1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 /* 512 entries per page on 64-bit archs, 64 pages max */
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT 9
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head cq_overflow_list;
379 struct xarray io_buffers;
380 struct xarray personalities;
382 unsigned sq_thread_idle;
383 } ____cacheline_aligned_in_smp;
385 /* IRQ completion list, under ->completion_lock */
386 struct list_head locked_free_list;
387 unsigned int locked_free_nr;
389 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
390 struct io_sq_data *sq_data; /* if using sq thread polling */
392 struct wait_queue_head sqo_sq_wait;
393 struct list_head sqd_list;
395 unsigned long check_cq_overflow;
398 unsigned cached_cq_tail;
400 struct eventfd_ctx *cq_ev_fd;
401 struct wait_queue_head poll_wait;
402 struct wait_queue_head cq_wait;
404 atomic_t cq_timeouts;
405 struct fasync_struct *cq_fasync;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
462 struct io_uring_task {
463 /* submission side */
466 struct wait_queue_head wait;
467 const struct io_ring_ctx *last;
469 struct percpu_counter inflight;
470 atomic_t inflight_tracked;
473 spinlock_t task_lock;
474 struct io_wq_work_list task_list;
475 struct callback_head task_work;
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
483 struct io_poll_iocb {
485 struct wait_queue_head *head;
489 struct wait_queue_entry wait;
492 struct io_poll_update {
498 bool update_user_data;
506 struct io_timeout_data {
507 struct io_kiocb *req;
508 struct hrtimer timer;
509 struct timespec64 ts;
510 enum hrtimer_mode mode;
515 struct sockaddr __user *addr;
516 int __user *addr_len;
518 unsigned long nofile;
538 struct list_head list;
539 /* head of the link, used by linked timeouts only */
540 struct io_kiocb *head;
541 /* for linked completions */
542 struct io_kiocb *prev;
545 struct io_timeout_rem {
550 struct timespec64 ts;
555 /* NOTE: kiocb has the file as the first member, so don't do it here */
563 struct sockaddr __user *addr;
570 struct compat_msghdr __user *umsg_compat;
571 struct user_msghdr __user *umsg;
577 struct io_buffer *kbuf;
583 struct filename *filename;
585 unsigned long nofile;
588 struct io_rsrc_update {
614 struct epoll_event event;
618 struct file *file_out;
619 struct file *file_in;
626 struct io_provide_buf {
640 const char __user *filename;
641 struct statx __user *buffer;
653 struct filename *oldpath;
654 struct filename *newpath;
662 struct filename *filename;
665 struct io_completion {
670 struct io_async_connect {
671 struct sockaddr_storage address;
674 struct io_async_msghdr {
675 struct iovec fast_iov[UIO_FASTIOV];
676 /* points to an allocated iov, if NULL we use fast_iov instead */
677 struct iovec *free_iov;
678 struct sockaddr __user *uaddr;
680 struct sockaddr_storage addr;
684 struct iovec fast_iov[UIO_FASTIOV];
685 const struct iovec *free_iovec;
686 struct iov_iter iter;
688 struct wait_page_queue wpq;
692 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
693 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
694 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
695 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
696 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
697 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
699 /* first byte is taken by user flags, shift it to not overlap */
704 REQ_F_LINK_TIMEOUT_BIT,
705 REQ_F_NEED_CLEANUP_BIT,
707 REQ_F_BUFFER_SELECTED_BIT,
708 REQ_F_LTIMEOUT_ACTIVE_BIT,
709 REQ_F_COMPLETE_INLINE_BIT,
711 REQ_F_DONT_REISSUE_BIT,
713 /* keep async read/write and isreg together and in order */
714 REQ_F_NOWAIT_READ_BIT,
715 REQ_F_NOWAIT_WRITE_BIT,
718 /* not a real bit, just to check we're not overflowing the space */
724 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
725 /* drain existing IO first */
726 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
728 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
729 /* doesn't sever on completion < 0 */
730 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
732 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
733 /* IOSQE_BUFFER_SELECT */
734 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
736 /* fail rest of links */
737 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
738 /* on inflight list, should be cancelled and waited on exit reliably */
739 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
740 /* read/write uses file position */
741 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
742 /* must not punt to workers */
743 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
744 /* has or had linked timeout */
745 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
747 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
748 /* already went through poll handler */
749 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
750 /* buffer already selected */
751 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
752 /* linked timeout is active, i.e. prepared by link's head */
753 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
754 /* completion is deferred through io_comp_state */
755 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
756 /* caller should reissue async */
757 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
758 /* don't attempt request reissue, see io_rw_reissue() */
759 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
760 /* supports async reads */
761 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
762 /* supports async writes */
763 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
765 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
766 /* has creds assigned */
767 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
771 struct io_poll_iocb poll;
772 struct io_poll_iocb *double_poll;
775 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
777 struct io_task_work {
779 struct io_wq_work_node node;
780 struct llist_node fallback_node;
782 io_req_tw_func_t func;
786 IORING_RSRC_FILE = 0,
787 IORING_RSRC_BUFFER = 1,
791 * NOTE! Each of the iocb union members has the file pointer
792 * as the first entry in their struct definition. So you can
793 * access the file pointer through any of the sub-structs,
794 * or directly as just 'ki_filp' in this struct.
800 struct io_poll_iocb poll;
801 struct io_poll_update poll_update;
802 struct io_accept accept;
804 struct io_cancel cancel;
805 struct io_timeout timeout;
806 struct io_timeout_rem timeout_rem;
807 struct io_connect connect;
808 struct io_sr_msg sr_msg;
810 struct io_close close;
811 struct io_rsrc_update rsrc_update;
812 struct io_fadvise fadvise;
813 struct io_madvise madvise;
814 struct io_epoll epoll;
815 struct io_splice splice;
816 struct io_provide_buf pbuf;
817 struct io_statx statx;
818 struct io_shutdown shutdown;
819 struct io_rename rename;
820 struct io_unlink unlink;
821 /* use only after cleaning per-op data, see io_clean_op() */
822 struct io_completion compl;
825 /* opcode allocated if it needs to store data for async defer */
828 /* polled IO has completed */
834 struct io_ring_ctx *ctx;
837 struct task_struct *task;
840 struct io_kiocb *link;
841 struct percpu_ref *fixed_rsrc_refs;
843 /* used with ctx->iopoll_list with reads/writes */
844 struct list_head inflight_entry;
845 struct io_task_work io_task_work;
846 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
847 struct hlist_node hash_node;
848 struct async_poll *apoll;
849 struct io_wq_work work;
850 const struct cred *creds;
852 /* store used ubuf, so we can prevent reloading */
853 struct io_mapped_ubuf *imu;
856 struct io_tctx_node {
857 struct list_head ctx_node;
858 struct task_struct *task;
859 struct io_ring_ctx *ctx;
862 struct io_defer_entry {
863 struct list_head list;
864 struct io_kiocb *req;
869 /* needs req->file assigned */
870 unsigned needs_file : 1;
871 /* hash wq insertion if file is a regular file */
872 unsigned hash_reg_file : 1;
873 /* unbound wq insertion if file is a non-regular file */
874 unsigned unbound_nonreg_file : 1;
875 /* opcode is not supported by this kernel */
876 unsigned not_supported : 1;
877 /* set if opcode supports polled "wait" */
879 unsigned pollout : 1;
880 /* op supports buffer selection */
881 unsigned buffer_select : 1;
882 /* do prep async if is going to be punted */
883 unsigned needs_async_setup : 1;
884 /* should block plug */
886 /* size of async data needed, if any */
887 unsigned short async_size;
890 static const struct io_op_def io_op_defs[] = {
891 [IORING_OP_NOP] = {},
892 [IORING_OP_READV] = {
894 .unbound_nonreg_file = 1,
897 .needs_async_setup = 1,
899 .async_size = sizeof(struct io_async_rw),
901 [IORING_OP_WRITEV] = {
904 .unbound_nonreg_file = 1,
906 .needs_async_setup = 1,
908 .async_size = sizeof(struct io_async_rw),
910 [IORING_OP_FSYNC] = {
913 [IORING_OP_READ_FIXED] = {
915 .unbound_nonreg_file = 1,
918 .async_size = sizeof(struct io_async_rw),
920 [IORING_OP_WRITE_FIXED] = {
923 .unbound_nonreg_file = 1,
926 .async_size = sizeof(struct io_async_rw),
928 [IORING_OP_POLL_ADD] = {
930 .unbound_nonreg_file = 1,
932 [IORING_OP_POLL_REMOVE] = {},
933 [IORING_OP_SYNC_FILE_RANGE] = {
936 [IORING_OP_SENDMSG] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_setup = 1,
941 .async_size = sizeof(struct io_async_msghdr),
943 [IORING_OP_RECVMSG] = {
945 .unbound_nonreg_file = 1,
948 .needs_async_setup = 1,
949 .async_size = sizeof(struct io_async_msghdr),
951 [IORING_OP_TIMEOUT] = {
952 .async_size = sizeof(struct io_timeout_data),
954 [IORING_OP_TIMEOUT_REMOVE] = {
955 /* used by timeout updates' prep() */
957 [IORING_OP_ACCEPT] = {
959 .unbound_nonreg_file = 1,
962 [IORING_OP_ASYNC_CANCEL] = {},
963 [IORING_OP_LINK_TIMEOUT] = {
964 .async_size = sizeof(struct io_timeout_data),
966 [IORING_OP_CONNECT] = {
968 .unbound_nonreg_file = 1,
970 .needs_async_setup = 1,
971 .async_size = sizeof(struct io_async_connect),
973 [IORING_OP_FALLOCATE] = {
976 [IORING_OP_OPENAT] = {},
977 [IORING_OP_CLOSE] = {},
978 [IORING_OP_FILES_UPDATE] = {},
979 [IORING_OP_STATX] = {},
982 .unbound_nonreg_file = 1,
986 .async_size = sizeof(struct io_async_rw),
988 [IORING_OP_WRITE] = {
990 .unbound_nonreg_file = 1,
993 .async_size = sizeof(struct io_async_rw),
995 [IORING_OP_FADVISE] = {
998 [IORING_OP_MADVISE] = {},
1001 .unbound_nonreg_file = 1,
1004 [IORING_OP_RECV] = {
1006 .unbound_nonreg_file = 1,
1010 [IORING_OP_OPENAT2] = {
1012 [IORING_OP_EPOLL_CTL] = {
1013 .unbound_nonreg_file = 1,
1015 [IORING_OP_SPLICE] = {
1018 .unbound_nonreg_file = 1,
1020 [IORING_OP_PROVIDE_BUFFERS] = {},
1021 [IORING_OP_REMOVE_BUFFERS] = {},
1025 .unbound_nonreg_file = 1,
1027 [IORING_OP_SHUTDOWN] = {
1030 [IORING_OP_RENAMEAT] = {},
1031 [IORING_OP_UNLINKAT] = {},
1034 static bool io_disarm_next(struct io_kiocb *req);
1035 static void io_uring_del_tctx_node(unsigned long index);
1036 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1037 struct task_struct *task,
1039 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1041 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1042 long res, unsigned int cflags);
1043 static void io_put_req(struct io_kiocb *req);
1044 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1045 static void io_dismantle_req(struct io_kiocb *req);
1046 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1047 static void io_queue_linked_timeout(struct io_kiocb *req);
1048 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1049 struct io_uring_rsrc_update2 *up,
1051 static void io_clean_op(struct io_kiocb *req);
1052 static struct file *io_file_get(struct io_ring_ctx *ctx,
1053 struct io_kiocb *req, int fd, bool fixed);
1054 static void __io_queue_sqe(struct io_kiocb *req);
1055 static void io_rsrc_put_work(struct work_struct *work);
1057 static void io_req_task_queue(struct io_kiocb *req);
1058 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1059 static int io_req_prep_async(struct io_kiocb *req);
1061 static struct kmem_cache *req_cachep;
1063 static const struct file_operations io_uring_fops;
1065 struct sock *io_uring_get_socket(struct file *file)
1067 #if defined(CONFIG_UNIX)
1068 if (file->f_op == &io_uring_fops) {
1069 struct io_ring_ctx *ctx = file->private_data;
1071 return ctx->ring_sock->sk;
1076 EXPORT_SYMBOL(io_uring_get_socket);
1078 #define io_for_each_link(pos, head) \
1079 for (pos = (head); pos; pos = pos->link)
1082 * Shamelessly stolen from the mm implementation of page reference checking,
1083 * see commit f958d7b528b1 for details.
1085 #define req_ref_zero_or_close_to_overflow(req) \
1086 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1088 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1090 return atomic_inc_not_zero(&req->refs);
1093 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1095 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1096 return atomic_sub_and_test(refs, &req->refs);
1099 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1101 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1102 return atomic_dec_and_test(&req->refs);
1105 static inline void req_ref_put(struct io_kiocb *req)
1107 WARN_ON_ONCE(req_ref_put_and_test(req));
1110 static inline void req_ref_get(struct io_kiocb *req)
1112 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1113 atomic_inc(&req->refs);
1116 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1118 struct io_ring_ctx *ctx = req->ctx;
1120 if (!req->fixed_rsrc_refs) {
1121 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1122 percpu_ref_get(req->fixed_rsrc_refs);
1126 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1128 bool got = percpu_ref_tryget(ref);
1130 /* already at zero, wait for ->release() */
1132 wait_for_completion(compl);
1133 percpu_ref_resurrect(ref);
1135 percpu_ref_put(ref);
1138 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1141 struct io_kiocb *req;
1143 if (task && head->task != task)
1148 io_for_each_link(req, head) {
1149 if (req->flags & REQ_F_INFLIGHT)
1155 static inline void req_set_fail(struct io_kiocb *req)
1157 req->flags |= REQ_F_FAIL;
1160 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1162 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1164 complete(&ctx->ref_comp);
1167 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1169 return !req->timeout.off;
1172 static void io_fallback_req_func(struct work_struct *work)
1174 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1175 fallback_work.work);
1176 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1177 struct io_kiocb *req, *tmp;
1179 percpu_ref_get(&ctx->refs);
1180 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1181 req->io_task_work.func(req);
1182 percpu_ref_put(&ctx->refs);
1185 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1187 struct io_ring_ctx *ctx;
1190 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1195 * Use 5 bits less than the max cq entries, that should give us around
1196 * 32 entries per hash list if totally full and uniformly spread.
1198 hash_bits = ilog2(p->cq_entries);
1202 ctx->cancel_hash_bits = hash_bits;
1203 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1205 if (!ctx->cancel_hash)
1207 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1209 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1210 if (!ctx->dummy_ubuf)
1212 /* set invalid range, so io_import_fixed() fails meeting it */
1213 ctx->dummy_ubuf->ubuf = -1UL;
1215 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1216 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1219 ctx->flags = p->flags;
1220 init_waitqueue_head(&ctx->sqo_sq_wait);
1221 INIT_LIST_HEAD(&ctx->sqd_list);
1222 init_waitqueue_head(&ctx->poll_wait);
1223 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1224 init_completion(&ctx->ref_comp);
1225 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1226 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1227 mutex_init(&ctx->uring_lock);
1228 init_waitqueue_head(&ctx->cq_wait);
1229 spin_lock_init(&ctx->completion_lock);
1230 spin_lock_init(&ctx->timeout_lock);
1231 INIT_LIST_HEAD(&ctx->iopoll_list);
1232 INIT_LIST_HEAD(&ctx->defer_list);
1233 INIT_LIST_HEAD(&ctx->timeout_list);
1234 spin_lock_init(&ctx->rsrc_ref_lock);
1235 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1236 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1237 init_llist_head(&ctx->rsrc_put_llist);
1238 INIT_LIST_HEAD(&ctx->tctx_list);
1239 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1240 INIT_LIST_HEAD(&ctx->locked_free_list);
1241 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1244 kfree(ctx->dummy_ubuf);
1245 kfree(ctx->cancel_hash);
1250 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1252 struct io_rings *r = ctx->rings;
1254 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1258 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1260 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1261 struct io_ring_ctx *ctx = req->ctx;
1263 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1269 #define FFS_ASYNC_READ 0x1UL
1270 #define FFS_ASYNC_WRITE 0x2UL
1272 #define FFS_ISREG 0x4UL
1274 #define FFS_ISREG 0x0UL
1276 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1278 static inline bool io_req_ffs_set(struct io_kiocb *req)
1280 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1283 static void io_req_track_inflight(struct io_kiocb *req)
1285 if (!(req->flags & REQ_F_INFLIGHT)) {
1286 req->flags |= REQ_F_INFLIGHT;
1287 atomic_inc(¤t->io_uring->inflight_tracked);
1291 static void io_prep_async_work(struct io_kiocb *req)
1293 const struct io_op_def *def = &io_op_defs[req->opcode];
1294 struct io_ring_ctx *ctx = req->ctx;
1296 if (!(req->flags & REQ_F_CREDS)) {
1297 req->flags |= REQ_F_CREDS;
1298 req->creds = get_current_cred();
1301 req->work.list.next = NULL;
1302 req->work.flags = 0;
1303 if (req->flags & REQ_F_FORCE_ASYNC)
1304 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1306 if (req->flags & REQ_F_ISREG) {
1307 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1308 io_wq_hash_work(&req->work, file_inode(req->file));
1309 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1310 if (def->unbound_nonreg_file)
1311 req->work.flags |= IO_WQ_WORK_UNBOUND;
1314 switch (req->opcode) {
1315 case IORING_OP_SPLICE:
1317 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1318 req->work.flags |= IO_WQ_WORK_UNBOUND;
1323 static void io_prep_async_link(struct io_kiocb *req)
1325 struct io_kiocb *cur;
1327 if (req->flags & REQ_F_LINK_TIMEOUT) {
1328 struct io_ring_ctx *ctx = req->ctx;
1330 spin_lock(&ctx->completion_lock);
1331 io_for_each_link(cur, req)
1332 io_prep_async_work(cur);
1333 spin_unlock(&ctx->completion_lock);
1335 io_for_each_link(cur, req)
1336 io_prep_async_work(cur);
1340 static void io_queue_async_work(struct io_kiocb *req)
1342 struct io_ring_ctx *ctx = req->ctx;
1343 struct io_kiocb *link = io_prep_linked_timeout(req);
1344 struct io_uring_task *tctx = req->task->io_uring;
1347 BUG_ON(!tctx->io_wq);
1349 /* init ->work of the whole link before punting */
1350 io_prep_async_link(req);
1353 * Not expected to happen, but if we do have a bug where this _can_
1354 * happen, catch it here and ensure the request is marked as
1355 * canceled. That will make io-wq go through the usual work cancel
1356 * procedure rather than attempt to run this request (or create a new
1359 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1360 req->work.flags |= IO_WQ_WORK_CANCEL;
1362 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1363 &req->work, req->flags);
1364 io_wq_enqueue(tctx->io_wq, &req->work);
1366 io_queue_linked_timeout(link);
1369 static void io_kill_timeout(struct io_kiocb *req, int status)
1370 __must_hold(&req->ctx->completion_lock)
1371 __must_hold(&req->ctx->timeout_lock)
1373 struct io_timeout_data *io = req->async_data;
1375 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1376 atomic_set(&req->ctx->cq_timeouts,
1377 atomic_read(&req->ctx->cq_timeouts) + 1);
1378 list_del_init(&req->timeout.list);
1379 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1380 io_put_req_deferred(req, 1);
1384 static void io_queue_deferred(struct io_ring_ctx *ctx)
1386 while (!list_empty(&ctx->defer_list)) {
1387 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1388 struct io_defer_entry, list);
1390 if (req_need_defer(de->req, de->seq))
1392 list_del_init(&de->list);
1393 io_req_task_queue(de->req);
1398 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1399 __must_hold(&ctx->completion_lock)
1401 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1403 spin_lock_irq(&ctx->timeout_lock);
1404 while (!list_empty(&ctx->timeout_list)) {
1405 u32 events_needed, events_got;
1406 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1407 struct io_kiocb, timeout.list);
1409 if (io_is_timeout_noseq(req))
1413 * Since seq can easily wrap around over time, subtract
1414 * the last seq at which timeouts were flushed before comparing.
1415 * Assuming not more than 2^31-1 events have happened since,
1416 * these subtractions won't have wrapped, so we can check if
1417 * target is in [last_seq, current_seq] by comparing the two.
1419 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1420 events_got = seq - ctx->cq_last_tm_flush;
1421 if (events_got < events_needed)
1424 list_del_init(&req->timeout.list);
1425 io_kill_timeout(req, 0);
1427 ctx->cq_last_tm_flush = seq;
1428 spin_unlock_irq(&ctx->timeout_lock);
1431 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1433 if (ctx->off_timeout_used)
1434 io_flush_timeouts(ctx);
1435 if (ctx->drain_active)
1436 io_queue_deferred(ctx);
1439 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1441 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1442 __io_commit_cqring_flush(ctx);
1443 /* order cqe stores with ring update */
1444 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1447 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1449 struct io_rings *r = ctx->rings;
1451 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1454 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1456 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1459 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1461 struct io_rings *rings = ctx->rings;
1462 unsigned tail, mask = ctx->cq_entries - 1;
1465 * writes to the cq entry need to come after reading head; the
1466 * control dependency is enough as we're using WRITE_ONCE to
1469 if (__io_cqring_events(ctx) == ctx->cq_entries)
1472 tail = ctx->cached_cq_tail++;
1473 return &rings->cqes[tail & mask];
1476 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1478 if (likely(!ctx->cq_ev_fd))
1480 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1482 return !ctx->eventfd_async || io_wq_current_is_worker();
1485 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1488 * wake_up_all() may seem excessive, but io_wake_function() and
1489 * io_should_wake() handle the termination of the loop and only
1490 * wake as many waiters as we need to.
1492 if (wq_has_sleeper(&ctx->cq_wait))
1493 wake_up_all(&ctx->cq_wait);
1494 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1495 wake_up(&ctx->sq_data->wait);
1496 if (io_should_trigger_evfd(ctx))
1497 eventfd_signal(ctx->cq_ev_fd, 1);
1498 if (waitqueue_active(&ctx->poll_wait)) {
1499 wake_up_interruptible(&ctx->poll_wait);
1500 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1504 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1506 if (ctx->flags & IORING_SETUP_SQPOLL) {
1507 if (wq_has_sleeper(&ctx->cq_wait))
1508 wake_up_all(&ctx->cq_wait);
1510 if (io_should_trigger_evfd(ctx))
1511 eventfd_signal(ctx->cq_ev_fd, 1);
1512 if (waitqueue_active(&ctx->poll_wait)) {
1513 wake_up_interruptible(&ctx->poll_wait);
1514 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1518 /* Returns true if there are no backlogged entries after the flush */
1519 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1521 bool all_flushed, posted;
1523 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1527 spin_lock(&ctx->completion_lock);
1528 while (!list_empty(&ctx->cq_overflow_list)) {
1529 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1530 struct io_overflow_cqe *ocqe;
1534 ocqe = list_first_entry(&ctx->cq_overflow_list,
1535 struct io_overflow_cqe, list);
1537 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1539 io_account_cq_overflow(ctx);
1542 list_del(&ocqe->list);
1546 all_flushed = list_empty(&ctx->cq_overflow_list);
1548 clear_bit(0, &ctx->check_cq_overflow);
1549 WRITE_ONCE(ctx->rings->sq_flags,
1550 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1554 io_commit_cqring(ctx);
1555 spin_unlock(&ctx->completion_lock);
1557 io_cqring_ev_posted(ctx);
1561 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1565 if (test_bit(0, &ctx->check_cq_overflow)) {
1566 /* iopoll syncs against uring_lock, not completion_lock */
1567 if (ctx->flags & IORING_SETUP_IOPOLL)
1568 mutex_lock(&ctx->uring_lock);
1569 ret = __io_cqring_overflow_flush(ctx, false);
1570 if (ctx->flags & IORING_SETUP_IOPOLL)
1571 mutex_unlock(&ctx->uring_lock);
1577 /* must to be called somewhat shortly after putting a request */
1578 static inline void io_put_task(struct task_struct *task, int nr)
1580 struct io_uring_task *tctx = task->io_uring;
1582 percpu_counter_sub(&tctx->inflight, nr);
1583 if (unlikely(atomic_read(&tctx->in_idle)))
1584 wake_up(&tctx->wait);
1585 put_task_struct_many(task, nr);
1588 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1589 long res, unsigned int cflags)
1591 struct io_overflow_cqe *ocqe;
1593 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1596 * If we're in ring overflow flush mode, or in task cancel mode,
1597 * or cannot allocate an overflow entry, then we need to drop it
1600 io_account_cq_overflow(ctx);
1603 if (list_empty(&ctx->cq_overflow_list)) {
1604 set_bit(0, &ctx->check_cq_overflow);
1605 WRITE_ONCE(ctx->rings->sq_flags,
1606 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1609 ocqe->cqe.user_data = user_data;
1610 ocqe->cqe.res = res;
1611 ocqe->cqe.flags = cflags;
1612 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1616 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1617 long res, unsigned int cflags)
1619 struct io_uring_cqe *cqe;
1621 trace_io_uring_complete(ctx, user_data, res, cflags);
1624 * If we can't get a cq entry, userspace overflowed the
1625 * submission (by quite a lot). Increment the overflow count in
1628 cqe = io_get_cqe(ctx);
1630 WRITE_ONCE(cqe->user_data, user_data);
1631 WRITE_ONCE(cqe->res, res);
1632 WRITE_ONCE(cqe->flags, cflags);
1635 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1638 /* not as hot to bloat with inlining */
1639 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1640 long res, unsigned int cflags)
1642 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1645 static void io_req_complete_post(struct io_kiocb *req, long res,
1646 unsigned int cflags)
1648 struct io_ring_ctx *ctx = req->ctx;
1650 spin_lock(&ctx->completion_lock);
1651 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1653 * If we're the last reference to this request, add to our locked
1656 if (req_ref_put_and_test(req)) {
1657 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1658 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1659 io_disarm_next(req);
1661 io_req_task_queue(req->link);
1665 io_dismantle_req(req);
1666 io_put_task(req->task, 1);
1667 list_add(&req->inflight_entry, &ctx->locked_free_list);
1668 ctx->locked_free_nr++;
1670 if (!percpu_ref_tryget(&ctx->refs))
1673 io_commit_cqring(ctx);
1674 spin_unlock(&ctx->completion_lock);
1677 io_cqring_ev_posted(ctx);
1678 percpu_ref_put(&ctx->refs);
1682 static inline bool io_req_needs_clean(struct io_kiocb *req)
1684 return req->flags & IO_REQ_CLEAN_FLAGS;
1687 static void io_req_complete_state(struct io_kiocb *req, long res,
1688 unsigned int cflags)
1690 if (io_req_needs_clean(req))
1693 req->compl.cflags = cflags;
1694 req->flags |= REQ_F_COMPLETE_INLINE;
1697 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1698 long res, unsigned cflags)
1700 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1701 io_req_complete_state(req, res, cflags);
1703 io_req_complete_post(req, res, cflags);
1706 static inline void io_req_complete(struct io_kiocb *req, long res)
1708 __io_req_complete(req, 0, res, 0);
1711 static void io_req_complete_failed(struct io_kiocb *req, long res)
1715 io_req_complete_post(req, res, 0);
1719 * Don't initialise the fields below on every allocation, but do that in
1720 * advance and keep them valid across allocations.
1722 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1726 req->async_data = NULL;
1727 /* not necessary, but safer to zero */
1731 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1732 struct io_submit_state *state)
1734 spin_lock(&ctx->completion_lock);
1735 list_splice_init(&ctx->locked_free_list, &state->free_list);
1736 ctx->locked_free_nr = 0;
1737 spin_unlock(&ctx->completion_lock);
1740 /* Returns true IFF there are requests in the cache */
1741 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1743 struct io_submit_state *state = &ctx->submit_state;
1747 * If we have more than a batch's worth of requests in our IRQ side
1748 * locked cache, grab the lock and move them over to our submission
1751 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1752 io_flush_cached_locked_reqs(ctx, state);
1754 nr = state->free_reqs;
1755 while (!list_empty(&state->free_list)) {
1756 struct io_kiocb *req = list_first_entry(&state->free_list,
1757 struct io_kiocb, inflight_entry);
1759 list_del(&req->inflight_entry);
1760 state->reqs[nr++] = req;
1761 if (nr == ARRAY_SIZE(state->reqs))
1765 state->free_reqs = nr;
1769 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1771 struct io_submit_state *state = &ctx->submit_state;
1772 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1775 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1777 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1780 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1784 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1785 * retry single alloc to be on the safe side.
1787 if (unlikely(ret <= 0)) {
1788 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1789 if (!state->reqs[0])
1794 for (i = 0; i < ret; i++)
1795 io_preinit_req(state->reqs[i], ctx);
1796 state->free_reqs = ret;
1799 return state->reqs[state->free_reqs];
1802 static inline void io_put_file(struct file *file)
1808 static void io_dismantle_req(struct io_kiocb *req)
1810 unsigned int flags = req->flags;
1812 if (io_req_needs_clean(req))
1814 if (!(flags & REQ_F_FIXED_FILE))
1815 io_put_file(req->file);
1816 if (req->fixed_rsrc_refs)
1817 percpu_ref_put(req->fixed_rsrc_refs);
1818 if (req->async_data) {
1819 kfree(req->async_data);
1820 req->async_data = NULL;
1824 static void __io_free_req(struct io_kiocb *req)
1826 struct io_ring_ctx *ctx = req->ctx;
1828 io_dismantle_req(req);
1829 io_put_task(req->task, 1);
1831 spin_lock(&ctx->completion_lock);
1832 list_add(&req->inflight_entry, &ctx->locked_free_list);
1833 ctx->locked_free_nr++;
1834 spin_unlock(&ctx->completion_lock);
1836 percpu_ref_put(&ctx->refs);
1839 static inline void io_remove_next_linked(struct io_kiocb *req)
1841 struct io_kiocb *nxt = req->link;
1843 req->link = nxt->link;
1847 static bool io_kill_linked_timeout(struct io_kiocb *req)
1848 __must_hold(&req->ctx->completion_lock)
1849 __must_hold(&req->ctx->timeout_lock)
1851 struct io_kiocb *link = req->link;
1854 * Can happen if a linked timeout fired and link had been like
1855 * req -> link t-out -> link t-out [-> ...]
1857 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1858 struct io_timeout_data *io = link->async_data;
1860 io_remove_next_linked(req);
1861 link->timeout.head = NULL;
1862 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1863 io_cqring_fill_event(link->ctx, link->user_data,
1865 io_put_req_deferred(link, 1);
1872 static void io_fail_links(struct io_kiocb *req)
1873 __must_hold(&req->ctx->completion_lock)
1875 struct io_kiocb *nxt, *link = req->link;
1882 trace_io_uring_fail_link(req, link);
1883 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1884 io_put_req_deferred(link, 2);
1889 static bool io_disarm_next(struct io_kiocb *req)
1890 __must_hold(&req->ctx->completion_lock)
1892 bool posted = false;
1894 if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1895 struct io_ring_ctx *ctx = req->ctx;
1897 spin_lock_irq(&ctx->timeout_lock);
1898 posted = io_kill_linked_timeout(req);
1899 spin_unlock_irq(&ctx->timeout_lock);
1901 if (unlikely((req->flags & REQ_F_FAIL) &&
1902 !(req->flags & REQ_F_HARDLINK))) {
1903 posted |= (req->link != NULL);
1909 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1911 struct io_kiocb *nxt;
1914 * If LINK is set, we have dependent requests in this chain. If we
1915 * didn't fail this request, queue the first one up, moving any other
1916 * dependencies to the next request. In case of failure, fail the rest
1919 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1920 struct io_ring_ctx *ctx = req->ctx;
1923 spin_lock(&ctx->completion_lock);
1924 posted = io_disarm_next(req);
1926 io_commit_cqring(req->ctx);
1927 spin_unlock(&ctx->completion_lock);
1929 io_cqring_ev_posted(ctx);
1936 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1938 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1940 return __io_req_find_next(req);
1943 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1947 if (ctx->submit_state.compl_nr) {
1948 mutex_lock(&ctx->uring_lock);
1949 io_submit_flush_completions(ctx);
1950 mutex_unlock(&ctx->uring_lock);
1952 percpu_ref_put(&ctx->refs);
1955 static void tctx_task_work(struct callback_head *cb)
1957 struct io_ring_ctx *ctx = NULL;
1958 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1962 struct io_wq_work_node *node;
1964 spin_lock_irq(&tctx->task_lock);
1965 node = tctx->task_list.first;
1966 INIT_WQ_LIST(&tctx->task_list);
1968 tctx->task_running = false;
1969 spin_unlock_irq(&tctx->task_lock);
1974 struct io_wq_work_node *next = node->next;
1975 struct io_kiocb *req = container_of(node, struct io_kiocb,
1978 if (req->ctx != ctx) {
1979 ctx_flush_and_put(ctx);
1981 percpu_ref_get(&ctx->refs);
1983 req->io_task_work.func(req);
1990 ctx_flush_and_put(ctx);
1993 static void io_req_task_work_add(struct io_kiocb *req)
1995 struct task_struct *tsk = req->task;
1996 struct io_uring_task *tctx = tsk->io_uring;
1997 enum task_work_notify_mode notify;
1998 struct io_wq_work_node *node;
1999 unsigned long flags;
2002 WARN_ON_ONCE(!tctx);
2004 spin_lock_irqsave(&tctx->task_lock, flags);
2005 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2006 running = tctx->task_running;
2008 tctx->task_running = true;
2009 spin_unlock_irqrestore(&tctx->task_lock, flags);
2011 /* task_work already pending, we're done */
2016 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2017 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2018 * processing task_work. There's no reliable way to tell if TWA_RESUME
2021 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2022 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2023 wake_up_process(tsk);
2027 spin_lock_irqsave(&tctx->task_lock, flags);
2028 tctx->task_running = false;
2029 node = tctx->task_list.first;
2030 INIT_WQ_LIST(&tctx->task_list);
2031 spin_unlock_irqrestore(&tctx->task_lock, flags);
2034 req = container_of(node, struct io_kiocb, io_task_work.node);
2036 if (llist_add(&req->io_task_work.fallback_node,
2037 &req->ctx->fallback_llist))
2038 schedule_delayed_work(&req->ctx->fallback_work, 1);
2042 static void io_req_task_cancel(struct io_kiocb *req)
2044 struct io_ring_ctx *ctx = req->ctx;
2046 /* ctx is guaranteed to stay alive while we hold uring_lock */
2047 mutex_lock(&ctx->uring_lock);
2048 io_req_complete_failed(req, req->result);
2049 mutex_unlock(&ctx->uring_lock);
2052 static void io_req_task_submit(struct io_kiocb *req)
2054 struct io_ring_ctx *ctx = req->ctx;
2056 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2057 mutex_lock(&ctx->uring_lock);
2058 if (likely(!(req->task->flags & PF_EXITING)))
2059 __io_queue_sqe(req);
2061 io_req_complete_failed(req, -EFAULT);
2062 mutex_unlock(&ctx->uring_lock);
2065 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2068 req->io_task_work.func = io_req_task_cancel;
2069 io_req_task_work_add(req);
2072 static void io_req_task_queue(struct io_kiocb *req)
2074 req->io_task_work.func = io_req_task_submit;
2075 io_req_task_work_add(req);
2078 static void io_req_task_queue_reissue(struct io_kiocb *req)
2080 req->io_task_work.func = io_queue_async_work;
2081 io_req_task_work_add(req);
2084 static inline void io_queue_next(struct io_kiocb *req)
2086 struct io_kiocb *nxt = io_req_find_next(req);
2089 io_req_task_queue(nxt);
2092 static void io_free_req(struct io_kiocb *req)
2099 struct task_struct *task;
2104 static inline void io_init_req_batch(struct req_batch *rb)
2111 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2112 struct req_batch *rb)
2115 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2116 if (rb->task == current)
2117 current->io_uring->cached_refs += rb->task_refs;
2119 io_put_task(rb->task, rb->task_refs);
2122 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2123 struct io_submit_state *state)
2126 io_dismantle_req(req);
2128 if (req->task != rb->task) {
2130 io_put_task(rb->task, rb->task_refs);
2131 rb->task = req->task;
2137 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2138 state->reqs[state->free_reqs++] = req;
2140 list_add(&req->inflight_entry, &state->free_list);
2143 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2144 __must_hold(&req->ctx->uring_lock)
2146 struct io_submit_state *state = &ctx->submit_state;
2147 int i, nr = state->compl_nr;
2148 struct req_batch rb;
2150 spin_lock(&ctx->completion_lock);
2151 for (i = 0; i < nr; i++) {
2152 struct io_kiocb *req = state->compl_reqs[i];
2154 __io_cqring_fill_event(ctx, req->user_data, req->result,
2157 io_commit_cqring(ctx);
2158 spin_unlock(&ctx->completion_lock);
2159 io_cqring_ev_posted(ctx);
2161 io_init_req_batch(&rb);
2162 for (i = 0; i < nr; i++) {
2163 struct io_kiocb *req = state->compl_reqs[i];
2165 /* submission and completion refs */
2166 if (req_ref_sub_and_test(req, 2))
2167 io_req_free_batch(&rb, req, &ctx->submit_state);
2170 io_req_free_batch_finish(ctx, &rb);
2171 state->compl_nr = 0;
2175 * Drop reference to request, return next in chain (if there is one) if this
2176 * was the last reference to this request.
2178 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2180 struct io_kiocb *nxt = NULL;
2182 if (req_ref_put_and_test(req)) {
2183 nxt = io_req_find_next(req);
2189 static inline void io_put_req(struct io_kiocb *req)
2191 if (req_ref_put_and_test(req))
2195 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2197 if (req_ref_sub_and_test(req, refs)) {
2198 req->io_task_work.func = io_free_req;
2199 io_req_task_work_add(req);
2203 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2205 /* See comment at the top of this file */
2207 return __io_cqring_events(ctx);
2210 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2212 struct io_rings *rings = ctx->rings;
2214 /* make sure SQ entry isn't read before tail */
2215 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2218 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2220 unsigned int cflags;
2222 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2223 cflags |= IORING_CQE_F_BUFFER;
2224 req->flags &= ~REQ_F_BUFFER_SELECTED;
2229 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2231 struct io_buffer *kbuf;
2233 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2234 return io_put_kbuf(req, kbuf);
2237 static inline bool io_run_task_work(void)
2239 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2240 __set_current_state(TASK_RUNNING);
2241 tracehook_notify_signal();
2249 * Find and free completed poll iocbs
2251 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2252 struct list_head *done, bool resubmit)
2254 struct req_batch rb;
2255 struct io_kiocb *req;
2257 /* order with ->result store in io_complete_rw_iopoll() */
2260 io_init_req_batch(&rb);
2261 while (!list_empty(done)) {
2264 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2265 list_del(&req->inflight_entry);
2267 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2268 !(req->flags & REQ_F_DONT_REISSUE)) {
2269 req->iopoll_completed = 0;
2271 io_req_task_queue_reissue(req);
2275 if (req->flags & REQ_F_BUFFER_SELECTED)
2276 cflags = io_put_rw_kbuf(req);
2278 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2281 if (req_ref_put_and_test(req))
2282 io_req_free_batch(&rb, req, &ctx->submit_state);
2285 io_commit_cqring(ctx);
2286 io_cqring_ev_posted_iopoll(ctx);
2287 io_req_free_batch_finish(ctx, &rb);
2290 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2291 long min, bool resubmit)
2293 struct io_kiocb *req, *tmp;
2298 * Only spin for completions if we don't have multiple devices hanging
2299 * off our complete list, and we're under the requested amount.
2301 spin = !ctx->poll_multi_queue && *nr_events < min;
2303 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2304 struct kiocb *kiocb = &req->rw.kiocb;
2308 * Move completed and retryable entries to our local lists.
2309 * If we find a request that requires polling, break out
2310 * and complete those lists first, if we have entries there.
2312 if (READ_ONCE(req->iopoll_completed)) {
2313 list_move_tail(&req->inflight_entry, &done);
2316 if (!list_empty(&done))
2319 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2320 if (unlikely(ret < 0))
2325 /* iopoll may have completed current req */
2326 if (READ_ONCE(req->iopoll_completed))
2327 list_move_tail(&req->inflight_entry, &done);
2330 if (!list_empty(&done))
2331 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2337 * We can't just wait for polled events to come to us, we have to actively
2338 * find and complete them.
2340 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2342 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2345 mutex_lock(&ctx->uring_lock);
2346 while (!list_empty(&ctx->iopoll_list)) {
2347 unsigned int nr_events = 0;
2349 io_do_iopoll(ctx, &nr_events, 0, false);
2351 /* let it sleep and repeat later if can't complete a request */
2355 * Ensure we allow local-to-the-cpu processing to take place,
2356 * in this case we need to ensure that we reap all events.
2357 * Also let task_work, etc. to progress by releasing the mutex
2359 if (need_resched()) {
2360 mutex_unlock(&ctx->uring_lock);
2362 mutex_lock(&ctx->uring_lock);
2365 mutex_unlock(&ctx->uring_lock);
2368 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2370 unsigned int nr_events = 0;
2374 * We disallow the app entering submit/complete with polling, but we
2375 * still need to lock the ring to prevent racing with polled issue
2376 * that got punted to a workqueue.
2378 mutex_lock(&ctx->uring_lock);
2380 * Don't enter poll loop if we already have events pending.
2381 * If we do, we can potentially be spinning for commands that
2382 * already triggered a CQE (eg in error).
2384 if (test_bit(0, &ctx->check_cq_overflow))
2385 __io_cqring_overflow_flush(ctx, false);
2386 if (io_cqring_events(ctx))
2390 * If a submit got punted to a workqueue, we can have the
2391 * application entering polling for a command before it gets
2392 * issued. That app will hold the uring_lock for the duration
2393 * of the poll right here, so we need to take a breather every
2394 * now and then to ensure that the issue has a chance to add
2395 * the poll to the issued list. Otherwise we can spin here
2396 * forever, while the workqueue is stuck trying to acquire the
2399 if (list_empty(&ctx->iopoll_list)) {
2400 u32 tail = ctx->cached_cq_tail;
2402 mutex_unlock(&ctx->uring_lock);
2404 mutex_lock(&ctx->uring_lock);
2406 /* some requests don't go through iopoll_list */
2407 if (tail != ctx->cached_cq_tail ||
2408 list_empty(&ctx->iopoll_list))
2411 ret = io_do_iopoll(ctx, &nr_events, min, true);
2412 } while (!ret && nr_events < min && !need_resched());
2414 mutex_unlock(&ctx->uring_lock);
2418 static void kiocb_end_write(struct io_kiocb *req)
2421 * Tell lockdep we inherited freeze protection from submission
2424 if (req->flags & REQ_F_ISREG) {
2425 struct super_block *sb = file_inode(req->file)->i_sb;
2427 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2433 static bool io_resubmit_prep(struct io_kiocb *req)
2435 struct io_async_rw *rw = req->async_data;
2438 return !io_req_prep_async(req);
2439 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2440 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2444 static bool io_rw_should_reissue(struct io_kiocb *req)
2446 umode_t mode = file_inode(req->file)->i_mode;
2447 struct io_ring_ctx *ctx = req->ctx;
2449 if (!S_ISBLK(mode) && !S_ISREG(mode))
2451 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2452 !(ctx->flags & IORING_SETUP_IOPOLL)))
2455 * If ref is dying, we might be running poll reap from the exit work.
2456 * Don't attempt to reissue from that path, just let it fail with
2459 if (percpu_ref_is_dying(&ctx->refs))
2462 * Play it safe and assume not safe to re-import and reissue if we're
2463 * not in the original thread group (or in task context).
2465 if (!same_thread_group(req->task, current) || !in_task())
2470 static bool io_resubmit_prep(struct io_kiocb *req)
2474 static bool io_rw_should_reissue(struct io_kiocb *req)
2480 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2482 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2483 kiocb_end_write(req);
2484 if (res != req->result) {
2485 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2486 io_rw_should_reissue(req)) {
2487 req->flags |= REQ_F_REISSUE;
2496 static void io_req_task_complete(struct io_kiocb *req)
2500 if (req->flags & REQ_F_BUFFER_SELECTED)
2501 cflags = io_put_rw_kbuf(req);
2502 __io_req_complete(req, 0, req->result, cflags);
2505 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2506 unsigned int issue_flags)
2508 if (__io_complete_rw_common(req, res))
2510 io_req_task_complete(req);
2513 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2515 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2517 if (__io_complete_rw_common(req, res))
2520 req->io_task_work.func = io_req_task_complete;
2521 io_req_task_work_add(req);
2524 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2526 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2528 if (kiocb->ki_flags & IOCB_WRITE)
2529 kiocb_end_write(req);
2530 if (unlikely(res != req->result)) {
2531 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2532 io_resubmit_prep(req))) {
2534 req->flags |= REQ_F_DONT_REISSUE;
2538 WRITE_ONCE(req->result, res);
2539 /* order with io_iopoll_complete() checking ->result */
2541 WRITE_ONCE(req->iopoll_completed, 1);
2545 * After the iocb has been issued, it's safe to be found on the poll list.
2546 * Adding the kiocb to the list AFTER submission ensures that we don't
2547 * find it from a io_do_iopoll() thread before the issuer is done
2548 * accessing the kiocb cookie.
2550 static void io_iopoll_req_issued(struct io_kiocb *req)
2552 struct io_ring_ctx *ctx = req->ctx;
2553 const bool in_async = io_wq_current_is_worker();
2555 /* workqueue context doesn't hold uring_lock, grab it now */
2556 if (unlikely(in_async))
2557 mutex_lock(&ctx->uring_lock);
2560 * Track whether we have multiple files in our lists. This will impact
2561 * how we do polling eventually, not spinning if we're on potentially
2562 * different devices.
2564 if (list_empty(&ctx->iopoll_list)) {
2565 ctx->poll_multi_queue = false;
2566 } else if (!ctx->poll_multi_queue) {
2567 struct io_kiocb *list_req;
2568 unsigned int queue_num0, queue_num1;
2570 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2573 if (list_req->file != req->file) {
2574 ctx->poll_multi_queue = true;
2576 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2577 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2578 if (queue_num0 != queue_num1)
2579 ctx->poll_multi_queue = true;
2584 * For fast devices, IO may have already completed. If it has, add
2585 * it to the front so we find it first.
2587 if (READ_ONCE(req->iopoll_completed))
2588 list_add(&req->inflight_entry, &ctx->iopoll_list);
2590 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2592 if (unlikely(in_async)) {
2594 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2595 * in sq thread task context or in io worker task context. If
2596 * current task context is sq thread, we don't need to check
2597 * whether should wake up sq thread.
2599 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2600 wq_has_sleeper(&ctx->sq_data->wait))
2601 wake_up(&ctx->sq_data->wait);
2603 mutex_unlock(&ctx->uring_lock);
2607 static bool io_bdev_nowait(struct block_device *bdev)
2609 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2613 * If we tracked the file through the SCM inflight mechanism, we could support
2614 * any file. For now, just ensure that anything potentially problematic is done
2617 static bool __io_file_supports_nowait(struct file *file, int rw)
2619 umode_t mode = file_inode(file)->i_mode;
2621 if (S_ISBLK(mode)) {
2622 if (IS_ENABLED(CONFIG_BLOCK) &&
2623 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2629 if (S_ISREG(mode)) {
2630 if (IS_ENABLED(CONFIG_BLOCK) &&
2631 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2632 file->f_op != &io_uring_fops)
2637 /* any ->read/write should understand O_NONBLOCK */
2638 if (file->f_flags & O_NONBLOCK)
2641 if (!(file->f_mode & FMODE_NOWAIT))
2645 return file->f_op->read_iter != NULL;
2647 return file->f_op->write_iter != NULL;
2650 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2652 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2654 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2657 return __io_file_supports_nowait(req->file, rw);
2660 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2662 struct io_ring_ctx *ctx = req->ctx;
2663 struct kiocb *kiocb = &req->rw.kiocb;
2664 struct file *file = req->file;
2668 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2669 req->flags |= REQ_F_ISREG;
2671 kiocb->ki_pos = READ_ONCE(sqe->off);
2672 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2673 req->flags |= REQ_F_CUR_POS;
2674 kiocb->ki_pos = file->f_pos;
2676 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2677 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2678 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2682 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2683 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2684 req->flags |= REQ_F_NOWAIT;
2686 ioprio = READ_ONCE(sqe->ioprio);
2688 ret = ioprio_check_cap(ioprio);
2692 kiocb->ki_ioprio = ioprio;
2694 kiocb->ki_ioprio = get_current_ioprio();
2696 if (ctx->flags & IORING_SETUP_IOPOLL) {
2697 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2698 !kiocb->ki_filp->f_op->iopoll)
2701 kiocb->ki_flags |= IOCB_HIPRI;
2702 kiocb->ki_complete = io_complete_rw_iopoll;
2703 req->iopoll_completed = 0;
2705 if (kiocb->ki_flags & IOCB_HIPRI)
2707 kiocb->ki_complete = io_complete_rw;
2710 if (req->opcode == IORING_OP_READ_FIXED ||
2711 req->opcode == IORING_OP_WRITE_FIXED) {
2713 io_req_set_rsrc_node(req);
2716 req->rw.addr = READ_ONCE(sqe->addr);
2717 req->rw.len = READ_ONCE(sqe->len);
2718 req->buf_index = READ_ONCE(sqe->buf_index);
2722 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2728 case -ERESTARTNOINTR:
2729 case -ERESTARTNOHAND:
2730 case -ERESTART_RESTARTBLOCK:
2732 * We can't just restart the syscall, since previously
2733 * submitted sqes may already be in progress. Just fail this
2739 kiocb->ki_complete(kiocb, ret, 0);
2743 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2744 unsigned int issue_flags)
2746 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2747 struct io_async_rw *io = req->async_data;
2748 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2750 /* add previously done IO, if any */
2751 if (io && io->bytes_done > 0) {
2753 ret = io->bytes_done;
2755 ret += io->bytes_done;
2758 if (req->flags & REQ_F_CUR_POS)
2759 req->file->f_pos = kiocb->ki_pos;
2760 if (ret >= 0 && check_reissue)
2761 __io_complete_rw(req, ret, 0, issue_flags);
2763 io_rw_done(kiocb, ret);
2765 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2766 req->flags &= ~REQ_F_REISSUE;
2767 if (io_resubmit_prep(req)) {
2769 io_req_task_queue_reissue(req);
2774 if (req->flags & REQ_F_BUFFER_SELECTED)
2775 cflags = io_put_rw_kbuf(req);
2776 __io_req_complete(req, issue_flags, ret, cflags);
2781 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2782 struct io_mapped_ubuf *imu)
2784 size_t len = req->rw.len;
2785 u64 buf_end, buf_addr = req->rw.addr;
2788 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2790 /* not inside the mapped region */
2791 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2795 * May not be a start of buffer, set size appropriately
2796 * and advance us to the beginning.
2798 offset = buf_addr - imu->ubuf;
2799 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2803 * Don't use iov_iter_advance() here, as it's really slow for
2804 * using the latter parts of a big fixed buffer - it iterates
2805 * over each segment manually. We can cheat a bit here, because
2808 * 1) it's a BVEC iter, we set it up
2809 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2810 * first and last bvec
2812 * So just find our index, and adjust the iterator afterwards.
2813 * If the offset is within the first bvec (or the whole first
2814 * bvec, just use iov_iter_advance(). This makes it easier
2815 * since we can just skip the first segment, which may not
2816 * be PAGE_SIZE aligned.
2818 const struct bio_vec *bvec = imu->bvec;
2820 if (offset <= bvec->bv_len) {
2821 iov_iter_advance(iter, offset);
2823 unsigned long seg_skip;
2825 /* skip first vec */
2826 offset -= bvec->bv_len;
2827 seg_skip = 1 + (offset >> PAGE_SHIFT);
2829 iter->bvec = bvec + seg_skip;
2830 iter->nr_segs -= seg_skip;
2831 iter->count -= bvec->bv_len + offset;
2832 iter->iov_offset = offset & ~PAGE_MASK;
2839 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2841 struct io_ring_ctx *ctx = req->ctx;
2842 struct io_mapped_ubuf *imu = req->imu;
2843 u16 index, buf_index = req->buf_index;
2846 if (unlikely(buf_index >= ctx->nr_user_bufs))
2848 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2849 imu = READ_ONCE(ctx->user_bufs[index]);
2852 return __io_import_fixed(req, rw, iter, imu);
2855 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2858 mutex_unlock(&ctx->uring_lock);
2861 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2864 * "Normal" inline submissions always hold the uring_lock, since we
2865 * grab it from the system call. Same is true for the SQPOLL offload.
2866 * The only exception is when we've detached the request and issue it
2867 * from an async worker thread, grab the lock for that case.
2870 mutex_lock(&ctx->uring_lock);
2873 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2874 int bgid, struct io_buffer *kbuf,
2877 struct io_buffer *head;
2879 if (req->flags & REQ_F_BUFFER_SELECTED)
2882 io_ring_submit_lock(req->ctx, needs_lock);
2884 lockdep_assert_held(&req->ctx->uring_lock);
2886 head = xa_load(&req->ctx->io_buffers, bgid);
2888 if (!list_empty(&head->list)) {
2889 kbuf = list_last_entry(&head->list, struct io_buffer,
2891 list_del(&kbuf->list);
2894 xa_erase(&req->ctx->io_buffers, bgid);
2896 if (*len > kbuf->len)
2899 kbuf = ERR_PTR(-ENOBUFS);
2902 io_ring_submit_unlock(req->ctx, needs_lock);
2907 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2910 struct io_buffer *kbuf;
2913 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2914 bgid = req->buf_index;
2915 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2918 req->rw.addr = (u64) (unsigned long) kbuf;
2919 req->flags |= REQ_F_BUFFER_SELECTED;
2920 return u64_to_user_ptr(kbuf->addr);
2923 #ifdef CONFIG_COMPAT
2924 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2927 struct compat_iovec __user *uiov;
2928 compat_ssize_t clen;
2932 uiov = u64_to_user_ptr(req->rw.addr);
2933 if (!access_ok(uiov, sizeof(*uiov)))
2935 if (__get_user(clen, &uiov->iov_len))
2941 buf = io_rw_buffer_select(req, &len, needs_lock);
2943 return PTR_ERR(buf);
2944 iov[0].iov_base = buf;
2945 iov[0].iov_len = (compat_size_t) len;
2950 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2953 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2957 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2960 len = iov[0].iov_len;
2963 buf = io_rw_buffer_select(req, &len, needs_lock);
2965 return PTR_ERR(buf);
2966 iov[0].iov_base = buf;
2967 iov[0].iov_len = len;
2971 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2974 if (req->flags & REQ_F_BUFFER_SELECTED) {
2975 struct io_buffer *kbuf;
2977 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2978 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2979 iov[0].iov_len = kbuf->len;
2982 if (req->rw.len != 1)
2985 #ifdef CONFIG_COMPAT
2986 if (req->ctx->compat)
2987 return io_compat_import(req, iov, needs_lock);
2990 return __io_iov_buffer_select(req, iov, needs_lock);
2993 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2994 struct iov_iter *iter, bool needs_lock)
2996 void __user *buf = u64_to_user_ptr(req->rw.addr);
2997 size_t sqe_len = req->rw.len;
2998 u8 opcode = req->opcode;
3001 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3003 return io_import_fixed(req, rw, iter);
3006 /* buffer index only valid with fixed read/write, or buffer select */
3007 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3010 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3011 if (req->flags & REQ_F_BUFFER_SELECT) {
3012 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3014 return PTR_ERR(buf);
3015 req->rw.len = sqe_len;
3018 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3023 if (req->flags & REQ_F_BUFFER_SELECT) {
3024 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3026 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3031 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3035 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3037 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3041 * For files that don't have ->read_iter() and ->write_iter(), handle them
3042 * by looping over ->read() or ->write() manually.
3044 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3046 struct kiocb *kiocb = &req->rw.kiocb;
3047 struct file *file = req->file;
3051 * Don't support polled IO through this interface, and we can't
3052 * support non-blocking either. For the latter, this just causes
3053 * the kiocb to be handled from an async context.
3055 if (kiocb->ki_flags & IOCB_HIPRI)
3057 if (kiocb->ki_flags & IOCB_NOWAIT)
3060 while (iov_iter_count(iter)) {
3064 if (!iov_iter_is_bvec(iter)) {
3065 iovec = iov_iter_iovec(iter);
3067 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3068 iovec.iov_len = req->rw.len;
3072 nr = file->f_op->read(file, iovec.iov_base,
3073 iovec.iov_len, io_kiocb_ppos(kiocb));
3075 nr = file->f_op->write(file, iovec.iov_base,
3076 iovec.iov_len, io_kiocb_ppos(kiocb));
3085 if (nr != iovec.iov_len)
3089 iov_iter_advance(iter, nr);
3095 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3096 const struct iovec *fast_iov, struct iov_iter *iter)
3098 struct io_async_rw *rw = req->async_data;
3100 memcpy(&rw->iter, iter, sizeof(*iter));
3101 rw->free_iovec = iovec;
3103 /* can only be fixed buffers, no need to do anything */
3104 if (iov_iter_is_bvec(iter))
3107 unsigned iov_off = 0;
3109 rw->iter.iov = rw->fast_iov;
3110 if (iter->iov != fast_iov) {
3111 iov_off = iter->iov - fast_iov;
3112 rw->iter.iov += iov_off;
3114 if (rw->fast_iov != fast_iov)
3115 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3116 sizeof(struct iovec) * iter->nr_segs);
3118 req->flags |= REQ_F_NEED_CLEANUP;
3122 static inline int io_alloc_async_data(struct io_kiocb *req)
3124 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3125 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3126 return req->async_data == NULL;
3129 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3130 const struct iovec *fast_iov,
3131 struct iov_iter *iter, bool force)
3133 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3135 if (!req->async_data) {
3136 if (io_alloc_async_data(req)) {
3141 io_req_map_rw(req, iovec, fast_iov, iter);
3146 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3148 struct io_async_rw *iorw = req->async_data;
3149 struct iovec *iov = iorw->fast_iov;
3152 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3153 if (unlikely(ret < 0))
3156 iorw->bytes_done = 0;
3157 iorw->free_iovec = iov;
3159 req->flags |= REQ_F_NEED_CLEANUP;
3163 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3165 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3167 return io_prep_rw(req, sqe);
3171 * This is our waitqueue callback handler, registered through lock_page_async()
3172 * when we initially tried to do the IO with the iocb armed our waitqueue.
3173 * This gets called when the page is unlocked, and we generally expect that to
3174 * happen when the page IO is completed and the page is now uptodate. This will
3175 * queue a task_work based retry of the operation, attempting to copy the data
3176 * again. If the latter fails because the page was NOT uptodate, then we will
3177 * do a thread based blocking retry of the operation. That's the unexpected
3180 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3181 int sync, void *arg)
3183 struct wait_page_queue *wpq;
3184 struct io_kiocb *req = wait->private;
3185 struct wait_page_key *key = arg;
3187 wpq = container_of(wait, struct wait_page_queue, wait);
3189 if (!wake_page_match(wpq, key))
3192 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3193 list_del_init(&wait->entry);
3195 /* submit ref gets dropped, acquire a new one */
3197 io_req_task_queue(req);
3202 * This controls whether a given IO request should be armed for async page
3203 * based retry. If we return false here, the request is handed to the async
3204 * worker threads for retry. If we're doing buffered reads on a regular file,
3205 * we prepare a private wait_page_queue entry and retry the operation. This
3206 * will either succeed because the page is now uptodate and unlocked, or it
3207 * will register a callback when the page is unlocked at IO completion. Through
3208 * that callback, io_uring uses task_work to setup a retry of the operation.
3209 * That retry will attempt the buffered read again. The retry will generally
3210 * succeed, or in rare cases where it fails, we then fall back to using the
3211 * async worker threads for a blocking retry.
3213 static bool io_rw_should_retry(struct io_kiocb *req)
3215 struct io_async_rw *rw = req->async_data;
3216 struct wait_page_queue *wait = &rw->wpq;
3217 struct kiocb *kiocb = &req->rw.kiocb;
3219 /* never retry for NOWAIT, we just complete with -EAGAIN */
3220 if (req->flags & REQ_F_NOWAIT)
3223 /* Only for buffered IO */
3224 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3228 * just use poll if we can, and don't attempt if the fs doesn't
3229 * support callback based unlocks
3231 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3234 wait->wait.func = io_async_buf_func;
3235 wait->wait.private = req;
3236 wait->wait.flags = 0;
3237 INIT_LIST_HEAD(&wait->wait.entry);
3238 kiocb->ki_flags |= IOCB_WAITQ;
3239 kiocb->ki_flags &= ~IOCB_NOWAIT;
3240 kiocb->ki_waitq = wait;
3244 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3246 if (req->file->f_op->read_iter)
3247 return call_read_iter(req->file, &req->rw.kiocb, iter);
3248 else if (req->file->f_op->read)
3249 return loop_rw_iter(READ, req, iter);
3254 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3256 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3257 struct kiocb *kiocb = &req->rw.kiocb;
3258 struct iov_iter __iter, *iter = &__iter;
3259 struct io_async_rw *rw = req->async_data;
3260 ssize_t io_size, ret, ret2;
3261 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3267 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3271 io_size = iov_iter_count(iter);
3272 req->result = io_size;
3274 /* Ensure we clear previously set non-block flag */
3275 if (!force_nonblock)
3276 kiocb->ki_flags &= ~IOCB_NOWAIT;
3278 kiocb->ki_flags |= IOCB_NOWAIT;
3280 /* If the file doesn't support async, just async punt */
3281 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3282 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3283 return ret ?: -EAGAIN;
3286 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3287 if (unlikely(ret)) {
3292 ret = io_iter_do_read(req, iter);
3294 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3295 req->flags &= ~REQ_F_REISSUE;
3296 /* IOPOLL retry should happen for io-wq threads */
3297 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3299 /* no retry on NONBLOCK nor RWF_NOWAIT */
3300 if (req->flags & REQ_F_NOWAIT)
3302 /* some cases will consume bytes even on error returns */
3303 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3305 } else if (ret == -EIOCBQUEUED) {
3307 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3308 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3309 /* read all, failed, already did sync or don't want to retry */
3313 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3318 rw = req->async_data;
3319 /* now use our persistent iterator, if we aren't already */
3324 rw->bytes_done += ret;
3325 /* if we can retry, do so with the callbacks armed */
3326 if (!io_rw_should_retry(req)) {
3327 kiocb->ki_flags &= ~IOCB_WAITQ;
3332 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3333 * we get -EIOCBQUEUED, then we'll get a notification when the
3334 * desired page gets unlocked. We can also get a partial read
3335 * here, and if we do, then just retry at the new offset.
3337 ret = io_iter_do_read(req, iter);
3338 if (ret == -EIOCBQUEUED)
3340 /* we got some bytes, but not all. retry. */
3341 kiocb->ki_flags &= ~IOCB_WAITQ;
3342 } while (ret > 0 && ret < io_size);
3344 kiocb_done(kiocb, ret, issue_flags);
3346 /* it's faster to check here then delegate to kfree */
3352 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3354 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3356 return io_prep_rw(req, sqe);
3359 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3361 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3362 struct kiocb *kiocb = &req->rw.kiocb;
3363 struct iov_iter __iter, *iter = &__iter;
3364 struct io_async_rw *rw = req->async_data;
3365 ssize_t ret, ret2, io_size;
3366 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3372 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3376 io_size = iov_iter_count(iter);
3377 req->result = io_size;
3379 /* Ensure we clear previously set non-block flag */
3380 if (!force_nonblock)
3381 kiocb->ki_flags &= ~IOCB_NOWAIT;
3383 kiocb->ki_flags |= IOCB_NOWAIT;
3385 /* If the file doesn't support async, just async punt */
3386 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3389 /* file path doesn't support NOWAIT for non-direct_IO */
3390 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3391 (req->flags & REQ_F_ISREG))
3394 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3399 * Open-code file_start_write here to grab freeze protection,
3400 * which will be released by another thread in
3401 * io_complete_rw(). Fool lockdep by telling it the lock got
3402 * released so that it doesn't complain about the held lock when
3403 * we return to userspace.
3405 if (req->flags & REQ_F_ISREG) {
3406 sb_start_write(file_inode(req->file)->i_sb);
3407 __sb_writers_release(file_inode(req->file)->i_sb,
3410 kiocb->ki_flags |= IOCB_WRITE;
3412 if (req->file->f_op->write_iter)
3413 ret2 = call_write_iter(req->file, kiocb, iter);
3414 else if (req->file->f_op->write)
3415 ret2 = loop_rw_iter(WRITE, req, iter);
3419 if (req->flags & REQ_F_REISSUE) {
3420 req->flags &= ~REQ_F_REISSUE;
3425 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3426 * retry them without IOCB_NOWAIT.
3428 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3430 /* no retry on NONBLOCK nor RWF_NOWAIT */
3431 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3433 if (!force_nonblock || ret2 != -EAGAIN) {
3434 /* IOPOLL retry should happen for io-wq threads */
3435 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3438 kiocb_done(kiocb, ret2, issue_flags);
3441 /* some cases will consume bytes even on error returns */
3442 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3443 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3444 return ret ?: -EAGAIN;
3447 /* it's reportedly faster than delegating the null check to kfree() */
3453 static int io_renameat_prep(struct io_kiocb *req,
3454 const struct io_uring_sqe *sqe)
3456 struct io_rename *ren = &req->rename;
3457 const char __user *oldf, *newf;
3459 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3461 if (sqe->ioprio || sqe->buf_index)
3463 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3466 ren->old_dfd = READ_ONCE(sqe->fd);
3467 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3468 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3469 ren->new_dfd = READ_ONCE(sqe->len);
3470 ren->flags = READ_ONCE(sqe->rename_flags);
3472 ren->oldpath = getname(oldf);
3473 if (IS_ERR(ren->oldpath))
3474 return PTR_ERR(ren->oldpath);
3476 ren->newpath = getname(newf);
3477 if (IS_ERR(ren->newpath)) {
3478 putname(ren->oldpath);
3479 return PTR_ERR(ren->newpath);
3482 req->flags |= REQ_F_NEED_CLEANUP;
3486 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3488 struct io_rename *ren = &req->rename;
3491 if (issue_flags & IO_URING_F_NONBLOCK)
3494 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3495 ren->newpath, ren->flags);
3497 req->flags &= ~REQ_F_NEED_CLEANUP;
3500 io_req_complete(req, ret);
3504 static int io_unlinkat_prep(struct io_kiocb *req,
3505 const struct io_uring_sqe *sqe)
3507 struct io_unlink *un = &req->unlink;
3508 const char __user *fname;
3510 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3512 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3514 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3517 un->dfd = READ_ONCE(sqe->fd);
3519 un->flags = READ_ONCE(sqe->unlink_flags);
3520 if (un->flags & ~AT_REMOVEDIR)
3523 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3524 un->filename = getname(fname);
3525 if (IS_ERR(un->filename))
3526 return PTR_ERR(un->filename);
3528 req->flags |= REQ_F_NEED_CLEANUP;
3532 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3534 struct io_unlink *un = &req->unlink;
3537 if (issue_flags & IO_URING_F_NONBLOCK)
3540 if (un->flags & AT_REMOVEDIR)
3541 ret = do_rmdir(un->dfd, un->filename);
3543 ret = do_unlinkat(un->dfd, un->filename);
3545 req->flags &= ~REQ_F_NEED_CLEANUP;
3548 io_req_complete(req, ret);
3552 static int io_shutdown_prep(struct io_kiocb *req,
3553 const struct io_uring_sqe *sqe)
3555 #if defined(CONFIG_NET)
3556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3558 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3562 req->shutdown.how = READ_ONCE(sqe->len);
3569 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3571 #if defined(CONFIG_NET)
3572 struct socket *sock;
3575 if (issue_flags & IO_URING_F_NONBLOCK)
3578 sock = sock_from_file(req->file);
3579 if (unlikely(!sock))
3582 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3585 io_req_complete(req, ret);
3592 static int __io_splice_prep(struct io_kiocb *req,
3593 const struct io_uring_sqe *sqe)
3595 struct io_splice *sp = &req->splice;
3596 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3598 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3602 sp->len = READ_ONCE(sqe->len);
3603 sp->flags = READ_ONCE(sqe->splice_flags);
3605 if (unlikely(sp->flags & ~valid_flags))
3608 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3609 (sp->flags & SPLICE_F_FD_IN_FIXED));
3612 req->flags |= REQ_F_NEED_CLEANUP;
3616 static int io_tee_prep(struct io_kiocb *req,
3617 const struct io_uring_sqe *sqe)
3619 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3621 return __io_splice_prep(req, sqe);
3624 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3626 struct io_splice *sp = &req->splice;
3627 struct file *in = sp->file_in;
3628 struct file *out = sp->file_out;
3629 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3632 if (issue_flags & IO_URING_F_NONBLOCK)
3635 ret = do_tee(in, out, sp->len, flags);
3637 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3639 req->flags &= ~REQ_F_NEED_CLEANUP;
3643 io_req_complete(req, ret);
3647 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3649 struct io_splice *sp = &req->splice;
3651 sp->off_in = READ_ONCE(sqe->splice_off_in);
3652 sp->off_out = READ_ONCE(sqe->off);
3653 return __io_splice_prep(req, sqe);
3656 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3658 struct io_splice *sp = &req->splice;
3659 struct file *in = sp->file_in;
3660 struct file *out = sp->file_out;
3661 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3662 loff_t *poff_in, *poff_out;
3665 if (issue_flags & IO_URING_F_NONBLOCK)
3668 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3669 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3672 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3674 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3676 req->flags &= ~REQ_F_NEED_CLEANUP;
3680 io_req_complete(req, ret);
3685 * IORING_OP_NOP just posts a completion event, nothing else.
3687 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3689 struct io_ring_ctx *ctx = req->ctx;
3691 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3694 __io_req_complete(req, issue_flags, 0, 0);
3698 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3700 struct io_ring_ctx *ctx = req->ctx;
3705 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3707 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3710 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3711 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3714 req->sync.off = READ_ONCE(sqe->off);
3715 req->sync.len = READ_ONCE(sqe->len);
3719 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3721 loff_t end = req->sync.off + req->sync.len;
3724 /* fsync always requires a blocking context */
3725 if (issue_flags & IO_URING_F_NONBLOCK)
3728 ret = vfs_fsync_range(req->file, req->sync.off,
3729 end > 0 ? end : LLONG_MAX,
3730 req->sync.flags & IORING_FSYNC_DATASYNC);
3733 io_req_complete(req, ret);
3737 static int io_fallocate_prep(struct io_kiocb *req,
3738 const struct io_uring_sqe *sqe)
3740 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3745 req->sync.off = READ_ONCE(sqe->off);
3746 req->sync.len = READ_ONCE(sqe->addr);
3747 req->sync.mode = READ_ONCE(sqe->len);
3751 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3755 /* fallocate always requiring blocking context */
3756 if (issue_flags & IO_URING_F_NONBLOCK)
3758 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3762 io_req_complete(req, ret);
3766 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3768 const char __user *fname;
3771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3773 if (unlikely(sqe->ioprio || sqe->buf_index))
3775 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3778 /* open.how should be already initialised */
3779 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3780 req->open.how.flags |= O_LARGEFILE;
3782 req->open.dfd = READ_ONCE(sqe->fd);
3783 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3784 req->open.filename = getname(fname);
3785 if (IS_ERR(req->open.filename)) {
3786 ret = PTR_ERR(req->open.filename);
3787 req->open.filename = NULL;
3790 req->open.nofile = rlimit(RLIMIT_NOFILE);
3791 req->flags |= REQ_F_NEED_CLEANUP;
3795 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3797 u64 mode = READ_ONCE(sqe->len);
3798 u64 flags = READ_ONCE(sqe->open_flags);
3800 req->open.how = build_open_how(flags, mode);
3801 return __io_openat_prep(req, sqe);
3804 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3806 struct open_how __user *how;
3810 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3811 len = READ_ONCE(sqe->len);
3812 if (len < OPEN_HOW_SIZE_VER0)
3815 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3820 return __io_openat_prep(req, sqe);
3823 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3825 struct open_flags op;
3828 bool resolve_nonblock;
3831 ret = build_open_flags(&req->open.how, &op);
3834 nonblock_set = op.open_flag & O_NONBLOCK;
3835 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3836 if (issue_flags & IO_URING_F_NONBLOCK) {
3838 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3839 * it'll always -EAGAIN
3841 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3843 op.lookup_flags |= LOOKUP_CACHED;
3844 op.open_flag |= O_NONBLOCK;
3847 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3851 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3854 * We could hang on to this 'fd' on retrying, but seems like
3855 * marginal gain for something that is now known to be a slower
3856 * path. So just put it, and we'll get a new one when we retry.
3860 ret = PTR_ERR(file);
3861 /* only retry if RESOLVE_CACHED wasn't already set by application */
3862 if (ret == -EAGAIN &&
3863 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3868 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3869 file->f_flags &= ~O_NONBLOCK;
3870 fsnotify_open(file);
3871 fd_install(ret, file);
3873 putname(req->open.filename);
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3877 __io_req_complete(req, issue_flags, ret, 0);
3881 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3883 return io_openat2(req, issue_flags);
3886 static int io_remove_buffers_prep(struct io_kiocb *req,
3887 const struct io_uring_sqe *sqe)
3889 struct io_provide_buf *p = &req->pbuf;
3892 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3895 tmp = READ_ONCE(sqe->fd);
3896 if (!tmp || tmp > USHRT_MAX)
3899 memset(p, 0, sizeof(*p));
3901 p->bgid = READ_ONCE(sqe->buf_group);
3905 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3906 int bgid, unsigned nbufs)
3910 /* shouldn't happen */
3914 /* the head kbuf is the list itself */
3915 while (!list_empty(&buf->list)) {
3916 struct io_buffer *nxt;
3918 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3919 list_del(&nxt->list);
3926 xa_erase(&ctx->io_buffers, bgid);
3931 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3933 struct io_provide_buf *p = &req->pbuf;
3934 struct io_ring_ctx *ctx = req->ctx;
3935 struct io_buffer *head;
3937 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3939 io_ring_submit_lock(ctx, !force_nonblock);
3941 lockdep_assert_held(&ctx->uring_lock);
3944 head = xa_load(&ctx->io_buffers, p->bgid);
3946 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3950 /* complete before unlock, IOPOLL may need the lock */
3951 __io_req_complete(req, issue_flags, ret, 0);
3952 io_ring_submit_unlock(ctx, !force_nonblock);
3956 static int io_provide_buffers_prep(struct io_kiocb *req,
3957 const struct io_uring_sqe *sqe)
3959 unsigned long size, tmp_check;
3960 struct io_provide_buf *p = &req->pbuf;
3963 if (sqe->ioprio || sqe->rw_flags)
3966 tmp = READ_ONCE(sqe->fd);
3967 if (!tmp || tmp > USHRT_MAX)
3970 p->addr = READ_ONCE(sqe->addr);
3971 p->len = READ_ONCE(sqe->len);
3973 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3976 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3979 size = (unsigned long)p->len * p->nbufs;
3980 if (!access_ok(u64_to_user_ptr(p->addr), size))
3983 p->bgid = READ_ONCE(sqe->buf_group);
3984 tmp = READ_ONCE(sqe->off);
3985 if (tmp > USHRT_MAX)
3991 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3993 struct io_buffer *buf;
3994 u64 addr = pbuf->addr;
3995 int i, bid = pbuf->bid;
3997 for (i = 0; i < pbuf->nbufs; i++) {
3998 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4003 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4008 INIT_LIST_HEAD(&buf->list);
4011 list_add_tail(&buf->list, &(*head)->list);
4015 return i ? i : -ENOMEM;
4018 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4020 struct io_provide_buf *p = &req->pbuf;
4021 struct io_ring_ctx *ctx = req->ctx;
4022 struct io_buffer *head, *list;
4024 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4026 io_ring_submit_lock(ctx, !force_nonblock);
4028 lockdep_assert_held(&ctx->uring_lock);
4030 list = head = xa_load(&ctx->io_buffers, p->bgid);
4032 ret = io_add_buffers(p, &head);
4033 if (ret >= 0 && !list) {
4034 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4036 __io_remove_buffers(ctx, head, p->bgid, -1U);
4040 /* complete before unlock, IOPOLL may need the lock */
4041 __io_req_complete(req, issue_flags, ret, 0);
4042 io_ring_submit_unlock(ctx, !force_nonblock);
4046 static int io_epoll_ctl_prep(struct io_kiocb *req,
4047 const struct io_uring_sqe *sqe)
4049 #if defined(CONFIG_EPOLL)
4050 if (sqe->ioprio || sqe->buf_index)
4052 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4055 req->epoll.epfd = READ_ONCE(sqe->fd);
4056 req->epoll.op = READ_ONCE(sqe->len);
4057 req->epoll.fd = READ_ONCE(sqe->off);
4059 if (ep_op_has_event(req->epoll.op)) {
4060 struct epoll_event __user *ev;
4062 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4063 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4073 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4075 #if defined(CONFIG_EPOLL)
4076 struct io_epoll *ie = &req->epoll;
4078 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4080 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4081 if (force_nonblock && ret == -EAGAIN)
4086 __io_req_complete(req, issue_flags, ret, 0);
4093 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4095 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4096 if (sqe->ioprio || sqe->buf_index || sqe->off)
4098 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4101 req->madvise.addr = READ_ONCE(sqe->addr);
4102 req->madvise.len = READ_ONCE(sqe->len);
4103 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4110 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4112 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4113 struct io_madvise *ma = &req->madvise;
4116 if (issue_flags & IO_URING_F_NONBLOCK)
4119 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4122 io_req_complete(req, ret);
4129 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4136 req->fadvise.offset = READ_ONCE(sqe->off);
4137 req->fadvise.len = READ_ONCE(sqe->len);
4138 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4142 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4144 struct io_fadvise *fa = &req->fadvise;
4147 if (issue_flags & IO_URING_F_NONBLOCK) {
4148 switch (fa->advice) {
4149 case POSIX_FADV_NORMAL:
4150 case POSIX_FADV_RANDOM:
4151 case POSIX_FADV_SEQUENTIAL:
4158 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4161 __io_req_complete(req, issue_flags, ret, 0);
4165 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4167 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4169 if (sqe->ioprio || sqe->buf_index)
4171 if (req->flags & REQ_F_FIXED_FILE)
4174 req->statx.dfd = READ_ONCE(sqe->fd);
4175 req->statx.mask = READ_ONCE(sqe->len);
4176 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4177 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4178 req->statx.flags = READ_ONCE(sqe->statx_flags);
4183 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4185 struct io_statx *ctx = &req->statx;
4188 if (issue_flags & IO_URING_F_NONBLOCK)
4191 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4196 io_req_complete(req, ret);
4200 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4204 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4205 sqe->rw_flags || sqe->buf_index)
4207 if (req->flags & REQ_F_FIXED_FILE)
4210 req->close.fd = READ_ONCE(sqe->fd);
4214 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4216 struct files_struct *files = current->files;
4217 struct io_close *close = &req->close;
4218 struct fdtable *fdt;
4219 struct file *file = NULL;
4222 spin_lock(&files->file_lock);
4223 fdt = files_fdtable(files);
4224 if (close->fd >= fdt->max_fds) {
4225 spin_unlock(&files->file_lock);
4228 file = fdt->fd[close->fd];
4229 if (!file || file->f_op == &io_uring_fops) {
4230 spin_unlock(&files->file_lock);
4235 /* if the file has a flush method, be safe and punt to async */
4236 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4237 spin_unlock(&files->file_lock);
4241 ret = __close_fd_get_file(close->fd, &file);
4242 spin_unlock(&files->file_lock);
4249 /* No ->flush() or already async, safely close from here */
4250 ret = filp_close(file, current->files);
4256 __io_req_complete(req, issue_flags, ret, 0);
4260 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4262 struct io_ring_ctx *ctx = req->ctx;
4264 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4266 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4269 req->sync.off = READ_ONCE(sqe->off);
4270 req->sync.len = READ_ONCE(sqe->len);
4271 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4275 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4279 /* sync_file_range always requires a blocking context */
4280 if (issue_flags & IO_URING_F_NONBLOCK)
4283 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4287 io_req_complete(req, ret);
4291 #if defined(CONFIG_NET)
4292 static int io_setup_async_msg(struct io_kiocb *req,
4293 struct io_async_msghdr *kmsg)
4295 struct io_async_msghdr *async_msg = req->async_data;
4299 if (io_alloc_async_data(req)) {
4300 kfree(kmsg->free_iov);
4303 async_msg = req->async_data;
4304 req->flags |= REQ_F_NEED_CLEANUP;
4305 memcpy(async_msg, kmsg, sizeof(*kmsg));
4306 async_msg->msg.msg_name = &async_msg->addr;
4307 /* if were using fast_iov, set it to the new one */
4308 if (!async_msg->free_iov)
4309 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4314 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4315 struct io_async_msghdr *iomsg)
4317 iomsg->msg.msg_name = &iomsg->addr;
4318 iomsg->free_iov = iomsg->fast_iov;
4319 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4320 req->sr_msg.msg_flags, &iomsg->free_iov);
4323 static int io_sendmsg_prep_async(struct io_kiocb *req)
4327 ret = io_sendmsg_copy_hdr(req, req->async_data);
4329 req->flags |= REQ_F_NEED_CLEANUP;
4333 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4335 struct io_sr_msg *sr = &req->sr_msg;
4337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4340 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4341 sr->len = READ_ONCE(sqe->len);
4342 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4343 if (sr->msg_flags & MSG_DONTWAIT)
4344 req->flags |= REQ_F_NOWAIT;
4346 #ifdef CONFIG_COMPAT
4347 if (req->ctx->compat)
4348 sr->msg_flags |= MSG_CMSG_COMPAT;
4353 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4355 struct io_async_msghdr iomsg, *kmsg;
4356 struct socket *sock;
4361 sock = sock_from_file(req->file);
4362 if (unlikely(!sock))
4365 kmsg = req->async_data;
4367 ret = io_sendmsg_copy_hdr(req, &iomsg);
4373 flags = req->sr_msg.msg_flags;
4374 if (issue_flags & IO_URING_F_NONBLOCK)
4375 flags |= MSG_DONTWAIT;
4376 if (flags & MSG_WAITALL)
4377 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4379 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4380 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4381 return io_setup_async_msg(req, kmsg);
4382 if (ret == -ERESTARTSYS)
4385 /* fast path, check for non-NULL to avoid function call */
4387 kfree(kmsg->free_iov);
4388 req->flags &= ~REQ_F_NEED_CLEANUP;
4391 __io_req_complete(req, issue_flags, ret, 0);
4395 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4397 struct io_sr_msg *sr = &req->sr_msg;
4400 struct socket *sock;
4405 sock = sock_from_file(req->file);
4406 if (unlikely(!sock))
4409 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4413 msg.msg_name = NULL;
4414 msg.msg_control = NULL;
4415 msg.msg_controllen = 0;
4416 msg.msg_namelen = 0;
4418 flags = req->sr_msg.msg_flags;
4419 if (issue_flags & IO_URING_F_NONBLOCK)
4420 flags |= MSG_DONTWAIT;
4421 if (flags & MSG_WAITALL)
4422 min_ret = iov_iter_count(&msg.msg_iter);
4424 msg.msg_flags = flags;
4425 ret = sock_sendmsg(sock, &msg);
4426 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4428 if (ret == -ERESTARTSYS)
4433 __io_req_complete(req, issue_flags, ret, 0);
4437 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4438 struct io_async_msghdr *iomsg)
4440 struct io_sr_msg *sr = &req->sr_msg;
4441 struct iovec __user *uiov;
4445 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4446 &iomsg->uaddr, &uiov, &iov_len);
4450 if (req->flags & REQ_F_BUFFER_SELECT) {
4453 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4455 sr->len = iomsg->fast_iov[0].iov_len;
4456 iomsg->free_iov = NULL;
4458 iomsg->free_iov = iomsg->fast_iov;
4459 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4460 &iomsg->free_iov, &iomsg->msg.msg_iter,
4469 #ifdef CONFIG_COMPAT
4470 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4471 struct io_async_msghdr *iomsg)
4473 struct io_sr_msg *sr = &req->sr_msg;
4474 struct compat_iovec __user *uiov;
4479 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4484 uiov = compat_ptr(ptr);
4485 if (req->flags & REQ_F_BUFFER_SELECT) {
4486 compat_ssize_t clen;
4490 if (!access_ok(uiov, sizeof(*uiov)))
4492 if (__get_user(clen, &uiov->iov_len))
4497 iomsg->free_iov = NULL;
4499 iomsg->free_iov = iomsg->fast_iov;
4500 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4501 UIO_FASTIOV, &iomsg->free_iov,
4502 &iomsg->msg.msg_iter, true);
4511 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4512 struct io_async_msghdr *iomsg)
4514 iomsg->msg.msg_name = &iomsg->addr;
4516 #ifdef CONFIG_COMPAT
4517 if (req->ctx->compat)
4518 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4521 return __io_recvmsg_copy_hdr(req, iomsg);
4524 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4527 struct io_sr_msg *sr = &req->sr_msg;
4528 struct io_buffer *kbuf;
4530 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4535 req->flags |= REQ_F_BUFFER_SELECTED;
4539 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4541 return io_put_kbuf(req, req->sr_msg.kbuf);
4544 static int io_recvmsg_prep_async(struct io_kiocb *req)
4548 ret = io_recvmsg_copy_hdr(req, req->async_data);
4550 req->flags |= REQ_F_NEED_CLEANUP;
4554 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4556 struct io_sr_msg *sr = &req->sr_msg;
4558 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4561 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4562 sr->len = READ_ONCE(sqe->len);
4563 sr->bgid = READ_ONCE(sqe->buf_group);
4564 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4565 if (sr->msg_flags & MSG_DONTWAIT)
4566 req->flags |= REQ_F_NOWAIT;
4568 #ifdef CONFIG_COMPAT
4569 if (req->ctx->compat)
4570 sr->msg_flags |= MSG_CMSG_COMPAT;
4575 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4577 struct io_async_msghdr iomsg, *kmsg;
4578 struct socket *sock;
4579 struct io_buffer *kbuf;
4582 int ret, cflags = 0;
4583 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4585 sock = sock_from_file(req->file);
4586 if (unlikely(!sock))
4589 kmsg = req->async_data;
4591 ret = io_recvmsg_copy_hdr(req, &iomsg);
4597 if (req->flags & REQ_F_BUFFER_SELECT) {
4598 kbuf = io_recv_buffer_select(req, !force_nonblock);
4600 return PTR_ERR(kbuf);
4601 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4602 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4603 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4604 1, req->sr_msg.len);
4607 flags = req->sr_msg.msg_flags;
4609 flags |= MSG_DONTWAIT;
4610 if (flags & MSG_WAITALL)
4611 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4613 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4614 kmsg->uaddr, flags);
4615 if (force_nonblock && ret == -EAGAIN)
4616 return io_setup_async_msg(req, kmsg);
4617 if (ret == -ERESTARTSYS)
4620 if (req->flags & REQ_F_BUFFER_SELECTED)
4621 cflags = io_put_recv_kbuf(req);
4622 /* fast path, check for non-NULL to avoid function call */
4624 kfree(kmsg->free_iov);
4625 req->flags &= ~REQ_F_NEED_CLEANUP;
4626 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4628 __io_req_complete(req, issue_flags, ret, cflags);
4632 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4634 struct io_buffer *kbuf;
4635 struct io_sr_msg *sr = &req->sr_msg;
4637 void __user *buf = sr->buf;
4638 struct socket *sock;
4642 int ret, cflags = 0;
4643 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4645 sock = sock_from_file(req->file);
4646 if (unlikely(!sock))
4649 if (req->flags & REQ_F_BUFFER_SELECT) {
4650 kbuf = io_recv_buffer_select(req, !force_nonblock);
4652 return PTR_ERR(kbuf);
4653 buf = u64_to_user_ptr(kbuf->addr);
4656 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4660 msg.msg_name = NULL;
4661 msg.msg_control = NULL;
4662 msg.msg_controllen = 0;
4663 msg.msg_namelen = 0;
4664 msg.msg_iocb = NULL;
4667 flags = req->sr_msg.msg_flags;
4669 flags |= MSG_DONTWAIT;
4670 if (flags & MSG_WAITALL)
4671 min_ret = iov_iter_count(&msg.msg_iter);
4673 ret = sock_recvmsg(sock, &msg, flags);
4674 if (force_nonblock && ret == -EAGAIN)
4676 if (ret == -ERESTARTSYS)
4679 if (req->flags & REQ_F_BUFFER_SELECTED)
4680 cflags = io_put_recv_kbuf(req);
4681 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4683 __io_req_complete(req, issue_flags, ret, cflags);
4687 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4689 struct io_accept *accept = &req->accept;
4691 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4693 if (sqe->ioprio || sqe->len || sqe->buf_index)
4696 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4697 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4698 accept->flags = READ_ONCE(sqe->accept_flags);
4699 accept->nofile = rlimit(RLIMIT_NOFILE);
4703 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4705 struct io_accept *accept = &req->accept;
4706 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4707 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4710 if (req->file->f_flags & O_NONBLOCK)
4711 req->flags |= REQ_F_NOWAIT;
4713 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4714 accept->addr_len, accept->flags,
4716 if (ret == -EAGAIN && force_nonblock)
4719 if (ret == -ERESTARTSYS)
4723 __io_req_complete(req, issue_flags, ret, 0);
4727 static int io_connect_prep_async(struct io_kiocb *req)
4729 struct io_async_connect *io = req->async_data;
4730 struct io_connect *conn = &req->connect;
4732 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4735 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4737 struct io_connect *conn = &req->connect;
4739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4741 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4744 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4745 conn->addr_len = READ_ONCE(sqe->addr2);
4749 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4751 struct io_async_connect __io, *io;
4752 unsigned file_flags;
4754 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4756 if (req->async_data) {
4757 io = req->async_data;
4759 ret = move_addr_to_kernel(req->connect.addr,
4760 req->connect.addr_len,
4767 file_flags = force_nonblock ? O_NONBLOCK : 0;
4769 ret = __sys_connect_file(req->file, &io->address,
4770 req->connect.addr_len, file_flags);
4771 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4772 if (req->async_data)
4774 if (io_alloc_async_data(req)) {
4778 memcpy(req->async_data, &__io, sizeof(__io));
4781 if (ret == -ERESTARTSYS)
4786 __io_req_complete(req, issue_flags, ret, 0);
4789 #else /* !CONFIG_NET */
4790 #define IO_NETOP_FN(op) \
4791 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4793 return -EOPNOTSUPP; \
4796 #define IO_NETOP_PREP(op) \
4798 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4800 return -EOPNOTSUPP; \
4803 #define IO_NETOP_PREP_ASYNC(op) \
4805 static int io_##op##_prep_async(struct io_kiocb *req) \
4807 return -EOPNOTSUPP; \
4810 IO_NETOP_PREP_ASYNC(sendmsg);
4811 IO_NETOP_PREP_ASYNC(recvmsg);
4812 IO_NETOP_PREP_ASYNC(connect);
4813 IO_NETOP_PREP(accept);
4816 #endif /* CONFIG_NET */
4818 struct io_poll_table {
4819 struct poll_table_struct pt;
4820 struct io_kiocb *req;
4825 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4826 __poll_t mask, io_req_tw_func_t func)
4828 /* for instances that support it check for an event match first: */
4829 if (mask && !(mask & poll->events))
4832 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4834 list_del_init(&poll->wait.entry);
4837 req->io_task_work.func = func;
4840 * If this fails, then the task is exiting. When a task exits, the
4841 * work gets canceled, so just cancel this request as well instead
4842 * of executing it. We can't safely execute it anyway, as we may not
4843 * have the needed state needed for it anyway.
4845 io_req_task_work_add(req);
4849 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4850 __acquires(&req->ctx->completion_lock)
4852 struct io_ring_ctx *ctx = req->ctx;
4854 if (unlikely(req->task->flags & PF_EXITING))
4855 WRITE_ONCE(poll->canceled, true);
4857 if (!req->result && !READ_ONCE(poll->canceled)) {
4858 struct poll_table_struct pt = { ._key = poll->events };
4860 req->result = vfs_poll(req->file, &pt) & poll->events;
4863 spin_lock(&ctx->completion_lock);
4864 if (!req->result && !READ_ONCE(poll->canceled)) {
4865 add_wait_queue(poll->head, &poll->wait);
4872 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4874 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4875 if (req->opcode == IORING_OP_POLL_ADD)
4876 return req->async_data;
4877 return req->apoll->double_poll;
4880 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4882 if (req->opcode == IORING_OP_POLL_ADD)
4884 return &req->apoll->poll;
4887 static void io_poll_remove_double(struct io_kiocb *req)
4888 __must_hold(&req->ctx->completion_lock)
4890 struct io_poll_iocb *poll = io_poll_get_double(req);
4892 lockdep_assert_held(&req->ctx->completion_lock);
4894 if (poll && poll->head) {
4895 struct wait_queue_head *head = poll->head;
4897 spin_lock_irq(&head->lock);
4898 list_del_init(&poll->wait.entry);
4899 if (poll->wait.private)
4902 spin_unlock_irq(&head->lock);
4906 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4907 __must_hold(&req->ctx->completion_lock)
4909 struct io_ring_ctx *ctx = req->ctx;
4910 unsigned flags = IORING_CQE_F_MORE;
4913 if (READ_ONCE(req->poll.canceled)) {
4915 req->poll.events |= EPOLLONESHOT;
4917 error = mangle_poll(mask);
4919 if (req->poll.events & EPOLLONESHOT)
4921 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4922 req->poll.done = true;
4925 if (flags & IORING_CQE_F_MORE)
4928 io_commit_cqring(ctx);
4929 return !(flags & IORING_CQE_F_MORE);
4932 static void io_poll_task_func(struct io_kiocb *req)
4934 struct io_ring_ctx *ctx = req->ctx;
4935 struct io_kiocb *nxt;
4937 if (io_poll_rewait(req, &req->poll)) {
4938 spin_unlock(&ctx->completion_lock);
4942 done = io_poll_complete(req, req->result);
4944 io_poll_remove_double(req);
4945 hash_del(&req->hash_node);
4948 add_wait_queue(req->poll.head, &req->poll.wait);
4950 spin_unlock(&ctx->completion_lock);
4951 io_cqring_ev_posted(ctx);
4954 nxt = io_put_req_find_next(req);
4956 io_req_task_submit(nxt);
4961 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4962 int sync, void *key)
4964 struct io_kiocb *req = wait->private;
4965 struct io_poll_iocb *poll = io_poll_get_single(req);
4966 __poll_t mask = key_to_poll(key);
4967 unsigned long flags;
4969 /* for instances that support it check for an event match first: */
4970 if (mask && !(mask & poll->events))
4972 if (!(poll->events & EPOLLONESHOT))
4973 return poll->wait.func(&poll->wait, mode, sync, key);
4975 list_del_init(&wait->entry);
4980 spin_lock_irqsave(&poll->head->lock, flags);
4981 done = list_empty(&poll->wait.entry);
4983 list_del_init(&poll->wait.entry);
4984 /* make sure double remove sees this as being gone */
4985 wait->private = NULL;
4986 spin_unlock_irqrestore(&poll->head->lock, flags);
4988 /* use wait func handler, so it matches the rq type */
4989 poll->wait.func(&poll->wait, mode, sync, key);
4996 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4997 wait_queue_func_t wake_func)
5001 poll->canceled = false;
5002 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5003 /* mask in events that we always want/need */
5004 poll->events = events | IO_POLL_UNMASK;
5005 INIT_LIST_HEAD(&poll->wait.entry);
5006 init_waitqueue_func_entry(&poll->wait, wake_func);
5009 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5010 struct wait_queue_head *head,
5011 struct io_poll_iocb **poll_ptr)
5013 struct io_kiocb *req = pt->req;
5016 * The file being polled uses multiple waitqueues for poll handling
5017 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5020 if (unlikely(pt->nr_entries)) {
5021 struct io_poll_iocb *poll_one = poll;
5023 /* already have a 2nd entry, fail a third attempt */
5025 pt->error = -EINVAL;
5029 * Can't handle multishot for double wait for now, turn it
5030 * into one-shot mode.
5032 if (!(poll_one->events & EPOLLONESHOT))
5033 poll_one->events |= EPOLLONESHOT;
5034 /* double add on the same waitqueue head, ignore */
5035 if (poll_one->head == head)
5037 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5039 pt->error = -ENOMEM;
5042 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5044 poll->wait.private = req;
5051 if (poll->events & EPOLLEXCLUSIVE)
5052 add_wait_queue_exclusive(head, &poll->wait);
5054 add_wait_queue(head, &poll->wait);
5057 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5058 struct poll_table_struct *p)
5060 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5061 struct async_poll *apoll = pt->req->apoll;
5063 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5066 static void io_async_task_func(struct io_kiocb *req)
5068 struct async_poll *apoll = req->apoll;
5069 struct io_ring_ctx *ctx = req->ctx;
5071 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5073 if (io_poll_rewait(req, &apoll->poll)) {
5074 spin_unlock(&ctx->completion_lock);
5078 hash_del(&req->hash_node);
5079 io_poll_remove_double(req);
5080 spin_unlock(&ctx->completion_lock);
5082 if (!READ_ONCE(apoll->poll.canceled))
5083 io_req_task_submit(req);
5085 io_req_complete_failed(req, -ECANCELED);
5088 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5091 struct io_kiocb *req = wait->private;
5092 struct io_poll_iocb *poll = &req->apoll->poll;
5094 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5097 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5100 static void io_poll_req_insert(struct io_kiocb *req)
5102 struct io_ring_ctx *ctx = req->ctx;
5103 struct hlist_head *list;
5105 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5106 hlist_add_head(&req->hash_node, list);
5109 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5110 struct io_poll_iocb *poll,
5111 struct io_poll_table *ipt, __poll_t mask,
5112 wait_queue_func_t wake_func)
5113 __acquires(&ctx->completion_lock)
5115 struct io_ring_ctx *ctx = req->ctx;
5116 bool cancel = false;
5118 INIT_HLIST_NODE(&req->hash_node);
5119 io_init_poll_iocb(poll, mask, wake_func);
5120 poll->file = req->file;
5121 poll->wait.private = req;
5123 ipt->pt._key = mask;
5126 ipt->nr_entries = 0;
5128 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5129 if (unlikely(!ipt->nr_entries) && !ipt->error)
5130 ipt->error = -EINVAL;
5132 spin_lock(&ctx->completion_lock);
5133 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5134 io_poll_remove_double(req);
5135 if (likely(poll->head)) {
5136 spin_lock_irq(&poll->head->lock);
5137 if (unlikely(list_empty(&poll->wait.entry))) {
5143 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5144 list_del_init(&poll->wait.entry);
5146 WRITE_ONCE(poll->canceled, true);
5147 else if (!poll->done) /* actually waiting for an event */
5148 io_poll_req_insert(req);
5149 spin_unlock_irq(&poll->head->lock);
5161 static int io_arm_poll_handler(struct io_kiocb *req)
5163 const struct io_op_def *def = &io_op_defs[req->opcode];
5164 struct io_ring_ctx *ctx = req->ctx;
5165 struct async_poll *apoll;
5166 struct io_poll_table ipt;
5167 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5170 if (!req->file || !file_can_poll(req->file))
5171 return IO_APOLL_ABORTED;
5172 if (req->flags & REQ_F_POLLED)
5173 return IO_APOLL_ABORTED;
5174 if (!def->pollin && !def->pollout)
5175 return IO_APOLL_ABORTED;
5179 mask |= POLLIN | POLLRDNORM;
5181 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5182 if ((req->opcode == IORING_OP_RECVMSG) &&
5183 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5187 mask |= POLLOUT | POLLWRNORM;
5190 /* if we can't nonblock try, then no point in arming a poll handler */
5191 if (!io_file_supports_nowait(req, rw))
5192 return IO_APOLL_ABORTED;
5194 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5195 if (unlikely(!apoll))
5196 return IO_APOLL_ABORTED;
5197 apoll->double_poll = NULL;
5199 req->flags |= REQ_F_POLLED;
5200 ipt.pt._qproc = io_async_queue_proc;
5202 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5204 if (ret || ipt.error) {
5205 spin_unlock(&ctx->completion_lock);
5207 return IO_APOLL_READY;
5208 return IO_APOLL_ABORTED;
5210 spin_unlock(&ctx->completion_lock);
5211 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5212 mask, apoll->poll.events);
5216 static bool __io_poll_remove_one(struct io_kiocb *req,
5217 struct io_poll_iocb *poll, bool do_cancel)
5218 __must_hold(&req->ctx->completion_lock)
5220 bool do_complete = false;
5224 spin_lock_irq(&poll->head->lock);
5226 WRITE_ONCE(poll->canceled, true);
5227 if (!list_empty(&poll->wait.entry)) {
5228 list_del_init(&poll->wait.entry);
5231 spin_unlock_irq(&poll->head->lock);
5232 hash_del(&req->hash_node);
5236 static bool io_poll_remove_one(struct io_kiocb *req)
5237 __must_hold(&req->ctx->completion_lock)
5242 io_poll_remove_double(req);
5243 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5246 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5247 io_commit_cqring(req->ctx);
5250 /* non-poll requests have submit ref still */
5251 refs = 1 + (req->opcode != IORING_OP_POLL_ADD);
5252 io_put_req_deferred(req, refs);
5258 * Returns true if we found and killed one or more poll requests
5260 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5263 struct hlist_node *tmp;
5264 struct io_kiocb *req;
5267 spin_lock(&ctx->completion_lock);
5268 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5269 struct hlist_head *list;
5271 list = &ctx->cancel_hash[i];
5272 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5273 if (io_match_task(req, tsk, cancel_all))
5274 posted += io_poll_remove_one(req);
5277 spin_unlock(&ctx->completion_lock);
5280 io_cqring_ev_posted(ctx);
5285 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5287 __must_hold(&ctx->completion_lock)
5289 struct hlist_head *list;
5290 struct io_kiocb *req;
5292 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5293 hlist_for_each_entry(req, list, hash_node) {
5294 if (sqe_addr != req->user_data)
5296 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5303 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5305 __must_hold(&ctx->completion_lock)
5307 struct io_kiocb *req;
5309 req = io_poll_find(ctx, sqe_addr, poll_only);
5312 if (io_poll_remove_one(req))
5318 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5323 events = READ_ONCE(sqe->poll32_events);
5325 events = swahw32(events);
5327 if (!(flags & IORING_POLL_ADD_MULTI))
5328 events |= EPOLLONESHOT;
5329 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5332 static int io_poll_update_prep(struct io_kiocb *req,
5333 const struct io_uring_sqe *sqe)
5335 struct io_poll_update *upd = &req->poll_update;
5338 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5340 if (sqe->ioprio || sqe->buf_index)
5342 flags = READ_ONCE(sqe->len);
5343 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5344 IORING_POLL_ADD_MULTI))
5346 /* meaningless without update */
5347 if (flags == IORING_POLL_ADD_MULTI)
5350 upd->old_user_data = READ_ONCE(sqe->addr);
5351 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5352 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5354 upd->new_user_data = READ_ONCE(sqe->off);
5355 if (!upd->update_user_data && upd->new_user_data)
5357 if (upd->update_events)
5358 upd->events = io_poll_parse_events(sqe, flags);
5359 else if (sqe->poll32_events)
5365 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5368 struct io_kiocb *req = wait->private;
5369 struct io_poll_iocb *poll = &req->poll;
5371 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5374 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5375 struct poll_table_struct *p)
5377 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5379 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5382 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5384 struct io_poll_iocb *poll = &req->poll;
5387 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5389 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5391 flags = READ_ONCE(sqe->len);
5392 if (flags & ~IORING_POLL_ADD_MULTI)
5395 poll->events = io_poll_parse_events(sqe, flags);
5399 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5401 struct io_poll_iocb *poll = &req->poll;
5402 struct io_ring_ctx *ctx = req->ctx;
5403 struct io_poll_table ipt;
5406 ipt.pt._qproc = io_poll_queue_proc;
5408 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5411 if (mask) { /* no async, we'd stolen it */
5413 io_poll_complete(req, mask);
5415 spin_unlock(&ctx->completion_lock);
5418 io_cqring_ev_posted(ctx);
5419 if (poll->events & EPOLLONESHOT)
5425 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5427 struct io_ring_ctx *ctx = req->ctx;
5428 struct io_kiocb *preq;
5432 spin_lock(&ctx->completion_lock);
5433 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5439 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5441 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5446 * Don't allow racy completion with singleshot, as we cannot safely
5447 * update those. For multishot, if we're racing with completion, just
5448 * let completion re-add it.
5450 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5451 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5455 /* we now have a detached poll request. reissue. */
5459 spin_unlock(&ctx->completion_lock);
5461 io_req_complete(req, ret);
5464 /* only mask one event flags, keep behavior flags */
5465 if (req->poll_update.update_events) {
5466 preq->poll.events &= ~0xffff;
5467 preq->poll.events |= req->poll_update.events & 0xffff;
5468 preq->poll.events |= IO_POLL_UNMASK;
5470 if (req->poll_update.update_user_data)
5471 preq->user_data = req->poll_update.new_user_data;
5472 spin_unlock(&ctx->completion_lock);
5474 /* complete update request, we're done with it */
5475 io_req_complete(req, ret);
5478 ret = io_poll_add(preq, issue_flags);
5481 io_req_complete(preq, ret);
5487 static void io_req_task_timeout(struct io_kiocb *req)
5489 struct io_ring_ctx *ctx = req->ctx;
5491 spin_lock(&ctx->completion_lock);
5492 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5493 io_commit_cqring(ctx);
5494 spin_unlock(&ctx->completion_lock);
5496 io_cqring_ev_posted(ctx);
5501 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5503 struct io_timeout_data *data = container_of(timer,
5504 struct io_timeout_data, timer);
5505 struct io_kiocb *req = data->req;
5506 struct io_ring_ctx *ctx = req->ctx;
5507 unsigned long flags;
5509 spin_lock_irqsave(&ctx->timeout_lock, flags);
5510 list_del_init(&req->timeout.list);
5511 atomic_set(&req->ctx->cq_timeouts,
5512 atomic_read(&req->ctx->cq_timeouts) + 1);
5513 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5515 req->io_task_work.func = io_req_task_timeout;
5516 io_req_task_work_add(req);
5517 return HRTIMER_NORESTART;
5520 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5522 __must_hold(&ctx->timeout_lock)
5524 struct io_timeout_data *io;
5525 struct io_kiocb *req;
5528 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5529 found = user_data == req->user_data;
5534 return ERR_PTR(-ENOENT);
5536 io = req->async_data;
5537 if (hrtimer_try_to_cancel(&io->timer) == -1)
5538 return ERR_PTR(-EALREADY);
5539 list_del_init(&req->timeout.list);
5543 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5544 __must_hold(&ctx->timeout_lock)
5546 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5549 return PTR_ERR(req);
5552 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5553 io_put_req_deferred(req, 1);
5557 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5558 struct timespec64 *ts, enum hrtimer_mode mode)
5559 __must_hold(&ctx->timeout_lock)
5561 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5562 struct io_timeout_data *data;
5565 return PTR_ERR(req);
5567 req->timeout.off = 0; /* noseq */
5568 data = req->async_data;
5569 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5570 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5571 data->timer.function = io_timeout_fn;
5572 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5576 static int io_timeout_remove_prep(struct io_kiocb *req,
5577 const struct io_uring_sqe *sqe)
5579 struct io_timeout_rem *tr = &req->timeout_rem;
5581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5583 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5585 if (sqe->ioprio || sqe->buf_index || sqe->len)
5588 tr->addr = READ_ONCE(sqe->addr);
5589 tr->flags = READ_ONCE(sqe->timeout_flags);
5590 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5591 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5593 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5595 } else if (tr->flags) {
5596 /* timeout removal doesn't support flags */
5603 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5605 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5610 * Remove or update an existing timeout command
5612 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5614 struct io_timeout_rem *tr = &req->timeout_rem;
5615 struct io_ring_ctx *ctx = req->ctx;
5618 spin_lock_irq(&ctx->timeout_lock);
5619 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5620 ret = io_timeout_cancel(ctx, tr->addr);
5622 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5623 io_translate_timeout_mode(tr->flags));
5624 spin_unlock_irq(&ctx->timeout_lock);
5626 spin_lock(&ctx->completion_lock);
5627 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5628 io_commit_cqring(ctx);
5629 spin_unlock(&ctx->completion_lock);
5630 io_cqring_ev_posted(ctx);
5637 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5638 bool is_timeout_link)
5640 struct io_timeout_data *data;
5642 u32 off = READ_ONCE(sqe->off);
5644 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5646 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5648 if (off && is_timeout_link)
5650 flags = READ_ONCE(sqe->timeout_flags);
5651 if (flags & ~IORING_TIMEOUT_ABS)
5654 req->timeout.off = off;
5655 if (unlikely(off && !req->ctx->off_timeout_used))
5656 req->ctx->off_timeout_used = true;
5658 if (!req->async_data && io_alloc_async_data(req))
5661 data = req->async_data;
5664 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5667 data->mode = io_translate_timeout_mode(flags);
5668 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5669 if (is_timeout_link)
5670 io_req_track_inflight(req);
5674 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5676 struct io_ring_ctx *ctx = req->ctx;
5677 struct io_timeout_data *data = req->async_data;
5678 struct list_head *entry;
5679 u32 tail, off = req->timeout.off;
5681 spin_lock_irq(&ctx->timeout_lock);
5684 * sqe->off holds how many events that need to occur for this
5685 * timeout event to be satisfied. If it isn't set, then this is
5686 * a pure timeout request, sequence isn't used.
5688 if (io_is_timeout_noseq(req)) {
5689 entry = ctx->timeout_list.prev;
5693 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5694 req->timeout.target_seq = tail + off;
5696 /* Update the last seq here in case io_flush_timeouts() hasn't.
5697 * This is safe because ->completion_lock is held, and submissions
5698 * and completions are never mixed in the same ->completion_lock section.
5700 ctx->cq_last_tm_flush = tail;
5703 * Insertion sort, ensuring the first entry in the list is always
5704 * the one we need first.
5706 list_for_each_prev(entry, &ctx->timeout_list) {
5707 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5710 if (io_is_timeout_noseq(nxt))
5712 /* nxt.seq is behind @tail, otherwise would've been completed */
5713 if (off >= nxt->timeout.target_seq - tail)
5717 list_add(&req->timeout.list, entry);
5718 data->timer.function = io_timeout_fn;
5719 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5720 spin_unlock_irq(&ctx->timeout_lock);
5724 struct io_cancel_data {
5725 struct io_ring_ctx *ctx;
5729 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5731 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5732 struct io_cancel_data *cd = data;
5734 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5737 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5738 struct io_ring_ctx *ctx)
5740 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5741 enum io_wq_cancel cancel_ret;
5744 if (!tctx || !tctx->io_wq)
5747 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5748 switch (cancel_ret) {
5749 case IO_WQ_CANCEL_OK:
5752 case IO_WQ_CANCEL_RUNNING:
5755 case IO_WQ_CANCEL_NOTFOUND:
5763 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5764 struct io_kiocb *req, __u64 sqe_addr,
5769 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5770 spin_lock(&ctx->completion_lock);
5773 spin_lock_irq(&ctx->timeout_lock);
5774 ret = io_timeout_cancel(ctx, sqe_addr);
5775 spin_unlock_irq(&ctx->timeout_lock);
5778 ret = io_poll_cancel(ctx, sqe_addr, false);
5782 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5783 io_commit_cqring(ctx);
5784 spin_unlock(&ctx->completion_lock);
5785 io_cqring_ev_posted(ctx);
5791 static int io_async_cancel_prep(struct io_kiocb *req,
5792 const struct io_uring_sqe *sqe)
5794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5796 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5798 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5801 req->cancel.addr = READ_ONCE(sqe->addr);
5805 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5807 struct io_ring_ctx *ctx = req->ctx;
5808 u64 sqe_addr = req->cancel.addr;
5809 struct io_tctx_node *node;
5812 /* tasks should wait for their io-wq threads, so safe w/o sync */
5813 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5814 spin_lock(&ctx->completion_lock);
5817 spin_lock_irq(&ctx->timeout_lock);
5818 ret = io_timeout_cancel(ctx, sqe_addr);
5819 spin_unlock_irq(&ctx->timeout_lock);
5822 ret = io_poll_cancel(ctx, sqe_addr, false);
5825 spin_unlock(&ctx->completion_lock);
5827 /* slow path, try all io-wq's */
5828 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5830 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5831 struct io_uring_task *tctx = node->task->io_uring;
5833 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5837 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5839 spin_lock(&ctx->completion_lock);
5841 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5842 io_commit_cqring(ctx);
5843 spin_unlock(&ctx->completion_lock);
5844 io_cqring_ev_posted(ctx);
5852 static int io_rsrc_update_prep(struct io_kiocb *req,
5853 const struct io_uring_sqe *sqe)
5855 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5857 if (sqe->ioprio || sqe->rw_flags)
5860 req->rsrc_update.offset = READ_ONCE(sqe->off);
5861 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5862 if (!req->rsrc_update.nr_args)
5864 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5868 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5870 struct io_ring_ctx *ctx = req->ctx;
5871 struct io_uring_rsrc_update2 up;
5874 if (issue_flags & IO_URING_F_NONBLOCK)
5877 up.offset = req->rsrc_update.offset;
5878 up.data = req->rsrc_update.arg;
5883 mutex_lock(&ctx->uring_lock);
5884 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5885 &up, req->rsrc_update.nr_args);
5886 mutex_unlock(&ctx->uring_lock);
5890 __io_req_complete(req, issue_flags, ret, 0);
5894 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5896 switch (req->opcode) {
5899 case IORING_OP_READV:
5900 case IORING_OP_READ_FIXED:
5901 case IORING_OP_READ:
5902 return io_read_prep(req, sqe);
5903 case IORING_OP_WRITEV:
5904 case IORING_OP_WRITE_FIXED:
5905 case IORING_OP_WRITE:
5906 return io_write_prep(req, sqe);
5907 case IORING_OP_POLL_ADD:
5908 return io_poll_add_prep(req, sqe);
5909 case IORING_OP_POLL_REMOVE:
5910 return io_poll_update_prep(req, sqe);
5911 case IORING_OP_FSYNC:
5912 return io_fsync_prep(req, sqe);
5913 case IORING_OP_SYNC_FILE_RANGE:
5914 return io_sfr_prep(req, sqe);
5915 case IORING_OP_SENDMSG:
5916 case IORING_OP_SEND:
5917 return io_sendmsg_prep(req, sqe);
5918 case IORING_OP_RECVMSG:
5919 case IORING_OP_RECV:
5920 return io_recvmsg_prep(req, sqe);
5921 case IORING_OP_CONNECT:
5922 return io_connect_prep(req, sqe);
5923 case IORING_OP_TIMEOUT:
5924 return io_timeout_prep(req, sqe, false);
5925 case IORING_OP_TIMEOUT_REMOVE:
5926 return io_timeout_remove_prep(req, sqe);
5927 case IORING_OP_ASYNC_CANCEL:
5928 return io_async_cancel_prep(req, sqe);
5929 case IORING_OP_LINK_TIMEOUT:
5930 return io_timeout_prep(req, sqe, true);
5931 case IORING_OP_ACCEPT:
5932 return io_accept_prep(req, sqe);
5933 case IORING_OP_FALLOCATE:
5934 return io_fallocate_prep(req, sqe);
5935 case IORING_OP_OPENAT:
5936 return io_openat_prep(req, sqe);
5937 case IORING_OP_CLOSE:
5938 return io_close_prep(req, sqe);
5939 case IORING_OP_FILES_UPDATE:
5940 return io_rsrc_update_prep(req, sqe);
5941 case IORING_OP_STATX:
5942 return io_statx_prep(req, sqe);
5943 case IORING_OP_FADVISE:
5944 return io_fadvise_prep(req, sqe);
5945 case IORING_OP_MADVISE:
5946 return io_madvise_prep(req, sqe);
5947 case IORING_OP_OPENAT2:
5948 return io_openat2_prep(req, sqe);
5949 case IORING_OP_EPOLL_CTL:
5950 return io_epoll_ctl_prep(req, sqe);
5951 case IORING_OP_SPLICE:
5952 return io_splice_prep(req, sqe);
5953 case IORING_OP_PROVIDE_BUFFERS:
5954 return io_provide_buffers_prep(req, sqe);
5955 case IORING_OP_REMOVE_BUFFERS:
5956 return io_remove_buffers_prep(req, sqe);
5958 return io_tee_prep(req, sqe);
5959 case IORING_OP_SHUTDOWN:
5960 return io_shutdown_prep(req, sqe);
5961 case IORING_OP_RENAMEAT:
5962 return io_renameat_prep(req, sqe);
5963 case IORING_OP_UNLINKAT:
5964 return io_unlinkat_prep(req, sqe);
5967 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5972 static int io_req_prep_async(struct io_kiocb *req)
5974 if (!io_op_defs[req->opcode].needs_async_setup)
5976 if (WARN_ON_ONCE(req->async_data))
5978 if (io_alloc_async_data(req))
5981 switch (req->opcode) {
5982 case IORING_OP_READV:
5983 return io_rw_prep_async(req, READ);
5984 case IORING_OP_WRITEV:
5985 return io_rw_prep_async(req, WRITE);
5986 case IORING_OP_SENDMSG:
5987 return io_sendmsg_prep_async(req);
5988 case IORING_OP_RECVMSG:
5989 return io_recvmsg_prep_async(req);
5990 case IORING_OP_CONNECT:
5991 return io_connect_prep_async(req);
5993 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5998 static u32 io_get_sequence(struct io_kiocb *req)
6000 u32 seq = req->ctx->cached_sq_head;
6002 /* need original cached_sq_head, but it was increased for each req */
6003 io_for_each_link(req, req)
6008 static bool io_drain_req(struct io_kiocb *req)
6010 struct io_kiocb *pos;
6011 struct io_ring_ctx *ctx = req->ctx;
6012 struct io_defer_entry *de;
6017 * If we need to drain a request in the middle of a link, drain the
6018 * head request and the next request/link after the current link.
6019 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6020 * maintained for every request of our link.
6022 if (ctx->drain_next) {
6023 req->flags |= REQ_F_IO_DRAIN;
6024 ctx->drain_next = false;
6026 /* not interested in head, start from the first linked */
6027 io_for_each_link(pos, req->link) {
6028 if (pos->flags & REQ_F_IO_DRAIN) {
6029 ctx->drain_next = true;
6030 req->flags |= REQ_F_IO_DRAIN;
6035 /* Still need defer if there is pending req in defer list. */
6036 if (likely(list_empty_careful(&ctx->defer_list) &&
6037 !(req->flags & REQ_F_IO_DRAIN))) {
6038 ctx->drain_active = false;
6042 seq = io_get_sequence(req);
6043 /* Still a chance to pass the sequence check */
6044 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6047 ret = io_req_prep_async(req);
6050 io_prep_async_link(req);
6051 de = kmalloc(sizeof(*de), GFP_KERNEL);
6055 io_req_complete_failed(req, ret);
6059 spin_lock(&ctx->completion_lock);
6060 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6061 spin_unlock(&ctx->completion_lock);
6063 io_queue_async_work(req);
6067 trace_io_uring_defer(ctx, req, req->user_data);
6070 list_add_tail(&de->list, &ctx->defer_list);
6071 spin_unlock(&ctx->completion_lock);
6075 static void io_clean_op(struct io_kiocb *req)
6077 if (req->flags & REQ_F_BUFFER_SELECTED) {
6078 switch (req->opcode) {
6079 case IORING_OP_READV:
6080 case IORING_OP_READ_FIXED:
6081 case IORING_OP_READ:
6082 kfree((void *)(unsigned long)req->rw.addr);
6084 case IORING_OP_RECVMSG:
6085 case IORING_OP_RECV:
6086 kfree(req->sr_msg.kbuf);
6091 if (req->flags & REQ_F_NEED_CLEANUP) {
6092 switch (req->opcode) {
6093 case IORING_OP_READV:
6094 case IORING_OP_READ_FIXED:
6095 case IORING_OP_READ:
6096 case IORING_OP_WRITEV:
6097 case IORING_OP_WRITE_FIXED:
6098 case IORING_OP_WRITE: {
6099 struct io_async_rw *io = req->async_data;
6101 kfree(io->free_iovec);
6104 case IORING_OP_RECVMSG:
6105 case IORING_OP_SENDMSG: {
6106 struct io_async_msghdr *io = req->async_data;
6108 kfree(io->free_iov);
6111 case IORING_OP_SPLICE:
6113 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6114 io_put_file(req->splice.file_in);
6116 case IORING_OP_OPENAT:
6117 case IORING_OP_OPENAT2:
6118 if (req->open.filename)
6119 putname(req->open.filename);
6121 case IORING_OP_RENAMEAT:
6122 putname(req->rename.oldpath);
6123 putname(req->rename.newpath);
6125 case IORING_OP_UNLINKAT:
6126 putname(req->unlink.filename);
6130 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6131 kfree(req->apoll->double_poll);
6135 if (req->flags & REQ_F_INFLIGHT) {
6136 struct io_uring_task *tctx = req->task->io_uring;
6138 atomic_dec(&tctx->inflight_tracked);
6140 if (req->flags & REQ_F_CREDS)
6141 put_cred(req->creds);
6143 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6146 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6148 struct io_ring_ctx *ctx = req->ctx;
6149 const struct cred *creds = NULL;
6152 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6153 creds = override_creds(req->creds);
6155 switch (req->opcode) {
6157 ret = io_nop(req, issue_flags);
6159 case IORING_OP_READV:
6160 case IORING_OP_READ_FIXED:
6161 case IORING_OP_READ:
6162 ret = io_read(req, issue_flags);
6164 case IORING_OP_WRITEV:
6165 case IORING_OP_WRITE_FIXED:
6166 case IORING_OP_WRITE:
6167 ret = io_write(req, issue_flags);
6169 case IORING_OP_FSYNC:
6170 ret = io_fsync(req, issue_flags);
6172 case IORING_OP_POLL_ADD:
6173 ret = io_poll_add(req, issue_flags);
6175 case IORING_OP_POLL_REMOVE:
6176 ret = io_poll_update(req, issue_flags);
6178 case IORING_OP_SYNC_FILE_RANGE:
6179 ret = io_sync_file_range(req, issue_flags);
6181 case IORING_OP_SENDMSG:
6182 ret = io_sendmsg(req, issue_flags);
6184 case IORING_OP_SEND:
6185 ret = io_send(req, issue_flags);
6187 case IORING_OP_RECVMSG:
6188 ret = io_recvmsg(req, issue_flags);
6190 case IORING_OP_RECV:
6191 ret = io_recv(req, issue_flags);
6193 case IORING_OP_TIMEOUT:
6194 ret = io_timeout(req, issue_flags);
6196 case IORING_OP_TIMEOUT_REMOVE:
6197 ret = io_timeout_remove(req, issue_flags);
6199 case IORING_OP_ACCEPT:
6200 ret = io_accept(req, issue_flags);
6202 case IORING_OP_CONNECT:
6203 ret = io_connect(req, issue_flags);
6205 case IORING_OP_ASYNC_CANCEL:
6206 ret = io_async_cancel(req, issue_flags);
6208 case IORING_OP_FALLOCATE:
6209 ret = io_fallocate(req, issue_flags);
6211 case IORING_OP_OPENAT:
6212 ret = io_openat(req, issue_flags);
6214 case IORING_OP_CLOSE:
6215 ret = io_close(req, issue_flags);
6217 case IORING_OP_FILES_UPDATE:
6218 ret = io_files_update(req, issue_flags);
6220 case IORING_OP_STATX:
6221 ret = io_statx(req, issue_flags);
6223 case IORING_OP_FADVISE:
6224 ret = io_fadvise(req, issue_flags);
6226 case IORING_OP_MADVISE:
6227 ret = io_madvise(req, issue_flags);
6229 case IORING_OP_OPENAT2:
6230 ret = io_openat2(req, issue_flags);
6232 case IORING_OP_EPOLL_CTL:
6233 ret = io_epoll_ctl(req, issue_flags);
6235 case IORING_OP_SPLICE:
6236 ret = io_splice(req, issue_flags);
6238 case IORING_OP_PROVIDE_BUFFERS:
6239 ret = io_provide_buffers(req, issue_flags);
6241 case IORING_OP_REMOVE_BUFFERS:
6242 ret = io_remove_buffers(req, issue_flags);
6245 ret = io_tee(req, issue_flags);
6247 case IORING_OP_SHUTDOWN:
6248 ret = io_shutdown(req, issue_flags);
6250 case IORING_OP_RENAMEAT:
6251 ret = io_renameat(req, issue_flags);
6253 case IORING_OP_UNLINKAT:
6254 ret = io_unlinkat(req, issue_flags);
6262 revert_creds(creds);
6265 /* If the op doesn't have a file, we're not polling for it */
6266 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6267 io_iopoll_req_issued(req);
6272 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6274 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6276 req = io_put_req_find_next(req);
6277 return req ? &req->work : NULL;
6280 static void io_wq_submit_work(struct io_wq_work *work)
6282 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6283 struct io_kiocb *timeout;
6286 timeout = io_prep_linked_timeout(req);
6288 io_queue_linked_timeout(timeout);
6290 if (work->flags & IO_WQ_WORK_CANCEL)
6295 ret = io_issue_sqe(req, 0);
6297 * We can get EAGAIN for polled IO even though we're
6298 * forcing a sync submission from here, since we can't
6299 * wait for request slots on the block side.
6307 /* avoid locking problems by failing it from a clean context */
6309 /* io-wq is going to take one down */
6311 io_req_task_queue_fail(req, ret);
6315 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6318 return &table->files[i];
6321 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6324 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6326 return (struct file *) (slot->file_ptr & FFS_MASK);
6329 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6331 unsigned long file_ptr = (unsigned long) file;
6333 if (__io_file_supports_nowait(file, READ))
6334 file_ptr |= FFS_ASYNC_READ;
6335 if (__io_file_supports_nowait(file, WRITE))
6336 file_ptr |= FFS_ASYNC_WRITE;
6337 if (S_ISREG(file_inode(file)->i_mode))
6338 file_ptr |= FFS_ISREG;
6339 file_slot->file_ptr = file_ptr;
6342 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6343 struct io_kiocb *req, int fd)
6346 unsigned long file_ptr;
6348 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6350 fd = array_index_nospec(fd, ctx->nr_user_files);
6351 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6352 file = (struct file *) (file_ptr & FFS_MASK);
6353 file_ptr &= ~FFS_MASK;
6354 /* mask in overlapping REQ_F and FFS bits */
6355 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6356 io_req_set_rsrc_node(req);
6360 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6361 struct io_kiocb *req, int fd)
6363 struct file *file = fget(fd);
6365 trace_io_uring_file_get(ctx, fd);
6367 /* we don't allow fixed io_uring files */
6368 if (file && unlikely(file->f_op == &io_uring_fops))
6369 io_req_track_inflight(req);
6373 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6374 struct io_kiocb *req, int fd, bool fixed)
6377 return io_file_get_fixed(ctx, req, fd);
6379 return io_file_get_normal(ctx, req, fd);
6382 static void io_req_task_link_timeout(struct io_kiocb *req)
6384 struct io_kiocb *prev = req->timeout.prev;
6385 struct io_ring_ctx *ctx = req->ctx;
6388 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6392 io_req_complete_post(req, -ETIME, 0);
6396 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6398 struct io_timeout_data *data = container_of(timer,
6399 struct io_timeout_data, timer);
6400 struct io_kiocb *prev, *req = data->req;
6401 struct io_ring_ctx *ctx = req->ctx;
6402 unsigned long flags;
6404 spin_lock_irqsave(&ctx->timeout_lock, flags);
6405 prev = req->timeout.head;
6406 req->timeout.head = NULL;
6409 * We don't expect the list to be empty, that will only happen if we
6410 * race with the completion of the linked work.
6413 io_remove_next_linked(prev);
6414 if (!req_ref_inc_not_zero(prev))
6417 req->timeout.prev = prev;
6418 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6420 req->io_task_work.func = io_req_task_link_timeout;
6421 io_req_task_work_add(req);
6422 return HRTIMER_NORESTART;
6425 static void io_queue_linked_timeout(struct io_kiocb *req)
6427 struct io_ring_ctx *ctx = req->ctx;
6429 spin_lock_irq(&ctx->timeout_lock);
6431 * If the back reference is NULL, then our linked request finished
6432 * before we got a chance to setup the timer
6434 if (req->timeout.head) {
6435 struct io_timeout_data *data = req->async_data;
6437 data->timer.function = io_link_timeout_fn;
6438 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6441 spin_unlock_irq(&ctx->timeout_lock);
6442 /* drop submission reference */
6446 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6448 struct io_kiocb *nxt = req->link;
6450 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6451 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6454 nxt->timeout.head = req;
6455 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6456 req->flags |= REQ_F_LINK_TIMEOUT;
6460 static void __io_queue_sqe(struct io_kiocb *req)
6461 __must_hold(&req->ctx->uring_lock)
6463 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6467 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6470 * We async punt it if the file wasn't marked NOWAIT, or if the file
6471 * doesn't support non-blocking read/write attempts
6474 /* drop submission reference */
6475 if (req->flags & REQ_F_COMPLETE_INLINE) {
6476 struct io_ring_ctx *ctx = req->ctx;
6477 struct io_submit_state *state = &ctx->submit_state;
6479 state->compl_reqs[state->compl_nr++] = req;
6480 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6481 io_submit_flush_completions(ctx);
6485 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6486 switch (io_arm_poll_handler(req)) {
6487 case IO_APOLL_READY:
6489 case IO_APOLL_ABORTED:
6491 * Queued up for async execution, worker will release
6492 * submit reference when the iocb is actually submitted.
6494 io_queue_async_work(req);
6498 io_req_complete_failed(req, ret);
6501 io_queue_linked_timeout(linked_timeout);
6504 static inline void io_queue_sqe(struct io_kiocb *req)
6505 __must_hold(&req->ctx->uring_lock)
6507 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6510 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6511 __io_queue_sqe(req);
6513 int ret = io_req_prep_async(req);
6516 io_req_complete_failed(req, ret);
6518 io_queue_async_work(req);
6523 * Check SQE restrictions (opcode and flags).
6525 * Returns 'true' if SQE is allowed, 'false' otherwise.
6527 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6528 struct io_kiocb *req,
6529 unsigned int sqe_flags)
6531 if (likely(!ctx->restricted))
6534 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6537 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6538 ctx->restrictions.sqe_flags_required)
6541 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6542 ctx->restrictions.sqe_flags_required))
6548 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6549 const struct io_uring_sqe *sqe)
6550 __must_hold(&ctx->uring_lock)
6552 struct io_submit_state *state;
6553 unsigned int sqe_flags;
6554 int personality, ret = 0;
6556 /* req is partially pre-initialised, see io_preinit_req() */
6557 req->opcode = READ_ONCE(sqe->opcode);
6558 /* same numerical values with corresponding REQ_F_*, safe to copy */
6559 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6560 req->user_data = READ_ONCE(sqe->user_data);
6562 req->fixed_rsrc_refs = NULL;
6563 /* one is dropped after submission, the other at completion */
6564 atomic_set(&req->refs, 2);
6565 req->task = current;
6567 /* enforce forwards compatibility on users */
6568 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6570 if (unlikely(req->opcode >= IORING_OP_LAST))
6572 if (!io_check_restriction(ctx, req, sqe_flags))
6575 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6576 !io_op_defs[req->opcode].buffer_select)
6578 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6579 ctx->drain_active = true;
6581 personality = READ_ONCE(sqe->personality);
6583 req->creds = xa_load(&ctx->personalities, personality);
6586 get_cred(req->creds);
6587 req->flags |= REQ_F_CREDS;
6589 state = &ctx->submit_state;
6592 * Plug now if we have more than 1 IO left after this, and the target
6593 * is potentially a read/write to block based storage.
6595 if (!state->plug_started && state->ios_left > 1 &&
6596 io_op_defs[req->opcode].plug) {
6597 blk_start_plug(&state->plug);
6598 state->plug_started = true;
6601 if (io_op_defs[req->opcode].needs_file) {
6602 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6603 (sqe_flags & IOSQE_FIXED_FILE));
6604 if (unlikely(!req->file))
6612 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6613 const struct io_uring_sqe *sqe)
6614 __must_hold(&ctx->uring_lock)
6616 struct io_submit_link *link = &ctx->submit_state.link;
6619 ret = io_init_req(ctx, req, sqe);
6620 if (unlikely(ret)) {
6623 /* fail even hard links since we don't submit */
6624 req_set_fail(link->head);
6625 io_req_complete_failed(link->head, -ECANCELED);
6628 io_req_complete_failed(req, ret);
6632 ret = io_req_prep(req, sqe);
6636 /* don't need @sqe from now on */
6637 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6639 ctx->flags & IORING_SETUP_SQPOLL);
6642 * If we already have a head request, queue this one for async
6643 * submittal once the head completes. If we don't have a head but
6644 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6645 * submitted sync once the chain is complete. If none of those
6646 * conditions are true (normal request), then just queue it.
6649 struct io_kiocb *head = link->head;
6651 ret = io_req_prep_async(req);
6654 trace_io_uring_link(ctx, req, head);
6655 link->last->link = req;
6658 /* last request of a link, enqueue the link */
6659 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6664 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6676 * Batched submission is done, ensure local IO is flushed out.
6678 static void io_submit_state_end(struct io_submit_state *state,
6679 struct io_ring_ctx *ctx)
6681 if (state->link.head)
6682 io_queue_sqe(state->link.head);
6683 if (state->compl_nr)
6684 io_submit_flush_completions(ctx);
6685 if (state->plug_started)
6686 blk_finish_plug(&state->plug);
6690 * Start submission side cache.
6692 static void io_submit_state_start(struct io_submit_state *state,
6693 unsigned int max_ios)
6695 state->plug_started = false;
6696 state->ios_left = max_ios;
6697 /* set only head, no need to init link_last in advance */
6698 state->link.head = NULL;
6701 static void io_commit_sqring(struct io_ring_ctx *ctx)
6703 struct io_rings *rings = ctx->rings;
6706 * Ensure any loads from the SQEs are done at this point,
6707 * since once we write the new head, the application could
6708 * write new data to them.
6710 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6714 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6715 * that is mapped by userspace. This means that care needs to be taken to
6716 * ensure that reads are stable, as we cannot rely on userspace always
6717 * being a good citizen. If members of the sqe are validated and then later
6718 * used, it's important that those reads are done through READ_ONCE() to
6719 * prevent a re-load down the line.
6721 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6723 unsigned head, mask = ctx->sq_entries - 1;
6724 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6727 * The cached sq head (or cq tail) serves two purposes:
6729 * 1) allows us to batch the cost of updating the user visible
6731 * 2) allows the kernel side to track the head on its own, even
6732 * though the application is the one updating it.
6734 head = READ_ONCE(ctx->sq_array[sq_idx]);
6735 if (likely(head < ctx->sq_entries))
6736 return &ctx->sq_sqes[head];
6738 /* drop invalid entries */
6740 WRITE_ONCE(ctx->rings->sq_dropped,
6741 READ_ONCE(ctx->rings->sq_dropped) + 1);
6745 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6746 __must_hold(&ctx->uring_lock)
6748 struct io_uring_task *tctx;
6751 /* make sure SQ entry isn't read before tail */
6752 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6753 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6756 tctx = current->io_uring;
6757 tctx->cached_refs -= nr;
6758 if (unlikely(tctx->cached_refs < 0)) {
6759 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6761 percpu_counter_add(&tctx->inflight, refill);
6762 refcount_add(refill, ¤t->usage);
6763 tctx->cached_refs += refill;
6765 io_submit_state_start(&ctx->submit_state, nr);
6767 while (submitted < nr) {
6768 const struct io_uring_sqe *sqe;
6769 struct io_kiocb *req;
6771 req = io_alloc_req(ctx);
6772 if (unlikely(!req)) {
6774 submitted = -EAGAIN;
6777 sqe = io_get_sqe(ctx);
6778 if (unlikely(!sqe)) {
6779 kmem_cache_free(req_cachep, req);
6782 /* will complete beyond this point, count as submitted */
6784 if (io_submit_sqe(ctx, req, sqe))
6788 if (unlikely(submitted != nr)) {
6789 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6790 int unused = nr - ref_used;
6792 current->io_uring->cached_refs += unused;
6793 percpu_ref_put_many(&ctx->refs, unused);
6796 io_submit_state_end(&ctx->submit_state, ctx);
6797 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6798 io_commit_sqring(ctx);
6803 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6805 return READ_ONCE(sqd->state);
6808 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6810 /* Tell userspace we may need a wakeup call */
6811 spin_lock(&ctx->completion_lock);
6812 WRITE_ONCE(ctx->rings->sq_flags,
6813 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6814 spin_unlock(&ctx->completion_lock);
6817 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6819 spin_lock(&ctx->completion_lock);
6820 WRITE_ONCE(ctx->rings->sq_flags,
6821 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6822 spin_unlock(&ctx->completion_lock);
6825 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6827 unsigned int to_submit;
6830 to_submit = io_sqring_entries(ctx);
6831 /* if we're handling multiple rings, cap submit size for fairness */
6832 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6833 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6835 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6836 unsigned nr_events = 0;
6837 const struct cred *creds = NULL;
6839 if (ctx->sq_creds != current_cred())
6840 creds = override_creds(ctx->sq_creds);
6842 mutex_lock(&ctx->uring_lock);
6843 if (!list_empty(&ctx->iopoll_list))
6844 io_do_iopoll(ctx, &nr_events, 0, true);
6847 * Don't submit if refs are dying, good for io_uring_register(),
6848 * but also it is relied upon by io_ring_exit_work()
6850 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6851 !(ctx->flags & IORING_SETUP_R_DISABLED))
6852 ret = io_submit_sqes(ctx, to_submit);
6853 mutex_unlock(&ctx->uring_lock);
6855 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6856 wake_up(&ctx->sqo_sq_wait);
6858 revert_creds(creds);
6864 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6866 struct io_ring_ctx *ctx;
6867 unsigned sq_thread_idle = 0;
6869 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6870 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6871 sqd->sq_thread_idle = sq_thread_idle;
6874 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6876 bool did_sig = false;
6877 struct ksignal ksig;
6879 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6880 signal_pending(current)) {
6881 mutex_unlock(&sqd->lock);
6882 if (signal_pending(current))
6883 did_sig = get_signal(&ksig);
6885 mutex_lock(&sqd->lock);
6887 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6890 static int io_sq_thread(void *data)
6892 struct io_sq_data *sqd = data;
6893 struct io_ring_ctx *ctx;
6894 unsigned long timeout = 0;
6895 char buf[TASK_COMM_LEN];
6898 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6899 set_task_comm(current, buf);
6901 if (sqd->sq_cpu != -1)
6902 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6904 set_cpus_allowed_ptr(current, cpu_online_mask);
6905 current->flags |= PF_NO_SETAFFINITY;
6907 mutex_lock(&sqd->lock);
6909 bool cap_entries, sqt_spin = false;
6911 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6912 if (io_sqd_handle_event(sqd))
6914 timeout = jiffies + sqd->sq_thread_idle;
6917 cap_entries = !list_is_singular(&sqd->ctx_list);
6918 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6919 int ret = __io_sq_thread(ctx, cap_entries);
6921 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6924 if (io_run_task_work())
6927 if (sqt_spin || !time_after(jiffies, timeout)) {
6930 timeout = jiffies + sqd->sq_thread_idle;
6934 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6935 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6936 bool needs_sched = true;
6938 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6939 io_ring_set_wakeup_flag(ctx);
6941 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6942 !list_empty_careful(&ctx->iopoll_list)) {
6943 needs_sched = false;
6946 if (io_sqring_entries(ctx)) {
6947 needs_sched = false;
6953 mutex_unlock(&sqd->lock);
6955 mutex_lock(&sqd->lock);
6957 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6958 io_ring_clear_wakeup_flag(ctx);
6961 finish_wait(&sqd->wait, &wait);
6962 timeout = jiffies + sqd->sq_thread_idle;
6965 io_uring_cancel_generic(true, sqd);
6967 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6968 io_ring_set_wakeup_flag(ctx);
6970 mutex_unlock(&sqd->lock);
6972 complete(&sqd->exited);
6976 struct io_wait_queue {
6977 struct wait_queue_entry wq;
6978 struct io_ring_ctx *ctx;
6980 unsigned nr_timeouts;
6983 static inline bool io_should_wake(struct io_wait_queue *iowq)
6985 struct io_ring_ctx *ctx = iowq->ctx;
6986 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
6989 * Wake up if we have enough events, or if a timeout occurred since we
6990 * started waiting. For timeouts, we always want to return to userspace,
6991 * regardless of event count.
6993 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6996 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6997 int wake_flags, void *key)
6999 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7003 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7004 * the task, and the next invocation will do it.
7006 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7007 return autoremove_wake_function(curr, mode, wake_flags, key);
7011 static int io_run_task_work_sig(void)
7013 if (io_run_task_work())
7015 if (!signal_pending(current))
7017 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7018 return -ERESTARTSYS;
7022 /* when returns >0, the caller should retry */
7023 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7024 struct io_wait_queue *iowq,
7025 signed long *timeout)
7029 /* make sure we run task_work before checking for signals */
7030 ret = io_run_task_work_sig();
7031 if (ret || io_should_wake(iowq))
7033 /* let the caller flush overflows, retry */
7034 if (test_bit(0, &ctx->check_cq_overflow))
7037 *timeout = schedule_timeout(*timeout);
7038 return !*timeout ? -ETIME : 1;
7042 * Wait until events become available, if we don't already have some. The
7043 * application must reap them itself, as they reside on the shared cq ring.
7045 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7046 const sigset_t __user *sig, size_t sigsz,
7047 struct __kernel_timespec __user *uts)
7049 struct io_wait_queue iowq;
7050 struct io_rings *rings = ctx->rings;
7051 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7055 io_cqring_overflow_flush(ctx);
7056 if (io_cqring_events(ctx) >= min_events)
7058 if (!io_run_task_work())
7063 #ifdef CONFIG_COMPAT
7064 if (in_compat_syscall())
7065 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7069 ret = set_user_sigmask(sig, sigsz);
7076 struct timespec64 ts;
7078 if (get_timespec64(&ts, uts))
7080 timeout = timespec64_to_jiffies(&ts);
7083 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7084 iowq.wq.private = current;
7085 INIT_LIST_HEAD(&iowq.wq.entry);
7087 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7088 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7090 trace_io_uring_cqring_wait(ctx, min_events);
7092 /* if we can't even flush overflow, don't wait for more */
7093 if (!io_cqring_overflow_flush(ctx)) {
7097 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7098 TASK_INTERRUPTIBLE);
7099 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7100 finish_wait(&ctx->cq_wait, &iowq.wq);
7104 restore_saved_sigmask_unless(ret == -EINTR);
7106 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7109 static void io_free_page_table(void **table, size_t size)
7111 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7113 for (i = 0; i < nr_tables; i++)
7118 static void **io_alloc_page_table(size_t size)
7120 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7121 size_t init_size = size;
7124 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7128 for (i = 0; i < nr_tables; i++) {
7129 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7131 table[i] = kzalloc(this_size, GFP_KERNEL);
7133 io_free_page_table(table, init_size);
7141 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7143 percpu_ref_exit(&ref_node->refs);
7147 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7149 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7150 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7151 unsigned long flags;
7152 bool first_add = false;
7154 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7157 while (!list_empty(&ctx->rsrc_ref_list)) {
7158 node = list_first_entry(&ctx->rsrc_ref_list,
7159 struct io_rsrc_node, node);
7160 /* recycle ref nodes in order */
7163 list_del(&node->node);
7164 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7166 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7169 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7172 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7174 struct io_rsrc_node *ref_node;
7176 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7180 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7185 INIT_LIST_HEAD(&ref_node->node);
7186 INIT_LIST_HEAD(&ref_node->rsrc_list);
7187 ref_node->done = false;
7191 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7192 struct io_rsrc_data *data_to_kill)
7194 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7195 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7198 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7200 rsrc_node->rsrc_data = data_to_kill;
7201 spin_lock_irq(&ctx->rsrc_ref_lock);
7202 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7203 spin_unlock_irq(&ctx->rsrc_ref_lock);
7205 atomic_inc(&data_to_kill->refs);
7206 percpu_ref_kill(&rsrc_node->refs);
7207 ctx->rsrc_node = NULL;
7210 if (!ctx->rsrc_node) {
7211 ctx->rsrc_node = ctx->rsrc_backup_node;
7212 ctx->rsrc_backup_node = NULL;
7216 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7218 if (ctx->rsrc_backup_node)
7220 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7221 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7224 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7228 /* As we may drop ->uring_lock, other task may have started quiesce */
7232 data->quiesce = true;
7234 ret = io_rsrc_node_switch_start(ctx);
7237 io_rsrc_node_switch(ctx, data);
7239 /* kill initial ref, already quiesced if zero */
7240 if (atomic_dec_and_test(&data->refs))
7242 mutex_unlock(&ctx->uring_lock);
7243 flush_delayed_work(&ctx->rsrc_put_work);
7244 ret = wait_for_completion_interruptible(&data->done);
7246 mutex_lock(&ctx->uring_lock);
7250 atomic_inc(&data->refs);
7251 /* wait for all works potentially completing data->done */
7252 flush_delayed_work(&ctx->rsrc_put_work);
7253 reinit_completion(&data->done);
7255 ret = io_run_task_work_sig();
7256 mutex_lock(&ctx->uring_lock);
7258 data->quiesce = false;
7263 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7265 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7266 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7268 return &data->tags[table_idx][off];
7271 static void io_rsrc_data_free(struct io_rsrc_data *data)
7273 size_t size = data->nr * sizeof(data->tags[0][0]);
7276 io_free_page_table((void **)data->tags, size);
7280 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7281 u64 __user *utags, unsigned nr,
7282 struct io_rsrc_data **pdata)
7284 struct io_rsrc_data *data;
7288 data = kzalloc(sizeof(*data), GFP_KERNEL);
7291 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7299 data->do_put = do_put;
7302 for (i = 0; i < nr; i++) {
7303 u64 *tag_slot = io_get_tag_slot(data, i);
7305 if (copy_from_user(tag_slot, &utags[i],
7311 atomic_set(&data->refs, 1);
7312 init_completion(&data->done);
7316 io_rsrc_data_free(data);
7320 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7322 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7323 return !!table->files;
7326 static void io_free_file_tables(struct io_file_table *table)
7328 kvfree(table->files);
7329 table->files = NULL;
7332 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7334 #if defined(CONFIG_UNIX)
7335 if (ctx->ring_sock) {
7336 struct sock *sock = ctx->ring_sock->sk;
7337 struct sk_buff *skb;
7339 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7345 for (i = 0; i < ctx->nr_user_files; i++) {
7348 file = io_file_from_index(ctx, i);
7353 io_free_file_tables(&ctx->file_table);
7354 io_rsrc_data_free(ctx->file_data);
7355 ctx->file_data = NULL;
7356 ctx->nr_user_files = 0;
7359 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7363 if (!ctx->file_data)
7365 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7367 __io_sqe_files_unregister(ctx);
7371 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7372 __releases(&sqd->lock)
7374 WARN_ON_ONCE(sqd->thread == current);
7377 * Do the dance but not conditional clear_bit() because it'd race with
7378 * other threads incrementing park_pending and setting the bit.
7380 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7381 if (atomic_dec_return(&sqd->park_pending))
7382 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7383 mutex_unlock(&sqd->lock);
7386 static void io_sq_thread_park(struct io_sq_data *sqd)
7387 __acquires(&sqd->lock)
7389 WARN_ON_ONCE(sqd->thread == current);
7391 atomic_inc(&sqd->park_pending);
7392 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7393 mutex_lock(&sqd->lock);
7395 wake_up_process(sqd->thread);
7398 static void io_sq_thread_stop(struct io_sq_data *sqd)
7400 WARN_ON_ONCE(sqd->thread == current);
7401 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7403 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7404 mutex_lock(&sqd->lock);
7406 wake_up_process(sqd->thread);
7407 mutex_unlock(&sqd->lock);
7408 wait_for_completion(&sqd->exited);
7411 static void io_put_sq_data(struct io_sq_data *sqd)
7413 if (refcount_dec_and_test(&sqd->refs)) {
7414 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7416 io_sq_thread_stop(sqd);
7421 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7423 struct io_sq_data *sqd = ctx->sq_data;
7426 io_sq_thread_park(sqd);
7427 list_del_init(&ctx->sqd_list);
7428 io_sqd_update_thread_idle(sqd);
7429 io_sq_thread_unpark(sqd);
7431 io_put_sq_data(sqd);
7432 ctx->sq_data = NULL;
7436 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7438 struct io_ring_ctx *ctx_attach;
7439 struct io_sq_data *sqd;
7442 f = fdget(p->wq_fd);
7444 return ERR_PTR(-ENXIO);
7445 if (f.file->f_op != &io_uring_fops) {
7447 return ERR_PTR(-EINVAL);
7450 ctx_attach = f.file->private_data;
7451 sqd = ctx_attach->sq_data;
7454 return ERR_PTR(-EINVAL);
7456 if (sqd->task_tgid != current->tgid) {
7458 return ERR_PTR(-EPERM);
7461 refcount_inc(&sqd->refs);
7466 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7469 struct io_sq_data *sqd;
7472 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7473 sqd = io_attach_sq_data(p);
7478 /* fall through for EPERM case, setup new sqd/task */
7479 if (PTR_ERR(sqd) != -EPERM)
7483 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7485 return ERR_PTR(-ENOMEM);
7487 atomic_set(&sqd->park_pending, 0);
7488 refcount_set(&sqd->refs, 1);
7489 INIT_LIST_HEAD(&sqd->ctx_list);
7490 mutex_init(&sqd->lock);
7491 init_waitqueue_head(&sqd->wait);
7492 init_completion(&sqd->exited);
7496 #if defined(CONFIG_UNIX)
7498 * Ensure the UNIX gc is aware of our file set, so we are certain that
7499 * the io_uring can be safely unregistered on process exit, even if we have
7500 * loops in the file referencing.
7502 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7504 struct sock *sk = ctx->ring_sock->sk;
7505 struct scm_fp_list *fpl;
7506 struct sk_buff *skb;
7509 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7513 skb = alloc_skb(0, GFP_KERNEL);
7522 fpl->user = get_uid(current_user());
7523 for (i = 0; i < nr; i++) {
7524 struct file *file = io_file_from_index(ctx, i + offset);
7528 fpl->fp[nr_files] = get_file(file);
7529 unix_inflight(fpl->user, fpl->fp[nr_files]);
7534 fpl->max = SCM_MAX_FD;
7535 fpl->count = nr_files;
7536 UNIXCB(skb).fp = fpl;
7537 skb->destructor = unix_destruct_scm;
7538 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7539 skb_queue_head(&sk->sk_receive_queue, skb);
7541 for (i = 0; i < nr_files; i++)
7552 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7553 * causes regular reference counting to break down. We rely on the UNIX
7554 * garbage collection to take care of this problem for us.
7556 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7558 unsigned left, total;
7562 left = ctx->nr_user_files;
7564 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7566 ret = __io_sqe_files_scm(ctx, this_files, total);
7570 total += this_files;
7576 while (total < ctx->nr_user_files) {
7577 struct file *file = io_file_from_index(ctx, total);
7587 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7593 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7595 struct file *file = prsrc->file;
7596 #if defined(CONFIG_UNIX)
7597 struct sock *sock = ctx->ring_sock->sk;
7598 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7599 struct sk_buff *skb;
7602 __skb_queue_head_init(&list);
7605 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7606 * remove this entry and rearrange the file array.
7608 skb = skb_dequeue(head);
7610 struct scm_fp_list *fp;
7612 fp = UNIXCB(skb).fp;
7613 for (i = 0; i < fp->count; i++) {
7616 if (fp->fp[i] != file)
7619 unix_notinflight(fp->user, fp->fp[i]);
7620 left = fp->count - 1 - i;
7622 memmove(&fp->fp[i], &fp->fp[i + 1],
7623 left * sizeof(struct file *));
7630 __skb_queue_tail(&list, skb);
7640 __skb_queue_tail(&list, skb);
7642 skb = skb_dequeue(head);
7645 if (skb_peek(&list)) {
7646 spin_lock_irq(&head->lock);
7647 while ((skb = __skb_dequeue(&list)) != NULL)
7648 __skb_queue_tail(head, skb);
7649 spin_unlock_irq(&head->lock);
7656 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7658 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7659 struct io_ring_ctx *ctx = rsrc_data->ctx;
7660 struct io_rsrc_put *prsrc, *tmp;
7662 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7663 list_del(&prsrc->list);
7666 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7668 io_ring_submit_lock(ctx, lock_ring);
7669 spin_lock(&ctx->completion_lock);
7670 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7672 io_commit_cqring(ctx);
7673 spin_unlock(&ctx->completion_lock);
7674 io_cqring_ev_posted(ctx);
7675 io_ring_submit_unlock(ctx, lock_ring);
7678 rsrc_data->do_put(ctx, prsrc);
7682 io_rsrc_node_destroy(ref_node);
7683 if (atomic_dec_and_test(&rsrc_data->refs))
7684 complete(&rsrc_data->done);
7687 static void io_rsrc_put_work(struct work_struct *work)
7689 struct io_ring_ctx *ctx;
7690 struct llist_node *node;
7692 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7693 node = llist_del_all(&ctx->rsrc_put_llist);
7696 struct io_rsrc_node *ref_node;
7697 struct llist_node *next = node->next;
7699 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7700 __io_rsrc_put_work(ref_node);
7705 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7706 unsigned nr_args, u64 __user *tags)
7708 __s32 __user *fds = (__s32 __user *) arg;
7717 if (nr_args > IORING_MAX_FIXED_FILES)
7719 ret = io_rsrc_node_switch_start(ctx);
7722 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7728 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7731 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7732 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7736 /* allow sparse sets */
7739 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7746 if (unlikely(!file))
7750 * Don't allow io_uring instances to be registered. If UNIX
7751 * isn't enabled, then this causes a reference cycle and this
7752 * instance can never get freed. If UNIX is enabled we'll
7753 * handle it just fine, but there's still no point in allowing
7754 * a ring fd as it doesn't support regular read/write anyway.
7756 if (file->f_op == &io_uring_fops) {
7760 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7763 ret = io_sqe_files_scm(ctx);
7765 __io_sqe_files_unregister(ctx);
7769 io_rsrc_node_switch(ctx, NULL);
7772 for (i = 0; i < ctx->nr_user_files; i++) {
7773 file = io_file_from_index(ctx, i);
7777 io_free_file_tables(&ctx->file_table);
7778 ctx->nr_user_files = 0;
7780 io_rsrc_data_free(ctx->file_data);
7781 ctx->file_data = NULL;
7785 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7788 #if defined(CONFIG_UNIX)
7789 struct sock *sock = ctx->ring_sock->sk;
7790 struct sk_buff_head *head = &sock->sk_receive_queue;
7791 struct sk_buff *skb;
7794 * See if we can merge this file into an existing skb SCM_RIGHTS
7795 * file set. If there's no room, fall back to allocating a new skb
7796 * and filling it in.
7798 spin_lock_irq(&head->lock);
7799 skb = skb_peek(head);
7801 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7803 if (fpl->count < SCM_MAX_FD) {
7804 __skb_unlink(skb, head);
7805 spin_unlock_irq(&head->lock);
7806 fpl->fp[fpl->count] = get_file(file);
7807 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7809 spin_lock_irq(&head->lock);
7810 __skb_queue_head(head, skb);
7815 spin_unlock_irq(&head->lock);
7822 return __io_sqe_files_scm(ctx, 1, index);
7828 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7829 struct io_rsrc_node *node, void *rsrc)
7831 struct io_rsrc_put *prsrc;
7833 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7837 prsrc->tag = *io_get_tag_slot(data, idx);
7839 list_add(&prsrc->list, &node->rsrc_list);
7843 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7844 struct io_uring_rsrc_update2 *up,
7847 u64 __user *tags = u64_to_user_ptr(up->tags);
7848 __s32 __user *fds = u64_to_user_ptr(up->data);
7849 struct io_rsrc_data *data = ctx->file_data;
7850 struct io_fixed_file *file_slot;
7854 bool needs_switch = false;
7856 if (!ctx->file_data)
7858 if (up->offset + nr_args > ctx->nr_user_files)
7861 for (done = 0; done < nr_args; done++) {
7864 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7865 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7869 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7873 if (fd == IORING_REGISTER_FILES_SKIP)
7876 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7877 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7879 if (file_slot->file_ptr) {
7880 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7881 err = io_queue_rsrc_removal(data, up->offset + done,
7882 ctx->rsrc_node, file);
7885 file_slot->file_ptr = 0;
7886 needs_switch = true;
7895 * Don't allow io_uring instances to be registered. If
7896 * UNIX isn't enabled, then this causes a reference
7897 * cycle and this instance can never get freed. If UNIX
7898 * is enabled we'll handle it just fine, but there's
7899 * still no point in allowing a ring fd as it doesn't
7900 * support regular read/write anyway.
7902 if (file->f_op == &io_uring_fops) {
7907 *io_get_tag_slot(data, up->offset + done) = tag;
7908 io_fixed_file_set(file_slot, file);
7909 err = io_sqe_file_register(ctx, file, i);
7911 file_slot->file_ptr = 0;
7919 io_rsrc_node_switch(ctx, data);
7920 return done ? done : err;
7923 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7924 struct task_struct *task)
7926 struct io_wq_hash *hash;
7927 struct io_wq_data data;
7928 unsigned int concurrency;
7930 mutex_lock(&ctx->uring_lock);
7931 hash = ctx->hash_map;
7933 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7935 mutex_unlock(&ctx->uring_lock);
7936 return ERR_PTR(-ENOMEM);
7938 refcount_set(&hash->refs, 1);
7939 init_waitqueue_head(&hash->wait);
7940 ctx->hash_map = hash;
7942 mutex_unlock(&ctx->uring_lock);
7946 data.free_work = io_wq_free_work;
7947 data.do_work = io_wq_submit_work;
7949 /* Do QD, or 4 * CPUS, whatever is smallest */
7950 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7952 return io_wq_create(concurrency, &data);
7955 static int io_uring_alloc_task_context(struct task_struct *task,
7956 struct io_ring_ctx *ctx)
7958 struct io_uring_task *tctx;
7961 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7962 if (unlikely(!tctx))
7965 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7966 if (unlikely(ret)) {
7971 tctx->io_wq = io_init_wq_offload(ctx, task);
7972 if (IS_ERR(tctx->io_wq)) {
7973 ret = PTR_ERR(tctx->io_wq);
7974 percpu_counter_destroy(&tctx->inflight);
7980 init_waitqueue_head(&tctx->wait);
7981 atomic_set(&tctx->in_idle, 0);
7982 atomic_set(&tctx->inflight_tracked, 0);
7983 task->io_uring = tctx;
7984 spin_lock_init(&tctx->task_lock);
7985 INIT_WQ_LIST(&tctx->task_list);
7986 init_task_work(&tctx->task_work, tctx_task_work);
7990 void __io_uring_free(struct task_struct *tsk)
7992 struct io_uring_task *tctx = tsk->io_uring;
7994 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7995 WARN_ON_ONCE(tctx->io_wq);
7996 WARN_ON_ONCE(tctx->cached_refs);
7998 percpu_counter_destroy(&tctx->inflight);
8000 tsk->io_uring = NULL;
8003 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8004 struct io_uring_params *p)
8008 /* Retain compatibility with failing for an invalid attach attempt */
8009 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8010 IORING_SETUP_ATTACH_WQ) {
8013 f = fdget(p->wq_fd);
8016 if (f.file->f_op != &io_uring_fops) {
8022 if (ctx->flags & IORING_SETUP_SQPOLL) {
8023 struct task_struct *tsk;
8024 struct io_sq_data *sqd;
8027 sqd = io_get_sq_data(p, &attached);
8033 ctx->sq_creds = get_current_cred();
8035 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8036 if (!ctx->sq_thread_idle)
8037 ctx->sq_thread_idle = HZ;
8039 io_sq_thread_park(sqd);
8040 list_add(&ctx->sqd_list, &sqd->ctx_list);
8041 io_sqd_update_thread_idle(sqd);
8042 /* don't attach to a dying SQPOLL thread, would be racy */
8043 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8044 io_sq_thread_unpark(sqd);
8051 if (p->flags & IORING_SETUP_SQ_AFF) {
8052 int cpu = p->sq_thread_cpu;
8055 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8062 sqd->task_pid = current->pid;
8063 sqd->task_tgid = current->tgid;
8064 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8071 ret = io_uring_alloc_task_context(tsk, ctx);
8072 wake_up_new_task(tsk);
8075 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8076 /* Can't have SQ_AFF without SQPOLL */
8083 complete(&ctx->sq_data->exited);
8085 io_sq_thread_finish(ctx);
8089 static inline void __io_unaccount_mem(struct user_struct *user,
8090 unsigned long nr_pages)
8092 atomic_long_sub(nr_pages, &user->locked_vm);
8095 static inline int __io_account_mem(struct user_struct *user,
8096 unsigned long nr_pages)
8098 unsigned long page_limit, cur_pages, new_pages;
8100 /* Don't allow more pages than we can safely lock */
8101 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8104 cur_pages = atomic_long_read(&user->locked_vm);
8105 new_pages = cur_pages + nr_pages;
8106 if (new_pages > page_limit)
8108 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8109 new_pages) != cur_pages);
8114 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8117 __io_unaccount_mem(ctx->user, nr_pages);
8119 if (ctx->mm_account)
8120 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8123 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8128 ret = __io_account_mem(ctx->user, nr_pages);
8133 if (ctx->mm_account)
8134 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8139 static void io_mem_free(void *ptr)
8146 page = virt_to_head_page(ptr);
8147 if (put_page_testzero(page))
8148 free_compound_page(page);
8151 static void *io_mem_alloc(size_t size)
8153 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8154 __GFP_NORETRY | __GFP_ACCOUNT;
8156 return (void *) __get_free_pages(gfp_flags, get_order(size));
8159 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8162 struct io_rings *rings;
8163 size_t off, sq_array_size;
8165 off = struct_size(rings, cqes, cq_entries);
8166 if (off == SIZE_MAX)
8170 off = ALIGN(off, SMP_CACHE_BYTES);
8178 sq_array_size = array_size(sizeof(u32), sq_entries);
8179 if (sq_array_size == SIZE_MAX)
8182 if (check_add_overflow(off, sq_array_size, &off))
8188 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8190 struct io_mapped_ubuf *imu = *slot;
8193 if (imu != ctx->dummy_ubuf) {
8194 for (i = 0; i < imu->nr_bvecs; i++)
8195 unpin_user_page(imu->bvec[i].bv_page);
8196 if (imu->acct_pages)
8197 io_unaccount_mem(ctx, imu->acct_pages);
8203 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8205 io_buffer_unmap(ctx, &prsrc->buf);
8209 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8213 for (i = 0; i < ctx->nr_user_bufs; i++)
8214 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8215 kfree(ctx->user_bufs);
8216 io_rsrc_data_free(ctx->buf_data);
8217 ctx->user_bufs = NULL;
8218 ctx->buf_data = NULL;
8219 ctx->nr_user_bufs = 0;
8222 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8229 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8231 __io_sqe_buffers_unregister(ctx);
8235 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8236 void __user *arg, unsigned index)
8238 struct iovec __user *src;
8240 #ifdef CONFIG_COMPAT
8242 struct compat_iovec __user *ciovs;
8243 struct compat_iovec ciov;
8245 ciovs = (struct compat_iovec __user *) arg;
8246 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8249 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8250 dst->iov_len = ciov.iov_len;
8254 src = (struct iovec __user *) arg;
8255 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8261 * Not super efficient, but this is just a registration time. And we do cache
8262 * the last compound head, so generally we'll only do a full search if we don't
8265 * We check if the given compound head page has already been accounted, to
8266 * avoid double accounting it. This allows us to account the full size of the
8267 * page, not just the constituent pages of a huge page.
8269 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8270 int nr_pages, struct page *hpage)
8274 /* check current page array */
8275 for (i = 0; i < nr_pages; i++) {
8276 if (!PageCompound(pages[i]))
8278 if (compound_head(pages[i]) == hpage)
8282 /* check previously registered pages */
8283 for (i = 0; i < ctx->nr_user_bufs; i++) {
8284 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8286 for (j = 0; j < imu->nr_bvecs; j++) {
8287 if (!PageCompound(imu->bvec[j].bv_page))
8289 if (compound_head(imu->bvec[j].bv_page) == hpage)
8297 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8298 int nr_pages, struct io_mapped_ubuf *imu,
8299 struct page **last_hpage)
8303 imu->acct_pages = 0;
8304 for (i = 0; i < nr_pages; i++) {
8305 if (!PageCompound(pages[i])) {
8310 hpage = compound_head(pages[i]);
8311 if (hpage == *last_hpage)
8313 *last_hpage = hpage;
8314 if (headpage_already_acct(ctx, pages, i, hpage))
8316 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8320 if (!imu->acct_pages)
8323 ret = io_account_mem(ctx, imu->acct_pages);
8325 imu->acct_pages = 0;
8329 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8330 struct io_mapped_ubuf **pimu,
8331 struct page **last_hpage)
8333 struct io_mapped_ubuf *imu = NULL;
8334 struct vm_area_struct **vmas = NULL;
8335 struct page **pages = NULL;
8336 unsigned long off, start, end, ubuf;
8338 int ret, pret, nr_pages, i;
8340 if (!iov->iov_base) {
8341 *pimu = ctx->dummy_ubuf;
8345 ubuf = (unsigned long) iov->iov_base;
8346 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8347 start = ubuf >> PAGE_SHIFT;
8348 nr_pages = end - start;
8353 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8357 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8362 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8367 mmap_read_lock(current->mm);
8368 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8370 if (pret == nr_pages) {
8371 /* don't support file backed memory */
8372 for (i = 0; i < nr_pages; i++) {
8373 struct vm_area_struct *vma = vmas[i];
8375 if (vma_is_shmem(vma))
8378 !is_file_hugepages(vma->vm_file)) {
8384 ret = pret < 0 ? pret : -EFAULT;
8386 mmap_read_unlock(current->mm);
8389 * if we did partial map, or found file backed vmas,
8390 * release any pages we did get
8393 unpin_user_pages(pages, pret);
8397 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8399 unpin_user_pages(pages, pret);
8403 off = ubuf & ~PAGE_MASK;
8404 size = iov->iov_len;
8405 for (i = 0; i < nr_pages; i++) {
8408 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8409 imu->bvec[i].bv_page = pages[i];
8410 imu->bvec[i].bv_len = vec_len;
8411 imu->bvec[i].bv_offset = off;
8415 /* store original address for later verification */
8417 imu->ubuf_end = ubuf + iov->iov_len;
8418 imu->nr_bvecs = nr_pages;
8429 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8431 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8432 return ctx->user_bufs ? 0 : -ENOMEM;
8435 static int io_buffer_validate(struct iovec *iov)
8437 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8440 * Don't impose further limits on the size and buffer
8441 * constraints here, we'll -EINVAL later when IO is
8442 * submitted if they are wrong.
8445 return iov->iov_len ? -EFAULT : 0;
8449 /* arbitrary limit, but we need something */
8450 if (iov->iov_len > SZ_1G)
8453 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8459 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8460 unsigned int nr_args, u64 __user *tags)
8462 struct page *last_hpage = NULL;
8463 struct io_rsrc_data *data;
8469 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8471 ret = io_rsrc_node_switch_start(ctx);
8474 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8477 ret = io_buffers_map_alloc(ctx, nr_args);
8479 io_rsrc_data_free(data);
8483 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8484 ret = io_copy_iov(ctx, &iov, arg, i);
8487 ret = io_buffer_validate(&iov);
8490 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8495 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8501 WARN_ON_ONCE(ctx->buf_data);
8503 ctx->buf_data = data;
8505 __io_sqe_buffers_unregister(ctx);
8507 io_rsrc_node_switch(ctx, NULL);
8511 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8512 struct io_uring_rsrc_update2 *up,
8513 unsigned int nr_args)
8515 u64 __user *tags = u64_to_user_ptr(up->tags);
8516 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8517 struct page *last_hpage = NULL;
8518 bool needs_switch = false;
8524 if (up->offset + nr_args > ctx->nr_user_bufs)
8527 for (done = 0; done < nr_args; done++) {
8528 struct io_mapped_ubuf *imu;
8529 int offset = up->offset + done;
8532 err = io_copy_iov(ctx, &iov, iovs, done);
8535 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8539 err = io_buffer_validate(&iov);
8542 if (!iov.iov_base && tag) {
8546 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8550 i = array_index_nospec(offset, ctx->nr_user_bufs);
8551 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8552 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8553 ctx->rsrc_node, ctx->user_bufs[i]);
8554 if (unlikely(err)) {
8555 io_buffer_unmap(ctx, &imu);
8558 ctx->user_bufs[i] = NULL;
8559 needs_switch = true;
8562 ctx->user_bufs[i] = imu;
8563 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8567 io_rsrc_node_switch(ctx, ctx->buf_data);
8568 return done ? done : err;
8571 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8573 __s32 __user *fds = arg;
8579 if (copy_from_user(&fd, fds, sizeof(*fds)))
8582 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8583 if (IS_ERR(ctx->cq_ev_fd)) {
8584 int ret = PTR_ERR(ctx->cq_ev_fd);
8586 ctx->cq_ev_fd = NULL;
8593 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8595 if (ctx->cq_ev_fd) {
8596 eventfd_ctx_put(ctx->cq_ev_fd);
8597 ctx->cq_ev_fd = NULL;
8604 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8606 struct io_buffer *buf;
8607 unsigned long index;
8609 xa_for_each(&ctx->io_buffers, index, buf)
8610 __io_remove_buffers(ctx, buf, index, -1U);
8613 static void io_req_cache_free(struct list_head *list)
8615 struct io_kiocb *req, *nxt;
8617 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8618 list_del(&req->inflight_entry);
8619 kmem_cache_free(req_cachep, req);
8623 static void io_req_caches_free(struct io_ring_ctx *ctx)
8625 struct io_submit_state *state = &ctx->submit_state;
8627 mutex_lock(&ctx->uring_lock);
8629 if (state->free_reqs) {
8630 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8631 state->free_reqs = 0;
8634 io_flush_cached_locked_reqs(ctx, state);
8635 io_req_cache_free(&state->free_list);
8636 mutex_unlock(&ctx->uring_lock);
8639 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8641 if (data && !atomic_dec_and_test(&data->refs))
8642 wait_for_completion(&data->done);
8645 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8647 io_sq_thread_finish(ctx);
8649 if (ctx->mm_account) {
8650 mmdrop(ctx->mm_account);
8651 ctx->mm_account = NULL;
8654 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8655 io_wait_rsrc_data(ctx->buf_data);
8656 io_wait_rsrc_data(ctx->file_data);
8658 mutex_lock(&ctx->uring_lock);
8660 __io_sqe_buffers_unregister(ctx);
8662 __io_sqe_files_unregister(ctx);
8664 __io_cqring_overflow_flush(ctx, true);
8665 mutex_unlock(&ctx->uring_lock);
8666 io_eventfd_unregister(ctx);
8667 io_destroy_buffers(ctx);
8669 put_cred(ctx->sq_creds);
8671 /* there are no registered resources left, nobody uses it */
8673 io_rsrc_node_destroy(ctx->rsrc_node);
8674 if (ctx->rsrc_backup_node)
8675 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8676 flush_delayed_work(&ctx->rsrc_put_work);
8678 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8679 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8681 #if defined(CONFIG_UNIX)
8682 if (ctx->ring_sock) {
8683 ctx->ring_sock->file = NULL; /* so that iput() is called */
8684 sock_release(ctx->ring_sock);
8688 io_mem_free(ctx->rings);
8689 io_mem_free(ctx->sq_sqes);
8691 percpu_ref_exit(&ctx->refs);
8692 free_uid(ctx->user);
8693 io_req_caches_free(ctx);
8695 io_wq_put_hash(ctx->hash_map);
8696 kfree(ctx->cancel_hash);
8697 kfree(ctx->dummy_ubuf);
8701 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8703 struct io_ring_ctx *ctx = file->private_data;
8706 poll_wait(file, &ctx->poll_wait, wait);
8708 * synchronizes with barrier from wq_has_sleeper call in
8712 if (!io_sqring_full(ctx))
8713 mask |= EPOLLOUT | EPOLLWRNORM;
8716 * Don't flush cqring overflow list here, just do a simple check.
8717 * Otherwise there could possible be ABBA deadlock:
8720 * lock(&ctx->uring_lock);
8722 * lock(&ctx->uring_lock);
8725 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8726 * pushs them to do the flush.
8728 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8729 mask |= EPOLLIN | EPOLLRDNORM;
8734 static int io_uring_fasync(int fd, struct file *file, int on)
8736 struct io_ring_ctx *ctx = file->private_data;
8738 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8741 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8743 const struct cred *creds;
8745 creds = xa_erase(&ctx->personalities, id);
8754 struct io_tctx_exit {
8755 struct callback_head task_work;
8756 struct completion completion;
8757 struct io_ring_ctx *ctx;
8760 static void io_tctx_exit_cb(struct callback_head *cb)
8762 struct io_uring_task *tctx = current->io_uring;
8763 struct io_tctx_exit *work;
8765 work = container_of(cb, struct io_tctx_exit, task_work);
8767 * When @in_idle, we're in cancellation and it's racy to remove the
8768 * node. It'll be removed by the end of cancellation, just ignore it.
8770 if (!atomic_read(&tctx->in_idle))
8771 io_uring_del_tctx_node((unsigned long)work->ctx);
8772 complete(&work->completion);
8775 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8779 return req->ctx == data;
8782 static void io_ring_exit_work(struct work_struct *work)
8784 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8785 unsigned long timeout = jiffies + HZ * 60 * 5;
8786 unsigned long interval = HZ / 20;
8787 struct io_tctx_exit exit;
8788 struct io_tctx_node *node;
8792 * If we're doing polled IO and end up having requests being
8793 * submitted async (out-of-line), then completions can come in while
8794 * we're waiting for refs to drop. We need to reap these manually,
8795 * as nobody else will be looking for them.
8798 io_uring_try_cancel_requests(ctx, NULL, true);
8800 struct io_sq_data *sqd = ctx->sq_data;
8801 struct task_struct *tsk;
8803 io_sq_thread_park(sqd);
8805 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8806 io_wq_cancel_cb(tsk->io_uring->io_wq,
8807 io_cancel_ctx_cb, ctx, true);
8808 io_sq_thread_unpark(sqd);
8811 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8812 /* there is little hope left, don't run it too often */
8815 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8817 init_completion(&exit.completion);
8818 init_task_work(&exit.task_work, io_tctx_exit_cb);
8821 * Some may use context even when all refs and requests have been put,
8822 * and they are free to do so while still holding uring_lock or
8823 * completion_lock, see io_req_task_submit(). Apart from other work,
8824 * this lock/unlock section also waits them to finish.
8826 mutex_lock(&ctx->uring_lock);
8827 while (!list_empty(&ctx->tctx_list)) {
8828 WARN_ON_ONCE(time_after(jiffies, timeout));
8830 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8832 /* don't spin on a single task if cancellation failed */
8833 list_rotate_left(&ctx->tctx_list);
8834 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8835 if (WARN_ON_ONCE(ret))
8837 wake_up_process(node->task);
8839 mutex_unlock(&ctx->uring_lock);
8840 wait_for_completion(&exit.completion);
8841 mutex_lock(&ctx->uring_lock);
8843 mutex_unlock(&ctx->uring_lock);
8844 spin_lock(&ctx->completion_lock);
8845 spin_unlock(&ctx->completion_lock);
8847 io_ring_ctx_free(ctx);
8850 /* Returns true if we found and killed one or more timeouts */
8851 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8854 struct io_kiocb *req, *tmp;
8857 spin_lock(&ctx->completion_lock);
8858 spin_lock_irq(&ctx->timeout_lock);
8859 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8860 if (io_match_task(req, tsk, cancel_all)) {
8861 io_kill_timeout(req, -ECANCELED);
8865 spin_unlock_irq(&ctx->timeout_lock);
8867 io_commit_cqring(ctx);
8868 spin_unlock(&ctx->completion_lock);
8870 io_cqring_ev_posted(ctx);
8871 return canceled != 0;
8874 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8876 unsigned long index;
8877 struct creds *creds;
8879 mutex_lock(&ctx->uring_lock);
8880 percpu_ref_kill(&ctx->refs);
8882 __io_cqring_overflow_flush(ctx, true);
8883 xa_for_each(&ctx->personalities, index, creds)
8884 io_unregister_personality(ctx, index);
8885 mutex_unlock(&ctx->uring_lock);
8887 io_kill_timeouts(ctx, NULL, true);
8888 io_poll_remove_all(ctx, NULL, true);
8890 /* if we failed setting up the ctx, we might not have any rings */
8891 io_iopoll_try_reap_events(ctx);
8893 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8895 * Use system_unbound_wq to avoid spawning tons of event kworkers
8896 * if we're exiting a ton of rings at the same time. It just adds
8897 * noise and overhead, there's no discernable change in runtime
8898 * over using system_wq.
8900 queue_work(system_unbound_wq, &ctx->exit_work);
8903 static int io_uring_release(struct inode *inode, struct file *file)
8905 struct io_ring_ctx *ctx = file->private_data;
8907 file->private_data = NULL;
8908 io_ring_ctx_wait_and_kill(ctx);
8912 struct io_task_cancel {
8913 struct task_struct *task;
8917 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8919 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8920 struct io_task_cancel *cancel = data;
8923 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8924 struct io_ring_ctx *ctx = req->ctx;
8926 /* protect against races with linked timeouts */
8927 spin_lock(&ctx->completion_lock);
8928 ret = io_match_task(req, cancel->task, cancel->all);
8929 spin_unlock(&ctx->completion_lock);
8931 ret = io_match_task(req, cancel->task, cancel->all);
8936 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8937 struct task_struct *task, bool cancel_all)
8939 struct io_defer_entry *de;
8942 spin_lock(&ctx->completion_lock);
8943 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8944 if (io_match_task(de->req, task, cancel_all)) {
8945 list_cut_position(&list, &ctx->defer_list, &de->list);
8949 spin_unlock(&ctx->completion_lock);
8950 if (list_empty(&list))
8953 while (!list_empty(&list)) {
8954 de = list_first_entry(&list, struct io_defer_entry, list);
8955 list_del_init(&de->list);
8956 io_req_complete_failed(de->req, -ECANCELED);
8962 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8964 struct io_tctx_node *node;
8965 enum io_wq_cancel cret;
8968 mutex_lock(&ctx->uring_lock);
8969 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8970 struct io_uring_task *tctx = node->task->io_uring;
8973 * io_wq will stay alive while we hold uring_lock, because it's
8974 * killed after ctx nodes, which requires to take the lock.
8976 if (!tctx || !tctx->io_wq)
8978 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8979 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8981 mutex_unlock(&ctx->uring_lock);
8986 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8987 struct task_struct *task,
8990 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8991 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8994 enum io_wq_cancel cret;
8998 ret |= io_uring_try_cancel_iowq(ctx);
8999 } else if (tctx && tctx->io_wq) {
9001 * Cancels requests of all rings, not only @ctx, but
9002 * it's fine as the task is in exit/exec.
9004 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9006 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9009 /* SQPOLL thread does its own polling */
9010 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9011 (ctx->sq_data && ctx->sq_data->thread == current)) {
9012 while (!list_empty_careful(&ctx->iopoll_list)) {
9013 io_iopoll_try_reap_events(ctx);
9018 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9019 ret |= io_poll_remove_all(ctx, task, cancel_all);
9020 ret |= io_kill_timeouts(ctx, task, cancel_all);
9022 ret |= io_run_task_work();
9029 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9031 struct io_uring_task *tctx = current->io_uring;
9032 struct io_tctx_node *node;
9035 if (unlikely(!tctx)) {
9036 ret = io_uring_alloc_task_context(current, ctx);
9039 tctx = current->io_uring;
9041 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9042 node = kmalloc(sizeof(*node), GFP_KERNEL);
9046 node->task = current;
9048 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9055 mutex_lock(&ctx->uring_lock);
9056 list_add(&node->ctx_node, &ctx->tctx_list);
9057 mutex_unlock(&ctx->uring_lock);
9064 * Note that this task has used io_uring. We use it for cancelation purposes.
9066 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9068 struct io_uring_task *tctx = current->io_uring;
9070 if (likely(tctx && tctx->last == ctx))
9072 return __io_uring_add_tctx_node(ctx);
9076 * Remove this io_uring_file -> task mapping.
9078 static void io_uring_del_tctx_node(unsigned long index)
9080 struct io_uring_task *tctx = current->io_uring;
9081 struct io_tctx_node *node;
9085 node = xa_erase(&tctx->xa, index);
9089 WARN_ON_ONCE(current != node->task);
9090 WARN_ON_ONCE(list_empty(&node->ctx_node));
9092 mutex_lock(&node->ctx->uring_lock);
9093 list_del(&node->ctx_node);
9094 mutex_unlock(&node->ctx->uring_lock);
9096 if (tctx->last == node->ctx)
9101 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9103 struct io_wq *wq = tctx->io_wq;
9104 struct io_tctx_node *node;
9105 unsigned long index;
9107 xa_for_each(&tctx->xa, index, node)
9108 io_uring_del_tctx_node(index);
9111 * Must be after io_uring_del_task_file() (removes nodes under
9112 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9115 io_wq_put_and_exit(wq);
9119 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9122 return atomic_read(&tctx->inflight_tracked);
9123 return percpu_counter_sum(&tctx->inflight);
9126 static void io_uring_drop_tctx_refs(struct task_struct *task)
9128 struct io_uring_task *tctx = task->io_uring;
9129 unsigned int refs = tctx->cached_refs;
9132 tctx->cached_refs = 0;
9133 percpu_counter_sub(&tctx->inflight, refs);
9134 put_task_struct_many(task, refs);
9139 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9140 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9142 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9144 struct io_uring_task *tctx = current->io_uring;
9145 struct io_ring_ctx *ctx;
9149 WARN_ON_ONCE(sqd && sqd->thread != current);
9151 if (!current->io_uring)
9154 io_wq_exit_start(tctx->io_wq);
9156 atomic_inc(&tctx->in_idle);
9158 io_uring_drop_tctx_refs(current);
9159 /* read completions before cancelations */
9160 inflight = tctx_inflight(tctx, !cancel_all);
9165 struct io_tctx_node *node;
9166 unsigned long index;
9168 xa_for_each(&tctx->xa, index, node) {
9169 /* sqpoll task will cancel all its requests */
9170 if (node->ctx->sq_data)
9172 io_uring_try_cancel_requests(node->ctx, current,
9176 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9177 io_uring_try_cancel_requests(ctx, current,
9181 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9182 io_uring_drop_tctx_refs(current);
9184 * If we've seen completions, retry without waiting. This
9185 * avoids a race where a completion comes in before we did
9186 * prepare_to_wait().
9188 if (inflight == tctx_inflight(tctx, !cancel_all))
9190 finish_wait(&tctx->wait, &wait);
9192 atomic_dec(&tctx->in_idle);
9194 io_uring_clean_tctx(tctx);
9196 /* for exec all current's requests should be gone, kill tctx */
9197 __io_uring_free(current);
9201 void __io_uring_cancel(struct files_struct *files)
9203 io_uring_cancel_generic(!files, NULL);
9206 static void *io_uring_validate_mmap_request(struct file *file,
9207 loff_t pgoff, size_t sz)
9209 struct io_ring_ctx *ctx = file->private_data;
9210 loff_t offset = pgoff << PAGE_SHIFT;
9215 case IORING_OFF_SQ_RING:
9216 case IORING_OFF_CQ_RING:
9219 case IORING_OFF_SQES:
9223 return ERR_PTR(-EINVAL);
9226 page = virt_to_head_page(ptr);
9227 if (sz > page_size(page))
9228 return ERR_PTR(-EINVAL);
9235 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9237 size_t sz = vma->vm_end - vma->vm_start;
9241 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9243 return PTR_ERR(ptr);
9245 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9246 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9249 #else /* !CONFIG_MMU */
9251 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9253 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9256 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9258 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9261 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9262 unsigned long addr, unsigned long len,
9263 unsigned long pgoff, unsigned long flags)
9267 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9269 return PTR_ERR(ptr);
9271 return (unsigned long) ptr;
9274 #endif /* !CONFIG_MMU */
9276 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9281 if (!io_sqring_full(ctx))
9283 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9285 if (!io_sqring_full(ctx))
9288 } while (!signal_pending(current));
9290 finish_wait(&ctx->sqo_sq_wait, &wait);
9294 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9295 struct __kernel_timespec __user **ts,
9296 const sigset_t __user **sig)
9298 struct io_uring_getevents_arg arg;
9301 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9302 * is just a pointer to the sigset_t.
9304 if (!(flags & IORING_ENTER_EXT_ARG)) {
9305 *sig = (const sigset_t __user *) argp;
9311 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9312 * timespec and sigset_t pointers if good.
9314 if (*argsz != sizeof(arg))
9316 if (copy_from_user(&arg, argp, sizeof(arg)))
9318 *sig = u64_to_user_ptr(arg.sigmask);
9319 *argsz = arg.sigmask_sz;
9320 *ts = u64_to_user_ptr(arg.ts);
9324 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9325 u32, min_complete, u32, flags, const void __user *, argp,
9328 struct io_ring_ctx *ctx;
9335 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9336 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9340 if (unlikely(!f.file))
9344 if (unlikely(f.file->f_op != &io_uring_fops))
9348 ctx = f.file->private_data;
9349 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9353 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9357 * For SQ polling, the thread will do all submissions and completions.
9358 * Just return the requested submit count, and wake the thread if
9362 if (ctx->flags & IORING_SETUP_SQPOLL) {
9363 io_cqring_overflow_flush(ctx);
9365 if (unlikely(ctx->sq_data->thread == NULL)) {
9369 if (flags & IORING_ENTER_SQ_WAKEUP)
9370 wake_up(&ctx->sq_data->wait);
9371 if (flags & IORING_ENTER_SQ_WAIT) {
9372 ret = io_sqpoll_wait_sq(ctx);
9376 submitted = to_submit;
9377 } else if (to_submit) {
9378 ret = io_uring_add_tctx_node(ctx);
9381 mutex_lock(&ctx->uring_lock);
9382 submitted = io_submit_sqes(ctx, to_submit);
9383 mutex_unlock(&ctx->uring_lock);
9385 if (submitted != to_submit)
9388 if (flags & IORING_ENTER_GETEVENTS) {
9389 const sigset_t __user *sig;
9390 struct __kernel_timespec __user *ts;
9392 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9396 min_complete = min(min_complete, ctx->cq_entries);
9399 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9400 * space applications don't need to do io completion events
9401 * polling again, they can rely on io_sq_thread to do polling
9402 * work, which can reduce cpu usage and uring_lock contention.
9404 if (ctx->flags & IORING_SETUP_IOPOLL &&
9405 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9406 ret = io_iopoll_check(ctx, min_complete);
9408 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9413 percpu_ref_put(&ctx->refs);
9416 return submitted ? submitted : ret;
9419 #ifdef CONFIG_PROC_FS
9420 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9421 const struct cred *cred)
9423 struct user_namespace *uns = seq_user_ns(m);
9424 struct group_info *gi;
9429 seq_printf(m, "%5d\n", id);
9430 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9431 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9432 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9433 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9434 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9435 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9436 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9437 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9438 seq_puts(m, "\n\tGroups:\t");
9439 gi = cred->group_info;
9440 for (g = 0; g < gi->ngroups; g++) {
9441 seq_put_decimal_ull(m, g ? " " : "",
9442 from_kgid_munged(uns, gi->gid[g]));
9444 seq_puts(m, "\n\tCapEff:\t");
9445 cap = cred->cap_effective;
9446 CAP_FOR_EACH_U32(__capi)
9447 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9452 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9454 struct io_sq_data *sq = NULL;
9459 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9460 * since fdinfo case grabs it in the opposite direction of normal use
9461 * cases. If we fail to get the lock, we just don't iterate any
9462 * structures that could be going away outside the io_uring mutex.
9464 has_lock = mutex_trylock(&ctx->uring_lock);
9466 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9472 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9473 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9474 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9475 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9476 struct file *f = io_file_from_index(ctx, i);
9479 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9481 seq_printf(m, "%5u: <none>\n", i);
9483 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9484 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9485 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9486 unsigned int len = buf->ubuf_end - buf->ubuf;
9488 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9490 if (has_lock && !xa_empty(&ctx->personalities)) {
9491 unsigned long index;
9492 const struct cred *cred;
9494 seq_printf(m, "Personalities:\n");
9495 xa_for_each(&ctx->personalities, index, cred)
9496 io_uring_show_cred(m, index, cred);
9498 seq_printf(m, "PollList:\n");
9499 spin_lock(&ctx->completion_lock);
9500 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9501 struct hlist_head *list = &ctx->cancel_hash[i];
9502 struct io_kiocb *req;
9504 hlist_for_each_entry(req, list, hash_node)
9505 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9506 req->task->task_works != NULL);
9508 spin_unlock(&ctx->completion_lock);
9510 mutex_unlock(&ctx->uring_lock);
9513 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9515 struct io_ring_ctx *ctx = f->private_data;
9517 if (percpu_ref_tryget(&ctx->refs)) {
9518 __io_uring_show_fdinfo(ctx, m);
9519 percpu_ref_put(&ctx->refs);
9524 static const struct file_operations io_uring_fops = {
9525 .release = io_uring_release,
9526 .mmap = io_uring_mmap,
9528 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9529 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9531 .poll = io_uring_poll,
9532 .fasync = io_uring_fasync,
9533 #ifdef CONFIG_PROC_FS
9534 .show_fdinfo = io_uring_show_fdinfo,
9538 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9539 struct io_uring_params *p)
9541 struct io_rings *rings;
9542 size_t size, sq_array_offset;
9544 /* make sure these are sane, as we already accounted them */
9545 ctx->sq_entries = p->sq_entries;
9546 ctx->cq_entries = p->cq_entries;
9548 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9549 if (size == SIZE_MAX)
9552 rings = io_mem_alloc(size);
9557 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9558 rings->sq_ring_mask = p->sq_entries - 1;
9559 rings->cq_ring_mask = p->cq_entries - 1;
9560 rings->sq_ring_entries = p->sq_entries;
9561 rings->cq_ring_entries = p->cq_entries;
9563 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9564 if (size == SIZE_MAX) {
9565 io_mem_free(ctx->rings);
9570 ctx->sq_sqes = io_mem_alloc(size);
9571 if (!ctx->sq_sqes) {
9572 io_mem_free(ctx->rings);
9580 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9584 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9588 ret = io_uring_add_tctx_node(ctx);
9593 fd_install(fd, file);
9598 * Allocate an anonymous fd, this is what constitutes the application
9599 * visible backing of an io_uring instance. The application mmaps this
9600 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9601 * we have to tie this fd to a socket for file garbage collection purposes.
9603 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9606 #if defined(CONFIG_UNIX)
9609 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9612 return ERR_PTR(ret);
9615 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9616 O_RDWR | O_CLOEXEC);
9617 #if defined(CONFIG_UNIX)
9619 sock_release(ctx->ring_sock);
9620 ctx->ring_sock = NULL;
9622 ctx->ring_sock->file = file;
9628 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9629 struct io_uring_params __user *params)
9631 struct io_ring_ctx *ctx;
9637 if (entries > IORING_MAX_ENTRIES) {
9638 if (!(p->flags & IORING_SETUP_CLAMP))
9640 entries = IORING_MAX_ENTRIES;
9644 * Use twice as many entries for the CQ ring. It's possible for the
9645 * application to drive a higher depth than the size of the SQ ring,
9646 * since the sqes are only used at submission time. This allows for
9647 * some flexibility in overcommitting a bit. If the application has
9648 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9649 * of CQ ring entries manually.
9651 p->sq_entries = roundup_pow_of_two(entries);
9652 if (p->flags & IORING_SETUP_CQSIZE) {
9654 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9655 * to a power-of-two, if it isn't already. We do NOT impose
9656 * any cq vs sq ring sizing.
9660 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9661 if (!(p->flags & IORING_SETUP_CLAMP))
9663 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9665 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9666 if (p->cq_entries < p->sq_entries)
9669 p->cq_entries = 2 * p->sq_entries;
9672 ctx = io_ring_ctx_alloc(p);
9675 ctx->compat = in_compat_syscall();
9676 if (!capable(CAP_IPC_LOCK))
9677 ctx->user = get_uid(current_user());
9680 * This is just grabbed for accounting purposes. When a process exits,
9681 * the mm is exited and dropped before the files, hence we need to hang
9682 * on to this mm purely for the purposes of being able to unaccount
9683 * memory (locked/pinned vm). It's not used for anything else.
9685 mmgrab(current->mm);
9686 ctx->mm_account = current->mm;
9688 ret = io_allocate_scq_urings(ctx, p);
9692 ret = io_sq_offload_create(ctx, p);
9695 /* always set a rsrc node */
9696 ret = io_rsrc_node_switch_start(ctx);
9699 io_rsrc_node_switch(ctx, NULL);
9701 memset(&p->sq_off, 0, sizeof(p->sq_off));
9702 p->sq_off.head = offsetof(struct io_rings, sq.head);
9703 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9704 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9705 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9706 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9707 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9708 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9710 memset(&p->cq_off, 0, sizeof(p->cq_off));
9711 p->cq_off.head = offsetof(struct io_rings, cq.head);
9712 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9713 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9714 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9715 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9716 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9717 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9719 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9720 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9721 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9722 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9723 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9724 IORING_FEAT_RSRC_TAGS;
9726 if (copy_to_user(params, p, sizeof(*p))) {
9731 file = io_uring_get_file(ctx);
9733 ret = PTR_ERR(file);
9738 * Install ring fd as the very last thing, so we don't risk someone
9739 * having closed it before we finish setup
9741 ret = io_uring_install_fd(ctx, file);
9743 /* fput will clean it up */
9748 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9751 io_ring_ctx_wait_and_kill(ctx);
9756 * Sets up an aio uring context, and returns the fd. Applications asks for a
9757 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9758 * params structure passed in.
9760 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9762 struct io_uring_params p;
9765 if (copy_from_user(&p, params, sizeof(p)))
9767 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9772 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9773 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9774 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9775 IORING_SETUP_R_DISABLED))
9778 return io_uring_create(entries, &p, params);
9781 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9782 struct io_uring_params __user *, params)
9784 return io_uring_setup(entries, params);
9787 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9789 struct io_uring_probe *p;
9793 size = struct_size(p, ops, nr_args);
9794 if (size == SIZE_MAX)
9796 p = kzalloc(size, GFP_KERNEL);
9801 if (copy_from_user(p, arg, size))
9804 if (memchr_inv(p, 0, size))
9807 p->last_op = IORING_OP_LAST - 1;
9808 if (nr_args > IORING_OP_LAST)
9809 nr_args = IORING_OP_LAST;
9811 for (i = 0; i < nr_args; i++) {
9813 if (!io_op_defs[i].not_supported)
9814 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9819 if (copy_to_user(arg, p, size))
9826 static int io_register_personality(struct io_ring_ctx *ctx)
9828 const struct cred *creds;
9832 creds = get_current_cred();
9834 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9835 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9843 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9844 unsigned int nr_args)
9846 struct io_uring_restriction *res;
9850 /* Restrictions allowed only if rings started disabled */
9851 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9854 /* We allow only a single restrictions registration */
9855 if (ctx->restrictions.registered)
9858 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9861 size = array_size(nr_args, sizeof(*res));
9862 if (size == SIZE_MAX)
9865 res = memdup_user(arg, size);
9867 return PTR_ERR(res);
9871 for (i = 0; i < nr_args; i++) {
9872 switch (res[i].opcode) {
9873 case IORING_RESTRICTION_REGISTER_OP:
9874 if (res[i].register_op >= IORING_REGISTER_LAST) {
9879 __set_bit(res[i].register_op,
9880 ctx->restrictions.register_op);
9882 case IORING_RESTRICTION_SQE_OP:
9883 if (res[i].sqe_op >= IORING_OP_LAST) {
9888 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9890 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9891 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9893 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9894 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9903 /* Reset all restrictions if an error happened */
9905 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9907 ctx->restrictions.registered = true;
9913 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9915 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9918 if (ctx->restrictions.registered)
9919 ctx->restricted = 1;
9921 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9922 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9923 wake_up(&ctx->sq_data->wait);
9927 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9928 struct io_uring_rsrc_update2 *up,
9936 if (check_add_overflow(up->offset, nr_args, &tmp))
9938 err = io_rsrc_node_switch_start(ctx);
9943 case IORING_RSRC_FILE:
9944 return __io_sqe_files_update(ctx, up, nr_args);
9945 case IORING_RSRC_BUFFER:
9946 return __io_sqe_buffers_update(ctx, up, nr_args);
9951 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9954 struct io_uring_rsrc_update2 up;
9958 memset(&up, 0, sizeof(up));
9959 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9961 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9964 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9965 unsigned size, unsigned type)
9967 struct io_uring_rsrc_update2 up;
9969 if (size != sizeof(up))
9971 if (copy_from_user(&up, arg, sizeof(up)))
9973 if (!up.nr || up.resv)
9975 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9978 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9979 unsigned int size, unsigned int type)
9981 struct io_uring_rsrc_register rr;
9983 /* keep it extendible */
9984 if (size != sizeof(rr))
9987 memset(&rr, 0, sizeof(rr));
9988 if (copy_from_user(&rr, arg, size))
9990 if (!rr.nr || rr.resv || rr.resv2)
9994 case IORING_RSRC_FILE:
9995 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9996 rr.nr, u64_to_user_ptr(rr.tags));
9997 case IORING_RSRC_BUFFER:
9998 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9999 rr.nr, u64_to_user_ptr(rr.tags));
10004 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10007 struct io_uring_task *tctx = current->io_uring;
10008 cpumask_var_t new_mask;
10011 if (!tctx || !tctx->io_wq)
10014 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10017 cpumask_clear(new_mask);
10018 if (len > cpumask_size())
10019 len = cpumask_size();
10021 if (copy_from_user(new_mask, arg, len)) {
10022 free_cpumask_var(new_mask);
10026 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10027 free_cpumask_var(new_mask);
10031 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10033 struct io_uring_task *tctx = current->io_uring;
10035 if (!tctx || !tctx->io_wq)
10038 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10041 static bool io_register_op_must_quiesce(int op)
10044 case IORING_REGISTER_BUFFERS:
10045 case IORING_UNREGISTER_BUFFERS:
10046 case IORING_REGISTER_FILES:
10047 case IORING_UNREGISTER_FILES:
10048 case IORING_REGISTER_FILES_UPDATE:
10049 case IORING_REGISTER_PROBE:
10050 case IORING_REGISTER_PERSONALITY:
10051 case IORING_UNREGISTER_PERSONALITY:
10052 case IORING_REGISTER_FILES2:
10053 case IORING_REGISTER_FILES_UPDATE2:
10054 case IORING_REGISTER_BUFFERS2:
10055 case IORING_REGISTER_BUFFERS_UPDATE:
10056 case IORING_REGISTER_IOWQ_AFF:
10057 case IORING_UNREGISTER_IOWQ_AFF:
10064 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10068 percpu_ref_kill(&ctx->refs);
10071 * Drop uring mutex before waiting for references to exit. If another
10072 * thread is currently inside io_uring_enter() it might need to grab the
10073 * uring_lock to make progress. If we hold it here across the drain
10074 * wait, then we can deadlock. It's safe to drop the mutex here, since
10075 * no new references will come in after we've killed the percpu ref.
10077 mutex_unlock(&ctx->uring_lock);
10079 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10082 ret = io_run_task_work_sig();
10083 } while (ret >= 0);
10084 mutex_lock(&ctx->uring_lock);
10087 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10091 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10092 void __user *arg, unsigned nr_args)
10093 __releases(ctx->uring_lock)
10094 __acquires(ctx->uring_lock)
10099 * We're inside the ring mutex, if the ref is already dying, then
10100 * someone else killed the ctx or is already going through
10101 * io_uring_register().
10103 if (percpu_ref_is_dying(&ctx->refs))
10106 if (ctx->restricted) {
10107 if (opcode >= IORING_REGISTER_LAST)
10109 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10110 if (!test_bit(opcode, ctx->restrictions.register_op))
10114 if (io_register_op_must_quiesce(opcode)) {
10115 ret = io_ctx_quiesce(ctx);
10121 case IORING_REGISTER_BUFFERS:
10122 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10124 case IORING_UNREGISTER_BUFFERS:
10126 if (arg || nr_args)
10128 ret = io_sqe_buffers_unregister(ctx);
10130 case IORING_REGISTER_FILES:
10131 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10133 case IORING_UNREGISTER_FILES:
10135 if (arg || nr_args)
10137 ret = io_sqe_files_unregister(ctx);
10139 case IORING_REGISTER_FILES_UPDATE:
10140 ret = io_register_files_update(ctx, arg, nr_args);
10142 case IORING_REGISTER_EVENTFD:
10143 case IORING_REGISTER_EVENTFD_ASYNC:
10147 ret = io_eventfd_register(ctx, arg);
10150 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10151 ctx->eventfd_async = 1;
10153 ctx->eventfd_async = 0;
10155 case IORING_UNREGISTER_EVENTFD:
10157 if (arg || nr_args)
10159 ret = io_eventfd_unregister(ctx);
10161 case IORING_REGISTER_PROBE:
10163 if (!arg || nr_args > 256)
10165 ret = io_probe(ctx, arg, nr_args);
10167 case IORING_REGISTER_PERSONALITY:
10169 if (arg || nr_args)
10171 ret = io_register_personality(ctx);
10173 case IORING_UNREGISTER_PERSONALITY:
10177 ret = io_unregister_personality(ctx, nr_args);
10179 case IORING_REGISTER_ENABLE_RINGS:
10181 if (arg || nr_args)
10183 ret = io_register_enable_rings(ctx);
10185 case IORING_REGISTER_RESTRICTIONS:
10186 ret = io_register_restrictions(ctx, arg, nr_args);
10188 case IORING_REGISTER_FILES2:
10189 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10191 case IORING_REGISTER_FILES_UPDATE2:
10192 ret = io_register_rsrc_update(ctx, arg, nr_args,
10195 case IORING_REGISTER_BUFFERS2:
10196 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10198 case IORING_REGISTER_BUFFERS_UPDATE:
10199 ret = io_register_rsrc_update(ctx, arg, nr_args,
10200 IORING_RSRC_BUFFER);
10202 case IORING_REGISTER_IOWQ_AFF:
10204 if (!arg || !nr_args)
10206 ret = io_register_iowq_aff(ctx, arg, nr_args);
10208 case IORING_UNREGISTER_IOWQ_AFF:
10210 if (arg || nr_args)
10212 ret = io_unregister_iowq_aff(ctx);
10219 if (io_register_op_must_quiesce(opcode)) {
10220 /* bring the ctx back to life */
10221 percpu_ref_reinit(&ctx->refs);
10222 reinit_completion(&ctx->ref_comp);
10227 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10228 void __user *, arg, unsigned int, nr_args)
10230 struct io_ring_ctx *ctx;
10239 if (f.file->f_op != &io_uring_fops)
10242 ctx = f.file->private_data;
10244 io_run_task_work();
10246 mutex_lock(&ctx->uring_lock);
10247 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10248 mutex_unlock(&ctx->uring_lock);
10249 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10250 ctx->cq_ev_fd != NULL, ret);
10256 static int __init io_uring_init(void)
10258 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10259 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10260 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10263 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10264 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10265 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10266 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10267 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10268 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10269 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10270 BUILD_BUG_SQE_ELEM(8, __u64, off);
10271 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10272 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10273 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10274 BUILD_BUG_SQE_ELEM(24, __u32, len);
10275 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10276 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10277 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10278 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10279 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10280 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10281 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10282 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10283 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10284 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10286 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10287 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10288 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10289 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10290 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10291 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10292 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10293 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10294 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10296 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10297 sizeof(struct io_uring_rsrc_update));
10298 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10299 sizeof(struct io_uring_rsrc_update2));
10300 /* should fit into one byte */
10301 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10303 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10304 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10306 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10310 __initcall(io_uring_init);