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);
1073 static inline void __io_req_init_async(struct io_kiocb *req)
1075 memset(&req->work, 0, sizeof(req->work));
1076 req->flags |= REQ_F_WORK_INITIALIZED;
1080 * Note: must call io_req_init_async() for the first time you
1081 * touch any members of io_wq_work.
1083 static inline void io_req_init_async(struct io_kiocb *req)
1085 struct io_uring_task *tctx = current->io_uring;
1087 if (req->flags & REQ_F_WORK_INITIALIZED)
1090 __io_req_init_async(req);
1092 /* Grab a ref if this isn't our static identity */
1093 req->work.identity = tctx->identity;
1094 if (tctx->identity != &tctx->__identity)
1095 refcount_inc(&req->work.identity->count);
1098 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1100 return ctx->flags & IORING_SETUP_SQPOLL;
1103 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1105 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1107 complete(&ctx->ref_comp);
1110 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1112 return !req->timeout.off;
1115 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1117 struct io_ring_ctx *ctx;
1120 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1124 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1125 if (!ctx->fallback_req)
1129 * Use 5 bits less than the max cq entries, that should give us around
1130 * 32 entries per hash list if totally full and uniformly spread.
1132 hash_bits = ilog2(p->cq_entries);
1136 ctx->cancel_hash_bits = hash_bits;
1137 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1139 if (!ctx->cancel_hash)
1141 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1143 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1144 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1147 ctx->flags = p->flags;
1148 init_waitqueue_head(&ctx->sqo_sq_wait);
1149 INIT_LIST_HEAD(&ctx->sqd_list);
1150 init_waitqueue_head(&ctx->cq_wait);
1151 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1152 init_completion(&ctx->ref_comp);
1153 init_completion(&ctx->sq_thread_comp);
1154 idr_init(&ctx->io_buffer_idr);
1155 idr_init(&ctx->personality_idr);
1156 mutex_init(&ctx->uring_lock);
1157 init_waitqueue_head(&ctx->wait);
1158 spin_lock_init(&ctx->completion_lock);
1159 INIT_LIST_HEAD(&ctx->iopoll_list);
1160 INIT_LIST_HEAD(&ctx->defer_list);
1161 INIT_LIST_HEAD(&ctx->timeout_list);
1162 init_waitqueue_head(&ctx->inflight_wait);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1166 init_llist_head(&ctx->file_put_llist);
1169 if (ctx->fallback_req)
1170 kmem_cache_free(req_cachep, ctx->fallback_req);
1171 kfree(ctx->cancel_hash);
1176 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1178 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1179 struct io_ring_ctx *ctx = req->ctx;
1181 return seq != ctx->cached_cq_tail
1182 + atomic_read(&ctx->cached_cq_overflow);
1188 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1190 struct io_rings *rings = ctx->rings;
1192 /* order cqe stores with ring update */
1193 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1195 if (wq_has_sleeper(&ctx->cq_wait)) {
1196 wake_up_interruptible(&ctx->cq_wait);
1197 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1201 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1203 if (req->work.identity == &tctx->__identity)
1205 if (refcount_dec_and_test(&req->work.identity->count))
1206 kfree(req->work.identity);
1209 static void io_req_clean_work(struct io_kiocb *req)
1211 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1214 req->flags &= ~REQ_F_WORK_INITIALIZED;
1216 if (req->work.flags & IO_WQ_WORK_MM) {
1217 mmdrop(req->work.identity->mm);
1218 req->work.flags &= ~IO_WQ_WORK_MM;
1220 #ifdef CONFIG_BLK_CGROUP
1221 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1222 css_put(req->work.identity->blkcg_css);
1223 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1226 if (req->work.flags & IO_WQ_WORK_CREDS) {
1227 put_cred(req->work.identity->creds);
1228 req->work.flags &= ~IO_WQ_WORK_CREDS;
1230 if (req->work.flags & IO_WQ_WORK_FS) {
1231 struct fs_struct *fs = req->work.identity->fs;
1233 spin_lock(&req->work.identity->fs->lock);
1236 spin_unlock(&req->work.identity->fs->lock);
1239 req->work.flags &= ~IO_WQ_WORK_FS;
1242 io_put_identity(req->task->io_uring, req);
1246 * Create a private copy of io_identity, since some fields don't match
1247 * the current context.
1249 static bool io_identity_cow(struct io_kiocb *req)
1251 struct io_uring_task *tctx = current->io_uring;
1252 const struct cred *creds = NULL;
1253 struct io_identity *id;
1255 if (req->work.flags & IO_WQ_WORK_CREDS)
1256 creds = req->work.identity->creds;
1258 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1259 if (unlikely(!id)) {
1260 req->work.flags |= IO_WQ_WORK_CANCEL;
1265 * We can safely just re-init the creds we copied Either the field
1266 * matches the current one, or we haven't grabbed it yet. The only
1267 * exception is ->creds, through registered personalities, so handle
1268 * that one separately.
1270 io_init_identity(id);
1272 req->work.identity->creds = creds;
1274 /* add one for this request */
1275 refcount_inc(&id->count);
1277 /* drop old identity, assign new one. one ref for req, one for tctx */
1278 if (req->work.identity != tctx->identity &&
1279 refcount_sub_and_test(2, &req->work.identity->count))
1280 kfree(req->work.identity);
1282 req->work.identity = id;
1283 tctx->identity = id;
1287 static bool io_grab_identity(struct io_kiocb *req)
1289 const struct io_op_def *def = &io_op_defs[req->opcode];
1290 struct io_identity *id = req->work.identity;
1291 struct io_ring_ctx *ctx = req->ctx;
1293 if (def->needs_fsize && id->fsize != rlimit(RLIMIT_FSIZE))
1296 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1297 (def->work_flags & IO_WQ_WORK_FILES) &&
1298 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1299 if (id->files != current->files ||
1300 id->nsproxy != current->nsproxy)
1302 atomic_inc(&id->files->count);
1303 get_nsproxy(id->nsproxy);
1304 req->flags |= REQ_F_INFLIGHT;
1306 spin_lock_irq(&ctx->inflight_lock);
1307 list_add(&req->inflight_entry, &ctx->inflight_list);
1308 spin_unlock_irq(&ctx->inflight_lock);
1309 req->work.flags |= IO_WQ_WORK_FILES;
1311 #ifdef CONFIG_BLK_CGROUP
1312 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1313 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1315 if (id->blkcg_css != blkcg_css()) {
1320 * This should be rare, either the cgroup is dying or the task
1321 * is moving cgroups. Just punt to root for the handful of ios.
1323 if (css_tryget_online(id->blkcg_css))
1324 req->work.flags |= IO_WQ_WORK_BLKCG;
1328 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1329 if (id->creds != current_cred())
1331 get_cred(id->creds);
1332 req->work.flags |= IO_WQ_WORK_CREDS;
1335 if (!uid_eq(current->loginuid, id->loginuid) ||
1336 current->sessionid != id->sessionid)
1339 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1340 (def->work_flags & IO_WQ_WORK_FS)) {
1341 if (current->fs != id->fs)
1343 spin_lock(&id->fs->lock);
1344 if (!id->fs->in_exec) {
1346 req->work.flags |= IO_WQ_WORK_FS;
1348 req->work.flags |= IO_WQ_WORK_CANCEL;
1350 spin_unlock(¤t->fs->lock);
1356 static void io_prep_async_work(struct io_kiocb *req)
1358 const struct io_op_def *def = &io_op_defs[req->opcode];
1359 struct io_ring_ctx *ctx = req->ctx;
1360 struct io_identity *id;
1362 io_req_init_async(req);
1363 id = req->work.identity;
1365 if (req->flags & REQ_F_ISREG) {
1366 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1367 io_wq_hash_work(&req->work, file_inode(req->file));
1369 if (def->unbound_nonreg_file)
1370 req->work.flags |= IO_WQ_WORK_UNBOUND;
1373 /* ->mm can never change on us */
1374 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1375 (def->work_flags & IO_WQ_WORK_MM)) {
1377 req->work.flags |= IO_WQ_WORK_MM;
1380 /* if we fail grabbing identity, we must COW, regrab, and retry */
1381 if (io_grab_identity(req))
1384 if (!io_identity_cow(req))
1387 /* can't fail at this point */
1388 if (!io_grab_identity(req))
1392 static void io_prep_async_link(struct io_kiocb *req)
1394 struct io_kiocb *cur;
1396 io_prep_async_work(req);
1397 if (req->flags & REQ_F_LINK_HEAD)
1398 list_for_each_entry(cur, &req->link_list, link_list)
1399 io_prep_async_work(cur);
1402 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1404 struct io_ring_ctx *ctx = req->ctx;
1405 struct io_kiocb *link = io_prep_linked_timeout(req);
1407 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1408 &req->work, req->flags);
1409 io_wq_enqueue(ctx->io_wq, &req->work);
1413 static void io_queue_async_work(struct io_kiocb *req)
1415 struct io_kiocb *link;
1417 /* init ->work of the whole link before punting */
1418 io_prep_async_link(req);
1419 link = __io_queue_async_work(req);
1422 io_queue_linked_timeout(link);
1425 static void io_kill_timeout(struct io_kiocb *req)
1427 struct io_timeout_data *io = req->async_data;
1430 ret = hrtimer_try_to_cancel(&io->timer);
1432 atomic_set(&req->ctx->cq_timeouts,
1433 atomic_read(&req->ctx->cq_timeouts) + 1);
1434 list_del_init(&req->timeout.list);
1435 io_cqring_fill_event(req, 0);
1436 io_put_req_deferred(req, 1);
1440 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1442 struct io_ring_ctx *ctx = req->ctx;
1444 if (!tsk || req->task == tsk)
1446 if (ctx->flags & IORING_SETUP_SQPOLL) {
1447 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1454 * Returns true if we found and killed one or more timeouts
1456 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1458 struct io_kiocb *req, *tmp;
1461 spin_lock_irq(&ctx->completion_lock);
1462 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1463 if (io_task_match(req, tsk)) {
1464 io_kill_timeout(req);
1468 spin_unlock_irq(&ctx->completion_lock);
1469 return canceled != 0;
1472 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1475 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1476 struct io_defer_entry, list);
1477 struct io_kiocb *link;
1479 if (req_need_defer(de->req, de->seq))
1481 list_del_init(&de->list);
1482 /* punt-init is done before queueing for defer */
1483 link = __io_queue_async_work(de->req);
1485 __io_queue_linked_timeout(link);
1486 /* drop submission reference */
1487 io_put_req_deferred(link, 1);
1490 } while (!list_empty(&ctx->defer_list));
1493 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1495 while (!list_empty(&ctx->timeout_list)) {
1496 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1497 struct io_kiocb, timeout.list);
1499 if (io_is_timeout_noseq(req))
1501 if (req->timeout.target_seq != ctx->cached_cq_tail
1502 - atomic_read(&ctx->cq_timeouts))
1505 list_del_init(&req->timeout.list);
1506 io_kill_timeout(req);
1510 static void io_commit_cqring(struct io_ring_ctx *ctx)
1512 io_flush_timeouts(ctx);
1513 __io_commit_cqring(ctx);
1515 if (unlikely(!list_empty(&ctx->defer_list)))
1516 __io_queue_deferred(ctx);
1519 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1521 struct io_rings *r = ctx->rings;
1523 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1526 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1528 struct io_rings *rings = ctx->rings;
1531 tail = ctx->cached_cq_tail;
1533 * writes to the cq entry need to come after reading head; the
1534 * control dependency is enough as we're using WRITE_ONCE to
1537 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1540 ctx->cached_cq_tail++;
1541 return &rings->cqes[tail & ctx->cq_mask];
1544 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1548 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1550 if (!ctx->eventfd_async)
1552 return io_wq_current_is_worker();
1555 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1557 if (waitqueue_active(&ctx->wait))
1558 wake_up(&ctx->wait);
1559 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1560 wake_up(&ctx->sq_data->wait);
1561 if (io_should_trigger_evfd(ctx))
1562 eventfd_signal(ctx->cq_ev_fd, 1);
1565 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1567 if (list_empty(&ctx->cq_overflow_list)) {
1568 clear_bit(0, &ctx->sq_check_overflow);
1569 clear_bit(0, &ctx->cq_check_overflow);
1570 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1574 static inline bool io_match_files(struct io_kiocb *req,
1575 struct files_struct *files)
1579 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1580 (req->work.flags & IO_WQ_WORK_FILES))
1581 return req->work.identity->files == files;
1585 /* Returns true if there are no backlogged entries after the flush */
1586 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1587 struct task_struct *tsk,
1588 struct files_struct *files)
1590 struct io_rings *rings = ctx->rings;
1591 struct io_kiocb *req, *tmp;
1592 struct io_uring_cqe *cqe;
1593 unsigned long flags;
1597 if (list_empty_careful(&ctx->cq_overflow_list))
1599 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1600 rings->cq_ring_entries))
1604 spin_lock_irqsave(&ctx->completion_lock, flags);
1606 /* if force is set, the ring is going away. always drop after that */
1608 ctx->cq_overflow_flushed = 1;
1611 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1612 if (tsk && req->task != tsk)
1614 if (!io_match_files(req, files))
1617 cqe = io_get_cqring(ctx);
1621 list_move(&req->compl.list, &list);
1623 WRITE_ONCE(cqe->user_data, req->user_data);
1624 WRITE_ONCE(cqe->res, req->result);
1625 WRITE_ONCE(cqe->flags, req->compl.cflags);
1627 WRITE_ONCE(ctx->rings->cq_overflow,
1628 atomic_inc_return(&ctx->cached_cq_overflow));
1632 io_commit_cqring(ctx);
1633 io_cqring_mark_overflow(ctx);
1635 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1636 io_cqring_ev_posted(ctx);
1638 while (!list_empty(&list)) {
1639 req = list_first_entry(&list, struct io_kiocb, compl.list);
1640 list_del(&req->compl.list);
1647 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1649 struct io_ring_ctx *ctx = req->ctx;
1650 struct io_uring_cqe *cqe;
1652 trace_io_uring_complete(ctx, req->user_data, res);
1655 * If we can't get a cq entry, userspace overflowed the
1656 * submission (by quite a lot). Increment the overflow count in
1659 cqe = io_get_cqring(ctx);
1661 WRITE_ONCE(cqe->user_data, req->user_data);
1662 WRITE_ONCE(cqe->res, res);
1663 WRITE_ONCE(cqe->flags, cflags);
1664 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1666 * If we're in ring overflow flush mode, or in task cancel mode,
1667 * then we cannot store the request for later flushing, we need
1668 * to drop it on the floor.
1670 WRITE_ONCE(ctx->rings->cq_overflow,
1671 atomic_inc_return(&ctx->cached_cq_overflow));
1673 if (list_empty(&ctx->cq_overflow_list)) {
1674 set_bit(0, &ctx->sq_check_overflow);
1675 set_bit(0, &ctx->cq_check_overflow);
1676 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1680 req->compl.cflags = cflags;
1681 refcount_inc(&req->refs);
1682 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1686 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1688 __io_cqring_fill_event(req, res, 0);
1691 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1693 struct io_ring_ctx *ctx = req->ctx;
1694 unsigned long flags;
1696 spin_lock_irqsave(&ctx->completion_lock, flags);
1697 __io_cqring_fill_event(req, res, cflags);
1698 io_commit_cqring(ctx);
1699 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1701 io_cqring_ev_posted(ctx);
1704 static void io_submit_flush_completions(struct io_comp_state *cs)
1706 struct io_ring_ctx *ctx = cs->ctx;
1708 spin_lock_irq(&ctx->completion_lock);
1709 while (!list_empty(&cs->list)) {
1710 struct io_kiocb *req;
1712 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1713 list_del(&req->compl.list);
1714 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1717 * io_free_req() doesn't care about completion_lock unless one
1718 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1719 * because of a potential deadlock with req->work.fs->lock
1721 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1722 |REQ_F_WORK_INITIALIZED)) {
1723 spin_unlock_irq(&ctx->completion_lock);
1725 spin_lock_irq(&ctx->completion_lock);
1730 io_commit_cqring(ctx);
1731 spin_unlock_irq(&ctx->completion_lock);
1733 io_cqring_ev_posted(ctx);
1737 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1738 struct io_comp_state *cs)
1741 io_cqring_add_event(req, res, cflags);
1746 req->compl.cflags = cflags;
1747 list_add_tail(&req->compl.list, &cs->list);
1749 io_submit_flush_completions(cs);
1753 static void io_req_complete(struct io_kiocb *req, long res)
1755 __io_req_complete(req, res, 0, NULL);
1758 static inline bool io_is_fallback_req(struct io_kiocb *req)
1760 return req == (struct io_kiocb *)
1761 ((unsigned long) req->ctx->fallback_req & ~1UL);
1764 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1766 struct io_kiocb *req;
1768 req = ctx->fallback_req;
1769 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1775 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1776 struct io_submit_state *state)
1778 if (!state->free_reqs) {
1779 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1783 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1784 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1787 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1788 * retry single alloc to be on the safe side.
1790 if (unlikely(ret <= 0)) {
1791 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1792 if (!state->reqs[0])
1796 state->free_reqs = ret;
1800 return state->reqs[state->free_reqs];
1802 return io_get_fallback_req(ctx);
1805 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1809 percpu_ref_put(req->fixed_file_refs);
1814 static void io_dismantle_req(struct io_kiocb *req)
1818 if (req->async_data)
1819 kfree(req->async_data);
1821 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1823 io_req_clean_work(req);
1826 static void __io_free_req(struct io_kiocb *req)
1828 struct io_uring_task *tctx = req->task->io_uring;
1829 struct io_ring_ctx *ctx = req->ctx;
1831 io_dismantle_req(req);
1833 percpu_counter_dec(&tctx->inflight);
1835 wake_up(&tctx->wait);
1836 put_task_struct(req->task);
1838 if (likely(!io_is_fallback_req(req)))
1839 kmem_cache_free(req_cachep, req);
1841 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1842 percpu_ref_put(&ctx->refs);
1845 static bool io_link_cancel_timeout(struct io_kiocb *req)
1847 struct io_timeout_data *io = req->async_data;
1848 struct io_ring_ctx *ctx = req->ctx;
1851 ret = hrtimer_try_to_cancel(&io->timer);
1853 io_cqring_fill_event(req, -ECANCELED);
1854 io_commit_cqring(ctx);
1855 req->flags &= ~REQ_F_LINK_HEAD;
1856 io_put_req_deferred(req, 1);
1863 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1865 struct io_kiocb *link;
1868 if (list_empty(&req->link_list))
1870 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1871 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1874 list_del_init(&link->link_list);
1875 wake_ev = io_link_cancel_timeout(link);
1876 req->flags &= ~REQ_F_LINK_TIMEOUT;
1880 static void io_kill_linked_timeout(struct io_kiocb *req)
1882 struct io_ring_ctx *ctx = req->ctx;
1883 unsigned long flags;
1886 spin_lock_irqsave(&ctx->completion_lock, flags);
1887 wake_ev = __io_kill_linked_timeout(req);
1888 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1891 io_cqring_ev_posted(ctx);
1894 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1896 struct io_kiocb *nxt;
1899 * The list should never be empty when we are called here. But could
1900 * potentially happen if the chain is messed up, check to be on the
1903 if (unlikely(list_empty(&req->link_list)))
1906 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1907 list_del_init(&req->link_list);
1908 if (!list_empty(&nxt->link_list))
1909 nxt->flags |= REQ_F_LINK_HEAD;
1914 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1916 static void __io_fail_links(struct io_kiocb *req)
1918 struct io_ring_ctx *ctx = req->ctx;
1920 while (!list_empty(&req->link_list)) {
1921 struct io_kiocb *link = list_first_entry(&req->link_list,
1922 struct io_kiocb, link_list);
1924 list_del_init(&link->link_list);
1925 trace_io_uring_fail_link(req, link);
1927 io_cqring_fill_event(link, -ECANCELED);
1930 * It's ok to free under spinlock as they're not linked anymore,
1931 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1934 if (link->flags & REQ_F_WORK_INITIALIZED)
1935 io_put_req_deferred(link, 2);
1937 io_double_put_req(link);
1940 io_commit_cqring(ctx);
1943 static void io_fail_links(struct io_kiocb *req)
1945 struct io_ring_ctx *ctx = req->ctx;
1946 unsigned long flags;
1948 spin_lock_irqsave(&ctx->completion_lock, flags);
1949 __io_fail_links(req);
1950 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1952 io_cqring_ev_posted(ctx);
1955 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1957 req->flags &= ~REQ_F_LINK_HEAD;
1958 if (req->flags & REQ_F_LINK_TIMEOUT)
1959 io_kill_linked_timeout(req);
1962 * If LINK is set, we have dependent requests in this chain. If we
1963 * didn't fail this request, queue the first one up, moving any other
1964 * dependencies to the next request. In case of failure, fail the rest
1967 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1968 return io_req_link_next(req);
1973 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1975 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1977 return __io_req_find_next(req);
1980 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1982 struct task_struct *tsk = req->task;
1983 struct io_ring_ctx *ctx = req->ctx;
1986 if (tsk->flags & PF_EXITING)
1990 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1991 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1992 * processing task_work. There's no reliable way to tell if TWA_RESUME
1996 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1997 notify = TWA_SIGNAL;
1999 ret = task_work_add(tsk, &req->task_work, notify);
2001 wake_up_process(tsk);
2006 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2008 struct io_ring_ctx *ctx = req->ctx;
2010 spin_lock_irq(&ctx->completion_lock);
2011 io_cqring_fill_event(req, error);
2012 io_commit_cqring(ctx);
2013 spin_unlock_irq(&ctx->completion_lock);
2015 io_cqring_ev_posted(ctx);
2016 req_set_fail_links(req);
2017 io_double_put_req(req);
2020 static void io_req_task_cancel(struct callback_head *cb)
2022 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2023 struct io_ring_ctx *ctx = req->ctx;
2025 __io_req_task_cancel(req, -ECANCELED);
2026 percpu_ref_put(&ctx->refs);
2029 static void __io_req_task_submit(struct io_kiocb *req)
2031 struct io_ring_ctx *ctx = req->ctx;
2033 if (!__io_sq_thread_acquire_mm(ctx)) {
2034 mutex_lock(&ctx->uring_lock);
2035 __io_queue_sqe(req, NULL);
2036 mutex_unlock(&ctx->uring_lock);
2038 __io_req_task_cancel(req, -EFAULT);
2042 static void io_req_task_submit(struct callback_head *cb)
2044 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2045 struct io_ring_ctx *ctx = req->ctx;
2047 __io_req_task_submit(req);
2048 percpu_ref_put(&ctx->refs);
2051 static void io_req_task_queue(struct io_kiocb *req)
2055 init_task_work(&req->task_work, io_req_task_submit);
2056 percpu_ref_get(&req->ctx->refs);
2058 ret = io_req_task_work_add(req, true);
2059 if (unlikely(ret)) {
2060 struct task_struct *tsk;
2062 init_task_work(&req->task_work, io_req_task_cancel);
2063 tsk = io_wq_get_task(req->ctx->io_wq);
2064 task_work_add(tsk, &req->task_work, 0);
2065 wake_up_process(tsk);
2069 static void io_queue_next(struct io_kiocb *req)
2071 struct io_kiocb *nxt = io_req_find_next(req);
2074 io_req_task_queue(nxt);
2077 static void io_free_req(struct io_kiocb *req)
2084 void *reqs[IO_IOPOLL_BATCH];
2087 struct task_struct *task;
2091 static inline void io_init_req_batch(struct req_batch *rb)
2098 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2099 struct req_batch *rb)
2101 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2102 percpu_ref_put_many(&ctx->refs, rb->to_free);
2106 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2107 struct req_batch *rb)
2110 __io_req_free_batch_flush(ctx, rb);
2112 struct io_uring_task *tctx = rb->task->io_uring;
2114 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2115 put_task_struct_many(rb->task, rb->task_refs);
2120 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2122 if (unlikely(io_is_fallback_req(req))) {
2126 if (req->flags & REQ_F_LINK_HEAD)
2129 if (req->task != rb->task) {
2131 struct io_uring_task *tctx = rb->task->io_uring;
2133 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2134 put_task_struct_many(rb->task, rb->task_refs);
2136 rb->task = req->task;
2141 io_dismantle_req(req);
2142 rb->reqs[rb->to_free++] = req;
2143 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2144 __io_req_free_batch_flush(req->ctx, rb);
2148 * Drop reference to request, return next in chain (if there is one) if this
2149 * was the last reference to this request.
2151 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2153 struct io_kiocb *nxt = NULL;
2155 if (refcount_dec_and_test(&req->refs)) {
2156 nxt = io_req_find_next(req);
2162 static void io_put_req(struct io_kiocb *req)
2164 if (refcount_dec_and_test(&req->refs))
2168 static void io_put_req_deferred_cb(struct callback_head *cb)
2170 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2175 static void io_free_req_deferred(struct io_kiocb *req)
2179 init_task_work(&req->task_work, io_put_req_deferred_cb);
2180 ret = io_req_task_work_add(req, true);
2181 if (unlikely(ret)) {
2182 struct task_struct *tsk;
2184 tsk = io_wq_get_task(req->ctx->io_wq);
2185 task_work_add(tsk, &req->task_work, 0);
2186 wake_up_process(tsk);
2190 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2192 if (refcount_sub_and_test(refs, &req->refs))
2193 io_free_req_deferred(req);
2196 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2198 struct io_kiocb *nxt;
2201 * A ref is owned by io-wq in which context we're. So, if that's the
2202 * last one, it's safe to steal next work. False negatives are Ok,
2203 * it just will be re-punted async in io_put_work()
2205 if (refcount_read(&req->refs) != 1)
2208 nxt = io_req_find_next(req);
2209 return nxt ? &nxt->work : NULL;
2212 static void io_double_put_req(struct io_kiocb *req)
2214 /* drop both submit and complete references */
2215 if (refcount_sub_and_test(2, &req->refs))
2219 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2221 struct io_rings *rings = ctx->rings;
2223 if (test_bit(0, &ctx->cq_check_overflow)) {
2225 * noflush == true is from the waitqueue handler, just ensure
2226 * we wake up the task, and the next invocation will flush the
2227 * entries. We cannot safely to it from here.
2229 if (noflush && !list_empty(&ctx->cq_overflow_list))
2232 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2235 /* See comment at the top of this file */
2237 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2240 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2242 struct io_rings *rings = ctx->rings;
2244 /* make sure SQ entry isn't read before tail */
2245 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2248 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2250 unsigned int cflags;
2252 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2253 cflags |= IORING_CQE_F_BUFFER;
2254 req->flags &= ~REQ_F_BUFFER_SELECTED;
2259 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2261 struct io_buffer *kbuf;
2263 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2264 return io_put_kbuf(req, kbuf);
2267 static inline bool io_run_task_work(void)
2270 * Not safe to run on exiting task, and the task_work handling will
2271 * not add work to such a task.
2273 if (unlikely(current->flags & PF_EXITING))
2275 if (current->task_works) {
2276 __set_current_state(TASK_RUNNING);
2284 static void io_iopoll_queue(struct list_head *again)
2286 struct io_kiocb *req;
2289 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2290 list_del(&req->inflight_entry);
2291 __io_complete_rw(req, -EAGAIN, 0, NULL);
2292 } while (!list_empty(again));
2296 * Find and free completed poll iocbs
2298 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2299 struct list_head *done)
2301 struct req_batch rb;
2302 struct io_kiocb *req;
2305 /* order with ->result store in io_complete_rw_iopoll() */
2308 io_init_req_batch(&rb);
2309 while (!list_empty(done)) {
2312 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2313 if (READ_ONCE(req->result) == -EAGAIN) {
2315 req->iopoll_completed = 0;
2316 list_move_tail(&req->inflight_entry, &again);
2319 list_del(&req->inflight_entry);
2321 if (req->flags & REQ_F_BUFFER_SELECTED)
2322 cflags = io_put_rw_kbuf(req);
2324 __io_cqring_fill_event(req, req->result, cflags);
2327 if (refcount_dec_and_test(&req->refs))
2328 io_req_free_batch(&rb, req);
2331 io_commit_cqring(ctx);
2332 if (ctx->flags & IORING_SETUP_SQPOLL)
2333 io_cqring_ev_posted(ctx);
2334 io_req_free_batch_finish(ctx, &rb);
2336 if (!list_empty(&again))
2337 io_iopoll_queue(&again);
2340 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2343 struct io_kiocb *req, *tmp;
2349 * Only spin for completions if we don't have multiple devices hanging
2350 * off our complete list, and we're under the requested amount.
2352 spin = !ctx->poll_multi_file && *nr_events < min;
2355 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2356 struct kiocb *kiocb = &req->rw.kiocb;
2359 * Move completed and retryable entries to our local lists.
2360 * If we find a request that requires polling, break out
2361 * and complete those lists first, if we have entries there.
2363 if (READ_ONCE(req->iopoll_completed)) {
2364 list_move_tail(&req->inflight_entry, &done);
2367 if (!list_empty(&done))
2370 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2374 /* iopoll may have completed current req */
2375 if (READ_ONCE(req->iopoll_completed))
2376 list_move_tail(&req->inflight_entry, &done);
2383 if (!list_empty(&done))
2384 io_iopoll_complete(ctx, nr_events, &done);
2390 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2391 * non-spinning poll check - we'll still enter the driver poll loop, but only
2392 * as a non-spinning completion check.
2394 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2397 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2400 ret = io_do_iopoll(ctx, nr_events, min);
2403 if (*nr_events >= min)
2411 * We can't just wait for polled events to come to us, we have to actively
2412 * find and complete them.
2414 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2416 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2419 mutex_lock(&ctx->uring_lock);
2420 while (!list_empty(&ctx->iopoll_list)) {
2421 unsigned int nr_events = 0;
2423 io_do_iopoll(ctx, &nr_events, 0);
2425 /* let it sleep and repeat later if can't complete a request */
2429 * Ensure we allow local-to-the-cpu processing to take place,
2430 * in this case we need to ensure that we reap all events.
2431 * Also let task_work, etc. to progress by releasing the mutex
2433 if (need_resched()) {
2434 mutex_unlock(&ctx->uring_lock);
2436 mutex_lock(&ctx->uring_lock);
2439 mutex_unlock(&ctx->uring_lock);
2442 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2444 unsigned int nr_events = 0;
2445 int iters = 0, ret = 0;
2448 * We disallow the app entering submit/complete with polling, but we
2449 * still need to lock the ring to prevent racing with polled issue
2450 * that got punted to a workqueue.
2452 mutex_lock(&ctx->uring_lock);
2455 * Don't enter poll loop if we already have events pending.
2456 * If we do, we can potentially be spinning for commands that
2457 * already triggered a CQE (eg in error).
2459 if (io_cqring_events(ctx, false))
2463 * If a submit got punted to a workqueue, we can have the
2464 * application entering polling for a command before it gets
2465 * issued. That app will hold the uring_lock for the duration
2466 * of the poll right here, so we need to take a breather every
2467 * now and then to ensure that the issue has a chance to add
2468 * the poll to the issued list. Otherwise we can spin here
2469 * forever, while the workqueue is stuck trying to acquire the
2472 if (!(++iters & 7)) {
2473 mutex_unlock(&ctx->uring_lock);
2475 mutex_lock(&ctx->uring_lock);
2478 ret = io_iopoll_getevents(ctx, &nr_events, min);
2482 } while (min && !nr_events && !need_resched());
2484 mutex_unlock(&ctx->uring_lock);
2488 static void kiocb_end_write(struct io_kiocb *req)
2491 * Tell lockdep we inherited freeze protection from submission
2494 if (req->flags & REQ_F_ISREG) {
2495 struct inode *inode = file_inode(req->file);
2497 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2499 file_end_write(req->file);
2502 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2503 struct io_comp_state *cs)
2505 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2508 if (kiocb->ki_flags & IOCB_WRITE)
2509 kiocb_end_write(req);
2511 if (res != req->result)
2512 req_set_fail_links(req);
2513 if (req->flags & REQ_F_BUFFER_SELECTED)
2514 cflags = io_put_rw_kbuf(req);
2515 __io_req_complete(req, res, cflags, cs);
2519 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2521 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2522 ssize_t ret = -ECANCELED;
2523 struct iov_iter iter;
2531 switch (req->opcode) {
2532 case IORING_OP_READV:
2533 case IORING_OP_READ_FIXED:
2534 case IORING_OP_READ:
2537 case IORING_OP_WRITEV:
2538 case IORING_OP_WRITE_FIXED:
2539 case IORING_OP_WRITE:
2543 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2548 if (!req->async_data) {
2549 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2552 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2560 req_set_fail_links(req);
2561 io_req_complete(req, ret);
2566 static bool io_rw_reissue(struct io_kiocb *req, long res)
2569 umode_t mode = file_inode(req->file)->i_mode;
2572 if (!S_ISBLK(mode) && !S_ISREG(mode))
2574 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2577 ret = io_sq_thread_acquire_mm(req->ctx, req);
2579 if (io_resubmit_prep(req, ret)) {
2580 refcount_inc(&req->refs);
2581 io_queue_async_work(req);
2589 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2590 struct io_comp_state *cs)
2592 if (!io_rw_reissue(req, res))
2593 io_complete_rw_common(&req->rw.kiocb, res, cs);
2596 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2598 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2600 __io_complete_rw(req, res, res2, NULL);
2603 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2605 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2607 if (kiocb->ki_flags & IOCB_WRITE)
2608 kiocb_end_write(req);
2610 if (res != -EAGAIN && res != req->result)
2611 req_set_fail_links(req);
2613 WRITE_ONCE(req->result, res);
2614 /* order with io_poll_complete() checking ->result */
2616 WRITE_ONCE(req->iopoll_completed, 1);
2620 * After the iocb has been issued, it's safe to be found on the poll list.
2621 * Adding the kiocb to the list AFTER submission ensures that we don't
2622 * find it from a io_iopoll_getevents() thread before the issuer is done
2623 * accessing the kiocb cookie.
2625 static void io_iopoll_req_issued(struct io_kiocb *req)
2627 struct io_ring_ctx *ctx = req->ctx;
2630 * Track whether we have multiple files in our lists. This will impact
2631 * how we do polling eventually, not spinning if we're on potentially
2632 * different devices.
2634 if (list_empty(&ctx->iopoll_list)) {
2635 ctx->poll_multi_file = false;
2636 } else if (!ctx->poll_multi_file) {
2637 struct io_kiocb *list_req;
2639 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2641 if (list_req->file != req->file)
2642 ctx->poll_multi_file = true;
2646 * For fast devices, IO may have already completed. If it has, add
2647 * it to the front so we find it first.
2649 if (READ_ONCE(req->iopoll_completed))
2650 list_add(&req->inflight_entry, &ctx->iopoll_list);
2652 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2654 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2655 wq_has_sleeper(&ctx->sq_data->wait))
2656 wake_up(&ctx->sq_data->wait);
2659 static void __io_state_file_put(struct io_submit_state *state)
2661 if (state->has_refs)
2662 fput_many(state->file, state->has_refs);
2666 static inline void io_state_file_put(struct io_submit_state *state)
2669 __io_state_file_put(state);
2673 * Get as many references to a file as we have IOs left in this submission,
2674 * assuming most submissions are for one file, or at least that each file
2675 * has more than one submission.
2677 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2683 if (state->fd == fd) {
2687 __io_state_file_put(state);
2689 state->file = fget_many(fd, state->ios_left);
2694 state->has_refs = state->ios_left - 1;
2698 static bool io_bdev_nowait(struct block_device *bdev)
2701 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2708 * If we tracked the file through the SCM inflight mechanism, we could support
2709 * any file. For now, just ensure that anything potentially problematic is done
2712 static bool io_file_supports_async(struct file *file, int rw)
2714 umode_t mode = file_inode(file)->i_mode;
2716 if (S_ISBLK(mode)) {
2717 if (io_bdev_nowait(file->f_inode->i_bdev))
2721 if (S_ISCHR(mode) || S_ISSOCK(mode))
2723 if (S_ISREG(mode)) {
2724 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2725 file->f_op != &io_uring_fops)
2730 /* any ->read/write should understand O_NONBLOCK */
2731 if (file->f_flags & O_NONBLOCK)
2734 if (!(file->f_mode & FMODE_NOWAIT))
2738 return file->f_op->read_iter != NULL;
2740 return file->f_op->write_iter != NULL;
2743 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2745 struct io_ring_ctx *ctx = req->ctx;
2746 struct kiocb *kiocb = &req->rw.kiocb;
2750 if (S_ISREG(file_inode(req->file)->i_mode))
2751 req->flags |= REQ_F_ISREG;
2753 kiocb->ki_pos = READ_ONCE(sqe->off);
2754 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2755 req->flags |= REQ_F_CUR_POS;
2756 kiocb->ki_pos = req->file->f_pos;
2758 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2759 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2760 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2764 ioprio = READ_ONCE(sqe->ioprio);
2766 ret = ioprio_check_cap(ioprio);
2770 kiocb->ki_ioprio = ioprio;
2772 kiocb->ki_ioprio = get_current_ioprio();
2774 /* don't allow async punt if RWF_NOWAIT was requested */
2775 if (kiocb->ki_flags & IOCB_NOWAIT)
2776 req->flags |= REQ_F_NOWAIT;
2778 if (ctx->flags & IORING_SETUP_IOPOLL) {
2779 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2780 !kiocb->ki_filp->f_op->iopoll)
2783 kiocb->ki_flags |= IOCB_HIPRI;
2784 kiocb->ki_complete = io_complete_rw_iopoll;
2785 req->iopoll_completed = 0;
2787 if (kiocb->ki_flags & IOCB_HIPRI)
2789 kiocb->ki_complete = io_complete_rw;
2792 req->rw.addr = READ_ONCE(sqe->addr);
2793 req->rw.len = READ_ONCE(sqe->len);
2794 req->buf_index = READ_ONCE(sqe->buf_index);
2798 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2804 case -ERESTARTNOINTR:
2805 case -ERESTARTNOHAND:
2806 case -ERESTART_RESTARTBLOCK:
2808 * We can't just restart the syscall, since previously
2809 * submitted sqes may already be in progress. Just fail this
2815 kiocb->ki_complete(kiocb, ret, 0);
2819 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2820 struct io_comp_state *cs)
2822 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2823 struct io_async_rw *io = req->async_data;
2825 /* add previously done IO, if any */
2826 if (io && io->bytes_done > 0) {
2828 ret = io->bytes_done;
2830 ret += io->bytes_done;
2833 if (req->flags & REQ_F_CUR_POS)
2834 req->file->f_pos = kiocb->ki_pos;
2835 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2836 __io_complete_rw(req, ret, 0, cs);
2838 io_rw_done(kiocb, ret);
2841 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2842 struct iov_iter *iter)
2844 struct io_ring_ctx *ctx = req->ctx;
2845 size_t len = req->rw.len;
2846 struct io_mapped_ubuf *imu;
2847 u16 index, buf_index = req->buf_index;
2851 if (unlikely(buf_index >= ctx->nr_user_bufs))
2853 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2854 imu = &ctx->user_bufs[index];
2855 buf_addr = req->rw.addr;
2858 if (buf_addr + len < buf_addr)
2860 /* not inside the mapped region */
2861 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2865 * May not be a start of buffer, set size appropriately
2866 * and advance us to the beginning.
2868 offset = buf_addr - imu->ubuf;
2869 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2873 * Don't use iov_iter_advance() here, as it's really slow for
2874 * using the latter parts of a big fixed buffer - it iterates
2875 * over each segment manually. We can cheat a bit here, because
2878 * 1) it's a BVEC iter, we set it up
2879 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2880 * first and last bvec
2882 * So just find our index, and adjust the iterator afterwards.
2883 * If the offset is within the first bvec (or the whole first
2884 * bvec, just use iov_iter_advance(). This makes it easier
2885 * since we can just skip the first segment, which may not
2886 * be PAGE_SIZE aligned.
2888 const struct bio_vec *bvec = imu->bvec;
2890 if (offset <= bvec->bv_len) {
2891 iov_iter_advance(iter, offset);
2893 unsigned long seg_skip;
2895 /* skip first vec */
2896 offset -= bvec->bv_len;
2897 seg_skip = 1 + (offset >> PAGE_SHIFT);
2899 iter->bvec = bvec + seg_skip;
2900 iter->nr_segs -= seg_skip;
2901 iter->count -= bvec->bv_len + offset;
2902 iter->iov_offset = offset & ~PAGE_MASK;
2909 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2912 mutex_unlock(&ctx->uring_lock);
2915 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2918 * "Normal" inline submissions always hold the uring_lock, since we
2919 * grab it from the system call. Same is true for the SQPOLL offload.
2920 * The only exception is when we've detached the request and issue it
2921 * from an async worker thread, grab the lock for that case.
2924 mutex_lock(&ctx->uring_lock);
2927 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2928 int bgid, struct io_buffer *kbuf,
2931 struct io_buffer *head;
2933 if (req->flags & REQ_F_BUFFER_SELECTED)
2936 io_ring_submit_lock(req->ctx, needs_lock);
2938 lockdep_assert_held(&req->ctx->uring_lock);
2940 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2942 if (!list_empty(&head->list)) {
2943 kbuf = list_last_entry(&head->list, struct io_buffer,
2945 list_del(&kbuf->list);
2948 idr_remove(&req->ctx->io_buffer_idr, bgid);
2950 if (*len > kbuf->len)
2953 kbuf = ERR_PTR(-ENOBUFS);
2956 io_ring_submit_unlock(req->ctx, needs_lock);
2961 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2964 struct io_buffer *kbuf;
2967 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2968 bgid = req->buf_index;
2969 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2972 req->rw.addr = (u64) (unsigned long) kbuf;
2973 req->flags |= REQ_F_BUFFER_SELECTED;
2974 return u64_to_user_ptr(kbuf->addr);
2977 #ifdef CONFIG_COMPAT
2978 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2981 struct compat_iovec __user *uiov;
2982 compat_ssize_t clen;
2986 uiov = u64_to_user_ptr(req->rw.addr);
2987 if (!access_ok(uiov, sizeof(*uiov)))
2989 if (__get_user(clen, &uiov->iov_len))
2995 buf = io_rw_buffer_select(req, &len, needs_lock);
2997 return PTR_ERR(buf);
2998 iov[0].iov_base = buf;
2999 iov[0].iov_len = (compat_size_t) len;
3004 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3007 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3011 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3014 len = iov[0].iov_len;
3017 buf = io_rw_buffer_select(req, &len, needs_lock);
3019 return PTR_ERR(buf);
3020 iov[0].iov_base = buf;
3021 iov[0].iov_len = len;
3025 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3028 if (req->flags & REQ_F_BUFFER_SELECTED) {
3029 struct io_buffer *kbuf;
3031 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3032 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3033 iov[0].iov_len = kbuf->len;
3038 else if (req->rw.len > 1)
3041 #ifdef CONFIG_COMPAT
3042 if (req->ctx->compat)
3043 return io_compat_import(req, iov, needs_lock);
3046 return __io_iov_buffer_select(req, iov, needs_lock);
3049 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3050 struct iovec **iovec, struct iov_iter *iter,
3053 void __user *buf = u64_to_user_ptr(req->rw.addr);
3054 size_t sqe_len = req->rw.len;
3058 opcode = req->opcode;
3059 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3061 return io_import_fixed(req, rw, iter);
3064 /* buffer index only valid with fixed read/write, or buffer select */
3065 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3068 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3069 if (req->flags & REQ_F_BUFFER_SELECT) {
3070 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3072 return PTR_ERR(buf);
3073 req->rw.len = sqe_len;
3076 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3078 return ret < 0 ? ret : sqe_len;
3081 if (req->flags & REQ_F_BUFFER_SELECT) {
3082 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3084 ret = (*iovec)->iov_len;
3085 iov_iter_init(iter, rw, *iovec, 1, ret);
3091 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3095 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3096 struct iovec **iovec, struct iov_iter *iter,
3099 struct io_async_rw *iorw = req->async_data;
3102 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3104 return iov_iter_count(&iorw->iter);
3107 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3109 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3113 * For files that don't have ->read_iter() and ->write_iter(), handle them
3114 * by looping over ->read() or ->write() manually.
3116 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3117 struct iov_iter *iter)
3122 * Don't support polled IO through this interface, and we can't
3123 * support non-blocking either. For the latter, this just causes
3124 * the kiocb to be handled from an async context.
3126 if (kiocb->ki_flags & IOCB_HIPRI)
3128 if (kiocb->ki_flags & IOCB_NOWAIT)
3131 while (iov_iter_count(iter)) {
3135 if (!iov_iter_is_bvec(iter)) {
3136 iovec = iov_iter_iovec(iter);
3138 /* fixed buffers import bvec */
3139 iovec.iov_base = kmap(iter->bvec->bv_page)
3141 iovec.iov_len = min(iter->count,
3142 iter->bvec->bv_len - iter->iov_offset);
3146 nr = file->f_op->read(file, iovec.iov_base,
3147 iovec.iov_len, io_kiocb_ppos(kiocb));
3149 nr = file->f_op->write(file, iovec.iov_base,
3150 iovec.iov_len, io_kiocb_ppos(kiocb));
3153 if (iov_iter_is_bvec(iter))
3154 kunmap(iter->bvec->bv_page);
3162 if (nr != iovec.iov_len)
3164 iov_iter_advance(iter, nr);
3170 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3171 const struct iovec *fast_iov, struct iov_iter *iter)
3173 struct io_async_rw *rw = req->async_data;
3175 memcpy(&rw->iter, iter, sizeof(*iter));
3176 rw->free_iovec = iovec;
3178 /* can only be fixed buffers, no need to do anything */
3179 if (iter->type == ITER_BVEC)
3182 unsigned iov_off = 0;
3184 rw->iter.iov = rw->fast_iov;
3185 if (iter->iov != fast_iov) {
3186 iov_off = iter->iov - fast_iov;
3187 rw->iter.iov += iov_off;
3189 if (rw->fast_iov != fast_iov)
3190 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3191 sizeof(struct iovec) * iter->nr_segs);
3193 req->flags |= REQ_F_NEED_CLEANUP;
3197 static inline int __io_alloc_async_data(struct io_kiocb *req)
3199 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3200 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3201 return req->async_data == NULL;
3204 static int io_alloc_async_data(struct io_kiocb *req)
3206 if (!io_op_defs[req->opcode].needs_async_data)
3209 return __io_alloc_async_data(req);
3212 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3213 const struct iovec *fast_iov,
3214 struct iov_iter *iter, bool force)
3216 if (!force && !io_op_defs[req->opcode].needs_async_data)
3218 if (!req->async_data) {
3219 if (__io_alloc_async_data(req))
3222 io_req_map_rw(req, iovec, fast_iov, iter);
3227 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3229 struct io_async_rw *iorw = req->async_data;
3230 struct iovec *iov = iorw->fast_iov;
3233 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3234 if (unlikely(ret < 0))
3237 iorw->bytes_done = 0;
3238 iorw->free_iovec = iov;
3240 req->flags |= REQ_F_NEED_CLEANUP;
3244 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3248 ret = io_prep_rw(req, sqe);
3252 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3255 /* either don't need iovec imported or already have it */
3256 if (!req->async_data)
3258 return io_rw_prep_async(req, READ);
3262 * This is our waitqueue callback handler, registered through lock_page_async()
3263 * when we initially tried to do the IO with the iocb armed our waitqueue.
3264 * This gets called when the page is unlocked, and we generally expect that to
3265 * happen when the page IO is completed and the page is now uptodate. This will
3266 * queue a task_work based retry of the operation, attempting to copy the data
3267 * again. If the latter fails because the page was NOT uptodate, then we will
3268 * do a thread based blocking retry of the operation. That's the unexpected
3271 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3272 int sync, void *arg)
3274 struct wait_page_queue *wpq;
3275 struct io_kiocb *req = wait->private;
3276 struct wait_page_key *key = arg;
3279 wpq = container_of(wait, struct wait_page_queue, wait);
3281 if (!wake_page_match(wpq, key))
3284 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3285 list_del_init(&wait->entry);
3287 init_task_work(&req->task_work, io_req_task_submit);
3288 percpu_ref_get(&req->ctx->refs);
3290 /* submit ref gets dropped, acquire a new one */
3291 refcount_inc(&req->refs);
3292 ret = io_req_task_work_add(req, true);
3293 if (unlikely(ret)) {
3294 struct task_struct *tsk;
3296 /* queue just for cancelation */
3297 init_task_work(&req->task_work, io_req_task_cancel);
3298 tsk = io_wq_get_task(req->ctx->io_wq);
3299 task_work_add(tsk, &req->task_work, 0);
3300 wake_up_process(tsk);
3306 * This controls whether a given IO request should be armed for async page
3307 * based retry. If we return false here, the request is handed to the async
3308 * worker threads for retry. If we're doing buffered reads on a regular file,
3309 * we prepare a private wait_page_queue entry and retry the operation. This
3310 * will either succeed because the page is now uptodate and unlocked, or it
3311 * will register a callback when the page is unlocked at IO completion. Through
3312 * that callback, io_uring uses task_work to setup a retry of the operation.
3313 * That retry will attempt the buffered read again. The retry will generally
3314 * succeed, or in rare cases where it fails, we then fall back to using the
3315 * async worker threads for a blocking retry.
3317 static bool io_rw_should_retry(struct io_kiocb *req)
3319 struct io_async_rw *rw = req->async_data;
3320 struct wait_page_queue *wait = &rw->wpq;
3321 struct kiocb *kiocb = &req->rw.kiocb;
3323 /* never retry for NOWAIT, we just complete with -EAGAIN */
3324 if (req->flags & REQ_F_NOWAIT)
3327 /* Only for buffered IO */
3328 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3332 * just use poll if we can, and don't attempt if the fs doesn't
3333 * support callback based unlocks
3335 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3338 wait->wait.func = io_async_buf_func;
3339 wait->wait.private = req;
3340 wait->wait.flags = 0;
3341 INIT_LIST_HEAD(&wait->wait.entry);
3342 kiocb->ki_flags |= IOCB_WAITQ;
3343 kiocb->ki_flags &= ~IOCB_NOWAIT;
3344 kiocb->ki_waitq = wait;
3348 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3350 if (req->file->f_op->read_iter)
3351 return call_read_iter(req->file, &req->rw.kiocb, iter);
3352 else if (req->file->f_op->read)
3353 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3358 static int io_read(struct io_kiocb *req, bool force_nonblock,
3359 struct io_comp_state *cs)
3361 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3362 struct kiocb *kiocb = &req->rw.kiocb;
3363 struct iov_iter __iter, *iter = &__iter;
3364 struct io_async_rw *rw = req->async_data;
3365 ssize_t io_size, ret, ret2;
3372 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3375 iov_count = iov_iter_count(iter);
3377 req->result = io_size;
3380 /* Ensure we clear previously set non-block flag */
3381 if (!force_nonblock)
3382 kiocb->ki_flags &= ~IOCB_NOWAIT;
3384 kiocb->ki_flags |= IOCB_NOWAIT;
3387 /* If the file doesn't support async, just async punt */
3388 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3392 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3396 ret = io_iter_do_read(req, iter);
3400 } else if (ret == -EIOCBQUEUED) {
3403 } else if (ret == -EAGAIN) {
3404 /* IOPOLL retry should happen for io-wq threads */
3405 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3407 /* no retry on NONBLOCK marked file */
3408 if (req->file->f_flags & O_NONBLOCK)
3410 /* some cases will consume bytes even on error returns */
3411 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3414 } else if (ret < 0) {
3415 /* make sure -ERESTARTSYS -> -EINTR is done */
3419 /* read it all, or we did blocking attempt. no retry. */
3420 if (!iov_iter_count(iter) || !force_nonblock ||
3421 (req->file->f_flags & O_NONBLOCK))
3426 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3433 rw = req->async_data;
3434 /* it's copied and will be cleaned with ->io */
3436 /* now use our persistent iterator, if we aren't already */
3439 rw->bytes_done += ret;
3440 /* if we can retry, do so with the callbacks armed */
3441 if (!io_rw_should_retry(req)) {
3442 kiocb->ki_flags &= ~IOCB_WAITQ;
3447 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3448 * get -EIOCBQUEUED, then we'll get a notification when the desired
3449 * page gets unlocked. We can also get a partial read here, and if we
3450 * do, then just retry at the new offset.
3452 ret = io_iter_do_read(req, iter);
3453 if (ret == -EIOCBQUEUED) {
3456 } else if (ret > 0 && ret < io_size) {
3457 /* we got some bytes, but not all. retry. */
3461 kiocb_done(kiocb, ret, cs);
3464 /* it's reportedly faster than delegating the null check to kfree() */
3470 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3474 ret = io_prep_rw(req, sqe);
3478 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3481 /* either don't need iovec imported or already have it */
3482 if (!req->async_data)
3484 return io_rw_prep_async(req, WRITE);
3487 static int io_write(struct io_kiocb *req, bool force_nonblock,
3488 struct io_comp_state *cs)
3490 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3491 struct kiocb *kiocb = &req->rw.kiocb;
3492 struct iov_iter __iter, *iter = &__iter;
3493 struct io_async_rw *rw = req->async_data;
3495 ssize_t ret, ret2, io_size;
3500 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3503 iov_count = iov_iter_count(iter);
3505 req->result = io_size;
3507 /* Ensure we clear previously set non-block flag */
3508 if (!force_nonblock)
3509 kiocb->ki_flags &= ~IOCB_NOWAIT;
3511 kiocb->ki_flags |= IOCB_NOWAIT;
3513 /* If the file doesn't support async, just async punt */
3514 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3517 /* file path doesn't support NOWAIT for non-direct_IO */
3518 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3519 (req->flags & REQ_F_ISREG))
3522 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3527 * Open-code file_start_write here to grab freeze protection,
3528 * which will be released by another thread in
3529 * io_complete_rw(). Fool lockdep by telling it the lock got
3530 * released so that it doesn't complain about the held lock when
3531 * we return to userspace.
3533 if (req->flags & REQ_F_ISREG) {
3534 __sb_start_write(file_inode(req->file)->i_sb,
3535 SB_FREEZE_WRITE, true);
3536 __sb_writers_release(file_inode(req->file)->i_sb,
3539 kiocb->ki_flags |= IOCB_WRITE;
3541 if (req->file->f_op->write_iter)
3542 ret2 = call_write_iter(req->file, kiocb, iter);
3543 else if (req->file->f_op->write)
3544 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3549 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3550 * retry them without IOCB_NOWAIT.
3552 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3554 /* no retry on NONBLOCK marked file */
3555 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3557 if (!force_nonblock || ret2 != -EAGAIN) {
3558 /* IOPOLL retry should happen for io-wq threads */
3559 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3562 kiocb_done(kiocb, ret2, cs);
3565 /* some cases will consume bytes even on error returns */
3566 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3567 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3572 /* it's reportedly faster than delegating the null check to kfree() */
3578 static int __io_splice_prep(struct io_kiocb *req,
3579 const struct io_uring_sqe *sqe)
3581 struct io_splice* sp = &req->splice;
3582 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3584 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3588 sp->len = READ_ONCE(sqe->len);
3589 sp->flags = READ_ONCE(sqe->splice_flags);
3591 if (unlikely(sp->flags & ~valid_flags))
3594 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3595 (sp->flags & SPLICE_F_FD_IN_FIXED));
3598 req->flags |= REQ_F_NEED_CLEANUP;
3600 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3602 * Splice operation will be punted aync, and here need to
3603 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3605 io_req_init_async(req);
3606 req->work.flags |= IO_WQ_WORK_UNBOUND;
3612 static int io_tee_prep(struct io_kiocb *req,
3613 const struct io_uring_sqe *sqe)
3615 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3617 return __io_splice_prep(req, sqe);
3620 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3622 struct io_splice *sp = &req->splice;
3623 struct file *in = sp->file_in;
3624 struct file *out = sp->file_out;
3625 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3631 ret = do_tee(in, out, sp->len, flags);
3633 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3634 req->flags &= ~REQ_F_NEED_CLEANUP;
3637 req_set_fail_links(req);
3638 io_req_complete(req, ret);
3642 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3644 struct io_splice* sp = &req->splice;
3646 sp->off_in = READ_ONCE(sqe->splice_off_in);
3647 sp->off_out = READ_ONCE(sqe->off);
3648 return __io_splice_prep(req, sqe);
3651 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3653 struct io_splice *sp = &req->splice;
3654 struct file *in = sp->file_in;
3655 struct file *out = sp->file_out;
3656 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3657 loff_t *poff_in, *poff_out;
3663 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3664 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3667 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3669 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3670 req->flags &= ~REQ_F_NEED_CLEANUP;
3673 req_set_fail_links(req);
3674 io_req_complete(req, ret);
3679 * IORING_OP_NOP just posts a completion event, nothing else.
3681 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3683 struct io_ring_ctx *ctx = req->ctx;
3685 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3688 __io_req_complete(req, 0, 0, cs);
3692 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694 struct io_ring_ctx *ctx = req->ctx;
3699 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3701 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3704 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3705 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3708 req->sync.off = READ_ONCE(sqe->off);
3709 req->sync.len = READ_ONCE(sqe->len);
3713 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3715 loff_t end = req->sync.off + req->sync.len;
3718 /* fsync always requires a blocking context */
3722 ret = vfs_fsync_range(req->file, req->sync.off,
3723 end > 0 ? end : LLONG_MAX,
3724 req->sync.flags & IORING_FSYNC_DATASYNC);
3726 req_set_fail_links(req);
3727 io_req_complete(req, ret);
3731 static int io_fallocate_prep(struct io_kiocb *req,
3732 const struct io_uring_sqe *sqe)
3734 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3739 req->sync.off = READ_ONCE(sqe->off);
3740 req->sync.len = READ_ONCE(sqe->addr);
3741 req->sync.mode = READ_ONCE(sqe->len);
3745 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3749 /* fallocate always requiring blocking context */
3752 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3755 req_set_fail_links(req);
3756 io_req_complete(req, ret);
3760 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3762 const char __user *fname;
3765 if (unlikely(sqe->ioprio || sqe->buf_index))
3767 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3770 /* open.how should be already initialised */
3771 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3772 req->open.how.flags |= O_LARGEFILE;
3774 req->open.dfd = READ_ONCE(sqe->fd);
3775 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3776 req->open.filename = getname(fname);
3777 if (IS_ERR(req->open.filename)) {
3778 ret = PTR_ERR(req->open.filename);
3779 req->open.filename = NULL;
3782 req->open.nofile = rlimit(RLIMIT_NOFILE);
3783 req->flags |= REQ_F_NEED_CLEANUP;
3787 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3791 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3793 mode = READ_ONCE(sqe->len);
3794 flags = READ_ONCE(sqe->open_flags);
3795 req->open.how = build_open_how(flags, mode);
3796 return __io_openat_prep(req, sqe);
3799 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3801 struct open_how __user *how;
3805 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3807 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3808 len = READ_ONCE(sqe->len);
3809 if (len < OPEN_HOW_SIZE_VER0)
3812 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3817 return __io_openat_prep(req, sqe);
3820 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3822 struct open_flags op;
3829 ret = build_open_flags(&req->open.how, &op);
3833 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3837 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3840 ret = PTR_ERR(file);
3842 fsnotify_open(file);
3843 fd_install(ret, file);
3846 putname(req->open.filename);
3847 req->flags &= ~REQ_F_NEED_CLEANUP;
3849 req_set_fail_links(req);
3850 io_req_complete(req, ret);
3854 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3856 return io_openat2(req, force_nonblock);
3859 static int io_remove_buffers_prep(struct io_kiocb *req,
3860 const struct io_uring_sqe *sqe)
3862 struct io_provide_buf *p = &req->pbuf;
3865 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3868 tmp = READ_ONCE(sqe->fd);
3869 if (!tmp || tmp > USHRT_MAX)
3872 memset(p, 0, sizeof(*p));
3874 p->bgid = READ_ONCE(sqe->buf_group);
3878 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3879 int bgid, unsigned nbufs)
3883 /* shouldn't happen */
3887 /* the head kbuf is the list itself */
3888 while (!list_empty(&buf->list)) {
3889 struct io_buffer *nxt;
3891 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3892 list_del(&nxt->list);
3899 idr_remove(&ctx->io_buffer_idr, bgid);
3904 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3905 struct io_comp_state *cs)
3907 struct io_provide_buf *p = &req->pbuf;
3908 struct io_ring_ctx *ctx = req->ctx;
3909 struct io_buffer *head;
3912 io_ring_submit_lock(ctx, !force_nonblock);
3914 lockdep_assert_held(&ctx->uring_lock);
3917 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3919 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3921 io_ring_submit_lock(ctx, !force_nonblock);
3923 req_set_fail_links(req);
3924 __io_req_complete(req, ret, 0, cs);
3928 static int io_provide_buffers_prep(struct io_kiocb *req,
3929 const struct io_uring_sqe *sqe)
3931 struct io_provide_buf *p = &req->pbuf;
3934 if (sqe->ioprio || sqe->rw_flags)
3937 tmp = READ_ONCE(sqe->fd);
3938 if (!tmp || tmp > USHRT_MAX)
3941 p->addr = READ_ONCE(sqe->addr);
3942 p->len = READ_ONCE(sqe->len);
3944 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3947 p->bgid = READ_ONCE(sqe->buf_group);
3948 tmp = READ_ONCE(sqe->off);
3949 if (tmp > USHRT_MAX)
3955 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3957 struct io_buffer *buf;
3958 u64 addr = pbuf->addr;
3959 int i, bid = pbuf->bid;
3961 for (i = 0; i < pbuf->nbufs; i++) {
3962 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3967 buf->len = pbuf->len;
3972 INIT_LIST_HEAD(&buf->list);
3975 list_add_tail(&buf->list, &(*head)->list);
3979 return i ? i : -ENOMEM;
3982 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3983 struct io_comp_state *cs)
3985 struct io_provide_buf *p = &req->pbuf;
3986 struct io_ring_ctx *ctx = req->ctx;
3987 struct io_buffer *head, *list;
3990 io_ring_submit_lock(ctx, !force_nonblock);
3992 lockdep_assert_held(&ctx->uring_lock);
3994 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3996 ret = io_add_buffers(p, &head);
4001 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4004 __io_remove_buffers(ctx, head, p->bgid, -1U);
4009 io_ring_submit_unlock(ctx, !force_nonblock);
4011 req_set_fail_links(req);
4012 __io_req_complete(req, ret, 0, cs);
4016 static int io_epoll_ctl_prep(struct io_kiocb *req,
4017 const struct io_uring_sqe *sqe)
4019 #if defined(CONFIG_EPOLL)
4020 if (sqe->ioprio || sqe->buf_index)
4022 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4025 req->epoll.epfd = READ_ONCE(sqe->fd);
4026 req->epoll.op = READ_ONCE(sqe->len);
4027 req->epoll.fd = READ_ONCE(sqe->off);
4029 if (ep_op_has_event(req->epoll.op)) {
4030 struct epoll_event __user *ev;
4032 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4033 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4043 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4044 struct io_comp_state *cs)
4046 #if defined(CONFIG_EPOLL)
4047 struct io_epoll *ie = &req->epoll;
4050 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4051 if (force_nonblock && ret == -EAGAIN)
4055 req_set_fail_links(req);
4056 __io_req_complete(req, ret, 0, cs);
4063 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4065 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4066 if (sqe->ioprio || sqe->buf_index || sqe->off)
4068 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4071 req->madvise.addr = READ_ONCE(sqe->addr);
4072 req->madvise.len = READ_ONCE(sqe->len);
4073 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4080 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4082 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4083 struct io_madvise *ma = &req->madvise;
4089 ret = do_madvise(ma->addr, ma->len, ma->advice);
4091 req_set_fail_links(req);
4092 io_req_complete(req, ret);
4099 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4101 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4103 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4106 req->fadvise.offset = READ_ONCE(sqe->off);
4107 req->fadvise.len = READ_ONCE(sqe->len);
4108 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4112 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4114 struct io_fadvise *fa = &req->fadvise;
4117 if (force_nonblock) {
4118 switch (fa->advice) {
4119 case POSIX_FADV_NORMAL:
4120 case POSIX_FADV_RANDOM:
4121 case POSIX_FADV_SEQUENTIAL:
4128 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4130 req_set_fail_links(req);
4131 io_req_complete(req, ret);
4135 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4139 if (sqe->ioprio || sqe->buf_index)
4141 if (req->flags & REQ_F_FIXED_FILE)
4144 req->statx.dfd = READ_ONCE(sqe->fd);
4145 req->statx.mask = READ_ONCE(sqe->len);
4146 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4147 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4148 req->statx.flags = READ_ONCE(sqe->statx_flags);
4153 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4155 struct io_statx *ctx = &req->statx;
4158 if (force_nonblock) {
4159 /* only need file table for an actual valid fd */
4160 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4161 req->flags |= REQ_F_NO_FILE_TABLE;
4165 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4169 req_set_fail_links(req);
4170 io_req_complete(req, ret);
4174 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4177 * If we queue this for async, it must not be cancellable. That would
4178 * leave the 'file' in an undeterminate state, and here need to modify
4179 * io_wq_work.flags, so initialize io_wq_work firstly.
4181 io_req_init_async(req);
4182 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4184 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4186 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4187 sqe->rw_flags || sqe->buf_index)
4189 if (req->flags & REQ_F_FIXED_FILE)
4192 req->close.fd = READ_ONCE(sqe->fd);
4193 if ((req->file && req->file->f_op == &io_uring_fops))
4196 req->close.put_file = NULL;
4200 static int io_close(struct io_kiocb *req, bool force_nonblock,
4201 struct io_comp_state *cs)
4203 struct io_close *close = &req->close;
4206 /* might be already done during nonblock submission */
4207 if (!close->put_file) {
4208 ret = __close_fd_get_file(close->fd, &close->put_file);
4210 return (ret == -ENOENT) ? -EBADF : ret;
4213 /* if the file has a flush method, be safe and punt to async */
4214 if (close->put_file->f_op->flush && force_nonblock) {
4215 /* was never set, but play safe */
4216 req->flags &= ~REQ_F_NOWAIT;
4217 /* avoid grabbing files - we don't need the files */
4218 req->flags |= REQ_F_NO_FILE_TABLE;
4222 /* No ->flush() or already async, safely close from here */
4223 ret = filp_close(close->put_file, req->work.identity->files);
4225 req_set_fail_links(req);
4226 fput(close->put_file);
4227 close->put_file = NULL;
4228 __io_req_complete(req, ret, 0, cs);
4232 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4234 struct io_ring_ctx *ctx = req->ctx;
4239 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4241 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4244 req->sync.off = READ_ONCE(sqe->off);
4245 req->sync.len = READ_ONCE(sqe->len);
4246 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4250 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4254 /* sync_file_range always requires a blocking context */
4258 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4261 req_set_fail_links(req);
4262 io_req_complete(req, ret);
4266 #if defined(CONFIG_NET)
4267 static int io_setup_async_msg(struct io_kiocb *req,
4268 struct io_async_msghdr *kmsg)
4270 struct io_async_msghdr *async_msg = req->async_data;
4274 if (io_alloc_async_data(req)) {
4275 if (kmsg->iov != kmsg->fast_iov)
4279 async_msg = req->async_data;
4280 req->flags |= REQ_F_NEED_CLEANUP;
4281 memcpy(async_msg, kmsg, sizeof(*kmsg));
4285 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4286 struct io_async_msghdr *iomsg)
4288 iomsg->iov = iomsg->fast_iov;
4289 iomsg->msg.msg_name = &iomsg->addr;
4290 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4291 req->sr_msg.msg_flags, &iomsg->iov);
4294 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4296 struct io_async_msghdr *async_msg = req->async_data;
4297 struct io_sr_msg *sr = &req->sr_msg;
4300 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4303 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4304 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4305 sr->len = READ_ONCE(sqe->len);
4307 #ifdef CONFIG_COMPAT
4308 if (req->ctx->compat)
4309 sr->msg_flags |= MSG_CMSG_COMPAT;
4312 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4314 ret = io_sendmsg_copy_hdr(req, async_msg);
4316 req->flags |= REQ_F_NEED_CLEANUP;
4320 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4321 struct io_comp_state *cs)
4323 struct io_async_msghdr iomsg, *kmsg;
4324 struct socket *sock;
4328 sock = sock_from_file(req->file, &ret);
4329 if (unlikely(!sock))
4332 if (req->async_data) {
4333 kmsg = req->async_data;
4334 kmsg->msg.msg_name = &kmsg->addr;
4335 /* if iov is set, it's allocated already */
4337 kmsg->iov = kmsg->fast_iov;
4338 kmsg->msg.msg_iter.iov = kmsg->iov;
4340 ret = io_sendmsg_copy_hdr(req, &iomsg);
4346 flags = req->sr_msg.msg_flags;
4347 if (flags & MSG_DONTWAIT)
4348 req->flags |= REQ_F_NOWAIT;
4349 else if (force_nonblock)
4350 flags |= MSG_DONTWAIT;
4352 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4353 if (force_nonblock && ret == -EAGAIN)
4354 return io_setup_async_msg(req, kmsg);
4355 if (ret == -ERESTARTSYS)
4358 if (kmsg->iov != kmsg->fast_iov)
4360 req->flags &= ~REQ_F_NEED_CLEANUP;
4362 req_set_fail_links(req);
4363 __io_req_complete(req, ret, 0, cs);
4367 static int io_send(struct io_kiocb *req, bool force_nonblock,
4368 struct io_comp_state *cs)
4370 struct io_sr_msg *sr = &req->sr_msg;
4373 struct socket *sock;
4377 sock = sock_from_file(req->file, &ret);
4378 if (unlikely(!sock))
4381 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4385 msg.msg_name = NULL;
4386 msg.msg_control = NULL;
4387 msg.msg_controllen = 0;
4388 msg.msg_namelen = 0;
4390 flags = req->sr_msg.msg_flags;
4391 if (flags & MSG_DONTWAIT)
4392 req->flags |= REQ_F_NOWAIT;
4393 else if (force_nonblock)
4394 flags |= MSG_DONTWAIT;
4396 msg.msg_flags = flags;
4397 ret = sock_sendmsg(sock, &msg);
4398 if (force_nonblock && ret == -EAGAIN)
4400 if (ret == -ERESTARTSYS)
4404 req_set_fail_links(req);
4405 __io_req_complete(req, ret, 0, cs);
4409 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4410 struct io_async_msghdr *iomsg)
4412 struct io_sr_msg *sr = &req->sr_msg;
4413 struct iovec __user *uiov;
4417 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4418 &iomsg->uaddr, &uiov, &iov_len);
4422 if (req->flags & REQ_F_BUFFER_SELECT) {
4425 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4427 sr->len = iomsg->iov[0].iov_len;
4428 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4432 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4433 &iomsg->iov, &iomsg->msg.msg_iter,
4442 #ifdef CONFIG_COMPAT
4443 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4444 struct io_async_msghdr *iomsg)
4446 struct compat_msghdr __user *msg_compat;
4447 struct io_sr_msg *sr = &req->sr_msg;
4448 struct compat_iovec __user *uiov;
4453 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4454 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4459 uiov = compat_ptr(ptr);
4460 if (req->flags & REQ_F_BUFFER_SELECT) {
4461 compat_ssize_t clen;
4465 if (!access_ok(uiov, sizeof(*uiov)))
4467 if (__get_user(clen, &uiov->iov_len))
4471 sr->len = iomsg->iov[0].iov_len;
4474 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4475 UIO_FASTIOV, &iomsg->iov,
4476 &iomsg->msg.msg_iter, true);
4485 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4486 struct io_async_msghdr *iomsg)
4488 iomsg->msg.msg_name = &iomsg->addr;
4489 iomsg->iov = iomsg->fast_iov;
4491 #ifdef CONFIG_COMPAT
4492 if (req->ctx->compat)
4493 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4496 return __io_recvmsg_copy_hdr(req, iomsg);
4499 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4502 struct io_sr_msg *sr = &req->sr_msg;
4503 struct io_buffer *kbuf;
4505 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4510 req->flags |= REQ_F_BUFFER_SELECTED;
4514 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4516 return io_put_kbuf(req, req->sr_msg.kbuf);
4519 static int io_recvmsg_prep(struct io_kiocb *req,
4520 const struct io_uring_sqe *sqe)
4522 struct io_async_msghdr *async_msg = req->async_data;
4523 struct io_sr_msg *sr = &req->sr_msg;
4526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4529 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4530 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4531 sr->len = READ_ONCE(sqe->len);
4532 sr->bgid = READ_ONCE(sqe->buf_group);
4534 #ifdef CONFIG_COMPAT
4535 if (req->ctx->compat)
4536 sr->msg_flags |= MSG_CMSG_COMPAT;
4539 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4541 ret = io_recvmsg_copy_hdr(req, async_msg);
4543 req->flags |= REQ_F_NEED_CLEANUP;
4547 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4548 struct io_comp_state *cs)
4550 struct io_async_msghdr iomsg, *kmsg;
4551 struct socket *sock;
4552 struct io_buffer *kbuf;
4554 int ret, cflags = 0;
4556 sock = sock_from_file(req->file, &ret);
4557 if (unlikely(!sock))
4560 if (req->async_data) {
4561 kmsg = req->async_data;
4562 kmsg->msg.msg_name = &kmsg->addr;
4563 /* if iov is set, it's allocated already */
4565 kmsg->iov = kmsg->fast_iov;
4566 kmsg->msg.msg_iter.iov = kmsg->iov;
4568 ret = io_recvmsg_copy_hdr(req, &iomsg);
4574 if (req->flags & REQ_F_BUFFER_SELECT) {
4575 kbuf = io_recv_buffer_select(req, !force_nonblock);
4577 return PTR_ERR(kbuf);
4578 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4579 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4580 1, req->sr_msg.len);
4583 flags = req->sr_msg.msg_flags;
4584 if (flags & MSG_DONTWAIT)
4585 req->flags |= REQ_F_NOWAIT;
4586 else if (force_nonblock)
4587 flags |= MSG_DONTWAIT;
4589 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4590 kmsg->uaddr, flags);
4591 if (force_nonblock && ret == -EAGAIN)
4592 return io_setup_async_msg(req, kmsg);
4593 if (ret == -ERESTARTSYS)
4596 if (req->flags & REQ_F_BUFFER_SELECTED)
4597 cflags = io_put_recv_kbuf(req);
4598 if (kmsg->iov != kmsg->fast_iov)
4600 req->flags &= ~REQ_F_NEED_CLEANUP;
4602 req_set_fail_links(req);
4603 __io_req_complete(req, ret, cflags, cs);
4607 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4608 struct io_comp_state *cs)
4610 struct io_buffer *kbuf;
4611 struct io_sr_msg *sr = &req->sr_msg;
4613 void __user *buf = sr->buf;
4614 struct socket *sock;
4617 int ret, cflags = 0;
4619 sock = sock_from_file(req->file, &ret);
4620 if (unlikely(!sock))
4623 if (req->flags & REQ_F_BUFFER_SELECT) {
4624 kbuf = io_recv_buffer_select(req, !force_nonblock);
4626 return PTR_ERR(kbuf);
4627 buf = u64_to_user_ptr(kbuf->addr);
4630 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4634 msg.msg_name = NULL;
4635 msg.msg_control = NULL;
4636 msg.msg_controllen = 0;
4637 msg.msg_namelen = 0;
4638 msg.msg_iocb = NULL;
4641 flags = req->sr_msg.msg_flags;
4642 if (flags & MSG_DONTWAIT)
4643 req->flags |= REQ_F_NOWAIT;
4644 else if (force_nonblock)
4645 flags |= MSG_DONTWAIT;
4647 ret = sock_recvmsg(sock, &msg, flags);
4648 if (force_nonblock && ret == -EAGAIN)
4650 if (ret == -ERESTARTSYS)
4653 if (req->flags & REQ_F_BUFFER_SELECTED)
4654 cflags = io_put_recv_kbuf(req);
4656 req_set_fail_links(req);
4657 __io_req_complete(req, ret, cflags, cs);
4661 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4663 struct io_accept *accept = &req->accept;
4665 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4667 if (sqe->ioprio || sqe->len || sqe->buf_index)
4670 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4671 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4672 accept->flags = READ_ONCE(sqe->accept_flags);
4673 accept->nofile = rlimit(RLIMIT_NOFILE);
4677 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4678 struct io_comp_state *cs)
4680 struct io_accept *accept = &req->accept;
4681 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4684 if (req->file->f_flags & O_NONBLOCK)
4685 req->flags |= REQ_F_NOWAIT;
4687 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4688 accept->addr_len, accept->flags,
4690 if (ret == -EAGAIN && force_nonblock)
4693 if (ret == -ERESTARTSYS)
4695 req_set_fail_links(req);
4697 __io_req_complete(req, ret, 0, cs);
4701 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4703 struct io_connect *conn = &req->connect;
4704 struct io_async_connect *io = req->async_data;
4706 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4708 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4711 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4712 conn->addr_len = READ_ONCE(sqe->addr2);
4717 return move_addr_to_kernel(conn->addr, conn->addr_len,
4721 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4722 struct io_comp_state *cs)
4724 struct io_async_connect __io, *io;
4725 unsigned file_flags;
4728 if (req->async_data) {
4729 io = req->async_data;
4731 ret = move_addr_to_kernel(req->connect.addr,
4732 req->connect.addr_len,
4739 file_flags = force_nonblock ? O_NONBLOCK : 0;
4741 ret = __sys_connect_file(req->file, &io->address,
4742 req->connect.addr_len, file_flags);
4743 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4744 if (req->async_data)
4746 if (io_alloc_async_data(req)) {
4750 io = req->async_data;
4751 memcpy(req->async_data, &__io, sizeof(__io));
4754 if (ret == -ERESTARTSYS)
4758 req_set_fail_links(req);
4759 __io_req_complete(req, ret, 0, cs);
4762 #else /* !CONFIG_NET */
4763 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4768 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4769 struct io_comp_state *cs)
4774 static int io_send(struct io_kiocb *req, bool force_nonblock,
4775 struct io_comp_state *cs)
4780 static int io_recvmsg_prep(struct io_kiocb *req,
4781 const struct io_uring_sqe *sqe)
4786 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4787 struct io_comp_state *cs)
4792 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4793 struct io_comp_state *cs)
4798 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4803 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4804 struct io_comp_state *cs)
4809 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4814 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4815 struct io_comp_state *cs)
4819 #endif /* CONFIG_NET */
4821 struct io_poll_table {
4822 struct poll_table_struct pt;
4823 struct io_kiocb *req;
4827 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4828 __poll_t mask, task_work_func_t func)
4833 /* for instances that support it check for an event match first: */
4834 if (mask && !(mask & poll->events))
4837 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4839 list_del_init(&poll->wait.entry);
4842 init_task_work(&req->task_work, func);
4843 percpu_ref_get(&req->ctx->refs);
4846 * If we using the signalfd wait_queue_head for this wakeup, then
4847 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4848 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4849 * either, as the normal wakeup will suffice.
4851 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4854 * If this fails, then the task is exiting. When a task exits, the
4855 * work gets canceled, so just cancel this request as well instead
4856 * of executing it. We can't safely execute it anyway, as we may not
4857 * have the needed state needed for it anyway.
4859 ret = io_req_task_work_add(req, twa_signal_ok);
4860 if (unlikely(ret)) {
4861 struct task_struct *tsk;
4863 WRITE_ONCE(poll->canceled, true);
4864 tsk = io_wq_get_task(req->ctx->io_wq);
4865 task_work_add(tsk, &req->task_work, 0);
4866 wake_up_process(tsk);
4871 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4872 __acquires(&req->ctx->completion_lock)
4874 struct io_ring_ctx *ctx = req->ctx;
4876 if (!req->result && !READ_ONCE(poll->canceled)) {
4877 struct poll_table_struct pt = { ._key = poll->events };
4879 req->result = vfs_poll(req->file, &pt) & poll->events;
4882 spin_lock_irq(&ctx->completion_lock);
4883 if (!req->result && !READ_ONCE(poll->canceled)) {
4884 add_wait_queue(poll->head, &poll->wait);
4891 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4893 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4894 if (req->opcode == IORING_OP_POLL_ADD)
4895 return req->async_data;
4896 return req->apoll->double_poll;
4899 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4901 if (req->opcode == IORING_OP_POLL_ADD)
4903 return &req->apoll->poll;
4906 static void io_poll_remove_double(struct io_kiocb *req)
4908 struct io_poll_iocb *poll = io_poll_get_double(req);
4910 lockdep_assert_held(&req->ctx->completion_lock);
4912 if (poll && poll->head) {
4913 struct wait_queue_head *head = poll->head;
4915 spin_lock(&head->lock);
4916 list_del_init(&poll->wait.entry);
4917 if (poll->wait.private)
4918 refcount_dec(&req->refs);
4920 spin_unlock(&head->lock);
4924 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4926 struct io_ring_ctx *ctx = req->ctx;
4928 io_poll_remove_double(req);
4929 req->poll.done = true;
4930 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4931 io_commit_cqring(ctx);
4934 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4936 struct io_ring_ctx *ctx = req->ctx;
4938 if (io_poll_rewait(req, &req->poll)) {
4939 spin_unlock_irq(&ctx->completion_lock);
4943 hash_del(&req->hash_node);
4944 io_poll_complete(req, req->result, 0);
4945 spin_unlock_irq(&ctx->completion_lock);
4947 *nxt = io_put_req_find_next(req);
4948 io_cqring_ev_posted(ctx);
4951 static void io_poll_task_func(struct callback_head *cb)
4953 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4954 struct io_ring_ctx *ctx = req->ctx;
4955 struct io_kiocb *nxt = NULL;
4957 io_poll_task_handler(req, &nxt);
4959 __io_req_task_submit(nxt);
4960 percpu_ref_put(&ctx->refs);
4963 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4964 int sync, void *key)
4966 struct io_kiocb *req = wait->private;
4967 struct io_poll_iocb *poll = io_poll_get_single(req);
4968 __poll_t mask = key_to_poll(key);
4970 /* for instances that support it check for an event match first: */
4971 if (mask && !(mask & poll->events))
4974 list_del_init(&wait->entry);
4976 if (poll && poll->head) {
4979 spin_lock(&poll->head->lock);
4980 done = list_empty(&poll->wait.entry);
4982 list_del_init(&poll->wait.entry);
4983 /* make sure double remove sees this as being gone */
4984 wait->private = NULL;
4985 spin_unlock(&poll->head->lock);
4987 __io_async_wake(req, poll, mask, io_poll_task_func);
4989 refcount_dec(&req->refs);
4993 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4994 wait_queue_func_t wake_func)
4998 poll->canceled = false;
4999 poll->events = events;
5000 INIT_LIST_HEAD(&poll->wait.entry);
5001 init_waitqueue_func_entry(&poll->wait, wake_func);
5004 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5005 struct wait_queue_head *head,
5006 struct io_poll_iocb **poll_ptr)
5008 struct io_kiocb *req = pt->req;
5011 * If poll->head is already set, it's because the file being polled
5012 * uses multiple waitqueues for poll handling (eg one for read, one
5013 * for write). Setup a separate io_poll_iocb if this happens.
5015 if (unlikely(poll->head)) {
5016 struct io_poll_iocb *poll_one = poll;
5018 /* already have a 2nd entry, fail a third attempt */
5020 pt->error = -EINVAL;
5023 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5025 pt->error = -ENOMEM;
5028 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5029 refcount_inc(&req->refs);
5030 poll->wait.private = req;
5037 if (poll->events & EPOLLEXCLUSIVE)
5038 add_wait_queue_exclusive(head, &poll->wait);
5040 add_wait_queue(head, &poll->wait);
5043 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5044 struct poll_table_struct *p)
5046 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5047 struct async_poll *apoll = pt->req->apoll;
5049 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5052 static void io_async_task_func(struct callback_head *cb)
5054 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5055 struct async_poll *apoll = req->apoll;
5056 struct io_ring_ctx *ctx = req->ctx;
5058 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5060 if (io_poll_rewait(req, &apoll->poll)) {
5061 spin_unlock_irq(&ctx->completion_lock);
5062 percpu_ref_put(&ctx->refs);
5066 /* If req is still hashed, it cannot have been canceled. Don't check. */
5067 if (hash_hashed(&req->hash_node))
5068 hash_del(&req->hash_node);
5070 io_poll_remove_double(req);
5071 spin_unlock_irq(&ctx->completion_lock);
5073 if (!READ_ONCE(apoll->poll.canceled))
5074 __io_req_task_submit(req);
5076 __io_req_task_cancel(req, -ECANCELED);
5078 percpu_ref_put(&ctx->refs);
5079 kfree(apoll->double_poll);
5083 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5086 struct io_kiocb *req = wait->private;
5087 struct io_poll_iocb *poll = &req->apoll->poll;
5089 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5092 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5095 static void io_poll_req_insert(struct io_kiocb *req)
5097 struct io_ring_ctx *ctx = req->ctx;
5098 struct hlist_head *list;
5100 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5101 hlist_add_head(&req->hash_node, list);
5104 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5105 struct io_poll_iocb *poll,
5106 struct io_poll_table *ipt, __poll_t mask,
5107 wait_queue_func_t wake_func)
5108 __acquires(&ctx->completion_lock)
5110 struct io_ring_ctx *ctx = req->ctx;
5111 bool cancel = false;
5113 io_init_poll_iocb(poll, mask, wake_func);
5114 poll->file = req->file;
5115 poll->wait.private = req;
5117 ipt->pt._key = mask;
5119 ipt->error = -EINVAL;
5121 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5123 spin_lock_irq(&ctx->completion_lock);
5124 if (likely(poll->head)) {
5125 spin_lock(&poll->head->lock);
5126 if (unlikely(list_empty(&poll->wait.entry))) {
5132 if (mask || ipt->error)
5133 list_del_init(&poll->wait.entry);
5135 WRITE_ONCE(poll->canceled, true);
5136 else if (!poll->done) /* actually waiting for an event */
5137 io_poll_req_insert(req);
5138 spin_unlock(&poll->head->lock);
5144 static bool io_arm_poll_handler(struct io_kiocb *req)
5146 const struct io_op_def *def = &io_op_defs[req->opcode];
5147 struct io_ring_ctx *ctx = req->ctx;
5148 struct async_poll *apoll;
5149 struct io_poll_table ipt;
5153 if (!req->file || !file_can_poll(req->file))
5155 if (req->flags & REQ_F_POLLED)
5159 else if (def->pollout)
5163 /* if we can't nonblock try, then no point in arming a poll handler */
5164 if (!io_file_supports_async(req->file, rw))
5167 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5168 if (unlikely(!apoll))
5170 apoll->double_poll = NULL;
5172 req->flags |= REQ_F_POLLED;
5174 INIT_HLIST_NODE(&req->hash_node);
5178 mask |= POLLIN | POLLRDNORM;
5180 mask |= POLLOUT | POLLWRNORM;
5182 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5183 if ((req->opcode == IORING_OP_RECVMSG) &&
5184 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5187 mask |= POLLERR | POLLPRI;
5189 ipt.pt._qproc = io_async_queue_proc;
5191 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5193 if (ret || ipt.error) {
5194 io_poll_remove_double(req);
5195 spin_unlock_irq(&ctx->completion_lock);
5196 kfree(apoll->double_poll);
5200 spin_unlock_irq(&ctx->completion_lock);
5201 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5202 apoll->poll.events);
5206 static bool __io_poll_remove_one(struct io_kiocb *req,
5207 struct io_poll_iocb *poll)
5209 bool do_complete = false;
5211 spin_lock(&poll->head->lock);
5212 WRITE_ONCE(poll->canceled, true);
5213 if (!list_empty(&poll->wait.entry)) {
5214 list_del_init(&poll->wait.entry);
5217 spin_unlock(&poll->head->lock);
5218 hash_del(&req->hash_node);
5222 static bool io_poll_remove_one(struct io_kiocb *req)
5226 io_poll_remove_double(req);
5228 if (req->opcode == IORING_OP_POLL_ADD) {
5229 do_complete = __io_poll_remove_one(req, &req->poll);
5231 struct async_poll *apoll = req->apoll;
5233 /* non-poll requests have submit ref still */
5234 do_complete = __io_poll_remove_one(req, &apoll->poll);
5237 kfree(apoll->double_poll);
5243 io_cqring_fill_event(req, -ECANCELED);
5244 io_commit_cqring(req->ctx);
5245 req_set_fail_links(req);
5246 io_put_req_deferred(req, 1);
5253 * Returns true if we found and killed one or more poll requests
5255 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5257 struct hlist_node *tmp;
5258 struct io_kiocb *req;
5261 spin_lock_irq(&ctx->completion_lock);
5262 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5263 struct hlist_head *list;
5265 list = &ctx->cancel_hash[i];
5266 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5267 if (io_task_match(req, tsk))
5268 posted += io_poll_remove_one(req);
5271 spin_unlock_irq(&ctx->completion_lock);
5274 io_cqring_ev_posted(ctx);
5279 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5281 struct hlist_head *list;
5282 struct io_kiocb *req;
5284 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5285 hlist_for_each_entry(req, list, hash_node) {
5286 if (sqe_addr != req->user_data)
5288 if (io_poll_remove_one(req))
5296 static int io_poll_remove_prep(struct io_kiocb *req,
5297 const struct io_uring_sqe *sqe)
5299 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5301 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5305 req->poll.addr = READ_ONCE(sqe->addr);
5310 * Find a running poll command that matches one specified in sqe->addr,
5311 * and remove it if found.
5313 static int io_poll_remove(struct io_kiocb *req)
5315 struct io_ring_ctx *ctx = req->ctx;
5319 addr = req->poll.addr;
5320 spin_lock_irq(&ctx->completion_lock);
5321 ret = io_poll_cancel(ctx, addr);
5322 spin_unlock_irq(&ctx->completion_lock);
5325 req_set_fail_links(req);
5326 io_req_complete(req, ret);
5330 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5333 struct io_kiocb *req = wait->private;
5334 struct io_poll_iocb *poll = &req->poll;
5336 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5339 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5340 struct poll_table_struct *p)
5342 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5344 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5347 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5349 struct io_poll_iocb *poll = &req->poll;
5352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5354 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5359 events = READ_ONCE(sqe->poll32_events);
5361 events = swahw32(events);
5363 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5364 (events & EPOLLEXCLUSIVE);
5368 static int io_poll_add(struct io_kiocb *req)
5370 struct io_poll_iocb *poll = &req->poll;
5371 struct io_ring_ctx *ctx = req->ctx;
5372 struct io_poll_table ipt;
5375 INIT_HLIST_NODE(&req->hash_node);
5376 ipt.pt._qproc = io_poll_queue_proc;
5378 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5381 if (mask) { /* no async, we'd stolen it */
5383 io_poll_complete(req, mask, 0);
5385 spin_unlock_irq(&ctx->completion_lock);
5388 io_cqring_ev_posted(ctx);
5394 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5396 struct io_timeout_data *data = container_of(timer,
5397 struct io_timeout_data, timer);
5398 struct io_kiocb *req = data->req;
5399 struct io_ring_ctx *ctx = req->ctx;
5400 unsigned long flags;
5402 spin_lock_irqsave(&ctx->completion_lock, flags);
5403 list_del_init(&req->timeout.list);
5404 atomic_set(&req->ctx->cq_timeouts,
5405 atomic_read(&req->ctx->cq_timeouts) + 1);
5407 io_cqring_fill_event(req, -ETIME);
5408 io_commit_cqring(ctx);
5409 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5411 io_cqring_ev_posted(ctx);
5412 req_set_fail_links(req);
5414 return HRTIMER_NORESTART;
5417 static int __io_timeout_cancel(struct io_kiocb *req)
5419 struct io_timeout_data *io = req->async_data;
5422 ret = hrtimer_try_to_cancel(&io->timer);
5425 list_del_init(&req->timeout.list);
5427 req_set_fail_links(req);
5428 io_cqring_fill_event(req, -ECANCELED);
5429 io_put_req_deferred(req, 1);
5433 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5435 struct io_kiocb *req;
5438 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5439 if (user_data == req->user_data) {
5448 return __io_timeout_cancel(req);
5451 static int io_timeout_remove_prep(struct io_kiocb *req,
5452 const struct io_uring_sqe *sqe)
5454 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5456 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5458 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5461 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5466 * Remove or update an existing timeout command
5468 static int io_timeout_remove(struct io_kiocb *req)
5470 struct io_ring_ctx *ctx = req->ctx;
5473 spin_lock_irq(&ctx->completion_lock);
5474 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5476 io_cqring_fill_event(req, ret);
5477 io_commit_cqring(ctx);
5478 spin_unlock_irq(&ctx->completion_lock);
5479 io_cqring_ev_posted(ctx);
5481 req_set_fail_links(req);
5486 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5487 bool is_timeout_link)
5489 struct io_timeout_data *data;
5491 u32 off = READ_ONCE(sqe->off);
5493 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5495 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5497 if (off && is_timeout_link)
5499 flags = READ_ONCE(sqe->timeout_flags);
5500 if (flags & ~IORING_TIMEOUT_ABS)
5503 req->timeout.off = off;
5505 if (!req->async_data && io_alloc_async_data(req))
5508 data = req->async_data;
5511 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5514 if (flags & IORING_TIMEOUT_ABS)
5515 data->mode = HRTIMER_MODE_ABS;
5517 data->mode = HRTIMER_MODE_REL;
5519 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5523 static int io_timeout(struct io_kiocb *req)
5525 struct io_ring_ctx *ctx = req->ctx;
5526 struct io_timeout_data *data = req->async_data;
5527 struct list_head *entry;
5528 u32 tail, off = req->timeout.off;
5530 spin_lock_irq(&ctx->completion_lock);
5533 * sqe->off holds how many events that need to occur for this
5534 * timeout event to be satisfied. If it isn't set, then this is
5535 * a pure timeout request, sequence isn't used.
5537 if (io_is_timeout_noseq(req)) {
5538 entry = ctx->timeout_list.prev;
5542 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5543 req->timeout.target_seq = tail + off;
5546 * Insertion sort, ensuring the first entry in the list is always
5547 * the one we need first.
5549 list_for_each_prev(entry, &ctx->timeout_list) {
5550 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5553 if (io_is_timeout_noseq(nxt))
5555 /* nxt.seq is behind @tail, otherwise would've been completed */
5556 if (off >= nxt->timeout.target_seq - tail)
5560 list_add(&req->timeout.list, entry);
5561 data->timer.function = io_timeout_fn;
5562 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5563 spin_unlock_irq(&ctx->completion_lock);
5567 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5569 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5571 return req->user_data == (unsigned long) data;
5574 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5576 enum io_wq_cancel cancel_ret;
5579 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5580 switch (cancel_ret) {
5581 case IO_WQ_CANCEL_OK:
5584 case IO_WQ_CANCEL_RUNNING:
5587 case IO_WQ_CANCEL_NOTFOUND:
5595 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5596 struct io_kiocb *req, __u64 sqe_addr,
5599 unsigned long flags;
5602 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5603 if (ret != -ENOENT) {
5604 spin_lock_irqsave(&ctx->completion_lock, flags);
5608 spin_lock_irqsave(&ctx->completion_lock, flags);
5609 ret = io_timeout_cancel(ctx, sqe_addr);
5612 ret = io_poll_cancel(ctx, sqe_addr);
5616 io_cqring_fill_event(req, ret);
5617 io_commit_cqring(ctx);
5618 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5619 io_cqring_ev_posted(ctx);
5622 req_set_fail_links(req);
5626 static int io_async_cancel_prep(struct io_kiocb *req,
5627 const struct io_uring_sqe *sqe)
5629 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5631 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5633 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5636 req->cancel.addr = READ_ONCE(sqe->addr);
5640 static int io_async_cancel(struct io_kiocb *req)
5642 struct io_ring_ctx *ctx = req->ctx;
5644 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5648 static int io_files_update_prep(struct io_kiocb *req,
5649 const struct io_uring_sqe *sqe)
5651 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5653 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5655 if (sqe->ioprio || sqe->rw_flags)
5658 req->files_update.offset = READ_ONCE(sqe->off);
5659 req->files_update.nr_args = READ_ONCE(sqe->len);
5660 if (!req->files_update.nr_args)
5662 req->files_update.arg = READ_ONCE(sqe->addr);
5666 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5667 struct io_comp_state *cs)
5669 struct io_ring_ctx *ctx = req->ctx;
5670 struct io_uring_files_update up;
5676 up.offset = req->files_update.offset;
5677 up.fds = req->files_update.arg;
5679 mutex_lock(&ctx->uring_lock);
5680 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5681 mutex_unlock(&ctx->uring_lock);
5684 req_set_fail_links(req);
5685 __io_req_complete(req, ret, 0, cs);
5689 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5691 switch (req->opcode) {
5694 case IORING_OP_READV:
5695 case IORING_OP_READ_FIXED:
5696 case IORING_OP_READ:
5697 return io_read_prep(req, sqe);
5698 case IORING_OP_WRITEV:
5699 case IORING_OP_WRITE_FIXED:
5700 case IORING_OP_WRITE:
5701 return io_write_prep(req, sqe);
5702 case IORING_OP_POLL_ADD:
5703 return io_poll_add_prep(req, sqe);
5704 case IORING_OP_POLL_REMOVE:
5705 return io_poll_remove_prep(req, sqe);
5706 case IORING_OP_FSYNC:
5707 return io_prep_fsync(req, sqe);
5708 case IORING_OP_SYNC_FILE_RANGE:
5709 return io_prep_sfr(req, sqe);
5710 case IORING_OP_SENDMSG:
5711 case IORING_OP_SEND:
5712 return io_sendmsg_prep(req, sqe);
5713 case IORING_OP_RECVMSG:
5714 case IORING_OP_RECV:
5715 return io_recvmsg_prep(req, sqe);
5716 case IORING_OP_CONNECT:
5717 return io_connect_prep(req, sqe);
5718 case IORING_OP_TIMEOUT:
5719 return io_timeout_prep(req, sqe, false);
5720 case IORING_OP_TIMEOUT_REMOVE:
5721 return io_timeout_remove_prep(req, sqe);
5722 case IORING_OP_ASYNC_CANCEL:
5723 return io_async_cancel_prep(req, sqe);
5724 case IORING_OP_LINK_TIMEOUT:
5725 return io_timeout_prep(req, sqe, true);
5726 case IORING_OP_ACCEPT:
5727 return io_accept_prep(req, sqe);
5728 case IORING_OP_FALLOCATE:
5729 return io_fallocate_prep(req, sqe);
5730 case IORING_OP_OPENAT:
5731 return io_openat_prep(req, sqe);
5732 case IORING_OP_CLOSE:
5733 return io_close_prep(req, sqe);
5734 case IORING_OP_FILES_UPDATE:
5735 return io_files_update_prep(req, sqe);
5736 case IORING_OP_STATX:
5737 return io_statx_prep(req, sqe);
5738 case IORING_OP_FADVISE:
5739 return io_fadvise_prep(req, sqe);
5740 case IORING_OP_MADVISE:
5741 return io_madvise_prep(req, sqe);
5742 case IORING_OP_OPENAT2:
5743 return io_openat2_prep(req, sqe);
5744 case IORING_OP_EPOLL_CTL:
5745 return io_epoll_ctl_prep(req, sqe);
5746 case IORING_OP_SPLICE:
5747 return io_splice_prep(req, sqe);
5748 case IORING_OP_PROVIDE_BUFFERS:
5749 return io_provide_buffers_prep(req, sqe);
5750 case IORING_OP_REMOVE_BUFFERS:
5751 return io_remove_buffers_prep(req, sqe);
5753 return io_tee_prep(req, sqe);
5756 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5761 static int io_req_defer_prep(struct io_kiocb *req,
5762 const struct io_uring_sqe *sqe)
5766 if (io_alloc_async_data(req))
5768 return io_req_prep(req, sqe);
5771 static u32 io_get_sequence(struct io_kiocb *req)
5773 struct io_kiocb *pos;
5774 struct io_ring_ctx *ctx = req->ctx;
5775 u32 total_submitted, nr_reqs = 1;
5777 if (req->flags & REQ_F_LINK_HEAD)
5778 list_for_each_entry(pos, &req->link_list, link_list)
5781 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5782 return total_submitted - nr_reqs;
5785 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5787 struct io_ring_ctx *ctx = req->ctx;
5788 struct io_defer_entry *de;
5792 /* Still need defer if there is pending req in defer list. */
5793 if (likely(list_empty_careful(&ctx->defer_list) &&
5794 !(req->flags & REQ_F_IO_DRAIN)))
5797 seq = io_get_sequence(req);
5798 /* Still a chance to pass the sequence check */
5799 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5802 if (!req->async_data) {
5803 ret = io_req_defer_prep(req, sqe);
5807 io_prep_async_link(req);
5808 de = kmalloc(sizeof(*de), GFP_KERNEL);
5812 spin_lock_irq(&ctx->completion_lock);
5813 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5814 spin_unlock_irq(&ctx->completion_lock);
5816 io_queue_async_work(req);
5817 return -EIOCBQUEUED;
5820 trace_io_uring_defer(ctx, req, req->user_data);
5823 list_add_tail(&de->list, &ctx->defer_list);
5824 spin_unlock_irq(&ctx->completion_lock);
5825 return -EIOCBQUEUED;
5828 static void io_req_drop_files(struct io_kiocb *req)
5830 struct io_ring_ctx *ctx = req->ctx;
5831 unsigned long flags;
5833 spin_lock_irqsave(&ctx->inflight_lock, flags);
5834 list_del(&req->inflight_entry);
5835 if (waitqueue_active(&ctx->inflight_wait))
5836 wake_up(&ctx->inflight_wait);
5837 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5838 req->flags &= ~REQ_F_INFLIGHT;
5839 put_files_struct(req->work.identity->files);
5840 put_nsproxy(req->work.identity->nsproxy);
5841 req->work.flags &= ~IO_WQ_WORK_FILES;
5844 static void __io_clean_op(struct io_kiocb *req)
5846 if (req->flags & REQ_F_BUFFER_SELECTED) {
5847 switch (req->opcode) {
5848 case IORING_OP_READV:
5849 case IORING_OP_READ_FIXED:
5850 case IORING_OP_READ:
5851 kfree((void *)(unsigned long)req->rw.addr);
5853 case IORING_OP_RECVMSG:
5854 case IORING_OP_RECV:
5855 kfree(req->sr_msg.kbuf);
5858 req->flags &= ~REQ_F_BUFFER_SELECTED;
5861 if (req->flags & REQ_F_NEED_CLEANUP) {
5862 switch (req->opcode) {
5863 case IORING_OP_READV:
5864 case IORING_OP_READ_FIXED:
5865 case IORING_OP_READ:
5866 case IORING_OP_WRITEV:
5867 case IORING_OP_WRITE_FIXED:
5868 case IORING_OP_WRITE: {
5869 struct io_async_rw *io = req->async_data;
5871 kfree(io->free_iovec);
5874 case IORING_OP_RECVMSG:
5875 case IORING_OP_SENDMSG: {
5876 struct io_async_msghdr *io = req->async_data;
5877 if (io->iov != io->fast_iov)
5881 case IORING_OP_SPLICE:
5883 io_put_file(req, req->splice.file_in,
5884 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5886 case IORING_OP_OPENAT:
5887 case IORING_OP_OPENAT2:
5888 if (req->open.filename)
5889 putname(req->open.filename);
5892 req->flags &= ~REQ_F_NEED_CLEANUP;
5895 if (req->flags & REQ_F_INFLIGHT)
5896 io_req_drop_files(req);
5899 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5900 struct io_comp_state *cs)
5902 struct io_ring_ctx *ctx = req->ctx;
5905 switch (req->opcode) {
5907 ret = io_nop(req, cs);
5909 case IORING_OP_READV:
5910 case IORING_OP_READ_FIXED:
5911 case IORING_OP_READ:
5912 ret = io_read(req, force_nonblock, cs);
5914 case IORING_OP_WRITEV:
5915 case IORING_OP_WRITE_FIXED:
5916 case IORING_OP_WRITE:
5917 ret = io_write(req, force_nonblock, cs);
5919 case IORING_OP_FSYNC:
5920 ret = io_fsync(req, force_nonblock);
5922 case IORING_OP_POLL_ADD:
5923 ret = io_poll_add(req);
5925 case IORING_OP_POLL_REMOVE:
5926 ret = io_poll_remove(req);
5928 case IORING_OP_SYNC_FILE_RANGE:
5929 ret = io_sync_file_range(req, force_nonblock);
5931 case IORING_OP_SENDMSG:
5932 ret = io_sendmsg(req, force_nonblock, cs);
5934 case IORING_OP_SEND:
5935 ret = io_send(req, force_nonblock, cs);
5937 case IORING_OP_RECVMSG:
5938 ret = io_recvmsg(req, force_nonblock, cs);
5940 case IORING_OP_RECV:
5941 ret = io_recv(req, force_nonblock, cs);
5943 case IORING_OP_TIMEOUT:
5944 ret = io_timeout(req);
5946 case IORING_OP_TIMEOUT_REMOVE:
5947 ret = io_timeout_remove(req);
5949 case IORING_OP_ACCEPT:
5950 ret = io_accept(req, force_nonblock, cs);
5952 case IORING_OP_CONNECT:
5953 ret = io_connect(req, force_nonblock, cs);
5955 case IORING_OP_ASYNC_CANCEL:
5956 ret = io_async_cancel(req);
5958 case IORING_OP_FALLOCATE:
5959 ret = io_fallocate(req, force_nonblock);
5961 case IORING_OP_OPENAT:
5962 ret = io_openat(req, force_nonblock);
5964 case IORING_OP_CLOSE:
5965 ret = io_close(req, force_nonblock, cs);
5967 case IORING_OP_FILES_UPDATE:
5968 ret = io_files_update(req, force_nonblock, cs);
5970 case IORING_OP_STATX:
5971 ret = io_statx(req, force_nonblock);
5973 case IORING_OP_FADVISE:
5974 ret = io_fadvise(req, force_nonblock);
5976 case IORING_OP_MADVISE:
5977 ret = io_madvise(req, force_nonblock);
5979 case IORING_OP_OPENAT2:
5980 ret = io_openat2(req, force_nonblock);
5982 case IORING_OP_EPOLL_CTL:
5983 ret = io_epoll_ctl(req, force_nonblock, cs);
5985 case IORING_OP_SPLICE:
5986 ret = io_splice(req, force_nonblock);
5988 case IORING_OP_PROVIDE_BUFFERS:
5989 ret = io_provide_buffers(req, force_nonblock, cs);
5991 case IORING_OP_REMOVE_BUFFERS:
5992 ret = io_remove_buffers(req, force_nonblock, cs);
5995 ret = io_tee(req, force_nonblock);
6005 /* If the op doesn't have a file, we're not polling for it */
6006 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6007 const bool in_async = io_wq_current_is_worker();
6009 /* workqueue context doesn't hold uring_lock, grab it now */
6011 mutex_lock(&ctx->uring_lock);
6013 io_iopoll_req_issued(req);
6016 mutex_unlock(&ctx->uring_lock);
6022 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6024 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6025 struct io_kiocb *timeout;
6028 timeout = io_prep_linked_timeout(req);
6030 io_queue_linked_timeout(timeout);
6032 /* if NO_CANCEL is set, we must still run the work */
6033 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6034 IO_WQ_WORK_CANCEL) {
6040 ret = io_issue_sqe(req, false, NULL);
6042 * We can get EAGAIN for polled IO even though we're
6043 * forcing a sync submission from here, since we can't
6044 * wait for request slots on the block side.
6053 req_set_fail_links(req);
6054 io_req_complete(req, ret);
6057 return io_steal_work(req);
6060 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6063 struct fixed_file_table *table;
6065 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6066 return table->files[index & IORING_FILE_TABLE_MASK];
6069 static struct file *io_file_get(struct io_submit_state *state,
6070 struct io_kiocb *req, int fd, bool fixed)
6072 struct io_ring_ctx *ctx = req->ctx;
6076 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6078 fd = array_index_nospec(fd, ctx->nr_user_files);
6079 file = io_file_from_index(ctx, fd);
6081 req->fixed_file_refs = &ctx->file_data->node->refs;
6082 percpu_ref_get(req->fixed_file_refs);
6085 trace_io_uring_file_get(ctx, fd);
6086 file = __io_file_get(state, fd);
6092 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6097 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6098 if (unlikely(!fixed && io_async_submit(req->ctx)))
6101 req->file = io_file_get(state, req, fd, fixed);
6102 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6107 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6109 struct io_timeout_data *data = container_of(timer,
6110 struct io_timeout_data, timer);
6111 struct io_kiocb *req = data->req;
6112 struct io_ring_ctx *ctx = req->ctx;
6113 struct io_kiocb *prev = NULL;
6114 unsigned long flags;
6116 spin_lock_irqsave(&ctx->completion_lock, flags);
6119 * We don't expect the list to be empty, that will only happen if we
6120 * race with the completion of the linked work.
6122 if (!list_empty(&req->link_list)) {
6123 prev = list_entry(req->link_list.prev, struct io_kiocb,
6125 if (refcount_inc_not_zero(&prev->refs)) {
6126 list_del_init(&req->link_list);
6127 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6132 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6135 req_set_fail_links(prev);
6136 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6139 io_req_complete(req, -ETIME);
6141 return HRTIMER_NORESTART;
6144 static void __io_queue_linked_timeout(struct io_kiocb *req)
6147 * If the list is now empty, then our linked request finished before
6148 * we got a chance to setup the timer
6150 if (!list_empty(&req->link_list)) {
6151 struct io_timeout_data *data = req->async_data;
6153 data->timer.function = io_link_timeout_fn;
6154 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6159 static void io_queue_linked_timeout(struct io_kiocb *req)
6161 struct io_ring_ctx *ctx = req->ctx;
6163 spin_lock_irq(&ctx->completion_lock);
6164 __io_queue_linked_timeout(req);
6165 spin_unlock_irq(&ctx->completion_lock);
6167 /* drop submission reference */
6171 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6173 struct io_kiocb *nxt;
6175 if (!(req->flags & REQ_F_LINK_HEAD))
6177 if (req->flags & REQ_F_LINK_TIMEOUT)
6180 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6182 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6185 req->flags |= REQ_F_LINK_TIMEOUT;
6189 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6191 struct io_kiocb *linked_timeout;
6192 struct io_kiocb *nxt;
6193 const struct cred *old_creds = NULL;
6197 linked_timeout = io_prep_linked_timeout(req);
6199 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6200 (req->work.flags & IO_WQ_WORK_CREDS) &&
6201 req->work.identity->creds != current_cred()) {
6203 revert_creds(old_creds);
6204 if (old_creds == req->work.identity->creds)
6205 old_creds = NULL; /* restored original creds */
6207 old_creds = override_creds(req->work.identity->creds);
6210 ret = io_issue_sqe(req, true, cs);
6213 * We async punt it if the file wasn't marked NOWAIT, or if the file
6214 * doesn't support non-blocking read/write attempts
6216 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6217 if (!io_arm_poll_handler(req)) {
6220 * Queued up for async execution, worker will release
6221 * submit reference when the iocb is actually submitted.
6223 io_queue_async_work(req);
6227 io_queue_linked_timeout(linked_timeout);
6231 if (unlikely(ret)) {
6232 /* un-prep timeout, so it'll be killed as any other linked */
6233 req->flags &= ~REQ_F_LINK_TIMEOUT;
6234 req_set_fail_links(req);
6236 io_req_complete(req, ret);
6240 /* drop submission reference */
6241 nxt = io_put_req_find_next(req);
6243 io_queue_linked_timeout(linked_timeout);
6248 if (req->flags & REQ_F_FORCE_ASYNC)
6254 revert_creds(old_creds);
6257 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6258 struct io_comp_state *cs)
6262 ret = io_req_defer(req, sqe);
6264 if (ret != -EIOCBQUEUED) {
6266 req_set_fail_links(req);
6268 io_req_complete(req, ret);
6270 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6271 if (!req->async_data) {
6272 ret = io_req_defer_prep(req, sqe);
6278 * Never try inline submit of IOSQE_ASYNC is set, go straight
6279 * to async execution.
6281 io_req_init_async(req);
6282 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6283 io_queue_async_work(req);
6286 ret = io_req_prep(req, sqe);
6290 __io_queue_sqe(req, cs);
6294 static inline void io_queue_link_head(struct io_kiocb *req,
6295 struct io_comp_state *cs)
6297 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6299 io_req_complete(req, -ECANCELED);
6301 io_queue_sqe(req, NULL, cs);
6304 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6305 struct io_kiocb **link, struct io_comp_state *cs)
6307 struct io_ring_ctx *ctx = req->ctx;
6311 * If we already have a head request, queue this one for async
6312 * submittal once the head completes. If we don't have a head but
6313 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6314 * submitted sync once the chain is complete. If none of those
6315 * conditions are true (normal request), then just queue it.
6318 struct io_kiocb *head = *link;
6321 * Taking sequential execution of a link, draining both sides
6322 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6323 * requests in the link. So, it drains the head and the
6324 * next after the link request. The last one is done via
6325 * drain_next flag to persist the effect across calls.
6327 if (req->flags & REQ_F_IO_DRAIN) {
6328 head->flags |= REQ_F_IO_DRAIN;
6329 ctx->drain_next = 1;
6331 ret = io_req_defer_prep(req, sqe);
6332 if (unlikely(ret)) {
6333 /* fail even hard links since we don't submit */
6334 head->flags |= REQ_F_FAIL_LINK;
6337 trace_io_uring_link(ctx, req, head);
6338 list_add_tail(&req->link_list, &head->link_list);
6340 /* last request of a link, enqueue the link */
6341 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6342 io_queue_link_head(head, cs);
6346 if (unlikely(ctx->drain_next)) {
6347 req->flags |= REQ_F_IO_DRAIN;
6348 ctx->drain_next = 0;
6350 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6351 req->flags |= REQ_F_LINK_HEAD;
6352 INIT_LIST_HEAD(&req->link_list);
6354 ret = io_req_defer_prep(req, sqe);
6356 req->flags |= REQ_F_FAIL_LINK;
6359 io_queue_sqe(req, sqe, cs);
6367 * Batched submission is done, ensure local IO is flushed out.
6369 static void io_submit_state_end(struct io_submit_state *state)
6371 if (!list_empty(&state->comp.list))
6372 io_submit_flush_completions(&state->comp);
6373 blk_finish_plug(&state->plug);
6374 io_state_file_put(state);
6375 if (state->free_reqs)
6376 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6380 * Start submission side cache.
6382 static void io_submit_state_start(struct io_submit_state *state,
6383 struct io_ring_ctx *ctx, unsigned int max_ios)
6385 blk_start_plug(&state->plug);
6387 INIT_LIST_HEAD(&state->comp.list);
6388 state->comp.ctx = ctx;
6389 state->free_reqs = 0;
6391 state->ios_left = max_ios;
6394 static void io_commit_sqring(struct io_ring_ctx *ctx)
6396 struct io_rings *rings = ctx->rings;
6399 * Ensure any loads from the SQEs are done at this point,
6400 * since once we write the new head, the application could
6401 * write new data to them.
6403 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6407 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6408 * that is mapped by userspace. This means that care needs to be taken to
6409 * ensure that reads are stable, as we cannot rely on userspace always
6410 * being a good citizen. If members of the sqe are validated and then later
6411 * used, it's important that those reads are done through READ_ONCE() to
6412 * prevent a re-load down the line.
6414 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6416 u32 *sq_array = ctx->sq_array;
6420 * The cached sq head (or cq tail) serves two purposes:
6422 * 1) allows us to batch the cost of updating the user visible
6424 * 2) allows the kernel side to track the head on its own, even
6425 * though the application is the one updating it.
6427 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6428 if (likely(head < ctx->sq_entries))
6429 return &ctx->sq_sqes[head];
6431 /* drop invalid entries */
6432 ctx->cached_sq_dropped++;
6433 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6437 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6439 ctx->cached_sq_head++;
6443 * Check SQE restrictions (opcode and flags).
6445 * Returns 'true' if SQE is allowed, 'false' otherwise.
6447 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6448 struct io_kiocb *req,
6449 unsigned int sqe_flags)
6451 if (!ctx->restricted)
6454 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6457 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6458 ctx->restrictions.sqe_flags_required)
6461 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6462 ctx->restrictions.sqe_flags_required))
6468 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6469 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6470 IOSQE_BUFFER_SELECT)
6472 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6473 const struct io_uring_sqe *sqe,
6474 struct io_submit_state *state)
6476 unsigned int sqe_flags;
6479 req->opcode = READ_ONCE(sqe->opcode);
6480 req->user_data = READ_ONCE(sqe->user_data);
6481 req->async_data = NULL;
6485 /* one is dropped after submission, the other at completion */
6486 refcount_set(&req->refs, 2);
6487 req->task = current;
6490 if (unlikely(req->opcode >= IORING_OP_LAST))
6493 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6496 sqe_flags = READ_ONCE(sqe->flags);
6497 /* enforce forwards compatibility on users */
6498 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6501 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6504 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6505 !io_op_defs[req->opcode].buffer_select)
6508 id = READ_ONCE(sqe->personality);
6510 struct io_identity *iod;
6512 iod = idr_find(&ctx->personality_idr, id);
6515 refcount_inc(&iod->count);
6517 __io_req_init_async(req);
6518 get_cred(iod->creds);
6519 req->work.identity = iod;
6520 req->work.flags |= IO_WQ_WORK_CREDS;
6523 /* same numerical values with corresponding REQ_F_*, safe to copy */
6524 req->flags |= sqe_flags;
6526 if (!io_op_defs[req->opcode].needs_file)
6529 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6534 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6536 struct io_submit_state state;
6537 struct io_kiocb *link = NULL;
6538 int i, submitted = 0;
6540 /* if we have a backlog and couldn't flush it all, return BUSY */
6541 if (test_bit(0, &ctx->sq_check_overflow)) {
6542 if (!list_empty(&ctx->cq_overflow_list) &&
6543 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6547 /* make sure SQ entry isn't read before tail */
6548 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6550 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6553 percpu_counter_add(¤t->io_uring->inflight, nr);
6554 refcount_add(nr, ¤t->usage);
6556 io_submit_state_start(&state, ctx, nr);
6558 for (i = 0; i < nr; i++) {
6559 const struct io_uring_sqe *sqe;
6560 struct io_kiocb *req;
6563 sqe = io_get_sqe(ctx);
6564 if (unlikely(!sqe)) {
6565 io_consume_sqe(ctx);
6568 req = io_alloc_req(ctx, &state);
6569 if (unlikely(!req)) {
6571 submitted = -EAGAIN;
6574 io_consume_sqe(ctx);
6575 /* will complete beyond this point, count as submitted */
6578 err = io_init_req(ctx, req, sqe, &state);
6579 if (unlikely(err)) {
6582 io_req_complete(req, err);
6586 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6587 true, io_async_submit(ctx));
6588 err = io_submit_sqe(req, sqe, &link, &state.comp);
6593 if (unlikely(submitted != nr)) {
6594 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6595 struct io_uring_task *tctx = current->io_uring;
6596 int unused = nr - ref_used;
6598 percpu_ref_put_many(&ctx->refs, unused);
6599 percpu_counter_sub(&tctx->inflight, unused);
6600 put_task_struct_many(current, unused);
6603 io_queue_link_head(link, &state.comp);
6604 io_submit_state_end(&state);
6606 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6607 io_commit_sqring(ctx);
6612 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6614 /* Tell userspace we may need a wakeup call */
6615 spin_lock_irq(&ctx->completion_lock);
6616 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6617 spin_unlock_irq(&ctx->completion_lock);
6620 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6622 spin_lock_irq(&ctx->completion_lock);
6623 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6624 spin_unlock_irq(&ctx->completion_lock);
6627 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6628 int sync, void *key)
6630 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6633 ret = autoremove_wake_function(wqe, mode, sync, key);
6635 unsigned long flags;
6637 spin_lock_irqsave(&ctx->completion_lock, flags);
6638 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6639 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6650 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6651 unsigned long start_jiffies, bool cap_entries)
6653 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6654 struct io_sq_data *sqd = ctx->sq_data;
6655 unsigned int to_submit;
6659 if (!list_empty(&ctx->iopoll_list)) {
6660 unsigned nr_events = 0;
6662 mutex_lock(&ctx->uring_lock);
6663 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6664 io_do_iopoll(ctx, &nr_events, 0);
6665 mutex_unlock(&ctx->uring_lock);
6668 to_submit = io_sqring_entries(ctx);
6671 * If submit got -EBUSY, flag us as needing the application
6672 * to enter the kernel to reap and flush events.
6674 if (!to_submit || ret == -EBUSY || need_resched()) {
6676 * Drop cur_mm before scheduling, we can't hold it for
6677 * long periods (or over schedule()). Do this before
6678 * adding ourselves to the waitqueue, as the unuse/drop
6681 io_sq_thread_drop_mm();
6684 * We're polling. If we're within the defined idle
6685 * period, then let us spin without work before going
6686 * to sleep. The exception is if we got EBUSY doing
6687 * more IO, we should wait for the application to
6688 * reap events and wake us up.
6690 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6691 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6692 !percpu_ref_is_dying(&ctx->refs)))
6695 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6696 TASK_INTERRUPTIBLE);
6699 * While doing polled IO, before going to sleep, we need
6700 * to check if there are new reqs added to iopoll_list,
6701 * it is because reqs may have been punted to io worker
6702 * and will be added to iopoll_list later, hence check
6703 * the iopoll_list again.
6705 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6706 !list_empty_careful(&ctx->iopoll_list)) {
6707 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6711 to_submit = io_sqring_entries(ctx);
6712 if (!to_submit || ret == -EBUSY)
6716 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6717 io_ring_clear_wakeup_flag(ctx);
6719 /* if we're handling multiple rings, cap submit size for fairness */
6720 if (cap_entries && to_submit > 8)
6723 mutex_lock(&ctx->uring_lock);
6724 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6725 ret = io_submit_sqes(ctx, to_submit);
6726 mutex_unlock(&ctx->uring_lock);
6728 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6729 wake_up(&ctx->sqo_sq_wait);
6731 return SQT_DID_WORK;
6734 static void io_sqd_init_new(struct io_sq_data *sqd)
6736 struct io_ring_ctx *ctx;
6738 while (!list_empty(&sqd->ctx_new_list)) {
6739 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6740 init_wait(&ctx->sqo_wait_entry);
6741 ctx->sqo_wait_entry.func = io_sq_wake_function;
6742 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6743 complete(&ctx->sq_thread_comp);
6747 static int io_sq_thread(void *data)
6749 struct cgroup_subsys_state *cur_css = NULL;
6750 const struct cred *old_cred = NULL;
6751 struct io_sq_data *sqd = data;
6752 struct io_ring_ctx *ctx;
6753 unsigned long start_jiffies;
6755 start_jiffies = jiffies;
6756 while (!kthread_should_stop()) {
6757 enum sq_ret ret = 0;
6761 * Any changes to the sqd lists are synchronized through the
6762 * kthread parking. This synchronizes the thread vs users,
6763 * the users are synchronized on the sqd->ctx_lock.
6765 if (kthread_should_park())
6768 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6769 io_sqd_init_new(sqd);
6771 cap_entries = !list_is_singular(&sqd->ctx_list);
6773 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6774 if (current->cred != ctx->creds) {
6776 revert_creds(old_cred);
6777 old_cred = override_creds(ctx->creds);
6779 io_sq_thread_associate_blkcg(ctx, &cur_css);
6781 current->loginuid = ctx->loginuid;
6782 current->sessionid = ctx->sessionid;
6785 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6787 io_sq_thread_drop_mm();
6790 if (ret & SQT_SPIN) {
6793 } else if (ret == SQT_IDLE) {
6794 if (kthread_should_park())
6796 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6797 io_ring_set_wakeup_flag(ctx);
6799 start_jiffies = jiffies;
6800 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6801 io_ring_clear_wakeup_flag(ctx);
6808 io_sq_thread_unassociate_blkcg();
6810 revert_creds(old_cred);
6817 struct io_wait_queue {
6818 struct wait_queue_entry wq;
6819 struct io_ring_ctx *ctx;
6821 unsigned nr_timeouts;
6824 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6826 struct io_ring_ctx *ctx = iowq->ctx;
6829 * Wake up if we have enough events, or if a timeout occurred since we
6830 * started waiting. For timeouts, we always want to return to userspace,
6831 * regardless of event count.
6833 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6834 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6837 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6838 int wake_flags, void *key)
6840 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6843 /* use noflush == true, as we can't safely rely on locking context */
6844 if (!io_should_wake(iowq, true))
6847 return autoremove_wake_function(curr, mode, wake_flags, key);
6850 static int io_run_task_work_sig(void)
6852 if (io_run_task_work())
6854 if (!signal_pending(current))
6856 if (current->jobctl & JOBCTL_TASK_WORK) {
6857 spin_lock_irq(¤t->sighand->siglock);
6858 current->jobctl &= ~JOBCTL_TASK_WORK;
6859 recalc_sigpending();
6860 spin_unlock_irq(¤t->sighand->siglock);
6867 * Wait until events become available, if we don't already have some. The
6868 * application must reap them itself, as they reside on the shared cq ring.
6870 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6871 const sigset_t __user *sig, size_t sigsz)
6873 struct io_wait_queue iowq = {
6876 .func = io_wake_function,
6877 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6880 .to_wait = min_events,
6882 struct io_rings *rings = ctx->rings;
6886 if (io_cqring_events(ctx, false) >= min_events)
6888 if (!io_run_task_work())
6893 #ifdef CONFIG_COMPAT
6894 if (in_compat_syscall())
6895 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6899 ret = set_user_sigmask(sig, sigsz);
6905 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6906 trace_io_uring_cqring_wait(ctx, min_events);
6908 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6909 TASK_INTERRUPTIBLE);
6910 /* make sure we run task_work before checking for signals */
6911 ret = io_run_task_work_sig();
6916 if (io_should_wake(&iowq, false))
6920 finish_wait(&ctx->wait, &iowq.wq);
6922 restore_saved_sigmask_unless(ret == -EINTR);
6924 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6927 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6929 #if defined(CONFIG_UNIX)
6930 if (ctx->ring_sock) {
6931 struct sock *sock = ctx->ring_sock->sk;
6932 struct sk_buff *skb;
6934 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6940 for (i = 0; i < ctx->nr_user_files; i++) {
6943 file = io_file_from_index(ctx, i);
6950 static void io_file_ref_kill(struct percpu_ref *ref)
6952 struct fixed_file_data *data;
6954 data = container_of(ref, struct fixed_file_data, refs);
6955 complete(&data->done);
6958 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6960 struct fixed_file_data *data = ctx->file_data;
6961 struct fixed_file_ref_node *ref_node = NULL;
6962 unsigned nr_tables, i;
6967 spin_lock(&data->lock);
6968 if (!list_empty(&data->ref_list))
6969 ref_node = list_first_entry(&data->ref_list,
6970 struct fixed_file_ref_node, node);
6971 spin_unlock(&data->lock);
6973 percpu_ref_kill(&ref_node->refs);
6975 percpu_ref_kill(&data->refs);
6977 /* wait for all refs nodes to complete */
6978 flush_delayed_work(&ctx->file_put_work);
6979 wait_for_completion(&data->done);
6981 __io_sqe_files_unregister(ctx);
6982 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6983 for (i = 0; i < nr_tables; i++)
6984 kfree(data->table[i].files);
6986 percpu_ref_exit(&data->refs);
6988 ctx->file_data = NULL;
6989 ctx->nr_user_files = 0;
6993 static void io_put_sq_data(struct io_sq_data *sqd)
6995 if (refcount_dec_and_test(&sqd->refs)) {
6997 * The park is a bit of a work-around, without it we get
6998 * warning spews on shutdown with SQPOLL set and affinity
6999 * set to a single CPU.
7002 kthread_park(sqd->thread);
7003 kthread_stop(sqd->thread);
7010 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7012 struct io_ring_ctx *ctx_attach;
7013 struct io_sq_data *sqd;
7016 f = fdget(p->wq_fd);
7018 return ERR_PTR(-ENXIO);
7019 if (f.file->f_op != &io_uring_fops) {
7021 return ERR_PTR(-EINVAL);
7024 ctx_attach = f.file->private_data;
7025 sqd = ctx_attach->sq_data;
7028 return ERR_PTR(-EINVAL);
7031 refcount_inc(&sqd->refs);
7036 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7038 struct io_sq_data *sqd;
7040 if (p->flags & IORING_SETUP_ATTACH_WQ)
7041 return io_attach_sq_data(p);
7043 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7045 return ERR_PTR(-ENOMEM);
7047 refcount_set(&sqd->refs, 1);
7048 INIT_LIST_HEAD(&sqd->ctx_list);
7049 INIT_LIST_HEAD(&sqd->ctx_new_list);
7050 mutex_init(&sqd->ctx_lock);
7051 mutex_init(&sqd->lock);
7052 init_waitqueue_head(&sqd->wait);
7056 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7057 __releases(&sqd->lock)
7061 kthread_unpark(sqd->thread);
7062 mutex_unlock(&sqd->lock);
7065 static void io_sq_thread_park(struct io_sq_data *sqd)
7066 __acquires(&sqd->lock)
7070 mutex_lock(&sqd->lock);
7071 kthread_park(sqd->thread);
7074 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7076 struct io_sq_data *sqd = ctx->sq_data;
7081 * We may arrive here from the error branch in
7082 * io_sq_offload_create() where the kthread is created
7083 * without being waked up, thus wake it up now to make
7084 * sure the wait will complete.
7086 wake_up_process(sqd->thread);
7087 wait_for_completion(&ctx->sq_thread_comp);
7089 io_sq_thread_park(sqd);
7092 mutex_lock(&sqd->ctx_lock);
7093 list_del(&ctx->sqd_list);
7094 mutex_unlock(&sqd->ctx_lock);
7097 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7098 io_sq_thread_unpark(sqd);
7101 io_put_sq_data(sqd);
7102 ctx->sq_data = NULL;
7106 static void io_finish_async(struct io_ring_ctx *ctx)
7108 io_sq_thread_stop(ctx);
7111 io_wq_destroy(ctx->io_wq);
7116 #if defined(CONFIG_UNIX)
7118 * Ensure the UNIX gc is aware of our file set, so we are certain that
7119 * the io_uring can be safely unregistered on process exit, even if we have
7120 * loops in the file referencing.
7122 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7124 struct sock *sk = ctx->ring_sock->sk;
7125 struct scm_fp_list *fpl;
7126 struct sk_buff *skb;
7129 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7133 skb = alloc_skb(0, GFP_KERNEL);
7142 fpl->user = get_uid(ctx->user);
7143 for (i = 0; i < nr; i++) {
7144 struct file *file = io_file_from_index(ctx, i + offset);
7148 fpl->fp[nr_files] = get_file(file);
7149 unix_inflight(fpl->user, fpl->fp[nr_files]);
7154 fpl->max = SCM_MAX_FD;
7155 fpl->count = nr_files;
7156 UNIXCB(skb).fp = fpl;
7157 skb->destructor = unix_destruct_scm;
7158 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7159 skb_queue_head(&sk->sk_receive_queue, skb);
7161 for (i = 0; i < nr_files; i++)
7172 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7173 * causes regular reference counting to break down. We rely on the UNIX
7174 * garbage collection to take care of this problem for us.
7176 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7178 unsigned left, total;
7182 left = ctx->nr_user_files;
7184 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7186 ret = __io_sqe_files_scm(ctx, this_files, total);
7190 total += this_files;
7196 while (total < ctx->nr_user_files) {
7197 struct file *file = io_file_from_index(ctx, total);
7207 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7213 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7214 unsigned nr_tables, unsigned nr_files)
7218 for (i = 0; i < nr_tables; i++) {
7219 struct fixed_file_table *table = &file_data->table[i];
7220 unsigned this_files;
7222 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7223 table->files = kcalloc(this_files, sizeof(struct file *),
7227 nr_files -= this_files;
7233 for (i = 0; i < nr_tables; i++) {
7234 struct fixed_file_table *table = &file_data->table[i];
7235 kfree(table->files);
7240 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7242 #if defined(CONFIG_UNIX)
7243 struct sock *sock = ctx->ring_sock->sk;
7244 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7245 struct sk_buff *skb;
7248 __skb_queue_head_init(&list);
7251 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7252 * remove this entry and rearrange the file array.
7254 skb = skb_dequeue(head);
7256 struct scm_fp_list *fp;
7258 fp = UNIXCB(skb).fp;
7259 for (i = 0; i < fp->count; i++) {
7262 if (fp->fp[i] != file)
7265 unix_notinflight(fp->user, fp->fp[i]);
7266 left = fp->count - 1 - i;
7268 memmove(&fp->fp[i], &fp->fp[i + 1],
7269 left * sizeof(struct file *));
7276 __skb_queue_tail(&list, skb);
7286 __skb_queue_tail(&list, skb);
7288 skb = skb_dequeue(head);
7291 if (skb_peek(&list)) {
7292 spin_lock_irq(&head->lock);
7293 while ((skb = __skb_dequeue(&list)) != NULL)
7294 __skb_queue_tail(head, skb);
7295 spin_unlock_irq(&head->lock);
7302 struct io_file_put {
7303 struct list_head list;
7307 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7309 struct fixed_file_data *file_data = ref_node->file_data;
7310 struct io_ring_ctx *ctx = file_data->ctx;
7311 struct io_file_put *pfile, *tmp;
7313 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7314 list_del(&pfile->list);
7315 io_ring_file_put(ctx, pfile->file);
7319 spin_lock(&file_data->lock);
7320 list_del(&ref_node->node);
7321 spin_unlock(&file_data->lock);
7323 percpu_ref_exit(&ref_node->refs);
7325 percpu_ref_put(&file_data->refs);
7328 static void io_file_put_work(struct work_struct *work)
7330 struct io_ring_ctx *ctx;
7331 struct llist_node *node;
7333 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7334 node = llist_del_all(&ctx->file_put_llist);
7337 struct fixed_file_ref_node *ref_node;
7338 struct llist_node *next = node->next;
7340 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7341 __io_file_put_work(ref_node);
7346 static void io_file_data_ref_zero(struct percpu_ref *ref)
7348 struct fixed_file_ref_node *ref_node;
7349 struct io_ring_ctx *ctx;
7353 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7354 ctx = ref_node->file_data->ctx;
7356 if (percpu_ref_is_dying(&ctx->file_data->refs))
7359 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7361 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7363 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7366 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7367 struct io_ring_ctx *ctx)
7369 struct fixed_file_ref_node *ref_node;
7371 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7373 return ERR_PTR(-ENOMEM);
7375 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7378 return ERR_PTR(-ENOMEM);
7380 INIT_LIST_HEAD(&ref_node->node);
7381 INIT_LIST_HEAD(&ref_node->file_list);
7382 ref_node->file_data = ctx->file_data;
7386 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7388 percpu_ref_exit(&ref_node->refs);
7392 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7395 __s32 __user *fds = (__s32 __user *) arg;
7396 unsigned nr_tables, i;
7398 int fd, ret = -ENOMEM;
7399 struct fixed_file_ref_node *ref_node;
7400 struct fixed_file_data *file_data;
7406 if (nr_args > IORING_MAX_FIXED_FILES)
7409 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7412 file_data->ctx = ctx;
7413 init_completion(&file_data->done);
7414 INIT_LIST_HEAD(&file_data->ref_list);
7415 spin_lock_init(&file_data->lock);
7417 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7418 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7420 if (!file_data->table)
7423 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7424 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7427 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7429 ctx->file_data = file_data;
7431 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7432 struct fixed_file_table *table;
7435 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7439 /* allow sparse sets */
7449 * Don't allow io_uring instances to be registered. If UNIX
7450 * isn't enabled, then this causes a reference cycle and this
7451 * instance can never get freed. If UNIX is enabled we'll
7452 * handle it just fine, but there's still no point in allowing
7453 * a ring fd as it doesn't support regular read/write anyway.
7455 if (file->f_op == &io_uring_fops) {
7459 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7460 index = i & IORING_FILE_TABLE_MASK;
7461 table->files[index] = file;
7464 ret = io_sqe_files_scm(ctx);
7466 io_sqe_files_unregister(ctx);
7470 ref_node = alloc_fixed_file_ref_node(ctx);
7471 if (IS_ERR(ref_node)) {
7472 io_sqe_files_unregister(ctx);
7473 return PTR_ERR(ref_node);
7476 file_data->node = ref_node;
7477 spin_lock(&file_data->lock);
7478 list_add(&ref_node->node, &file_data->ref_list);
7479 spin_unlock(&file_data->lock);
7480 percpu_ref_get(&file_data->refs);
7483 for (i = 0; i < ctx->nr_user_files; i++) {
7484 file = io_file_from_index(ctx, i);
7488 for (i = 0; i < nr_tables; i++)
7489 kfree(file_data->table[i].files);
7490 ctx->nr_user_files = 0;
7492 percpu_ref_exit(&file_data->refs);
7494 kfree(file_data->table);
7496 ctx->file_data = NULL;
7500 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7503 #if defined(CONFIG_UNIX)
7504 struct sock *sock = ctx->ring_sock->sk;
7505 struct sk_buff_head *head = &sock->sk_receive_queue;
7506 struct sk_buff *skb;
7509 * See if we can merge this file into an existing skb SCM_RIGHTS
7510 * file set. If there's no room, fall back to allocating a new skb
7511 * and filling it in.
7513 spin_lock_irq(&head->lock);
7514 skb = skb_peek(head);
7516 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7518 if (fpl->count < SCM_MAX_FD) {
7519 __skb_unlink(skb, head);
7520 spin_unlock_irq(&head->lock);
7521 fpl->fp[fpl->count] = get_file(file);
7522 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7524 spin_lock_irq(&head->lock);
7525 __skb_queue_head(head, skb);
7530 spin_unlock_irq(&head->lock);
7537 return __io_sqe_files_scm(ctx, 1, index);
7543 static int io_queue_file_removal(struct fixed_file_data *data,
7546 struct io_file_put *pfile;
7547 struct fixed_file_ref_node *ref_node = data->node;
7549 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7554 list_add(&pfile->list, &ref_node->file_list);
7559 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7560 struct io_uring_files_update *up,
7563 struct fixed_file_data *data = ctx->file_data;
7564 struct fixed_file_ref_node *ref_node;
7569 bool needs_switch = false;
7571 if (check_add_overflow(up->offset, nr_args, &done))
7573 if (done > ctx->nr_user_files)
7576 ref_node = alloc_fixed_file_ref_node(ctx);
7577 if (IS_ERR(ref_node))
7578 return PTR_ERR(ref_node);
7581 fds = u64_to_user_ptr(up->fds);
7583 struct fixed_file_table *table;
7587 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7591 i = array_index_nospec(up->offset, ctx->nr_user_files);
7592 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7593 index = i & IORING_FILE_TABLE_MASK;
7594 if (table->files[index]) {
7595 file = table->files[index];
7596 err = io_queue_file_removal(data, file);
7599 table->files[index] = NULL;
7600 needs_switch = true;
7609 * Don't allow io_uring instances to be registered. If
7610 * UNIX isn't enabled, then this causes a reference
7611 * cycle and this instance can never get freed. If UNIX
7612 * is enabled we'll handle it just fine, but there's
7613 * still no point in allowing a ring fd as it doesn't
7614 * support regular read/write anyway.
7616 if (file->f_op == &io_uring_fops) {
7621 table->files[index] = file;
7622 err = io_sqe_file_register(ctx, file, i);
7624 table->files[index] = NULL;
7635 percpu_ref_kill(&data->node->refs);
7636 spin_lock(&data->lock);
7637 list_add(&ref_node->node, &data->ref_list);
7638 data->node = ref_node;
7639 spin_unlock(&data->lock);
7640 percpu_ref_get(&ctx->file_data->refs);
7642 destroy_fixed_file_ref_node(ref_node);
7644 return done ? done : err;
7647 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7650 struct io_uring_files_update up;
7652 if (!ctx->file_data)
7656 if (copy_from_user(&up, arg, sizeof(up)))
7661 return __io_sqe_files_update(ctx, &up, nr_args);
7664 static void io_free_work(struct io_wq_work *work)
7666 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7668 /* Consider that io_steal_work() relies on this ref */
7672 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7673 struct io_uring_params *p)
7675 struct io_wq_data data;
7677 struct io_ring_ctx *ctx_attach;
7678 unsigned int concurrency;
7681 data.user = ctx->user;
7682 data.free_work = io_free_work;
7683 data.do_work = io_wq_submit_work;
7685 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7686 /* Do QD, or 4 * CPUS, whatever is smallest */
7687 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7689 ctx->io_wq = io_wq_create(concurrency, &data);
7690 if (IS_ERR(ctx->io_wq)) {
7691 ret = PTR_ERR(ctx->io_wq);
7697 f = fdget(p->wq_fd);
7701 if (f.file->f_op != &io_uring_fops) {
7706 ctx_attach = f.file->private_data;
7707 /* @io_wq is protected by holding the fd */
7708 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7713 ctx->io_wq = ctx_attach->io_wq;
7719 static int io_uring_alloc_task_context(struct task_struct *task)
7721 struct io_uring_task *tctx;
7724 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7725 if (unlikely(!tctx))
7728 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7729 if (unlikely(ret)) {
7735 init_waitqueue_head(&tctx->wait);
7738 io_init_identity(&tctx->__identity);
7739 tctx->identity = &tctx->__identity;
7740 task->io_uring = tctx;
7744 void __io_uring_free(struct task_struct *tsk)
7746 struct io_uring_task *tctx = tsk->io_uring;
7748 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7749 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7750 if (tctx->identity != &tctx->__identity)
7751 kfree(tctx->identity);
7752 percpu_counter_destroy(&tctx->inflight);
7754 tsk->io_uring = NULL;
7757 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7758 struct io_uring_params *p)
7762 if (ctx->flags & IORING_SETUP_SQPOLL) {
7763 struct io_sq_data *sqd;
7766 if (!capable(CAP_SYS_ADMIN))
7769 sqd = io_get_sq_data(p);
7776 io_sq_thread_park(sqd);
7777 mutex_lock(&sqd->ctx_lock);
7778 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7779 mutex_unlock(&sqd->ctx_lock);
7780 io_sq_thread_unpark(sqd);
7782 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7783 if (!ctx->sq_thread_idle)
7784 ctx->sq_thread_idle = HZ;
7789 if (p->flags & IORING_SETUP_SQ_AFF) {
7790 int cpu = p->sq_thread_cpu;
7793 if (cpu >= nr_cpu_ids)
7795 if (!cpu_online(cpu))
7798 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7799 cpu, "io_uring-sq");
7801 sqd->thread = kthread_create(io_sq_thread, sqd,
7804 if (IS_ERR(sqd->thread)) {
7805 ret = PTR_ERR(sqd->thread);
7809 ret = io_uring_alloc_task_context(sqd->thread);
7812 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7813 /* Can't have SQ_AFF without SQPOLL */
7819 ret = io_init_wq_offload(ctx, p);
7825 io_finish_async(ctx);
7829 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7831 struct io_sq_data *sqd = ctx->sq_data;
7833 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7834 wake_up_process(sqd->thread);
7837 static inline void __io_unaccount_mem(struct user_struct *user,
7838 unsigned long nr_pages)
7840 atomic_long_sub(nr_pages, &user->locked_vm);
7843 static inline int __io_account_mem(struct user_struct *user,
7844 unsigned long nr_pages)
7846 unsigned long page_limit, cur_pages, new_pages;
7848 /* Don't allow more pages than we can safely lock */
7849 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7852 cur_pages = atomic_long_read(&user->locked_vm);
7853 new_pages = cur_pages + nr_pages;
7854 if (new_pages > page_limit)
7856 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7857 new_pages) != cur_pages);
7862 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7863 enum io_mem_account acct)
7866 __io_unaccount_mem(ctx->user, nr_pages);
7868 if (ctx->mm_account) {
7869 if (acct == ACCT_LOCKED)
7870 ctx->mm_account->locked_vm -= nr_pages;
7871 else if (acct == ACCT_PINNED)
7872 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7876 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7877 enum io_mem_account acct)
7881 if (ctx->limit_mem) {
7882 ret = __io_account_mem(ctx->user, nr_pages);
7887 if (ctx->mm_account) {
7888 if (acct == ACCT_LOCKED)
7889 ctx->mm_account->locked_vm += nr_pages;
7890 else if (acct == ACCT_PINNED)
7891 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7897 static void io_mem_free(void *ptr)
7904 page = virt_to_head_page(ptr);
7905 if (put_page_testzero(page))
7906 free_compound_page(page);
7909 static void *io_mem_alloc(size_t size)
7911 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7914 return (void *) __get_free_pages(gfp_flags, get_order(size));
7917 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7920 struct io_rings *rings;
7921 size_t off, sq_array_size;
7923 off = struct_size(rings, cqes, cq_entries);
7924 if (off == SIZE_MAX)
7928 off = ALIGN(off, SMP_CACHE_BYTES);
7936 sq_array_size = array_size(sizeof(u32), sq_entries);
7937 if (sq_array_size == SIZE_MAX)
7940 if (check_add_overflow(off, sq_array_size, &off))
7946 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7950 pages = (size_t)1 << get_order(
7951 rings_size(sq_entries, cq_entries, NULL));
7952 pages += (size_t)1 << get_order(
7953 array_size(sizeof(struct io_uring_sqe), sq_entries));
7958 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7962 if (!ctx->user_bufs)
7965 for (i = 0; i < ctx->nr_user_bufs; i++) {
7966 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7968 for (j = 0; j < imu->nr_bvecs; j++)
7969 unpin_user_page(imu->bvec[j].bv_page);
7971 if (imu->acct_pages)
7972 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7977 kfree(ctx->user_bufs);
7978 ctx->user_bufs = NULL;
7979 ctx->nr_user_bufs = 0;
7983 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7984 void __user *arg, unsigned index)
7986 struct iovec __user *src;
7988 #ifdef CONFIG_COMPAT
7990 struct compat_iovec __user *ciovs;
7991 struct compat_iovec ciov;
7993 ciovs = (struct compat_iovec __user *) arg;
7994 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7997 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7998 dst->iov_len = ciov.iov_len;
8002 src = (struct iovec __user *) arg;
8003 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8009 * Not super efficient, but this is just a registration time. And we do cache
8010 * the last compound head, so generally we'll only do a full search if we don't
8013 * We check if the given compound head page has already been accounted, to
8014 * avoid double accounting it. This allows us to account the full size of the
8015 * page, not just the constituent pages of a huge page.
8017 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8018 int nr_pages, struct page *hpage)
8022 /* check current page array */
8023 for (i = 0; i < nr_pages; i++) {
8024 if (!PageCompound(pages[i]))
8026 if (compound_head(pages[i]) == hpage)
8030 /* check previously registered pages */
8031 for (i = 0; i < ctx->nr_user_bufs; i++) {
8032 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8034 for (j = 0; j < imu->nr_bvecs; j++) {
8035 if (!PageCompound(imu->bvec[j].bv_page))
8037 if (compound_head(imu->bvec[j].bv_page) == hpage)
8045 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8046 int nr_pages, struct io_mapped_ubuf *imu,
8047 struct page **last_hpage)
8051 for (i = 0; i < nr_pages; i++) {
8052 if (!PageCompound(pages[i])) {
8057 hpage = compound_head(pages[i]);
8058 if (hpage == *last_hpage)
8060 *last_hpage = hpage;
8061 if (headpage_already_acct(ctx, pages, i, hpage))
8063 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8067 if (!imu->acct_pages)
8070 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8072 imu->acct_pages = 0;
8076 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8079 struct vm_area_struct **vmas = NULL;
8080 struct page **pages = NULL;
8081 struct page *last_hpage = NULL;
8082 int i, j, got_pages = 0;
8087 if (!nr_args || nr_args > UIO_MAXIOV)
8090 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8092 if (!ctx->user_bufs)
8095 for (i = 0; i < nr_args; i++) {
8096 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8097 unsigned long off, start, end, ubuf;
8102 ret = io_copy_iov(ctx, &iov, arg, i);
8107 * Don't impose further limits on the size and buffer
8108 * constraints here, we'll -EINVAL later when IO is
8109 * submitted if they are wrong.
8112 if (!iov.iov_base || !iov.iov_len)
8115 /* arbitrary limit, but we need something */
8116 if (iov.iov_len > SZ_1G)
8119 ubuf = (unsigned long) iov.iov_base;
8120 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8121 start = ubuf >> PAGE_SHIFT;
8122 nr_pages = end - start;
8125 if (!pages || nr_pages > got_pages) {
8128 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8130 vmas = kvmalloc_array(nr_pages,
8131 sizeof(struct vm_area_struct *),
8133 if (!pages || !vmas) {
8137 got_pages = nr_pages;
8140 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8147 mmap_read_lock(current->mm);
8148 pret = pin_user_pages(ubuf, nr_pages,
8149 FOLL_WRITE | FOLL_LONGTERM,
8151 if (pret == nr_pages) {
8152 /* don't support file backed memory */
8153 for (j = 0; j < nr_pages; j++) {
8154 struct vm_area_struct *vma = vmas[j];
8157 !is_file_hugepages(vma->vm_file)) {
8163 ret = pret < 0 ? pret : -EFAULT;
8165 mmap_read_unlock(current->mm);
8168 * if we did partial map, or found file backed vmas,
8169 * release any pages we did get
8172 unpin_user_pages(pages, pret);
8177 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8179 unpin_user_pages(pages, pret);
8184 off = ubuf & ~PAGE_MASK;
8186 for (j = 0; j < nr_pages; j++) {
8189 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8190 imu->bvec[j].bv_page = pages[j];
8191 imu->bvec[j].bv_len = vec_len;
8192 imu->bvec[j].bv_offset = off;
8196 /* store original address for later verification */
8198 imu->len = iov.iov_len;
8199 imu->nr_bvecs = nr_pages;
8201 ctx->nr_user_bufs++;
8209 io_sqe_buffer_unregister(ctx);
8213 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8215 __s32 __user *fds = arg;
8221 if (copy_from_user(&fd, fds, sizeof(*fds)))
8224 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8225 if (IS_ERR(ctx->cq_ev_fd)) {
8226 int ret = PTR_ERR(ctx->cq_ev_fd);
8227 ctx->cq_ev_fd = NULL;
8234 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8236 if (ctx->cq_ev_fd) {
8237 eventfd_ctx_put(ctx->cq_ev_fd);
8238 ctx->cq_ev_fd = NULL;
8245 static int __io_destroy_buffers(int id, void *p, void *data)
8247 struct io_ring_ctx *ctx = data;
8248 struct io_buffer *buf = p;
8250 __io_remove_buffers(ctx, buf, id, -1U);
8254 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8256 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8257 idr_destroy(&ctx->io_buffer_idr);
8260 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8262 io_finish_async(ctx);
8263 io_sqe_buffer_unregister(ctx);
8265 if (ctx->sqo_task) {
8266 put_task_struct(ctx->sqo_task);
8267 ctx->sqo_task = NULL;
8268 mmdrop(ctx->mm_account);
8269 ctx->mm_account = NULL;
8272 #ifdef CONFIG_BLK_CGROUP
8273 if (ctx->sqo_blkcg_css)
8274 css_put(ctx->sqo_blkcg_css);
8277 io_sqe_files_unregister(ctx);
8278 io_eventfd_unregister(ctx);
8279 io_destroy_buffers(ctx);
8280 idr_destroy(&ctx->personality_idr);
8282 #if defined(CONFIG_UNIX)
8283 if (ctx->ring_sock) {
8284 ctx->ring_sock->file = NULL; /* so that iput() is called */
8285 sock_release(ctx->ring_sock);
8289 io_mem_free(ctx->rings);
8290 io_mem_free(ctx->sq_sqes);
8292 percpu_ref_exit(&ctx->refs);
8293 free_uid(ctx->user);
8294 put_cred(ctx->creds);
8295 kfree(ctx->cancel_hash);
8296 kmem_cache_free(req_cachep, ctx->fallback_req);
8300 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8302 struct io_ring_ctx *ctx = file->private_data;
8305 poll_wait(file, &ctx->cq_wait, wait);
8307 * synchronizes with barrier from wq_has_sleeper call in
8311 if (!io_sqring_full(ctx))
8312 mask |= EPOLLOUT | EPOLLWRNORM;
8313 if (io_cqring_events(ctx, false))
8314 mask |= EPOLLIN | EPOLLRDNORM;
8319 static int io_uring_fasync(int fd, struct file *file, int on)
8321 struct io_ring_ctx *ctx = file->private_data;
8323 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8326 static int io_remove_personalities(int id, void *p, void *data)
8328 struct io_ring_ctx *ctx = data;
8329 struct io_identity *iod;
8331 iod = idr_remove(&ctx->personality_idr, id);
8333 put_cred(iod->creds);
8334 if (refcount_dec_and_test(&iod->count))
8340 static void io_ring_exit_work(struct work_struct *work)
8342 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8346 * If we're doing polled IO and end up having requests being
8347 * submitted async (out-of-line), then completions can come in while
8348 * we're waiting for refs to drop. We need to reap these manually,
8349 * as nobody else will be looking for them.
8353 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8354 io_iopoll_try_reap_events(ctx);
8355 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8356 io_ring_ctx_free(ctx);
8359 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8361 mutex_lock(&ctx->uring_lock);
8362 percpu_ref_kill(&ctx->refs);
8363 mutex_unlock(&ctx->uring_lock);
8365 io_kill_timeouts(ctx, NULL);
8366 io_poll_remove_all(ctx, NULL);
8369 io_wq_cancel_all(ctx->io_wq);
8371 /* if we failed setting up the ctx, we might not have any rings */
8373 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8374 io_iopoll_try_reap_events(ctx);
8375 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8378 * Do this upfront, so we won't have a grace period where the ring
8379 * is closed but resources aren't reaped yet. This can cause
8380 * spurious failure in setting up a new ring.
8382 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8385 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8387 * Use system_unbound_wq to avoid spawning tons of event kworkers
8388 * if we're exiting a ton of rings at the same time. It just adds
8389 * noise and overhead, there's no discernable change in runtime
8390 * over using system_wq.
8392 queue_work(system_unbound_wq, &ctx->exit_work);
8395 static int io_uring_release(struct inode *inode, struct file *file)
8397 struct io_ring_ctx *ctx = file->private_data;
8399 file->private_data = NULL;
8400 io_ring_ctx_wait_and_kill(ctx);
8404 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8406 struct files_struct *files = data;
8408 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8409 work->identity->files == files);
8413 * Returns true if 'preq' is the link parent of 'req'
8415 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8417 struct io_kiocb *link;
8419 if (!(preq->flags & REQ_F_LINK_HEAD))
8422 list_for_each_entry(link, &preq->link_list, link_list) {
8430 static bool io_match_link_files(struct io_kiocb *req,
8431 struct files_struct *files)
8433 struct io_kiocb *link;
8435 if (io_match_files(req, files))
8437 if (req->flags & REQ_F_LINK_HEAD) {
8438 list_for_each_entry(link, &req->link_list, link_list) {
8439 if (io_match_files(link, files))
8447 * We're looking to cancel 'req' because it's holding on to our files, but
8448 * 'req' could be a link to another request. See if it is, and cancel that
8449 * parent request if so.
8451 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8453 struct hlist_node *tmp;
8454 struct io_kiocb *preq;
8458 spin_lock_irq(&ctx->completion_lock);
8459 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8460 struct hlist_head *list;
8462 list = &ctx->cancel_hash[i];
8463 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8464 found = io_match_link(preq, req);
8466 io_poll_remove_one(preq);
8471 spin_unlock_irq(&ctx->completion_lock);
8475 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8476 struct io_kiocb *req)
8478 struct io_kiocb *preq;
8481 spin_lock_irq(&ctx->completion_lock);
8482 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8483 found = io_match_link(preq, req);
8485 __io_timeout_cancel(preq);
8489 spin_unlock_irq(&ctx->completion_lock);
8493 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8495 return io_match_link(container_of(work, struct io_kiocb, work), data);
8498 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8500 enum io_wq_cancel cret;
8502 /* cancel this particular work, if it's running */
8503 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8504 if (cret != IO_WQ_CANCEL_NOTFOUND)
8507 /* find links that hold this pending, cancel those */
8508 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8509 if (cret != IO_WQ_CANCEL_NOTFOUND)
8512 /* if we have a poll link holding this pending, cancel that */
8513 if (io_poll_remove_link(ctx, req))
8516 /* final option, timeout link is holding this req pending */
8517 io_timeout_remove_link(ctx, req);
8520 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8521 struct files_struct *files)
8523 struct io_defer_entry *de = NULL;
8526 spin_lock_irq(&ctx->completion_lock);
8527 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8528 if (io_match_link_files(de->req, files)) {
8529 list_cut_position(&list, &ctx->defer_list, &de->list);
8533 spin_unlock_irq(&ctx->completion_lock);
8535 while (!list_empty(&list)) {
8536 de = list_first_entry(&list, struct io_defer_entry, list);
8537 list_del_init(&de->list);
8538 req_set_fail_links(de->req);
8539 io_put_req(de->req);
8540 io_req_complete(de->req, -ECANCELED);
8546 * Returns true if we found and killed one or more files pinning requests
8548 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8549 struct files_struct *files)
8551 if (list_empty_careful(&ctx->inflight_list))
8554 io_cancel_defer_files(ctx, files);
8555 /* cancel all at once, should be faster than doing it one by one*/
8556 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8558 while (!list_empty_careful(&ctx->inflight_list)) {
8559 struct io_kiocb *cancel_req = NULL, *req;
8562 spin_lock_irq(&ctx->inflight_lock);
8563 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8564 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8565 req->work.identity->files != files)
8567 /* req is being completed, ignore */
8568 if (!refcount_inc_not_zero(&req->refs))
8574 prepare_to_wait(&ctx->inflight_wait, &wait,
8575 TASK_UNINTERRUPTIBLE);
8576 spin_unlock_irq(&ctx->inflight_lock);
8578 /* We need to keep going until we don't find a matching req */
8581 /* cancel this request, or head link requests */
8582 io_attempt_cancel(ctx, cancel_req);
8583 io_put_req(cancel_req);
8584 /* cancellations _may_ trigger task work */
8587 finish_wait(&ctx->inflight_wait, &wait);
8593 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8595 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8596 struct task_struct *task = data;
8598 return io_task_match(req, task);
8601 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8602 struct task_struct *task,
8603 struct files_struct *files)
8607 ret = io_uring_cancel_files(ctx, files);
8609 enum io_wq_cancel cret;
8611 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8612 if (cret != IO_WQ_CANCEL_NOTFOUND)
8615 /* SQPOLL thread does its own polling */
8616 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8617 while (!list_empty_careful(&ctx->iopoll_list)) {
8618 io_iopoll_try_reap_events(ctx);
8623 ret |= io_poll_remove_all(ctx, task);
8624 ret |= io_kill_timeouts(ctx, task);
8631 * We need to iteratively cancel requests, in case a request has dependent
8632 * hard links. These persist even for failure of cancelations, hence keep
8633 * looping until none are found.
8635 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8636 struct files_struct *files)
8638 struct task_struct *task = current;
8640 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8641 task = ctx->sq_data->thread;
8643 io_cqring_overflow_flush(ctx, true, task, files);
8645 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8652 * Note that this task has used io_uring. We use it for cancelation purposes.
8654 static int io_uring_add_task_file(struct file *file)
8656 struct io_uring_task *tctx = current->io_uring;
8658 if (unlikely(!tctx)) {
8661 ret = io_uring_alloc_task_context(current);
8664 tctx = current->io_uring;
8666 if (tctx->last != file) {
8667 void *old = xa_load(&tctx->xa, (unsigned long)file);
8671 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8680 * Remove this io_uring_file -> task mapping.
8682 static void io_uring_del_task_file(struct file *file)
8684 struct io_uring_task *tctx = current->io_uring;
8686 if (tctx->last == file)
8688 file = xa_erase(&tctx->xa, (unsigned long)file);
8693 static void __io_uring_attempt_task_drop(struct file *file)
8695 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8698 io_uring_del_task_file(file);
8702 * Drop task note for this file if we're the only ones that hold it after
8705 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8707 if (!current->io_uring)
8710 * fput() is pending, will be 2 if the only other ref is our potential
8711 * task file note. If the task is exiting, drop regardless of count.
8713 if (!exiting && atomic_long_read(&file->f_count) != 2)
8716 __io_uring_attempt_task_drop(file);
8719 void __io_uring_files_cancel(struct files_struct *files)
8721 struct io_uring_task *tctx = current->io_uring;
8723 unsigned long index;
8725 /* make sure overflow events are dropped */
8726 tctx->in_idle = true;
8728 xa_for_each(&tctx->xa, index, file) {
8729 struct io_ring_ctx *ctx = file->private_data;
8731 io_uring_cancel_task_requests(ctx, files);
8733 io_uring_del_task_file(file);
8738 * Find any io_uring fd that this task has registered or done IO on, and cancel
8741 void __io_uring_task_cancel(void)
8743 struct io_uring_task *tctx = current->io_uring;
8747 /* make sure overflow events are dropped */
8748 tctx->in_idle = true;
8751 /* read completions before cancelations */
8752 inflight = percpu_counter_sum(&tctx->inflight);
8755 __io_uring_files_cancel(NULL);
8757 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8760 * If we've seen completions, retry. This avoids a race where
8761 * a completion comes in before we did prepare_to_wait().
8763 if (inflight != percpu_counter_sum(&tctx->inflight))
8768 finish_wait(&tctx->wait, &wait);
8769 tctx->in_idle = false;
8772 static int io_uring_flush(struct file *file, void *data)
8774 struct io_ring_ctx *ctx = file->private_data;
8777 * If the task is going away, cancel work it may have pending
8779 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8782 io_uring_cancel_task_requests(ctx, data);
8783 io_uring_attempt_task_drop(file, !data);
8787 static void *io_uring_validate_mmap_request(struct file *file,
8788 loff_t pgoff, size_t sz)
8790 struct io_ring_ctx *ctx = file->private_data;
8791 loff_t offset = pgoff << PAGE_SHIFT;
8796 case IORING_OFF_SQ_RING:
8797 case IORING_OFF_CQ_RING:
8800 case IORING_OFF_SQES:
8804 return ERR_PTR(-EINVAL);
8807 page = virt_to_head_page(ptr);
8808 if (sz > page_size(page))
8809 return ERR_PTR(-EINVAL);
8816 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8818 size_t sz = vma->vm_end - vma->vm_start;
8822 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8824 return PTR_ERR(ptr);
8826 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8827 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8830 #else /* !CONFIG_MMU */
8832 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8834 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8837 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8839 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8842 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8843 unsigned long addr, unsigned long len,
8844 unsigned long pgoff, unsigned long flags)
8848 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8850 return PTR_ERR(ptr);
8852 return (unsigned long) ptr;
8855 #endif /* !CONFIG_MMU */
8857 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8862 if (!io_sqring_full(ctx))
8865 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8867 if (!io_sqring_full(ctx))
8871 } while (!signal_pending(current));
8873 finish_wait(&ctx->sqo_sq_wait, &wait);
8876 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8877 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8880 struct io_ring_ctx *ctx;
8887 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8888 IORING_ENTER_SQ_WAIT))
8896 if (f.file->f_op != &io_uring_fops)
8900 ctx = f.file->private_data;
8901 if (!percpu_ref_tryget(&ctx->refs))
8905 if (ctx->flags & IORING_SETUP_R_DISABLED)
8909 * For SQ polling, the thread will do all submissions and completions.
8910 * Just return the requested submit count, and wake the thread if
8914 if (ctx->flags & IORING_SETUP_SQPOLL) {
8915 if (!list_empty_careful(&ctx->cq_overflow_list))
8916 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8917 if (flags & IORING_ENTER_SQ_WAKEUP)
8918 wake_up(&ctx->sq_data->wait);
8919 if (flags & IORING_ENTER_SQ_WAIT)
8920 io_sqpoll_wait_sq(ctx);
8921 submitted = to_submit;
8922 } else if (to_submit) {
8923 ret = io_uring_add_task_file(f.file);
8926 mutex_lock(&ctx->uring_lock);
8927 submitted = io_submit_sqes(ctx, to_submit);
8928 mutex_unlock(&ctx->uring_lock);
8930 if (submitted != to_submit)
8933 if (flags & IORING_ENTER_GETEVENTS) {
8934 min_complete = min(min_complete, ctx->cq_entries);
8937 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8938 * space applications don't need to do io completion events
8939 * polling again, they can rely on io_sq_thread to do polling
8940 * work, which can reduce cpu usage and uring_lock contention.
8942 if (ctx->flags & IORING_SETUP_IOPOLL &&
8943 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8944 ret = io_iopoll_check(ctx, min_complete);
8946 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8951 percpu_ref_put(&ctx->refs);
8954 return submitted ? submitted : ret;
8957 #ifdef CONFIG_PROC_FS
8958 static int io_uring_show_cred(int id, void *p, void *data)
8960 const struct cred *cred = p;
8961 struct seq_file *m = data;
8962 struct user_namespace *uns = seq_user_ns(m);
8963 struct group_info *gi;
8968 seq_printf(m, "%5d\n", id);
8969 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8970 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8971 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8972 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8973 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8974 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8975 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8976 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8977 seq_puts(m, "\n\tGroups:\t");
8978 gi = cred->group_info;
8979 for (g = 0; g < gi->ngroups; g++) {
8980 seq_put_decimal_ull(m, g ? " " : "",
8981 from_kgid_munged(uns, gi->gid[g]));
8983 seq_puts(m, "\n\tCapEff:\t");
8984 cap = cred->cap_effective;
8985 CAP_FOR_EACH_U32(__capi)
8986 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8991 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8993 struct io_sq_data *sq = NULL;
8998 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8999 * since fdinfo case grabs it in the opposite direction of normal use
9000 * cases. If we fail to get the lock, we just don't iterate any
9001 * structures that could be going away outside the io_uring mutex.
9003 has_lock = mutex_trylock(&ctx->uring_lock);
9005 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9008 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9009 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9010 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9011 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9012 struct fixed_file_table *table;
9015 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9016 f = table->files[i & IORING_FILE_TABLE_MASK];
9018 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9020 seq_printf(m, "%5u: <none>\n", i);
9022 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9023 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9024 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9026 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9027 (unsigned int) buf->len);
9029 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9030 seq_printf(m, "Personalities:\n");
9031 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9033 seq_printf(m, "PollList:\n");
9034 spin_lock_irq(&ctx->completion_lock);
9035 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9036 struct hlist_head *list = &ctx->cancel_hash[i];
9037 struct io_kiocb *req;
9039 hlist_for_each_entry(req, list, hash_node)
9040 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9041 req->task->task_works != NULL);
9043 spin_unlock_irq(&ctx->completion_lock);
9045 mutex_unlock(&ctx->uring_lock);
9048 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9050 struct io_ring_ctx *ctx = f->private_data;
9052 if (percpu_ref_tryget(&ctx->refs)) {
9053 __io_uring_show_fdinfo(ctx, m);
9054 percpu_ref_put(&ctx->refs);
9059 static const struct file_operations io_uring_fops = {
9060 .release = io_uring_release,
9061 .flush = io_uring_flush,
9062 .mmap = io_uring_mmap,
9064 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9065 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9067 .poll = io_uring_poll,
9068 .fasync = io_uring_fasync,
9069 #ifdef CONFIG_PROC_FS
9070 .show_fdinfo = io_uring_show_fdinfo,
9074 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9075 struct io_uring_params *p)
9077 struct io_rings *rings;
9078 size_t size, sq_array_offset;
9080 /* make sure these are sane, as we already accounted them */
9081 ctx->sq_entries = p->sq_entries;
9082 ctx->cq_entries = p->cq_entries;
9084 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9085 if (size == SIZE_MAX)
9088 rings = io_mem_alloc(size);
9093 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9094 rings->sq_ring_mask = p->sq_entries - 1;
9095 rings->cq_ring_mask = p->cq_entries - 1;
9096 rings->sq_ring_entries = p->sq_entries;
9097 rings->cq_ring_entries = p->cq_entries;
9098 ctx->sq_mask = rings->sq_ring_mask;
9099 ctx->cq_mask = rings->cq_ring_mask;
9101 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9102 if (size == SIZE_MAX) {
9103 io_mem_free(ctx->rings);
9108 ctx->sq_sqes = io_mem_alloc(size);
9109 if (!ctx->sq_sqes) {
9110 io_mem_free(ctx->rings);
9119 * Allocate an anonymous fd, this is what constitutes the application
9120 * visible backing of an io_uring instance. The application mmaps this
9121 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9122 * we have to tie this fd to a socket for file garbage collection purposes.
9124 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9129 #if defined(CONFIG_UNIX)
9130 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9136 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9140 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9141 O_RDWR | O_CLOEXEC);
9145 ret = PTR_ERR(file);
9149 #if defined(CONFIG_UNIX)
9150 ctx->ring_sock->file = file;
9152 if (unlikely(io_uring_add_task_file(file))) {
9153 file = ERR_PTR(-ENOMEM);
9156 fd_install(ret, file);
9159 #if defined(CONFIG_UNIX)
9160 sock_release(ctx->ring_sock);
9161 ctx->ring_sock = NULL;
9166 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9167 struct io_uring_params __user *params)
9169 struct user_struct *user = NULL;
9170 struct io_ring_ctx *ctx;
9176 if (entries > IORING_MAX_ENTRIES) {
9177 if (!(p->flags & IORING_SETUP_CLAMP))
9179 entries = IORING_MAX_ENTRIES;
9183 * Use twice as many entries for the CQ ring. It's possible for the
9184 * application to drive a higher depth than the size of the SQ ring,
9185 * since the sqes are only used at submission time. This allows for
9186 * some flexibility in overcommitting a bit. If the application has
9187 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9188 * of CQ ring entries manually.
9190 p->sq_entries = roundup_pow_of_two(entries);
9191 if (p->flags & IORING_SETUP_CQSIZE) {
9193 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9194 * to a power-of-two, if it isn't already. We do NOT impose
9195 * any cq vs sq ring sizing.
9197 if (p->cq_entries < p->sq_entries)
9199 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9200 if (!(p->flags & IORING_SETUP_CLAMP))
9202 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9204 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9206 p->cq_entries = 2 * p->sq_entries;
9209 user = get_uid(current_user());
9210 limit_mem = !capable(CAP_IPC_LOCK);
9213 ret = __io_account_mem(user,
9214 ring_pages(p->sq_entries, p->cq_entries));
9221 ctx = io_ring_ctx_alloc(p);
9224 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9229 ctx->compat = in_compat_syscall();
9231 ctx->creds = get_current_cred();
9233 ctx->loginuid = current->loginuid;
9234 ctx->sessionid = current->sessionid;
9236 ctx->sqo_task = get_task_struct(current);
9239 * This is just grabbed for accounting purposes. When a process exits,
9240 * the mm is exited and dropped before the files, hence we need to hang
9241 * on to this mm purely for the purposes of being able to unaccount
9242 * memory (locked/pinned vm). It's not used for anything else.
9244 mmgrab(current->mm);
9245 ctx->mm_account = current->mm;
9247 #ifdef CONFIG_BLK_CGROUP
9249 * The sq thread will belong to the original cgroup it was inited in.
9250 * If the cgroup goes offline (e.g. disabling the io controller), then
9251 * issued bios will be associated with the closest cgroup later in the
9255 ctx->sqo_blkcg_css = blkcg_css();
9256 ret = css_tryget_online(ctx->sqo_blkcg_css);
9259 /* don't init against a dying cgroup, have the user try again */
9260 ctx->sqo_blkcg_css = NULL;
9267 * Account memory _before_ installing the file descriptor. Once
9268 * the descriptor is installed, it can get closed at any time. Also
9269 * do this before hitting the general error path, as ring freeing
9270 * will un-account as well.
9272 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9274 ctx->limit_mem = limit_mem;
9276 ret = io_allocate_scq_urings(ctx, p);
9280 ret = io_sq_offload_create(ctx, p);
9284 if (!(p->flags & IORING_SETUP_R_DISABLED))
9285 io_sq_offload_start(ctx);
9287 memset(&p->sq_off, 0, sizeof(p->sq_off));
9288 p->sq_off.head = offsetof(struct io_rings, sq.head);
9289 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9290 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9291 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9292 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9293 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9294 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9296 memset(&p->cq_off, 0, sizeof(p->cq_off));
9297 p->cq_off.head = offsetof(struct io_rings, cq.head);
9298 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9299 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9300 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9301 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9302 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9303 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9305 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9306 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9307 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9308 IORING_FEAT_POLL_32BITS;
9310 if (copy_to_user(params, p, sizeof(*p))) {
9316 * Install ring fd as the very last thing, so we don't risk someone
9317 * having closed it before we finish setup
9319 ret = io_uring_get_fd(ctx);
9323 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9326 io_ring_ctx_wait_and_kill(ctx);
9331 * Sets up an aio uring context, and returns the fd. Applications asks for a
9332 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9333 * params structure passed in.
9335 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9337 struct io_uring_params p;
9340 if (copy_from_user(&p, params, sizeof(p)))
9342 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9347 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9348 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9349 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9350 IORING_SETUP_R_DISABLED))
9353 return io_uring_create(entries, &p, params);
9356 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9357 struct io_uring_params __user *, params)
9359 return io_uring_setup(entries, params);
9362 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9364 struct io_uring_probe *p;
9368 size = struct_size(p, ops, nr_args);
9369 if (size == SIZE_MAX)
9371 p = kzalloc(size, GFP_KERNEL);
9376 if (copy_from_user(p, arg, size))
9379 if (memchr_inv(p, 0, size))
9382 p->last_op = IORING_OP_LAST - 1;
9383 if (nr_args > IORING_OP_LAST)
9384 nr_args = IORING_OP_LAST;
9386 for (i = 0; i < nr_args; i++) {
9388 if (!io_op_defs[i].not_supported)
9389 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9394 if (copy_to_user(arg, p, size))
9401 static int io_register_personality(struct io_ring_ctx *ctx)
9403 struct io_identity *id;
9406 id = kmalloc(sizeof(*id), GFP_KERNEL);
9410 io_init_identity(id);
9411 id->creds = get_current_cred();
9413 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9415 put_cred(id->creds);
9421 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9423 struct io_identity *iod;
9425 iod = idr_remove(&ctx->personality_idr, id);
9427 put_cred(iod->creds);
9428 if (refcount_dec_and_test(&iod->count))
9436 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9437 unsigned int nr_args)
9439 struct io_uring_restriction *res;
9443 /* Restrictions allowed only if rings started disabled */
9444 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9447 /* We allow only a single restrictions registration */
9448 if (ctx->restrictions.registered)
9451 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9454 size = array_size(nr_args, sizeof(*res));
9455 if (size == SIZE_MAX)
9458 res = memdup_user(arg, size);
9460 return PTR_ERR(res);
9464 for (i = 0; i < nr_args; i++) {
9465 switch (res[i].opcode) {
9466 case IORING_RESTRICTION_REGISTER_OP:
9467 if (res[i].register_op >= IORING_REGISTER_LAST) {
9472 __set_bit(res[i].register_op,
9473 ctx->restrictions.register_op);
9475 case IORING_RESTRICTION_SQE_OP:
9476 if (res[i].sqe_op >= IORING_OP_LAST) {
9481 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9483 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9484 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9486 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9487 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9496 /* Reset all restrictions if an error happened */
9498 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9500 ctx->restrictions.registered = true;
9506 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9508 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9511 if (ctx->restrictions.registered)
9512 ctx->restricted = 1;
9514 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9516 io_sq_offload_start(ctx);
9521 static bool io_register_op_must_quiesce(int op)
9524 case IORING_UNREGISTER_FILES:
9525 case IORING_REGISTER_FILES_UPDATE:
9526 case IORING_REGISTER_PROBE:
9527 case IORING_REGISTER_PERSONALITY:
9528 case IORING_UNREGISTER_PERSONALITY:
9535 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9536 void __user *arg, unsigned nr_args)
9537 __releases(ctx->uring_lock)
9538 __acquires(ctx->uring_lock)
9543 * We're inside the ring mutex, if the ref is already dying, then
9544 * someone else killed the ctx or is already going through
9545 * io_uring_register().
9547 if (percpu_ref_is_dying(&ctx->refs))
9550 if (io_register_op_must_quiesce(opcode)) {
9551 percpu_ref_kill(&ctx->refs);
9554 * Drop uring mutex before waiting for references to exit. If
9555 * another thread is currently inside io_uring_enter() it might
9556 * need to grab the uring_lock to make progress. If we hold it
9557 * here across the drain wait, then we can deadlock. It's safe
9558 * to drop the mutex here, since no new references will come in
9559 * after we've killed the percpu ref.
9561 mutex_unlock(&ctx->uring_lock);
9563 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9566 ret = io_run_task_work_sig();
9571 mutex_lock(&ctx->uring_lock);
9574 percpu_ref_resurrect(&ctx->refs);
9579 if (ctx->restricted) {
9580 if (opcode >= IORING_REGISTER_LAST) {
9585 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9592 case IORING_REGISTER_BUFFERS:
9593 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9595 case IORING_UNREGISTER_BUFFERS:
9599 ret = io_sqe_buffer_unregister(ctx);
9601 case IORING_REGISTER_FILES:
9602 ret = io_sqe_files_register(ctx, arg, nr_args);
9604 case IORING_UNREGISTER_FILES:
9608 ret = io_sqe_files_unregister(ctx);
9610 case IORING_REGISTER_FILES_UPDATE:
9611 ret = io_sqe_files_update(ctx, arg, nr_args);
9613 case IORING_REGISTER_EVENTFD:
9614 case IORING_REGISTER_EVENTFD_ASYNC:
9618 ret = io_eventfd_register(ctx, arg);
9621 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9622 ctx->eventfd_async = 1;
9624 ctx->eventfd_async = 0;
9626 case IORING_UNREGISTER_EVENTFD:
9630 ret = io_eventfd_unregister(ctx);
9632 case IORING_REGISTER_PROBE:
9634 if (!arg || nr_args > 256)
9636 ret = io_probe(ctx, arg, nr_args);
9638 case IORING_REGISTER_PERSONALITY:
9642 ret = io_register_personality(ctx);
9644 case IORING_UNREGISTER_PERSONALITY:
9648 ret = io_unregister_personality(ctx, nr_args);
9650 case IORING_REGISTER_ENABLE_RINGS:
9654 ret = io_register_enable_rings(ctx);
9656 case IORING_REGISTER_RESTRICTIONS:
9657 ret = io_register_restrictions(ctx, arg, nr_args);
9665 if (io_register_op_must_quiesce(opcode)) {
9666 /* bring the ctx back to life */
9667 percpu_ref_reinit(&ctx->refs);
9669 reinit_completion(&ctx->ref_comp);
9674 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9675 void __user *, arg, unsigned int, nr_args)
9677 struct io_ring_ctx *ctx;
9686 if (f.file->f_op != &io_uring_fops)
9689 ctx = f.file->private_data;
9691 mutex_lock(&ctx->uring_lock);
9692 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9693 mutex_unlock(&ctx->uring_lock);
9694 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9695 ctx->cq_ev_fd != NULL, ret);
9701 static int __init io_uring_init(void)
9703 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9704 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9705 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9708 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9709 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9710 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9711 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9712 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9713 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9714 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9715 BUILD_BUG_SQE_ELEM(8, __u64, off);
9716 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9717 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9718 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9719 BUILD_BUG_SQE_ELEM(24, __u32, len);
9720 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9721 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9722 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9723 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9724 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9725 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9726 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9727 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9728 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9729 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9730 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9731 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9732 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9733 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9734 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9735 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9736 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9737 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9738 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9740 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9741 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9742 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9745 __initcall(io_uring_init);