4583 ret = io_recvmsg_copy_hdr(req, req->async_data);
4585 req->flags |= REQ_F_NEED_CLEANUP;
4589 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4591 struct io_sr_msg *sr = &req->sr_msg;
4593 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4596 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4597 sr->len = READ_ONCE(sqe->len);
4598 sr->bgid = READ_ONCE(sqe->buf_group);
4599 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4600 if (sr->msg_flags & MSG_DONTWAIT)
4601 req->flags |= REQ_F_NOWAIT;
4603 #ifdef CONFIG_COMPAT
4604 if (req->ctx->compat)
4605 sr->msg_flags |= MSG_CMSG_COMPAT;
4610 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4612 struct io_async_msghdr iomsg, *kmsg;
4613 struct socket *sock;
4614 struct io_buffer *kbuf;
4617 int ret, cflags = 0;
4618 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4620 sock = sock_from_file(req->file);
4621 if (unlikely(!sock))
4624 kmsg = req->async_data;
4626 ret = io_recvmsg_copy_hdr(req, &iomsg);
4632 if (req->flags & REQ_F_BUFFER_SELECT) {
4633 kbuf = io_recv_buffer_select(req, !force_nonblock);
4635 return PTR_ERR(kbuf);
4636 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4637 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4638 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4639 1, req->sr_msg.len);
4642 flags = req->sr_msg.msg_flags;
4644 flags |= MSG_DONTWAIT;
4645 if (flags & MSG_WAITALL)
4646 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4648 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4649 kmsg->uaddr, flags);
4650 if (force_nonblock && ret == -EAGAIN)
4651 return io_setup_async_msg(req, kmsg);
4652 if (ret == -ERESTARTSYS)
4655 if (req->flags & REQ_F_BUFFER_SELECTED)
4656 cflags = io_put_recv_kbuf(req);
4657 /* fast path, check for non-NULL to avoid function call */
4659 kfree(kmsg->free_iov);
4660 req->flags &= ~REQ_F_NEED_CLEANUP;
4661 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4663 __io_req_complete(req, issue_flags, ret, cflags);
4667 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4669 struct io_buffer *kbuf;
4670 struct io_sr_msg *sr = &req->sr_msg;
4672 void __user *buf = sr->buf;
4673 struct socket *sock;
4677 int ret, cflags = 0;
4678 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4680 sock = sock_from_file(req->file);
4681 if (unlikely(!sock))
4684 if (req->flags & REQ_F_BUFFER_SELECT) {
4685 kbuf = io_recv_buffer_select(req, !force_nonblock);
4687 return PTR_ERR(kbuf);
4688 buf = u64_to_user_ptr(kbuf->addr);
4691 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4695 msg.msg_name = NULL;
4696 msg.msg_control = NULL;
4697 msg.msg_controllen = 0;
4698 msg.msg_namelen = 0;
4699 msg.msg_iocb = NULL;
4702 flags = req->sr_msg.msg_flags;
4704 flags |= MSG_DONTWAIT;
4705 if (flags & MSG_WAITALL)
4706 min_ret = iov_iter_count(&msg.msg_iter);
4708 ret = sock_recvmsg(sock, &msg, flags);
4709 if (force_nonblock && ret == -EAGAIN)
4711 if (ret == -ERESTARTSYS)
4714 if (req->flags & REQ_F_BUFFER_SELECTED)
4715 cflags = io_put_recv_kbuf(req);
4716 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4718 __io_req_complete(req, issue_flags, ret, cflags);
4722 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4724 struct io_accept *accept = &req->accept;
4726 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4728 if (sqe->ioprio || sqe->len || sqe->buf_index)
4731 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4732 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4733 accept->flags = READ_ONCE(sqe->accept_flags);
4734 accept->nofile = rlimit(RLIMIT_NOFILE);
4738 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4740 struct io_accept *accept = &req->accept;
4741 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4742 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4745 if (req->file->f_flags & O_NONBLOCK)
4746 req->flags |= REQ_F_NOWAIT;
4748 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4749 accept->addr_len, accept->flags,
4751 if (ret == -EAGAIN && force_nonblock)
4754 if (ret == -ERESTARTSYS)
4758 __io_req_complete(req, issue_flags, ret, 0);
4762 static int io_connect_prep_async(struct io_kiocb *req)
4764 struct io_async_connect *io = req->async_data;
4765 struct io_connect *conn = &req->connect;
4767 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4770 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4772 struct io_connect *conn = &req->connect;
4774 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4776 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4779 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4780 conn->addr_len = READ_ONCE(sqe->addr2);
4784 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4786 struct io_async_connect __io, *io;
4787 unsigned file_flags;
4789 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4791 if (req->async_data) {
4792 io = req->async_data;
4794 ret = move_addr_to_kernel(req->connect.addr,
4795 req->connect.addr_len,
4802 file_flags = force_nonblock ? O_NONBLOCK : 0;
4804 ret = __sys_connect_file(req->file, &io->address,
4805 req->connect.addr_len, file_flags);
4806 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4807 if (req->async_data)
4809 if (io_alloc_async_data(req)) {
4813 memcpy(req->async_data, &__io, sizeof(__io));
4816 if (ret == -ERESTARTSYS)
4821 __io_req_complete(req, issue_flags, ret, 0);
4824 #else /* !CONFIG_NET */
4825 #define IO_NETOP_FN(op) \
4826 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4828 return -EOPNOTSUPP; \
4831 #define IO_NETOP_PREP(op) \
4833 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4835 return -EOPNOTSUPP; \
4838 #define IO_NETOP_PREP_ASYNC(op) \
4840 static int io_##op##_prep_async(struct io_kiocb *req) \
4842 return -EOPNOTSUPP; \
4845 IO_NETOP_PREP_ASYNC(sendmsg);
4846 IO_NETOP_PREP_ASYNC(recvmsg);
4847 IO_NETOP_PREP_ASYNC(connect);
4848 IO_NETOP_PREP(accept);
4851 #endif /* CONFIG_NET */
4853 struct io_poll_table {
4854 struct poll_table_struct pt;
4855 struct io_kiocb *req;
4860 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4861 __poll_t mask, io_req_tw_func_t func)
4863 /* for instances that support it check for an event match first: */
4864 if (mask && !(mask & poll->events))
4867 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4869 list_del_init(&poll->wait.entry);
4872 req->io_task_work.func = func;
4875 * If this fails, then the task is exiting. When a task exits, the
4876 * work gets canceled, so just cancel this request as well instead
4877 * of executing it. We can't safely execute it anyway, as we may not
4878 * have the needed state needed for it anyway.
4880 io_req_task_work_add(req);
4884 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4885 __acquires(&req->ctx->completion_lock)
4887 struct io_ring_ctx *ctx = req->ctx;
4889 if (unlikely(req->task->flags & PF_EXITING))
4890 WRITE_ONCE(poll->canceled, true);
4892 if (!req->result && !READ_ONCE(poll->canceled)) {
4893 struct poll_table_struct pt = { ._key = poll->events };
4895 req->result = vfs_poll(req->file, &pt) & poll->events;
4898 spin_lock(&ctx->completion_lock);
4899 if (!req->result && !READ_ONCE(poll->canceled)) {
4900 add_wait_queue(poll->head, &poll->wait);
4907 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4909 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4910 if (req->opcode == IORING_OP_POLL_ADD)
4911 return req->async_data;
4912 return req->apoll->double_poll;
4915 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4917 if (req->opcode == IORING_OP_POLL_ADD)
4919 return &req->apoll->poll;
4922 static void io_poll_remove_double(struct io_kiocb *req)
4923 __must_hold(&req->ctx->completion_lock)
4925 struct io_poll_iocb *poll = io_poll_get_double(req);
4927 lockdep_assert_held(&req->ctx->completion_lock);
4929 if (poll && poll->head) {
4930 struct wait_queue_head *head = poll->head;
4932 spin_lock_irq(&head->lock);
4933 list_del_init(&poll->wait.entry);
4934 if (poll->wait.private)
4937 spin_unlock_irq(&head->lock);
4941 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4942 __must_hold(&req->ctx->completion_lock)
4944 struct io_ring_ctx *ctx = req->ctx;
4945 unsigned flags = IORING_CQE_F_MORE;
4948 if (READ_ONCE(req->poll.canceled)) {
4950 req->poll.events |= EPOLLONESHOT;
4952 error = mangle_poll(mask);
4954 if (req->poll.events & EPOLLONESHOT)
4956 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4957 req->poll.done = true;
4960 if (flags & IORING_CQE_F_MORE)
4963 io_commit_cqring(ctx);
4964 return !(flags & IORING_CQE_F_MORE);
4967 static void io_poll_task_func(struct io_kiocb *req)
4969 struct io_ring_ctx *ctx = req->ctx;
4970 struct io_kiocb *nxt;
4972 if (io_poll_rewait(req, &req->poll)) {
4973 spin_unlock(&ctx->completion_lock);
4977 done = io_poll_complete(req, req->result);
4979 io_poll_remove_double(req);
4980 hash_del(&req->hash_node);
4983 add_wait_queue(req->poll.head, &req->poll.wait);
4985 spin_unlock(&ctx->completion_lock);
4986 io_cqring_ev_posted(ctx);
4989 nxt = io_put_req_find_next(req);
4991 io_req_task_submit(nxt);
4996 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4997 int sync, void *key)
4999 struct io_kiocb *req = wait->private;
5000 struct io_poll_iocb *poll = io_poll_get_single(req);
5001 __poll_t mask = key_to_poll(key);
5002 unsigned long flags;
5004 /* for instances that support it check for an event match first: */
5005 if (mask && !(mask & poll->events))
5007 if (!(poll->events & EPOLLONESHOT))
5008 return poll->wait.func(&poll->wait, mode, sync, key);
5010 list_del_init(&wait->entry);
5015 spin_lock_irqsave(&poll->head->lock, flags);
5016 done = list_empty(&poll->wait.entry);
5018 list_del_init(&poll->wait.entry);
5019 /* make sure double remove sees this as being gone */
5020 wait->private = NULL;
5021 spin_unlock_irqrestore(&poll->head->lock, flags);
5023 /* use wait func handler, so it matches the rq type */
5024 poll->wait.func(&poll->wait, mode, sync, key);
5031 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5032 wait_queue_func_t wake_func)
5036 poll->canceled = false;
5037 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5038 /* mask in events that we always want/need */
5039 poll->events = events | IO_POLL_UNMASK;
5040 INIT_LIST_HEAD(&poll->wait.entry);
5041 init_waitqueue_func_entry(&poll->wait, wake_func);
5044 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5045 struct wait_queue_head *head,
5046 struct io_poll_iocb **poll_ptr)
5048 struct io_kiocb *req = pt->req;
5051 * The file being polled uses multiple waitqueues for poll handling
5052 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5055 if (unlikely(pt->nr_entries)) {
5056 struct io_poll_iocb *poll_one = poll;
5058 /* already have a 2nd entry, fail a third attempt */
5060 pt->error = -EINVAL;
5064 * Can't handle multishot for double wait for now, turn it
5065 * into one-shot mode.
5067 if (!(poll_one->events & EPOLLONESHOT))
5068 poll_one->events |= EPOLLONESHOT;
5069 /* double add on the same waitqueue head, ignore */
5070 if (poll_one->head == head)
5072 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5074 pt->error = -ENOMEM;
5077 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5079 poll->wait.private = req;
5086 if (poll->events & EPOLLEXCLUSIVE)
5087 add_wait_queue_exclusive(head, &poll->wait);
5089 add_wait_queue(head, &poll->wait);
5092 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5093 struct poll_table_struct *p)
5095 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5096 struct async_poll *apoll = pt->req->apoll;
5098 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5101 static void io_async_task_func(struct io_kiocb *req)
5103 struct async_poll *apoll = req->apoll;
5104 struct io_ring_ctx *ctx = req->ctx;
5106 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5108 if (io_poll_rewait(req, &apoll->poll)) {
5109 spin_unlock(&ctx->completion_lock);
5113 hash_del(&req->hash_node);
5114 io_poll_remove_double(req);
5115 spin_unlock(&ctx->completion_lock);
5117 if (!READ_ONCE(apoll->poll.canceled))
5118 io_req_task_submit(req);
5120 io_req_complete_failed(req, -ECANCELED);
5123 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5126 struct io_kiocb *req = wait->private;
5127 struct io_poll_iocb *poll = &req->apoll->poll;
5129 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5132 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5135 static void io_poll_req_insert(struct io_kiocb *req)
5137 struct io_ring_ctx *ctx = req->ctx;
5138 struct hlist_head *list;
5140 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5141 hlist_add_head(&req->hash_node, list);
5144 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5145 struct io_poll_iocb *poll,
5146 struct io_poll_table *ipt, __poll_t mask,
5147 wait_queue_func_t wake_func)
5148 __acquires(&ctx->completion_lock)
5150 struct io_ring_ctx *ctx = req->ctx;
5151 bool cancel = false;
5153 INIT_HLIST_NODE(&req->hash_node);
5154 io_init_poll_iocb(poll, mask, wake_func);
5155 poll->file = req->file;
5156 poll->wait.private = req;
5158 ipt->pt._key = mask;
5161 ipt->nr_entries = 0;
5163 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5164 if (unlikely(!ipt->nr_entries) && !ipt->error)
5165 ipt->error = -EINVAL;
5167 spin_lock(&ctx->completion_lock);
5168 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5169 io_poll_remove_double(req);
5170 if (likely(poll->head)) {
5171 spin_lock_irq(&poll->head->lock);
5172 if (unlikely(list_empty(&poll->wait.entry))) {
5178 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5179 list_del_init(&poll->wait.entry);
5181 WRITE_ONCE(poll->canceled, true);
5182 else if (!poll->done) /* actually waiting for an event */
5183 io_poll_req_insert(req);
5184 spin_unlock_irq(&poll->head->lock);
5196 static int io_arm_poll_handler(struct io_kiocb *req)
5198 const struct io_op_def *def = &io_op_defs[req->opcode];
5199 struct io_ring_ctx *ctx = req->ctx;
5200 struct async_poll *apoll;
5201 struct io_poll_table ipt;
5202 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5205 if (!req->file || !file_can_poll(req->file))
5206 return IO_APOLL_ABORTED;
5207 if (req->flags & REQ_F_POLLED)
5208 return IO_APOLL_ABORTED;
5209 if (!def->pollin && !def->pollout)
5210 return IO_APOLL_ABORTED;
5214 mask |= POLLIN | POLLRDNORM;
5216 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5217 if ((req->opcode == IORING_OP_RECVMSG) &&
5218 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5222 mask |= POLLOUT | POLLWRNORM;
5225 /* if we can't nonblock try, then no point in arming a poll handler */
5226 if (!io_file_supports_nowait(req, rw))
5227 return IO_APOLL_ABORTED;
5229 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5230 if (unlikely(!apoll))
5231 return IO_APOLL_ABORTED;
5232 apoll->double_poll = NULL;
5234 req->flags |= REQ_F_POLLED;
5235 ipt.pt._qproc = io_async_queue_proc;
5236 io_req_refcount(req);
5238 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5240 spin_unlock(&ctx->completion_lock);
5241 if (ret || ipt.error)
5242 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5244 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5245 mask, apoll->poll.events);
5249 static bool __io_poll_remove_one(struct io_kiocb *req,
5250 struct io_poll_iocb *poll, bool do_cancel)
5251 __must_hold(&req->ctx->completion_lock)
5253 bool do_complete = false;
5257 spin_lock_irq(&poll->head->lock);
5259 WRITE_ONCE(poll->canceled, true);
5260 if (!list_empty(&poll->wait.entry)) {
5261 list_del_init(&poll->wait.entry);
5264 spin_unlock_irq(&poll->head->lock);
5265 hash_del(&req->hash_node);
5269 static bool io_poll_remove_one(struct io_kiocb *req)
5270 __must_hold(&req->ctx->completion_lock)
5274 io_poll_remove_double(req);
5275 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5278 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5279 io_commit_cqring(req->ctx);
5281 io_put_req_deferred(req);
5287 * Returns true if we found and killed one or more poll requests
5289 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5292 struct hlist_node *tmp;
5293 struct io_kiocb *req;
5296 spin_lock(&ctx->completion_lock);
5297 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298 struct hlist_head *list;
5300 list = &ctx->cancel_hash[i];
5301 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302 if (io_match_task(req, tsk, cancel_all))
5303 posted += io_poll_remove_one(req);
5306 spin_unlock(&ctx->completion_lock);
5309 io_cqring_ev_posted(ctx);
5314 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5316 __must_hold(&ctx->completion_lock)
5318 struct hlist_head *list;
5319 struct io_kiocb *req;
5321 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322 hlist_for_each_entry(req, list, hash_node) {
5323 if (sqe_addr != req->user_data)
5325 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5332 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5334 __must_hold(&ctx->completion_lock)
5336 struct io_kiocb *req;
5338 req = io_poll_find(ctx, sqe_addr, poll_only);
5341 if (io_poll_remove_one(req))
5347 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5352 events = READ_ONCE(sqe->poll32_events);
5354 events = swahw32(events);
5356 if (!(flags & IORING_POLL_ADD_MULTI))
5357 events |= EPOLLONESHOT;
5358 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5361 static int io_poll_update_prep(struct io_kiocb *req,
5362 const struct io_uring_sqe *sqe)
5364 struct io_poll_update *upd = &req->poll_update;
5367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5369 if (sqe->ioprio || sqe->buf_index)
5371 flags = READ_ONCE(sqe->len);
5372 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373 IORING_POLL_ADD_MULTI))
5375 /* meaningless without update */
5376 if (flags == IORING_POLL_ADD_MULTI)
5379 upd->old_user_data = READ_ONCE(sqe->addr);
5380 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5383 upd->new_user_data = READ_ONCE(sqe->off);
5384 if (!upd->update_user_data && upd->new_user_data)
5386 if (upd->update_events)
5387 upd->events = io_poll_parse_events(sqe, flags);
5388 else if (sqe->poll32_events)
5394 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5397 struct io_kiocb *req = wait->private;
5398 struct io_poll_iocb *poll = &req->poll;
5400 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5403 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404 struct poll_table_struct *p)
5406 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5408 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5411 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5413 struct io_poll_iocb *poll = &req->poll;
5416 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5418 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5420 flags = READ_ONCE(sqe->len);
5421 if (flags & ~IORING_POLL_ADD_MULTI)
5424 io_req_refcount(req);
5425 poll->events = io_poll_parse_events(sqe, flags);
5429 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5431 struct io_poll_iocb *poll = &req->poll;
5432 struct io_ring_ctx *ctx = req->ctx;
5433 struct io_poll_table ipt;
5436 ipt.pt._qproc = io_poll_queue_proc;
5438 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5441 if (mask) { /* no async, we'd stolen it */
5443 io_poll_complete(req, mask);
5445 spin_unlock(&ctx->completion_lock);
5448 io_cqring_ev_posted(ctx);
5449 if (poll->events & EPOLLONESHOT)
5455 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5457 struct io_ring_ctx *ctx = req->ctx;
5458 struct io_kiocb *preq;
5462 spin_lock(&ctx->completion_lock);
5463 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5469 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5471 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5476 * Don't allow racy completion with singleshot, as we cannot safely
5477 * update those. For multishot, if we're racing with completion, just
5478 * let completion re-add it.
5480 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5481 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5485 /* we now have a detached poll request. reissue. */
5489 spin_unlock(&ctx->completion_lock);
5491 io_req_complete(req, ret);
5494 /* only mask one event flags, keep behavior flags */
5495 if (req->poll_update.update_events) {
5496 preq->poll.events &= ~0xffff;
5497 preq->poll.events |= req->poll_update.events & 0xffff;
5498 preq->poll.events |= IO_POLL_UNMASK;
5500 if (req->poll_update.update_user_data)
5501 preq->user_data = req->poll_update.new_user_data;
5502 spin_unlock(&ctx->completion_lock);
5504 /* complete update request, we're done with it */
5505 io_req_complete(req, ret);
5508 ret = io_poll_add(preq, issue_flags);
5511 io_req_complete(preq, ret);
5517 static void io_req_task_timeout(struct io_kiocb *req)
5519 struct io_ring_ctx *ctx = req->ctx;
5521 spin_lock(&ctx->completion_lock);
5522 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5523 io_commit_cqring(ctx);
5524 spin_unlock(&ctx->completion_lock);
5526 io_cqring_ev_posted(ctx);
5531 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5533 struct io_timeout_data *data = container_of(timer,
5534 struct io_timeout_data, timer);
5535 struct io_kiocb *req = data->req;
5536 struct io_ring_ctx *ctx = req->ctx;
5537 unsigned long flags;
5539 spin_lock_irqsave(&ctx->timeout_lock, flags);
5540 list_del_init(&req->timeout.list);
5541 atomic_set(&req->ctx->cq_timeouts,
5542 atomic_read(&req->ctx->cq_timeouts) + 1);
5543 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5545 req->io_task_work.func = io_req_task_timeout;
5546 io_req_task_work_add(req);
5547 return HRTIMER_NORESTART;
5550 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5552 __must_hold(&ctx->timeout_lock)
5554 struct io_timeout_data *io;
5555 struct io_kiocb *req;
5558 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5559 found = user_data == req->user_data;
5564 return ERR_PTR(-ENOENT);
5566 io = req->async_data;
5567 if (hrtimer_try_to_cancel(&io->timer) == -1)
5568 return ERR_PTR(-EALREADY);
5569 list_del_init(&req->timeout.list);
5573 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5574 __must_hold(&ctx->timeout_lock)
5576 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5579 return PTR_ERR(req);
5582 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5583 io_put_req_deferred(req);
5587 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5588 struct timespec64 *ts, enum hrtimer_mode mode)
5589 __must_hold(&ctx->timeout_lock)
5591 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5592 struct io_timeout_data *data;
5595 return PTR_ERR(req);
5597 req->timeout.off = 0; /* noseq */
5598 data = req->async_data;
5599 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5600 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5601 data->timer.function = io_timeout_fn;
5602 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5606 static int io_timeout_remove_prep(struct io_kiocb *req,
5607 const struct io_uring_sqe *sqe)
5609 struct io_timeout_rem *tr = &req->timeout_rem;
5611 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5613 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5615 if (sqe->ioprio || sqe->buf_index || sqe->len)
5618 tr->addr = READ_ONCE(sqe->addr);
5619 tr->flags = READ_ONCE(sqe->timeout_flags);
5620 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5621 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5623 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5625 } else if (tr->flags) {
5626 /* timeout removal doesn't support flags */
5633 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5635 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5640 * Remove or update an existing timeout command
5642 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5644 struct io_timeout_rem *tr = &req->timeout_rem;
5645 struct io_ring_ctx *ctx = req->ctx;
5648 spin_lock_irq(&ctx->timeout_lock);
5649 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5650 ret = io_timeout_cancel(ctx, tr->addr);
5652 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5653 io_translate_timeout_mode(tr->flags));
5654 spin_unlock_irq(&ctx->timeout_lock);
5656 spin_lock(&ctx->completion_lock);
5657 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5658 io_commit_cqring(ctx);
5659 spin_unlock(&ctx->completion_lock);
5660 io_cqring_ev_posted(ctx);
5667 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5668 bool is_timeout_link)
5670 struct io_timeout_data *data;
5672 u32 off = READ_ONCE(sqe->off);
5674 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5676 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5678 if (off && is_timeout_link)
5680 flags = READ_ONCE(sqe->timeout_flags);
5681 if (flags & ~IORING_TIMEOUT_ABS)
5684 req->timeout.off = off;
5685 if (unlikely(off && !req->ctx->off_timeout_used))
5686 req->ctx->off_timeout_used = true;
5688 if (!req->async_data && io_alloc_async_data(req))
5691 data = req->async_data;
5694 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5697 data->mode = io_translate_timeout_mode(flags);
5698 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5699 if (is_timeout_link)
5700 io_req_track_inflight(req);
5704 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5706 struct io_ring_ctx *ctx = req->ctx;
5707 struct io_timeout_data *data = req->async_data;
5708 struct list_head *entry;
5709 u32 tail, off = req->timeout.off;
5711 spin_lock_irq(&ctx->timeout_lock);
5714 * sqe->off holds how many events that need to occur for this
5715 * timeout event to be satisfied. If it isn't set, then this is
5716 * a pure timeout request, sequence isn't used.
5718 if (io_is_timeout_noseq(req)) {
5719 entry = ctx->timeout_list.prev;
5723 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5724 req->timeout.target_seq = tail + off;
5726 /* Update the last seq here in case io_flush_timeouts() hasn't.
5727 * This is safe because ->completion_lock is held, and submissions
5728 * and completions are never mixed in the same ->completion_lock section.
5730 ctx->cq_last_tm_flush = tail;
5733 * Insertion sort, ensuring the first entry in the list is always
5734 * the one we need first.
5736 list_for_each_prev(entry, &ctx->timeout_list) {
5737 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5740 if (io_is_timeout_noseq(nxt))
5742 /* nxt.seq is behind @tail, otherwise would've been completed */
5743 if (off >= nxt->timeout.target_seq - tail)
5747 list_add(&req->timeout.list, entry);
5748 data->timer.function = io_timeout_fn;
5749 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5750 spin_unlock_irq(&ctx->timeout_lock);
5754 struct io_cancel_data {
5755 struct io_ring_ctx *ctx;
5759 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5761 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5762 struct io_cancel_data *cd = data;
5764 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5767 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5768 struct io_ring_ctx *ctx)
5770 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5771 enum io_wq_cancel cancel_ret;
5774 if (!tctx || !tctx->io_wq)
5777 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5778 switch (cancel_ret) {
5779 case IO_WQ_CANCEL_OK:
5782 case IO_WQ_CANCEL_RUNNING:
5785 case IO_WQ_CANCEL_NOTFOUND:
5793 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5794 struct io_kiocb *req, __u64 sqe_addr,
5799 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5800 spin_lock(&ctx->completion_lock);
5803 spin_lock_irq(&ctx->timeout_lock);
5804 ret = io_timeout_cancel(ctx, sqe_addr);
5805 spin_unlock_irq(&ctx->timeout_lock);
5808 ret = io_poll_cancel(ctx, sqe_addr, false);
5812 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5813 io_commit_cqring(ctx);
5814 spin_unlock(&ctx->completion_lock);
5815 io_cqring_ev_posted(ctx);
5821 static int io_async_cancel_prep(struct io_kiocb *req,
5822 const struct io_uring_sqe *sqe)
5824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5826 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5828 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5831 req->cancel.addr = READ_ONCE(sqe->addr);
5835 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5837 struct io_ring_ctx *ctx = req->ctx;
5838 u64 sqe_addr = req->cancel.addr;
5839 struct io_tctx_node *node;
5842 /* tasks should wait for their io-wq threads, so safe w/o sync */
5843 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5844 spin_lock(&ctx->completion_lock);
5847 spin_lock_irq(&ctx->timeout_lock);
5848 ret = io_timeout_cancel(ctx, sqe_addr);
5849 spin_unlock_irq(&ctx->timeout_lock);
5852 ret = io_poll_cancel(ctx, sqe_addr, false);
5855 spin_unlock(&ctx->completion_lock);
5857 /* slow path, try all io-wq's */
5858 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5860 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5861 struct io_uring_task *tctx = node->task->io_uring;
5863 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5867 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5869 spin_lock(&ctx->completion_lock);
5871 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5872 io_commit_cqring(ctx);
5873 spin_unlock(&ctx->completion_lock);
5874 io_cqring_ev_posted(ctx);
5882 static int io_rsrc_update_prep(struct io_kiocb *req,
5883 const struct io_uring_sqe *sqe)
5885 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5887 if (sqe->ioprio || sqe->rw_flags)
5890 req->rsrc_update.offset = READ_ONCE(sqe->off);
5891 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5892 if (!req->rsrc_update.nr_args)
5894 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5898 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5900 struct io_ring_ctx *ctx = req->ctx;
5901 struct io_uring_rsrc_update2 up;
5904 if (issue_flags & IO_URING_F_NONBLOCK)
5907 up.offset = req->rsrc_update.offset;
5908 up.data = req->rsrc_update.arg;
5913 mutex_lock(&ctx->uring_lock);
5914 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5915 &up, req->rsrc_update.nr_args);
5916 mutex_unlock(&ctx->uring_lock);
5920 __io_req_complete(req, issue_flags, ret, 0);
5924 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5926 switch (req->opcode) {
5929 case IORING_OP_READV:
5930 case IORING_OP_READ_FIXED:
5931 case IORING_OP_READ:
5932 return io_read_prep(req, sqe);
5933 case IORING_OP_WRITEV:
5934 case IORING_OP_WRITE_FIXED:
5935 case IORING_OP_WRITE:
5936 return io_write_prep(req, sqe);
5937 case IORING_OP_POLL_ADD:
5938 return io_poll_add_prep(req, sqe);
5939 case IORING_OP_POLL_REMOVE:
5940 return io_poll_update_prep(req, sqe);
5941 case IORING_OP_FSYNC:
5942 return io_fsync_prep(req, sqe);
5943 case IORING_OP_SYNC_FILE_RANGE:
5944 return io_sfr_prep(req, sqe);
5945 case IORING_OP_SENDMSG:
5946 case IORING_OP_SEND:
5947 return io_sendmsg_prep(req, sqe);
5948 case IORING_OP_RECVMSG:
5949 case IORING_OP_RECV:
5950 return io_recvmsg_prep(req, sqe);
5951 case IORING_OP_CONNECT:
5952 return io_connect_prep(req, sqe);
5953 case IORING_OP_TIMEOUT:
5954 return io_timeout_prep(req, sqe, false);
5955 case IORING_OP_TIMEOUT_REMOVE:
5956 return io_timeout_remove_prep(req, sqe);
5957 case IORING_OP_ASYNC_CANCEL:
5958 return io_async_cancel_prep(req, sqe);
5959 case IORING_OP_LINK_TIMEOUT:
5960 return io_timeout_prep(req, sqe, true);
5961 case IORING_OP_ACCEPT:
5962 return io_accept_prep(req, sqe);
5963 case IORING_OP_FALLOCATE:
5964 return io_fallocate_prep(req, sqe);
5965 case IORING_OP_OPENAT:
5966 return io_openat_prep(req, sqe);
5967 case IORING_OP_CLOSE:
5968 return io_close_prep(req, sqe);
5969 case IORING_OP_FILES_UPDATE:
5970 return io_rsrc_update_prep(req, sqe);
5971 case IORING_OP_STATX:
5972 return io_statx_prep(req, sqe);
5973 case IORING_OP_FADVISE:
5974 return io_fadvise_prep(req, sqe);
5975 case IORING_OP_MADVISE:
5976 return io_madvise_prep(req, sqe);
5977 case IORING_OP_OPENAT2:
5978 return io_openat2_prep(req, sqe);
5979 case IORING_OP_EPOLL_CTL:
5980 return io_epoll_ctl_prep(req, sqe);
5981 case IORING_OP_SPLICE:
5982 return io_splice_prep(req, sqe);
5983 case IORING_OP_PROVIDE_BUFFERS:
5984 return io_provide_buffers_prep(req, sqe);
5985 case IORING_OP_REMOVE_BUFFERS:
5986 return io_remove_buffers_prep(req, sqe);
5988 return io_tee_prep(req, sqe);
5989 case IORING_OP_SHUTDOWN:
5990 return io_shutdown_prep(req, sqe);
5991 case IORING_OP_RENAMEAT:
5992 return io_renameat_prep(req, sqe);
5993 case IORING_OP_UNLINKAT:
5994 return io_unlinkat_prep(req, sqe);
5997 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6002 static int io_req_prep_async(struct io_kiocb *req)
6004 if (!io_op_defs[req->opcode].needs_async_setup)
6006 if (WARN_ON_ONCE(req->async_data))
6008 if (io_alloc_async_data(req))
6011 switch (req->opcode) {
6012 case IORING_OP_READV:
6013 return io_rw_prep_async(req, READ);
6014 case IORING_OP_WRITEV:
6015 return io_rw_prep_async(req, WRITE);
6016 case IORING_OP_SENDMSG:
6017 return io_sendmsg_prep_async(req);
6018 case IORING_OP_RECVMSG:
6019 return io_recvmsg_prep_async(req);
6020 case IORING_OP_CONNECT:
6021 return io_connect_prep_async(req);
6023 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6028 static u32 io_get_sequence(struct io_kiocb *req)
6030 u32 seq = req->ctx->cached_sq_head;
6032 /* need original cached_sq_head, but it was increased for each req */
6033 io_for_each_link(req, req)
6038 static bool io_drain_req(struct io_kiocb *req)
6040 struct io_kiocb *pos;
6041 struct io_ring_ctx *ctx = req->ctx;
6042 struct io_defer_entry *de;
6047 * If we need to drain a request in the middle of a link, drain the
6048 * head request and the next request/link after the current link.
6049 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6050 * maintained for every request of our link.
6052 if (ctx->drain_next) {
6053 req->flags |= REQ_F_IO_DRAIN;
6054 ctx->drain_next = false;
6056 /* not interested in head, start from the first linked */
6057 io_for_each_link(pos, req->link) {
6058 if (pos->flags & REQ_F_IO_DRAIN) {
6059 ctx->drain_next = true;
6060 req->flags |= REQ_F_IO_DRAIN;
6065 /* Still need defer if there is pending req in defer list. */
6066 if (likely(list_empty_careful(&ctx->defer_list) &&
6067 !(req->flags & REQ_F_IO_DRAIN))) {
6068 ctx->drain_active = false;
6072 seq = io_get_sequence(req);
6073 /* Still a chance to pass the sequence check */
6074 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6077 ret = io_req_prep_async(req);
6080 io_prep_async_link(req);
6081 de = kmalloc(sizeof(*de), GFP_KERNEL);
6085 io_req_complete_failed(req, ret);
6089 spin_lock(&ctx->completion_lock);
6090 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6091 spin_unlock(&ctx->completion_lock);
6093 io_queue_async_work(req);
6097 trace_io_uring_defer(ctx, req, req->user_data);
6100 list_add_tail(&de->list, &ctx->defer_list);
6101 spin_unlock(&ctx->completion_lock);
6105 static void io_clean_op(struct io_kiocb *req)
6107 if (req->flags & REQ_F_BUFFER_SELECTED) {
6108 switch (req->opcode) {
6109 case IORING_OP_READV:
6110 case IORING_OP_READ_FIXED:
6111 case IORING_OP_READ:
6112 kfree((void *)(unsigned long)req->rw.addr);
6114 case IORING_OP_RECVMSG:
6115 case IORING_OP_RECV:
6116 kfree(req->sr_msg.kbuf);
6121 if (req->flags & REQ_F_NEED_CLEANUP) {
6122 switch (req->opcode) {
6123 case IORING_OP_READV:
6124 case IORING_OP_READ_FIXED:
6125 case IORING_OP_READ:
6126 case IORING_OP_WRITEV:
6127 case IORING_OP_WRITE_FIXED:
6128 case IORING_OP_WRITE: {
6129 struct io_async_rw *io = req->async_data;
6131 kfree(io->free_iovec);
6134 case IORING_OP_RECVMSG:
6135 case IORING_OP_SENDMSG: {
6136 struct io_async_msghdr *io = req->async_data;
6138 kfree(io->free_iov);
6141 case IORING_OP_SPLICE:
6143 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6144 io_put_file(req->splice.file_in);
6146 case IORING_OP_OPENAT:
6147 case IORING_OP_OPENAT2:
6148 if (req->open.filename)
6149 putname(req->open.filename);
6151 case IORING_OP_RENAMEAT:
6152 putname(req->rename.oldpath);
6153 putname(req->rename.newpath);
6155 case IORING_OP_UNLINKAT:
6156 putname(req->unlink.filename);
6160 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6161 kfree(req->apoll->double_poll);
6165 if (req->flags & REQ_F_INFLIGHT) {
6166 struct io_uring_task *tctx = req->task->io_uring;
6168 atomic_dec(&tctx->inflight_tracked);
6170 if (req->flags & REQ_F_CREDS)
6171 put_cred(req->creds);
6173 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6176 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6178 struct io_ring_ctx *ctx = req->ctx;
6179 const struct cred *creds = NULL;
6182 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6183 creds = override_creds(req->creds);
6185 switch (req->opcode) {
6187 ret = io_nop(req, issue_flags);
6189 case IORING_OP_READV:
6190 case IORING_OP_READ_FIXED:
6191 case IORING_OP_READ:
6192 ret = io_read(req, issue_flags);
6194 case IORING_OP_WRITEV:
6195 case IORING_OP_WRITE_FIXED:
6196 case IORING_OP_WRITE:
6197 ret = io_write(req, issue_flags);
6199 case IORING_OP_FSYNC:
6200 ret = io_fsync(req, issue_flags);
6202 case IORING_OP_POLL_ADD:
6203 ret = io_poll_add(req, issue_flags);
6205 case IORING_OP_POLL_REMOVE:
6206 ret = io_poll_update(req, issue_flags);
6208 case IORING_OP_SYNC_FILE_RANGE:
6209 ret = io_sync_file_range(req, issue_flags);
6211 case IORING_OP_SENDMSG:
6212 ret = io_sendmsg(req, issue_flags);
6214 case IORING_OP_SEND:
6215 ret = io_send(req, issue_flags);
6217 case IORING_OP_RECVMSG:
6218 ret = io_recvmsg(req, issue_flags);
6220 case IORING_OP_RECV:
6221 ret = io_recv(req, issue_flags);
6223 case IORING_OP_TIMEOUT:
6224 ret = io_timeout(req, issue_flags);
6226 case IORING_OP_TIMEOUT_REMOVE:
6227 ret = io_timeout_remove(req, issue_flags);
6229 case IORING_OP_ACCEPT:
6230 ret = io_accept(req, issue_flags);
6232 case IORING_OP_CONNECT:
6233 ret = io_connect(req, issue_flags);
6235 case IORING_OP_ASYNC_CANCEL:
6236 ret = io_async_cancel(req, issue_flags);
6238 case IORING_OP_FALLOCATE:
6239 ret = io_fallocate(req, issue_flags);
6241 case IORING_OP_OPENAT:
6242 ret = io_openat(req, issue_flags);
6244 case IORING_OP_CLOSE:
6245 ret = io_close(req, issue_flags);
6247 case IORING_OP_FILES_UPDATE:
6248 ret = io_files_update(req, issue_flags);
6250 case IORING_OP_STATX:
6251 ret = io_statx(req, issue_flags);
6253 case IORING_OP_FADVISE:
6254 ret = io_fadvise(req, issue_flags);
6256 case IORING_OP_MADVISE:
6257 ret = io_madvise(req, issue_flags);
6259 case IORING_OP_OPENAT2:
6260 ret = io_openat2(req, issue_flags);
6262 case IORING_OP_EPOLL_CTL:
6263 ret = io_epoll_ctl(req, issue_flags);
6265 case IORING_OP_SPLICE:
6266 ret = io_splice(req, issue_flags);
6268 case IORING_OP_PROVIDE_BUFFERS:
6269 ret = io_provide_buffers(req, issue_flags);
6271 case IORING_OP_REMOVE_BUFFERS:
6272 ret = io_remove_buffers(req, issue_flags);
6275 ret = io_tee(req, issue_flags);
6277 case IORING_OP_SHUTDOWN:
6278 ret = io_shutdown(req, issue_flags);
6280 case IORING_OP_RENAMEAT:
6281 ret = io_renameat(req, issue_flags);
6283 case IORING_OP_UNLINKAT:
6284 ret = io_unlinkat(req, issue_flags);
6292 revert_creds(creds);
6295 /* If the op doesn't have a file, we're not polling for it */
6296 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6297 io_iopoll_req_issued(req);
6302 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6304 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6306 req = io_put_req_find_next(req);
6307 return req ? &req->work : NULL;
6310 static void io_wq_submit_work(struct io_wq_work *work)
6312 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6313 struct io_kiocb *timeout;
6316 io_req_refcount(req);
6317 /* will be dropped by ->io_free_work() after returning to io-wq */
6320 timeout = io_prep_linked_timeout(req);
6322 io_queue_linked_timeout(timeout);
6324 if (work->flags & IO_WQ_WORK_CANCEL)
6329 ret = io_issue_sqe(req, 0);
6331 * We can get EAGAIN for polled IO even though we're
6332 * forcing a sync submission from here, since we can't
6333 * wait for request slots on the block side.
6341 /* avoid locking problems by failing it from a clean context */
6343 io_req_task_queue_fail(req, ret);
6346 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6349 return &table->files[i];
6352 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6355 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6357 return (struct file *) (slot->file_ptr & FFS_MASK);
6360 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6362 unsigned long file_ptr = (unsigned long) file;
6364 if (__io_file_supports_nowait(file, READ))
6365 file_ptr |= FFS_ASYNC_READ;
6366 if (__io_file_supports_nowait(file, WRITE))
6367 file_ptr |= FFS_ASYNC_WRITE;
6368 if (S_ISREG(file_inode(file)->i_mode))
6369 file_ptr |= FFS_ISREG;
6370 file_slot->file_ptr = file_ptr;
6373 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6374 struct io_kiocb *req, int fd)
6377 unsigned long file_ptr;
6379 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6381 fd = array_index_nospec(fd, ctx->nr_user_files);
6382 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6383 file = (struct file *) (file_ptr & FFS_MASK);
6384 file_ptr &= ~FFS_MASK;
6385 /* mask in overlapping REQ_F and FFS bits */
6386 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6387 io_req_set_rsrc_node(req);
6391 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6392 struct io_kiocb *req, int fd)
6394 struct file *file = fget(fd);
6396 trace_io_uring_file_get(ctx, fd);
6398 /* we don't allow fixed io_uring files */
6399 if (file && unlikely(file->f_op == &io_uring_fops))
6400 io_req_track_inflight(req);
6404 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6405 struct io_kiocb *req, int fd, bool fixed)
6408 return io_file_get_fixed(ctx, req, fd);
6410 return io_file_get_normal(ctx, req, fd);
6413 static void io_req_task_link_timeout(struct io_kiocb *req)
6415 struct io_kiocb *prev = req->timeout.prev;
6416 struct io_ring_ctx *ctx = req->ctx;
6419 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6423 io_req_complete_post(req, -ETIME, 0);
6427 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6429 struct io_timeout_data *data = container_of(timer,
6430 struct io_timeout_data, timer);
6431 struct io_kiocb *prev, *req = data->req;
6432 struct io_ring_ctx *ctx = req->ctx;
6433 unsigned long flags;
6435 spin_lock_irqsave(&ctx->timeout_lock, flags);
6436 prev = req->timeout.head;
6437 req->timeout.head = NULL;
6440 * We don't expect the list to be empty, that will only happen if we
6441 * race with the completion of the linked work.
6444 io_remove_next_linked(prev);
6445 if (!req_ref_inc_not_zero(prev))
6448 req->timeout.prev = prev;
6449 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6451 req->io_task_work.func = io_req_task_link_timeout;
6452 io_req_task_work_add(req);
6453 return HRTIMER_NORESTART;
6456 static void io_queue_linked_timeout(struct io_kiocb *req)
6458 struct io_ring_ctx *ctx = req->ctx;
6460 spin_lock_irq(&ctx->timeout_lock);
6462 * If the back reference is NULL, then our linked request finished
6463 * before we got a chance to setup the timer
6465 if (req->timeout.head) {
6466 struct io_timeout_data *data = req->async_data;
6468 data->timer.function = io_link_timeout_fn;
6469 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6472 spin_unlock_irq(&ctx->timeout_lock);
6473 /* drop submission reference */
6477 static void __io_queue_sqe(struct io_kiocb *req)
6478 __must_hold(&req->ctx->uring_lock)
6480 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6484 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6487 * We async punt it if the file wasn't marked NOWAIT, or if the file
6488 * doesn't support non-blocking read/write attempts
6491 if (req->flags & REQ_F_COMPLETE_INLINE) {
6492 struct io_ring_ctx *ctx = req->ctx;
6493 struct io_submit_state *state = &ctx->submit_state;
6495 state->compl_reqs[state->compl_nr++] = req;
6496 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6497 io_submit_flush_completions(ctx);
6499 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6500 switch (io_arm_poll_handler(req)) {
6501 case IO_APOLL_READY:
6503 case IO_APOLL_ABORTED:
6505 * Queued up for async execution, worker will release
6506 * submit reference when the iocb is actually submitted.
6508 io_queue_async_work(req);
6512 io_req_complete_failed(req, ret);
6515 io_queue_linked_timeout(linked_timeout);
6518 static inline void io_queue_sqe(struct io_kiocb *req)
6519 __must_hold(&req->ctx->uring_lock)
6521 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6524 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6525 __io_queue_sqe(req);
6527 int ret = io_req_prep_async(req);
6530 io_req_complete_failed(req, ret);
6532 io_queue_async_work(req);
6537 * Check SQE restrictions (opcode and flags).
6539 * Returns 'true' if SQE is allowed, 'false' otherwise.
6541 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6542 struct io_kiocb *req,
6543 unsigned int sqe_flags)
6545 if (likely(!ctx->restricted))
6548 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6551 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6552 ctx->restrictions.sqe_flags_required)
6555 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6556 ctx->restrictions.sqe_flags_required))
6562 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6563 const struct io_uring_sqe *sqe)
6564 __must_hold(&ctx->uring_lock)
6566 struct io_submit_state *state;
6567 unsigned int sqe_flags;
6568 int personality, ret = 0;
6570 /* req is partially pre-initialised, see io_preinit_req() */
6571 req->opcode = READ_ONCE(sqe->opcode);
6572 /* same numerical values with corresponding REQ_F_*, safe to copy */
6573 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6574 req->user_data = READ_ONCE(sqe->user_data);
6576 req->fixed_rsrc_refs = NULL;
6577 req->task = current;
6579 /* enforce forwards compatibility on users */
6580 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6582 if (unlikely(req->opcode >= IORING_OP_LAST))
6584 if (!io_check_restriction(ctx, req, sqe_flags))
6587 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6588 !io_op_defs[req->opcode].buffer_select)
6590 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6591 ctx->drain_active = true;
6593 personality = READ_ONCE(sqe->personality);
6595 req->creds = xa_load(&ctx->personalities, personality);
6598 get_cred(req->creds);
6599 req->flags |= REQ_F_CREDS;
6601 state = &ctx->submit_state;
6604 * Plug now if we have more than 1 IO left after this, and the target
6605 * is potentially a read/write to block based storage.
6607 if (!state->plug_started && state->ios_left > 1 &&
6608 io_op_defs[req->opcode].plug) {
6609 blk_start_plug(&state->plug);
6610 state->plug_started = true;
6613 if (io_op_defs[req->opcode].needs_file) {
6614 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6615 (sqe_flags & IOSQE_FIXED_FILE));
6616 if (unlikely(!req->file))
6624 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6625 const struct io_uring_sqe *sqe)
6626 __must_hold(&ctx->uring_lock)
6628 struct io_submit_link *link = &ctx->submit_state.link;
6631 ret = io_init_req(ctx, req, sqe);
6632 if (unlikely(ret)) {
6635 /* fail even hard links since we don't submit */
6636 req_set_fail(link->head);
6637 io_req_complete_failed(link->head, -ECANCELED);
6640 io_req_complete_failed(req, ret);
6644 ret = io_req_prep(req, sqe);
6648 /* don't need @sqe from now on */
6649 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6651 ctx->flags & IORING_SETUP_SQPOLL);
6654 * If we already have a head request, queue this one for async
6655 * submittal once the head completes. If we don't have a head but
6656 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6657 * submitted sync once the chain is complete. If none of those
6658 * conditions are true (normal request), then just queue it.
6661 struct io_kiocb *head = link->head;
6663 ret = io_req_prep_async(req);
6666 trace_io_uring_link(ctx, req, head);
6667 link->last->link = req;
6670 /* last request of a link, enqueue the link */
6671 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6676 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6688 * Batched submission is done, ensure local IO is flushed out.
6690 static void io_submit_state_end(struct io_submit_state *state,
6691 struct io_ring_ctx *ctx)
6693 if (state->link.head)
6694 io_queue_sqe(state->link.head);
6695 if (state->compl_nr)
6696 io_submit_flush_completions(ctx);
6697 if (state->plug_started)
6698 blk_finish_plug(&state->plug);
6702 * Start submission side cache.
6704 static void io_submit_state_start(struct io_submit_state *state,
6705 unsigned int max_ios)
6707 state->plug_started = false;
6708 state->ios_left = max_ios;
6709 /* set only head, no need to init link_last in advance */
6710 state->link.head = NULL;
6713 static void io_commit_sqring(struct io_ring_ctx *ctx)
6715 struct io_rings *rings = ctx->rings;
6718 * Ensure any loads from the SQEs are done at this point,
6719 * since once we write the new head, the application could
6720 * write new data to them.
6722 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6726 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6727 * that is mapped by userspace. This means that care needs to be taken to
6728 * ensure that reads are stable, as we cannot rely on userspace always
6729 * being a good citizen. If members of the sqe are validated and then later
6730 * used, it's important that those reads are done through READ_ONCE() to
6731 * prevent a re-load down the line.
6733 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6735 unsigned head, mask = ctx->sq_entries - 1;
6736 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6739 * The cached sq head (or cq tail) serves two purposes:
6741 * 1) allows us to batch the cost of updating the user visible
6743 * 2) allows the kernel side to track the head on its own, even
6744 * though the application is the one updating it.
6746 head = READ_ONCE(ctx->sq_array[sq_idx]);
6747 if (likely(head < ctx->sq_entries))
6748 return &ctx->sq_sqes[head];
6750 /* drop invalid entries */
6752 WRITE_ONCE(ctx->rings->sq_dropped,
6753 READ_ONCE(ctx->rings->sq_dropped) + 1);
6757 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6758 __must_hold(&ctx->uring_lock)
6760 struct io_uring_task *tctx;
6763 /* make sure SQ entry isn't read before tail */
6764 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6765 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6768 tctx = current->io_uring;
6769 tctx->cached_refs -= nr;
6770 if (unlikely(tctx->cached_refs < 0)) {
6771 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6773 percpu_counter_add(&tctx->inflight, refill);
6774 refcount_add(refill, ¤t->usage);
6775 tctx->cached_refs += refill;
6777 io_submit_state_start(&ctx->submit_state, nr);
6779 while (submitted < nr) {
6780 const struct io_uring_sqe *sqe;
6781 struct io_kiocb *req;
6783 req = io_alloc_req(ctx);
6784 if (unlikely(!req)) {
6786 submitted = -EAGAIN;
6789 sqe = io_get_sqe(ctx);
6790 if (unlikely(!sqe)) {
6791 kmem_cache_free(req_cachep, req);
6794 /* will complete beyond this point, count as submitted */
6796 if (io_submit_sqe(ctx, req, sqe))
6800 if (unlikely(submitted != nr)) {
6801 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6802 int unused = nr - ref_used;
6804 current->io_uring->cached_refs += unused;
6805 percpu_ref_put_many(&ctx->refs, unused);
6808 io_submit_state_end(&ctx->submit_state, ctx);
6809 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6810 io_commit_sqring(ctx);
6815 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6817 return READ_ONCE(sqd->state);
6820 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6822 /* Tell userspace we may need a wakeup call */
6823 spin_lock(&ctx->completion_lock);
6824 WRITE_ONCE(ctx->rings->sq_flags,
6825 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6826 spin_unlock(&ctx->completion_lock);
6829 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6831 spin_lock(&ctx->completion_lock);
6832 WRITE_ONCE(ctx->rings->sq_flags,
6833 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6834 spin_unlock(&ctx->completion_lock);
6837 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6839 unsigned int to_submit;
6842 to_submit = io_sqring_entries(ctx);
6843 /* if we're handling multiple rings, cap submit size for fairness */
6844 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6845 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6847 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6848 unsigned nr_events = 0;
6849 const struct cred *creds = NULL;
6851 if (ctx->sq_creds != current_cred())
6852 creds = override_creds(ctx->sq_creds);
6854 mutex_lock(&ctx->uring_lock);
6855 if (!list_empty(&ctx->iopoll_list))
6856 io_do_iopoll(ctx, &nr_events, 0, true);
6859 * Don't submit if refs are dying, good for io_uring_register(),
6860 * but also it is relied upon by io_ring_exit_work()
6862 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6863 !(ctx->flags & IORING_SETUP_R_DISABLED))
6864 ret = io_submit_sqes(ctx, to_submit);
6865 mutex_unlock(&ctx->uring_lock);
6867 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6868 wake_up(&ctx->sqo_sq_wait);
6870 revert_creds(creds);
6876 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6878 struct io_ring_ctx *ctx;
6879 unsigned sq_thread_idle = 0;
6881 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6882 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6883 sqd->sq_thread_idle = sq_thread_idle;
6886 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6888 bool did_sig = false;
6889 struct ksignal ksig;
6891 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6892 signal_pending(current)) {
6893 mutex_unlock(&sqd->lock);
6894 if (signal_pending(current))
6895 did_sig = get_signal(&ksig);
6897 mutex_lock(&sqd->lock);
6899 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6902 static int io_sq_thread(void *data)
6904 struct io_sq_data *sqd = data;
6905 struct io_ring_ctx *ctx;
6906 unsigned long timeout = 0;
6907 char buf[TASK_COMM_LEN];
6910 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6911 set_task_comm(current, buf);
6913 if (sqd->sq_cpu != -1)
6914 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6916 set_cpus_allowed_ptr(current, cpu_online_mask);
6917 current->flags |= PF_NO_SETAFFINITY;
6919 mutex_lock(&sqd->lock);
6921 bool cap_entries, sqt_spin = false;
6923 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6924 if (io_sqd_handle_event(sqd))
6926 timeout = jiffies + sqd->sq_thread_idle;
6929 cap_entries = !list_is_singular(&sqd->ctx_list);
6930 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6931 int ret = __io_sq_thread(ctx, cap_entries);
6933 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6936 if (io_run_task_work())
6939 if (sqt_spin || !time_after(jiffies, timeout)) {
6942 timeout = jiffies + sqd->sq_thread_idle;
6946 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6947 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6948 bool needs_sched = true;
6950 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6951 io_ring_set_wakeup_flag(ctx);
6953 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6954 !list_empty_careful(&ctx->iopoll_list)) {
6955 needs_sched = false;
6958 if (io_sqring_entries(ctx)) {
6959 needs_sched = false;
6965 mutex_unlock(&sqd->lock);
6967 mutex_lock(&sqd->lock);
6969 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6970 io_ring_clear_wakeup_flag(ctx);
6973 finish_wait(&sqd->wait, &wait);
6974 timeout = jiffies + sqd->sq_thread_idle;
6977 io_uring_cancel_generic(true, sqd);
6979 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6980 io_ring_set_wakeup_flag(ctx);
6982 mutex_unlock(&sqd->lock);
6984 complete(&sqd->exited);
6988 struct io_wait_queue {
6989 struct wait_queue_entry wq;
6990 struct io_ring_ctx *ctx;
6992 unsigned nr_timeouts;
6995 static inline bool io_should_wake(struct io_wait_queue *iowq)
6997 struct io_ring_ctx *ctx = iowq->ctx;
6998 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7001 * Wake up if we have enough events, or if a timeout occurred since we
7002 * started waiting. For timeouts, we always want to return to userspace,
7003 * regardless of event count.
7005 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7008 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7009 int wake_flags, void *key)
7011 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7015 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7016 * the task, and the next invocation will do it.
7018 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7019 return autoremove_wake_function(curr, mode, wake_flags, key);
7023 static int io_run_task_work_sig(void)
7025 if (io_run_task_work())
7027 if (!signal_pending(current))
7029 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7030 return -ERESTARTSYS;
7034 /* when returns >0, the caller should retry */
7035 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7036 struct io_wait_queue *iowq,
7037 signed long *timeout)
7041 /* make sure we run task_work before checking for signals */
7042 ret = io_run_task_work_sig();
7043 if (ret || io_should_wake(iowq))
7045 /* let the caller flush overflows, retry */
7046 if (test_bit(0, &ctx->check_cq_overflow))
7049 *timeout = schedule_timeout(*timeout);
7050 return !*timeout ? -ETIME : 1;
7054 * Wait until events become available, if we don't already have some. The
7055 * application must reap them itself, as they reside on the shared cq ring.
7057 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7058 const sigset_t __user *sig, size_t sigsz,
7059 struct __kernel_timespec __user *uts)
7061 struct io_wait_queue iowq;
7062 struct io_rings *rings = ctx->rings;
7063 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7067 io_cqring_overflow_flush(ctx);
7068 if (io_cqring_events(ctx) >= min_events)
7070 if (!io_run_task_work())
7075 #ifdef CONFIG_COMPAT
7076 if (in_compat_syscall())
7077 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7081 ret = set_user_sigmask(sig, sigsz);
7088 struct timespec64 ts;
7090 if (get_timespec64(&ts, uts))
7092 timeout = timespec64_to_jiffies(&ts);
7095 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7096 iowq.wq.private = current;
7097 INIT_LIST_HEAD(&iowq.wq.entry);
7099 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7100 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7102 trace_io_uring_cqring_wait(ctx, min_events);
7104 /* if we can't even flush overflow, don't wait for more */
7105 if (!io_cqring_overflow_flush(ctx)) {
7109 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7110 TASK_INTERRUPTIBLE);
7111 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7112 finish_wait(&ctx->cq_wait, &iowq.wq);
7116 restore_saved_sigmask_unless(ret == -EINTR);
7118 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7121 static void io_free_page_table(void **table, size_t size)
7123 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7125 for (i = 0; i < nr_tables; i++)
7130 static void **io_alloc_page_table(size_t size)
7132 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7133 size_t init_size = size;
7136 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7140 for (i = 0; i < nr_tables; i++) {
7141 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7143 table[i] = kzalloc(this_size, GFP_KERNEL);
7145 io_free_page_table(table, init_size);
7153 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7155 percpu_ref_exit(&ref_node->refs);
7159 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7161 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7162 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7163 unsigned long flags;
7164 bool first_add = false;
7166 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7169 while (!list_empty(&ctx->rsrc_ref_list)) {
7170 node = list_first_entry(&ctx->rsrc_ref_list,
7171 struct io_rsrc_node, node);
7172 /* recycle ref nodes in order */
7175 list_del(&node->node);
7176 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7178 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7181 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7184 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7186 struct io_rsrc_node *ref_node;
7188 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7192 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7197 INIT_LIST_HEAD(&ref_node->node);
7198 INIT_LIST_HEAD(&ref_node->rsrc_list);
7199 ref_node->done = false;
7203 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7204 struct io_rsrc_data *data_to_kill)
7206 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7207 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7210 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7212 rsrc_node->rsrc_data = data_to_kill;
7213 spin_lock_irq(&ctx->rsrc_ref_lock);
7214 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7215 spin_unlock_irq(&ctx->rsrc_ref_lock);
7217 atomic_inc(&data_to_kill->refs);
7218 percpu_ref_kill(&rsrc_node->refs);
7219 ctx->rsrc_node = NULL;
7222 if (!ctx->rsrc_node) {
7223 ctx->rsrc_node = ctx->rsrc_backup_node;
7224 ctx->rsrc_backup_node = NULL;
7228 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7230 if (ctx->rsrc_backup_node)
7232 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7233 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7236 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7240 /* As we may drop ->uring_lock, other task may have started quiesce */
7244 data->quiesce = true;
7246 ret = io_rsrc_node_switch_start(ctx);
7249 io_rsrc_node_switch(ctx, data);
7251 /* kill initial ref, already quiesced if zero */
7252 if (atomic_dec_and_test(&data->refs))
7254 mutex_unlock(&ctx->uring_lock);
7255 flush_delayed_work(&ctx->rsrc_put_work);
7256 ret = wait_for_completion_interruptible(&data->done);
7258 mutex_lock(&ctx->uring_lock);
7262 atomic_inc(&data->refs);
7263 /* wait for all works potentially completing data->done */
7264 flush_delayed_work(&ctx->rsrc_put_work);
7265 reinit_completion(&data->done);
7267 ret = io_run_task_work_sig();
7268 mutex_lock(&ctx->uring_lock);
7270 data->quiesce = false;
7275 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7277 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7278 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7280 return &data->tags[table_idx][off];
7283 static void io_rsrc_data_free(struct io_rsrc_data *data)
7285 size_t size = data->nr * sizeof(data->tags[0][0]);
7288 io_free_page_table((void **)data->tags, size);
7292 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7293 u64 __user *utags, unsigned nr,
7294 struct io_rsrc_data **pdata)
7296 struct io_rsrc_data *data;
7300 data = kzalloc(sizeof(*data), GFP_KERNEL);
7303 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7311 data->do_put = do_put;
7314 for (i = 0; i < nr; i++) {
7315 u64 *tag_slot = io_get_tag_slot(data, i);
7317 if (copy_from_user(tag_slot, &utags[i],
7323 atomic_set(&data->refs, 1);
7324 init_completion(&data->done);
7328 io_rsrc_data_free(data);
7332 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7334 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7335 return !!table->files;
7338 static void io_free_file_tables(struct io_file_table *table)
7340 kvfree(table->files);
7341 table->files = NULL;
7344 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7346 #if defined(CONFIG_UNIX)
7347 if (ctx->ring_sock) {
7348 struct sock *sock = ctx->ring_sock->sk;
7349 struct sk_buff *skb;
7351 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7357 for (i = 0; i < ctx->nr_user_files; i++) {
7360 file = io_file_from_index(ctx, i);
7365 io_free_file_tables(&ctx->file_table);
7366 io_rsrc_data_free(ctx->file_data);
7367 ctx->file_data = NULL;
7368 ctx->nr_user_files = 0;
7371 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7375 if (!ctx->file_data)
7377 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7379 __io_sqe_files_unregister(ctx);
7383 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7384 __releases(&sqd->lock)
7386 WARN_ON_ONCE(sqd->thread == current);
7389 * Do the dance but not conditional clear_bit() because it'd race with
7390 * other threads incrementing park_pending and setting the bit.
7392 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7393 if (atomic_dec_return(&sqd->park_pending))
7394 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7395 mutex_unlock(&sqd->lock);
7398 static void io_sq_thread_park(struct io_sq_data *sqd)
7399 __acquires(&sqd->lock)
7401 WARN_ON_ONCE(sqd->thread == current);
7403 atomic_inc(&sqd->park_pending);
7404 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7405 mutex_lock(&sqd->lock);
7407 wake_up_process(sqd->thread);
7410 static void io_sq_thread_stop(struct io_sq_data *sqd)
7412 WARN_ON_ONCE(sqd->thread == current);
7413 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7415 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7416 mutex_lock(&sqd->lock);
7418 wake_up_process(sqd->thread);
7419 mutex_unlock(&sqd->lock);
7420 wait_for_completion(&sqd->exited);
7423 static void io_put_sq_data(struct io_sq_data *sqd)
7425 if (refcount_dec_and_test(&sqd->refs)) {
7426 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7428 io_sq_thread_stop(sqd);
7433 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7435 struct io_sq_data *sqd = ctx->sq_data;
7438 io_sq_thread_park(sqd);
7439 list_del_init(&ctx->sqd_list);
7440 io_sqd_update_thread_idle(sqd);
7441 io_sq_thread_unpark(sqd);
7443 io_put_sq_data(sqd);
7444 ctx->sq_data = NULL;
7448 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7450 struct io_ring_ctx *ctx_attach;
7451 struct io_sq_data *sqd;
7454 f = fdget(p->wq_fd);
7456 return ERR_PTR(-ENXIO);
7457 if (f.file->f_op != &io_uring_fops) {
7459 return ERR_PTR(-EINVAL);
7462 ctx_attach = f.file->private_data;
7463 sqd = ctx_attach->sq_data;
7466 return ERR_PTR(-EINVAL);
7468 if (sqd->task_tgid != current->tgid) {
7470 return ERR_PTR(-EPERM);
7473 refcount_inc(&sqd->refs);
7478 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7481 struct io_sq_data *sqd;
7484 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7485 sqd = io_attach_sq_data(p);
7490 /* fall through for EPERM case, setup new sqd/task */
7491 if (PTR_ERR(sqd) != -EPERM)
7495 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7497 return ERR_PTR(-ENOMEM);
7499 atomic_set(&sqd->park_pending, 0);
7500 refcount_set(&sqd->refs, 1);
7501 INIT_LIST_HEAD(&sqd->ctx_list);
7502 mutex_init(&sqd->lock);
7503 init_waitqueue_head(&sqd->wait);
7504 init_completion(&sqd->exited);
7508 #if defined(CONFIG_UNIX)
7510 * Ensure the UNIX gc is aware of our file set, so we are certain that
7511 * the io_uring can be safely unregistered on process exit, even if we have
7512 * loops in the file referencing.
7514 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7516 struct sock *sk = ctx->ring_sock->sk;
7517 struct scm_fp_list *fpl;
7518 struct sk_buff *skb;
7521 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7525 skb = alloc_skb(0, GFP_KERNEL);
7534 fpl->user = get_uid(current_user());
7535 for (i = 0; i < nr; i++) {
7536 struct file *file = io_file_from_index(ctx, i + offset);
7540 fpl->fp[nr_files] = get_file(file);
7541 unix_inflight(fpl->user, fpl->fp[nr_files]);
7546 fpl->max = SCM_MAX_FD;
7547 fpl->count = nr_files;
7548 UNIXCB(skb).fp = fpl;
7549 skb->destructor = unix_destruct_scm;
7550 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7551 skb_queue_head(&sk->sk_receive_queue, skb);
7553 for (i = 0; i < nr_files; i++)
7564 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7565 * causes regular reference counting to break down. We rely on the UNIX
7566 * garbage collection to take care of this problem for us.
7568 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7570 unsigned left, total;
7574 left = ctx->nr_user_files;
7576 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7578 ret = __io_sqe_files_scm(ctx, this_files, total);
7582 total += this_files;
7588 while (total < ctx->nr_user_files) {
7589 struct file *file = io_file_from_index(ctx, total);
7599 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7605 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7607 struct file *file = prsrc->file;
7608 #if defined(CONFIG_UNIX)
7609 struct sock *sock = ctx->ring_sock->sk;
7610 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7611 struct sk_buff *skb;
7614 __skb_queue_head_init(&list);
7617 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7618 * remove this entry and rearrange the file array.
7620 skb = skb_dequeue(head);
7622 struct scm_fp_list *fp;
7624 fp = UNIXCB(skb).fp;
7625 for (i = 0; i < fp->count; i++) {
7628 if (fp->fp[i] != file)
7631 unix_notinflight(fp->user, fp->fp[i]);
7632 left = fp->count - 1 - i;
7634 memmove(&fp->fp[i], &fp->fp[i + 1],
7635 left * sizeof(struct file *));
7642 __skb_queue_tail(&list, skb);
7652 __skb_queue_tail(&list, skb);
7654 skb = skb_dequeue(head);
7657 if (skb_peek(&list)) {
7658 spin_lock_irq(&head->lock);
7659 while ((skb = __skb_dequeue(&list)) != NULL)
7660 __skb_queue_tail(head, skb);
7661 spin_unlock_irq(&head->lock);
7668 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7670 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7671 struct io_ring_ctx *ctx = rsrc_data->ctx;
7672 struct io_rsrc_put *prsrc, *tmp;
7674 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7675 list_del(&prsrc->list);
7678 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7680 io_ring_submit_lock(ctx, lock_ring);
7681 spin_lock(&ctx->completion_lock);
7682 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7684 io_commit_cqring(ctx);
7685 spin_unlock(&ctx->completion_lock);
7686 io_cqring_ev_posted(ctx);
7687 io_ring_submit_unlock(ctx, lock_ring);
7690 rsrc_data->do_put(ctx, prsrc);
7694 io_rsrc_node_destroy(ref_node);
7695 if (atomic_dec_and_test(&rsrc_data->refs))
7696 complete(&rsrc_data->done);
7699 static void io_rsrc_put_work(struct work_struct *work)
7701 struct io_ring_ctx *ctx;
7702 struct llist_node *node;
7704 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7705 node = llist_del_all(&ctx->rsrc_put_llist);
7708 struct io_rsrc_node *ref_node;
7709 struct llist_node *next = node->next;
7711 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7712 __io_rsrc_put_work(ref_node);
7717 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7718 unsigned nr_args, u64 __user *tags)
7720 __s32 __user *fds = (__s32 __user *) arg;
7729 if (nr_args > IORING_MAX_FIXED_FILES)
7731 ret = io_rsrc_node_switch_start(ctx);
7734 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7740 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7743 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7744 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7748 /* allow sparse sets */
7751 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7758 if (unlikely(!file))
7762 * Don't allow io_uring instances to be registered. If UNIX
7763 * isn't enabled, then this causes a reference cycle and this
7764 * instance can never get freed. If UNIX is enabled we'll
7765 * handle it just fine, but there's still no point in allowing
7766 * a ring fd as it doesn't support regular read/write anyway.
7768 if (file->f_op == &io_uring_fops) {
7772 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7775 ret = io_sqe_files_scm(ctx);
7777 __io_sqe_files_unregister(ctx);
7781 io_rsrc_node_switch(ctx, NULL);
7784 for (i = 0; i < ctx->nr_user_files; i++) {
7785 file = io_file_from_index(ctx, i);
7789 io_free_file_tables(&ctx->file_table);
7790 ctx->nr_user_files = 0;
7792 io_rsrc_data_free(ctx->file_data);
7793 ctx->file_data = NULL;
7797 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7800 #if defined(CONFIG_UNIX)
7801 struct sock *sock = ctx->ring_sock->sk;
7802 struct sk_buff_head *head = &sock->sk_receive_queue;
7803 struct sk_buff *skb;
7806 * See if we can merge this file into an existing skb SCM_RIGHTS
7807 * file set. If there's no room, fall back to allocating a new skb
7808 * and filling it in.
7810 spin_lock_irq(&head->lock);
7811 skb = skb_peek(head);
7813 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7815 if (fpl->count < SCM_MAX_FD) {
7816 __skb_unlink(skb, head);
7817 spin_unlock_irq(&head->lock);
7818 fpl->fp[fpl->count] = get_file(file);
7819 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7821 spin_lock_irq(&head->lock);
7822 __skb_queue_head(head, skb);
7827 spin_unlock_irq(&head->lock);
7834 return __io_sqe_files_scm(ctx, 1, index);
7840 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7841 struct io_rsrc_node *node, void *rsrc)
7843 struct io_rsrc_put *prsrc;
7845 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7849 prsrc->tag = *io_get_tag_slot(data, idx);
7851 list_add(&prsrc->list, &node->rsrc_list);
7855 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7856 struct io_uring_rsrc_update2 *up,
7859 u64 __user *tags = u64_to_user_ptr(up->tags);
7860 __s32 __user *fds = u64_to_user_ptr(up->data);
7861 struct io_rsrc_data *data = ctx->file_data;
7862 struct io_fixed_file *file_slot;
7866 bool needs_switch = false;
7868 if (!ctx->file_data)
7870 if (up->offset + nr_args > ctx->nr_user_files)
7873 for (done = 0; done < nr_args; done++) {
7876 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7877 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7881 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7885 if (fd == IORING_REGISTER_FILES_SKIP)
7888 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7889 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7891 if (file_slot->file_ptr) {
7892 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7893 err = io_queue_rsrc_removal(data, up->offset + done,
7894 ctx->rsrc_node, file);
7897 file_slot->file_ptr = 0;
7898 needs_switch = true;
7907 * Don't allow io_uring instances to be registered. If
7908 * UNIX isn't enabled, then this causes a reference
7909 * cycle and this instance can never get freed. If UNIX
7910 * is enabled we'll handle it just fine, but there's
7911 * still no point in allowing a ring fd as it doesn't
7912 * support regular read/write anyway.
7914 if (file->f_op == &io_uring_fops) {
7919 *io_get_tag_slot(data, up->offset + done) = tag;
7920 io_fixed_file_set(file_slot, file);
7921 err = io_sqe_file_register(ctx, file, i);
7923 file_slot->file_ptr = 0;
7931 io_rsrc_node_switch(ctx, data);
7932 return done ? done : err;
7935 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7936 struct task_struct *task)
7938 struct io_wq_hash *hash;
7939 struct io_wq_data data;
7940 unsigned int concurrency;
7942 mutex_lock(&ctx->uring_lock);
7943 hash = ctx->hash_map;
7945 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7947 mutex_unlock(&ctx->uring_lock);
7948 return ERR_PTR(-ENOMEM);
7950 refcount_set(&hash->refs, 1);
7951 init_waitqueue_head(&hash->wait);
7952 ctx->hash_map = hash;
7954 mutex_unlock(&ctx->uring_lock);
7958 data.free_work = io_wq_free_work;
7959 data.do_work = io_wq_submit_work;
7961 /* Do QD, or 4 * CPUS, whatever is smallest */
7962 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7964 return io_wq_create(concurrency, &data);
7967 static int io_uring_alloc_task_context(struct task_struct *task,
7968 struct io_ring_ctx *ctx)
7970 struct io_uring_task *tctx;
7973 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7974 if (unlikely(!tctx))
7977 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7978 if (unlikely(ret)) {
7983 tctx->io_wq = io_init_wq_offload(ctx, task);
7984 if (IS_ERR(tctx->io_wq)) {
7985 ret = PTR_ERR(tctx->io_wq);
7986 percpu_counter_destroy(&tctx->inflight);
7992 init_waitqueue_head(&tctx->wait);
7993 atomic_set(&tctx->in_idle, 0);
7994 atomic_set(&tctx->inflight_tracked, 0);
7995 task->io_uring = tctx;
7996 spin_lock_init(&tctx->task_lock);
7997 INIT_WQ_LIST(&tctx->task_list);
7998 init_task_work(&tctx->task_work, tctx_task_work);
8002 void __io_uring_free(struct task_struct *tsk)
8004 struct io_uring_task *tctx = tsk->io_uring;
8006 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8007 WARN_ON_ONCE(tctx->io_wq);
8008 WARN_ON_ONCE(tctx->cached_refs);
8010 percpu_counter_destroy(&tctx->inflight);
8012 tsk->io_uring = NULL;
8015 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8016 struct io_uring_params *p)
8020 /* Retain compatibility with failing for an invalid attach attempt */
8021 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8022 IORING_SETUP_ATTACH_WQ) {
8025 f = fdget(p->wq_fd);
8028 if (f.file->f_op != &io_uring_fops) {
8034 if (ctx->flags & IORING_SETUP_SQPOLL) {
8035 struct task_struct *tsk;
8036 struct io_sq_data *sqd;
8039 sqd = io_get_sq_data(p, &attached);
8045 ctx->sq_creds = get_current_cred();
8047 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8048 if (!ctx->sq_thread_idle)
8049 ctx->sq_thread_idle = HZ;
8051 io_sq_thread_park(sqd);
8052 list_add(&ctx->sqd_list, &sqd->ctx_list);
8053 io_sqd_update_thread_idle(sqd);
8054 /* don't attach to a dying SQPOLL thread, would be racy */
8055 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8056 io_sq_thread_unpark(sqd);
8063 if (p->flags & IORING_SETUP_SQ_AFF) {
8064 int cpu = p->sq_thread_cpu;
8067 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8074 sqd->task_pid = current->pid;
8075 sqd->task_tgid = current->tgid;
8076 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8083 ret = io_uring_alloc_task_context(tsk, ctx);
8084 wake_up_new_task(tsk);
8087 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8088 /* Can't have SQ_AFF without SQPOLL */
8095 complete(&ctx->sq_data->exited);
8097 io_sq_thread_finish(ctx);
8101 static inline void __io_unaccount_mem(struct user_struct *user,
8102 unsigned long nr_pages)
8104 atomic_long_sub(nr_pages, &user->locked_vm);
8107 static inline int __io_account_mem(struct user_struct *user,
8108 unsigned long nr_pages)
8110 unsigned long page_limit, cur_pages, new_pages;
8112 /* Don't allow more pages than we can safely lock */
8113 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8116 cur_pages = atomic_long_read(&user->locked_vm);
8117 new_pages = cur_pages + nr_pages;
8118 if (new_pages > page_limit)
8120 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8121 new_pages) != cur_pages);
8126 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8129 __io_unaccount_mem(ctx->user, nr_pages);
8131 if (ctx->mm_account)
8132 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8135 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8140 ret = __io_account_mem(ctx->user, nr_pages);
8145 if (ctx->mm_account)
8146 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8151 static void io_mem_free(void *ptr)
8158 page = virt_to_head_page(ptr);
8159 if (put_page_testzero(page))
8160 free_compound_page(page);
8163 static void *io_mem_alloc(size_t size)
8165 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8166 __GFP_NORETRY | __GFP_ACCOUNT;
8168 return (void *) __get_free_pages(gfp_flags, get_order(size));
8171 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8174 struct io_rings *rings;
8175 size_t off, sq_array_size;
8177 off = struct_size(rings, cqes, cq_entries);
8178 if (off == SIZE_MAX)
8182 off = ALIGN(off, SMP_CACHE_BYTES);
8190 sq_array_size = array_size(sizeof(u32), sq_entries);
8191 if (sq_array_size == SIZE_MAX)
8194 if (check_add_overflow(off, sq_array_size, &off))
8200 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8202 struct io_mapped_ubuf *imu = *slot;
8205 if (imu != ctx->dummy_ubuf) {
8206 for (i = 0; i < imu->nr_bvecs; i++)
8207 unpin_user_page(imu->bvec[i].bv_page);
8208 if (imu->acct_pages)
8209 io_unaccount_mem(ctx, imu->acct_pages);
8215 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8217 io_buffer_unmap(ctx, &prsrc->buf);
8221 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8225 for (i = 0; i < ctx->nr_user_bufs; i++)
8226 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8227 kfree(ctx->user_bufs);
8228 io_rsrc_data_free(ctx->buf_data);
8229 ctx->user_bufs = NULL;
8230 ctx->buf_data = NULL;
8231 ctx->nr_user_bufs = 0;
8234 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8241 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8243 __io_sqe_buffers_unregister(ctx);
8247 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8248 void __user *arg, unsigned index)
8250 struct iovec __user *src;
8252 #ifdef CONFIG_COMPAT
8254 struct compat_iovec __user *ciovs;
8255 struct compat_iovec ciov;
8257 ciovs = (struct compat_iovec __user *) arg;
8258 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8261 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8262 dst->iov_len = ciov.iov_len;
8266 src = (struct iovec __user *) arg;
8267 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8273 * Not super efficient, but this is just a registration time. And we do cache
8274 * the last compound head, so generally we'll only do a full search if we don't
8277 * We check if the given compound head page has already been accounted, to
8278 * avoid double accounting it. This allows us to account the full size of the
8279 * page, not just the constituent pages of a huge page.
8281 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8282 int nr_pages, struct page *hpage)
8286 /* check current page array */
8287 for (i = 0; i < nr_pages; i++) {
8288 if (!PageCompound(pages[i]))
8290 if (compound_head(pages[i]) == hpage)
8294 /* check previously registered pages */
8295 for (i = 0; i < ctx->nr_user_bufs; i++) {
8296 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8298 for (j = 0; j < imu->nr_bvecs; j++) {
8299 if (!PageCompound(imu->bvec[j].bv_page))
8301 if (compound_head(imu->bvec[j].bv_page) == hpage)
8309 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8310 int nr_pages, struct io_mapped_ubuf *imu,
8311 struct page **last_hpage)
8315 imu->acct_pages = 0;
8316 for (i = 0; i < nr_pages; i++) {
8317 if (!PageCompound(pages[i])) {
8322 hpage = compound_head(pages[i]);
8323 if (hpage == *last_hpage)
8325 *last_hpage = hpage;
8326 if (headpage_already_acct(ctx, pages, i, hpage))
8328 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8332 if (!imu->acct_pages)
8335 ret = io_account_mem(ctx, imu->acct_pages);
8337 imu->acct_pages = 0;
8341 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8342 struct io_mapped_ubuf **pimu,
8343 struct page **last_hpage)
8345 struct io_mapped_ubuf *imu = NULL;
8346 struct vm_area_struct **vmas = NULL;
8347 struct page **pages = NULL;
8348 unsigned long off, start, end, ubuf;
8350 int ret, pret, nr_pages, i;
8352 if (!iov->iov_base) {
8353 *pimu = ctx->dummy_ubuf;
8357 ubuf = (unsigned long) iov->iov_base;
8358 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8359 start = ubuf >> PAGE_SHIFT;
8360 nr_pages = end - start;
8365 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8369 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8374 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8379 mmap_read_lock(current->mm);
8380 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8382 if (pret == nr_pages) {
8383 /* don't support file backed memory */
8384 for (i = 0; i < nr_pages; i++) {
8385 struct vm_area_struct *vma = vmas[i];
8387 if (vma_is_shmem(vma))
8390 !is_file_hugepages(vma->vm_file)) {
8396 ret = pret < 0 ? pret : -EFAULT;
8398 mmap_read_unlock(current->mm);
8401 * if we did partial map, or found file backed vmas,
8402 * release any pages we did get
8405 unpin_user_pages(pages, pret);
8409 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8411 unpin_user_pages(pages, pret);
8415 off = ubuf & ~PAGE_MASK;
8416 size = iov->iov_len;
8417 for (i = 0; i < nr_pages; i++) {
8420 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8421 imu->bvec[i].bv_page = pages[i];
8422 imu->bvec[i].bv_len = vec_len;
8423 imu->bvec[i].bv_offset = off;
8427 /* store original address for later verification */
8429 imu->ubuf_end = ubuf + iov->iov_len;
8430 imu->nr_bvecs = nr_pages;
8441 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8443 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8444 return ctx->user_bufs ? 0 : -ENOMEM;
8447 static int io_buffer_validate(struct iovec *iov)
8449 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8452 * Don't impose further limits on the size and buffer
8453 * constraints here, we'll -EINVAL later when IO is
8454 * submitted if they are wrong.
8457 return iov->iov_len ? -EFAULT : 0;
8461 /* arbitrary limit, but we need something */
8462 if (iov->iov_len > SZ_1G)
8465 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8471 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8472 unsigned int nr_args, u64 __user *tags)
8474 struct page *last_hpage = NULL;
8475 struct io_rsrc_data *data;
8481 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8483 ret = io_rsrc_node_switch_start(ctx);
8486 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8489 ret = io_buffers_map_alloc(ctx, nr_args);
8491 io_rsrc_data_free(data);
8495 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8496 ret = io_copy_iov(ctx, &iov, arg, i);
8499 ret = io_buffer_validate(&iov);
8502 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8507 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8513 WARN_ON_ONCE(ctx->buf_data);
8515 ctx->buf_data = data;
8517 __io_sqe_buffers_unregister(ctx);
8519 io_rsrc_node_switch(ctx, NULL);
8523 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8524 struct io_uring_rsrc_update2 *up,
8525 unsigned int nr_args)
8527 u64 __user *tags = u64_to_user_ptr(up->tags);
8528 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8529 struct page *last_hpage = NULL;
8530 bool needs_switch = false;
8536 if (up->offset + nr_args > ctx->nr_user_bufs)
8539 for (done = 0; done < nr_args; done++) {
8540 struct io_mapped_ubuf *imu;
8541 int offset = up->offset + done;
8544 err = io_copy_iov(ctx, &iov, iovs, done);
8547 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8551 err = io_buffer_validate(&iov);
8554 if (!iov.iov_base && tag) {
8558 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8562 i = array_index_nospec(offset, ctx->nr_user_bufs);
8563 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8564 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8565 ctx->rsrc_node, ctx->user_bufs[i]);
8566 if (unlikely(err)) {
8567 io_buffer_unmap(ctx, &imu);
8570 ctx->user_bufs[i] = NULL;
8571 needs_switch = true;
8574 ctx->user_bufs[i] = imu;
8575 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8579 io_rsrc_node_switch(ctx, ctx->buf_data);
8580 return done ? done : err;
8583 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8585 __s32 __user *fds = arg;
8591 if (copy_from_user(&fd, fds, sizeof(*fds)))
8594 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8595 if (IS_ERR(ctx->cq_ev_fd)) {
8596 int ret = PTR_ERR(ctx->cq_ev_fd);
8598 ctx->cq_ev_fd = NULL;
8605 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8607 if (ctx->cq_ev_fd) {
8608 eventfd_ctx_put(ctx->cq_ev_fd);
8609 ctx->cq_ev_fd = NULL;
8616 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8618 struct io_buffer *buf;
8619 unsigned long index;
8621 xa_for_each(&ctx->io_buffers, index, buf)
8622 __io_remove_buffers(ctx, buf, index, -1U);
8625 static void io_req_cache_free(struct list_head *list)
8627 struct io_kiocb *req, *nxt;
8629 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8630 list_del(&req->inflight_entry);
8631 kmem_cache_free(req_cachep, req);
8635 static void io_req_caches_free(struct io_ring_ctx *ctx)
8637 struct io_submit_state *state = &ctx->submit_state;
8639 mutex_lock(&ctx->uring_lock);
8641 if (state->free_reqs) {
8642 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8643 state->free_reqs = 0;
8646 io_flush_cached_locked_reqs(ctx, state);
8647 io_req_cache_free(&state->free_list);
8648 mutex_unlock(&ctx->uring_lock);
8651 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8653 if (data && !atomic_dec_and_test(&data->refs))
8654 wait_for_completion(&data->done);
8657 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8659 io_sq_thread_finish(ctx);
8661 if (ctx->mm_account) {
8662 mmdrop(ctx->mm_account);
8663 ctx->mm_account = NULL;
8666 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8667 io_wait_rsrc_data(ctx->buf_data);
8668 io_wait_rsrc_data(ctx->file_data);
8670 mutex_lock(&ctx->uring_lock);
8672 __io_sqe_buffers_unregister(ctx);
8674 __io_sqe_files_unregister(ctx);
8676 __io_cqring_overflow_flush(ctx, true);
8677 mutex_unlock(&ctx->uring_lock);
8678 io_eventfd_unregister(ctx);
8679 io_destroy_buffers(ctx);
8681 put_cred(ctx->sq_creds);
8683 /* there are no registered resources left, nobody uses it */
8685 io_rsrc_node_destroy(ctx->rsrc_node);
8686 if (ctx->rsrc_backup_node)
8687 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8688 flush_delayed_work(&ctx->rsrc_put_work);
8690 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8691 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8693 #if defined(CONFIG_UNIX)
8694 if (ctx->ring_sock) {
8695 ctx->ring_sock->file = NULL; /* so that iput() is called */
8696 sock_release(ctx->ring_sock);
8700 io_mem_free(ctx->rings);
8701 io_mem_free(ctx->sq_sqes);
8703 percpu_ref_exit(&ctx->refs);
8704 free_uid(ctx->user);
8705 io_req_caches_free(ctx);
8707 io_wq_put_hash(ctx->hash_map);
8708 kfree(ctx->cancel_hash);
8709 kfree(ctx->dummy_ubuf);
8713 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8715 struct io_ring_ctx *ctx = file->private_data;
8718 poll_wait(file, &ctx->poll_wait, wait);
8720 * synchronizes with barrier from wq_has_sleeper call in
8724 if (!io_sqring_full(ctx))
8725 mask |= EPOLLOUT | EPOLLWRNORM;
8728 * Don't flush cqring overflow list here, just do a simple check.
8729 * Otherwise there could possible be ABBA deadlock:
8732 * lock(&ctx->uring_lock);
8734 * lock(&ctx->uring_lock);
8737 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8738 * pushs them to do the flush.
8740 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8741 mask |= EPOLLIN | EPOLLRDNORM;
8746 static int io_uring_fasync(int fd, struct file *file, int on)
8748 struct io_ring_ctx *ctx = file->private_data;
8750 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8753 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8755 const struct cred *creds;
8757 creds = xa_erase(&ctx->personalities, id);
8766 struct io_tctx_exit {
8767 struct callback_head task_work;
8768 struct completion completion;
8769 struct io_ring_ctx *ctx;
8772 static void io_tctx_exit_cb(struct callback_head *cb)
8774 struct io_uring_task *tctx = current->io_uring;
8775 struct io_tctx_exit *work;
8777 work = container_of(cb, struct io_tctx_exit, task_work);
8779 * When @in_idle, we're in cancellation and it's racy to remove the
8780 * node. It'll be removed by the end of cancellation, just ignore it.
8782 if (!atomic_read(&tctx->in_idle))
8783 io_uring_del_tctx_node((unsigned long)work->ctx);
8784 complete(&work->completion);
8787 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8789 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8791 return req->ctx == data;
8794 static void io_ring_exit_work(struct work_struct *work)
8796 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8797 unsigned long timeout = jiffies + HZ * 60 * 5;
8798 unsigned long interval = HZ / 20;
8799 struct io_tctx_exit exit;
8800 struct io_tctx_node *node;
8804 * If we're doing polled IO and end up having requests being
8805 * submitted async (out-of-line), then completions can come in while
8806 * we're waiting for refs to drop. We need to reap these manually,
8807 * as nobody else will be looking for them.
8810 io_uring_try_cancel_requests(ctx, NULL, true);
8812 struct io_sq_data *sqd = ctx->sq_data;
8813 struct task_struct *tsk;
8815 io_sq_thread_park(sqd);
8817 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8818 io_wq_cancel_cb(tsk->io_uring->io_wq,
8819 io_cancel_ctx_cb, ctx, true);
8820 io_sq_thread_unpark(sqd);
8823 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8824 /* there is little hope left, don't run it too often */
8827 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8829 init_completion(&exit.completion);
8830 init_task_work(&exit.task_work, io_tctx_exit_cb);
8833 * Some may use context even when all refs and requests have been put,
8834 * and they are free to do so while still holding uring_lock or
8835 * completion_lock, see io_req_task_submit(). Apart from other work,
8836 * this lock/unlock section also waits them to finish.
8838 mutex_lock(&ctx->uring_lock);
8839 while (!list_empty(&ctx->tctx_list)) {
8840 WARN_ON_ONCE(time_after(jiffies, timeout));
8842 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8844 /* don't spin on a single task if cancellation failed */
8845 list_rotate_left(&ctx->tctx_list);
8846 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8847 if (WARN_ON_ONCE(ret))
8849 wake_up_process(node->task);
8851 mutex_unlock(&ctx->uring_lock);
8852 wait_for_completion(&exit.completion);
8853 mutex_lock(&ctx->uring_lock);
8855 mutex_unlock(&ctx->uring_lock);
8856 spin_lock(&ctx->completion_lock);
8857 spin_unlock(&ctx->completion_lock);
8859 io_ring_ctx_free(ctx);
8862 /* Returns true if we found and killed one or more timeouts */
8863 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8866 struct io_kiocb *req, *tmp;
8869 spin_lock(&ctx->completion_lock);
8870 spin_lock_irq(&ctx->timeout_lock);
8871 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8872 if (io_match_task(req, tsk, cancel_all)) {
8873 io_kill_timeout(req, -ECANCELED);
8877 spin_unlock_irq(&ctx->timeout_lock);
8879 io_commit_cqring(ctx);
8880 spin_unlock(&ctx->completion_lock);
8882 io_cqring_ev_posted(ctx);
8883 return canceled != 0;
8886 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8888 unsigned long index;
8889 struct creds *creds;
8891 mutex_lock(&ctx->uring_lock);
8892 percpu_ref_kill(&ctx->refs);
8894 __io_cqring_overflow_flush(ctx, true);
8895 xa_for_each(&ctx->personalities, index, creds)
8896 io_unregister_personality(ctx, index);
8897 mutex_unlock(&ctx->uring_lock);
8899 io_kill_timeouts(ctx, NULL, true);
8900 io_poll_remove_all(ctx, NULL, true);
8902 /* if we failed setting up the ctx, we might not have any rings */
8903 io_iopoll_try_reap_events(ctx);
8905 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8907 * Use system_unbound_wq to avoid spawning tons of event kworkers
8908 * if we're exiting a ton of rings at the same time. It just adds
8909 * noise and overhead, there's no discernable change in runtime
8910 * over using system_wq.
8912 queue_work(system_unbound_wq, &ctx->exit_work);
8915 static int io_uring_release(struct inode *inode, struct file *file)
8917 struct io_ring_ctx *ctx = file->private_data;
8919 file->private_data = NULL;
8920 io_ring_ctx_wait_and_kill(ctx);
8924 struct io_task_cancel {
8925 struct task_struct *task;
8929 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8931 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8932 struct io_task_cancel *cancel = data;
8935 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8936 struct io_ring_ctx *ctx = req->ctx;
8938 /* protect against races with linked timeouts */
8939 spin_lock(&ctx->completion_lock);
8940 ret = io_match_task(req, cancel->task, cancel->all);
8941 spin_unlock(&ctx->completion_lock);
8943 ret = io_match_task(req, cancel->task, cancel->all);
8948 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8949 struct task_struct *task, bool cancel_all)
8951 struct io_defer_entry *de;
8954 spin_lock(&ctx->completion_lock);
8955 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8956 if (io_match_task(de->req, task, cancel_all)) {
8957 list_cut_position(&list, &ctx->defer_list, &de->list);
8961 spin_unlock(&ctx->completion_lock);
8962 if (list_empty(&list))
8965 while (!list_empty(&list)) {
8966 de = list_first_entry(&list, struct io_defer_entry, list);
8967 list_del_init(&de->list);
8968 io_req_complete_failed(de->req, -ECANCELED);
8974 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8976 struct io_tctx_node *node;
8977 enum io_wq_cancel cret;
8980 mutex_lock(&ctx->uring_lock);
8981 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8982 struct io_uring_task *tctx = node->task->io_uring;
8985 * io_wq will stay alive while we hold uring_lock, because it's
8986 * killed after ctx nodes, which requires to take the lock.
8988 if (!tctx || !tctx->io_wq)
8990 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8991 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8993 mutex_unlock(&ctx->uring_lock);
8998 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8999 struct task_struct *task,
9002 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9003 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9006 enum io_wq_cancel cret;
9010 ret |= io_uring_try_cancel_iowq(ctx);
9011 } else if (tctx && tctx->io_wq) {
9013 * Cancels requests of all rings, not only @ctx, but
9014 * it's fine as the task is in exit/exec.
9016 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9018 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9021 /* SQPOLL thread does its own polling */
9022 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9023 (ctx->sq_data && ctx->sq_data->thread == current)) {
9024 while (!list_empty_careful(&ctx->iopoll_list)) {
9025 io_iopoll_try_reap_events(ctx);
9030 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9031 ret |= io_poll_remove_all(ctx, task, cancel_all);
9032 ret |= io_kill_timeouts(ctx, task, cancel_all);
9034 ret |= io_run_task_work();
9041 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9043 struct io_uring_task *tctx = current->io_uring;
9044 struct io_tctx_node *node;
9047 if (unlikely(!tctx)) {
9048 ret = io_uring_alloc_task_context(current, ctx);
9051 tctx = current->io_uring;
9053 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9054 node = kmalloc(sizeof(*node), GFP_KERNEL);
9058 node->task = current;
9060 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9067 mutex_lock(&ctx->uring_lock);
9068 list_add(&node->ctx_node, &ctx->tctx_list);
9069 mutex_unlock(&ctx->uring_lock);
9076 * Note that this task has used io_uring. We use it for cancelation purposes.
9078 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9080 struct io_uring_task *tctx = current->io_uring;
9082 if (likely(tctx && tctx->last == ctx))
9084 return __io_uring_add_tctx_node(ctx);
9088 * Remove this io_uring_file -> task mapping.
9090 static void io_uring_del_tctx_node(unsigned long index)
9092 struct io_uring_task *tctx = current->io_uring;
9093 struct io_tctx_node *node;
9097 node = xa_erase(&tctx->xa, index);
9101 WARN_ON_ONCE(current != node->task);
9102 WARN_ON_ONCE(list_empty(&node->ctx_node));
9104 mutex_lock(&node->ctx->uring_lock);
9105 list_del(&node->ctx_node);
9106 mutex_unlock(&node->ctx->uring_lock);
9108 if (tctx->last == node->ctx)
9113 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9115 struct io_wq *wq = tctx->io_wq;
9116 struct io_tctx_node *node;
9117 unsigned long index;
9119 xa_for_each(&tctx->xa, index, node)
9120 io_uring_del_tctx_node(index);
9123 * Must be after io_uring_del_task_file() (removes nodes under
9124 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9127 io_wq_put_and_exit(wq);
9131 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9134 return atomic_read(&tctx->inflight_tracked);
9135 return percpu_counter_sum(&tctx->inflight);
9138 static void io_uring_drop_tctx_refs(struct task_struct *task)
9140 struct io_uring_task *tctx = task->io_uring;
9141 unsigned int refs = tctx->cached_refs;
9144 tctx->cached_refs = 0;
9145 percpu_counter_sub(&tctx->inflight, refs);
9146 put_task_struct_many(task, refs);
9151 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9152 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9154 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9156 struct io_uring_task *tctx = current->io_uring;
9157 struct io_ring_ctx *ctx;
9161 WARN_ON_ONCE(sqd && sqd->thread != current);
9163 if (!current->io_uring)
9166 io_wq_exit_start(tctx->io_wq);
9168 atomic_inc(&tctx->in_idle);
9170 io_uring_drop_tctx_refs(current);
9171 /* read completions before cancelations */
9172 inflight = tctx_inflight(tctx, !cancel_all);
9177 struct io_tctx_node *node;
9178 unsigned long index;
9180 xa_for_each(&tctx->xa, index, node) {
9181 /* sqpoll task will cancel all its requests */
9182 if (node->ctx->sq_data)
9184 io_uring_try_cancel_requests(node->ctx, current,
9188 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9189 io_uring_try_cancel_requests(ctx, current,
9193 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9194 io_uring_drop_tctx_refs(current);
9196 * If we've seen completions, retry without waiting. This
9197 * avoids a race where a completion comes in before we did
9198 * prepare_to_wait().
9200 if (inflight == tctx_inflight(tctx, !cancel_all))
9202 finish_wait(&tctx->wait, &wait);
9204 atomic_dec(&tctx->in_idle);
9206 io_uring_clean_tctx(tctx);
9208 /* for exec all current's requests should be gone, kill tctx */
9209 __io_uring_free(current);
9213 void __io_uring_cancel(bool cancel_all)
9215 io_uring_cancel_generic(cancel_all, NULL);
9218 static void *io_uring_validate_mmap_request(struct file *file,
9219 loff_t pgoff, size_t sz)
9221 struct io_ring_ctx *ctx = file->private_data;
9222 loff_t offset = pgoff << PAGE_SHIFT;
9227 case IORING_OFF_SQ_RING:
9228 case IORING_OFF_CQ_RING:
9231 case IORING_OFF_SQES:
9235 return ERR_PTR(-EINVAL);
9238 page = virt_to_head_page(ptr);
9239 if (sz > page_size(page))
9240 return ERR_PTR(-EINVAL);
9247 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9249 size_t sz = vma->vm_end - vma->vm_start;
9253 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9255 return PTR_ERR(ptr);
9257 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9258 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9261 #else /* !CONFIG_MMU */
9263 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9265 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9268 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9270 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9273 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9274 unsigned long addr, unsigned long len,
9275 unsigned long pgoff, unsigned long flags)
9279 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9281 return PTR_ERR(ptr);
9283 return (unsigned long) ptr;
9286 #endif /* !CONFIG_MMU */
9288 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9293 if (!io_sqring_full(ctx))
9295 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9297 if (!io_sqring_full(ctx))
9300 } while (!signal_pending(current));
9302 finish_wait(&ctx->sqo_sq_wait, &wait);
9306 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9307 struct __kernel_timespec __user **ts,
9308 const sigset_t __user **sig)
9310 struct io_uring_getevents_arg arg;
9313 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9314 * is just a pointer to the sigset_t.
9316 if (!(flags & IORING_ENTER_EXT_ARG)) {
9317 *sig = (const sigset_t __user *) argp;
9323 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9324 * timespec and sigset_t pointers if good.
9326 if (*argsz != sizeof(arg))
9328 if (copy_from_user(&arg, argp, sizeof(arg)))
9330 *sig = u64_to_user_ptr(arg.sigmask);
9331 *argsz = arg.sigmask_sz;
9332 *ts = u64_to_user_ptr(arg.ts);
9336 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9337 u32, min_complete, u32, flags, const void __user *, argp,
9340 struct io_ring_ctx *ctx;
9347 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9348 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9352 if (unlikely(!f.file))
9356 if (unlikely(f.file->f_op != &io_uring_fops))
9360 ctx = f.file->private_data;
9361 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9365 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9369 * For SQ polling, the thread will do all submissions and completions.
9370 * Just return the requested submit count, and wake the thread if
9374 if (ctx->flags & IORING_SETUP_SQPOLL) {
9375 io_cqring_overflow_flush(ctx);
9377 if (unlikely(ctx->sq_data->thread == NULL)) {
9381 if (flags & IORING_ENTER_SQ_WAKEUP)
9382 wake_up(&ctx->sq_data->wait);
9383 if (flags & IORING_ENTER_SQ_WAIT) {
9384 ret = io_sqpoll_wait_sq(ctx);
9388 submitted = to_submit;
9389 } else if (to_submit) {
9390 ret = io_uring_add_tctx_node(ctx);
9393 mutex_lock(&ctx->uring_lock);
9394 submitted = io_submit_sqes(ctx, to_submit);
9395 mutex_unlock(&ctx->uring_lock);
9397 if (submitted != to_submit)
9400 if (flags & IORING_ENTER_GETEVENTS) {
9401 const sigset_t __user *sig;
9402 struct __kernel_timespec __user *ts;
9404 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9408 min_complete = min(min_complete, ctx->cq_entries);
9411 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9412 * space applications don't need to do io completion events
9413 * polling again, they can rely on io_sq_thread to do polling
9414 * work, which can reduce cpu usage and uring_lock contention.
9416 if (ctx->flags & IORING_SETUP_IOPOLL &&
9417 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9418 ret = io_iopoll_check(ctx, min_complete);
9420 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9425 percpu_ref_put(&ctx->refs);
9428 return submitted ? submitted : ret;
9431 #ifdef CONFIG_PROC_FS
9432 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9433 const struct cred *cred)
9435 struct user_namespace *uns = seq_user_ns(m);
9436 struct group_info *gi;
9441 seq_printf(m, "%5d\n", id);
9442 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9443 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9444 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9445 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9446 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9447 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9448 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9449 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9450 seq_puts(m, "\n\tGroups:\t");
9451 gi = cred->group_info;
9452 for (g = 0; g < gi->ngroups; g++) {
9453 seq_put_decimal_ull(m, g ? " " : "",
9454 from_kgid_munged(uns, gi->gid[g]));
9456 seq_puts(m, "\n\tCapEff:\t");
9457 cap = cred->cap_effective;
9458 CAP_FOR_EACH_U32(__capi)
9459 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9464 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9466 struct io_sq_data *sq = NULL;
9471 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9472 * since fdinfo case grabs it in the opposite direction of normal use
9473 * cases. If we fail to get the lock, we just don't iterate any
9474 * structures that could be going away outside the io_uring mutex.
9476 has_lock = mutex_trylock(&ctx->uring_lock);
9478 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9484 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9485 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9486 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9487 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9488 struct file *f = io_file_from_index(ctx, i);
9491 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9493 seq_printf(m, "%5u: <none>\n", i);
9495 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9496 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9497 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9498 unsigned int len = buf->ubuf_end - buf->ubuf;
9500 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9502 if (has_lock && !xa_empty(&ctx->personalities)) {
9503 unsigned long index;
9504 const struct cred *cred;
9506 seq_printf(m, "Personalities:\n");
9507 xa_for_each(&ctx->personalities, index, cred)
9508 io_uring_show_cred(m, index, cred);
9510 seq_printf(m, "PollList:\n");
9511 spin_lock(&ctx->completion_lock);
9512 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9513 struct hlist_head *list = &ctx->cancel_hash[i];
9514 struct io_kiocb *req;
9516 hlist_for_each_entry(req, list, hash_node)
9517 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9518 req->task->task_works != NULL);
9520 spin_unlock(&ctx->completion_lock);
9522 mutex_unlock(&ctx->uring_lock);
9525 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9527 struct io_ring_ctx *ctx = f->private_data;
9529 if (percpu_ref_tryget(&ctx->refs)) {
9530 __io_uring_show_fdinfo(ctx, m);
9531 percpu_ref_put(&ctx->refs);
9536 static const struct file_operations io_uring_fops = {
9537 .release = io_uring_release,
9538 .mmap = io_uring_mmap,
9540 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9541 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9543 .poll = io_uring_poll,
9544 .fasync = io_uring_fasync,
9545 #ifdef CONFIG_PROC_FS
9546 .show_fdinfo = io_uring_show_fdinfo,
9550 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9551 struct io_uring_params *p)
9553 struct io_rings *rings;
9554 size_t size, sq_array_offset;
9556 /* make sure these are sane, as we already accounted them */
9557 ctx->sq_entries = p->sq_entries;
9558 ctx->cq_entries = p->cq_entries;
9560 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9561 if (size == SIZE_MAX)
9564 rings = io_mem_alloc(size);
9569 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9570 rings->sq_ring_mask = p->sq_entries - 1;
9571 rings->cq_ring_mask = p->cq_entries - 1;
9572 rings->sq_ring_entries = p->sq_entries;
9573 rings->cq_ring_entries = p->cq_entries;
9575 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9576 if (size == SIZE_MAX) {
9577 io_mem_free(ctx->rings);
9582 ctx->sq_sqes = io_mem_alloc(size);
9583 if (!ctx->sq_sqes) {
9584 io_mem_free(ctx->rings);
9592 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9596 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9600 ret = io_uring_add_tctx_node(ctx);
9605 fd_install(fd, file);
9610 * Allocate an anonymous fd, this is what constitutes the application
9611 * visible backing of an io_uring instance. The application mmaps this
9612 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9613 * we have to tie this fd to a socket for file garbage collection purposes.
9615 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9618 #if defined(CONFIG_UNIX)
9621 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9624 return ERR_PTR(ret);
9627 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9628 O_RDWR | O_CLOEXEC);
9629 #if defined(CONFIG_UNIX)
9631 sock_release(ctx->ring_sock);
9632 ctx->ring_sock = NULL;
9634 ctx->ring_sock->file = file;
9640 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9641 struct io_uring_params __user *params)
9643 struct io_ring_ctx *ctx;
9649 if (entries > IORING_MAX_ENTRIES) {
9650 if (!(p->flags & IORING_SETUP_CLAMP))
9652 entries = IORING_MAX_ENTRIES;
9656 * Use twice as many entries for the CQ ring. It's possible for the
9657 * application to drive a higher depth than the size of the SQ ring,
9658 * since the sqes are only used at submission time. This allows for
9659 * some flexibility in overcommitting a bit. If the application has
9660 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9661 * of CQ ring entries manually.
9663 p->sq_entries = roundup_pow_of_two(entries);
9664 if (p->flags & IORING_SETUP_CQSIZE) {
9666 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9667 * to a power-of-two, if it isn't already. We do NOT impose
9668 * any cq vs sq ring sizing.
9672 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9673 if (!(p->flags & IORING_SETUP_CLAMP))
9675 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9677 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9678 if (p->cq_entries < p->sq_entries)
9681 p->cq_entries = 2 * p->sq_entries;
9684 ctx = io_ring_ctx_alloc(p);
9687 ctx->compat = in_compat_syscall();
9688 if (!capable(CAP_IPC_LOCK))
9689 ctx->user = get_uid(current_user());
9692 * This is just grabbed for accounting purposes. When a process exits,
9693 * the mm is exited and dropped before the files, hence we need to hang
9694 * on to this mm purely for the purposes of being able to unaccount
9695 * memory (locked/pinned vm). It's not used for anything else.
9697 mmgrab(current->mm);
9698 ctx->mm_account = current->mm;
9700 ret = io_allocate_scq_urings(ctx, p);
9704 ret = io_sq_offload_create(ctx, p);
9707 /* always set a rsrc node */
9708 ret = io_rsrc_node_switch_start(ctx);
9711 io_rsrc_node_switch(ctx, NULL);
9713 memset(&p->sq_off, 0, sizeof(p->sq_off));
9714 p->sq_off.head = offsetof(struct io_rings, sq.head);
9715 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9716 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9717 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9718 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9719 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9720 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9722 memset(&p->cq_off, 0, sizeof(p->cq_off));
9723 p->cq_off.head = offsetof(struct io_rings, cq.head);
9724 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9725 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9726 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9727 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9728 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9729 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9731 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9732 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9733 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9734 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9735 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9736 IORING_FEAT_RSRC_TAGS;
9738 if (copy_to_user(params, p, sizeof(*p))) {
9743 file = io_uring_get_file(ctx);
9745 ret = PTR_ERR(file);
9750 * Install ring fd as the very last thing, so we don't risk someone
9751 * having closed it before we finish setup
9753 ret = io_uring_install_fd(ctx, file);
9755 /* fput will clean it up */
9760 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9763 io_ring_ctx_wait_and_kill(ctx);
9768 * Sets up an aio uring context, and returns the fd. Applications asks for a
9769 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9770 * params structure passed in.
9772 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9774 struct io_uring_params p;
9777 if (copy_from_user(&p, params, sizeof(p)))
9779 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9784 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9785 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9786 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9787 IORING_SETUP_R_DISABLED))
9790 return io_uring_create(entries, &p, params);
9793 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9794 struct io_uring_params __user *, params)
9796 return io_uring_setup(entries, params);
9799 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9801 struct io_uring_probe *p;
9805 size = struct_size(p, ops, nr_args);
9806 if (size == SIZE_MAX)
9808 p = kzalloc(size, GFP_KERNEL);
9813 if (copy_from_user(p, arg, size))
9816 if (memchr_inv(p, 0, size))
9819 p->last_op = IORING_OP_LAST - 1;
9820 if (nr_args > IORING_OP_LAST)
9821 nr_args = IORING_OP_LAST;
9823 for (i = 0; i < nr_args; i++) {
9825 if (!io_op_defs[i].not_supported)
9826 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9831 if (copy_to_user(arg, p, size))
9838 static int io_register_personality(struct io_ring_ctx *ctx)
9840 const struct cred *creds;
9844 creds = get_current_cred();
9846 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9847 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9855 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9856 unsigned int nr_args)
9858 struct io_uring_restriction *res;
9862 /* Restrictions allowed only if rings started disabled */
9863 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9866 /* We allow only a single restrictions registration */
9867 if (ctx->restrictions.registered)
9870 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9873 size = array_size(nr_args, sizeof(*res));
9874 if (size == SIZE_MAX)
9877 res = memdup_user(arg, size);
9879 return PTR_ERR(res);
9883 for (i = 0; i < nr_args; i++) {
9884 switch (res[i].opcode) {
9885 case IORING_RESTRICTION_REGISTER_OP:
9886 if (res[i].register_op >= IORING_REGISTER_LAST) {
9891 __set_bit(res[i].register_op,
9892 ctx->restrictions.register_op);
9894 case IORING_RESTRICTION_SQE_OP:
9895 if (res[i].sqe_op >= IORING_OP_LAST) {
9900 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9902 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9903 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9905 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9906 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9915 /* Reset all restrictions if an error happened */
9917 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9919 ctx->restrictions.registered = true;
9925 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9927 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9930 if (ctx->restrictions.registered)
9931 ctx->restricted = 1;
9933 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9934 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9935 wake_up(&ctx->sq_data->wait);
9939 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9940 struct io_uring_rsrc_update2 *up,
9948 if (check_add_overflow(up->offset, nr_args, &tmp))
9950 err = io_rsrc_node_switch_start(ctx);
9955 case IORING_RSRC_FILE:
9956 return __io_sqe_files_update(ctx, up, nr_args);
9957 case IORING_RSRC_BUFFER:
9958 return __io_sqe_buffers_update(ctx, up, nr_args);
9963 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9966 struct io_uring_rsrc_update2 up;
9970 memset(&up, 0, sizeof(up));
9971 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9973 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9976 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9977 unsigned size, unsigned type)
9979 struct io_uring_rsrc_update2 up;
9981 if (size != sizeof(up))
9983 if (copy_from_user(&up, arg, sizeof(up)))
9985 if (!up.nr || up.resv)
9987 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9990 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9991 unsigned int size, unsigned int type)
9993 struct io_uring_rsrc_register rr;
9995 /* keep it extendible */
9996 if (size != sizeof(rr))
9999 memset(&rr, 0, sizeof(rr));
10000 if (copy_from_user(&rr, arg, size))
10002 if (!rr.nr || rr.resv || rr.resv2)
10006 case IORING_RSRC_FILE:
10007 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10008 rr.nr, u64_to_user_ptr(rr.tags));
10009 case IORING_RSRC_BUFFER:
10010 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10011 rr.nr, u64_to_user_ptr(rr.tags));
10016 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10019 struct io_uring_task *tctx = current->io_uring;
10020 cpumask_var_t new_mask;
10023 if (!tctx || !tctx->io_wq)
10026 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10029 cpumask_clear(new_mask);
10030 if (len > cpumask_size())
10031 len = cpumask_size();
10033 if (copy_from_user(new_mask, arg, len)) {
10034 free_cpumask_var(new_mask);
10038 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10039 free_cpumask_var(new_mask);
10043 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10045 struct io_uring_task *tctx = current->io_uring;
10047 if (!tctx || !tctx->io_wq)
10050 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10053 static bool io_register_op_must_quiesce(int op)
10056 case IORING_REGISTER_BUFFERS:
10057 case IORING_UNREGISTER_BUFFERS:
10058 case IORING_REGISTER_FILES:
10059 case IORING_UNREGISTER_FILES:
10060 case IORING_REGISTER_FILES_UPDATE:
10061 case IORING_REGISTER_PROBE:
10062 case IORING_REGISTER_PERSONALITY:
10063 case IORING_UNREGISTER_PERSONALITY:
10064 case IORING_REGISTER_FILES2:
10065 case IORING_REGISTER_FILES_UPDATE2:
10066 case IORING_REGISTER_BUFFERS2:
10067 case IORING_REGISTER_BUFFERS_UPDATE:
10068 case IORING_REGISTER_IOWQ_AFF:
10069 case IORING_UNREGISTER_IOWQ_AFF:
10076 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10080 percpu_ref_kill(&ctx->refs);
10083 * Drop uring mutex before waiting for references to exit. If another
10084 * thread is currently inside io_uring_enter() it might need to grab the
10085 * uring_lock to make progress. If we hold it here across the drain
10086 * wait, then we can deadlock. It's safe to drop the mutex here, since
10087 * no new references will come in after we've killed the percpu ref.
10089 mutex_unlock(&ctx->uring_lock);
10091 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10094 ret = io_run_task_work_sig();
10095 } while (ret >= 0);
10096 mutex_lock(&ctx->uring_lock);
10099 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10103 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10104 void __user *arg, unsigned nr_args)
10105 __releases(ctx->uring_lock)
10106 __acquires(ctx->uring_lock)
10111 * We're inside the ring mutex, if the ref is already dying, then
10112 * someone else killed the ctx or is already going through
10113 * io_uring_register().
10115 if (percpu_ref_is_dying(&ctx->refs))
10118 if (ctx->restricted) {
10119 if (opcode >= IORING_REGISTER_LAST)
10121 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10122 if (!test_bit(opcode, ctx->restrictions.register_op))
10126 if (io_register_op_must_quiesce(opcode)) {
10127 ret = io_ctx_quiesce(ctx);
10133 case IORING_REGISTER_BUFFERS:
10134 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10136 case IORING_UNREGISTER_BUFFERS:
10138 if (arg || nr_args)
10140 ret = io_sqe_buffers_unregister(ctx);
10142 case IORING_REGISTER_FILES:
10143 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10145 case IORING_UNREGISTER_FILES:
10147 if (arg || nr_args)
10149 ret = io_sqe_files_unregister(ctx);
10151 case IORING_REGISTER_FILES_UPDATE:
10152 ret = io_register_files_update(ctx, arg, nr_args);
10154 case IORING_REGISTER_EVENTFD:
10155 case IORING_REGISTER_EVENTFD_ASYNC:
10159 ret = io_eventfd_register(ctx, arg);
10162 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10163 ctx->eventfd_async = 1;
10165 ctx->eventfd_async = 0;
10167 case IORING_UNREGISTER_EVENTFD:
10169 if (arg || nr_args)
10171 ret = io_eventfd_unregister(ctx);
10173 case IORING_REGISTER_PROBE:
10175 if (!arg || nr_args > 256)
10177 ret = io_probe(ctx, arg, nr_args);
10179 case IORING_REGISTER_PERSONALITY:
10181 if (arg || nr_args)
10183 ret = io_register_personality(ctx);
10185 case IORING_UNREGISTER_PERSONALITY:
10189 ret = io_unregister_personality(ctx, nr_args);
10191 case IORING_REGISTER_ENABLE_RINGS:
10193 if (arg || nr_args)
10195 ret = io_register_enable_rings(ctx);
10197 case IORING_REGISTER_RESTRICTIONS:
10198 ret = io_register_restrictions(ctx, arg, nr_args);
10200 case IORING_REGISTER_FILES2:
10201 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10203 case IORING_REGISTER_FILES_UPDATE2:
10204 ret = io_register_rsrc_update(ctx, arg, nr_args,
10207 case IORING_REGISTER_BUFFERS2:
10208 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10210 case IORING_REGISTER_BUFFERS_UPDATE:
10211 ret = io_register_rsrc_update(ctx, arg, nr_args,
10212 IORING_RSRC_BUFFER);
10214 case IORING_REGISTER_IOWQ_AFF:
10216 if (!arg || !nr_args)
10218 ret = io_register_iowq_aff(ctx, arg, nr_args);
10220 case IORING_UNREGISTER_IOWQ_AFF:
10222 if (arg || nr_args)
10224 ret = io_unregister_iowq_aff(ctx);
10231 if (io_register_op_must_quiesce(opcode)) {
10232 /* bring the ctx back to life */
10233 percpu_ref_reinit(&ctx->refs);
10234 reinit_completion(&ctx->ref_comp);
10239 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10240 void __user *, arg, unsigned int, nr_args)
10242 struct io_ring_ctx *ctx;
10251 if (f.file->f_op != &io_uring_fops)
10254 ctx = f.file->private_data;
10256 io_run_task_work();
10258 mutex_lock(&ctx->uring_lock);
10259 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10260 mutex_unlock(&ctx->uring_lock);
10261 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10262 ctx->cq_ev_fd != NULL, ret);
10268 static int __init io_uring_init(void)
10270 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10271 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10272 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10275 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10276 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10277 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10278 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10279 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10280 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10281 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10282 BUILD_BUG_SQE_ELEM(8, __u64, off);
10283 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10284 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10285 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10286 BUILD_BUG_SQE_ELEM(24, __u32, len);
10287 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10288 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10289 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10290 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10291 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10292 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10293 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10296 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10299 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10300 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10301 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10302 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10303 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10304 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10305 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10306 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10308 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10309 sizeof(struct io_uring_rsrc_update));
10310 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10311 sizeof(struct io_uring_rsrc_update2));
10312 /* should fit into one byte */
10313 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10315 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10316 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10318 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10322 __initcall(io_uring_init);