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_cqring (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/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
210 struct fixed_file_data {
211 struct fixed_file_table *table;
212 struct io_ring_ctx *ctx;
214 struct fixed_file_ref_node *node;
215 struct percpu_ref refs;
216 struct completion done;
217 struct list_head ref_list;
222 struct list_head list;
228 struct io_restriction {
229 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
230 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
231 u8 sqe_flags_allowed;
232 u8 sqe_flags_required;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list;
242 struct list_head ctx_new_list;
243 struct mutex ctx_lock;
245 struct task_struct *thread;
246 struct wait_queue_head wait;
251 struct percpu_ref refs;
252 } ____cacheline_aligned_in_smp;
256 unsigned int compat: 1;
257 unsigned int limit_mem: 1;
258 unsigned int cq_overflow_flushed: 1;
259 unsigned int drain_next: 1;
260 unsigned int eventfd_async: 1;
261 unsigned int restricted: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head;
278 unsigned sq_thread_idle;
279 unsigned cached_sq_dropped;
280 atomic_t cached_cq_overflow;
281 unsigned long sq_check_overflow;
283 struct list_head defer_list;
284 struct list_head timeout_list;
285 struct list_head cq_overflow_list;
287 wait_queue_head_t inflight_wait;
288 struct io_uring_sqe *sq_sqes;
289 } ____cacheline_aligned_in_smp;
291 struct io_rings *rings;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct *sqo_task;
302 /* Only used for accounting purposes */
303 struct mm_struct *mm_account;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state *sqo_blkcg_css;
309 struct io_sq_data *sq_data; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait;
312 struct wait_queue_entry sqo_wait_entry;
313 struct list_head sqd_list;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data *file_data;
321 unsigned nr_user_files;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs;
325 struct io_mapped_ubuf *user_bufs;
327 struct user_struct *user;
329 const struct cred *creds;
333 unsigned int sessionid;
336 struct completion ref_comp;
337 struct completion sq_thread_comp;
339 /* if all else fails... */
340 struct io_kiocb *fallback_req;
342 #if defined(CONFIG_UNIX)
343 struct socket *ring_sock;
346 struct idr io_buffer_idr;
348 struct idr personality_idr;
351 unsigned cached_cq_tail;
354 atomic_t cq_timeouts;
355 unsigned long cq_check_overflow;
356 struct wait_queue_head cq_wait;
357 struct fasync_struct *cq_fasync;
358 struct eventfd_ctx *cq_ev_fd;
359 } ____cacheline_aligned_in_smp;
362 struct mutex uring_lock;
363 wait_queue_head_t wait;
364 } ____cacheline_aligned_in_smp;
367 spinlock_t completion_lock;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list;
376 struct hlist_head *cancel_hash;
377 unsigned cancel_hash_bits;
378 bool poll_multi_file;
380 spinlock_t inflight_lock;
381 struct list_head inflight_list;
382 } ____cacheline_aligned_in_smp;
384 struct delayed_work file_put_work;
385 struct llist_head file_put_llist;
387 struct work_struct exit_work;
388 struct io_restriction restrictions;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb {
398 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
409 struct file *put_file;
413 struct io_timeout_data {
414 struct io_kiocb *req;
415 struct hrtimer timer;
416 struct timespec64 ts;
417 enum hrtimer_mode mode;
422 struct sockaddr __user *addr;
423 int __user *addr_len;
425 unsigned long nofile;
445 struct list_head list;
448 struct io_timeout_rem {
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user *addr;
469 struct user_msghdr __user *umsg;
475 struct io_buffer *kbuf;
481 struct filename *filename;
483 unsigned long nofile;
486 struct io_files_update {
512 struct epoll_event event;
516 struct file *file_out;
517 struct file *file_in;
524 struct io_provide_buf {
538 const char __user *filename;
539 struct statx __user *buffer;
542 struct io_completion {
544 struct list_head list;
548 struct io_async_connect {
549 struct sockaddr_storage address;
552 struct io_async_msghdr {
553 struct iovec fast_iov[UIO_FASTIOV];
555 struct sockaddr __user *uaddr;
557 struct sockaddr_storage addr;
561 struct iovec fast_iov[UIO_FASTIOV];
562 const struct iovec *free_iovec;
563 struct iov_iter iter;
565 struct wait_page_queue wpq;
569 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
570 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
571 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
572 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
573 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
574 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
581 REQ_F_LINK_TIMEOUT_BIT,
583 REQ_F_NEED_CLEANUP_BIT,
585 REQ_F_BUFFER_SELECTED_BIT,
586 REQ_F_NO_FILE_TABLE_BIT,
587 REQ_F_WORK_INITIALIZED_BIT,
589 /* not a real bit, just to check we're not overflowing the space */
595 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
596 /* drain existing IO first */
597 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
599 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
600 /* doesn't sever on completion < 0 */
601 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
603 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
604 /* IOSQE_BUFFER_SELECT */
605 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
608 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
609 /* fail rest of links */
610 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
611 /* on inflight list */
612 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
613 /* read/write uses file position */
614 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
615 /* must not punt to workers */
616 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
617 /* has linked timeout */
618 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
620 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
622 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
623 /* already went through poll handler */
624 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
625 /* buffer already selected */
626 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
627 /* doesn't need file table for this request */
628 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
629 /* io_wq_work is initialized */
630 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
634 struct io_poll_iocb poll;
635 struct io_poll_iocb *double_poll;
639 * NOTE! Each of the iocb union members has the file pointer
640 * as the first entry in their struct definition. So you can
641 * access the file pointer through any of the sub-structs,
642 * or directly as just 'ki_filp' in this struct.
648 struct io_poll_iocb poll;
649 struct io_accept accept;
651 struct io_cancel cancel;
652 struct io_timeout timeout;
653 struct io_timeout_rem timeout_rem;
654 struct io_connect connect;
655 struct io_sr_msg sr_msg;
657 struct io_close close;
658 struct io_files_update files_update;
659 struct io_fadvise fadvise;
660 struct io_madvise madvise;
661 struct io_epoll epoll;
662 struct io_splice splice;
663 struct io_provide_buf pbuf;
664 struct io_statx statx;
665 /* use only after cleaning per-op data, see io_clean_op() */
666 struct io_completion compl;
669 /* opcode allocated if it needs to store data for async defer */
672 /* polled IO has completed */
678 struct io_ring_ctx *ctx;
681 struct task_struct *task;
684 struct list_head link_list;
687 * 1. used with ctx->iopoll_list with reads/writes
688 * 2. to track reqs with ->files (see io_op_def::file_table)
690 struct list_head inflight_entry;
692 struct percpu_ref *fixed_file_refs;
693 struct callback_head task_work;
694 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
695 struct hlist_node hash_node;
696 struct async_poll *apoll;
697 struct io_wq_work work;
700 struct io_defer_entry {
701 struct list_head list;
702 struct io_kiocb *req;
706 #define IO_IOPOLL_BATCH 8
708 struct io_comp_state {
710 struct list_head list;
711 struct io_ring_ctx *ctx;
714 struct io_submit_state {
715 struct blk_plug plug;
718 * io_kiocb alloc cache
720 void *reqs[IO_IOPOLL_BATCH];
721 unsigned int free_reqs;
724 * Batch completion logic
726 struct io_comp_state comp;
729 * File reference cache
733 unsigned int has_refs;
734 unsigned int ios_left;
738 /* needs req->file assigned */
739 unsigned needs_file : 1;
740 /* don't fail if file grab fails */
741 unsigned needs_file_no_error : 1;
742 /* hash wq insertion if file is a regular file */
743 unsigned hash_reg_file : 1;
744 /* unbound wq insertion if file is a non-regular file */
745 unsigned unbound_nonreg_file : 1;
746 /* opcode is not supported by this kernel */
747 unsigned not_supported : 1;
748 /* set if opcode supports polled "wait" */
750 unsigned pollout : 1;
751 /* op supports buffer selection */
752 unsigned buffer_select : 1;
753 /* needs rlimit(RLIMIT_FSIZE) assigned */
754 unsigned needs_fsize : 1;
755 /* must always have async data allocated */
756 unsigned needs_async_data : 1;
757 /* size of async data needed, if any */
758 unsigned short async_size;
762 static const struct io_op_def io_op_defs[] = {
763 [IORING_OP_NOP] = {},
764 [IORING_OP_READV] = {
766 .unbound_nonreg_file = 1,
769 .needs_async_data = 1,
770 .async_size = sizeof(struct io_async_rw),
771 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
773 [IORING_OP_WRITEV] = {
776 .unbound_nonreg_file = 1,
779 .needs_async_data = 1,
780 .async_size = sizeof(struct io_async_rw),
781 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
783 [IORING_OP_FSYNC] = {
785 .work_flags = IO_WQ_WORK_BLKCG,
787 [IORING_OP_READ_FIXED] = {
789 .unbound_nonreg_file = 1,
791 .async_size = sizeof(struct io_async_rw),
792 .work_flags = IO_WQ_WORK_BLKCG,
794 [IORING_OP_WRITE_FIXED] = {
797 .unbound_nonreg_file = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_BLKCG,
803 [IORING_OP_POLL_ADD] = {
805 .unbound_nonreg_file = 1,
807 [IORING_OP_POLL_REMOVE] = {},
808 [IORING_OP_SYNC_FILE_RANGE] = {
810 .work_flags = IO_WQ_WORK_BLKCG,
812 [IORING_OP_SENDMSG] = {
814 .unbound_nonreg_file = 1,
816 .needs_async_data = 1,
817 .async_size = sizeof(struct io_async_msghdr),
818 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
821 [IORING_OP_RECVMSG] = {
823 .unbound_nonreg_file = 1,
826 .needs_async_data = 1,
827 .async_size = sizeof(struct io_async_msghdr),
828 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
831 [IORING_OP_TIMEOUT] = {
832 .needs_async_data = 1,
833 .async_size = sizeof(struct io_timeout_data),
834 .work_flags = IO_WQ_WORK_MM,
836 [IORING_OP_TIMEOUT_REMOVE] = {},
837 [IORING_OP_ACCEPT] = {
839 .unbound_nonreg_file = 1,
841 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
843 [IORING_OP_ASYNC_CANCEL] = {},
844 [IORING_OP_LINK_TIMEOUT] = {
845 .needs_async_data = 1,
846 .async_size = sizeof(struct io_timeout_data),
847 .work_flags = IO_WQ_WORK_MM,
849 [IORING_OP_CONNECT] = {
851 .unbound_nonreg_file = 1,
853 .needs_async_data = 1,
854 .async_size = sizeof(struct io_async_connect),
855 .work_flags = IO_WQ_WORK_MM,
857 [IORING_OP_FALLOCATE] = {
860 .work_flags = IO_WQ_WORK_BLKCG,
862 [IORING_OP_OPENAT] = {
863 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
866 [IORING_OP_CLOSE] = {
868 .needs_file_no_error = 1,
869 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
871 [IORING_OP_FILES_UPDATE] = {
872 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
874 [IORING_OP_STATX] = {
875 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
876 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
880 .unbound_nonreg_file = 1,
883 .async_size = sizeof(struct io_async_rw),
884 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
886 [IORING_OP_WRITE] = {
888 .unbound_nonreg_file = 1,
891 .async_size = sizeof(struct io_async_rw),
892 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
894 [IORING_OP_FADVISE] = {
896 .work_flags = IO_WQ_WORK_BLKCG,
898 [IORING_OP_MADVISE] = {
899 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
903 .unbound_nonreg_file = 1,
905 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
909 .unbound_nonreg_file = 1,
912 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 [IORING_OP_OPENAT2] = {
915 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
918 [IORING_OP_EPOLL_CTL] = {
919 .unbound_nonreg_file = 1,
920 .work_flags = IO_WQ_WORK_FILES,
922 [IORING_OP_SPLICE] = {
925 .unbound_nonreg_file = 1,
926 .work_flags = IO_WQ_WORK_BLKCG,
928 [IORING_OP_PROVIDE_BUFFERS] = {},
929 [IORING_OP_REMOVE_BUFFERS] = {},
933 .unbound_nonreg_file = 1,
937 enum io_mem_account {
942 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
943 struct io_comp_state *cs);
944 static void io_cqring_fill_event(struct io_kiocb *req, long res);
945 static void io_put_req(struct io_kiocb *req);
946 static void io_put_req_deferred(struct io_kiocb *req, int nr);
947 static void io_double_put_req(struct io_kiocb *req);
948 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
949 static void __io_queue_linked_timeout(struct io_kiocb *req);
950 static void io_queue_linked_timeout(struct io_kiocb *req);
951 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
952 struct io_uring_files_update *ip,
954 static void __io_clean_op(struct io_kiocb *req);
955 static struct file *io_file_get(struct io_submit_state *state,
956 struct io_kiocb *req, int fd, bool fixed);
957 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
958 static void io_file_put_work(struct work_struct *work);
960 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
961 struct iovec **iovec, struct iov_iter *iter,
963 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
964 const struct iovec *fast_iov,
965 struct iov_iter *iter, bool force);
967 static struct kmem_cache *req_cachep;
969 static const struct file_operations io_uring_fops;
971 struct sock *io_uring_get_socket(struct file *file)
973 #if defined(CONFIG_UNIX)
974 if (file->f_op == &io_uring_fops) {
975 struct io_ring_ctx *ctx = file->private_data;
977 return ctx->ring_sock->sk;
982 EXPORT_SYMBOL(io_uring_get_socket);
984 static inline void io_clean_op(struct io_kiocb *req)
986 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
991 static void io_sq_thread_drop_mm(void)
993 struct mm_struct *mm = current->mm;
996 kthread_unuse_mm(mm);
1001 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1004 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1005 !ctx->sqo_task->mm ||
1006 !mmget_not_zero(ctx->sqo_task->mm)))
1008 kthread_use_mm(ctx->sqo_task->mm);
1014 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1015 struct io_kiocb *req)
1017 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1019 return __io_sq_thread_acquire_mm(ctx);
1022 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1023 struct cgroup_subsys_state **cur_css)
1026 #ifdef CONFIG_BLK_CGROUP
1027 /* puts the old one when swapping */
1028 if (*cur_css != ctx->sqo_blkcg_css) {
1029 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1030 *cur_css = ctx->sqo_blkcg_css;
1035 static void io_sq_thread_unassociate_blkcg(void)
1037 #ifdef CONFIG_BLK_CGROUP
1038 kthread_associate_blkcg(NULL);
1042 static inline void req_set_fail_links(struct io_kiocb *req)
1044 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1045 req->flags |= REQ_F_FAIL_LINK;
1049 * None of these are dereferenced, they are simply used to check if any of
1050 * them have changed. If we're under current and check they are still the
1051 * same, we're fine to grab references to them for actual out-of-line use.
1053 static void io_init_identity(struct io_identity *id)
1055 id->files = current->files;
1056 id->mm = current->mm;
1057 #ifdef CONFIG_BLK_CGROUP
1059 id->blkcg_css = blkcg_css();
1062 id->creds = current_cred();
1063 id->nsproxy = current->nsproxy;
1064 id->fs = current->fs;
1065 id->fsize = rlimit(RLIMIT_FSIZE);
1067 id->loginuid = current->loginuid;
1068 id->sessionid = current->sessionid;
1070 refcount_set(&id->count, 1);
1074 * Note: must call io_req_init_async() for the first time you
1075 * touch any members of io_wq_work.
1077 static inline void io_req_init_async(struct io_kiocb *req)
1079 struct io_uring_task *tctx = current->io_uring;
1081 if (req->flags & REQ_F_WORK_INITIALIZED)
1084 memset(&req->work, 0, sizeof(req->work));
1085 req->flags |= REQ_F_WORK_INITIALIZED;
1087 /* Grab a ref if this isn't our static identity */
1088 req->work.identity = tctx->identity;
1089 if (tctx->identity != &tctx->__identity)
1090 refcount_inc(&req->work.identity->count);
1093 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1095 return ctx->flags & IORING_SETUP_SQPOLL;
1098 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1100 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1102 complete(&ctx->ref_comp);
1105 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1107 return !req->timeout.off;
1110 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1112 struct io_ring_ctx *ctx;
1115 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1119 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1120 if (!ctx->fallback_req)
1124 * Use 5 bits less than the max cq entries, that should give us around
1125 * 32 entries per hash list if totally full and uniformly spread.
1127 hash_bits = ilog2(p->cq_entries);
1131 ctx->cancel_hash_bits = hash_bits;
1132 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1134 if (!ctx->cancel_hash)
1136 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1138 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1139 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1142 ctx->flags = p->flags;
1143 init_waitqueue_head(&ctx->sqo_sq_wait);
1144 INIT_LIST_HEAD(&ctx->sqd_list);
1145 init_waitqueue_head(&ctx->cq_wait);
1146 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1147 init_completion(&ctx->ref_comp);
1148 init_completion(&ctx->sq_thread_comp);
1149 idr_init(&ctx->io_buffer_idr);
1150 idr_init(&ctx->personality_idr);
1151 mutex_init(&ctx->uring_lock);
1152 init_waitqueue_head(&ctx->wait);
1153 spin_lock_init(&ctx->completion_lock);
1154 INIT_LIST_HEAD(&ctx->iopoll_list);
1155 INIT_LIST_HEAD(&ctx->defer_list);
1156 INIT_LIST_HEAD(&ctx->timeout_list);
1157 init_waitqueue_head(&ctx->inflight_wait);
1158 spin_lock_init(&ctx->inflight_lock);
1159 INIT_LIST_HEAD(&ctx->inflight_list);
1160 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1161 init_llist_head(&ctx->file_put_llist);
1164 if (ctx->fallback_req)
1165 kmem_cache_free(req_cachep, ctx->fallback_req);
1166 kfree(ctx->cancel_hash);
1171 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1173 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1174 struct io_ring_ctx *ctx = req->ctx;
1176 return seq != ctx->cached_cq_tail
1177 + atomic_read(&ctx->cached_cq_overflow);
1183 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1185 struct io_rings *rings = ctx->rings;
1187 /* order cqe stores with ring update */
1188 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1190 if (wq_has_sleeper(&ctx->cq_wait)) {
1191 wake_up_interruptible(&ctx->cq_wait);
1192 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1196 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1198 if (req->work.identity == &tctx->__identity)
1200 if (refcount_dec_and_test(&req->work.identity->count))
1201 kfree(req->work.identity);
1204 static void io_req_clean_work(struct io_kiocb *req)
1206 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1209 req->flags &= ~REQ_F_WORK_INITIALIZED;
1211 if (req->work.flags & IO_WQ_WORK_MM) {
1212 mmdrop(req->work.identity->mm);
1213 req->work.flags &= ~IO_WQ_WORK_MM;
1215 #ifdef CONFIG_BLK_CGROUP
1216 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1217 css_put(req->work.identity->blkcg_css);
1218 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1221 if (req->work.flags & IO_WQ_WORK_CREDS) {
1222 put_cred(req->work.identity->creds);
1223 req->work.flags &= ~IO_WQ_WORK_CREDS;
1225 if (req->work.flags & IO_WQ_WORK_FS) {
1226 struct fs_struct *fs = req->work.identity->fs;
1228 spin_lock(&req->work.identity->fs->lock);
1231 spin_unlock(&req->work.identity->fs->lock);
1234 req->work.flags &= ~IO_WQ_WORK_FS;
1237 io_put_identity(req->task->io_uring, req);
1241 * Create a private copy of io_identity, since some fields don't match
1242 * the current context.
1244 static bool io_identity_cow(struct io_kiocb *req)
1246 struct io_uring_task *tctx = current->io_uring;
1247 const struct cred *creds = NULL;
1248 struct io_identity *id;
1250 if (req->work.flags & IO_WQ_WORK_CREDS)
1251 creds = req->work.identity->creds;
1253 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1254 if (unlikely(!id)) {
1255 req->work.flags |= IO_WQ_WORK_CANCEL;
1260 * We can safely just re-init the creds we copied Either the field
1261 * matches the current one, or we haven't grabbed it yet. The only
1262 * exception is ->creds, through registered personalities, so handle
1263 * that one separately.
1265 io_init_identity(id);
1267 req->work.identity->creds = creds;
1269 /* add one for this request */
1270 refcount_inc(&id->count);
1272 /* drop old identity, assign new one. one ref for req, one for tctx */
1273 if (req->work.identity != tctx->identity &&
1274 refcount_sub_and_test(2, &req->work.identity->count))
1275 kfree(req->work.identity);
1277 req->work.identity = id;
1278 tctx->identity = id;
1282 static bool io_grab_identity(struct io_kiocb *req)
1284 const struct io_op_def *def = &io_op_defs[req->opcode];
1285 struct io_identity *id = req->work.identity;
1286 struct io_ring_ctx *ctx = req->ctx;
1288 if (def->needs_fsize && id->fsize != rlimit(RLIMIT_FSIZE))
1291 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1292 (def->work_flags & IO_WQ_WORK_FILES) &&
1293 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1294 if (id->files != current->files ||
1295 id->nsproxy != current->nsproxy)
1297 atomic_inc(&id->files->count);
1298 get_nsproxy(id->nsproxy);
1299 req->flags |= REQ_F_INFLIGHT;
1301 spin_lock_irq(&ctx->inflight_lock);
1302 list_add(&req->inflight_entry, &ctx->inflight_list);
1303 spin_unlock_irq(&ctx->inflight_lock);
1304 req->work.flags |= IO_WQ_WORK_FILES;
1306 #ifdef CONFIG_BLK_CGROUP
1307 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1308 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1310 if (id->blkcg_css != blkcg_css()) {
1315 * This should be rare, either the cgroup is dying or the task
1316 * is moving cgroups. Just punt to root for the handful of ios.
1318 if (css_tryget_online(id->blkcg_css))
1319 req->work.flags |= IO_WQ_WORK_BLKCG;
1323 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1324 if (id->creds != current_cred())
1326 get_cred(id->creds);
1327 req->work.flags |= IO_WQ_WORK_CREDS;
1330 if (!uid_eq(current->loginuid, id->loginuid) ||
1331 current->sessionid != id->sessionid)
1334 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1335 (def->work_flags & IO_WQ_WORK_FS)) {
1336 if (current->fs != id->fs)
1338 spin_lock(&id->fs->lock);
1339 if (!id->fs->in_exec) {
1341 req->work.flags |= IO_WQ_WORK_FS;
1343 req->work.flags |= IO_WQ_WORK_CANCEL;
1345 spin_unlock(¤t->fs->lock);
1351 static void io_prep_async_work(struct io_kiocb *req)
1353 const struct io_op_def *def = &io_op_defs[req->opcode];
1354 struct io_ring_ctx *ctx = req->ctx;
1355 struct io_identity *id;
1357 io_req_init_async(req);
1358 id = req->work.identity;
1360 if (req->flags & REQ_F_ISREG) {
1361 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1362 io_wq_hash_work(&req->work, file_inode(req->file));
1364 if (def->unbound_nonreg_file)
1365 req->work.flags |= IO_WQ_WORK_UNBOUND;
1368 /* ->mm can never change on us */
1369 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1370 (def->work_flags & IO_WQ_WORK_MM)) {
1372 req->work.flags |= IO_WQ_WORK_MM;
1375 /* if we fail grabbing identity, we must COW, regrab, and retry */
1376 if (io_grab_identity(req))
1379 if (!io_identity_cow(req))
1382 /* can't fail at this point */
1383 if (!io_grab_identity(req))
1387 static void io_prep_async_link(struct io_kiocb *req)
1389 struct io_kiocb *cur;
1391 io_prep_async_work(req);
1392 if (req->flags & REQ_F_LINK_HEAD)
1393 list_for_each_entry(cur, &req->link_list, link_list)
1394 io_prep_async_work(cur);
1397 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1399 struct io_ring_ctx *ctx = req->ctx;
1400 struct io_kiocb *link = io_prep_linked_timeout(req);
1402 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1403 &req->work, req->flags);
1404 io_wq_enqueue(ctx->io_wq, &req->work);
1408 static void io_queue_async_work(struct io_kiocb *req)
1410 struct io_kiocb *link;
1412 /* init ->work of the whole link before punting */
1413 io_prep_async_link(req);
1414 link = __io_queue_async_work(req);
1417 io_queue_linked_timeout(link);
1420 static void io_kill_timeout(struct io_kiocb *req)
1422 struct io_timeout_data *io = req->async_data;
1425 ret = hrtimer_try_to_cancel(&io->timer);
1427 atomic_set(&req->ctx->cq_timeouts,
1428 atomic_read(&req->ctx->cq_timeouts) + 1);
1429 list_del_init(&req->timeout.list);
1430 io_cqring_fill_event(req, 0);
1431 io_put_req_deferred(req, 1);
1435 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1437 struct io_ring_ctx *ctx = req->ctx;
1439 if (!tsk || req->task == tsk)
1441 if (ctx->flags & IORING_SETUP_SQPOLL) {
1442 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1449 * Returns true if we found and killed one or more timeouts
1451 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1453 struct io_kiocb *req, *tmp;
1456 spin_lock_irq(&ctx->completion_lock);
1457 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1458 if (io_task_match(req, tsk)) {
1459 io_kill_timeout(req);
1463 spin_unlock_irq(&ctx->completion_lock);
1464 return canceled != 0;
1467 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1470 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1471 struct io_defer_entry, list);
1472 struct io_kiocb *link;
1474 if (req_need_defer(de->req, de->seq))
1476 list_del_init(&de->list);
1477 /* punt-init is done before queueing for defer */
1478 link = __io_queue_async_work(de->req);
1480 __io_queue_linked_timeout(link);
1481 /* drop submission reference */
1482 io_put_req_deferred(link, 1);
1485 } while (!list_empty(&ctx->defer_list));
1488 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1490 while (!list_empty(&ctx->timeout_list)) {
1491 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1492 struct io_kiocb, timeout.list);
1494 if (io_is_timeout_noseq(req))
1496 if (req->timeout.target_seq != ctx->cached_cq_tail
1497 - atomic_read(&ctx->cq_timeouts))
1500 list_del_init(&req->timeout.list);
1501 io_kill_timeout(req);
1505 static void io_commit_cqring(struct io_ring_ctx *ctx)
1507 io_flush_timeouts(ctx);
1508 __io_commit_cqring(ctx);
1510 if (unlikely(!list_empty(&ctx->defer_list)))
1511 __io_queue_deferred(ctx);
1514 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1516 struct io_rings *r = ctx->rings;
1518 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1521 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1523 struct io_rings *rings = ctx->rings;
1526 tail = ctx->cached_cq_tail;
1528 * writes to the cq entry need to come after reading head; the
1529 * control dependency is enough as we're using WRITE_ONCE to
1532 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1535 ctx->cached_cq_tail++;
1536 return &rings->cqes[tail & ctx->cq_mask];
1539 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1543 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1545 if (!ctx->eventfd_async)
1547 return io_wq_current_is_worker();
1550 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1552 if (waitqueue_active(&ctx->wait))
1553 wake_up(&ctx->wait);
1554 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1555 wake_up(&ctx->sq_data->wait);
1556 if (io_should_trigger_evfd(ctx))
1557 eventfd_signal(ctx->cq_ev_fd, 1);
1560 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1562 if (list_empty(&ctx->cq_overflow_list)) {
1563 clear_bit(0, &ctx->sq_check_overflow);
1564 clear_bit(0, &ctx->cq_check_overflow);
1565 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1569 static inline bool io_match_files(struct io_kiocb *req,
1570 struct files_struct *files)
1574 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1575 (req->work.flags & IO_WQ_WORK_FILES))
1576 return req->work.identity->files == files;
1580 /* Returns true if there are no backlogged entries after the flush */
1581 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1582 struct task_struct *tsk,
1583 struct files_struct *files)
1585 struct io_rings *rings = ctx->rings;
1586 struct io_kiocb *req, *tmp;
1587 struct io_uring_cqe *cqe;
1588 unsigned long flags;
1592 if (list_empty_careful(&ctx->cq_overflow_list))
1594 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1595 rings->cq_ring_entries))
1599 spin_lock_irqsave(&ctx->completion_lock, flags);
1601 /* if force is set, the ring is going away. always drop after that */
1603 ctx->cq_overflow_flushed = 1;
1606 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1607 if (tsk && req->task != tsk)
1609 if (!io_match_files(req, files))
1612 cqe = io_get_cqring(ctx);
1616 list_move(&req->compl.list, &list);
1618 WRITE_ONCE(cqe->user_data, req->user_data);
1619 WRITE_ONCE(cqe->res, req->result);
1620 WRITE_ONCE(cqe->flags, req->compl.cflags);
1622 WRITE_ONCE(ctx->rings->cq_overflow,
1623 atomic_inc_return(&ctx->cached_cq_overflow));
1627 io_commit_cqring(ctx);
1628 io_cqring_mark_overflow(ctx);
1630 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1631 io_cqring_ev_posted(ctx);
1633 while (!list_empty(&list)) {
1634 req = list_first_entry(&list, struct io_kiocb, compl.list);
1635 list_del(&req->compl.list);
1642 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1644 struct io_ring_ctx *ctx = req->ctx;
1645 struct io_uring_cqe *cqe;
1647 trace_io_uring_complete(ctx, req->user_data, res);
1650 * If we can't get a cq entry, userspace overflowed the
1651 * submission (by quite a lot). Increment the overflow count in
1654 cqe = io_get_cqring(ctx);
1656 WRITE_ONCE(cqe->user_data, req->user_data);
1657 WRITE_ONCE(cqe->res, res);
1658 WRITE_ONCE(cqe->flags, cflags);
1659 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1661 * If we're in ring overflow flush mode, or in task cancel mode,
1662 * then we cannot store the request for later flushing, we need
1663 * to drop it on the floor.
1665 WRITE_ONCE(ctx->rings->cq_overflow,
1666 atomic_inc_return(&ctx->cached_cq_overflow));
1668 if (list_empty(&ctx->cq_overflow_list)) {
1669 set_bit(0, &ctx->sq_check_overflow);
1670 set_bit(0, &ctx->cq_check_overflow);
1671 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1675 req->compl.cflags = cflags;
1676 refcount_inc(&req->refs);
1677 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1681 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1683 __io_cqring_fill_event(req, res, 0);
1686 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1688 struct io_ring_ctx *ctx = req->ctx;
1689 unsigned long flags;
1691 spin_lock_irqsave(&ctx->completion_lock, flags);
1692 __io_cqring_fill_event(req, res, cflags);
1693 io_commit_cqring(ctx);
1694 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1696 io_cqring_ev_posted(ctx);
1699 static void io_submit_flush_completions(struct io_comp_state *cs)
1701 struct io_ring_ctx *ctx = cs->ctx;
1703 spin_lock_irq(&ctx->completion_lock);
1704 while (!list_empty(&cs->list)) {
1705 struct io_kiocb *req;
1707 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1708 list_del(&req->compl.list);
1709 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1712 * io_free_req() doesn't care about completion_lock unless one
1713 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1714 * because of a potential deadlock with req->work.fs->lock
1716 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1717 |REQ_F_WORK_INITIALIZED)) {
1718 spin_unlock_irq(&ctx->completion_lock);
1720 spin_lock_irq(&ctx->completion_lock);
1725 io_commit_cqring(ctx);
1726 spin_unlock_irq(&ctx->completion_lock);
1728 io_cqring_ev_posted(ctx);
1732 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1733 struct io_comp_state *cs)
1736 io_cqring_add_event(req, res, cflags);
1741 req->compl.cflags = cflags;
1742 list_add_tail(&req->compl.list, &cs->list);
1744 io_submit_flush_completions(cs);
1748 static void io_req_complete(struct io_kiocb *req, long res)
1750 __io_req_complete(req, res, 0, NULL);
1753 static inline bool io_is_fallback_req(struct io_kiocb *req)
1755 return req == (struct io_kiocb *)
1756 ((unsigned long) req->ctx->fallback_req & ~1UL);
1759 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1761 struct io_kiocb *req;
1763 req = ctx->fallback_req;
1764 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1770 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1771 struct io_submit_state *state)
1773 if (!state->free_reqs) {
1774 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1778 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1779 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1782 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1783 * retry single alloc to be on the safe side.
1785 if (unlikely(ret <= 0)) {
1786 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1787 if (!state->reqs[0])
1791 state->free_reqs = ret;
1795 return state->reqs[state->free_reqs];
1797 return io_get_fallback_req(ctx);
1800 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1804 percpu_ref_put(req->fixed_file_refs);
1809 static void io_dismantle_req(struct io_kiocb *req)
1813 if (req->async_data)
1814 kfree(req->async_data);
1816 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1818 io_req_clean_work(req);
1821 static void __io_free_req(struct io_kiocb *req)
1823 struct io_uring_task *tctx = req->task->io_uring;
1824 struct io_ring_ctx *ctx = req->ctx;
1826 io_dismantle_req(req);
1828 percpu_counter_dec(&tctx->inflight);
1830 wake_up(&tctx->wait);
1831 put_task_struct(req->task);
1833 if (likely(!io_is_fallback_req(req)))
1834 kmem_cache_free(req_cachep, req);
1836 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1837 percpu_ref_put(&ctx->refs);
1840 static bool io_link_cancel_timeout(struct io_kiocb *req)
1842 struct io_timeout_data *io = req->async_data;
1843 struct io_ring_ctx *ctx = req->ctx;
1846 ret = hrtimer_try_to_cancel(&io->timer);
1848 io_cqring_fill_event(req, -ECANCELED);
1849 io_commit_cqring(ctx);
1850 req->flags &= ~REQ_F_LINK_HEAD;
1851 io_put_req_deferred(req, 1);
1858 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1860 struct io_kiocb *link;
1863 if (list_empty(&req->link_list))
1865 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1866 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1869 list_del_init(&link->link_list);
1870 wake_ev = io_link_cancel_timeout(link);
1871 req->flags &= ~REQ_F_LINK_TIMEOUT;
1875 static void io_kill_linked_timeout(struct io_kiocb *req)
1877 struct io_ring_ctx *ctx = req->ctx;
1878 unsigned long flags;
1881 spin_lock_irqsave(&ctx->completion_lock, flags);
1882 wake_ev = __io_kill_linked_timeout(req);
1883 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1886 io_cqring_ev_posted(ctx);
1889 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1891 struct io_kiocb *nxt;
1894 * The list should never be empty when we are called here. But could
1895 * potentially happen if the chain is messed up, check to be on the
1898 if (unlikely(list_empty(&req->link_list)))
1901 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1902 list_del_init(&req->link_list);
1903 if (!list_empty(&nxt->link_list))
1904 nxt->flags |= REQ_F_LINK_HEAD;
1909 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1911 static void __io_fail_links(struct io_kiocb *req)
1913 struct io_ring_ctx *ctx = req->ctx;
1915 while (!list_empty(&req->link_list)) {
1916 struct io_kiocb *link = list_first_entry(&req->link_list,
1917 struct io_kiocb, link_list);
1919 list_del_init(&link->link_list);
1920 trace_io_uring_fail_link(req, link);
1922 io_cqring_fill_event(link, -ECANCELED);
1925 * It's ok to free under spinlock as they're not linked anymore,
1926 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1929 if (link->flags & REQ_F_WORK_INITIALIZED)
1930 io_put_req_deferred(link, 2);
1932 io_double_put_req(link);
1935 io_commit_cqring(ctx);
1938 static void io_fail_links(struct io_kiocb *req)
1940 struct io_ring_ctx *ctx = req->ctx;
1941 unsigned long flags;
1943 spin_lock_irqsave(&ctx->completion_lock, flags);
1944 __io_fail_links(req);
1945 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1947 io_cqring_ev_posted(ctx);
1950 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1952 req->flags &= ~REQ_F_LINK_HEAD;
1953 if (req->flags & REQ_F_LINK_TIMEOUT)
1954 io_kill_linked_timeout(req);
1957 * If LINK is set, we have dependent requests in this chain. If we
1958 * didn't fail this request, queue the first one up, moving any other
1959 * dependencies to the next request. In case of failure, fail the rest
1962 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1963 return io_req_link_next(req);
1968 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1970 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1972 return __io_req_find_next(req);
1975 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1977 struct task_struct *tsk = req->task;
1978 struct io_ring_ctx *ctx = req->ctx;
1981 if (tsk->flags & PF_EXITING)
1985 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1986 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1987 * processing task_work. There's no reliable way to tell if TWA_RESUME
1991 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1992 notify = TWA_SIGNAL;
1994 ret = task_work_add(tsk, &req->task_work, notify);
1996 wake_up_process(tsk);
2001 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2003 struct io_ring_ctx *ctx = req->ctx;
2005 spin_lock_irq(&ctx->completion_lock);
2006 io_cqring_fill_event(req, error);
2007 io_commit_cqring(ctx);
2008 spin_unlock_irq(&ctx->completion_lock);
2010 io_cqring_ev_posted(ctx);
2011 req_set_fail_links(req);
2012 io_double_put_req(req);
2015 static void io_req_task_cancel(struct callback_head *cb)
2017 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2018 struct io_ring_ctx *ctx = req->ctx;
2020 __io_req_task_cancel(req, -ECANCELED);
2021 percpu_ref_put(&ctx->refs);
2024 static void __io_req_task_submit(struct io_kiocb *req)
2026 struct io_ring_ctx *ctx = req->ctx;
2028 if (!__io_sq_thread_acquire_mm(ctx)) {
2029 mutex_lock(&ctx->uring_lock);
2030 __io_queue_sqe(req, NULL);
2031 mutex_unlock(&ctx->uring_lock);
2033 __io_req_task_cancel(req, -EFAULT);
2037 static void io_req_task_submit(struct callback_head *cb)
2039 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2040 struct io_ring_ctx *ctx = req->ctx;
2042 __io_req_task_submit(req);
2043 percpu_ref_put(&ctx->refs);
2046 static void io_req_task_queue(struct io_kiocb *req)
2050 init_task_work(&req->task_work, io_req_task_submit);
2051 percpu_ref_get(&req->ctx->refs);
2053 ret = io_req_task_work_add(req, true);
2054 if (unlikely(ret)) {
2055 struct task_struct *tsk;
2057 init_task_work(&req->task_work, io_req_task_cancel);
2058 tsk = io_wq_get_task(req->ctx->io_wq);
2059 task_work_add(tsk, &req->task_work, 0);
2060 wake_up_process(tsk);
2064 static void io_queue_next(struct io_kiocb *req)
2066 struct io_kiocb *nxt = io_req_find_next(req);
2069 io_req_task_queue(nxt);
2072 static void io_free_req(struct io_kiocb *req)
2079 void *reqs[IO_IOPOLL_BATCH];
2082 struct task_struct *task;
2086 static inline void io_init_req_batch(struct req_batch *rb)
2093 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2094 struct req_batch *rb)
2096 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2097 percpu_ref_put_many(&ctx->refs, rb->to_free);
2101 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2102 struct req_batch *rb)
2105 __io_req_free_batch_flush(ctx, rb);
2107 struct io_uring_task *tctx = rb->task->io_uring;
2109 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2110 put_task_struct_many(rb->task, rb->task_refs);
2115 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2117 if (unlikely(io_is_fallback_req(req))) {
2121 if (req->flags & REQ_F_LINK_HEAD)
2124 if (req->task != rb->task) {
2126 struct io_uring_task *tctx = rb->task->io_uring;
2128 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2129 put_task_struct_many(rb->task, rb->task_refs);
2131 rb->task = req->task;
2136 io_dismantle_req(req);
2137 rb->reqs[rb->to_free++] = req;
2138 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2139 __io_req_free_batch_flush(req->ctx, rb);
2143 * Drop reference to request, return next in chain (if there is one) if this
2144 * was the last reference to this request.
2146 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2148 struct io_kiocb *nxt = NULL;
2150 if (refcount_dec_and_test(&req->refs)) {
2151 nxt = io_req_find_next(req);
2157 static void io_put_req(struct io_kiocb *req)
2159 if (refcount_dec_and_test(&req->refs))
2163 static void io_put_req_deferred_cb(struct callback_head *cb)
2165 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2170 static void io_free_req_deferred(struct io_kiocb *req)
2174 init_task_work(&req->task_work, io_put_req_deferred_cb);
2175 ret = io_req_task_work_add(req, true);
2176 if (unlikely(ret)) {
2177 struct task_struct *tsk;
2179 tsk = io_wq_get_task(req->ctx->io_wq);
2180 task_work_add(tsk, &req->task_work, 0);
2181 wake_up_process(tsk);
2185 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2187 if (refcount_sub_and_test(refs, &req->refs))
2188 io_free_req_deferred(req);
2191 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2193 struct io_kiocb *nxt;
2196 * A ref is owned by io-wq in which context we're. So, if that's the
2197 * last one, it's safe to steal next work. False negatives are Ok,
2198 * it just will be re-punted async in io_put_work()
2200 if (refcount_read(&req->refs) != 1)
2203 nxt = io_req_find_next(req);
2204 return nxt ? &nxt->work : NULL;
2207 static void io_double_put_req(struct io_kiocb *req)
2209 /* drop both submit and complete references */
2210 if (refcount_sub_and_test(2, &req->refs))
2214 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2216 struct io_rings *rings = ctx->rings;
2218 if (test_bit(0, &ctx->cq_check_overflow)) {
2220 * noflush == true is from the waitqueue handler, just ensure
2221 * we wake up the task, and the next invocation will flush the
2222 * entries. We cannot safely to it from here.
2224 if (noflush && !list_empty(&ctx->cq_overflow_list))
2227 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2230 /* See comment at the top of this file */
2232 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2235 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2237 struct io_rings *rings = ctx->rings;
2239 /* make sure SQ entry isn't read before tail */
2240 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2243 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2245 unsigned int cflags;
2247 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2248 cflags |= IORING_CQE_F_BUFFER;
2249 req->flags &= ~REQ_F_BUFFER_SELECTED;
2254 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2256 struct io_buffer *kbuf;
2258 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2259 return io_put_kbuf(req, kbuf);
2262 static inline bool io_run_task_work(void)
2265 * Not safe to run on exiting task, and the task_work handling will
2266 * not add work to such a task.
2268 if (unlikely(current->flags & PF_EXITING))
2270 if (current->task_works) {
2271 __set_current_state(TASK_RUNNING);
2279 static void io_iopoll_queue(struct list_head *again)
2281 struct io_kiocb *req;
2284 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2285 list_del(&req->inflight_entry);
2286 __io_complete_rw(req, -EAGAIN, 0, NULL);
2287 } while (!list_empty(again));
2291 * Find and free completed poll iocbs
2293 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2294 struct list_head *done)
2296 struct req_batch rb;
2297 struct io_kiocb *req;
2300 /* order with ->result store in io_complete_rw_iopoll() */
2303 io_init_req_batch(&rb);
2304 while (!list_empty(done)) {
2307 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2308 if (READ_ONCE(req->result) == -EAGAIN) {
2310 req->iopoll_completed = 0;
2311 list_move_tail(&req->inflight_entry, &again);
2314 list_del(&req->inflight_entry);
2316 if (req->flags & REQ_F_BUFFER_SELECTED)
2317 cflags = io_put_rw_kbuf(req);
2319 __io_cqring_fill_event(req, req->result, cflags);
2322 if (refcount_dec_and_test(&req->refs))
2323 io_req_free_batch(&rb, req);
2326 io_commit_cqring(ctx);
2327 if (ctx->flags & IORING_SETUP_SQPOLL)
2328 io_cqring_ev_posted(ctx);
2329 io_req_free_batch_finish(ctx, &rb);
2331 if (!list_empty(&again))
2332 io_iopoll_queue(&again);
2335 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2338 struct io_kiocb *req, *tmp;
2344 * Only spin for completions if we don't have multiple devices hanging
2345 * off our complete list, and we're under the requested amount.
2347 spin = !ctx->poll_multi_file && *nr_events < min;
2350 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2351 struct kiocb *kiocb = &req->rw.kiocb;
2354 * Move completed and retryable entries to our local lists.
2355 * If we find a request that requires polling, break out
2356 * and complete those lists first, if we have entries there.
2358 if (READ_ONCE(req->iopoll_completed)) {
2359 list_move_tail(&req->inflight_entry, &done);
2362 if (!list_empty(&done))
2365 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2369 /* iopoll may have completed current req */
2370 if (READ_ONCE(req->iopoll_completed))
2371 list_move_tail(&req->inflight_entry, &done);
2378 if (!list_empty(&done))
2379 io_iopoll_complete(ctx, nr_events, &done);
2385 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2386 * non-spinning poll check - we'll still enter the driver poll loop, but only
2387 * as a non-spinning completion check.
2389 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2392 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2395 ret = io_do_iopoll(ctx, nr_events, min);
2398 if (*nr_events >= min)
2406 * We can't just wait for polled events to come to us, we have to actively
2407 * find and complete them.
2409 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2411 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2414 mutex_lock(&ctx->uring_lock);
2415 while (!list_empty(&ctx->iopoll_list)) {
2416 unsigned int nr_events = 0;
2418 io_do_iopoll(ctx, &nr_events, 0);
2420 /* let it sleep and repeat later if can't complete a request */
2424 * Ensure we allow local-to-the-cpu processing to take place,
2425 * in this case we need to ensure that we reap all events.
2426 * Also let task_work, etc. to progress by releasing the mutex
2428 if (need_resched()) {
2429 mutex_unlock(&ctx->uring_lock);
2431 mutex_lock(&ctx->uring_lock);
2434 mutex_unlock(&ctx->uring_lock);
2437 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2439 unsigned int nr_events = 0;
2440 int iters = 0, ret = 0;
2443 * We disallow the app entering submit/complete with polling, but we
2444 * still need to lock the ring to prevent racing with polled issue
2445 * that got punted to a workqueue.
2447 mutex_lock(&ctx->uring_lock);
2450 * Don't enter poll loop if we already have events pending.
2451 * If we do, we can potentially be spinning for commands that
2452 * already triggered a CQE (eg in error).
2454 if (io_cqring_events(ctx, false))
2458 * If a submit got punted to a workqueue, we can have the
2459 * application entering polling for a command before it gets
2460 * issued. That app will hold the uring_lock for the duration
2461 * of the poll right here, so we need to take a breather every
2462 * now and then to ensure that the issue has a chance to add
2463 * the poll to the issued list. Otherwise we can spin here
2464 * forever, while the workqueue is stuck trying to acquire the
2467 if (!(++iters & 7)) {
2468 mutex_unlock(&ctx->uring_lock);
2470 mutex_lock(&ctx->uring_lock);
2473 ret = io_iopoll_getevents(ctx, &nr_events, min);
2477 } while (min && !nr_events && !need_resched());
2479 mutex_unlock(&ctx->uring_lock);
2483 static void kiocb_end_write(struct io_kiocb *req)
2486 * Tell lockdep we inherited freeze protection from submission
2489 if (req->flags & REQ_F_ISREG) {
2490 struct inode *inode = file_inode(req->file);
2492 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2494 file_end_write(req->file);
2497 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2498 struct io_comp_state *cs)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 if (kiocb->ki_flags & IOCB_WRITE)
2504 kiocb_end_write(req);
2506 if (res != req->result)
2507 req_set_fail_links(req);
2508 if (req->flags & REQ_F_BUFFER_SELECTED)
2509 cflags = io_put_rw_kbuf(req);
2510 __io_req_complete(req, res, cflags, cs);
2514 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2516 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2517 ssize_t ret = -ECANCELED;
2518 struct iov_iter iter;
2526 switch (req->opcode) {
2527 case IORING_OP_READV:
2528 case IORING_OP_READ_FIXED:
2529 case IORING_OP_READ:
2532 case IORING_OP_WRITEV:
2533 case IORING_OP_WRITE_FIXED:
2534 case IORING_OP_WRITE:
2538 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2543 if (!req->async_data) {
2544 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2547 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2555 req_set_fail_links(req);
2556 io_req_complete(req, ret);
2561 static bool io_rw_reissue(struct io_kiocb *req, long res)
2564 umode_t mode = file_inode(req->file)->i_mode;
2567 if (!S_ISBLK(mode) && !S_ISREG(mode))
2569 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2572 ret = io_sq_thread_acquire_mm(req->ctx, req);
2574 if (io_resubmit_prep(req, ret)) {
2575 refcount_inc(&req->refs);
2576 io_queue_async_work(req);
2584 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2585 struct io_comp_state *cs)
2587 if (!io_rw_reissue(req, res))
2588 io_complete_rw_common(&req->rw.kiocb, res, cs);
2591 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2593 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2595 __io_complete_rw(req, res, res2, NULL);
2598 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2600 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2602 if (kiocb->ki_flags & IOCB_WRITE)
2603 kiocb_end_write(req);
2605 if (res != -EAGAIN && res != req->result)
2606 req_set_fail_links(req);
2608 WRITE_ONCE(req->result, res);
2609 /* order with io_poll_complete() checking ->result */
2611 WRITE_ONCE(req->iopoll_completed, 1);
2615 * After the iocb has been issued, it's safe to be found on the poll list.
2616 * Adding the kiocb to the list AFTER submission ensures that we don't
2617 * find it from a io_iopoll_getevents() thread before the issuer is done
2618 * accessing the kiocb cookie.
2620 static void io_iopoll_req_issued(struct io_kiocb *req)
2622 struct io_ring_ctx *ctx = req->ctx;
2625 * Track whether we have multiple files in our lists. This will impact
2626 * how we do polling eventually, not spinning if we're on potentially
2627 * different devices.
2629 if (list_empty(&ctx->iopoll_list)) {
2630 ctx->poll_multi_file = false;
2631 } else if (!ctx->poll_multi_file) {
2632 struct io_kiocb *list_req;
2634 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2636 if (list_req->file != req->file)
2637 ctx->poll_multi_file = true;
2641 * For fast devices, IO may have already completed. If it has, add
2642 * it to the front so we find it first.
2644 if (READ_ONCE(req->iopoll_completed))
2645 list_add(&req->inflight_entry, &ctx->iopoll_list);
2647 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2649 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2650 wq_has_sleeper(&ctx->sq_data->wait))
2651 wake_up(&ctx->sq_data->wait);
2654 static void __io_state_file_put(struct io_submit_state *state)
2656 if (state->has_refs)
2657 fput_many(state->file, state->has_refs);
2661 static inline void io_state_file_put(struct io_submit_state *state)
2664 __io_state_file_put(state);
2668 * Get as many references to a file as we have IOs left in this submission,
2669 * assuming most submissions are for one file, or at least that each file
2670 * has more than one submission.
2672 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2678 if (state->fd == fd) {
2682 __io_state_file_put(state);
2684 state->file = fget_many(fd, state->ios_left);
2689 state->has_refs = state->ios_left - 1;
2693 static bool io_bdev_nowait(struct block_device *bdev)
2696 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2703 * If we tracked the file through the SCM inflight mechanism, we could support
2704 * any file. For now, just ensure that anything potentially problematic is done
2707 static bool io_file_supports_async(struct file *file, int rw)
2709 umode_t mode = file_inode(file)->i_mode;
2711 if (S_ISBLK(mode)) {
2712 if (io_bdev_nowait(file->f_inode->i_bdev))
2716 if (S_ISCHR(mode) || S_ISSOCK(mode))
2718 if (S_ISREG(mode)) {
2719 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2720 file->f_op != &io_uring_fops)
2725 /* any ->read/write should understand O_NONBLOCK */
2726 if (file->f_flags & O_NONBLOCK)
2729 if (!(file->f_mode & FMODE_NOWAIT))
2733 return file->f_op->read_iter != NULL;
2735 return file->f_op->write_iter != NULL;
2738 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2740 struct io_ring_ctx *ctx = req->ctx;
2741 struct kiocb *kiocb = &req->rw.kiocb;
2745 if (S_ISREG(file_inode(req->file)->i_mode))
2746 req->flags |= REQ_F_ISREG;
2748 kiocb->ki_pos = READ_ONCE(sqe->off);
2749 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2750 req->flags |= REQ_F_CUR_POS;
2751 kiocb->ki_pos = req->file->f_pos;
2753 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2754 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2755 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2759 ioprio = READ_ONCE(sqe->ioprio);
2761 ret = ioprio_check_cap(ioprio);
2765 kiocb->ki_ioprio = ioprio;
2767 kiocb->ki_ioprio = get_current_ioprio();
2769 /* don't allow async punt if RWF_NOWAIT was requested */
2770 if (kiocb->ki_flags & IOCB_NOWAIT)
2771 req->flags |= REQ_F_NOWAIT;
2773 if (ctx->flags & IORING_SETUP_IOPOLL) {
2774 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2775 !kiocb->ki_filp->f_op->iopoll)
2778 kiocb->ki_flags |= IOCB_HIPRI;
2779 kiocb->ki_complete = io_complete_rw_iopoll;
2780 req->iopoll_completed = 0;
2782 if (kiocb->ki_flags & IOCB_HIPRI)
2784 kiocb->ki_complete = io_complete_rw;
2787 req->rw.addr = READ_ONCE(sqe->addr);
2788 req->rw.len = READ_ONCE(sqe->len);
2789 req->buf_index = READ_ONCE(sqe->buf_index);
2793 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2799 case -ERESTARTNOINTR:
2800 case -ERESTARTNOHAND:
2801 case -ERESTART_RESTARTBLOCK:
2803 * We can't just restart the syscall, since previously
2804 * submitted sqes may already be in progress. Just fail this
2810 kiocb->ki_complete(kiocb, ret, 0);
2814 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2815 struct io_comp_state *cs)
2817 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2818 struct io_async_rw *io = req->async_data;
2820 /* add previously done IO, if any */
2821 if (io && io->bytes_done > 0) {
2823 ret = io->bytes_done;
2825 ret += io->bytes_done;
2828 if (req->flags & REQ_F_CUR_POS)
2829 req->file->f_pos = kiocb->ki_pos;
2830 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2831 __io_complete_rw(req, ret, 0, cs);
2833 io_rw_done(kiocb, ret);
2836 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2837 struct iov_iter *iter)
2839 struct io_ring_ctx *ctx = req->ctx;
2840 size_t len = req->rw.len;
2841 struct io_mapped_ubuf *imu;
2842 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 = &ctx->user_bufs[index];
2850 buf_addr = req->rw.addr;
2853 if (buf_addr + len < buf_addr)
2855 /* not inside the mapped region */
2856 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2860 * May not be a start of buffer, set size appropriately
2861 * and advance us to the beginning.
2863 offset = buf_addr - imu->ubuf;
2864 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2868 * Don't use iov_iter_advance() here, as it's really slow for
2869 * using the latter parts of a big fixed buffer - it iterates
2870 * over each segment manually. We can cheat a bit here, because
2873 * 1) it's a BVEC iter, we set it up
2874 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2875 * first and last bvec
2877 * So just find our index, and adjust the iterator afterwards.
2878 * If the offset is within the first bvec (or the whole first
2879 * bvec, just use iov_iter_advance(). This makes it easier
2880 * since we can just skip the first segment, which may not
2881 * be PAGE_SIZE aligned.
2883 const struct bio_vec *bvec = imu->bvec;
2885 if (offset <= bvec->bv_len) {
2886 iov_iter_advance(iter, offset);
2888 unsigned long seg_skip;
2890 /* skip first vec */
2891 offset -= bvec->bv_len;
2892 seg_skip = 1 + (offset >> PAGE_SHIFT);
2894 iter->bvec = bvec + seg_skip;
2895 iter->nr_segs -= seg_skip;
2896 iter->count -= bvec->bv_len + offset;
2897 iter->iov_offset = offset & ~PAGE_MASK;
2904 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2907 mutex_unlock(&ctx->uring_lock);
2910 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2913 * "Normal" inline submissions always hold the uring_lock, since we
2914 * grab it from the system call. Same is true for the SQPOLL offload.
2915 * The only exception is when we've detached the request and issue it
2916 * from an async worker thread, grab the lock for that case.
2919 mutex_lock(&ctx->uring_lock);
2922 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2923 int bgid, struct io_buffer *kbuf,
2926 struct io_buffer *head;
2928 if (req->flags & REQ_F_BUFFER_SELECTED)
2931 io_ring_submit_lock(req->ctx, needs_lock);
2933 lockdep_assert_held(&req->ctx->uring_lock);
2935 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2937 if (!list_empty(&head->list)) {
2938 kbuf = list_last_entry(&head->list, struct io_buffer,
2940 list_del(&kbuf->list);
2943 idr_remove(&req->ctx->io_buffer_idr, bgid);
2945 if (*len > kbuf->len)
2948 kbuf = ERR_PTR(-ENOBUFS);
2951 io_ring_submit_unlock(req->ctx, needs_lock);
2956 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2959 struct io_buffer *kbuf;
2962 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2963 bgid = req->buf_index;
2964 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2967 req->rw.addr = (u64) (unsigned long) kbuf;
2968 req->flags |= REQ_F_BUFFER_SELECTED;
2969 return u64_to_user_ptr(kbuf->addr);
2972 #ifdef CONFIG_COMPAT
2973 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2976 struct compat_iovec __user *uiov;
2977 compat_ssize_t clen;
2981 uiov = u64_to_user_ptr(req->rw.addr);
2982 if (!access_ok(uiov, sizeof(*uiov)))
2984 if (__get_user(clen, &uiov->iov_len))
2990 buf = io_rw_buffer_select(req, &len, needs_lock);
2992 return PTR_ERR(buf);
2993 iov[0].iov_base = buf;
2994 iov[0].iov_len = (compat_size_t) len;
2999 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3002 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3006 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3009 len = iov[0].iov_len;
3012 buf = io_rw_buffer_select(req, &len, needs_lock);
3014 return PTR_ERR(buf);
3015 iov[0].iov_base = buf;
3016 iov[0].iov_len = len;
3020 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3023 if (req->flags & REQ_F_BUFFER_SELECTED) {
3024 struct io_buffer *kbuf;
3026 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3027 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3028 iov[0].iov_len = kbuf->len;
3033 else if (req->rw.len > 1)
3036 #ifdef CONFIG_COMPAT
3037 if (req->ctx->compat)
3038 return io_compat_import(req, iov, needs_lock);
3041 return __io_iov_buffer_select(req, iov, needs_lock);
3044 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3045 struct iovec **iovec, struct iov_iter *iter,
3048 void __user *buf = u64_to_user_ptr(req->rw.addr);
3049 size_t sqe_len = req->rw.len;
3053 opcode = req->opcode;
3054 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3056 return io_import_fixed(req, rw, iter);
3059 /* buffer index only valid with fixed read/write, or buffer select */
3060 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3063 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3064 if (req->flags & REQ_F_BUFFER_SELECT) {
3065 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3067 return PTR_ERR(buf);
3068 req->rw.len = sqe_len;
3071 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3073 return ret < 0 ? ret : sqe_len;
3076 if (req->flags & REQ_F_BUFFER_SELECT) {
3077 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3079 ret = (*iovec)->iov_len;
3080 iov_iter_init(iter, rw, *iovec, 1, ret);
3086 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3090 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3091 struct iovec **iovec, struct iov_iter *iter,
3094 struct io_async_rw *iorw = req->async_data;
3097 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3099 return iov_iter_count(&iorw->iter);
3102 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3104 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3108 * For files that don't have ->read_iter() and ->write_iter(), handle them
3109 * by looping over ->read() or ->write() manually.
3111 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3112 struct iov_iter *iter)
3117 * Don't support polled IO through this interface, and we can't
3118 * support non-blocking either. For the latter, this just causes
3119 * the kiocb to be handled from an async context.
3121 if (kiocb->ki_flags & IOCB_HIPRI)
3123 if (kiocb->ki_flags & IOCB_NOWAIT)
3126 while (iov_iter_count(iter)) {
3130 if (!iov_iter_is_bvec(iter)) {
3131 iovec = iov_iter_iovec(iter);
3133 /* fixed buffers import bvec */
3134 iovec.iov_base = kmap(iter->bvec->bv_page)
3136 iovec.iov_len = min(iter->count,
3137 iter->bvec->bv_len - iter->iov_offset);
3141 nr = file->f_op->read(file, iovec.iov_base,
3142 iovec.iov_len, io_kiocb_ppos(kiocb));
3144 nr = file->f_op->write(file, iovec.iov_base,
3145 iovec.iov_len, io_kiocb_ppos(kiocb));
3148 if (iov_iter_is_bvec(iter))
3149 kunmap(iter->bvec->bv_page);
3157 if (nr != iovec.iov_len)
3159 iov_iter_advance(iter, nr);
3165 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3166 const struct iovec *fast_iov, struct iov_iter *iter)
3168 struct io_async_rw *rw = req->async_data;
3170 memcpy(&rw->iter, iter, sizeof(*iter));
3171 rw->free_iovec = iovec;
3173 /* can only be fixed buffers, no need to do anything */
3174 if (iter->type == ITER_BVEC)
3177 unsigned iov_off = 0;
3179 rw->iter.iov = rw->fast_iov;
3180 if (iter->iov != fast_iov) {
3181 iov_off = iter->iov - fast_iov;
3182 rw->iter.iov += iov_off;
3184 if (rw->fast_iov != fast_iov)
3185 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3186 sizeof(struct iovec) * iter->nr_segs);
3188 req->flags |= REQ_F_NEED_CLEANUP;
3192 static inline int __io_alloc_async_data(struct io_kiocb *req)
3194 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3195 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3196 return req->async_data == NULL;
3199 static int io_alloc_async_data(struct io_kiocb *req)
3201 if (!io_op_defs[req->opcode].needs_async_data)
3204 return __io_alloc_async_data(req);
3207 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3208 const struct iovec *fast_iov,
3209 struct iov_iter *iter, bool force)
3211 if (!force && !io_op_defs[req->opcode].needs_async_data)
3213 if (!req->async_data) {
3214 if (__io_alloc_async_data(req))
3217 io_req_map_rw(req, iovec, fast_iov, iter);
3222 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3224 struct io_async_rw *iorw = req->async_data;
3225 struct iovec *iov = iorw->fast_iov;
3228 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3229 if (unlikely(ret < 0))
3232 iorw->bytes_done = 0;
3233 iorw->free_iovec = iov;
3235 req->flags |= REQ_F_NEED_CLEANUP;
3239 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3243 ret = io_prep_rw(req, sqe);
3247 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3250 /* either don't need iovec imported or already have it */
3251 if (!req->async_data)
3253 return io_rw_prep_async(req, READ);
3257 * This is our waitqueue callback handler, registered through lock_page_async()
3258 * when we initially tried to do the IO with the iocb armed our waitqueue.
3259 * This gets called when the page is unlocked, and we generally expect that to
3260 * happen when the page IO is completed and the page is now uptodate. This will
3261 * queue a task_work based retry of the operation, attempting to copy the data
3262 * again. If the latter fails because the page was NOT uptodate, then we will
3263 * do a thread based blocking retry of the operation. That's the unexpected
3266 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3267 int sync, void *arg)
3269 struct wait_page_queue *wpq;
3270 struct io_kiocb *req = wait->private;
3271 struct wait_page_key *key = arg;
3274 wpq = container_of(wait, struct wait_page_queue, wait);
3276 if (!wake_page_match(wpq, key))
3279 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3280 list_del_init(&wait->entry);
3282 init_task_work(&req->task_work, io_req_task_submit);
3283 percpu_ref_get(&req->ctx->refs);
3285 /* submit ref gets dropped, acquire a new one */
3286 refcount_inc(&req->refs);
3287 ret = io_req_task_work_add(req, true);
3288 if (unlikely(ret)) {
3289 struct task_struct *tsk;
3291 /* queue just for cancelation */
3292 init_task_work(&req->task_work, io_req_task_cancel);
3293 tsk = io_wq_get_task(req->ctx->io_wq);
3294 task_work_add(tsk, &req->task_work, 0);
3295 wake_up_process(tsk);
3301 * This controls whether a given IO request should be armed for async page
3302 * based retry. If we return false here, the request is handed to the async
3303 * worker threads for retry. If we're doing buffered reads on a regular file,
3304 * we prepare a private wait_page_queue entry and retry the operation. This
3305 * will either succeed because the page is now uptodate and unlocked, or it
3306 * will register a callback when the page is unlocked at IO completion. Through
3307 * that callback, io_uring uses task_work to setup a retry of the operation.
3308 * That retry will attempt the buffered read again. The retry will generally
3309 * succeed, or in rare cases where it fails, we then fall back to using the
3310 * async worker threads for a blocking retry.
3312 static bool io_rw_should_retry(struct io_kiocb *req)
3314 struct io_async_rw *rw = req->async_data;
3315 struct wait_page_queue *wait = &rw->wpq;
3316 struct kiocb *kiocb = &req->rw.kiocb;
3318 /* never retry for NOWAIT, we just complete with -EAGAIN */
3319 if (req->flags & REQ_F_NOWAIT)
3322 /* Only for buffered IO */
3323 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3327 * just use poll if we can, and don't attempt if the fs doesn't
3328 * support callback based unlocks
3330 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3333 wait->wait.func = io_async_buf_func;
3334 wait->wait.private = req;
3335 wait->wait.flags = 0;
3336 INIT_LIST_HEAD(&wait->wait.entry);
3337 kiocb->ki_flags |= IOCB_WAITQ;
3338 kiocb->ki_flags &= ~IOCB_NOWAIT;
3339 kiocb->ki_waitq = wait;
3343 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3345 if (req->file->f_op->read_iter)
3346 return call_read_iter(req->file, &req->rw.kiocb, iter);
3347 else if (req->file->f_op->read)
3348 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3353 static int io_read(struct io_kiocb *req, bool force_nonblock,
3354 struct io_comp_state *cs)
3356 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3357 struct kiocb *kiocb = &req->rw.kiocb;
3358 struct iov_iter __iter, *iter = &__iter;
3359 struct io_async_rw *rw = req->async_data;
3360 ssize_t io_size, ret, ret2;
3367 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3370 iov_count = iov_iter_count(iter);
3372 req->result = io_size;
3375 /* Ensure we clear previously set non-block flag */
3376 if (!force_nonblock)
3377 kiocb->ki_flags &= ~IOCB_NOWAIT;
3379 kiocb->ki_flags |= IOCB_NOWAIT;
3382 /* If the file doesn't support async, just async punt */
3383 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3387 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3391 ret = io_iter_do_read(req, iter);
3395 } else if (ret == -EIOCBQUEUED) {
3398 } else if (ret == -EAGAIN) {
3399 /* IOPOLL retry should happen for io-wq threads */
3400 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3402 /* no retry on NONBLOCK marked file */
3403 if (req->file->f_flags & O_NONBLOCK)
3405 /* some cases will consume bytes even on error returns */
3406 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3409 } else if (ret < 0) {
3410 /* make sure -ERESTARTSYS -> -EINTR is done */
3414 /* read it all, or we did blocking attempt. no retry. */
3415 if (!iov_iter_count(iter) || !force_nonblock ||
3416 (req->file->f_flags & O_NONBLOCK))
3421 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3428 rw = req->async_data;
3429 /* it's copied and will be cleaned with ->io */
3431 /* now use our persistent iterator, if we aren't already */
3434 rw->bytes_done += ret;
3435 /* if we can retry, do so with the callbacks armed */
3436 if (!io_rw_should_retry(req)) {
3437 kiocb->ki_flags &= ~IOCB_WAITQ;
3442 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3443 * get -EIOCBQUEUED, then we'll get a notification when the desired
3444 * page gets unlocked. We can also get a partial read here, and if we
3445 * do, then just retry at the new offset.
3447 ret = io_iter_do_read(req, iter);
3448 if (ret == -EIOCBQUEUED) {
3451 } else if (ret > 0 && ret < io_size) {
3452 /* we got some bytes, but not all. retry. */
3456 kiocb_done(kiocb, ret, cs);
3459 /* it's reportedly faster than delegating the null check to kfree() */
3465 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3469 ret = io_prep_rw(req, sqe);
3473 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3476 /* either don't need iovec imported or already have it */
3477 if (!req->async_data)
3479 return io_rw_prep_async(req, WRITE);
3482 static int io_write(struct io_kiocb *req, bool force_nonblock,
3483 struct io_comp_state *cs)
3485 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3486 struct kiocb *kiocb = &req->rw.kiocb;
3487 struct iov_iter __iter, *iter = &__iter;
3488 struct io_async_rw *rw = req->async_data;
3490 ssize_t ret, ret2, io_size;
3495 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3498 iov_count = iov_iter_count(iter);
3500 req->result = io_size;
3502 /* Ensure we clear previously set non-block flag */
3503 if (!force_nonblock)
3504 kiocb->ki_flags &= ~IOCB_NOWAIT;
3506 kiocb->ki_flags |= IOCB_NOWAIT;
3508 /* If the file doesn't support async, just async punt */
3509 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3512 /* file path doesn't support NOWAIT for non-direct_IO */
3513 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3514 (req->flags & REQ_F_ISREG))
3517 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3522 * Open-code file_start_write here to grab freeze protection,
3523 * which will be released by another thread in
3524 * io_complete_rw(). Fool lockdep by telling it the lock got
3525 * released so that it doesn't complain about the held lock when
3526 * we return to userspace.
3528 if (req->flags & REQ_F_ISREG) {
3529 __sb_start_write(file_inode(req->file)->i_sb,
3530 SB_FREEZE_WRITE, true);
3531 __sb_writers_release(file_inode(req->file)->i_sb,
3534 kiocb->ki_flags |= IOCB_WRITE;
3536 if (req->file->f_op->write_iter)
3537 ret2 = call_write_iter(req->file, kiocb, iter);
3538 else if (req->file->f_op->write)
3539 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3544 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3545 * retry them without IOCB_NOWAIT.
3547 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3549 /* no retry on NONBLOCK marked file */
3550 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3552 if (!force_nonblock || ret2 != -EAGAIN) {
3553 /* IOPOLL retry should happen for io-wq threads */
3554 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3557 kiocb_done(kiocb, ret2, cs);
3560 /* some cases will consume bytes even on error returns */
3561 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3562 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3567 /* it's reportedly faster than delegating the null check to kfree() */
3573 static int __io_splice_prep(struct io_kiocb *req,
3574 const struct io_uring_sqe *sqe)
3576 struct io_splice* sp = &req->splice;
3577 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3579 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3583 sp->len = READ_ONCE(sqe->len);
3584 sp->flags = READ_ONCE(sqe->splice_flags);
3586 if (unlikely(sp->flags & ~valid_flags))
3589 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3590 (sp->flags & SPLICE_F_FD_IN_FIXED));
3593 req->flags |= REQ_F_NEED_CLEANUP;
3595 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3597 * Splice operation will be punted aync, and here need to
3598 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3600 io_req_init_async(req);
3601 req->work.flags |= IO_WQ_WORK_UNBOUND;
3607 static int io_tee_prep(struct io_kiocb *req,
3608 const struct io_uring_sqe *sqe)
3610 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3612 return __io_splice_prep(req, sqe);
3615 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3617 struct io_splice *sp = &req->splice;
3618 struct file *in = sp->file_in;
3619 struct file *out = sp->file_out;
3620 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3626 ret = do_tee(in, out, sp->len, flags);
3628 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3629 req->flags &= ~REQ_F_NEED_CLEANUP;
3632 req_set_fail_links(req);
3633 io_req_complete(req, ret);
3637 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3639 struct io_splice* sp = &req->splice;
3641 sp->off_in = READ_ONCE(sqe->splice_off_in);
3642 sp->off_out = READ_ONCE(sqe->off);
3643 return __io_splice_prep(req, sqe);
3646 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3648 struct io_splice *sp = &req->splice;
3649 struct file *in = sp->file_in;
3650 struct file *out = sp->file_out;
3651 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3652 loff_t *poff_in, *poff_out;
3658 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3659 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3662 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3664 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3665 req->flags &= ~REQ_F_NEED_CLEANUP;
3668 req_set_fail_links(req);
3669 io_req_complete(req, ret);
3674 * IORING_OP_NOP just posts a completion event, nothing else.
3676 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3678 struct io_ring_ctx *ctx = req->ctx;
3680 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3683 __io_req_complete(req, 0, 0, cs);
3687 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3689 struct io_ring_ctx *ctx = req->ctx;
3694 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3696 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3699 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3700 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3703 req->sync.off = READ_ONCE(sqe->off);
3704 req->sync.len = READ_ONCE(sqe->len);
3708 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3710 loff_t end = req->sync.off + req->sync.len;
3713 /* fsync always requires a blocking context */
3717 ret = vfs_fsync_range(req->file, req->sync.off,
3718 end > 0 ? end : LLONG_MAX,
3719 req->sync.flags & IORING_FSYNC_DATASYNC);
3721 req_set_fail_links(req);
3722 io_req_complete(req, ret);
3726 static int io_fallocate_prep(struct io_kiocb *req,
3727 const struct io_uring_sqe *sqe)
3729 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3734 req->sync.off = READ_ONCE(sqe->off);
3735 req->sync.len = READ_ONCE(sqe->addr);
3736 req->sync.mode = READ_ONCE(sqe->len);
3740 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3744 /* fallocate always requiring blocking context */
3747 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3750 req_set_fail_links(req);
3751 io_req_complete(req, ret);
3755 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3757 const char __user *fname;
3760 if (unlikely(sqe->ioprio || sqe->buf_index))
3762 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3765 /* open.how should be already initialised */
3766 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3767 req->open.how.flags |= O_LARGEFILE;
3769 req->open.dfd = READ_ONCE(sqe->fd);
3770 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3771 req->open.filename = getname(fname);
3772 if (IS_ERR(req->open.filename)) {
3773 ret = PTR_ERR(req->open.filename);
3774 req->open.filename = NULL;
3777 req->open.nofile = rlimit(RLIMIT_NOFILE);
3778 req->flags |= REQ_F_NEED_CLEANUP;
3782 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3786 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3788 mode = READ_ONCE(sqe->len);
3789 flags = READ_ONCE(sqe->open_flags);
3790 req->open.how = build_open_how(flags, mode);
3791 return __io_openat_prep(req, sqe);
3794 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796 struct open_how __user *how;
3800 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3802 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3803 len = READ_ONCE(sqe->len);
3804 if (len < OPEN_HOW_SIZE_VER0)
3807 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3817 struct open_flags op;
3824 ret = build_open_flags(&req->open.how, &op);
3828 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3832 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3835 ret = PTR_ERR(file);
3837 fsnotify_open(file);
3838 fd_install(ret, file);
3841 putname(req->open.filename);
3842 req->flags &= ~REQ_F_NEED_CLEANUP;
3844 req_set_fail_links(req);
3845 io_req_complete(req, ret);
3849 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3851 return io_openat2(req, force_nonblock);
3854 static int io_remove_buffers_prep(struct io_kiocb *req,
3855 const struct io_uring_sqe *sqe)
3857 struct io_provide_buf *p = &req->pbuf;
3860 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3863 tmp = READ_ONCE(sqe->fd);
3864 if (!tmp || tmp > USHRT_MAX)
3867 memset(p, 0, sizeof(*p));
3869 p->bgid = READ_ONCE(sqe->buf_group);
3873 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3874 int bgid, unsigned nbufs)
3878 /* shouldn't happen */
3882 /* the head kbuf is the list itself */
3883 while (!list_empty(&buf->list)) {
3884 struct io_buffer *nxt;
3886 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3887 list_del(&nxt->list);
3894 idr_remove(&ctx->io_buffer_idr, bgid);
3899 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3900 struct io_comp_state *cs)
3902 struct io_provide_buf *p = &req->pbuf;
3903 struct io_ring_ctx *ctx = req->ctx;
3904 struct io_buffer *head;
3907 io_ring_submit_lock(ctx, !force_nonblock);
3909 lockdep_assert_held(&ctx->uring_lock);
3912 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3914 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3916 io_ring_submit_lock(ctx, !force_nonblock);
3918 req_set_fail_links(req);
3919 __io_req_complete(req, ret, 0, cs);
3923 static int io_provide_buffers_prep(struct io_kiocb *req,
3924 const struct io_uring_sqe *sqe)
3926 struct io_provide_buf *p = &req->pbuf;
3929 if (sqe->ioprio || sqe->rw_flags)
3932 tmp = READ_ONCE(sqe->fd);
3933 if (!tmp || tmp > USHRT_MAX)
3936 p->addr = READ_ONCE(sqe->addr);
3937 p->len = READ_ONCE(sqe->len);
3939 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3942 p->bgid = READ_ONCE(sqe->buf_group);
3943 tmp = READ_ONCE(sqe->off);
3944 if (tmp > USHRT_MAX)
3950 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3952 struct io_buffer *buf;
3953 u64 addr = pbuf->addr;
3954 int i, bid = pbuf->bid;
3956 for (i = 0; i < pbuf->nbufs; i++) {
3957 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3962 buf->len = pbuf->len;
3967 INIT_LIST_HEAD(&buf->list);
3970 list_add_tail(&buf->list, &(*head)->list);
3974 return i ? i : -ENOMEM;
3977 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3978 struct io_comp_state *cs)
3980 struct io_provide_buf *p = &req->pbuf;
3981 struct io_ring_ctx *ctx = req->ctx;
3982 struct io_buffer *head, *list;
3985 io_ring_submit_lock(ctx, !force_nonblock);
3987 lockdep_assert_held(&ctx->uring_lock);
3989 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3991 ret = io_add_buffers(p, &head);
3996 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3999 __io_remove_buffers(ctx, head, p->bgid, -1U);
4004 io_ring_submit_unlock(ctx, !force_nonblock);
4006 req_set_fail_links(req);
4007 __io_req_complete(req, ret, 0, cs);
4011 static int io_epoll_ctl_prep(struct io_kiocb *req,
4012 const struct io_uring_sqe *sqe)
4014 #if defined(CONFIG_EPOLL)
4015 if (sqe->ioprio || sqe->buf_index)
4017 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4020 req->epoll.epfd = READ_ONCE(sqe->fd);
4021 req->epoll.op = READ_ONCE(sqe->len);
4022 req->epoll.fd = READ_ONCE(sqe->off);
4024 if (ep_op_has_event(req->epoll.op)) {
4025 struct epoll_event __user *ev;
4027 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4028 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4038 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4039 struct io_comp_state *cs)
4041 #if defined(CONFIG_EPOLL)
4042 struct io_epoll *ie = &req->epoll;
4045 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4046 if (force_nonblock && ret == -EAGAIN)
4050 req_set_fail_links(req);
4051 __io_req_complete(req, ret, 0, cs);
4058 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4060 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4061 if (sqe->ioprio || sqe->buf_index || sqe->off)
4063 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4066 req->madvise.addr = READ_ONCE(sqe->addr);
4067 req->madvise.len = READ_ONCE(sqe->len);
4068 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4075 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4077 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4078 struct io_madvise *ma = &req->madvise;
4084 ret = do_madvise(ma->addr, ma->len, ma->advice);
4086 req_set_fail_links(req);
4087 io_req_complete(req, ret);
4094 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4096 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4098 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4101 req->fadvise.offset = READ_ONCE(sqe->off);
4102 req->fadvise.len = READ_ONCE(sqe->len);
4103 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4107 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4109 struct io_fadvise *fa = &req->fadvise;
4112 if (force_nonblock) {
4113 switch (fa->advice) {
4114 case POSIX_FADV_NORMAL:
4115 case POSIX_FADV_RANDOM:
4116 case POSIX_FADV_SEQUENTIAL:
4123 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4125 req_set_fail_links(req);
4126 io_req_complete(req, ret);
4130 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4132 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4134 if (sqe->ioprio || sqe->buf_index)
4136 if (req->flags & REQ_F_FIXED_FILE)
4139 req->statx.dfd = READ_ONCE(sqe->fd);
4140 req->statx.mask = READ_ONCE(sqe->len);
4141 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4142 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4143 req->statx.flags = READ_ONCE(sqe->statx_flags);
4148 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4150 struct io_statx *ctx = &req->statx;
4153 if (force_nonblock) {
4154 /* only need file table for an actual valid fd */
4155 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4156 req->flags |= REQ_F_NO_FILE_TABLE;
4160 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4164 req_set_fail_links(req);
4165 io_req_complete(req, ret);
4169 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4172 * If we queue this for async, it must not be cancellable. That would
4173 * leave the 'file' in an undeterminate state, and here need to modify
4174 * io_wq_work.flags, so initialize io_wq_work firstly.
4176 io_req_init_async(req);
4177 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4179 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4181 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4182 sqe->rw_flags || sqe->buf_index)
4184 if (req->flags & REQ_F_FIXED_FILE)
4187 req->close.fd = READ_ONCE(sqe->fd);
4188 if ((req->file && req->file->f_op == &io_uring_fops))
4191 req->close.put_file = NULL;
4195 static int io_close(struct io_kiocb *req, bool force_nonblock,
4196 struct io_comp_state *cs)
4198 struct io_close *close = &req->close;
4201 /* might be already done during nonblock submission */
4202 if (!close->put_file) {
4203 ret = __close_fd_get_file(close->fd, &close->put_file);
4205 return (ret == -ENOENT) ? -EBADF : ret;
4208 /* if the file has a flush method, be safe and punt to async */
4209 if (close->put_file->f_op->flush && force_nonblock) {
4210 /* was never set, but play safe */
4211 req->flags &= ~REQ_F_NOWAIT;
4212 /* avoid grabbing files - we don't need the files */
4213 req->flags |= REQ_F_NO_FILE_TABLE;
4217 /* No ->flush() or already async, safely close from here */
4218 ret = filp_close(close->put_file, req->work.identity->files);
4220 req_set_fail_links(req);
4221 fput(close->put_file);
4222 close->put_file = NULL;
4223 __io_req_complete(req, ret, 0, cs);
4227 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4229 struct io_ring_ctx *ctx = req->ctx;
4234 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4236 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4239 req->sync.off = READ_ONCE(sqe->off);
4240 req->sync.len = READ_ONCE(sqe->len);
4241 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4245 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4249 /* sync_file_range always requires a blocking context */
4253 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4256 req_set_fail_links(req);
4257 io_req_complete(req, ret);
4261 #if defined(CONFIG_NET)
4262 static int io_setup_async_msg(struct io_kiocb *req,
4263 struct io_async_msghdr *kmsg)
4265 struct io_async_msghdr *async_msg = req->async_data;
4269 if (io_alloc_async_data(req)) {
4270 if (kmsg->iov != kmsg->fast_iov)
4274 async_msg = req->async_data;
4275 req->flags |= REQ_F_NEED_CLEANUP;
4276 memcpy(async_msg, kmsg, sizeof(*kmsg));
4280 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4281 struct io_async_msghdr *iomsg)
4283 iomsg->iov = iomsg->fast_iov;
4284 iomsg->msg.msg_name = &iomsg->addr;
4285 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4286 req->sr_msg.msg_flags, &iomsg->iov);
4289 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4291 struct io_async_msghdr *async_msg = req->async_data;
4292 struct io_sr_msg *sr = &req->sr_msg;
4295 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4298 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4299 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4300 sr->len = READ_ONCE(sqe->len);
4302 #ifdef CONFIG_COMPAT
4303 if (req->ctx->compat)
4304 sr->msg_flags |= MSG_CMSG_COMPAT;
4307 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4309 ret = io_sendmsg_copy_hdr(req, async_msg);
4311 req->flags |= REQ_F_NEED_CLEANUP;
4315 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4316 struct io_comp_state *cs)
4318 struct io_async_msghdr iomsg, *kmsg;
4319 struct socket *sock;
4323 sock = sock_from_file(req->file, &ret);
4324 if (unlikely(!sock))
4327 if (req->async_data) {
4328 kmsg = req->async_data;
4329 kmsg->msg.msg_name = &kmsg->addr;
4330 /* if iov is set, it's allocated already */
4332 kmsg->iov = kmsg->fast_iov;
4333 kmsg->msg.msg_iter.iov = kmsg->iov;
4335 ret = io_sendmsg_copy_hdr(req, &iomsg);
4341 flags = req->sr_msg.msg_flags;
4342 if (flags & MSG_DONTWAIT)
4343 req->flags |= REQ_F_NOWAIT;
4344 else if (force_nonblock)
4345 flags |= MSG_DONTWAIT;
4347 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4348 if (force_nonblock && ret == -EAGAIN)
4349 return io_setup_async_msg(req, kmsg);
4350 if (ret == -ERESTARTSYS)
4353 if (kmsg->iov != kmsg->fast_iov)
4355 req->flags &= ~REQ_F_NEED_CLEANUP;
4357 req_set_fail_links(req);
4358 __io_req_complete(req, ret, 0, cs);
4362 static int io_send(struct io_kiocb *req, bool force_nonblock,
4363 struct io_comp_state *cs)
4365 struct io_sr_msg *sr = &req->sr_msg;
4368 struct socket *sock;
4372 sock = sock_from_file(req->file, &ret);
4373 if (unlikely(!sock))
4376 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4380 msg.msg_name = NULL;
4381 msg.msg_control = NULL;
4382 msg.msg_controllen = 0;
4383 msg.msg_namelen = 0;
4385 flags = req->sr_msg.msg_flags;
4386 if (flags & MSG_DONTWAIT)
4387 req->flags |= REQ_F_NOWAIT;
4388 else if (force_nonblock)
4389 flags |= MSG_DONTWAIT;
4391 msg.msg_flags = flags;
4392 ret = sock_sendmsg(sock, &msg);
4393 if (force_nonblock && ret == -EAGAIN)
4395 if (ret == -ERESTARTSYS)
4399 req_set_fail_links(req);
4400 __io_req_complete(req, ret, 0, cs);
4404 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4405 struct io_async_msghdr *iomsg)
4407 struct io_sr_msg *sr = &req->sr_msg;
4408 struct iovec __user *uiov;
4412 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4413 &iomsg->uaddr, &uiov, &iov_len);
4417 if (req->flags & REQ_F_BUFFER_SELECT) {
4420 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4422 sr->len = iomsg->iov[0].iov_len;
4423 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4427 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4428 &iomsg->iov, &iomsg->msg.msg_iter,
4437 #ifdef CONFIG_COMPAT
4438 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4439 struct io_async_msghdr *iomsg)
4441 struct compat_msghdr __user *msg_compat;
4442 struct io_sr_msg *sr = &req->sr_msg;
4443 struct compat_iovec __user *uiov;
4448 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4449 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4454 uiov = compat_ptr(ptr);
4455 if (req->flags & REQ_F_BUFFER_SELECT) {
4456 compat_ssize_t clen;
4460 if (!access_ok(uiov, sizeof(*uiov)))
4462 if (__get_user(clen, &uiov->iov_len))
4466 sr->len = iomsg->iov[0].iov_len;
4469 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4470 UIO_FASTIOV, &iomsg->iov,
4471 &iomsg->msg.msg_iter, true);
4480 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4481 struct io_async_msghdr *iomsg)
4483 iomsg->msg.msg_name = &iomsg->addr;
4484 iomsg->iov = iomsg->fast_iov;
4486 #ifdef CONFIG_COMPAT
4487 if (req->ctx->compat)
4488 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4491 return __io_recvmsg_copy_hdr(req, iomsg);
4494 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4497 struct io_sr_msg *sr = &req->sr_msg;
4498 struct io_buffer *kbuf;
4500 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4505 req->flags |= REQ_F_BUFFER_SELECTED;
4509 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4511 return io_put_kbuf(req, req->sr_msg.kbuf);
4514 static int io_recvmsg_prep(struct io_kiocb *req,
4515 const struct io_uring_sqe *sqe)
4517 struct io_async_msghdr *async_msg = req->async_data;
4518 struct io_sr_msg *sr = &req->sr_msg;
4521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4524 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4525 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4526 sr->len = READ_ONCE(sqe->len);
4527 sr->bgid = READ_ONCE(sqe->buf_group);
4529 #ifdef CONFIG_COMPAT
4530 if (req->ctx->compat)
4531 sr->msg_flags |= MSG_CMSG_COMPAT;
4534 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4536 ret = io_recvmsg_copy_hdr(req, async_msg);
4538 req->flags |= REQ_F_NEED_CLEANUP;
4542 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4543 struct io_comp_state *cs)
4545 struct io_async_msghdr iomsg, *kmsg;
4546 struct socket *sock;
4547 struct io_buffer *kbuf;
4549 int ret, cflags = 0;
4551 sock = sock_from_file(req->file, &ret);
4552 if (unlikely(!sock))
4555 if (req->async_data) {
4556 kmsg = req->async_data;
4557 kmsg->msg.msg_name = &kmsg->addr;
4558 /* if iov is set, it's allocated already */
4560 kmsg->iov = kmsg->fast_iov;
4561 kmsg->msg.msg_iter.iov = kmsg->iov;
4563 ret = io_recvmsg_copy_hdr(req, &iomsg);
4569 if (req->flags & REQ_F_BUFFER_SELECT) {
4570 kbuf = io_recv_buffer_select(req, !force_nonblock);
4572 return PTR_ERR(kbuf);
4573 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4574 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4575 1, req->sr_msg.len);
4578 flags = req->sr_msg.msg_flags;
4579 if (flags & MSG_DONTWAIT)
4580 req->flags |= REQ_F_NOWAIT;
4581 else if (force_nonblock)
4582 flags |= MSG_DONTWAIT;
4584 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4585 kmsg->uaddr, flags);
4586 if (force_nonblock && ret == -EAGAIN)
4587 return io_setup_async_msg(req, kmsg);
4588 if (ret == -ERESTARTSYS)
4591 if (req->flags & REQ_F_BUFFER_SELECTED)
4592 cflags = io_put_recv_kbuf(req);
4593 if (kmsg->iov != kmsg->fast_iov)
4595 req->flags &= ~REQ_F_NEED_CLEANUP;
4597 req_set_fail_links(req);
4598 __io_req_complete(req, ret, cflags, cs);
4602 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4603 struct io_comp_state *cs)
4605 struct io_buffer *kbuf;
4606 struct io_sr_msg *sr = &req->sr_msg;
4608 void __user *buf = sr->buf;
4609 struct socket *sock;
4612 int ret, cflags = 0;
4614 sock = sock_from_file(req->file, &ret);
4615 if (unlikely(!sock))
4618 if (req->flags & REQ_F_BUFFER_SELECT) {
4619 kbuf = io_recv_buffer_select(req, !force_nonblock);
4621 return PTR_ERR(kbuf);
4622 buf = u64_to_user_ptr(kbuf->addr);
4625 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4629 msg.msg_name = NULL;
4630 msg.msg_control = NULL;
4631 msg.msg_controllen = 0;
4632 msg.msg_namelen = 0;
4633 msg.msg_iocb = NULL;
4636 flags = req->sr_msg.msg_flags;
4637 if (flags & MSG_DONTWAIT)
4638 req->flags |= REQ_F_NOWAIT;
4639 else if (force_nonblock)
4640 flags |= MSG_DONTWAIT;
4642 ret = sock_recvmsg(sock, &msg, flags);
4643 if (force_nonblock && ret == -EAGAIN)
4645 if (ret == -ERESTARTSYS)
4648 if (req->flags & REQ_F_BUFFER_SELECTED)
4649 cflags = io_put_recv_kbuf(req);
4651 req_set_fail_links(req);
4652 __io_req_complete(req, ret, cflags, cs);
4656 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4658 struct io_accept *accept = &req->accept;
4660 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4662 if (sqe->ioprio || sqe->len || sqe->buf_index)
4665 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4666 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4667 accept->flags = READ_ONCE(sqe->accept_flags);
4668 accept->nofile = rlimit(RLIMIT_NOFILE);
4672 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4673 struct io_comp_state *cs)
4675 struct io_accept *accept = &req->accept;
4676 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4679 if (req->file->f_flags & O_NONBLOCK)
4680 req->flags |= REQ_F_NOWAIT;
4682 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4683 accept->addr_len, accept->flags,
4685 if (ret == -EAGAIN && force_nonblock)
4688 if (ret == -ERESTARTSYS)
4690 req_set_fail_links(req);
4692 __io_req_complete(req, ret, 0, cs);
4696 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4698 struct io_connect *conn = &req->connect;
4699 struct io_async_connect *io = req->async_data;
4701 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4703 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4706 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4707 conn->addr_len = READ_ONCE(sqe->addr2);
4712 return move_addr_to_kernel(conn->addr, conn->addr_len,
4716 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4717 struct io_comp_state *cs)
4719 struct io_async_connect __io, *io;
4720 unsigned file_flags;
4723 if (req->async_data) {
4724 io = req->async_data;
4726 ret = move_addr_to_kernel(req->connect.addr,
4727 req->connect.addr_len,
4734 file_flags = force_nonblock ? O_NONBLOCK : 0;
4736 ret = __sys_connect_file(req->file, &io->address,
4737 req->connect.addr_len, file_flags);
4738 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4739 if (req->async_data)
4741 if (io_alloc_async_data(req)) {
4745 io = req->async_data;
4746 memcpy(req->async_data, &__io, sizeof(__io));
4749 if (ret == -ERESTARTSYS)
4753 req_set_fail_links(req);
4754 __io_req_complete(req, ret, 0, cs);
4757 #else /* !CONFIG_NET */
4758 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4763 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4764 struct io_comp_state *cs)
4769 static int io_send(struct io_kiocb *req, bool force_nonblock,
4770 struct io_comp_state *cs)
4775 static int io_recvmsg_prep(struct io_kiocb *req,
4776 const struct io_uring_sqe *sqe)
4781 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4782 struct io_comp_state *cs)
4787 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4788 struct io_comp_state *cs)
4793 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4798 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4799 struct io_comp_state *cs)
4804 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4809 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4810 struct io_comp_state *cs)
4814 #endif /* CONFIG_NET */
4816 struct io_poll_table {
4817 struct poll_table_struct pt;
4818 struct io_kiocb *req;
4822 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4823 __poll_t mask, task_work_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 init_task_work(&req->task_work, func);
4838 percpu_ref_get(&req->ctx->refs);
4841 * If we using the signalfd wait_queue_head for this wakeup, then
4842 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4843 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4844 * either, as the normal wakeup will suffice.
4846 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4849 * If this fails, then the task is exiting. When a task exits, the
4850 * work gets canceled, so just cancel this request as well instead
4851 * of executing it. We can't safely execute it anyway, as we may not
4852 * have the needed state needed for it anyway.
4854 ret = io_req_task_work_add(req, twa_signal_ok);
4855 if (unlikely(ret)) {
4856 struct task_struct *tsk;
4858 WRITE_ONCE(poll->canceled, true);
4859 tsk = io_wq_get_task(req->ctx->io_wq);
4860 task_work_add(tsk, &req->task_work, 0);
4861 wake_up_process(tsk);
4866 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4867 __acquires(&req->ctx->completion_lock)
4869 struct io_ring_ctx *ctx = req->ctx;
4871 if (!req->result && !READ_ONCE(poll->canceled)) {
4872 struct poll_table_struct pt = { ._key = poll->events };
4874 req->result = vfs_poll(req->file, &pt) & poll->events;
4877 spin_lock_irq(&ctx->completion_lock);
4878 if (!req->result && !READ_ONCE(poll->canceled)) {
4879 add_wait_queue(poll->head, &poll->wait);
4886 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4888 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4889 if (req->opcode == IORING_OP_POLL_ADD)
4890 return req->async_data;
4891 return req->apoll->double_poll;
4894 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4896 if (req->opcode == IORING_OP_POLL_ADD)
4898 return &req->apoll->poll;
4901 static void io_poll_remove_double(struct io_kiocb *req)
4903 struct io_poll_iocb *poll = io_poll_get_double(req);
4905 lockdep_assert_held(&req->ctx->completion_lock);
4907 if (poll && poll->head) {
4908 struct wait_queue_head *head = poll->head;
4910 spin_lock(&head->lock);
4911 list_del_init(&poll->wait.entry);
4912 if (poll->wait.private)
4913 refcount_dec(&req->refs);
4915 spin_unlock(&head->lock);
4919 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4921 struct io_ring_ctx *ctx = req->ctx;
4923 io_poll_remove_double(req);
4924 req->poll.done = true;
4925 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4926 io_commit_cqring(ctx);
4929 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4931 struct io_ring_ctx *ctx = req->ctx;
4933 if (io_poll_rewait(req, &req->poll)) {
4934 spin_unlock_irq(&ctx->completion_lock);
4938 hash_del(&req->hash_node);
4939 io_poll_complete(req, req->result, 0);
4940 spin_unlock_irq(&ctx->completion_lock);
4942 *nxt = io_put_req_find_next(req);
4943 io_cqring_ev_posted(ctx);
4946 static void io_poll_task_func(struct callback_head *cb)
4948 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4949 struct io_ring_ctx *ctx = req->ctx;
4950 struct io_kiocb *nxt = NULL;
4952 io_poll_task_handler(req, &nxt);
4954 __io_req_task_submit(nxt);
4955 percpu_ref_put(&ctx->refs);
4958 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4959 int sync, void *key)
4961 struct io_kiocb *req = wait->private;
4962 struct io_poll_iocb *poll = io_poll_get_single(req);
4963 __poll_t mask = key_to_poll(key);
4965 /* for instances that support it check for an event match first: */
4966 if (mask && !(mask & poll->events))
4969 list_del_init(&wait->entry);
4971 if (poll && poll->head) {
4974 spin_lock(&poll->head->lock);
4975 done = list_empty(&poll->wait.entry);
4977 list_del_init(&poll->wait.entry);
4978 /* make sure double remove sees this as being gone */
4979 wait->private = NULL;
4980 spin_unlock(&poll->head->lock);
4982 __io_async_wake(req, poll, mask, io_poll_task_func);
4984 refcount_dec(&req->refs);
4988 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4989 wait_queue_func_t wake_func)
4993 poll->canceled = false;
4994 poll->events = events;
4995 INIT_LIST_HEAD(&poll->wait.entry);
4996 init_waitqueue_func_entry(&poll->wait, wake_func);
4999 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5000 struct wait_queue_head *head,
5001 struct io_poll_iocb **poll_ptr)
5003 struct io_kiocb *req = pt->req;
5006 * If poll->head is already set, it's because the file being polled
5007 * uses multiple waitqueues for poll handling (eg one for read, one
5008 * for write). Setup a separate io_poll_iocb if this happens.
5010 if (unlikely(poll->head)) {
5011 struct io_poll_iocb *poll_one = poll;
5013 /* already have a 2nd entry, fail a third attempt */
5015 pt->error = -EINVAL;
5018 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5020 pt->error = -ENOMEM;
5023 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5024 refcount_inc(&req->refs);
5025 poll->wait.private = req;
5032 if (poll->events & EPOLLEXCLUSIVE)
5033 add_wait_queue_exclusive(head, &poll->wait);
5035 add_wait_queue(head, &poll->wait);
5038 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5039 struct poll_table_struct *p)
5041 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5042 struct async_poll *apoll = pt->req->apoll;
5044 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5047 static void io_async_task_func(struct callback_head *cb)
5049 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5050 struct async_poll *apoll = req->apoll;
5051 struct io_ring_ctx *ctx = req->ctx;
5053 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5055 if (io_poll_rewait(req, &apoll->poll)) {
5056 spin_unlock_irq(&ctx->completion_lock);
5057 percpu_ref_put(&ctx->refs);
5061 /* If req is still hashed, it cannot have been canceled. Don't check. */
5062 if (hash_hashed(&req->hash_node))
5063 hash_del(&req->hash_node);
5065 io_poll_remove_double(req);
5066 spin_unlock_irq(&ctx->completion_lock);
5068 if (!READ_ONCE(apoll->poll.canceled))
5069 __io_req_task_submit(req);
5071 __io_req_task_cancel(req, -ECANCELED);
5073 percpu_ref_put(&ctx->refs);
5074 kfree(apoll->double_poll);
5078 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5081 struct io_kiocb *req = wait->private;
5082 struct io_poll_iocb *poll = &req->apoll->poll;
5084 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5087 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5090 static void io_poll_req_insert(struct io_kiocb *req)
5092 struct io_ring_ctx *ctx = req->ctx;
5093 struct hlist_head *list;
5095 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5096 hlist_add_head(&req->hash_node, list);
5099 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5100 struct io_poll_iocb *poll,
5101 struct io_poll_table *ipt, __poll_t mask,
5102 wait_queue_func_t wake_func)
5103 __acquires(&ctx->completion_lock)
5105 struct io_ring_ctx *ctx = req->ctx;
5106 bool cancel = false;
5108 io_init_poll_iocb(poll, mask, wake_func);
5109 poll->file = req->file;
5110 poll->wait.private = req;
5112 ipt->pt._key = mask;
5114 ipt->error = -EINVAL;
5116 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5118 spin_lock_irq(&ctx->completion_lock);
5119 if (likely(poll->head)) {
5120 spin_lock(&poll->head->lock);
5121 if (unlikely(list_empty(&poll->wait.entry))) {
5127 if (mask || ipt->error)
5128 list_del_init(&poll->wait.entry);
5130 WRITE_ONCE(poll->canceled, true);
5131 else if (!poll->done) /* actually waiting for an event */
5132 io_poll_req_insert(req);
5133 spin_unlock(&poll->head->lock);
5139 static bool io_arm_poll_handler(struct io_kiocb *req)
5141 const struct io_op_def *def = &io_op_defs[req->opcode];
5142 struct io_ring_ctx *ctx = req->ctx;
5143 struct async_poll *apoll;
5144 struct io_poll_table ipt;
5148 if (!req->file || !file_can_poll(req->file))
5150 if (req->flags & REQ_F_POLLED)
5154 else if (def->pollout)
5158 /* if we can't nonblock try, then no point in arming a poll handler */
5159 if (!io_file_supports_async(req->file, rw))
5162 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5163 if (unlikely(!apoll))
5165 apoll->double_poll = NULL;
5167 req->flags |= REQ_F_POLLED;
5169 INIT_HLIST_NODE(&req->hash_node);
5173 mask |= POLLIN | POLLRDNORM;
5175 mask |= POLLOUT | POLLWRNORM;
5177 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5178 if ((req->opcode == IORING_OP_RECVMSG) &&
5179 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5182 mask |= POLLERR | POLLPRI;
5184 ipt.pt._qproc = io_async_queue_proc;
5186 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5188 if (ret || ipt.error) {
5189 io_poll_remove_double(req);
5190 spin_unlock_irq(&ctx->completion_lock);
5191 kfree(apoll->double_poll);
5195 spin_unlock_irq(&ctx->completion_lock);
5196 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5197 apoll->poll.events);
5201 static bool __io_poll_remove_one(struct io_kiocb *req,
5202 struct io_poll_iocb *poll)
5204 bool do_complete = false;
5206 spin_lock(&poll->head->lock);
5207 WRITE_ONCE(poll->canceled, true);
5208 if (!list_empty(&poll->wait.entry)) {
5209 list_del_init(&poll->wait.entry);
5212 spin_unlock(&poll->head->lock);
5213 hash_del(&req->hash_node);
5217 static bool io_poll_remove_one(struct io_kiocb *req)
5221 io_poll_remove_double(req);
5223 if (req->opcode == IORING_OP_POLL_ADD) {
5224 do_complete = __io_poll_remove_one(req, &req->poll);
5226 struct async_poll *apoll = req->apoll;
5228 /* non-poll requests have submit ref still */
5229 do_complete = __io_poll_remove_one(req, &apoll->poll);
5232 kfree(apoll->double_poll);
5238 io_cqring_fill_event(req, -ECANCELED);
5239 io_commit_cqring(req->ctx);
5240 req_set_fail_links(req);
5241 io_put_req_deferred(req, 1);
5248 * Returns true if we found and killed one or more poll requests
5250 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5252 struct hlist_node *tmp;
5253 struct io_kiocb *req;
5256 spin_lock_irq(&ctx->completion_lock);
5257 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5258 struct hlist_head *list;
5260 list = &ctx->cancel_hash[i];
5261 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5262 if (io_task_match(req, tsk))
5263 posted += io_poll_remove_one(req);
5266 spin_unlock_irq(&ctx->completion_lock);
5269 io_cqring_ev_posted(ctx);
5274 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5276 struct hlist_head *list;
5277 struct io_kiocb *req;
5279 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5280 hlist_for_each_entry(req, list, hash_node) {
5281 if (sqe_addr != req->user_data)
5283 if (io_poll_remove_one(req))
5291 static int io_poll_remove_prep(struct io_kiocb *req,
5292 const struct io_uring_sqe *sqe)
5294 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5296 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5300 req->poll.addr = READ_ONCE(sqe->addr);
5305 * Find a running poll command that matches one specified in sqe->addr,
5306 * and remove it if found.
5308 static int io_poll_remove(struct io_kiocb *req)
5310 struct io_ring_ctx *ctx = req->ctx;
5314 addr = req->poll.addr;
5315 spin_lock_irq(&ctx->completion_lock);
5316 ret = io_poll_cancel(ctx, addr);
5317 spin_unlock_irq(&ctx->completion_lock);
5320 req_set_fail_links(req);
5321 io_req_complete(req, ret);
5325 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5328 struct io_kiocb *req = wait->private;
5329 struct io_poll_iocb *poll = &req->poll;
5331 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5334 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5335 struct poll_table_struct *p)
5337 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5339 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5342 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5344 struct io_poll_iocb *poll = &req->poll;
5347 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5349 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5354 events = READ_ONCE(sqe->poll32_events);
5356 events = swahw32(events);
5358 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5359 (events & EPOLLEXCLUSIVE);
5363 static int io_poll_add(struct io_kiocb *req)
5365 struct io_poll_iocb *poll = &req->poll;
5366 struct io_ring_ctx *ctx = req->ctx;
5367 struct io_poll_table ipt;
5370 INIT_HLIST_NODE(&req->hash_node);
5371 ipt.pt._qproc = io_poll_queue_proc;
5373 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5376 if (mask) { /* no async, we'd stolen it */
5378 io_poll_complete(req, mask, 0);
5380 spin_unlock_irq(&ctx->completion_lock);
5383 io_cqring_ev_posted(ctx);
5389 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5391 struct io_timeout_data *data = container_of(timer,
5392 struct io_timeout_data, timer);
5393 struct io_kiocb *req = data->req;
5394 struct io_ring_ctx *ctx = req->ctx;
5395 unsigned long flags;
5397 spin_lock_irqsave(&ctx->completion_lock, flags);
5398 list_del_init(&req->timeout.list);
5399 atomic_set(&req->ctx->cq_timeouts,
5400 atomic_read(&req->ctx->cq_timeouts) + 1);
5402 io_cqring_fill_event(req, -ETIME);
5403 io_commit_cqring(ctx);
5404 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5406 io_cqring_ev_posted(ctx);
5407 req_set_fail_links(req);
5409 return HRTIMER_NORESTART;
5412 static int __io_timeout_cancel(struct io_kiocb *req)
5414 struct io_timeout_data *io = req->async_data;
5417 ret = hrtimer_try_to_cancel(&io->timer);
5420 list_del_init(&req->timeout.list);
5422 req_set_fail_links(req);
5423 io_cqring_fill_event(req, -ECANCELED);
5424 io_put_req_deferred(req, 1);
5428 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5430 struct io_kiocb *req;
5433 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5434 if (user_data == req->user_data) {
5443 return __io_timeout_cancel(req);
5446 static int io_timeout_remove_prep(struct io_kiocb *req,
5447 const struct io_uring_sqe *sqe)
5449 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5451 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5453 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5456 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5461 * Remove or update an existing timeout command
5463 static int io_timeout_remove(struct io_kiocb *req)
5465 struct io_ring_ctx *ctx = req->ctx;
5468 spin_lock_irq(&ctx->completion_lock);
5469 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5471 io_cqring_fill_event(req, ret);
5472 io_commit_cqring(ctx);
5473 spin_unlock_irq(&ctx->completion_lock);
5474 io_cqring_ev_posted(ctx);
5476 req_set_fail_links(req);
5481 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5482 bool is_timeout_link)
5484 struct io_timeout_data *data;
5486 u32 off = READ_ONCE(sqe->off);
5488 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5490 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5492 if (off && is_timeout_link)
5494 flags = READ_ONCE(sqe->timeout_flags);
5495 if (flags & ~IORING_TIMEOUT_ABS)
5498 req->timeout.off = off;
5500 if (!req->async_data && io_alloc_async_data(req))
5503 data = req->async_data;
5506 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5509 if (flags & IORING_TIMEOUT_ABS)
5510 data->mode = HRTIMER_MODE_ABS;
5512 data->mode = HRTIMER_MODE_REL;
5514 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5518 static int io_timeout(struct io_kiocb *req)
5520 struct io_ring_ctx *ctx = req->ctx;
5521 struct io_timeout_data *data = req->async_data;
5522 struct list_head *entry;
5523 u32 tail, off = req->timeout.off;
5525 spin_lock_irq(&ctx->completion_lock);
5528 * sqe->off holds how many events that need to occur for this
5529 * timeout event to be satisfied. If it isn't set, then this is
5530 * a pure timeout request, sequence isn't used.
5532 if (io_is_timeout_noseq(req)) {
5533 entry = ctx->timeout_list.prev;
5537 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5538 req->timeout.target_seq = tail + off;
5541 * Insertion sort, ensuring the first entry in the list is always
5542 * the one we need first.
5544 list_for_each_prev(entry, &ctx->timeout_list) {
5545 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5548 if (io_is_timeout_noseq(nxt))
5550 /* nxt.seq is behind @tail, otherwise would've been completed */
5551 if (off >= nxt->timeout.target_seq - tail)
5555 list_add(&req->timeout.list, entry);
5556 data->timer.function = io_timeout_fn;
5557 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5558 spin_unlock_irq(&ctx->completion_lock);
5562 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5564 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5566 return req->user_data == (unsigned long) data;
5569 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5571 enum io_wq_cancel cancel_ret;
5574 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5575 switch (cancel_ret) {
5576 case IO_WQ_CANCEL_OK:
5579 case IO_WQ_CANCEL_RUNNING:
5582 case IO_WQ_CANCEL_NOTFOUND:
5590 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5591 struct io_kiocb *req, __u64 sqe_addr,
5594 unsigned long flags;
5597 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5598 if (ret != -ENOENT) {
5599 spin_lock_irqsave(&ctx->completion_lock, flags);
5603 spin_lock_irqsave(&ctx->completion_lock, flags);
5604 ret = io_timeout_cancel(ctx, sqe_addr);
5607 ret = io_poll_cancel(ctx, sqe_addr);
5611 io_cqring_fill_event(req, ret);
5612 io_commit_cqring(ctx);
5613 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5614 io_cqring_ev_posted(ctx);
5617 req_set_fail_links(req);
5621 static int io_async_cancel_prep(struct io_kiocb *req,
5622 const struct io_uring_sqe *sqe)
5624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5626 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5628 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5631 req->cancel.addr = READ_ONCE(sqe->addr);
5635 static int io_async_cancel(struct io_kiocb *req)
5637 struct io_ring_ctx *ctx = req->ctx;
5639 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5643 static int io_files_update_prep(struct io_kiocb *req,
5644 const struct io_uring_sqe *sqe)
5646 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5648 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5650 if (sqe->ioprio || sqe->rw_flags)
5653 req->files_update.offset = READ_ONCE(sqe->off);
5654 req->files_update.nr_args = READ_ONCE(sqe->len);
5655 if (!req->files_update.nr_args)
5657 req->files_update.arg = READ_ONCE(sqe->addr);
5661 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5662 struct io_comp_state *cs)
5664 struct io_ring_ctx *ctx = req->ctx;
5665 struct io_uring_files_update up;
5671 up.offset = req->files_update.offset;
5672 up.fds = req->files_update.arg;
5674 mutex_lock(&ctx->uring_lock);
5675 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5676 mutex_unlock(&ctx->uring_lock);
5679 req_set_fail_links(req);
5680 __io_req_complete(req, ret, 0, cs);
5684 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5686 switch (req->opcode) {
5689 case IORING_OP_READV:
5690 case IORING_OP_READ_FIXED:
5691 case IORING_OP_READ:
5692 return io_read_prep(req, sqe);
5693 case IORING_OP_WRITEV:
5694 case IORING_OP_WRITE_FIXED:
5695 case IORING_OP_WRITE:
5696 return io_write_prep(req, sqe);
5697 case IORING_OP_POLL_ADD:
5698 return io_poll_add_prep(req, sqe);
5699 case IORING_OP_POLL_REMOVE:
5700 return io_poll_remove_prep(req, sqe);
5701 case IORING_OP_FSYNC:
5702 return io_prep_fsync(req, sqe);
5703 case IORING_OP_SYNC_FILE_RANGE:
5704 return io_prep_sfr(req, sqe);
5705 case IORING_OP_SENDMSG:
5706 case IORING_OP_SEND:
5707 return io_sendmsg_prep(req, sqe);
5708 case IORING_OP_RECVMSG:
5709 case IORING_OP_RECV:
5710 return io_recvmsg_prep(req, sqe);
5711 case IORING_OP_CONNECT:
5712 return io_connect_prep(req, sqe);
5713 case IORING_OP_TIMEOUT:
5714 return io_timeout_prep(req, sqe, false);
5715 case IORING_OP_TIMEOUT_REMOVE:
5716 return io_timeout_remove_prep(req, sqe);
5717 case IORING_OP_ASYNC_CANCEL:
5718 return io_async_cancel_prep(req, sqe);
5719 case IORING_OP_LINK_TIMEOUT:
5720 return io_timeout_prep(req, sqe, true);
5721 case IORING_OP_ACCEPT:
5722 return io_accept_prep(req, sqe);
5723 case IORING_OP_FALLOCATE:
5724 return io_fallocate_prep(req, sqe);
5725 case IORING_OP_OPENAT:
5726 return io_openat_prep(req, sqe);
5727 case IORING_OP_CLOSE:
5728 return io_close_prep(req, sqe);
5729 case IORING_OP_FILES_UPDATE:
5730 return io_files_update_prep(req, sqe);
5731 case IORING_OP_STATX:
5732 return io_statx_prep(req, sqe);
5733 case IORING_OP_FADVISE:
5734 return io_fadvise_prep(req, sqe);
5735 case IORING_OP_MADVISE:
5736 return io_madvise_prep(req, sqe);
5737 case IORING_OP_OPENAT2:
5738 return io_openat2_prep(req, sqe);
5739 case IORING_OP_EPOLL_CTL:
5740 return io_epoll_ctl_prep(req, sqe);
5741 case IORING_OP_SPLICE:
5742 return io_splice_prep(req, sqe);
5743 case IORING_OP_PROVIDE_BUFFERS:
5744 return io_provide_buffers_prep(req, sqe);
5745 case IORING_OP_REMOVE_BUFFERS:
5746 return io_remove_buffers_prep(req, sqe);
5748 return io_tee_prep(req, sqe);
5751 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5756 static int io_req_defer_prep(struct io_kiocb *req,
5757 const struct io_uring_sqe *sqe)
5761 if (io_alloc_async_data(req))
5763 return io_req_prep(req, sqe);
5766 static u32 io_get_sequence(struct io_kiocb *req)
5768 struct io_kiocb *pos;
5769 struct io_ring_ctx *ctx = req->ctx;
5770 u32 total_submitted, nr_reqs = 1;
5772 if (req->flags & REQ_F_LINK_HEAD)
5773 list_for_each_entry(pos, &req->link_list, link_list)
5776 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5777 return total_submitted - nr_reqs;
5780 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5782 struct io_ring_ctx *ctx = req->ctx;
5783 struct io_defer_entry *de;
5787 /* Still need defer if there is pending req in defer list. */
5788 if (likely(list_empty_careful(&ctx->defer_list) &&
5789 !(req->flags & REQ_F_IO_DRAIN)))
5792 seq = io_get_sequence(req);
5793 /* Still a chance to pass the sequence check */
5794 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5797 if (!req->async_data) {
5798 ret = io_req_defer_prep(req, sqe);
5802 io_prep_async_link(req);
5803 de = kmalloc(sizeof(*de), GFP_KERNEL);
5807 spin_lock_irq(&ctx->completion_lock);
5808 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5809 spin_unlock_irq(&ctx->completion_lock);
5811 io_queue_async_work(req);
5812 return -EIOCBQUEUED;
5815 trace_io_uring_defer(ctx, req, req->user_data);
5818 list_add_tail(&de->list, &ctx->defer_list);
5819 spin_unlock_irq(&ctx->completion_lock);
5820 return -EIOCBQUEUED;
5823 static void io_req_drop_files(struct io_kiocb *req)
5825 struct io_ring_ctx *ctx = req->ctx;
5826 unsigned long flags;
5828 spin_lock_irqsave(&ctx->inflight_lock, flags);
5829 list_del(&req->inflight_entry);
5830 if (waitqueue_active(&ctx->inflight_wait))
5831 wake_up(&ctx->inflight_wait);
5832 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5833 req->flags &= ~REQ_F_INFLIGHT;
5834 put_files_struct(req->work.identity->files);
5835 put_nsproxy(req->work.identity->nsproxy);
5836 req->work.flags &= ~IO_WQ_WORK_FILES;
5839 static void __io_clean_op(struct io_kiocb *req)
5841 if (req->flags & REQ_F_BUFFER_SELECTED) {
5842 switch (req->opcode) {
5843 case IORING_OP_READV:
5844 case IORING_OP_READ_FIXED:
5845 case IORING_OP_READ:
5846 kfree((void *)(unsigned long)req->rw.addr);
5848 case IORING_OP_RECVMSG:
5849 case IORING_OP_RECV:
5850 kfree(req->sr_msg.kbuf);
5853 req->flags &= ~REQ_F_BUFFER_SELECTED;
5856 if (req->flags & REQ_F_NEED_CLEANUP) {
5857 switch (req->opcode) {
5858 case IORING_OP_READV:
5859 case IORING_OP_READ_FIXED:
5860 case IORING_OP_READ:
5861 case IORING_OP_WRITEV:
5862 case IORING_OP_WRITE_FIXED:
5863 case IORING_OP_WRITE: {
5864 struct io_async_rw *io = req->async_data;
5866 kfree(io->free_iovec);
5869 case IORING_OP_RECVMSG:
5870 case IORING_OP_SENDMSG: {
5871 struct io_async_msghdr *io = req->async_data;
5872 if (io->iov != io->fast_iov)
5876 case IORING_OP_SPLICE:
5878 io_put_file(req, req->splice.file_in,
5879 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5881 case IORING_OP_OPENAT:
5882 case IORING_OP_OPENAT2:
5883 if (req->open.filename)
5884 putname(req->open.filename);
5887 req->flags &= ~REQ_F_NEED_CLEANUP;
5890 if (req->flags & REQ_F_INFLIGHT)
5891 io_req_drop_files(req);
5894 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5895 struct io_comp_state *cs)
5897 struct io_ring_ctx *ctx = req->ctx;
5900 switch (req->opcode) {
5902 ret = io_nop(req, cs);
5904 case IORING_OP_READV:
5905 case IORING_OP_READ_FIXED:
5906 case IORING_OP_READ:
5907 ret = io_read(req, force_nonblock, cs);
5909 case IORING_OP_WRITEV:
5910 case IORING_OP_WRITE_FIXED:
5911 case IORING_OP_WRITE:
5912 ret = io_write(req, force_nonblock, cs);
5914 case IORING_OP_FSYNC:
5915 ret = io_fsync(req, force_nonblock);
5917 case IORING_OP_POLL_ADD:
5918 ret = io_poll_add(req);
5920 case IORING_OP_POLL_REMOVE:
5921 ret = io_poll_remove(req);
5923 case IORING_OP_SYNC_FILE_RANGE:
5924 ret = io_sync_file_range(req, force_nonblock);
5926 case IORING_OP_SENDMSG:
5927 ret = io_sendmsg(req, force_nonblock, cs);
5929 case IORING_OP_SEND:
5930 ret = io_send(req, force_nonblock, cs);
5932 case IORING_OP_RECVMSG:
5933 ret = io_recvmsg(req, force_nonblock, cs);
5935 case IORING_OP_RECV:
5936 ret = io_recv(req, force_nonblock, cs);
5938 case IORING_OP_TIMEOUT:
5939 ret = io_timeout(req);
5941 case IORING_OP_TIMEOUT_REMOVE:
5942 ret = io_timeout_remove(req);
5944 case IORING_OP_ACCEPT:
5945 ret = io_accept(req, force_nonblock, cs);
5947 case IORING_OP_CONNECT:
5948 ret = io_connect(req, force_nonblock, cs);
5950 case IORING_OP_ASYNC_CANCEL:
5951 ret = io_async_cancel(req);
5953 case IORING_OP_FALLOCATE:
5954 ret = io_fallocate(req, force_nonblock);
5956 case IORING_OP_OPENAT:
5957 ret = io_openat(req, force_nonblock);
5959 case IORING_OP_CLOSE:
5960 ret = io_close(req, force_nonblock, cs);
5962 case IORING_OP_FILES_UPDATE:
5963 ret = io_files_update(req, force_nonblock, cs);
5965 case IORING_OP_STATX:
5966 ret = io_statx(req, force_nonblock);
5968 case IORING_OP_FADVISE:
5969 ret = io_fadvise(req, force_nonblock);
5971 case IORING_OP_MADVISE:
5972 ret = io_madvise(req, force_nonblock);
5974 case IORING_OP_OPENAT2:
5975 ret = io_openat2(req, force_nonblock);
5977 case IORING_OP_EPOLL_CTL:
5978 ret = io_epoll_ctl(req, force_nonblock, cs);
5980 case IORING_OP_SPLICE:
5981 ret = io_splice(req, force_nonblock);
5983 case IORING_OP_PROVIDE_BUFFERS:
5984 ret = io_provide_buffers(req, force_nonblock, cs);
5986 case IORING_OP_REMOVE_BUFFERS:
5987 ret = io_remove_buffers(req, force_nonblock, cs);
5990 ret = io_tee(req, force_nonblock);
6000 /* If the op doesn't have a file, we're not polling for it */
6001 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6002 const bool in_async = io_wq_current_is_worker();
6004 /* workqueue context doesn't hold uring_lock, grab it now */
6006 mutex_lock(&ctx->uring_lock);
6008 io_iopoll_req_issued(req);
6011 mutex_unlock(&ctx->uring_lock);
6017 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6019 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6020 struct io_kiocb *timeout;
6023 timeout = io_prep_linked_timeout(req);
6025 io_queue_linked_timeout(timeout);
6027 /* if NO_CANCEL is set, we must still run the work */
6028 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6029 IO_WQ_WORK_CANCEL) {
6035 ret = io_issue_sqe(req, false, NULL);
6037 * We can get EAGAIN for polled IO even though we're
6038 * forcing a sync submission from here, since we can't
6039 * wait for request slots on the block side.
6048 req_set_fail_links(req);
6049 io_req_complete(req, ret);
6052 return io_steal_work(req);
6055 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6058 struct fixed_file_table *table;
6060 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6061 return table->files[index & IORING_FILE_TABLE_MASK];
6064 static struct file *io_file_get(struct io_submit_state *state,
6065 struct io_kiocb *req, int fd, bool fixed)
6067 struct io_ring_ctx *ctx = req->ctx;
6071 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6073 fd = array_index_nospec(fd, ctx->nr_user_files);
6074 file = io_file_from_index(ctx, fd);
6076 req->fixed_file_refs = &ctx->file_data->node->refs;
6077 percpu_ref_get(req->fixed_file_refs);
6080 trace_io_uring_file_get(ctx, fd);
6081 file = __io_file_get(state, fd);
6087 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6092 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6093 if (unlikely(!fixed && io_async_submit(req->ctx)))
6096 req->file = io_file_get(state, req, fd, fixed);
6097 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6102 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6104 struct io_timeout_data *data = container_of(timer,
6105 struct io_timeout_data, timer);
6106 struct io_kiocb *req = data->req;
6107 struct io_ring_ctx *ctx = req->ctx;
6108 struct io_kiocb *prev = NULL;
6109 unsigned long flags;
6111 spin_lock_irqsave(&ctx->completion_lock, flags);
6114 * We don't expect the list to be empty, that will only happen if we
6115 * race with the completion of the linked work.
6117 if (!list_empty(&req->link_list)) {
6118 prev = list_entry(req->link_list.prev, struct io_kiocb,
6120 if (refcount_inc_not_zero(&prev->refs)) {
6121 list_del_init(&req->link_list);
6122 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6127 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6130 req_set_fail_links(prev);
6131 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6134 io_req_complete(req, -ETIME);
6136 return HRTIMER_NORESTART;
6139 static void __io_queue_linked_timeout(struct io_kiocb *req)
6142 * If the list is now empty, then our linked request finished before
6143 * we got a chance to setup the timer
6145 if (!list_empty(&req->link_list)) {
6146 struct io_timeout_data *data = req->async_data;
6148 data->timer.function = io_link_timeout_fn;
6149 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6154 static void io_queue_linked_timeout(struct io_kiocb *req)
6156 struct io_ring_ctx *ctx = req->ctx;
6158 spin_lock_irq(&ctx->completion_lock);
6159 __io_queue_linked_timeout(req);
6160 spin_unlock_irq(&ctx->completion_lock);
6162 /* drop submission reference */
6166 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6168 struct io_kiocb *nxt;
6170 if (!(req->flags & REQ_F_LINK_HEAD))
6172 if (req->flags & REQ_F_LINK_TIMEOUT)
6175 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6177 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6180 req->flags |= REQ_F_LINK_TIMEOUT;
6184 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6186 struct io_kiocb *linked_timeout;
6187 struct io_kiocb *nxt;
6188 const struct cred *old_creds = NULL;
6192 linked_timeout = io_prep_linked_timeout(req);
6194 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.identity->creds &&
6195 req->work.identity->creds != current_cred()) {
6197 revert_creds(old_creds);
6198 if (old_creds == req->work.identity->creds)
6199 old_creds = NULL; /* restored original creds */
6201 old_creds = override_creds(req->work.identity->creds);
6202 req->work.flags |= IO_WQ_WORK_CREDS;
6205 ret = io_issue_sqe(req, true, cs);
6208 * We async punt it if the file wasn't marked NOWAIT, or if the file
6209 * doesn't support non-blocking read/write attempts
6211 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6212 if (!io_arm_poll_handler(req)) {
6215 * Queued up for async execution, worker will release
6216 * submit reference when the iocb is actually submitted.
6218 io_queue_async_work(req);
6222 io_queue_linked_timeout(linked_timeout);
6226 if (unlikely(ret)) {
6227 /* un-prep timeout, so it'll be killed as any other linked */
6228 req->flags &= ~REQ_F_LINK_TIMEOUT;
6229 req_set_fail_links(req);
6231 io_req_complete(req, ret);
6235 /* drop submission reference */
6236 nxt = io_put_req_find_next(req);
6238 io_queue_linked_timeout(linked_timeout);
6243 if (req->flags & REQ_F_FORCE_ASYNC)
6249 revert_creds(old_creds);
6252 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6253 struct io_comp_state *cs)
6257 ret = io_req_defer(req, sqe);
6259 if (ret != -EIOCBQUEUED) {
6261 req_set_fail_links(req);
6263 io_req_complete(req, ret);
6265 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6266 if (!req->async_data) {
6267 ret = io_req_defer_prep(req, sqe);
6273 * Never try inline submit of IOSQE_ASYNC is set, go straight
6274 * to async execution.
6276 io_req_init_async(req);
6277 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6278 io_queue_async_work(req);
6281 ret = io_req_prep(req, sqe);
6285 __io_queue_sqe(req, cs);
6289 static inline void io_queue_link_head(struct io_kiocb *req,
6290 struct io_comp_state *cs)
6292 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6294 io_req_complete(req, -ECANCELED);
6296 io_queue_sqe(req, NULL, cs);
6299 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6300 struct io_kiocb **link, struct io_comp_state *cs)
6302 struct io_ring_ctx *ctx = req->ctx;
6306 * If we already have a head request, queue this one for async
6307 * submittal once the head completes. If we don't have a head but
6308 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6309 * submitted sync once the chain is complete. If none of those
6310 * conditions are true (normal request), then just queue it.
6313 struct io_kiocb *head = *link;
6316 * Taking sequential execution of a link, draining both sides
6317 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6318 * requests in the link. So, it drains the head and the
6319 * next after the link request. The last one is done via
6320 * drain_next flag to persist the effect across calls.
6322 if (req->flags & REQ_F_IO_DRAIN) {
6323 head->flags |= REQ_F_IO_DRAIN;
6324 ctx->drain_next = 1;
6326 ret = io_req_defer_prep(req, sqe);
6327 if (unlikely(ret)) {
6328 /* fail even hard links since we don't submit */
6329 head->flags |= REQ_F_FAIL_LINK;
6332 trace_io_uring_link(ctx, req, head);
6333 list_add_tail(&req->link_list, &head->link_list);
6335 /* last request of a link, enqueue the link */
6336 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6337 io_queue_link_head(head, cs);
6341 if (unlikely(ctx->drain_next)) {
6342 req->flags |= REQ_F_IO_DRAIN;
6343 ctx->drain_next = 0;
6345 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6346 req->flags |= REQ_F_LINK_HEAD;
6347 INIT_LIST_HEAD(&req->link_list);
6349 ret = io_req_defer_prep(req, sqe);
6351 req->flags |= REQ_F_FAIL_LINK;
6354 io_queue_sqe(req, sqe, cs);
6362 * Batched submission is done, ensure local IO is flushed out.
6364 static void io_submit_state_end(struct io_submit_state *state)
6366 if (!list_empty(&state->comp.list))
6367 io_submit_flush_completions(&state->comp);
6368 blk_finish_plug(&state->plug);
6369 io_state_file_put(state);
6370 if (state->free_reqs)
6371 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6375 * Start submission side cache.
6377 static void io_submit_state_start(struct io_submit_state *state,
6378 struct io_ring_ctx *ctx, unsigned int max_ios)
6380 blk_start_plug(&state->plug);
6382 INIT_LIST_HEAD(&state->comp.list);
6383 state->comp.ctx = ctx;
6384 state->free_reqs = 0;
6386 state->ios_left = max_ios;
6389 static void io_commit_sqring(struct io_ring_ctx *ctx)
6391 struct io_rings *rings = ctx->rings;
6394 * Ensure any loads from the SQEs are done at this point,
6395 * since once we write the new head, the application could
6396 * write new data to them.
6398 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6402 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6403 * that is mapped by userspace. This means that care needs to be taken to
6404 * ensure that reads are stable, as we cannot rely on userspace always
6405 * being a good citizen. If members of the sqe are validated and then later
6406 * used, it's important that those reads are done through READ_ONCE() to
6407 * prevent a re-load down the line.
6409 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6411 u32 *sq_array = ctx->sq_array;
6415 * The cached sq head (or cq tail) serves two purposes:
6417 * 1) allows us to batch the cost of updating the user visible
6419 * 2) allows the kernel side to track the head on its own, even
6420 * though the application is the one updating it.
6422 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6423 if (likely(head < ctx->sq_entries))
6424 return &ctx->sq_sqes[head];
6426 /* drop invalid entries */
6427 ctx->cached_sq_dropped++;
6428 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6432 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6434 ctx->cached_sq_head++;
6438 * Check SQE restrictions (opcode and flags).
6440 * Returns 'true' if SQE is allowed, 'false' otherwise.
6442 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6443 struct io_kiocb *req,
6444 unsigned int sqe_flags)
6446 if (!ctx->restricted)
6449 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6452 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6453 ctx->restrictions.sqe_flags_required)
6456 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6457 ctx->restrictions.sqe_flags_required))
6463 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6464 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6465 IOSQE_BUFFER_SELECT)
6467 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6468 const struct io_uring_sqe *sqe,
6469 struct io_submit_state *state)
6471 unsigned int sqe_flags;
6474 req->opcode = READ_ONCE(sqe->opcode);
6475 req->user_data = READ_ONCE(sqe->user_data);
6476 req->async_data = NULL;
6480 /* one is dropped after submission, the other at completion */
6481 refcount_set(&req->refs, 2);
6482 req->task = current;
6485 if (unlikely(req->opcode >= IORING_OP_LAST))
6488 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6491 sqe_flags = READ_ONCE(sqe->flags);
6492 /* enforce forwards compatibility on users */
6493 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6496 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6499 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6500 !io_op_defs[req->opcode].buffer_select)
6503 id = READ_ONCE(sqe->personality);
6505 struct io_identity *iod;
6507 io_req_init_async(req);
6508 iod = idr_find(&ctx->personality_idr, id);
6511 refcount_inc(&iod->count);
6512 io_put_identity(current->io_uring, req);
6513 get_cred(iod->creds);
6514 req->work.identity = iod;
6515 req->work.flags |= IO_WQ_WORK_CREDS;
6518 /* same numerical values with corresponding REQ_F_*, safe to copy */
6519 req->flags |= sqe_flags;
6521 if (!io_op_defs[req->opcode].needs_file)
6524 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6529 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6531 struct io_submit_state state;
6532 struct io_kiocb *link = NULL;
6533 int i, submitted = 0;
6535 /* if we have a backlog and couldn't flush it all, return BUSY */
6536 if (test_bit(0, &ctx->sq_check_overflow)) {
6537 if (!list_empty(&ctx->cq_overflow_list) &&
6538 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6542 /* make sure SQ entry isn't read before tail */
6543 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6545 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6548 percpu_counter_add(¤t->io_uring->inflight, nr);
6549 refcount_add(nr, ¤t->usage);
6551 io_submit_state_start(&state, ctx, nr);
6553 for (i = 0; i < nr; i++) {
6554 const struct io_uring_sqe *sqe;
6555 struct io_kiocb *req;
6558 sqe = io_get_sqe(ctx);
6559 if (unlikely(!sqe)) {
6560 io_consume_sqe(ctx);
6563 req = io_alloc_req(ctx, &state);
6564 if (unlikely(!req)) {
6566 submitted = -EAGAIN;
6569 io_consume_sqe(ctx);
6570 /* will complete beyond this point, count as submitted */
6573 err = io_init_req(ctx, req, sqe, &state);
6574 if (unlikely(err)) {
6577 io_req_complete(req, err);
6581 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6582 true, io_async_submit(ctx));
6583 err = io_submit_sqe(req, sqe, &link, &state.comp);
6588 if (unlikely(submitted != nr)) {
6589 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6590 struct io_uring_task *tctx = current->io_uring;
6591 int unused = nr - ref_used;
6593 percpu_ref_put_many(&ctx->refs, unused);
6594 percpu_counter_sub(&tctx->inflight, unused);
6595 put_task_struct_many(current, unused);
6598 io_queue_link_head(link, &state.comp);
6599 io_submit_state_end(&state);
6601 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6602 io_commit_sqring(ctx);
6607 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6609 /* Tell userspace we may need a wakeup call */
6610 spin_lock_irq(&ctx->completion_lock);
6611 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6612 spin_unlock_irq(&ctx->completion_lock);
6615 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6617 spin_lock_irq(&ctx->completion_lock);
6618 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6619 spin_unlock_irq(&ctx->completion_lock);
6622 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6623 int sync, void *key)
6625 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6628 ret = autoremove_wake_function(wqe, mode, sync, key);
6630 unsigned long flags;
6632 spin_lock_irqsave(&ctx->completion_lock, flags);
6633 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6634 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6645 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6646 unsigned long start_jiffies, bool cap_entries)
6648 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6649 struct io_sq_data *sqd = ctx->sq_data;
6650 unsigned int to_submit;
6654 if (!list_empty(&ctx->iopoll_list)) {
6655 unsigned nr_events = 0;
6657 mutex_lock(&ctx->uring_lock);
6658 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6659 io_do_iopoll(ctx, &nr_events, 0);
6660 mutex_unlock(&ctx->uring_lock);
6663 to_submit = io_sqring_entries(ctx);
6666 * If submit got -EBUSY, flag us as needing the application
6667 * to enter the kernel to reap and flush events.
6669 if (!to_submit || ret == -EBUSY || need_resched()) {
6671 * Drop cur_mm before scheduling, we can't hold it for
6672 * long periods (or over schedule()). Do this before
6673 * adding ourselves to the waitqueue, as the unuse/drop
6676 io_sq_thread_drop_mm();
6679 * We're polling. If we're within the defined idle
6680 * period, then let us spin without work before going
6681 * to sleep. The exception is if we got EBUSY doing
6682 * more IO, we should wait for the application to
6683 * reap events and wake us up.
6685 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6686 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6687 !percpu_ref_is_dying(&ctx->refs)))
6690 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6691 TASK_INTERRUPTIBLE);
6694 * While doing polled IO, before going to sleep, we need
6695 * to check if there are new reqs added to iopoll_list,
6696 * it is because reqs may have been punted to io worker
6697 * and will be added to iopoll_list later, hence check
6698 * the iopoll_list again.
6700 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6701 !list_empty_careful(&ctx->iopoll_list)) {
6702 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6706 to_submit = io_sqring_entries(ctx);
6707 if (!to_submit || ret == -EBUSY)
6711 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6712 io_ring_clear_wakeup_flag(ctx);
6714 /* if we're handling multiple rings, cap submit size for fairness */
6715 if (cap_entries && to_submit > 8)
6718 mutex_lock(&ctx->uring_lock);
6719 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6720 ret = io_submit_sqes(ctx, to_submit);
6721 mutex_unlock(&ctx->uring_lock);
6723 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6724 wake_up(&ctx->sqo_sq_wait);
6726 return SQT_DID_WORK;
6729 static void io_sqd_init_new(struct io_sq_data *sqd)
6731 struct io_ring_ctx *ctx;
6733 while (!list_empty(&sqd->ctx_new_list)) {
6734 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6735 init_wait(&ctx->sqo_wait_entry);
6736 ctx->sqo_wait_entry.func = io_sq_wake_function;
6737 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6738 complete(&ctx->sq_thread_comp);
6742 static int io_sq_thread(void *data)
6744 struct cgroup_subsys_state *cur_css = NULL;
6745 const struct cred *old_cred = NULL;
6746 struct io_sq_data *sqd = data;
6747 struct io_ring_ctx *ctx;
6748 unsigned long start_jiffies;
6750 start_jiffies = jiffies;
6751 while (!kthread_should_stop()) {
6752 enum sq_ret ret = 0;
6756 * Any changes to the sqd lists are synchronized through the
6757 * kthread parking. This synchronizes the thread vs users,
6758 * the users are synchronized on the sqd->ctx_lock.
6760 if (kthread_should_park())
6763 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6764 io_sqd_init_new(sqd);
6766 cap_entries = !list_is_singular(&sqd->ctx_list);
6768 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6769 if (current->cred != ctx->creds) {
6771 revert_creds(old_cred);
6772 old_cred = override_creds(ctx->creds);
6774 io_sq_thread_associate_blkcg(ctx, &cur_css);
6776 current->loginuid = ctx->loginuid;
6777 current->sessionid = ctx->sessionid;
6780 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6782 io_sq_thread_drop_mm();
6785 if (ret & SQT_SPIN) {
6788 } else if (ret == SQT_IDLE) {
6789 if (kthread_should_park())
6791 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6792 io_ring_set_wakeup_flag(ctx);
6794 start_jiffies = jiffies;
6795 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6796 io_ring_clear_wakeup_flag(ctx);
6803 io_sq_thread_unassociate_blkcg();
6805 revert_creds(old_cred);
6812 struct io_wait_queue {
6813 struct wait_queue_entry wq;
6814 struct io_ring_ctx *ctx;
6816 unsigned nr_timeouts;
6819 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6821 struct io_ring_ctx *ctx = iowq->ctx;
6824 * Wake up if we have enough events, or if a timeout occurred since we
6825 * started waiting. For timeouts, we always want to return to userspace,
6826 * regardless of event count.
6828 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6829 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6832 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6833 int wake_flags, void *key)
6835 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6838 /* use noflush == true, as we can't safely rely on locking context */
6839 if (!io_should_wake(iowq, true))
6842 return autoremove_wake_function(curr, mode, wake_flags, key);
6845 static int io_run_task_work_sig(void)
6847 if (io_run_task_work())
6849 if (!signal_pending(current))
6851 if (current->jobctl & JOBCTL_TASK_WORK) {
6852 spin_lock_irq(¤t->sighand->siglock);
6853 current->jobctl &= ~JOBCTL_TASK_WORK;
6854 recalc_sigpending();
6855 spin_unlock_irq(¤t->sighand->siglock);
6862 * Wait until events become available, if we don't already have some. The
6863 * application must reap them itself, as they reside on the shared cq ring.
6865 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6866 const sigset_t __user *sig, size_t sigsz)
6868 struct io_wait_queue iowq = {
6871 .func = io_wake_function,
6872 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6875 .to_wait = min_events,
6877 struct io_rings *rings = ctx->rings;
6881 if (io_cqring_events(ctx, false) >= min_events)
6883 if (!io_run_task_work())
6888 #ifdef CONFIG_COMPAT
6889 if (in_compat_syscall())
6890 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6894 ret = set_user_sigmask(sig, sigsz);
6900 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6901 trace_io_uring_cqring_wait(ctx, min_events);
6903 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6904 TASK_INTERRUPTIBLE);
6905 /* make sure we run task_work before checking for signals */
6906 ret = io_run_task_work_sig();
6911 if (io_should_wake(&iowq, false))
6915 finish_wait(&ctx->wait, &iowq.wq);
6917 restore_saved_sigmask_unless(ret == -EINTR);
6919 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6922 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6924 #if defined(CONFIG_UNIX)
6925 if (ctx->ring_sock) {
6926 struct sock *sock = ctx->ring_sock->sk;
6927 struct sk_buff *skb;
6929 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6935 for (i = 0; i < ctx->nr_user_files; i++) {
6938 file = io_file_from_index(ctx, i);
6945 static void io_file_ref_kill(struct percpu_ref *ref)
6947 struct fixed_file_data *data;
6949 data = container_of(ref, struct fixed_file_data, refs);
6950 complete(&data->done);
6953 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6955 struct fixed_file_data *data = ctx->file_data;
6956 struct fixed_file_ref_node *ref_node = NULL;
6957 unsigned nr_tables, i;
6962 spin_lock(&data->lock);
6963 if (!list_empty(&data->ref_list))
6964 ref_node = list_first_entry(&data->ref_list,
6965 struct fixed_file_ref_node, node);
6966 spin_unlock(&data->lock);
6968 percpu_ref_kill(&ref_node->refs);
6970 percpu_ref_kill(&data->refs);
6972 /* wait for all refs nodes to complete */
6973 flush_delayed_work(&ctx->file_put_work);
6974 wait_for_completion(&data->done);
6976 __io_sqe_files_unregister(ctx);
6977 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6978 for (i = 0; i < nr_tables; i++)
6979 kfree(data->table[i].files);
6981 percpu_ref_exit(&data->refs);
6983 ctx->file_data = NULL;
6984 ctx->nr_user_files = 0;
6988 static void io_put_sq_data(struct io_sq_data *sqd)
6990 if (refcount_dec_and_test(&sqd->refs)) {
6992 * The park is a bit of a work-around, without it we get
6993 * warning spews on shutdown with SQPOLL set and affinity
6994 * set to a single CPU.
6997 kthread_park(sqd->thread);
6998 kthread_stop(sqd->thread);
7005 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7007 struct io_ring_ctx *ctx_attach;
7008 struct io_sq_data *sqd;
7011 f = fdget(p->wq_fd);
7013 return ERR_PTR(-ENXIO);
7014 if (f.file->f_op != &io_uring_fops) {
7016 return ERR_PTR(-EINVAL);
7019 ctx_attach = f.file->private_data;
7020 sqd = ctx_attach->sq_data;
7023 return ERR_PTR(-EINVAL);
7026 refcount_inc(&sqd->refs);
7031 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7033 struct io_sq_data *sqd;
7035 if (p->flags & IORING_SETUP_ATTACH_WQ)
7036 return io_attach_sq_data(p);
7038 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7040 return ERR_PTR(-ENOMEM);
7042 refcount_set(&sqd->refs, 1);
7043 INIT_LIST_HEAD(&sqd->ctx_list);
7044 INIT_LIST_HEAD(&sqd->ctx_new_list);
7045 mutex_init(&sqd->ctx_lock);
7046 mutex_init(&sqd->lock);
7047 init_waitqueue_head(&sqd->wait);
7051 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7052 __releases(&sqd->lock)
7056 kthread_unpark(sqd->thread);
7057 mutex_unlock(&sqd->lock);
7060 static void io_sq_thread_park(struct io_sq_data *sqd)
7061 __acquires(&sqd->lock)
7065 mutex_lock(&sqd->lock);
7066 kthread_park(sqd->thread);
7069 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7071 struct io_sq_data *sqd = ctx->sq_data;
7076 * We may arrive here from the error branch in
7077 * io_sq_offload_create() where the kthread is created
7078 * without being waked up, thus wake it up now to make
7079 * sure the wait will complete.
7081 wake_up_process(sqd->thread);
7082 wait_for_completion(&ctx->sq_thread_comp);
7084 io_sq_thread_park(sqd);
7087 mutex_lock(&sqd->ctx_lock);
7088 list_del(&ctx->sqd_list);
7089 mutex_unlock(&sqd->ctx_lock);
7092 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7093 io_sq_thread_unpark(sqd);
7096 io_put_sq_data(sqd);
7097 ctx->sq_data = NULL;
7101 static void io_finish_async(struct io_ring_ctx *ctx)
7103 io_sq_thread_stop(ctx);
7106 io_wq_destroy(ctx->io_wq);
7111 #if defined(CONFIG_UNIX)
7113 * Ensure the UNIX gc is aware of our file set, so we are certain that
7114 * the io_uring can be safely unregistered on process exit, even if we have
7115 * loops in the file referencing.
7117 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7119 struct sock *sk = ctx->ring_sock->sk;
7120 struct scm_fp_list *fpl;
7121 struct sk_buff *skb;
7124 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7128 skb = alloc_skb(0, GFP_KERNEL);
7137 fpl->user = get_uid(ctx->user);
7138 for (i = 0; i < nr; i++) {
7139 struct file *file = io_file_from_index(ctx, i + offset);
7143 fpl->fp[nr_files] = get_file(file);
7144 unix_inflight(fpl->user, fpl->fp[nr_files]);
7149 fpl->max = SCM_MAX_FD;
7150 fpl->count = nr_files;
7151 UNIXCB(skb).fp = fpl;
7152 skb->destructor = unix_destruct_scm;
7153 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7154 skb_queue_head(&sk->sk_receive_queue, skb);
7156 for (i = 0; i < nr_files; i++)
7167 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7168 * causes regular reference counting to break down. We rely on the UNIX
7169 * garbage collection to take care of this problem for us.
7171 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7173 unsigned left, total;
7177 left = ctx->nr_user_files;
7179 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7181 ret = __io_sqe_files_scm(ctx, this_files, total);
7185 total += this_files;
7191 while (total < ctx->nr_user_files) {
7192 struct file *file = io_file_from_index(ctx, total);
7202 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7208 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7209 unsigned nr_tables, unsigned nr_files)
7213 for (i = 0; i < nr_tables; i++) {
7214 struct fixed_file_table *table = &file_data->table[i];
7215 unsigned this_files;
7217 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7218 table->files = kcalloc(this_files, sizeof(struct file *),
7222 nr_files -= this_files;
7228 for (i = 0; i < nr_tables; i++) {
7229 struct fixed_file_table *table = &file_data->table[i];
7230 kfree(table->files);
7235 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7237 #if defined(CONFIG_UNIX)
7238 struct sock *sock = ctx->ring_sock->sk;
7239 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7240 struct sk_buff *skb;
7243 __skb_queue_head_init(&list);
7246 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7247 * remove this entry and rearrange the file array.
7249 skb = skb_dequeue(head);
7251 struct scm_fp_list *fp;
7253 fp = UNIXCB(skb).fp;
7254 for (i = 0; i < fp->count; i++) {
7257 if (fp->fp[i] != file)
7260 unix_notinflight(fp->user, fp->fp[i]);
7261 left = fp->count - 1 - i;
7263 memmove(&fp->fp[i], &fp->fp[i + 1],
7264 left * sizeof(struct file *));
7271 __skb_queue_tail(&list, skb);
7281 __skb_queue_tail(&list, skb);
7283 skb = skb_dequeue(head);
7286 if (skb_peek(&list)) {
7287 spin_lock_irq(&head->lock);
7288 while ((skb = __skb_dequeue(&list)) != NULL)
7289 __skb_queue_tail(head, skb);
7290 spin_unlock_irq(&head->lock);
7297 struct io_file_put {
7298 struct list_head list;
7302 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7304 struct fixed_file_data *file_data = ref_node->file_data;
7305 struct io_ring_ctx *ctx = file_data->ctx;
7306 struct io_file_put *pfile, *tmp;
7308 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7309 list_del(&pfile->list);
7310 io_ring_file_put(ctx, pfile->file);
7314 spin_lock(&file_data->lock);
7315 list_del(&ref_node->node);
7316 spin_unlock(&file_data->lock);
7318 percpu_ref_exit(&ref_node->refs);
7320 percpu_ref_put(&file_data->refs);
7323 static void io_file_put_work(struct work_struct *work)
7325 struct io_ring_ctx *ctx;
7326 struct llist_node *node;
7328 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7329 node = llist_del_all(&ctx->file_put_llist);
7332 struct fixed_file_ref_node *ref_node;
7333 struct llist_node *next = node->next;
7335 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7336 __io_file_put_work(ref_node);
7341 static void io_file_data_ref_zero(struct percpu_ref *ref)
7343 struct fixed_file_ref_node *ref_node;
7344 struct io_ring_ctx *ctx;
7348 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7349 ctx = ref_node->file_data->ctx;
7351 if (percpu_ref_is_dying(&ctx->file_data->refs))
7354 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7356 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7358 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7361 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7362 struct io_ring_ctx *ctx)
7364 struct fixed_file_ref_node *ref_node;
7366 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7368 return ERR_PTR(-ENOMEM);
7370 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7373 return ERR_PTR(-ENOMEM);
7375 INIT_LIST_HEAD(&ref_node->node);
7376 INIT_LIST_HEAD(&ref_node->file_list);
7377 ref_node->file_data = ctx->file_data;
7381 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7383 percpu_ref_exit(&ref_node->refs);
7387 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7390 __s32 __user *fds = (__s32 __user *) arg;
7391 unsigned nr_tables, i;
7393 int fd, ret = -ENOMEM;
7394 struct fixed_file_ref_node *ref_node;
7395 struct fixed_file_data *file_data;
7401 if (nr_args > IORING_MAX_FIXED_FILES)
7404 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7407 file_data->ctx = ctx;
7408 init_completion(&file_data->done);
7409 INIT_LIST_HEAD(&file_data->ref_list);
7410 spin_lock_init(&file_data->lock);
7412 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7413 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7415 if (!file_data->table)
7418 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7419 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7422 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7424 ctx->file_data = file_data;
7426 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7427 struct fixed_file_table *table;
7430 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7434 /* allow sparse sets */
7444 * Don't allow io_uring instances to be registered. If UNIX
7445 * isn't enabled, then this causes a reference cycle and this
7446 * instance can never get freed. If UNIX is enabled we'll
7447 * handle it just fine, but there's still no point in allowing
7448 * a ring fd as it doesn't support regular read/write anyway.
7450 if (file->f_op == &io_uring_fops) {
7454 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7455 index = i & IORING_FILE_TABLE_MASK;
7456 table->files[index] = file;
7459 ret = io_sqe_files_scm(ctx);
7461 io_sqe_files_unregister(ctx);
7465 ref_node = alloc_fixed_file_ref_node(ctx);
7466 if (IS_ERR(ref_node)) {
7467 io_sqe_files_unregister(ctx);
7468 return PTR_ERR(ref_node);
7471 file_data->node = ref_node;
7472 spin_lock(&file_data->lock);
7473 list_add(&ref_node->node, &file_data->ref_list);
7474 spin_unlock(&file_data->lock);
7475 percpu_ref_get(&file_data->refs);
7478 for (i = 0; i < ctx->nr_user_files; i++) {
7479 file = io_file_from_index(ctx, i);
7483 for (i = 0; i < nr_tables; i++)
7484 kfree(file_data->table[i].files);
7485 ctx->nr_user_files = 0;
7487 percpu_ref_exit(&file_data->refs);
7489 kfree(file_data->table);
7491 ctx->file_data = NULL;
7495 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7498 #if defined(CONFIG_UNIX)
7499 struct sock *sock = ctx->ring_sock->sk;
7500 struct sk_buff_head *head = &sock->sk_receive_queue;
7501 struct sk_buff *skb;
7504 * See if we can merge this file into an existing skb SCM_RIGHTS
7505 * file set. If there's no room, fall back to allocating a new skb
7506 * and filling it in.
7508 spin_lock_irq(&head->lock);
7509 skb = skb_peek(head);
7511 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7513 if (fpl->count < SCM_MAX_FD) {
7514 __skb_unlink(skb, head);
7515 spin_unlock_irq(&head->lock);
7516 fpl->fp[fpl->count] = get_file(file);
7517 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7519 spin_lock_irq(&head->lock);
7520 __skb_queue_head(head, skb);
7525 spin_unlock_irq(&head->lock);
7532 return __io_sqe_files_scm(ctx, 1, index);
7538 static int io_queue_file_removal(struct fixed_file_data *data,
7541 struct io_file_put *pfile;
7542 struct fixed_file_ref_node *ref_node = data->node;
7544 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7549 list_add(&pfile->list, &ref_node->file_list);
7554 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7555 struct io_uring_files_update *up,
7558 struct fixed_file_data *data = ctx->file_data;
7559 struct fixed_file_ref_node *ref_node;
7564 bool needs_switch = false;
7566 if (check_add_overflow(up->offset, nr_args, &done))
7568 if (done > ctx->nr_user_files)
7571 ref_node = alloc_fixed_file_ref_node(ctx);
7572 if (IS_ERR(ref_node))
7573 return PTR_ERR(ref_node);
7576 fds = u64_to_user_ptr(up->fds);
7578 struct fixed_file_table *table;
7582 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7586 i = array_index_nospec(up->offset, ctx->nr_user_files);
7587 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7588 index = i & IORING_FILE_TABLE_MASK;
7589 if (table->files[index]) {
7590 file = table->files[index];
7591 err = io_queue_file_removal(data, file);
7594 table->files[index] = NULL;
7595 needs_switch = true;
7604 * Don't allow io_uring instances to be registered. If
7605 * UNIX isn't enabled, then this causes a reference
7606 * cycle and this instance can never get freed. If UNIX
7607 * is enabled we'll handle it just fine, but there's
7608 * still no point in allowing a ring fd as it doesn't
7609 * support regular read/write anyway.
7611 if (file->f_op == &io_uring_fops) {
7616 table->files[index] = file;
7617 err = io_sqe_file_register(ctx, file, i);
7619 table->files[index] = NULL;
7630 percpu_ref_kill(&data->node->refs);
7631 spin_lock(&data->lock);
7632 list_add(&ref_node->node, &data->ref_list);
7633 data->node = ref_node;
7634 spin_unlock(&data->lock);
7635 percpu_ref_get(&ctx->file_data->refs);
7637 destroy_fixed_file_ref_node(ref_node);
7639 return done ? done : err;
7642 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7645 struct io_uring_files_update up;
7647 if (!ctx->file_data)
7651 if (copy_from_user(&up, arg, sizeof(up)))
7656 return __io_sqe_files_update(ctx, &up, nr_args);
7659 static void io_free_work(struct io_wq_work *work)
7661 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7663 /* Consider that io_steal_work() relies on this ref */
7667 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7668 struct io_uring_params *p)
7670 struct io_wq_data data;
7672 struct io_ring_ctx *ctx_attach;
7673 unsigned int concurrency;
7676 data.user = ctx->user;
7677 data.free_work = io_free_work;
7678 data.do_work = io_wq_submit_work;
7680 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7681 /* Do QD, or 4 * CPUS, whatever is smallest */
7682 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7684 ctx->io_wq = io_wq_create(concurrency, &data);
7685 if (IS_ERR(ctx->io_wq)) {
7686 ret = PTR_ERR(ctx->io_wq);
7692 f = fdget(p->wq_fd);
7696 if (f.file->f_op != &io_uring_fops) {
7701 ctx_attach = f.file->private_data;
7702 /* @io_wq is protected by holding the fd */
7703 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7708 ctx->io_wq = ctx_attach->io_wq;
7714 static int io_uring_alloc_task_context(struct task_struct *task)
7716 struct io_uring_task *tctx;
7719 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7720 if (unlikely(!tctx))
7723 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7724 if (unlikely(ret)) {
7730 init_waitqueue_head(&tctx->wait);
7733 io_init_identity(&tctx->__identity);
7734 tctx->identity = &tctx->__identity;
7735 task->io_uring = tctx;
7739 void __io_uring_free(struct task_struct *tsk)
7741 struct io_uring_task *tctx = tsk->io_uring;
7743 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7744 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7745 if (tctx->identity != &tctx->__identity)
7746 kfree(tctx->identity);
7747 percpu_counter_destroy(&tctx->inflight);
7749 tsk->io_uring = NULL;
7752 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7753 struct io_uring_params *p)
7757 if (ctx->flags & IORING_SETUP_SQPOLL) {
7758 struct io_sq_data *sqd;
7761 if (!capable(CAP_SYS_ADMIN))
7764 sqd = io_get_sq_data(p);
7771 io_sq_thread_park(sqd);
7772 mutex_lock(&sqd->ctx_lock);
7773 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7774 mutex_unlock(&sqd->ctx_lock);
7775 io_sq_thread_unpark(sqd);
7777 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7778 if (!ctx->sq_thread_idle)
7779 ctx->sq_thread_idle = HZ;
7784 if (p->flags & IORING_SETUP_SQ_AFF) {
7785 int cpu = p->sq_thread_cpu;
7788 if (cpu >= nr_cpu_ids)
7790 if (!cpu_online(cpu))
7793 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7794 cpu, "io_uring-sq");
7796 sqd->thread = kthread_create(io_sq_thread, sqd,
7799 if (IS_ERR(sqd->thread)) {
7800 ret = PTR_ERR(sqd->thread);
7804 ret = io_uring_alloc_task_context(sqd->thread);
7807 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7808 /* Can't have SQ_AFF without SQPOLL */
7814 ret = io_init_wq_offload(ctx, p);
7820 io_finish_async(ctx);
7824 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7826 struct io_sq_data *sqd = ctx->sq_data;
7828 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7829 wake_up_process(sqd->thread);
7832 static inline void __io_unaccount_mem(struct user_struct *user,
7833 unsigned long nr_pages)
7835 atomic_long_sub(nr_pages, &user->locked_vm);
7838 static inline int __io_account_mem(struct user_struct *user,
7839 unsigned long nr_pages)
7841 unsigned long page_limit, cur_pages, new_pages;
7843 /* Don't allow more pages than we can safely lock */
7844 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7847 cur_pages = atomic_long_read(&user->locked_vm);
7848 new_pages = cur_pages + nr_pages;
7849 if (new_pages > page_limit)
7851 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7852 new_pages) != cur_pages);
7857 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7858 enum io_mem_account acct)
7861 __io_unaccount_mem(ctx->user, nr_pages);
7863 if (ctx->mm_account) {
7864 if (acct == ACCT_LOCKED)
7865 ctx->mm_account->locked_vm -= nr_pages;
7866 else if (acct == ACCT_PINNED)
7867 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7871 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7872 enum io_mem_account acct)
7876 if (ctx->limit_mem) {
7877 ret = __io_account_mem(ctx->user, nr_pages);
7882 if (ctx->mm_account) {
7883 if (acct == ACCT_LOCKED)
7884 ctx->mm_account->locked_vm += nr_pages;
7885 else if (acct == ACCT_PINNED)
7886 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7892 static void io_mem_free(void *ptr)
7899 page = virt_to_head_page(ptr);
7900 if (put_page_testzero(page))
7901 free_compound_page(page);
7904 static void *io_mem_alloc(size_t size)
7906 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7909 return (void *) __get_free_pages(gfp_flags, get_order(size));
7912 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7915 struct io_rings *rings;
7916 size_t off, sq_array_size;
7918 off = struct_size(rings, cqes, cq_entries);
7919 if (off == SIZE_MAX)
7923 off = ALIGN(off, SMP_CACHE_BYTES);
7931 sq_array_size = array_size(sizeof(u32), sq_entries);
7932 if (sq_array_size == SIZE_MAX)
7935 if (check_add_overflow(off, sq_array_size, &off))
7941 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7945 pages = (size_t)1 << get_order(
7946 rings_size(sq_entries, cq_entries, NULL));
7947 pages += (size_t)1 << get_order(
7948 array_size(sizeof(struct io_uring_sqe), sq_entries));
7953 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7957 if (!ctx->user_bufs)
7960 for (i = 0; i < ctx->nr_user_bufs; i++) {
7961 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7963 for (j = 0; j < imu->nr_bvecs; j++)
7964 unpin_user_page(imu->bvec[j].bv_page);
7966 if (imu->acct_pages)
7967 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7972 kfree(ctx->user_bufs);
7973 ctx->user_bufs = NULL;
7974 ctx->nr_user_bufs = 0;
7978 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7979 void __user *arg, unsigned index)
7981 struct iovec __user *src;
7983 #ifdef CONFIG_COMPAT
7985 struct compat_iovec __user *ciovs;
7986 struct compat_iovec ciov;
7988 ciovs = (struct compat_iovec __user *) arg;
7989 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7992 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7993 dst->iov_len = ciov.iov_len;
7997 src = (struct iovec __user *) arg;
7998 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8004 * Not super efficient, but this is just a registration time. And we do cache
8005 * the last compound head, so generally we'll only do a full search if we don't
8008 * We check if the given compound head page has already been accounted, to
8009 * avoid double accounting it. This allows us to account the full size of the
8010 * page, not just the constituent pages of a huge page.
8012 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8013 int nr_pages, struct page *hpage)
8017 /* check current page array */
8018 for (i = 0; i < nr_pages; i++) {
8019 if (!PageCompound(pages[i]))
8021 if (compound_head(pages[i]) == hpage)
8025 /* check previously registered pages */
8026 for (i = 0; i < ctx->nr_user_bufs; i++) {
8027 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8029 for (j = 0; j < imu->nr_bvecs; j++) {
8030 if (!PageCompound(imu->bvec[j].bv_page))
8032 if (compound_head(imu->bvec[j].bv_page) == hpage)
8040 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8041 int nr_pages, struct io_mapped_ubuf *imu,
8042 struct page **last_hpage)
8046 for (i = 0; i < nr_pages; i++) {
8047 if (!PageCompound(pages[i])) {
8052 hpage = compound_head(pages[i]);
8053 if (hpage == *last_hpage)
8055 *last_hpage = hpage;
8056 if (headpage_already_acct(ctx, pages, i, hpage))
8058 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8062 if (!imu->acct_pages)
8065 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8067 imu->acct_pages = 0;
8071 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8074 struct vm_area_struct **vmas = NULL;
8075 struct page **pages = NULL;
8076 struct page *last_hpage = NULL;
8077 int i, j, got_pages = 0;
8082 if (!nr_args || nr_args > UIO_MAXIOV)
8085 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8087 if (!ctx->user_bufs)
8090 for (i = 0; i < nr_args; i++) {
8091 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8092 unsigned long off, start, end, ubuf;
8097 ret = io_copy_iov(ctx, &iov, arg, i);
8102 * Don't impose further limits on the size and buffer
8103 * constraints here, we'll -EINVAL later when IO is
8104 * submitted if they are wrong.
8107 if (!iov.iov_base || !iov.iov_len)
8110 /* arbitrary limit, but we need something */
8111 if (iov.iov_len > SZ_1G)
8114 ubuf = (unsigned long) iov.iov_base;
8115 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8116 start = ubuf >> PAGE_SHIFT;
8117 nr_pages = end - start;
8120 if (!pages || nr_pages > got_pages) {
8123 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8125 vmas = kvmalloc_array(nr_pages,
8126 sizeof(struct vm_area_struct *),
8128 if (!pages || !vmas) {
8132 got_pages = nr_pages;
8135 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8142 mmap_read_lock(current->mm);
8143 pret = pin_user_pages(ubuf, nr_pages,
8144 FOLL_WRITE | FOLL_LONGTERM,
8146 if (pret == nr_pages) {
8147 /* don't support file backed memory */
8148 for (j = 0; j < nr_pages; j++) {
8149 struct vm_area_struct *vma = vmas[j];
8152 !is_file_hugepages(vma->vm_file)) {
8158 ret = pret < 0 ? pret : -EFAULT;
8160 mmap_read_unlock(current->mm);
8163 * if we did partial map, or found file backed vmas,
8164 * release any pages we did get
8167 unpin_user_pages(pages, pret);
8172 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8174 unpin_user_pages(pages, pret);
8179 off = ubuf & ~PAGE_MASK;
8181 for (j = 0; j < nr_pages; j++) {
8184 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8185 imu->bvec[j].bv_page = pages[j];
8186 imu->bvec[j].bv_len = vec_len;
8187 imu->bvec[j].bv_offset = off;
8191 /* store original address for later verification */
8193 imu->len = iov.iov_len;
8194 imu->nr_bvecs = nr_pages;
8196 ctx->nr_user_bufs++;
8204 io_sqe_buffer_unregister(ctx);
8208 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8210 __s32 __user *fds = arg;
8216 if (copy_from_user(&fd, fds, sizeof(*fds)))
8219 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8220 if (IS_ERR(ctx->cq_ev_fd)) {
8221 int ret = PTR_ERR(ctx->cq_ev_fd);
8222 ctx->cq_ev_fd = NULL;
8229 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8231 if (ctx->cq_ev_fd) {
8232 eventfd_ctx_put(ctx->cq_ev_fd);
8233 ctx->cq_ev_fd = NULL;
8240 static int __io_destroy_buffers(int id, void *p, void *data)
8242 struct io_ring_ctx *ctx = data;
8243 struct io_buffer *buf = p;
8245 __io_remove_buffers(ctx, buf, id, -1U);
8249 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8251 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8252 idr_destroy(&ctx->io_buffer_idr);
8255 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8257 io_finish_async(ctx);
8258 io_sqe_buffer_unregister(ctx);
8260 if (ctx->sqo_task) {
8261 put_task_struct(ctx->sqo_task);
8262 ctx->sqo_task = NULL;
8263 mmdrop(ctx->mm_account);
8264 ctx->mm_account = NULL;
8267 #ifdef CONFIG_BLK_CGROUP
8268 if (ctx->sqo_blkcg_css)
8269 css_put(ctx->sqo_blkcg_css);
8272 io_sqe_files_unregister(ctx);
8273 io_eventfd_unregister(ctx);
8274 io_destroy_buffers(ctx);
8275 idr_destroy(&ctx->personality_idr);
8277 #if defined(CONFIG_UNIX)
8278 if (ctx->ring_sock) {
8279 ctx->ring_sock->file = NULL; /* so that iput() is called */
8280 sock_release(ctx->ring_sock);
8284 io_mem_free(ctx->rings);
8285 io_mem_free(ctx->sq_sqes);
8287 percpu_ref_exit(&ctx->refs);
8288 free_uid(ctx->user);
8289 put_cred(ctx->creds);
8290 kfree(ctx->cancel_hash);
8291 kmem_cache_free(req_cachep, ctx->fallback_req);
8295 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8297 struct io_ring_ctx *ctx = file->private_data;
8300 poll_wait(file, &ctx->cq_wait, wait);
8302 * synchronizes with barrier from wq_has_sleeper call in
8306 if (!io_sqring_full(ctx))
8307 mask |= EPOLLOUT | EPOLLWRNORM;
8308 if (io_cqring_events(ctx, false))
8309 mask |= EPOLLIN | EPOLLRDNORM;
8314 static int io_uring_fasync(int fd, struct file *file, int on)
8316 struct io_ring_ctx *ctx = file->private_data;
8318 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8321 static int io_remove_personalities(int id, void *p, void *data)
8323 struct io_ring_ctx *ctx = data;
8324 struct io_identity *iod;
8326 iod = idr_remove(&ctx->personality_idr, id);
8328 put_cred(iod->creds);
8329 if (refcount_dec_and_test(&iod->count))
8335 static void io_ring_exit_work(struct work_struct *work)
8337 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8341 * If we're doing polled IO and end up having requests being
8342 * submitted async (out-of-line), then completions can come in while
8343 * we're waiting for refs to drop. We need to reap these manually,
8344 * as nobody else will be looking for them.
8348 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8349 io_iopoll_try_reap_events(ctx);
8350 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8351 io_ring_ctx_free(ctx);
8354 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8356 mutex_lock(&ctx->uring_lock);
8357 percpu_ref_kill(&ctx->refs);
8358 mutex_unlock(&ctx->uring_lock);
8360 io_kill_timeouts(ctx, NULL);
8361 io_poll_remove_all(ctx, NULL);
8364 io_wq_cancel_all(ctx->io_wq);
8366 /* if we failed setting up the ctx, we might not have any rings */
8368 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8369 io_iopoll_try_reap_events(ctx);
8370 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8373 * Do this upfront, so we won't have a grace period where the ring
8374 * is closed but resources aren't reaped yet. This can cause
8375 * spurious failure in setting up a new ring.
8377 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8380 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8382 * Use system_unbound_wq to avoid spawning tons of event kworkers
8383 * if we're exiting a ton of rings at the same time. It just adds
8384 * noise and overhead, there's no discernable change in runtime
8385 * over using system_wq.
8387 queue_work(system_unbound_wq, &ctx->exit_work);
8390 static int io_uring_release(struct inode *inode, struct file *file)
8392 struct io_ring_ctx *ctx = file->private_data;
8394 file->private_data = NULL;
8395 io_ring_ctx_wait_and_kill(ctx);
8399 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8401 struct files_struct *files = data;
8403 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8404 work->identity->files == files);
8408 * Returns true if 'preq' is the link parent of 'req'
8410 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8412 struct io_kiocb *link;
8414 if (!(preq->flags & REQ_F_LINK_HEAD))
8417 list_for_each_entry(link, &preq->link_list, link_list) {
8425 static bool io_match_link_files(struct io_kiocb *req,
8426 struct files_struct *files)
8428 struct io_kiocb *link;
8430 if (io_match_files(req, files))
8432 if (req->flags & REQ_F_LINK_HEAD) {
8433 list_for_each_entry(link, &req->link_list, link_list) {
8434 if (io_match_files(link, files))
8442 * We're looking to cancel 'req' because it's holding on to our files, but
8443 * 'req' could be a link to another request. See if it is, and cancel that
8444 * parent request if so.
8446 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8448 struct hlist_node *tmp;
8449 struct io_kiocb *preq;
8453 spin_lock_irq(&ctx->completion_lock);
8454 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8455 struct hlist_head *list;
8457 list = &ctx->cancel_hash[i];
8458 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8459 found = io_match_link(preq, req);
8461 io_poll_remove_one(preq);
8466 spin_unlock_irq(&ctx->completion_lock);
8470 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8471 struct io_kiocb *req)
8473 struct io_kiocb *preq;
8476 spin_lock_irq(&ctx->completion_lock);
8477 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8478 found = io_match_link(preq, req);
8480 __io_timeout_cancel(preq);
8484 spin_unlock_irq(&ctx->completion_lock);
8488 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8490 return io_match_link(container_of(work, struct io_kiocb, work), data);
8493 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8495 enum io_wq_cancel cret;
8497 /* cancel this particular work, if it's running */
8498 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8499 if (cret != IO_WQ_CANCEL_NOTFOUND)
8502 /* find links that hold this pending, cancel those */
8503 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8504 if (cret != IO_WQ_CANCEL_NOTFOUND)
8507 /* if we have a poll link holding this pending, cancel that */
8508 if (io_poll_remove_link(ctx, req))
8511 /* final option, timeout link is holding this req pending */
8512 io_timeout_remove_link(ctx, req);
8515 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8516 struct files_struct *files)
8518 struct io_defer_entry *de = NULL;
8521 spin_lock_irq(&ctx->completion_lock);
8522 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8523 if (io_match_link_files(de->req, files)) {
8524 list_cut_position(&list, &ctx->defer_list, &de->list);
8528 spin_unlock_irq(&ctx->completion_lock);
8530 while (!list_empty(&list)) {
8531 de = list_first_entry(&list, struct io_defer_entry, list);
8532 list_del_init(&de->list);
8533 req_set_fail_links(de->req);
8534 io_put_req(de->req);
8535 io_req_complete(de->req, -ECANCELED);
8541 * Returns true if we found and killed one or more files pinning requests
8543 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8544 struct files_struct *files)
8546 if (list_empty_careful(&ctx->inflight_list))
8549 io_cancel_defer_files(ctx, files);
8550 /* cancel all at once, should be faster than doing it one by one*/
8551 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8553 while (!list_empty_careful(&ctx->inflight_list)) {
8554 struct io_kiocb *cancel_req = NULL, *req;
8557 spin_lock_irq(&ctx->inflight_lock);
8558 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8559 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8560 req->work.identity->files != files)
8562 /* req is being completed, ignore */
8563 if (!refcount_inc_not_zero(&req->refs))
8569 prepare_to_wait(&ctx->inflight_wait, &wait,
8570 TASK_UNINTERRUPTIBLE);
8571 spin_unlock_irq(&ctx->inflight_lock);
8573 /* We need to keep going until we don't find a matching req */
8576 /* cancel this request, or head link requests */
8577 io_attempt_cancel(ctx, cancel_req);
8578 io_put_req(cancel_req);
8579 /* cancellations _may_ trigger task work */
8582 finish_wait(&ctx->inflight_wait, &wait);
8588 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8590 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8591 struct task_struct *task = data;
8593 return io_task_match(req, task);
8596 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8597 struct task_struct *task,
8598 struct files_struct *files)
8602 ret = io_uring_cancel_files(ctx, files);
8604 enum io_wq_cancel cret;
8606 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8607 if (cret != IO_WQ_CANCEL_NOTFOUND)
8610 /* SQPOLL thread does its own polling */
8611 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8612 while (!list_empty_careful(&ctx->iopoll_list)) {
8613 io_iopoll_try_reap_events(ctx);
8618 ret |= io_poll_remove_all(ctx, task);
8619 ret |= io_kill_timeouts(ctx, task);
8626 * We need to iteratively cancel requests, in case a request has dependent
8627 * hard links. These persist even for failure of cancelations, hence keep
8628 * looping until none are found.
8630 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8631 struct files_struct *files)
8633 struct task_struct *task = current;
8635 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8636 task = ctx->sq_data->thread;
8638 io_cqring_overflow_flush(ctx, true, task, files);
8640 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8647 * Note that this task has used io_uring. We use it for cancelation purposes.
8649 static int io_uring_add_task_file(struct file *file)
8651 struct io_uring_task *tctx = current->io_uring;
8653 if (unlikely(!tctx)) {
8656 ret = io_uring_alloc_task_context(current);
8659 tctx = current->io_uring;
8661 if (tctx->last != file) {
8662 void *old = xa_load(&tctx->xa, (unsigned long)file);
8666 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8675 * Remove this io_uring_file -> task mapping.
8677 static void io_uring_del_task_file(struct file *file)
8679 struct io_uring_task *tctx = current->io_uring;
8681 if (tctx->last == file)
8683 file = xa_erase(&tctx->xa, (unsigned long)file);
8688 static void __io_uring_attempt_task_drop(struct file *file)
8690 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8693 io_uring_del_task_file(file);
8697 * Drop task note for this file if we're the only ones that hold it after
8700 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8702 if (!current->io_uring)
8705 * fput() is pending, will be 2 if the only other ref is our potential
8706 * task file note. If the task is exiting, drop regardless of count.
8708 if (!exiting && atomic_long_read(&file->f_count) != 2)
8711 __io_uring_attempt_task_drop(file);
8714 void __io_uring_files_cancel(struct files_struct *files)
8716 struct io_uring_task *tctx = current->io_uring;
8718 unsigned long index;
8720 /* make sure overflow events are dropped */
8721 tctx->in_idle = true;
8723 xa_for_each(&tctx->xa, index, file) {
8724 struct io_ring_ctx *ctx = file->private_data;
8726 io_uring_cancel_task_requests(ctx, files);
8728 io_uring_del_task_file(file);
8733 * Find any io_uring fd that this task has registered or done IO on, and cancel
8736 void __io_uring_task_cancel(void)
8738 struct io_uring_task *tctx = current->io_uring;
8742 /* make sure overflow events are dropped */
8743 tctx->in_idle = true;
8746 /* read completions before cancelations */
8747 inflight = percpu_counter_sum(&tctx->inflight);
8750 __io_uring_files_cancel(NULL);
8752 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8755 * If we've seen completions, retry. This avoids a race where
8756 * a completion comes in before we did prepare_to_wait().
8758 if (inflight != percpu_counter_sum(&tctx->inflight))
8763 finish_wait(&tctx->wait, &wait);
8764 tctx->in_idle = false;
8767 static int io_uring_flush(struct file *file, void *data)
8769 struct io_ring_ctx *ctx = file->private_data;
8772 * If the task is going away, cancel work it may have pending
8774 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8777 io_uring_cancel_task_requests(ctx, data);
8778 io_uring_attempt_task_drop(file, !data);
8782 static void *io_uring_validate_mmap_request(struct file *file,
8783 loff_t pgoff, size_t sz)
8785 struct io_ring_ctx *ctx = file->private_data;
8786 loff_t offset = pgoff << PAGE_SHIFT;
8791 case IORING_OFF_SQ_RING:
8792 case IORING_OFF_CQ_RING:
8795 case IORING_OFF_SQES:
8799 return ERR_PTR(-EINVAL);
8802 page = virt_to_head_page(ptr);
8803 if (sz > page_size(page))
8804 return ERR_PTR(-EINVAL);
8811 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8813 size_t sz = vma->vm_end - vma->vm_start;
8817 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8819 return PTR_ERR(ptr);
8821 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8822 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8825 #else /* !CONFIG_MMU */
8827 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8829 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8832 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8834 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8837 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8838 unsigned long addr, unsigned long len,
8839 unsigned long pgoff, unsigned long flags)
8843 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8845 return PTR_ERR(ptr);
8847 return (unsigned long) ptr;
8850 #endif /* !CONFIG_MMU */
8852 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8857 if (!io_sqring_full(ctx))
8860 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8862 if (!io_sqring_full(ctx))
8866 } while (!signal_pending(current));
8868 finish_wait(&ctx->sqo_sq_wait, &wait);
8871 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8872 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8875 struct io_ring_ctx *ctx;
8882 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8883 IORING_ENTER_SQ_WAIT))
8891 if (f.file->f_op != &io_uring_fops)
8895 ctx = f.file->private_data;
8896 if (!percpu_ref_tryget(&ctx->refs))
8900 if (ctx->flags & IORING_SETUP_R_DISABLED)
8904 * For SQ polling, the thread will do all submissions and completions.
8905 * Just return the requested submit count, and wake the thread if
8909 if (ctx->flags & IORING_SETUP_SQPOLL) {
8910 if (!list_empty_careful(&ctx->cq_overflow_list))
8911 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8912 if (flags & IORING_ENTER_SQ_WAKEUP)
8913 wake_up(&ctx->sq_data->wait);
8914 if (flags & IORING_ENTER_SQ_WAIT)
8915 io_sqpoll_wait_sq(ctx);
8916 submitted = to_submit;
8917 } else if (to_submit) {
8918 ret = io_uring_add_task_file(f.file);
8921 mutex_lock(&ctx->uring_lock);
8922 submitted = io_submit_sqes(ctx, to_submit);
8923 mutex_unlock(&ctx->uring_lock);
8925 if (submitted != to_submit)
8928 if (flags & IORING_ENTER_GETEVENTS) {
8929 min_complete = min(min_complete, ctx->cq_entries);
8932 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8933 * space applications don't need to do io completion events
8934 * polling again, they can rely on io_sq_thread to do polling
8935 * work, which can reduce cpu usage and uring_lock contention.
8937 if (ctx->flags & IORING_SETUP_IOPOLL &&
8938 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8939 ret = io_iopoll_check(ctx, min_complete);
8941 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8946 percpu_ref_put(&ctx->refs);
8949 return submitted ? submitted : ret;
8952 #ifdef CONFIG_PROC_FS
8953 static int io_uring_show_cred(int id, void *p, void *data)
8955 const struct cred *cred = p;
8956 struct seq_file *m = data;
8957 struct user_namespace *uns = seq_user_ns(m);
8958 struct group_info *gi;
8963 seq_printf(m, "%5d\n", id);
8964 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8965 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8966 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8967 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8968 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8969 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8970 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8971 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8972 seq_puts(m, "\n\tGroups:\t");
8973 gi = cred->group_info;
8974 for (g = 0; g < gi->ngroups; g++) {
8975 seq_put_decimal_ull(m, g ? " " : "",
8976 from_kgid_munged(uns, gi->gid[g]));
8978 seq_puts(m, "\n\tCapEff:\t");
8979 cap = cred->cap_effective;
8980 CAP_FOR_EACH_U32(__capi)
8981 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8986 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8988 struct io_sq_data *sq = NULL;
8993 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8994 * since fdinfo case grabs it in the opposite direction of normal use
8995 * cases. If we fail to get the lock, we just don't iterate any
8996 * structures that could be going away outside the io_uring mutex.
8998 has_lock = mutex_trylock(&ctx->uring_lock);
9000 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9003 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9004 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9005 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9006 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9007 struct fixed_file_table *table;
9010 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9011 f = table->files[i & IORING_FILE_TABLE_MASK];
9013 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9015 seq_printf(m, "%5u: <none>\n", i);
9017 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9018 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9019 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9021 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9022 (unsigned int) buf->len);
9024 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9025 seq_printf(m, "Personalities:\n");
9026 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9028 seq_printf(m, "PollList:\n");
9029 spin_lock_irq(&ctx->completion_lock);
9030 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9031 struct hlist_head *list = &ctx->cancel_hash[i];
9032 struct io_kiocb *req;
9034 hlist_for_each_entry(req, list, hash_node)
9035 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9036 req->task->task_works != NULL);
9038 spin_unlock_irq(&ctx->completion_lock);
9040 mutex_unlock(&ctx->uring_lock);
9043 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9045 struct io_ring_ctx *ctx = f->private_data;
9047 if (percpu_ref_tryget(&ctx->refs)) {
9048 __io_uring_show_fdinfo(ctx, m);
9049 percpu_ref_put(&ctx->refs);
9054 static const struct file_operations io_uring_fops = {
9055 .release = io_uring_release,
9056 .flush = io_uring_flush,
9057 .mmap = io_uring_mmap,
9059 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9060 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9062 .poll = io_uring_poll,
9063 .fasync = io_uring_fasync,
9064 #ifdef CONFIG_PROC_FS
9065 .show_fdinfo = io_uring_show_fdinfo,
9069 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9070 struct io_uring_params *p)
9072 struct io_rings *rings;
9073 size_t size, sq_array_offset;
9075 /* make sure these are sane, as we already accounted them */
9076 ctx->sq_entries = p->sq_entries;
9077 ctx->cq_entries = p->cq_entries;
9079 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9080 if (size == SIZE_MAX)
9083 rings = io_mem_alloc(size);
9088 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9089 rings->sq_ring_mask = p->sq_entries - 1;
9090 rings->cq_ring_mask = p->cq_entries - 1;
9091 rings->sq_ring_entries = p->sq_entries;
9092 rings->cq_ring_entries = p->cq_entries;
9093 ctx->sq_mask = rings->sq_ring_mask;
9094 ctx->cq_mask = rings->cq_ring_mask;
9096 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9097 if (size == SIZE_MAX) {
9098 io_mem_free(ctx->rings);
9103 ctx->sq_sqes = io_mem_alloc(size);
9104 if (!ctx->sq_sqes) {
9105 io_mem_free(ctx->rings);
9114 * Allocate an anonymous fd, this is what constitutes the application
9115 * visible backing of an io_uring instance. The application mmaps this
9116 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9117 * we have to tie this fd to a socket for file garbage collection purposes.
9119 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9124 #if defined(CONFIG_UNIX)
9125 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9131 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9135 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9136 O_RDWR | O_CLOEXEC);
9140 ret = PTR_ERR(file);
9144 #if defined(CONFIG_UNIX)
9145 ctx->ring_sock->file = file;
9147 if (unlikely(io_uring_add_task_file(file))) {
9148 file = ERR_PTR(-ENOMEM);
9151 fd_install(ret, file);
9154 #if defined(CONFIG_UNIX)
9155 sock_release(ctx->ring_sock);
9156 ctx->ring_sock = NULL;
9161 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9162 struct io_uring_params __user *params)
9164 struct user_struct *user = NULL;
9165 struct io_ring_ctx *ctx;
9171 if (entries > IORING_MAX_ENTRIES) {
9172 if (!(p->flags & IORING_SETUP_CLAMP))
9174 entries = IORING_MAX_ENTRIES;
9178 * Use twice as many entries for the CQ ring. It's possible for the
9179 * application to drive a higher depth than the size of the SQ ring,
9180 * since the sqes are only used at submission time. This allows for
9181 * some flexibility in overcommitting a bit. If the application has
9182 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9183 * of CQ ring entries manually.
9185 p->sq_entries = roundup_pow_of_two(entries);
9186 if (p->flags & IORING_SETUP_CQSIZE) {
9188 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9189 * to a power-of-two, if it isn't already. We do NOT impose
9190 * any cq vs sq ring sizing.
9192 if (p->cq_entries < p->sq_entries)
9194 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9195 if (!(p->flags & IORING_SETUP_CLAMP))
9197 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9199 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9201 p->cq_entries = 2 * p->sq_entries;
9204 user = get_uid(current_user());
9205 limit_mem = !capable(CAP_IPC_LOCK);
9208 ret = __io_account_mem(user,
9209 ring_pages(p->sq_entries, p->cq_entries));
9216 ctx = io_ring_ctx_alloc(p);
9219 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9224 ctx->compat = in_compat_syscall();
9226 ctx->creds = get_current_cred();
9228 ctx->loginuid = current->loginuid;
9229 ctx->sessionid = current->sessionid;
9231 ctx->sqo_task = get_task_struct(current);
9234 * This is just grabbed for accounting purposes. When a process exits,
9235 * the mm is exited and dropped before the files, hence we need to hang
9236 * on to this mm purely for the purposes of being able to unaccount
9237 * memory (locked/pinned vm). It's not used for anything else.
9239 mmgrab(current->mm);
9240 ctx->mm_account = current->mm;
9242 #ifdef CONFIG_BLK_CGROUP
9244 * The sq thread will belong to the original cgroup it was inited in.
9245 * If the cgroup goes offline (e.g. disabling the io controller), then
9246 * issued bios will be associated with the closest cgroup later in the
9250 ctx->sqo_blkcg_css = blkcg_css();
9251 ret = css_tryget_online(ctx->sqo_blkcg_css);
9254 /* don't init against a dying cgroup, have the user try again */
9255 ctx->sqo_blkcg_css = NULL;
9262 * Account memory _before_ installing the file descriptor. Once
9263 * the descriptor is installed, it can get closed at any time. Also
9264 * do this before hitting the general error path, as ring freeing
9265 * will un-account as well.
9267 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9269 ctx->limit_mem = limit_mem;
9271 ret = io_allocate_scq_urings(ctx, p);
9275 ret = io_sq_offload_create(ctx, p);
9279 if (!(p->flags & IORING_SETUP_R_DISABLED))
9280 io_sq_offload_start(ctx);
9282 memset(&p->sq_off, 0, sizeof(p->sq_off));
9283 p->sq_off.head = offsetof(struct io_rings, sq.head);
9284 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9285 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9286 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9287 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9288 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9289 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9291 memset(&p->cq_off, 0, sizeof(p->cq_off));
9292 p->cq_off.head = offsetof(struct io_rings, cq.head);
9293 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9294 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9295 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9296 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9297 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9298 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9300 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9301 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9302 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9303 IORING_FEAT_POLL_32BITS;
9305 if (copy_to_user(params, p, sizeof(*p))) {
9311 * Install ring fd as the very last thing, so we don't risk someone
9312 * having closed it before we finish setup
9314 ret = io_uring_get_fd(ctx);
9318 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9321 io_ring_ctx_wait_and_kill(ctx);
9326 * Sets up an aio uring context, and returns the fd. Applications asks for a
9327 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9328 * params structure passed in.
9330 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9332 struct io_uring_params p;
9335 if (copy_from_user(&p, params, sizeof(p)))
9337 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9342 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9343 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9344 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9345 IORING_SETUP_R_DISABLED))
9348 return io_uring_create(entries, &p, params);
9351 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9352 struct io_uring_params __user *, params)
9354 return io_uring_setup(entries, params);
9357 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9359 struct io_uring_probe *p;
9363 size = struct_size(p, ops, nr_args);
9364 if (size == SIZE_MAX)
9366 p = kzalloc(size, GFP_KERNEL);
9371 if (copy_from_user(p, arg, size))
9374 if (memchr_inv(p, 0, size))
9377 p->last_op = IORING_OP_LAST - 1;
9378 if (nr_args > IORING_OP_LAST)
9379 nr_args = IORING_OP_LAST;
9381 for (i = 0; i < nr_args; i++) {
9383 if (!io_op_defs[i].not_supported)
9384 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9389 if (copy_to_user(arg, p, size))
9396 static int io_register_personality(struct io_ring_ctx *ctx)
9398 struct io_identity *id;
9401 id = kmalloc(sizeof(*id), GFP_KERNEL);
9405 io_init_identity(id);
9406 id->creds = get_current_cred();
9408 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9410 put_cred(id->creds);
9416 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9418 struct io_identity *iod;
9420 iod = idr_remove(&ctx->personality_idr, id);
9422 put_cred(iod->creds);
9423 if (refcount_dec_and_test(&iod->count))
9431 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9432 unsigned int nr_args)
9434 struct io_uring_restriction *res;
9438 /* Restrictions allowed only if rings started disabled */
9439 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9442 /* We allow only a single restrictions registration */
9443 if (ctx->restrictions.registered)
9446 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9449 size = array_size(nr_args, sizeof(*res));
9450 if (size == SIZE_MAX)
9453 res = memdup_user(arg, size);
9455 return PTR_ERR(res);
9459 for (i = 0; i < nr_args; i++) {
9460 switch (res[i].opcode) {
9461 case IORING_RESTRICTION_REGISTER_OP:
9462 if (res[i].register_op >= IORING_REGISTER_LAST) {
9467 __set_bit(res[i].register_op,
9468 ctx->restrictions.register_op);
9470 case IORING_RESTRICTION_SQE_OP:
9471 if (res[i].sqe_op >= IORING_OP_LAST) {
9476 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9478 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9479 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9481 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9482 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9491 /* Reset all restrictions if an error happened */
9493 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9495 ctx->restrictions.registered = true;
9501 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9503 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9506 if (ctx->restrictions.registered)
9507 ctx->restricted = 1;
9509 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9511 io_sq_offload_start(ctx);
9516 static bool io_register_op_must_quiesce(int op)
9519 case IORING_UNREGISTER_FILES:
9520 case IORING_REGISTER_FILES_UPDATE:
9521 case IORING_REGISTER_PROBE:
9522 case IORING_REGISTER_PERSONALITY:
9523 case IORING_UNREGISTER_PERSONALITY:
9530 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9531 void __user *arg, unsigned nr_args)
9532 __releases(ctx->uring_lock)
9533 __acquires(ctx->uring_lock)
9538 * We're inside the ring mutex, if the ref is already dying, then
9539 * someone else killed the ctx or is already going through
9540 * io_uring_register().
9542 if (percpu_ref_is_dying(&ctx->refs))
9545 if (io_register_op_must_quiesce(opcode)) {
9546 percpu_ref_kill(&ctx->refs);
9549 * Drop uring mutex before waiting for references to exit. If
9550 * another thread is currently inside io_uring_enter() it might
9551 * need to grab the uring_lock to make progress. If we hold it
9552 * here across the drain wait, then we can deadlock. It's safe
9553 * to drop the mutex here, since no new references will come in
9554 * after we've killed the percpu ref.
9556 mutex_unlock(&ctx->uring_lock);
9558 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9561 ret = io_run_task_work_sig();
9566 mutex_lock(&ctx->uring_lock);
9569 percpu_ref_resurrect(&ctx->refs);
9574 if (ctx->restricted) {
9575 if (opcode >= IORING_REGISTER_LAST) {
9580 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9587 case IORING_REGISTER_BUFFERS:
9588 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9590 case IORING_UNREGISTER_BUFFERS:
9594 ret = io_sqe_buffer_unregister(ctx);
9596 case IORING_REGISTER_FILES:
9597 ret = io_sqe_files_register(ctx, arg, nr_args);
9599 case IORING_UNREGISTER_FILES:
9603 ret = io_sqe_files_unregister(ctx);
9605 case IORING_REGISTER_FILES_UPDATE:
9606 ret = io_sqe_files_update(ctx, arg, nr_args);
9608 case IORING_REGISTER_EVENTFD:
9609 case IORING_REGISTER_EVENTFD_ASYNC:
9613 ret = io_eventfd_register(ctx, arg);
9616 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9617 ctx->eventfd_async = 1;
9619 ctx->eventfd_async = 0;
9621 case IORING_UNREGISTER_EVENTFD:
9625 ret = io_eventfd_unregister(ctx);
9627 case IORING_REGISTER_PROBE:
9629 if (!arg || nr_args > 256)
9631 ret = io_probe(ctx, arg, nr_args);
9633 case IORING_REGISTER_PERSONALITY:
9637 ret = io_register_personality(ctx);
9639 case IORING_UNREGISTER_PERSONALITY:
9643 ret = io_unregister_personality(ctx, nr_args);
9645 case IORING_REGISTER_ENABLE_RINGS:
9649 ret = io_register_enable_rings(ctx);
9651 case IORING_REGISTER_RESTRICTIONS:
9652 ret = io_register_restrictions(ctx, arg, nr_args);
9660 if (io_register_op_must_quiesce(opcode)) {
9661 /* bring the ctx back to life */
9662 percpu_ref_reinit(&ctx->refs);
9664 reinit_completion(&ctx->ref_comp);
9669 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9670 void __user *, arg, unsigned int, nr_args)
9672 struct io_ring_ctx *ctx;
9681 if (f.file->f_op != &io_uring_fops)
9684 ctx = f.file->private_data;
9686 mutex_lock(&ctx->uring_lock);
9687 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9688 mutex_unlock(&ctx->uring_lock);
9689 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9690 ctx->cq_ev_fd != NULL, ret);
9696 static int __init io_uring_init(void)
9698 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9699 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9700 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9703 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9704 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9705 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9706 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9707 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9708 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9709 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9710 BUILD_BUG_SQE_ELEM(8, __u64, off);
9711 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9712 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9713 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9714 BUILD_BUG_SQE_ELEM(24, __u32, len);
9715 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9716 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9717 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9718 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9719 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9720 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9721 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9722 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9723 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9724 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9725 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9726 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9727 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9728 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9729 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9730 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9731 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9732 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9733 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9735 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9736 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9737 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9740 __initcall(io_uring_init);