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
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 (PAGE_SHIFT - 3)
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_COMPLETE_INLINE_BIT,
710 REQ_F_DONT_REISSUE_BIT,
713 REQ_F_ARM_LTIMEOUT_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 /* completion is deferred through io_comp_state */
754 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
755 /* caller should reissue async */
756 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
757 /* don't attempt request reissue, see io_rw_reissue() */
758 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
759 /* supports async reads */
760 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
761 /* supports async writes */
762 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
764 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
765 /* has creds assigned */
766 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
767 /* skip refcounting if not set */
768 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
769 /* there is a linked timeout that has to be armed */
770 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_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 /* requests with any of those set should undergo io_disarm_next() */
1038 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1040 static bool io_disarm_next(struct io_kiocb *req);
1041 static void io_uring_del_tctx_node(unsigned long index);
1042 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1043 struct task_struct *task,
1045 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1047 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1048 long res, unsigned int cflags);
1049 static void io_put_req(struct io_kiocb *req);
1050 static void io_put_req_deferred(struct io_kiocb *req);
1051 static void io_dismantle_req(struct io_kiocb *req);
1052 static void io_queue_linked_timeout(struct io_kiocb *req);
1053 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1054 struct io_uring_rsrc_update2 *up,
1056 static void io_clean_op(struct io_kiocb *req);
1057 static struct file *io_file_get(struct io_ring_ctx *ctx,
1058 struct io_kiocb *req, int fd, bool fixed);
1059 static void __io_queue_sqe(struct io_kiocb *req);
1060 static void io_rsrc_put_work(struct work_struct *work);
1062 static void io_req_task_queue(struct io_kiocb *req);
1063 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1064 static int io_req_prep_async(struct io_kiocb *req);
1066 static struct kmem_cache *req_cachep;
1068 static const struct file_operations io_uring_fops;
1070 struct sock *io_uring_get_socket(struct file *file)
1072 #if defined(CONFIG_UNIX)
1073 if (file->f_op == &io_uring_fops) {
1074 struct io_ring_ctx *ctx = file->private_data;
1076 return ctx->ring_sock->sk;
1081 EXPORT_SYMBOL(io_uring_get_socket);
1083 #define io_for_each_link(pos, head) \
1084 for (pos = (head); pos; pos = pos->link)
1087 * Shamelessly stolen from the mm implementation of page reference checking,
1088 * see commit f958d7b528b1 for details.
1090 #define req_ref_zero_or_close_to_overflow(req) \
1091 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1093 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1095 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1096 return atomic_inc_not_zero(&req->refs);
1099 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1101 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1104 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1105 return atomic_dec_and_test(&req->refs);
1108 static inline void req_ref_put(struct io_kiocb *req)
1110 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1111 WARN_ON_ONCE(req_ref_put_and_test(req));
1114 static inline void req_ref_get(struct io_kiocb *req)
1116 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1117 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1118 atomic_inc(&req->refs);
1121 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1123 if (!(req->flags & REQ_F_REFCOUNT)) {
1124 req->flags |= REQ_F_REFCOUNT;
1125 atomic_set(&req->refs, nr);
1129 static inline void io_req_set_refcount(struct io_kiocb *req)
1131 __io_req_set_refcount(req, 1);
1134 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1136 struct io_ring_ctx *ctx = req->ctx;
1138 if (!req->fixed_rsrc_refs) {
1139 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1140 percpu_ref_get(req->fixed_rsrc_refs);
1144 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1146 bool got = percpu_ref_tryget(ref);
1148 /* already at zero, wait for ->release() */
1150 wait_for_completion(compl);
1151 percpu_ref_resurrect(ref);
1153 percpu_ref_put(ref);
1156 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1159 struct io_kiocb *req;
1161 if (task && head->task != task)
1166 io_for_each_link(req, head) {
1167 if (req->flags & REQ_F_INFLIGHT)
1173 static inline void req_set_fail(struct io_kiocb *req)
1175 req->flags |= REQ_F_FAIL;
1178 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1180 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1182 complete(&ctx->ref_comp);
1185 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1187 return !req->timeout.off;
1190 static void io_fallback_req_func(struct work_struct *work)
1192 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1193 fallback_work.work);
1194 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1195 struct io_kiocb *req, *tmp;
1197 percpu_ref_get(&ctx->refs);
1198 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1199 req->io_task_work.func(req);
1200 percpu_ref_put(&ctx->refs);
1203 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1205 struct io_ring_ctx *ctx;
1208 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1213 * Use 5 bits less than the max cq entries, that should give us around
1214 * 32 entries per hash list if totally full and uniformly spread.
1216 hash_bits = ilog2(p->cq_entries);
1220 ctx->cancel_hash_bits = hash_bits;
1221 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1223 if (!ctx->cancel_hash)
1225 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1227 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1228 if (!ctx->dummy_ubuf)
1230 /* set invalid range, so io_import_fixed() fails meeting it */
1231 ctx->dummy_ubuf->ubuf = -1UL;
1233 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1234 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1237 ctx->flags = p->flags;
1238 init_waitqueue_head(&ctx->sqo_sq_wait);
1239 INIT_LIST_HEAD(&ctx->sqd_list);
1240 init_waitqueue_head(&ctx->poll_wait);
1241 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1242 init_completion(&ctx->ref_comp);
1243 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1244 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1245 mutex_init(&ctx->uring_lock);
1246 init_waitqueue_head(&ctx->cq_wait);
1247 spin_lock_init(&ctx->completion_lock);
1248 spin_lock_init(&ctx->timeout_lock);
1249 INIT_LIST_HEAD(&ctx->iopoll_list);
1250 INIT_LIST_HEAD(&ctx->defer_list);
1251 INIT_LIST_HEAD(&ctx->timeout_list);
1252 spin_lock_init(&ctx->rsrc_ref_lock);
1253 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1254 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1255 init_llist_head(&ctx->rsrc_put_llist);
1256 INIT_LIST_HEAD(&ctx->tctx_list);
1257 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1258 INIT_LIST_HEAD(&ctx->locked_free_list);
1259 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1262 kfree(ctx->dummy_ubuf);
1263 kfree(ctx->cancel_hash);
1268 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1270 struct io_rings *r = ctx->rings;
1272 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1276 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1278 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1279 struct io_ring_ctx *ctx = req->ctx;
1281 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1287 #define FFS_ASYNC_READ 0x1UL
1288 #define FFS_ASYNC_WRITE 0x2UL
1290 #define FFS_ISREG 0x4UL
1292 #define FFS_ISREG 0x0UL
1294 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1296 static inline bool io_req_ffs_set(struct io_kiocb *req)
1298 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1301 static void io_req_track_inflight(struct io_kiocb *req)
1303 if (!(req->flags & REQ_F_INFLIGHT)) {
1304 req->flags |= REQ_F_INFLIGHT;
1305 atomic_inc(¤t->io_uring->inflight_tracked);
1309 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1311 req->flags &= ~REQ_F_LINK_TIMEOUT;
1314 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1316 if (WARN_ON_ONCE(!req->link))
1319 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1320 req->flags |= REQ_F_LINK_TIMEOUT;
1322 /* linked timeouts should have two refs once prep'ed */
1323 io_req_set_refcount(req);
1324 __io_req_set_refcount(req->link, 2);
1328 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1330 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1332 return __io_prep_linked_timeout(req);
1335 static void io_prep_async_work(struct io_kiocb *req)
1337 const struct io_op_def *def = &io_op_defs[req->opcode];
1338 struct io_ring_ctx *ctx = req->ctx;
1340 if (!(req->flags & REQ_F_CREDS)) {
1341 req->flags |= REQ_F_CREDS;
1342 req->creds = get_current_cred();
1345 req->work.list.next = NULL;
1346 req->work.flags = 0;
1347 if (req->flags & REQ_F_FORCE_ASYNC)
1348 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1350 if (req->flags & REQ_F_ISREG) {
1351 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1352 io_wq_hash_work(&req->work, file_inode(req->file));
1353 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1354 if (def->unbound_nonreg_file)
1355 req->work.flags |= IO_WQ_WORK_UNBOUND;
1358 switch (req->opcode) {
1359 case IORING_OP_SPLICE:
1361 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1362 req->work.flags |= IO_WQ_WORK_UNBOUND;
1367 static void io_prep_async_link(struct io_kiocb *req)
1369 struct io_kiocb *cur;
1371 if (req->flags & REQ_F_LINK_TIMEOUT) {
1372 struct io_ring_ctx *ctx = req->ctx;
1374 spin_lock(&ctx->completion_lock);
1375 io_for_each_link(cur, req)
1376 io_prep_async_work(cur);
1377 spin_unlock(&ctx->completion_lock);
1379 io_for_each_link(cur, req)
1380 io_prep_async_work(cur);
1384 static void io_queue_async_work(struct io_kiocb *req)
1386 struct io_ring_ctx *ctx = req->ctx;
1387 struct io_kiocb *link = io_prep_linked_timeout(req);
1388 struct io_uring_task *tctx = req->task->io_uring;
1391 BUG_ON(!tctx->io_wq);
1393 /* init ->work of the whole link before punting */
1394 io_prep_async_link(req);
1397 * Not expected to happen, but if we do have a bug where this _can_
1398 * happen, catch it here and ensure the request is marked as
1399 * canceled. That will make io-wq go through the usual work cancel
1400 * procedure rather than attempt to run this request (or create a new
1403 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1404 req->work.flags |= IO_WQ_WORK_CANCEL;
1406 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1407 &req->work, req->flags);
1408 io_wq_enqueue(tctx->io_wq, &req->work);
1410 io_queue_linked_timeout(link);
1413 static void io_kill_timeout(struct io_kiocb *req, int status)
1414 __must_hold(&req->ctx->completion_lock)
1415 __must_hold(&req->ctx->timeout_lock)
1417 struct io_timeout_data *io = req->async_data;
1419 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1420 atomic_set(&req->ctx->cq_timeouts,
1421 atomic_read(&req->ctx->cq_timeouts) + 1);
1422 list_del_init(&req->timeout.list);
1423 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1424 io_put_req_deferred(req);
1428 static void io_queue_deferred(struct io_ring_ctx *ctx)
1430 while (!list_empty(&ctx->defer_list)) {
1431 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1432 struct io_defer_entry, list);
1434 if (req_need_defer(de->req, de->seq))
1436 list_del_init(&de->list);
1437 io_req_task_queue(de->req);
1442 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1443 __must_hold(&ctx->completion_lock)
1445 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1447 spin_lock_irq(&ctx->timeout_lock);
1448 while (!list_empty(&ctx->timeout_list)) {
1449 u32 events_needed, events_got;
1450 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1451 struct io_kiocb, timeout.list);
1453 if (io_is_timeout_noseq(req))
1457 * Since seq can easily wrap around over time, subtract
1458 * the last seq at which timeouts were flushed before comparing.
1459 * Assuming not more than 2^31-1 events have happened since,
1460 * these subtractions won't have wrapped, so we can check if
1461 * target is in [last_seq, current_seq] by comparing the two.
1463 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1464 events_got = seq - ctx->cq_last_tm_flush;
1465 if (events_got < events_needed)
1468 list_del_init(&req->timeout.list);
1469 io_kill_timeout(req, 0);
1471 ctx->cq_last_tm_flush = seq;
1472 spin_unlock_irq(&ctx->timeout_lock);
1475 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1477 if (ctx->off_timeout_used)
1478 io_flush_timeouts(ctx);
1479 if (ctx->drain_active)
1480 io_queue_deferred(ctx);
1483 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1485 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1486 __io_commit_cqring_flush(ctx);
1487 /* order cqe stores with ring update */
1488 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1491 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1493 struct io_rings *r = ctx->rings;
1495 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1498 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1500 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1503 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1505 struct io_rings *rings = ctx->rings;
1506 unsigned tail, mask = ctx->cq_entries - 1;
1509 * writes to the cq entry need to come after reading head; the
1510 * control dependency is enough as we're using WRITE_ONCE to
1513 if (__io_cqring_events(ctx) == ctx->cq_entries)
1516 tail = ctx->cached_cq_tail++;
1517 return &rings->cqes[tail & mask];
1520 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1522 if (likely(!ctx->cq_ev_fd))
1524 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1526 return !ctx->eventfd_async || io_wq_current_is_worker();
1529 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1532 * wake_up_all() may seem excessive, but io_wake_function() and
1533 * io_should_wake() handle the termination of the loop and only
1534 * wake as many waiters as we need to.
1536 if (wq_has_sleeper(&ctx->cq_wait))
1537 wake_up_all(&ctx->cq_wait);
1538 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1539 wake_up(&ctx->sq_data->wait);
1540 if (io_should_trigger_evfd(ctx))
1541 eventfd_signal(ctx->cq_ev_fd, 1);
1542 if (waitqueue_active(&ctx->poll_wait)) {
1543 wake_up_interruptible(&ctx->poll_wait);
1544 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1548 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1550 if (ctx->flags & IORING_SETUP_SQPOLL) {
1551 if (wq_has_sleeper(&ctx->cq_wait))
1552 wake_up_all(&ctx->cq_wait);
1554 if (io_should_trigger_evfd(ctx))
1555 eventfd_signal(ctx->cq_ev_fd, 1);
1556 if (waitqueue_active(&ctx->poll_wait)) {
1557 wake_up_interruptible(&ctx->poll_wait);
1558 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1562 /* Returns true if there are no backlogged entries after the flush */
1563 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1565 bool all_flushed, posted;
1567 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1571 spin_lock(&ctx->completion_lock);
1572 while (!list_empty(&ctx->cq_overflow_list)) {
1573 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1574 struct io_overflow_cqe *ocqe;
1578 ocqe = list_first_entry(&ctx->cq_overflow_list,
1579 struct io_overflow_cqe, list);
1581 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1583 io_account_cq_overflow(ctx);
1586 list_del(&ocqe->list);
1590 all_flushed = list_empty(&ctx->cq_overflow_list);
1592 clear_bit(0, &ctx->check_cq_overflow);
1593 WRITE_ONCE(ctx->rings->sq_flags,
1594 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1598 io_commit_cqring(ctx);
1599 spin_unlock(&ctx->completion_lock);
1601 io_cqring_ev_posted(ctx);
1605 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1609 if (test_bit(0, &ctx->check_cq_overflow)) {
1610 /* iopoll syncs against uring_lock, not completion_lock */
1611 if (ctx->flags & IORING_SETUP_IOPOLL)
1612 mutex_lock(&ctx->uring_lock);
1613 ret = __io_cqring_overflow_flush(ctx, false);
1614 if (ctx->flags & IORING_SETUP_IOPOLL)
1615 mutex_unlock(&ctx->uring_lock);
1621 /* must to be called somewhat shortly after putting a request */
1622 static inline void io_put_task(struct task_struct *task, int nr)
1624 struct io_uring_task *tctx = task->io_uring;
1626 if (likely(task == current)) {
1627 tctx->cached_refs += nr;
1629 percpu_counter_sub(&tctx->inflight, nr);
1630 if (unlikely(atomic_read(&tctx->in_idle)))
1631 wake_up(&tctx->wait);
1632 put_task_struct_many(task, nr);
1636 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1637 long res, unsigned int cflags)
1639 struct io_overflow_cqe *ocqe;
1641 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1644 * If we're in ring overflow flush mode, or in task cancel mode,
1645 * or cannot allocate an overflow entry, then we need to drop it
1648 io_account_cq_overflow(ctx);
1651 if (list_empty(&ctx->cq_overflow_list)) {
1652 set_bit(0, &ctx->check_cq_overflow);
1653 WRITE_ONCE(ctx->rings->sq_flags,
1654 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1657 ocqe->cqe.user_data = user_data;
1658 ocqe->cqe.res = res;
1659 ocqe->cqe.flags = cflags;
1660 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1664 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1665 long res, unsigned int cflags)
1667 struct io_uring_cqe *cqe;
1669 trace_io_uring_complete(ctx, user_data, res, cflags);
1672 * If we can't get a cq entry, userspace overflowed the
1673 * submission (by quite a lot). Increment the overflow count in
1676 cqe = io_get_cqe(ctx);
1678 WRITE_ONCE(cqe->user_data, user_data);
1679 WRITE_ONCE(cqe->res, res);
1680 WRITE_ONCE(cqe->flags, cflags);
1683 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1686 /* not as hot to bloat with inlining */
1687 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1688 long res, unsigned int cflags)
1690 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1693 static void io_req_complete_post(struct io_kiocb *req, long res,
1694 unsigned int cflags)
1696 struct io_ring_ctx *ctx = req->ctx;
1698 spin_lock(&ctx->completion_lock);
1699 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1701 * If we're the last reference to this request, add to our locked
1704 if (req_ref_put_and_test(req)) {
1705 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1706 if (req->flags & IO_DISARM_MASK)
1707 io_disarm_next(req);
1709 io_req_task_queue(req->link);
1713 io_dismantle_req(req);
1714 io_put_task(req->task, 1);
1715 list_add(&req->inflight_entry, &ctx->locked_free_list);
1716 ctx->locked_free_nr++;
1718 if (!percpu_ref_tryget(&ctx->refs))
1721 io_commit_cqring(ctx);
1722 spin_unlock(&ctx->completion_lock);
1725 io_cqring_ev_posted(ctx);
1726 percpu_ref_put(&ctx->refs);
1730 static inline bool io_req_needs_clean(struct io_kiocb *req)
1732 return req->flags & IO_REQ_CLEAN_FLAGS;
1735 static void io_req_complete_state(struct io_kiocb *req, long res,
1736 unsigned int cflags)
1738 if (io_req_needs_clean(req))
1741 req->compl.cflags = cflags;
1742 req->flags |= REQ_F_COMPLETE_INLINE;
1745 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1746 long res, unsigned cflags)
1748 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1749 io_req_complete_state(req, res, cflags);
1751 io_req_complete_post(req, res, cflags);
1754 static inline void io_req_complete(struct io_kiocb *req, long res)
1756 __io_req_complete(req, 0, res, 0);
1759 static void io_req_complete_failed(struct io_kiocb *req, long res)
1762 io_req_complete_post(req, res, 0);
1766 * Don't initialise the fields below on every allocation, but do that in
1767 * advance and keep them valid across allocations.
1769 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1773 req->async_data = NULL;
1774 /* not necessary, but safer to zero */
1778 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1779 struct io_submit_state *state)
1781 spin_lock(&ctx->completion_lock);
1782 list_splice_init(&ctx->locked_free_list, &state->free_list);
1783 ctx->locked_free_nr = 0;
1784 spin_unlock(&ctx->completion_lock);
1787 /* Returns true IFF there are requests in the cache */
1788 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1790 struct io_submit_state *state = &ctx->submit_state;
1794 * If we have more than a batch's worth of requests in our IRQ side
1795 * locked cache, grab the lock and move them over to our submission
1798 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1799 io_flush_cached_locked_reqs(ctx, state);
1801 nr = state->free_reqs;
1802 while (!list_empty(&state->free_list)) {
1803 struct io_kiocb *req = list_first_entry(&state->free_list,
1804 struct io_kiocb, inflight_entry);
1806 list_del(&req->inflight_entry);
1807 state->reqs[nr++] = req;
1808 if (nr == ARRAY_SIZE(state->reqs))
1812 state->free_reqs = nr;
1817 * A request might get retired back into the request caches even before opcode
1818 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1819 * Because of that, io_alloc_req() should be called only under ->uring_lock
1820 * and with extra caution to not get a request that is still worked on.
1822 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1823 __must_hold(&ctx->uring_lock)
1825 struct io_submit_state *state = &ctx->submit_state;
1826 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1829 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1831 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1834 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1838 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1839 * retry single alloc to be on the safe side.
1841 if (unlikely(ret <= 0)) {
1842 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1843 if (!state->reqs[0])
1848 for (i = 0; i < ret; i++)
1849 io_preinit_req(state->reqs[i], ctx);
1850 state->free_reqs = ret;
1853 return state->reqs[state->free_reqs];
1856 static inline void io_put_file(struct file *file)
1862 static void io_dismantle_req(struct io_kiocb *req)
1864 unsigned int flags = req->flags;
1866 if (io_req_needs_clean(req))
1868 if (!(flags & REQ_F_FIXED_FILE))
1869 io_put_file(req->file);
1870 if (req->fixed_rsrc_refs)
1871 percpu_ref_put(req->fixed_rsrc_refs);
1872 if (req->async_data) {
1873 kfree(req->async_data);
1874 req->async_data = NULL;
1878 static void __io_free_req(struct io_kiocb *req)
1880 struct io_ring_ctx *ctx = req->ctx;
1882 io_dismantle_req(req);
1883 io_put_task(req->task, 1);
1885 spin_lock(&ctx->completion_lock);
1886 list_add(&req->inflight_entry, &ctx->locked_free_list);
1887 ctx->locked_free_nr++;
1888 spin_unlock(&ctx->completion_lock);
1890 percpu_ref_put(&ctx->refs);
1893 static inline void io_remove_next_linked(struct io_kiocb *req)
1895 struct io_kiocb *nxt = req->link;
1897 req->link = nxt->link;
1901 static bool io_kill_linked_timeout(struct io_kiocb *req)
1902 __must_hold(&req->ctx->completion_lock)
1903 __must_hold(&req->ctx->timeout_lock)
1905 struct io_kiocb *link = req->link;
1907 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1908 struct io_timeout_data *io = link->async_data;
1910 io_remove_next_linked(req);
1911 link->timeout.head = NULL;
1912 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1913 io_cqring_fill_event(link->ctx, link->user_data,
1915 io_put_req_deferred(link);
1922 static void io_fail_links(struct io_kiocb *req)
1923 __must_hold(&req->ctx->completion_lock)
1925 struct io_kiocb *nxt, *link = req->link;
1932 trace_io_uring_fail_link(req, link);
1933 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1934 io_put_req_deferred(link);
1939 static bool io_disarm_next(struct io_kiocb *req)
1940 __must_hold(&req->ctx->completion_lock)
1942 bool posted = false;
1944 if (req->flags & REQ_F_ARM_LTIMEOUT) {
1945 struct io_kiocb *link = req->link;
1947 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1948 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1949 io_remove_next_linked(req);
1950 io_cqring_fill_event(link->ctx, link->user_data,
1952 io_put_req_deferred(link);
1955 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
1956 struct io_ring_ctx *ctx = req->ctx;
1958 spin_lock_irq(&ctx->timeout_lock);
1959 posted = io_kill_linked_timeout(req);
1960 spin_unlock_irq(&ctx->timeout_lock);
1962 if (unlikely((req->flags & REQ_F_FAIL) &&
1963 !(req->flags & REQ_F_HARDLINK))) {
1964 posted |= (req->link != NULL);
1970 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1972 struct io_kiocb *nxt;
1975 * If LINK is set, we have dependent requests in this chain. If we
1976 * didn't fail this request, queue the first one up, moving any other
1977 * dependencies to the next request. In case of failure, fail the rest
1980 if (req->flags & IO_DISARM_MASK) {
1981 struct io_ring_ctx *ctx = req->ctx;
1984 spin_lock(&ctx->completion_lock);
1985 posted = io_disarm_next(req);
1987 io_commit_cqring(req->ctx);
1988 spin_unlock(&ctx->completion_lock);
1990 io_cqring_ev_posted(ctx);
1997 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1999 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2001 return __io_req_find_next(req);
2004 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2008 if (ctx->submit_state.compl_nr) {
2009 mutex_lock(&ctx->uring_lock);
2010 io_submit_flush_completions(ctx);
2011 mutex_unlock(&ctx->uring_lock);
2013 percpu_ref_put(&ctx->refs);
2016 static void tctx_task_work(struct callback_head *cb)
2018 struct io_ring_ctx *ctx = NULL;
2019 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2023 struct io_wq_work_node *node;
2025 spin_lock_irq(&tctx->task_lock);
2026 node = tctx->task_list.first;
2027 INIT_WQ_LIST(&tctx->task_list);
2029 tctx->task_running = false;
2030 spin_unlock_irq(&tctx->task_lock);
2035 struct io_wq_work_node *next = node->next;
2036 struct io_kiocb *req = container_of(node, struct io_kiocb,
2039 if (req->ctx != ctx) {
2040 ctx_flush_and_put(ctx);
2042 percpu_ref_get(&ctx->refs);
2044 req->io_task_work.func(req);
2051 ctx_flush_and_put(ctx);
2054 static void io_req_task_work_add(struct io_kiocb *req)
2056 struct task_struct *tsk = req->task;
2057 struct io_uring_task *tctx = tsk->io_uring;
2058 enum task_work_notify_mode notify;
2059 struct io_wq_work_node *node;
2060 unsigned long flags;
2063 WARN_ON_ONCE(!tctx);
2065 spin_lock_irqsave(&tctx->task_lock, flags);
2066 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2067 running = tctx->task_running;
2069 tctx->task_running = true;
2070 spin_unlock_irqrestore(&tctx->task_lock, flags);
2072 /* task_work already pending, we're done */
2077 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2078 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2079 * processing task_work. There's no reliable way to tell if TWA_RESUME
2082 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2083 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2084 wake_up_process(tsk);
2088 spin_lock_irqsave(&tctx->task_lock, flags);
2089 tctx->task_running = false;
2090 node = tctx->task_list.first;
2091 INIT_WQ_LIST(&tctx->task_list);
2092 spin_unlock_irqrestore(&tctx->task_lock, flags);
2095 req = container_of(node, struct io_kiocb, io_task_work.node);
2097 if (llist_add(&req->io_task_work.fallback_node,
2098 &req->ctx->fallback_llist))
2099 schedule_delayed_work(&req->ctx->fallback_work, 1);
2103 static void io_req_task_cancel(struct io_kiocb *req)
2105 struct io_ring_ctx *ctx = req->ctx;
2107 /* ctx is guaranteed to stay alive while we hold uring_lock */
2108 mutex_lock(&ctx->uring_lock);
2109 io_req_complete_failed(req, req->result);
2110 mutex_unlock(&ctx->uring_lock);
2113 static void io_req_task_submit(struct io_kiocb *req)
2115 struct io_ring_ctx *ctx = req->ctx;
2117 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2118 mutex_lock(&ctx->uring_lock);
2119 /* req->task == current here, checking PF_EXITING is safe */
2120 if (likely(!(req->task->flags & PF_EXITING)))
2121 __io_queue_sqe(req);
2123 io_req_complete_failed(req, -EFAULT);
2124 mutex_unlock(&ctx->uring_lock);
2127 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2130 req->io_task_work.func = io_req_task_cancel;
2131 io_req_task_work_add(req);
2134 static void io_req_task_queue(struct io_kiocb *req)
2136 req->io_task_work.func = io_req_task_submit;
2137 io_req_task_work_add(req);
2140 static void io_req_task_queue_reissue(struct io_kiocb *req)
2142 req->io_task_work.func = io_queue_async_work;
2143 io_req_task_work_add(req);
2146 static inline void io_queue_next(struct io_kiocb *req)
2148 struct io_kiocb *nxt = io_req_find_next(req);
2151 io_req_task_queue(nxt);
2154 static void io_free_req(struct io_kiocb *req)
2161 struct task_struct *task;
2166 static inline void io_init_req_batch(struct req_batch *rb)
2173 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2174 struct req_batch *rb)
2177 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2179 io_put_task(rb->task, rb->task_refs);
2182 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2183 struct io_submit_state *state)
2186 io_dismantle_req(req);
2188 if (req->task != rb->task) {
2190 io_put_task(rb->task, rb->task_refs);
2191 rb->task = req->task;
2197 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2198 state->reqs[state->free_reqs++] = req;
2200 list_add(&req->inflight_entry, &state->free_list);
2203 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2204 __must_hold(&ctx->uring_lock)
2206 struct io_submit_state *state = &ctx->submit_state;
2207 int i, nr = state->compl_nr;
2208 struct req_batch rb;
2210 spin_lock(&ctx->completion_lock);
2211 for (i = 0; i < nr; i++) {
2212 struct io_kiocb *req = state->compl_reqs[i];
2214 __io_cqring_fill_event(ctx, req->user_data, req->result,
2217 io_commit_cqring(ctx);
2218 spin_unlock(&ctx->completion_lock);
2219 io_cqring_ev_posted(ctx);
2221 io_init_req_batch(&rb);
2222 for (i = 0; i < nr; i++) {
2223 struct io_kiocb *req = state->compl_reqs[i];
2225 if (req_ref_put_and_test(req))
2226 io_req_free_batch(&rb, req, &ctx->submit_state);
2229 io_req_free_batch_finish(ctx, &rb);
2230 state->compl_nr = 0;
2234 * Drop reference to request, return next in chain (if there is one) if this
2235 * was the last reference to this request.
2237 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2239 struct io_kiocb *nxt = NULL;
2241 if (req_ref_put_and_test(req)) {
2242 nxt = io_req_find_next(req);
2248 static inline void io_put_req(struct io_kiocb *req)
2250 if (req_ref_put_and_test(req))
2254 static inline void io_put_req_deferred(struct io_kiocb *req)
2256 if (req_ref_put_and_test(req)) {
2257 req->io_task_work.func = io_free_req;
2258 io_req_task_work_add(req);
2262 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2264 /* See comment at the top of this file */
2266 return __io_cqring_events(ctx);
2269 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2271 struct io_rings *rings = ctx->rings;
2273 /* make sure SQ entry isn't read before tail */
2274 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2277 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2279 unsigned int cflags;
2281 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2282 cflags |= IORING_CQE_F_BUFFER;
2283 req->flags &= ~REQ_F_BUFFER_SELECTED;
2288 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2290 struct io_buffer *kbuf;
2292 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2294 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2295 return io_put_kbuf(req, kbuf);
2298 static inline bool io_run_task_work(void)
2300 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2301 __set_current_state(TASK_RUNNING);
2302 tracehook_notify_signal();
2310 * Find and free completed poll iocbs
2312 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2313 struct list_head *done)
2315 struct req_batch rb;
2316 struct io_kiocb *req;
2318 /* order with ->result store in io_complete_rw_iopoll() */
2321 io_init_req_batch(&rb);
2322 while (!list_empty(done)) {
2323 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2324 list_del(&req->inflight_entry);
2326 if (READ_ONCE(req->result) == -EAGAIN &&
2327 !(req->flags & REQ_F_DONT_REISSUE)) {
2328 req->iopoll_completed = 0;
2329 io_req_task_queue_reissue(req);
2333 __io_cqring_fill_event(ctx, req->user_data, req->result,
2334 io_put_rw_kbuf(req));
2337 if (req_ref_put_and_test(req))
2338 io_req_free_batch(&rb, req, &ctx->submit_state);
2341 io_commit_cqring(ctx);
2342 io_cqring_ev_posted_iopoll(ctx);
2343 io_req_free_batch_finish(ctx, &rb);
2346 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2349 struct io_kiocb *req, *tmp;
2354 * Only spin for completions if we don't have multiple devices hanging
2355 * off our complete list, and we're under the requested amount.
2357 spin = !ctx->poll_multi_queue && *nr_events < min;
2359 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2360 struct kiocb *kiocb = &req->rw.kiocb;
2364 * Move completed and retryable entries to our local lists.
2365 * If we find a request that requires polling, break out
2366 * and complete those lists first, if we have entries there.
2368 if (READ_ONCE(req->iopoll_completed)) {
2369 list_move_tail(&req->inflight_entry, &done);
2372 if (!list_empty(&done))
2375 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2376 if (unlikely(ret < 0))
2381 /* iopoll may have completed current req */
2382 if (READ_ONCE(req->iopoll_completed))
2383 list_move_tail(&req->inflight_entry, &done);
2386 if (!list_empty(&done))
2387 io_iopoll_complete(ctx, nr_events, &done);
2393 * We can't just wait for polled events to come to us, we have to actively
2394 * find and complete them.
2396 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2398 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2401 mutex_lock(&ctx->uring_lock);
2402 while (!list_empty(&ctx->iopoll_list)) {
2403 unsigned int nr_events = 0;
2405 io_do_iopoll(ctx, &nr_events, 0);
2407 /* let it sleep and repeat later if can't complete a request */
2411 * Ensure we allow local-to-the-cpu processing to take place,
2412 * in this case we need to ensure that we reap all events.
2413 * Also let task_work, etc. to progress by releasing the mutex
2415 if (need_resched()) {
2416 mutex_unlock(&ctx->uring_lock);
2418 mutex_lock(&ctx->uring_lock);
2421 mutex_unlock(&ctx->uring_lock);
2424 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2426 unsigned int nr_events = 0;
2430 * We disallow the app entering submit/complete with polling, but we
2431 * still need to lock the ring to prevent racing with polled issue
2432 * that got punted to a workqueue.
2434 mutex_lock(&ctx->uring_lock);
2436 * Don't enter poll loop if we already have events pending.
2437 * If we do, we can potentially be spinning for commands that
2438 * already triggered a CQE (eg in error).
2440 if (test_bit(0, &ctx->check_cq_overflow))
2441 __io_cqring_overflow_flush(ctx, false);
2442 if (io_cqring_events(ctx))
2446 * If a submit got punted to a workqueue, we can have the
2447 * application entering polling for a command before it gets
2448 * issued. That app will hold the uring_lock for the duration
2449 * of the poll right here, so we need to take a breather every
2450 * now and then to ensure that the issue has a chance to add
2451 * the poll to the issued list. Otherwise we can spin here
2452 * forever, while the workqueue is stuck trying to acquire the
2455 if (list_empty(&ctx->iopoll_list)) {
2456 u32 tail = ctx->cached_cq_tail;
2458 mutex_unlock(&ctx->uring_lock);
2460 mutex_lock(&ctx->uring_lock);
2462 /* some requests don't go through iopoll_list */
2463 if (tail != ctx->cached_cq_tail ||
2464 list_empty(&ctx->iopoll_list))
2467 ret = io_do_iopoll(ctx, &nr_events, min);
2468 } while (!ret && nr_events < min && !need_resched());
2470 mutex_unlock(&ctx->uring_lock);
2474 static void kiocb_end_write(struct io_kiocb *req)
2477 * Tell lockdep we inherited freeze protection from submission
2480 if (req->flags & REQ_F_ISREG) {
2481 struct super_block *sb = file_inode(req->file)->i_sb;
2483 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2489 static bool io_resubmit_prep(struct io_kiocb *req)
2491 struct io_async_rw *rw = req->async_data;
2494 return !io_req_prep_async(req);
2495 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2496 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2500 static bool io_rw_should_reissue(struct io_kiocb *req)
2502 umode_t mode = file_inode(req->file)->i_mode;
2503 struct io_ring_ctx *ctx = req->ctx;
2505 if (!S_ISBLK(mode) && !S_ISREG(mode))
2507 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2508 !(ctx->flags & IORING_SETUP_IOPOLL)))
2511 * If ref is dying, we might be running poll reap from the exit work.
2512 * Don't attempt to reissue from that path, just let it fail with
2515 if (percpu_ref_is_dying(&ctx->refs))
2518 * Play it safe and assume not safe to re-import and reissue if we're
2519 * not in the original thread group (or in task context).
2521 if (!same_thread_group(req->task, current) || !in_task())
2526 static bool io_resubmit_prep(struct io_kiocb *req)
2530 static bool io_rw_should_reissue(struct io_kiocb *req)
2536 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2538 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2539 kiocb_end_write(req);
2540 if (res != req->result) {
2541 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2542 io_rw_should_reissue(req)) {
2543 req->flags |= REQ_F_REISSUE;
2552 static void io_req_task_complete(struct io_kiocb *req)
2554 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2557 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2558 unsigned int issue_flags)
2560 if (__io_complete_rw_common(req, res))
2562 io_req_task_complete(req);
2565 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2567 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2569 if (__io_complete_rw_common(req, res))
2572 req->io_task_work.func = io_req_task_complete;
2573 io_req_task_work_add(req);
2576 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2578 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2580 if (kiocb->ki_flags & IOCB_WRITE)
2581 kiocb_end_write(req);
2582 if (unlikely(res != req->result)) {
2583 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2584 io_resubmit_prep(req))) {
2586 req->flags |= REQ_F_DONT_REISSUE;
2590 WRITE_ONCE(req->result, res);
2591 /* order with io_iopoll_complete() checking ->result */
2593 WRITE_ONCE(req->iopoll_completed, 1);
2597 * After the iocb has been issued, it's safe to be found on the poll list.
2598 * Adding the kiocb to the list AFTER submission ensures that we don't
2599 * find it from a io_do_iopoll() thread before the issuer is done
2600 * accessing the kiocb cookie.
2602 static void io_iopoll_req_issued(struct io_kiocb *req)
2604 struct io_ring_ctx *ctx = req->ctx;
2605 const bool in_async = io_wq_current_is_worker();
2607 /* workqueue context doesn't hold uring_lock, grab it now */
2608 if (unlikely(in_async))
2609 mutex_lock(&ctx->uring_lock);
2612 * Track whether we have multiple files in our lists. This will impact
2613 * how we do polling eventually, not spinning if we're on potentially
2614 * different devices.
2616 if (list_empty(&ctx->iopoll_list)) {
2617 ctx->poll_multi_queue = false;
2618 } else if (!ctx->poll_multi_queue) {
2619 struct io_kiocb *list_req;
2620 unsigned int queue_num0, queue_num1;
2622 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2625 if (list_req->file != req->file) {
2626 ctx->poll_multi_queue = true;
2628 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2629 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2630 if (queue_num0 != queue_num1)
2631 ctx->poll_multi_queue = true;
2636 * For fast devices, IO may have already completed. If it has, add
2637 * it to the front so we find it first.
2639 if (READ_ONCE(req->iopoll_completed))
2640 list_add(&req->inflight_entry, &ctx->iopoll_list);
2642 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2644 if (unlikely(in_async)) {
2646 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2647 * in sq thread task context or in io worker task context. If
2648 * current task context is sq thread, we don't need to check
2649 * whether should wake up sq thread.
2651 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2652 wq_has_sleeper(&ctx->sq_data->wait))
2653 wake_up(&ctx->sq_data->wait);
2655 mutex_unlock(&ctx->uring_lock);
2659 static bool io_bdev_nowait(struct block_device *bdev)
2661 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2665 * If we tracked the file through the SCM inflight mechanism, we could support
2666 * any file. For now, just ensure that anything potentially problematic is done
2669 static bool __io_file_supports_nowait(struct file *file, int rw)
2671 umode_t mode = file_inode(file)->i_mode;
2673 if (S_ISBLK(mode)) {
2674 if (IS_ENABLED(CONFIG_BLOCK) &&
2675 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2681 if (S_ISREG(mode)) {
2682 if (IS_ENABLED(CONFIG_BLOCK) &&
2683 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2684 file->f_op != &io_uring_fops)
2689 /* any ->read/write should understand O_NONBLOCK */
2690 if (file->f_flags & O_NONBLOCK)
2693 if (!(file->f_mode & FMODE_NOWAIT))
2697 return file->f_op->read_iter != NULL;
2699 return file->f_op->write_iter != NULL;
2702 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2704 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2706 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2709 return __io_file_supports_nowait(req->file, rw);
2712 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2714 struct io_ring_ctx *ctx = req->ctx;
2715 struct kiocb *kiocb = &req->rw.kiocb;
2716 struct file *file = req->file;
2720 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2721 req->flags |= REQ_F_ISREG;
2723 kiocb->ki_pos = READ_ONCE(sqe->off);
2724 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2725 req->flags |= REQ_F_CUR_POS;
2726 kiocb->ki_pos = file->f_pos;
2728 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2729 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2730 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2734 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2735 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2736 req->flags |= REQ_F_NOWAIT;
2738 ioprio = READ_ONCE(sqe->ioprio);
2740 ret = ioprio_check_cap(ioprio);
2744 kiocb->ki_ioprio = ioprio;
2746 kiocb->ki_ioprio = get_current_ioprio();
2748 if (ctx->flags & IORING_SETUP_IOPOLL) {
2749 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2750 !kiocb->ki_filp->f_op->iopoll)
2753 kiocb->ki_flags |= IOCB_HIPRI;
2754 kiocb->ki_complete = io_complete_rw_iopoll;
2755 req->iopoll_completed = 0;
2757 if (kiocb->ki_flags & IOCB_HIPRI)
2759 kiocb->ki_complete = io_complete_rw;
2762 if (req->opcode == IORING_OP_READ_FIXED ||
2763 req->opcode == IORING_OP_WRITE_FIXED) {
2765 io_req_set_rsrc_node(req);
2768 req->rw.addr = READ_ONCE(sqe->addr);
2769 req->rw.len = READ_ONCE(sqe->len);
2770 req->buf_index = READ_ONCE(sqe->buf_index);
2774 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2780 case -ERESTARTNOINTR:
2781 case -ERESTARTNOHAND:
2782 case -ERESTART_RESTARTBLOCK:
2784 * We can't just restart the syscall, since previously
2785 * submitted sqes may already be in progress. Just fail this
2791 kiocb->ki_complete(kiocb, ret, 0);
2795 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2796 unsigned int issue_flags)
2798 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2799 struct io_async_rw *io = req->async_data;
2800 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2802 /* add previously done IO, if any */
2803 if (io && io->bytes_done > 0) {
2805 ret = io->bytes_done;
2807 ret += io->bytes_done;
2810 if (req->flags & REQ_F_CUR_POS)
2811 req->file->f_pos = kiocb->ki_pos;
2812 if (ret >= 0 && check_reissue)
2813 __io_complete_rw(req, ret, 0, issue_flags);
2815 io_rw_done(kiocb, ret);
2817 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2818 req->flags &= ~REQ_F_REISSUE;
2819 if (io_resubmit_prep(req)) {
2820 io_req_task_queue_reissue(req);
2823 __io_req_complete(req, issue_flags, ret,
2824 io_put_rw_kbuf(req));
2829 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2830 struct io_mapped_ubuf *imu)
2832 size_t len = req->rw.len;
2833 u64 buf_end, buf_addr = req->rw.addr;
2836 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2838 /* not inside the mapped region */
2839 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2843 * May not be a start of buffer, set size appropriately
2844 * and advance us to the beginning.
2846 offset = buf_addr - imu->ubuf;
2847 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2851 * Don't use iov_iter_advance() here, as it's really slow for
2852 * using the latter parts of a big fixed buffer - it iterates
2853 * over each segment manually. We can cheat a bit here, because
2856 * 1) it's a BVEC iter, we set it up
2857 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2858 * first and last bvec
2860 * So just find our index, and adjust the iterator afterwards.
2861 * If the offset is within the first bvec (or the whole first
2862 * bvec, just use iov_iter_advance(). This makes it easier
2863 * since we can just skip the first segment, which may not
2864 * be PAGE_SIZE aligned.
2866 const struct bio_vec *bvec = imu->bvec;
2868 if (offset <= bvec->bv_len) {
2869 iov_iter_advance(iter, offset);
2871 unsigned long seg_skip;
2873 /* skip first vec */
2874 offset -= bvec->bv_len;
2875 seg_skip = 1 + (offset >> PAGE_SHIFT);
2877 iter->bvec = bvec + seg_skip;
2878 iter->nr_segs -= seg_skip;
2879 iter->count -= bvec->bv_len + offset;
2880 iter->iov_offset = offset & ~PAGE_MASK;
2887 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2889 struct io_ring_ctx *ctx = req->ctx;
2890 struct io_mapped_ubuf *imu = req->imu;
2891 u16 index, buf_index = req->buf_index;
2894 if (unlikely(buf_index >= ctx->nr_user_bufs))
2896 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2897 imu = READ_ONCE(ctx->user_bufs[index]);
2900 return __io_import_fixed(req, rw, iter, imu);
2903 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2906 mutex_unlock(&ctx->uring_lock);
2909 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2912 * "Normal" inline submissions always hold the uring_lock, since we
2913 * grab it from the system call. Same is true for the SQPOLL offload.
2914 * The only exception is when we've detached the request and issue it
2915 * from an async worker thread, grab the lock for that case.
2918 mutex_lock(&ctx->uring_lock);
2921 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2922 int bgid, struct io_buffer *kbuf,
2925 struct io_buffer *head;
2927 if (req->flags & REQ_F_BUFFER_SELECTED)
2930 io_ring_submit_lock(req->ctx, needs_lock);
2932 lockdep_assert_held(&req->ctx->uring_lock);
2934 head = xa_load(&req->ctx->io_buffers, bgid);
2936 if (!list_empty(&head->list)) {
2937 kbuf = list_last_entry(&head->list, struct io_buffer,
2939 list_del(&kbuf->list);
2942 xa_erase(&req->ctx->io_buffers, bgid);
2944 if (*len > kbuf->len)
2947 kbuf = ERR_PTR(-ENOBUFS);
2950 io_ring_submit_unlock(req->ctx, needs_lock);
2955 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2958 struct io_buffer *kbuf;
2961 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2962 bgid = req->buf_index;
2963 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2966 req->rw.addr = (u64) (unsigned long) kbuf;
2967 req->flags |= REQ_F_BUFFER_SELECTED;
2968 return u64_to_user_ptr(kbuf->addr);
2971 #ifdef CONFIG_COMPAT
2972 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2975 struct compat_iovec __user *uiov;
2976 compat_ssize_t clen;
2980 uiov = u64_to_user_ptr(req->rw.addr);
2981 if (!access_ok(uiov, sizeof(*uiov)))
2983 if (__get_user(clen, &uiov->iov_len))
2989 buf = io_rw_buffer_select(req, &len, needs_lock);
2991 return PTR_ERR(buf);
2992 iov[0].iov_base = buf;
2993 iov[0].iov_len = (compat_size_t) len;
2998 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3001 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3005 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3008 len = iov[0].iov_len;
3011 buf = io_rw_buffer_select(req, &len, needs_lock);
3013 return PTR_ERR(buf);
3014 iov[0].iov_base = buf;
3015 iov[0].iov_len = len;
3019 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3022 if (req->flags & REQ_F_BUFFER_SELECTED) {
3023 struct io_buffer *kbuf;
3025 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3026 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3027 iov[0].iov_len = kbuf->len;
3030 if (req->rw.len != 1)
3033 #ifdef CONFIG_COMPAT
3034 if (req->ctx->compat)
3035 return io_compat_import(req, iov, needs_lock);
3038 return __io_iov_buffer_select(req, iov, needs_lock);
3041 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3042 struct iov_iter *iter, bool needs_lock)
3044 void __user *buf = u64_to_user_ptr(req->rw.addr);
3045 size_t sqe_len = req->rw.len;
3046 u8 opcode = req->opcode;
3049 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3051 return io_import_fixed(req, rw, iter);
3054 /* buffer index only valid with fixed read/write, or buffer select */
3055 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3058 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3059 if (req->flags & REQ_F_BUFFER_SELECT) {
3060 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3062 return PTR_ERR(buf);
3063 req->rw.len = sqe_len;
3066 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3071 if (req->flags & REQ_F_BUFFER_SELECT) {
3072 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3074 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3079 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3083 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3085 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3089 * For files that don't have ->read_iter() and ->write_iter(), handle them
3090 * by looping over ->read() or ->write() manually.
3092 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3094 struct kiocb *kiocb = &req->rw.kiocb;
3095 struct file *file = req->file;
3099 * Don't support polled IO through this interface, and we can't
3100 * support non-blocking either. For the latter, this just causes
3101 * the kiocb to be handled from an async context.
3103 if (kiocb->ki_flags & IOCB_HIPRI)
3105 if (kiocb->ki_flags & IOCB_NOWAIT)
3108 while (iov_iter_count(iter)) {
3112 if (!iov_iter_is_bvec(iter)) {
3113 iovec = iov_iter_iovec(iter);
3115 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3116 iovec.iov_len = req->rw.len;
3120 nr = file->f_op->read(file, iovec.iov_base,
3121 iovec.iov_len, io_kiocb_ppos(kiocb));
3123 nr = file->f_op->write(file, iovec.iov_base,
3124 iovec.iov_len, io_kiocb_ppos(kiocb));
3133 if (nr != iovec.iov_len)
3137 iov_iter_advance(iter, nr);
3143 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3144 const struct iovec *fast_iov, struct iov_iter *iter)
3146 struct io_async_rw *rw = req->async_data;
3148 memcpy(&rw->iter, iter, sizeof(*iter));
3149 rw->free_iovec = iovec;
3151 /* can only be fixed buffers, no need to do anything */
3152 if (iov_iter_is_bvec(iter))
3155 unsigned iov_off = 0;
3157 rw->iter.iov = rw->fast_iov;
3158 if (iter->iov != fast_iov) {
3159 iov_off = iter->iov - fast_iov;
3160 rw->iter.iov += iov_off;
3162 if (rw->fast_iov != fast_iov)
3163 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3164 sizeof(struct iovec) * iter->nr_segs);
3166 req->flags |= REQ_F_NEED_CLEANUP;
3170 static inline int io_alloc_async_data(struct io_kiocb *req)
3172 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3173 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3174 return req->async_data == NULL;
3177 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3178 const struct iovec *fast_iov,
3179 struct iov_iter *iter, bool force)
3181 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3183 if (!req->async_data) {
3184 if (io_alloc_async_data(req)) {
3189 io_req_map_rw(req, iovec, fast_iov, iter);
3194 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3196 struct io_async_rw *iorw = req->async_data;
3197 struct iovec *iov = iorw->fast_iov;
3200 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3201 if (unlikely(ret < 0))
3204 iorw->bytes_done = 0;
3205 iorw->free_iovec = iov;
3207 req->flags |= REQ_F_NEED_CLEANUP;
3211 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3213 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3215 return io_prep_rw(req, sqe);
3219 * This is our waitqueue callback handler, registered through lock_page_async()
3220 * when we initially tried to do the IO with the iocb armed our waitqueue.
3221 * This gets called when the page is unlocked, and we generally expect that to
3222 * happen when the page IO is completed and the page is now uptodate. This will
3223 * queue a task_work based retry of the operation, attempting to copy the data
3224 * again. If the latter fails because the page was NOT uptodate, then we will
3225 * do a thread based blocking retry of the operation. That's the unexpected
3228 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3229 int sync, void *arg)
3231 struct wait_page_queue *wpq;
3232 struct io_kiocb *req = wait->private;
3233 struct wait_page_key *key = arg;
3235 wpq = container_of(wait, struct wait_page_queue, wait);
3237 if (!wake_page_match(wpq, key))
3240 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3241 list_del_init(&wait->entry);
3242 io_req_task_queue(req);
3247 * This controls whether a given IO request should be armed for async page
3248 * based retry. If we return false here, the request is handed to the async
3249 * worker threads for retry. If we're doing buffered reads on a regular file,
3250 * we prepare a private wait_page_queue entry and retry the operation. This
3251 * will either succeed because the page is now uptodate and unlocked, or it
3252 * will register a callback when the page is unlocked at IO completion. Through
3253 * that callback, io_uring uses task_work to setup a retry of the operation.
3254 * That retry will attempt the buffered read again. The retry will generally
3255 * succeed, or in rare cases where it fails, we then fall back to using the
3256 * async worker threads for a blocking retry.
3258 static bool io_rw_should_retry(struct io_kiocb *req)
3260 struct io_async_rw *rw = req->async_data;
3261 struct wait_page_queue *wait = &rw->wpq;
3262 struct kiocb *kiocb = &req->rw.kiocb;
3264 /* never retry for NOWAIT, we just complete with -EAGAIN */
3265 if (req->flags & REQ_F_NOWAIT)
3268 /* Only for buffered IO */
3269 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3273 * just use poll if we can, and don't attempt if the fs doesn't
3274 * support callback based unlocks
3276 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3279 wait->wait.func = io_async_buf_func;
3280 wait->wait.private = req;
3281 wait->wait.flags = 0;
3282 INIT_LIST_HEAD(&wait->wait.entry);
3283 kiocb->ki_flags |= IOCB_WAITQ;
3284 kiocb->ki_flags &= ~IOCB_NOWAIT;
3285 kiocb->ki_waitq = wait;
3289 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3291 if (req->file->f_op->read_iter)
3292 return call_read_iter(req->file, &req->rw.kiocb, iter);
3293 else if (req->file->f_op->read)
3294 return loop_rw_iter(READ, req, iter);
3299 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3301 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3302 struct kiocb *kiocb = &req->rw.kiocb;
3303 struct iov_iter __iter, *iter = &__iter;
3304 struct io_async_rw *rw = req->async_data;
3305 ssize_t io_size, ret, ret2;
3306 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3312 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3316 io_size = iov_iter_count(iter);
3317 req->result = io_size;
3319 /* Ensure we clear previously set non-block flag */
3320 if (!force_nonblock)
3321 kiocb->ki_flags &= ~IOCB_NOWAIT;
3323 kiocb->ki_flags |= IOCB_NOWAIT;
3325 /* If the file doesn't support async, just async punt */
3326 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3327 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3328 return ret ?: -EAGAIN;
3331 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3332 if (unlikely(ret)) {
3337 ret = io_iter_do_read(req, iter);
3339 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3340 req->flags &= ~REQ_F_REISSUE;
3341 /* IOPOLL retry should happen for io-wq threads */
3342 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3344 /* no retry on NONBLOCK nor RWF_NOWAIT */
3345 if (req->flags & REQ_F_NOWAIT)
3347 /* some cases will consume bytes even on error returns */
3348 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3350 } else if (ret == -EIOCBQUEUED) {
3352 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3353 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3354 /* read all, failed, already did sync or don't want to retry */
3358 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3363 rw = req->async_data;
3364 /* now use our persistent iterator, if we aren't already */
3369 rw->bytes_done += ret;
3370 /* if we can retry, do so with the callbacks armed */
3371 if (!io_rw_should_retry(req)) {
3372 kiocb->ki_flags &= ~IOCB_WAITQ;
3377 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3378 * we get -EIOCBQUEUED, then we'll get a notification when the
3379 * desired page gets unlocked. We can also get a partial read
3380 * here, and if we do, then just retry at the new offset.
3382 ret = io_iter_do_read(req, iter);
3383 if (ret == -EIOCBQUEUED)
3385 /* we got some bytes, but not all. retry. */
3386 kiocb->ki_flags &= ~IOCB_WAITQ;
3387 } while (ret > 0 && ret < io_size);
3389 kiocb_done(kiocb, ret, issue_flags);
3391 /* it's faster to check here then delegate to kfree */
3397 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3399 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3401 return io_prep_rw(req, sqe);
3404 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3406 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3407 struct kiocb *kiocb = &req->rw.kiocb;
3408 struct iov_iter __iter, *iter = &__iter;
3409 struct io_async_rw *rw = req->async_data;
3410 ssize_t ret, ret2, io_size;
3411 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3417 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3421 io_size = iov_iter_count(iter);
3422 req->result = io_size;
3424 /* Ensure we clear previously set non-block flag */
3425 if (!force_nonblock)
3426 kiocb->ki_flags &= ~IOCB_NOWAIT;
3428 kiocb->ki_flags |= IOCB_NOWAIT;
3430 /* If the file doesn't support async, just async punt */
3431 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3434 /* file path doesn't support NOWAIT for non-direct_IO */
3435 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3436 (req->flags & REQ_F_ISREG))
3439 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3444 * Open-code file_start_write here to grab freeze protection,
3445 * which will be released by another thread in
3446 * io_complete_rw(). Fool lockdep by telling it the lock got
3447 * released so that it doesn't complain about the held lock when
3448 * we return to userspace.
3450 if (req->flags & REQ_F_ISREG) {
3451 sb_start_write(file_inode(req->file)->i_sb);
3452 __sb_writers_release(file_inode(req->file)->i_sb,
3455 kiocb->ki_flags |= IOCB_WRITE;
3457 if (req->file->f_op->write_iter)
3458 ret2 = call_write_iter(req->file, kiocb, iter);
3459 else if (req->file->f_op->write)
3460 ret2 = loop_rw_iter(WRITE, req, iter);
3464 if (req->flags & REQ_F_REISSUE) {
3465 req->flags &= ~REQ_F_REISSUE;
3470 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3471 * retry them without IOCB_NOWAIT.
3473 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3475 /* no retry on NONBLOCK nor RWF_NOWAIT */
3476 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3478 if (!force_nonblock || ret2 != -EAGAIN) {
3479 /* IOPOLL retry should happen for io-wq threads */
3480 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3483 kiocb_done(kiocb, ret2, issue_flags);
3486 /* some cases will consume bytes even on error returns */
3487 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3488 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3489 return ret ?: -EAGAIN;
3492 /* it's reportedly faster than delegating the null check to kfree() */
3498 static int io_renameat_prep(struct io_kiocb *req,
3499 const struct io_uring_sqe *sqe)
3501 struct io_rename *ren = &req->rename;
3502 const char __user *oldf, *newf;
3504 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3506 if (sqe->ioprio || sqe->buf_index)
3508 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3511 ren->old_dfd = READ_ONCE(sqe->fd);
3512 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3513 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3514 ren->new_dfd = READ_ONCE(sqe->len);
3515 ren->flags = READ_ONCE(sqe->rename_flags);
3517 ren->oldpath = getname(oldf);
3518 if (IS_ERR(ren->oldpath))
3519 return PTR_ERR(ren->oldpath);
3521 ren->newpath = getname(newf);
3522 if (IS_ERR(ren->newpath)) {
3523 putname(ren->oldpath);
3524 return PTR_ERR(ren->newpath);
3527 req->flags |= REQ_F_NEED_CLEANUP;
3531 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3533 struct io_rename *ren = &req->rename;
3536 if (issue_flags & IO_URING_F_NONBLOCK)
3539 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3540 ren->newpath, ren->flags);
3542 req->flags &= ~REQ_F_NEED_CLEANUP;
3545 io_req_complete(req, ret);
3549 static int io_unlinkat_prep(struct io_kiocb *req,
3550 const struct io_uring_sqe *sqe)
3552 struct io_unlink *un = &req->unlink;
3553 const char __user *fname;
3555 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3557 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3559 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3562 un->dfd = READ_ONCE(sqe->fd);
3564 un->flags = READ_ONCE(sqe->unlink_flags);
3565 if (un->flags & ~AT_REMOVEDIR)
3568 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3569 un->filename = getname(fname);
3570 if (IS_ERR(un->filename))
3571 return PTR_ERR(un->filename);
3573 req->flags |= REQ_F_NEED_CLEANUP;
3577 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3579 struct io_unlink *un = &req->unlink;
3582 if (issue_flags & IO_URING_F_NONBLOCK)
3585 if (un->flags & AT_REMOVEDIR)
3586 ret = do_rmdir(un->dfd, un->filename);
3588 ret = do_unlinkat(un->dfd, un->filename);
3590 req->flags &= ~REQ_F_NEED_CLEANUP;
3593 io_req_complete(req, ret);
3597 static int io_shutdown_prep(struct io_kiocb *req,
3598 const struct io_uring_sqe *sqe)
3600 #if defined(CONFIG_NET)
3601 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3603 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3607 req->shutdown.how = READ_ONCE(sqe->len);
3614 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3616 #if defined(CONFIG_NET)
3617 struct socket *sock;
3620 if (issue_flags & IO_URING_F_NONBLOCK)
3623 sock = sock_from_file(req->file);
3624 if (unlikely(!sock))
3627 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3630 io_req_complete(req, ret);
3637 static int __io_splice_prep(struct io_kiocb *req,
3638 const struct io_uring_sqe *sqe)
3640 struct io_splice *sp = &req->splice;
3641 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3643 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3647 sp->len = READ_ONCE(sqe->len);
3648 sp->flags = READ_ONCE(sqe->splice_flags);
3650 if (unlikely(sp->flags & ~valid_flags))
3653 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3654 (sp->flags & SPLICE_F_FD_IN_FIXED));
3657 req->flags |= REQ_F_NEED_CLEANUP;
3661 static int io_tee_prep(struct io_kiocb *req,
3662 const struct io_uring_sqe *sqe)
3664 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3666 return __io_splice_prep(req, sqe);
3669 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3671 struct io_splice *sp = &req->splice;
3672 struct file *in = sp->file_in;
3673 struct file *out = sp->file_out;
3674 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3677 if (issue_flags & IO_URING_F_NONBLOCK)
3680 ret = do_tee(in, out, sp->len, flags);
3682 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3684 req->flags &= ~REQ_F_NEED_CLEANUP;
3688 io_req_complete(req, ret);
3692 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694 struct io_splice *sp = &req->splice;
3696 sp->off_in = READ_ONCE(sqe->splice_off_in);
3697 sp->off_out = READ_ONCE(sqe->off);
3698 return __io_splice_prep(req, sqe);
3701 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3703 struct io_splice *sp = &req->splice;
3704 struct file *in = sp->file_in;
3705 struct file *out = sp->file_out;
3706 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3707 loff_t *poff_in, *poff_out;
3710 if (issue_flags & IO_URING_F_NONBLOCK)
3713 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3714 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3717 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3719 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3721 req->flags &= ~REQ_F_NEED_CLEANUP;
3725 io_req_complete(req, ret);
3730 * IORING_OP_NOP just posts a completion event, nothing else.
3732 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3734 struct io_ring_ctx *ctx = req->ctx;
3736 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3739 __io_req_complete(req, issue_flags, 0, 0);
3743 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3745 struct io_ring_ctx *ctx = req->ctx;
3750 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3752 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3755 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3756 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3759 req->sync.off = READ_ONCE(sqe->off);
3760 req->sync.len = READ_ONCE(sqe->len);
3764 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3766 loff_t end = req->sync.off + req->sync.len;
3769 /* fsync always requires a blocking context */
3770 if (issue_flags & IO_URING_F_NONBLOCK)
3773 ret = vfs_fsync_range(req->file, req->sync.off,
3774 end > 0 ? end : LLONG_MAX,
3775 req->sync.flags & IORING_FSYNC_DATASYNC);
3778 io_req_complete(req, ret);
3782 static int io_fallocate_prep(struct io_kiocb *req,
3783 const struct io_uring_sqe *sqe)
3785 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3787 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3790 req->sync.off = READ_ONCE(sqe->off);
3791 req->sync.len = READ_ONCE(sqe->addr);
3792 req->sync.mode = READ_ONCE(sqe->len);
3796 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3800 /* fallocate always requiring blocking context */
3801 if (issue_flags & IO_URING_F_NONBLOCK)
3803 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3807 io_req_complete(req, ret);
3811 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3813 const char __user *fname;
3816 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3818 if (unlikely(sqe->ioprio || sqe->buf_index))
3820 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3823 /* open.how should be already initialised */
3824 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3825 req->open.how.flags |= O_LARGEFILE;
3827 req->open.dfd = READ_ONCE(sqe->fd);
3828 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3829 req->open.filename = getname(fname);
3830 if (IS_ERR(req->open.filename)) {
3831 ret = PTR_ERR(req->open.filename);
3832 req->open.filename = NULL;
3835 req->open.nofile = rlimit(RLIMIT_NOFILE);
3836 req->flags |= REQ_F_NEED_CLEANUP;
3840 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3842 u64 mode = READ_ONCE(sqe->len);
3843 u64 flags = READ_ONCE(sqe->open_flags);
3845 req->open.how = build_open_how(flags, mode);
3846 return __io_openat_prep(req, sqe);
3849 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3851 struct open_how __user *how;
3855 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3856 len = READ_ONCE(sqe->len);
3857 if (len < OPEN_HOW_SIZE_VER0)
3860 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3865 return __io_openat_prep(req, sqe);
3868 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3870 struct open_flags op;
3873 bool resolve_nonblock;
3876 ret = build_open_flags(&req->open.how, &op);
3879 nonblock_set = op.open_flag & O_NONBLOCK;
3880 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3881 if (issue_flags & IO_URING_F_NONBLOCK) {
3883 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3884 * it'll always -EAGAIN
3886 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3888 op.lookup_flags |= LOOKUP_CACHED;
3889 op.open_flag |= O_NONBLOCK;
3892 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3896 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3899 * We could hang on to this 'fd' on retrying, but seems like
3900 * marginal gain for something that is now known to be a slower
3901 * path. So just put it, and we'll get a new one when we retry.
3905 ret = PTR_ERR(file);
3906 /* only retry if RESOLVE_CACHED wasn't already set by application */
3907 if (ret == -EAGAIN &&
3908 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3913 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3914 file->f_flags &= ~O_NONBLOCK;
3915 fsnotify_open(file);
3916 fd_install(ret, file);
3918 putname(req->open.filename);
3919 req->flags &= ~REQ_F_NEED_CLEANUP;
3922 __io_req_complete(req, issue_flags, ret, 0);
3926 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3928 return io_openat2(req, issue_flags);
3931 static int io_remove_buffers_prep(struct io_kiocb *req,
3932 const struct io_uring_sqe *sqe)
3934 struct io_provide_buf *p = &req->pbuf;
3937 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3940 tmp = READ_ONCE(sqe->fd);
3941 if (!tmp || tmp > USHRT_MAX)
3944 memset(p, 0, sizeof(*p));
3946 p->bgid = READ_ONCE(sqe->buf_group);
3950 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3951 int bgid, unsigned nbufs)
3955 /* shouldn't happen */
3959 /* the head kbuf is the list itself */
3960 while (!list_empty(&buf->list)) {
3961 struct io_buffer *nxt;
3963 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3964 list_del(&nxt->list);
3971 xa_erase(&ctx->io_buffers, bgid);
3976 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3978 struct io_provide_buf *p = &req->pbuf;
3979 struct io_ring_ctx *ctx = req->ctx;
3980 struct io_buffer *head;
3982 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3984 io_ring_submit_lock(ctx, !force_nonblock);
3986 lockdep_assert_held(&ctx->uring_lock);
3989 head = xa_load(&ctx->io_buffers, p->bgid);
3991 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3995 /* complete before unlock, IOPOLL may need the lock */
3996 __io_req_complete(req, issue_flags, ret, 0);
3997 io_ring_submit_unlock(ctx, !force_nonblock);
4001 static int io_provide_buffers_prep(struct io_kiocb *req,
4002 const struct io_uring_sqe *sqe)
4004 unsigned long size, tmp_check;
4005 struct io_provide_buf *p = &req->pbuf;
4008 if (sqe->ioprio || sqe->rw_flags)
4011 tmp = READ_ONCE(sqe->fd);
4012 if (!tmp || tmp > USHRT_MAX)
4015 p->addr = READ_ONCE(sqe->addr);
4016 p->len = READ_ONCE(sqe->len);
4018 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4021 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4024 size = (unsigned long)p->len * p->nbufs;
4025 if (!access_ok(u64_to_user_ptr(p->addr), size))
4028 p->bgid = READ_ONCE(sqe->buf_group);
4029 tmp = READ_ONCE(sqe->off);
4030 if (tmp > USHRT_MAX)
4036 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4038 struct io_buffer *buf;
4039 u64 addr = pbuf->addr;
4040 int i, bid = pbuf->bid;
4042 for (i = 0; i < pbuf->nbufs; i++) {
4043 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4048 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4053 INIT_LIST_HEAD(&buf->list);
4056 list_add_tail(&buf->list, &(*head)->list);
4060 return i ? i : -ENOMEM;
4063 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4065 struct io_provide_buf *p = &req->pbuf;
4066 struct io_ring_ctx *ctx = req->ctx;
4067 struct io_buffer *head, *list;
4069 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4071 io_ring_submit_lock(ctx, !force_nonblock);
4073 lockdep_assert_held(&ctx->uring_lock);
4075 list = head = xa_load(&ctx->io_buffers, p->bgid);
4077 ret = io_add_buffers(p, &head);
4078 if (ret >= 0 && !list) {
4079 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4081 __io_remove_buffers(ctx, head, p->bgid, -1U);
4085 /* complete before unlock, IOPOLL may need the lock */
4086 __io_req_complete(req, issue_flags, ret, 0);
4087 io_ring_submit_unlock(ctx, !force_nonblock);
4091 static int io_epoll_ctl_prep(struct io_kiocb *req,
4092 const struct io_uring_sqe *sqe)
4094 #if defined(CONFIG_EPOLL)
4095 if (sqe->ioprio || sqe->buf_index)
4097 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4100 req->epoll.epfd = READ_ONCE(sqe->fd);
4101 req->epoll.op = READ_ONCE(sqe->len);
4102 req->epoll.fd = READ_ONCE(sqe->off);
4104 if (ep_op_has_event(req->epoll.op)) {
4105 struct epoll_event __user *ev;
4107 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4108 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4118 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4120 #if defined(CONFIG_EPOLL)
4121 struct io_epoll *ie = &req->epoll;
4123 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4125 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4126 if (force_nonblock && ret == -EAGAIN)
4131 __io_req_complete(req, issue_flags, ret, 0);
4138 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4140 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4141 if (sqe->ioprio || sqe->buf_index || sqe->off)
4143 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4146 req->madvise.addr = READ_ONCE(sqe->addr);
4147 req->madvise.len = READ_ONCE(sqe->len);
4148 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4155 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4157 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4158 struct io_madvise *ma = &req->madvise;
4161 if (issue_flags & IO_URING_F_NONBLOCK)
4164 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4167 io_req_complete(req, ret);
4174 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4176 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4178 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4181 req->fadvise.offset = READ_ONCE(sqe->off);
4182 req->fadvise.len = READ_ONCE(sqe->len);
4183 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4187 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4189 struct io_fadvise *fa = &req->fadvise;
4192 if (issue_flags & IO_URING_F_NONBLOCK) {
4193 switch (fa->advice) {
4194 case POSIX_FADV_NORMAL:
4195 case POSIX_FADV_RANDOM:
4196 case POSIX_FADV_SEQUENTIAL:
4203 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4206 __io_req_complete(req, issue_flags, ret, 0);
4210 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4212 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4214 if (sqe->ioprio || sqe->buf_index)
4216 if (req->flags & REQ_F_FIXED_FILE)
4219 req->statx.dfd = READ_ONCE(sqe->fd);
4220 req->statx.mask = READ_ONCE(sqe->len);
4221 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4222 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4223 req->statx.flags = READ_ONCE(sqe->statx_flags);
4228 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4230 struct io_statx *ctx = &req->statx;
4233 if (issue_flags & IO_URING_F_NONBLOCK)
4236 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4241 io_req_complete(req, ret);
4245 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4247 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4249 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4250 sqe->rw_flags || sqe->buf_index)
4252 if (req->flags & REQ_F_FIXED_FILE)
4255 req->close.fd = READ_ONCE(sqe->fd);
4259 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4261 struct files_struct *files = current->files;
4262 struct io_close *close = &req->close;
4263 struct fdtable *fdt;
4264 struct file *file = NULL;
4267 spin_lock(&files->file_lock);
4268 fdt = files_fdtable(files);
4269 if (close->fd >= fdt->max_fds) {
4270 spin_unlock(&files->file_lock);
4273 file = fdt->fd[close->fd];
4274 if (!file || file->f_op == &io_uring_fops) {
4275 spin_unlock(&files->file_lock);
4280 /* if the file has a flush method, be safe and punt to async */
4281 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4282 spin_unlock(&files->file_lock);
4286 ret = __close_fd_get_file(close->fd, &file);
4287 spin_unlock(&files->file_lock);
4294 /* No ->flush() or already async, safely close from here */
4295 ret = filp_close(file, current->files);
4301 __io_req_complete(req, issue_flags, ret, 0);
4305 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4307 struct io_ring_ctx *ctx = req->ctx;
4309 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4311 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4314 req->sync.off = READ_ONCE(sqe->off);
4315 req->sync.len = READ_ONCE(sqe->len);
4316 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4320 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4324 /* sync_file_range always requires a blocking context */
4325 if (issue_flags & IO_URING_F_NONBLOCK)
4328 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4332 io_req_complete(req, ret);
4336 #if defined(CONFIG_NET)
4337 static int io_setup_async_msg(struct io_kiocb *req,
4338 struct io_async_msghdr *kmsg)
4340 struct io_async_msghdr *async_msg = req->async_data;
4344 if (io_alloc_async_data(req)) {
4345 kfree(kmsg->free_iov);
4348 async_msg = req->async_data;
4349 req->flags |= REQ_F_NEED_CLEANUP;
4350 memcpy(async_msg, kmsg, sizeof(*kmsg));
4351 async_msg->msg.msg_name = &async_msg->addr;
4352 /* if were using fast_iov, set it to the new one */
4353 if (!async_msg->free_iov)
4354 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4359 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4360 struct io_async_msghdr *iomsg)
4362 iomsg->msg.msg_name = &iomsg->addr;
4363 iomsg->free_iov = iomsg->fast_iov;
4364 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4365 req->sr_msg.msg_flags, &iomsg->free_iov);
4368 static int io_sendmsg_prep_async(struct io_kiocb *req)
4372 ret = io_sendmsg_copy_hdr(req, req->async_data);
4374 req->flags |= REQ_F_NEED_CLEANUP;
4378 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4380 struct io_sr_msg *sr = &req->sr_msg;
4382 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4385 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4386 sr->len = READ_ONCE(sqe->len);
4387 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4388 if (sr->msg_flags & MSG_DONTWAIT)
4389 req->flags |= REQ_F_NOWAIT;
4391 #ifdef CONFIG_COMPAT
4392 if (req->ctx->compat)
4393 sr->msg_flags |= MSG_CMSG_COMPAT;
4398 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4400 struct io_async_msghdr iomsg, *kmsg;
4401 struct socket *sock;
4406 sock = sock_from_file(req->file);
4407 if (unlikely(!sock))
4410 kmsg = req->async_data;
4412 ret = io_sendmsg_copy_hdr(req, &iomsg);
4418 flags = req->sr_msg.msg_flags;
4419 if (issue_flags & IO_URING_F_NONBLOCK)
4420 flags |= MSG_DONTWAIT;
4421 if (flags & MSG_WAITALL)
4422 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4424 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4425 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4426 return io_setup_async_msg(req, kmsg);
4427 if (ret == -ERESTARTSYS)
4430 /* fast path, check for non-NULL to avoid function call */
4432 kfree(kmsg->free_iov);
4433 req->flags &= ~REQ_F_NEED_CLEANUP;
4436 __io_req_complete(req, issue_flags, ret, 0);
4440 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4442 struct io_sr_msg *sr = &req->sr_msg;
4445 struct socket *sock;
4450 sock = sock_from_file(req->file);
4451 if (unlikely(!sock))
4454 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4458 msg.msg_name = NULL;
4459 msg.msg_control = NULL;
4460 msg.msg_controllen = 0;
4461 msg.msg_namelen = 0;
4463 flags = req->sr_msg.msg_flags;
4464 if (issue_flags & IO_URING_F_NONBLOCK)
4465 flags |= MSG_DONTWAIT;
4466 if (flags & MSG_WAITALL)
4467 min_ret = iov_iter_count(&msg.msg_iter);
4469 msg.msg_flags = flags;
4470 ret = sock_sendmsg(sock, &msg);
4471 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4473 if (ret == -ERESTARTSYS)
4478 __io_req_complete(req, issue_flags, ret, 0);
4482 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4483 struct io_async_msghdr *iomsg)
4485 struct io_sr_msg *sr = &req->sr_msg;
4486 struct iovec __user *uiov;
4490 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4491 &iomsg->uaddr, &uiov, &iov_len);
4495 if (req->flags & REQ_F_BUFFER_SELECT) {
4498 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4500 sr->len = iomsg->fast_iov[0].iov_len;
4501 iomsg->free_iov = NULL;
4503 iomsg->free_iov = iomsg->fast_iov;
4504 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4505 &iomsg->free_iov, &iomsg->msg.msg_iter,
4514 #ifdef CONFIG_COMPAT
4515 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4516 struct io_async_msghdr *iomsg)
4518 struct io_sr_msg *sr = &req->sr_msg;
4519 struct compat_iovec __user *uiov;
4524 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4529 uiov = compat_ptr(ptr);
4530 if (req->flags & REQ_F_BUFFER_SELECT) {
4531 compat_ssize_t clen;
4535 if (!access_ok(uiov, sizeof(*uiov)))
4537 if (__get_user(clen, &uiov->iov_len))
4542 iomsg->free_iov = NULL;
4544 iomsg->free_iov = iomsg->fast_iov;
4545 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4546 UIO_FASTIOV, &iomsg->free_iov,
4547 &iomsg->msg.msg_iter, true);
4556 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4557 struct io_async_msghdr *iomsg)
4559 iomsg->msg.msg_name = &iomsg->addr;
4561 #ifdef CONFIG_COMPAT
4562 if (req->ctx->compat)
4563 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4566 return __io_recvmsg_copy_hdr(req, iomsg);
4569 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4572 struct io_sr_msg *sr = &req->sr_msg;
4573 struct io_buffer *kbuf;
4575 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4580 req->flags |= REQ_F_BUFFER_SELECTED;
4584 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4586 return io_put_kbuf(req, req->sr_msg.kbuf);
4589 static int io_recvmsg_prep_async(struct io_kiocb *req)
4593 ret = io_recvmsg_copy_hdr(req, req->async_data);
4595 req->flags |= REQ_F_NEED_CLEANUP;
4599 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4601 struct io_sr_msg *sr = &req->sr_msg;
4603 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4606 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4607 sr->len = READ_ONCE(sqe->len);
4608 sr->bgid = READ_ONCE(sqe->buf_group);
4609 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4610 if (sr->msg_flags & MSG_DONTWAIT)
4611 req->flags |= REQ_F_NOWAIT;
4613 #ifdef CONFIG_COMPAT
4614 if (req->ctx->compat)
4615 sr->msg_flags |= MSG_CMSG_COMPAT;
4620 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4622 struct io_async_msghdr iomsg, *kmsg;
4623 struct socket *sock;
4624 struct io_buffer *kbuf;
4627 int ret, cflags = 0;
4628 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4630 sock = sock_from_file(req->file);
4631 if (unlikely(!sock))
4634 kmsg = req->async_data;
4636 ret = io_recvmsg_copy_hdr(req, &iomsg);
4642 if (req->flags & REQ_F_BUFFER_SELECT) {
4643 kbuf = io_recv_buffer_select(req, !force_nonblock);
4645 return PTR_ERR(kbuf);
4646 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4647 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4648 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4649 1, req->sr_msg.len);
4652 flags = req->sr_msg.msg_flags;
4654 flags |= MSG_DONTWAIT;
4655 if (flags & MSG_WAITALL)
4656 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4658 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4659 kmsg->uaddr, flags);
4660 if (force_nonblock && ret == -EAGAIN)
4661 return io_setup_async_msg(req, kmsg);
4662 if (ret == -ERESTARTSYS)
4665 if (req->flags & REQ_F_BUFFER_SELECTED)
4666 cflags = io_put_recv_kbuf(req);
4667 /* fast path, check for non-NULL to avoid function call */
4669 kfree(kmsg->free_iov);
4670 req->flags &= ~REQ_F_NEED_CLEANUP;
4671 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4673 __io_req_complete(req, issue_flags, ret, cflags);
4677 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4679 struct io_buffer *kbuf;
4680 struct io_sr_msg *sr = &req->sr_msg;
4682 void __user *buf = sr->buf;
4683 struct socket *sock;
4687 int ret, cflags = 0;
4688 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4690 sock = sock_from_file(req->file);
4691 if (unlikely(!sock))
4694 if (req->flags & REQ_F_BUFFER_SELECT) {
4695 kbuf = io_recv_buffer_select(req, !force_nonblock);
4697 return PTR_ERR(kbuf);
4698 buf = u64_to_user_ptr(kbuf->addr);
4701 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4705 msg.msg_name = NULL;
4706 msg.msg_control = NULL;
4707 msg.msg_controllen = 0;
4708 msg.msg_namelen = 0;
4709 msg.msg_iocb = NULL;
4712 flags = req->sr_msg.msg_flags;
4714 flags |= MSG_DONTWAIT;
4715 if (flags & MSG_WAITALL)
4716 min_ret = iov_iter_count(&msg.msg_iter);
4718 ret = sock_recvmsg(sock, &msg, flags);
4719 if (force_nonblock && ret == -EAGAIN)
4721 if (ret == -ERESTARTSYS)
4724 if (req->flags & REQ_F_BUFFER_SELECTED)
4725 cflags = io_put_recv_kbuf(req);
4726 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4728 __io_req_complete(req, issue_flags, ret, cflags);
4732 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4734 struct io_accept *accept = &req->accept;
4736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4738 if (sqe->ioprio || sqe->len || sqe->buf_index)
4741 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4742 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4743 accept->flags = READ_ONCE(sqe->accept_flags);
4744 accept->nofile = rlimit(RLIMIT_NOFILE);
4748 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4750 struct io_accept *accept = &req->accept;
4751 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4752 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4755 if (req->file->f_flags & O_NONBLOCK)
4756 req->flags |= REQ_F_NOWAIT;
4758 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4759 accept->addr_len, accept->flags,
4761 if (ret == -EAGAIN && force_nonblock)
4764 if (ret == -ERESTARTSYS)
4768 __io_req_complete(req, issue_flags, ret, 0);
4772 static int io_connect_prep_async(struct io_kiocb *req)
4774 struct io_async_connect *io = req->async_data;
4775 struct io_connect *conn = &req->connect;
4777 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4780 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4782 struct io_connect *conn = &req->connect;
4784 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4786 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4789 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4790 conn->addr_len = READ_ONCE(sqe->addr2);
4794 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4796 struct io_async_connect __io, *io;
4797 unsigned file_flags;
4799 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4801 if (req->async_data) {
4802 io = req->async_data;
4804 ret = move_addr_to_kernel(req->connect.addr,
4805 req->connect.addr_len,
4812 file_flags = force_nonblock ? O_NONBLOCK : 0;
4814 ret = __sys_connect_file(req->file, &io->address,
4815 req->connect.addr_len, file_flags);
4816 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4817 if (req->async_data)
4819 if (io_alloc_async_data(req)) {
4823 memcpy(req->async_data, &__io, sizeof(__io));
4826 if (ret == -ERESTARTSYS)
4831 __io_req_complete(req, issue_flags, ret, 0);
4834 #else /* !CONFIG_NET */
4835 #define IO_NETOP_FN(op) \
4836 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4838 return -EOPNOTSUPP; \
4841 #define IO_NETOP_PREP(op) \
4843 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4845 return -EOPNOTSUPP; \
4848 #define IO_NETOP_PREP_ASYNC(op) \
4850 static int io_##op##_prep_async(struct io_kiocb *req) \
4852 return -EOPNOTSUPP; \
4855 IO_NETOP_PREP_ASYNC(sendmsg);
4856 IO_NETOP_PREP_ASYNC(recvmsg);
4857 IO_NETOP_PREP_ASYNC(connect);
4858 IO_NETOP_PREP(accept);
4861 #endif /* CONFIG_NET */
4863 struct io_poll_table {
4864 struct poll_table_struct pt;
4865 struct io_kiocb *req;
4870 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4871 __poll_t mask, io_req_tw_func_t func)
4873 /* for instances that support it check for an event match first: */
4874 if (mask && !(mask & poll->events))
4877 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4879 list_del_init(&poll->wait.entry);
4882 req->io_task_work.func = func;
4885 * If this fails, then the task is exiting. When a task exits, the
4886 * work gets canceled, so just cancel this request as well instead
4887 * of executing it. We can't safely execute it anyway, as we may not
4888 * have the needed state needed for it anyway.
4890 io_req_task_work_add(req);
4894 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4895 __acquires(&req->ctx->completion_lock)
4897 struct io_ring_ctx *ctx = req->ctx;
4899 /* req->task == current here, checking PF_EXITING is safe */
4900 if (unlikely(req->task->flags & PF_EXITING))
4901 WRITE_ONCE(poll->canceled, true);
4903 if (!req->result && !READ_ONCE(poll->canceled)) {
4904 struct poll_table_struct pt = { ._key = poll->events };
4906 req->result = vfs_poll(req->file, &pt) & poll->events;
4909 spin_lock(&ctx->completion_lock);
4910 if (!req->result && !READ_ONCE(poll->canceled)) {
4911 add_wait_queue(poll->head, &poll->wait);
4918 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4920 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4921 if (req->opcode == IORING_OP_POLL_ADD)
4922 return req->async_data;
4923 return req->apoll->double_poll;
4926 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4928 if (req->opcode == IORING_OP_POLL_ADD)
4930 return &req->apoll->poll;
4933 static void io_poll_remove_double(struct io_kiocb *req)
4934 __must_hold(&req->ctx->completion_lock)
4936 struct io_poll_iocb *poll = io_poll_get_double(req);
4938 lockdep_assert_held(&req->ctx->completion_lock);
4940 if (poll && poll->head) {
4941 struct wait_queue_head *head = poll->head;
4943 spin_lock_irq(&head->lock);
4944 list_del_init(&poll->wait.entry);
4945 if (poll->wait.private)
4948 spin_unlock_irq(&head->lock);
4952 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4953 __must_hold(&req->ctx->completion_lock)
4955 struct io_ring_ctx *ctx = req->ctx;
4956 unsigned flags = IORING_CQE_F_MORE;
4959 if (READ_ONCE(req->poll.canceled)) {
4961 req->poll.events |= EPOLLONESHOT;
4963 error = mangle_poll(mask);
4965 if (req->poll.events & EPOLLONESHOT)
4967 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4968 req->poll.done = true;
4971 if (flags & IORING_CQE_F_MORE)
4974 io_commit_cqring(ctx);
4975 return !(flags & IORING_CQE_F_MORE);
4978 static void io_poll_task_func(struct io_kiocb *req)
4980 struct io_ring_ctx *ctx = req->ctx;
4981 struct io_kiocb *nxt;
4983 if (io_poll_rewait(req, &req->poll)) {
4984 spin_unlock(&ctx->completion_lock);
4988 done = io_poll_complete(req, req->result);
4990 io_poll_remove_double(req);
4991 hash_del(&req->hash_node);
4994 add_wait_queue(req->poll.head, &req->poll.wait);
4996 spin_unlock(&ctx->completion_lock);
4997 io_cqring_ev_posted(ctx);
5000 nxt = io_put_req_find_next(req);
5002 io_req_task_submit(nxt);
5007 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5008 int sync, void *key)
5010 struct io_kiocb *req = wait->private;
5011 struct io_poll_iocb *poll = io_poll_get_single(req);
5012 __poll_t mask = key_to_poll(key);
5013 unsigned long flags;
5015 /* for instances that support it check for an event match first: */
5016 if (mask && !(mask & poll->events))
5018 if (!(poll->events & EPOLLONESHOT))
5019 return poll->wait.func(&poll->wait, mode, sync, key);
5021 list_del_init(&wait->entry);
5026 spin_lock_irqsave(&poll->head->lock, flags);
5027 done = list_empty(&poll->wait.entry);
5029 list_del_init(&poll->wait.entry);
5030 /* make sure double remove sees this as being gone */
5031 wait->private = NULL;
5032 spin_unlock_irqrestore(&poll->head->lock, flags);
5034 /* use wait func handler, so it matches the rq type */
5035 poll->wait.func(&poll->wait, mode, sync, key);
5042 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5043 wait_queue_func_t wake_func)
5047 poll->canceled = false;
5048 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5049 /* mask in events that we always want/need */
5050 poll->events = events | IO_POLL_UNMASK;
5051 INIT_LIST_HEAD(&poll->wait.entry);
5052 init_waitqueue_func_entry(&poll->wait, wake_func);
5055 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5056 struct wait_queue_head *head,
5057 struct io_poll_iocb **poll_ptr)
5059 struct io_kiocb *req = pt->req;
5062 * The file being polled uses multiple waitqueues for poll handling
5063 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5066 if (unlikely(pt->nr_entries)) {
5067 struct io_poll_iocb *poll_one = poll;
5069 /* double add on the same waitqueue head, ignore */
5070 if (poll_one->head == head)
5072 /* already have a 2nd entry, fail a third attempt */
5074 if ((*poll_ptr)->head == head)
5076 pt->error = -EINVAL;
5080 * Can't handle multishot for double wait for now, turn it
5081 * into one-shot mode.
5083 if (!(poll_one->events & EPOLLONESHOT))
5084 poll_one->events |= EPOLLONESHOT;
5085 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5087 pt->error = -ENOMEM;
5090 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5092 poll->wait.private = req;
5099 if (poll->events & EPOLLEXCLUSIVE)
5100 add_wait_queue_exclusive(head, &poll->wait);
5102 add_wait_queue(head, &poll->wait);
5105 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5106 struct poll_table_struct *p)
5108 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5109 struct async_poll *apoll = pt->req->apoll;
5111 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5114 static void io_async_task_func(struct io_kiocb *req)
5116 struct async_poll *apoll = req->apoll;
5117 struct io_ring_ctx *ctx = req->ctx;
5119 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5121 if (io_poll_rewait(req, &apoll->poll)) {
5122 spin_unlock(&ctx->completion_lock);
5126 hash_del(&req->hash_node);
5127 io_poll_remove_double(req);
5128 spin_unlock(&ctx->completion_lock);
5130 if (!READ_ONCE(apoll->poll.canceled))
5131 io_req_task_submit(req);
5133 io_req_complete_failed(req, -ECANCELED);
5136 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5139 struct io_kiocb *req = wait->private;
5140 struct io_poll_iocb *poll = &req->apoll->poll;
5142 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5145 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5148 static void io_poll_req_insert(struct io_kiocb *req)
5150 struct io_ring_ctx *ctx = req->ctx;
5151 struct hlist_head *list;
5153 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5154 hlist_add_head(&req->hash_node, list);
5157 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5158 struct io_poll_iocb *poll,
5159 struct io_poll_table *ipt, __poll_t mask,
5160 wait_queue_func_t wake_func)
5161 __acquires(&ctx->completion_lock)
5163 struct io_ring_ctx *ctx = req->ctx;
5164 bool cancel = false;
5166 INIT_HLIST_NODE(&req->hash_node);
5167 io_init_poll_iocb(poll, mask, wake_func);
5168 poll->file = req->file;
5169 poll->wait.private = req;
5171 ipt->pt._key = mask;
5174 ipt->nr_entries = 0;
5176 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5177 if (unlikely(!ipt->nr_entries) && !ipt->error)
5178 ipt->error = -EINVAL;
5180 spin_lock(&ctx->completion_lock);
5181 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5182 io_poll_remove_double(req);
5183 if (likely(poll->head)) {
5184 spin_lock_irq(&poll->head->lock);
5185 if (unlikely(list_empty(&poll->wait.entry))) {
5191 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5192 list_del_init(&poll->wait.entry);
5194 WRITE_ONCE(poll->canceled, true);
5195 else if (!poll->done) /* actually waiting for an event */
5196 io_poll_req_insert(req);
5197 spin_unlock_irq(&poll->head->lock);
5209 static int io_arm_poll_handler(struct io_kiocb *req)
5211 const struct io_op_def *def = &io_op_defs[req->opcode];
5212 struct io_ring_ctx *ctx = req->ctx;
5213 struct async_poll *apoll;
5214 struct io_poll_table ipt;
5215 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5218 if (!req->file || !file_can_poll(req->file))
5219 return IO_APOLL_ABORTED;
5220 if (req->flags & REQ_F_POLLED)
5221 return IO_APOLL_ABORTED;
5222 if (!def->pollin && !def->pollout)
5223 return IO_APOLL_ABORTED;
5227 mask |= POLLIN | POLLRDNORM;
5229 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5230 if ((req->opcode == IORING_OP_RECVMSG) &&
5231 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5235 mask |= POLLOUT | POLLWRNORM;
5238 /* if we can't nonblock try, then no point in arming a poll handler */
5239 if (!io_file_supports_nowait(req, rw))
5240 return IO_APOLL_ABORTED;
5242 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5243 if (unlikely(!apoll))
5244 return IO_APOLL_ABORTED;
5245 apoll->double_poll = NULL;
5247 req->flags |= REQ_F_POLLED;
5248 ipt.pt._qproc = io_async_queue_proc;
5249 io_req_set_refcount(req);
5251 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5253 spin_unlock(&ctx->completion_lock);
5254 if (ret || ipt.error)
5255 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5257 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5258 mask, apoll->poll.events);
5262 static bool __io_poll_remove_one(struct io_kiocb *req,
5263 struct io_poll_iocb *poll, bool do_cancel)
5264 __must_hold(&req->ctx->completion_lock)
5266 bool do_complete = false;
5270 spin_lock_irq(&poll->head->lock);
5272 WRITE_ONCE(poll->canceled, true);
5273 if (!list_empty(&poll->wait.entry)) {
5274 list_del_init(&poll->wait.entry);
5277 spin_unlock_irq(&poll->head->lock);
5278 hash_del(&req->hash_node);
5282 static bool io_poll_remove_one(struct io_kiocb *req)
5283 __must_hold(&req->ctx->completion_lock)
5287 io_poll_remove_double(req);
5288 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5291 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5292 io_commit_cqring(req->ctx);
5294 io_put_req_deferred(req);
5300 * Returns true if we found and killed one or more poll requests
5302 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5305 struct hlist_node *tmp;
5306 struct io_kiocb *req;
5309 spin_lock(&ctx->completion_lock);
5310 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5311 struct hlist_head *list;
5313 list = &ctx->cancel_hash[i];
5314 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5315 if (io_match_task(req, tsk, cancel_all))
5316 posted += io_poll_remove_one(req);
5319 spin_unlock(&ctx->completion_lock);
5322 io_cqring_ev_posted(ctx);
5327 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5329 __must_hold(&ctx->completion_lock)
5331 struct hlist_head *list;
5332 struct io_kiocb *req;
5334 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5335 hlist_for_each_entry(req, list, hash_node) {
5336 if (sqe_addr != req->user_data)
5338 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5345 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5347 __must_hold(&ctx->completion_lock)
5349 struct io_kiocb *req;
5351 req = io_poll_find(ctx, sqe_addr, poll_only);
5354 if (io_poll_remove_one(req))
5360 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5365 events = READ_ONCE(sqe->poll32_events);
5367 events = swahw32(events);
5369 if (!(flags & IORING_POLL_ADD_MULTI))
5370 events |= EPOLLONESHOT;
5371 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5374 static int io_poll_update_prep(struct io_kiocb *req,
5375 const struct io_uring_sqe *sqe)
5377 struct io_poll_update *upd = &req->poll_update;
5380 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5382 if (sqe->ioprio || sqe->buf_index)
5384 flags = READ_ONCE(sqe->len);
5385 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5386 IORING_POLL_ADD_MULTI))
5388 /* meaningless without update */
5389 if (flags == IORING_POLL_ADD_MULTI)
5392 upd->old_user_data = READ_ONCE(sqe->addr);
5393 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5394 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5396 upd->new_user_data = READ_ONCE(sqe->off);
5397 if (!upd->update_user_data && upd->new_user_data)
5399 if (upd->update_events)
5400 upd->events = io_poll_parse_events(sqe, flags);
5401 else if (sqe->poll32_events)
5407 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5410 struct io_kiocb *req = wait->private;
5411 struct io_poll_iocb *poll = &req->poll;
5413 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5416 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5417 struct poll_table_struct *p)
5419 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5421 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5424 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5426 struct io_poll_iocb *poll = &req->poll;
5429 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5431 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5433 flags = READ_ONCE(sqe->len);
5434 if (flags & ~IORING_POLL_ADD_MULTI)
5437 io_req_set_refcount(req);
5438 poll->events = io_poll_parse_events(sqe, flags);
5442 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5444 struct io_poll_iocb *poll = &req->poll;
5445 struct io_ring_ctx *ctx = req->ctx;
5446 struct io_poll_table ipt;
5449 ipt.pt._qproc = io_poll_queue_proc;
5451 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5454 if (mask) { /* no async, we'd stolen it */
5456 io_poll_complete(req, mask);
5458 spin_unlock(&ctx->completion_lock);
5461 io_cqring_ev_posted(ctx);
5462 if (poll->events & EPOLLONESHOT)
5468 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5470 struct io_ring_ctx *ctx = req->ctx;
5471 struct io_kiocb *preq;
5475 spin_lock(&ctx->completion_lock);
5476 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5482 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5484 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5489 * Don't allow racy completion with singleshot, as we cannot safely
5490 * update those. For multishot, if we're racing with completion, just
5491 * let completion re-add it.
5493 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5494 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5498 /* we now have a detached poll request. reissue. */
5502 spin_unlock(&ctx->completion_lock);
5504 io_req_complete(req, ret);
5507 /* only mask one event flags, keep behavior flags */
5508 if (req->poll_update.update_events) {
5509 preq->poll.events &= ~0xffff;
5510 preq->poll.events |= req->poll_update.events & 0xffff;
5511 preq->poll.events |= IO_POLL_UNMASK;
5513 if (req->poll_update.update_user_data)
5514 preq->user_data = req->poll_update.new_user_data;
5515 spin_unlock(&ctx->completion_lock);
5517 /* complete update request, we're done with it */
5518 io_req_complete(req, ret);
5521 ret = io_poll_add(preq, issue_flags);
5524 io_req_complete(preq, ret);
5530 static void io_req_task_timeout(struct io_kiocb *req)
5533 io_req_complete_post(req, -ETIME, 0);
5536 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5538 struct io_timeout_data *data = container_of(timer,
5539 struct io_timeout_data, timer);
5540 struct io_kiocb *req = data->req;
5541 struct io_ring_ctx *ctx = req->ctx;
5542 unsigned long flags;
5544 spin_lock_irqsave(&ctx->timeout_lock, flags);
5545 list_del_init(&req->timeout.list);
5546 atomic_set(&req->ctx->cq_timeouts,
5547 atomic_read(&req->ctx->cq_timeouts) + 1);
5548 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5550 req->io_task_work.func = io_req_task_timeout;
5551 io_req_task_work_add(req);
5552 return HRTIMER_NORESTART;
5555 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5557 __must_hold(&ctx->timeout_lock)
5559 struct io_timeout_data *io;
5560 struct io_kiocb *req;
5563 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5564 found = user_data == req->user_data;
5569 return ERR_PTR(-ENOENT);
5571 io = req->async_data;
5572 if (hrtimer_try_to_cancel(&io->timer) == -1)
5573 return ERR_PTR(-EALREADY);
5574 list_del_init(&req->timeout.list);
5578 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5579 __must_hold(&ctx->completion_lock)
5580 __must_hold(&ctx->timeout_lock)
5582 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5585 return PTR_ERR(req);
5588 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5589 io_put_req_deferred(req);
5593 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5594 struct timespec64 *ts, enum hrtimer_mode mode)
5595 __must_hold(&ctx->timeout_lock)
5597 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5598 struct io_timeout_data *data;
5601 return PTR_ERR(req);
5603 req->timeout.off = 0; /* noseq */
5604 data = req->async_data;
5605 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5606 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5607 data->timer.function = io_timeout_fn;
5608 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5612 static int io_timeout_remove_prep(struct io_kiocb *req,
5613 const struct io_uring_sqe *sqe)
5615 struct io_timeout_rem *tr = &req->timeout_rem;
5617 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5619 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5621 if (sqe->ioprio || sqe->buf_index || sqe->len)
5624 tr->addr = READ_ONCE(sqe->addr);
5625 tr->flags = READ_ONCE(sqe->timeout_flags);
5626 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5627 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5629 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5631 } else if (tr->flags) {
5632 /* timeout removal doesn't support flags */
5639 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5641 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5646 * Remove or update an existing timeout command
5648 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5650 struct io_timeout_rem *tr = &req->timeout_rem;
5651 struct io_ring_ctx *ctx = req->ctx;
5654 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5655 spin_lock(&ctx->completion_lock);
5656 spin_lock_irq(&ctx->timeout_lock);
5657 ret = io_timeout_cancel(ctx, tr->addr);
5658 spin_unlock_irq(&ctx->timeout_lock);
5659 spin_unlock(&ctx->completion_lock);
5661 spin_lock_irq(&ctx->timeout_lock);
5662 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5663 io_translate_timeout_mode(tr->flags));
5664 spin_unlock_irq(&ctx->timeout_lock);
5669 io_req_complete_post(req, ret, 0);
5673 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5674 bool is_timeout_link)
5676 struct io_timeout_data *data;
5678 u32 off = READ_ONCE(sqe->off);
5680 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5682 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5684 if (off && is_timeout_link)
5686 flags = READ_ONCE(sqe->timeout_flags);
5687 if (flags & ~IORING_TIMEOUT_ABS)
5690 req->timeout.off = off;
5691 if (unlikely(off && !req->ctx->off_timeout_used))
5692 req->ctx->off_timeout_used = true;
5694 if (!req->async_data && io_alloc_async_data(req))
5697 data = req->async_data;
5700 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5703 data->mode = io_translate_timeout_mode(flags);
5704 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5706 if (is_timeout_link) {
5707 struct io_submit_link *link = &req->ctx->submit_state.link;
5711 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5713 req->timeout.head = link->last;
5714 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5719 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5721 struct io_ring_ctx *ctx = req->ctx;
5722 struct io_timeout_data *data = req->async_data;
5723 struct list_head *entry;
5724 u32 tail, off = req->timeout.off;
5726 spin_lock_irq(&ctx->timeout_lock);
5729 * sqe->off holds how many events that need to occur for this
5730 * timeout event to be satisfied. If it isn't set, then this is
5731 * a pure timeout request, sequence isn't used.
5733 if (io_is_timeout_noseq(req)) {
5734 entry = ctx->timeout_list.prev;
5738 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5739 req->timeout.target_seq = tail + off;
5741 /* Update the last seq here in case io_flush_timeouts() hasn't.
5742 * This is safe because ->completion_lock is held, and submissions
5743 * and completions are never mixed in the same ->completion_lock section.
5745 ctx->cq_last_tm_flush = tail;
5748 * Insertion sort, ensuring the first entry in the list is always
5749 * the one we need first.
5751 list_for_each_prev(entry, &ctx->timeout_list) {
5752 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5755 if (io_is_timeout_noseq(nxt))
5757 /* nxt.seq is behind @tail, otherwise would've been completed */
5758 if (off >= nxt->timeout.target_seq - tail)
5762 list_add(&req->timeout.list, entry);
5763 data->timer.function = io_timeout_fn;
5764 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5765 spin_unlock_irq(&ctx->timeout_lock);
5769 struct io_cancel_data {
5770 struct io_ring_ctx *ctx;
5774 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5776 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5777 struct io_cancel_data *cd = data;
5779 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5782 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5783 struct io_ring_ctx *ctx)
5785 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5786 enum io_wq_cancel cancel_ret;
5789 if (!tctx || !tctx->io_wq)
5792 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5793 switch (cancel_ret) {
5794 case IO_WQ_CANCEL_OK:
5797 case IO_WQ_CANCEL_RUNNING:
5800 case IO_WQ_CANCEL_NOTFOUND:
5808 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5810 struct io_ring_ctx *ctx = req->ctx;
5813 WARN_ON_ONCE(req->task != current);
5815 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5819 spin_lock(&ctx->completion_lock);
5820 spin_lock_irq(&ctx->timeout_lock);
5821 ret = io_timeout_cancel(ctx, sqe_addr);
5822 spin_unlock_irq(&ctx->timeout_lock);
5825 ret = io_poll_cancel(ctx, sqe_addr, false);
5827 spin_unlock(&ctx->completion_lock);
5831 static int io_async_cancel_prep(struct io_kiocb *req,
5832 const struct io_uring_sqe *sqe)
5834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5836 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5838 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5841 req->cancel.addr = READ_ONCE(sqe->addr);
5845 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5847 struct io_ring_ctx *ctx = req->ctx;
5848 u64 sqe_addr = req->cancel.addr;
5849 struct io_tctx_node *node;
5852 ret = io_try_cancel_userdata(req, sqe_addr);
5856 /* slow path, try all io-wq's */
5857 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5859 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5860 struct io_uring_task *tctx = node->task->io_uring;
5862 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5866 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5870 io_req_complete_post(req, ret, 0);
5874 static int io_rsrc_update_prep(struct io_kiocb *req,
5875 const struct io_uring_sqe *sqe)
5877 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5879 if (sqe->ioprio || sqe->rw_flags)
5882 req->rsrc_update.offset = READ_ONCE(sqe->off);
5883 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5884 if (!req->rsrc_update.nr_args)
5886 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5890 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5892 struct io_ring_ctx *ctx = req->ctx;
5893 struct io_uring_rsrc_update2 up;
5896 if (issue_flags & IO_URING_F_NONBLOCK)
5899 up.offset = req->rsrc_update.offset;
5900 up.data = req->rsrc_update.arg;
5905 mutex_lock(&ctx->uring_lock);
5906 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5907 &up, req->rsrc_update.nr_args);
5908 mutex_unlock(&ctx->uring_lock);
5912 __io_req_complete(req, issue_flags, ret, 0);
5916 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5918 switch (req->opcode) {
5921 case IORING_OP_READV:
5922 case IORING_OP_READ_FIXED:
5923 case IORING_OP_READ:
5924 return io_read_prep(req, sqe);
5925 case IORING_OP_WRITEV:
5926 case IORING_OP_WRITE_FIXED:
5927 case IORING_OP_WRITE:
5928 return io_write_prep(req, sqe);
5929 case IORING_OP_POLL_ADD:
5930 return io_poll_add_prep(req, sqe);
5931 case IORING_OP_POLL_REMOVE:
5932 return io_poll_update_prep(req, sqe);
5933 case IORING_OP_FSYNC:
5934 return io_fsync_prep(req, sqe);
5935 case IORING_OP_SYNC_FILE_RANGE:
5936 return io_sfr_prep(req, sqe);
5937 case IORING_OP_SENDMSG:
5938 case IORING_OP_SEND:
5939 return io_sendmsg_prep(req, sqe);
5940 case IORING_OP_RECVMSG:
5941 case IORING_OP_RECV:
5942 return io_recvmsg_prep(req, sqe);
5943 case IORING_OP_CONNECT:
5944 return io_connect_prep(req, sqe);
5945 case IORING_OP_TIMEOUT:
5946 return io_timeout_prep(req, sqe, false);
5947 case IORING_OP_TIMEOUT_REMOVE:
5948 return io_timeout_remove_prep(req, sqe);
5949 case IORING_OP_ASYNC_CANCEL:
5950 return io_async_cancel_prep(req, sqe);
5951 case IORING_OP_LINK_TIMEOUT:
5952 return io_timeout_prep(req, sqe, true);
5953 case IORING_OP_ACCEPT:
5954 return io_accept_prep(req, sqe);
5955 case IORING_OP_FALLOCATE:
5956 return io_fallocate_prep(req, sqe);
5957 case IORING_OP_OPENAT:
5958 return io_openat_prep(req, sqe);
5959 case IORING_OP_CLOSE:
5960 return io_close_prep(req, sqe);
5961 case IORING_OP_FILES_UPDATE:
5962 return io_rsrc_update_prep(req, sqe);
5963 case IORING_OP_STATX:
5964 return io_statx_prep(req, sqe);
5965 case IORING_OP_FADVISE:
5966 return io_fadvise_prep(req, sqe);
5967 case IORING_OP_MADVISE:
5968 return io_madvise_prep(req, sqe);
5969 case IORING_OP_OPENAT2:
5970 return io_openat2_prep(req, sqe);
5971 case IORING_OP_EPOLL_CTL:
5972 return io_epoll_ctl_prep(req, sqe);
5973 case IORING_OP_SPLICE:
5974 return io_splice_prep(req, sqe);
5975 case IORING_OP_PROVIDE_BUFFERS:
5976 return io_provide_buffers_prep(req, sqe);
5977 case IORING_OP_REMOVE_BUFFERS:
5978 return io_remove_buffers_prep(req, sqe);
5980 return io_tee_prep(req, sqe);
5981 case IORING_OP_SHUTDOWN:
5982 return io_shutdown_prep(req, sqe);
5983 case IORING_OP_RENAMEAT:
5984 return io_renameat_prep(req, sqe);
5985 case IORING_OP_UNLINKAT:
5986 return io_unlinkat_prep(req, sqe);
5989 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5994 static int io_req_prep_async(struct io_kiocb *req)
5996 if (!io_op_defs[req->opcode].needs_async_setup)
5998 if (WARN_ON_ONCE(req->async_data))
6000 if (io_alloc_async_data(req))
6003 switch (req->opcode) {
6004 case IORING_OP_READV:
6005 return io_rw_prep_async(req, READ);
6006 case IORING_OP_WRITEV:
6007 return io_rw_prep_async(req, WRITE);
6008 case IORING_OP_SENDMSG:
6009 return io_sendmsg_prep_async(req);
6010 case IORING_OP_RECVMSG:
6011 return io_recvmsg_prep_async(req);
6012 case IORING_OP_CONNECT:
6013 return io_connect_prep_async(req);
6015 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6020 static u32 io_get_sequence(struct io_kiocb *req)
6022 u32 seq = req->ctx->cached_sq_head;
6024 /* need original cached_sq_head, but it was increased for each req */
6025 io_for_each_link(req, req)
6030 static bool io_drain_req(struct io_kiocb *req)
6032 struct io_kiocb *pos;
6033 struct io_ring_ctx *ctx = req->ctx;
6034 struct io_defer_entry *de;
6039 * If we need to drain a request in the middle of a link, drain the
6040 * head request and the next request/link after the current link.
6041 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6042 * maintained for every request of our link.
6044 if (ctx->drain_next) {
6045 req->flags |= REQ_F_IO_DRAIN;
6046 ctx->drain_next = false;
6048 /* not interested in head, start from the first linked */
6049 io_for_each_link(pos, req->link) {
6050 if (pos->flags & REQ_F_IO_DRAIN) {
6051 ctx->drain_next = true;
6052 req->flags |= REQ_F_IO_DRAIN;
6057 /* Still need defer if there is pending req in defer list. */
6058 if (likely(list_empty_careful(&ctx->defer_list) &&
6059 !(req->flags & REQ_F_IO_DRAIN))) {
6060 ctx->drain_active = false;
6064 seq = io_get_sequence(req);
6065 /* Still a chance to pass the sequence check */
6066 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6069 ret = io_req_prep_async(req);
6072 io_prep_async_link(req);
6073 de = kmalloc(sizeof(*de), GFP_KERNEL);
6077 io_req_complete_failed(req, ret);
6081 spin_lock(&ctx->completion_lock);
6082 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6083 spin_unlock(&ctx->completion_lock);
6085 io_queue_async_work(req);
6089 trace_io_uring_defer(ctx, req, req->user_data);
6092 list_add_tail(&de->list, &ctx->defer_list);
6093 spin_unlock(&ctx->completion_lock);
6097 static void io_clean_op(struct io_kiocb *req)
6099 if (req->flags & REQ_F_BUFFER_SELECTED) {
6100 switch (req->opcode) {
6101 case IORING_OP_READV:
6102 case IORING_OP_READ_FIXED:
6103 case IORING_OP_READ:
6104 kfree((void *)(unsigned long)req->rw.addr);
6106 case IORING_OP_RECVMSG:
6107 case IORING_OP_RECV:
6108 kfree(req->sr_msg.kbuf);
6113 if (req->flags & REQ_F_NEED_CLEANUP) {
6114 switch (req->opcode) {
6115 case IORING_OP_READV:
6116 case IORING_OP_READ_FIXED:
6117 case IORING_OP_READ:
6118 case IORING_OP_WRITEV:
6119 case IORING_OP_WRITE_FIXED:
6120 case IORING_OP_WRITE: {
6121 struct io_async_rw *io = req->async_data;
6123 kfree(io->free_iovec);
6126 case IORING_OP_RECVMSG:
6127 case IORING_OP_SENDMSG: {
6128 struct io_async_msghdr *io = req->async_data;
6130 kfree(io->free_iov);
6133 case IORING_OP_SPLICE:
6135 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6136 io_put_file(req->splice.file_in);
6138 case IORING_OP_OPENAT:
6139 case IORING_OP_OPENAT2:
6140 if (req->open.filename)
6141 putname(req->open.filename);
6143 case IORING_OP_RENAMEAT:
6144 putname(req->rename.oldpath);
6145 putname(req->rename.newpath);
6147 case IORING_OP_UNLINKAT:
6148 putname(req->unlink.filename);
6152 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6153 kfree(req->apoll->double_poll);
6157 if (req->flags & REQ_F_INFLIGHT) {
6158 struct io_uring_task *tctx = req->task->io_uring;
6160 atomic_dec(&tctx->inflight_tracked);
6162 if (req->flags & REQ_F_CREDS)
6163 put_cred(req->creds);
6165 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6168 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6170 struct io_ring_ctx *ctx = req->ctx;
6171 const struct cred *creds = NULL;
6174 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6175 creds = override_creds(req->creds);
6177 switch (req->opcode) {
6179 ret = io_nop(req, issue_flags);
6181 case IORING_OP_READV:
6182 case IORING_OP_READ_FIXED:
6183 case IORING_OP_READ:
6184 ret = io_read(req, issue_flags);
6186 case IORING_OP_WRITEV:
6187 case IORING_OP_WRITE_FIXED:
6188 case IORING_OP_WRITE:
6189 ret = io_write(req, issue_flags);
6191 case IORING_OP_FSYNC:
6192 ret = io_fsync(req, issue_flags);
6194 case IORING_OP_POLL_ADD:
6195 ret = io_poll_add(req, issue_flags);
6197 case IORING_OP_POLL_REMOVE:
6198 ret = io_poll_update(req, issue_flags);
6200 case IORING_OP_SYNC_FILE_RANGE:
6201 ret = io_sync_file_range(req, issue_flags);
6203 case IORING_OP_SENDMSG:
6204 ret = io_sendmsg(req, issue_flags);
6206 case IORING_OP_SEND:
6207 ret = io_send(req, issue_flags);
6209 case IORING_OP_RECVMSG:
6210 ret = io_recvmsg(req, issue_flags);
6212 case IORING_OP_RECV:
6213 ret = io_recv(req, issue_flags);
6215 case IORING_OP_TIMEOUT:
6216 ret = io_timeout(req, issue_flags);
6218 case IORING_OP_TIMEOUT_REMOVE:
6219 ret = io_timeout_remove(req, issue_flags);
6221 case IORING_OP_ACCEPT:
6222 ret = io_accept(req, issue_flags);
6224 case IORING_OP_CONNECT:
6225 ret = io_connect(req, issue_flags);
6227 case IORING_OP_ASYNC_CANCEL:
6228 ret = io_async_cancel(req, issue_flags);
6230 case IORING_OP_FALLOCATE:
6231 ret = io_fallocate(req, issue_flags);
6233 case IORING_OP_OPENAT:
6234 ret = io_openat(req, issue_flags);
6236 case IORING_OP_CLOSE:
6237 ret = io_close(req, issue_flags);
6239 case IORING_OP_FILES_UPDATE:
6240 ret = io_files_update(req, issue_flags);
6242 case IORING_OP_STATX:
6243 ret = io_statx(req, issue_flags);
6245 case IORING_OP_FADVISE:
6246 ret = io_fadvise(req, issue_flags);
6248 case IORING_OP_MADVISE:
6249 ret = io_madvise(req, issue_flags);
6251 case IORING_OP_OPENAT2:
6252 ret = io_openat2(req, issue_flags);
6254 case IORING_OP_EPOLL_CTL:
6255 ret = io_epoll_ctl(req, issue_flags);
6257 case IORING_OP_SPLICE:
6258 ret = io_splice(req, issue_flags);
6260 case IORING_OP_PROVIDE_BUFFERS:
6261 ret = io_provide_buffers(req, issue_flags);
6263 case IORING_OP_REMOVE_BUFFERS:
6264 ret = io_remove_buffers(req, issue_flags);
6267 ret = io_tee(req, issue_flags);
6269 case IORING_OP_SHUTDOWN:
6270 ret = io_shutdown(req, issue_flags);
6272 case IORING_OP_RENAMEAT:
6273 ret = io_renameat(req, issue_flags);
6275 case IORING_OP_UNLINKAT:
6276 ret = io_unlinkat(req, issue_flags);
6284 revert_creds(creds);
6287 /* If the op doesn't have a file, we're not polling for it */
6288 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6289 io_iopoll_req_issued(req);
6294 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6296 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6298 req = io_put_req_find_next(req);
6299 return req ? &req->work : NULL;
6302 static void io_wq_submit_work(struct io_wq_work *work)
6304 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6305 struct io_kiocb *timeout;
6308 /* one will be dropped by ->io_free_work() after returning to io-wq */
6309 if (!(req->flags & REQ_F_REFCOUNT))
6310 __io_req_set_refcount(req, 2);
6314 timeout = io_prep_linked_timeout(req);
6316 io_queue_linked_timeout(timeout);
6318 if (work->flags & IO_WQ_WORK_CANCEL)
6323 ret = io_issue_sqe(req, 0);
6325 * We can get EAGAIN for polled IO even though we're
6326 * forcing a sync submission from here, since we can't
6327 * wait for request slots on the block side.
6335 /* avoid locking problems by failing it from a clean context */
6337 io_req_task_queue_fail(req, ret);
6340 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6343 return &table->files[i];
6346 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6349 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6351 return (struct file *) (slot->file_ptr & FFS_MASK);
6354 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6356 unsigned long file_ptr = (unsigned long) file;
6358 if (__io_file_supports_nowait(file, READ))
6359 file_ptr |= FFS_ASYNC_READ;
6360 if (__io_file_supports_nowait(file, WRITE))
6361 file_ptr |= FFS_ASYNC_WRITE;
6362 if (S_ISREG(file_inode(file)->i_mode))
6363 file_ptr |= FFS_ISREG;
6364 file_slot->file_ptr = file_ptr;
6367 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6368 struct io_kiocb *req, int fd)
6371 unsigned long file_ptr;
6373 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6375 fd = array_index_nospec(fd, ctx->nr_user_files);
6376 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6377 file = (struct file *) (file_ptr & FFS_MASK);
6378 file_ptr &= ~FFS_MASK;
6379 /* mask in overlapping REQ_F and FFS bits */
6380 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6381 io_req_set_rsrc_node(req);
6385 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6386 struct io_kiocb *req, int fd)
6388 struct file *file = fget(fd);
6390 trace_io_uring_file_get(ctx, fd);
6392 /* we don't allow fixed io_uring files */
6393 if (file && unlikely(file->f_op == &io_uring_fops))
6394 io_req_track_inflight(req);
6398 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6399 struct io_kiocb *req, int fd, bool fixed)
6402 return io_file_get_fixed(ctx, req, fd);
6404 return io_file_get_normal(ctx, req, fd);
6407 static void io_req_task_link_timeout(struct io_kiocb *req)
6409 struct io_kiocb *prev = req->timeout.prev;
6413 ret = io_try_cancel_userdata(req, prev->user_data);
6414 io_req_complete_post(req, ret ?: -ETIME, 0);
6417 io_req_complete_post(req, -ETIME, 0);
6421 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6423 struct io_timeout_data *data = container_of(timer,
6424 struct io_timeout_data, timer);
6425 struct io_kiocb *prev, *req = data->req;
6426 struct io_ring_ctx *ctx = req->ctx;
6427 unsigned long flags;
6429 spin_lock_irqsave(&ctx->timeout_lock, flags);
6430 prev = req->timeout.head;
6431 req->timeout.head = NULL;
6434 * We don't expect the list to be empty, that will only happen if we
6435 * race with the completion of the linked work.
6438 io_remove_next_linked(prev);
6439 if (!req_ref_inc_not_zero(prev))
6442 req->timeout.prev = prev;
6443 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6445 req->io_task_work.func = io_req_task_link_timeout;
6446 io_req_task_work_add(req);
6447 return HRTIMER_NORESTART;
6450 static void io_queue_linked_timeout(struct io_kiocb *req)
6452 struct io_ring_ctx *ctx = req->ctx;
6454 spin_lock_irq(&ctx->timeout_lock);
6456 * If the back reference is NULL, then our linked request finished
6457 * before we got a chance to setup the timer
6459 if (req->timeout.head) {
6460 struct io_timeout_data *data = req->async_data;
6462 data->timer.function = io_link_timeout_fn;
6463 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6466 spin_unlock_irq(&ctx->timeout_lock);
6467 /* drop submission reference */
6471 static void __io_queue_sqe(struct io_kiocb *req)
6472 __must_hold(&req->ctx->uring_lock)
6474 struct io_kiocb *linked_timeout;
6478 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6481 * We async punt it if the file wasn't marked NOWAIT, or if the file
6482 * doesn't support non-blocking read/write attempts
6485 if (req->flags & REQ_F_COMPLETE_INLINE) {
6486 struct io_ring_ctx *ctx = req->ctx;
6487 struct io_submit_state *state = &ctx->submit_state;
6489 state->compl_reqs[state->compl_nr++] = req;
6490 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6491 io_submit_flush_completions(ctx);
6495 linked_timeout = io_prep_linked_timeout(req);
6497 io_queue_linked_timeout(linked_timeout);
6498 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6499 linked_timeout = io_prep_linked_timeout(req);
6501 switch (io_arm_poll_handler(req)) {
6502 case IO_APOLL_READY:
6504 io_unprep_linked_timeout(req);
6506 case IO_APOLL_ABORTED:
6508 * Queued up for async execution, worker will release
6509 * submit reference when the iocb is actually submitted.
6511 io_queue_async_work(req);
6516 io_queue_linked_timeout(linked_timeout);
6518 io_req_complete_failed(req, ret);
6522 static inline void io_queue_sqe(struct io_kiocb *req)
6523 __must_hold(&req->ctx->uring_lock)
6525 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6528 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6529 __io_queue_sqe(req);
6531 int ret = io_req_prep_async(req);
6534 io_req_complete_failed(req, ret);
6536 io_queue_async_work(req);
6541 * Check SQE restrictions (opcode and flags).
6543 * Returns 'true' if SQE is allowed, 'false' otherwise.
6545 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6546 struct io_kiocb *req,
6547 unsigned int sqe_flags)
6549 if (likely(!ctx->restricted))
6552 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6555 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6556 ctx->restrictions.sqe_flags_required)
6559 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6560 ctx->restrictions.sqe_flags_required))
6566 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6567 const struct io_uring_sqe *sqe)
6568 __must_hold(&ctx->uring_lock)
6570 struct io_submit_state *state;
6571 unsigned int sqe_flags;
6572 int personality, ret = 0;
6574 /* req is partially pre-initialised, see io_preinit_req() */
6575 req->opcode = READ_ONCE(sqe->opcode);
6576 /* same numerical values with corresponding REQ_F_*, safe to copy */
6577 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6578 req->user_data = READ_ONCE(sqe->user_data);
6580 req->fixed_rsrc_refs = NULL;
6581 req->task = current;
6583 /* enforce forwards compatibility on users */
6584 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6586 if (unlikely(req->opcode >= IORING_OP_LAST))
6588 if (!io_check_restriction(ctx, req, sqe_flags))
6591 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6592 !io_op_defs[req->opcode].buffer_select)
6594 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6595 ctx->drain_active = true;
6597 personality = READ_ONCE(sqe->personality);
6599 req->creds = xa_load(&ctx->personalities, personality);
6602 get_cred(req->creds);
6603 req->flags |= REQ_F_CREDS;
6605 state = &ctx->submit_state;
6608 * Plug now if we have more than 1 IO left after this, and the target
6609 * is potentially a read/write to block based storage.
6611 if (!state->plug_started && state->ios_left > 1 &&
6612 io_op_defs[req->opcode].plug) {
6613 blk_start_plug(&state->plug);
6614 state->plug_started = true;
6617 if (io_op_defs[req->opcode].needs_file) {
6618 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6619 (sqe_flags & IOSQE_FIXED_FILE));
6620 if (unlikely(!req->file))
6628 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6629 const struct io_uring_sqe *sqe)
6630 __must_hold(&ctx->uring_lock)
6632 struct io_submit_link *link = &ctx->submit_state.link;
6635 ret = io_init_req(ctx, req, sqe);
6636 if (unlikely(ret)) {
6639 /* fail even hard links since we don't submit */
6640 req_set_fail(link->head);
6641 io_req_complete_failed(link->head, -ECANCELED);
6644 io_req_complete_failed(req, ret);
6648 ret = io_req_prep(req, sqe);
6652 /* don't need @sqe from now on */
6653 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6655 ctx->flags & IORING_SETUP_SQPOLL);
6658 * If we already have a head request, queue this one for async
6659 * submittal once the head completes. If we don't have a head but
6660 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6661 * submitted sync once the chain is complete. If none of those
6662 * conditions are true (normal request), then just queue it.
6665 struct io_kiocb *head = link->head;
6667 ret = io_req_prep_async(req);
6670 trace_io_uring_link(ctx, req, head);
6671 link->last->link = req;
6674 /* last request of a link, enqueue the link */
6675 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6680 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6692 * Batched submission is done, ensure local IO is flushed out.
6694 static void io_submit_state_end(struct io_submit_state *state,
6695 struct io_ring_ctx *ctx)
6697 if (state->link.head)
6698 io_queue_sqe(state->link.head);
6699 if (state->compl_nr)
6700 io_submit_flush_completions(ctx);
6701 if (state->plug_started)
6702 blk_finish_plug(&state->plug);
6706 * Start submission side cache.
6708 static void io_submit_state_start(struct io_submit_state *state,
6709 unsigned int max_ios)
6711 state->plug_started = false;
6712 state->ios_left = max_ios;
6713 /* set only head, no need to init link_last in advance */
6714 state->link.head = NULL;
6717 static void io_commit_sqring(struct io_ring_ctx *ctx)
6719 struct io_rings *rings = ctx->rings;
6722 * Ensure any loads from the SQEs are done at this point,
6723 * since once we write the new head, the application could
6724 * write new data to them.
6726 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6730 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6731 * that is mapped by userspace. This means that care needs to be taken to
6732 * ensure that reads are stable, as we cannot rely on userspace always
6733 * being a good citizen. If members of the sqe are validated and then later
6734 * used, it's important that those reads are done through READ_ONCE() to
6735 * prevent a re-load down the line.
6737 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6739 unsigned head, mask = ctx->sq_entries - 1;
6740 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6743 * The cached sq head (or cq tail) serves two purposes:
6745 * 1) allows us to batch the cost of updating the user visible
6747 * 2) allows the kernel side to track the head on its own, even
6748 * though the application is the one updating it.
6750 head = READ_ONCE(ctx->sq_array[sq_idx]);
6751 if (likely(head < ctx->sq_entries))
6752 return &ctx->sq_sqes[head];
6754 /* drop invalid entries */
6756 WRITE_ONCE(ctx->rings->sq_dropped,
6757 READ_ONCE(ctx->rings->sq_dropped) + 1);
6761 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6762 __must_hold(&ctx->uring_lock)
6764 struct io_uring_task *tctx;
6767 /* make sure SQ entry isn't read before tail */
6768 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6769 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6772 tctx = current->io_uring;
6773 tctx->cached_refs -= nr;
6774 if (unlikely(tctx->cached_refs < 0)) {
6775 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6777 percpu_counter_add(&tctx->inflight, refill);
6778 refcount_add(refill, ¤t->usage);
6779 tctx->cached_refs += refill;
6781 io_submit_state_start(&ctx->submit_state, nr);
6783 while (submitted < nr) {
6784 const struct io_uring_sqe *sqe;
6785 struct io_kiocb *req;
6787 req = io_alloc_req(ctx);
6788 if (unlikely(!req)) {
6790 submitted = -EAGAIN;
6793 sqe = io_get_sqe(ctx);
6794 if (unlikely(!sqe)) {
6795 kmem_cache_free(req_cachep, req);
6798 /* will complete beyond this point, count as submitted */
6800 if (io_submit_sqe(ctx, req, sqe))
6804 if (unlikely(submitted != nr)) {
6805 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6806 int unused = nr - ref_used;
6808 current->io_uring->cached_refs += unused;
6809 percpu_ref_put_many(&ctx->refs, unused);
6812 io_submit_state_end(&ctx->submit_state, ctx);
6813 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6814 io_commit_sqring(ctx);
6819 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6821 return READ_ONCE(sqd->state);
6824 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6826 /* Tell userspace we may need a wakeup call */
6827 spin_lock(&ctx->completion_lock);
6828 WRITE_ONCE(ctx->rings->sq_flags,
6829 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6830 spin_unlock(&ctx->completion_lock);
6833 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6835 spin_lock(&ctx->completion_lock);
6836 WRITE_ONCE(ctx->rings->sq_flags,
6837 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6838 spin_unlock(&ctx->completion_lock);
6841 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6843 unsigned int to_submit;
6846 to_submit = io_sqring_entries(ctx);
6847 /* if we're handling multiple rings, cap submit size for fairness */
6848 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6849 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6851 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6852 unsigned nr_events = 0;
6853 const struct cred *creds = NULL;
6855 if (ctx->sq_creds != current_cred())
6856 creds = override_creds(ctx->sq_creds);
6858 mutex_lock(&ctx->uring_lock);
6859 if (!list_empty(&ctx->iopoll_list))
6860 io_do_iopoll(ctx, &nr_events, 0);
6863 * Don't submit if refs are dying, good for io_uring_register(),
6864 * but also it is relied upon by io_ring_exit_work()
6866 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6867 !(ctx->flags & IORING_SETUP_R_DISABLED))
6868 ret = io_submit_sqes(ctx, to_submit);
6869 mutex_unlock(&ctx->uring_lock);
6871 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6872 wake_up(&ctx->sqo_sq_wait);
6874 revert_creds(creds);
6880 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6882 struct io_ring_ctx *ctx;
6883 unsigned sq_thread_idle = 0;
6885 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6886 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6887 sqd->sq_thread_idle = sq_thread_idle;
6890 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6892 bool did_sig = false;
6893 struct ksignal ksig;
6895 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6896 signal_pending(current)) {
6897 mutex_unlock(&sqd->lock);
6898 if (signal_pending(current))
6899 did_sig = get_signal(&ksig);
6901 mutex_lock(&sqd->lock);
6903 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6906 static int io_sq_thread(void *data)
6908 struct io_sq_data *sqd = data;
6909 struct io_ring_ctx *ctx;
6910 unsigned long timeout = 0;
6911 char buf[TASK_COMM_LEN];
6914 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6915 set_task_comm(current, buf);
6917 if (sqd->sq_cpu != -1)
6918 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6920 set_cpus_allowed_ptr(current, cpu_online_mask);
6921 current->flags |= PF_NO_SETAFFINITY;
6923 mutex_lock(&sqd->lock);
6925 bool cap_entries, sqt_spin = false;
6927 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6928 if (io_sqd_handle_event(sqd))
6930 timeout = jiffies + sqd->sq_thread_idle;
6933 cap_entries = !list_is_singular(&sqd->ctx_list);
6934 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6935 int ret = __io_sq_thread(ctx, cap_entries);
6937 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6940 if (io_run_task_work())
6943 if (sqt_spin || !time_after(jiffies, timeout)) {
6946 timeout = jiffies + sqd->sq_thread_idle;
6950 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6951 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6952 bool needs_sched = true;
6954 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6955 io_ring_set_wakeup_flag(ctx);
6957 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6958 !list_empty_careful(&ctx->iopoll_list)) {
6959 needs_sched = false;
6962 if (io_sqring_entries(ctx)) {
6963 needs_sched = false;
6969 mutex_unlock(&sqd->lock);
6971 mutex_lock(&sqd->lock);
6973 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6974 io_ring_clear_wakeup_flag(ctx);
6977 finish_wait(&sqd->wait, &wait);
6978 timeout = jiffies + sqd->sq_thread_idle;
6981 io_uring_cancel_generic(true, sqd);
6983 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6984 io_ring_set_wakeup_flag(ctx);
6986 mutex_unlock(&sqd->lock);
6988 complete(&sqd->exited);
6992 struct io_wait_queue {
6993 struct wait_queue_entry wq;
6994 struct io_ring_ctx *ctx;
6996 unsigned nr_timeouts;
6999 static inline bool io_should_wake(struct io_wait_queue *iowq)
7001 struct io_ring_ctx *ctx = iowq->ctx;
7002 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7005 * Wake up if we have enough events, or if a timeout occurred since we
7006 * started waiting. For timeouts, we always want to return to userspace,
7007 * regardless of event count.
7009 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7012 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7013 int wake_flags, void *key)
7015 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7019 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7020 * the task, and the next invocation will do it.
7022 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7023 return autoremove_wake_function(curr, mode, wake_flags, key);
7027 static int io_run_task_work_sig(void)
7029 if (io_run_task_work())
7031 if (!signal_pending(current))
7033 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7034 return -ERESTARTSYS;
7038 /* when returns >0, the caller should retry */
7039 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7040 struct io_wait_queue *iowq,
7041 signed long *timeout)
7045 /* make sure we run task_work before checking for signals */
7046 ret = io_run_task_work_sig();
7047 if (ret || io_should_wake(iowq))
7049 /* let the caller flush overflows, retry */
7050 if (test_bit(0, &ctx->check_cq_overflow))
7053 *timeout = schedule_timeout(*timeout);
7054 return !*timeout ? -ETIME : 1;
7058 * Wait until events become available, if we don't already have some. The
7059 * application must reap them itself, as they reside on the shared cq ring.
7061 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7062 const sigset_t __user *sig, size_t sigsz,
7063 struct __kernel_timespec __user *uts)
7065 struct io_wait_queue iowq;
7066 struct io_rings *rings = ctx->rings;
7067 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7071 io_cqring_overflow_flush(ctx);
7072 if (io_cqring_events(ctx) >= min_events)
7074 if (!io_run_task_work())
7079 #ifdef CONFIG_COMPAT
7080 if (in_compat_syscall())
7081 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7085 ret = set_user_sigmask(sig, sigsz);
7092 struct timespec64 ts;
7094 if (get_timespec64(&ts, uts))
7096 timeout = timespec64_to_jiffies(&ts);
7099 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7100 iowq.wq.private = current;
7101 INIT_LIST_HEAD(&iowq.wq.entry);
7103 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7104 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7106 trace_io_uring_cqring_wait(ctx, min_events);
7108 /* if we can't even flush overflow, don't wait for more */
7109 if (!io_cqring_overflow_flush(ctx)) {
7113 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7114 TASK_INTERRUPTIBLE);
7115 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7116 finish_wait(&ctx->cq_wait, &iowq.wq);
7120 restore_saved_sigmask_unless(ret == -EINTR);
7122 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7125 static void io_free_page_table(void **table, size_t size)
7127 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7129 for (i = 0; i < nr_tables; i++)
7134 static void **io_alloc_page_table(size_t size)
7136 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7137 size_t init_size = size;
7140 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7144 for (i = 0; i < nr_tables; i++) {
7145 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7147 table[i] = kzalloc(this_size, GFP_KERNEL);
7149 io_free_page_table(table, init_size);
7157 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7159 percpu_ref_exit(&ref_node->refs);
7163 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7165 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7166 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7167 unsigned long flags;
7168 bool first_add = false;
7170 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7173 while (!list_empty(&ctx->rsrc_ref_list)) {
7174 node = list_first_entry(&ctx->rsrc_ref_list,
7175 struct io_rsrc_node, node);
7176 /* recycle ref nodes in order */
7179 list_del(&node->node);
7180 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7182 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7185 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7188 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7190 struct io_rsrc_node *ref_node;
7192 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7196 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7201 INIT_LIST_HEAD(&ref_node->node);
7202 INIT_LIST_HEAD(&ref_node->rsrc_list);
7203 ref_node->done = false;
7207 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7208 struct io_rsrc_data *data_to_kill)
7210 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7211 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7214 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7216 rsrc_node->rsrc_data = data_to_kill;
7217 spin_lock_irq(&ctx->rsrc_ref_lock);
7218 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7219 spin_unlock_irq(&ctx->rsrc_ref_lock);
7221 atomic_inc(&data_to_kill->refs);
7222 percpu_ref_kill(&rsrc_node->refs);
7223 ctx->rsrc_node = NULL;
7226 if (!ctx->rsrc_node) {
7227 ctx->rsrc_node = ctx->rsrc_backup_node;
7228 ctx->rsrc_backup_node = NULL;
7232 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7234 if (ctx->rsrc_backup_node)
7236 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7237 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7240 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7244 /* As we may drop ->uring_lock, other task may have started quiesce */
7248 data->quiesce = true;
7250 ret = io_rsrc_node_switch_start(ctx);
7253 io_rsrc_node_switch(ctx, data);
7255 /* kill initial ref, already quiesced if zero */
7256 if (atomic_dec_and_test(&data->refs))
7258 mutex_unlock(&ctx->uring_lock);
7259 flush_delayed_work(&ctx->rsrc_put_work);
7260 ret = wait_for_completion_interruptible(&data->done);
7262 mutex_lock(&ctx->uring_lock);
7266 atomic_inc(&data->refs);
7267 /* wait for all works potentially completing data->done */
7268 flush_delayed_work(&ctx->rsrc_put_work);
7269 reinit_completion(&data->done);
7271 ret = io_run_task_work_sig();
7272 mutex_lock(&ctx->uring_lock);
7274 data->quiesce = false;
7279 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7281 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7282 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7284 return &data->tags[table_idx][off];
7287 static void io_rsrc_data_free(struct io_rsrc_data *data)
7289 size_t size = data->nr * sizeof(data->tags[0][0]);
7292 io_free_page_table((void **)data->tags, size);
7296 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7297 u64 __user *utags, unsigned nr,
7298 struct io_rsrc_data **pdata)
7300 struct io_rsrc_data *data;
7304 data = kzalloc(sizeof(*data), GFP_KERNEL);
7307 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7315 data->do_put = do_put;
7318 for (i = 0; i < nr; i++) {
7319 u64 *tag_slot = io_get_tag_slot(data, i);
7321 if (copy_from_user(tag_slot, &utags[i],
7327 atomic_set(&data->refs, 1);
7328 init_completion(&data->done);
7332 io_rsrc_data_free(data);
7336 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7338 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7339 return !!table->files;
7342 static void io_free_file_tables(struct io_file_table *table)
7344 kvfree(table->files);
7345 table->files = NULL;
7348 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7350 #if defined(CONFIG_UNIX)
7351 if (ctx->ring_sock) {
7352 struct sock *sock = ctx->ring_sock->sk;
7353 struct sk_buff *skb;
7355 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7361 for (i = 0; i < ctx->nr_user_files; i++) {
7364 file = io_file_from_index(ctx, i);
7369 io_free_file_tables(&ctx->file_table);
7370 io_rsrc_data_free(ctx->file_data);
7371 ctx->file_data = NULL;
7372 ctx->nr_user_files = 0;
7375 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7379 if (!ctx->file_data)
7381 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7383 __io_sqe_files_unregister(ctx);
7387 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7388 __releases(&sqd->lock)
7390 WARN_ON_ONCE(sqd->thread == current);
7393 * Do the dance but not conditional clear_bit() because it'd race with
7394 * other threads incrementing park_pending and setting the bit.
7396 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7397 if (atomic_dec_return(&sqd->park_pending))
7398 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7399 mutex_unlock(&sqd->lock);
7402 static void io_sq_thread_park(struct io_sq_data *sqd)
7403 __acquires(&sqd->lock)
7405 WARN_ON_ONCE(sqd->thread == current);
7407 atomic_inc(&sqd->park_pending);
7408 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7409 mutex_lock(&sqd->lock);
7411 wake_up_process(sqd->thread);
7414 static void io_sq_thread_stop(struct io_sq_data *sqd)
7416 WARN_ON_ONCE(sqd->thread == current);
7417 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7419 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7420 mutex_lock(&sqd->lock);
7422 wake_up_process(sqd->thread);
7423 mutex_unlock(&sqd->lock);
7424 wait_for_completion(&sqd->exited);
7427 static void io_put_sq_data(struct io_sq_data *sqd)
7429 if (refcount_dec_and_test(&sqd->refs)) {
7430 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7432 io_sq_thread_stop(sqd);
7437 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7439 struct io_sq_data *sqd = ctx->sq_data;
7442 io_sq_thread_park(sqd);
7443 list_del_init(&ctx->sqd_list);
7444 io_sqd_update_thread_idle(sqd);
7445 io_sq_thread_unpark(sqd);
7447 io_put_sq_data(sqd);
7448 ctx->sq_data = NULL;
7452 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7454 struct io_ring_ctx *ctx_attach;
7455 struct io_sq_data *sqd;
7458 f = fdget(p->wq_fd);
7460 return ERR_PTR(-ENXIO);
7461 if (f.file->f_op != &io_uring_fops) {
7463 return ERR_PTR(-EINVAL);
7466 ctx_attach = f.file->private_data;
7467 sqd = ctx_attach->sq_data;
7470 return ERR_PTR(-EINVAL);
7472 if (sqd->task_tgid != current->tgid) {
7474 return ERR_PTR(-EPERM);
7477 refcount_inc(&sqd->refs);
7482 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7485 struct io_sq_data *sqd;
7488 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7489 sqd = io_attach_sq_data(p);
7494 /* fall through for EPERM case, setup new sqd/task */
7495 if (PTR_ERR(sqd) != -EPERM)
7499 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7501 return ERR_PTR(-ENOMEM);
7503 atomic_set(&sqd->park_pending, 0);
7504 refcount_set(&sqd->refs, 1);
7505 INIT_LIST_HEAD(&sqd->ctx_list);
7506 mutex_init(&sqd->lock);
7507 init_waitqueue_head(&sqd->wait);
7508 init_completion(&sqd->exited);
7512 #if defined(CONFIG_UNIX)
7514 * Ensure the UNIX gc is aware of our file set, so we are certain that
7515 * the io_uring can be safely unregistered on process exit, even if we have
7516 * loops in the file referencing.
7518 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7520 struct sock *sk = ctx->ring_sock->sk;
7521 struct scm_fp_list *fpl;
7522 struct sk_buff *skb;
7525 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7529 skb = alloc_skb(0, GFP_KERNEL);
7538 fpl->user = get_uid(current_user());
7539 for (i = 0; i < nr; i++) {
7540 struct file *file = io_file_from_index(ctx, i + offset);
7544 fpl->fp[nr_files] = get_file(file);
7545 unix_inflight(fpl->user, fpl->fp[nr_files]);
7550 fpl->max = SCM_MAX_FD;
7551 fpl->count = nr_files;
7552 UNIXCB(skb).fp = fpl;
7553 skb->destructor = unix_destruct_scm;
7554 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7555 skb_queue_head(&sk->sk_receive_queue, skb);
7557 for (i = 0; i < nr_files; i++)
7568 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7569 * causes regular reference counting to break down. We rely on the UNIX
7570 * garbage collection to take care of this problem for us.
7572 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7574 unsigned left, total;
7578 left = ctx->nr_user_files;
7580 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7582 ret = __io_sqe_files_scm(ctx, this_files, total);
7586 total += this_files;
7592 while (total < ctx->nr_user_files) {
7593 struct file *file = io_file_from_index(ctx, total);
7603 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7609 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7611 struct file *file = prsrc->file;
7612 #if defined(CONFIG_UNIX)
7613 struct sock *sock = ctx->ring_sock->sk;
7614 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7615 struct sk_buff *skb;
7618 __skb_queue_head_init(&list);
7621 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7622 * remove this entry and rearrange the file array.
7624 skb = skb_dequeue(head);
7626 struct scm_fp_list *fp;
7628 fp = UNIXCB(skb).fp;
7629 for (i = 0; i < fp->count; i++) {
7632 if (fp->fp[i] != file)
7635 unix_notinflight(fp->user, fp->fp[i]);
7636 left = fp->count - 1 - i;
7638 memmove(&fp->fp[i], &fp->fp[i + 1],
7639 left * sizeof(struct file *));
7646 __skb_queue_tail(&list, skb);
7656 __skb_queue_tail(&list, skb);
7658 skb = skb_dequeue(head);
7661 if (skb_peek(&list)) {
7662 spin_lock_irq(&head->lock);
7663 while ((skb = __skb_dequeue(&list)) != NULL)
7664 __skb_queue_tail(head, skb);
7665 spin_unlock_irq(&head->lock);
7672 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7674 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7675 struct io_ring_ctx *ctx = rsrc_data->ctx;
7676 struct io_rsrc_put *prsrc, *tmp;
7678 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7679 list_del(&prsrc->list);
7682 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7684 io_ring_submit_lock(ctx, lock_ring);
7685 spin_lock(&ctx->completion_lock);
7686 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7688 io_commit_cqring(ctx);
7689 spin_unlock(&ctx->completion_lock);
7690 io_cqring_ev_posted(ctx);
7691 io_ring_submit_unlock(ctx, lock_ring);
7694 rsrc_data->do_put(ctx, prsrc);
7698 io_rsrc_node_destroy(ref_node);
7699 if (atomic_dec_and_test(&rsrc_data->refs))
7700 complete(&rsrc_data->done);
7703 static void io_rsrc_put_work(struct work_struct *work)
7705 struct io_ring_ctx *ctx;
7706 struct llist_node *node;
7708 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7709 node = llist_del_all(&ctx->rsrc_put_llist);
7712 struct io_rsrc_node *ref_node;
7713 struct llist_node *next = node->next;
7715 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7716 __io_rsrc_put_work(ref_node);
7721 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7722 unsigned nr_args, u64 __user *tags)
7724 __s32 __user *fds = (__s32 __user *) arg;
7733 if (nr_args > IORING_MAX_FIXED_FILES)
7735 ret = io_rsrc_node_switch_start(ctx);
7738 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7744 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7747 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7748 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7752 /* allow sparse sets */
7755 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7762 if (unlikely(!file))
7766 * Don't allow io_uring instances to be registered. If UNIX
7767 * isn't enabled, then this causes a reference cycle and this
7768 * instance can never get freed. If UNIX is enabled we'll
7769 * handle it just fine, but there's still no point in allowing
7770 * a ring fd as it doesn't support regular read/write anyway.
7772 if (file->f_op == &io_uring_fops) {
7776 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7779 ret = io_sqe_files_scm(ctx);
7781 __io_sqe_files_unregister(ctx);
7785 io_rsrc_node_switch(ctx, NULL);
7788 for (i = 0; i < ctx->nr_user_files; i++) {
7789 file = io_file_from_index(ctx, i);
7793 io_free_file_tables(&ctx->file_table);
7794 ctx->nr_user_files = 0;
7796 io_rsrc_data_free(ctx->file_data);
7797 ctx->file_data = NULL;
7801 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7804 #if defined(CONFIG_UNIX)
7805 struct sock *sock = ctx->ring_sock->sk;
7806 struct sk_buff_head *head = &sock->sk_receive_queue;
7807 struct sk_buff *skb;
7810 * See if we can merge this file into an existing skb SCM_RIGHTS
7811 * file set. If there's no room, fall back to allocating a new skb
7812 * and filling it in.
7814 spin_lock_irq(&head->lock);
7815 skb = skb_peek(head);
7817 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7819 if (fpl->count < SCM_MAX_FD) {
7820 __skb_unlink(skb, head);
7821 spin_unlock_irq(&head->lock);
7822 fpl->fp[fpl->count] = get_file(file);
7823 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7825 spin_lock_irq(&head->lock);
7826 __skb_queue_head(head, skb);
7831 spin_unlock_irq(&head->lock);
7838 return __io_sqe_files_scm(ctx, 1, index);
7844 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7845 struct io_rsrc_node *node, void *rsrc)
7847 struct io_rsrc_put *prsrc;
7849 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7853 prsrc->tag = *io_get_tag_slot(data, idx);
7855 list_add(&prsrc->list, &node->rsrc_list);
7859 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7860 struct io_uring_rsrc_update2 *up,
7863 u64 __user *tags = u64_to_user_ptr(up->tags);
7864 __s32 __user *fds = u64_to_user_ptr(up->data);
7865 struct io_rsrc_data *data = ctx->file_data;
7866 struct io_fixed_file *file_slot;
7870 bool needs_switch = false;
7872 if (!ctx->file_data)
7874 if (up->offset + nr_args > ctx->nr_user_files)
7877 for (done = 0; done < nr_args; done++) {
7880 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7881 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7885 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7889 if (fd == IORING_REGISTER_FILES_SKIP)
7892 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7893 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7895 if (file_slot->file_ptr) {
7896 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7897 err = io_queue_rsrc_removal(data, up->offset + done,
7898 ctx->rsrc_node, file);
7901 file_slot->file_ptr = 0;
7902 needs_switch = true;
7911 * Don't allow io_uring instances to be registered. If
7912 * UNIX isn't enabled, then this causes a reference
7913 * cycle and this instance can never get freed. If UNIX
7914 * is enabled we'll handle it just fine, but there's
7915 * still no point in allowing a ring fd as it doesn't
7916 * support regular read/write anyway.
7918 if (file->f_op == &io_uring_fops) {
7923 *io_get_tag_slot(data, up->offset + done) = tag;
7924 io_fixed_file_set(file_slot, file);
7925 err = io_sqe_file_register(ctx, file, i);
7927 file_slot->file_ptr = 0;
7935 io_rsrc_node_switch(ctx, data);
7936 return done ? done : err;
7939 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7940 struct task_struct *task)
7942 struct io_wq_hash *hash;
7943 struct io_wq_data data;
7944 unsigned int concurrency;
7946 mutex_lock(&ctx->uring_lock);
7947 hash = ctx->hash_map;
7949 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7951 mutex_unlock(&ctx->uring_lock);
7952 return ERR_PTR(-ENOMEM);
7954 refcount_set(&hash->refs, 1);
7955 init_waitqueue_head(&hash->wait);
7956 ctx->hash_map = hash;
7958 mutex_unlock(&ctx->uring_lock);
7962 data.free_work = io_wq_free_work;
7963 data.do_work = io_wq_submit_work;
7965 /* Do QD, or 4 * CPUS, whatever is smallest */
7966 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7968 return io_wq_create(concurrency, &data);
7971 static int io_uring_alloc_task_context(struct task_struct *task,
7972 struct io_ring_ctx *ctx)
7974 struct io_uring_task *tctx;
7977 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7978 if (unlikely(!tctx))
7981 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7982 if (unlikely(ret)) {
7987 tctx->io_wq = io_init_wq_offload(ctx, task);
7988 if (IS_ERR(tctx->io_wq)) {
7989 ret = PTR_ERR(tctx->io_wq);
7990 percpu_counter_destroy(&tctx->inflight);
7996 init_waitqueue_head(&tctx->wait);
7997 atomic_set(&tctx->in_idle, 0);
7998 atomic_set(&tctx->inflight_tracked, 0);
7999 task->io_uring = tctx;
8000 spin_lock_init(&tctx->task_lock);
8001 INIT_WQ_LIST(&tctx->task_list);
8002 init_task_work(&tctx->task_work, tctx_task_work);
8006 void __io_uring_free(struct task_struct *tsk)
8008 struct io_uring_task *tctx = tsk->io_uring;
8010 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8011 WARN_ON_ONCE(tctx->io_wq);
8012 WARN_ON_ONCE(tctx->cached_refs);
8014 percpu_counter_destroy(&tctx->inflight);
8016 tsk->io_uring = NULL;
8019 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8020 struct io_uring_params *p)
8024 /* Retain compatibility with failing for an invalid attach attempt */
8025 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8026 IORING_SETUP_ATTACH_WQ) {
8029 f = fdget(p->wq_fd);
8032 if (f.file->f_op != &io_uring_fops) {
8038 if (ctx->flags & IORING_SETUP_SQPOLL) {
8039 struct task_struct *tsk;
8040 struct io_sq_data *sqd;
8043 sqd = io_get_sq_data(p, &attached);
8049 ctx->sq_creds = get_current_cred();
8051 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8052 if (!ctx->sq_thread_idle)
8053 ctx->sq_thread_idle = HZ;
8055 io_sq_thread_park(sqd);
8056 list_add(&ctx->sqd_list, &sqd->ctx_list);
8057 io_sqd_update_thread_idle(sqd);
8058 /* don't attach to a dying SQPOLL thread, would be racy */
8059 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8060 io_sq_thread_unpark(sqd);
8067 if (p->flags & IORING_SETUP_SQ_AFF) {
8068 int cpu = p->sq_thread_cpu;
8071 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8078 sqd->task_pid = current->pid;
8079 sqd->task_tgid = current->tgid;
8080 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8087 ret = io_uring_alloc_task_context(tsk, ctx);
8088 wake_up_new_task(tsk);
8091 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8092 /* Can't have SQ_AFF without SQPOLL */
8099 complete(&ctx->sq_data->exited);
8101 io_sq_thread_finish(ctx);
8105 static inline void __io_unaccount_mem(struct user_struct *user,
8106 unsigned long nr_pages)
8108 atomic_long_sub(nr_pages, &user->locked_vm);
8111 static inline int __io_account_mem(struct user_struct *user,
8112 unsigned long nr_pages)
8114 unsigned long page_limit, cur_pages, new_pages;
8116 /* Don't allow more pages than we can safely lock */
8117 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8120 cur_pages = atomic_long_read(&user->locked_vm);
8121 new_pages = cur_pages + nr_pages;
8122 if (new_pages > page_limit)
8124 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8125 new_pages) != cur_pages);
8130 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8133 __io_unaccount_mem(ctx->user, nr_pages);
8135 if (ctx->mm_account)
8136 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8139 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8144 ret = __io_account_mem(ctx->user, nr_pages);
8149 if (ctx->mm_account)
8150 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8155 static void io_mem_free(void *ptr)
8162 page = virt_to_head_page(ptr);
8163 if (put_page_testzero(page))
8164 free_compound_page(page);
8167 static void *io_mem_alloc(size_t size)
8169 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8170 __GFP_NORETRY | __GFP_ACCOUNT;
8172 return (void *) __get_free_pages(gfp_flags, get_order(size));
8175 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8178 struct io_rings *rings;
8179 size_t off, sq_array_size;
8181 off = struct_size(rings, cqes, cq_entries);
8182 if (off == SIZE_MAX)
8186 off = ALIGN(off, SMP_CACHE_BYTES);
8194 sq_array_size = array_size(sizeof(u32), sq_entries);
8195 if (sq_array_size == SIZE_MAX)
8198 if (check_add_overflow(off, sq_array_size, &off))
8204 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8206 struct io_mapped_ubuf *imu = *slot;
8209 if (imu != ctx->dummy_ubuf) {
8210 for (i = 0; i < imu->nr_bvecs; i++)
8211 unpin_user_page(imu->bvec[i].bv_page);
8212 if (imu->acct_pages)
8213 io_unaccount_mem(ctx, imu->acct_pages);
8219 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8221 io_buffer_unmap(ctx, &prsrc->buf);
8225 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8229 for (i = 0; i < ctx->nr_user_bufs; i++)
8230 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8231 kfree(ctx->user_bufs);
8232 io_rsrc_data_free(ctx->buf_data);
8233 ctx->user_bufs = NULL;
8234 ctx->buf_data = NULL;
8235 ctx->nr_user_bufs = 0;
8238 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8245 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8247 __io_sqe_buffers_unregister(ctx);
8251 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8252 void __user *arg, unsigned index)
8254 struct iovec __user *src;
8256 #ifdef CONFIG_COMPAT
8258 struct compat_iovec __user *ciovs;
8259 struct compat_iovec ciov;
8261 ciovs = (struct compat_iovec __user *) arg;
8262 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8265 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8266 dst->iov_len = ciov.iov_len;
8270 src = (struct iovec __user *) arg;
8271 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8277 * Not super efficient, but this is just a registration time. And we do cache
8278 * the last compound head, so generally we'll only do a full search if we don't
8281 * We check if the given compound head page has already been accounted, to
8282 * avoid double accounting it. This allows us to account the full size of the
8283 * page, not just the constituent pages of a huge page.
8285 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8286 int nr_pages, struct page *hpage)
8290 /* check current page array */
8291 for (i = 0; i < nr_pages; i++) {
8292 if (!PageCompound(pages[i]))
8294 if (compound_head(pages[i]) == hpage)
8298 /* check previously registered pages */
8299 for (i = 0; i < ctx->nr_user_bufs; i++) {
8300 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8302 for (j = 0; j < imu->nr_bvecs; j++) {
8303 if (!PageCompound(imu->bvec[j].bv_page))
8305 if (compound_head(imu->bvec[j].bv_page) == hpage)
8313 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8314 int nr_pages, struct io_mapped_ubuf *imu,
8315 struct page **last_hpage)
8319 imu->acct_pages = 0;
8320 for (i = 0; i < nr_pages; i++) {
8321 if (!PageCompound(pages[i])) {
8326 hpage = compound_head(pages[i]);
8327 if (hpage == *last_hpage)
8329 *last_hpage = hpage;
8330 if (headpage_already_acct(ctx, pages, i, hpage))
8332 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8336 if (!imu->acct_pages)
8339 ret = io_account_mem(ctx, imu->acct_pages);
8341 imu->acct_pages = 0;
8345 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8346 struct io_mapped_ubuf **pimu,
8347 struct page **last_hpage)
8349 struct io_mapped_ubuf *imu = NULL;
8350 struct vm_area_struct **vmas = NULL;
8351 struct page **pages = NULL;
8352 unsigned long off, start, end, ubuf;
8354 int ret, pret, nr_pages, i;
8356 if (!iov->iov_base) {
8357 *pimu = ctx->dummy_ubuf;
8361 ubuf = (unsigned long) iov->iov_base;
8362 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8363 start = ubuf >> PAGE_SHIFT;
8364 nr_pages = end - start;
8369 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8373 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8378 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8383 mmap_read_lock(current->mm);
8384 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8386 if (pret == nr_pages) {
8387 /* don't support file backed memory */
8388 for (i = 0; i < nr_pages; i++) {
8389 struct vm_area_struct *vma = vmas[i];
8391 if (vma_is_shmem(vma))
8394 !is_file_hugepages(vma->vm_file)) {
8400 ret = pret < 0 ? pret : -EFAULT;
8402 mmap_read_unlock(current->mm);
8405 * if we did partial map, or found file backed vmas,
8406 * release any pages we did get
8409 unpin_user_pages(pages, pret);
8413 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8415 unpin_user_pages(pages, pret);
8419 off = ubuf & ~PAGE_MASK;
8420 size = iov->iov_len;
8421 for (i = 0; i < nr_pages; i++) {
8424 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8425 imu->bvec[i].bv_page = pages[i];
8426 imu->bvec[i].bv_len = vec_len;
8427 imu->bvec[i].bv_offset = off;
8431 /* store original address for later verification */
8433 imu->ubuf_end = ubuf + iov->iov_len;
8434 imu->nr_bvecs = nr_pages;
8445 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8447 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8448 return ctx->user_bufs ? 0 : -ENOMEM;
8451 static int io_buffer_validate(struct iovec *iov)
8453 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8456 * Don't impose further limits on the size and buffer
8457 * constraints here, we'll -EINVAL later when IO is
8458 * submitted if they are wrong.
8461 return iov->iov_len ? -EFAULT : 0;
8465 /* arbitrary limit, but we need something */
8466 if (iov->iov_len > SZ_1G)
8469 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8475 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8476 unsigned int nr_args, u64 __user *tags)
8478 struct page *last_hpage = NULL;
8479 struct io_rsrc_data *data;
8485 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8487 ret = io_rsrc_node_switch_start(ctx);
8490 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8493 ret = io_buffers_map_alloc(ctx, nr_args);
8495 io_rsrc_data_free(data);
8499 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8500 ret = io_copy_iov(ctx, &iov, arg, i);
8503 ret = io_buffer_validate(&iov);
8506 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8511 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8517 WARN_ON_ONCE(ctx->buf_data);
8519 ctx->buf_data = data;
8521 __io_sqe_buffers_unregister(ctx);
8523 io_rsrc_node_switch(ctx, NULL);
8527 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8528 struct io_uring_rsrc_update2 *up,
8529 unsigned int nr_args)
8531 u64 __user *tags = u64_to_user_ptr(up->tags);
8532 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8533 struct page *last_hpage = NULL;
8534 bool needs_switch = false;
8540 if (up->offset + nr_args > ctx->nr_user_bufs)
8543 for (done = 0; done < nr_args; done++) {
8544 struct io_mapped_ubuf *imu;
8545 int offset = up->offset + done;
8548 err = io_copy_iov(ctx, &iov, iovs, done);
8551 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8555 err = io_buffer_validate(&iov);
8558 if (!iov.iov_base && tag) {
8562 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8566 i = array_index_nospec(offset, ctx->nr_user_bufs);
8567 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8568 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8569 ctx->rsrc_node, ctx->user_bufs[i]);
8570 if (unlikely(err)) {
8571 io_buffer_unmap(ctx, &imu);
8574 ctx->user_bufs[i] = NULL;
8575 needs_switch = true;
8578 ctx->user_bufs[i] = imu;
8579 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8583 io_rsrc_node_switch(ctx, ctx->buf_data);
8584 return done ? done : err;
8587 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8589 __s32 __user *fds = arg;
8595 if (copy_from_user(&fd, fds, sizeof(*fds)))
8598 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8599 if (IS_ERR(ctx->cq_ev_fd)) {
8600 int ret = PTR_ERR(ctx->cq_ev_fd);
8602 ctx->cq_ev_fd = NULL;
8609 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8611 if (ctx->cq_ev_fd) {
8612 eventfd_ctx_put(ctx->cq_ev_fd);
8613 ctx->cq_ev_fd = NULL;
8620 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8622 struct io_buffer *buf;
8623 unsigned long index;
8625 xa_for_each(&ctx->io_buffers, index, buf)
8626 __io_remove_buffers(ctx, buf, index, -1U);
8629 static void io_req_cache_free(struct list_head *list)
8631 struct io_kiocb *req, *nxt;
8633 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8634 list_del(&req->inflight_entry);
8635 kmem_cache_free(req_cachep, req);
8639 static void io_req_caches_free(struct io_ring_ctx *ctx)
8641 struct io_submit_state *state = &ctx->submit_state;
8643 mutex_lock(&ctx->uring_lock);
8645 if (state->free_reqs) {
8646 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8647 state->free_reqs = 0;
8650 io_flush_cached_locked_reqs(ctx, state);
8651 io_req_cache_free(&state->free_list);
8652 mutex_unlock(&ctx->uring_lock);
8655 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8657 if (data && !atomic_dec_and_test(&data->refs))
8658 wait_for_completion(&data->done);
8661 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8663 io_sq_thread_finish(ctx);
8665 if (ctx->mm_account) {
8666 mmdrop(ctx->mm_account);
8667 ctx->mm_account = NULL;
8670 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8671 io_wait_rsrc_data(ctx->buf_data);
8672 io_wait_rsrc_data(ctx->file_data);
8674 mutex_lock(&ctx->uring_lock);
8676 __io_sqe_buffers_unregister(ctx);
8678 __io_sqe_files_unregister(ctx);
8680 __io_cqring_overflow_flush(ctx, true);
8681 mutex_unlock(&ctx->uring_lock);
8682 io_eventfd_unregister(ctx);
8683 io_destroy_buffers(ctx);
8685 put_cred(ctx->sq_creds);
8687 /* there are no registered resources left, nobody uses it */
8689 io_rsrc_node_destroy(ctx->rsrc_node);
8690 if (ctx->rsrc_backup_node)
8691 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8692 flush_delayed_work(&ctx->rsrc_put_work);
8694 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8695 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8697 #if defined(CONFIG_UNIX)
8698 if (ctx->ring_sock) {
8699 ctx->ring_sock->file = NULL; /* so that iput() is called */
8700 sock_release(ctx->ring_sock);
8704 io_mem_free(ctx->rings);
8705 io_mem_free(ctx->sq_sqes);
8707 percpu_ref_exit(&ctx->refs);
8708 free_uid(ctx->user);
8709 io_req_caches_free(ctx);
8711 io_wq_put_hash(ctx->hash_map);
8712 kfree(ctx->cancel_hash);
8713 kfree(ctx->dummy_ubuf);
8717 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8719 struct io_ring_ctx *ctx = file->private_data;
8722 poll_wait(file, &ctx->poll_wait, wait);
8724 * synchronizes with barrier from wq_has_sleeper call in
8728 if (!io_sqring_full(ctx))
8729 mask |= EPOLLOUT | EPOLLWRNORM;
8732 * Don't flush cqring overflow list here, just do a simple check.
8733 * Otherwise there could possible be ABBA deadlock:
8736 * lock(&ctx->uring_lock);
8738 * lock(&ctx->uring_lock);
8741 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8742 * pushs them to do the flush.
8744 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8745 mask |= EPOLLIN | EPOLLRDNORM;
8750 static int io_uring_fasync(int fd, struct file *file, int on)
8752 struct io_ring_ctx *ctx = file->private_data;
8754 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8757 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8759 const struct cred *creds;
8761 creds = xa_erase(&ctx->personalities, id);
8770 struct io_tctx_exit {
8771 struct callback_head task_work;
8772 struct completion completion;
8773 struct io_ring_ctx *ctx;
8776 static void io_tctx_exit_cb(struct callback_head *cb)
8778 struct io_uring_task *tctx = current->io_uring;
8779 struct io_tctx_exit *work;
8781 work = container_of(cb, struct io_tctx_exit, task_work);
8783 * When @in_idle, we're in cancellation and it's racy to remove the
8784 * node. It'll be removed by the end of cancellation, just ignore it.
8786 if (!atomic_read(&tctx->in_idle))
8787 io_uring_del_tctx_node((unsigned long)work->ctx);
8788 complete(&work->completion);
8791 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8793 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8795 return req->ctx == data;
8798 static void io_ring_exit_work(struct work_struct *work)
8800 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8801 unsigned long timeout = jiffies + HZ * 60 * 5;
8802 unsigned long interval = HZ / 20;
8803 struct io_tctx_exit exit;
8804 struct io_tctx_node *node;
8808 * If we're doing polled IO and end up having requests being
8809 * submitted async (out-of-line), then completions can come in while
8810 * we're waiting for refs to drop. We need to reap these manually,
8811 * as nobody else will be looking for them.
8814 io_uring_try_cancel_requests(ctx, NULL, true);
8816 struct io_sq_data *sqd = ctx->sq_data;
8817 struct task_struct *tsk;
8819 io_sq_thread_park(sqd);
8821 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8822 io_wq_cancel_cb(tsk->io_uring->io_wq,
8823 io_cancel_ctx_cb, ctx, true);
8824 io_sq_thread_unpark(sqd);
8827 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8828 /* there is little hope left, don't run it too often */
8831 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8833 init_completion(&exit.completion);
8834 init_task_work(&exit.task_work, io_tctx_exit_cb);
8837 * Some may use context even when all refs and requests have been put,
8838 * and they are free to do so while still holding uring_lock or
8839 * completion_lock, see io_req_task_submit(). Apart from other work,
8840 * this lock/unlock section also waits them to finish.
8842 mutex_lock(&ctx->uring_lock);
8843 while (!list_empty(&ctx->tctx_list)) {
8844 WARN_ON_ONCE(time_after(jiffies, timeout));
8846 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8848 /* don't spin on a single task if cancellation failed */
8849 list_rotate_left(&ctx->tctx_list);
8850 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8851 if (WARN_ON_ONCE(ret))
8853 wake_up_process(node->task);
8855 mutex_unlock(&ctx->uring_lock);
8856 wait_for_completion(&exit.completion);
8857 mutex_lock(&ctx->uring_lock);
8859 mutex_unlock(&ctx->uring_lock);
8860 spin_lock(&ctx->completion_lock);
8861 spin_unlock(&ctx->completion_lock);
8863 io_ring_ctx_free(ctx);
8866 /* Returns true if we found and killed one or more timeouts */
8867 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8870 struct io_kiocb *req, *tmp;
8873 spin_lock(&ctx->completion_lock);
8874 spin_lock_irq(&ctx->timeout_lock);
8875 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8876 if (io_match_task(req, tsk, cancel_all)) {
8877 io_kill_timeout(req, -ECANCELED);
8881 spin_unlock_irq(&ctx->timeout_lock);
8883 io_commit_cqring(ctx);
8884 spin_unlock(&ctx->completion_lock);
8886 io_cqring_ev_posted(ctx);
8887 return canceled != 0;
8890 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8892 unsigned long index;
8893 struct creds *creds;
8895 mutex_lock(&ctx->uring_lock);
8896 percpu_ref_kill(&ctx->refs);
8898 __io_cqring_overflow_flush(ctx, true);
8899 xa_for_each(&ctx->personalities, index, creds)
8900 io_unregister_personality(ctx, index);
8901 mutex_unlock(&ctx->uring_lock);
8903 io_kill_timeouts(ctx, NULL, true);
8904 io_poll_remove_all(ctx, NULL, true);
8906 /* if we failed setting up the ctx, we might not have any rings */
8907 io_iopoll_try_reap_events(ctx);
8909 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8911 * Use system_unbound_wq to avoid spawning tons of event kworkers
8912 * if we're exiting a ton of rings at the same time. It just adds
8913 * noise and overhead, there's no discernable change in runtime
8914 * over using system_wq.
8916 queue_work(system_unbound_wq, &ctx->exit_work);
8919 static int io_uring_release(struct inode *inode, struct file *file)
8921 struct io_ring_ctx *ctx = file->private_data;
8923 file->private_data = NULL;
8924 io_ring_ctx_wait_and_kill(ctx);
8928 struct io_task_cancel {
8929 struct task_struct *task;
8933 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8935 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8936 struct io_task_cancel *cancel = data;
8939 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8940 struct io_ring_ctx *ctx = req->ctx;
8942 /* protect against races with linked timeouts */
8943 spin_lock(&ctx->completion_lock);
8944 ret = io_match_task(req, cancel->task, cancel->all);
8945 spin_unlock(&ctx->completion_lock);
8947 ret = io_match_task(req, cancel->task, cancel->all);
8952 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8953 struct task_struct *task, bool cancel_all)
8955 struct io_defer_entry *de;
8958 spin_lock(&ctx->completion_lock);
8959 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8960 if (io_match_task(de->req, task, cancel_all)) {
8961 list_cut_position(&list, &ctx->defer_list, &de->list);
8965 spin_unlock(&ctx->completion_lock);
8966 if (list_empty(&list))
8969 while (!list_empty(&list)) {
8970 de = list_first_entry(&list, struct io_defer_entry, list);
8971 list_del_init(&de->list);
8972 io_req_complete_failed(de->req, -ECANCELED);
8978 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8980 struct io_tctx_node *node;
8981 enum io_wq_cancel cret;
8984 mutex_lock(&ctx->uring_lock);
8985 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8986 struct io_uring_task *tctx = node->task->io_uring;
8989 * io_wq will stay alive while we hold uring_lock, because it's
8990 * killed after ctx nodes, which requires to take the lock.
8992 if (!tctx || !tctx->io_wq)
8994 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8995 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8997 mutex_unlock(&ctx->uring_lock);
9002 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9003 struct task_struct *task,
9006 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9007 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9010 enum io_wq_cancel cret;
9014 ret |= io_uring_try_cancel_iowq(ctx);
9015 } else if (tctx && tctx->io_wq) {
9017 * Cancels requests of all rings, not only @ctx, but
9018 * it's fine as the task is in exit/exec.
9020 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9022 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9025 /* SQPOLL thread does its own polling */
9026 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9027 (ctx->sq_data && ctx->sq_data->thread == current)) {
9028 while (!list_empty_careful(&ctx->iopoll_list)) {
9029 io_iopoll_try_reap_events(ctx);
9034 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9035 ret |= io_poll_remove_all(ctx, task, cancel_all);
9036 ret |= io_kill_timeouts(ctx, task, cancel_all);
9038 ret |= io_run_task_work();
9045 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9047 struct io_uring_task *tctx = current->io_uring;
9048 struct io_tctx_node *node;
9051 if (unlikely(!tctx)) {
9052 ret = io_uring_alloc_task_context(current, ctx);
9055 tctx = current->io_uring;
9057 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9058 node = kmalloc(sizeof(*node), GFP_KERNEL);
9062 node->task = current;
9064 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9071 mutex_lock(&ctx->uring_lock);
9072 list_add(&node->ctx_node, &ctx->tctx_list);
9073 mutex_unlock(&ctx->uring_lock);
9080 * Note that this task has used io_uring. We use it for cancelation purposes.
9082 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9084 struct io_uring_task *tctx = current->io_uring;
9086 if (likely(tctx && tctx->last == ctx))
9088 return __io_uring_add_tctx_node(ctx);
9092 * Remove this io_uring_file -> task mapping.
9094 static void io_uring_del_tctx_node(unsigned long index)
9096 struct io_uring_task *tctx = current->io_uring;
9097 struct io_tctx_node *node;
9101 node = xa_erase(&tctx->xa, index);
9105 WARN_ON_ONCE(current != node->task);
9106 WARN_ON_ONCE(list_empty(&node->ctx_node));
9108 mutex_lock(&node->ctx->uring_lock);
9109 list_del(&node->ctx_node);
9110 mutex_unlock(&node->ctx->uring_lock);
9112 if (tctx->last == node->ctx)
9117 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9119 struct io_wq *wq = tctx->io_wq;
9120 struct io_tctx_node *node;
9121 unsigned long index;
9123 xa_for_each(&tctx->xa, index, node)
9124 io_uring_del_tctx_node(index);
9127 * Must be after io_uring_del_task_file() (removes nodes under
9128 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9131 io_wq_put_and_exit(wq);
9135 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9138 return atomic_read(&tctx->inflight_tracked);
9139 return percpu_counter_sum(&tctx->inflight);
9142 static void io_uring_drop_tctx_refs(struct task_struct *task)
9144 struct io_uring_task *tctx = task->io_uring;
9145 unsigned int refs = tctx->cached_refs;
9148 tctx->cached_refs = 0;
9149 percpu_counter_sub(&tctx->inflight, refs);
9150 put_task_struct_many(task, refs);
9155 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9156 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9158 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9160 struct io_uring_task *tctx = current->io_uring;
9161 struct io_ring_ctx *ctx;
9165 WARN_ON_ONCE(sqd && sqd->thread != current);
9167 if (!current->io_uring)
9170 io_wq_exit_start(tctx->io_wq);
9172 atomic_inc(&tctx->in_idle);
9174 io_uring_drop_tctx_refs(current);
9175 /* read completions before cancelations */
9176 inflight = tctx_inflight(tctx, !cancel_all);
9181 struct io_tctx_node *node;
9182 unsigned long index;
9184 xa_for_each(&tctx->xa, index, node) {
9185 /* sqpoll task will cancel all its requests */
9186 if (node->ctx->sq_data)
9188 io_uring_try_cancel_requests(node->ctx, current,
9192 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9193 io_uring_try_cancel_requests(ctx, current,
9197 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9198 io_uring_drop_tctx_refs(current);
9200 * If we've seen completions, retry without waiting. This
9201 * avoids a race where a completion comes in before we did
9202 * prepare_to_wait().
9204 if (inflight == tctx_inflight(tctx, !cancel_all))
9206 finish_wait(&tctx->wait, &wait);
9208 atomic_dec(&tctx->in_idle);
9210 io_uring_clean_tctx(tctx);
9212 /* for exec all current's requests should be gone, kill tctx */
9213 __io_uring_free(current);
9217 void __io_uring_cancel(bool cancel_all)
9219 io_uring_cancel_generic(cancel_all, NULL);
9222 static void *io_uring_validate_mmap_request(struct file *file,
9223 loff_t pgoff, size_t sz)
9225 struct io_ring_ctx *ctx = file->private_data;
9226 loff_t offset = pgoff << PAGE_SHIFT;
9231 case IORING_OFF_SQ_RING:
9232 case IORING_OFF_CQ_RING:
9235 case IORING_OFF_SQES:
9239 return ERR_PTR(-EINVAL);
9242 page = virt_to_head_page(ptr);
9243 if (sz > page_size(page))
9244 return ERR_PTR(-EINVAL);
9251 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9253 size_t sz = vma->vm_end - vma->vm_start;
9257 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9259 return PTR_ERR(ptr);
9261 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9262 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9265 #else /* !CONFIG_MMU */
9267 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9269 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9272 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9274 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9277 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9278 unsigned long addr, unsigned long len,
9279 unsigned long pgoff, unsigned long flags)
9283 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9285 return PTR_ERR(ptr);
9287 return (unsigned long) ptr;
9290 #endif /* !CONFIG_MMU */
9292 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9297 if (!io_sqring_full(ctx))
9299 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9301 if (!io_sqring_full(ctx))
9304 } while (!signal_pending(current));
9306 finish_wait(&ctx->sqo_sq_wait, &wait);
9310 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9311 struct __kernel_timespec __user **ts,
9312 const sigset_t __user **sig)
9314 struct io_uring_getevents_arg arg;
9317 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9318 * is just a pointer to the sigset_t.
9320 if (!(flags & IORING_ENTER_EXT_ARG)) {
9321 *sig = (const sigset_t __user *) argp;
9327 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9328 * timespec and sigset_t pointers if good.
9330 if (*argsz != sizeof(arg))
9332 if (copy_from_user(&arg, argp, sizeof(arg)))
9334 *sig = u64_to_user_ptr(arg.sigmask);
9335 *argsz = arg.sigmask_sz;
9336 *ts = u64_to_user_ptr(arg.ts);
9340 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9341 u32, min_complete, u32, flags, const void __user *, argp,
9344 struct io_ring_ctx *ctx;
9351 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9352 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9356 if (unlikely(!f.file))
9360 if (unlikely(f.file->f_op != &io_uring_fops))
9364 ctx = f.file->private_data;
9365 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9369 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9373 * For SQ polling, the thread will do all submissions and completions.
9374 * Just return the requested submit count, and wake the thread if
9378 if (ctx->flags & IORING_SETUP_SQPOLL) {
9379 io_cqring_overflow_flush(ctx);
9381 if (unlikely(ctx->sq_data->thread == NULL)) {
9385 if (flags & IORING_ENTER_SQ_WAKEUP)
9386 wake_up(&ctx->sq_data->wait);
9387 if (flags & IORING_ENTER_SQ_WAIT) {
9388 ret = io_sqpoll_wait_sq(ctx);
9392 submitted = to_submit;
9393 } else if (to_submit) {
9394 ret = io_uring_add_tctx_node(ctx);
9397 mutex_lock(&ctx->uring_lock);
9398 submitted = io_submit_sqes(ctx, to_submit);
9399 mutex_unlock(&ctx->uring_lock);
9401 if (submitted != to_submit)
9404 if (flags & IORING_ENTER_GETEVENTS) {
9405 const sigset_t __user *sig;
9406 struct __kernel_timespec __user *ts;
9408 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9412 min_complete = min(min_complete, ctx->cq_entries);
9415 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9416 * space applications don't need to do io completion events
9417 * polling again, they can rely on io_sq_thread to do polling
9418 * work, which can reduce cpu usage and uring_lock contention.
9420 if (ctx->flags & IORING_SETUP_IOPOLL &&
9421 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9422 ret = io_iopoll_check(ctx, min_complete);
9424 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9429 percpu_ref_put(&ctx->refs);
9432 return submitted ? submitted : ret;
9435 #ifdef CONFIG_PROC_FS
9436 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9437 const struct cred *cred)
9439 struct user_namespace *uns = seq_user_ns(m);
9440 struct group_info *gi;
9445 seq_printf(m, "%5d\n", id);
9446 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9447 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9448 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9449 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9450 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9451 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9452 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9453 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9454 seq_puts(m, "\n\tGroups:\t");
9455 gi = cred->group_info;
9456 for (g = 0; g < gi->ngroups; g++) {
9457 seq_put_decimal_ull(m, g ? " " : "",
9458 from_kgid_munged(uns, gi->gid[g]));
9460 seq_puts(m, "\n\tCapEff:\t");
9461 cap = cred->cap_effective;
9462 CAP_FOR_EACH_U32(__capi)
9463 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9468 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9470 struct io_sq_data *sq = NULL;
9475 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9476 * since fdinfo case grabs it in the opposite direction of normal use
9477 * cases. If we fail to get the lock, we just don't iterate any
9478 * structures that could be going away outside the io_uring mutex.
9480 has_lock = mutex_trylock(&ctx->uring_lock);
9482 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9488 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9489 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9490 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9491 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9492 struct file *f = io_file_from_index(ctx, i);
9495 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9497 seq_printf(m, "%5u: <none>\n", i);
9499 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9500 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9501 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9502 unsigned int len = buf->ubuf_end - buf->ubuf;
9504 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9506 if (has_lock && !xa_empty(&ctx->personalities)) {
9507 unsigned long index;
9508 const struct cred *cred;
9510 seq_printf(m, "Personalities:\n");
9511 xa_for_each(&ctx->personalities, index, cred)
9512 io_uring_show_cred(m, index, cred);
9514 seq_printf(m, "PollList:\n");
9515 spin_lock(&ctx->completion_lock);
9516 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9517 struct hlist_head *list = &ctx->cancel_hash[i];
9518 struct io_kiocb *req;
9520 hlist_for_each_entry(req, list, hash_node)
9521 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9522 req->task->task_works != NULL);
9524 spin_unlock(&ctx->completion_lock);
9526 mutex_unlock(&ctx->uring_lock);
9529 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9531 struct io_ring_ctx *ctx = f->private_data;
9533 if (percpu_ref_tryget(&ctx->refs)) {
9534 __io_uring_show_fdinfo(ctx, m);
9535 percpu_ref_put(&ctx->refs);
9540 static const struct file_operations io_uring_fops = {
9541 .release = io_uring_release,
9542 .mmap = io_uring_mmap,
9544 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9545 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9547 .poll = io_uring_poll,
9548 .fasync = io_uring_fasync,
9549 #ifdef CONFIG_PROC_FS
9550 .show_fdinfo = io_uring_show_fdinfo,
9554 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9555 struct io_uring_params *p)
9557 struct io_rings *rings;
9558 size_t size, sq_array_offset;
9560 /* make sure these are sane, as we already accounted them */
9561 ctx->sq_entries = p->sq_entries;
9562 ctx->cq_entries = p->cq_entries;
9564 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9565 if (size == SIZE_MAX)
9568 rings = io_mem_alloc(size);
9573 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9574 rings->sq_ring_mask = p->sq_entries - 1;
9575 rings->cq_ring_mask = p->cq_entries - 1;
9576 rings->sq_ring_entries = p->sq_entries;
9577 rings->cq_ring_entries = p->cq_entries;
9579 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9580 if (size == SIZE_MAX) {
9581 io_mem_free(ctx->rings);
9586 ctx->sq_sqes = io_mem_alloc(size);
9587 if (!ctx->sq_sqes) {
9588 io_mem_free(ctx->rings);
9596 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9600 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9604 ret = io_uring_add_tctx_node(ctx);
9609 fd_install(fd, file);
9614 * Allocate an anonymous fd, this is what constitutes the application
9615 * visible backing of an io_uring instance. The application mmaps this
9616 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9617 * we have to tie this fd to a socket for file garbage collection purposes.
9619 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9622 #if defined(CONFIG_UNIX)
9625 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9628 return ERR_PTR(ret);
9631 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9632 O_RDWR | O_CLOEXEC);
9633 #if defined(CONFIG_UNIX)
9635 sock_release(ctx->ring_sock);
9636 ctx->ring_sock = NULL;
9638 ctx->ring_sock->file = file;
9644 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9645 struct io_uring_params __user *params)
9647 struct io_ring_ctx *ctx;
9653 if (entries > IORING_MAX_ENTRIES) {
9654 if (!(p->flags & IORING_SETUP_CLAMP))
9656 entries = IORING_MAX_ENTRIES;
9660 * Use twice as many entries for the CQ ring. It's possible for the
9661 * application to drive a higher depth than the size of the SQ ring,
9662 * since the sqes are only used at submission time. This allows for
9663 * some flexibility in overcommitting a bit. If the application has
9664 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9665 * of CQ ring entries manually.
9667 p->sq_entries = roundup_pow_of_two(entries);
9668 if (p->flags & IORING_SETUP_CQSIZE) {
9670 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9671 * to a power-of-two, if it isn't already. We do NOT impose
9672 * any cq vs sq ring sizing.
9676 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9677 if (!(p->flags & IORING_SETUP_CLAMP))
9679 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9681 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9682 if (p->cq_entries < p->sq_entries)
9685 p->cq_entries = 2 * p->sq_entries;
9688 ctx = io_ring_ctx_alloc(p);
9691 ctx->compat = in_compat_syscall();
9692 if (!capable(CAP_IPC_LOCK))
9693 ctx->user = get_uid(current_user());
9696 * This is just grabbed for accounting purposes. When a process exits,
9697 * the mm is exited and dropped before the files, hence we need to hang
9698 * on to this mm purely for the purposes of being able to unaccount
9699 * memory (locked/pinned vm). It's not used for anything else.
9701 mmgrab(current->mm);
9702 ctx->mm_account = current->mm;
9704 ret = io_allocate_scq_urings(ctx, p);
9708 ret = io_sq_offload_create(ctx, p);
9711 /* always set a rsrc node */
9712 ret = io_rsrc_node_switch_start(ctx);
9715 io_rsrc_node_switch(ctx, NULL);
9717 memset(&p->sq_off, 0, sizeof(p->sq_off));
9718 p->sq_off.head = offsetof(struct io_rings, sq.head);
9719 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9720 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9721 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9722 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9723 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9724 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9726 memset(&p->cq_off, 0, sizeof(p->cq_off));
9727 p->cq_off.head = offsetof(struct io_rings, cq.head);
9728 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9729 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9730 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9731 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9732 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9733 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9735 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9736 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9737 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9738 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9739 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9740 IORING_FEAT_RSRC_TAGS;
9742 if (copy_to_user(params, p, sizeof(*p))) {
9747 file = io_uring_get_file(ctx);
9749 ret = PTR_ERR(file);
9754 * Install ring fd as the very last thing, so we don't risk someone
9755 * having closed it before we finish setup
9757 ret = io_uring_install_fd(ctx, file);
9759 /* fput will clean it up */
9764 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9767 io_ring_ctx_wait_and_kill(ctx);
9772 * Sets up an aio uring context, and returns the fd. Applications asks for a
9773 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9774 * params structure passed in.
9776 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9778 struct io_uring_params p;
9781 if (copy_from_user(&p, params, sizeof(p)))
9783 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9788 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9789 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9790 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9791 IORING_SETUP_R_DISABLED))
9794 return io_uring_create(entries, &p, params);
9797 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9798 struct io_uring_params __user *, params)
9800 return io_uring_setup(entries, params);
9803 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9805 struct io_uring_probe *p;
9809 size = struct_size(p, ops, nr_args);
9810 if (size == SIZE_MAX)
9812 p = kzalloc(size, GFP_KERNEL);
9817 if (copy_from_user(p, arg, size))
9820 if (memchr_inv(p, 0, size))
9823 p->last_op = IORING_OP_LAST - 1;
9824 if (nr_args > IORING_OP_LAST)
9825 nr_args = IORING_OP_LAST;
9827 for (i = 0; i < nr_args; i++) {
9829 if (!io_op_defs[i].not_supported)
9830 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9835 if (copy_to_user(arg, p, size))
9842 static int io_register_personality(struct io_ring_ctx *ctx)
9844 const struct cred *creds;
9848 creds = get_current_cred();
9850 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9851 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9859 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9860 unsigned int nr_args)
9862 struct io_uring_restriction *res;
9866 /* Restrictions allowed only if rings started disabled */
9867 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9870 /* We allow only a single restrictions registration */
9871 if (ctx->restrictions.registered)
9874 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9877 size = array_size(nr_args, sizeof(*res));
9878 if (size == SIZE_MAX)
9881 res = memdup_user(arg, size);
9883 return PTR_ERR(res);
9887 for (i = 0; i < nr_args; i++) {
9888 switch (res[i].opcode) {
9889 case IORING_RESTRICTION_REGISTER_OP:
9890 if (res[i].register_op >= IORING_REGISTER_LAST) {
9895 __set_bit(res[i].register_op,
9896 ctx->restrictions.register_op);
9898 case IORING_RESTRICTION_SQE_OP:
9899 if (res[i].sqe_op >= IORING_OP_LAST) {
9904 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9906 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9907 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9909 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9910 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9919 /* Reset all restrictions if an error happened */
9921 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9923 ctx->restrictions.registered = true;
9929 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9931 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9934 if (ctx->restrictions.registered)
9935 ctx->restricted = 1;
9937 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9938 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9939 wake_up(&ctx->sq_data->wait);
9943 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9944 struct io_uring_rsrc_update2 *up,
9952 if (check_add_overflow(up->offset, nr_args, &tmp))
9954 err = io_rsrc_node_switch_start(ctx);
9959 case IORING_RSRC_FILE:
9960 return __io_sqe_files_update(ctx, up, nr_args);
9961 case IORING_RSRC_BUFFER:
9962 return __io_sqe_buffers_update(ctx, up, nr_args);
9967 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9970 struct io_uring_rsrc_update2 up;
9974 memset(&up, 0, sizeof(up));
9975 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9977 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9980 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9981 unsigned size, unsigned type)
9983 struct io_uring_rsrc_update2 up;
9985 if (size != sizeof(up))
9987 if (copy_from_user(&up, arg, sizeof(up)))
9989 if (!up.nr || up.resv)
9991 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9994 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9995 unsigned int size, unsigned int type)
9997 struct io_uring_rsrc_register rr;
9999 /* keep it extendible */
10000 if (size != sizeof(rr))
10003 memset(&rr, 0, sizeof(rr));
10004 if (copy_from_user(&rr, arg, size))
10006 if (!rr.nr || rr.resv || rr.resv2)
10010 case IORING_RSRC_FILE:
10011 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10012 rr.nr, u64_to_user_ptr(rr.tags));
10013 case IORING_RSRC_BUFFER:
10014 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10015 rr.nr, u64_to_user_ptr(rr.tags));
10020 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10023 struct io_uring_task *tctx = current->io_uring;
10024 cpumask_var_t new_mask;
10027 if (!tctx || !tctx->io_wq)
10030 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10033 cpumask_clear(new_mask);
10034 if (len > cpumask_size())
10035 len = cpumask_size();
10037 if (copy_from_user(new_mask, arg, len)) {
10038 free_cpumask_var(new_mask);
10042 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10043 free_cpumask_var(new_mask);
10047 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10049 struct io_uring_task *tctx = current->io_uring;
10051 if (!tctx || !tctx->io_wq)
10054 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10057 static bool io_register_op_must_quiesce(int op)
10060 case IORING_REGISTER_BUFFERS:
10061 case IORING_UNREGISTER_BUFFERS:
10062 case IORING_REGISTER_FILES:
10063 case IORING_UNREGISTER_FILES:
10064 case IORING_REGISTER_FILES_UPDATE:
10065 case IORING_REGISTER_PROBE:
10066 case IORING_REGISTER_PERSONALITY:
10067 case IORING_UNREGISTER_PERSONALITY:
10068 case IORING_REGISTER_FILES2:
10069 case IORING_REGISTER_FILES_UPDATE2:
10070 case IORING_REGISTER_BUFFERS2:
10071 case IORING_REGISTER_BUFFERS_UPDATE:
10072 case IORING_REGISTER_IOWQ_AFF:
10073 case IORING_UNREGISTER_IOWQ_AFF:
10080 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10084 percpu_ref_kill(&ctx->refs);
10087 * Drop uring mutex before waiting for references to exit. If another
10088 * thread is currently inside io_uring_enter() it might need to grab the
10089 * uring_lock to make progress. If we hold it here across the drain
10090 * wait, then we can deadlock. It's safe to drop the mutex here, since
10091 * no new references will come in after we've killed the percpu ref.
10093 mutex_unlock(&ctx->uring_lock);
10095 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10098 ret = io_run_task_work_sig();
10099 } while (ret >= 0);
10100 mutex_lock(&ctx->uring_lock);
10103 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10107 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10108 void __user *arg, unsigned nr_args)
10109 __releases(ctx->uring_lock)
10110 __acquires(ctx->uring_lock)
10115 * We're inside the ring mutex, if the ref is already dying, then
10116 * someone else killed the ctx or is already going through
10117 * io_uring_register().
10119 if (percpu_ref_is_dying(&ctx->refs))
10122 if (ctx->restricted) {
10123 if (opcode >= IORING_REGISTER_LAST)
10125 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10126 if (!test_bit(opcode, ctx->restrictions.register_op))
10130 if (io_register_op_must_quiesce(opcode)) {
10131 ret = io_ctx_quiesce(ctx);
10137 case IORING_REGISTER_BUFFERS:
10138 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10140 case IORING_UNREGISTER_BUFFERS:
10142 if (arg || nr_args)
10144 ret = io_sqe_buffers_unregister(ctx);
10146 case IORING_REGISTER_FILES:
10147 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10149 case IORING_UNREGISTER_FILES:
10151 if (arg || nr_args)
10153 ret = io_sqe_files_unregister(ctx);
10155 case IORING_REGISTER_FILES_UPDATE:
10156 ret = io_register_files_update(ctx, arg, nr_args);
10158 case IORING_REGISTER_EVENTFD:
10159 case IORING_REGISTER_EVENTFD_ASYNC:
10163 ret = io_eventfd_register(ctx, arg);
10166 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10167 ctx->eventfd_async = 1;
10169 ctx->eventfd_async = 0;
10171 case IORING_UNREGISTER_EVENTFD:
10173 if (arg || nr_args)
10175 ret = io_eventfd_unregister(ctx);
10177 case IORING_REGISTER_PROBE:
10179 if (!arg || nr_args > 256)
10181 ret = io_probe(ctx, arg, nr_args);
10183 case IORING_REGISTER_PERSONALITY:
10185 if (arg || nr_args)
10187 ret = io_register_personality(ctx);
10189 case IORING_UNREGISTER_PERSONALITY:
10193 ret = io_unregister_personality(ctx, nr_args);
10195 case IORING_REGISTER_ENABLE_RINGS:
10197 if (arg || nr_args)
10199 ret = io_register_enable_rings(ctx);
10201 case IORING_REGISTER_RESTRICTIONS:
10202 ret = io_register_restrictions(ctx, arg, nr_args);
10204 case IORING_REGISTER_FILES2:
10205 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10207 case IORING_REGISTER_FILES_UPDATE2:
10208 ret = io_register_rsrc_update(ctx, arg, nr_args,
10211 case IORING_REGISTER_BUFFERS2:
10212 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10214 case IORING_REGISTER_BUFFERS_UPDATE:
10215 ret = io_register_rsrc_update(ctx, arg, nr_args,
10216 IORING_RSRC_BUFFER);
10218 case IORING_REGISTER_IOWQ_AFF:
10220 if (!arg || !nr_args)
10222 ret = io_register_iowq_aff(ctx, arg, nr_args);
10224 case IORING_UNREGISTER_IOWQ_AFF:
10226 if (arg || nr_args)
10228 ret = io_unregister_iowq_aff(ctx);
10235 if (io_register_op_must_quiesce(opcode)) {
10236 /* bring the ctx back to life */
10237 percpu_ref_reinit(&ctx->refs);
10238 reinit_completion(&ctx->ref_comp);
10243 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10244 void __user *, arg, unsigned int, nr_args)
10246 struct io_ring_ctx *ctx;
10255 if (f.file->f_op != &io_uring_fops)
10258 ctx = f.file->private_data;
10260 io_run_task_work();
10262 mutex_lock(&ctx->uring_lock);
10263 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10264 mutex_unlock(&ctx->uring_lock);
10265 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10266 ctx->cq_ev_fd != NULL, ret);
10272 static int __init io_uring_init(void)
10274 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10275 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10276 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10279 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10280 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10281 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10282 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10283 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10284 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10285 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10286 BUILD_BUG_SQE_ELEM(8, __u64, off);
10287 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10288 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10289 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10290 BUILD_BUG_SQE_ELEM(24, __u32, len);
10291 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10292 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10293 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10295 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10296 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10297 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10299 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10300 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10301 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10302 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10303 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10304 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10305 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10306 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10307 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10308 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10309 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10310 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10312 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10313 sizeof(struct io_uring_rsrc_update));
10314 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10315 sizeof(struct io_uring_rsrc_update2));
10316 /* should fit into one byte */
10317 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10319 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10320 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10322 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10326 __initcall(io_uring_init);