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;
1979 enum task_work_notify_mode notify;
1982 if (tsk->flags & PF_EXITING)
1986 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1987 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1988 * processing task_work. There's no reliable way to tell if TWA_RESUME
1992 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1993 notify = TWA_SIGNAL;
1995 ret = task_work_add(tsk, &req->task_work, notify);
1997 wake_up_process(tsk);
2002 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2004 struct io_ring_ctx *ctx = req->ctx;
2006 spin_lock_irq(&ctx->completion_lock);
2007 io_cqring_fill_event(req, error);
2008 io_commit_cqring(ctx);
2009 spin_unlock_irq(&ctx->completion_lock);
2011 io_cqring_ev_posted(ctx);
2012 req_set_fail_links(req);
2013 io_double_put_req(req);
2016 static void io_req_task_cancel(struct callback_head *cb)
2018 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2019 struct io_ring_ctx *ctx = req->ctx;
2021 __io_req_task_cancel(req, -ECANCELED);
2022 percpu_ref_put(&ctx->refs);
2025 static void __io_req_task_submit(struct io_kiocb *req)
2027 struct io_ring_ctx *ctx = req->ctx;
2029 if (!__io_sq_thread_acquire_mm(ctx)) {
2030 mutex_lock(&ctx->uring_lock);
2031 __io_queue_sqe(req, NULL);
2032 mutex_unlock(&ctx->uring_lock);
2034 __io_req_task_cancel(req, -EFAULT);
2038 static void io_req_task_submit(struct callback_head *cb)
2040 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2041 struct io_ring_ctx *ctx = req->ctx;
2043 __io_req_task_submit(req);
2044 percpu_ref_put(&ctx->refs);
2047 static void io_req_task_queue(struct io_kiocb *req)
2051 init_task_work(&req->task_work, io_req_task_submit);
2052 percpu_ref_get(&req->ctx->refs);
2054 ret = io_req_task_work_add(req, true);
2055 if (unlikely(ret)) {
2056 struct task_struct *tsk;
2058 init_task_work(&req->task_work, io_req_task_cancel);
2059 tsk = io_wq_get_task(req->ctx->io_wq);
2060 task_work_add(tsk, &req->task_work, TWA_NONE);
2061 wake_up_process(tsk);
2065 static void io_queue_next(struct io_kiocb *req)
2067 struct io_kiocb *nxt = io_req_find_next(req);
2070 io_req_task_queue(nxt);
2073 static void io_free_req(struct io_kiocb *req)
2080 void *reqs[IO_IOPOLL_BATCH];
2083 struct task_struct *task;
2087 static inline void io_init_req_batch(struct req_batch *rb)
2094 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2095 struct req_batch *rb)
2097 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2098 percpu_ref_put_many(&ctx->refs, rb->to_free);
2102 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2103 struct req_batch *rb)
2106 __io_req_free_batch_flush(ctx, rb);
2108 struct io_uring_task *tctx = rb->task->io_uring;
2110 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2111 put_task_struct_many(rb->task, rb->task_refs);
2116 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2118 if (unlikely(io_is_fallback_req(req))) {
2122 if (req->flags & REQ_F_LINK_HEAD)
2125 if (req->task != rb->task) {
2127 struct io_uring_task *tctx = rb->task->io_uring;
2129 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2130 put_task_struct_many(rb->task, rb->task_refs);
2132 rb->task = req->task;
2137 io_dismantle_req(req);
2138 rb->reqs[rb->to_free++] = req;
2139 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2140 __io_req_free_batch_flush(req->ctx, rb);
2144 * Drop reference to request, return next in chain (if there is one) if this
2145 * was the last reference to this request.
2147 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2149 struct io_kiocb *nxt = NULL;
2151 if (refcount_dec_and_test(&req->refs)) {
2152 nxt = io_req_find_next(req);
2158 static void io_put_req(struct io_kiocb *req)
2160 if (refcount_dec_and_test(&req->refs))
2164 static void io_put_req_deferred_cb(struct callback_head *cb)
2166 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2171 static void io_free_req_deferred(struct io_kiocb *req)
2175 init_task_work(&req->task_work, io_put_req_deferred_cb);
2176 ret = io_req_task_work_add(req, true);
2177 if (unlikely(ret)) {
2178 struct task_struct *tsk;
2180 tsk = io_wq_get_task(req->ctx->io_wq);
2181 task_work_add(tsk, &req->task_work, TWA_NONE);
2182 wake_up_process(tsk);
2186 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2188 if (refcount_sub_and_test(refs, &req->refs))
2189 io_free_req_deferred(req);
2192 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2194 struct io_kiocb *nxt;
2197 * A ref is owned by io-wq in which context we're. So, if that's the
2198 * last one, it's safe to steal next work. False negatives are Ok,
2199 * it just will be re-punted async in io_put_work()
2201 if (refcount_read(&req->refs) != 1)
2204 nxt = io_req_find_next(req);
2205 return nxt ? &nxt->work : NULL;
2208 static void io_double_put_req(struct io_kiocb *req)
2210 /* drop both submit and complete references */
2211 if (refcount_sub_and_test(2, &req->refs))
2215 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2217 struct io_rings *rings = ctx->rings;
2219 if (test_bit(0, &ctx->cq_check_overflow)) {
2221 * noflush == true is from the waitqueue handler, just ensure
2222 * we wake up the task, and the next invocation will flush the
2223 * entries. We cannot safely to it from here.
2225 if (noflush && !list_empty(&ctx->cq_overflow_list))
2228 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2231 /* See comment at the top of this file */
2233 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2236 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2238 struct io_rings *rings = ctx->rings;
2240 /* make sure SQ entry isn't read before tail */
2241 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2244 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2246 unsigned int cflags;
2248 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2249 cflags |= IORING_CQE_F_BUFFER;
2250 req->flags &= ~REQ_F_BUFFER_SELECTED;
2255 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2257 struct io_buffer *kbuf;
2259 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2260 return io_put_kbuf(req, kbuf);
2263 static inline bool io_run_task_work(void)
2266 * Not safe to run on exiting task, and the task_work handling will
2267 * not add work to such a task.
2269 if (unlikely(current->flags & PF_EXITING))
2271 if (current->task_works) {
2272 __set_current_state(TASK_RUNNING);
2280 static void io_iopoll_queue(struct list_head *again)
2282 struct io_kiocb *req;
2285 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2286 list_del(&req->inflight_entry);
2287 __io_complete_rw(req, -EAGAIN, 0, NULL);
2288 } while (!list_empty(again));
2292 * Find and free completed poll iocbs
2294 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2295 struct list_head *done)
2297 struct req_batch rb;
2298 struct io_kiocb *req;
2301 /* order with ->result store in io_complete_rw_iopoll() */
2304 io_init_req_batch(&rb);
2305 while (!list_empty(done)) {
2308 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2309 if (READ_ONCE(req->result) == -EAGAIN) {
2311 req->iopoll_completed = 0;
2312 list_move_tail(&req->inflight_entry, &again);
2315 list_del(&req->inflight_entry);
2317 if (req->flags & REQ_F_BUFFER_SELECTED)
2318 cflags = io_put_rw_kbuf(req);
2320 __io_cqring_fill_event(req, req->result, cflags);
2323 if (refcount_dec_and_test(&req->refs))
2324 io_req_free_batch(&rb, req);
2327 io_commit_cqring(ctx);
2328 if (ctx->flags & IORING_SETUP_SQPOLL)
2329 io_cqring_ev_posted(ctx);
2330 io_req_free_batch_finish(ctx, &rb);
2332 if (!list_empty(&again))
2333 io_iopoll_queue(&again);
2336 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2339 struct io_kiocb *req, *tmp;
2345 * Only spin for completions if we don't have multiple devices hanging
2346 * off our complete list, and we're under the requested amount.
2348 spin = !ctx->poll_multi_file && *nr_events < min;
2351 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2352 struct kiocb *kiocb = &req->rw.kiocb;
2355 * Move completed and retryable entries to our local lists.
2356 * If we find a request that requires polling, break out
2357 * and complete those lists first, if we have entries there.
2359 if (READ_ONCE(req->iopoll_completed)) {
2360 list_move_tail(&req->inflight_entry, &done);
2363 if (!list_empty(&done))
2366 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2370 /* iopoll may have completed current req */
2371 if (READ_ONCE(req->iopoll_completed))
2372 list_move_tail(&req->inflight_entry, &done);
2379 if (!list_empty(&done))
2380 io_iopoll_complete(ctx, nr_events, &done);
2386 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2387 * non-spinning poll check - we'll still enter the driver poll loop, but only
2388 * as a non-spinning completion check.
2390 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2393 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2396 ret = io_do_iopoll(ctx, nr_events, min);
2399 if (*nr_events >= min)
2407 * We can't just wait for polled events to come to us, we have to actively
2408 * find and complete them.
2410 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2412 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2415 mutex_lock(&ctx->uring_lock);
2416 while (!list_empty(&ctx->iopoll_list)) {
2417 unsigned int nr_events = 0;
2419 io_do_iopoll(ctx, &nr_events, 0);
2421 /* let it sleep and repeat later if can't complete a request */
2425 * Ensure we allow local-to-the-cpu processing to take place,
2426 * in this case we need to ensure that we reap all events.
2427 * Also let task_work, etc. to progress by releasing the mutex
2429 if (need_resched()) {
2430 mutex_unlock(&ctx->uring_lock);
2432 mutex_lock(&ctx->uring_lock);
2435 mutex_unlock(&ctx->uring_lock);
2438 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2440 unsigned int nr_events = 0;
2441 int iters = 0, ret = 0;
2444 * We disallow the app entering submit/complete with polling, but we
2445 * still need to lock the ring to prevent racing with polled issue
2446 * that got punted to a workqueue.
2448 mutex_lock(&ctx->uring_lock);
2451 * Don't enter poll loop if we already have events pending.
2452 * If we do, we can potentially be spinning for commands that
2453 * already triggered a CQE (eg in error).
2455 if (io_cqring_events(ctx, false))
2459 * If a submit got punted to a workqueue, we can have the
2460 * application entering polling for a command before it gets
2461 * issued. That app will hold the uring_lock for the duration
2462 * of the poll right here, so we need to take a breather every
2463 * now and then to ensure that the issue has a chance to add
2464 * the poll to the issued list. Otherwise we can spin here
2465 * forever, while the workqueue is stuck trying to acquire the
2468 if (!(++iters & 7)) {
2469 mutex_unlock(&ctx->uring_lock);
2471 mutex_lock(&ctx->uring_lock);
2474 ret = io_iopoll_getevents(ctx, &nr_events, min);
2478 } while (min && !nr_events && !need_resched());
2480 mutex_unlock(&ctx->uring_lock);
2484 static void kiocb_end_write(struct io_kiocb *req)
2487 * Tell lockdep we inherited freeze protection from submission
2490 if (req->flags & REQ_F_ISREG) {
2491 struct inode *inode = file_inode(req->file);
2493 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2495 file_end_write(req->file);
2498 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2499 struct io_comp_state *cs)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2504 if (kiocb->ki_flags & IOCB_WRITE)
2505 kiocb_end_write(req);
2507 if (res != req->result)
2508 req_set_fail_links(req);
2509 if (req->flags & REQ_F_BUFFER_SELECTED)
2510 cflags = io_put_rw_kbuf(req);
2511 __io_req_complete(req, res, cflags, cs);
2515 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2517 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2518 ssize_t ret = -ECANCELED;
2519 struct iov_iter iter;
2527 switch (req->opcode) {
2528 case IORING_OP_READV:
2529 case IORING_OP_READ_FIXED:
2530 case IORING_OP_READ:
2533 case IORING_OP_WRITEV:
2534 case IORING_OP_WRITE_FIXED:
2535 case IORING_OP_WRITE:
2539 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2544 if (!req->async_data) {
2545 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2548 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2556 req_set_fail_links(req);
2557 io_req_complete(req, ret);
2562 static bool io_rw_reissue(struct io_kiocb *req, long res)
2565 umode_t mode = file_inode(req->file)->i_mode;
2568 if (!S_ISBLK(mode) && !S_ISREG(mode))
2570 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2573 ret = io_sq_thread_acquire_mm(req->ctx, req);
2575 if (io_resubmit_prep(req, ret)) {
2576 refcount_inc(&req->refs);
2577 io_queue_async_work(req);
2585 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2586 struct io_comp_state *cs)
2588 if (!io_rw_reissue(req, res))
2589 io_complete_rw_common(&req->rw.kiocb, res, cs);
2592 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2594 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2596 __io_complete_rw(req, res, res2, NULL);
2599 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2601 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2603 if (kiocb->ki_flags & IOCB_WRITE)
2604 kiocb_end_write(req);
2606 if (res != -EAGAIN && res != req->result)
2607 req_set_fail_links(req);
2609 WRITE_ONCE(req->result, res);
2610 /* order with io_poll_complete() checking ->result */
2612 WRITE_ONCE(req->iopoll_completed, 1);
2616 * After the iocb has been issued, it's safe to be found on the poll list.
2617 * Adding the kiocb to the list AFTER submission ensures that we don't
2618 * find it from a io_iopoll_getevents() thread before the issuer is done
2619 * accessing the kiocb cookie.
2621 static void io_iopoll_req_issued(struct io_kiocb *req)
2623 struct io_ring_ctx *ctx = req->ctx;
2626 * Track whether we have multiple files in our lists. This will impact
2627 * how we do polling eventually, not spinning if we're on potentially
2628 * different devices.
2630 if (list_empty(&ctx->iopoll_list)) {
2631 ctx->poll_multi_file = false;
2632 } else if (!ctx->poll_multi_file) {
2633 struct io_kiocb *list_req;
2635 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2637 if (list_req->file != req->file)
2638 ctx->poll_multi_file = true;
2642 * For fast devices, IO may have already completed. If it has, add
2643 * it to the front so we find it first.
2645 if (READ_ONCE(req->iopoll_completed))
2646 list_add(&req->inflight_entry, &ctx->iopoll_list);
2648 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2650 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2651 wq_has_sleeper(&ctx->sq_data->wait))
2652 wake_up(&ctx->sq_data->wait);
2655 static void __io_state_file_put(struct io_submit_state *state)
2657 if (state->has_refs)
2658 fput_many(state->file, state->has_refs);
2662 static inline void io_state_file_put(struct io_submit_state *state)
2665 __io_state_file_put(state);
2669 * Get as many references to a file as we have IOs left in this submission,
2670 * assuming most submissions are for one file, or at least that each file
2671 * has more than one submission.
2673 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2679 if (state->fd == fd) {
2683 __io_state_file_put(state);
2685 state->file = fget_many(fd, state->ios_left);
2690 state->has_refs = state->ios_left - 1;
2694 static bool io_bdev_nowait(struct block_device *bdev)
2697 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2704 * If we tracked the file through the SCM inflight mechanism, we could support
2705 * any file. For now, just ensure that anything potentially problematic is done
2708 static bool io_file_supports_async(struct file *file, int rw)
2710 umode_t mode = file_inode(file)->i_mode;
2712 if (S_ISBLK(mode)) {
2713 if (io_bdev_nowait(file->f_inode->i_bdev))
2717 if (S_ISCHR(mode) || S_ISSOCK(mode))
2719 if (S_ISREG(mode)) {
2720 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2721 file->f_op != &io_uring_fops)
2726 /* any ->read/write should understand O_NONBLOCK */
2727 if (file->f_flags & O_NONBLOCK)
2730 if (!(file->f_mode & FMODE_NOWAIT))
2734 return file->f_op->read_iter != NULL;
2736 return file->f_op->write_iter != NULL;
2739 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2741 struct io_ring_ctx *ctx = req->ctx;
2742 struct kiocb *kiocb = &req->rw.kiocb;
2746 if (S_ISREG(file_inode(req->file)->i_mode))
2747 req->flags |= REQ_F_ISREG;
2749 kiocb->ki_pos = READ_ONCE(sqe->off);
2750 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2751 req->flags |= REQ_F_CUR_POS;
2752 kiocb->ki_pos = req->file->f_pos;
2754 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2755 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2756 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2760 ioprio = READ_ONCE(sqe->ioprio);
2762 ret = ioprio_check_cap(ioprio);
2766 kiocb->ki_ioprio = ioprio;
2768 kiocb->ki_ioprio = get_current_ioprio();
2770 /* don't allow async punt if RWF_NOWAIT was requested */
2771 if (kiocb->ki_flags & IOCB_NOWAIT)
2772 req->flags |= REQ_F_NOWAIT;
2774 if (ctx->flags & IORING_SETUP_IOPOLL) {
2775 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2776 !kiocb->ki_filp->f_op->iopoll)
2779 kiocb->ki_flags |= IOCB_HIPRI;
2780 kiocb->ki_complete = io_complete_rw_iopoll;
2781 req->iopoll_completed = 0;
2783 if (kiocb->ki_flags & IOCB_HIPRI)
2785 kiocb->ki_complete = io_complete_rw;
2788 req->rw.addr = READ_ONCE(sqe->addr);
2789 req->rw.len = READ_ONCE(sqe->len);
2790 req->buf_index = READ_ONCE(sqe->buf_index);
2794 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2800 case -ERESTARTNOINTR:
2801 case -ERESTARTNOHAND:
2802 case -ERESTART_RESTARTBLOCK:
2804 * We can't just restart the syscall, since previously
2805 * submitted sqes may already be in progress. Just fail this
2811 kiocb->ki_complete(kiocb, ret, 0);
2815 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2816 struct io_comp_state *cs)
2818 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2819 struct io_async_rw *io = req->async_data;
2821 /* add previously done IO, if any */
2822 if (io && io->bytes_done > 0) {
2824 ret = io->bytes_done;
2826 ret += io->bytes_done;
2829 if (req->flags & REQ_F_CUR_POS)
2830 req->file->f_pos = kiocb->ki_pos;
2831 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2832 __io_complete_rw(req, ret, 0, cs);
2834 io_rw_done(kiocb, ret);
2837 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2838 struct iov_iter *iter)
2840 struct io_ring_ctx *ctx = req->ctx;
2841 size_t len = req->rw.len;
2842 struct io_mapped_ubuf *imu;
2843 u16 index, buf_index = req->buf_index;
2847 if (unlikely(buf_index >= ctx->nr_user_bufs))
2849 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2850 imu = &ctx->user_bufs[index];
2851 buf_addr = req->rw.addr;
2854 if (buf_addr + len < buf_addr)
2856 /* not inside the mapped region */
2857 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2861 * May not be a start of buffer, set size appropriately
2862 * and advance us to the beginning.
2864 offset = buf_addr - imu->ubuf;
2865 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2869 * Don't use iov_iter_advance() here, as it's really slow for
2870 * using the latter parts of a big fixed buffer - it iterates
2871 * over each segment manually. We can cheat a bit here, because
2874 * 1) it's a BVEC iter, we set it up
2875 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2876 * first and last bvec
2878 * So just find our index, and adjust the iterator afterwards.
2879 * If the offset is within the first bvec (or the whole first
2880 * bvec, just use iov_iter_advance(). This makes it easier
2881 * since we can just skip the first segment, which may not
2882 * be PAGE_SIZE aligned.
2884 const struct bio_vec *bvec = imu->bvec;
2886 if (offset <= bvec->bv_len) {
2887 iov_iter_advance(iter, offset);
2889 unsigned long seg_skip;
2891 /* skip first vec */
2892 offset -= bvec->bv_len;
2893 seg_skip = 1 + (offset >> PAGE_SHIFT);
2895 iter->bvec = bvec + seg_skip;
2896 iter->nr_segs -= seg_skip;
2897 iter->count -= bvec->bv_len + offset;
2898 iter->iov_offset = offset & ~PAGE_MASK;
2905 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2908 mutex_unlock(&ctx->uring_lock);
2911 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2914 * "Normal" inline submissions always hold the uring_lock, since we
2915 * grab it from the system call. Same is true for the SQPOLL offload.
2916 * The only exception is when we've detached the request and issue it
2917 * from an async worker thread, grab the lock for that case.
2920 mutex_lock(&ctx->uring_lock);
2923 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2924 int bgid, struct io_buffer *kbuf,
2927 struct io_buffer *head;
2929 if (req->flags & REQ_F_BUFFER_SELECTED)
2932 io_ring_submit_lock(req->ctx, needs_lock);
2934 lockdep_assert_held(&req->ctx->uring_lock);
2936 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2938 if (!list_empty(&head->list)) {
2939 kbuf = list_last_entry(&head->list, struct io_buffer,
2941 list_del(&kbuf->list);
2944 idr_remove(&req->ctx->io_buffer_idr, bgid);
2946 if (*len > kbuf->len)
2949 kbuf = ERR_PTR(-ENOBUFS);
2952 io_ring_submit_unlock(req->ctx, needs_lock);
2957 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2960 struct io_buffer *kbuf;
2963 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2964 bgid = req->buf_index;
2965 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2968 req->rw.addr = (u64) (unsigned long) kbuf;
2969 req->flags |= REQ_F_BUFFER_SELECTED;
2970 return u64_to_user_ptr(kbuf->addr);
2973 #ifdef CONFIG_COMPAT
2974 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2977 struct compat_iovec __user *uiov;
2978 compat_ssize_t clen;
2982 uiov = u64_to_user_ptr(req->rw.addr);
2983 if (!access_ok(uiov, sizeof(*uiov)))
2985 if (__get_user(clen, &uiov->iov_len))
2991 buf = io_rw_buffer_select(req, &len, needs_lock);
2993 return PTR_ERR(buf);
2994 iov[0].iov_base = buf;
2995 iov[0].iov_len = (compat_size_t) len;
3000 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3003 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3007 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3010 len = iov[0].iov_len;
3013 buf = io_rw_buffer_select(req, &len, needs_lock);
3015 return PTR_ERR(buf);
3016 iov[0].iov_base = buf;
3017 iov[0].iov_len = len;
3021 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3024 if (req->flags & REQ_F_BUFFER_SELECTED) {
3025 struct io_buffer *kbuf;
3027 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3028 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3029 iov[0].iov_len = kbuf->len;
3034 else if (req->rw.len > 1)
3037 #ifdef CONFIG_COMPAT
3038 if (req->ctx->compat)
3039 return io_compat_import(req, iov, needs_lock);
3042 return __io_iov_buffer_select(req, iov, needs_lock);
3045 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3046 struct iovec **iovec, struct iov_iter *iter,
3049 void __user *buf = u64_to_user_ptr(req->rw.addr);
3050 size_t sqe_len = req->rw.len;
3054 opcode = req->opcode;
3055 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3057 return io_import_fixed(req, rw, iter);
3060 /* buffer index only valid with fixed read/write, or buffer select */
3061 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3064 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3065 if (req->flags & REQ_F_BUFFER_SELECT) {
3066 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3068 return PTR_ERR(buf);
3069 req->rw.len = sqe_len;
3072 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3074 return ret < 0 ? ret : sqe_len;
3077 if (req->flags & REQ_F_BUFFER_SELECT) {
3078 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3080 ret = (*iovec)->iov_len;
3081 iov_iter_init(iter, rw, *iovec, 1, ret);
3087 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3091 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3092 struct iovec **iovec, struct iov_iter *iter,
3095 struct io_async_rw *iorw = req->async_data;
3098 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3100 return iov_iter_count(&iorw->iter);
3103 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3105 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3109 * For files that don't have ->read_iter() and ->write_iter(), handle them
3110 * by looping over ->read() or ->write() manually.
3112 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3113 struct iov_iter *iter)
3118 * Don't support polled IO through this interface, and we can't
3119 * support non-blocking either. For the latter, this just causes
3120 * the kiocb to be handled from an async context.
3122 if (kiocb->ki_flags & IOCB_HIPRI)
3124 if (kiocb->ki_flags & IOCB_NOWAIT)
3127 while (iov_iter_count(iter)) {
3131 if (!iov_iter_is_bvec(iter)) {
3132 iovec = iov_iter_iovec(iter);
3134 /* fixed buffers import bvec */
3135 iovec.iov_base = kmap(iter->bvec->bv_page)
3137 iovec.iov_len = min(iter->count,
3138 iter->bvec->bv_len - iter->iov_offset);
3142 nr = file->f_op->read(file, iovec.iov_base,
3143 iovec.iov_len, io_kiocb_ppos(kiocb));
3145 nr = file->f_op->write(file, iovec.iov_base,
3146 iovec.iov_len, io_kiocb_ppos(kiocb));
3149 if (iov_iter_is_bvec(iter))
3150 kunmap(iter->bvec->bv_page);
3158 if (nr != iovec.iov_len)
3160 iov_iter_advance(iter, nr);
3166 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3167 const struct iovec *fast_iov, struct iov_iter *iter)
3169 struct io_async_rw *rw = req->async_data;
3171 memcpy(&rw->iter, iter, sizeof(*iter));
3172 rw->free_iovec = iovec;
3174 /* can only be fixed buffers, no need to do anything */
3175 if (iter->type == ITER_BVEC)
3178 unsigned iov_off = 0;
3180 rw->iter.iov = rw->fast_iov;
3181 if (iter->iov != fast_iov) {
3182 iov_off = iter->iov - fast_iov;
3183 rw->iter.iov += iov_off;
3185 if (rw->fast_iov != fast_iov)
3186 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3187 sizeof(struct iovec) * iter->nr_segs);
3189 req->flags |= REQ_F_NEED_CLEANUP;
3193 static inline int __io_alloc_async_data(struct io_kiocb *req)
3195 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3196 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3197 return req->async_data == NULL;
3200 static int io_alloc_async_data(struct io_kiocb *req)
3202 if (!io_op_defs[req->opcode].needs_async_data)
3205 return __io_alloc_async_data(req);
3208 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3209 const struct iovec *fast_iov,
3210 struct iov_iter *iter, bool force)
3212 if (!force && !io_op_defs[req->opcode].needs_async_data)
3214 if (!req->async_data) {
3215 if (__io_alloc_async_data(req))
3218 io_req_map_rw(req, iovec, fast_iov, iter);
3223 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3225 struct io_async_rw *iorw = req->async_data;
3226 struct iovec *iov = iorw->fast_iov;
3229 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3230 if (unlikely(ret < 0))
3233 iorw->bytes_done = 0;
3234 iorw->free_iovec = iov;
3236 req->flags |= REQ_F_NEED_CLEANUP;
3240 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3244 ret = io_prep_rw(req, sqe);
3248 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3251 /* either don't need iovec imported or already have it */
3252 if (!req->async_data)
3254 return io_rw_prep_async(req, READ);
3258 * This is our waitqueue callback handler, registered through lock_page_async()
3259 * when we initially tried to do the IO with the iocb armed our waitqueue.
3260 * This gets called when the page is unlocked, and we generally expect that to
3261 * happen when the page IO is completed and the page is now uptodate. This will
3262 * queue a task_work based retry of the operation, attempting to copy the data
3263 * again. If the latter fails because the page was NOT uptodate, then we will
3264 * do a thread based blocking retry of the operation. That's the unexpected
3267 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3268 int sync, void *arg)
3270 struct wait_page_queue *wpq;
3271 struct io_kiocb *req = wait->private;
3272 struct wait_page_key *key = arg;
3275 wpq = container_of(wait, struct wait_page_queue, wait);
3277 if (!wake_page_match(wpq, key))
3280 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3281 list_del_init(&wait->entry);
3283 init_task_work(&req->task_work, io_req_task_submit);
3284 percpu_ref_get(&req->ctx->refs);
3286 /* submit ref gets dropped, acquire a new one */
3287 refcount_inc(&req->refs);
3288 ret = io_req_task_work_add(req, true);
3289 if (unlikely(ret)) {
3290 struct task_struct *tsk;
3292 /* queue just for cancelation */
3293 init_task_work(&req->task_work, io_req_task_cancel);
3294 tsk = io_wq_get_task(req->ctx->io_wq);
3295 task_work_add(tsk, &req->task_work, TWA_NONE);
3296 wake_up_process(tsk);
3302 * This controls whether a given IO request should be armed for async page
3303 * based retry. If we return false here, the request is handed to the async
3304 * worker threads for retry. If we're doing buffered reads on a regular file,
3305 * we prepare a private wait_page_queue entry and retry the operation. This
3306 * will either succeed because the page is now uptodate and unlocked, or it
3307 * will register a callback when the page is unlocked at IO completion. Through
3308 * that callback, io_uring uses task_work to setup a retry of the operation.
3309 * That retry will attempt the buffered read again. The retry will generally
3310 * succeed, or in rare cases where it fails, we then fall back to using the
3311 * async worker threads for a blocking retry.
3313 static bool io_rw_should_retry(struct io_kiocb *req)
3315 struct io_async_rw *rw = req->async_data;
3316 struct wait_page_queue *wait = &rw->wpq;
3317 struct kiocb *kiocb = &req->rw.kiocb;
3319 /* never retry for NOWAIT, we just complete with -EAGAIN */
3320 if (req->flags & REQ_F_NOWAIT)
3323 /* Only for buffered IO */
3324 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3328 * just use poll if we can, and don't attempt if the fs doesn't
3329 * support callback based unlocks
3331 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3334 wait->wait.func = io_async_buf_func;
3335 wait->wait.private = req;
3336 wait->wait.flags = 0;
3337 INIT_LIST_HEAD(&wait->wait.entry);
3338 kiocb->ki_flags |= IOCB_WAITQ;
3339 kiocb->ki_flags &= ~IOCB_NOWAIT;
3340 kiocb->ki_waitq = wait;
3344 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3346 if (req->file->f_op->read_iter)
3347 return call_read_iter(req->file, &req->rw.kiocb, iter);
3348 else if (req->file->f_op->read)
3349 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3354 static int io_read(struct io_kiocb *req, bool force_nonblock,
3355 struct io_comp_state *cs)
3357 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3358 struct kiocb *kiocb = &req->rw.kiocb;
3359 struct iov_iter __iter, *iter = &__iter;
3360 struct io_async_rw *rw = req->async_data;
3361 ssize_t io_size, ret, ret2;
3368 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3371 iov_count = iov_iter_count(iter);
3373 req->result = io_size;
3376 /* Ensure we clear previously set non-block flag */
3377 if (!force_nonblock)
3378 kiocb->ki_flags &= ~IOCB_NOWAIT;
3380 kiocb->ki_flags |= IOCB_NOWAIT;
3383 /* If the file doesn't support async, just async punt */
3384 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3388 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3392 ret = io_iter_do_read(req, iter);
3396 } else if (ret == -EIOCBQUEUED) {
3399 } else if (ret == -EAGAIN) {
3400 /* IOPOLL retry should happen for io-wq threads */
3401 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3403 /* no retry on NONBLOCK marked file */
3404 if (req->file->f_flags & O_NONBLOCK)
3406 /* some cases will consume bytes even on error returns */
3407 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3410 } else if (ret < 0) {
3411 /* make sure -ERESTARTSYS -> -EINTR is done */
3415 /* read it all, or we did blocking attempt. no retry. */
3416 if (!iov_iter_count(iter) || !force_nonblock ||
3417 (req->file->f_flags & O_NONBLOCK))
3422 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3429 rw = req->async_data;
3430 /* it's copied and will be cleaned with ->io */
3432 /* now use our persistent iterator, if we aren't already */
3435 rw->bytes_done += ret;
3436 /* if we can retry, do so with the callbacks armed */
3437 if (!io_rw_should_retry(req)) {
3438 kiocb->ki_flags &= ~IOCB_WAITQ;
3443 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3444 * get -EIOCBQUEUED, then we'll get a notification when the desired
3445 * page gets unlocked. We can also get a partial read here, and if we
3446 * do, then just retry at the new offset.
3448 ret = io_iter_do_read(req, iter);
3449 if (ret == -EIOCBQUEUED) {
3452 } else if (ret > 0 && ret < io_size) {
3453 /* we got some bytes, but not all. retry. */
3457 kiocb_done(kiocb, ret, cs);
3460 /* it's reportedly faster than delegating the null check to kfree() */
3466 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3470 ret = io_prep_rw(req, sqe);
3474 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3477 /* either don't need iovec imported or already have it */
3478 if (!req->async_data)
3480 return io_rw_prep_async(req, WRITE);
3483 static int io_write(struct io_kiocb *req, bool force_nonblock,
3484 struct io_comp_state *cs)
3486 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3487 struct kiocb *kiocb = &req->rw.kiocb;
3488 struct iov_iter __iter, *iter = &__iter;
3489 struct io_async_rw *rw = req->async_data;
3491 ssize_t ret, ret2, io_size;
3496 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3499 iov_count = iov_iter_count(iter);
3501 req->result = io_size;
3503 /* Ensure we clear previously set non-block flag */
3504 if (!force_nonblock)
3505 kiocb->ki_flags &= ~IOCB_NOWAIT;
3507 kiocb->ki_flags |= IOCB_NOWAIT;
3509 /* If the file doesn't support async, just async punt */
3510 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3513 /* file path doesn't support NOWAIT for non-direct_IO */
3514 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3515 (req->flags & REQ_F_ISREG))
3518 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3523 * Open-code file_start_write here to grab freeze protection,
3524 * which will be released by another thread in
3525 * io_complete_rw(). Fool lockdep by telling it the lock got
3526 * released so that it doesn't complain about the held lock when
3527 * we return to userspace.
3529 if (req->flags & REQ_F_ISREG) {
3530 __sb_start_write(file_inode(req->file)->i_sb,
3531 SB_FREEZE_WRITE, true);
3532 __sb_writers_release(file_inode(req->file)->i_sb,
3535 kiocb->ki_flags |= IOCB_WRITE;
3537 if (req->file->f_op->write_iter)
3538 ret2 = call_write_iter(req->file, kiocb, iter);
3539 else if (req->file->f_op->write)
3540 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3545 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3546 * retry them without IOCB_NOWAIT.
3548 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3550 /* no retry on NONBLOCK marked file */
3551 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3553 if (!force_nonblock || ret2 != -EAGAIN) {
3554 /* IOPOLL retry should happen for io-wq threads */
3555 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3558 kiocb_done(kiocb, ret2, cs);
3561 /* some cases will consume bytes even on error returns */
3562 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3563 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3568 /* it's reportedly faster than delegating the null check to kfree() */
3574 static int __io_splice_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 struct io_splice* sp = &req->splice;
3578 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3584 sp->len = READ_ONCE(sqe->len);
3585 sp->flags = READ_ONCE(sqe->splice_flags);
3587 if (unlikely(sp->flags & ~valid_flags))
3590 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3591 (sp->flags & SPLICE_F_FD_IN_FIXED));
3594 req->flags |= REQ_F_NEED_CLEANUP;
3596 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3598 * Splice operation will be punted aync, and here need to
3599 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3601 io_req_init_async(req);
3602 req->work.flags |= IO_WQ_WORK_UNBOUND;
3608 static int io_tee_prep(struct io_kiocb *req,
3609 const struct io_uring_sqe *sqe)
3611 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3613 return __io_splice_prep(req, sqe);
3616 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3618 struct io_splice *sp = &req->splice;
3619 struct file *in = sp->file_in;
3620 struct file *out = sp->file_out;
3621 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3627 ret = do_tee(in, out, sp->len, flags);
3629 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3630 req->flags &= ~REQ_F_NEED_CLEANUP;
3633 req_set_fail_links(req);
3634 io_req_complete(req, ret);
3638 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3640 struct io_splice* sp = &req->splice;
3642 sp->off_in = READ_ONCE(sqe->splice_off_in);
3643 sp->off_out = READ_ONCE(sqe->off);
3644 return __io_splice_prep(req, sqe);
3647 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3649 struct io_splice *sp = &req->splice;
3650 struct file *in = sp->file_in;
3651 struct file *out = sp->file_out;
3652 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3653 loff_t *poff_in, *poff_out;
3659 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3660 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3663 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3665 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3666 req->flags &= ~REQ_F_NEED_CLEANUP;
3669 req_set_fail_links(req);
3670 io_req_complete(req, ret);
3675 * IORING_OP_NOP just posts a completion event, nothing else.
3677 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3679 struct io_ring_ctx *ctx = req->ctx;
3681 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3684 __io_req_complete(req, 0, 0, cs);
3688 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3690 struct io_ring_ctx *ctx = req->ctx;
3695 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3697 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3700 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3701 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3704 req->sync.off = READ_ONCE(sqe->off);
3705 req->sync.len = READ_ONCE(sqe->len);
3709 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3711 loff_t end = req->sync.off + req->sync.len;
3714 /* fsync always requires a blocking context */
3718 ret = vfs_fsync_range(req->file, req->sync.off,
3719 end > 0 ? end : LLONG_MAX,
3720 req->sync.flags & IORING_FSYNC_DATASYNC);
3722 req_set_fail_links(req);
3723 io_req_complete(req, ret);
3727 static int io_fallocate_prep(struct io_kiocb *req,
3728 const struct io_uring_sqe *sqe)
3730 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3732 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3735 req->sync.off = READ_ONCE(sqe->off);
3736 req->sync.len = READ_ONCE(sqe->addr);
3737 req->sync.mode = READ_ONCE(sqe->len);
3741 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3745 /* fallocate always requiring blocking context */
3748 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3751 req_set_fail_links(req);
3752 io_req_complete(req, ret);
3756 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3758 const char __user *fname;
3761 if (unlikely(sqe->ioprio || sqe->buf_index))
3763 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3766 /* open.how should be already initialised */
3767 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3768 req->open.how.flags |= O_LARGEFILE;
3770 req->open.dfd = READ_ONCE(sqe->fd);
3771 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3772 req->open.filename = getname(fname);
3773 if (IS_ERR(req->open.filename)) {
3774 ret = PTR_ERR(req->open.filename);
3775 req->open.filename = NULL;
3778 req->open.nofile = rlimit(RLIMIT_NOFILE);
3779 req->flags |= REQ_F_NEED_CLEANUP;
3783 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3787 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3789 mode = READ_ONCE(sqe->len);
3790 flags = READ_ONCE(sqe->open_flags);
3791 req->open.how = build_open_how(flags, mode);
3792 return __io_openat_prep(req, sqe);
3795 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3797 struct open_how __user *how;
3801 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3803 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3804 len = READ_ONCE(sqe->len);
3805 if (len < OPEN_HOW_SIZE_VER0)
3808 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3813 return __io_openat_prep(req, sqe);
3816 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3818 struct open_flags op;
3825 ret = build_open_flags(&req->open.how, &op);
3829 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3833 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3836 ret = PTR_ERR(file);
3838 fsnotify_open(file);
3839 fd_install(ret, file);
3842 putname(req->open.filename);
3843 req->flags &= ~REQ_F_NEED_CLEANUP;
3845 req_set_fail_links(req);
3846 io_req_complete(req, ret);
3850 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3852 return io_openat2(req, force_nonblock);
3855 static int io_remove_buffers_prep(struct io_kiocb *req,
3856 const struct io_uring_sqe *sqe)
3858 struct io_provide_buf *p = &req->pbuf;
3861 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3864 tmp = READ_ONCE(sqe->fd);
3865 if (!tmp || tmp > USHRT_MAX)
3868 memset(p, 0, sizeof(*p));
3870 p->bgid = READ_ONCE(sqe->buf_group);
3874 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3875 int bgid, unsigned nbufs)
3879 /* shouldn't happen */
3883 /* the head kbuf is the list itself */
3884 while (!list_empty(&buf->list)) {
3885 struct io_buffer *nxt;
3887 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3888 list_del(&nxt->list);
3895 idr_remove(&ctx->io_buffer_idr, bgid);
3900 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3901 struct io_comp_state *cs)
3903 struct io_provide_buf *p = &req->pbuf;
3904 struct io_ring_ctx *ctx = req->ctx;
3905 struct io_buffer *head;
3908 io_ring_submit_lock(ctx, !force_nonblock);
3910 lockdep_assert_held(&ctx->uring_lock);
3913 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3915 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3917 io_ring_submit_lock(ctx, !force_nonblock);
3919 req_set_fail_links(req);
3920 __io_req_complete(req, ret, 0, cs);
3924 static int io_provide_buffers_prep(struct io_kiocb *req,
3925 const struct io_uring_sqe *sqe)
3927 struct io_provide_buf *p = &req->pbuf;
3930 if (sqe->ioprio || sqe->rw_flags)
3933 tmp = READ_ONCE(sqe->fd);
3934 if (!tmp || tmp > USHRT_MAX)
3937 p->addr = READ_ONCE(sqe->addr);
3938 p->len = READ_ONCE(sqe->len);
3940 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3943 p->bgid = READ_ONCE(sqe->buf_group);
3944 tmp = READ_ONCE(sqe->off);
3945 if (tmp > USHRT_MAX)
3951 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3953 struct io_buffer *buf;
3954 u64 addr = pbuf->addr;
3955 int i, bid = pbuf->bid;
3957 for (i = 0; i < pbuf->nbufs; i++) {
3958 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3963 buf->len = pbuf->len;
3968 INIT_LIST_HEAD(&buf->list);
3971 list_add_tail(&buf->list, &(*head)->list);
3975 return i ? i : -ENOMEM;
3978 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3979 struct io_comp_state *cs)
3981 struct io_provide_buf *p = &req->pbuf;
3982 struct io_ring_ctx *ctx = req->ctx;
3983 struct io_buffer *head, *list;
3986 io_ring_submit_lock(ctx, !force_nonblock);
3988 lockdep_assert_held(&ctx->uring_lock);
3990 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3992 ret = io_add_buffers(p, &head);
3997 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4000 __io_remove_buffers(ctx, head, p->bgid, -1U);
4005 io_ring_submit_unlock(ctx, !force_nonblock);
4007 req_set_fail_links(req);
4008 __io_req_complete(req, ret, 0, cs);
4012 static int io_epoll_ctl_prep(struct io_kiocb *req,
4013 const struct io_uring_sqe *sqe)
4015 #if defined(CONFIG_EPOLL)
4016 if (sqe->ioprio || sqe->buf_index)
4018 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4021 req->epoll.epfd = READ_ONCE(sqe->fd);
4022 req->epoll.op = READ_ONCE(sqe->len);
4023 req->epoll.fd = READ_ONCE(sqe->off);
4025 if (ep_op_has_event(req->epoll.op)) {
4026 struct epoll_event __user *ev;
4028 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4029 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4039 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4040 struct io_comp_state *cs)
4042 #if defined(CONFIG_EPOLL)
4043 struct io_epoll *ie = &req->epoll;
4046 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4047 if (force_nonblock && ret == -EAGAIN)
4051 req_set_fail_links(req);
4052 __io_req_complete(req, ret, 0, cs);
4059 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4061 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4062 if (sqe->ioprio || sqe->buf_index || sqe->off)
4064 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4067 req->madvise.addr = READ_ONCE(sqe->addr);
4068 req->madvise.len = READ_ONCE(sqe->len);
4069 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4076 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4078 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4079 struct io_madvise *ma = &req->madvise;
4085 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4087 req_set_fail_links(req);
4088 io_req_complete(req, ret);
4095 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4097 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4099 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4102 req->fadvise.offset = READ_ONCE(sqe->off);
4103 req->fadvise.len = READ_ONCE(sqe->len);
4104 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4108 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4110 struct io_fadvise *fa = &req->fadvise;
4113 if (force_nonblock) {
4114 switch (fa->advice) {
4115 case POSIX_FADV_NORMAL:
4116 case POSIX_FADV_RANDOM:
4117 case POSIX_FADV_SEQUENTIAL:
4124 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4126 req_set_fail_links(req);
4127 io_req_complete(req, ret);
4131 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4133 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4135 if (sqe->ioprio || sqe->buf_index)
4137 if (req->flags & REQ_F_FIXED_FILE)
4140 req->statx.dfd = READ_ONCE(sqe->fd);
4141 req->statx.mask = READ_ONCE(sqe->len);
4142 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4143 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4144 req->statx.flags = READ_ONCE(sqe->statx_flags);
4149 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4151 struct io_statx *ctx = &req->statx;
4154 if (force_nonblock) {
4155 /* only need file table for an actual valid fd */
4156 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4157 req->flags |= REQ_F_NO_FILE_TABLE;
4161 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4165 req_set_fail_links(req);
4166 io_req_complete(req, ret);
4170 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4173 * If we queue this for async, it must not be cancellable. That would
4174 * leave the 'file' in an undeterminate state, and here need to modify
4175 * io_wq_work.flags, so initialize io_wq_work firstly.
4177 io_req_init_async(req);
4178 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4180 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4182 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4183 sqe->rw_flags || sqe->buf_index)
4185 if (req->flags & REQ_F_FIXED_FILE)
4188 req->close.fd = READ_ONCE(sqe->fd);
4189 if ((req->file && req->file->f_op == &io_uring_fops))
4192 req->close.put_file = NULL;
4196 static int io_close(struct io_kiocb *req, bool force_nonblock,
4197 struct io_comp_state *cs)
4199 struct io_close *close = &req->close;
4202 /* might be already done during nonblock submission */
4203 if (!close->put_file) {
4204 ret = __close_fd_get_file(close->fd, &close->put_file);
4206 return (ret == -ENOENT) ? -EBADF : ret;
4209 /* if the file has a flush method, be safe and punt to async */
4210 if (close->put_file->f_op->flush && force_nonblock) {
4211 /* was never set, but play safe */
4212 req->flags &= ~REQ_F_NOWAIT;
4213 /* avoid grabbing files - we don't need the files */
4214 req->flags |= REQ_F_NO_FILE_TABLE;
4218 /* No ->flush() or already async, safely close from here */
4219 ret = filp_close(close->put_file, req->work.identity->files);
4221 req_set_fail_links(req);
4222 fput(close->put_file);
4223 close->put_file = NULL;
4224 __io_req_complete(req, ret, 0, cs);
4228 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4230 struct io_ring_ctx *ctx = req->ctx;
4235 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4237 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4240 req->sync.off = READ_ONCE(sqe->off);
4241 req->sync.len = READ_ONCE(sqe->len);
4242 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4246 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4250 /* sync_file_range always requires a blocking context */
4254 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4257 req_set_fail_links(req);
4258 io_req_complete(req, ret);
4262 #if defined(CONFIG_NET)
4263 static int io_setup_async_msg(struct io_kiocb *req,
4264 struct io_async_msghdr *kmsg)
4266 struct io_async_msghdr *async_msg = req->async_data;
4270 if (io_alloc_async_data(req)) {
4271 if (kmsg->iov != kmsg->fast_iov)
4275 async_msg = req->async_data;
4276 req->flags |= REQ_F_NEED_CLEANUP;
4277 memcpy(async_msg, kmsg, sizeof(*kmsg));
4281 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4282 struct io_async_msghdr *iomsg)
4284 iomsg->iov = iomsg->fast_iov;
4285 iomsg->msg.msg_name = &iomsg->addr;
4286 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4287 req->sr_msg.msg_flags, &iomsg->iov);
4290 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4292 struct io_async_msghdr *async_msg = req->async_data;
4293 struct io_sr_msg *sr = &req->sr_msg;
4296 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4299 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4300 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4301 sr->len = READ_ONCE(sqe->len);
4303 #ifdef CONFIG_COMPAT
4304 if (req->ctx->compat)
4305 sr->msg_flags |= MSG_CMSG_COMPAT;
4308 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4310 ret = io_sendmsg_copy_hdr(req, async_msg);
4312 req->flags |= REQ_F_NEED_CLEANUP;
4316 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4317 struct io_comp_state *cs)
4319 struct io_async_msghdr iomsg, *kmsg;
4320 struct socket *sock;
4324 sock = sock_from_file(req->file, &ret);
4325 if (unlikely(!sock))
4328 if (req->async_data) {
4329 kmsg = req->async_data;
4330 kmsg->msg.msg_name = &kmsg->addr;
4331 /* if iov is set, it's allocated already */
4333 kmsg->iov = kmsg->fast_iov;
4334 kmsg->msg.msg_iter.iov = kmsg->iov;
4336 ret = io_sendmsg_copy_hdr(req, &iomsg);
4342 flags = req->sr_msg.msg_flags;
4343 if (flags & MSG_DONTWAIT)
4344 req->flags |= REQ_F_NOWAIT;
4345 else if (force_nonblock)
4346 flags |= MSG_DONTWAIT;
4348 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4349 if (force_nonblock && ret == -EAGAIN)
4350 return io_setup_async_msg(req, kmsg);
4351 if (ret == -ERESTARTSYS)
4354 if (kmsg->iov != kmsg->fast_iov)
4356 req->flags &= ~REQ_F_NEED_CLEANUP;
4358 req_set_fail_links(req);
4359 __io_req_complete(req, ret, 0, cs);
4363 static int io_send(struct io_kiocb *req, bool force_nonblock,
4364 struct io_comp_state *cs)
4366 struct io_sr_msg *sr = &req->sr_msg;
4369 struct socket *sock;
4373 sock = sock_from_file(req->file, &ret);
4374 if (unlikely(!sock))
4377 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4381 msg.msg_name = NULL;
4382 msg.msg_control = NULL;
4383 msg.msg_controllen = 0;
4384 msg.msg_namelen = 0;
4386 flags = req->sr_msg.msg_flags;
4387 if (flags & MSG_DONTWAIT)
4388 req->flags |= REQ_F_NOWAIT;
4389 else if (force_nonblock)
4390 flags |= MSG_DONTWAIT;
4392 msg.msg_flags = flags;
4393 ret = sock_sendmsg(sock, &msg);
4394 if (force_nonblock && ret == -EAGAIN)
4396 if (ret == -ERESTARTSYS)
4400 req_set_fail_links(req);
4401 __io_req_complete(req, ret, 0, cs);
4405 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4406 struct io_async_msghdr *iomsg)
4408 struct io_sr_msg *sr = &req->sr_msg;
4409 struct iovec __user *uiov;
4413 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4414 &iomsg->uaddr, &uiov, &iov_len);
4418 if (req->flags & REQ_F_BUFFER_SELECT) {
4421 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4423 sr->len = iomsg->iov[0].iov_len;
4424 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4428 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4429 &iomsg->iov, &iomsg->msg.msg_iter,
4438 #ifdef CONFIG_COMPAT
4439 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4440 struct io_async_msghdr *iomsg)
4442 struct compat_msghdr __user *msg_compat;
4443 struct io_sr_msg *sr = &req->sr_msg;
4444 struct compat_iovec __user *uiov;
4449 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4450 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4455 uiov = compat_ptr(ptr);
4456 if (req->flags & REQ_F_BUFFER_SELECT) {
4457 compat_ssize_t clen;
4461 if (!access_ok(uiov, sizeof(*uiov)))
4463 if (__get_user(clen, &uiov->iov_len))
4467 sr->len = iomsg->iov[0].iov_len;
4470 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4471 UIO_FASTIOV, &iomsg->iov,
4472 &iomsg->msg.msg_iter, true);
4481 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4482 struct io_async_msghdr *iomsg)
4484 iomsg->msg.msg_name = &iomsg->addr;
4485 iomsg->iov = iomsg->fast_iov;
4487 #ifdef CONFIG_COMPAT
4488 if (req->ctx->compat)
4489 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4492 return __io_recvmsg_copy_hdr(req, iomsg);
4495 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4498 struct io_sr_msg *sr = &req->sr_msg;
4499 struct io_buffer *kbuf;
4501 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4506 req->flags |= REQ_F_BUFFER_SELECTED;
4510 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4512 return io_put_kbuf(req, req->sr_msg.kbuf);
4515 static int io_recvmsg_prep(struct io_kiocb *req,
4516 const struct io_uring_sqe *sqe)
4518 struct io_async_msghdr *async_msg = req->async_data;
4519 struct io_sr_msg *sr = &req->sr_msg;
4522 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4525 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4526 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4527 sr->len = READ_ONCE(sqe->len);
4528 sr->bgid = READ_ONCE(sqe->buf_group);
4530 #ifdef CONFIG_COMPAT
4531 if (req->ctx->compat)
4532 sr->msg_flags |= MSG_CMSG_COMPAT;
4535 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4537 ret = io_recvmsg_copy_hdr(req, async_msg);
4539 req->flags |= REQ_F_NEED_CLEANUP;
4543 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4544 struct io_comp_state *cs)
4546 struct io_async_msghdr iomsg, *kmsg;
4547 struct socket *sock;
4548 struct io_buffer *kbuf;
4550 int ret, cflags = 0;
4552 sock = sock_from_file(req->file, &ret);
4553 if (unlikely(!sock))
4556 if (req->async_data) {
4557 kmsg = req->async_data;
4558 kmsg->msg.msg_name = &kmsg->addr;
4559 /* if iov is set, it's allocated already */
4561 kmsg->iov = kmsg->fast_iov;
4562 kmsg->msg.msg_iter.iov = kmsg->iov;
4564 ret = io_recvmsg_copy_hdr(req, &iomsg);
4570 if (req->flags & REQ_F_BUFFER_SELECT) {
4571 kbuf = io_recv_buffer_select(req, !force_nonblock);
4573 return PTR_ERR(kbuf);
4574 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4575 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4576 1, req->sr_msg.len);
4579 flags = req->sr_msg.msg_flags;
4580 if (flags & MSG_DONTWAIT)
4581 req->flags |= REQ_F_NOWAIT;
4582 else if (force_nonblock)
4583 flags |= MSG_DONTWAIT;
4585 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4586 kmsg->uaddr, flags);
4587 if (force_nonblock && ret == -EAGAIN)
4588 return io_setup_async_msg(req, kmsg);
4589 if (ret == -ERESTARTSYS)
4592 if (req->flags & REQ_F_BUFFER_SELECTED)
4593 cflags = io_put_recv_kbuf(req);
4594 if (kmsg->iov != kmsg->fast_iov)
4596 req->flags &= ~REQ_F_NEED_CLEANUP;
4598 req_set_fail_links(req);
4599 __io_req_complete(req, ret, cflags, cs);
4603 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4604 struct io_comp_state *cs)
4606 struct io_buffer *kbuf;
4607 struct io_sr_msg *sr = &req->sr_msg;
4609 void __user *buf = sr->buf;
4610 struct socket *sock;
4613 int ret, cflags = 0;
4615 sock = sock_from_file(req->file, &ret);
4616 if (unlikely(!sock))
4619 if (req->flags & REQ_F_BUFFER_SELECT) {
4620 kbuf = io_recv_buffer_select(req, !force_nonblock);
4622 return PTR_ERR(kbuf);
4623 buf = u64_to_user_ptr(kbuf->addr);
4626 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4630 msg.msg_name = NULL;
4631 msg.msg_control = NULL;
4632 msg.msg_controllen = 0;
4633 msg.msg_namelen = 0;
4634 msg.msg_iocb = NULL;
4637 flags = req->sr_msg.msg_flags;
4638 if (flags & MSG_DONTWAIT)
4639 req->flags |= REQ_F_NOWAIT;
4640 else if (force_nonblock)
4641 flags |= MSG_DONTWAIT;
4643 ret = sock_recvmsg(sock, &msg, flags);
4644 if (force_nonblock && ret == -EAGAIN)
4646 if (ret == -ERESTARTSYS)
4649 if (req->flags & REQ_F_BUFFER_SELECTED)
4650 cflags = io_put_recv_kbuf(req);
4652 req_set_fail_links(req);
4653 __io_req_complete(req, ret, cflags, cs);
4657 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4659 struct io_accept *accept = &req->accept;
4661 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4663 if (sqe->ioprio || sqe->len || sqe->buf_index)
4666 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4667 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4668 accept->flags = READ_ONCE(sqe->accept_flags);
4669 accept->nofile = rlimit(RLIMIT_NOFILE);
4673 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4674 struct io_comp_state *cs)
4676 struct io_accept *accept = &req->accept;
4677 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4680 if (req->file->f_flags & O_NONBLOCK)
4681 req->flags |= REQ_F_NOWAIT;
4683 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4684 accept->addr_len, accept->flags,
4686 if (ret == -EAGAIN && force_nonblock)
4689 if (ret == -ERESTARTSYS)
4691 req_set_fail_links(req);
4693 __io_req_complete(req, ret, 0, cs);
4697 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4699 struct io_connect *conn = &req->connect;
4700 struct io_async_connect *io = req->async_data;
4702 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4704 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4707 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4708 conn->addr_len = READ_ONCE(sqe->addr2);
4713 return move_addr_to_kernel(conn->addr, conn->addr_len,
4717 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4718 struct io_comp_state *cs)
4720 struct io_async_connect __io, *io;
4721 unsigned file_flags;
4724 if (req->async_data) {
4725 io = req->async_data;
4727 ret = move_addr_to_kernel(req->connect.addr,
4728 req->connect.addr_len,
4735 file_flags = force_nonblock ? O_NONBLOCK : 0;
4737 ret = __sys_connect_file(req->file, &io->address,
4738 req->connect.addr_len, file_flags);
4739 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4740 if (req->async_data)
4742 if (io_alloc_async_data(req)) {
4746 io = req->async_data;
4747 memcpy(req->async_data, &__io, sizeof(__io));
4750 if (ret == -ERESTARTSYS)
4754 req_set_fail_links(req);
4755 __io_req_complete(req, ret, 0, cs);
4758 #else /* !CONFIG_NET */
4759 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4764 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4765 struct io_comp_state *cs)
4770 static int io_send(struct io_kiocb *req, bool force_nonblock,
4771 struct io_comp_state *cs)
4776 static int io_recvmsg_prep(struct io_kiocb *req,
4777 const struct io_uring_sqe *sqe)
4782 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4783 struct io_comp_state *cs)
4788 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4789 struct io_comp_state *cs)
4794 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4799 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4800 struct io_comp_state *cs)
4805 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4810 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4811 struct io_comp_state *cs)
4815 #endif /* CONFIG_NET */
4817 struct io_poll_table {
4818 struct poll_table_struct pt;
4819 struct io_kiocb *req;
4823 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4824 __poll_t mask, task_work_func_t func)
4829 /* for instances that support it check for an event match first: */
4830 if (mask && !(mask & poll->events))
4833 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4835 list_del_init(&poll->wait.entry);
4838 init_task_work(&req->task_work, func);
4839 percpu_ref_get(&req->ctx->refs);
4842 * If we using the signalfd wait_queue_head for this wakeup, then
4843 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4844 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4845 * either, as the normal wakeup will suffice.
4847 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4850 * If this fails, then the task is exiting. When a task exits, the
4851 * work gets canceled, so just cancel this request as well instead
4852 * of executing it. We can't safely execute it anyway, as we may not
4853 * have the needed state needed for it anyway.
4855 ret = io_req_task_work_add(req, twa_signal_ok);
4856 if (unlikely(ret)) {
4857 struct task_struct *tsk;
4859 WRITE_ONCE(poll->canceled, true);
4860 tsk = io_wq_get_task(req->ctx->io_wq);
4861 task_work_add(tsk, &req->task_work, TWA_NONE);
4862 wake_up_process(tsk);
4867 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4868 __acquires(&req->ctx->completion_lock)
4870 struct io_ring_ctx *ctx = req->ctx;
4872 if (!req->result && !READ_ONCE(poll->canceled)) {
4873 struct poll_table_struct pt = { ._key = poll->events };
4875 req->result = vfs_poll(req->file, &pt) & poll->events;
4878 spin_lock_irq(&ctx->completion_lock);
4879 if (!req->result && !READ_ONCE(poll->canceled)) {
4880 add_wait_queue(poll->head, &poll->wait);
4887 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4889 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4890 if (req->opcode == IORING_OP_POLL_ADD)
4891 return req->async_data;
4892 return req->apoll->double_poll;
4895 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4897 if (req->opcode == IORING_OP_POLL_ADD)
4899 return &req->apoll->poll;
4902 static void io_poll_remove_double(struct io_kiocb *req)
4904 struct io_poll_iocb *poll = io_poll_get_double(req);
4906 lockdep_assert_held(&req->ctx->completion_lock);
4908 if (poll && poll->head) {
4909 struct wait_queue_head *head = poll->head;
4911 spin_lock(&head->lock);
4912 list_del_init(&poll->wait.entry);
4913 if (poll->wait.private)
4914 refcount_dec(&req->refs);
4916 spin_unlock(&head->lock);
4920 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4922 struct io_ring_ctx *ctx = req->ctx;
4924 io_poll_remove_double(req);
4925 req->poll.done = true;
4926 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4927 io_commit_cqring(ctx);
4930 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4932 struct io_ring_ctx *ctx = req->ctx;
4934 if (io_poll_rewait(req, &req->poll)) {
4935 spin_unlock_irq(&ctx->completion_lock);
4939 hash_del(&req->hash_node);
4940 io_poll_complete(req, req->result, 0);
4941 spin_unlock_irq(&ctx->completion_lock);
4943 *nxt = io_put_req_find_next(req);
4944 io_cqring_ev_posted(ctx);
4947 static void io_poll_task_func(struct callback_head *cb)
4949 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4950 struct io_ring_ctx *ctx = req->ctx;
4951 struct io_kiocb *nxt = NULL;
4953 io_poll_task_handler(req, &nxt);
4955 __io_req_task_submit(nxt);
4956 percpu_ref_put(&ctx->refs);
4959 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4960 int sync, void *key)
4962 struct io_kiocb *req = wait->private;
4963 struct io_poll_iocb *poll = io_poll_get_single(req);
4964 __poll_t mask = key_to_poll(key);
4966 /* for instances that support it check for an event match first: */
4967 if (mask && !(mask & poll->events))
4970 list_del_init(&wait->entry);
4972 if (poll && poll->head) {
4975 spin_lock(&poll->head->lock);
4976 done = list_empty(&poll->wait.entry);
4978 list_del_init(&poll->wait.entry);
4979 /* make sure double remove sees this as being gone */
4980 wait->private = NULL;
4981 spin_unlock(&poll->head->lock);
4983 __io_async_wake(req, poll, mask, io_poll_task_func);
4985 refcount_dec(&req->refs);
4989 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4990 wait_queue_func_t wake_func)
4994 poll->canceled = false;
4995 poll->events = events;
4996 INIT_LIST_HEAD(&poll->wait.entry);
4997 init_waitqueue_func_entry(&poll->wait, wake_func);
5000 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5001 struct wait_queue_head *head,
5002 struct io_poll_iocb **poll_ptr)
5004 struct io_kiocb *req = pt->req;
5007 * If poll->head is already set, it's because the file being polled
5008 * uses multiple waitqueues for poll handling (eg one for read, one
5009 * for write). Setup a separate io_poll_iocb if this happens.
5011 if (unlikely(poll->head)) {
5012 struct io_poll_iocb *poll_one = poll;
5014 /* already have a 2nd entry, fail a third attempt */
5016 pt->error = -EINVAL;
5019 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5021 pt->error = -ENOMEM;
5024 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5025 refcount_inc(&req->refs);
5026 poll->wait.private = req;
5033 if (poll->events & EPOLLEXCLUSIVE)
5034 add_wait_queue_exclusive(head, &poll->wait);
5036 add_wait_queue(head, &poll->wait);
5039 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5040 struct poll_table_struct *p)
5042 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5043 struct async_poll *apoll = pt->req->apoll;
5045 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5048 static void io_async_task_func(struct callback_head *cb)
5050 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5051 struct async_poll *apoll = req->apoll;
5052 struct io_ring_ctx *ctx = req->ctx;
5054 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5056 if (io_poll_rewait(req, &apoll->poll)) {
5057 spin_unlock_irq(&ctx->completion_lock);
5058 percpu_ref_put(&ctx->refs);
5062 /* If req is still hashed, it cannot have been canceled. Don't check. */
5063 if (hash_hashed(&req->hash_node))
5064 hash_del(&req->hash_node);
5066 io_poll_remove_double(req);
5067 spin_unlock_irq(&ctx->completion_lock);
5069 if (!READ_ONCE(apoll->poll.canceled))
5070 __io_req_task_submit(req);
5072 __io_req_task_cancel(req, -ECANCELED);
5074 percpu_ref_put(&ctx->refs);
5075 kfree(apoll->double_poll);
5079 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5082 struct io_kiocb *req = wait->private;
5083 struct io_poll_iocb *poll = &req->apoll->poll;
5085 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5088 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5091 static void io_poll_req_insert(struct io_kiocb *req)
5093 struct io_ring_ctx *ctx = req->ctx;
5094 struct hlist_head *list;
5096 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5097 hlist_add_head(&req->hash_node, list);
5100 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5101 struct io_poll_iocb *poll,
5102 struct io_poll_table *ipt, __poll_t mask,
5103 wait_queue_func_t wake_func)
5104 __acquires(&ctx->completion_lock)
5106 struct io_ring_ctx *ctx = req->ctx;
5107 bool cancel = false;
5109 io_init_poll_iocb(poll, mask, wake_func);
5110 poll->file = req->file;
5111 poll->wait.private = req;
5113 ipt->pt._key = mask;
5115 ipt->error = -EINVAL;
5117 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5119 spin_lock_irq(&ctx->completion_lock);
5120 if (likely(poll->head)) {
5121 spin_lock(&poll->head->lock);
5122 if (unlikely(list_empty(&poll->wait.entry))) {
5128 if (mask || ipt->error)
5129 list_del_init(&poll->wait.entry);
5131 WRITE_ONCE(poll->canceled, true);
5132 else if (!poll->done) /* actually waiting for an event */
5133 io_poll_req_insert(req);
5134 spin_unlock(&poll->head->lock);
5140 static bool io_arm_poll_handler(struct io_kiocb *req)
5142 const struct io_op_def *def = &io_op_defs[req->opcode];
5143 struct io_ring_ctx *ctx = req->ctx;
5144 struct async_poll *apoll;
5145 struct io_poll_table ipt;
5149 if (!req->file || !file_can_poll(req->file))
5151 if (req->flags & REQ_F_POLLED)
5155 else if (def->pollout)
5159 /* if we can't nonblock try, then no point in arming a poll handler */
5160 if (!io_file_supports_async(req->file, rw))
5163 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5164 if (unlikely(!apoll))
5166 apoll->double_poll = NULL;
5168 req->flags |= REQ_F_POLLED;
5170 INIT_HLIST_NODE(&req->hash_node);
5174 mask |= POLLIN | POLLRDNORM;
5176 mask |= POLLOUT | POLLWRNORM;
5178 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5179 if ((req->opcode == IORING_OP_RECVMSG) &&
5180 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5183 mask |= POLLERR | POLLPRI;
5185 ipt.pt._qproc = io_async_queue_proc;
5187 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5189 if (ret || ipt.error) {
5190 io_poll_remove_double(req);
5191 spin_unlock_irq(&ctx->completion_lock);
5192 kfree(apoll->double_poll);
5196 spin_unlock_irq(&ctx->completion_lock);
5197 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5198 apoll->poll.events);
5202 static bool __io_poll_remove_one(struct io_kiocb *req,
5203 struct io_poll_iocb *poll)
5205 bool do_complete = false;
5207 spin_lock(&poll->head->lock);
5208 WRITE_ONCE(poll->canceled, true);
5209 if (!list_empty(&poll->wait.entry)) {
5210 list_del_init(&poll->wait.entry);
5213 spin_unlock(&poll->head->lock);
5214 hash_del(&req->hash_node);
5218 static bool io_poll_remove_one(struct io_kiocb *req)
5222 io_poll_remove_double(req);
5224 if (req->opcode == IORING_OP_POLL_ADD) {
5225 do_complete = __io_poll_remove_one(req, &req->poll);
5227 struct async_poll *apoll = req->apoll;
5229 /* non-poll requests have submit ref still */
5230 do_complete = __io_poll_remove_one(req, &apoll->poll);
5233 kfree(apoll->double_poll);
5239 io_cqring_fill_event(req, -ECANCELED);
5240 io_commit_cqring(req->ctx);
5241 req_set_fail_links(req);
5242 io_put_req_deferred(req, 1);
5249 * Returns true if we found and killed one or more poll requests
5251 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5253 struct hlist_node *tmp;
5254 struct io_kiocb *req;
5257 spin_lock_irq(&ctx->completion_lock);
5258 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5259 struct hlist_head *list;
5261 list = &ctx->cancel_hash[i];
5262 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5263 if (io_task_match(req, tsk))
5264 posted += io_poll_remove_one(req);
5267 spin_unlock_irq(&ctx->completion_lock);
5270 io_cqring_ev_posted(ctx);
5275 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5277 struct hlist_head *list;
5278 struct io_kiocb *req;
5280 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5281 hlist_for_each_entry(req, list, hash_node) {
5282 if (sqe_addr != req->user_data)
5284 if (io_poll_remove_one(req))
5292 static int io_poll_remove_prep(struct io_kiocb *req,
5293 const struct io_uring_sqe *sqe)
5295 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5297 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5301 req->poll.addr = READ_ONCE(sqe->addr);
5306 * Find a running poll command that matches one specified in sqe->addr,
5307 * and remove it if found.
5309 static int io_poll_remove(struct io_kiocb *req)
5311 struct io_ring_ctx *ctx = req->ctx;
5315 addr = req->poll.addr;
5316 spin_lock_irq(&ctx->completion_lock);
5317 ret = io_poll_cancel(ctx, addr);
5318 spin_unlock_irq(&ctx->completion_lock);
5321 req_set_fail_links(req);
5322 io_req_complete(req, ret);
5326 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5329 struct io_kiocb *req = wait->private;
5330 struct io_poll_iocb *poll = &req->poll;
5332 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5335 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5336 struct poll_table_struct *p)
5338 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5340 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5343 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5345 struct io_poll_iocb *poll = &req->poll;
5348 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5350 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5355 events = READ_ONCE(sqe->poll32_events);
5357 events = swahw32(events);
5359 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5360 (events & EPOLLEXCLUSIVE);
5364 static int io_poll_add(struct io_kiocb *req)
5366 struct io_poll_iocb *poll = &req->poll;
5367 struct io_ring_ctx *ctx = req->ctx;
5368 struct io_poll_table ipt;
5371 INIT_HLIST_NODE(&req->hash_node);
5372 ipt.pt._qproc = io_poll_queue_proc;
5374 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5377 if (mask) { /* no async, we'd stolen it */
5379 io_poll_complete(req, mask, 0);
5381 spin_unlock_irq(&ctx->completion_lock);
5384 io_cqring_ev_posted(ctx);
5390 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5392 struct io_timeout_data *data = container_of(timer,
5393 struct io_timeout_data, timer);
5394 struct io_kiocb *req = data->req;
5395 struct io_ring_ctx *ctx = req->ctx;
5396 unsigned long flags;
5398 spin_lock_irqsave(&ctx->completion_lock, flags);
5399 list_del_init(&req->timeout.list);
5400 atomic_set(&req->ctx->cq_timeouts,
5401 atomic_read(&req->ctx->cq_timeouts) + 1);
5403 io_cqring_fill_event(req, -ETIME);
5404 io_commit_cqring(ctx);
5405 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5407 io_cqring_ev_posted(ctx);
5408 req_set_fail_links(req);
5410 return HRTIMER_NORESTART;
5413 static int __io_timeout_cancel(struct io_kiocb *req)
5415 struct io_timeout_data *io = req->async_data;
5418 ret = hrtimer_try_to_cancel(&io->timer);
5421 list_del_init(&req->timeout.list);
5423 req_set_fail_links(req);
5424 io_cqring_fill_event(req, -ECANCELED);
5425 io_put_req_deferred(req, 1);
5429 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5431 struct io_kiocb *req;
5434 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5435 if (user_data == req->user_data) {
5444 return __io_timeout_cancel(req);
5447 static int io_timeout_remove_prep(struct io_kiocb *req,
5448 const struct io_uring_sqe *sqe)
5450 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5452 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5454 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5457 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5462 * Remove or update an existing timeout command
5464 static int io_timeout_remove(struct io_kiocb *req)
5466 struct io_ring_ctx *ctx = req->ctx;
5469 spin_lock_irq(&ctx->completion_lock);
5470 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5472 io_cqring_fill_event(req, ret);
5473 io_commit_cqring(ctx);
5474 spin_unlock_irq(&ctx->completion_lock);
5475 io_cqring_ev_posted(ctx);
5477 req_set_fail_links(req);
5482 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5483 bool is_timeout_link)
5485 struct io_timeout_data *data;
5487 u32 off = READ_ONCE(sqe->off);
5489 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5491 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5493 if (off && is_timeout_link)
5495 flags = READ_ONCE(sqe->timeout_flags);
5496 if (flags & ~IORING_TIMEOUT_ABS)
5499 req->timeout.off = off;
5501 if (!req->async_data && io_alloc_async_data(req))
5504 data = req->async_data;
5507 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5510 if (flags & IORING_TIMEOUT_ABS)
5511 data->mode = HRTIMER_MODE_ABS;
5513 data->mode = HRTIMER_MODE_REL;
5515 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5519 static int io_timeout(struct io_kiocb *req)
5521 struct io_ring_ctx *ctx = req->ctx;
5522 struct io_timeout_data *data = req->async_data;
5523 struct list_head *entry;
5524 u32 tail, off = req->timeout.off;
5526 spin_lock_irq(&ctx->completion_lock);
5529 * sqe->off holds how many events that need to occur for this
5530 * timeout event to be satisfied. If it isn't set, then this is
5531 * a pure timeout request, sequence isn't used.
5533 if (io_is_timeout_noseq(req)) {
5534 entry = ctx->timeout_list.prev;
5538 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5539 req->timeout.target_seq = tail + off;
5542 * Insertion sort, ensuring the first entry in the list is always
5543 * the one we need first.
5545 list_for_each_prev(entry, &ctx->timeout_list) {
5546 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5549 if (io_is_timeout_noseq(nxt))
5551 /* nxt.seq is behind @tail, otherwise would've been completed */
5552 if (off >= nxt->timeout.target_seq - tail)
5556 list_add(&req->timeout.list, entry);
5557 data->timer.function = io_timeout_fn;
5558 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5559 spin_unlock_irq(&ctx->completion_lock);
5563 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5565 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5567 return req->user_data == (unsigned long) data;
5570 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5572 enum io_wq_cancel cancel_ret;
5575 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5576 switch (cancel_ret) {
5577 case IO_WQ_CANCEL_OK:
5580 case IO_WQ_CANCEL_RUNNING:
5583 case IO_WQ_CANCEL_NOTFOUND:
5591 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5592 struct io_kiocb *req, __u64 sqe_addr,
5595 unsigned long flags;
5598 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5599 if (ret != -ENOENT) {
5600 spin_lock_irqsave(&ctx->completion_lock, flags);
5604 spin_lock_irqsave(&ctx->completion_lock, flags);
5605 ret = io_timeout_cancel(ctx, sqe_addr);
5608 ret = io_poll_cancel(ctx, sqe_addr);
5612 io_cqring_fill_event(req, ret);
5613 io_commit_cqring(ctx);
5614 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5615 io_cqring_ev_posted(ctx);
5618 req_set_fail_links(req);
5622 static int io_async_cancel_prep(struct io_kiocb *req,
5623 const struct io_uring_sqe *sqe)
5625 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5627 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5629 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5632 req->cancel.addr = READ_ONCE(sqe->addr);
5636 static int io_async_cancel(struct io_kiocb *req)
5638 struct io_ring_ctx *ctx = req->ctx;
5640 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5644 static int io_files_update_prep(struct io_kiocb *req,
5645 const struct io_uring_sqe *sqe)
5647 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5649 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5651 if (sqe->ioprio || sqe->rw_flags)
5654 req->files_update.offset = READ_ONCE(sqe->off);
5655 req->files_update.nr_args = READ_ONCE(sqe->len);
5656 if (!req->files_update.nr_args)
5658 req->files_update.arg = READ_ONCE(sqe->addr);
5662 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5663 struct io_comp_state *cs)
5665 struct io_ring_ctx *ctx = req->ctx;
5666 struct io_uring_files_update up;
5672 up.offset = req->files_update.offset;
5673 up.fds = req->files_update.arg;
5675 mutex_lock(&ctx->uring_lock);
5676 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5677 mutex_unlock(&ctx->uring_lock);
5680 req_set_fail_links(req);
5681 __io_req_complete(req, ret, 0, cs);
5685 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5687 switch (req->opcode) {
5690 case IORING_OP_READV:
5691 case IORING_OP_READ_FIXED:
5692 case IORING_OP_READ:
5693 return io_read_prep(req, sqe);
5694 case IORING_OP_WRITEV:
5695 case IORING_OP_WRITE_FIXED:
5696 case IORING_OP_WRITE:
5697 return io_write_prep(req, sqe);
5698 case IORING_OP_POLL_ADD:
5699 return io_poll_add_prep(req, sqe);
5700 case IORING_OP_POLL_REMOVE:
5701 return io_poll_remove_prep(req, sqe);
5702 case IORING_OP_FSYNC:
5703 return io_prep_fsync(req, sqe);
5704 case IORING_OP_SYNC_FILE_RANGE:
5705 return io_prep_sfr(req, sqe);
5706 case IORING_OP_SENDMSG:
5707 case IORING_OP_SEND:
5708 return io_sendmsg_prep(req, sqe);
5709 case IORING_OP_RECVMSG:
5710 case IORING_OP_RECV:
5711 return io_recvmsg_prep(req, sqe);
5712 case IORING_OP_CONNECT:
5713 return io_connect_prep(req, sqe);
5714 case IORING_OP_TIMEOUT:
5715 return io_timeout_prep(req, sqe, false);
5716 case IORING_OP_TIMEOUT_REMOVE:
5717 return io_timeout_remove_prep(req, sqe);
5718 case IORING_OP_ASYNC_CANCEL:
5719 return io_async_cancel_prep(req, sqe);
5720 case IORING_OP_LINK_TIMEOUT:
5721 return io_timeout_prep(req, sqe, true);
5722 case IORING_OP_ACCEPT:
5723 return io_accept_prep(req, sqe);
5724 case IORING_OP_FALLOCATE:
5725 return io_fallocate_prep(req, sqe);
5726 case IORING_OP_OPENAT:
5727 return io_openat_prep(req, sqe);
5728 case IORING_OP_CLOSE:
5729 return io_close_prep(req, sqe);
5730 case IORING_OP_FILES_UPDATE:
5731 return io_files_update_prep(req, sqe);
5732 case IORING_OP_STATX:
5733 return io_statx_prep(req, sqe);
5734 case IORING_OP_FADVISE:
5735 return io_fadvise_prep(req, sqe);
5736 case IORING_OP_MADVISE:
5737 return io_madvise_prep(req, sqe);
5738 case IORING_OP_OPENAT2:
5739 return io_openat2_prep(req, sqe);
5740 case IORING_OP_EPOLL_CTL:
5741 return io_epoll_ctl_prep(req, sqe);
5742 case IORING_OP_SPLICE:
5743 return io_splice_prep(req, sqe);
5744 case IORING_OP_PROVIDE_BUFFERS:
5745 return io_provide_buffers_prep(req, sqe);
5746 case IORING_OP_REMOVE_BUFFERS:
5747 return io_remove_buffers_prep(req, sqe);
5749 return io_tee_prep(req, sqe);
5752 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5757 static int io_req_defer_prep(struct io_kiocb *req,
5758 const struct io_uring_sqe *sqe)
5762 if (io_alloc_async_data(req))
5764 return io_req_prep(req, sqe);
5767 static u32 io_get_sequence(struct io_kiocb *req)
5769 struct io_kiocb *pos;
5770 struct io_ring_ctx *ctx = req->ctx;
5771 u32 total_submitted, nr_reqs = 1;
5773 if (req->flags & REQ_F_LINK_HEAD)
5774 list_for_each_entry(pos, &req->link_list, link_list)
5777 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5778 return total_submitted - nr_reqs;
5781 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5783 struct io_ring_ctx *ctx = req->ctx;
5784 struct io_defer_entry *de;
5788 /* Still need defer if there is pending req in defer list. */
5789 if (likely(list_empty_careful(&ctx->defer_list) &&
5790 !(req->flags & REQ_F_IO_DRAIN)))
5793 seq = io_get_sequence(req);
5794 /* Still a chance to pass the sequence check */
5795 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5798 if (!req->async_data) {
5799 ret = io_req_defer_prep(req, sqe);
5803 io_prep_async_link(req);
5804 de = kmalloc(sizeof(*de), GFP_KERNEL);
5808 spin_lock_irq(&ctx->completion_lock);
5809 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5810 spin_unlock_irq(&ctx->completion_lock);
5812 io_queue_async_work(req);
5813 return -EIOCBQUEUED;
5816 trace_io_uring_defer(ctx, req, req->user_data);
5819 list_add_tail(&de->list, &ctx->defer_list);
5820 spin_unlock_irq(&ctx->completion_lock);
5821 return -EIOCBQUEUED;
5824 static void io_req_drop_files(struct io_kiocb *req)
5826 struct io_ring_ctx *ctx = req->ctx;
5827 unsigned long flags;
5829 spin_lock_irqsave(&ctx->inflight_lock, flags);
5830 list_del(&req->inflight_entry);
5831 if (waitqueue_active(&ctx->inflight_wait))
5832 wake_up(&ctx->inflight_wait);
5833 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5834 req->flags &= ~REQ_F_INFLIGHT;
5835 put_files_struct(req->work.identity->files);
5836 put_nsproxy(req->work.identity->nsproxy);
5837 req->work.flags &= ~IO_WQ_WORK_FILES;
5840 static void __io_clean_op(struct io_kiocb *req)
5842 if (req->flags & REQ_F_BUFFER_SELECTED) {
5843 switch (req->opcode) {
5844 case IORING_OP_READV:
5845 case IORING_OP_READ_FIXED:
5846 case IORING_OP_READ:
5847 kfree((void *)(unsigned long)req->rw.addr);
5849 case IORING_OP_RECVMSG:
5850 case IORING_OP_RECV:
5851 kfree(req->sr_msg.kbuf);
5854 req->flags &= ~REQ_F_BUFFER_SELECTED;
5857 if (req->flags & REQ_F_NEED_CLEANUP) {
5858 switch (req->opcode) {
5859 case IORING_OP_READV:
5860 case IORING_OP_READ_FIXED:
5861 case IORING_OP_READ:
5862 case IORING_OP_WRITEV:
5863 case IORING_OP_WRITE_FIXED:
5864 case IORING_OP_WRITE: {
5865 struct io_async_rw *io = req->async_data;
5867 kfree(io->free_iovec);
5870 case IORING_OP_RECVMSG:
5871 case IORING_OP_SENDMSG: {
5872 struct io_async_msghdr *io = req->async_data;
5873 if (io->iov != io->fast_iov)
5877 case IORING_OP_SPLICE:
5879 io_put_file(req, req->splice.file_in,
5880 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5882 case IORING_OP_OPENAT:
5883 case IORING_OP_OPENAT2:
5884 if (req->open.filename)
5885 putname(req->open.filename);
5888 req->flags &= ~REQ_F_NEED_CLEANUP;
5891 if (req->flags & REQ_F_INFLIGHT)
5892 io_req_drop_files(req);
5895 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5896 struct io_comp_state *cs)
5898 struct io_ring_ctx *ctx = req->ctx;
5901 switch (req->opcode) {
5903 ret = io_nop(req, cs);
5905 case IORING_OP_READV:
5906 case IORING_OP_READ_FIXED:
5907 case IORING_OP_READ:
5908 ret = io_read(req, force_nonblock, cs);
5910 case IORING_OP_WRITEV:
5911 case IORING_OP_WRITE_FIXED:
5912 case IORING_OP_WRITE:
5913 ret = io_write(req, force_nonblock, cs);
5915 case IORING_OP_FSYNC:
5916 ret = io_fsync(req, force_nonblock);
5918 case IORING_OP_POLL_ADD:
5919 ret = io_poll_add(req);
5921 case IORING_OP_POLL_REMOVE:
5922 ret = io_poll_remove(req);
5924 case IORING_OP_SYNC_FILE_RANGE:
5925 ret = io_sync_file_range(req, force_nonblock);
5927 case IORING_OP_SENDMSG:
5928 ret = io_sendmsg(req, force_nonblock, cs);
5930 case IORING_OP_SEND:
5931 ret = io_send(req, force_nonblock, cs);
5933 case IORING_OP_RECVMSG:
5934 ret = io_recvmsg(req, force_nonblock, cs);
5936 case IORING_OP_RECV:
5937 ret = io_recv(req, force_nonblock, cs);
5939 case IORING_OP_TIMEOUT:
5940 ret = io_timeout(req);
5942 case IORING_OP_TIMEOUT_REMOVE:
5943 ret = io_timeout_remove(req);
5945 case IORING_OP_ACCEPT:
5946 ret = io_accept(req, force_nonblock, cs);
5948 case IORING_OP_CONNECT:
5949 ret = io_connect(req, force_nonblock, cs);
5951 case IORING_OP_ASYNC_CANCEL:
5952 ret = io_async_cancel(req);
5954 case IORING_OP_FALLOCATE:
5955 ret = io_fallocate(req, force_nonblock);
5957 case IORING_OP_OPENAT:
5958 ret = io_openat(req, force_nonblock);
5960 case IORING_OP_CLOSE:
5961 ret = io_close(req, force_nonblock, cs);
5963 case IORING_OP_FILES_UPDATE:
5964 ret = io_files_update(req, force_nonblock, cs);
5966 case IORING_OP_STATX:
5967 ret = io_statx(req, force_nonblock);
5969 case IORING_OP_FADVISE:
5970 ret = io_fadvise(req, force_nonblock);
5972 case IORING_OP_MADVISE:
5973 ret = io_madvise(req, force_nonblock);
5975 case IORING_OP_OPENAT2:
5976 ret = io_openat2(req, force_nonblock);
5978 case IORING_OP_EPOLL_CTL:
5979 ret = io_epoll_ctl(req, force_nonblock, cs);
5981 case IORING_OP_SPLICE:
5982 ret = io_splice(req, force_nonblock);
5984 case IORING_OP_PROVIDE_BUFFERS:
5985 ret = io_provide_buffers(req, force_nonblock, cs);
5987 case IORING_OP_REMOVE_BUFFERS:
5988 ret = io_remove_buffers(req, force_nonblock, cs);
5991 ret = io_tee(req, force_nonblock);
6001 /* If the op doesn't have a file, we're not polling for it */
6002 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6003 const bool in_async = io_wq_current_is_worker();
6005 /* workqueue context doesn't hold uring_lock, grab it now */
6007 mutex_lock(&ctx->uring_lock);
6009 io_iopoll_req_issued(req);
6012 mutex_unlock(&ctx->uring_lock);
6018 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6020 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6021 struct io_kiocb *timeout;
6024 timeout = io_prep_linked_timeout(req);
6026 io_queue_linked_timeout(timeout);
6028 /* if NO_CANCEL is set, we must still run the work */
6029 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6030 IO_WQ_WORK_CANCEL) {
6036 ret = io_issue_sqe(req, false, NULL);
6038 * We can get EAGAIN for polled IO even though we're
6039 * forcing a sync submission from here, since we can't
6040 * wait for request slots on the block side.
6049 req_set_fail_links(req);
6050 io_req_complete(req, ret);
6053 return io_steal_work(req);
6056 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6059 struct fixed_file_table *table;
6061 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6062 return table->files[index & IORING_FILE_TABLE_MASK];
6065 static struct file *io_file_get(struct io_submit_state *state,
6066 struct io_kiocb *req, int fd, bool fixed)
6068 struct io_ring_ctx *ctx = req->ctx;
6072 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6074 fd = array_index_nospec(fd, ctx->nr_user_files);
6075 file = io_file_from_index(ctx, fd);
6077 req->fixed_file_refs = &ctx->file_data->node->refs;
6078 percpu_ref_get(req->fixed_file_refs);
6081 trace_io_uring_file_get(ctx, fd);
6082 file = __io_file_get(state, fd);
6088 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6093 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6094 if (unlikely(!fixed && io_async_submit(req->ctx)))
6097 req->file = io_file_get(state, req, fd, fixed);
6098 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6103 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6105 struct io_timeout_data *data = container_of(timer,
6106 struct io_timeout_data, timer);
6107 struct io_kiocb *req = data->req;
6108 struct io_ring_ctx *ctx = req->ctx;
6109 struct io_kiocb *prev = NULL;
6110 unsigned long flags;
6112 spin_lock_irqsave(&ctx->completion_lock, flags);
6115 * We don't expect the list to be empty, that will only happen if we
6116 * race with the completion of the linked work.
6118 if (!list_empty(&req->link_list)) {
6119 prev = list_entry(req->link_list.prev, struct io_kiocb,
6121 if (refcount_inc_not_zero(&prev->refs)) {
6122 list_del_init(&req->link_list);
6123 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6128 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6131 req_set_fail_links(prev);
6132 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6135 io_req_complete(req, -ETIME);
6137 return HRTIMER_NORESTART;
6140 static void __io_queue_linked_timeout(struct io_kiocb *req)
6143 * If the list is now empty, then our linked request finished before
6144 * we got a chance to setup the timer
6146 if (!list_empty(&req->link_list)) {
6147 struct io_timeout_data *data = req->async_data;
6149 data->timer.function = io_link_timeout_fn;
6150 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6155 static void io_queue_linked_timeout(struct io_kiocb *req)
6157 struct io_ring_ctx *ctx = req->ctx;
6159 spin_lock_irq(&ctx->completion_lock);
6160 __io_queue_linked_timeout(req);
6161 spin_unlock_irq(&ctx->completion_lock);
6163 /* drop submission reference */
6167 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6169 struct io_kiocb *nxt;
6171 if (!(req->flags & REQ_F_LINK_HEAD))
6173 if (req->flags & REQ_F_LINK_TIMEOUT)
6176 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6178 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6181 req->flags |= REQ_F_LINK_TIMEOUT;
6185 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6187 struct io_kiocb *linked_timeout;
6188 struct io_kiocb *nxt;
6189 const struct cred *old_creds = NULL;
6193 linked_timeout = io_prep_linked_timeout(req);
6195 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.identity->creds &&
6196 req->work.identity->creds != current_cred()) {
6198 revert_creds(old_creds);
6199 if (old_creds == req->work.identity->creds)
6200 old_creds = NULL; /* restored original creds */
6202 old_creds = override_creds(req->work.identity->creds);
6203 req->work.flags |= IO_WQ_WORK_CREDS;
6206 ret = io_issue_sqe(req, true, cs);
6209 * We async punt it if the file wasn't marked NOWAIT, or if the file
6210 * doesn't support non-blocking read/write attempts
6212 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6213 if (!io_arm_poll_handler(req)) {
6216 * Queued up for async execution, worker will release
6217 * submit reference when the iocb is actually submitted.
6219 io_queue_async_work(req);
6223 io_queue_linked_timeout(linked_timeout);
6227 if (unlikely(ret)) {
6228 /* un-prep timeout, so it'll be killed as any other linked */
6229 req->flags &= ~REQ_F_LINK_TIMEOUT;
6230 req_set_fail_links(req);
6232 io_req_complete(req, ret);
6236 /* drop submission reference */
6237 nxt = io_put_req_find_next(req);
6239 io_queue_linked_timeout(linked_timeout);
6244 if (req->flags & REQ_F_FORCE_ASYNC)
6250 revert_creds(old_creds);
6253 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6254 struct io_comp_state *cs)
6258 ret = io_req_defer(req, sqe);
6260 if (ret != -EIOCBQUEUED) {
6262 req_set_fail_links(req);
6264 io_req_complete(req, ret);
6266 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6267 if (!req->async_data) {
6268 ret = io_req_defer_prep(req, sqe);
6274 * Never try inline submit of IOSQE_ASYNC is set, go straight
6275 * to async execution.
6277 io_req_init_async(req);
6278 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6279 io_queue_async_work(req);
6282 ret = io_req_prep(req, sqe);
6286 __io_queue_sqe(req, cs);
6290 static inline void io_queue_link_head(struct io_kiocb *req,
6291 struct io_comp_state *cs)
6293 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6295 io_req_complete(req, -ECANCELED);
6297 io_queue_sqe(req, NULL, cs);
6300 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6301 struct io_kiocb **link, struct io_comp_state *cs)
6303 struct io_ring_ctx *ctx = req->ctx;
6307 * If we already have a head request, queue this one for async
6308 * submittal once the head completes. If we don't have a head but
6309 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6310 * submitted sync once the chain is complete. If none of those
6311 * conditions are true (normal request), then just queue it.
6314 struct io_kiocb *head = *link;
6317 * Taking sequential execution of a link, draining both sides
6318 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6319 * requests in the link. So, it drains the head and the
6320 * next after the link request. The last one is done via
6321 * drain_next flag to persist the effect across calls.
6323 if (req->flags & REQ_F_IO_DRAIN) {
6324 head->flags |= REQ_F_IO_DRAIN;
6325 ctx->drain_next = 1;
6327 ret = io_req_defer_prep(req, sqe);
6328 if (unlikely(ret)) {
6329 /* fail even hard links since we don't submit */
6330 head->flags |= REQ_F_FAIL_LINK;
6333 trace_io_uring_link(ctx, req, head);
6334 list_add_tail(&req->link_list, &head->link_list);
6336 /* last request of a link, enqueue the link */
6337 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6338 io_queue_link_head(head, cs);
6342 if (unlikely(ctx->drain_next)) {
6343 req->flags |= REQ_F_IO_DRAIN;
6344 ctx->drain_next = 0;
6346 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6347 req->flags |= REQ_F_LINK_HEAD;
6348 INIT_LIST_HEAD(&req->link_list);
6350 ret = io_req_defer_prep(req, sqe);
6352 req->flags |= REQ_F_FAIL_LINK;
6355 io_queue_sqe(req, sqe, cs);
6363 * Batched submission is done, ensure local IO is flushed out.
6365 static void io_submit_state_end(struct io_submit_state *state)
6367 if (!list_empty(&state->comp.list))
6368 io_submit_flush_completions(&state->comp);
6369 blk_finish_plug(&state->plug);
6370 io_state_file_put(state);
6371 if (state->free_reqs)
6372 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6376 * Start submission side cache.
6378 static void io_submit_state_start(struct io_submit_state *state,
6379 struct io_ring_ctx *ctx, unsigned int max_ios)
6381 blk_start_plug(&state->plug);
6383 INIT_LIST_HEAD(&state->comp.list);
6384 state->comp.ctx = ctx;
6385 state->free_reqs = 0;
6387 state->ios_left = max_ios;
6390 static void io_commit_sqring(struct io_ring_ctx *ctx)
6392 struct io_rings *rings = ctx->rings;
6395 * Ensure any loads from the SQEs are done at this point,
6396 * since once we write the new head, the application could
6397 * write new data to them.
6399 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6403 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6404 * that is mapped by userspace. This means that care needs to be taken to
6405 * ensure that reads are stable, as we cannot rely on userspace always
6406 * being a good citizen. If members of the sqe are validated and then later
6407 * used, it's important that those reads are done through READ_ONCE() to
6408 * prevent a re-load down the line.
6410 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6412 u32 *sq_array = ctx->sq_array;
6416 * The cached sq head (or cq tail) serves two purposes:
6418 * 1) allows us to batch the cost of updating the user visible
6420 * 2) allows the kernel side to track the head on its own, even
6421 * though the application is the one updating it.
6423 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6424 if (likely(head < ctx->sq_entries))
6425 return &ctx->sq_sqes[head];
6427 /* drop invalid entries */
6428 ctx->cached_sq_dropped++;
6429 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6433 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6435 ctx->cached_sq_head++;
6439 * Check SQE restrictions (opcode and flags).
6441 * Returns 'true' if SQE is allowed, 'false' otherwise.
6443 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6444 struct io_kiocb *req,
6445 unsigned int sqe_flags)
6447 if (!ctx->restricted)
6450 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6453 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6454 ctx->restrictions.sqe_flags_required)
6457 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6458 ctx->restrictions.sqe_flags_required))
6464 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6465 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6466 IOSQE_BUFFER_SELECT)
6468 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6469 const struct io_uring_sqe *sqe,
6470 struct io_submit_state *state)
6472 unsigned int sqe_flags;
6475 req->opcode = READ_ONCE(sqe->opcode);
6476 req->user_data = READ_ONCE(sqe->user_data);
6477 req->async_data = NULL;
6481 /* one is dropped after submission, the other at completion */
6482 refcount_set(&req->refs, 2);
6483 req->task = current;
6486 if (unlikely(req->opcode >= IORING_OP_LAST))
6489 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6492 sqe_flags = READ_ONCE(sqe->flags);
6493 /* enforce forwards compatibility on users */
6494 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6497 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6500 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6501 !io_op_defs[req->opcode].buffer_select)
6504 id = READ_ONCE(sqe->personality);
6506 struct io_identity *iod;
6508 io_req_init_async(req);
6509 iod = idr_find(&ctx->personality_idr, id);
6512 refcount_inc(&iod->count);
6513 io_put_identity(current->io_uring, req);
6514 get_cred(iod->creds);
6515 req->work.identity = iod;
6516 req->work.flags |= IO_WQ_WORK_CREDS;
6519 /* same numerical values with corresponding REQ_F_*, safe to copy */
6520 req->flags |= sqe_flags;
6522 if (!io_op_defs[req->opcode].needs_file)
6525 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6530 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6532 struct io_submit_state state;
6533 struct io_kiocb *link = NULL;
6534 int i, submitted = 0;
6536 /* if we have a backlog and couldn't flush it all, return BUSY */
6537 if (test_bit(0, &ctx->sq_check_overflow)) {
6538 if (!list_empty(&ctx->cq_overflow_list) &&
6539 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6543 /* make sure SQ entry isn't read before tail */
6544 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6546 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6549 percpu_counter_add(¤t->io_uring->inflight, nr);
6550 refcount_add(nr, ¤t->usage);
6552 io_submit_state_start(&state, ctx, nr);
6554 for (i = 0; i < nr; i++) {
6555 const struct io_uring_sqe *sqe;
6556 struct io_kiocb *req;
6559 sqe = io_get_sqe(ctx);
6560 if (unlikely(!sqe)) {
6561 io_consume_sqe(ctx);
6564 req = io_alloc_req(ctx, &state);
6565 if (unlikely(!req)) {
6567 submitted = -EAGAIN;
6570 io_consume_sqe(ctx);
6571 /* will complete beyond this point, count as submitted */
6574 err = io_init_req(ctx, req, sqe, &state);
6575 if (unlikely(err)) {
6578 io_req_complete(req, err);
6582 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6583 true, io_async_submit(ctx));
6584 err = io_submit_sqe(req, sqe, &link, &state.comp);
6589 if (unlikely(submitted != nr)) {
6590 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6591 struct io_uring_task *tctx = current->io_uring;
6592 int unused = nr - ref_used;
6594 percpu_ref_put_many(&ctx->refs, unused);
6595 percpu_counter_sub(&tctx->inflight, unused);
6596 put_task_struct_many(current, unused);
6599 io_queue_link_head(link, &state.comp);
6600 io_submit_state_end(&state);
6602 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6603 io_commit_sqring(ctx);
6608 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6610 /* Tell userspace we may need a wakeup call */
6611 spin_lock_irq(&ctx->completion_lock);
6612 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6613 spin_unlock_irq(&ctx->completion_lock);
6616 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6618 spin_lock_irq(&ctx->completion_lock);
6619 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6620 spin_unlock_irq(&ctx->completion_lock);
6623 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6624 int sync, void *key)
6626 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6629 ret = autoremove_wake_function(wqe, mode, sync, key);
6631 unsigned long flags;
6633 spin_lock_irqsave(&ctx->completion_lock, flags);
6634 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6635 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6646 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6647 unsigned long start_jiffies, bool cap_entries)
6649 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6650 struct io_sq_data *sqd = ctx->sq_data;
6651 unsigned int to_submit;
6655 if (!list_empty(&ctx->iopoll_list)) {
6656 unsigned nr_events = 0;
6658 mutex_lock(&ctx->uring_lock);
6659 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6660 io_do_iopoll(ctx, &nr_events, 0);
6661 mutex_unlock(&ctx->uring_lock);
6664 to_submit = io_sqring_entries(ctx);
6667 * If submit got -EBUSY, flag us as needing the application
6668 * to enter the kernel to reap and flush events.
6670 if (!to_submit || ret == -EBUSY || need_resched()) {
6672 * Drop cur_mm before scheduling, we can't hold it for
6673 * long periods (or over schedule()). Do this before
6674 * adding ourselves to the waitqueue, as the unuse/drop
6677 io_sq_thread_drop_mm();
6680 * We're polling. If we're within the defined idle
6681 * period, then let us spin without work before going
6682 * to sleep. The exception is if we got EBUSY doing
6683 * more IO, we should wait for the application to
6684 * reap events and wake us up.
6686 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6687 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6688 !percpu_ref_is_dying(&ctx->refs)))
6691 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6692 TASK_INTERRUPTIBLE);
6695 * While doing polled IO, before going to sleep, we need
6696 * to check if there are new reqs added to iopoll_list,
6697 * it is because reqs may have been punted to io worker
6698 * and will be added to iopoll_list later, hence check
6699 * the iopoll_list again.
6701 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6702 !list_empty_careful(&ctx->iopoll_list)) {
6703 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6707 to_submit = io_sqring_entries(ctx);
6708 if (!to_submit || ret == -EBUSY)
6712 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6713 io_ring_clear_wakeup_flag(ctx);
6715 /* if we're handling multiple rings, cap submit size for fairness */
6716 if (cap_entries && to_submit > 8)
6719 mutex_lock(&ctx->uring_lock);
6720 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6721 ret = io_submit_sqes(ctx, to_submit);
6722 mutex_unlock(&ctx->uring_lock);
6724 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6725 wake_up(&ctx->sqo_sq_wait);
6727 return SQT_DID_WORK;
6730 static void io_sqd_init_new(struct io_sq_data *sqd)
6732 struct io_ring_ctx *ctx;
6734 while (!list_empty(&sqd->ctx_new_list)) {
6735 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6736 init_wait(&ctx->sqo_wait_entry);
6737 ctx->sqo_wait_entry.func = io_sq_wake_function;
6738 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6739 complete(&ctx->sq_thread_comp);
6743 static int io_sq_thread(void *data)
6745 struct cgroup_subsys_state *cur_css = NULL;
6746 const struct cred *old_cred = NULL;
6747 struct io_sq_data *sqd = data;
6748 struct io_ring_ctx *ctx;
6749 unsigned long start_jiffies;
6751 start_jiffies = jiffies;
6752 while (!kthread_should_stop()) {
6753 enum sq_ret ret = 0;
6757 * Any changes to the sqd lists are synchronized through the
6758 * kthread parking. This synchronizes the thread vs users,
6759 * the users are synchronized on the sqd->ctx_lock.
6761 if (kthread_should_park())
6764 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6765 io_sqd_init_new(sqd);
6767 cap_entries = !list_is_singular(&sqd->ctx_list);
6769 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6770 if (current->cred != ctx->creds) {
6772 revert_creds(old_cred);
6773 old_cred = override_creds(ctx->creds);
6775 io_sq_thread_associate_blkcg(ctx, &cur_css);
6777 current->loginuid = ctx->loginuid;
6778 current->sessionid = ctx->sessionid;
6781 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6783 io_sq_thread_drop_mm();
6786 if (ret & SQT_SPIN) {
6789 } else if (ret == SQT_IDLE) {
6790 if (kthread_should_park())
6792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6793 io_ring_set_wakeup_flag(ctx);
6795 start_jiffies = jiffies;
6796 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6797 io_ring_clear_wakeup_flag(ctx);
6804 io_sq_thread_unassociate_blkcg();
6806 revert_creds(old_cred);
6813 struct io_wait_queue {
6814 struct wait_queue_entry wq;
6815 struct io_ring_ctx *ctx;
6817 unsigned nr_timeouts;
6820 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6822 struct io_ring_ctx *ctx = iowq->ctx;
6825 * Wake up if we have enough events, or if a timeout occurred since we
6826 * started waiting. For timeouts, we always want to return to userspace,
6827 * regardless of event count.
6829 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6830 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6833 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6834 int wake_flags, void *key)
6836 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6839 /* use noflush == true, as we can't safely rely on locking context */
6840 if (!io_should_wake(iowq, true))
6843 return autoremove_wake_function(curr, mode, wake_flags, key);
6846 static int io_run_task_work_sig(void)
6848 if (io_run_task_work())
6850 if (!signal_pending(current))
6852 if (current->jobctl & JOBCTL_TASK_WORK) {
6853 spin_lock_irq(¤t->sighand->siglock);
6854 current->jobctl &= ~JOBCTL_TASK_WORK;
6855 recalc_sigpending();
6856 spin_unlock_irq(¤t->sighand->siglock);
6863 * Wait until events become available, if we don't already have some. The
6864 * application must reap them itself, as they reside on the shared cq ring.
6866 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6867 const sigset_t __user *sig, size_t sigsz)
6869 struct io_wait_queue iowq = {
6872 .func = io_wake_function,
6873 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6876 .to_wait = min_events,
6878 struct io_rings *rings = ctx->rings;
6882 if (io_cqring_events(ctx, false) >= min_events)
6884 if (!io_run_task_work())
6889 #ifdef CONFIG_COMPAT
6890 if (in_compat_syscall())
6891 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6895 ret = set_user_sigmask(sig, sigsz);
6901 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6902 trace_io_uring_cqring_wait(ctx, min_events);
6904 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6905 TASK_INTERRUPTIBLE);
6906 /* make sure we run task_work before checking for signals */
6907 ret = io_run_task_work_sig();
6912 if (io_should_wake(&iowq, false))
6916 finish_wait(&ctx->wait, &iowq.wq);
6918 restore_saved_sigmask_unless(ret == -EINTR);
6920 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6923 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6925 #if defined(CONFIG_UNIX)
6926 if (ctx->ring_sock) {
6927 struct sock *sock = ctx->ring_sock->sk;
6928 struct sk_buff *skb;
6930 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6936 for (i = 0; i < ctx->nr_user_files; i++) {
6939 file = io_file_from_index(ctx, i);
6946 static void io_file_ref_kill(struct percpu_ref *ref)
6948 struct fixed_file_data *data;
6950 data = container_of(ref, struct fixed_file_data, refs);
6951 complete(&data->done);
6954 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6956 struct fixed_file_data *data = ctx->file_data;
6957 struct fixed_file_ref_node *ref_node = NULL;
6958 unsigned nr_tables, i;
6963 spin_lock(&data->lock);
6964 if (!list_empty(&data->ref_list))
6965 ref_node = list_first_entry(&data->ref_list,
6966 struct fixed_file_ref_node, node);
6967 spin_unlock(&data->lock);
6969 percpu_ref_kill(&ref_node->refs);
6971 percpu_ref_kill(&data->refs);
6973 /* wait for all refs nodes to complete */
6974 flush_delayed_work(&ctx->file_put_work);
6975 wait_for_completion(&data->done);
6977 __io_sqe_files_unregister(ctx);
6978 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6979 for (i = 0; i < nr_tables; i++)
6980 kfree(data->table[i].files);
6982 percpu_ref_exit(&data->refs);
6984 ctx->file_data = NULL;
6985 ctx->nr_user_files = 0;
6989 static void io_put_sq_data(struct io_sq_data *sqd)
6991 if (refcount_dec_and_test(&sqd->refs)) {
6993 * The park is a bit of a work-around, without it we get
6994 * warning spews on shutdown with SQPOLL set and affinity
6995 * set to a single CPU.
6998 kthread_park(sqd->thread);
6999 kthread_stop(sqd->thread);
7006 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7008 struct io_ring_ctx *ctx_attach;
7009 struct io_sq_data *sqd;
7012 f = fdget(p->wq_fd);
7014 return ERR_PTR(-ENXIO);
7015 if (f.file->f_op != &io_uring_fops) {
7017 return ERR_PTR(-EINVAL);
7020 ctx_attach = f.file->private_data;
7021 sqd = ctx_attach->sq_data;
7024 return ERR_PTR(-EINVAL);
7027 refcount_inc(&sqd->refs);
7032 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7034 struct io_sq_data *sqd;
7036 if (p->flags & IORING_SETUP_ATTACH_WQ)
7037 return io_attach_sq_data(p);
7039 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7041 return ERR_PTR(-ENOMEM);
7043 refcount_set(&sqd->refs, 1);
7044 INIT_LIST_HEAD(&sqd->ctx_list);
7045 INIT_LIST_HEAD(&sqd->ctx_new_list);
7046 mutex_init(&sqd->ctx_lock);
7047 mutex_init(&sqd->lock);
7048 init_waitqueue_head(&sqd->wait);
7052 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7053 __releases(&sqd->lock)
7057 kthread_unpark(sqd->thread);
7058 mutex_unlock(&sqd->lock);
7061 static void io_sq_thread_park(struct io_sq_data *sqd)
7062 __acquires(&sqd->lock)
7066 mutex_lock(&sqd->lock);
7067 kthread_park(sqd->thread);
7070 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7072 struct io_sq_data *sqd = ctx->sq_data;
7077 * We may arrive here from the error branch in
7078 * io_sq_offload_create() where the kthread is created
7079 * without being waked up, thus wake it up now to make
7080 * sure the wait will complete.
7082 wake_up_process(sqd->thread);
7083 wait_for_completion(&ctx->sq_thread_comp);
7085 io_sq_thread_park(sqd);
7088 mutex_lock(&sqd->ctx_lock);
7089 list_del(&ctx->sqd_list);
7090 mutex_unlock(&sqd->ctx_lock);
7093 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7094 io_sq_thread_unpark(sqd);
7097 io_put_sq_data(sqd);
7098 ctx->sq_data = NULL;
7102 static void io_finish_async(struct io_ring_ctx *ctx)
7104 io_sq_thread_stop(ctx);
7107 io_wq_destroy(ctx->io_wq);
7112 #if defined(CONFIG_UNIX)
7114 * Ensure the UNIX gc is aware of our file set, so we are certain that
7115 * the io_uring can be safely unregistered on process exit, even if we have
7116 * loops in the file referencing.
7118 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7120 struct sock *sk = ctx->ring_sock->sk;
7121 struct scm_fp_list *fpl;
7122 struct sk_buff *skb;
7125 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7129 skb = alloc_skb(0, GFP_KERNEL);
7138 fpl->user = get_uid(ctx->user);
7139 for (i = 0; i < nr; i++) {
7140 struct file *file = io_file_from_index(ctx, i + offset);
7144 fpl->fp[nr_files] = get_file(file);
7145 unix_inflight(fpl->user, fpl->fp[nr_files]);
7150 fpl->max = SCM_MAX_FD;
7151 fpl->count = nr_files;
7152 UNIXCB(skb).fp = fpl;
7153 skb->destructor = unix_destruct_scm;
7154 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7155 skb_queue_head(&sk->sk_receive_queue, skb);
7157 for (i = 0; i < nr_files; i++)
7168 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7169 * causes regular reference counting to break down. We rely on the UNIX
7170 * garbage collection to take care of this problem for us.
7172 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7174 unsigned left, total;
7178 left = ctx->nr_user_files;
7180 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7182 ret = __io_sqe_files_scm(ctx, this_files, total);
7186 total += this_files;
7192 while (total < ctx->nr_user_files) {
7193 struct file *file = io_file_from_index(ctx, total);
7203 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7209 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7210 unsigned nr_tables, unsigned nr_files)
7214 for (i = 0; i < nr_tables; i++) {
7215 struct fixed_file_table *table = &file_data->table[i];
7216 unsigned this_files;
7218 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7219 table->files = kcalloc(this_files, sizeof(struct file *),
7223 nr_files -= this_files;
7229 for (i = 0; i < nr_tables; i++) {
7230 struct fixed_file_table *table = &file_data->table[i];
7231 kfree(table->files);
7236 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7238 #if defined(CONFIG_UNIX)
7239 struct sock *sock = ctx->ring_sock->sk;
7240 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7241 struct sk_buff *skb;
7244 __skb_queue_head_init(&list);
7247 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7248 * remove this entry and rearrange the file array.
7250 skb = skb_dequeue(head);
7252 struct scm_fp_list *fp;
7254 fp = UNIXCB(skb).fp;
7255 for (i = 0; i < fp->count; i++) {
7258 if (fp->fp[i] != file)
7261 unix_notinflight(fp->user, fp->fp[i]);
7262 left = fp->count - 1 - i;
7264 memmove(&fp->fp[i], &fp->fp[i + 1],
7265 left * sizeof(struct file *));
7272 __skb_queue_tail(&list, skb);
7282 __skb_queue_tail(&list, skb);
7284 skb = skb_dequeue(head);
7287 if (skb_peek(&list)) {
7288 spin_lock_irq(&head->lock);
7289 while ((skb = __skb_dequeue(&list)) != NULL)
7290 __skb_queue_tail(head, skb);
7291 spin_unlock_irq(&head->lock);
7298 struct io_file_put {
7299 struct list_head list;
7303 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7305 struct fixed_file_data *file_data = ref_node->file_data;
7306 struct io_ring_ctx *ctx = file_data->ctx;
7307 struct io_file_put *pfile, *tmp;
7309 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7310 list_del(&pfile->list);
7311 io_ring_file_put(ctx, pfile->file);
7315 spin_lock(&file_data->lock);
7316 list_del(&ref_node->node);
7317 spin_unlock(&file_data->lock);
7319 percpu_ref_exit(&ref_node->refs);
7321 percpu_ref_put(&file_data->refs);
7324 static void io_file_put_work(struct work_struct *work)
7326 struct io_ring_ctx *ctx;
7327 struct llist_node *node;
7329 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7330 node = llist_del_all(&ctx->file_put_llist);
7333 struct fixed_file_ref_node *ref_node;
7334 struct llist_node *next = node->next;
7336 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7337 __io_file_put_work(ref_node);
7342 static void io_file_data_ref_zero(struct percpu_ref *ref)
7344 struct fixed_file_ref_node *ref_node;
7345 struct io_ring_ctx *ctx;
7349 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7350 ctx = ref_node->file_data->ctx;
7352 if (percpu_ref_is_dying(&ctx->file_data->refs))
7355 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7357 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7359 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7362 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7363 struct io_ring_ctx *ctx)
7365 struct fixed_file_ref_node *ref_node;
7367 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7369 return ERR_PTR(-ENOMEM);
7371 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7374 return ERR_PTR(-ENOMEM);
7376 INIT_LIST_HEAD(&ref_node->node);
7377 INIT_LIST_HEAD(&ref_node->file_list);
7378 ref_node->file_data = ctx->file_data;
7382 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7384 percpu_ref_exit(&ref_node->refs);
7388 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7391 __s32 __user *fds = (__s32 __user *) arg;
7392 unsigned nr_tables, i;
7394 int fd, ret = -ENOMEM;
7395 struct fixed_file_ref_node *ref_node;
7396 struct fixed_file_data *file_data;
7402 if (nr_args > IORING_MAX_FIXED_FILES)
7405 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7408 file_data->ctx = ctx;
7409 init_completion(&file_data->done);
7410 INIT_LIST_HEAD(&file_data->ref_list);
7411 spin_lock_init(&file_data->lock);
7413 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7414 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7416 if (!file_data->table)
7419 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7420 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7423 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7425 ctx->file_data = file_data;
7427 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7428 struct fixed_file_table *table;
7431 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7435 /* allow sparse sets */
7445 * Don't allow io_uring instances to be registered. If UNIX
7446 * isn't enabled, then this causes a reference cycle and this
7447 * instance can never get freed. If UNIX is enabled we'll
7448 * handle it just fine, but there's still no point in allowing
7449 * a ring fd as it doesn't support regular read/write anyway.
7451 if (file->f_op == &io_uring_fops) {
7455 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7456 index = i & IORING_FILE_TABLE_MASK;
7457 table->files[index] = file;
7460 ret = io_sqe_files_scm(ctx);
7462 io_sqe_files_unregister(ctx);
7466 ref_node = alloc_fixed_file_ref_node(ctx);
7467 if (IS_ERR(ref_node)) {
7468 io_sqe_files_unregister(ctx);
7469 return PTR_ERR(ref_node);
7472 file_data->node = ref_node;
7473 spin_lock(&file_data->lock);
7474 list_add(&ref_node->node, &file_data->ref_list);
7475 spin_unlock(&file_data->lock);
7476 percpu_ref_get(&file_data->refs);
7479 for (i = 0; i < ctx->nr_user_files; i++) {
7480 file = io_file_from_index(ctx, i);
7484 for (i = 0; i < nr_tables; i++)
7485 kfree(file_data->table[i].files);
7486 ctx->nr_user_files = 0;
7488 percpu_ref_exit(&file_data->refs);
7490 kfree(file_data->table);
7492 ctx->file_data = NULL;
7496 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7499 #if defined(CONFIG_UNIX)
7500 struct sock *sock = ctx->ring_sock->sk;
7501 struct sk_buff_head *head = &sock->sk_receive_queue;
7502 struct sk_buff *skb;
7505 * See if we can merge this file into an existing skb SCM_RIGHTS
7506 * file set. If there's no room, fall back to allocating a new skb
7507 * and filling it in.
7509 spin_lock_irq(&head->lock);
7510 skb = skb_peek(head);
7512 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7514 if (fpl->count < SCM_MAX_FD) {
7515 __skb_unlink(skb, head);
7516 spin_unlock_irq(&head->lock);
7517 fpl->fp[fpl->count] = get_file(file);
7518 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7520 spin_lock_irq(&head->lock);
7521 __skb_queue_head(head, skb);
7526 spin_unlock_irq(&head->lock);
7533 return __io_sqe_files_scm(ctx, 1, index);
7539 static int io_queue_file_removal(struct fixed_file_data *data,
7542 struct io_file_put *pfile;
7543 struct fixed_file_ref_node *ref_node = data->node;
7545 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7550 list_add(&pfile->list, &ref_node->file_list);
7555 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7556 struct io_uring_files_update *up,
7559 struct fixed_file_data *data = ctx->file_data;
7560 struct fixed_file_ref_node *ref_node;
7565 bool needs_switch = false;
7567 if (check_add_overflow(up->offset, nr_args, &done))
7569 if (done > ctx->nr_user_files)
7572 ref_node = alloc_fixed_file_ref_node(ctx);
7573 if (IS_ERR(ref_node))
7574 return PTR_ERR(ref_node);
7577 fds = u64_to_user_ptr(up->fds);
7579 struct fixed_file_table *table;
7583 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7587 i = array_index_nospec(up->offset, ctx->nr_user_files);
7588 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7589 index = i & IORING_FILE_TABLE_MASK;
7590 if (table->files[index]) {
7591 file = table->files[index];
7592 err = io_queue_file_removal(data, file);
7595 table->files[index] = NULL;
7596 needs_switch = true;
7605 * Don't allow io_uring instances to be registered. If
7606 * UNIX isn't enabled, then this causes a reference
7607 * cycle and this instance can never get freed. If UNIX
7608 * is enabled we'll handle it just fine, but there's
7609 * still no point in allowing a ring fd as it doesn't
7610 * support regular read/write anyway.
7612 if (file->f_op == &io_uring_fops) {
7617 table->files[index] = file;
7618 err = io_sqe_file_register(ctx, file, i);
7620 table->files[index] = NULL;
7631 percpu_ref_kill(&data->node->refs);
7632 spin_lock(&data->lock);
7633 list_add(&ref_node->node, &data->ref_list);
7634 data->node = ref_node;
7635 spin_unlock(&data->lock);
7636 percpu_ref_get(&ctx->file_data->refs);
7638 destroy_fixed_file_ref_node(ref_node);
7640 return done ? done : err;
7643 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7646 struct io_uring_files_update up;
7648 if (!ctx->file_data)
7652 if (copy_from_user(&up, arg, sizeof(up)))
7657 return __io_sqe_files_update(ctx, &up, nr_args);
7660 static void io_free_work(struct io_wq_work *work)
7662 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7664 /* Consider that io_steal_work() relies on this ref */
7668 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7669 struct io_uring_params *p)
7671 struct io_wq_data data;
7673 struct io_ring_ctx *ctx_attach;
7674 unsigned int concurrency;
7677 data.user = ctx->user;
7678 data.free_work = io_free_work;
7679 data.do_work = io_wq_submit_work;
7681 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7682 /* Do QD, or 4 * CPUS, whatever is smallest */
7683 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7685 ctx->io_wq = io_wq_create(concurrency, &data);
7686 if (IS_ERR(ctx->io_wq)) {
7687 ret = PTR_ERR(ctx->io_wq);
7693 f = fdget(p->wq_fd);
7697 if (f.file->f_op != &io_uring_fops) {
7702 ctx_attach = f.file->private_data;
7703 /* @io_wq is protected by holding the fd */
7704 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7709 ctx->io_wq = ctx_attach->io_wq;
7715 static int io_uring_alloc_task_context(struct task_struct *task)
7717 struct io_uring_task *tctx;
7720 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7721 if (unlikely(!tctx))
7724 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7725 if (unlikely(ret)) {
7731 init_waitqueue_head(&tctx->wait);
7734 io_init_identity(&tctx->__identity);
7735 tctx->identity = &tctx->__identity;
7736 task->io_uring = tctx;
7740 void __io_uring_free(struct task_struct *tsk)
7742 struct io_uring_task *tctx = tsk->io_uring;
7744 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7745 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7746 if (tctx->identity != &tctx->__identity)
7747 kfree(tctx->identity);
7748 percpu_counter_destroy(&tctx->inflight);
7750 tsk->io_uring = NULL;
7753 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7754 struct io_uring_params *p)
7758 if (ctx->flags & IORING_SETUP_SQPOLL) {
7759 struct io_sq_data *sqd;
7762 if (!capable(CAP_SYS_ADMIN))
7765 sqd = io_get_sq_data(p);
7772 io_sq_thread_park(sqd);
7773 mutex_lock(&sqd->ctx_lock);
7774 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7775 mutex_unlock(&sqd->ctx_lock);
7776 io_sq_thread_unpark(sqd);
7778 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7779 if (!ctx->sq_thread_idle)
7780 ctx->sq_thread_idle = HZ;
7785 if (p->flags & IORING_SETUP_SQ_AFF) {
7786 int cpu = p->sq_thread_cpu;
7789 if (cpu >= nr_cpu_ids)
7791 if (!cpu_online(cpu))
7794 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7795 cpu, "io_uring-sq");
7797 sqd->thread = kthread_create(io_sq_thread, sqd,
7800 if (IS_ERR(sqd->thread)) {
7801 ret = PTR_ERR(sqd->thread);
7805 ret = io_uring_alloc_task_context(sqd->thread);
7808 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7809 /* Can't have SQ_AFF without SQPOLL */
7815 ret = io_init_wq_offload(ctx, p);
7821 io_finish_async(ctx);
7825 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7827 struct io_sq_data *sqd = ctx->sq_data;
7829 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7830 wake_up_process(sqd->thread);
7833 static inline void __io_unaccount_mem(struct user_struct *user,
7834 unsigned long nr_pages)
7836 atomic_long_sub(nr_pages, &user->locked_vm);
7839 static inline int __io_account_mem(struct user_struct *user,
7840 unsigned long nr_pages)
7842 unsigned long page_limit, cur_pages, new_pages;
7844 /* Don't allow more pages than we can safely lock */
7845 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7848 cur_pages = atomic_long_read(&user->locked_vm);
7849 new_pages = cur_pages + nr_pages;
7850 if (new_pages > page_limit)
7852 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7853 new_pages) != cur_pages);
7858 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7859 enum io_mem_account acct)
7862 __io_unaccount_mem(ctx->user, nr_pages);
7864 if (ctx->mm_account) {
7865 if (acct == ACCT_LOCKED)
7866 ctx->mm_account->locked_vm -= nr_pages;
7867 else if (acct == ACCT_PINNED)
7868 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7872 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7873 enum io_mem_account acct)
7877 if (ctx->limit_mem) {
7878 ret = __io_account_mem(ctx->user, nr_pages);
7883 if (ctx->mm_account) {
7884 if (acct == ACCT_LOCKED)
7885 ctx->mm_account->locked_vm += nr_pages;
7886 else if (acct == ACCT_PINNED)
7887 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7893 static void io_mem_free(void *ptr)
7900 page = virt_to_head_page(ptr);
7901 if (put_page_testzero(page))
7902 free_compound_page(page);
7905 static void *io_mem_alloc(size_t size)
7907 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7910 return (void *) __get_free_pages(gfp_flags, get_order(size));
7913 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7916 struct io_rings *rings;
7917 size_t off, sq_array_size;
7919 off = struct_size(rings, cqes, cq_entries);
7920 if (off == SIZE_MAX)
7924 off = ALIGN(off, SMP_CACHE_BYTES);
7932 sq_array_size = array_size(sizeof(u32), sq_entries);
7933 if (sq_array_size == SIZE_MAX)
7936 if (check_add_overflow(off, sq_array_size, &off))
7942 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7946 pages = (size_t)1 << get_order(
7947 rings_size(sq_entries, cq_entries, NULL));
7948 pages += (size_t)1 << get_order(
7949 array_size(sizeof(struct io_uring_sqe), sq_entries));
7954 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7958 if (!ctx->user_bufs)
7961 for (i = 0; i < ctx->nr_user_bufs; i++) {
7962 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7964 for (j = 0; j < imu->nr_bvecs; j++)
7965 unpin_user_page(imu->bvec[j].bv_page);
7967 if (imu->acct_pages)
7968 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7973 kfree(ctx->user_bufs);
7974 ctx->user_bufs = NULL;
7975 ctx->nr_user_bufs = 0;
7979 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7980 void __user *arg, unsigned index)
7982 struct iovec __user *src;
7984 #ifdef CONFIG_COMPAT
7986 struct compat_iovec __user *ciovs;
7987 struct compat_iovec ciov;
7989 ciovs = (struct compat_iovec __user *) arg;
7990 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7993 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7994 dst->iov_len = ciov.iov_len;
7998 src = (struct iovec __user *) arg;
7999 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8005 * Not super efficient, but this is just a registration time. And we do cache
8006 * the last compound head, so generally we'll only do a full search if we don't
8009 * We check if the given compound head page has already been accounted, to
8010 * avoid double accounting it. This allows us to account the full size of the
8011 * page, not just the constituent pages of a huge page.
8013 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8014 int nr_pages, struct page *hpage)
8018 /* check current page array */
8019 for (i = 0; i < nr_pages; i++) {
8020 if (!PageCompound(pages[i]))
8022 if (compound_head(pages[i]) == hpage)
8026 /* check previously registered pages */
8027 for (i = 0; i < ctx->nr_user_bufs; i++) {
8028 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8030 for (j = 0; j < imu->nr_bvecs; j++) {
8031 if (!PageCompound(imu->bvec[j].bv_page))
8033 if (compound_head(imu->bvec[j].bv_page) == hpage)
8041 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8042 int nr_pages, struct io_mapped_ubuf *imu,
8043 struct page **last_hpage)
8047 for (i = 0; i < nr_pages; i++) {
8048 if (!PageCompound(pages[i])) {
8053 hpage = compound_head(pages[i]);
8054 if (hpage == *last_hpage)
8056 *last_hpage = hpage;
8057 if (headpage_already_acct(ctx, pages, i, hpage))
8059 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8063 if (!imu->acct_pages)
8066 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8068 imu->acct_pages = 0;
8072 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8075 struct vm_area_struct **vmas = NULL;
8076 struct page **pages = NULL;
8077 struct page *last_hpage = NULL;
8078 int i, j, got_pages = 0;
8083 if (!nr_args || nr_args > UIO_MAXIOV)
8086 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8088 if (!ctx->user_bufs)
8091 for (i = 0; i < nr_args; i++) {
8092 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8093 unsigned long off, start, end, ubuf;
8098 ret = io_copy_iov(ctx, &iov, arg, i);
8103 * Don't impose further limits on the size and buffer
8104 * constraints here, we'll -EINVAL later when IO is
8105 * submitted if they are wrong.
8108 if (!iov.iov_base || !iov.iov_len)
8111 /* arbitrary limit, but we need something */
8112 if (iov.iov_len > SZ_1G)
8115 ubuf = (unsigned long) iov.iov_base;
8116 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8117 start = ubuf >> PAGE_SHIFT;
8118 nr_pages = end - start;
8121 if (!pages || nr_pages > got_pages) {
8124 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8126 vmas = kvmalloc_array(nr_pages,
8127 sizeof(struct vm_area_struct *),
8129 if (!pages || !vmas) {
8133 got_pages = nr_pages;
8136 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8143 mmap_read_lock(current->mm);
8144 pret = pin_user_pages(ubuf, nr_pages,
8145 FOLL_WRITE | FOLL_LONGTERM,
8147 if (pret == nr_pages) {
8148 /* don't support file backed memory */
8149 for (j = 0; j < nr_pages; j++) {
8150 struct vm_area_struct *vma = vmas[j];
8153 !is_file_hugepages(vma->vm_file)) {
8159 ret = pret < 0 ? pret : -EFAULT;
8161 mmap_read_unlock(current->mm);
8164 * if we did partial map, or found file backed vmas,
8165 * release any pages we did get
8168 unpin_user_pages(pages, pret);
8173 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8175 unpin_user_pages(pages, pret);
8180 off = ubuf & ~PAGE_MASK;
8182 for (j = 0; j < nr_pages; j++) {
8185 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8186 imu->bvec[j].bv_page = pages[j];
8187 imu->bvec[j].bv_len = vec_len;
8188 imu->bvec[j].bv_offset = off;
8192 /* store original address for later verification */
8194 imu->len = iov.iov_len;
8195 imu->nr_bvecs = nr_pages;
8197 ctx->nr_user_bufs++;
8205 io_sqe_buffer_unregister(ctx);
8209 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8211 __s32 __user *fds = arg;
8217 if (copy_from_user(&fd, fds, sizeof(*fds)))
8220 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8221 if (IS_ERR(ctx->cq_ev_fd)) {
8222 int ret = PTR_ERR(ctx->cq_ev_fd);
8223 ctx->cq_ev_fd = NULL;
8230 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8232 if (ctx->cq_ev_fd) {
8233 eventfd_ctx_put(ctx->cq_ev_fd);
8234 ctx->cq_ev_fd = NULL;
8241 static int __io_destroy_buffers(int id, void *p, void *data)
8243 struct io_ring_ctx *ctx = data;
8244 struct io_buffer *buf = p;
8246 __io_remove_buffers(ctx, buf, id, -1U);
8250 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8252 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8253 idr_destroy(&ctx->io_buffer_idr);
8256 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8258 io_finish_async(ctx);
8259 io_sqe_buffer_unregister(ctx);
8261 if (ctx->sqo_task) {
8262 put_task_struct(ctx->sqo_task);
8263 ctx->sqo_task = NULL;
8264 mmdrop(ctx->mm_account);
8265 ctx->mm_account = NULL;
8268 #ifdef CONFIG_BLK_CGROUP
8269 if (ctx->sqo_blkcg_css)
8270 css_put(ctx->sqo_blkcg_css);
8273 io_sqe_files_unregister(ctx);
8274 io_eventfd_unregister(ctx);
8275 io_destroy_buffers(ctx);
8276 idr_destroy(&ctx->personality_idr);
8278 #if defined(CONFIG_UNIX)
8279 if (ctx->ring_sock) {
8280 ctx->ring_sock->file = NULL; /* so that iput() is called */
8281 sock_release(ctx->ring_sock);
8285 io_mem_free(ctx->rings);
8286 io_mem_free(ctx->sq_sqes);
8288 percpu_ref_exit(&ctx->refs);
8289 free_uid(ctx->user);
8290 put_cred(ctx->creds);
8291 kfree(ctx->cancel_hash);
8292 kmem_cache_free(req_cachep, ctx->fallback_req);
8296 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8298 struct io_ring_ctx *ctx = file->private_data;
8301 poll_wait(file, &ctx->cq_wait, wait);
8303 * synchronizes with barrier from wq_has_sleeper call in
8307 if (!io_sqring_full(ctx))
8308 mask |= EPOLLOUT | EPOLLWRNORM;
8309 if (io_cqring_events(ctx, false))
8310 mask |= EPOLLIN | EPOLLRDNORM;
8315 static int io_uring_fasync(int fd, struct file *file, int on)
8317 struct io_ring_ctx *ctx = file->private_data;
8319 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8322 static int io_remove_personalities(int id, void *p, void *data)
8324 struct io_ring_ctx *ctx = data;
8325 struct io_identity *iod;
8327 iod = idr_remove(&ctx->personality_idr, id);
8329 put_cred(iod->creds);
8330 if (refcount_dec_and_test(&iod->count))
8336 static void io_ring_exit_work(struct work_struct *work)
8338 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8342 * If we're doing polled IO and end up having requests being
8343 * submitted async (out-of-line), then completions can come in while
8344 * we're waiting for refs to drop. We need to reap these manually,
8345 * as nobody else will be looking for them.
8349 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8350 io_iopoll_try_reap_events(ctx);
8351 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8352 io_ring_ctx_free(ctx);
8355 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8357 mutex_lock(&ctx->uring_lock);
8358 percpu_ref_kill(&ctx->refs);
8359 mutex_unlock(&ctx->uring_lock);
8361 io_kill_timeouts(ctx, NULL);
8362 io_poll_remove_all(ctx, NULL);
8365 io_wq_cancel_all(ctx->io_wq);
8367 /* if we failed setting up the ctx, we might not have any rings */
8369 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8370 io_iopoll_try_reap_events(ctx);
8371 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8374 * Do this upfront, so we won't have a grace period where the ring
8375 * is closed but resources aren't reaped yet. This can cause
8376 * spurious failure in setting up a new ring.
8378 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8381 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8383 * Use system_unbound_wq to avoid spawning tons of event kworkers
8384 * if we're exiting a ton of rings at the same time. It just adds
8385 * noise and overhead, there's no discernable change in runtime
8386 * over using system_wq.
8388 queue_work(system_unbound_wq, &ctx->exit_work);
8391 static int io_uring_release(struct inode *inode, struct file *file)
8393 struct io_ring_ctx *ctx = file->private_data;
8395 file->private_data = NULL;
8396 io_ring_ctx_wait_and_kill(ctx);
8400 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8402 struct files_struct *files = data;
8404 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8405 work->identity->files == files);
8409 * Returns true if 'preq' is the link parent of 'req'
8411 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8413 struct io_kiocb *link;
8415 if (!(preq->flags & REQ_F_LINK_HEAD))
8418 list_for_each_entry(link, &preq->link_list, link_list) {
8426 static bool io_match_link_files(struct io_kiocb *req,
8427 struct files_struct *files)
8429 struct io_kiocb *link;
8431 if (io_match_files(req, files))
8433 if (req->flags & REQ_F_LINK_HEAD) {
8434 list_for_each_entry(link, &req->link_list, link_list) {
8435 if (io_match_files(link, files))
8443 * We're looking to cancel 'req' because it's holding on to our files, but
8444 * 'req' could be a link to another request. See if it is, and cancel that
8445 * parent request if so.
8447 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8449 struct hlist_node *tmp;
8450 struct io_kiocb *preq;
8454 spin_lock_irq(&ctx->completion_lock);
8455 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8456 struct hlist_head *list;
8458 list = &ctx->cancel_hash[i];
8459 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8460 found = io_match_link(preq, req);
8462 io_poll_remove_one(preq);
8467 spin_unlock_irq(&ctx->completion_lock);
8471 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8472 struct io_kiocb *req)
8474 struct io_kiocb *preq;
8477 spin_lock_irq(&ctx->completion_lock);
8478 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8479 found = io_match_link(preq, req);
8481 __io_timeout_cancel(preq);
8485 spin_unlock_irq(&ctx->completion_lock);
8489 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8491 return io_match_link(container_of(work, struct io_kiocb, work), data);
8494 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8496 enum io_wq_cancel cret;
8498 /* cancel this particular work, if it's running */
8499 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8500 if (cret != IO_WQ_CANCEL_NOTFOUND)
8503 /* find links that hold this pending, cancel those */
8504 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8505 if (cret != IO_WQ_CANCEL_NOTFOUND)
8508 /* if we have a poll link holding this pending, cancel that */
8509 if (io_poll_remove_link(ctx, req))
8512 /* final option, timeout link is holding this req pending */
8513 io_timeout_remove_link(ctx, req);
8516 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8517 struct files_struct *files)
8519 struct io_defer_entry *de = NULL;
8522 spin_lock_irq(&ctx->completion_lock);
8523 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8524 if (io_match_link_files(de->req, files)) {
8525 list_cut_position(&list, &ctx->defer_list, &de->list);
8529 spin_unlock_irq(&ctx->completion_lock);
8531 while (!list_empty(&list)) {
8532 de = list_first_entry(&list, struct io_defer_entry, list);
8533 list_del_init(&de->list);
8534 req_set_fail_links(de->req);
8535 io_put_req(de->req);
8536 io_req_complete(de->req, -ECANCELED);
8542 * Returns true if we found and killed one or more files pinning requests
8544 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8545 struct files_struct *files)
8547 if (list_empty_careful(&ctx->inflight_list))
8550 io_cancel_defer_files(ctx, files);
8551 /* cancel all at once, should be faster than doing it one by one*/
8552 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8554 while (!list_empty_careful(&ctx->inflight_list)) {
8555 struct io_kiocb *cancel_req = NULL, *req;
8558 spin_lock_irq(&ctx->inflight_lock);
8559 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8560 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8561 req->work.identity->files != files)
8563 /* req is being completed, ignore */
8564 if (!refcount_inc_not_zero(&req->refs))
8570 prepare_to_wait(&ctx->inflight_wait, &wait,
8571 TASK_UNINTERRUPTIBLE);
8572 spin_unlock_irq(&ctx->inflight_lock);
8574 /* We need to keep going until we don't find a matching req */
8577 /* cancel this request, or head link requests */
8578 io_attempt_cancel(ctx, cancel_req);
8579 io_put_req(cancel_req);
8580 /* cancellations _may_ trigger task work */
8583 finish_wait(&ctx->inflight_wait, &wait);
8589 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8591 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8592 struct task_struct *task = data;
8594 return io_task_match(req, task);
8597 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8598 struct task_struct *task,
8599 struct files_struct *files)
8603 ret = io_uring_cancel_files(ctx, files);
8605 enum io_wq_cancel cret;
8607 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8608 if (cret != IO_WQ_CANCEL_NOTFOUND)
8611 /* SQPOLL thread does its own polling */
8612 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8613 while (!list_empty_careful(&ctx->iopoll_list)) {
8614 io_iopoll_try_reap_events(ctx);
8619 ret |= io_poll_remove_all(ctx, task);
8620 ret |= io_kill_timeouts(ctx, task);
8627 * We need to iteratively cancel requests, in case a request has dependent
8628 * hard links. These persist even for failure of cancelations, hence keep
8629 * looping until none are found.
8631 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8632 struct files_struct *files)
8634 struct task_struct *task = current;
8636 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8637 task = ctx->sq_data->thread;
8639 io_cqring_overflow_flush(ctx, true, task, files);
8641 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8648 * Note that this task has used io_uring. We use it for cancelation purposes.
8650 static int io_uring_add_task_file(struct file *file)
8652 struct io_uring_task *tctx = current->io_uring;
8654 if (unlikely(!tctx)) {
8657 ret = io_uring_alloc_task_context(current);
8660 tctx = current->io_uring;
8662 if (tctx->last != file) {
8663 void *old = xa_load(&tctx->xa, (unsigned long)file);
8667 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8676 * Remove this io_uring_file -> task mapping.
8678 static void io_uring_del_task_file(struct file *file)
8680 struct io_uring_task *tctx = current->io_uring;
8682 if (tctx->last == file)
8684 file = xa_erase(&tctx->xa, (unsigned long)file);
8689 static void __io_uring_attempt_task_drop(struct file *file)
8691 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8694 io_uring_del_task_file(file);
8698 * Drop task note for this file if we're the only ones that hold it after
8701 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8703 if (!current->io_uring)
8706 * fput() is pending, will be 2 if the only other ref is our potential
8707 * task file note. If the task is exiting, drop regardless of count.
8709 if (!exiting && atomic_long_read(&file->f_count) != 2)
8712 __io_uring_attempt_task_drop(file);
8715 void __io_uring_files_cancel(struct files_struct *files)
8717 struct io_uring_task *tctx = current->io_uring;
8719 unsigned long index;
8721 /* make sure overflow events are dropped */
8722 tctx->in_idle = true;
8724 xa_for_each(&tctx->xa, index, file) {
8725 struct io_ring_ctx *ctx = file->private_data;
8727 io_uring_cancel_task_requests(ctx, files);
8729 io_uring_del_task_file(file);
8734 * Find any io_uring fd that this task has registered or done IO on, and cancel
8737 void __io_uring_task_cancel(void)
8739 struct io_uring_task *tctx = current->io_uring;
8743 /* make sure overflow events are dropped */
8744 tctx->in_idle = true;
8747 /* read completions before cancelations */
8748 inflight = percpu_counter_sum(&tctx->inflight);
8751 __io_uring_files_cancel(NULL);
8753 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8756 * If we've seen completions, retry. This avoids a race where
8757 * a completion comes in before we did prepare_to_wait().
8759 if (inflight != percpu_counter_sum(&tctx->inflight))
8764 finish_wait(&tctx->wait, &wait);
8765 tctx->in_idle = false;
8768 static int io_uring_flush(struct file *file, void *data)
8770 struct io_ring_ctx *ctx = file->private_data;
8773 * If the task is going away, cancel work it may have pending
8775 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8778 io_uring_cancel_task_requests(ctx, data);
8779 io_uring_attempt_task_drop(file, !data);
8783 static void *io_uring_validate_mmap_request(struct file *file,
8784 loff_t pgoff, size_t sz)
8786 struct io_ring_ctx *ctx = file->private_data;
8787 loff_t offset = pgoff << PAGE_SHIFT;
8792 case IORING_OFF_SQ_RING:
8793 case IORING_OFF_CQ_RING:
8796 case IORING_OFF_SQES:
8800 return ERR_PTR(-EINVAL);
8803 page = virt_to_head_page(ptr);
8804 if (sz > page_size(page))
8805 return ERR_PTR(-EINVAL);
8812 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8814 size_t sz = vma->vm_end - vma->vm_start;
8818 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8820 return PTR_ERR(ptr);
8822 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8823 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8826 #else /* !CONFIG_MMU */
8828 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8830 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8833 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8835 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8838 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8839 unsigned long addr, unsigned long len,
8840 unsigned long pgoff, unsigned long flags)
8844 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8846 return PTR_ERR(ptr);
8848 return (unsigned long) ptr;
8851 #endif /* !CONFIG_MMU */
8853 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8858 if (!io_sqring_full(ctx))
8861 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8863 if (!io_sqring_full(ctx))
8867 } while (!signal_pending(current));
8869 finish_wait(&ctx->sqo_sq_wait, &wait);
8872 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8873 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8876 struct io_ring_ctx *ctx;
8883 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8884 IORING_ENTER_SQ_WAIT))
8892 if (f.file->f_op != &io_uring_fops)
8896 ctx = f.file->private_data;
8897 if (!percpu_ref_tryget(&ctx->refs))
8901 if (ctx->flags & IORING_SETUP_R_DISABLED)
8905 * For SQ polling, the thread will do all submissions and completions.
8906 * Just return the requested submit count, and wake the thread if
8910 if (ctx->flags & IORING_SETUP_SQPOLL) {
8911 if (!list_empty_careful(&ctx->cq_overflow_list))
8912 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8913 if (flags & IORING_ENTER_SQ_WAKEUP)
8914 wake_up(&ctx->sq_data->wait);
8915 if (flags & IORING_ENTER_SQ_WAIT)
8916 io_sqpoll_wait_sq(ctx);
8917 submitted = to_submit;
8918 } else if (to_submit) {
8919 ret = io_uring_add_task_file(f.file);
8922 mutex_lock(&ctx->uring_lock);
8923 submitted = io_submit_sqes(ctx, to_submit);
8924 mutex_unlock(&ctx->uring_lock);
8926 if (submitted != to_submit)
8929 if (flags & IORING_ENTER_GETEVENTS) {
8930 min_complete = min(min_complete, ctx->cq_entries);
8933 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8934 * space applications don't need to do io completion events
8935 * polling again, they can rely on io_sq_thread to do polling
8936 * work, which can reduce cpu usage and uring_lock contention.
8938 if (ctx->flags & IORING_SETUP_IOPOLL &&
8939 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8940 ret = io_iopoll_check(ctx, min_complete);
8942 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8947 percpu_ref_put(&ctx->refs);
8950 return submitted ? submitted : ret;
8953 #ifdef CONFIG_PROC_FS
8954 static int io_uring_show_cred(int id, void *p, void *data)
8956 const struct cred *cred = p;
8957 struct seq_file *m = data;
8958 struct user_namespace *uns = seq_user_ns(m);
8959 struct group_info *gi;
8964 seq_printf(m, "%5d\n", id);
8965 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8966 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8967 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8968 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8969 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8970 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8971 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8972 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8973 seq_puts(m, "\n\tGroups:\t");
8974 gi = cred->group_info;
8975 for (g = 0; g < gi->ngroups; g++) {
8976 seq_put_decimal_ull(m, g ? " " : "",
8977 from_kgid_munged(uns, gi->gid[g]));
8979 seq_puts(m, "\n\tCapEff:\t");
8980 cap = cred->cap_effective;
8981 CAP_FOR_EACH_U32(__capi)
8982 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8987 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8989 struct io_sq_data *sq = NULL;
8994 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8995 * since fdinfo case grabs it in the opposite direction of normal use
8996 * cases. If we fail to get the lock, we just don't iterate any
8997 * structures that could be going away outside the io_uring mutex.
8999 has_lock = mutex_trylock(&ctx->uring_lock);
9001 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9004 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9005 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9006 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9007 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9008 struct fixed_file_table *table;
9011 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9012 f = table->files[i & IORING_FILE_TABLE_MASK];
9014 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9016 seq_printf(m, "%5u: <none>\n", i);
9018 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9019 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9020 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9022 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9023 (unsigned int) buf->len);
9025 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9026 seq_printf(m, "Personalities:\n");
9027 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9029 seq_printf(m, "PollList:\n");
9030 spin_lock_irq(&ctx->completion_lock);
9031 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9032 struct hlist_head *list = &ctx->cancel_hash[i];
9033 struct io_kiocb *req;
9035 hlist_for_each_entry(req, list, hash_node)
9036 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9037 req->task->task_works != NULL);
9039 spin_unlock_irq(&ctx->completion_lock);
9041 mutex_unlock(&ctx->uring_lock);
9044 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9046 struct io_ring_ctx *ctx = f->private_data;
9048 if (percpu_ref_tryget(&ctx->refs)) {
9049 __io_uring_show_fdinfo(ctx, m);
9050 percpu_ref_put(&ctx->refs);
9055 static const struct file_operations io_uring_fops = {
9056 .release = io_uring_release,
9057 .flush = io_uring_flush,
9058 .mmap = io_uring_mmap,
9060 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9061 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9063 .poll = io_uring_poll,
9064 .fasync = io_uring_fasync,
9065 #ifdef CONFIG_PROC_FS
9066 .show_fdinfo = io_uring_show_fdinfo,
9070 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9071 struct io_uring_params *p)
9073 struct io_rings *rings;
9074 size_t size, sq_array_offset;
9076 /* make sure these are sane, as we already accounted them */
9077 ctx->sq_entries = p->sq_entries;
9078 ctx->cq_entries = p->cq_entries;
9080 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9081 if (size == SIZE_MAX)
9084 rings = io_mem_alloc(size);
9089 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9090 rings->sq_ring_mask = p->sq_entries - 1;
9091 rings->cq_ring_mask = p->cq_entries - 1;
9092 rings->sq_ring_entries = p->sq_entries;
9093 rings->cq_ring_entries = p->cq_entries;
9094 ctx->sq_mask = rings->sq_ring_mask;
9095 ctx->cq_mask = rings->cq_ring_mask;
9097 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9098 if (size == SIZE_MAX) {
9099 io_mem_free(ctx->rings);
9104 ctx->sq_sqes = io_mem_alloc(size);
9105 if (!ctx->sq_sqes) {
9106 io_mem_free(ctx->rings);
9115 * Allocate an anonymous fd, this is what constitutes the application
9116 * visible backing of an io_uring instance. The application mmaps this
9117 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9118 * we have to tie this fd to a socket for file garbage collection purposes.
9120 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9125 #if defined(CONFIG_UNIX)
9126 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9132 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9136 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9137 O_RDWR | O_CLOEXEC);
9141 ret = PTR_ERR(file);
9145 #if defined(CONFIG_UNIX)
9146 ctx->ring_sock->file = file;
9148 if (unlikely(io_uring_add_task_file(file))) {
9149 file = ERR_PTR(-ENOMEM);
9152 fd_install(ret, file);
9155 #if defined(CONFIG_UNIX)
9156 sock_release(ctx->ring_sock);
9157 ctx->ring_sock = NULL;
9162 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9163 struct io_uring_params __user *params)
9165 struct user_struct *user = NULL;
9166 struct io_ring_ctx *ctx;
9172 if (entries > IORING_MAX_ENTRIES) {
9173 if (!(p->flags & IORING_SETUP_CLAMP))
9175 entries = IORING_MAX_ENTRIES;
9179 * Use twice as many entries for the CQ ring. It's possible for the
9180 * application to drive a higher depth than the size of the SQ ring,
9181 * since the sqes are only used at submission time. This allows for
9182 * some flexibility in overcommitting a bit. If the application has
9183 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9184 * of CQ ring entries manually.
9186 p->sq_entries = roundup_pow_of_two(entries);
9187 if (p->flags & IORING_SETUP_CQSIZE) {
9189 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9190 * to a power-of-two, if it isn't already. We do NOT impose
9191 * any cq vs sq ring sizing.
9193 if (p->cq_entries < p->sq_entries)
9195 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9196 if (!(p->flags & IORING_SETUP_CLAMP))
9198 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9200 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9202 p->cq_entries = 2 * p->sq_entries;
9205 user = get_uid(current_user());
9206 limit_mem = !capable(CAP_IPC_LOCK);
9209 ret = __io_account_mem(user,
9210 ring_pages(p->sq_entries, p->cq_entries));
9217 ctx = io_ring_ctx_alloc(p);
9220 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9225 ctx->compat = in_compat_syscall();
9227 ctx->creds = get_current_cred();
9229 ctx->loginuid = current->loginuid;
9230 ctx->sessionid = current->sessionid;
9232 ctx->sqo_task = get_task_struct(current);
9235 * This is just grabbed for accounting purposes. When a process exits,
9236 * the mm is exited and dropped before the files, hence we need to hang
9237 * on to this mm purely for the purposes of being able to unaccount
9238 * memory (locked/pinned vm). It's not used for anything else.
9240 mmgrab(current->mm);
9241 ctx->mm_account = current->mm;
9243 #ifdef CONFIG_BLK_CGROUP
9245 * The sq thread will belong to the original cgroup it was inited in.
9246 * If the cgroup goes offline (e.g. disabling the io controller), then
9247 * issued bios will be associated with the closest cgroup later in the
9251 ctx->sqo_blkcg_css = blkcg_css();
9252 ret = css_tryget_online(ctx->sqo_blkcg_css);
9255 /* don't init against a dying cgroup, have the user try again */
9256 ctx->sqo_blkcg_css = NULL;
9263 * Account memory _before_ installing the file descriptor. Once
9264 * the descriptor is installed, it can get closed at any time. Also
9265 * do this before hitting the general error path, as ring freeing
9266 * will un-account as well.
9268 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9270 ctx->limit_mem = limit_mem;
9272 ret = io_allocate_scq_urings(ctx, p);
9276 ret = io_sq_offload_create(ctx, p);
9280 if (!(p->flags & IORING_SETUP_R_DISABLED))
9281 io_sq_offload_start(ctx);
9283 memset(&p->sq_off, 0, sizeof(p->sq_off));
9284 p->sq_off.head = offsetof(struct io_rings, sq.head);
9285 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9286 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9287 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9288 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9289 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9290 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9292 memset(&p->cq_off, 0, sizeof(p->cq_off));
9293 p->cq_off.head = offsetof(struct io_rings, cq.head);
9294 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9295 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9296 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9297 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9298 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9299 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9301 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9302 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9303 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9304 IORING_FEAT_POLL_32BITS;
9306 if (copy_to_user(params, p, sizeof(*p))) {
9312 * Install ring fd as the very last thing, so we don't risk someone
9313 * having closed it before we finish setup
9315 ret = io_uring_get_fd(ctx);
9319 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9322 io_ring_ctx_wait_and_kill(ctx);
9327 * Sets up an aio uring context, and returns the fd. Applications asks for a
9328 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9329 * params structure passed in.
9331 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9333 struct io_uring_params p;
9336 if (copy_from_user(&p, params, sizeof(p)))
9338 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9343 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9344 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9345 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9346 IORING_SETUP_R_DISABLED))
9349 return io_uring_create(entries, &p, params);
9352 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9353 struct io_uring_params __user *, params)
9355 return io_uring_setup(entries, params);
9358 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9360 struct io_uring_probe *p;
9364 size = struct_size(p, ops, nr_args);
9365 if (size == SIZE_MAX)
9367 p = kzalloc(size, GFP_KERNEL);
9372 if (copy_from_user(p, arg, size))
9375 if (memchr_inv(p, 0, size))
9378 p->last_op = IORING_OP_LAST - 1;
9379 if (nr_args > IORING_OP_LAST)
9380 nr_args = IORING_OP_LAST;
9382 for (i = 0; i < nr_args; i++) {
9384 if (!io_op_defs[i].not_supported)
9385 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9390 if (copy_to_user(arg, p, size))
9397 static int io_register_personality(struct io_ring_ctx *ctx)
9399 struct io_identity *id;
9402 id = kmalloc(sizeof(*id), GFP_KERNEL);
9406 io_init_identity(id);
9407 id->creds = get_current_cred();
9409 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9411 put_cred(id->creds);
9417 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9419 struct io_identity *iod;
9421 iod = idr_remove(&ctx->personality_idr, id);
9423 put_cred(iod->creds);
9424 if (refcount_dec_and_test(&iod->count))
9432 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9433 unsigned int nr_args)
9435 struct io_uring_restriction *res;
9439 /* Restrictions allowed only if rings started disabled */
9440 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9443 /* We allow only a single restrictions registration */
9444 if (ctx->restrictions.registered)
9447 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9450 size = array_size(nr_args, sizeof(*res));
9451 if (size == SIZE_MAX)
9454 res = memdup_user(arg, size);
9456 return PTR_ERR(res);
9460 for (i = 0; i < nr_args; i++) {
9461 switch (res[i].opcode) {
9462 case IORING_RESTRICTION_REGISTER_OP:
9463 if (res[i].register_op >= IORING_REGISTER_LAST) {
9468 __set_bit(res[i].register_op,
9469 ctx->restrictions.register_op);
9471 case IORING_RESTRICTION_SQE_OP:
9472 if (res[i].sqe_op >= IORING_OP_LAST) {
9477 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9479 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9480 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9482 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9483 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9492 /* Reset all restrictions if an error happened */
9494 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9496 ctx->restrictions.registered = true;
9502 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9504 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9507 if (ctx->restrictions.registered)
9508 ctx->restricted = 1;
9510 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9512 io_sq_offload_start(ctx);
9517 static bool io_register_op_must_quiesce(int op)
9520 case IORING_UNREGISTER_FILES:
9521 case IORING_REGISTER_FILES_UPDATE:
9522 case IORING_REGISTER_PROBE:
9523 case IORING_REGISTER_PERSONALITY:
9524 case IORING_UNREGISTER_PERSONALITY:
9531 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9532 void __user *arg, unsigned nr_args)
9533 __releases(ctx->uring_lock)
9534 __acquires(ctx->uring_lock)
9539 * We're inside the ring mutex, if the ref is already dying, then
9540 * someone else killed the ctx or is already going through
9541 * io_uring_register().
9543 if (percpu_ref_is_dying(&ctx->refs))
9546 if (io_register_op_must_quiesce(opcode)) {
9547 percpu_ref_kill(&ctx->refs);
9550 * Drop uring mutex before waiting for references to exit. If
9551 * another thread is currently inside io_uring_enter() it might
9552 * need to grab the uring_lock to make progress. If we hold it
9553 * here across the drain wait, then we can deadlock. It's safe
9554 * to drop the mutex here, since no new references will come in
9555 * after we've killed the percpu ref.
9557 mutex_unlock(&ctx->uring_lock);
9559 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9562 ret = io_run_task_work_sig();
9567 mutex_lock(&ctx->uring_lock);
9570 percpu_ref_resurrect(&ctx->refs);
9575 if (ctx->restricted) {
9576 if (opcode >= IORING_REGISTER_LAST) {
9581 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9588 case IORING_REGISTER_BUFFERS:
9589 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9591 case IORING_UNREGISTER_BUFFERS:
9595 ret = io_sqe_buffer_unregister(ctx);
9597 case IORING_REGISTER_FILES:
9598 ret = io_sqe_files_register(ctx, arg, nr_args);
9600 case IORING_UNREGISTER_FILES:
9604 ret = io_sqe_files_unregister(ctx);
9606 case IORING_REGISTER_FILES_UPDATE:
9607 ret = io_sqe_files_update(ctx, arg, nr_args);
9609 case IORING_REGISTER_EVENTFD:
9610 case IORING_REGISTER_EVENTFD_ASYNC:
9614 ret = io_eventfd_register(ctx, arg);
9617 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9618 ctx->eventfd_async = 1;
9620 ctx->eventfd_async = 0;
9622 case IORING_UNREGISTER_EVENTFD:
9626 ret = io_eventfd_unregister(ctx);
9628 case IORING_REGISTER_PROBE:
9630 if (!arg || nr_args > 256)
9632 ret = io_probe(ctx, arg, nr_args);
9634 case IORING_REGISTER_PERSONALITY:
9638 ret = io_register_personality(ctx);
9640 case IORING_UNREGISTER_PERSONALITY:
9644 ret = io_unregister_personality(ctx, nr_args);
9646 case IORING_REGISTER_ENABLE_RINGS:
9650 ret = io_register_enable_rings(ctx);
9652 case IORING_REGISTER_RESTRICTIONS:
9653 ret = io_register_restrictions(ctx, arg, nr_args);
9661 if (io_register_op_must_quiesce(opcode)) {
9662 /* bring the ctx back to life */
9663 percpu_ref_reinit(&ctx->refs);
9665 reinit_completion(&ctx->ref_comp);
9670 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9671 void __user *, arg, unsigned int, nr_args)
9673 struct io_ring_ctx *ctx;
9682 if (f.file->f_op != &io_uring_fops)
9685 ctx = f.file->private_data;
9687 mutex_lock(&ctx->uring_lock);
9688 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9689 mutex_unlock(&ctx->uring_lock);
9690 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9691 ctx->cq_ev_fd != NULL, ret);
9697 static int __init io_uring_init(void)
9699 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9700 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9701 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9704 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9705 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9706 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9707 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9708 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9709 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9710 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9711 BUILD_BUG_SQE_ELEM(8, __u64, off);
9712 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9713 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9714 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9715 BUILD_BUG_SQE_ELEM(24, __u32, len);
9716 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9717 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9718 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9719 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9720 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9721 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9722 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9723 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9724 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9725 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9726 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9727 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9728 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9729 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9730 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9731 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9732 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9733 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9734 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9736 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9737 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9738 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9741 __initcall(io_uring_init);