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);
1201 mutex_lock(&ctx->uring_lock);
1202 if (ctx->submit_state.compl_nr)
1203 io_submit_flush_completions(ctx);
1204 mutex_unlock(&ctx->uring_lock);
1205 percpu_ref_put(&ctx->refs);
1208 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1210 struct io_ring_ctx *ctx;
1213 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1218 * Use 5 bits less than the max cq entries, that should give us around
1219 * 32 entries per hash list if totally full and uniformly spread.
1221 hash_bits = ilog2(p->cq_entries);
1225 ctx->cancel_hash_bits = hash_bits;
1226 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1228 if (!ctx->cancel_hash)
1230 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1232 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1233 if (!ctx->dummy_ubuf)
1235 /* set invalid range, so io_import_fixed() fails meeting it */
1236 ctx->dummy_ubuf->ubuf = -1UL;
1238 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1239 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1242 ctx->flags = p->flags;
1243 init_waitqueue_head(&ctx->sqo_sq_wait);
1244 INIT_LIST_HEAD(&ctx->sqd_list);
1245 init_waitqueue_head(&ctx->poll_wait);
1246 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1247 init_completion(&ctx->ref_comp);
1248 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1249 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1250 mutex_init(&ctx->uring_lock);
1251 init_waitqueue_head(&ctx->cq_wait);
1252 spin_lock_init(&ctx->completion_lock);
1253 spin_lock_init(&ctx->timeout_lock);
1254 INIT_LIST_HEAD(&ctx->iopoll_list);
1255 INIT_LIST_HEAD(&ctx->defer_list);
1256 INIT_LIST_HEAD(&ctx->timeout_list);
1257 spin_lock_init(&ctx->rsrc_ref_lock);
1258 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1259 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1260 init_llist_head(&ctx->rsrc_put_llist);
1261 INIT_LIST_HEAD(&ctx->tctx_list);
1262 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1263 INIT_LIST_HEAD(&ctx->locked_free_list);
1264 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1267 kfree(ctx->dummy_ubuf);
1268 kfree(ctx->cancel_hash);
1273 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1275 struct io_rings *r = ctx->rings;
1277 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1281 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1283 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1284 struct io_ring_ctx *ctx = req->ctx;
1286 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1292 #define FFS_ASYNC_READ 0x1UL
1293 #define FFS_ASYNC_WRITE 0x2UL
1295 #define FFS_ISREG 0x4UL
1297 #define FFS_ISREG 0x0UL
1299 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1301 static inline bool io_req_ffs_set(struct io_kiocb *req)
1303 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1306 static void io_req_track_inflight(struct io_kiocb *req)
1308 if (!(req->flags & REQ_F_INFLIGHT)) {
1309 req->flags |= REQ_F_INFLIGHT;
1310 atomic_inc(¤t->io_uring->inflight_tracked);
1314 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1316 req->flags &= ~REQ_F_LINK_TIMEOUT;
1319 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1321 if (WARN_ON_ONCE(!req->link))
1324 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1325 req->flags |= REQ_F_LINK_TIMEOUT;
1327 /* linked timeouts should have two refs once prep'ed */
1328 io_req_set_refcount(req);
1329 __io_req_set_refcount(req->link, 2);
1333 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1335 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1337 return __io_prep_linked_timeout(req);
1340 static void io_prep_async_work(struct io_kiocb *req)
1342 const struct io_op_def *def = &io_op_defs[req->opcode];
1343 struct io_ring_ctx *ctx = req->ctx;
1345 if (!(req->flags & REQ_F_CREDS)) {
1346 req->flags |= REQ_F_CREDS;
1347 req->creds = get_current_cred();
1350 req->work.list.next = NULL;
1351 req->work.flags = 0;
1352 if (req->flags & REQ_F_FORCE_ASYNC)
1353 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1355 if (req->flags & REQ_F_ISREG) {
1356 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1357 io_wq_hash_work(&req->work, file_inode(req->file));
1358 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1359 if (def->unbound_nonreg_file)
1360 req->work.flags |= IO_WQ_WORK_UNBOUND;
1363 switch (req->opcode) {
1364 case IORING_OP_SPLICE:
1366 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1367 req->work.flags |= IO_WQ_WORK_UNBOUND;
1372 static void io_prep_async_link(struct io_kiocb *req)
1374 struct io_kiocb *cur;
1376 if (req->flags & REQ_F_LINK_TIMEOUT) {
1377 struct io_ring_ctx *ctx = req->ctx;
1379 spin_lock(&ctx->completion_lock);
1380 io_for_each_link(cur, req)
1381 io_prep_async_work(cur);
1382 spin_unlock(&ctx->completion_lock);
1384 io_for_each_link(cur, req)
1385 io_prep_async_work(cur);
1389 static void io_queue_async_work(struct io_kiocb *req)
1391 struct io_ring_ctx *ctx = req->ctx;
1392 struct io_kiocb *link = io_prep_linked_timeout(req);
1393 struct io_uring_task *tctx = req->task->io_uring;
1396 BUG_ON(!tctx->io_wq);
1398 /* init ->work of the whole link before punting */
1399 io_prep_async_link(req);
1402 * Not expected to happen, but if we do have a bug where this _can_
1403 * happen, catch it here and ensure the request is marked as
1404 * canceled. That will make io-wq go through the usual work cancel
1405 * procedure rather than attempt to run this request (or create a new
1408 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1409 req->work.flags |= IO_WQ_WORK_CANCEL;
1411 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1412 &req->work, req->flags);
1413 io_wq_enqueue(tctx->io_wq, &req->work);
1415 io_queue_linked_timeout(link);
1418 static void io_kill_timeout(struct io_kiocb *req, int status)
1419 __must_hold(&req->ctx->completion_lock)
1420 __must_hold(&req->ctx->timeout_lock)
1422 struct io_timeout_data *io = req->async_data;
1424 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1425 atomic_set(&req->ctx->cq_timeouts,
1426 atomic_read(&req->ctx->cq_timeouts) + 1);
1427 list_del_init(&req->timeout.list);
1428 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1429 io_put_req_deferred(req);
1433 static void io_queue_deferred(struct io_ring_ctx *ctx)
1435 while (!list_empty(&ctx->defer_list)) {
1436 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1437 struct io_defer_entry, list);
1439 if (req_need_defer(de->req, de->seq))
1441 list_del_init(&de->list);
1442 io_req_task_queue(de->req);
1447 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1448 __must_hold(&ctx->completion_lock)
1450 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1452 spin_lock_irq(&ctx->timeout_lock);
1453 while (!list_empty(&ctx->timeout_list)) {
1454 u32 events_needed, events_got;
1455 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1456 struct io_kiocb, timeout.list);
1458 if (io_is_timeout_noseq(req))
1462 * Since seq can easily wrap around over time, subtract
1463 * the last seq at which timeouts were flushed before comparing.
1464 * Assuming not more than 2^31-1 events have happened since,
1465 * these subtractions won't have wrapped, so we can check if
1466 * target is in [last_seq, current_seq] by comparing the two.
1468 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1469 events_got = seq - ctx->cq_last_tm_flush;
1470 if (events_got < events_needed)
1473 list_del_init(&req->timeout.list);
1474 io_kill_timeout(req, 0);
1476 ctx->cq_last_tm_flush = seq;
1477 spin_unlock_irq(&ctx->timeout_lock);
1480 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1482 if (ctx->off_timeout_used)
1483 io_flush_timeouts(ctx);
1484 if (ctx->drain_active)
1485 io_queue_deferred(ctx);
1488 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1490 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1491 __io_commit_cqring_flush(ctx);
1492 /* order cqe stores with ring update */
1493 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1496 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1498 struct io_rings *r = ctx->rings;
1500 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1503 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1505 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1508 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1510 struct io_rings *rings = ctx->rings;
1511 unsigned tail, mask = ctx->cq_entries - 1;
1514 * writes to the cq entry need to come after reading head; the
1515 * control dependency is enough as we're using WRITE_ONCE to
1518 if (__io_cqring_events(ctx) == ctx->cq_entries)
1521 tail = ctx->cached_cq_tail++;
1522 return &rings->cqes[tail & mask];
1525 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1527 if (likely(!ctx->cq_ev_fd))
1529 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1531 return !ctx->eventfd_async || io_wq_current_is_worker();
1535 * This should only get called when at least one event has been posted.
1536 * Some applications rely on the eventfd notification count only changing
1537 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1538 * 1:1 relationship between how many times this function is called (and
1539 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1541 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1544 * wake_up_all() may seem excessive, but io_wake_function() and
1545 * io_should_wake() handle the termination of the loop and only
1546 * wake as many waiters as we need to.
1548 if (wq_has_sleeper(&ctx->cq_wait))
1549 wake_up_all(&ctx->cq_wait);
1550 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1551 wake_up(&ctx->sq_data->wait);
1552 if (io_should_trigger_evfd(ctx))
1553 eventfd_signal(ctx->cq_ev_fd, 1);
1554 if (waitqueue_active(&ctx->poll_wait)) {
1555 wake_up_interruptible(&ctx->poll_wait);
1556 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1560 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1562 if (ctx->flags & IORING_SETUP_SQPOLL) {
1563 if (wq_has_sleeper(&ctx->cq_wait))
1564 wake_up_all(&ctx->cq_wait);
1566 if (io_should_trigger_evfd(ctx))
1567 eventfd_signal(ctx->cq_ev_fd, 1);
1568 if (waitqueue_active(&ctx->poll_wait)) {
1569 wake_up_interruptible(&ctx->poll_wait);
1570 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1574 /* Returns true if there are no backlogged entries after the flush */
1575 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1577 bool all_flushed, posted;
1579 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1583 spin_lock(&ctx->completion_lock);
1584 while (!list_empty(&ctx->cq_overflow_list)) {
1585 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1586 struct io_overflow_cqe *ocqe;
1590 ocqe = list_first_entry(&ctx->cq_overflow_list,
1591 struct io_overflow_cqe, list);
1593 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1595 io_account_cq_overflow(ctx);
1598 list_del(&ocqe->list);
1602 all_flushed = list_empty(&ctx->cq_overflow_list);
1604 clear_bit(0, &ctx->check_cq_overflow);
1605 WRITE_ONCE(ctx->rings->sq_flags,
1606 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1610 io_commit_cqring(ctx);
1611 spin_unlock(&ctx->completion_lock);
1613 io_cqring_ev_posted(ctx);
1617 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1621 if (test_bit(0, &ctx->check_cq_overflow)) {
1622 /* iopoll syncs against uring_lock, not completion_lock */
1623 if (ctx->flags & IORING_SETUP_IOPOLL)
1624 mutex_lock(&ctx->uring_lock);
1625 ret = __io_cqring_overflow_flush(ctx, false);
1626 if (ctx->flags & IORING_SETUP_IOPOLL)
1627 mutex_unlock(&ctx->uring_lock);
1633 /* must to be called somewhat shortly after putting a request */
1634 static inline void io_put_task(struct task_struct *task, int nr)
1636 struct io_uring_task *tctx = task->io_uring;
1638 if (likely(task == current)) {
1639 tctx->cached_refs += nr;
1641 percpu_counter_sub(&tctx->inflight, nr);
1642 if (unlikely(atomic_read(&tctx->in_idle)))
1643 wake_up(&tctx->wait);
1644 put_task_struct_many(task, nr);
1648 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1649 long res, unsigned int cflags)
1651 struct io_overflow_cqe *ocqe;
1653 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1656 * If we're in ring overflow flush mode, or in task cancel mode,
1657 * or cannot allocate an overflow entry, then we need to drop it
1660 io_account_cq_overflow(ctx);
1663 if (list_empty(&ctx->cq_overflow_list)) {
1664 set_bit(0, &ctx->check_cq_overflow);
1665 WRITE_ONCE(ctx->rings->sq_flags,
1666 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1669 ocqe->cqe.user_data = user_data;
1670 ocqe->cqe.res = res;
1671 ocqe->cqe.flags = cflags;
1672 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1676 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1677 long res, unsigned int cflags)
1679 struct io_uring_cqe *cqe;
1681 trace_io_uring_complete(ctx, user_data, res, cflags);
1684 * If we can't get a cq entry, userspace overflowed the
1685 * submission (by quite a lot). Increment the overflow count in
1688 cqe = io_get_cqe(ctx);
1690 WRITE_ONCE(cqe->user_data, user_data);
1691 WRITE_ONCE(cqe->res, res);
1692 WRITE_ONCE(cqe->flags, cflags);
1695 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1698 /* not as hot to bloat with inlining */
1699 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1700 long res, unsigned int cflags)
1702 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1705 static void io_req_complete_post(struct io_kiocb *req, long res,
1706 unsigned int cflags)
1708 struct io_ring_ctx *ctx = req->ctx;
1710 spin_lock(&ctx->completion_lock);
1711 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1713 * If we're the last reference to this request, add to our locked
1716 if (req_ref_put_and_test(req)) {
1717 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1718 if (req->flags & IO_DISARM_MASK)
1719 io_disarm_next(req);
1721 io_req_task_queue(req->link);
1725 io_dismantle_req(req);
1726 io_put_task(req->task, 1);
1727 list_add(&req->inflight_entry, &ctx->locked_free_list);
1728 ctx->locked_free_nr++;
1730 if (!percpu_ref_tryget(&ctx->refs))
1733 io_commit_cqring(ctx);
1734 spin_unlock(&ctx->completion_lock);
1737 io_cqring_ev_posted(ctx);
1738 percpu_ref_put(&ctx->refs);
1742 static inline bool io_req_needs_clean(struct io_kiocb *req)
1744 return req->flags & IO_REQ_CLEAN_FLAGS;
1747 static void io_req_complete_state(struct io_kiocb *req, long res,
1748 unsigned int cflags)
1750 if (io_req_needs_clean(req))
1753 req->compl.cflags = cflags;
1754 req->flags |= REQ_F_COMPLETE_INLINE;
1757 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1758 long res, unsigned cflags)
1760 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1761 io_req_complete_state(req, res, cflags);
1763 io_req_complete_post(req, res, cflags);
1766 static inline void io_req_complete(struct io_kiocb *req, long res)
1768 __io_req_complete(req, 0, res, 0);
1771 static void io_req_complete_failed(struct io_kiocb *req, long res)
1774 io_req_complete_post(req, res, 0);
1778 * Don't initialise the fields below on every allocation, but do that in
1779 * advance and keep them valid across allocations.
1781 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1785 req->async_data = NULL;
1786 /* not necessary, but safer to zero */
1790 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1791 struct io_submit_state *state)
1793 spin_lock(&ctx->completion_lock);
1794 list_splice_init(&ctx->locked_free_list, &state->free_list);
1795 ctx->locked_free_nr = 0;
1796 spin_unlock(&ctx->completion_lock);
1799 /* Returns true IFF there are requests in the cache */
1800 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1802 struct io_submit_state *state = &ctx->submit_state;
1806 * If we have more than a batch's worth of requests in our IRQ side
1807 * locked cache, grab the lock and move them over to our submission
1810 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1811 io_flush_cached_locked_reqs(ctx, state);
1813 nr = state->free_reqs;
1814 while (!list_empty(&state->free_list)) {
1815 struct io_kiocb *req = list_first_entry(&state->free_list,
1816 struct io_kiocb, inflight_entry);
1818 list_del(&req->inflight_entry);
1819 state->reqs[nr++] = req;
1820 if (nr == ARRAY_SIZE(state->reqs))
1824 state->free_reqs = nr;
1829 * A request might get retired back into the request caches even before opcode
1830 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1831 * Because of that, io_alloc_req() should be called only under ->uring_lock
1832 * and with extra caution to not get a request that is still worked on.
1834 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1835 __must_hold(&ctx->uring_lock)
1837 struct io_submit_state *state = &ctx->submit_state;
1838 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1841 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1843 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1846 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1850 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1851 * retry single alloc to be on the safe side.
1853 if (unlikely(ret <= 0)) {
1854 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1855 if (!state->reqs[0])
1860 for (i = 0; i < ret; i++)
1861 io_preinit_req(state->reqs[i], ctx);
1862 state->free_reqs = ret;
1865 return state->reqs[state->free_reqs];
1868 static inline void io_put_file(struct file *file)
1874 static void io_dismantle_req(struct io_kiocb *req)
1876 unsigned int flags = req->flags;
1878 if (io_req_needs_clean(req))
1880 if (!(flags & REQ_F_FIXED_FILE))
1881 io_put_file(req->file);
1882 if (req->fixed_rsrc_refs)
1883 percpu_ref_put(req->fixed_rsrc_refs);
1884 if (req->async_data) {
1885 kfree(req->async_data);
1886 req->async_data = NULL;
1890 static void __io_free_req(struct io_kiocb *req)
1892 struct io_ring_ctx *ctx = req->ctx;
1894 io_dismantle_req(req);
1895 io_put_task(req->task, 1);
1897 spin_lock(&ctx->completion_lock);
1898 list_add(&req->inflight_entry, &ctx->locked_free_list);
1899 ctx->locked_free_nr++;
1900 spin_unlock(&ctx->completion_lock);
1902 percpu_ref_put(&ctx->refs);
1905 static inline void io_remove_next_linked(struct io_kiocb *req)
1907 struct io_kiocb *nxt = req->link;
1909 req->link = nxt->link;
1913 static bool io_kill_linked_timeout(struct io_kiocb *req)
1914 __must_hold(&req->ctx->completion_lock)
1915 __must_hold(&req->ctx->timeout_lock)
1917 struct io_kiocb *link = req->link;
1919 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1920 struct io_timeout_data *io = link->async_data;
1922 io_remove_next_linked(req);
1923 link->timeout.head = NULL;
1924 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1925 io_cqring_fill_event(link->ctx, link->user_data,
1927 io_put_req_deferred(link);
1934 static void io_fail_links(struct io_kiocb *req)
1935 __must_hold(&req->ctx->completion_lock)
1937 struct io_kiocb *nxt, *link = req->link;
1944 trace_io_uring_fail_link(req, link);
1945 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1946 io_put_req_deferred(link);
1951 static bool io_disarm_next(struct io_kiocb *req)
1952 __must_hold(&req->ctx->completion_lock)
1954 bool posted = false;
1956 if (req->flags & REQ_F_ARM_LTIMEOUT) {
1957 struct io_kiocb *link = req->link;
1959 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1960 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1961 io_remove_next_linked(req);
1962 io_cqring_fill_event(link->ctx, link->user_data,
1964 io_put_req_deferred(link);
1967 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
1968 struct io_ring_ctx *ctx = req->ctx;
1970 spin_lock_irq(&ctx->timeout_lock);
1971 posted = io_kill_linked_timeout(req);
1972 spin_unlock_irq(&ctx->timeout_lock);
1974 if (unlikely((req->flags & REQ_F_FAIL) &&
1975 !(req->flags & REQ_F_HARDLINK))) {
1976 posted |= (req->link != NULL);
1982 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1984 struct io_kiocb *nxt;
1987 * If LINK is set, we have dependent requests in this chain. If we
1988 * didn't fail this request, queue the first one up, moving any other
1989 * dependencies to the next request. In case of failure, fail the rest
1992 if (req->flags & IO_DISARM_MASK) {
1993 struct io_ring_ctx *ctx = req->ctx;
1996 spin_lock(&ctx->completion_lock);
1997 posted = io_disarm_next(req);
1999 io_commit_cqring(req->ctx);
2000 spin_unlock(&ctx->completion_lock);
2002 io_cqring_ev_posted(ctx);
2009 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2011 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2013 return __io_req_find_next(req);
2016 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2020 if (ctx->submit_state.compl_nr) {
2021 mutex_lock(&ctx->uring_lock);
2022 if (ctx->submit_state.compl_nr)
2023 io_submit_flush_completions(ctx);
2024 mutex_unlock(&ctx->uring_lock);
2026 percpu_ref_put(&ctx->refs);
2029 static void tctx_task_work(struct callback_head *cb)
2031 struct io_ring_ctx *ctx = NULL;
2032 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2036 struct io_wq_work_node *node;
2038 spin_lock_irq(&tctx->task_lock);
2039 node = tctx->task_list.first;
2040 INIT_WQ_LIST(&tctx->task_list);
2042 tctx->task_running = false;
2043 spin_unlock_irq(&tctx->task_lock);
2048 struct io_wq_work_node *next = node->next;
2049 struct io_kiocb *req = container_of(node, struct io_kiocb,
2052 if (req->ctx != ctx) {
2053 ctx_flush_and_put(ctx);
2055 percpu_ref_get(&ctx->refs);
2057 req->io_task_work.func(req);
2064 ctx_flush_and_put(ctx);
2067 static void io_req_task_work_add(struct io_kiocb *req)
2069 struct task_struct *tsk = req->task;
2070 struct io_uring_task *tctx = tsk->io_uring;
2071 enum task_work_notify_mode notify;
2072 struct io_wq_work_node *node;
2073 unsigned long flags;
2076 WARN_ON_ONCE(!tctx);
2078 spin_lock_irqsave(&tctx->task_lock, flags);
2079 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2080 running = tctx->task_running;
2082 tctx->task_running = true;
2083 spin_unlock_irqrestore(&tctx->task_lock, flags);
2085 /* task_work already pending, we're done */
2090 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2091 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2092 * processing task_work. There's no reliable way to tell if TWA_RESUME
2095 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2096 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2097 wake_up_process(tsk);
2101 spin_lock_irqsave(&tctx->task_lock, flags);
2102 tctx->task_running = false;
2103 node = tctx->task_list.first;
2104 INIT_WQ_LIST(&tctx->task_list);
2105 spin_unlock_irqrestore(&tctx->task_lock, flags);
2108 req = container_of(node, struct io_kiocb, io_task_work.node);
2110 if (llist_add(&req->io_task_work.fallback_node,
2111 &req->ctx->fallback_llist))
2112 schedule_delayed_work(&req->ctx->fallback_work, 1);
2116 static void io_req_task_cancel(struct io_kiocb *req)
2118 struct io_ring_ctx *ctx = req->ctx;
2120 /* ctx is guaranteed to stay alive while we hold uring_lock */
2121 mutex_lock(&ctx->uring_lock);
2122 io_req_complete_failed(req, req->result);
2123 mutex_unlock(&ctx->uring_lock);
2126 static void io_req_task_submit(struct io_kiocb *req)
2128 struct io_ring_ctx *ctx = req->ctx;
2130 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2131 mutex_lock(&ctx->uring_lock);
2132 /* req->task == current here, checking PF_EXITING is safe */
2133 if (likely(!(req->task->flags & PF_EXITING)))
2134 __io_queue_sqe(req);
2136 io_req_complete_failed(req, -EFAULT);
2137 mutex_unlock(&ctx->uring_lock);
2140 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2143 req->io_task_work.func = io_req_task_cancel;
2144 io_req_task_work_add(req);
2147 static void io_req_task_queue(struct io_kiocb *req)
2149 req->io_task_work.func = io_req_task_submit;
2150 io_req_task_work_add(req);
2153 static void io_req_task_queue_reissue(struct io_kiocb *req)
2155 req->io_task_work.func = io_queue_async_work;
2156 io_req_task_work_add(req);
2159 static inline void io_queue_next(struct io_kiocb *req)
2161 struct io_kiocb *nxt = io_req_find_next(req);
2164 io_req_task_queue(nxt);
2167 static void io_free_req(struct io_kiocb *req)
2174 struct task_struct *task;
2179 static inline void io_init_req_batch(struct req_batch *rb)
2186 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2187 struct req_batch *rb)
2190 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2192 io_put_task(rb->task, rb->task_refs);
2195 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2196 struct io_submit_state *state)
2199 io_dismantle_req(req);
2201 if (req->task != rb->task) {
2203 io_put_task(rb->task, rb->task_refs);
2204 rb->task = req->task;
2210 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2211 state->reqs[state->free_reqs++] = req;
2213 list_add(&req->inflight_entry, &state->free_list);
2216 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2217 __must_hold(&ctx->uring_lock)
2219 struct io_submit_state *state = &ctx->submit_state;
2220 int i, nr = state->compl_nr;
2221 struct req_batch rb;
2223 spin_lock(&ctx->completion_lock);
2224 for (i = 0; i < nr; i++) {
2225 struct io_kiocb *req = state->compl_reqs[i];
2227 __io_cqring_fill_event(ctx, req->user_data, req->result,
2230 io_commit_cqring(ctx);
2231 spin_unlock(&ctx->completion_lock);
2232 io_cqring_ev_posted(ctx);
2234 io_init_req_batch(&rb);
2235 for (i = 0; i < nr; i++) {
2236 struct io_kiocb *req = state->compl_reqs[i];
2238 if (req_ref_put_and_test(req))
2239 io_req_free_batch(&rb, req, &ctx->submit_state);
2242 io_req_free_batch_finish(ctx, &rb);
2243 state->compl_nr = 0;
2247 * Drop reference to request, return next in chain (if there is one) if this
2248 * was the last reference to this request.
2250 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2252 struct io_kiocb *nxt = NULL;
2254 if (req_ref_put_and_test(req)) {
2255 nxt = io_req_find_next(req);
2261 static inline void io_put_req(struct io_kiocb *req)
2263 if (req_ref_put_and_test(req))
2267 static inline void io_put_req_deferred(struct io_kiocb *req)
2269 if (req_ref_put_and_test(req)) {
2270 req->io_task_work.func = io_free_req;
2271 io_req_task_work_add(req);
2275 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2277 /* See comment at the top of this file */
2279 return __io_cqring_events(ctx);
2282 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2284 struct io_rings *rings = ctx->rings;
2286 /* make sure SQ entry isn't read before tail */
2287 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2290 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2292 unsigned int cflags;
2294 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2295 cflags |= IORING_CQE_F_BUFFER;
2296 req->flags &= ~REQ_F_BUFFER_SELECTED;
2301 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2303 struct io_buffer *kbuf;
2305 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2307 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2308 return io_put_kbuf(req, kbuf);
2311 static inline bool io_run_task_work(void)
2313 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2314 __set_current_state(TASK_RUNNING);
2315 tracehook_notify_signal();
2323 * Find and free completed poll iocbs
2325 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2326 struct list_head *done)
2328 struct req_batch rb;
2329 struct io_kiocb *req;
2331 /* order with ->result store in io_complete_rw_iopoll() */
2334 io_init_req_batch(&rb);
2335 while (!list_empty(done)) {
2336 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2337 list_del(&req->inflight_entry);
2339 if (READ_ONCE(req->result) == -EAGAIN &&
2340 !(req->flags & REQ_F_DONT_REISSUE)) {
2341 req->iopoll_completed = 0;
2342 io_req_task_queue_reissue(req);
2346 __io_cqring_fill_event(ctx, req->user_data, req->result,
2347 io_put_rw_kbuf(req));
2350 if (req_ref_put_and_test(req))
2351 io_req_free_batch(&rb, req, &ctx->submit_state);
2354 io_commit_cqring(ctx);
2355 io_cqring_ev_posted_iopoll(ctx);
2356 io_req_free_batch_finish(ctx, &rb);
2359 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2362 struct io_kiocb *req, *tmp;
2367 * Only spin for completions if we don't have multiple devices hanging
2368 * off our complete list, and we're under the requested amount.
2370 spin = !ctx->poll_multi_queue && *nr_events < min;
2372 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2373 struct kiocb *kiocb = &req->rw.kiocb;
2377 * Move completed and retryable entries to our local lists.
2378 * If we find a request that requires polling, break out
2379 * and complete those lists first, if we have entries there.
2381 if (READ_ONCE(req->iopoll_completed)) {
2382 list_move_tail(&req->inflight_entry, &done);
2385 if (!list_empty(&done))
2388 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2389 if (unlikely(ret < 0))
2394 /* iopoll may have completed current req */
2395 if (READ_ONCE(req->iopoll_completed))
2396 list_move_tail(&req->inflight_entry, &done);
2399 if (!list_empty(&done))
2400 io_iopoll_complete(ctx, nr_events, &done);
2406 * We can't just wait for polled events to come to us, we have to actively
2407 * find and complete them.
2409 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2411 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2414 mutex_lock(&ctx->uring_lock);
2415 while (!list_empty(&ctx->iopoll_list)) {
2416 unsigned int nr_events = 0;
2418 io_do_iopoll(ctx, &nr_events, 0);
2420 /* let it sleep and repeat later if can't complete a request */
2424 * Ensure we allow local-to-the-cpu processing to take place,
2425 * in this case we need to ensure that we reap all events.
2426 * Also let task_work, etc. to progress by releasing the mutex
2428 if (need_resched()) {
2429 mutex_unlock(&ctx->uring_lock);
2431 mutex_lock(&ctx->uring_lock);
2434 mutex_unlock(&ctx->uring_lock);
2437 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2439 unsigned int nr_events = 0;
2443 * We disallow the app entering submit/complete with polling, but we
2444 * still need to lock the ring to prevent racing with polled issue
2445 * that got punted to a workqueue.
2447 mutex_lock(&ctx->uring_lock);
2449 * Don't enter poll loop if we already have events pending.
2450 * If we do, we can potentially be spinning for commands that
2451 * already triggered a CQE (eg in error).
2453 if (test_bit(0, &ctx->check_cq_overflow))
2454 __io_cqring_overflow_flush(ctx, false);
2455 if (io_cqring_events(ctx))
2459 * If a submit got punted to a workqueue, we can have the
2460 * application entering polling for a command before it gets
2461 * issued. That app will hold the uring_lock for the duration
2462 * of the poll right here, so we need to take a breather every
2463 * now and then to ensure that the issue has a chance to add
2464 * the poll to the issued list. Otherwise we can spin here
2465 * forever, while the workqueue is stuck trying to acquire the
2468 if (list_empty(&ctx->iopoll_list)) {
2469 u32 tail = ctx->cached_cq_tail;
2471 mutex_unlock(&ctx->uring_lock);
2473 mutex_lock(&ctx->uring_lock);
2475 /* some requests don't go through iopoll_list */
2476 if (tail != ctx->cached_cq_tail ||
2477 list_empty(&ctx->iopoll_list))
2480 ret = io_do_iopoll(ctx, &nr_events, min);
2481 } while (!ret && nr_events < min && !need_resched());
2483 mutex_unlock(&ctx->uring_lock);
2487 static void kiocb_end_write(struct io_kiocb *req)
2490 * Tell lockdep we inherited freeze protection from submission
2493 if (req->flags & REQ_F_ISREG) {
2494 struct super_block *sb = file_inode(req->file)->i_sb;
2496 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2502 static bool io_resubmit_prep(struct io_kiocb *req)
2504 struct io_async_rw *rw = req->async_data;
2507 return !io_req_prep_async(req);
2508 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2509 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2513 static bool io_rw_should_reissue(struct io_kiocb *req)
2515 umode_t mode = file_inode(req->file)->i_mode;
2516 struct io_ring_ctx *ctx = req->ctx;
2518 if (!S_ISBLK(mode) && !S_ISREG(mode))
2520 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2521 !(ctx->flags & IORING_SETUP_IOPOLL)))
2524 * If ref is dying, we might be running poll reap from the exit work.
2525 * Don't attempt to reissue from that path, just let it fail with
2528 if (percpu_ref_is_dying(&ctx->refs))
2531 * Play it safe and assume not safe to re-import and reissue if we're
2532 * not in the original thread group (or in task context).
2534 if (!same_thread_group(req->task, current) || !in_task())
2539 static bool io_resubmit_prep(struct io_kiocb *req)
2543 static bool io_rw_should_reissue(struct io_kiocb *req)
2549 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2551 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2552 kiocb_end_write(req);
2553 if (res != req->result) {
2554 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2555 io_rw_should_reissue(req)) {
2556 req->flags |= REQ_F_REISSUE;
2565 static void io_req_task_complete(struct io_kiocb *req)
2567 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2570 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2571 unsigned int issue_flags)
2573 if (__io_complete_rw_common(req, res))
2575 io_req_task_complete(req);
2578 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2580 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2582 if (__io_complete_rw_common(req, res))
2585 req->io_task_work.func = io_req_task_complete;
2586 io_req_task_work_add(req);
2589 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2591 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2593 if (kiocb->ki_flags & IOCB_WRITE)
2594 kiocb_end_write(req);
2595 if (unlikely(res != req->result)) {
2596 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2597 io_resubmit_prep(req))) {
2599 req->flags |= REQ_F_DONT_REISSUE;
2603 WRITE_ONCE(req->result, res);
2604 /* order with io_iopoll_complete() checking ->result */
2606 WRITE_ONCE(req->iopoll_completed, 1);
2610 * After the iocb has been issued, it's safe to be found on the poll list.
2611 * Adding the kiocb to the list AFTER submission ensures that we don't
2612 * find it from a io_do_iopoll() thread before the issuer is done
2613 * accessing the kiocb cookie.
2615 static void io_iopoll_req_issued(struct io_kiocb *req)
2617 struct io_ring_ctx *ctx = req->ctx;
2618 const bool in_async = io_wq_current_is_worker();
2620 /* workqueue context doesn't hold uring_lock, grab it now */
2621 if (unlikely(in_async))
2622 mutex_lock(&ctx->uring_lock);
2625 * Track whether we have multiple files in our lists. This will impact
2626 * how we do polling eventually, not spinning if we're on potentially
2627 * different devices.
2629 if (list_empty(&ctx->iopoll_list)) {
2630 ctx->poll_multi_queue = false;
2631 } else if (!ctx->poll_multi_queue) {
2632 struct io_kiocb *list_req;
2633 unsigned int queue_num0, queue_num1;
2635 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2638 if (list_req->file != req->file) {
2639 ctx->poll_multi_queue = true;
2641 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2642 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2643 if (queue_num0 != queue_num1)
2644 ctx->poll_multi_queue = true;
2649 * For fast devices, IO may have already completed. If it has, add
2650 * it to the front so we find it first.
2652 if (READ_ONCE(req->iopoll_completed))
2653 list_add(&req->inflight_entry, &ctx->iopoll_list);
2655 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2657 if (unlikely(in_async)) {
2659 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2660 * in sq thread task context or in io worker task context. If
2661 * current task context is sq thread, we don't need to check
2662 * whether should wake up sq thread.
2664 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2665 wq_has_sleeper(&ctx->sq_data->wait))
2666 wake_up(&ctx->sq_data->wait);
2668 mutex_unlock(&ctx->uring_lock);
2672 static bool io_bdev_nowait(struct block_device *bdev)
2674 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2678 * If we tracked the file through the SCM inflight mechanism, we could support
2679 * any file. For now, just ensure that anything potentially problematic is done
2682 static bool __io_file_supports_nowait(struct file *file, int rw)
2684 umode_t mode = file_inode(file)->i_mode;
2686 if (S_ISBLK(mode)) {
2687 if (IS_ENABLED(CONFIG_BLOCK) &&
2688 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2694 if (S_ISREG(mode)) {
2695 if (IS_ENABLED(CONFIG_BLOCK) &&
2696 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2697 file->f_op != &io_uring_fops)
2702 /* any ->read/write should understand O_NONBLOCK */
2703 if (file->f_flags & O_NONBLOCK)
2706 if (!(file->f_mode & FMODE_NOWAIT))
2710 return file->f_op->read_iter != NULL;
2712 return file->f_op->write_iter != NULL;
2715 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2717 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2719 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2722 return __io_file_supports_nowait(req->file, rw);
2725 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2727 struct io_ring_ctx *ctx = req->ctx;
2728 struct kiocb *kiocb = &req->rw.kiocb;
2729 struct file *file = req->file;
2733 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2734 req->flags |= REQ_F_ISREG;
2736 kiocb->ki_pos = READ_ONCE(sqe->off);
2737 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2738 req->flags |= REQ_F_CUR_POS;
2739 kiocb->ki_pos = file->f_pos;
2741 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2742 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2743 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2747 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2748 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2749 req->flags |= REQ_F_NOWAIT;
2751 ioprio = READ_ONCE(sqe->ioprio);
2753 ret = ioprio_check_cap(ioprio);
2757 kiocb->ki_ioprio = ioprio;
2759 kiocb->ki_ioprio = get_current_ioprio();
2761 if (ctx->flags & IORING_SETUP_IOPOLL) {
2762 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2763 !kiocb->ki_filp->f_op->iopoll)
2766 kiocb->ki_flags |= IOCB_HIPRI;
2767 kiocb->ki_complete = io_complete_rw_iopoll;
2768 req->iopoll_completed = 0;
2770 if (kiocb->ki_flags & IOCB_HIPRI)
2772 kiocb->ki_complete = io_complete_rw;
2775 if (req->opcode == IORING_OP_READ_FIXED ||
2776 req->opcode == IORING_OP_WRITE_FIXED) {
2778 io_req_set_rsrc_node(req);
2781 req->rw.addr = READ_ONCE(sqe->addr);
2782 req->rw.len = READ_ONCE(sqe->len);
2783 req->buf_index = READ_ONCE(sqe->buf_index);
2787 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2793 case -ERESTARTNOINTR:
2794 case -ERESTARTNOHAND:
2795 case -ERESTART_RESTARTBLOCK:
2797 * We can't just restart the syscall, since previously
2798 * submitted sqes may already be in progress. Just fail this
2804 kiocb->ki_complete(kiocb, ret, 0);
2808 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2809 unsigned int issue_flags)
2811 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2812 struct io_async_rw *io = req->async_data;
2813 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2815 /* add previously done IO, if any */
2816 if (io && io->bytes_done > 0) {
2818 ret = io->bytes_done;
2820 ret += io->bytes_done;
2823 if (req->flags & REQ_F_CUR_POS)
2824 req->file->f_pos = kiocb->ki_pos;
2825 if (ret >= 0 && check_reissue)
2826 __io_complete_rw(req, ret, 0, issue_flags);
2828 io_rw_done(kiocb, ret);
2830 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2831 req->flags &= ~REQ_F_REISSUE;
2832 if (io_resubmit_prep(req)) {
2833 io_req_task_queue_reissue(req);
2836 __io_req_complete(req, issue_flags, ret,
2837 io_put_rw_kbuf(req));
2842 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2843 struct io_mapped_ubuf *imu)
2845 size_t len = req->rw.len;
2846 u64 buf_end, buf_addr = req->rw.addr;
2849 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2851 /* not inside the mapped region */
2852 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2856 * May not be a start of buffer, set size appropriately
2857 * and advance us to the beginning.
2859 offset = buf_addr - imu->ubuf;
2860 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2864 * Don't use iov_iter_advance() here, as it's really slow for
2865 * using the latter parts of a big fixed buffer - it iterates
2866 * over each segment manually. We can cheat a bit here, because
2869 * 1) it's a BVEC iter, we set it up
2870 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2871 * first and last bvec
2873 * So just find our index, and adjust the iterator afterwards.
2874 * If the offset is within the first bvec (or the whole first
2875 * bvec, just use iov_iter_advance(). This makes it easier
2876 * since we can just skip the first segment, which may not
2877 * be PAGE_SIZE aligned.
2879 const struct bio_vec *bvec = imu->bvec;
2881 if (offset <= bvec->bv_len) {
2882 iov_iter_advance(iter, offset);
2884 unsigned long seg_skip;
2886 /* skip first vec */
2887 offset -= bvec->bv_len;
2888 seg_skip = 1 + (offset >> PAGE_SHIFT);
2890 iter->bvec = bvec + seg_skip;
2891 iter->nr_segs -= seg_skip;
2892 iter->count -= bvec->bv_len + offset;
2893 iter->iov_offset = offset & ~PAGE_MASK;
2900 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2902 struct io_ring_ctx *ctx = req->ctx;
2903 struct io_mapped_ubuf *imu = req->imu;
2904 u16 index, buf_index = req->buf_index;
2907 if (unlikely(buf_index >= ctx->nr_user_bufs))
2909 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2910 imu = READ_ONCE(ctx->user_bufs[index]);
2913 return __io_import_fixed(req, rw, iter, imu);
2916 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2919 mutex_unlock(&ctx->uring_lock);
2922 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2925 * "Normal" inline submissions always hold the uring_lock, since we
2926 * grab it from the system call. Same is true for the SQPOLL offload.
2927 * The only exception is when we've detached the request and issue it
2928 * from an async worker thread, grab the lock for that case.
2931 mutex_lock(&ctx->uring_lock);
2934 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2935 int bgid, struct io_buffer *kbuf,
2938 struct io_buffer *head;
2940 if (req->flags & REQ_F_BUFFER_SELECTED)
2943 io_ring_submit_lock(req->ctx, needs_lock);
2945 lockdep_assert_held(&req->ctx->uring_lock);
2947 head = xa_load(&req->ctx->io_buffers, bgid);
2949 if (!list_empty(&head->list)) {
2950 kbuf = list_last_entry(&head->list, struct io_buffer,
2952 list_del(&kbuf->list);
2955 xa_erase(&req->ctx->io_buffers, bgid);
2957 if (*len > kbuf->len)
2960 kbuf = ERR_PTR(-ENOBUFS);
2963 io_ring_submit_unlock(req->ctx, needs_lock);
2968 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2971 struct io_buffer *kbuf;
2974 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2975 bgid = req->buf_index;
2976 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2979 req->rw.addr = (u64) (unsigned long) kbuf;
2980 req->flags |= REQ_F_BUFFER_SELECTED;
2981 return u64_to_user_ptr(kbuf->addr);
2984 #ifdef CONFIG_COMPAT
2985 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2988 struct compat_iovec __user *uiov;
2989 compat_ssize_t clen;
2993 uiov = u64_to_user_ptr(req->rw.addr);
2994 if (!access_ok(uiov, sizeof(*uiov)))
2996 if (__get_user(clen, &uiov->iov_len))
3002 buf = io_rw_buffer_select(req, &len, needs_lock);
3004 return PTR_ERR(buf);
3005 iov[0].iov_base = buf;
3006 iov[0].iov_len = (compat_size_t) len;
3011 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3014 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3018 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3021 len = iov[0].iov_len;
3024 buf = io_rw_buffer_select(req, &len, needs_lock);
3026 return PTR_ERR(buf);
3027 iov[0].iov_base = buf;
3028 iov[0].iov_len = len;
3032 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3035 if (req->flags & REQ_F_BUFFER_SELECTED) {
3036 struct io_buffer *kbuf;
3038 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3039 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3040 iov[0].iov_len = kbuf->len;
3043 if (req->rw.len != 1)
3046 #ifdef CONFIG_COMPAT
3047 if (req->ctx->compat)
3048 return io_compat_import(req, iov, needs_lock);
3051 return __io_iov_buffer_select(req, iov, needs_lock);
3054 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3055 struct iov_iter *iter, bool needs_lock)
3057 void __user *buf = u64_to_user_ptr(req->rw.addr);
3058 size_t sqe_len = req->rw.len;
3059 u8 opcode = req->opcode;
3062 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3064 return io_import_fixed(req, rw, iter);
3067 /* buffer index only valid with fixed read/write, or buffer select */
3068 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3071 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3072 if (req->flags & REQ_F_BUFFER_SELECT) {
3073 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3075 return PTR_ERR(buf);
3076 req->rw.len = sqe_len;
3079 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3084 if (req->flags & REQ_F_BUFFER_SELECT) {
3085 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3087 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3092 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3096 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3098 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3102 * For files that don't have ->read_iter() and ->write_iter(), handle them
3103 * by looping over ->read() or ->write() manually.
3105 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3107 struct kiocb *kiocb = &req->rw.kiocb;
3108 struct file *file = req->file;
3112 * Don't support polled IO through this interface, and we can't
3113 * support non-blocking either. For the latter, this just causes
3114 * the kiocb to be handled from an async context.
3116 if (kiocb->ki_flags & IOCB_HIPRI)
3118 if (kiocb->ki_flags & IOCB_NOWAIT)
3121 while (iov_iter_count(iter)) {
3125 if (!iov_iter_is_bvec(iter)) {
3126 iovec = iov_iter_iovec(iter);
3128 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3129 iovec.iov_len = req->rw.len;
3133 nr = file->f_op->read(file, iovec.iov_base,
3134 iovec.iov_len, io_kiocb_ppos(kiocb));
3136 nr = file->f_op->write(file, iovec.iov_base,
3137 iovec.iov_len, io_kiocb_ppos(kiocb));
3146 if (nr != iovec.iov_len)
3150 iov_iter_advance(iter, nr);
3156 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3157 const struct iovec *fast_iov, struct iov_iter *iter)
3159 struct io_async_rw *rw = req->async_data;
3161 memcpy(&rw->iter, iter, sizeof(*iter));
3162 rw->free_iovec = iovec;
3164 /* can only be fixed buffers, no need to do anything */
3165 if (iov_iter_is_bvec(iter))
3168 unsigned iov_off = 0;
3170 rw->iter.iov = rw->fast_iov;
3171 if (iter->iov != fast_iov) {
3172 iov_off = iter->iov - fast_iov;
3173 rw->iter.iov += iov_off;
3175 if (rw->fast_iov != fast_iov)
3176 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3177 sizeof(struct iovec) * iter->nr_segs);
3179 req->flags |= REQ_F_NEED_CLEANUP;
3183 static inline int io_alloc_async_data(struct io_kiocb *req)
3185 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3186 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3187 return req->async_data == NULL;
3190 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3191 const struct iovec *fast_iov,
3192 struct iov_iter *iter, bool force)
3194 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3196 if (!req->async_data) {
3197 if (io_alloc_async_data(req)) {
3202 io_req_map_rw(req, iovec, fast_iov, iter);
3207 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3209 struct io_async_rw *iorw = req->async_data;
3210 struct iovec *iov = iorw->fast_iov;
3213 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3214 if (unlikely(ret < 0))
3217 iorw->bytes_done = 0;
3218 iorw->free_iovec = iov;
3220 req->flags |= REQ_F_NEED_CLEANUP;
3224 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3226 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3228 return io_prep_rw(req, sqe);
3232 * This is our waitqueue callback handler, registered through lock_page_async()
3233 * when we initially tried to do the IO with the iocb armed our waitqueue.
3234 * This gets called when the page is unlocked, and we generally expect that to
3235 * happen when the page IO is completed and the page is now uptodate. This will
3236 * queue a task_work based retry of the operation, attempting to copy the data
3237 * again. If the latter fails because the page was NOT uptodate, then we will
3238 * do a thread based blocking retry of the operation. That's the unexpected
3241 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3242 int sync, void *arg)
3244 struct wait_page_queue *wpq;
3245 struct io_kiocb *req = wait->private;
3246 struct wait_page_key *key = arg;
3248 wpq = container_of(wait, struct wait_page_queue, wait);
3250 if (!wake_page_match(wpq, key))
3253 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3254 list_del_init(&wait->entry);
3255 io_req_task_queue(req);
3260 * This controls whether a given IO request should be armed for async page
3261 * based retry. If we return false here, the request is handed to the async
3262 * worker threads for retry. If we're doing buffered reads on a regular file,
3263 * we prepare a private wait_page_queue entry and retry the operation. This
3264 * will either succeed because the page is now uptodate and unlocked, or it
3265 * will register a callback when the page is unlocked at IO completion. Through
3266 * that callback, io_uring uses task_work to setup a retry of the operation.
3267 * That retry will attempt the buffered read again. The retry will generally
3268 * succeed, or in rare cases where it fails, we then fall back to using the
3269 * async worker threads for a blocking retry.
3271 static bool io_rw_should_retry(struct io_kiocb *req)
3273 struct io_async_rw *rw = req->async_data;
3274 struct wait_page_queue *wait = &rw->wpq;
3275 struct kiocb *kiocb = &req->rw.kiocb;
3277 /* never retry for NOWAIT, we just complete with -EAGAIN */
3278 if (req->flags & REQ_F_NOWAIT)
3281 /* Only for buffered IO */
3282 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3286 * just use poll if we can, and don't attempt if the fs doesn't
3287 * support callback based unlocks
3289 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3292 wait->wait.func = io_async_buf_func;
3293 wait->wait.private = req;
3294 wait->wait.flags = 0;
3295 INIT_LIST_HEAD(&wait->wait.entry);
3296 kiocb->ki_flags |= IOCB_WAITQ;
3297 kiocb->ki_flags &= ~IOCB_NOWAIT;
3298 kiocb->ki_waitq = wait;
3302 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3304 if (req->file->f_op->read_iter)
3305 return call_read_iter(req->file, &req->rw.kiocb, iter);
3306 else if (req->file->f_op->read)
3307 return loop_rw_iter(READ, req, iter);
3312 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3314 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3315 struct kiocb *kiocb = &req->rw.kiocb;
3316 struct iov_iter __iter, *iter = &__iter;
3317 struct io_async_rw *rw = req->async_data;
3318 ssize_t io_size, ret, ret2;
3319 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3325 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3329 io_size = iov_iter_count(iter);
3330 req->result = io_size;
3332 /* Ensure we clear previously set non-block flag */
3333 if (!force_nonblock)
3334 kiocb->ki_flags &= ~IOCB_NOWAIT;
3336 kiocb->ki_flags |= IOCB_NOWAIT;
3338 /* If the file doesn't support async, just async punt */
3339 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3340 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3341 return ret ?: -EAGAIN;
3344 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3345 if (unlikely(ret)) {
3350 ret = io_iter_do_read(req, iter);
3352 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3353 req->flags &= ~REQ_F_REISSUE;
3354 /* IOPOLL retry should happen for io-wq threads */
3355 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3357 /* no retry on NONBLOCK nor RWF_NOWAIT */
3358 if (req->flags & REQ_F_NOWAIT)
3360 /* some cases will consume bytes even on error returns */
3361 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3363 } else if (ret == -EIOCBQUEUED) {
3365 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3366 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3367 /* read all, failed, already did sync or don't want to retry */
3371 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3376 rw = req->async_data;
3377 /* now use our persistent iterator, if we aren't already */
3382 rw->bytes_done += ret;
3383 /* if we can retry, do so with the callbacks armed */
3384 if (!io_rw_should_retry(req)) {
3385 kiocb->ki_flags &= ~IOCB_WAITQ;
3390 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3391 * we get -EIOCBQUEUED, then we'll get a notification when the
3392 * desired page gets unlocked. We can also get a partial read
3393 * here, and if we do, then just retry at the new offset.
3395 ret = io_iter_do_read(req, iter);
3396 if (ret == -EIOCBQUEUED)
3398 /* we got some bytes, but not all. retry. */
3399 kiocb->ki_flags &= ~IOCB_WAITQ;
3400 } while (ret > 0 && ret < io_size);
3402 kiocb_done(kiocb, ret, issue_flags);
3404 /* it's faster to check here then delegate to kfree */
3410 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3412 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3414 return io_prep_rw(req, sqe);
3417 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3419 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3420 struct kiocb *kiocb = &req->rw.kiocb;
3421 struct iov_iter __iter, *iter = &__iter;
3422 struct io_async_rw *rw = req->async_data;
3423 ssize_t ret, ret2, io_size;
3424 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3430 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3434 io_size = iov_iter_count(iter);
3435 req->result = io_size;
3437 /* Ensure we clear previously set non-block flag */
3438 if (!force_nonblock)
3439 kiocb->ki_flags &= ~IOCB_NOWAIT;
3441 kiocb->ki_flags |= IOCB_NOWAIT;
3443 /* If the file doesn't support async, just async punt */
3444 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3447 /* file path doesn't support NOWAIT for non-direct_IO */
3448 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3449 (req->flags & REQ_F_ISREG))
3452 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3457 * Open-code file_start_write here to grab freeze protection,
3458 * which will be released by another thread in
3459 * io_complete_rw(). Fool lockdep by telling it the lock got
3460 * released so that it doesn't complain about the held lock when
3461 * we return to userspace.
3463 if (req->flags & REQ_F_ISREG) {
3464 sb_start_write(file_inode(req->file)->i_sb);
3465 __sb_writers_release(file_inode(req->file)->i_sb,
3468 kiocb->ki_flags |= IOCB_WRITE;
3470 if (req->file->f_op->write_iter)
3471 ret2 = call_write_iter(req->file, kiocb, iter);
3472 else if (req->file->f_op->write)
3473 ret2 = loop_rw_iter(WRITE, req, iter);
3477 if (req->flags & REQ_F_REISSUE) {
3478 req->flags &= ~REQ_F_REISSUE;
3483 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3484 * retry them without IOCB_NOWAIT.
3486 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3488 /* no retry on NONBLOCK nor RWF_NOWAIT */
3489 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3491 if (!force_nonblock || ret2 != -EAGAIN) {
3492 /* IOPOLL retry should happen for io-wq threads */
3493 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3496 kiocb_done(kiocb, ret2, issue_flags);
3499 /* some cases will consume bytes even on error returns */
3500 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3501 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3502 return ret ?: -EAGAIN;
3505 /* it's reportedly faster than delegating the null check to kfree() */
3511 static int io_renameat_prep(struct io_kiocb *req,
3512 const struct io_uring_sqe *sqe)
3514 struct io_rename *ren = &req->rename;
3515 const char __user *oldf, *newf;
3517 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3519 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3521 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3524 ren->old_dfd = READ_ONCE(sqe->fd);
3525 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3526 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3527 ren->new_dfd = READ_ONCE(sqe->len);
3528 ren->flags = READ_ONCE(sqe->rename_flags);
3530 ren->oldpath = getname(oldf);
3531 if (IS_ERR(ren->oldpath))
3532 return PTR_ERR(ren->oldpath);
3534 ren->newpath = getname(newf);
3535 if (IS_ERR(ren->newpath)) {
3536 putname(ren->oldpath);
3537 return PTR_ERR(ren->newpath);
3540 req->flags |= REQ_F_NEED_CLEANUP;
3544 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3546 struct io_rename *ren = &req->rename;
3549 if (issue_flags & IO_URING_F_NONBLOCK)
3552 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3553 ren->newpath, ren->flags);
3555 req->flags &= ~REQ_F_NEED_CLEANUP;
3558 io_req_complete(req, ret);
3562 static int io_unlinkat_prep(struct io_kiocb *req,
3563 const struct io_uring_sqe *sqe)
3565 struct io_unlink *un = &req->unlink;
3566 const char __user *fname;
3568 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3570 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3573 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3576 un->dfd = READ_ONCE(sqe->fd);
3578 un->flags = READ_ONCE(sqe->unlink_flags);
3579 if (un->flags & ~AT_REMOVEDIR)
3582 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3583 un->filename = getname(fname);
3584 if (IS_ERR(un->filename))
3585 return PTR_ERR(un->filename);
3587 req->flags |= REQ_F_NEED_CLEANUP;
3591 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3593 struct io_unlink *un = &req->unlink;
3596 if (issue_flags & IO_URING_F_NONBLOCK)
3599 if (un->flags & AT_REMOVEDIR)
3600 ret = do_rmdir(un->dfd, un->filename);
3602 ret = do_unlinkat(un->dfd, un->filename);
3604 req->flags &= ~REQ_F_NEED_CLEANUP;
3607 io_req_complete(req, ret);
3611 static int io_shutdown_prep(struct io_kiocb *req,
3612 const struct io_uring_sqe *sqe)
3614 #if defined(CONFIG_NET)
3615 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3617 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3618 sqe->buf_index || sqe->splice_fd_in))
3621 req->shutdown.how = READ_ONCE(sqe->len);
3628 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3630 #if defined(CONFIG_NET)
3631 struct socket *sock;
3634 if (issue_flags & IO_URING_F_NONBLOCK)
3637 sock = sock_from_file(req->file);
3638 if (unlikely(!sock))
3641 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3644 io_req_complete(req, ret);
3651 static int __io_splice_prep(struct io_kiocb *req,
3652 const struct io_uring_sqe *sqe)
3654 struct io_splice *sp = &req->splice;
3655 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3657 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3661 sp->len = READ_ONCE(sqe->len);
3662 sp->flags = READ_ONCE(sqe->splice_flags);
3664 if (unlikely(sp->flags & ~valid_flags))
3667 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3668 (sp->flags & SPLICE_F_FD_IN_FIXED));
3671 req->flags |= REQ_F_NEED_CLEANUP;
3675 static int io_tee_prep(struct io_kiocb *req,
3676 const struct io_uring_sqe *sqe)
3678 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3680 return __io_splice_prep(req, sqe);
3683 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3685 struct io_splice *sp = &req->splice;
3686 struct file *in = sp->file_in;
3687 struct file *out = sp->file_out;
3688 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3691 if (issue_flags & IO_URING_F_NONBLOCK)
3694 ret = do_tee(in, out, sp->len, flags);
3696 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3698 req->flags &= ~REQ_F_NEED_CLEANUP;
3702 io_req_complete(req, ret);
3706 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3708 struct io_splice *sp = &req->splice;
3710 sp->off_in = READ_ONCE(sqe->splice_off_in);
3711 sp->off_out = READ_ONCE(sqe->off);
3712 return __io_splice_prep(req, sqe);
3715 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3717 struct io_splice *sp = &req->splice;
3718 struct file *in = sp->file_in;
3719 struct file *out = sp->file_out;
3720 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3721 loff_t *poff_in, *poff_out;
3724 if (issue_flags & IO_URING_F_NONBLOCK)
3727 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3728 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3731 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3733 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3735 req->flags &= ~REQ_F_NEED_CLEANUP;
3739 io_req_complete(req, ret);
3744 * IORING_OP_NOP just posts a completion event, nothing else.
3746 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3748 struct io_ring_ctx *ctx = req->ctx;
3750 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3753 __io_req_complete(req, issue_flags, 0, 0);
3757 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3759 struct io_ring_ctx *ctx = req->ctx;
3764 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3766 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3770 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3771 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3774 req->sync.off = READ_ONCE(sqe->off);
3775 req->sync.len = READ_ONCE(sqe->len);
3779 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3781 loff_t end = req->sync.off + req->sync.len;
3784 /* fsync always requires a blocking context */
3785 if (issue_flags & IO_URING_F_NONBLOCK)
3788 ret = vfs_fsync_range(req->file, req->sync.off,
3789 end > 0 ? end : LLONG_MAX,
3790 req->sync.flags & IORING_FSYNC_DATASYNC);
3793 io_req_complete(req, ret);
3797 static int io_fallocate_prep(struct io_kiocb *req,
3798 const struct io_uring_sqe *sqe)
3800 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3806 req->sync.off = READ_ONCE(sqe->off);
3807 req->sync.len = READ_ONCE(sqe->addr);
3808 req->sync.mode = READ_ONCE(sqe->len);
3812 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3816 /* fallocate always requiring blocking context */
3817 if (issue_flags & IO_URING_F_NONBLOCK)
3819 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3823 io_req_complete(req, ret);
3827 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3829 const char __user *fname;
3832 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3834 if (unlikely(sqe->ioprio || sqe->buf_index || sqe->splice_fd_in))
3836 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3839 /* open.how should be already initialised */
3840 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3841 req->open.how.flags |= O_LARGEFILE;
3843 req->open.dfd = READ_ONCE(sqe->fd);
3844 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3845 req->open.filename = getname(fname);
3846 if (IS_ERR(req->open.filename)) {
3847 ret = PTR_ERR(req->open.filename);
3848 req->open.filename = NULL;
3851 req->open.nofile = rlimit(RLIMIT_NOFILE);
3852 req->flags |= REQ_F_NEED_CLEANUP;
3856 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3858 u64 mode = READ_ONCE(sqe->len);
3859 u64 flags = READ_ONCE(sqe->open_flags);
3861 req->open.how = build_open_how(flags, mode);
3862 return __io_openat_prep(req, sqe);
3865 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3867 struct open_how __user *how;
3871 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3872 len = READ_ONCE(sqe->len);
3873 if (len < OPEN_HOW_SIZE_VER0)
3876 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3881 return __io_openat_prep(req, sqe);
3884 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3886 struct open_flags op;
3889 bool resolve_nonblock;
3892 ret = build_open_flags(&req->open.how, &op);
3895 nonblock_set = op.open_flag & O_NONBLOCK;
3896 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3897 if (issue_flags & IO_URING_F_NONBLOCK) {
3899 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3900 * it'll always -EAGAIN
3902 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3904 op.lookup_flags |= LOOKUP_CACHED;
3905 op.open_flag |= O_NONBLOCK;
3908 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3912 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3915 * We could hang on to this 'fd' on retrying, but seems like
3916 * marginal gain for something that is now known to be a slower
3917 * path. So just put it, and we'll get a new one when we retry.
3921 ret = PTR_ERR(file);
3922 /* only retry if RESOLVE_CACHED wasn't already set by application */
3923 if (ret == -EAGAIN &&
3924 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3929 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3930 file->f_flags &= ~O_NONBLOCK;
3931 fsnotify_open(file);
3932 fd_install(ret, file);
3934 putname(req->open.filename);
3935 req->flags &= ~REQ_F_NEED_CLEANUP;
3938 __io_req_complete(req, issue_flags, ret, 0);
3942 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3944 return io_openat2(req, issue_flags);
3947 static int io_remove_buffers_prep(struct io_kiocb *req,
3948 const struct io_uring_sqe *sqe)
3950 struct io_provide_buf *p = &req->pbuf;
3953 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
3957 tmp = READ_ONCE(sqe->fd);
3958 if (!tmp || tmp > USHRT_MAX)
3961 memset(p, 0, sizeof(*p));
3963 p->bgid = READ_ONCE(sqe->buf_group);
3967 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3968 int bgid, unsigned nbufs)
3972 /* shouldn't happen */
3976 /* the head kbuf is the list itself */
3977 while (!list_empty(&buf->list)) {
3978 struct io_buffer *nxt;
3980 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3981 list_del(&nxt->list);
3988 xa_erase(&ctx->io_buffers, bgid);
3993 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3995 struct io_provide_buf *p = &req->pbuf;
3996 struct io_ring_ctx *ctx = req->ctx;
3997 struct io_buffer *head;
3999 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4001 io_ring_submit_lock(ctx, !force_nonblock);
4003 lockdep_assert_held(&ctx->uring_lock);
4006 head = xa_load(&ctx->io_buffers, p->bgid);
4008 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4012 /* complete before unlock, IOPOLL may need the lock */
4013 __io_req_complete(req, issue_flags, ret, 0);
4014 io_ring_submit_unlock(ctx, !force_nonblock);
4018 static int io_provide_buffers_prep(struct io_kiocb *req,
4019 const struct io_uring_sqe *sqe)
4021 unsigned long size, tmp_check;
4022 struct io_provide_buf *p = &req->pbuf;
4025 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4028 tmp = READ_ONCE(sqe->fd);
4029 if (!tmp || tmp > USHRT_MAX)
4032 p->addr = READ_ONCE(sqe->addr);
4033 p->len = READ_ONCE(sqe->len);
4035 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4038 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4041 size = (unsigned long)p->len * p->nbufs;
4042 if (!access_ok(u64_to_user_ptr(p->addr), size))
4045 p->bgid = READ_ONCE(sqe->buf_group);
4046 tmp = READ_ONCE(sqe->off);
4047 if (tmp > USHRT_MAX)
4053 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4055 struct io_buffer *buf;
4056 u64 addr = pbuf->addr;
4057 int i, bid = pbuf->bid;
4059 for (i = 0; i < pbuf->nbufs; i++) {
4060 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4065 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4070 INIT_LIST_HEAD(&buf->list);
4073 list_add_tail(&buf->list, &(*head)->list);
4077 return i ? i : -ENOMEM;
4080 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4082 struct io_provide_buf *p = &req->pbuf;
4083 struct io_ring_ctx *ctx = req->ctx;
4084 struct io_buffer *head, *list;
4086 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4088 io_ring_submit_lock(ctx, !force_nonblock);
4090 lockdep_assert_held(&ctx->uring_lock);
4092 list = head = xa_load(&ctx->io_buffers, p->bgid);
4094 ret = io_add_buffers(p, &head);
4095 if (ret >= 0 && !list) {
4096 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4098 __io_remove_buffers(ctx, head, p->bgid, -1U);
4102 /* complete before unlock, IOPOLL may need the lock */
4103 __io_req_complete(req, issue_flags, ret, 0);
4104 io_ring_submit_unlock(ctx, !force_nonblock);
4108 static int io_epoll_ctl_prep(struct io_kiocb *req,
4109 const struct io_uring_sqe *sqe)
4111 #if defined(CONFIG_EPOLL)
4112 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4114 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4117 req->epoll.epfd = READ_ONCE(sqe->fd);
4118 req->epoll.op = READ_ONCE(sqe->len);
4119 req->epoll.fd = READ_ONCE(sqe->off);
4121 if (ep_op_has_event(req->epoll.op)) {
4122 struct epoll_event __user *ev;
4124 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4125 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4135 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4137 #if defined(CONFIG_EPOLL)
4138 struct io_epoll *ie = &req->epoll;
4140 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4142 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4143 if (force_nonblock && ret == -EAGAIN)
4148 __io_req_complete(req, issue_flags, ret, 0);
4155 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4157 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4158 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4160 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4163 req->madvise.addr = READ_ONCE(sqe->addr);
4164 req->madvise.len = READ_ONCE(sqe->len);
4165 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4172 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4174 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4175 struct io_madvise *ma = &req->madvise;
4178 if (issue_flags & IO_URING_F_NONBLOCK)
4181 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4184 io_req_complete(req, ret);
4191 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4193 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4195 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4198 req->fadvise.offset = READ_ONCE(sqe->off);
4199 req->fadvise.len = READ_ONCE(sqe->len);
4200 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4204 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4206 struct io_fadvise *fa = &req->fadvise;
4209 if (issue_flags & IO_URING_F_NONBLOCK) {
4210 switch (fa->advice) {
4211 case POSIX_FADV_NORMAL:
4212 case POSIX_FADV_RANDOM:
4213 case POSIX_FADV_SEQUENTIAL:
4220 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4223 __io_req_complete(req, issue_flags, ret, 0);
4227 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4229 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4231 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4233 if (req->flags & REQ_F_FIXED_FILE)
4236 req->statx.dfd = READ_ONCE(sqe->fd);
4237 req->statx.mask = READ_ONCE(sqe->len);
4238 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4239 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4240 req->statx.flags = READ_ONCE(sqe->statx_flags);
4245 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4247 struct io_statx *ctx = &req->statx;
4250 if (issue_flags & IO_URING_F_NONBLOCK)
4253 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4258 io_req_complete(req, ret);
4262 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4264 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4266 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4267 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4269 if (req->flags & REQ_F_FIXED_FILE)
4272 req->close.fd = READ_ONCE(sqe->fd);
4276 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4278 struct files_struct *files = current->files;
4279 struct io_close *close = &req->close;
4280 struct fdtable *fdt;
4281 struct file *file = NULL;
4284 spin_lock(&files->file_lock);
4285 fdt = files_fdtable(files);
4286 if (close->fd >= fdt->max_fds) {
4287 spin_unlock(&files->file_lock);
4290 file = fdt->fd[close->fd];
4291 if (!file || file->f_op == &io_uring_fops) {
4292 spin_unlock(&files->file_lock);
4297 /* if the file has a flush method, be safe and punt to async */
4298 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4299 spin_unlock(&files->file_lock);
4303 ret = __close_fd_get_file(close->fd, &file);
4304 spin_unlock(&files->file_lock);
4311 /* No ->flush() or already async, safely close from here */
4312 ret = filp_close(file, current->files);
4318 __io_req_complete(req, issue_flags, ret, 0);
4322 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4324 struct io_ring_ctx *ctx = req->ctx;
4326 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4328 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4332 req->sync.off = READ_ONCE(sqe->off);
4333 req->sync.len = READ_ONCE(sqe->len);
4334 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4338 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4342 /* sync_file_range always requires a blocking context */
4343 if (issue_flags & IO_URING_F_NONBLOCK)
4346 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4350 io_req_complete(req, ret);
4354 #if defined(CONFIG_NET)
4355 static int io_setup_async_msg(struct io_kiocb *req,
4356 struct io_async_msghdr *kmsg)
4358 struct io_async_msghdr *async_msg = req->async_data;
4362 if (io_alloc_async_data(req)) {
4363 kfree(kmsg->free_iov);
4366 async_msg = req->async_data;
4367 req->flags |= REQ_F_NEED_CLEANUP;
4368 memcpy(async_msg, kmsg, sizeof(*kmsg));
4369 async_msg->msg.msg_name = &async_msg->addr;
4370 /* if were using fast_iov, set it to the new one */
4371 if (!async_msg->free_iov)
4372 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4377 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4378 struct io_async_msghdr *iomsg)
4380 iomsg->msg.msg_name = &iomsg->addr;
4381 iomsg->free_iov = iomsg->fast_iov;
4382 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4383 req->sr_msg.msg_flags, &iomsg->free_iov);
4386 static int io_sendmsg_prep_async(struct io_kiocb *req)
4390 ret = io_sendmsg_copy_hdr(req, req->async_data);
4392 req->flags |= REQ_F_NEED_CLEANUP;
4396 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4398 struct io_sr_msg *sr = &req->sr_msg;
4400 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4403 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4404 sr->len = READ_ONCE(sqe->len);
4405 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4406 if (sr->msg_flags & MSG_DONTWAIT)
4407 req->flags |= REQ_F_NOWAIT;
4409 #ifdef CONFIG_COMPAT
4410 if (req->ctx->compat)
4411 sr->msg_flags |= MSG_CMSG_COMPAT;
4416 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4418 struct io_async_msghdr iomsg, *kmsg;
4419 struct socket *sock;
4424 sock = sock_from_file(req->file);
4425 if (unlikely(!sock))
4428 kmsg = req->async_data;
4430 ret = io_sendmsg_copy_hdr(req, &iomsg);
4436 flags = req->sr_msg.msg_flags;
4437 if (issue_flags & IO_URING_F_NONBLOCK)
4438 flags |= MSG_DONTWAIT;
4439 if (flags & MSG_WAITALL)
4440 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4442 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4443 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4444 return io_setup_async_msg(req, kmsg);
4445 if (ret == -ERESTARTSYS)
4448 /* fast path, check for non-NULL to avoid function call */
4450 kfree(kmsg->free_iov);
4451 req->flags &= ~REQ_F_NEED_CLEANUP;
4454 __io_req_complete(req, issue_flags, ret, 0);
4458 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4460 struct io_sr_msg *sr = &req->sr_msg;
4463 struct socket *sock;
4468 sock = sock_from_file(req->file);
4469 if (unlikely(!sock))
4472 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4476 msg.msg_name = NULL;
4477 msg.msg_control = NULL;
4478 msg.msg_controllen = 0;
4479 msg.msg_namelen = 0;
4481 flags = req->sr_msg.msg_flags;
4482 if (issue_flags & IO_URING_F_NONBLOCK)
4483 flags |= MSG_DONTWAIT;
4484 if (flags & MSG_WAITALL)
4485 min_ret = iov_iter_count(&msg.msg_iter);
4487 msg.msg_flags = flags;
4488 ret = sock_sendmsg(sock, &msg);
4489 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4491 if (ret == -ERESTARTSYS)
4496 __io_req_complete(req, issue_flags, ret, 0);
4500 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4501 struct io_async_msghdr *iomsg)
4503 struct io_sr_msg *sr = &req->sr_msg;
4504 struct iovec __user *uiov;
4508 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4509 &iomsg->uaddr, &uiov, &iov_len);
4513 if (req->flags & REQ_F_BUFFER_SELECT) {
4516 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4518 sr->len = iomsg->fast_iov[0].iov_len;
4519 iomsg->free_iov = NULL;
4521 iomsg->free_iov = iomsg->fast_iov;
4522 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4523 &iomsg->free_iov, &iomsg->msg.msg_iter,
4532 #ifdef CONFIG_COMPAT
4533 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4534 struct io_async_msghdr *iomsg)
4536 struct io_sr_msg *sr = &req->sr_msg;
4537 struct compat_iovec __user *uiov;
4542 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4547 uiov = compat_ptr(ptr);
4548 if (req->flags & REQ_F_BUFFER_SELECT) {
4549 compat_ssize_t clen;
4553 if (!access_ok(uiov, sizeof(*uiov)))
4555 if (__get_user(clen, &uiov->iov_len))
4560 iomsg->free_iov = NULL;
4562 iomsg->free_iov = iomsg->fast_iov;
4563 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4564 UIO_FASTIOV, &iomsg->free_iov,
4565 &iomsg->msg.msg_iter, true);
4574 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4575 struct io_async_msghdr *iomsg)
4577 iomsg->msg.msg_name = &iomsg->addr;
4579 #ifdef CONFIG_COMPAT
4580 if (req->ctx->compat)
4581 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4584 return __io_recvmsg_copy_hdr(req, iomsg);
4587 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4590 struct io_sr_msg *sr = &req->sr_msg;
4591 struct io_buffer *kbuf;
4593 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4598 req->flags |= REQ_F_BUFFER_SELECTED;
4602 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4604 return io_put_kbuf(req, req->sr_msg.kbuf);
4607 static int io_recvmsg_prep_async(struct io_kiocb *req)
4611 ret = io_recvmsg_copy_hdr(req, req->async_data);
4613 req->flags |= REQ_F_NEED_CLEANUP;
4617 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4619 struct io_sr_msg *sr = &req->sr_msg;
4621 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4624 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4625 sr->len = READ_ONCE(sqe->len);
4626 sr->bgid = READ_ONCE(sqe->buf_group);
4627 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4628 if (sr->msg_flags & MSG_DONTWAIT)
4629 req->flags |= REQ_F_NOWAIT;
4631 #ifdef CONFIG_COMPAT
4632 if (req->ctx->compat)
4633 sr->msg_flags |= MSG_CMSG_COMPAT;
4638 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4640 struct io_async_msghdr iomsg, *kmsg;
4641 struct socket *sock;
4642 struct io_buffer *kbuf;
4645 int ret, cflags = 0;
4646 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4648 sock = sock_from_file(req->file);
4649 if (unlikely(!sock))
4652 kmsg = req->async_data;
4654 ret = io_recvmsg_copy_hdr(req, &iomsg);
4660 if (req->flags & REQ_F_BUFFER_SELECT) {
4661 kbuf = io_recv_buffer_select(req, !force_nonblock);
4663 return PTR_ERR(kbuf);
4664 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4665 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4666 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4667 1, req->sr_msg.len);
4670 flags = req->sr_msg.msg_flags;
4672 flags |= MSG_DONTWAIT;
4673 if (flags & MSG_WAITALL)
4674 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4676 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4677 kmsg->uaddr, flags);
4678 if (force_nonblock && ret == -EAGAIN)
4679 return io_setup_async_msg(req, kmsg);
4680 if (ret == -ERESTARTSYS)
4683 if (req->flags & REQ_F_BUFFER_SELECTED)
4684 cflags = io_put_recv_kbuf(req);
4685 /* fast path, check for non-NULL to avoid function call */
4687 kfree(kmsg->free_iov);
4688 req->flags &= ~REQ_F_NEED_CLEANUP;
4689 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4691 __io_req_complete(req, issue_flags, ret, cflags);
4695 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4697 struct io_buffer *kbuf;
4698 struct io_sr_msg *sr = &req->sr_msg;
4700 void __user *buf = sr->buf;
4701 struct socket *sock;
4705 int ret, cflags = 0;
4706 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4708 sock = sock_from_file(req->file);
4709 if (unlikely(!sock))
4712 if (req->flags & REQ_F_BUFFER_SELECT) {
4713 kbuf = io_recv_buffer_select(req, !force_nonblock);
4715 return PTR_ERR(kbuf);
4716 buf = u64_to_user_ptr(kbuf->addr);
4719 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4723 msg.msg_name = NULL;
4724 msg.msg_control = NULL;
4725 msg.msg_controllen = 0;
4726 msg.msg_namelen = 0;
4727 msg.msg_iocb = NULL;
4730 flags = req->sr_msg.msg_flags;
4732 flags |= MSG_DONTWAIT;
4733 if (flags & MSG_WAITALL)
4734 min_ret = iov_iter_count(&msg.msg_iter);
4736 ret = sock_recvmsg(sock, &msg, flags);
4737 if (force_nonblock && ret == -EAGAIN)
4739 if (ret == -ERESTARTSYS)
4742 if (req->flags & REQ_F_BUFFER_SELECTED)
4743 cflags = io_put_recv_kbuf(req);
4744 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4746 __io_req_complete(req, issue_flags, ret, cflags);
4750 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4752 struct io_accept *accept = &req->accept;
4754 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4756 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->splice_fd_in)
4759 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4760 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4761 accept->flags = READ_ONCE(sqe->accept_flags);
4762 accept->nofile = rlimit(RLIMIT_NOFILE);
4766 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4768 struct io_accept *accept = &req->accept;
4769 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4770 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4773 if (req->file->f_flags & O_NONBLOCK)
4774 req->flags |= REQ_F_NOWAIT;
4776 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4777 accept->addr_len, accept->flags,
4779 if (ret == -EAGAIN && force_nonblock)
4782 if (ret == -ERESTARTSYS)
4786 __io_req_complete(req, issue_flags, ret, 0);
4790 static int io_connect_prep_async(struct io_kiocb *req)
4792 struct io_async_connect *io = req->async_data;
4793 struct io_connect *conn = &req->connect;
4795 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4798 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4800 struct io_connect *conn = &req->connect;
4802 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4804 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4808 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4809 conn->addr_len = READ_ONCE(sqe->addr2);
4813 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4815 struct io_async_connect __io, *io;
4816 unsigned file_flags;
4818 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4820 if (req->async_data) {
4821 io = req->async_data;
4823 ret = move_addr_to_kernel(req->connect.addr,
4824 req->connect.addr_len,
4831 file_flags = force_nonblock ? O_NONBLOCK : 0;
4833 ret = __sys_connect_file(req->file, &io->address,
4834 req->connect.addr_len, file_flags);
4835 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4836 if (req->async_data)
4838 if (io_alloc_async_data(req)) {
4842 memcpy(req->async_data, &__io, sizeof(__io));
4845 if (ret == -ERESTARTSYS)
4850 __io_req_complete(req, issue_flags, ret, 0);
4853 #else /* !CONFIG_NET */
4854 #define IO_NETOP_FN(op) \
4855 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4857 return -EOPNOTSUPP; \
4860 #define IO_NETOP_PREP(op) \
4862 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4864 return -EOPNOTSUPP; \
4867 #define IO_NETOP_PREP_ASYNC(op) \
4869 static int io_##op##_prep_async(struct io_kiocb *req) \
4871 return -EOPNOTSUPP; \
4874 IO_NETOP_PREP_ASYNC(sendmsg);
4875 IO_NETOP_PREP_ASYNC(recvmsg);
4876 IO_NETOP_PREP_ASYNC(connect);
4877 IO_NETOP_PREP(accept);
4880 #endif /* CONFIG_NET */
4882 struct io_poll_table {
4883 struct poll_table_struct pt;
4884 struct io_kiocb *req;
4889 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4890 __poll_t mask, io_req_tw_func_t func)
4892 /* for instances that support it check for an event match first: */
4893 if (mask && !(mask & poll->events))
4896 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4898 list_del_init(&poll->wait.entry);
4901 req->io_task_work.func = func;
4904 * If this fails, then the task is exiting. When a task exits, the
4905 * work gets canceled, so just cancel this request as well instead
4906 * of executing it. We can't safely execute it anyway, as we may not
4907 * have the needed state needed for it anyway.
4909 io_req_task_work_add(req);
4913 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4914 __acquires(&req->ctx->completion_lock)
4916 struct io_ring_ctx *ctx = req->ctx;
4918 /* req->task == current here, checking PF_EXITING is safe */
4919 if (unlikely(req->task->flags & PF_EXITING))
4920 WRITE_ONCE(poll->canceled, true);
4922 if (!req->result && !READ_ONCE(poll->canceled)) {
4923 struct poll_table_struct pt = { ._key = poll->events };
4925 req->result = vfs_poll(req->file, &pt) & poll->events;
4928 spin_lock(&ctx->completion_lock);
4929 if (!req->result && !READ_ONCE(poll->canceled)) {
4930 add_wait_queue(poll->head, &poll->wait);
4937 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4939 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4940 if (req->opcode == IORING_OP_POLL_ADD)
4941 return req->async_data;
4942 return req->apoll->double_poll;
4945 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4947 if (req->opcode == IORING_OP_POLL_ADD)
4949 return &req->apoll->poll;
4952 static void io_poll_remove_double(struct io_kiocb *req)
4953 __must_hold(&req->ctx->completion_lock)
4955 struct io_poll_iocb *poll = io_poll_get_double(req);
4957 lockdep_assert_held(&req->ctx->completion_lock);
4959 if (poll && poll->head) {
4960 struct wait_queue_head *head = poll->head;
4962 spin_lock_irq(&head->lock);
4963 list_del_init(&poll->wait.entry);
4964 if (poll->wait.private)
4967 spin_unlock_irq(&head->lock);
4971 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4972 __must_hold(&req->ctx->completion_lock)
4974 struct io_ring_ctx *ctx = req->ctx;
4975 unsigned flags = IORING_CQE_F_MORE;
4978 if (READ_ONCE(req->poll.canceled)) {
4980 req->poll.events |= EPOLLONESHOT;
4982 error = mangle_poll(mask);
4984 if (req->poll.events & EPOLLONESHOT)
4986 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4987 req->poll.done = true;
4990 if (flags & IORING_CQE_F_MORE)
4993 io_commit_cqring(ctx);
4994 return !(flags & IORING_CQE_F_MORE);
4997 static void io_poll_task_func(struct io_kiocb *req)
4999 struct io_ring_ctx *ctx = req->ctx;
5000 struct io_kiocb *nxt;
5002 if (io_poll_rewait(req, &req->poll)) {
5003 spin_unlock(&ctx->completion_lock);
5007 done = io_poll_complete(req, req->result);
5009 io_poll_remove_double(req);
5010 hash_del(&req->hash_node);
5013 add_wait_queue(req->poll.head, &req->poll.wait);
5015 spin_unlock(&ctx->completion_lock);
5016 io_cqring_ev_posted(ctx);
5019 nxt = io_put_req_find_next(req);
5021 io_req_task_submit(nxt);
5026 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5027 int sync, void *key)
5029 struct io_kiocb *req = wait->private;
5030 struct io_poll_iocb *poll = io_poll_get_single(req);
5031 __poll_t mask = key_to_poll(key);
5032 unsigned long flags;
5034 /* for instances that support it check for an event match first: */
5035 if (mask && !(mask & poll->events))
5037 if (!(poll->events & EPOLLONESHOT))
5038 return poll->wait.func(&poll->wait, mode, sync, key);
5040 list_del_init(&wait->entry);
5045 spin_lock_irqsave(&poll->head->lock, flags);
5046 done = list_empty(&poll->wait.entry);
5048 list_del_init(&poll->wait.entry);
5049 /* make sure double remove sees this as being gone */
5050 wait->private = NULL;
5051 spin_unlock_irqrestore(&poll->head->lock, flags);
5053 /* use wait func handler, so it matches the rq type */
5054 poll->wait.func(&poll->wait, mode, sync, key);
5061 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5062 wait_queue_func_t wake_func)
5066 poll->canceled = false;
5067 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5068 /* mask in events that we always want/need */
5069 poll->events = events | IO_POLL_UNMASK;
5070 INIT_LIST_HEAD(&poll->wait.entry);
5071 init_waitqueue_func_entry(&poll->wait, wake_func);
5074 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5075 struct wait_queue_head *head,
5076 struct io_poll_iocb **poll_ptr)
5078 struct io_kiocb *req = pt->req;
5081 * The file being polled uses multiple waitqueues for poll handling
5082 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5085 if (unlikely(pt->nr_entries)) {
5086 struct io_poll_iocb *poll_one = poll;
5088 /* double add on the same waitqueue head, ignore */
5089 if (poll_one->head == head)
5091 /* already have a 2nd entry, fail a third attempt */
5093 if ((*poll_ptr)->head == head)
5095 pt->error = -EINVAL;
5099 * Can't handle multishot for double wait for now, turn it
5100 * into one-shot mode.
5102 if (!(poll_one->events & EPOLLONESHOT))
5103 poll_one->events |= EPOLLONESHOT;
5104 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5106 pt->error = -ENOMEM;
5109 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5111 poll->wait.private = req;
5118 if (poll->events & EPOLLEXCLUSIVE)
5119 add_wait_queue_exclusive(head, &poll->wait);
5121 add_wait_queue(head, &poll->wait);
5124 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5125 struct poll_table_struct *p)
5127 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5128 struct async_poll *apoll = pt->req->apoll;
5130 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5133 static void io_async_task_func(struct io_kiocb *req)
5135 struct async_poll *apoll = req->apoll;
5136 struct io_ring_ctx *ctx = req->ctx;
5138 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5140 if (io_poll_rewait(req, &apoll->poll)) {
5141 spin_unlock(&ctx->completion_lock);
5145 hash_del(&req->hash_node);
5146 io_poll_remove_double(req);
5147 spin_unlock(&ctx->completion_lock);
5149 if (!READ_ONCE(apoll->poll.canceled))
5150 io_req_task_submit(req);
5152 io_req_complete_failed(req, -ECANCELED);
5155 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5158 struct io_kiocb *req = wait->private;
5159 struct io_poll_iocb *poll = &req->apoll->poll;
5161 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5164 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5167 static void io_poll_req_insert(struct io_kiocb *req)
5169 struct io_ring_ctx *ctx = req->ctx;
5170 struct hlist_head *list;
5172 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5173 hlist_add_head(&req->hash_node, list);
5176 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5177 struct io_poll_iocb *poll,
5178 struct io_poll_table *ipt, __poll_t mask,
5179 wait_queue_func_t wake_func)
5180 __acquires(&ctx->completion_lock)
5182 struct io_ring_ctx *ctx = req->ctx;
5183 bool cancel = false;
5185 INIT_HLIST_NODE(&req->hash_node);
5186 io_init_poll_iocb(poll, mask, wake_func);
5187 poll->file = req->file;
5188 poll->wait.private = req;
5190 ipt->pt._key = mask;
5193 ipt->nr_entries = 0;
5195 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5196 if (unlikely(!ipt->nr_entries) && !ipt->error)
5197 ipt->error = -EINVAL;
5199 spin_lock(&ctx->completion_lock);
5200 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5201 io_poll_remove_double(req);
5202 if (likely(poll->head)) {
5203 spin_lock_irq(&poll->head->lock);
5204 if (unlikely(list_empty(&poll->wait.entry))) {
5210 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5211 list_del_init(&poll->wait.entry);
5213 WRITE_ONCE(poll->canceled, true);
5214 else if (!poll->done) /* actually waiting for an event */
5215 io_poll_req_insert(req);
5216 spin_unlock_irq(&poll->head->lock);
5228 static int io_arm_poll_handler(struct io_kiocb *req)
5230 const struct io_op_def *def = &io_op_defs[req->opcode];
5231 struct io_ring_ctx *ctx = req->ctx;
5232 struct async_poll *apoll;
5233 struct io_poll_table ipt;
5234 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5237 if (!req->file || !file_can_poll(req->file))
5238 return IO_APOLL_ABORTED;
5239 if (req->flags & REQ_F_POLLED)
5240 return IO_APOLL_ABORTED;
5241 if (!def->pollin && !def->pollout)
5242 return IO_APOLL_ABORTED;
5246 mask |= POLLIN | POLLRDNORM;
5248 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5249 if ((req->opcode == IORING_OP_RECVMSG) &&
5250 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5254 mask |= POLLOUT | POLLWRNORM;
5257 /* if we can't nonblock try, then no point in arming a poll handler */
5258 if (!io_file_supports_nowait(req, rw))
5259 return IO_APOLL_ABORTED;
5261 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5262 if (unlikely(!apoll))
5263 return IO_APOLL_ABORTED;
5264 apoll->double_poll = NULL;
5266 req->flags |= REQ_F_POLLED;
5267 ipt.pt._qproc = io_async_queue_proc;
5268 io_req_set_refcount(req);
5270 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5272 spin_unlock(&ctx->completion_lock);
5273 if (ret || ipt.error)
5274 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5276 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5277 mask, apoll->poll.events);
5281 static bool __io_poll_remove_one(struct io_kiocb *req,
5282 struct io_poll_iocb *poll, bool do_cancel)
5283 __must_hold(&req->ctx->completion_lock)
5285 bool do_complete = false;
5289 spin_lock_irq(&poll->head->lock);
5291 WRITE_ONCE(poll->canceled, true);
5292 if (!list_empty(&poll->wait.entry)) {
5293 list_del_init(&poll->wait.entry);
5296 spin_unlock_irq(&poll->head->lock);
5297 hash_del(&req->hash_node);
5301 static bool io_poll_remove_one(struct io_kiocb *req)
5302 __must_hold(&req->ctx->completion_lock)
5306 io_poll_remove_double(req);
5307 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5310 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5311 io_commit_cqring(req->ctx);
5313 io_put_req_deferred(req);
5319 * Returns true if we found and killed one or more poll requests
5321 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5324 struct hlist_node *tmp;
5325 struct io_kiocb *req;
5328 spin_lock(&ctx->completion_lock);
5329 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5330 struct hlist_head *list;
5332 list = &ctx->cancel_hash[i];
5333 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5334 if (io_match_task(req, tsk, cancel_all))
5335 posted += io_poll_remove_one(req);
5338 spin_unlock(&ctx->completion_lock);
5341 io_cqring_ev_posted(ctx);
5346 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5348 __must_hold(&ctx->completion_lock)
5350 struct hlist_head *list;
5351 struct io_kiocb *req;
5353 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5354 hlist_for_each_entry(req, list, hash_node) {
5355 if (sqe_addr != req->user_data)
5357 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5364 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5366 __must_hold(&ctx->completion_lock)
5368 struct io_kiocb *req;
5370 req = io_poll_find(ctx, sqe_addr, poll_only);
5373 if (io_poll_remove_one(req))
5379 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5384 events = READ_ONCE(sqe->poll32_events);
5386 events = swahw32(events);
5388 if (!(flags & IORING_POLL_ADD_MULTI))
5389 events |= EPOLLONESHOT;
5390 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5393 static int io_poll_update_prep(struct io_kiocb *req,
5394 const struct io_uring_sqe *sqe)
5396 struct io_poll_update *upd = &req->poll_update;
5399 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5401 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5403 flags = READ_ONCE(sqe->len);
5404 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5405 IORING_POLL_ADD_MULTI))
5407 /* meaningless without update */
5408 if (flags == IORING_POLL_ADD_MULTI)
5411 upd->old_user_data = READ_ONCE(sqe->addr);
5412 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5413 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5415 upd->new_user_data = READ_ONCE(sqe->off);
5416 if (!upd->update_user_data && upd->new_user_data)
5418 if (upd->update_events)
5419 upd->events = io_poll_parse_events(sqe, flags);
5420 else if (sqe->poll32_events)
5426 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5429 struct io_kiocb *req = wait->private;
5430 struct io_poll_iocb *poll = &req->poll;
5432 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5435 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5436 struct poll_table_struct *p)
5438 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5440 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5443 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5445 struct io_poll_iocb *poll = &req->poll;
5448 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5450 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5452 flags = READ_ONCE(sqe->len);
5453 if (flags & ~IORING_POLL_ADD_MULTI)
5456 io_req_set_refcount(req);
5457 poll->events = io_poll_parse_events(sqe, flags);
5461 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5463 struct io_poll_iocb *poll = &req->poll;
5464 struct io_ring_ctx *ctx = req->ctx;
5465 struct io_poll_table ipt;
5468 ipt.pt._qproc = io_poll_queue_proc;
5470 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5473 if (mask) { /* no async, we'd stolen it */
5475 io_poll_complete(req, mask);
5477 spin_unlock(&ctx->completion_lock);
5480 io_cqring_ev_posted(ctx);
5481 if (poll->events & EPOLLONESHOT)
5487 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5489 struct io_ring_ctx *ctx = req->ctx;
5490 struct io_kiocb *preq;
5494 spin_lock(&ctx->completion_lock);
5495 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5501 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5503 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5508 * Don't allow racy completion with singleshot, as we cannot safely
5509 * update those. For multishot, if we're racing with completion, just
5510 * let completion re-add it.
5512 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5513 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5517 /* we now have a detached poll request. reissue. */
5521 spin_unlock(&ctx->completion_lock);
5523 io_req_complete(req, ret);
5526 /* only mask one event flags, keep behavior flags */
5527 if (req->poll_update.update_events) {
5528 preq->poll.events &= ~0xffff;
5529 preq->poll.events |= req->poll_update.events & 0xffff;
5530 preq->poll.events |= IO_POLL_UNMASK;
5532 if (req->poll_update.update_user_data)
5533 preq->user_data = req->poll_update.new_user_data;
5534 spin_unlock(&ctx->completion_lock);
5536 /* complete update request, we're done with it */
5537 io_req_complete(req, ret);
5540 ret = io_poll_add(preq, issue_flags);
5543 io_req_complete(preq, ret);
5549 static void io_req_task_timeout(struct io_kiocb *req)
5552 io_req_complete_post(req, -ETIME, 0);
5555 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5557 struct io_timeout_data *data = container_of(timer,
5558 struct io_timeout_data, timer);
5559 struct io_kiocb *req = data->req;
5560 struct io_ring_ctx *ctx = req->ctx;
5561 unsigned long flags;
5563 spin_lock_irqsave(&ctx->timeout_lock, flags);
5564 list_del_init(&req->timeout.list);
5565 atomic_set(&req->ctx->cq_timeouts,
5566 atomic_read(&req->ctx->cq_timeouts) + 1);
5567 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5569 req->io_task_work.func = io_req_task_timeout;
5570 io_req_task_work_add(req);
5571 return HRTIMER_NORESTART;
5574 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5576 __must_hold(&ctx->timeout_lock)
5578 struct io_timeout_data *io;
5579 struct io_kiocb *req;
5582 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5583 found = user_data == req->user_data;
5588 return ERR_PTR(-ENOENT);
5590 io = req->async_data;
5591 if (hrtimer_try_to_cancel(&io->timer) == -1)
5592 return ERR_PTR(-EALREADY);
5593 list_del_init(&req->timeout.list);
5597 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5598 __must_hold(&ctx->completion_lock)
5599 __must_hold(&ctx->timeout_lock)
5601 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5604 return PTR_ERR(req);
5607 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5608 io_put_req_deferred(req);
5612 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5613 struct timespec64 *ts, enum hrtimer_mode mode)
5614 __must_hold(&ctx->timeout_lock)
5616 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5617 struct io_timeout_data *data;
5620 return PTR_ERR(req);
5622 req->timeout.off = 0; /* noseq */
5623 data = req->async_data;
5624 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5625 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5626 data->timer.function = io_timeout_fn;
5627 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5631 static int io_timeout_remove_prep(struct io_kiocb *req,
5632 const struct io_uring_sqe *sqe)
5634 struct io_timeout_rem *tr = &req->timeout_rem;
5636 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5638 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5640 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5643 tr->addr = READ_ONCE(sqe->addr);
5644 tr->flags = READ_ONCE(sqe->timeout_flags);
5645 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5646 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5648 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5650 } else if (tr->flags) {
5651 /* timeout removal doesn't support flags */
5658 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5660 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5665 * Remove or update an existing timeout command
5667 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5669 struct io_timeout_rem *tr = &req->timeout_rem;
5670 struct io_ring_ctx *ctx = req->ctx;
5673 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5674 spin_lock(&ctx->completion_lock);
5675 spin_lock_irq(&ctx->timeout_lock);
5676 ret = io_timeout_cancel(ctx, tr->addr);
5677 spin_unlock_irq(&ctx->timeout_lock);
5678 spin_unlock(&ctx->completion_lock);
5680 spin_lock_irq(&ctx->timeout_lock);
5681 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5682 io_translate_timeout_mode(tr->flags));
5683 spin_unlock_irq(&ctx->timeout_lock);
5688 io_req_complete_post(req, ret, 0);
5692 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5693 bool is_timeout_link)
5695 struct io_timeout_data *data;
5697 u32 off = READ_ONCE(sqe->off);
5699 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5701 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5704 if (off && is_timeout_link)
5706 flags = READ_ONCE(sqe->timeout_flags);
5707 if (flags & ~IORING_TIMEOUT_ABS)
5710 req->timeout.off = off;
5711 if (unlikely(off && !req->ctx->off_timeout_used))
5712 req->ctx->off_timeout_used = true;
5714 if (!req->async_data && io_alloc_async_data(req))
5717 data = req->async_data;
5720 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5723 data->mode = io_translate_timeout_mode(flags);
5724 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5726 if (is_timeout_link) {
5727 struct io_submit_link *link = &req->ctx->submit_state.link;
5731 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5733 req->timeout.head = link->last;
5734 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5739 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5741 struct io_ring_ctx *ctx = req->ctx;
5742 struct io_timeout_data *data = req->async_data;
5743 struct list_head *entry;
5744 u32 tail, off = req->timeout.off;
5746 spin_lock_irq(&ctx->timeout_lock);
5749 * sqe->off holds how many events that need to occur for this
5750 * timeout event to be satisfied. If it isn't set, then this is
5751 * a pure timeout request, sequence isn't used.
5753 if (io_is_timeout_noseq(req)) {
5754 entry = ctx->timeout_list.prev;
5758 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5759 req->timeout.target_seq = tail + off;
5761 /* Update the last seq here in case io_flush_timeouts() hasn't.
5762 * This is safe because ->completion_lock is held, and submissions
5763 * and completions are never mixed in the same ->completion_lock section.
5765 ctx->cq_last_tm_flush = tail;
5768 * Insertion sort, ensuring the first entry in the list is always
5769 * the one we need first.
5771 list_for_each_prev(entry, &ctx->timeout_list) {
5772 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5775 if (io_is_timeout_noseq(nxt))
5777 /* nxt.seq is behind @tail, otherwise would've been completed */
5778 if (off >= nxt->timeout.target_seq - tail)
5782 list_add(&req->timeout.list, entry);
5783 data->timer.function = io_timeout_fn;
5784 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5785 spin_unlock_irq(&ctx->timeout_lock);
5789 struct io_cancel_data {
5790 struct io_ring_ctx *ctx;
5794 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5796 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5797 struct io_cancel_data *cd = data;
5799 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5802 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5803 struct io_ring_ctx *ctx)
5805 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5806 enum io_wq_cancel cancel_ret;
5809 if (!tctx || !tctx->io_wq)
5812 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5813 switch (cancel_ret) {
5814 case IO_WQ_CANCEL_OK:
5817 case IO_WQ_CANCEL_RUNNING:
5820 case IO_WQ_CANCEL_NOTFOUND:
5828 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5830 struct io_ring_ctx *ctx = req->ctx;
5833 WARN_ON_ONCE(req->task != current);
5835 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5839 spin_lock(&ctx->completion_lock);
5840 spin_lock_irq(&ctx->timeout_lock);
5841 ret = io_timeout_cancel(ctx, sqe_addr);
5842 spin_unlock_irq(&ctx->timeout_lock);
5845 ret = io_poll_cancel(ctx, sqe_addr, false);
5847 spin_unlock(&ctx->completion_lock);
5851 static int io_async_cancel_prep(struct io_kiocb *req,
5852 const struct io_uring_sqe *sqe)
5854 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5856 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5858 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5862 req->cancel.addr = READ_ONCE(sqe->addr);
5866 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5868 struct io_ring_ctx *ctx = req->ctx;
5869 u64 sqe_addr = req->cancel.addr;
5870 struct io_tctx_node *node;
5873 ret = io_try_cancel_userdata(req, sqe_addr);
5877 /* slow path, try all io-wq's */
5878 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5880 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5881 struct io_uring_task *tctx = node->task->io_uring;
5883 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5887 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5891 io_req_complete_post(req, ret, 0);
5895 static int io_rsrc_update_prep(struct io_kiocb *req,
5896 const struct io_uring_sqe *sqe)
5898 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5900 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
5903 req->rsrc_update.offset = READ_ONCE(sqe->off);
5904 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5905 if (!req->rsrc_update.nr_args)
5907 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5911 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5913 struct io_ring_ctx *ctx = req->ctx;
5914 struct io_uring_rsrc_update2 up;
5917 if (issue_flags & IO_URING_F_NONBLOCK)
5920 up.offset = req->rsrc_update.offset;
5921 up.data = req->rsrc_update.arg;
5926 mutex_lock(&ctx->uring_lock);
5927 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5928 &up, req->rsrc_update.nr_args);
5929 mutex_unlock(&ctx->uring_lock);
5933 __io_req_complete(req, issue_flags, ret, 0);
5937 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5939 switch (req->opcode) {
5942 case IORING_OP_READV:
5943 case IORING_OP_READ_FIXED:
5944 case IORING_OP_READ:
5945 return io_read_prep(req, sqe);
5946 case IORING_OP_WRITEV:
5947 case IORING_OP_WRITE_FIXED:
5948 case IORING_OP_WRITE:
5949 return io_write_prep(req, sqe);
5950 case IORING_OP_POLL_ADD:
5951 return io_poll_add_prep(req, sqe);
5952 case IORING_OP_POLL_REMOVE:
5953 return io_poll_update_prep(req, sqe);
5954 case IORING_OP_FSYNC:
5955 return io_fsync_prep(req, sqe);
5956 case IORING_OP_SYNC_FILE_RANGE:
5957 return io_sfr_prep(req, sqe);
5958 case IORING_OP_SENDMSG:
5959 case IORING_OP_SEND:
5960 return io_sendmsg_prep(req, sqe);
5961 case IORING_OP_RECVMSG:
5962 case IORING_OP_RECV:
5963 return io_recvmsg_prep(req, sqe);
5964 case IORING_OP_CONNECT:
5965 return io_connect_prep(req, sqe);
5966 case IORING_OP_TIMEOUT:
5967 return io_timeout_prep(req, sqe, false);
5968 case IORING_OP_TIMEOUT_REMOVE:
5969 return io_timeout_remove_prep(req, sqe);
5970 case IORING_OP_ASYNC_CANCEL:
5971 return io_async_cancel_prep(req, sqe);
5972 case IORING_OP_LINK_TIMEOUT:
5973 return io_timeout_prep(req, sqe, true);
5974 case IORING_OP_ACCEPT:
5975 return io_accept_prep(req, sqe);
5976 case IORING_OP_FALLOCATE:
5977 return io_fallocate_prep(req, sqe);
5978 case IORING_OP_OPENAT:
5979 return io_openat_prep(req, sqe);
5980 case IORING_OP_CLOSE:
5981 return io_close_prep(req, sqe);
5982 case IORING_OP_FILES_UPDATE:
5983 return io_rsrc_update_prep(req, sqe);
5984 case IORING_OP_STATX:
5985 return io_statx_prep(req, sqe);
5986 case IORING_OP_FADVISE:
5987 return io_fadvise_prep(req, sqe);
5988 case IORING_OP_MADVISE:
5989 return io_madvise_prep(req, sqe);
5990 case IORING_OP_OPENAT2:
5991 return io_openat2_prep(req, sqe);
5992 case IORING_OP_EPOLL_CTL:
5993 return io_epoll_ctl_prep(req, sqe);
5994 case IORING_OP_SPLICE:
5995 return io_splice_prep(req, sqe);
5996 case IORING_OP_PROVIDE_BUFFERS:
5997 return io_provide_buffers_prep(req, sqe);
5998 case IORING_OP_REMOVE_BUFFERS:
5999 return io_remove_buffers_prep(req, sqe);
6001 return io_tee_prep(req, sqe);
6002 case IORING_OP_SHUTDOWN:
6003 return io_shutdown_prep(req, sqe);
6004 case IORING_OP_RENAMEAT:
6005 return io_renameat_prep(req, sqe);
6006 case IORING_OP_UNLINKAT:
6007 return io_unlinkat_prep(req, sqe);
6010 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6015 static int io_req_prep_async(struct io_kiocb *req)
6017 if (!io_op_defs[req->opcode].needs_async_setup)
6019 if (WARN_ON_ONCE(req->async_data))
6021 if (io_alloc_async_data(req))
6024 switch (req->opcode) {
6025 case IORING_OP_READV:
6026 return io_rw_prep_async(req, READ);
6027 case IORING_OP_WRITEV:
6028 return io_rw_prep_async(req, WRITE);
6029 case IORING_OP_SENDMSG:
6030 return io_sendmsg_prep_async(req);
6031 case IORING_OP_RECVMSG:
6032 return io_recvmsg_prep_async(req);
6033 case IORING_OP_CONNECT:
6034 return io_connect_prep_async(req);
6036 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6041 static u32 io_get_sequence(struct io_kiocb *req)
6043 u32 seq = req->ctx->cached_sq_head;
6045 /* need original cached_sq_head, but it was increased for each req */
6046 io_for_each_link(req, req)
6051 static bool io_drain_req(struct io_kiocb *req)
6053 struct io_kiocb *pos;
6054 struct io_ring_ctx *ctx = req->ctx;
6055 struct io_defer_entry *de;
6060 * If we need to drain a request in the middle of a link, drain the
6061 * head request and the next request/link after the current link.
6062 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6063 * maintained for every request of our link.
6065 if (ctx->drain_next) {
6066 req->flags |= REQ_F_IO_DRAIN;
6067 ctx->drain_next = false;
6069 /* not interested in head, start from the first linked */
6070 io_for_each_link(pos, req->link) {
6071 if (pos->flags & REQ_F_IO_DRAIN) {
6072 ctx->drain_next = true;
6073 req->flags |= REQ_F_IO_DRAIN;
6078 /* Still need defer if there is pending req in defer list. */
6079 if (likely(list_empty_careful(&ctx->defer_list) &&
6080 !(req->flags & REQ_F_IO_DRAIN))) {
6081 ctx->drain_active = false;
6085 seq = io_get_sequence(req);
6086 /* Still a chance to pass the sequence check */
6087 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6090 ret = io_req_prep_async(req);
6093 io_prep_async_link(req);
6094 de = kmalloc(sizeof(*de), GFP_KERNEL);
6098 io_req_complete_failed(req, ret);
6102 spin_lock(&ctx->completion_lock);
6103 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6104 spin_unlock(&ctx->completion_lock);
6106 io_queue_async_work(req);
6110 trace_io_uring_defer(ctx, req, req->user_data);
6113 list_add_tail(&de->list, &ctx->defer_list);
6114 spin_unlock(&ctx->completion_lock);
6118 static void io_clean_op(struct io_kiocb *req)
6120 if (req->flags & REQ_F_BUFFER_SELECTED) {
6121 switch (req->opcode) {
6122 case IORING_OP_READV:
6123 case IORING_OP_READ_FIXED:
6124 case IORING_OP_READ:
6125 kfree((void *)(unsigned long)req->rw.addr);
6127 case IORING_OP_RECVMSG:
6128 case IORING_OP_RECV:
6129 kfree(req->sr_msg.kbuf);
6134 if (req->flags & REQ_F_NEED_CLEANUP) {
6135 switch (req->opcode) {
6136 case IORING_OP_READV:
6137 case IORING_OP_READ_FIXED:
6138 case IORING_OP_READ:
6139 case IORING_OP_WRITEV:
6140 case IORING_OP_WRITE_FIXED:
6141 case IORING_OP_WRITE: {
6142 struct io_async_rw *io = req->async_data;
6144 kfree(io->free_iovec);
6147 case IORING_OP_RECVMSG:
6148 case IORING_OP_SENDMSG: {
6149 struct io_async_msghdr *io = req->async_data;
6151 kfree(io->free_iov);
6154 case IORING_OP_SPLICE:
6156 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6157 io_put_file(req->splice.file_in);
6159 case IORING_OP_OPENAT:
6160 case IORING_OP_OPENAT2:
6161 if (req->open.filename)
6162 putname(req->open.filename);
6164 case IORING_OP_RENAMEAT:
6165 putname(req->rename.oldpath);
6166 putname(req->rename.newpath);
6168 case IORING_OP_UNLINKAT:
6169 putname(req->unlink.filename);
6173 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6174 kfree(req->apoll->double_poll);
6178 if (req->flags & REQ_F_INFLIGHT) {
6179 struct io_uring_task *tctx = req->task->io_uring;
6181 atomic_dec(&tctx->inflight_tracked);
6183 if (req->flags & REQ_F_CREDS)
6184 put_cred(req->creds);
6186 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6189 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6191 struct io_ring_ctx *ctx = req->ctx;
6192 const struct cred *creds = NULL;
6195 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6196 creds = override_creds(req->creds);
6198 switch (req->opcode) {
6200 ret = io_nop(req, issue_flags);
6202 case IORING_OP_READV:
6203 case IORING_OP_READ_FIXED:
6204 case IORING_OP_READ:
6205 ret = io_read(req, issue_flags);
6207 case IORING_OP_WRITEV:
6208 case IORING_OP_WRITE_FIXED:
6209 case IORING_OP_WRITE:
6210 ret = io_write(req, issue_flags);
6212 case IORING_OP_FSYNC:
6213 ret = io_fsync(req, issue_flags);
6215 case IORING_OP_POLL_ADD:
6216 ret = io_poll_add(req, issue_flags);
6218 case IORING_OP_POLL_REMOVE:
6219 ret = io_poll_update(req, issue_flags);
6221 case IORING_OP_SYNC_FILE_RANGE:
6222 ret = io_sync_file_range(req, issue_flags);
6224 case IORING_OP_SENDMSG:
6225 ret = io_sendmsg(req, issue_flags);
6227 case IORING_OP_SEND:
6228 ret = io_send(req, issue_flags);
6230 case IORING_OP_RECVMSG:
6231 ret = io_recvmsg(req, issue_flags);
6233 case IORING_OP_RECV:
6234 ret = io_recv(req, issue_flags);
6236 case IORING_OP_TIMEOUT:
6237 ret = io_timeout(req, issue_flags);
6239 case IORING_OP_TIMEOUT_REMOVE:
6240 ret = io_timeout_remove(req, issue_flags);
6242 case IORING_OP_ACCEPT:
6243 ret = io_accept(req, issue_flags);
6245 case IORING_OP_CONNECT:
6246 ret = io_connect(req, issue_flags);
6248 case IORING_OP_ASYNC_CANCEL:
6249 ret = io_async_cancel(req, issue_flags);
6251 case IORING_OP_FALLOCATE:
6252 ret = io_fallocate(req, issue_flags);
6254 case IORING_OP_OPENAT:
6255 ret = io_openat(req, issue_flags);
6257 case IORING_OP_CLOSE:
6258 ret = io_close(req, issue_flags);
6260 case IORING_OP_FILES_UPDATE:
6261 ret = io_files_update(req, issue_flags);
6263 case IORING_OP_STATX:
6264 ret = io_statx(req, issue_flags);
6266 case IORING_OP_FADVISE:
6267 ret = io_fadvise(req, issue_flags);
6269 case IORING_OP_MADVISE:
6270 ret = io_madvise(req, issue_flags);
6272 case IORING_OP_OPENAT2:
6273 ret = io_openat2(req, issue_flags);
6275 case IORING_OP_EPOLL_CTL:
6276 ret = io_epoll_ctl(req, issue_flags);
6278 case IORING_OP_SPLICE:
6279 ret = io_splice(req, issue_flags);
6281 case IORING_OP_PROVIDE_BUFFERS:
6282 ret = io_provide_buffers(req, issue_flags);
6284 case IORING_OP_REMOVE_BUFFERS:
6285 ret = io_remove_buffers(req, issue_flags);
6288 ret = io_tee(req, issue_flags);
6290 case IORING_OP_SHUTDOWN:
6291 ret = io_shutdown(req, issue_flags);
6293 case IORING_OP_RENAMEAT:
6294 ret = io_renameat(req, issue_flags);
6296 case IORING_OP_UNLINKAT:
6297 ret = io_unlinkat(req, issue_flags);
6305 revert_creds(creds);
6308 /* If the op doesn't have a file, we're not polling for it */
6309 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6310 io_iopoll_req_issued(req);
6315 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6317 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6319 req = io_put_req_find_next(req);
6320 return req ? &req->work : NULL;
6323 static void io_wq_submit_work(struct io_wq_work *work)
6325 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6326 struct io_kiocb *timeout;
6329 /* one will be dropped by ->io_free_work() after returning to io-wq */
6330 if (!(req->flags & REQ_F_REFCOUNT))
6331 __io_req_set_refcount(req, 2);
6335 timeout = io_prep_linked_timeout(req);
6337 io_queue_linked_timeout(timeout);
6339 if (work->flags & IO_WQ_WORK_CANCEL)
6344 ret = io_issue_sqe(req, 0);
6346 * We can get EAGAIN for polled IO even though we're
6347 * forcing a sync submission from here, since we can't
6348 * wait for request slots on the block side.
6356 /* avoid locking problems by failing it from a clean context */
6358 io_req_task_queue_fail(req, ret);
6361 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6364 return &table->files[i];
6367 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6370 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6372 return (struct file *) (slot->file_ptr & FFS_MASK);
6375 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6377 unsigned long file_ptr = (unsigned long) file;
6379 if (__io_file_supports_nowait(file, READ))
6380 file_ptr |= FFS_ASYNC_READ;
6381 if (__io_file_supports_nowait(file, WRITE))
6382 file_ptr |= FFS_ASYNC_WRITE;
6383 if (S_ISREG(file_inode(file)->i_mode))
6384 file_ptr |= FFS_ISREG;
6385 file_slot->file_ptr = file_ptr;
6388 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6389 struct io_kiocb *req, int fd)
6392 unsigned long file_ptr;
6394 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6396 fd = array_index_nospec(fd, ctx->nr_user_files);
6397 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6398 file = (struct file *) (file_ptr & FFS_MASK);
6399 file_ptr &= ~FFS_MASK;
6400 /* mask in overlapping REQ_F and FFS bits */
6401 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6402 io_req_set_rsrc_node(req);
6406 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6407 struct io_kiocb *req, int fd)
6409 struct file *file = fget(fd);
6411 trace_io_uring_file_get(ctx, fd);
6413 /* we don't allow fixed io_uring files */
6414 if (file && unlikely(file->f_op == &io_uring_fops))
6415 io_req_track_inflight(req);
6419 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6420 struct io_kiocb *req, int fd, bool fixed)
6423 return io_file_get_fixed(ctx, req, fd);
6425 return io_file_get_normal(ctx, req, fd);
6428 static void io_req_task_link_timeout(struct io_kiocb *req)
6430 struct io_kiocb *prev = req->timeout.prev;
6434 ret = io_try_cancel_userdata(req, prev->user_data);
6435 io_req_complete_post(req, ret ?: -ETIME, 0);
6438 io_req_complete_post(req, -ETIME, 0);
6442 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6444 struct io_timeout_data *data = container_of(timer,
6445 struct io_timeout_data, timer);
6446 struct io_kiocb *prev, *req = data->req;
6447 struct io_ring_ctx *ctx = req->ctx;
6448 unsigned long flags;
6450 spin_lock_irqsave(&ctx->timeout_lock, flags);
6451 prev = req->timeout.head;
6452 req->timeout.head = NULL;
6455 * We don't expect the list to be empty, that will only happen if we
6456 * race with the completion of the linked work.
6459 io_remove_next_linked(prev);
6460 if (!req_ref_inc_not_zero(prev))
6463 req->timeout.prev = prev;
6464 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6466 req->io_task_work.func = io_req_task_link_timeout;
6467 io_req_task_work_add(req);
6468 return HRTIMER_NORESTART;
6471 static void io_queue_linked_timeout(struct io_kiocb *req)
6473 struct io_ring_ctx *ctx = req->ctx;
6475 spin_lock_irq(&ctx->timeout_lock);
6477 * If the back reference is NULL, then our linked request finished
6478 * before we got a chance to setup the timer
6480 if (req->timeout.head) {
6481 struct io_timeout_data *data = req->async_data;
6483 data->timer.function = io_link_timeout_fn;
6484 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6487 spin_unlock_irq(&ctx->timeout_lock);
6488 /* drop submission reference */
6492 static void __io_queue_sqe(struct io_kiocb *req)
6493 __must_hold(&req->ctx->uring_lock)
6495 struct io_kiocb *linked_timeout;
6499 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6502 * We async punt it if the file wasn't marked NOWAIT, or if the file
6503 * doesn't support non-blocking read/write attempts
6506 if (req->flags & REQ_F_COMPLETE_INLINE) {
6507 struct io_ring_ctx *ctx = req->ctx;
6508 struct io_submit_state *state = &ctx->submit_state;
6510 state->compl_reqs[state->compl_nr++] = req;
6511 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6512 io_submit_flush_completions(ctx);
6516 linked_timeout = io_prep_linked_timeout(req);
6518 io_queue_linked_timeout(linked_timeout);
6519 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6520 linked_timeout = io_prep_linked_timeout(req);
6522 switch (io_arm_poll_handler(req)) {
6523 case IO_APOLL_READY:
6525 io_unprep_linked_timeout(req);
6527 case IO_APOLL_ABORTED:
6529 * Queued up for async execution, worker will release
6530 * submit reference when the iocb is actually submitted.
6532 io_queue_async_work(req);
6537 io_queue_linked_timeout(linked_timeout);
6539 io_req_complete_failed(req, ret);
6543 static inline void io_queue_sqe(struct io_kiocb *req)
6544 __must_hold(&req->ctx->uring_lock)
6546 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6549 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6550 __io_queue_sqe(req);
6552 int ret = io_req_prep_async(req);
6555 io_req_complete_failed(req, ret);
6557 io_queue_async_work(req);
6562 * Check SQE restrictions (opcode and flags).
6564 * Returns 'true' if SQE is allowed, 'false' otherwise.
6566 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6567 struct io_kiocb *req,
6568 unsigned int sqe_flags)
6570 if (likely(!ctx->restricted))
6573 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6576 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6577 ctx->restrictions.sqe_flags_required)
6580 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6581 ctx->restrictions.sqe_flags_required))
6587 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6588 const struct io_uring_sqe *sqe)
6589 __must_hold(&ctx->uring_lock)
6591 struct io_submit_state *state;
6592 unsigned int sqe_flags;
6593 int personality, ret = 0;
6595 /* req is partially pre-initialised, see io_preinit_req() */
6596 req->opcode = READ_ONCE(sqe->opcode);
6597 /* same numerical values with corresponding REQ_F_*, safe to copy */
6598 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6599 req->user_data = READ_ONCE(sqe->user_data);
6601 req->fixed_rsrc_refs = NULL;
6602 req->task = current;
6604 /* enforce forwards compatibility on users */
6605 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6607 if (unlikely(req->opcode >= IORING_OP_LAST))
6609 if (!io_check_restriction(ctx, req, sqe_flags))
6612 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6613 !io_op_defs[req->opcode].buffer_select)
6615 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6616 ctx->drain_active = true;
6618 personality = READ_ONCE(sqe->personality);
6620 req->creds = xa_load(&ctx->personalities, personality);
6623 get_cred(req->creds);
6624 req->flags |= REQ_F_CREDS;
6626 state = &ctx->submit_state;
6629 * Plug now if we have more than 1 IO left after this, and the target
6630 * is potentially a read/write to block based storage.
6632 if (!state->plug_started && state->ios_left > 1 &&
6633 io_op_defs[req->opcode].plug) {
6634 blk_start_plug(&state->plug);
6635 state->plug_started = true;
6638 if (io_op_defs[req->opcode].needs_file) {
6639 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6640 (sqe_flags & IOSQE_FIXED_FILE));
6641 if (unlikely(!req->file))
6649 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6650 const struct io_uring_sqe *sqe)
6651 __must_hold(&ctx->uring_lock)
6653 struct io_submit_link *link = &ctx->submit_state.link;
6656 ret = io_init_req(ctx, req, sqe);
6657 if (unlikely(ret)) {
6660 /* fail even hard links since we don't submit */
6661 req_set_fail(link->head);
6662 io_req_complete_failed(link->head, -ECANCELED);
6665 io_req_complete_failed(req, ret);
6669 ret = io_req_prep(req, sqe);
6673 /* don't need @sqe from now on */
6674 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6676 ctx->flags & IORING_SETUP_SQPOLL);
6679 * If we already have a head request, queue this one for async
6680 * submittal once the head completes. If we don't have a head but
6681 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6682 * submitted sync once the chain is complete. If none of those
6683 * conditions are true (normal request), then just queue it.
6686 struct io_kiocb *head = link->head;
6688 ret = io_req_prep_async(req);
6691 trace_io_uring_link(ctx, req, head);
6692 link->last->link = req;
6695 /* last request of a link, enqueue the link */
6696 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6701 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6713 * Batched submission is done, ensure local IO is flushed out.
6715 static void io_submit_state_end(struct io_submit_state *state,
6716 struct io_ring_ctx *ctx)
6718 if (state->link.head)
6719 io_queue_sqe(state->link.head);
6720 if (state->compl_nr)
6721 io_submit_flush_completions(ctx);
6722 if (state->plug_started)
6723 blk_finish_plug(&state->plug);
6727 * Start submission side cache.
6729 static void io_submit_state_start(struct io_submit_state *state,
6730 unsigned int max_ios)
6732 state->plug_started = false;
6733 state->ios_left = max_ios;
6734 /* set only head, no need to init link_last in advance */
6735 state->link.head = NULL;
6738 static void io_commit_sqring(struct io_ring_ctx *ctx)
6740 struct io_rings *rings = ctx->rings;
6743 * Ensure any loads from the SQEs are done at this point,
6744 * since once we write the new head, the application could
6745 * write new data to them.
6747 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6751 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6752 * that is mapped by userspace. This means that care needs to be taken to
6753 * ensure that reads are stable, as we cannot rely on userspace always
6754 * being a good citizen. If members of the sqe are validated and then later
6755 * used, it's important that those reads are done through READ_ONCE() to
6756 * prevent a re-load down the line.
6758 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6760 unsigned head, mask = ctx->sq_entries - 1;
6761 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6764 * The cached sq head (or cq tail) serves two purposes:
6766 * 1) allows us to batch the cost of updating the user visible
6768 * 2) allows the kernel side to track the head on its own, even
6769 * though the application is the one updating it.
6771 head = READ_ONCE(ctx->sq_array[sq_idx]);
6772 if (likely(head < ctx->sq_entries))
6773 return &ctx->sq_sqes[head];
6775 /* drop invalid entries */
6777 WRITE_ONCE(ctx->rings->sq_dropped,
6778 READ_ONCE(ctx->rings->sq_dropped) + 1);
6782 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6783 __must_hold(&ctx->uring_lock)
6785 struct io_uring_task *tctx;
6788 /* make sure SQ entry isn't read before tail */
6789 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6790 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6793 tctx = current->io_uring;
6794 tctx->cached_refs -= nr;
6795 if (unlikely(tctx->cached_refs < 0)) {
6796 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6798 percpu_counter_add(&tctx->inflight, refill);
6799 refcount_add(refill, ¤t->usage);
6800 tctx->cached_refs += refill;
6802 io_submit_state_start(&ctx->submit_state, nr);
6804 while (submitted < nr) {
6805 const struct io_uring_sqe *sqe;
6806 struct io_kiocb *req;
6808 req = io_alloc_req(ctx);
6809 if (unlikely(!req)) {
6811 submitted = -EAGAIN;
6814 sqe = io_get_sqe(ctx);
6815 if (unlikely(!sqe)) {
6816 kmem_cache_free(req_cachep, req);
6819 /* will complete beyond this point, count as submitted */
6821 if (io_submit_sqe(ctx, req, sqe))
6825 if (unlikely(submitted != nr)) {
6826 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6827 int unused = nr - ref_used;
6829 current->io_uring->cached_refs += unused;
6830 percpu_ref_put_many(&ctx->refs, unused);
6833 io_submit_state_end(&ctx->submit_state, ctx);
6834 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6835 io_commit_sqring(ctx);
6840 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6842 return READ_ONCE(sqd->state);
6845 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6847 /* Tell userspace we may need a wakeup call */
6848 spin_lock(&ctx->completion_lock);
6849 WRITE_ONCE(ctx->rings->sq_flags,
6850 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6851 spin_unlock(&ctx->completion_lock);
6854 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6856 spin_lock(&ctx->completion_lock);
6857 WRITE_ONCE(ctx->rings->sq_flags,
6858 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6859 spin_unlock(&ctx->completion_lock);
6862 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6864 unsigned int to_submit;
6867 to_submit = io_sqring_entries(ctx);
6868 /* if we're handling multiple rings, cap submit size for fairness */
6869 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6870 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6872 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6873 unsigned nr_events = 0;
6874 const struct cred *creds = NULL;
6876 if (ctx->sq_creds != current_cred())
6877 creds = override_creds(ctx->sq_creds);
6879 mutex_lock(&ctx->uring_lock);
6880 if (!list_empty(&ctx->iopoll_list))
6881 io_do_iopoll(ctx, &nr_events, 0);
6884 * Don't submit if refs are dying, good for io_uring_register(),
6885 * but also it is relied upon by io_ring_exit_work()
6887 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6888 !(ctx->flags & IORING_SETUP_R_DISABLED))
6889 ret = io_submit_sqes(ctx, to_submit);
6890 mutex_unlock(&ctx->uring_lock);
6892 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6893 wake_up(&ctx->sqo_sq_wait);
6895 revert_creds(creds);
6901 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6903 struct io_ring_ctx *ctx;
6904 unsigned sq_thread_idle = 0;
6906 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6907 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6908 sqd->sq_thread_idle = sq_thread_idle;
6911 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6913 bool did_sig = false;
6914 struct ksignal ksig;
6916 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6917 signal_pending(current)) {
6918 mutex_unlock(&sqd->lock);
6919 if (signal_pending(current))
6920 did_sig = get_signal(&ksig);
6922 mutex_lock(&sqd->lock);
6924 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6927 static int io_sq_thread(void *data)
6929 struct io_sq_data *sqd = data;
6930 struct io_ring_ctx *ctx;
6931 unsigned long timeout = 0;
6932 char buf[TASK_COMM_LEN];
6935 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6936 set_task_comm(current, buf);
6938 if (sqd->sq_cpu != -1)
6939 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6941 set_cpus_allowed_ptr(current, cpu_online_mask);
6942 current->flags |= PF_NO_SETAFFINITY;
6944 mutex_lock(&sqd->lock);
6946 bool cap_entries, sqt_spin = false;
6948 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6949 if (io_sqd_handle_event(sqd))
6951 timeout = jiffies + sqd->sq_thread_idle;
6954 cap_entries = !list_is_singular(&sqd->ctx_list);
6955 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6956 int ret = __io_sq_thread(ctx, cap_entries);
6958 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6961 if (io_run_task_work())
6964 if (sqt_spin || !time_after(jiffies, timeout)) {
6967 timeout = jiffies + sqd->sq_thread_idle;
6971 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6972 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6973 bool needs_sched = true;
6975 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6976 io_ring_set_wakeup_flag(ctx);
6978 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6979 !list_empty_careful(&ctx->iopoll_list)) {
6980 needs_sched = false;
6983 if (io_sqring_entries(ctx)) {
6984 needs_sched = false;
6990 mutex_unlock(&sqd->lock);
6992 mutex_lock(&sqd->lock);
6994 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6995 io_ring_clear_wakeup_flag(ctx);
6998 finish_wait(&sqd->wait, &wait);
6999 timeout = jiffies + sqd->sq_thread_idle;
7002 io_uring_cancel_generic(true, sqd);
7004 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7005 io_ring_set_wakeup_flag(ctx);
7007 mutex_unlock(&sqd->lock);
7009 complete(&sqd->exited);
7013 struct io_wait_queue {
7014 struct wait_queue_entry wq;
7015 struct io_ring_ctx *ctx;
7017 unsigned nr_timeouts;
7020 static inline bool io_should_wake(struct io_wait_queue *iowq)
7022 struct io_ring_ctx *ctx = iowq->ctx;
7023 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7026 * Wake up if we have enough events, or if a timeout occurred since we
7027 * started waiting. For timeouts, we always want to return to userspace,
7028 * regardless of event count.
7030 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7033 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7034 int wake_flags, void *key)
7036 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7040 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7041 * the task, and the next invocation will do it.
7043 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7044 return autoremove_wake_function(curr, mode, wake_flags, key);
7048 static int io_run_task_work_sig(void)
7050 if (io_run_task_work())
7052 if (!signal_pending(current))
7054 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7055 return -ERESTARTSYS;
7059 /* when returns >0, the caller should retry */
7060 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7061 struct io_wait_queue *iowq,
7062 signed long *timeout)
7066 /* make sure we run task_work before checking for signals */
7067 ret = io_run_task_work_sig();
7068 if (ret || io_should_wake(iowq))
7070 /* let the caller flush overflows, retry */
7071 if (test_bit(0, &ctx->check_cq_overflow))
7074 *timeout = schedule_timeout(*timeout);
7075 return !*timeout ? -ETIME : 1;
7079 * Wait until events become available, if we don't already have some. The
7080 * application must reap them itself, as they reside on the shared cq ring.
7082 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7083 const sigset_t __user *sig, size_t sigsz,
7084 struct __kernel_timespec __user *uts)
7086 struct io_wait_queue iowq;
7087 struct io_rings *rings = ctx->rings;
7088 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7092 io_cqring_overflow_flush(ctx);
7093 if (io_cqring_events(ctx) >= min_events)
7095 if (!io_run_task_work())
7100 #ifdef CONFIG_COMPAT
7101 if (in_compat_syscall())
7102 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7106 ret = set_user_sigmask(sig, sigsz);
7113 struct timespec64 ts;
7115 if (get_timespec64(&ts, uts))
7117 timeout = timespec64_to_jiffies(&ts);
7120 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7121 iowq.wq.private = current;
7122 INIT_LIST_HEAD(&iowq.wq.entry);
7124 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7125 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7127 trace_io_uring_cqring_wait(ctx, min_events);
7129 /* if we can't even flush overflow, don't wait for more */
7130 if (!io_cqring_overflow_flush(ctx)) {
7134 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7135 TASK_INTERRUPTIBLE);
7136 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7137 finish_wait(&ctx->cq_wait, &iowq.wq);
7141 restore_saved_sigmask_unless(ret == -EINTR);
7143 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7146 static void io_free_page_table(void **table, size_t size)
7148 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7150 for (i = 0; i < nr_tables; i++)
7155 static void **io_alloc_page_table(size_t size)
7157 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7158 size_t init_size = size;
7161 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7165 for (i = 0; i < nr_tables; i++) {
7166 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7168 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7170 io_free_page_table(table, init_size);
7178 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7180 percpu_ref_exit(&ref_node->refs);
7184 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7186 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7187 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7188 unsigned long flags;
7189 bool first_add = false;
7191 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7194 while (!list_empty(&ctx->rsrc_ref_list)) {
7195 node = list_first_entry(&ctx->rsrc_ref_list,
7196 struct io_rsrc_node, node);
7197 /* recycle ref nodes in order */
7200 list_del(&node->node);
7201 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7203 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7206 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7209 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7211 struct io_rsrc_node *ref_node;
7213 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7217 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7222 INIT_LIST_HEAD(&ref_node->node);
7223 INIT_LIST_HEAD(&ref_node->rsrc_list);
7224 ref_node->done = false;
7228 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7229 struct io_rsrc_data *data_to_kill)
7231 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7232 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7235 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7237 rsrc_node->rsrc_data = data_to_kill;
7238 spin_lock_irq(&ctx->rsrc_ref_lock);
7239 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7240 spin_unlock_irq(&ctx->rsrc_ref_lock);
7242 atomic_inc(&data_to_kill->refs);
7243 percpu_ref_kill(&rsrc_node->refs);
7244 ctx->rsrc_node = NULL;
7247 if (!ctx->rsrc_node) {
7248 ctx->rsrc_node = ctx->rsrc_backup_node;
7249 ctx->rsrc_backup_node = NULL;
7253 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7255 if (ctx->rsrc_backup_node)
7257 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7258 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7261 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7265 /* As we may drop ->uring_lock, other task may have started quiesce */
7269 data->quiesce = true;
7271 ret = io_rsrc_node_switch_start(ctx);
7274 io_rsrc_node_switch(ctx, data);
7276 /* kill initial ref, already quiesced if zero */
7277 if (atomic_dec_and_test(&data->refs))
7279 mutex_unlock(&ctx->uring_lock);
7280 flush_delayed_work(&ctx->rsrc_put_work);
7281 ret = wait_for_completion_interruptible(&data->done);
7283 mutex_lock(&ctx->uring_lock);
7287 atomic_inc(&data->refs);
7288 /* wait for all works potentially completing data->done */
7289 flush_delayed_work(&ctx->rsrc_put_work);
7290 reinit_completion(&data->done);
7292 ret = io_run_task_work_sig();
7293 mutex_lock(&ctx->uring_lock);
7295 data->quiesce = false;
7300 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7302 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7303 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7305 return &data->tags[table_idx][off];
7308 static void io_rsrc_data_free(struct io_rsrc_data *data)
7310 size_t size = data->nr * sizeof(data->tags[0][0]);
7313 io_free_page_table((void **)data->tags, size);
7317 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7318 u64 __user *utags, unsigned nr,
7319 struct io_rsrc_data **pdata)
7321 struct io_rsrc_data *data;
7325 data = kzalloc(sizeof(*data), GFP_KERNEL);
7328 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7336 data->do_put = do_put;
7339 for (i = 0; i < nr; i++) {
7340 u64 *tag_slot = io_get_tag_slot(data, i);
7342 if (copy_from_user(tag_slot, &utags[i],
7348 atomic_set(&data->refs, 1);
7349 init_completion(&data->done);
7353 io_rsrc_data_free(data);
7357 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7359 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7360 GFP_KERNEL_ACCOUNT);
7361 return !!table->files;
7364 static void io_free_file_tables(struct io_file_table *table)
7366 kvfree(table->files);
7367 table->files = NULL;
7370 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7372 #if defined(CONFIG_UNIX)
7373 if (ctx->ring_sock) {
7374 struct sock *sock = ctx->ring_sock->sk;
7375 struct sk_buff *skb;
7377 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7383 for (i = 0; i < ctx->nr_user_files; i++) {
7386 file = io_file_from_index(ctx, i);
7391 io_free_file_tables(&ctx->file_table);
7392 io_rsrc_data_free(ctx->file_data);
7393 ctx->file_data = NULL;
7394 ctx->nr_user_files = 0;
7397 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7401 if (!ctx->file_data)
7403 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7405 __io_sqe_files_unregister(ctx);
7409 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7410 __releases(&sqd->lock)
7412 WARN_ON_ONCE(sqd->thread == current);
7415 * Do the dance but not conditional clear_bit() because it'd race with
7416 * other threads incrementing park_pending and setting the bit.
7418 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7419 if (atomic_dec_return(&sqd->park_pending))
7420 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7421 mutex_unlock(&sqd->lock);
7424 static void io_sq_thread_park(struct io_sq_data *sqd)
7425 __acquires(&sqd->lock)
7427 WARN_ON_ONCE(sqd->thread == current);
7429 atomic_inc(&sqd->park_pending);
7430 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7431 mutex_lock(&sqd->lock);
7433 wake_up_process(sqd->thread);
7436 static void io_sq_thread_stop(struct io_sq_data *sqd)
7438 WARN_ON_ONCE(sqd->thread == current);
7439 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7441 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7442 mutex_lock(&sqd->lock);
7444 wake_up_process(sqd->thread);
7445 mutex_unlock(&sqd->lock);
7446 wait_for_completion(&sqd->exited);
7449 static void io_put_sq_data(struct io_sq_data *sqd)
7451 if (refcount_dec_and_test(&sqd->refs)) {
7452 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7454 io_sq_thread_stop(sqd);
7459 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7461 struct io_sq_data *sqd = ctx->sq_data;
7464 io_sq_thread_park(sqd);
7465 list_del_init(&ctx->sqd_list);
7466 io_sqd_update_thread_idle(sqd);
7467 io_sq_thread_unpark(sqd);
7469 io_put_sq_data(sqd);
7470 ctx->sq_data = NULL;
7474 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7476 struct io_ring_ctx *ctx_attach;
7477 struct io_sq_data *sqd;
7480 f = fdget(p->wq_fd);
7482 return ERR_PTR(-ENXIO);
7483 if (f.file->f_op != &io_uring_fops) {
7485 return ERR_PTR(-EINVAL);
7488 ctx_attach = f.file->private_data;
7489 sqd = ctx_attach->sq_data;
7492 return ERR_PTR(-EINVAL);
7494 if (sqd->task_tgid != current->tgid) {
7496 return ERR_PTR(-EPERM);
7499 refcount_inc(&sqd->refs);
7504 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7507 struct io_sq_data *sqd;
7510 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7511 sqd = io_attach_sq_data(p);
7516 /* fall through for EPERM case, setup new sqd/task */
7517 if (PTR_ERR(sqd) != -EPERM)
7521 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7523 return ERR_PTR(-ENOMEM);
7525 atomic_set(&sqd->park_pending, 0);
7526 refcount_set(&sqd->refs, 1);
7527 INIT_LIST_HEAD(&sqd->ctx_list);
7528 mutex_init(&sqd->lock);
7529 init_waitqueue_head(&sqd->wait);
7530 init_completion(&sqd->exited);
7534 #if defined(CONFIG_UNIX)
7536 * Ensure the UNIX gc is aware of our file set, so we are certain that
7537 * the io_uring can be safely unregistered on process exit, even if we have
7538 * loops in the file referencing.
7540 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7542 struct sock *sk = ctx->ring_sock->sk;
7543 struct scm_fp_list *fpl;
7544 struct sk_buff *skb;
7547 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7551 skb = alloc_skb(0, GFP_KERNEL);
7560 fpl->user = get_uid(current_user());
7561 for (i = 0; i < nr; i++) {
7562 struct file *file = io_file_from_index(ctx, i + offset);
7566 fpl->fp[nr_files] = get_file(file);
7567 unix_inflight(fpl->user, fpl->fp[nr_files]);
7572 fpl->max = SCM_MAX_FD;
7573 fpl->count = nr_files;
7574 UNIXCB(skb).fp = fpl;
7575 skb->destructor = unix_destruct_scm;
7576 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7577 skb_queue_head(&sk->sk_receive_queue, skb);
7579 for (i = 0; i < nr_files; i++)
7590 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7591 * causes regular reference counting to break down. We rely on the UNIX
7592 * garbage collection to take care of this problem for us.
7594 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7596 unsigned left, total;
7600 left = ctx->nr_user_files;
7602 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7604 ret = __io_sqe_files_scm(ctx, this_files, total);
7608 total += this_files;
7614 while (total < ctx->nr_user_files) {
7615 struct file *file = io_file_from_index(ctx, total);
7625 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7631 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7633 struct file *file = prsrc->file;
7634 #if defined(CONFIG_UNIX)
7635 struct sock *sock = ctx->ring_sock->sk;
7636 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7637 struct sk_buff *skb;
7640 __skb_queue_head_init(&list);
7643 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7644 * remove this entry and rearrange the file array.
7646 skb = skb_dequeue(head);
7648 struct scm_fp_list *fp;
7650 fp = UNIXCB(skb).fp;
7651 for (i = 0; i < fp->count; i++) {
7654 if (fp->fp[i] != file)
7657 unix_notinflight(fp->user, fp->fp[i]);
7658 left = fp->count - 1 - i;
7660 memmove(&fp->fp[i], &fp->fp[i + 1],
7661 left * sizeof(struct file *));
7668 __skb_queue_tail(&list, skb);
7678 __skb_queue_tail(&list, skb);
7680 skb = skb_dequeue(head);
7683 if (skb_peek(&list)) {
7684 spin_lock_irq(&head->lock);
7685 while ((skb = __skb_dequeue(&list)) != NULL)
7686 __skb_queue_tail(head, skb);
7687 spin_unlock_irq(&head->lock);
7694 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7696 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7697 struct io_ring_ctx *ctx = rsrc_data->ctx;
7698 struct io_rsrc_put *prsrc, *tmp;
7700 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7701 list_del(&prsrc->list);
7704 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7706 io_ring_submit_lock(ctx, lock_ring);
7707 spin_lock(&ctx->completion_lock);
7708 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7710 io_commit_cqring(ctx);
7711 spin_unlock(&ctx->completion_lock);
7712 io_cqring_ev_posted(ctx);
7713 io_ring_submit_unlock(ctx, lock_ring);
7716 rsrc_data->do_put(ctx, prsrc);
7720 io_rsrc_node_destroy(ref_node);
7721 if (atomic_dec_and_test(&rsrc_data->refs))
7722 complete(&rsrc_data->done);
7725 static void io_rsrc_put_work(struct work_struct *work)
7727 struct io_ring_ctx *ctx;
7728 struct llist_node *node;
7730 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7731 node = llist_del_all(&ctx->rsrc_put_llist);
7734 struct io_rsrc_node *ref_node;
7735 struct llist_node *next = node->next;
7737 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7738 __io_rsrc_put_work(ref_node);
7743 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7744 unsigned nr_args, u64 __user *tags)
7746 __s32 __user *fds = (__s32 __user *) arg;
7755 if (nr_args > IORING_MAX_FIXED_FILES)
7757 if (nr_args > rlimit(RLIMIT_NOFILE))
7759 ret = io_rsrc_node_switch_start(ctx);
7762 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7768 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7771 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7772 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7776 /* allow sparse sets */
7779 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7786 if (unlikely(!file))
7790 * Don't allow io_uring instances to be registered. If UNIX
7791 * isn't enabled, then this causes a reference cycle and this
7792 * instance can never get freed. If UNIX is enabled we'll
7793 * handle it just fine, but there's still no point in allowing
7794 * a ring fd as it doesn't support regular read/write anyway.
7796 if (file->f_op == &io_uring_fops) {
7800 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7803 ret = io_sqe_files_scm(ctx);
7805 __io_sqe_files_unregister(ctx);
7809 io_rsrc_node_switch(ctx, NULL);
7812 for (i = 0; i < ctx->nr_user_files; i++) {
7813 file = io_file_from_index(ctx, i);
7817 io_free_file_tables(&ctx->file_table);
7818 ctx->nr_user_files = 0;
7820 io_rsrc_data_free(ctx->file_data);
7821 ctx->file_data = NULL;
7825 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7828 #if defined(CONFIG_UNIX)
7829 struct sock *sock = ctx->ring_sock->sk;
7830 struct sk_buff_head *head = &sock->sk_receive_queue;
7831 struct sk_buff *skb;
7834 * See if we can merge this file into an existing skb SCM_RIGHTS
7835 * file set. If there's no room, fall back to allocating a new skb
7836 * and filling it in.
7838 spin_lock_irq(&head->lock);
7839 skb = skb_peek(head);
7841 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7843 if (fpl->count < SCM_MAX_FD) {
7844 __skb_unlink(skb, head);
7845 spin_unlock_irq(&head->lock);
7846 fpl->fp[fpl->count] = get_file(file);
7847 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7849 spin_lock_irq(&head->lock);
7850 __skb_queue_head(head, skb);
7855 spin_unlock_irq(&head->lock);
7862 return __io_sqe_files_scm(ctx, 1, index);
7868 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7869 struct io_rsrc_node *node, void *rsrc)
7871 struct io_rsrc_put *prsrc;
7873 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7877 prsrc->tag = *io_get_tag_slot(data, idx);
7879 list_add(&prsrc->list, &node->rsrc_list);
7883 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7884 struct io_uring_rsrc_update2 *up,
7887 u64 __user *tags = u64_to_user_ptr(up->tags);
7888 __s32 __user *fds = u64_to_user_ptr(up->data);
7889 struct io_rsrc_data *data = ctx->file_data;
7890 struct io_fixed_file *file_slot;
7894 bool needs_switch = false;
7896 if (!ctx->file_data)
7898 if (up->offset + nr_args > ctx->nr_user_files)
7901 for (done = 0; done < nr_args; done++) {
7904 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7905 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7909 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7913 if (fd == IORING_REGISTER_FILES_SKIP)
7916 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7917 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7919 if (file_slot->file_ptr) {
7920 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7921 err = io_queue_rsrc_removal(data, up->offset + done,
7922 ctx->rsrc_node, file);
7925 file_slot->file_ptr = 0;
7926 needs_switch = true;
7935 * Don't allow io_uring instances to be registered. If
7936 * UNIX isn't enabled, then this causes a reference
7937 * cycle and this instance can never get freed. If UNIX
7938 * is enabled we'll handle it just fine, but there's
7939 * still no point in allowing a ring fd as it doesn't
7940 * support regular read/write anyway.
7942 if (file->f_op == &io_uring_fops) {
7947 *io_get_tag_slot(data, up->offset + done) = tag;
7948 io_fixed_file_set(file_slot, file);
7949 err = io_sqe_file_register(ctx, file, i);
7951 file_slot->file_ptr = 0;
7959 io_rsrc_node_switch(ctx, data);
7960 return done ? done : err;
7963 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7964 struct task_struct *task)
7966 struct io_wq_hash *hash;
7967 struct io_wq_data data;
7968 unsigned int concurrency;
7970 mutex_lock(&ctx->uring_lock);
7971 hash = ctx->hash_map;
7973 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7975 mutex_unlock(&ctx->uring_lock);
7976 return ERR_PTR(-ENOMEM);
7978 refcount_set(&hash->refs, 1);
7979 init_waitqueue_head(&hash->wait);
7980 ctx->hash_map = hash;
7982 mutex_unlock(&ctx->uring_lock);
7986 data.free_work = io_wq_free_work;
7987 data.do_work = io_wq_submit_work;
7989 /* Do QD, or 4 * CPUS, whatever is smallest */
7990 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7992 return io_wq_create(concurrency, &data);
7995 static int io_uring_alloc_task_context(struct task_struct *task,
7996 struct io_ring_ctx *ctx)
7998 struct io_uring_task *tctx;
8001 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8002 if (unlikely(!tctx))
8005 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8006 if (unlikely(ret)) {
8011 tctx->io_wq = io_init_wq_offload(ctx, task);
8012 if (IS_ERR(tctx->io_wq)) {
8013 ret = PTR_ERR(tctx->io_wq);
8014 percpu_counter_destroy(&tctx->inflight);
8020 init_waitqueue_head(&tctx->wait);
8021 atomic_set(&tctx->in_idle, 0);
8022 atomic_set(&tctx->inflight_tracked, 0);
8023 task->io_uring = tctx;
8024 spin_lock_init(&tctx->task_lock);
8025 INIT_WQ_LIST(&tctx->task_list);
8026 init_task_work(&tctx->task_work, tctx_task_work);
8030 void __io_uring_free(struct task_struct *tsk)
8032 struct io_uring_task *tctx = tsk->io_uring;
8034 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8035 WARN_ON_ONCE(tctx->io_wq);
8036 WARN_ON_ONCE(tctx->cached_refs);
8038 percpu_counter_destroy(&tctx->inflight);
8040 tsk->io_uring = NULL;
8043 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8044 struct io_uring_params *p)
8048 /* Retain compatibility with failing for an invalid attach attempt */
8049 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8050 IORING_SETUP_ATTACH_WQ) {
8053 f = fdget(p->wq_fd);
8056 if (f.file->f_op != &io_uring_fops) {
8062 if (ctx->flags & IORING_SETUP_SQPOLL) {
8063 struct task_struct *tsk;
8064 struct io_sq_data *sqd;
8067 sqd = io_get_sq_data(p, &attached);
8073 ctx->sq_creds = get_current_cred();
8075 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8076 if (!ctx->sq_thread_idle)
8077 ctx->sq_thread_idle = HZ;
8079 io_sq_thread_park(sqd);
8080 list_add(&ctx->sqd_list, &sqd->ctx_list);
8081 io_sqd_update_thread_idle(sqd);
8082 /* don't attach to a dying SQPOLL thread, would be racy */
8083 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8084 io_sq_thread_unpark(sqd);
8091 if (p->flags & IORING_SETUP_SQ_AFF) {
8092 int cpu = p->sq_thread_cpu;
8095 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8102 sqd->task_pid = current->pid;
8103 sqd->task_tgid = current->tgid;
8104 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8111 ret = io_uring_alloc_task_context(tsk, ctx);
8112 wake_up_new_task(tsk);
8115 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8116 /* Can't have SQ_AFF without SQPOLL */
8123 complete(&ctx->sq_data->exited);
8125 io_sq_thread_finish(ctx);
8129 static inline void __io_unaccount_mem(struct user_struct *user,
8130 unsigned long nr_pages)
8132 atomic_long_sub(nr_pages, &user->locked_vm);
8135 static inline int __io_account_mem(struct user_struct *user,
8136 unsigned long nr_pages)
8138 unsigned long page_limit, cur_pages, new_pages;
8140 /* Don't allow more pages than we can safely lock */
8141 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8144 cur_pages = atomic_long_read(&user->locked_vm);
8145 new_pages = cur_pages + nr_pages;
8146 if (new_pages > page_limit)
8148 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8149 new_pages) != cur_pages);
8154 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8157 __io_unaccount_mem(ctx->user, nr_pages);
8159 if (ctx->mm_account)
8160 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8163 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8168 ret = __io_account_mem(ctx->user, nr_pages);
8173 if (ctx->mm_account)
8174 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8179 static void io_mem_free(void *ptr)
8186 page = virt_to_head_page(ptr);
8187 if (put_page_testzero(page))
8188 free_compound_page(page);
8191 static void *io_mem_alloc(size_t size)
8193 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8194 __GFP_NORETRY | __GFP_ACCOUNT;
8196 return (void *) __get_free_pages(gfp_flags, get_order(size));
8199 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8202 struct io_rings *rings;
8203 size_t off, sq_array_size;
8205 off = struct_size(rings, cqes, cq_entries);
8206 if (off == SIZE_MAX)
8210 off = ALIGN(off, SMP_CACHE_BYTES);
8218 sq_array_size = array_size(sizeof(u32), sq_entries);
8219 if (sq_array_size == SIZE_MAX)
8222 if (check_add_overflow(off, sq_array_size, &off))
8228 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8230 struct io_mapped_ubuf *imu = *slot;
8233 if (imu != ctx->dummy_ubuf) {
8234 for (i = 0; i < imu->nr_bvecs; i++)
8235 unpin_user_page(imu->bvec[i].bv_page);
8236 if (imu->acct_pages)
8237 io_unaccount_mem(ctx, imu->acct_pages);
8243 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8245 io_buffer_unmap(ctx, &prsrc->buf);
8249 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8253 for (i = 0; i < ctx->nr_user_bufs; i++)
8254 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8255 kfree(ctx->user_bufs);
8256 io_rsrc_data_free(ctx->buf_data);
8257 ctx->user_bufs = NULL;
8258 ctx->buf_data = NULL;
8259 ctx->nr_user_bufs = 0;
8262 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8269 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8271 __io_sqe_buffers_unregister(ctx);
8275 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8276 void __user *arg, unsigned index)
8278 struct iovec __user *src;
8280 #ifdef CONFIG_COMPAT
8282 struct compat_iovec __user *ciovs;
8283 struct compat_iovec ciov;
8285 ciovs = (struct compat_iovec __user *) arg;
8286 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8289 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8290 dst->iov_len = ciov.iov_len;
8294 src = (struct iovec __user *) arg;
8295 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8301 * Not super efficient, but this is just a registration time. And we do cache
8302 * the last compound head, so generally we'll only do a full search if we don't
8305 * We check if the given compound head page has already been accounted, to
8306 * avoid double accounting it. This allows us to account the full size of the
8307 * page, not just the constituent pages of a huge page.
8309 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8310 int nr_pages, struct page *hpage)
8314 /* check current page array */
8315 for (i = 0; i < nr_pages; i++) {
8316 if (!PageCompound(pages[i]))
8318 if (compound_head(pages[i]) == hpage)
8322 /* check previously registered pages */
8323 for (i = 0; i < ctx->nr_user_bufs; i++) {
8324 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8326 for (j = 0; j < imu->nr_bvecs; j++) {
8327 if (!PageCompound(imu->bvec[j].bv_page))
8329 if (compound_head(imu->bvec[j].bv_page) == hpage)
8337 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8338 int nr_pages, struct io_mapped_ubuf *imu,
8339 struct page **last_hpage)
8343 imu->acct_pages = 0;
8344 for (i = 0; i < nr_pages; i++) {
8345 if (!PageCompound(pages[i])) {
8350 hpage = compound_head(pages[i]);
8351 if (hpage == *last_hpage)
8353 *last_hpage = hpage;
8354 if (headpage_already_acct(ctx, pages, i, hpage))
8356 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8360 if (!imu->acct_pages)
8363 ret = io_account_mem(ctx, imu->acct_pages);
8365 imu->acct_pages = 0;
8369 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8370 struct io_mapped_ubuf **pimu,
8371 struct page **last_hpage)
8373 struct io_mapped_ubuf *imu = NULL;
8374 struct vm_area_struct **vmas = NULL;
8375 struct page **pages = NULL;
8376 unsigned long off, start, end, ubuf;
8378 int ret, pret, nr_pages, i;
8380 if (!iov->iov_base) {
8381 *pimu = ctx->dummy_ubuf;
8385 ubuf = (unsigned long) iov->iov_base;
8386 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8387 start = ubuf >> PAGE_SHIFT;
8388 nr_pages = end - start;
8393 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8397 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8402 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8407 mmap_read_lock(current->mm);
8408 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8410 if (pret == nr_pages) {
8411 /* don't support file backed memory */
8412 for (i = 0; i < nr_pages; i++) {
8413 struct vm_area_struct *vma = vmas[i];
8415 if (vma_is_shmem(vma))
8418 !is_file_hugepages(vma->vm_file)) {
8424 ret = pret < 0 ? pret : -EFAULT;
8426 mmap_read_unlock(current->mm);
8429 * if we did partial map, or found file backed vmas,
8430 * release any pages we did get
8433 unpin_user_pages(pages, pret);
8437 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8439 unpin_user_pages(pages, pret);
8443 off = ubuf & ~PAGE_MASK;
8444 size = iov->iov_len;
8445 for (i = 0; i < nr_pages; i++) {
8448 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8449 imu->bvec[i].bv_page = pages[i];
8450 imu->bvec[i].bv_len = vec_len;
8451 imu->bvec[i].bv_offset = off;
8455 /* store original address for later verification */
8457 imu->ubuf_end = ubuf + iov->iov_len;
8458 imu->nr_bvecs = nr_pages;
8469 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8471 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8472 return ctx->user_bufs ? 0 : -ENOMEM;
8475 static int io_buffer_validate(struct iovec *iov)
8477 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8480 * Don't impose further limits on the size and buffer
8481 * constraints here, we'll -EINVAL later when IO is
8482 * submitted if they are wrong.
8485 return iov->iov_len ? -EFAULT : 0;
8489 /* arbitrary limit, but we need something */
8490 if (iov->iov_len > SZ_1G)
8493 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8499 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8500 unsigned int nr_args, u64 __user *tags)
8502 struct page *last_hpage = NULL;
8503 struct io_rsrc_data *data;
8509 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8511 ret = io_rsrc_node_switch_start(ctx);
8514 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8517 ret = io_buffers_map_alloc(ctx, nr_args);
8519 io_rsrc_data_free(data);
8523 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8524 ret = io_copy_iov(ctx, &iov, arg, i);
8527 ret = io_buffer_validate(&iov);
8530 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8535 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8541 WARN_ON_ONCE(ctx->buf_data);
8543 ctx->buf_data = data;
8545 __io_sqe_buffers_unregister(ctx);
8547 io_rsrc_node_switch(ctx, NULL);
8551 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8552 struct io_uring_rsrc_update2 *up,
8553 unsigned int nr_args)
8555 u64 __user *tags = u64_to_user_ptr(up->tags);
8556 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8557 struct page *last_hpage = NULL;
8558 bool needs_switch = false;
8564 if (up->offset + nr_args > ctx->nr_user_bufs)
8567 for (done = 0; done < nr_args; done++) {
8568 struct io_mapped_ubuf *imu;
8569 int offset = up->offset + done;
8572 err = io_copy_iov(ctx, &iov, iovs, done);
8575 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8579 err = io_buffer_validate(&iov);
8582 if (!iov.iov_base && tag) {
8586 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8590 i = array_index_nospec(offset, ctx->nr_user_bufs);
8591 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8592 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8593 ctx->rsrc_node, ctx->user_bufs[i]);
8594 if (unlikely(err)) {
8595 io_buffer_unmap(ctx, &imu);
8598 ctx->user_bufs[i] = NULL;
8599 needs_switch = true;
8602 ctx->user_bufs[i] = imu;
8603 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8607 io_rsrc_node_switch(ctx, ctx->buf_data);
8608 return done ? done : err;
8611 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8613 __s32 __user *fds = arg;
8619 if (copy_from_user(&fd, fds, sizeof(*fds)))
8622 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8623 if (IS_ERR(ctx->cq_ev_fd)) {
8624 int ret = PTR_ERR(ctx->cq_ev_fd);
8626 ctx->cq_ev_fd = NULL;
8633 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8635 if (ctx->cq_ev_fd) {
8636 eventfd_ctx_put(ctx->cq_ev_fd);
8637 ctx->cq_ev_fd = NULL;
8644 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8646 struct io_buffer *buf;
8647 unsigned long index;
8649 xa_for_each(&ctx->io_buffers, index, buf)
8650 __io_remove_buffers(ctx, buf, index, -1U);
8653 static void io_req_cache_free(struct list_head *list)
8655 struct io_kiocb *req, *nxt;
8657 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8658 list_del(&req->inflight_entry);
8659 kmem_cache_free(req_cachep, req);
8663 static void io_req_caches_free(struct io_ring_ctx *ctx)
8665 struct io_submit_state *state = &ctx->submit_state;
8667 mutex_lock(&ctx->uring_lock);
8669 if (state->free_reqs) {
8670 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8671 state->free_reqs = 0;
8674 io_flush_cached_locked_reqs(ctx, state);
8675 io_req_cache_free(&state->free_list);
8676 mutex_unlock(&ctx->uring_lock);
8679 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8681 if (data && !atomic_dec_and_test(&data->refs))
8682 wait_for_completion(&data->done);
8685 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8687 io_sq_thread_finish(ctx);
8689 if (ctx->mm_account) {
8690 mmdrop(ctx->mm_account);
8691 ctx->mm_account = NULL;
8694 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8695 io_wait_rsrc_data(ctx->buf_data);
8696 io_wait_rsrc_data(ctx->file_data);
8698 mutex_lock(&ctx->uring_lock);
8700 __io_sqe_buffers_unregister(ctx);
8702 __io_sqe_files_unregister(ctx);
8704 __io_cqring_overflow_flush(ctx, true);
8705 mutex_unlock(&ctx->uring_lock);
8706 io_eventfd_unregister(ctx);
8707 io_destroy_buffers(ctx);
8709 put_cred(ctx->sq_creds);
8711 /* there are no registered resources left, nobody uses it */
8713 io_rsrc_node_destroy(ctx->rsrc_node);
8714 if (ctx->rsrc_backup_node)
8715 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8716 flush_delayed_work(&ctx->rsrc_put_work);
8718 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8719 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8721 #if defined(CONFIG_UNIX)
8722 if (ctx->ring_sock) {
8723 ctx->ring_sock->file = NULL; /* so that iput() is called */
8724 sock_release(ctx->ring_sock);
8728 io_mem_free(ctx->rings);
8729 io_mem_free(ctx->sq_sqes);
8731 percpu_ref_exit(&ctx->refs);
8732 free_uid(ctx->user);
8733 io_req_caches_free(ctx);
8735 io_wq_put_hash(ctx->hash_map);
8736 kfree(ctx->cancel_hash);
8737 kfree(ctx->dummy_ubuf);
8741 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8743 struct io_ring_ctx *ctx = file->private_data;
8746 poll_wait(file, &ctx->poll_wait, wait);
8748 * synchronizes with barrier from wq_has_sleeper call in
8752 if (!io_sqring_full(ctx))
8753 mask |= EPOLLOUT | EPOLLWRNORM;
8756 * Don't flush cqring overflow list here, just do a simple check.
8757 * Otherwise there could possible be ABBA deadlock:
8760 * lock(&ctx->uring_lock);
8762 * lock(&ctx->uring_lock);
8765 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8766 * pushs them to do the flush.
8768 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8769 mask |= EPOLLIN | EPOLLRDNORM;
8774 static int io_uring_fasync(int fd, struct file *file, int on)
8776 struct io_ring_ctx *ctx = file->private_data;
8778 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8781 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8783 const struct cred *creds;
8785 creds = xa_erase(&ctx->personalities, id);
8794 struct io_tctx_exit {
8795 struct callback_head task_work;
8796 struct completion completion;
8797 struct io_ring_ctx *ctx;
8800 static void io_tctx_exit_cb(struct callback_head *cb)
8802 struct io_uring_task *tctx = current->io_uring;
8803 struct io_tctx_exit *work;
8805 work = container_of(cb, struct io_tctx_exit, task_work);
8807 * When @in_idle, we're in cancellation and it's racy to remove the
8808 * node. It'll be removed by the end of cancellation, just ignore it.
8810 if (!atomic_read(&tctx->in_idle))
8811 io_uring_del_tctx_node((unsigned long)work->ctx);
8812 complete(&work->completion);
8815 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8817 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8819 return req->ctx == data;
8822 static void io_ring_exit_work(struct work_struct *work)
8824 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8825 unsigned long timeout = jiffies + HZ * 60 * 5;
8826 unsigned long interval = HZ / 20;
8827 struct io_tctx_exit exit;
8828 struct io_tctx_node *node;
8832 * If we're doing polled IO and end up having requests being
8833 * submitted async (out-of-line), then completions can come in while
8834 * we're waiting for refs to drop. We need to reap these manually,
8835 * as nobody else will be looking for them.
8838 io_uring_try_cancel_requests(ctx, NULL, true);
8840 struct io_sq_data *sqd = ctx->sq_data;
8841 struct task_struct *tsk;
8843 io_sq_thread_park(sqd);
8845 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8846 io_wq_cancel_cb(tsk->io_uring->io_wq,
8847 io_cancel_ctx_cb, ctx, true);
8848 io_sq_thread_unpark(sqd);
8851 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8852 /* there is little hope left, don't run it too often */
8855 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8857 init_completion(&exit.completion);
8858 init_task_work(&exit.task_work, io_tctx_exit_cb);
8861 * Some may use context even when all refs and requests have been put,
8862 * and they are free to do so while still holding uring_lock or
8863 * completion_lock, see io_req_task_submit(). Apart from other work,
8864 * this lock/unlock section also waits them to finish.
8866 mutex_lock(&ctx->uring_lock);
8867 while (!list_empty(&ctx->tctx_list)) {
8868 WARN_ON_ONCE(time_after(jiffies, timeout));
8870 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8872 /* don't spin on a single task if cancellation failed */
8873 list_rotate_left(&ctx->tctx_list);
8874 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8875 if (WARN_ON_ONCE(ret))
8877 wake_up_process(node->task);
8879 mutex_unlock(&ctx->uring_lock);
8880 wait_for_completion(&exit.completion);
8881 mutex_lock(&ctx->uring_lock);
8883 mutex_unlock(&ctx->uring_lock);
8884 spin_lock(&ctx->completion_lock);
8885 spin_unlock(&ctx->completion_lock);
8887 io_ring_ctx_free(ctx);
8890 /* Returns true if we found and killed one or more timeouts */
8891 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8894 struct io_kiocb *req, *tmp;
8897 spin_lock(&ctx->completion_lock);
8898 spin_lock_irq(&ctx->timeout_lock);
8899 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8900 if (io_match_task(req, tsk, cancel_all)) {
8901 io_kill_timeout(req, -ECANCELED);
8905 spin_unlock_irq(&ctx->timeout_lock);
8907 io_commit_cqring(ctx);
8908 spin_unlock(&ctx->completion_lock);
8910 io_cqring_ev_posted(ctx);
8911 return canceled != 0;
8914 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8916 unsigned long index;
8917 struct creds *creds;
8919 mutex_lock(&ctx->uring_lock);
8920 percpu_ref_kill(&ctx->refs);
8922 __io_cqring_overflow_flush(ctx, true);
8923 xa_for_each(&ctx->personalities, index, creds)
8924 io_unregister_personality(ctx, index);
8925 mutex_unlock(&ctx->uring_lock);
8927 io_kill_timeouts(ctx, NULL, true);
8928 io_poll_remove_all(ctx, NULL, true);
8930 /* if we failed setting up the ctx, we might not have any rings */
8931 io_iopoll_try_reap_events(ctx);
8933 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8935 * Use system_unbound_wq to avoid spawning tons of event kworkers
8936 * if we're exiting a ton of rings at the same time. It just adds
8937 * noise and overhead, there's no discernable change in runtime
8938 * over using system_wq.
8940 queue_work(system_unbound_wq, &ctx->exit_work);
8943 static int io_uring_release(struct inode *inode, struct file *file)
8945 struct io_ring_ctx *ctx = file->private_data;
8947 file->private_data = NULL;
8948 io_ring_ctx_wait_and_kill(ctx);
8952 struct io_task_cancel {
8953 struct task_struct *task;
8957 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8959 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8960 struct io_task_cancel *cancel = data;
8963 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8964 struct io_ring_ctx *ctx = req->ctx;
8966 /* protect against races with linked timeouts */
8967 spin_lock(&ctx->completion_lock);
8968 ret = io_match_task(req, cancel->task, cancel->all);
8969 spin_unlock(&ctx->completion_lock);
8971 ret = io_match_task(req, cancel->task, cancel->all);
8976 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8977 struct task_struct *task, bool cancel_all)
8979 struct io_defer_entry *de;
8982 spin_lock(&ctx->completion_lock);
8983 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8984 if (io_match_task(de->req, task, cancel_all)) {
8985 list_cut_position(&list, &ctx->defer_list, &de->list);
8989 spin_unlock(&ctx->completion_lock);
8990 if (list_empty(&list))
8993 while (!list_empty(&list)) {
8994 de = list_first_entry(&list, struct io_defer_entry, list);
8995 list_del_init(&de->list);
8996 io_req_complete_failed(de->req, -ECANCELED);
9002 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9004 struct io_tctx_node *node;
9005 enum io_wq_cancel cret;
9008 mutex_lock(&ctx->uring_lock);
9009 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9010 struct io_uring_task *tctx = node->task->io_uring;
9013 * io_wq will stay alive while we hold uring_lock, because it's
9014 * killed after ctx nodes, which requires to take the lock.
9016 if (!tctx || !tctx->io_wq)
9018 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9019 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9021 mutex_unlock(&ctx->uring_lock);
9026 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9027 struct task_struct *task,
9030 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9031 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9034 enum io_wq_cancel cret;
9038 ret |= io_uring_try_cancel_iowq(ctx);
9039 } else if (tctx && tctx->io_wq) {
9041 * Cancels requests of all rings, not only @ctx, but
9042 * it's fine as the task is in exit/exec.
9044 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9046 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9049 /* SQPOLL thread does its own polling */
9050 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9051 (ctx->sq_data && ctx->sq_data->thread == current)) {
9052 while (!list_empty_careful(&ctx->iopoll_list)) {
9053 io_iopoll_try_reap_events(ctx);
9058 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9059 ret |= io_poll_remove_all(ctx, task, cancel_all);
9060 ret |= io_kill_timeouts(ctx, task, cancel_all);
9062 ret |= io_run_task_work();
9069 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9071 struct io_uring_task *tctx = current->io_uring;
9072 struct io_tctx_node *node;
9075 if (unlikely(!tctx)) {
9076 ret = io_uring_alloc_task_context(current, ctx);
9079 tctx = current->io_uring;
9081 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9082 node = kmalloc(sizeof(*node), GFP_KERNEL);
9086 node->task = current;
9088 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9095 mutex_lock(&ctx->uring_lock);
9096 list_add(&node->ctx_node, &ctx->tctx_list);
9097 mutex_unlock(&ctx->uring_lock);
9104 * Note that this task has used io_uring. We use it for cancelation purposes.
9106 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9108 struct io_uring_task *tctx = current->io_uring;
9110 if (likely(tctx && tctx->last == ctx))
9112 return __io_uring_add_tctx_node(ctx);
9116 * Remove this io_uring_file -> task mapping.
9118 static void io_uring_del_tctx_node(unsigned long index)
9120 struct io_uring_task *tctx = current->io_uring;
9121 struct io_tctx_node *node;
9125 node = xa_erase(&tctx->xa, index);
9129 WARN_ON_ONCE(current != node->task);
9130 WARN_ON_ONCE(list_empty(&node->ctx_node));
9132 mutex_lock(&node->ctx->uring_lock);
9133 list_del(&node->ctx_node);
9134 mutex_unlock(&node->ctx->uring_lock);
9136 if (tctx->last == node->ctx)
9141 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9143 struct io_wq *wq = tctx->io_wq;
9144 struct io_tctx_node *node;
9145 unsigned long index;
9147 xa_for_each(&tctx->xa, index, node)
9148 io_uring_del_tctx_node(index);
9151 * Must be after io_uring_del_task_file() (removes nodes under
9152 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9155 io_wq_put_and_exit(wq);
9159 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9162 return atomic_read(&tctx->inflight_tracked);
9163 return percpu_counter_sum(&tctx->inflight);
9166 static void io_uring_drop_tctx_refs(struct task_struct *task)
9168 struct io_uring_task *tctx = task->io_uring;
9169 unsigned int refs = tctx->cached_refs;
9172 tctx->cached_refs = 0;
9173 percpu_counter_sub(&tctx->inflight, refs);
9174 put_task_struct_many(task, refs);
9179 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9180 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9182 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9184 struct io_uring_task *tctx = current->io_uring;
9185 struct io_ring_ctx *ctx;
9189 WARN_ON_ONCE(sqd && sqd->thread != current);
9191 if (!current->io_uring)
9194 io_wq_exit_start(tctx->io_wq);
9196 atomic_inc(&tctx->in_idle);
9198 io_uring_drop_tctx_refs(current);
9199 /* read completions before cancelations */
9200 inflight = tctx_inflight(tctx, !cancel_all);
9205 struct io_tctx_node *node;
9206 unsigned long index;
9208 xa_for_each(&tctx->xa, index, node) {
9209 /* sqpoll task will cancel all its requests */
9210 if (node->ctx->sq_data)
9212 io_uring_try_cancel_requests(node->ctx, current,
9216 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9217 io_uring_try_cancel_requests(ctx, current,
9221 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9222 io_uring_drop_tctx_refs(current);
9224 * If we've seen completions, retry without waiting. This
9225 * avoids a race where a completion comes in before we did
9226 * prepare_to_wait().
9228 if (inflight == tctx_inflight(tctx, !cancel_all))
9230 finish_wait(&tctx->wait, &wait);
9232 atomic_dec(&tctx->in_idle);
9234 io_uring_clean_tctx(tctx);
9236 /* for exec all current's requests should be gone, kill tctx */
9237 __io_uring_free(current);
9241 void __io_uring_cancel(bool cancel_all)
9243 io_uring_cancel_generic(cancel_all, NULL);
9246 static void *io_uring_validate_mmap_request(struct file *file,
9247 loff_t pgoff, size_t sz)
9249 struct io_ring_ctx *ctx = file->private_data;
9250 loff_t offset = pgoff << PAGE_SHIFT;
9255 case IORING_OFF_SQ_RING:
9256 case IORING_OFF_CQ_RING:
9259 case IORING_OFF_SQES:
9263 return ERR_PTR(-EINVAL);
9266 page = virt_to_head_page(ptr);
9267 if (sz > page_size(page))
9268 return ERR_PTR(-EINVAL);
9275 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9277 size_t sz = vma->vm_end - vma->vm_start;
9281 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9283 return PTR_ERR(ptr);
9285 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9286 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9289 #else /* !CONFIG_MMU */
9291 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9293 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9296 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9298 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9301 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9302 unsigned long addr, unsigned long len,
9303 unsigned long pgoff, unsigned long flags)
9307 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9309 return PTR_ERR(ptr);
9311 return (unsigned long) ptr;
9314 #endif /* !CONFIG_MMU */
9316 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9321 if (!io_sqring_full(ctx))
9323 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9325 if (!io_sqring_full(ctx))
9328 } while (!signal_pending(current));
9330 finish_wait(&ctx->sqo_sq_wait, &wait);
9334 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9335 struct __kernel_timespec __user **ts,
9336 const sigset_t __user **sig)
9338 struct io_uring_getevents_arg arg;
9341 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9342 * is just a pointer to the sigset_t.
9344 if (!(flags & IORING_ENTER_EXT_ARG)) {
9345 *sig = (const sigset_t __user *) argp;
9351 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9352 * timespec and sigset_t pointers if good.
9354 if (*argsz != sizeof(arg))
9356 if (copy_from_user(&arg, argp, sizeof(arg)))
9358 *sig = u64_to_user_ptr(arg.sigmask);
9359 *argsz = arg.sigmask_sz;
9360 *ts = u64_to_user_ptr(arg.ts);
9364 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9365 u32, min_complete, u32, flags, const void __user *, argp,
9368 struct io_ring_ctx *ctx;
9375 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9376 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9380 if (unlikely(!f.file))
9384 if (unlikely(f.file->f_op != &io_uring_fops))
9388 ctx = f.file->private_data;
9389 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9393 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9397 * For SQ polling, the thread will do all submissions and completions.
9398 * Just return the requested submit count, and wake the thread if
9402 if (ctx->flags & IORING_SETUP_SQPOLL) {
9403 io_cqring_overflow_flush(ctx);
9405 if (unlikely(ctx->sq_data->thread == NULL)) {
9409 if (flags & IORING_ENTER_SQ_WAKEUP)
9410 wake_up(&ctx->sq_data->wait);
9411 if (flags & IORING_ENTER_SQ_WAIT) {
9412 ret = io_sqpoll_wait_sq(ctx);
9416 submitted = to_submit;
9417 } else if (to_submit) {
9418 ret = io_uring_add_tctx_node(ctx);
9421 mutex_lock(&ctx->uring_lock);
9422 submitted = io_submit_sqes(ctx, to_submit);
9423 mutex_unlock(&ctx->uring_lock);
9425 if (submitted != to_submit)
9428 if (flags & IORING_ENTER_GETEVENTS) {
9429 const sigset_t __user *sig;
9430 struct __kernel_timespec __user *ts;
9432 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9436 min_complete = min(min_complete, ctx->cq_entries);
9439 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9440 * space applications don't need to do io completion events
9441 * polling again, they can rely on io_sq_thread to do polling
9442 * work, which can reduce cpu usage and uring_lock contention.
9444 if (ctx->flags & IORING_SETUP_IOPOLL &&
9445 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9446 ret = io_iopoll_check(ctx, min_complete);
9448 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9453 percpu_ref_put(&ctx->refs);
9456 return submitted ? submitted : ret;
9459 #ifdef CONFIG_PROC_FS
9460 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9461 const struct cred *cred)
9463 struct user_namespace *uns = seq_user_ns(m);
9464 struct group_info *gi;
9469 seq_printf(m, "%5d\n", id);
9470 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9471 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9472 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9473 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9474 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9475 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9476 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9477 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9478 seq_puts(m, "\n\tGroups:\t");
9479 gi = cred->group_info;
9480 for (g = 0; g < gi->ngroups; g++) {
9481 seq_put_decimal_ull(m, g ? " " : "",
9482 from_kgid_munged(uns, gi->gid[g]));
9484 seq_puts(m, "\n\tCapEff:\t");
9485 cap = cred->cap_effective;
9486 CAP_FOR_EACH_U32(__capi)
9487 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9492 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9494 struct io_sq_data *sq = NULL;
9499 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9500 * since fdinfo case grabs it in the opposite direction of normal use
9501 * cases. If we fail to get the lock, we just don't iterate any
9502 * structures that could be going away outside the io_uring mutex.
9504 has_lock = mutex_trylock(&ctx->uring_lock);
9506 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9512 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9513 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9514 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9515 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9516 struct file *f = io_file_from_index(ctx, i);
9519 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9521 seq_printf(m, "%5u: <none>\n", i);
9523 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9524 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9525 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9526 unsigned int len = buf->ubuf_end - buf->ubuf;
9528 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9530 if (has_lock && !xa_empty(&ctx->personalities)) {
9531 unsigned long index;
9532 const struct cred *cred;
9534 seq_printf(m, "Personalities:\n");
9535 xa_for_each(&ctx->personalities, index, cred)
9536 io_uring_show_cred(m, index, cred);
9538 seq_printf(m, "PollList:\n");
9539 spin_lock(&ctx->completion_lock);
9540 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9541 struct hlist_head *list = &ctx->cancel_hash[i];
9542 struct io_kiocb *req;
9544 hlist_for_each_entry(req, list, hash_node)
9545 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9546 req->task->task_works != NULL);
9548 spin_unlock(&ctx->completion_lock);
9550 mutex_unlock(&ctx->uring_lock);
9553 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9555 struct io_ring_ctx *ctx = f->private_data;
9557 if (percpu_ref_tryget(&ctx->refs)) {
9558 __io_uring_show_fdinfo(ctx, m);
9559 percpu_ref_put(&ctx->refs);
9564 static const struct file_operations io_uring_fops = {
9565 .release = io_uring_release,
9566 .mmap = io_uring_mmap,
9568 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9569 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9571 .poll = io_uring_poll,
9572 .fasync = io_uring_fasync,
9573 #ifdef CONFIG_PROC_FS
9574 .show_fdinfo = io_uring_show_fdinfo,
9578 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9579 struct io_uring_params *p)
9581 struct io_rings *rings;
9582 size_t size, sq_array_offset;
9584 /* make sure these are sane, as we already accounted them */
9585 ctx->sq_entries = p->sq_entries;
9586 ctx->cq_entries = p->cq_entries;
9588 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9589 if (size == SIZE_MAX)
9592 rings = io_mem_alloc(size);
9597 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9598 rings->sq_ring_mask = p->sq_entries - 1;
9599 rings->cq_ring_mask = p->cq_entries - 1;
9600 rings->sq_ring_entries = p->sq_entries;
9601 rings->cq_ring_entries = p->cq_entries;
9603 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9604 if (size == SIZE_MAX) {
9605 io_mem_free(ctx->rings);
9610 ctx->sq_sqes = io_mem_alloc(size);
9611 if (!ctx->sq_sqes) {
9612 io_mem_free(ctx->rings);
9620 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9624 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9628 ret = io_uring_add_tctx_node(ctx);
9633 fd_install(fd, file);
9638 * Allocate an anonymous fd, this is what constitutes the application
9639 * visible backing of an io_uring instance. The application mmaps this
9640 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9641 * we have to tie this fd to a socket for file garbage collection purposes.
9643 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9646 #if defined(CONFIG_UNIX)
9649 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9652 return ERR_PTR(ret);
9655 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9656 O_RDWR | O_CLOEXEC);
9657 #if defined(CONFIG_UNIX)
9659 sock_release(ctx->ring_sock);
9660 ctx->ring_sock = NULL;
9662 ctx->ring_sock->file = file;
9668 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9669 struct io_uring_params __user *params)
9671 struct io_ring_ctx *ctx;
9677 if (entries > IORING_MAX_ENTRIES) {
9678 if (!(p->flags & IORING_SETUP_CLAMP))
9680 entries = IORING_MAX_ENTRIES;
9684 * Use twice as many entries for the CQ ring. It's possible for the
9685 * application to drive a higher depth than the size of the SQ ring,
9686 * since the sqes are only used at submission time. This allows for
9687 * some flexibility in overcommitting a bit. If the application has
9688 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9689 * of CQ ring entries manually.
9691 p->sq_entries = roundup_pow_of_two(entries);
9692 if (p->flags & IORING_SETUP_CQSIZE) {
9694 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9695 * to a power-of-two, if it isn't already. We do NOT impose
9696 * any cq vs sq ring sizing.
9700 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9701 if (!(p->flags & IORING_SETUP_CLAMP))
9703 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9705 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9706 if (p->cq_entries < p->sq_entries)
9709 p->cq_entries = 2 * p->sq_entries;
9712 ctx = io_ring_ctx_alloc(p);
9715 ctx->compat = in_compat_syscall();
9716 if (!capable(CAP_IPC_LOCK))
9717 ctx->user = get_uid(current_user());
9720 * This is just grabbed for accounting purposes. When a process exits,
9721 * the mm is exited and dropped before the files, hence we need to hang
9722 * on to this mm purely for the purposes of being able to unaccount
9723 * memory (locked/pinned vm). It's not used for anything else.
9725 mmgrab(current->mm);
9726 ctx->mm_account = current->mm;
9728 ret = io_allocate_scq_urings(ctx, p);
9732 ret = io_sq_offload_create(ctx, p);
9735 /* always set a rsrc node */
9736 ret = io_rsrc_node_switch_start(ctx);
9739 io_rsrc_node_switch(ctx, NULL);
9741 memset(&p->sq_off, 0, sizeof(p->sq_off));
9742 p->sq_off.head = offsetof(struct io_rings, sq.head);
9743 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9744 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9745 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9746 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9747 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9748 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9750 memset(&p->cq_off, 0, sizeof(p->cq_off));
9751 p->cq_off.head = offsetof(struct io_rings, cq.head);
9752 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9753 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9754 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9755 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9756 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9757 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9759 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9760 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9761 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9762 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9763 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9764 IORING_FEAT_RSRC_TAGS;
9766 if (copy_to_user(params, p, sizeof(*p))) {
9771 file = io_uring_get_file(ctx);
9773 ret = PTR_ERR(file);
9778 * Install ring fd as the very last thing, so we don't risk someone
9779 * having closed it before we finish setup
9781 ret = io_uring_install_fd(ctx, file);
9783 /* fput will clean it up */
9788 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9791 io_ring_ctx_wait_and_kill(ctx);
9796 * Sets up an aio uring context, and returns the fd. Applications asks for a
9797 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9798 * params structure passed in.
9800 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9802 struct io_uring_params p;
9805 if (copy_from_user(&p, params, sizeof(p)))
9807 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9812 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9813 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9814 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9815 IORING_SETUP_R_DISABLED))
9818 return io_uring_create(entries, &p, params);
9821 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9822 struct io_uring_params __user *, params)
9824 return io_uring_setup(entries, params);
9827 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9829 struct io_uring_probe *p;
9833 size = struct_size(p, ops, nr_args);
9834 if (size == SIZE_MAX)
9836 p = kzalloc(size, GFP_KERNEL);
9841 if (copy_from_user(p, arg, size))
9844 if (memchr_inv(p, 0, size))
9847 p->last_op = IORING_OP_LAST - 1;
9848 if (nr_args > IORING_OP_LAST)
9849 nr_args = IORING_OP_LAST;
9851 for (i = 0; i < nr_args; i++) {
9853 if (!io_op_defs[i].not_supported)
9854 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9859 if (copy_to_user(arg, p, size))
9866 static int io_register_personality(struct io_ring_ctx *ctx)
9868 const struct cred *creds;
9872 creds = get_current_cred();
9874 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9875 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9883 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9884 unsigned int nr_args)
9886 struct io_uring_restriction *res;
9890 /* Restrictions allowed only if rings started disabled */
9891 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9894 /* We allow only a single restrictions registration */
9895 if (ctx->restrictions.registered)
9898 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9901 size = array_size(nr_args, sizeof(*res));
9902 if (size == SIZE_MAX)
9905 res = memdup_user(arg, size);
9907 return PTR_ERR(res);
9911 for (i = 0; i < nr_args; i++) {
9912 switch (res[i].opcode) {
9913 case IORING_RESTRICTION_REGISTER_OP:
9914 if (res[i].register_op >= IORING_REGISTER_LAST) {
9919 __set_bit(res[i].register_op,
9920 ctx->restrictions.register_op);
9922 case IORING_RESTRICTION_SQE_OP:
9923 if (res[i].sqe_op >= IORING_OP_LAST) {
9928 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9930 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9931 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9933 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9934 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9943 /* Reset all restrictions if an error happened */
9945 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9947 ctx->restrictions.registered = true;
9953 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9955 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9958 if (ctx->restrictions.registered)
9959 ctx->restricted = 1;
9961 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9962 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9963 wake_up(&ctx->sq_data->wait);
9967 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9968 struct io_uring_rsrc_update2 *up,
9976 if (check_add_overflow(up->offset, nr_args, &tmp))
9978 err = io_rsrc_node_switch_start(ctx);
9983 case IORING_RSRC_FILE:
9984 return __io_sqe_files_update(ctx, up, nr_args);
9985 case IORING_RSRC_BUFFER:
9986 return __io_sqe_buffers_update(ctx, up, nr_args);
9991 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9994 struct io_uring_rsrc_update2 up;
9998 memset(&up, 0, sizeof(up));
9999 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10001 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10004 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10005 unsigned size, unsigned type)
10007 struct io_uring_rsrc_update2 up;
10009 if (size != sizeof(up))
10011 if (copy_from_user(&up, arg, sizeof(up)))
10013 if (!up.nr || up.resv)
10015 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10018 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10019 unsigned int size, unsigned int type)
10021 struct io_uring_rsrc_register rr;
10023 /* keep it extendible */
10024 if (size != sizeof(rr))
10027 memset(&rr, 0, sizeof(rr));
10028 if (copy_from_user(&rr, arg, size))
10030 if (!rr.nr || rr.resv || rr.resv2)
10034 case IORING_RSRC_FILE:
10035 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10036 rr.nr, u64_to_user_ptr(rr.tags));
10037 case IORING_RSRC_BUFFER:
10038 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10039 rr.nr, u64_to_user_ptr(rr.tags));
10044 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10047 struct io_uring_task *tctx = current->io_uring;
10048 cpumask_var_t new_mask;
10051 if (!tctx || !tctx->io_wq)
10054 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10057 cpumask_clear(new_mask);
10058 if (len > cpumask_size())
10059 len = cpumask_size();
10061 if (copy_from_user(new_mask, arg, len)) {
10062 free_cpumask_var(new_mask);
10066 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10067 free_cpumask_var(new_mask);
10071 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10073 struct io_uring_task *tctx = current->io_uring;
10075 if (!tctx || !tctx->io_wq)
10078 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10081 static bool io_register_op_must_quiesce(int op)
10084 case IORING_REGISTER_BUFFERS:
10085 case IORING_UNREGISTER_BUFFERS:
10086 case IORING_REGISTER_FILES:
10087 case IORING_UNREGISTER_FILES:
10088 case IORING_REGISTER_FILES_UPDATE:
10089 case IORING_REGISTER_PROBE:
10090 case IORING_REGISTER_PERSONALITY:
10091 case IORING_UNREGISTER_PERSONALITY:
10092 case IORING_REGISTER_FILES2:
10093 case IORING_REGISTER_FILES_UPDATE2:
10094 case IORING_REGISTER_BUFFERS2:
10095 case IORING_REGISTER_BUFFERS_UPDATE:
10096 case IORING_REGISTER_IOWQ_AFF:
10097 case IORING_UNREGISTER_IOWQ_AFF:
10104 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10108 percpu_ref_kill(&ctx->refs);
10111 * Drop uring mutex before waiting for references to exit. If another
10112 * thread is currently inside io_uring_enter() it might need to grab the
10113 * uring_lock to make progress. If we hold it here across the drain
10114 * wait, then we can deadlock. It's safe to drop the mutex here, since
10115 * no new references will come in after we've killed the percpu ref.
10117 mutex_unlock(&ctx->uring_lock);
10119 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10122 ret = io_run_task_work_sig();
10123 } while (ret >= 0);
10124 mutex_lock(&ctx->uring_lock);
10127 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10131 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10132 void __user *arg, unsigned nr_args)
10133 __releases(ctx->uring_lock)
10134 __acquires(ctx->uring_lock)
10139 * We're inside the ring mutex, if the ref is already dying, then
10140 * someone else killed the ctx or is already going through
10141 * io_uring_register().
10143 if (percpu_ref_is_dying(&ctx->refs))
10146 if (ctx->restricted) {
10147 if (opcode >= IORING_REGISTER_LAST)
10149 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10150 if (!test_bit(opcode, ctx->restrictions.register_op))
10154 if (io_register_op_must_quiesce(opcode)) {
10155 ret = io_ctx_quiesce(ctx);
10161 case IORING_REGISTER_BUFFERS:
10162 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10164 case IORING_UNREGISTER_BUFFERS:
10166 if (arg || nr_args)
10168 ret = io_sqe_buffers_unregister(ctx);
10170 case IORING_REGISTER_FILES:
10171 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10173 case IORING_UNREGISTER_FILES:
10175 if (arg || nr_args)
10177 ret = io_sqe_files_unregister(ctx);
10179 case IORING_REGISTER_FILES_UPDATE:
10180 ret = io_register_files_update(ctx, arg, nr_args);
10182 case IORING_REGISTER_EVENTFD:
10183 case IORING_REGISTER_EVENTFD_ASYNC:
10187 ret = io_eventfd_register(ctx, arg);
10190 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10191 ctx->eventfd_async = 1;
10193 ctx->eventfd_async = 0;
10195 case IORING_UNREGISTER_EVENTFD:
10197 if (arg || nr_args)
10199 ret = io_eventfd_unregister(ctx);
10201 case IORING_REGISTER_PROBE:
10203 if (!arg || nr_args > 256)
10205 ret = io_probe(ctx, arg, nr_args);
10207 case IORING_REGISTER_PERSONALITY:
10209 if (arg || nr_args)
10211 ret = io_register_personality(ctx);
10213 case IORING_UNREGISTER_PERSONALITY:
10217 ret = io_unregister_personality(ctx, nr_args);
10219 case IORING_REGISTER_ENABLE_RINGS:
10221 if (arg || nr_args)
10223 ret = io_register_enable_rings(ctx);
10225 case IORING_REGISTER_RESTRICTIONS:
10226 ret = io_register_restrictions(ctx, arg, nr_args);
10228 case IORING_REGISTER_FILES2:
10229 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10231 case IORING_REGISTER_FILES_UPDATE2:
10232 ret = io_register_rsrc_update(ctx, arg, nr_args,
10235 case IORING_REGISTER_BUFFERS2:
10236 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10238 case IORING_REGISTER_BUFFERS_UPDATE:
10239 ret = io_register_rsrc_update(ctx, arg, nr_args,
10240 IORING_RSRC_BUFFER);
10242 case IORING_REGISTER_IOWQ_AFF:
10244 if (!arg || !nr_args)
10246 ret = io_register_iowq_aff(ctx, arg, nr_args);
10248 case IORING_UNREGISTER_IOWQ_AFF:
10250 if (arg || nr_args)
10252 ret = io_unregister_iowq_aff(ctx);
10259 if (io_register_op_must_quiesce(opcode)) {
10260 /* bring the ctx back to life */
10261 percpu_ref_reinit(&ctx->refs);
10262 reinit_completion(&ctx->ref_comp);
10267 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10268 void __user *, arg, unsigned int, nr_args)
10270 struct io_ring_ctx *ctx;
10279 if (f.file->f_op != &io_uring_fops)
10282 ctx = f.file->private_data;
10284 io_run_task_work();
10286 mutex_lock(&ctx->uring_lock);
10287 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10288 mutex_unlock(&ctx->uring_lock);
10289 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10290 ctx->cq_ev_fd != NULL, ret);
10296 static int __init io_uring_init(void)
10298 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10299 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10300 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10303 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10304 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10305 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10306 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10307 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10308 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10309 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10310 BUILD_BUG_SQE_ELEM(8, __u64, off);
10311 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10312 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10313 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10314 BUILD_BUG_SQE_ELEM(24, __u32, len);
10315 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10316 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10317 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10318 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10319 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10320 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10321 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10322 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10323 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10324 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10325 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10326 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10327 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10328 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10329 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10330 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10331 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10332 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10333 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10334 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10336 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10337 sizeof(struct io_uring_rsrc_update));
10338 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10339 sizeof(struct io_uring_rsrc_update2));
10340 /* should fit into one byte */
10341 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10343 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10344 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10346 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10350 __initcall(io_uring_init);