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();
1530 * This should only get called when at least one event has been posted.
1531 * Some applications rely on the eventfd notification count only changing
1532 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1533 * 1:1 relationship between how many times this function is called (and
1534 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1536 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1539 * wake_up_all() may seem excessive, but io_wake_function() and
1540 * io_should_wake() handle the termination of the loop and only
1541 * wake as many waiters as we need to.
1543 if (wq_has_sleeper(&ctx->cq_wait))
1544 wake_up_all(&ctx->cq_wait);
1545 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1546 wake_up(&ctx->sq_data->wait);
1547 if (io_should_trigger_evfd(ctx))
1548 eventfd_signal(ctx->cq_ev_fd, 1);
1549 if (waitqueue_active(&ctx->poll_wait)) {
1550 wake_up_interruptible(&ctx->poll_wait);
1551 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1555 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1557 if (ctx->flags & IORING_SETUP_SQPOLL) {
1558 if (wq_has_sleeper(&ctx->cq_wait))
1559 wake_up_all(&ctx->cq_wait);
1561 if (io_should_trigger_evfd(ctx))
1562 eventfd_signal(ctx->cq_ev_fd, 1);
1563 if (waitqueue_active(&ctx->poll_wait)) {
1564 wake_up_interruptible(&ctx->poll_wait);
1565 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1569 /* Returns true if there are no backlogged entries after the flush */
1570 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1572 bool all_flushed, posted;
1574 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1578 spin_lock(&ctx->completion_lock);
1579 while (!list_empty(&ctx->cq_overflow_list)) {
1580 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1581 struct io_overflow_cqe *ocqe;
1585 ocqe = list_first_entry(&ctx->cq_overflow_list,
1586 struct io_overflow_cqe, list);
1588 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1590 io_account_cq_overflow(ctx);
1593 list_del(&ocqe->list);
1597 all_flushed = list_empty(&ctx->cq_overflow_list);
1599 clear_bit(0, &ctx->check_cq_overflow);
1600 WRITE_ONCE(ctx->rings->sq_flags,
1601 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1605 io_commit_cqring(ctx);
1606 spin_unlock(&ctx->completion_lock);
1608 io_cqring_ev_posted(ctx);
1612 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1616 if (test_bit(0, &ctx->check_cq_overflow)) {
1617 /* iopoll syncs against uring_lock, not completion_lock */
1618 if (ctx->flags & IORING_SETUP_IOPOLL)
1619 mutex_lock(&ctx->uring_lock);
1620 ret = __io_cqring_overflow_flush(ctx, false);
1621 if (ctx->flags & IORING_SETUP_IOPOLL)
1622 mutex_unlock(&ctx->uring_lock);
1628 /* must to be called somewhat shortly after putting a request */
1629 static inline void io_put_task(struct task_struct *task, int nr)
1631 struct io_uring_task *tctx = task->io_uring;
1633 if (likely(task == current)) {
1634 tctx->cached_refs += nr;
1636 percpu_counter_sub(&tctx->inflight, nr);
1637 if (unlikely(atomic_read(&tctx->in_idle)))
1638 wake_up(&tctx->wait);
1639 put_task_struct_many(task, nr);
1643 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1644 long res, unsigned int cflags)
1646 struct io_overflow_cqe *ocqe;
1648 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1651 * If we're in ring overflow flush mode, or in task cancel mode,
1652 * or cannot allocate an overflow entry, then we need to drop it
1655 io_account_cq_overflow(ctx);
1658 if (list_empty(&ctx->cq_overflow_list)) {
1659 set_bit(0, &ctx->check_cq_overflow);
1660 WRITE_ONCE(ctx->rings->sq_flags,
1661 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1664 ocqe->cqe.user_data = user_data;
1665 ocqe->cqe.res = res;
1666 ocqe->cqe.flags = cflags;
1667 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1671 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1672 long res, unsigned int cflags)
1674 struct io_uring_cqe *cqe;
1676 trace_io_uring_complete(ctx, user_data, res, cflags);
1679 * If we can't get a cq entry, userspace overflowed the
1680 * submission (by quite a lot). Increment the overflow count in
1683 cqe = io_get_cqe(ctx);
1685 WRITE_ONCE(cqe->user_data, user_data);
1686 WRITE_ONCE(cqe->res, res);
1687 WRITE_ONCE(cqe->flags, cflags);
1690 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1693 /* not as hot to bloat with inlining */
1694 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1695 long res, unsigned int cflags)
1697 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1700 static void io_req_complete_post(struct io_kiocb *req, long res,
1701 unsigned int cflags)
1703 struct io_ring_ctx *ctx = req->ctx;
1705 spin_lock(&ctx->completion_lock);
1706 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1708 * If we're the last reference to this request, add to our locked
1711 if (req_ref_put_and_test(req)) {
1712 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1713 if (req->flags & IO_DISARM_MASK)
1714 io_disarm_next(req);
1716 io_req_task_queue(req->link);
1720 io_dismantle_req(req);
1721 io_put_task(req->task, 1);
1722 list_add(&req->inflight_entry, &ctx->locked_free_list);
1723 ctx->locked_free_nr++;
1725 if (!percpu_ref_tryget(&ctx->refs))
1728 io_commit_cqring(ctx);
1729 spin_unlock(&ctx->completion_lock);
1732 io_cqring_ev_posted(ctx);
1733 percpu_ref_put(&ctx->refs);
1737 static inline bool io_req_needs_clean(struct io_kiocb *req)
1739 return req->flags & IO_REQ_CLEAN_FLAGS;
1742 static void io_req_complete_state(struct io_kiocb *req, long res,
1743 unsigned int cflags)
1745 if (io_req_needs_clean(req))
1748 req->compl.cflags = cflags;
1749 req->flags |= REQ_F_COMPLETE_INLINE;
1752 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1753 long res, unsigned cflags)
1755 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1756 io_req_complete_state(req, res, cflags);
1758 io_req_complete_post(req, res, cflags);
1761 static inline void io_req_complete(struct io_kiocb *req, long res)
1763 __io_req_complete(req, 0, res, 0);
1766 static void io_req_complete_failed(struct io_kiocb *req, long res)
1769 io_req_complete_post(req, res, 0);
1773 * Don't initialise the fields below on every allocation, but do that in
1774 * advance and keep them valid across allocations.
1776 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1780 req->async_data = NULL;
1781 /* not necessary, but safer to zero */
1785 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1786 struct io_submit_state *state)
1788 spin_lock(&ctx->completion_lock);
1789 list_splice_init(&ctx->locked_free_list, &state->free_list);
1790 ctx->locked_free_nr = 0;
1791 spin_unlock(&ctx->completion_lock);
1794 /* Returns true IFF there are requests in the cache */
1795 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1797 struct io_submit_state *state = &ctx->submit_state;
1801 * If we have more than a batch's worth of requests in our IRQ side
1802 * locked cache, grab the lock and move them over to our submission
1805 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1806 io_flush_cached_locked_reqs(ctx, state);
1808 nr = state->free_reqs;
1809 while (!list_empty(&state->free_list)) {
1810 struct io_kiocb *req = list_first_entry(&state->free_list,
1811 struct io_kiocb, inflight_entry);
1813 list_del(&req->inflight_entry);
1814 state->reqs[nr++] = req;
1815 if (nr == ARRAY_SIZE(state->reqs))
1819 state->free_reqs = nr;
1824 * A request might get retired back into the request caches even before opcode
1825 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1826 * Because of that, io_alloc_req() should be called only under ->uring_lock
1827 * and with extra caution to not get a request that is still worked on.
1829 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1830 __must_hold(&ctx->uring_lock)
1832 struct io_submit_state *state = &ctx->submit_state;
1833 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1836 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1838 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1841 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1845 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1846 * retry single alloc to be on the safe side.
1848 if (unlikely(ret <= 0)) {
1849 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1850 if (!state->reqs[0])
1855 for (i = 0; i < ret; i++)
1856 io_preinit_req(state->reqs[i], ctx);
1857 state->free_reqs = ret;
1860 return state->reqs[state->free_reqs];
1863 static inline void io_put_file(struct file *file)
1869 static void io_dismantle_req(struct io_kiocb *req)
1871 unsigned int flags = req->flags;
1873 if (io_req_needs_clean(req))
1875 if (!(flags & REQ_F_FIXED_FILE))
1876 io_put_file(req->file);
1877 if (req->fixed_rsrc_refs)
1878 percpu_ref_put(req->fixed_rsrc_refs);
1879 if (req->async_data) {
1880 kfree(req->async_data);
1881 req->async_data = NULL;
1885 static void __io_free_req(struct io_kiocb *req)
1887 struct io_ring_ctx *ctx = req->ctx;
1889 io_dismantle_req(req);
1890 io_put_task(req->task, 1);
1892 spin_lock(&ctx->completion_lock);
1893 list_add(&req->inflight_entry, &ctx->locked_free_list);
1894 ctx->locked_free_nr++;
1895 spin_unlock(&ctx->completion_lock);
1897 percpu_ref_put(&ctx->refs);
1900 static inline void io_remove_next_linked(struct io_kiocb *req)
1902 struct io_kiocb *nxt = req->link;
1904 req->link = nxt->link;
1908 static bool io_kill_linked_timeout(struct io_kiocb *req)
1909 __must_hold(&req->ctx->completion_lock)
1910 __must_hold(&req->ctx->timeout_lock)
1912 struct io_kiocb *link = req->link;
1914 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1915 struct io_timeout_data *io = link->async_data;
1917 io_remove_next_linked(req);
1918 link->timeout.head = NULL;
1919 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1920 io_cqring_fill_event(link->ctx, link->user_data,
1922 io_put_req_deferred(link);
1929 static void io_fail_links(struct io_kiocb *req)
1930 __must_hold(&req->ctx->completion_lock)
1932 struct io_kiocb *nxt, *link = req->link;
1939 trace_io_uring_fail_link(req, link);
1940 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1941 io_put_req_deferred(link);
1946 static bool io_disarm_next(struct io_kiocb *req)
1947 __must_hold(&req->ctx->completion_lock)
1949 bool posted = false;
1951 if (req->flags & REQ_F_ARM_LTIMEOUT) {
1952 struct io_kiocb *link = req->link;
1954 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1955 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1956 io_remove_next_linked(req);
1957 io_cqring_fill_event(link->ctx, link->user_data,
1959 io_put_req_deferred(link);
1962 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
1963 struct io_ring_ctx *ctx = req->ctx;
1965 spin_lock_irq(&ctx->timeout_lock);
1966 posted = io_kill_linked_timeout(req);
1967 spin_unlock_irq(&ctx->timeout_lock);
1969 if (unlikely((req->flags & REQ_F_FAIL) &&
1970 !(req->flags & REQ_F_HARDLINK))) {
1971 posted |= (req->link != NULL);
1977 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1979 struct io_kiocb *nxt;
1982 * If LINK is set, we have dependent requests in this chain. If we
1983 * didn't fail this request, queue the first one up, moving any other
1984 * dependencies to the next request. In case of failure, fail the rest
1987 if (req->flags & IO_DISARM_MASK) {
1988 struct io_ring_ctx *ctx = req->ctx;
1991 spin_lock(&ctx->completion_lock);
1992 posted = io_disarm_next(req);
1994 io_commit_cqring(req->ctx);
1995 spin_unlock(&ctx->completion_lock);
1997 io_cqring_ev_posted(ctx);
2004 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2006 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2008 return __io_req_find_next(req);
2011 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2015 if (ctx->submit_state.compl_nr) {
2016 mutex_lock(&ctx->uring_lock);
2017 if (ctx->submit_state.compl_nr)
2018 io_submit_flush_completions(ctx);
2019 mutex_unlock(&ctx->uring_lock);
2021 percpu_ref_put(&ctx->refs);
2024 static void tctx_task_work(struct callback_head *cb)
2026 struct io_ring_ctx *ctx = NULL;
2027 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2031 struct io_wq_work_node *node;
2033 spin_lock_irq(&tctx->task_lock);
2034 node = tctx->task_list.first;
2035 INIT_WQ_LIST(&tctx->task_list);
2037 tctx->task_running = false;
2038 spin_unlock_irq(&tctx->task_lock);
2043 struct io_wq_work_node *next = node->next;
2044 struct io_kiocb *req = container_of(node, struct io_kiocb,
2047 if (req->ctx != ctx) {
2048 ctx_flush_and_put(ctx);
2050 percpu_ref_get(&ctx->refs);
2052 req->io_task_work.func(req);
2059 ctx_flush_and_put(ctx);
2062 static void io_req_task_work_add(struct io_kiocb *req)
2064 struct task_struct *tsk = req->task;
2065 struct io_uring_task *tctx = tsk->io_uring;
2066 enum task_work_notify_mode notify;
2067 struct io_wq_work_node *node;
2068 unsigned long flags;
2071 WARN_ON_ONCE(!tctx);
2073 spin_lock_irqsave(&tctx->task_lock, flags);
2074 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2075 running = tctx->task_running;
2077 tctx->task_running = true;
2078 spin_unlock_irqrestore(&tctx->task_lock, flags);
2080 /* task_work already pending, we're done */
2085 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2086 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2087 * processing task_work. There's no reliable way to tell if TWA_RESUME
2090 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2091 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2092 wake_up_process(tsk);
2096 spin_lock_irqsave(&tctx->task_lock, flags);
2097 tctx->task_running = false;
2098 node = tctx->task_list.first;
2099 INIT_WQ_LIST(&tctx->task_list);
2100 spin_unlock_irqrestore(&tctx->task_lock, flags);
2103 req = container_of(node, struct io_kiocb, io_task_work.node);
2105 if (llist_add(&req->io_task_work.fallback_node,
2106 &req->ctx->fallback_llist))
2107 schedule_delayed_work(&req->ctx->fallback_work, 1);
2111 static void io_req_task_cancel(struct io_kiocb *req)
2113 struct io_ring_ctx *ctx = req->ctx;
2115 /* ctx is guaranteed to stay alive while we hold uring_lock */
2116 mutex_lock(&ctx->uring_lock);
2117 io_req_complete_failed(req, req->result);
2118 mutex_unlock(&ctx->uring_lock);
2121 static void io_req_task_submit(struct io_kiocb *req)
2123 struct io_ring_ctx *ctx = req->ctx;
2125 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2126 mutex_lock(&ctx->uring_lock);
2127 /* req->task == current here, checking PF_EXITING is safe */
2128 if (likely(!(req->task->flags & PF_EXITING)))
2129 __io_queue_sqe(req);
2131 io_req_complete_failed(req, -EFAULT);
2132 mutex_unlock(&ctx->uring_lock);
2135 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2138 req->io_task_work.func = io_req_task_cancel;
2139 io_req_task_work_add(req);
2142 static void io_req_task_queue(struct io_kiocb *req)
2144 req->io_task_work.func = io_req_task_submit;
2145 io_req_task_work_add(req);
2148 static void io_req_task_queue_reissue(struct io_kiocb *req)
2150 req->io_task_work.func = io_queue_async_work;
2151 io_req_task_work_add(req);
2154 static inline void io_queue_next(struct io_kiocb *req)
2156 struct io_kiocb *nxt = io_req_find_next(req);
2159 io_req_task_queue(nxt);
2162 static void io_free_req(struct io_kiocb *req)
2169 struct task_struct *task;
2174 static inline void io_init_req_batch(struct req_batch *rb)
2181 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2182 struct req_batch *rb)
2185 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2187 io_put_task(rb->task, rb->task_refs);
2190 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2191 struct io_submit_state *state)
2194 io_dismantle_req(req);
2196 if (req->task != rb->task) {
2198 io_put_task(rb->task, rb->task_refs);
2199 rb->task = req->task;
2205 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2206 state->reqs[state->free_reqs++] = req;
2208 list_add(&req->inflight_entry, &state->free_list);
2211 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2212 __must_hold(&ctx->uring_lock)
2214 struct io_submit_state *state = &ctx->submit_state;
2215 int i, nr = state->compl_nr;
2216 struct req_batch rb;
2218 spin_lock(&ctx->completion_lock);
2219 for (i = 0; i < nr; i++) {
2220 struct io_kiocb *req = state->compl_reqs[i];
2222 __io_cqring_fill_event(ctx, req->user_data, req->result,
2225 io_commit_cqring(ctx);
2226 spin_unlock(&ctx->completion_lock);
2227 io_cqring_ev_posted(ctx);
2229 io_init_req_batch(&rb);
2230 for (i = 0; i < nr; i++) {
2231 struct io_kiocb *req = state->compl_reqs[i];
2233 if (req_ref_put_and_test(req))
2234 io_req_free_batch(&rb, req, &ctx->submit_state);
2237 io_req_free_batch_finish(ctx, &rb);
2238 state->compl_nr = 0;
2242 * Drop reference to request, return next in chain (if there is one) if this
2243 * was the last reference to this request.
2245 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2247 struct io_kiocb *nxt = NULL;
2249 if (req_ref_put_and_test(req)) {
2250 nxt = io_req_find_next(req);
2256 static inline void io_put_req(struct io_kiocb *req)
2258 if (req_ref_put_and_test(req))
2262 static inline void io_put_req_deferred(struct io_kiocb *req)
2264 if (req_ref_put_and_test(req)) {
2265 req->io_task_work.func = io_free_req;
2266 io_req_task_work_add(req);
2270 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2272 /* See comment at the top of this file */
2274 return __io_cqring_events(ctx);
2277 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2279 struct io_rings *rings = ctx->rings;
2281 /* make sure SQ entry isn't read before tail */
2282 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2285 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2287 unsigned int cflags;
2289 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2290 cflags |= IORING_CQE_F_BUFFER;
2291 req->flags &= ~REQ_F_BUFFER_SELECTED;
2296 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2298 struct io_buffer *kbuf;
2300 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2302 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2303 return io_put_kbuf(req, kbuf);
2306 static inline bool io_run_task_work(void)
2308 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2309 __set_current_state(TASK_RUNNING);
2310 tracehook_notify_signal();
2318 * Find and free completed poll iocbs
2320 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2321 struct list_head *done)
2323 struct req_batch rb;
2324 struct io_kiocb *req;
2326 /* order with ->result store in io_complete_rw_iopoll() */
2329 io_init_req_batch(&rb);
2330 while (!list_empty(done)) {
2331 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2332 list_del(&req->inflight_entry);
2334 if (READ_ONCE(req->result) == -EAGAIN &&
2335 !(req->flags & REQ_F_DONT_REISSUE)) {
2336 req->iopoll_completed = 0;
2337 io_req_task_queue_reissue(req);
2341 __io_cqring_fill_event(ctx, req->user_data, req->result,
2342 io_put_rw_kbuf(req));
2345 if (req_ref_put_and_test(req))
2346 io_req_free_batch(&rb, req, &ctx->submit_state);
2349 io_commit_cqring(ctx);
2350 io_cqring_ev_posted_iopoll(ctx);
2351 io_req_free_batch_finish(ctx, &rb);
2354 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2357 struct io_kiocb *req, *tmp;
2362 * Only spin for completions if we don't have multiple devices hanging
2363 * off our complete list, and we're under the requested amount.
2365 spin = !ctx->poll_multi_queue && *nr_events < min;
2367 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2368 struct kiocb *kiocb = &req->rw.kiocb;
2372 * Move completed and retryable entries to our local lists.
2373 * If we find a request that requires polling, break out
2374 * and complete those lists first, if we have entries there.
2376 if (READ_ONCE(req->iopoll_completed)) {
2377 list_move_tail(&req->inflight_entry, &done);
2380 if (!list_empty(&done))
2383 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2384 if (unlikely(ret < 0))
2389 /* iopoll may have completed current req */
2390 if (READ_ONCE(req->iopoll_completed))
2391 list_move_tail(&req->inflight_entry, &done);
2394 if (!list_empty(&done))
2395 io_iopoll_complete(ctx, nr_events, &done);
2401 * We can't just wait for polled events to come to us, we have to actively
2402 * find and complete them.
2404 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2406 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2409 mutex_lock(&ctx->uring_lock);
2410 while (!list_empty(&ctx->iopoll_list)) {
2411 unsigned int nr_events = 0;
2413 io_do_iopoll(ctx, &nr_events, 0);
2415 /* let it sleep and repeat later if can't complete a request */
2419 * Ensure we allow local-to-the-cpu processing to take place,
2420 * in this case we need to ensure that we reap all events.
2421 * Also let task_work, etc. to progress by releasing the mutex
2423 if (need_resched()) {
2424 mutex_unlock(&ctx->uring_lock);
2426 mutex_lock(&ctx->uring_lock);
2429 mutex_unlock(&ctx->uring_lock);
2432 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2434 unsigned int nr_events = 0;
2438 * We disallow the app entering submit/complete with polling, but we
2439 * still need to lock the ring to prevent racing with polled issue
2440 * that got punted to a workqueue.
2442 mutex_lock(&ctx->uring_lock);
2444 * Don't enter poll loop if we already have events pending.
2445 * If we do, we can potentially be spinning for commands that
2446 * already triggered a CQE (eg in error).
2448 if (test_bit(0, &ctx->check_cq_overflow))
2449 __io_cqring_overflow_flush(ctx, false);
2450 if (io_cqring_events(ctx))
2454 * If a submit got punted to a workqueue, we can have the
2455 * application entering polling for a command before it gets
2456 * issued. That app will hold the uring_lock for the duration
2457 * of the poll right here, so we need to take a breather every
2458 * now and then to ensure that the issue has a chance to add
2459 * the poll to the issued list. Otherwise we can spin here
2460 * forever, while the workqueue is stuck trying to acquire the
2463 if (list_empty(&ctx->iopoll_list)) {
2464 u32 tail = ctx->cached_cq_tail;
2466 mutex_unlock(&ctx->uring_lock);
2468 mutex_lock(&ctx->uring_lock);
2470 /* some requests don't go through iopoll_list */
2471 if (tail != ctx->cached_cq_tail ||
2472 list_empty(&ctx->iopoll_list))
2475 ret = io_do_iopoll(ctx, &nr_events, min);
2476 } while (!ret && nr_events < min && !need_resched());
2478 mutex_unlock(&ctx->uring_lock);
2482 static void kiocb_end_write(struct io_kiocb *req)
2485 * Tell lockdep we inherited freeze protection from submission
2488 if (req->flags & REQ_F_ISREG) {
2489 struct super_block *sb = file_inode(req->file)->i_sb;
2491 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2497 static bool io_resubmit_prep(struct io_kiocb *req)
2499 struct io_async_rw *rw = req->async_data;
2502 return !io_req_prep_async(req);
2503 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2504 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2508 static bool io_rw_should_reissue(struct io_kiocb *req)
2510 umode_t mode = file_inode(req->file)->i_mode;
2511 struct io_ring_ctx *ctx = req->ctx;
2513 if (!S_ISBLK(mode) && !S_ISREG(mode))
2515 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2516 !(ctx->flags & IORING_SETUP_IOPOLL)))
2519 * If ref is dying, we might be running poll reap from the exit work.
2520 * Don't attempt to reissue from that path, just let it fail with
2523 if (percpu_ref_is_dying(&ctx->refs))
2526 * Play it safe and assume not safe to re-import and reissue if we're
2527 * not in the original thread group (or in task context).
2529 if (!same_thread_group(req->task, current) || !in_task())
2534 static bool io_resubmit_prep(struct io_kiocb *req)
2538 static bool io_rw_should_reissue(struct io_kiocb *req)
2544 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2546 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2547 kiocb_end_write(req);
2548 if (res != req->result) {
2549 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2550 io_rw_should_reissue(req)) {
2551 req->flags |= REQ_F_REISSUE;
2560 static void io_req_task_complete(struct io_kiocb *req)
2562 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2565 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2566 unsigned int issue_flags)
2568 if (__io_complete_rw_common(req, res))
2570 io_req_task_complete(req);
2573 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2575 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2577 if (__io_complete_rw_common(req, res))
2580 req->io_task_work.func = io_req_task_complete;
2581 io_req_task_work_add(req);
2584 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2586 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2588 if (kiocb->ki_flags & IOCB_WRITE)
2589 kiocb_end_write(req);
2590 if (unlikely(res != req->result)) {
2591 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2592 io_resubmit_prep(req))) {
2594 req->flags |= REQ_F_DONT_REISSUE;
2598 WRITE_ONCE(req->result, res);
2599 /* order with io_iopoll_complete() checking ->result */
2601 WRITE_ONCE(req->iopoll_completed, 1);
2605 * After the iocb has been issued, it's safe to be found on the poll list.
2606 * Adding the kiocb to the list AFTER submission ensures that we don't
2607 * find it from a io_do_iopoll() thread before the issuer is done
2608 * accessing the kiocb cookie.
2610 static void io_iopoll_req_issued(struct io_kiocb *req)
2612 struct io_ring_ctx *ctx = req->ctx;
2613 const bool in_async = io_wq_current_is_worker();
2615 /* workqueue context doesn't hold uring_lock, grab it now */
2616 if (unlikely(in_async))
2617 mutex_lock(&ctx->uring_lock);
2620 * Track whether we have multiple files in our lists. This will impact
2621 * how we do polling eventually, not spinning if we're on potentially
2622 * different devices.
2624 if (list_empty(&ctx->iopoll_list)) {
2625 ctx->poll_multi_queue = false;
2626 } else if (!ctx->poll_multi_queue) {
2627 struct io_kiocb *list_req;
2628 unsigned int queue_num0, queue_num1;
2630 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2633 if (list_req->file != req->file) {
2634 ctx->poll_multi_queue = true;
2636 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2637 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2638 if (queue_num0 != queue_num1)
2639 ctx->poll_multi_queue = true;
2644 * For fast devices, IO may have already completed. If it has, add
2645 * it to the front so we find it first.
2647 if (READ_ONCE(req->iopoll_completed))
2648 list_add(&req->inflight_entry, &ctx->iopoll_list);
2650 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2652 if (unlikely(in_async)) {
2654 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2655 * in sq thread task context or in io worker task context. If
2656 * current task context is sq thread, we don't need to check
2657 * whether should wake up sq thread.
2659 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2660 wq_has_sleeper(&ctx->sq_data->wait))
2661 wake_up(&ctx->sq_data->wait);
2663 mutex_unlock(&ctx->uring_lock);
2667 static bool io_bdev_nowait(struct block_device *bdev)
2669 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2673 * If we tracked the file through the SCM inflight mechanism, we could support
2674 * any file. For now, just ensure that anything potentially problematic is done
2677 static bool __io_file_supports_nowait(struct file *file, int rw)
2679 umode_t mode = file_inode(file)->i_mode;
2681 if (S_ISBLK(mode)) {
2682 if (IS_ENABLED(CONFIG_BLOCK) &&
2683 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2689 if (S_ISREG(mode)) {
2690 if (IS_ENABLED(CONFIG_BLOCK) &&
2691 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2692 file->f_op != &io_uring_fops)
2697 /* any ->read/write should understand O_NONBLOCK */
2698 if (file->f_flags & O_NONBLOCK)
2701 if (!(file->f_mode & FMODE_NOWAIT))
2705 return file->f_op->read_iter != NULL;
2707 return file->f_op->write_iter != NULL;
2710 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2712 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2714 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2717 return __io_file_supports_nowait(req->file, rw);
2720 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2722 struct io_ring_ctx *ctx = req->ctx;
2723 struct kiocb *kiocb = &req->rw.kiocb;
2724 struct file *file = req->file;
2728 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2729 req->flags |= REQ_F_ISREG;
2731 kiocb->ki_pos = READ_ONCE(sqe->off);
2732 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2733 req->flags |= REQ_F_CUR_POS;
2734 kiocb->ki_pos = file->f_pos;
2736 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2737 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2738 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2742 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2743 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2744 req->flags |= REQ_F_NOWAIT;
2746 ioprio = READ_ONCE(sqe->ioprio);
2748 ret = ioprio_check_cap(ioprio);
2752 kiocb->ki_ioprio = ioprio;
2754 kiocb->ki_ioprio = get_current_ioprio();
2756 if (ctx->flags & IORING_SETUP_IOPOLL) {
2757 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2758 !kiocb->ki_filp->f_op->iopoll)
2761 kiocb->ki_flags |= IOCB_HIPRI;
2762 kiocb->ki_complete = io_complete_rw_iopoll;
2763 req->iopoll_completed = 0;
2765 if (kiocb->ki_flags & IOCB_HIPRI)
2767 kiocb->ki_complete = io_complete_rw;
2770 if (req->opcode == IORING_OP_READ_FIXED ||
2771 req->opcode == IORING_OP_WRITE_FIXED) {
2773 io_req_set_rsrc_node(req);
2776 req->rw.addr = READ_ONCE(sqe->addr);
2777 req->rw.len = READ_ONCE(sqe->len);
2778 req->buf_index = READ_ONCE(sqe->buf_index);
2782 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2788 case -ERESTARTNOINTR:
2789 case -ERESTARTNOHAND:
2790 case -ERESTART_RESTARTBLOCK:
2792 * We can't just restart the syscall, since previously
2793 * submitted sqes may already be in progress. Just fail this
2799 kiocb->ki_complete(kiocb, ret, 0);
2803 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2804 unsigned int issue_flags)
2806 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2807 struct io_async_rw *io = req->async_data;
2808 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2810 /* add previously done IO, if any */
2811 if (io && io->bytes_done > 0) {
2813 ret = io->bytes_done;
2815 ret += io->bytes_done;
2818 if (req->flags & REQ_F_CUR_POS)
2819 req->file->f_pos = kiocb->ki_pos;
2820 if (ret >= 0 && check_reissue)
2821 __io_complete_rw(req, ret, 0, issue_flags);
2823 io_rw_done(kiocb, ret);
2825 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2826 req->flags &= ~REQ_F_REISSUE;
2827 if (io_resubmit_prep(req)) {
2828 io_req_task_queue_reissue(req);
2831 __io_req_complete(req, issue_flags, ret,
2832 io_put_rw_kbuf(req));
2837 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2838 struct io_mapped_ubuf *imu)
2840 size_t len = req->rw.len;
2841 u64 buf_end, buf_addr = req->rw.addr;
2844 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2846 /* not inside the mapped region */
2847 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2851 * May not be a start of buffer, set size appropriately
2852 * and advance us to the beginning.
2854 offset = buf_addr - imu->ubuf;
2855 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2859 * Don't use iov_iter_advance() here, as it's really slow for
2860 * using the latter parts of a big fixed buffer - it iterates
2861 * over each segment manually. We can cheat a bit here, because
2864 * 1) it's a BVEC iter, we set it up
2865 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2866 * first and last bvec
2868 * So just find our index, and adjust the iterator afterwards.
2869 * If the offset is within the first bvec (or the whole first
2870 * bvec, just use iov_iter_advance(). This makes it easier
2871 * since we can just skip the first segment, which may not
2872 * be PAGE_SIZE aligned.
2874 const struct bio_vec *bvec = imu->bvec;
2876 if (offset <= bvec->bv_len) {
2877 iov_iter_advance(iter, offset);
2879 unsigned long seg_skip;
2881 /* skip first vec */
2882 offset -= bvec->bv_len;
2883 seg_skip = 1 + (offset >> PAGE_SHIFT);
2885 iter->bvec = bvec + seg_skip;
2886 iter->nr_segs -= seg_skip;
2887 iter->count -= bvec->bv_len + offset;
2888 iter->iov_offset = offset & ~PAGE_MASK;
2895 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2897 struct io_ring_ctx *ctx = req->ctx;
2898 struct io_mapped_ubuf *imu = req->imu;
2899 u16 index, buf_index = req->buf_index;
2902 if (unlikely(buf_index >= ctx->nr_user_bufs))
2904 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2905 imu = READ_ONCE(ctx->user_bufs[index]);
2908 return __io_import_fixed(req, rw, iter, imu);
2911 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2914 mutex_unlock(&ctx->uring_lock);
2917 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2920 * "Normal" inline submissions always hold the uring_lock, since we
2921 * grab it from the system call. Same is true for the SQPOLL offload.
2922 * The only exception is when we've detached the request and issue it
2923 * from an async worker thread, grab the lock for that case.
2926 mutex_lock(&ctx->uring_lock);
2929 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2930 int bgid, struct io_buffer *kbuf,
2933 struct io_buffer *head;
2935 if (req->flags & REQ_F_BUFFER_SELECTED)
2938 io_ring_submit_lock(req->ctx, needs_lock);
2940 lockdep_assert_held(&req->ctx->uring_lock);
2942 head = xa_load(&req->ctx->io_buffers, bgid);
2944 if (!list_empty(&head->list)) {
2945 kbuf = list_last_entry(&head->list, struct io_buffer,
2947 list_del(&kbuf->list);
2950 xa_erase(&req->ctx->io_buffers, bgid);
2952 if (*len > kbuf->len)
2955 kbuf = ERR_PTR(-ENOBUFS);
2958 io_ring_submit_unlock(req->ctx, needs_lock);
2963 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2966 struct io_buffer *kbuf;
2969 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2970 bgid = req->buf_index;
2971 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2974 req->rw.addr = (u64) (unsigned long) kbuf;
2975 req->flags |= REQ_F_BUFFER_SELECTED;
2976 return u64_to_user_ptr(kbuf->addr);
2979 #ifdef CONFIG_COMPAT
2980 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2983 struct compat_iovec __user *uiov;
2984 compat_ssize_t clen;
2988 uiov = u64_to_user_ptr(req->rw.addr);
2989 if (!access_ok(uiov, sizeof(*uiov)))
2991 if (__get_user(clen, &uiov->iov_len))
2997 buf = io_rw_buffer_select(req, &len, needs_lock);
2999 return PTR_ERR(buf);
3000 iov[0].iov_base = buf;
3001 iov[0].iov_len = (compat_size_t) len;
3006 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3009 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3013 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3016 len = iov[0].iov_len;
3019 buf = io_rw_buffer_select(req, &len, needs_lock);
3021 return PTR_ERR(buf);
3022 iov[0].iov_base = buf;
3023 iov[0].iov_len = len;
3027 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3030 if (req->flags & REQ_F_BUFFER_SELECTED) {
3031 struct io_buffer *kbuf;
3033 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3034 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3035 iov[0].iov_len = kbuf->len;
3038 if (req->rw.len != 1)
3041 #ifdef CONFIG_COMPAT
3042 if (req->ctx->compat)
3043 return io_compat_import(req, iov, needs_lock);
3046 return __io_iov_buffer_select(req, iov, needs_lock);
3049 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3050 struct iov_iter *iter, bool needs_lock)
3052 void __user *buf = u64_to_user_ptr(req->rw.addr);
3053 size_t sqe_len = req->rw.len;
3054 u8 opcode = req->opcode;
3057 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3059 return io_import_fixed(req, rw, iter);
3062 /* buffer index only valid with fixed read/write, or buffer select */
3063 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3066 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3067 if (req->flags & REQ_F_BUFFER_SELECT) {
3068 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3070 return PTR_ERR(buf);
3071 req->rw.len = sqe_len;
3074 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3079 if (req->flags & REQ_F_BUFFER_SELECT) {
3080 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3082 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3087 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3091 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3093 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3097 * For files that don't have ->read_iter() and ->write_iter(), handle them
3098 * by looping over ->read() or ->write() manually.
3100 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3102 struct kiocb *kiocb = &req->rw.kiocb;
3103 struct file *file = req->file;
3107 * Don't support polled IO through this interface, and we can't
3108 * support non-blocking either. For the latter, this just causes
3109 * the kiocb to be handled from an async context.
3111 if (kiocb->ki_flags & IOCB_HIPRI)
3113 if (kiocb->ki_flags & IOCB_NOWAIT)
3116 while (iov_iter_count(iter)) {
3120 if (!iov_iter_is_bvec(iter)) {
3121 iovec = iov_iter_iovec(iter);
3123 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3124 iovec.iov_len = req->rw.len;
3128 nr = file->f_op->read(file, iovec.iov_base,
3129 iovec.iov_len, io_kiocb_ppos(kiocb));
3131 nr = file->f_op->write(file, iovec.iov_base,
3132 iovec.iov_len, io_kiocb_ppos(kiocb));
3141 if (nr != iovec.iov_len)
3145 iov_iter_advance(iter, nr);
3151 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3152 const struct iovec *fast_iov, struct iov_iter *iter)
3154 struct io_async_rw *rw = req->async_data;
3156 memcpy(&rw->iter, iter, sizeof(*iter));
3157 rw->free_iovec = iovec;
3159 /* can only be fixed buffers, no need to do anything */
3160 if (iov_iter_is_bvec(iter))
3163 unsigned iov_off = 0;
3165 rw->iter.iov = rw->fast_iov;
3166 if (iter->iov != fast_iov) {
3167 iov_off = iter->iov - fast_iov;
3168 rw->iter.iov += iov_off;
3170 if (rw->fast_iov != fast_iov)
3171 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3172 sizeof(struct iovec) * iter->nr_segs);
3174 req->flags |= REQ_F_NEED_CLEANUP;
3178 static inline int io_alloc_async_data(struct io_kiocb *req)
3180 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3181 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3182 return req->async_data == NULL;
3185 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3186 const struct iovec *fast_iov,
3187 struct iov_iter *iter, bool force)
3189 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3191 if (!req->async_data) {
3192 if (io_alloc_async_data(req)) {
3197 io_req_map_rw(req, iovec, fast_iov, iter);
3202 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3204 struct io_async_rw *iorw = req->async_data;
3205 struct iovec *iov = iorw->fast_iov;
3208 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3209 if (unlikely(ret < 0))
3212 iorw->bytes_done = 0;
3213 iorw->free_iovec = iov;
3215 req->flags |= REQ_F_NEED_CLEANUP;
3219 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3221 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3223 return io_prep_rw(req, sqe);
3227 * This is our waitqueue callback handler, registered through lock_page_async()
3228 * when we initially tried to do the IO with the iocb armed our waitqueue.
3229 * This gets called when the page is unlocked, and we generally expect that to
3230 * happen when the page IO is completed and the page is now uptodate. This will
3231 * queue a task_work based retry of the operation, attempting to copy the data
3232 * again. If the latter fails because the page was NOT uptodate, then we will
3233 * do a thread based blocking retry of the operation. That's the unexpected
3236 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3237 int sync, void *arg)
3239 struct wait_page_queue *wpq;
3240 struct io_kiocb *req = wait->private;
3241 struct wait_page_key *key = arg;
3243 wpq = container_of(wait, struct wait_page_queue, wait);
3245 if (!wake_page_match(wpq, key))
3248 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3249 list_del_init(&wait->entry);
3250 io_req_task_queue(req);
3255 * This controls whether a given IO request should be armed for async page
3256 * based retry. If we return false here, the request is handed to the async
3257 * worker threads for retry. If we're doing buffered reads on a regular file,
3258 * we prepare a private wait_page_queue entry and retry the operation. This
3259 * will either succeed because the page is now uptodate and unlocked, or it
3260 * will register a callback when the page is unlocked at IO completion. Through
3261 * that callback, io_uring uses task_work to setup a retry of the operation.
3262 * That retry will attempt the buffered read again. The retry will generally
3263 * succeed, or in rare cases where it fails, we then fall back to using the
3264 * async worker threads for a blocking retry.
3266 static bool io_rw_should_retry(struct io_kiocb *req)
3268 struct io_async_rw *rw = req->async_data;
3269 struct wait_page_queue *wait = &rw->wpq;
3270 struct kiocb *kiocb = &req->rw.kiocb;
3272 /* never retry for NOWAIT, we just complete with -EAGAIN */
3273 if (req->flags & REQ_F_NOWAIT)
3276 /* Only for buffered IO */
3277 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3281 * just use poll if we can, and don't attempt if the fs doesn't
3282 * support callback based unlocks
3284 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3287 wait->wait.func = io_async_buf_func;
3288 wait->wait.private = req;
3289 wait->wait.flags = 0;
3290 INIT_LIST_HEAD(&wait->wait.entry);
3291 kiocb->ki_flags |= IOCB_WAITQ;
3292 kiocb->ki_flags &= ~IOCB_NOWAIT;
3293 kiocb->ki_waitq = wait;
3297 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3299 if (req->file->f_op->read_iter)
3300 return call_read_iter(req->file, &req->rw.kiocb, iter);
3301 else if (req->file->f_op->read)
3302 return loop_rw_iter(READ, req, iter);
3307 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3309 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3310 struct kiocb *kiocb = &req->rw.kiocb;
3311 struct iov_iter __iter, *iter = &__iter;
3312 struct io_async_rw *rw = req->async_data;
3313 ssize_t io_size, ret, ret2;
3314 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3320 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3324 io_size = iov_iter_count(iter);
3325 req->result = io_size;
3327 /* Ensure we clear previously set non-block flag */
3328 if (!force_nonblock)
3329 kiocb->ki_flags &= ~IOCB_NOWAIT;
3331 kiocb->ki_flags |= IOCB_NOWAIT;
3333 /* If the file doesn't support async, just async punt */
3334 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3335 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3336 return ret ?: -EAGAIN;
3339 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3340 if (unlikely(ret)) {
3345 ret = io_iter_do_read(req, iter);
3347 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3348 req->flags &= ~REQ_F_REISSUE;
3349 /* IOPOLL retry should happen for io-wq threads */
3350 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3352 /* no retry on NONBLOCK nor RWF_NOWAIT */
3353 if (req->flags & REQ_F_NOWAIT)
3355 /* some cases will consume bytes even on error returns */
3356 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3358 } else if (ret == -EIOCBQUEUED) {
3360 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3361 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3362 /* read all, failed, already did sync or don't want to retry */
3366 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3371 rw = req->async_data;
3372 /* now use our persistent iterator, if we aren't already */
3377 rw->bytes_done += ret;
3378 /* if we can retry, do so with the callbacks armed */
3379 if (!io_rw_should_retry(req)) {
3380 kiocb->ki_flags &= ~IOCB_WAITQ;
3385 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3386 * we get -EIOCBQUEUED, then we'll get a notification when the
3387 * desired page gets unlocked. We can also get a partial read
3388 * here, and if we do, then just retry at the new offset.
3390 ret = io_iter_do_read(req, iter);
3391 if (ret == -EIOCBQUEUED)
3393 /* we got some bytes, but not all. retry. */
3394 kiocb->ki_flags &= ~IOCB_WAITQ;
3395 } while (ret > 0 && ret < io_size);
3397 kiocb_done(kiocb, ret, issue_flags);
3399 /* it's faster to check here then delegate to kfree */
3405 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3407 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3409 return io_prep_rw(req, sqe);
3412 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3414 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3415 struct kiocb *kiocb = &req->rw.kiocb;
3416 struct iov_iter __iter, *iter = &__iter;
3417 struct io_async_rw *rw = req->async_data;
3418 ssize_t ret, ret2, io_size;
3419 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3425 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3429 io_size = iov_iter_count(iter);
3430 req->result = io_size;
3432 /* Ensure we clear previously set non-block flag */
3433 if (!force_nonblock)
3434 kiocb->ki_flags &= ~IOCB_NOWAIT;
3436 kiocb->ki_flags |= IOCB_NOWAIT;
3438 /* If the file doesn't support async, just async punt */
3439 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3442 /* file path doesn't support NOWAIT for non-direct_IO */
3443 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3444 (req->flags & REQ_F_ISREG))
3447 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3452 * Open-code file_start_write here to grab freeze protection,
3453 * which will be released by another thread in
3454 * io_complete_rw(). Fool lockdep by telling it the lock got
3455 * released so that it doesn't complain about the held lock when
3456 * we return to userspace.
3458 if (req->flags & REQ_F_ISREG) {
3459 sb_start_write(file_inode(req->file)->i_sb);
3460 __sb_writers_release(file_inode(req->file)->i_sb,
3463 kiocb->ki_flags |= IOCB_WRITE;
3465 if (req->file->f_op->write_iter)
3466 ret2 = call_write_iter(req->file, kiocb, iter);
3467 else if (req->file->f_op->write)
3468 ret2 = loop_rw_iter(WRITE, req, iter);
3472 if (req->flags & REQ_F_REISSUE) {
3473 req->flags &= ~REQ_F_REISSUE;
3478 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3479 * retry them without IOCB_NOWAIT.
3481 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3483 /* no retry on NONBLOCK nor RWF_NOWAIT */
3484 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3486 if (!force_nonblock || ret2 != -EAGAIN) {
3487 /* IOPOLL retry should happen for io-wq threads */
3488 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3491 kiocb_done(kiocb, ret2, issue_flags);
3494 /* some cases will consume bytes even on error returns */
3495 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3496 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3497 return ret ?: -EAGAIN;
3500 /* it's reportedly faster than delegating the null check to kfree() */
3506 static int io_renameat_prep(struct io_kiocb *req,
3507 const struct io_uring_sqe *sqe)
3509 struct io_rename *ren = &req->rename;
3510 const char __user *oldf, *newf;
3512 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3514 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3516 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3519 ren->old_dfd = READ_ONCE(sqe->fd);
3520 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3521 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3522 ren->new_dfd = READ_ONCE(sqe->len);
3523 ren->flags = READ_ONCE(sqe->rename_flags);
3525 ren->oldpath = getname(oldf);
3526 if (IS_ERR(ren->oldpath))
3527 return PTR_ERR(ren->oldpath);
3529 ren->newpath = getname(newf);
3530 if (IS_ERR(ren->newpath)) {
3531 putname(ren->oldpath);
3532 return PTR_ERR(ren->newpath);
3535 req->flags |= REQ_F_NEED_CLEANUP;
3539 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3541 struct io_rename *ren = &req->rename;
3544 if (issue_flags & IO_URING_F_NONBLOCK)
3547 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3548 ren->newpath, ren->flags);
3550 req->flags &= ~REQ_F_NEED_CLEANUP;
3553 io_req_complete(req, ret);
3557 static int io_unlinkat_prep(struct io_kiocb *req,
3558 const struct io_uring_sqe *sqe)
3560 struct io_unlink *un = &req->unlink;
3561 const char __user *fname;
3563 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3565 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3568 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3571 un->dfd = READ_ONCE(sqe->fd);
3573 un->flags = READ_ONCE(sqe->unlink_flags);
3574 if (un->flags & ~AT_REMOVEDIR)
3577 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3578 un->filename = getname(fname);
3579 if (IS_ERR(un->filename))
3580 return PTR_ERR(un->filename);
3582 req->flags |= REQ_F_NEED_CLEANUP;
3586 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3588 struct io_unlink *un = &req->unlink;
3591 if (issue_flags & IO_URING_F_NONBLOCK)
3594 if (un->flags & AT_REMOVEDIR)
3595 ret = do_rmdir(un->dfd, un->filename);
3597 ret = do_unlinkat(un->dfd, un->filename);
3599 req->flags &= ~REQ_F_NEED_CLEANUP;
3602 io_req_complete(req, ret);
3606 static int io_shutdown_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 #if defined(CONFIG_NET)
3610 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3612 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3613 sqe->buf_index || sqe->splice_fd_in))
3616 req->shutdown.how = READ_ONCE(sqe->len);
3623 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3625 #if defined(CONFIG_NET)
3626 struct socket *sock;
3629 if (issue_flags & IO_URING_F_NONBLOCK)
3632 sock = sock_from_file(req->file);
3633 if (unlikely(!sock))
3636 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3639 io_req_complete(req, ret);
3646 static int __io_splice_prep(struct io_kiocb *req,
3647 const struct io_uring_sqe *sqe)
3649 struct io_splice *sp = &req->splice;
3650 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3652 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3656 sp->len = READ_ONCE(sqe->len);
3657 sp->flags = READ_ONCE(sqe->splice_flags);
3659 if (unlikely(sp->flags & ~valid_flags))
3662 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3663 (sp->flags & SPLICE_F_FD_IN_FIXED));
3666 req->flags |= REQ_F_NEED_CLEANUP;
3670 static int io_tee_prep(struct io_kiocb *req,
3671 const struct io_uring_sqe *sqe)
3673 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3675 return __io_splice_prep(req, sqe);
3678 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3680 struct io_splice *sp = &req->splice;
3681 struct file *in = sp->file_in;
3682 struct file *out = sp->file_out;
3683 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3686 if (issue_flags & IO_URING_F_NONBLOCK)
3689 ret = do_tee(in, out, sp->len, flags);
3691 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3693 req->flags &= ~REQ_F_NEED_CLEANUP;
3697 io_req_complete(req, ret);
3701 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3703 struct io_splice *sp = &req->splice;
3705 sp->off_in = READ_ONCE(sqe->splice_off_in);
3706 sp->off_out = READ_ONCE(sqe->off);
3707 return __io_splice_prep(req, sqe);
3710 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3712 struct io_splice *sp = &req->splice;
3713 struct file *in = sp->file_in;
3714 struct file *out = sp->file_out;
3715 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3716 loff_t *poff_in, *poff_out;
3719 if (issue_flags & IO_URING_F_NONBLOCK)
3722 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3723 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3726 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3728 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3730 req->flags &= ~REQ_F_NEED_CLEANUP;
3734 io_req_complete(req, ret);
3739 * IORING_OP_NOP just posts a completion event, nothing else.
3741 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3743 struct io_ring_ctx *ctx = req->ctx;
3745 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3748 __io_req_complete(req, issue_flags, 0, 0);
3752 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3754 struct io_ring_ctx *ctx = req->ctx;
3759 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3761 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3765 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3766 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3769 req->sync.off = READ_ONCE(sqe->off);
3770 req->sync.len = READ_ONCE(sqe->len);
3774 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3776 loff_t end = req->sync.off + req->sync.len;
3779 /* fsync always requires a blocking context */
3780 if (issue_flags & IO_URING_F_NONBLOCK)
3783 ret = vfs_fsync_range(req->file, req->sync.off,
3784 end > 0 ? end : LLONG_MAX,
3785 req->sync.flags & IORING_FSYNC_DATASYNC);
3788 io_req_complete(req, ret);
3792 static int io_fallocate_prep(struct io_kiocb *req,
3793 const struct io_uring_sqe *sqe)
3795 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3798 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3801 req->sync.off = READ_ONCE(sqe->off);
3802 req->sync.len = READ_ONCE(sqe->addr);
3803 req->sync.mode = READ_ONCE(sqe->len);
3807 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3811 /* fallocate always requiring blocking context */
3812 if (issue_flags & IO_URING_F_NONBLOCK)
3814 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3818 io_req_complete(req, ret);
3822 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3824 const char __user *fname;
3827 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3829 if (unlikely(sqe->ioprio || sqe->buf_index || sqe->splice_fd_in))
3831 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3834 /* open.how should be already initialised */
3835 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3836 req->open.how.flags |= O_LARGEFILE;
3838 req->open.dfd = READ_ONCE(sqe->fd);
3839 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3840 req->open.filename = getname(fname);
3841 if (IS_ERR(req->open.filename)) {
3842 ret = PTR_ERR(req->open.filename);
3843 req->open.filename = NULL;
3846 req->open.nofile = rlimit(RLIMIT_NOFILE);
3847 req->flags |= REQ_F_NEED_CLEANUP;
3851 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3853 u64 mode = READ_ONCE(sqe->len);
3854 u64 flags = READ_ONCE(sqe->open_flags);
3856 req->open.how = build_open_how(flags, mode);
3857 return __io_openat_prep(req, sqe);
3860 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3862 struct open_how __user *how;
3866 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3867 len = READ_ONCE(sqe->len);
3868 if (len < OPEN_HOW_SIZE_VER0)
3871 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3876 return __io_openat_prep(req, sqe);
3879 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3881 struct open_flags op;
3884 bool resolve_nonblock;
3887 ret = build_open_flags(&req->open.how, &op);
3890 nonblock_set = op.open_flag & O_NONBLOCK;
3891 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3892 if (issue_flags & IO_URING_F_NONBLOCK) {
3894 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3895 * it'll always -EAGAIN
3897 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3899 op.lookup_flags |= LOOKUP_CACHED;
3900 op.open_flag |= O_NONBLOCK;
3903 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3907 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3910 * We could hang on to this 'fd' on retrying, but seems like
3911 * marginal gain for something that is now known to be a slower
3912 * path. So just put it, and we'll get a new one when we retry.
3916 ret = PTR_ERR(file);
3917 /* only retry if RESOLVE_CACHED wasn't already set by application */
3918 if (ret == -EAGAIN &&
3919 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3924 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3925 file->f_flags &= ~O_NONBLOCK;
3926 fsnotify_open(file);
3927 fd_install(ret, file);
3929 putname(req->open.filename);
3930 req->flags &= ~REQ_F_NEED_CLEANUP;
3933 __io_req_complete(req, issue_flags, ret, 0);
3937 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3939 return io_openat2(req, issue_flags);
3942 static int io_remove_buffers_prep(struct io_kiocb *req,
3943 const struct io_uring_sqe *sqe)
3945 struct io_provide_buf *p = &req->pbuf;
3948 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
3952 tmp = READ_ONCE(sqe->fd);
3953 if (!tmp || tmp > USHRT_MAX)
3956 memset(p, 0, sizeof(*p));
3958 p->bgid = READ_ONCE(sqe->buf_group);
3962 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3963 int bgid, unsigned nbufs)
3967 /* shouldn't happen */
3971 /* the head kbuf is the list itself */
3972 while (!list_empty(&buf->list)) {
3973 struct io_buffer *nxt;
3975 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3976 list_del(&nxt->list);
3983 xa_erase(&ctx->io_buffers, bgid);
3988 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3990 struct io_provide_buf *p = &req->pbuf;
3991 struct io_ring_ctx *ctx = req->ctx;
3992 struct io_buffer *head;
3994 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3996 io_ring_submit_lock(ctx, !force_nonblock);
3998 lockdep_assert_held(&ctx->uring_lock);
4001 head = xa_load(&ctx->io_buffers, p->bgid);
4003 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4007 /* complete before unlock, IOPOLL may need the lock */
4008 __io_req_complete(req, issue_flags, ret, 0);
4009 io_ring_submit_unlock(ctx, !force_nonblock);
4013 static int io_provide_buffers_prep(struct io_kiocb *req,
4014 const struct io_uring_sqe *sqe)
4016 unsigned long size, tmp_check;
4017 struct io_provide_buf *p = &req->pbuf;
4020 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4023 tmp = READ_ONCE(sqe->fd);
4024 if (!tmp || tmp > USHRT_MAX)
4027 p->addr = READ_ONCE(sqe->addr);
4028 p->len = READ_ONCE(sqe->len);
4030 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4033 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4036 size = (unsigned long)p->len * p->nbufs;
4037 if (!access_ok(u64_to_user_ptr(p->addr), size))
4040 p->bgid = READ_ONCE(sqe->buf_group);
4041 tmp = READ_ONCE(sqe->off);
4042 if (tmp > USHRT_MAX)
4048 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4050 struct io_buffer *buf;
4051 u64 addr = pbuf->addr;
4052 int i, bid = pbuf->bid;
4054 for (i = 0; i < pbuf->nbufs; i++) {
4055 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4060 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4065 INIT_LIST_HEAD(&buf->list);
4068 list_add_tail(&buf->list, &(*head)->list);
4072 return i ? i : -ENOMEM;
4075 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4077 struct io_provide_buf *p = &req->pbuf;
4078 struct io_ring_ctx *ctx = req->ctx;
4079 struct io_buffer *head, *list;
4081 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4083 io_ring_submit_lock(ctx, !force_nonblock);
4085 lockdep_assert_held(&ctx->uring_lock);
4087 list = head = xa_load(&ctx->io_buffers, p->bgid);
4089 ret = io_add_buffers(p, &head);
4090 if (ret >= 0 && !list) {
4091 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4093 __io_remove_buffers(ctx, head, p->bgid, -1U);
4097 /* complete before unlock, IOPOLL may need the lock */
4098 __io_req_complete(req, issue_flags, ret, 0);
4099 io_ring_submit_unlock(ctx, !force_nonblock);
4103 static int io_epoll_ctl_prep(struct io_kiocb *req,
4104 const struct io_uring_sqe *sqe)
4106 #if defined(CONFIG_EPOLL)
4107 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4109 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4112 req->epoll.epfd = READ_ONCE(sqe->fd);
4113 req->epoll.op = READ_ONCE(sqe->len);
4114 req->epoll.fd = READ_ONCE(sqe->off);
4116 if (ep_op_has_event(req->epoll.op)) {
4117 struct epoll_event __user *ev;
4119 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4120 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4130 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_EPOLL)
4133 struct io_epoll *ie = &req->epoll;
4135 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4137 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4138 if (force_nonblock && ret == -EAGAIN)
4143 __io_req_complete(req, issue_flags, ret, 0);
4150 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4152 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4153 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4155 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4158 req->madvise.addr = READ_ONCE(sqe->addr);
4159 req->madvise.len = READ_ONCE(sqe->len);
4160 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4167 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4169 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4170 struct io_madvise *ma = &req->madvise;
4173 if (issue_flags & IO_URING_F_NONBLOCK)
4176 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4179 io_req_complete(req, ret);
4186 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4188 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4190 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4193 req->fadvise.offset = READ_ONCE(sqe->off);
4194 req->fadvise.len = READ_ONCE(sqe->len);
4195 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4199 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4201 struct io_fadvise *fa = &req->fadvise;
4204 if (issue_flags & IO_URING_F_NONBLOCK) {
4205 switch (fa->advice) {
4206 case POSIX_FADV_NORMAL:
4207 case POSIX_FADV_RANDOM:
4208 case POSIX_FADV_SEQUENTIAL:
4215 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4218 __io_req_complete(req, issue_flags, ret, 0);
4222 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4224 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4226 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4228 if (req->flags & REQ_F_FIXED_FILE)
4231 req->statx.dfd = READ_ONCE(sqe->fd);
4232 req->statx.mask = READ_ONCE(sqe->len);
4233 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4234 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4235 req->statx.flags = READ_ONCE(sqe->statx_flags);
4240 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4242 struct io_statx *ctx = &req->statx;
4245 if (issue_flags & IO_URING_F_NONBLOCK)
4248 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4253 io_req_complete(req, ret);
4257 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4259 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4261 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4262 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4264 if (req->flags & REQ_F_FIXED_FILE)
4267 req->close.fd = READ_ONCE(sqe->fd);
4271 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4273 struct files_struct *files = current->files;
4274 struct io_close *close = &req->close;
4275 struct fdtable *fdt;
4276 struct file *file = NULL;
4279 spin_lock(&files->file_lock);
4280 fdt = files_fdtable(files);
4281 if (close->fd >= fdt->max_fds) {
4282 spin_unlock(&files->file_lock);
4285 file = fdt->fd[close->fd];
4286 if (!file || file->f_op == &io_uring_fops) {
4287 spin_unlock(&files->file_lock);
4292 /* if the file has a flush method, be safe and punt to async */
4293 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4294 spin_unlock(&files->file_lock);
4298 ret = __close_fd_get_file(close->fd, &file);
4299 spin_unlock(&files->file_lock);
4306 /* No ->flush() or already async, safely close from here */
4307 ret = filp_close(file, current->files);
4313 __io_req_complete(req, issue_flags, ret, 0);
4317 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4319 struct io_ring_ctx *ctx = req->ctx;
4321 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4323 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4327 req->sync.off = READ_ONCE(sqe->off);
4328 req->sync.len = READ_ONCE(sqe->len);
4329 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4333 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4337 /* sync_file_range always requires a blocking context */
4338 if (issue_flags & IO_URING_F_NONBLOCK)
4341 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4345 io_req_complete(req, ret);
4349 #if defined(CONFIG_NET)
4350 static int io_setup_async_msg(struct io_kiocb *req,
4351 struct io_async_msghdr *kmsg)
4353 struct io_async_msghdr *async_msg = req->async_data;
4357 if (io_alloc_async_data(req)) {
4358 kfree(kmsg->free_iov);
4361 async_msg = req->async_data;
4362 req->flags |= REQ_F_NEED_CLEANUP;
4363 memcpy(async_msg, kmsg, sizeof(*kmsg));
4364 async_msg->msg.msg_name = &async_msg->addr;
4365 /* if were using fast_iov, set it to the new one */
4366 if (!async_msg->free_iov)
4367 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4372 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4373 struct io_async_msghdr *iomsg)
4375 iomsg->msg.msg_name = &iomsg->addr;
4376 iomsg->free_iov = iomsg->fast_iov;
4377 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4378 req->sr_msg.msg_flags, &iomsg->free_iov);
4381 static int io_sendmsg_prep_async(struct io_kiocb *req)
4385 ret = io_sendmsg_copy_hdr(req, req->async_data);
4387 req->flags |= REQ_F_NEED_CLEANUP;
4391 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4393 struct io_sr_msg *sr = &req->sr_msg;
4395 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4398 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4399 sr->len = READ_ONCE(sqe->len);
4400 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4401 if (sr->msg_flags & MSG_DONTWAIT)
4402 req->flags |= REQ_F_NOWAIT;
4404 #ifdef CONFIG_COMPAT
4405 if (req->ctx->compat)
4406 sr->msg_flags |= MSG_CMSG_COMPAT;
4411 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4413 struct io_async_msghdr iomsg, *kmsg;
4414 struct socket *sock;
4419 sock = sock_from_file(req->file);
4420 if (unlikely(!sock))
4423 kmsg = req->async_data;
4425 ret = io_sendmsg_copy_hdr(req, &iomsg);
4431 flags = req->sr_msg.msg_flags;
4432 if (issue_flags & IO_URING_F_NONBLOCK)
4433 flags |= MSG_DONTWAIT;
4434 if (flags & MSG_WAITALL)
4435 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4437 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4438 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4439 return io_setup_async_msg(req, kmsg);
4440 if (ret == -ERESTARTSYS)
4443 /* fast path, check for non-NULL to avoid function call */
4445 kfree(kmsg->free_iov);
4446 req->flags &= ~REQ_F_NEED_CLEANUP;
4449 __io_req_complete(req, issue_flags, ret, 0);
4453 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4455 struct io_sr_msg *sr = &req->sr_msg;
4458 struct socket *sock;
4463 sock = sock_from_file(req->file);
4464 if (unlikely(!sock))
4467 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4471 msg.msg_name = NULL;
4472 msg.msg_control = NULL;
4473 msg.msg_controllen = 0;
4474 msg.msg_namelen = 0;
4476 flags = req->sr_msg.msg_flags;
4477 if (issue_flags & IO_URING_F_NONBLOCK)
4478 flags |= MSG_DONTWAIT;
4479 if (flags & MSG_WAITALL)
4480 min_ret = iov_iter_count(&msg.msg_iter);
4482 msg.msg_flags = flags;
4483 ret = sock_sendmsg(sock, &msg);
4484 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4486 if (ret == -ERESTARTSYS)
4491 __io_req_complete(req, issue_flags, ret, 0);
4495 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4496 struct io_async_msghdr *iomsg)
4498 struct io_sr_msg *sr = &req->sr_msg;
4499 struct iovec __user *uiov;
4503 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4504 &iomsg->uaddr, &uiov, &iov_len);
4508 if (req->flags & REQ_F_BUFFER_SELECT) {
4511 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4513 sr->len = iomsg->fast_iov[0].iov_len;
4514 iomsg->free_iov = NULL;
4516 iomsg->free_iov = iomsg->fast_iov;
4517 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4518 &iomsg->free_iov, &iomsg->msg.msg_iter,
4527 #ifdef CONFIG_COMPAT
4528 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4529 struct io_async_msghdr *iomsg)
4531 struct io_sr_msg *sr = &req->sr_msg;
4532 struct compat_iovec __user *uiov;
4537 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4542 uiov = compat_ptr(ptr);
4543 if (req->flags & REQ_F_BUFFER_SELECT) {
4544 compat_ssize_t clen;
4548 if (!access_ok(uiov, sizeof(*uiov)))
4550 if (__get_user(clen, &uiov->iov_len))
4555 iomsg->free_iov = NULL;
4557 iomsg->free_iov = iomsg->fast_iov;
4558 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4559 UIO_FASTIOV, &iomsg->free_iov,
4560 &iomsg->msg.msg_iter, true);
4569 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4570 struct io_async_msghdr *iomsg)
4572 iomsg->msg.msg_name = &iomsg->addr;
4574 #ifdef CONFIG_COMPAT
4575 if (req->ctx->compat)
4576 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4579 return __io_recvmsg_copy_hdr(req, iomsg);
4582 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4585 struct io_sr_msg *sr = &req->sr_msg;
4586 struct io_buffer *kbuf;
4588 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4593 req->flags |= REQ_F_BUFFER_SELECTED;
4597 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4599 return io_put_kbuf(req, req->sr_msg.kbuf);
4602 static int io_recvmsg_prep_async(struct io_kiocb *req)
4606 ret = io_recvmsg_copy_hdr(req, req->async_data);
4608 req->flags |= REQ_F_NEED_CLEANUP;
4612 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4614 struct io_sr_msg *sr = &req->sr_msg;
4616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4619 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4620 sr->len = READ_ONCE(sqe->len);
4621 sr->bgid = READ_ONCE(sqe->buf_group);
4622 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4623 if (sr->msg_flags & MSG_DONTWAIT)
4624 req->flags |= REQ_F_NOWAIT;
4626 #ifdef CONFIG_COMPAT
4627 if (req->ctx->compat)
4628 sr->msg_flags |= MSG_CMSG_COMPAT;
4633 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4635 struct io_async_msghdr iomsg, *kmsg;
4636 struct socket *sock;
4637 struct io_buffer *kbuf;
4640 int ret, cflags = 0;
4641 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4643 sock = sock_from_file(req->file);
4644 if (unlikely(!sock))
4647 kmsg = req->async_data;
4649 ret = io_recvmsg_copy_hdr(req, &iomsg);
4655 if (req->flags & REQ_F_BUFFER_SELECT) {
4656 kbuf = io_recv_buffer_select(req, !force_nonblock);
4658 return PTR_ERR(kbuf);
4659 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4660 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4661 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4662 1, req->sr_msg.len);
4665 flags = req->sr_msg.msg_flags;
4667 flags |= MSG_DONTWAIT;
4668 if (flags & MSG_WAITALL)
4669 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4671 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4672 kmsg->uaddr, flags);
4673 if (force_nonblock && ret == -EAGAIN)
4674 return io_setup_async_msg(req, kmsg);
4675 if (ret == -ERESTARTSYS)
4678 if (req->flags & REQ_F_BUFFER_SELECTED)
4679 cflags = io_put_recv_kbuf(req);
4680 /* fast path, check for non-NULL to avoid function call */
4682 kfree(kmsg->free_iov);
4683 req->flags &= ~REQ_F_NEED_CLEANUP;
4684 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4686 __io_req_complete(req, issue_flags, ret, cflags);
4690 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4692 struct io_buffer *kbuf;
4693 struct io_sr_msg *sr = &req->sr_msg;
4695 void __user *buf = sr->buf;
4696 struct socket *sock;
4700 int ret, cflags = 0;
4701 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4703 sock = sock_from_file(req->file);
4704 if (unlikely(!sock))
4707 if (req->flags & REQ_F_BUFFER_SELECT) {
4708 kbuf = io_recv_buffer_select(req, !force_nonblock);
4710 return PTR_ERR(kbuf);
4711 buf = u64_to_user_ptr(kbuf->addr);
4714 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4718 msg.msg_name = NULL;
4719 msg.msg_control = NULL;
4720 msg.msg_controllen = 0;
4721 msg.msg_namelen = 0;
4722 msg.msg_iocb = NULL;
4725 flags = req->sr_msg.msg_flags;
4727 flags |= MSG_DONTWAIT;
4728 if (flags & MSG_WAITALL)
4729 min_ret = iov_iter_count(&msg.msg_iter);
4731 ret = sock_recvmsg(sock, &msg, flags);
4732 if (force_nonblock && ret == -EAGAIN)
4734 if (ret == -ERESTARTSYS)
4737 if (req->flags & REQ_F_BUFFER_SELECTED)
4738 cflags = io_put_recv_kbuf(req);
4739 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4741 __io_req_complete(req, issue_flags, ret, cflags);
4745 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4747 struct io_accept *accept = &req->accept;
4749 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4751 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->splice_fd_in)
4754 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4755 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4756 accept->flags = READ_ONCE(sqe->accept_flags);
4757 accept->nofile = rlimit(RLIMIT_NOFILE);
4761 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4763 struct io_accept *accept = &req->accept;
4764 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4765 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4768 if (req->file->f_flags & O_NONBLOCK)
4769 req->flags |= REQ_F_NOWAIT;
4771 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4772 accept->addr_len, accept->flags,
4774 if (ret == -EAGAIN && force_nonblock)
4777 if (ret == -ERESTARTSYS)
4781 __io_req_complete(req, issue_flags, ret, 0);
4785 static int io_connect_prep_async(struct io_kiocb *req)
4787 struct io_async_connect *io = req->async_data;
4788 struct io_connect *conn = &req->connect;
4790 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4793 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4795 struct io_connect *conn = &req->connect;
4797 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4799 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4803 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4804 conn->addr_len = READ_ONCE(sqe->addr2);
4808 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4810 struct io_async_connect __io, *io;
4811 unsigned file_flags;
4813 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4815 if (req->async_data) {
4816 io = req->async_data;
4818 ret = move_addr_to_kernel(req->connect.addr,
4819 req->connect.addr_len,
4826 file_flags = force_nonblock ? O_NONBLOCK : 0;
4828 ret = __sys_connect_file(req->file, &io->address,
4829 req->connect.addr_len, file_flags);
4830 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4831 if (req->async_data)
4833 if (io_alloc_async_data(req)) {
4837 memcpy(req->async_data, &__io, sizeof(__io));
4840 if (ret == -ERESTARTSYS)
4845 __io_req_complete(req, issue_flags, ret, 0);
4848 #else /* !CONFIG_NET */
4849 #define IO_NETOP_FN(op) \
4850 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4852 return -EOPNOTSUPP; \
4855 #define IO_NETOP_PREP(op) \
4857 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4859 return -EOPNOTSUPP; \
4862 #define IO_NETOP_PREP_ASYNC(op) \
4864 static int io_##op##_prep_async(struct io_kiocb *req) \
4866 return -EOPNOTSUPP; \
4869 IO_NETOP_PREP_ASYNC(sendmsg);
4870 IO_NETOP_PREP_ASYNC(recvmsg);
4871 IO_NETOP_PREP_ASYNC(connect);
4872 IO_NETOP_PREP(accept);
4875 #endif /* CONFIG_NET */
4877 struct io_poll_table {
4878 struct poll_table_struct pt;
4879 struct io_kiocb *req;
4884 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4885 __poll_t mask, io_req_tw_func_t func)
4887 /* for instances that support it check for an event match first: */
4888 if (mask && !(mask & poll->events))
4891 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4893 list_del_init(&poll->wait.entry);
4896 req->io_task_work.func = func;
4899 * If this fails, then the task is exiting. When a task exits, the
4900 * work gets canceled, so just cancel this request as well instead
4901 * of executing it. We can't safely execute it anyway, as we may not
4902 * have the needed state needed for it anyway.
4904 io_req_task_work_add(req);
4908 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4909 __acquires(&req->ctx->completion_lock)
4911 struct io_ring_ctx *ctx = req->ctx;
4913 /* req->task == current here, checking PF_EXITING is safe */
4914 if (unlikely(req->task->flags & PF_EXITING))
4915 WRITE_ONCE(poll->canceled, true);
4917 if (!req->result && !READ_ONCE(poll->canceled)) {
4918 struct poll_table_struct pt = { ._key = poll->events };
4920 req->result = vfs_poll(req->file, &pt) & poll->events;
4923 spin_lock(&ctx->completion_lock);
4924 if (!req->result && !READ_ONCE(poll->canceled)) {
4925 add_wait_queue(poll->head, &poll->wait);
4932 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4934 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4935 if (req->opcode == IORING_OP_POLL_ADD)
4936 return req->async_data;
4937 return req->apoll->double_poll;
4940 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4942 if (req->opcode == IORING_OP_POLL_ADD)
4944 return &req->apoll->poll;
4947 static void io_poll_remove_double(struct io_kiocb *req)
4948 __must_hold(&req->ctx->completion_lock)
4950 struct io_poll_iocb *poll = io_poll_get_double(req);
4952 lockdep_assert_held(&req->ctx->completion_lock);
4954 if (poll && poll->head) {
4955 struct wait_queue_head *head = poll->head;
4957 spin_lock_irq(&head->lock);
4958 list_del_init(&poll->wait.entry);
4959 if (poll->wait.private)
4962 spin_unlock_irq(&head->lock);
4966 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4967 __must_hold(&req->ctx->completion_lock)
4969 struct io_ring_ctx *ctx = req->ctx;
4970 unsigned flags = IORING_CQE_F_MORE;
4973 if (READ_ONCE(req->poll.canceled)) {
4975 req->poll.events |= EPOLLONESHOT;
4977 error = mangle_poll(mask);
4979 if (req->poll.events & EPOLLONESHOT)
4981 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4982 req->poll.done = true;
4985 if (flags & IORING_CQE_F_MORE)
4988 io_commit_cqring(ctx);
4989 return !(flags & IORING_CQE_F_MORE);
4992 static void io_poll_task_func(struct io_kiocb *req)
4994 struct io_ring_ctx *ctx = req->ctx;
4995 struct io_kiocb *nxt;
4997 if (io_poll_rewait(req, &req->poll)) {
4998 spin_unlock(&ctx->completion_lock);
5002 done = io_poll_complete(req, req->result);
5004 io_poll_remove_double(req);
5005 hash_del(&req->hash_node);
5008 add_wait_queue(req->poll.head, &req->poll.wait);
5010 spin_unlock(&ctx->completion_lock);
5011 io_cqring_ev_posted(ctx);
5014 nxt = io_put_req_find_next(req);
5016 io_req_task_submit(nxt);
5021 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5022 int sync, void *key)
5024 struct io_kiocb *req = wait->private;
5025 struct io_poll_iocb *poll = io_poll_get_single(req);
5026 __poll_t mask = key_to_poll(key);
5027 unsigned long flags;
5029 /* for instances that support it check for an event match first: */
5030 if (mask && !(mask & poll->events))
5032 if (!(poll->events & EPOLLONESHOT))
5033 return poll->wait.func(&poll->wait, mode, sync, key);
5035 list_del_init(&wait->entry);
5040 spin_lock_irqsave(&poll->head->lock, flags);
5041 done = list_empty(&poll->wait.entry);
5043 list_del_init(&poll->wait.entry);
5044 /* make sure double remove sees this as being gone */
5045 wait->private = NULL;
5046 spin_unlock_irqrestore(&poll->head->lock, flags);
5048 /* use wait func handler, so it matches the rq type */
5049 poll->wait.func(&poll->wait, mode, sync, key);
5056 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5057 wait_queue_func_t wake_func)
5061 poll->canceled = false;
5062 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5063 /* mask in events that we always want/need */
5064 poll->events = events | IO_POLL_UNMASK;
5065 INIT_LIST_HEAD(&poll->wait.entry);
5066 init_waitqueue_func_entry(&poll->wait, wake_func);
5069 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5070 struct wait_queue_head *head,
5071 struct io_poll_iocb **poll_ptr)
5073 struct io_kiocb *req = pt->req;
5076 * The file being polled uses multiple waitqueues for poll handling
5077 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5080 if (unlikely(pt->nr_entries)) {
5081 struct io_poll_iocb *poll_one = poll;
5083 /* double add on the same waitqueue head, ignore */
5084 if (poll_one->head == head)
5086 /* already have a 2nd entry, fail a third attempt */
5088 if ((*poll_ptr)->head == head)
5090 pt->error = -EINVAL;
5094 * Can't handle multishot for double wait for now, turn it
5095 * into one-shot mode.
5097 if (!(poll_one->events & EPOLLONESHOT))
5098 poll_one->events |= EPOLLONESHOT;
5099 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5101 pt->error = -ENOMEM;
5104 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5106 poll->wait.private = req;
5113 if (poll->events & EPOLLEXCLUSIVE)
5114 add_wait_queue_exclusive(head, &poll->wait);
5116 add_wait_queue(head, &poll->wait);
5119 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5120 struct poll_table_struct *p)
5122 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5123 struct async_poll *apoll = pt->req->apoll;
5125 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5128 static void io_async_task_func(struct io_kiocb *req)
5130 struct async_poll *apoll = req->apoll;
5131 struct io_ring_ctx *ctx = req->ctx;
5133 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5135 if (io_poll_rewait(req, &apoll->poll)) {
5136 spin_unlock(&ctx->completion_lock);
5140 hash_del(&req->hash_node);
5141 io_poll_remove_double(req);
5142 spin_unlock(&ctx->completion_lock);
5144 if (!READ_ONCE(apoll->poll.canceled))
5145 io_req_task_submit(req);
5147 io_req_complete_failed(req, -ECANCELED);
5150 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5153 struct io_kiocb *req = wait->private;
5154 struct io_poll_iocb *poll = &req->apoll->poll;
5156 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5159 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5162 static void io_poll_req_insert(struct io_kiocb *req)
5164 struct io_ring_ctx *ctx = req->ctx;
5165 struct hlist_head *list;
5167 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5168 hlist_add_head(&req->hash_node, list);
5171 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5172 struct io_poll_iocb *poll,
5173 struct io_poll_table *ipt, __poll_t mask,
5174 wait_queue_func_t wake_func)
5175 __acquires(&ctx->completion_lock)
5177 struct io_ring_ctx *ctx = req->ctx;
5178 bool cancel = false;
5180 INIT_HLIST_NODE(&req->hash_node);
5181 io_init_poll_iocb(poll, mask, wake_func);
5182 poll->file = req->file;
5183 poll->wait.private = req;
5185 ipt->pt._key = mask;
5188 ipt->nr_entries = 0;
5190 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5191 if (unlikely(!ipt->nr_entries) && !ipt->error)
5192 ipt->error = -EINVAL;
5194 spin_lock(&ctx->completion_lock);
5195 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5196 io_poll_remove_double(req);
5197 if (likely(poll->head)) {
5198 spin_lock_irq(&poll->head->lock);
5199 if (unlikely(list_empty(&poll->wait.entry))) {
5205 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5206 list_del_init(&poll->wait.entry);
5208 WRITE_ONCE(poll->canceled, true);
5209 else if (!poll->done) /* actually waiting for an event */
5210 io_poll_req_insert(req);
5211 spin_unlock_irq(&poll->head->lock);
5223 static int io_arm_poll_handler(struct io_kiocb *req)
5225 const struct io_op_def *def = &io_op_defs[req->opcode];
5226 struct io_ring_ctx *ctx = req->ctx;
5227 struct async_poll *apoll;
5228 struct io_poll_table ipt;
5229 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5232 if (!req->file || !file_can_poll(req->file))
5233 return IO_APOLL_ABORTED;
5234 if (req->flags & REQ_F_POLLED)
5235 return IO_APOLL_ABORTED;
5236 if (!def->pollin && !def->pollout)
5237 return IO_APOLL_ABORTED;
5241 mask |= POLLIN | POLLRDNORM;
5243 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5244 if ((req->opcode == IORING_OP_RECVMSG) &&
5245 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5249 mask |= POLLOUT | POLLWRNORM;
5252 /* if we can't nonblock try, then no point in arming a poll handler */
5253 if (!io_file_supports_nowait(req, rw))
5254 return IO_APOLL_ABORTED;
5256 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5257 if (unlikely(!apoll))
5258 return IO_APOLL_ABORTED;
5259 apoll->double_poll = NULL;
5261 req->flags |= REQ_F_POLLED;
5262 ipt.pt._qproc = io_async_queue_proc;
5263 io_req_set_refcount(req);
5265 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5267 spin_unlock(&ctx->completion_lock);
5268 if (ret || ipt.error)
5269 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5271 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5272 mask, apoll->poll.events);
5276 static bool __io_poll_remove_one(struct io_kiocb *req,
5277 struct io_poll_iocb *poll, bool do_cancel)
5278 __must_hold(&req->ctx->completion_lock)
5280 bool do_complete = false;
5284 spin_lock_irq(&poll->head->lock);
5286 WRITE_ONCE(poll->canceled, true);
5287 if (!list_empty(&poll->wait.entry)) {
5288 list_del_init(&poll->wait.entry);
5291 spin_unlock_irq(&poll->head->lock);
5292 hash_del(&req->hash_node);
5296 static bool io_poll_remove_one(struct io_kiocb *req)
5297 __must_hold(&req->ctx->completion_lock)
5301 io_poll_remove_double(req);
5302 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5305 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5306 io_commit_cqring(req->ctx);
5308 io_put_req_deferred(req);
5314 * Returns true if we found and killed one or more poll requests
5316 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5319 struct hlist_node *tmp;
5320 struct io_kiocb *req;
5323 spin_lock(&ctx->completion_lock);
5324 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5325 struct hlist_head *list;
5327 list = &ctx->cancel_hash[i];
5328 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5329 if (io_match_task(req, tsk, cancel_all))
5330 posted += io_poll_remove_one(req);
5333 spin_unlock(&ctx->completion_lock);
5336 io_cqring_ev_posted(ctx);
5341 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5343 __must_hold(&ctx->completion_lock)
5345 struct hlist_head *list;
5346 struct io_kiocb *req;
5348 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5349 hlist_for_each_entry(req, list, hash_node) {
5350 if (sqe_addr != req->user_data)
5352 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5359 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5361 __must_hold(&ctx->completion_lock)
5363 struct io_kiocb *req;
5365 req = io_poll_find(ctx, sqe_addr, poll_only);
5368 if (io_poll_remove_one(req))
5374 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5379 events = READ_ONCE(sqe->poll32_events);
5381 events = swahw32(events);
5383 if (!(flags & IORING_POLL_ADD_MULTI))
5384 events |= EPOLLONESHOT;
5385 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5388 static int io_poll_update_prep(struct io_kiocb *req,
5389 const struct io_uring_sqe *sqe)
5391 struct io_poll_update *upd = &req->poll_update;
5394 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5396 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5398 flags = READ_ONCE(sqe->len);
5399 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5400 IORING_POLL_ADD_MULTI))
5402 /* meaningless without update */
5403 if (flags == IORING_POLL_ADD_MULTI)
5406 upd->old_user_data = READ_ONCE(sqe->addr);
5407 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5408 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5410 upd->new_user_data = READ_ONCE(sqe->off);
5411 if (!upd->update_user_data && upd->new_user_data)
5413 if (upd->update_events)
5414 upd->events = io_poll_parse_events(sqe, flags);
5415 else if (sqe->poll32_events)
5421 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5424 struct io_kiocb *req = wait->private;
5425 struct io_poll_iocb *poll = &req->poll;
5427 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5430 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5431 struct poll_table_struct *p)
5433 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5435 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5438 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5440 struct io_poll_iocb *poll = &req->poll;
5443 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5445 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5447 flags = READ_ONCE(sqe->len);
5448 if (flags & ~IORING_POLL_ADD_MULTI)
5451 io_req_set_refcount(req);
5452 poll->events = io_poll_parse_events(sqe, flags);
5456 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5458 struct io_poll_iocb *poll = &req->poll;
5459 struct io_ring_ctx *ctx = req->ctx;
5460 struct io_poll_table ipt;
5463 ipt.pt._qproc = io_poll_queue_proc;
5465 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5468 if (mask) { /* no async, we'd stolen it */
5470 io_poll_complete(req, mask);
5472 spin_unlock(&ctx->completion_lock);
5475 io_cqring_ev_posted(ctx);
5476 if (poll->events & EPOLLONESHOT)
5482 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5484 struct io_ring_ctx *ctx = req->ctx;
5485 struct io_kiocb *preq;
5489 spin_lock(&ctx->completion_lock);
5490 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5496 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5498 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5503 * Don't allow racy completion with singleshot, as we cannot safely
5504 * update those. For multishot, if we're racing with completion, just
5505 * let completion re-add it.
5507 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5508 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5512 /* we now have a detached poll request. reissue. */
5516 spin_unlock(&ctx->completion_lock);
5518 io_req_complete(req, ret);
5521 /* only mask one event flags, keep behavior flags */
5522 if (req->poll_update.update_events) {
5523 preq->poll.events &= ~0xffff;
5524 preq->poll.events |= req->poll_update.events & 0xffff;
5525 preq->poll.events |= IO_POLL_UNMASK;
5527 if (req->poll_update.update_user_data)
5528 preq->user_data = req->poll_update.new_user_data;
5529 spin_unlock(&ctx->completion_lock);
5531 /* complete update request, we're done with it */
5532 io_req_complete(req, ret);
5535 ret = io_poll_add(preq, issue_flags);
5538 io_req_complete(preq, ret);
5544 static void io_req_task_timeout(struct io_kiocb *req)
5547 io_req_complete_post(req, -ETIME, 0);
5550 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5552 struct io_timeout_data *data = container_of(timer,
5553 struct io_timeout_data, timer);
5554 struct io_kiocb *req = data->req;
5555 struct io_ring_ctx *ctx = req->ctx;
5556 unsigned long flags;
5558 spin_lock_irqsave(&ctx->timeout_lock, flags);
5559 list_del_init(&req->timeout.list);
5560 atomic_set(&req->ctx->cq_timeouts,
5561 atomic_read(&req->ctx->cq_timeouts) + 1);
5562 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5564 req->io_task_work.func = io_req_task_timeout;
5565 io_req_task_work_add(req);
5566 return HRTIMER_NORESTART;
5569 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5571 __must_hold(&ctx->timeout_lock)
5573 struct io_timeout_data *io;
5574 struct io_kiocb *req;
5577 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5578 found = user_data == req->user_data;
5583 return ERR_PTR(-ENOENT);
5585 io = req->async_data;
5586 if (hrtimer_try_to_cancel(&io->timer) == -1)
5587 return ERR_PTR(-EALREADY);
5588 list_del_init(&req->timeout.list);
5592 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5593 __must_hold(&ctx->completion_lock)
5594 __must_hold(&ctx->timeout_lock)
5596 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5599 return PTR_ERR(req);
5602 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5603 io_put_req_deferred(req);
5607 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5608 struct timespec64 *ts, enum hrtimer_mode mode)
5609 __must_hold(&ctx->timeout_lock)
5611 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5612 struct io_timeout_data *data;
5615 return PTR_ERR(req);
5617 req->timeout.off = 0; /* noseq */
5618 data = req->async_data;
5619 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5620 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5621 data->timer.function = io_timeout_fn;
5622 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5626 static int io_timeout_remove_prep(struct io_kiocb *req,
5627 const struct io_uring_sqe *sqe)
5629 struct io_timeout_rem *tr = &req->timeout_rem;
5631 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5633 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5635 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5638 tr->addr = READ_ONCE(sqe->addr);
5639 tr->flags = READ_ONCE(sqe->timeout_flags);
5640 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5641 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5643 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5645 } else if (tr->flags) {
5646 /* timeout removal doesn't support flags */
5653 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5655 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5660 * Remove or update an existing timeout command
5662 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5664 struct io_timeout_rem *tr = &req->timeout_rem;
5665 struct io_ring_ctx *ctx = req->ctx;
5668 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5669 spin_lock(&ctx->completion_lock);
5670 spin_lock_irq(&ctx->timeout_lock);
5671 ret = io_timeout_cancel(ctx, tr->addr);
5672 spin_unlock_irq(&ctx->timeout_lock);
5673 spin_unlock(&ctx->completion_lock);
5675 spin_lock_irq(&ctx->timeout_lock);
5676 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5677 io_translate_timeout_mode(tr->flags));
5678 spin_unlock_irq(&ctx->timeout_lock);
5683 io_req_complete_post(req, ret, 0);
5687 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5688 bool is_timeout_link)
5690 struct io_timeout_data *data;
5692 u32 off = READ_ONCE(sqe->off);
5694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5696 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5699 if (off && is_timeout_link)
5701 flags = READ_ONCE(sqe->timeout_flags);
5702 if (flags & ~IORING_TIMEOUT_ABS)
5705 req->timeout.off = off;
5706 if (unlikely(off && !req->ctx->off_timeout_used))
5707 req->ctx->off_timeout_used = true;
5709 if (!req->async_data && io_alloc_async_data(req))
5712 data = req->async_data;
5715 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5718 data->mode = io_translate_timeout_mode(flags);
5719 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5721 if (is_timeout_link) {
5722 struct io_submit_link *link = &req->ctx->submit_state.link;
5726 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5728 req->timeout.head = link->last;
5729 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5734 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5736 struct io_ring_ctx *ctx = req->ctx;
5737 struct io_timeout_data *data = req->async_data;
5738 struct list_head *entry;
5739 u32 tail, off = req->timeout.off;
5741 spin_lock_irq(&ctx->timeout_lock);
5744 * sqe->off holds how many events that need to occur for this
5745 * timeout event to be satisfied. If it isn't set, then this is
5746 * a pure timeout request, sequence isn't used.
5748 if (io_is_timeout_noseq(req)) {
5749 entry = ctx->timeout_list.prev;
5753 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5754 req->timeout.target_seq = tail + off;
5756 /* Update the last seq here in case io_flush_timeouts() hasn't.
5757 * This is safe because ->completion_lock is held, and submissions
5758 * and completions are never mixed in the same ->completion_lock section.
5760 ctx->cq_last_tm_flush = tail;
5763 * Insertion sort, ensuring the first entry in the list is always
5764 * the one we need first.
5766 list_for_each_prev(entry, &ctx->timeout_list) {
5767 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5770 if (io_is_timeout_noseq(nxt))
5772 /* nxt.seq is behind @tail, otherwise would've been completed */
5773 if (off >= nxt->timeout.target_seq - tail)
5777 list_add(&req->timeout.list, entry);
5778 data->timer.function = io_timeout_fn;
5779 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5780 spin_unlock_irq(&ctx->timeout_lock);
5784 struct io_cancel_data {
5785 struct io_ring_ctx *ctx;
5789 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5791 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5792 struct io_cancel_data *cd = data;
5794 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5797 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5798 struct io_ring_ctx *ctx)
5800 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5801 enum io_wq_cancel cancel_ret;
5804 if (!tctx || !tctx->io_wq)
5807 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5808 switch (cancel_ret) {
5809 case IO_WQ_CANCEL_OK:
5812 case IO_WQ_CANCEL_RUNNING:
5815 case IO_WQ_CANCEL_NOTFOUND:
5823 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5825 struct io_ring_ctx *ctx = req->ctx;
5828 WARN_ON_ONCE(req->task != current);
5830 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5834 spin_lock(&ctx->completion_lock);
5835 spin_lock_irq(&ctx->timeout_lock);
5836 ret = io_timeout_cancel(ctx, sqe_addr);
5837 spin_unlock_irq(&ctx->timeout_lock);
5840 ret = io_poll_cancel(ctx, sqe_addr, false);
5842 spin_unlock(&ctx->completion_lock);
5846 static int io_async_cancel_prep(struct io_kiocb *req,
5847 const struct io_uring_sqe *sqe)
5849 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5851 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5853 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5857 req->cancel.addr = READ_ONCE(sqe->addr);
5861 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5863 struct io_ring_ctx *ctx = req->ctx;
5864 u64 sqe_addr = req->cancel.addr;
5865 struct io_tctx_node *node;
5868 ret = io_try_cancel_userdata(req, sqe_addr);
5872 /* slow path, try all io-wq's */
5873 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5875 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5876 struct io_uring_task *tctx = node->task->io_uring;
5878 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5882 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5886 io_req_complete_post(req, ret, 0);
5890 static int io_rsrc_update_prep(struct io_kiocb *req,
5891 const struct io_uring_sqe *sqe)
5893 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5895 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
5898 req->rsrc_update.offset = READ_ONCE(sqe->off);
5899 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5900 if (!req->rsrc_update.nr_args)
5902 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5906 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5908 struct io_ring_ctx *ctx = req->ctx;
5909 struct io_uring_rsrc_update2 up;
5912 if (issue_flags & IO_URING_F_NONBLOCK)
5915 up.offset = req->rsrc_update.offset;
5916 up.data = req->rsrc_update.arg;
5921 mutex_lock(&ctx->uring_lock);
5922 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5923 &up, req->rsrc_update.nr_args);
5924 mutex_unlock(&ctx->uring_lock);
5928 __io_req_complete(req, issue_flags, ret, 0);
5932 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5934 switch (req->opcode) {
5937 case IORING_OP_READV:
5938 case IORING_OP_READ_FIXED:
5939 case IORING_OP_READ:
5940 return io_read_prep(req, sqe);
5941 case IORING_OP_WRITEV:
5942 case IORING_OP_WRITE_FIXED:
5943 case IORING_OP_WRITE:
5944 return io_write_prep(req, sqe);
5945 case IORING_OP_POLL_ADD:
5946 return io_poll_add_prep(req, sqe);
5947 case IORING_OP_POLL_REMOVE:
5948 return io_poll_update_prep(req, sqe);
5949 case IORING_OP_FSYNC:
5950 return io_fsync_prep(req, sqe);
5951 case IORING_OP_SYNC_FILE_RANGE:
5952 return io_sfr_prep(req, sqe);
5953 case IORING_OP_SENDMSG:
5954 case IORING_OP_SEND:
5955 return io_sendmsg_prep(req, sqe);
5956 case IORING_OP_RECVMSG:
5957 case IORING_OP_RECV:
5958 return io_recvmsg_prep(req, sqe);
5959 case IORING_OP_CONNECT:
5960 return io_connect_prep(req, sqe);
5961 case IORING_OP_TIMEOUT:
5962 return io_timeout_prep(req, sqe, false);
5963 case IORING_OP_TIMEOUT_REMOVE:
5964 return io_timeout_remove_prep(req, sqe);
5965 case IORING_OP_ASYNC_CANCEL:
5966 return io_async_cancel_prep(req, sqe);
5967 case IORING_OP_LINK_TIMEOUT:
5968 return io_timeout_prep(req, sqe, true);
5969 case IORING_OP_ACCEPT:
5970 return io_accept_prep(req, sqe);
5971 case IORING_OP_FALLOCATE:
5972 return io_fallocate_prep(req, sqe);
5973 case IORING_OP_OPENAT:
5974 return io_openat_prep(req, sqe);
5975 case IORING_OP_CLOSE:
5976 return io_close_prep(req, sqe);
5977 case IORING_OP_FILES_UPDATE:
5978 return io_rsrc_update_prep(req, sqe);
5979 case IORING_OP_STATX:
5980 return io_statx_prep(req, sqe);
5981 case IORING_OP_FADVISE:
5982 return io_fadvise_prep(req, sqe);
5983 case IORING_OP_MADVISE:
5984 return io_madvise_prep(req, sqe);
5985 case IORING_OP_OPENAT2:
5986 return io_openat2_prep(req, sqe);
5987 case IORING_OP_EPOLL_CTL:
5988 return io_epoll_ctl_prep(req, sqe);
5989 case IORING_OP_SPLICE:
5990 return io_splice_prep(req, sqe);
5991 case IORING_OP_PROVIDE_BUFFERS:
5992 return io_provide_buffers_prep(req, sqe);
5993 case IORING_OP_REMOVE_BUFFERS:
5994 return io_remove_buffers_prep(req, sqe);
5996 return io_tee_prep(req, sqe);
5997 case IORING_OP_SHUTDOWN:
5998 return io_shutdown_prep(req, sqe);
5999 case IORING_OP_RENAMEAT:
6000 return io_renameat_prep(req, sqe);
6001 case IORING_OP_UNLINKAT:
6002 return io_unlinkat_prep(req, sqe);
6005 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6010 static int io_req_prep_async(struct io_kiocb *req)
6012 if (!io_op_defs[req->opcode].needs_async_setup)
6014 if (WARN_ON_ONCE(req->async_data))
6016 if (io_alloc_async_data(req))
6019 switch (req->opcode) {
6020 case IORING_OP_READV:
6021 return io_rw_prep_async(req, READ);
6022 case IORING_OP_WRITEV:
6023 return io_rw_prep_async(req, WRITE);
6024 case IORING_OP_SENDMSG:
6025 return io_sendmsg_prep_async(req);
6026 case IORING_OP_RECVMSG:
6027 return io_recvmsg_prep_async(req);
6028 case IORING_OP_CONNECT:
6029 return io_connect_prep_async(req);
6031 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6036 static u32 io_get_sequence(struct io_kiocb *req)
6038 u32 seq = req->ctx->cached_sq_head;
6040 /* need original cached_sq_head, but it was increased for each req */
6041 io_for_each_link(req, req)
6046 static bool io_drain_req(struct io_kiocb *req)
6048 struct io_kiocb *pos;
6049 struct io_ring_ctx *ctx = req->ctx;
6050 struct io_defer_entry *de;
6055 * If we need to drain a request in the middle of a link, drain the
6056 * head request and the next request/link after the current link.
6057 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6058 * maintained for every request of our link.
6060 if (ctx->drain_next) {
6061 req->flags |= REQ_F_IO_DRAIN;
6062 ctx->drain_next = false;
6064 /* not interested in head, start from the first linked */
6065 io_for_each_link(pos, req->link) {
6066 if (pos->flags & REQ_F_IO_DRAIN) {
6067 ctx->drain_next = true;
6068 req->flags |= REQ_F_IO_DRAIN;
6073 /* Still need defer if there is pending req in defer list. */
6074 if (likely(list_empty_careful(&ctx->defer_list) &&
6075 !(req->flags & REQ_F_IO_DRAIN))) {
6076 ctx->drain_active = false;
6080 seq = io_get_sequence(req);
6081 /* Still a chance to pass the sequence check */
6082 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6085 ret = io_req_prep_async(req);
6088 io_prep_async_link(req);
6089 de = kmalloc(sizeof(*de), GFP_KERNEL);
6093 io_req_complete_failed(req, ret);
6097 spin_lock(&ctx->completion_lock);
6098 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6099 spin_unlock(&ctx->completion_lock);
6101 io_queue_async_work(req);
6105 trace_io_uring_defer(ctx, req, req->user_data);
6108 list_add_tail(&de->list, &ctx->defer_list);
6109 spin_unlock(&ctx->completion_lock);
6113 static void io_clean_op(struct io_kiocb *req)
6115 if (req->flags & REQ_F_BUFFER_SELECTED) {
6116 switch (req->opcode) {
6117 case IORING_OP_READV:
6118 case IORING_OP_READ_FIXED:
6119 case IORING_OP_READ:
6120 kfree((void *)(unsigned long)req->rw.addr);
6122 case IORING_OP_RECVMSG:
6123 case IORING_OP_RECV:
6124 kfree(req->sr_msg.kbuf);
6129 if (req->flags & REQ_F_NEED_CLEANUP) {
6130 switch (req->opcode) {
6131 case IORING_OP_READV:
6132 case IORING_OP_READ_FIXED:
6133 case IORING_OP_READ:
6134 case IORING_OP_WRITEV:
6135 case IORING_OP_WRITE_FIXED:
6136 case IORING_OP_WRITE: {
6137 struct io_async_rw *io = req->async_data;
6139 kfree(io->free_iovec);
6142 case IORING_OP_RECVMSG:
6143 case IORING_OP_SENDMSG: {
6144 struct io_async_msghdr *io = req->async_data;
6146 kfree(io->free_iov);
6149 case IORING_OP_SPLICE:
6151 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6152 io_put_file(req->splice.file_in);
6154 case IORING_OP_OPENAT:
6155 case IORING_OP_OPENAT2:
6156 if (req->open.filename)
6157 putname(req->open.filename);
6159 case IORING_OP_RENAMEAT:
6160 putname(req->rename.oldpath);
6161 putname(req->rename.newpath);
6163 case IORING_OP_UNLINKAT:
6164 putname(req->unlink.filename);
6168 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6169 kfree(req->apoll->double_poll);
6173 if (req->flags & REQ_F_INFLIGHT) {
6174 struct io_uring_task *tctx = req->task->io_uring;
6176 atomic_dec(&tctx->inflight_tracked);
6178 if (req->flags & REQ_F_CREDS)
6179 put_cred(req->creds);
6181 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6184 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6186 struct io_ring_ctx *ctx = req->ctx;
6187 const struct cred *creds = NULL;
6190 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6191 creds = override_creds(req->creds);
6193 switch (req->opcode) {
6195 ret = io_nop(req, issue_flags);
6197 case IORING_OP_READV:
6198 case IORING_OP_READ_FIXED:
6199 case IORING_OP_READ:
6200 ret = io_read(req, issue_flags);
6202 case IORING_OP_WRITEV:
6203 case IORING_OP_WRITE_FIXED:
6204 case IORING_OP_WRITE:
6205 ret = io_write(req, issue_flags);
6207 case IORING_OP_FSYNC:
6208 ret = io_fsync(req, issue_flags);
6210 case IORING_OP_POLL_ADD:
6211 ret = io_poll_add(req, issue_flags);
6213 case IORING_OP_POLL_REMOVE:
6214 ret = io_poll_update(req, issue_flags);
6216 case IORING_OP_SYNC_FILE_RANGE:
6217 ret = io_sync_file_range(req, issue_flags);
6219 case IORING_OP_SENDMSG:
6220 ret = io_sendmsg(req, issue_flags);
6222 case IORING_OP_SEND:
6223 ret = io_send(req, issue_flags);
6225 case IORING_OP_RECVMSG:
6226 ret = io_recvmsg(req, issue_flags);
6228 case IORING_OP_RECV:
6229 ret = io_recv(req, issue_flags);
6231 case IORING_OP_TIMEOUT:
6232 ret = io_timeout(req, issue_flags);
6234 case IORING_OP_TIMEOUT_REMOVE:
6235 ret = io_timeout_remove(req, issue_flags);
6237 case IORING_OP_ACCEPT:
6238 ret = io_accept(req, issue_flags);
6240 case IORING_OP_CONNECT:
6241 ret = io_connect(req, issue_flags);
6243 case IORING_OP_ASYNC_CANCEL:
6244 ret = io_async_cancel(req, issue_flags);
6246 case IORING_OP_FALLOCATE:
6247 ret = io_fallocate(req, issue_flags);
6249 case IORING_OP_OPENAT:
6250 ret = io_openat(req, issue_flags);
6252 case IORING_OP_CLOSE:
6253 ret = io_close(req, issue_flags);
6255 case IORING_OP_FILES_UPDATE:
6256 ret = io_files_update(req, issue_flags);
6258 case IORING_OP_STATX:
6259 ret = io_statx(req, issue_flags);
6261 case IORING_OP_FADVISE:
6262 ret = io_fadvise(req, issue_flags);
6264 case IORING_OP_MADVISE:
6265 ret = io_madvise(req, issue_flags);
6267 case IORING_OP_OPENAT2:
6268 ret = io_openat2(req, issue_flags);
6270 case IORING_OP_EPOLL_CTL:
6271 ret = io_epoll_ctl(req, issue_flags);
6273 case IORING_OP_SPLICE:
6274 ret = io_splice(req, issue_flags);
6276 case IORING_OP_PROVIDE_BUFFERS:
6277 ret = io_provide_buffers(req, issue_flags);
6279 case IORING_OP_REMOVE_BUFFERS:
6280 ret = io_remove_buffers(req, issue_flags);
6283 ret = io_tee(req, issue_flags);
6285 case IORING_OP_SHUTDOWN:
6286 ret = io_shutdown(req, issue_flags);
6288 case IORING_OP_RENAMEAT:
6289 ret = io_renameat(req, issue_flags);
6291 case IORING_OP_UNLINKAT:
6292 ret = io_unlinkat(req, issue_flags);
6300 revert_creds(creds);
6303 /* If the op doesn't have a file, we're not polling for it */
6304 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6305 io_iopoll_req_issued(req);
6310 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6312 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6314 req = io_put_req_find_next(req);
6315 return req ? &req->work : NULL;
6318 static void io_wq_submit_work(struct io_wq_work *work)
6320 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6321 struct io_kiocb *timeout;
6324 /* one will be dropped by ->io_free_work() after returning to io-wq */
6325 if (!(req->flags & REQ_F_REFCOUNT))
6326 __io_req_set_refcount(req, 2);
6330 timeout = io_prep_linked_timeout(req);
6332 io_queue_linked_timeout(timeout);
6334 if (work->flags & IO_WQ_WORK_CANCEL)
6339 ret = io_issue_sqe(req, 0);
6341 * We can get EAGAIN for polled IO even though we're
6342 * forcing a sync submission from here, since we can't
6343 * wait for request slots on the block side.
6351 /* avoid locking problems by failing it from a clean context */
6353 io_req_task_queue_fail(req, ret);
6356 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6359 return &table->files[i];
6362 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6365 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6367 return (struct file *) (slot->file_ptr & FFS_MASK);
6370 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6372 unsigned long file_ptr = (unsigned long) file;
6374 if (__io_file_supports_nowait(file, READ))
6375 file_ptr |= FFS_ASYNC_READ;
6376 if (__io_file_supports_nowait(file, WRITE))
6377 file_ptr |= FFS_ASYNC_WRITE;
6378 if (S_ISREG(file_inode(file)->i_mode))
6379 file_ptr |= FFS_ISREG;
6380 file_slot->file_ptr = file_ptr;
6383 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6384 struct io_kiocb *req, int fd)
6387 unsigned long file_ptr;
6389 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6391 fd = array_index_nospec(fd, ctx->nr_user_files);
6392 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6393 file = (struct file *) (file_ptr & FFS_MASK);
6394 file_ptr &= ~FFS_MASK;
6395 /* mask in overlapping REQ_F and FFS bits */
6396 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6397 io_req_set_rsrc_node(req);
6401 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6402 struct io_kiocb *req, int fd)
6404 struct file *file = fget(fd);
6406 trace_io_uring_file_get(ctx, fd);
6408 /* we don't allow fixed io_uring files */
6409 if (file && unlikely(file->f_op == &io_uring_fops))
6410 io_req_track_inflight(req);
6414 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6415 struct io_kiocb *req, int fd, bool fixed)
6418 return io_file_get_fixed(ctx, req, fd);
6420 return io_file_get_normal(ctx, req, fd);
6423 static void io_req_task_link_timeout(struct io_kiocb *req)
6425 struct io_kiocb *prev = req->timeout.prev;
6429 ret = io_try_cancel_userdata(req, prev->user_data);
6430 io_req_complete_post(req, ret ?: -ETIME, 0);
6433 io_req_complete_post(req, -ETIME, 0);
6437 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6439 struct io_timeout_data *data = container_of(timer,
6440 struct io_timeout_data, timer);
6441 struct io_kiocb *prev, *req = data->req;
6442 struct io_ring_ctx *ctx = req->ctx;
6443 unsigned long flags;
6445 spin_lock_irqsave(&ctx->timeout_lock, flags);
6446 prev = req->timeout.head;
6447 req->timeout.head = NULL;
6450 * We don't expect the list to be empty, that will only happen if we
6451 * race with the completion of the linked work.
6454 io_remove_next_linked(prev);
6455 if (!req_ref_inc_not_zero(prev))
6458 req->timeout.prev = prev;
6459 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6461 req->io_task_work.func = io_req_task_link_timeout;
6462 io_req_task_work_add(req);
6463 return HRTIMER_NORESTART;
6466 static void io_queue_linked_timeout(struct io_kiocb *req)
6468 struct io_ring_ctx *ctx = req->ctx;
6470 spin_lock_irq(&ctx->timeout_lock);
6472 * If the back reference is NULL, then our linked request finished
6473 * before we got a chance to setup the timer
6475 if (req->timeout.head) {
6476 struct io_timeout_data *data = req->async_data;
6478 data->timer.function = io_link_timeout_fn;
6479 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6482 spin_unlock_irq(&ctx->timeout_lock);
6483 /* drop submission reference */
6487 static void __io_queue_sqe(struct io_kiocb *req)
6488 __must_hold(&req->ctx->uring_lock)
6490 struct io_kiocb *linked_timeout;
6494 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6497 * We async punt it if the file wasn't marked NOWAIT, or if the file
6498 * doesn't support non-blocking read/write attempts
6501 if (req->flags & REQ_F_COMPLETE_INLINE) {
6502 struct io_ring_ctx *ctx = req->ctx;
6503 struct io_submit_state *state = &ctx->submit_state;
6505 state->compl_reqs[state->compl_nr++] = req;
6506 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6507 io_submit_flush_completions(ctx);
6511 linked_timeout = io_prep_linked_timeout(req);
6513 io_queue_linked_timeout(linked_timeout);
6514 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6515 linked_timeout = io_prep_linked_timeout(req);
6517 switch (io_arm_poll_handler(req)) {
6518 case IO_APOLL_READY:
6520 io_unprep_linked_timeout(req);
6522 case IO_APOLL_ABORTED:
6524 * Queued up for async execution, worker will release
6525 * submit reference when the iocb is actually submitted.
6527 io_queue_async_work(req);
6532 io_queue_linked_timeout(linked_timeout);
6534 io_req_complete_failed(req, ret);
6538 static inline void io_queue_sqe(struct io_kiocb *req)
6539 __must_hold(&req->ctx->uring_lock)
6541 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6544 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6545 __io_queue_sqe(req);
6547 int ret = io_req_prep_async(req);
6550 io_req_complete_failed(req, ret);
6552 io_queue_async_work(req);
6557 * Check SQE restrictions (opcode and flags).
6559 * Returns 'true' if SQE is allowed, 'false' otherwise.
6561 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6562 struct io_kiocb *req,
6563 unsigned int sqe_flags)
6565 if (likely(!ctx->restricted))
6568 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6571 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6572 ctx->restrictions.sqe_flags_required)
6575 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6576 ctx->restrictions.sqe_flags_required))
6582 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6583 const struct io_uring_sqe *sqe)
6584 __must_hold(&ctx->uring_lock)
6586 struct io_submit_state *state;
6587 unsigned int sqe_flags;
6588 int personality, ret = 0;
6590 /* req is partially pre-initialised, see io_preinit_req() */
6591 req->opcode = READ_ONCE(sqe->opcode);
6592 /* same numerical values with corresponding REQ_F_*, safe to copy */
6593 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6594 req->user_data = READ_ONCE(sqe->user_data);
6596 req->fixed_rsrc_refs = NULL;
6597 req->task = current;
6599 /* enforce forwards compatibility on users */
6600 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6602 if (unlikely(req->opcode >= IORING_OP_LAST))
6604 if (!io_check_restriction(ctx, req, sqe_flags))
6607 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6608 !io_op_defs[req->opcode].buffer_select)
6610 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6611 ctx->drain_active = true;
6613 personality = READ_ONCE(sqe->personality);
6615 req->creds = xa_load(&ctx->personalities, personality);
6618 get_cred(req->creds);
6619 req->flags |= REQ_F_CREDS;
6621 state = &ctx->submit_state;
6624 * Plug now if we have more than 1 IO left after this, and the target
6625 * is potentially a read/write to block based storage.
6627 if (!state->plug_started && state->ios_left > 1 &&
6628 io_op_defs[req->opcode].plug) {
6629 blk_start_plug(&state->plug);
6630 state->plug_started = true;
6633 if (io_op_defs[req->opcode].needs_file) {
6634 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6635 (sqe_flags & IOSQE_FIXED_FILE));
6636 if (unlikely(!req->file))
6644 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6645 const struct io_uring_sqe *sqe)
6646 __must_hold(&ctx->uring_lock)
6648 struct io_submit_link *link = &ctx->submit_state.link;
6651 ret = io_init_req(ctx, req, sqe);
6652 if (unlikely(ret)) {
6655 /* fail even hard links since we don't submit */
6656 req_set_fail(link->head);
6657 io_req_complete_failed(link->head, -ECANCELED);
6660 io_req_complete_failed(req, ret);
6664 ret = io_req_prep(req, sqe);
6668 /* don't need @sqe from now on */
6669 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6671 ctx->flags & IORING_SETUP_SQPOLL);
6674 * If we already have a head request, queue this one for async
6675 * submittal once the head completes. If we don't have a head but
6676 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6677 * submitted sync once the chain is complete. If none of those
6678 * conditions are true (normal request), then just queue it.
6681 struct io_kiocb *head = link->head;
6683 ret = io_req_prep_async(req);
6686 trace_io_uring_link(ctx, req, head);
6687 link->last->link = req;
6690 /* last request of a link, enqueue the link */
6691 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6696 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6708 * Batched submission is done, ensure local IO is flushed out.
6710 static void io_submit_state_end(struct io_submit_state *state,
6711 struct io_ring_ctx *ctx)
6713 if (state->link.head)
6714 io_queue_sqe(state->link.head);
6715 if (state->compl_nr)
6716 io_submit_flush_completions(ctx);
6717 if (state->plug_started)
6718 blk_finish_plug(&state->plug);
6722 * Start submission side cache.
6724 static void io_submit_state_start(struct io_submit_state *state,
6725 unsigned int max_ios)
6727 state->plug_started = false;
6728 state->ios_left = max_ios;
6729 /* set only head, no need to init link_last in advance */
6730 state->link.head = NULL;
6733 static void io_commit_sqring(struct io_ring_ctx *ctx)
6735 struct io_rings *rings = ctx->rings;
6738 * Ensure any loads from the SQEs are done at this point,
6739 * since once we write the new head, the application could
6740 * write new data to them.
6742 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6746 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6747 * that is mapped by userspace. This means that care needs to be taken to
6748 * ensure that reads are stable, as we cannot rely on userspace always
6749 * being a good citizen. If members of the sqe are validated and then later
6750 * used, it's important that those reads are done through READ_ONCE() to
6751 * prevent a re-load down the line.
6753 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6755 unsigned head, mask = ctx->sq_entries - 1;
6756 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6759 * The cached sq head (or cq tail) serves two purposes:
6761 * 1) allows us to batch the cost of updating the user visible
6763 * 2) allows the kernel side to track the head on its own, even
6764 * though the application is the one updating it.
6766 head = READ_ONCE(ctx->sq_array[sq_idx]);
6767 if (likely(head < ctx->sq_entries))
6768 return &ctx->sq_sqes[head];
6770 /* drop invalid entries */
6772 WRITE_ONCE(ctx->rings->sq_dropped,
6773 READ_ONCE(ctx->rings->sq_dropped) + 1);
6777 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6778 __must_hold(&ctx->uring_lock)
6780 struct io_uring_task *tctx;
6783 /* make sure SQ entry isn't read before tail */
6784 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6785 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6788 tctx = current->io_uring;
6789 tctx->cached_refs -= nr;
6790 if (unlikely(tctx->cached_refs < 0)) {
6791 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6793 percpu_counter_add(&tctx->inflight, refill);
6794 refcount_add(refill, ¤t->usage);
6795 tctx->cached_refs += refill;
6797 io_submit_state_start(&ctx->submit_state, nr);
6799 while (submitted < nr) {
6800 const struct io_uring_sqe *sqe;
6801 struct io_kiocb *req;
6803 req = io_alloc_req(ctx);
6804 if (unlikely(!req)) {
6806 submitted = -EAGAIN;
6809 sqe = io_get_sqe(ctx);
6810 if (unlikely(!sqe)) {
6811 kmem_cache_free(req_cachep, req);
6814 /* will complete beyond this point, count as submitted */
6816 if (io_submit_sqe(ctx, req, sqe))
6820 if (unlikely(submitted != nr)) {
6821 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6822 int unused = nr - ref_used;
6824 current->io_uring->cached_refs += unused;
6825 percpu_ref_put_many(&ctx->refs, unused);
6828 io_submit_state_end(&ctx->submit_state, ctx);
6829 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6830 io_commit_sqring(ctx);
6835 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6837 return READ_ONCE(sqd->state);
6840 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6842 /* Tell userspace we may need a wakeup call */
6843 spin_lock(&ctx->completion_lock);
6844 WRITE_ONCE(ctx->rings->sq_flags,
6845 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6846 spin_unlock(&ctx->completion_lock);
6849 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6851 spin_lock(&ctx->completion_lock);
6852 WRITE_ONCE(ctx->rings->sq_flags,
6853 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6854 spin_unlock(&ctx->completion_lock);
6857 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6859 unsigned int to_submit;
6862 to_submit = io_sqring_entries(ctx);
6863 /* if we're handling multiple rings, cap submit size for fairness */
6864 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6865 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6867 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6868 unsigned nr_events = 0;
6869 const struct cred *creds = NULL;
6871 if (ctx->sq_creds != current_cred())
6872 creds = override_creds(ctx->sq_creds);
6874 mutex_lock(&ctx->uring_lock);
6875 if (!list_empty(&ctx->iopoll_list))
6876 io_do_iopoll(ctx, &nr_events, 0);
6879 * Don't submit if refs are dying, good for io_uring_register(),
6880 * but also it is relied upon by io_ring_exit_work()
6882 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6883 !(ctx->flags & IORING_SETUP_R_DISABLED))
6884 ret = io_submit_sqes(ctx, to_submit);
6885 mutex_unlock(&ctx->uring_lock);
6887 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6888 wake_up(&ctx->sqo_sq_wait);
6890 revert_creds(creds);
6896 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6898 struct io_ring_ctx *ctx;
6899 unsigned sq_thread_idle = 0;
6901 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6902 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6903 sqd->sq_thread_idle = sq_thread_idle;
6906 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6908 bool did_sig = false;
6909 struct ksignal ksig;
6911 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6912 signal_pending(current)) {
6913 mutex_unlock(&sqd->lock);
6914 if (signal_pending(current))
6915 did_sig = get_signal(&ksig);
6917 mutex_lock(&sqd->lock);
6919 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6922 static int io_sq_thread(void *data)
6924 struct io_sq_data *sqd = data;
6925 struct io_ring_ctx *ctx;
6926 unsigned long timeout = 0;
6927 char buf[TASK_COMM_LEN];
6930 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6931 set_task_comm(current, buf);
6933 if (sqd->sq_cpu != -1)
6934 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6936 set_cpus_allowed_ptr(current, cpu_online_mask);
6937 current->flags |= PF_NO_SETAFFINITY;
6939 mutex_lock(&sqd->lock);
6941 bool cap_entries, sqt_spin = false;
6943 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6944 if (io_sqd_handle_event(sqd))
6946 timeout = jiffies + sqd->sq_thread_idle;
6949 cap_entries = !list_is_singular(&sqd->ctx_list);
6950 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6951 int ret = __io_sq_thread(ctx, cap_entries);
6953 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6956 if (io_run_task_work())
6959 if (sqt_spin || !time_after(jiffies, timeout)) {
6962 timeout = jiffies + sqd->sq_thread_idle;
6966 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6967 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6968 bool needs_sched = true;
6970 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6971 io_ring_set_wakeup_flag(ctx);
6973 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6974 !list_empty_careful(&ctx->iopoll_list)) {
6975 needs_sched = false;
6978 if (io_sqring_entries(ctx)) {
6979 needs_sched = false;
6985 mutex_unlock(&sqd->lock);
6987 mutex_lock(&sqd->lock);
6989 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6990 io_ring_clear_wakeup_flag(ctx);
6993 finish_wait(&sqd->wait, &wait);
6994 timeout = jiffies + sqd->sq_thread_idle;
6997 io_uring_cancel_generic(true, sqd);
6999 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7000 io_ring_set_wakeup_flag(ctx);
7002 mutex_unlock(&sqd->lock);
7004 complete(&sqd->exited);
7008 struct io_wait_queue {
7009 struct wait_queue_entry wq;
7010 struct io_ring_ctx *ctx;
7012 unsigned nr_timeouts;
7015 static inline bool io_should_wake(struct io_wait_queue *iowq)
7017 struct io_ring_ctx *ctx = iowq->ctx;
7018 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7021 * Wake up if we have enough events, or if a timeout occurred since we
7022 * started waiting. For timeouts, we always want to return to userspace,
7023 * regardless of event count.
7025 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7028 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7029 int wake_flags, void *key)
7031 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7035 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7036 * the task, and the next invocation will do it.
7038 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7039 return autoremove_wake_function(curr, mode, wake_flags, key);
7043 static int io_run_task_work_sig(void)
7045 if (io_run_task_work())
7047 if (!signal_pending(current))
7049 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7050 return -ERESTARTSYS;
7054 /* when returns >0, the caller should retry */
7055 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7056 struct io_wait_queue *iowq,
7057 signed long *timeout)
7061 /* make sure we run task_work before checking for signals */
7062 ret = io_run_task_work_sig();
7063 if (ret || io_should_wake(iowq))
7065 /* let the caller flush overflows, retry */
7066 if (test_bit(0, &ctx->check_cq_overflow))
7069 *timeout = schedule_timeout(*timeout);
7070 return !*timeout ? -ETIME : 1;
7074 * Wait until events become available, if we don't already have some. The
7075 * application must reap them itself, as they reside on the shared cq ring.
7077 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7078 const sigset_t __user *sig, size_t sigsz,
7079 struct __kernel_timespec __user *uts)
7081 struct io_wait_queue iowq;
7082 struct io_rings *rings = ctx->rings;
7083 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7087 io_cqring_overflow_flush(ctx);
7088 if (io_cqring_events(ctx) >= min_events)
7090 if (!io_run_task_work())
7095 #ifdef CONFIG_COMPAT
7096 if (in_compat_syscall())
7097 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7101 ret = set_user_sigmask(sig, sigsz);
7108 struct timespec64 ts;
7110 if (get_timespec64(&ts, uts))
7112 timeout = timespec64_to_jiffies(&ts);
7115 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7116 iowq.wq.private = current;
7117 INIT_LIST_HEAD(&iowq.wq.entry);
7119 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7120 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7122 trace_io_uring_cqring_wait(ctx, min_events);
7124 /* if we can't even flush overflow, don't wait for more */
7125 if (!io_cqring_overflow_flush(ctx)) {
7129 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7130 TASK_INTERRUPTIBLE);
7131 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7132 finish_wait(&ctx->cq_wait, &iowq.wq);
7136 restore_saved_sigmask_unless(ret == -EINTR);
7138 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7141 static void io_free_page_table(void **table, size_t size)
7143 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7145 for (i = 0; i < nr_tables; i++)
7150 static void **io_alloc_page_table(size_t size)
7152 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7153 size_t init_size = size;
7156 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7160 for (i = 0; i < nr_tables; i++) {
7161 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7163 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7165 io_free_page_table(table, init_size);
7173 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7175 percpu_ref_exit(&ref_node->refs);
7179 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7181 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7182 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7183 unsigned long flags;
7184 bool first_add = false;
7186 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7189 while (!list_empty(&ctx->rsrc_ref_list)) {
7190 node = list_first_entry(&ctx->rsrc_ref_list,
7191 struct io_rsrc_node, node);
7192 /* recycle ref nodes in order */
7195 list_del(&node->node);
7196 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7198 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7201 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7204 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7206 struct io_rsrc_node *ref_node;
7208 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7212 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7217 INIT_LIST_HEAD(&ref_node->node);
7218 INIT_LIST_HEAD(&ref_node->rsrc_list);
7219 ref_node->done = false;
7223 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7224 struct io_rsrc_data *data_to_kill)
7226 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7227 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7230 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7232 rsrc_node->rsrc_data = data_to_kill;
7233 spin_lock_irq(&ctx->rsrc_ref_lock);
7234 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7235 spin_unlock_irq(&ctx->rsrc_ref_lock);
7237 atomic_inc(&data_to_kill->refs);
7238 percpu_ref_kill(&rsrc_node->refs);
7239 ctx->rsrc_node = NULL;
7242 if (!ctx->rsrc_node) {
7243 ctx->rsrc_node = ctx->rsrc_backup_node;
7244 ctx->rsrc_backup_node = NULL;
7248 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7250 if (ctx->rsrc_backup_node)
7252 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7253 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7256 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7260 /* As we may drop ->uring_lock, other task may have started quiesce */
7264 data->quiesce = true;
7266 ret = io_rsrc_node_switch_start(ctx);
7269 io_rsrc_node_switch(ctx, data);
7271 /* kill initial ref, already quiesced if zero */
7272 if (atomic_dec_and_test(&data->refs))
7274 mutex_unlock(&ctx->uring_lock);
7275 flush_delayed_work(&ctx->rsrc_put_work);
7276 ret = wait_for_completion_interruptible(&data->done);
7278 mutex_lock(&ctx->uring_lock);
7282 atomic_inc(&data->refs);
7283 /* wait for all works potentially completing data->done */
7284 flush_delayed_work(&ctx->rsrc_put_work);
7285 reinit_completion(&data->done);
7287 ret = io_run_task_work_sig();
7288 mutex_lock(&ctx->uring_lock);
7290 data->quiesce = false;
7295 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7297 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7298 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7300 return &data->tags[table_idx][off];
7303 static void io_rsrc_data_free(struct io_rsrc_data *data)
7305 size_t size = data->nr * sizeof(data->tags[0][0]);
7308 io_free_page_table((void **)data->tags, size);
7312 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7313 u64 __user *utags, unsigned nr,
7314 struct io_rsrc_data **pdata)
7316 struct io_rsrc_data *data;
7320 data = kzalloc(sizeof(*data), GFP_KERNEL);
7323 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7331 data->do_put = do_put;
7334 for (i = 0; i < nr; i++) {
7335 u64 *tag_slot = io_get_tag_slot(data, i);
7337 if (copy_from_user(tag_slot, &utags[i],
7343 atomic_set(&data->refs, 1);
7344 init_completion(&data->done);
7348 io_rsrc_data_free(data);
7352 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7354 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7355 GFP_KERNEL_ACCOUNT);
7356 return !!table->files;
7359 static void io_free_file_tables(struct io_file_table *table)
7361 kvfree(table->files);
7362 table->files = NULL;
7365 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7367 #if defined(CONFIG_UNIX)
7368 if (ctx->ring_sock) {
7369 struct sock *sock = ctx->ring_sock->sk;
7370 struct sk_buff *skb;
7372 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7378 for (i = 0; i < ctx->nr_user_files; i++) {
7381 file = io_file_from_index(ctx, i);
7386 io_free_file_tables(&ctx->file_table);
7387 io_rsrc_data_free(ctx->file_data);
7388 ctx->file_data = NULL;
7389 ctx->nr_user_files = 0;
7392 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7396 if (!ctx->file_data)
7398 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7400 __io_sqe_files_unregister(ctx);
7404 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7405 __releases(&sqd->lock)
7407 WARN_ON_ONCE(sqd->thread == current);
7410 * Do the dance but not conditional clear_bit() because it'd race with
7411 * other threads incrementing park_pending and setting the bit.
7413 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7414 if (atomic_dec_return(&sqd->park_pending))
7415 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7416 mutex_unlock(&sqd->lock);
7419 static void io_sq_thread_park(struct io_sq_data *sqd)
7420 __acquires(&sqd->lock)
7422 WARN_ON_ONCE(sqd->thread == current);
7424 atomic_inc(&sqd->park_pending);
7425 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7426 mutex_lock(&sqd->lock);
7428 wake_up_process(sqd->thread);
7431 static void io_sq_thread_stop(struct io_sq_data *sqd)
7433 WARN_ON_ONCE(sqd->thread == current);
7434 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7436 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7437 mutex_lock(&sqd->lock);
7439 wake_up_process(sqd->thread);
7440 mutex_unlock(&sqd->lock);
7441 wait_for_completion(&sqd->exited);
7444 static void io_put_sq_data(struct io_sq_data *sqd)
7446 if (refcount_dec_and_test(&sqd->refs)) {
7447 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7449 io_sq_thread_stop(sqd);
7454 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7456 struct io_sq_data *sqd = ctx->sq_data;
7459 io_sq_thread_park(sqd);
7460 list_del_init(&ctx->sqd_list);
7461 io_sqd_update_thread_idle(sqd);
7462 io_sq_thread_unpark(sqd);
7464 io_put_sq_data(sqd);
7465 ctx->sq_data = NULL;
7469 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7471 struct io_ring_ctx *ctx_attach;
7472 struct io_sq_data *sqd;
7475 f = fdget(p->wq_fd);
7477 return ERR_PTR(-ENXIO);
7478 if (f.file->f_op != &io_uring_fops) {
7480 return ERR_PTR(-EINVAL);
7483 ctx_attach = f.file->private_data;
7484 sqd = ctx_attach->sq_data;
7487 return ERR_PTR(-EINVAL);
7489 if (sqd->task_tgid != current->tgid) {
7491 return ERR_PTR(-EPERM);
7494 refcount_inc(&sqd->refs);
7499 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7502 struct io_sq_data *sqd;
7505 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7506 sqd = io_attach_sq_data(p);
7511 /* fall through for EPERM case, setup new sqd/task */
7512 if (PTR_ERR(sqd) != -EPERM)
7516 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7518 return ERR_PTR(-ENOMEM);
7520 atomic_set(&sqd->park_pending, 0);
7521 refcount_set(&sqd->refs, 1);
7522 INIT_LIST_HEAD(&sqd->ctx_list);
7523 mutex_init(&sqd->lock);
7524 init_waitqueue_head(&sqd->wait);
7525 init_completion(&sqd->exited);
7529 #if defined(CONFIG_UNIX)
7531 * Ensure the UNIX gc is aware of our file set, so we are certain that
7532 * the io_uring can be safely unregistered on process exit, even if we have
7533 * loops in the file referencing.
7535 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7537 struct sock *sk = ctx->ring_sock->sk;
7538 struct scm_fp_list *fpl;
7539 struct sk_buff *skb;
7542 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7546 skb = alloc_skb(0, GFP_KERNEL);
7555 fpl->user = get_uid(current_user());
7556 for (i = 0; i < nr; i++) {
7557 struct file *file = io_file_from_index(ctx, i + offset);
7561 fpl->fp[nr_files] = get_file(file);
7562 unix_inflight(fpl->user, fpl->fp[nr_files]);
7567 fpl->max = SCM_MAX_FD;
7568 fpl->count = nr_files;
7569 UNIXCB(skb).fp = fpl;
7570 skb->destructor = unix_destruct_scm;
7571 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7572 skb_queue_head(&sk->sk_receive_queue, skb);
7574 for (i = 0; i < nr_files; i++)
7585 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7586 * causes regular reference counting to break down. We rely on the UNIX
7587 * garbage collection to take care of this problem for us.
7589 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7591 unsigned left, total;
7595 left = ctx->nr_user_files;
7597 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7599 ret = __io_sqe_files_scm(ctx, this_files, total);
7603 total += this_files;
7609 while (total < ctx->nr_user_files) {
7610 struct file *file = io_file_from_index(ctx, total);
7620 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7626 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7628 struct file *file = prsrc->file;
7629 #if defined(CONFIG_UNIX)
7630 struct sock *sock = ctx->ring_sock->sk;
7631 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7632 struct sk_buff *skb;
7635 __skb_queue_head_init(&list);
7638 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7639 * remove this entry and rearrange the file array.
7641 skb = skb_dequeue(head);
7643 struct scm_fp_list *fp;
7645 fp = UNIXCB(skb).fp;
7646 for (i = 0; i < fp->count; i++) {
7649 if (fp->fp[i] != file)
7652 unix_notinflight(fp->user, fp->fp[i]);
7653 left = fp->count - 1 - i;
7655 memmove(&fp->fp[i], &fp->fp[i + 1],
7656 left * sizeof(struct file *));
7663 __skb_queue_tail(&list, skb);
7673 __skb_queue_tail(&list, skb);
7675 skb = skb_dequeue(head);
7678 if (skb_peek(&list)) {
7679 spin_lock_irq(&head->lock);
7680 while ((skb = __skb_dequeue(&list)) != NULL)
7681 __skb_queue_tail(head, skb);
7682 spin_unlock_irq(&head->lock);
7689 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7691 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7692 struct io_ring_ctx *ctx = rsrc_data->ctx;
7693 struct io_rsrc_put *prsrc, *tmp;
7695 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7696 list_del(&prsrc->list);
7699 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7701 io_ring_submit_lock(ctx, lock_ring);
7702 spin_lock(&ctx->completion_lock);
7703 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7705 io_commit_cqring(ctx);
7706 spin_unlock(&ctx->completion_lock);
7707 io_cqring_ev_posted(ctx);
7708 io_ring_submit_unlock(ctx, lock_ring);
7711 rsrc_data->do_put(ctx, prsrc);
7715 io_rsrc_node_destroy(ref_node);
7716 if (atomic_dec_and_test(&rsrc_data->refs))
7717 complete(&rsrc_data->done);
7720 static void io_rsrc_put_work(struct work_struct *work)
7722 struct io_ring_ctx *ctx;
7723 struct llist_node *node;
7725 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7726 node = llist_del_all(&ctx->rsrc_put_llist);
7729 struct io_rsrc_node *ref_node;
7730 struct llist_node *next = node->next;
7732 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7733 __io_rsrc_put_work(ref_node);
7738 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7739 unsigned nr_args, u64 __user *tags)
7741 __s32 __user *fds = (__s32 __user *) arg;
7750 if (nr_args > IORING_MAX_FIXED_FILES)
7752 if (nr_args > rlimit(RLIMIT_NOFILE))
7754 ret = io_rsrc_node_switch_start(ctx);
7757 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7763 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7766 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7767 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7771 /* allow sparse sets */
7774 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7781 if (unlikely(!file))
7785 * Don't allow io_uring instances to be registered. If UNIX
7786 * isn't enabled, then this causes a reference cycle and this
7787 * instance can never get freed. If UNIX is enabled we'll
7788 * handle it just fine, but there's still no point in allowing
7789 * a ring fd as it doesn't support regular read/write anyway.
7791 if (file->f_op == &io_uring_fops) {
7795 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7798 ret = io_sqe_files_scm(ctx);
7800 __io_sqe_files_unregister(ctx);
7804 io_rsrc_node_switch(ctx, NULL);
7807 for (i = 0; i < ctx->nr_user_files; i++) {
7808 file = io_file_from_index(ctx, i);
7812 io_free_file_tables(&ctx->file_table);
7813 ctx->nr_user_files = 0;
7815 io_rsrc_data_free(ctx->file_data);
7816 ctx->file_data = NULL;
7820 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7823 #if defined(CONFIG_UNIX)
7824 struct sock *sock = ctx->ring_sock->sk;
7825 struct sk_buff_head *head = &sock->sk_receive_queue;
7826 struct sk_buff *skb;
7829 * See if we can merge this file into an existing skb SCM_RIGHTS
7830 * file set. If there's no room, fall back to allocating a new skb
7831 * and filling it in.
7833 spin_lock_irq(&head->lock);
7834 skb = skb_peek(head);
7836 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7838 if (fpl->count < SCM_MAX_FD) {
7839 __skb_unlink(skb, head);
7840 spin_unlock_irq(&head->lock);
7841 fpl->fp[fpl->count] = get_file(file);
7842 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7844 spin_lock_irq(&head->lock);
7845 __skb_queue_head(head, skb);
7850 spin_unlock_irq(&head->lock);
7857 return __io_sqe_files_scm(ctx, 1, index);
7863 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7864 struct io_rsrc_node *node, void *rsrc)
7866 struct io_rsrc_put *prsrc;
7868 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7872 prsrc->tag = *io_get_tag_slot(data, idx);
7874 list_add(&prsrc->list, &node->rsrc_list);
7878 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7879 struct io_uring_rsrc_update2 *up,
7882 u64 __user *tags = u64_to_user_ptr(up->tags);
7883 __s32 __user *fds = u64_to_user_ptr(up->data);
7884 struct io_rsrc_data *data = ctx->file_data;
7885 struct io_fixed_file *file_slot;
7889 bool needs_switch = false;
7891 if (!ctx->file_data)
7893 if (up->offset + nr_args > ctx->nr_user_files)
7896 for (done = 0; done < nr_args; done++) {
7899 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7900 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7904 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7908 if (fd == IORING_REGISTER_FILES_SKIP)
7911 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7912 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7914 if (file_slot->file_ptr) {
7915 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7916 err = io_queue_rsrc_removal(data, up->offset + done,
7917 ctx->rsrc_node, file);
7920 file_slot->file_ptr = 0;
7921 needs_switch = true;
7930 * Don't allow io_uring instances to be registered. If
7931 * UNIX isn't enabled, then this causes a reference
7932 * cycle and this instance can never get freed. If UNIX
7933 * is enabled we'll handle it just fine, but there's
7934 * still no point in allowing a ring fd as it doesn't
7935 * support regular read/write anyway.
7937 if (file->f_op == &io_uring_fops) {
7942 *io_get_tag_slot(data, up->offset + done) = tag;
7943 io_fixed_file_set(file_slot, file);
7944 err = io_sqe_file_register(ctx, file, i);
7946 file_slot->file_ptr = 0;
7954 io_rsrc_node_switch(ctx, data);
7955 return done ? done : err;
7958 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7959 struct task_struct *task)
7961 struct io_wq_hash *hash;
7962 struct io_wq_data data;
7963 unsigned int concurrency;
7965 mutex_lock(&ctx->uring_lock);
7966 hash = ctx->hash_map;
7968 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7970 mutex_unlock(&ctx->uring_lock);
7971 return ERR_PTR(-ENOMEM);
7973 refcount_set(&hash->refs, 1);
7974 init_waitqueue_head(&hash->wait);
7975 ctx->hash_map = hash;
7977 mutex_unlock(&ctx->uring_lock);
7981 data.free_work = io_wq_free_work;
7982 data.do_work = io_wq_submit_work;
7984 /* Do QD, or 4 * CPUS, whatever is smallest */
7985 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7987 return io_wq_create(concurrency, &data);
7990 static int io_uring_alloc_task_context(struct task_struct *task,
7991 struct io_ring_ctx *ctx)
7993 struct io_uring_task *tctx;
7996 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7997 if (unlikely(!tctx))
8000 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8001 if (unlikely(ret)) {
8006 tctx->io_wq = io_init_wq_offload(ctx, task);
8007 if (IS_ERR(tctx->io_wq)) {
8008 ret = PTR_ERR(tctx->io_wq);
8009 percpu_counter_destroy(&tctx->inflight);
8015 init_waitqueue_head(&tctx->wait);
8016 atomic_set(&tctx->in_idle, 0);
8017 atomic_set(&tctx->inflight_tracked, 0);
8018 task->io_uring = tctx;
8019 spin_lock_init(&tctx->task_lock);
8020 INIT_WQ_LIST(&tctx->task_list);
8021 init_task_work(&tctx->task_work, tctx_task_work);
8025 void __io_uring_free(struct task_struct *tsk)
8027 struct io_uring_task *tctx = tsk->io_uring;
8029 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8030 WARN_ON_ONCE(tctx->io_wq);
8031 WARN_ON_ONCE(tctx->cached_refs);
8033 percpu_counter_destroy(&tctx->inflight);
8035 tsk->io_uring = NULL;
8038 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8039 struct io_uring_params *p)
8043 /* Retain compatibility with failing for an invalid attach attempt */
8044 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8045 IORING_SETUP_ATTACH_WQ) {
8048 f = fdget(p->wq_fd);
8051 if (f.file->f_op != &io_uring_fops) {
8057 if (ctx->flags & IORING_SETUP_SQPOLL) {
8058 struct task_struct *tsk;
8059 struct io_sq_data *sqd;
8062 sqd = io_get_sq_data(p, &attached);
8068 ctx->sq_creds = get_current_cred();
8070 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8071 if (!ctx->sq_thread_idle)
8072 ctx->sq_thread_idle = HZ;
8074 io_sq_thread_park(sqd);
8075 list_add(&ctx->sqd_list, &sqd->ctx_list);
8076 io_sqd_update_thread_idle(sqd);
8077 /* don't attach to a dying SQPOLL thread, would be racy */
8078 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8079 io_sq_thread_unpark(sqd);
8086 if (p->flags & IORING_SETUP_SQ_AFF) {
8087 int cpu = p->sq_thread_cpu;
8090 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8097 sqd->task_pid = current->pid;
8098 sqd->task_tgid = current->tgid;
8099 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8106 ret = io_uring_alloc_task_context(tsk, ctx);
8107 wake_up_new_task(tsk);
8110 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8111 /* Can't have SQ_AFF without SQPOLL */
8118 complete(&ctx->sq_data->exited);
8120 io_sq_thread_finish(ctx);
8124 static inline void __io_unaccount_mem(struct user_struct *user,
8125 unsigned long nr_pages)
8127 atomic_long_sub(nr_pages, &user->locked_vm);
8130 static inline int __io_account_mem(struct user_struct *user,
8131 unsigned long nr_pages)
8133 unsigned long page_limit, cur_pages, new_pages;
8135 /* Don't allow more pages than we can safely lock */
8136 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8139 cur_pages = atomic_long_read(&user->locked_vm);
8140 new_pages = cur_pages + nr_pages;
8141 if (new_pages > page_limit)
8143 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8144 new_pages) != cur_pages);
8149 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8152 __io_unaccount_mem(ctx->user, nr_pages);
8154 if (ctx->mm_account)
8155 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8158 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8163 ret = __io_account_mem(ctx->user, nr_pages);
8168 if (ctx->mm_account)
8169 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8174 static void io_mem_free(void *ptr)
8181 page = virt_to_head_page(ptr);
8182 if (put_page_testzero(page))
8183 free_compound_page(page);
8186 static void *io_mem_alloc(size_t size)
8188 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8189 __GFP_NORETRY | __GFP_ACCOUNT;
8191 return (void *) __get_free_pages(gfp_flags, get_order(size));
8194 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8197 struct io_rings *rings;
8198 size_t off, sq_array_size;
8200 off = struct_size(rings, cqes, cq_entries);
8201 if (off == SIZE_MAX)
8205 off = ALIGN(off, SMP_CACHE_BYTES);
8213 sq_array_size = array_size(sizeof(u32), sq_entries);
8214 if (sq_array_size == SIZE_MAX)
8217 if (check_add_overflow(off, sq_array_size, &off))
8223 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8225 struct io_mapped_ubuf *imu = *slot;
8228 if (imu != ctx->dummy_ubuf) {
8229 for (i = 0; i < imu->nr_bvecs; i++)
8230 unpin_user_page(imu->bvec[i].bv_page);
8231 if (imu->acct_pages)
8232 io_unaccount_mem(ctx, imu->acct_pages);
8238 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8240 io_buffer_unmap(ctx, &prsrc->buf);
8244 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8248 for (i = 0; i < ctx->nr_user_bufs; i++)
8249 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8250 kfree(ctx->user_bufs);
8251 io_rsrc_data_free(ctx->buf_data);
8252 ctx->user_bufs = NULL;
8253 ctx->buf_data = NULL;
8254 ctx->nr_user_bufs = 0;
8257 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8264 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8266 __io_sqe_buffers_unregister(ctx);
8270 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8271 void __user *arg, unsigned index)
8273 struct iovec __user *src;
8275 #ifdef CONFIG_COMPAT
8277 struct compat_iovec __user *ciovs;
8278 struct compat_iovec ciov;
8280 ciovs = (struct compat_iovec __user *) arg;
8281 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8284 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8285 dst->iov_len = ciov.iov_len;
8289 src = (struct iovec __user *) arg;
8290 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8296 * Not super efficient, but this is just a registration time. And we do cache
8297 * the last compound head, so generally we'll only do a full search if we don't
8300 * We check if the given compound head page has already been accounted, to
8301 * avoid double accounting it. This allows us to account the full size of the
8302 * page, not just the constituent pages of a huge page.
8304 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8305 int nr_pages, struct page *hpage)
8309 /* check current page array */
8310 for (i = 0; i < nr_pages; i++) {
8311 if (!PageCompound(pages[i]))
8313 if (compound_head(pages[i]) == hpage)
8317 /* check previously registered pages */
8318 for (i = 0; i < ctx->nr_user_bufs; i++) {
8319 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8321 for (j = 0; j < imu->nr_bvecs; j++) {
8322 if (!PageCompound(imu->bvec[j].bv_page))
8324 if (compound_head(imu->bvec[j].bv_page) == hpage)
8332 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8333 int nr_pages, struct io_mapped_ubuf *imu,
8334 struct page **last_hpage)
8338 imu->acct_pages = 0;
8339 for (i = 0; i < nr_pages; i++) {
8340 if (!PageCompound(pages[i])) {
8345 hpage = compound_head(pages[i]);
8346 if (hpage == *last_hpage)
8348 *last_hpage = hpage;
8349 if (headpage_already_acct(ctx, pages, i, hpage))
8351 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8355 if (!imu->acct_pages)
8358 ret = io_account_mem(ctx, imu->acct_pages);
8360 imu->acct_pages = 0;
8364 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8365 struct io_mapped_ubuf **pimu,
8366 struct page **last_hpage)
8368 struct io_mapped_ubuf *imu = NULL;
8369 struct vm_area_struct **vmas = NULL;
8370 struct page **pages = NULL;
8371 unsigned long off, start, end, ubuf;
8373 int ret, pret, nr_pages, i;
8375 if (!iov->iov_base) {
8376 *pimu = ctx->dummy_ubuf;
8380 ubuf = (unsigned long) iov->iov_base;
8381 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8382 start = ubuf >> PAGE_SHIFT;
8383 nr_pages = end - start;
8388 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8392 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8397 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8402 mmap_read_lock(current->mm);
8403 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8405 if (pret == nr_pages) {
8406 /* don't support file backed memory */
8407 for (i = 0; i < nr_pages; i++) {
8408 struct vm_area_struct *vma = vmas[i];
8410 if (vma_is_shmem(vma))
8413 !is_file_hugepages(vma->vm_file)) {
8419 ret = pret < 0 ? pret : -EFAULT;
8421 mmap_read_unlock(current->mm);
8424 * if we did partial map, or found file backed vmas,
8425 * release any pages we did get
8428 unpin_user_pages(pages, pret);
8432 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8434 unpin_user_pages(pages, pret);
8438 off = ubuf & ~PAGE_MASK;
8439 size = iov->iov_len;
8440 for (i = 0; i < nr_pages; i++) {
8443 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8444 imu->bvec[i].bv_page = pages[i];
8445 imu->bvec[i].bv_len = vec_len;
8446 imu->bvec[i].bv_offset = off;
8450 /* store original address for later verification */
8452 imu->ubuf_end = ubuf + iov->iov_len;
8453 imu->nr_bvecs = nr_pages;
8464 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8466 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8467 return ctx->user_bufs ? 0 : -ENOMEM;
8470 static int io_buffer_validate(struct iovec *iov)
8472 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8475 * Don't impose further limits on the size and buffer
8476 * constraints here, we'll -EINVAL later when IO is
8477 * submitted if they are wrong.
8480 return iov->iov_len ? -EFAULT : 0;
8484 /* arbitrary limit, but we need something */
8485 if (iov->iov_len > SZ_1G)
8488 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8494 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8495 unsigned int nr_args, u64 __user *tags)
8497 struct page *last_hpage = NULL;
8498 struct io_rsrc_data *data;
8504 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8506 ret = io_rsrc_node_switch_start(ctx);
8509 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8512 ret = io_buffers_map_alloc(ctx, nr_args);
8514 io_rsrc_data_free(data);
8518 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8519 ret = io_copy_iov(ctx, &iov, arg, i);
8522 ret = io_buffer_validate(&iov);
8525 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8530 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8536 WARN_ON_ONCE(ctx->buf_data);
8538 ctx->buf_data = data;
8540 __io_sqe_buffers_unregister(ctx);
8542 io_rsrc_node_switch(ctx, NULL);
8546 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8547 struct io_uring_rsrc_update2 *up,
8548 unsigned int nr_args)
8550 u64 __user *tags = u64_to_user_ptr(up->tags);
8551 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8552 struct page *last_hpage = NULL;
8553 bool needs_switch = false;
8559 if (up->offset + nr_args > ctx->nr_user_bufs)
8562 for (done = 0; done < nr_args; done++) {
8563 struct io_mapped_ubuf *imu;
8564 int offset = up->offset + done;
8567 err = io_copy_iov(ctx, &iov, iovs, done);
8570 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8574 err = io_buffer_validate(&iov);
8577 if (!iov.iov_base && tag) {
8581 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8585 i = array_index_nospec(offset, ctx->nr_user_bufs);
8586 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8587 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8588 ctx->rsrc_node, ctx->user_bufs[i]);
8589 if (unlikely(err)) {
8590 io_buffer_unmap(ctx, &imu);
8593 ctx->user_bufs[i] = NULL;
8594 needs_switch = true;
8597 ctx->user_bufs[i] = imu;
8598 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8602 io_rsrc_node_switch(ctx, ctx->buf_data);
8603 return done ? done : err;
8606 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8608 __s32 __user *fds = arg;
8614 if (copy_from_user(&fd, fds, sizeof(*fds)))
8617 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8618 if (IS_ERR(ctx->cq_ev_fd)) {
8619 int ret = PTR_ERR(ctx->cq_ev_fd);
8621 ctx->cq_ev_fd = NULL;
8628 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8630 if (ctx->cq_ev_fd) {
8631 eventfd_ctx_put(ctx->cq_ev_fd);
8632 ctx->cq_ev_fd = NULL;
8639 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8641 struct io_buffer *buf;
8642 unsigned long index;
8644 xa_for_each(&ctx->io_buffers, index, buf)
8645 __io_remove_buffers(ctx, buf, index, -1U);
8648 static void io_req_cache_free(struct list_head *list)
8650 struct io_kiocb *req, *nxt;
8652 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8653 list_del(&req->inflight_entry);
8654 kmem_cache_free(req_cachep, req);
8658 static void io_req_caches_free(struct io_ring_ctx *ctx)
8660 struct io_submit_state *state = &ctx->submit_state;
8662 mutex_lock(&ctx->uring_lock);
8664 if (state->free_reqs) {
8665 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8666 state->free_reqs = 0;
8669 io_flush_cached_locked_reqs(ctx, state);
8670 io_req_cache_free(&state->free_list);
8671 mutex_unlock(&ctx->uring_lock);
8674 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8676 if (data && !atomic_dec_and_test(&data->refs))
8677 wait_for_completion(&data->done);
8680 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8682 io_sq_thread_finish(ctx);
8684 if (ctx->mm_account) {
8685 mmdrop(ctx->mm_account);
8686 ctx->mm_account = NULL;
8689 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8690 io_wait_rsrc_data(ctx->buf_data);
8691 io_wait_rsrc_data(ctx->file_data);
8693 mutex_lock(&ctx->uring_lock);
8695 __io_sqe_buffers_unregister(ctx);
8697 __io_sqe_files_unregister(ctx);
8699 __io_cqring_overflow_flush(ctx, true);
8700 mutex_unlock(&ctx->uring_lock);
8701 io_eventfd_unregister(ctx);
8702 io_destroy_buffers(ctx);
8704 put_cred(ctx->sq_creds);
8706 /* there are no registered resources left, nobody uses it */
8708 io_rsrc_node_destroy(ctx->rsrc_node);
8709 if (ctx->rsrc_backup_node)
8710 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8711 flush_delayed_work(&ctx->rsrc_put_work);
8713 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8714 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8716 #if defined(CONFIG_UNIX)
8717 if (ctx->ring_sock) {
8718 ctx->ring_sock->file = NULL; /* so that iput() is called */
8719 sock_release(ctx->ring_sock);
8723 io_mem_free(ctx->rings);
8724 io_mem_free(ctx->sq_sqes);
8726 percpu_ref_exit(&ctx->refs);
8727 free_uid(ctx->user);
8728 io_req_caches_free(ctx);
8730 io_wq_put_hash(ctx->hash_map);
8731 kfree(ctx->cancel_hash);
8732 kfree(ctx->dummy_ubuf);
8736 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8738 struct io_ring_ctx *ctx = file->private_data;
8741 poll_wait(file, &ctx->poll_wait, wait);
8743 * synchronizes with barrier from wq_has_sleeper call in
8747 if (!io_sqring_full(ctx))
8748 mask |= EPOLLOUT | EPOLLWRNORM;
8751 * Don't flush cqring overflow list here, just do a simple check.
8752 * Otherwise there could possible be ABBA deadlock:
8755 * lock(&ctx->uring_lock);
8757 * lock(&ctx->uring_lock);
8760 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8761 * pushs them to do the flush.
8763 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8764 mask |= EPOLLIN | EPOLLRDNORM;
8769 static int io_uring_fasync(int fd, struct file *file, int on)
8771 struct io_ring_ctx *ctx = file->private_data;
8773 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8776 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8778 const struct cred *creds;
8780 creds = xa_erase(&ctx->personalities, id);
8789 struct io_tctx_exit {
8790 struct callback_head task_work;
8791 struct completion completion;
8792 struct io_ring_ctx *ctx;
8795 static void io_tctx_exit_cb(struct callback_head *cb)
8797 struct io_uring_task *tctx = current->io_uring;
8798 struct io_tctx_exit *work;
8800 work = container_of(cb, struct io_tctx_exit, task_work);
8802 * When @in_idle, we're in cancellation and it's racy to remove the
8803 * node. It'll be removed by the end of cancellation, just ignore it.
8805 if (!atomic_read(&tctx->in_idle))
8806 io_uring_del_tctx_node((unsigned long)work->ctx);
8807 complete(&work->completion);
8810 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8812 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8814 return req->ctx == data;
8817 static void io_ring_exit_work(struct work_struct *work)
8819 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8820 unsigned long timeout = jiffies + HZ * 60 * 5;
8821 unsigned long interval = HZ / 20;
8822 struct io_tctx_exit exit;
8823 struct io_tctx_node *node;
8827 * If we're doing polled IO and end up having requests being
8828 * submitted async (out-of-line), then completions can come in while
8829 * we're waiting for refs to drop. We need to reap these manually,
8830 * as nobody else will be looking for them.
8833 io_uring_try_cancel_requests(ctx, NULL, true);
8835 struct io_sq_data *sqd = ctx->sq_data;
8836 struct task_struct *tsk;
8838 io_sq_thread_park(sqd);
8840 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8841 io_wq_cancel_cb(tsk->io_uring->io_wq,
8842 io_cancel_ctx_cb, ctx, true);
8843 io_sq_thread_unpark(sqd);
8846 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8847 /* there is little hope left, don't run it too often */
8850 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8852 init_completion(&exit.completion);
8853 init_task_work(&exit.task_work, io_tctx_exit_cb);
8856 * Some may use context even when all refs and requests have been put,
8857 * and they are free to do so while still holding uring_lock or
8858 * completion_lock, see io_req_task_submit(). Apart from other work,
8859 * this lock/unlock section also waits them to finish.
8861 mutex_lock(&ctx->uring_lock);
8862 while (!list_empty(&ctx->tctx_list)) {
8863 WARN_ON_ONCE(time_after(jiffies, timeout));
8865 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8867 /* don't spin on a single task if cancellation failed */
8868 list_rotate_left(&ctx->tctx_list);
8869 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8870 if (WARN_ON_ONCE(ret))
8872 wake_up_process(node->task);
8874 mutex_unlock(&ctx->uring_lock);
8875 wait_for_completion(&exit.completion);
8876 mutex_lock(&ctx->uring_lock);
8878 mutex_unlock(&ctx->uring_lock);
8879 spin_lock(&ctx->completion_lock);
8880 spin_unlock(&ctx->completion_lock);
8882 io_ring_ctx_free(ctx);
8885 /* Returns true if we found and killed one or more timeouts */
8886 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8889 struct io_kiocb *req, *tmp;
8892 spin_lock(&ctx->completion_lock);
8893 spin_lock_irq(&ctx->timeout_lock);
8894 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8895 if (io_match_task(req, tsk, cancel_all)) {
8896 io_kill_timeout(req, -ECANCELED);
8900 spin_unlock_irq(&ctx->timeout_lock);
8902 io_commit_cqring(ctx);
8903 spin_unlock(&ctx->completion_lock);
8905 io_cqring_ev_posted(ctx);
8906 return canceled != 0;
8909 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8911 unsigned long index;
8912 struct creds *creds;
8914 mutex_lock(&ctx->uring_lock);
8915 percpu_ref_kill(&ctx->refs);
8917 __io_cqring_overflow_flush(ctx, true);
8918 xa_for_each(&ctx->personalities, index, creds)
8919 io_unregister_personality(ctx, index);
8920 mutex_unlock(&ctx->uring_lock);
8922 io_kill_timeouts(ctx, NULL, true);
8923 io_poll_remove_all(ctx, NULL, true);
8925 /* if we failed setting up the ctx, we might not have any rings */
8926 io_iopoll_try_reap_events(ctx);
8928 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8930 * Use system_unbound_wq to avoid spawning tons of event kworkers
8931 * if we're exiting a ton of rings at the same time. It just adds
8932 * noise and overhead, there's no discernable change in runtime
8933 * over using system_wq.
8935 queue_work(system_unbound_wq, &ctx->exit_work);
8938 static int io_uring_release(struct inode *inode, struct file *file)
8940 struct io_ring_ctx *ctx = file->private_data;
8942 file->private_data = NULL;
8943 io_ring_ctx_wait_and_kill(ctx);
8947 struct io_task_cancel {
8948 struct task_struct *task;
8952 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8954 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8955 struct io_task_cancel *cancel = data;
8958 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8959 struct io_ring_ctx *ctx = req->ctx;
8961 /* protect against races with linked timeouts */
8962 spin_lock(&ctx->completion_lock);
8963 ret = io_match_task(req, cancel->task, cancel->all);
8964 spin_unlock(&ctx->completion_lock);
8966 ret = io_match_task(req, cancel->task, cancel->all);
8971 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8972 struct task_struct *task, bool cancel_all)
8974 struct io_defer_entry *de;
8977 spin_lock(&ctx->completion_lock);
8978 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8979 if (io_match_task(de->req, task, cancel_all)) {
8980 list_cut_position(&list, &ctx->defer_list, &de->list);
8984 spin_unlock(&ctx->completion_lock);
8985 if (list_empty(&list))
8988 while (!list_empty(&list)) {
8989 de = list_first_entry(&list, struct io_defer_entry, list);
8990 list_del_init(&de->list);
8991 io_req_complete_failed(de->req, -ECANCELED);
8997 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8999 struct io_tctx_node *node;
9000 enum io_wq_cancel cret;
9003 mutex_lock(&ctx->uring_lock);
9004 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9005 struct io_uring_task *tctx = node->task->io_uring;
9008 * io_wq will stay alive while we hold uring_lock, because it's
9009 * killed after ctx nodes, which requires to take the lock.
9011 if (!tctx || !tctx->io_wq)
9013 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9014 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9016 mutex_unlock(&ctx->uring_lock);
9021 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9022 struct task_struct *task,
9025 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9026 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9029 enum io_wq_cancel cret;
9033 ret |= io_uring_try_cancel_iowq(ctx);
9034 } else if (tctx && tctx->io_wq) {
9036 * Cancels requests of all rings, not only @ctx, but
9037 * it's fine as the task is in exit/exec.
9039 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9041 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9044 /* SQPOLL thread does its own polling */
9045 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9046 (ctx->sq_data && ctx->sq_data->thread == current)) {
9047 while (!list_empty_careful(&ctx->iopoll_list)) {
9048 io_iopoll_try_reap_events(ctx);
9053 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9054 ret |= io_poll_remove_all(ctx, task, cancel_all);
9055 ret |= io_kill_timeouts(ctx, task, cancel_all);
9057 ret |= io_run_task_work();
9064 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9066 struct io_uring_task *tctx = current->io_uring;
9067 struct io_tctx_node *node;
9070 if (unlikely(!tctx)) {
9071 ret = io_uring_alloc_task_context(current, ctx);
9074 tctx = current->io_uring;
9076 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9077 node = kmalloc(sizeof(*node), GFP_KERNEL);
9081 node->task = current;
9083 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9090 mutex_lock(&ctx->uring_lock);
9091 list_add(&node->ctx_node, &ctx->tctx_list);
9092 mutex_unlock(&ctx->uring_lock);
9099 * Note that this task has used io_uring. We use it for cancelation purposes.
9101 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9103 struct io_uring_task *tctx = current->io_uring;
9105 if (likely(tctx && tctx->last == ctx))
9107 return __io_uring_add_tctx_node(ctx);
9111 * Remove this io_uring_file -> task mapping.
9113 static void io_uring_del_tctx_node(unsigned long index)
9115 struct io_uring_task *tctx = current->io_uring;
9116 struct io_tctx_node *node;
9120 node = xa_erase(&tctx->xa, index);
9124 WARN_ON_ONCE(current != node->task);
9125 WARN_ON_ONCE(list_empty(&node->ctx_node));
9127 mutex_lock(&node->ctx->uring_lock);
9128 list_del(&node->ctx_node);
9129 mutex_unlock(&node->ctx->uring_lock);
9131 if (tctx->last == node->ctx)
9136 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9138 struct io_wq *wq = tctx->io_wq;
9139 struct io_tctx_node *node;
9140 unsigned long index;
9142 xa_for_each(&tctx->xa, index, node)
9143 io_uring_del_tctx_node(index);
9146 * Must be after io_uring_del_task_file() (removes nodes under
9147 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9150 io_wq_put_and_exit(wq);
9154 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9157 return atomic_read(&tctx->inflight_tracked);
9158 return percpu_counter_sum(&tctx->inflight);
9161 static void io_uring_drop_tctx_refs(struct task_struct *task)
9163 struct io_uring_task *tctx = task->io_uring;
9164 unsigned int refs = tctx->cached_refs;
9167 tctx->cached_refs = 0;
9168 percpu_counter_sub(&tctx->inflight, refs);
9169 put_task_struct_many(task, refs);
9174 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9175 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9177 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9179 struct io_uring_task *tctx = current->io_uring;
9180 struct io_ring_ctx *ctx;
9184 WARN_ON_ONCE(sqd && sqd->thread != current);
9186 if (!current->io_uring)
9189 io_wq_exit_start(tctx->io_wq);
9191 atomic_inc(&tctx->in_idle);
9193 io_uring_drop_tctx_refs(current);
9194 /* read completions before cancelations */
9195 inflight = tctx_inflight(tctx, !cancel_all);
9200 struct io_tctx_node *node;
9201 unsigned long index;
9203 xa_for_each(&tctx->xa, index, node) {
9204 /* sqpoll task will cancel all its requests */
9205 if (node->ctx->sq_data)
9207 io_uring_try_cancel_requests(node->ctx, current,
9211 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9212 io_uring_try_cancel_requests(ctx, current,
9216 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9217 io_uring_drop_tctx_refs(current);
9219 * If we've seen completions, retry without waiting. This
9220 * avoids a race where a completion comes in before we did
9221 * prepare_to_wait().
9223 if (inflight == tctx_inflight(tctx, !cancel_all))
9225 finish_wait(&tctx->wait, &wait);
9227 atomic_dec(&tctx->in_idle);
9229 io_uring_clean_tctx(tctx);
9231 /* for exec all current's requests should be gone, kill tctx */
9232 __io_uring_free(current);
9236 void __io_uring_cancel(bool cancel_all)
9238 io_uring_cancel_generic(cancel_all, NULL);
9241 static void *io_uring_validate_mmap_request(struct file *file,
9242 loff_t pgoff, size_t sz)
9244 struct io_ring_ctx *ctx = file->private_data;
9245 loff_t offset = pgoff << PAGE_SHIFT;
9250 case IORING_OFF_SQ_RING:
9251 case IORING_OFF_CQ_RING:
9254 case IORING_OFF_SQES:
9258 return ERR_PTR(-EINVAL);
9261 page = virt_to_head_page(ptr);
9262 if (sz > page_size(page))
9263 return ERR_PTR(-EINVAL);
9270 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9272 size_t sz = vma->vm_end - vma->vm_start;
9276 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9278 return PTR_ERR(ptr);
9280 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9281 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9284 #else /* !CONFIG_MMU */
9286 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9288 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9291 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9293 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9296 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9297 unsigned long addr, unsigned long len,
9298 unsigned long pgoff, unsigned long flags)
9302 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9304 return PTR_ERR(ptr);
9306 return (unsigned long) ptr;
9309 #endif /* !CONFIG_MMU */
9311 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9316 if (!io_sqring_full(ctx))
9318 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9320 if (!io_sqring_full(ctx))
9323 } while (!signal_pending(current));
9325 finish_wait(&ctx->sqo_sq_wait, &wait);
9329 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9330 struct __kernel_timespec __user **ts,
9331 const sigset_t __user **sig)
9333 struct io_uring_getevents_arg arg;
9336 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9337 * is just a pointer to the sigset_t.
9339 if (!(flags & IORING_ENTER_EXT_ARG)) {
9340 *sig = (const sigset_t __user *) argp;
9346 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9347 * timespec and sigset_t pointers if good.
9349 if (*argsz != sizeof(arg))
9351 if (copy_from_user(&arg, argp, sizeof(arg)))
9353 *sig = u64_to_user_ptr(arg.sigmask);
9354 *argsz = arg.sigmask_sz;
9355 *ts = u64_to_user_ptr(arg.ts);
9359 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9360 u32, min_complete, u32, flags, const void __user *, argp,
9363 struct io_ring_ctx *ctx;
9370 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9371 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9375 if (unlikely(!f.file))
9379 if (unlikely(f.file->f_op != &io_uring_fops))
9383 ctx = f.file->private_data;
9384 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9388 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9392 * For SQ polling, the thread will do all submissions and completions.
9393 * Just return the requested submit count, and wake the thread if
9397 if (ctx->flags & IORING_SETUP_SQPOLL) {
9398 io_cqring_overflow_flush(ctx);
9400 if (unlikely(ctx->sq_data->thread == NULL)) {
9404 if (flags & IORING_ENTER_SQ_WAKEUP)
9405 wake_up(&ctx->sq_data->wait);
9406 if (flags & IORING_ENTER_SQ_WAIT) {
9407 ret = io_sqpoll_wait_sq(ctx);
9411 submitted = to_submit;
9412 } else if (to_submit) {
9413 ret = io_uring_add_tctx_node(ctx);
9416 mutex_lock(&ctx->uring_lock);
9417 submitted = io_submit_sqes(ctx, to_submit);
9418 mutex_unlock(&ctx->uring_lock);
9420 if (submitted != to_submit)
9423 if (flags & IORING_ENTER_GETEVENTS) {
9424 const sigset_t __user *sig;
9425 struct __kernel_timespec __user *ts;
9427 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9431 min_complete = min(min_complete, ctx->cq_entries);
9434 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9435 * space applications don't need to do io completion events
9436 * polling again, they can rely on io_sq_thread to do polling
9437 * work, which can reduce cpu usage and uring_lock contention.
9439 if (ctx->flags & IORING_SETUP_IOPOLL &&
9440 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9441 ret = io_iopoll_check(ctx, min_complete);
9443 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9448 percpu_ref_put(&ctx->refs);
9451 return submitted ? submitted : ret;
9454 #ifdef CONFIG_PROC_FS
9455 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9456 const struct cred *cred)
9458 struct user_namespace *uns = seq_user_ns(m);
9459 struct group_info *gi;
9464 seq_printf(m, "%5d\n", id);
9465 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9466 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9467 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9468 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9469 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9470 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9471 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9472 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9473 seq_puts(m, "\n\tGroups:\t");
9474 gi = cred->group_info;
9475 for (g = 0; g < gi->ngroups; g++) {
9476 seq_put_decimal_ull(m, g ? " " : "",
9477 from_kgid_munged(uns, gi->gid[g]));
9479 seq_puts(m, "\n\tCapEff:\t");
9480 cap = cred->cap_effective;
9481 CAP_FOR_EACH_U32(__capi)
9482 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9487 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9489 struct io_sq_data *sq = NULL;
9494 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9495 * since fdinfo case grabs it in the opposite direction of normal use
9496 * cases. If we fail to get the lock, we just don't iterate any
9497 * structures that could be going away outside the io_uring mutex.
9499 has_lock = mutex_trylock(&ctx->uring_lock);
9501 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9507 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9508 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9509 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9510 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9511 struct file *f = io_file_from_index(ctx, i);
9514 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9516 seq_printf(m, "%5u: <none>\n", i);
9518 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9519 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9520 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9521 unsigned int len = buf->ubuf_end - buf->ubuf;
9523 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9525 if (has_lock && !xa_empty(&ctx->personalities)) {
9526 unsigned long index;
9527 const struct cred *cred;
9529 seq_printf(m, "Personalities:\n");
9530 xa_for_each(&ctx->personalities, index, cred)
9531 io_uring_show_cred(m, index, cred);
9533 seq_printf(m, "PollList:\n");
9534 spin_lock(&ctx->completion_lock);
9535 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9536 struct hlist_head *list = &ctx->cancel_hash[i];
9537 struct io_kiocb *req;
9539 hlist_for_each_entry(req, list, hash_node)
9540 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9541 req->task->task_works != NULL);
9543 spin_unlock(&ctx->completion_lock);
9545 mutex_unlock(&ctx->uring_lock);
9548 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9550 struct io_ring_ctx *ctx = f->private_data;
9552 if (percpu_ref_tryget(&ctx->refs)) {
9553 __io_uring_show_fdinfo(ctx, m);
9554 percpu_ref_put(&ctx->refs);
9559 static const struct file_operations io_uring_fops = {
9560 .release = io_uring_release,
9561 .mmap = io_uring_mmap,
9563 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9564 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9566 .poll = io_uring_poll,
9567 .fasync = io_uring_fasync,
9568 #ifdef CONFIG_PROC_FS
9569 .show_fdinfo = io_uring_show_fdinfo,
9573 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9574 struct io_uring_params *p)
9576 struct io_rings *rings;
9577 size_t size, sq_array_offset;
9579 /* make sure these are sane, as we already accounted them */
9580 ctx->sq_entries = p->sq_entries;
9581 ctx->cq_entries = p->cq_entries;
9583 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9584 if (size == SIZE_MAX)
9587 rings = io_mem_alloc(size);
9592 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9593 rings->sq_ring_mask = p->sq_entries - 1;
9594 rings->cq_ring_mask = p->cq_entries - 1;
9595 rings->sq_ring_entries = p->sq_entries;
9596 rings->cq_ring_entries = p->cq_entries;
9598 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9599 if (size == SIZE_MAX) {
9600 io_mem_free(ctx->rings);
9605 ctx->sq_sqes = io_mem_alloc(size);
9606 if (!ctx->sq_sqes) {
9607 io_mem_free(ctx->rings);
9615 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9619 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9623 ret = io_uring_add_tctx_node(ctx);
9628 fd_install(fd, file);
9633 * Allocate an anonymous fd, this is what constitutes the application
9634 * visible backing of an io_uring instance. The application mmaps this
9635 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9636 * we have to tie this fd to a socket for file garbage collection purposes.
9638 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9641 #if defined(CONFIG_UNIX)
9644 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9647 return ERR_PTR(ret);
9650 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9651 O_RDWR | O_CLOEXEC);
9652 #if defined(CONFIG_UNIX)
9654 sock_release(ctx->ring_sock);
9655 ctx->ring_sock = NULL;
9657 ctx->ring_sock->file = file;
9663 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9664 struct io_uring_params __user *params)
9666 struct io_ring_ctx *ctx;
9672 if (entries > IORING_MAX_ENTRIES) {
9673 if (!(p->flags & IORING_SETUP_CLAMP))
9675 entries = IORING_MAX_ENTRIES;
9679 * Use twice as many entries for the CQ ring. It's possible for the
9680 * application to drive a higher depth than the size of the SQ ring,
9681 * since the sqes are only used at submission time. This allows for
9682 * some flexibility in overcommitting a bit. If the application has
9683 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9684 * of CQ ring entries manually.
9686 p->sq_entries = roundup_pow_of_two(entries);
9687 if (p->flags & IORING_SETUP_CQSIZE) {
9689 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9690 * to a power-of-two, if it isn't already. We do NOT impose
9691 * any cq vs sq ring sizing.
9695 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9696 if (!(p->flags & IORING_SETUP_CLAMP))
9698 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9700 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9701 if (p->cq_entries < p->sq_entries)
9704 p->cq_entries = 2 * p->sq_entries;
9707 ctx = io_ring_ctx_alloc(p);
9710 ctx->compat = in_compat_syscall();
9711 if (!capable(CAP_IPC_LOCK))
9712 ctx->user = get_uid(current_user());
9715 * This is just grabbed for accounting purposes. When a process exits,
9716 * the mm is exited and dropped before the files, hence we need to hang
9717 * on to this mm purely for the purposes of being able to unaccount
9718 * memory (locked/pinned vm). It's not used for anything else.
9720 mmgrab(current->mm);
9721 ctx->mm_account = current->mm;
9723 ret = io_allocate_scq_urings(ctx, p);
9727 ret = io_sq_offload_create(ctx, p);
9730 /* always set a rsrc node */
9731 ret = io_rsrc_node_switch_start(ctx);
9734 io_rsrc_node_switch(ctx, NULL);
9736 memset(&p->sq_off, 0, sizeof(p->sq_off));
9737 p->sq_off.head = offsetof(struct io_rings, sq.head);
9738 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9739 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9740 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9741 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9742 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9743 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9745 memset(&p->cq_off, 0, sizeof(p->cq_off));
9746 p->cq_off.head = offsetof(struct io_rings, cq.head);
9747 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9748 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9749 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9750 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9751 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9752 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9754 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9755 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9756 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9757 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9758 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9759 IORING_FEAT_RSRC_TAGS;
9761 if (copy_to_user(params, p, sizeof(*p))) {
9766 file = io_uring_get_file(ctx);
9768 ret = PTR_ERR(file);
9773 * Install ring fd as the very last thing, so we don't risk someone
9774 * having closed it before we finish setup
9776 ret = io_uring_install_fd(ctx, file);
9778 /* fput will clean it up */
9783 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9786 io_ring_ctx_wait_and_kill(ctx);
9791 * Sets up an aio uring context, and returns the fd. Applications asks for a
9792 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9793 * params structure passed in.
9795 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9797 struct io_uring_params p;
9800 if (copy_from_user(&p, params, sizeof(p)))
9802 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9807 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9808 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9809 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9810 IORING_SETUP_R_DISABLED))
9813 return io_uring_create(entries, &p, params);
9816 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9817 struct io_uring_params __user *, params)
9819 return io_uring_setup(entries, params);
9822 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9824 struct io_uring_probe *p;
9828 size = struct_size(p, ops, nr_args);
9829 if (size == SIZE_MAX)
9831 p = kzalloc(size, GFP_KERNEL);
9836 if (copy_from_user(p, arg, size))
9839 if (memchr_inv(p, 0, size))
9842 p->last_op = IORING_OP_LAST - 1;
9843 if (nr_args > IORING_OP_LAST)
9844 nr_args = IORING_OP_LAST;
9846 for (i = 0; i < nr_args; i++) {
9848 if (!io_op_defs[i].not_supported)
9849 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9854 if (copy_to_user(arg, p, size))
9861 static int io_register_personality(struct io_ring_ctx *ctx)
9863 const struct cred *creds;
9867 creds = get_current_cred();
9869 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9870 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9878 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9879 unsigned int nr_args)
9881 struct io_uring_restriction *res;
9885 /* Restrictions allowed only if rings started disabled */
9886 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9889 /* We allow only a single restrictions registration */
9890 if (ctx->restrictions.registered)
9893 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9896 size = array_size(nr_args, sizeof(*res));
9897 if (size == SIZE_MAX)
9900 res = memdup_user(arg, size);
9902 return PTR_ERR(res);
9906 for (i = 0; i < nr_args; i++) {
9907 switch (res[i].opcode) {
9908 case IORING_RESTRICTION_REGISTER_OP:
9909 if (res[i].register_op >= IORING_REGISTER_LAST) {
9914 __set_bit(res[i].register_op,
9915 ctx->restrictions.register_op);
9917 case IORING_RESTRICTION_SQE_OP:
9918 if (res[i].sqe_op >= IORING_OP_LAST) {
9923 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9925 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9926 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9928 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9929 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9938 /* Reset all restrictions if an error happened */
9940 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9942 ctx->restrictions.registered = true;
9948 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9950 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9953 if (ctx->restrictions.registered)
9954 ctx->restricted = 1;
9956 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9957 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9958 wake_up(&ctx->sq_data->wait);
9962 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9963 struct io_uring_rsrc_update2 *up,
9971 if (check_add_overflow(up->offset, nr_args, &tmp))
9973 err = io_rsrc_node_switch_start(ctx);
9978 case IORING_RSRC_FILE:
9979 return __io_sqe_files_update(ctx, up, nr_args);
9980 case IORING_RSRC_BUFFER:
9981 return __io_sqe_buffers_update(ctx, up, nr_args);
9986 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9989 struct io_uring_rsrc_update2 up;
9993 memset(&up, 0, sizeof(up));
9994 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9996 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9999 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10000 unsigned size, unsigned type)
10002 struct io_uring_rsrc_update2 up;
10004 if (size != sizeof(up))
10006 if (copy_from_user(&up, arg, sizeof(up)))
10008 if (!up.nr || up.resv)
10010 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10013 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10014 unsigned int size, unsigned int type)
10016 struct io_uring_rsrc_register rr;
10018 /* keep it extendible */
10019 if (size != sizeof(rr))
10022 memset(&rr, 0, sizeof(rr));
10023 if (copy_from_user(&rr, arg, size))
10025 if (!rr.nr || rr.resv || rr.resv2)
10029 case IORING_RSRC_FILE:
10030 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10031 rr.nr, u64_to_user_ptr(rr.tags));
10032 case IORING_RSRC_BUFFER:
10033 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10034 rr.nr, u64_to_user_ptr(rr.tags));
10039 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10042 struct io_uring_task *tctx = current->io_uring;
10043 cpumask_var_t new_mask;
10046 if (!tctx || !tctx->io_wq)
10049 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10052 cpumask_clear(new_mask);
10053 if (len > cpumask_size())
10054 len = cpumask_size();
10056 if (copy_from_user(new_mask, arg, len)) {
10057 free_cpumask_var(new_mask);
10061 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10062 free_cpumask_var(new_mask);
10066 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10068 struct io_uring_task *tctx = current->io_uring;
10070 if (!tctx || !tctx->io_wq)
10073 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10076 static bool io_register_op_must_quiesce(int op)
10079 case IORING_REGISTER_BUFFERS:
10080 case IORING_UNREGISTER_BUFFERS:
10081 case IORING_REGISTER_FILES:
10082 case IORING_UNREGISTER_FILES:
10083 case IORING_REGISTER_FILES_UPDATE:
10084 case IORING_REGISTER_PROBE:
10085 case IORING_REGISTER_PERSONALITY:
10086 case IORING_UNREGISTER_PERSONALITY:
10087 case IORING_REGISTER_FILES2:
10088 case IORING_REGISTER_FILES_UPDATE2:
10089 case IORING_REGISTER_BUFFERS2:
10090 case IORING_REGISTER_BUFFERS_UPDATE:
10091 case IORING_REGISTER_IOWQ_AFF:
10092 case IORING_UNREGISTER_IOWQ_AFF:
10099 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10103 percpu_ref_kill(&ctx->refs);
10106 * Drop uring mutex before waiting for references to exit. If another
10107 * thread is currently inside io_uring_enter() it might need to grab the
10108 * uring_lock to make progress. If we hold it here across the drain
10109 * wait, then we can deadlock. It's safe to drop the mutex here, since
10110 * no new references will come in after we've killed the percpu ref.
10112 mutex_unlock(&ctx->uring_lock);
10114 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10117 ret = io_run_task_work_sig();
10118 } while (ret >= 0);
10119 mutex_lock(&ctx->uring_lock);
10122 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10126 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10127 void __user *arg, unsigned nr_args)
10128 __releases(ctx->uring_lock)
10129 __acquires(ctx->uring_lock)
10134 * We're inside the ring mutex, if the ref is already dying, then
10135 * someone else killed the ctx or is already going through
10136 * io_uring_register().
10138 if (percpu_ref_is_dying(&ctx->refs))
10141 if (ctx->restricted) {
10142 if (opcode >= IORING_REGISTER_LAST)
10144 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10145 if (!test_bit(opcode, ctx->restrictions.register_op))
10149 if (io_register_op_must_quiesce(opcode)) {
10150 ret = io_ctx_quiesce(ctx);
10156 case IORING_REGISTER_BUFFERS:
10157 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10159 case IORING_UNREGISTER_BUFFERS:
10161 if (arg || nr_args)
10163 ret = io_sqe_buffers_unregister(ctx);
10165 case IORING_REGISTER_FILES:
10166 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10168 case IORING_UNREGISTER_FILES:
10170 if (arg || nr_args)
10172 ret = io_sqe_files_unregister(ctx);
10174 case IORING_REGISTER_FILES_UPDATE:
10175 ret = io_register_files_update(ctx, arg, nr_args);
10177 case IORING_REGISTER_EVENTFD:
10178 case IORING_REGISTER_EVENTFD_ASYNC:
10182 ret = io_eventfd_register(ctx, arg);
10185 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10186 ctx->eventfd_async = 1;
10188 ctx->eventfd_async = 0;
10190 case IORING_UNREGISTER_EVENTFD:
10192 if (arg || nr_args)
10194 ret = io_eventfd_unregister(ctx);
10196 case IORING_REGISTER_PROBE:
10198 if (!arg || nr_args > 256)
10200 ret = io_probe(ctx, arg, nr_args);
10202 case IORING_REGISTER_PERSONALITY:
10204 if (arg || nr_args)
10206 ret = io_register_personality(ctx);
10208 case IORING_UNREGISTER_PERSONALITY:
10212 ret = io_unregister_personality(ctx, nr_args);
10214 case IORING_REGISTER_ENABLE_RINGS:
10216 if (arg || nr_args)
10218 ret = io_register_enable_rings(ctx);
10220 case IORING_REGISTER_RESTRICTIONS:
10221 ret = io_register_restrictions(ctx, arg, nr_args);
10223 case IORING_REGISTER_FILES2:
10224 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10226 case IORING_REGISTER_FILES_UPDATE2:
10227 ret = io_register_rsrc_update(ctx, arg, nr_args,
10230 case IORING_REGISTER_BUFFERS2:
10231 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10233 case IORING_REGISTER_BUFFERS_UPDATE:
10234 ret = io_register_rsrc_update(ctx, arg, nr_args,
10235 IORING_RSRC_BUFFER);
10237 case IORING_REGISTER_IOWQ_AFF:
10239 if (!arg || !nr_args)
10241 ret = io_register_iowq_aff(ctx, arg, nr_args);
10243 case IORING_UNREGISTER_IOWQ_AFF:
10245 if (arg || nr_args)
10247 ret = io_unregister_iowq_aff(ctx);
10254 if (io_register_op_must_quiesce(opcode)) {
10255 /* bring the ctx back to life */
10256 percpu_ref_reinit(&ctx->refs);
10257 reinit_completion(&ctx->ref_comp);
10262 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10263 void __user *, arg, unsigned int, nr_args)
10265 struct io_ring_ctx *ctx;
10274 if (f.file->f_op != &io_uring_fops)
10277 ctx = f.file->private_data;
10279 io_run_task_work();
10281 mutex_lock(&ctx->uring_lock);
10282 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10283 mutex_unlock(&ctx->uring_lock);
10284 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10285 ctx->cq_ev_fd != NULL, ret);
10291 static int __init io_uring_init(void)
10293 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10294 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10295 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10298 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10299 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10300 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10301 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10302 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10303 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10304 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10305 BUILD_BUG_SQE_ELEM(8, __u64, off);
10306 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10307 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10308 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10309 BUILD_BUG_SQE_ELEM(24, __u32, len);
10310 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10311 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10312 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10313 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10314 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10315 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10316 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10317 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10318 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10319 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10320 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10321 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10322 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10323 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10324 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10325 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10326 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10327 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10328 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10329 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10331 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10332 sizeof(struct io_uring_rsrc_update));
10333 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10334 sizeof(struct io_uring_rsrc_update2));
10335 /* should fit into one byte */
10336 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10338 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10339 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10341 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10345 __initcall(io_uring_init);