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>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
107 u32 head ____cacheline_aligned_in_smp;
108 u32 tail ____cacheline_aligned_in_smp;
112 * This data is shared with the application through the mmap at offsets
113 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
115 * The offsets to the member fields are published through struct
116 * io_sqring_offsets when calling io_uring_setup.
120 * Head and tail offsets into the ring; the offsets need to be
121 * masked to get valid indices.
123 * The kernel controls head of the sq ring and the tail of the cq ring,
124 * and the application controls tail of the sq ring and the head of the
127 struct io_uring sq, cq;
129 * Bitmasks to apply to head and tail offsets (constant, equals
132 u32 sq_ring_mask, cq_ring_mask;
133 /* Ring sizes (constant, power of 2) */
134 u32 sq_ring_entries, cq_ring_entries;
136 * Number of invalid entries dropped by the kernel due to
137 * invalid index stored in array
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * After a new SQ head value was read by the application this
144 * counter includes all submissions that were dropped reaching
145 * the new SQ head (and possibly more).
151 * Written by the kernel, shouldn't be modified by the
154 * The application needs a full memory barrier before checking
155 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
161 * Written by the application, shouldn't be modified by the
166 * Number of completion events lost because the queue was full;
167 * this should be avoided by the application by making sure
168 * there are not more requests pending than there is space in
169 * the completion queue.
171 * Written by the kernel, shouldn't be modified by the
172 * application (i.e. get number of "new events" by comparing to
175 * As completion events come in out of order this counter is not
176 * ordered with any other data.
180 * Ring buffer of completion events.
182 * The kernel writes completion events fresh every time they are
183 * produced, so the application is allowed to modify pending
186 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
189 struct io_mapped_ubuf {
192 struct bio_vec *bvec;
193 unsigned int nr_bvecs;
194 unsigned long acct_pages;
197 struct fixed_file_table {
201 struct fixed_file_ref_node {
202 struct percpu_ref refs;
203 struct list_head node;
204 struct list_head file_list;
205 struct fixed_file_data *file_data;
206 struct llist_node llist;
209 struct fixed_file_data {
210 struct fixed_file_table *table;
211 struct io_ring_ctx *ctx;
213 struct fixed_file_ref_node *node;
214 struct percpu_ref refs;
215 struct completion done;
216 struct list_head ref_list;
221 struct list_head list;
227 struct io_restriction {
228 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
229 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
230 u8 sqe_flags_allowed;
231 u8 sqe_flags_required;
239 /* ctx's that are using this sqd */
240 struct list_head ctx_list;
241 struct list_head ctx_new_list;
242 struct mutex ctx_lock;
244 struct task_struct *thread;
245 struct wait_queue_head wait;
250 struct percpu_ref refs;
251 } ____cacheline_aligned_in_smp;
255 unsigned int compat: 1;
256 unsigned int limit_mem: 1;
257 unsigned int cq_overflow_flushed: 1;
258 unsigned int drain_next: 1;
259 unsigned int eventfd_async: 1;
260 unsigned int restricted: 1;
263 * Ring buffer of indices into array of io_uring_sqe, which is
264 * mmapped by the application using the IORING_OFF_SQES offset.
266 * This indirection could e.g. be used to assign fixed
267 * io_uring_sqe entries to operations and only submit them to
268 * the queue when needed.
270 * The kernel modifies neither the indices array nor the entries
274 unsigned cached_sq_head;
277 unsigned sq_thread_idle;
278 unsigned cached_sq_dropped;
279 atomic_t cached_cq_overflow;
280 unsigned long sq_check_overflow;
282 struct list_head defer_list;
283 struct list_head timeout_list;
284 struct list_head cq_overflow_list;
286 wait_queue_head_t inflight_wait;
287 struct io_uring_sqe *sq_sqes;
288 } ____cacheline_aligned_in_smp;
290 struct io_rings *rings;
296 * For SQPOLL usage - we hold a reference to the parent task, so we
297 * have access to the ->files
299 struct task_struct *sqo_task;
301 /* Only used for accounting purposes */
302 struct mm_struct *mm_account;
304 #ifdef CONFIG_BLK_CGROUP
305 struct cgroup_subsys_state *sqo_blkcg_css;
308 struct io_sq_data *sq_data; /* if using sq thread polling */
310 struct wait_queue_head sqo_sq_wait;
311 struct wait_queue_entry sqo_wait_entry;
312 struct list_head sqd_list;
315 * If used, fixed file set. Writers must ensure that ->refs is dead,
316 * readers must ensure that ->refs is alive as long as the file* is
317 * used. Only updated through io_uring_register(2).
319 struct fixed_file_data *file_data;
320 unsigned nr_user_files;
322 /* if used, fixed mapped user buffers */
323 unsigned nr_user_bufs;
324 struct io_mapped_ubuf *user_bufs;
326 struct user_struct *user;
328 const struct cred *creds;
330 struct completion ref_comp;
331 struct completion sq_thread_comp;
333 /* if all else fails... */
334 struct io_kiocb *fallback_req;
336 #if defined(CONFIG_UNIX)
337 struct socket *ring_sock;
340 struct idr io_buffer_idr;
342 struct idr personality_idr;
345 unsigned cached_cq_tail;
348 atomic_t cq_timeouts;
349 unsigned long cq_check_overflow;
350 struct wait_queue_head cq_wait;
351 struct fasync_struct *cq_fasync;
352 struct eventfd_ctx *cq_ev_fd;
353 } ____cacheline_aligned_in_smp;
356 struct mutex uring_lock;
357 wait_queue_head_t wait;
358 } ____cacheline_aligned_in_smp;
361 spinlock_t completion_lock;
364 * ->iopoll_list is protected by the ctx->uring_lock for
365 * io_uring instances that don't use IORING_SETUP_SQPOLL.
366 * For SQPOLL, only the single threaded io_sq_thread() will
367 * manipulate the list, hence no extra locking is needed there.
369 struct list_head iopoll_list;
370 struct hlist_head *cancel_hash;
371 unsigned cancel_hash_bits;
372 bool poll_multi_file;
374 spinlock_t inflight_lock;
375 struct list_head inflight_list;
376 } ____cacheline_aligned_in_smp;
378 struct delayed_work file_put_work;
379 struct llist_head file_put_llist;
381 struct work_struct exit_work;
382 struct io_restriction restrictions;
386 * First field must be the file pointer in all the
387 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
389 struct io_poll_iocb {
392 struct wait_queue_head *head;
398 struct wait_queue_entry wait;
403 struct file *put_file;
407 struct io_timeout_data {
408 struct io_kiocb *req;
409 struct hrtimer timer;
410 struct timespec64 ts;
411 enum hrtimer_mode mode;
416 struct sockaddr __user *addr;
417 int __user *addr_len;
419 unsigned long nofile;
439 struct list_head list;
442 struct io_timeout_rem {
448 /* NOTE: kiocb has the file as the first member, so don't do it here */
456 struct sockaddr __user *addr;
463 struct user_msghdr __user *umsg;
469 struct io_buffer *kbuf;
475 struct filename *filename;
477 unsigned long nofile;
480 struct io_files_update {
506 struct epoll_event event;
510 struct file *file_out;
511 struct file *file_in;
518 struct io_provide_buf {
532 const char __user *filename;
533 struct statx __user *buffer;
536 struct io_completion {
538 struct list_head list;
542 struct io_async_connect {
543 struct sockaddr_storage address;
546 struct io_async_msghdr {
547 struct iovec fast_iov[UIO_FASTIOV];
549 struct sockaddr __user *uaddr;
551 struct sockaddr_storage addr;
555 struct iovec fast_iov[UIO_FASTIOV];
556 const struct iovec *free_iovec;
557 struct iov_iter iter;
559 struct wait_page_queue wpq;
563 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
564 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
565 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
566 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
567 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
568 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
575 REQ_F_LINK_TIMEOUT_BIT,
577 REQ_F_NEED_CLEANUP_BIT,
579 REQ_F_BUFFER_SELECTED_BIT,
580 REQ_F_NO_FILE_TABLE_BIT,
581 REQ_F_WORK_INITIALIZED_BIT,
583 /* not a real bit, just to check we're not overflowing the space */
589 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
590 /* drain existing IO first */
591 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
593 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
594 /* doesn't sever on completion < 0 */
595 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
597 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
598 /* IOSQE_BUFFER_SELECT */
599 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
602 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
603 /* fail rest of links */
604 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
605 /* on inflight list */
606 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
607 /* read/write uses file position */
608 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
609 /* must not punt to workers */
610 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
611 /* has linked timeout */
612 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
614 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
616 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
617 /* already went through poll handler */
618 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
619 /* buffer already selected */
620 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
621 /* doesn't need file table for this request */
622 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
623 /* io_wq_work is initialized */
624 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
628 struct io_poll_iocb poll;
629 struct io_poll_iocb *double_poll;
633 * NOTE! Each of the iocb union members has the file pointer
634 * as the first entry in their struct definition. So you can
635 * access the file pointer through any of the sub-structs,
636 * or directly as just 'ki_filp' in this struct.
642 struct io_poll_iocb poll;
643 struct io_accept accept;
645 struct io_cancel cancel;
646 struct io_timeout timeout;
647 struct io_timeout_rem timeout_rem;
648 struct io_connect connect;
649 struct io_sr_msg sr_msg;
651 struct io_close close;
652 struct io_files_update files_update;
653 struct io_fadvise fadvise;
654 struct io_madvise madvise;
655 struct io_epoll epoll;
656 struct io_splice splice;
657 struct io_provide_buf pbuf;
658 struct io_statx statx;
659 /* use only after cleaning per-op data, see io_clean_op() */
660 struct io_completion compl;
663 /* opcode allocated if it needs to store data for async defer */
666 /* polled IO has completed */
672 struct io_ring_ctx *ctx;
675 struct task_struct *task;
678 struct list_head link_list;
681 * 1. used with ctx->iopoll_list with reads/writes
682 * 2. to track reqs with ->files (see io_op_def::file_table)
684 struct list_head inflight_entry;
686 struct percpu_ref *fixed_file_refs;
687 struct callback_head task_work;
688 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
689 struct hlist_node hash_node;
690 struct async_poll *apoll;
691 struct io_wq_work work;
692 struct io_identity identity;
695 struct io_defer_entry {
696 struct list_head list;
697 struct io_kiocb *req;
701 #define IO_IOPOLL_BATCH 8
703 struct io_comp_state {
705 struct list_head list;
706 struct io_ring_ctx *ctx;
709 struct io_submit_state {
710 struct blk_plug plug;
713 * io_kiocb alloc cache
715 void *reqs[IO_IOPOLL_BATCH];
716 unsigned int free_reqs;
719 * Batch completion logic
721 struct io_comp_state comp;
724 * File reference cache
728 unsigned int has_refs;
729 unsigned int ios_left;
733 /* needs req->file assigned */
734 unsigned needs_file : 1;
735 /* don't fail if file grab fails */
736 unsigned needs_file_no_error : 1;
737 /* hash wq insertion if file is a regular file */
738 unsigned hash_reg_file : 1;
739 /* unbound wq insertion if file is a non-regular file */
740 unsigned unbound_nonreg_file : 1;
741 /* opcode is not supported by this kernel */
742 unsigned not_supported : 1;
743 /* set if opcode supports polled "wait" */
745 unsigned pollout : 1;
746 /* op supports buffer selection */
747 unsigned buffer_select : 1;
748 /* needs rlimit(RLIMIT_FSIZE) assigned */
749 unsigned needs_fsize : 1;
750 /* must always have async data allocated */
751 unsigned needs_async_data : 1;
752 /* size of async data needed, if any */
753 unsigned short async_size;
757 static const struct io_op_def io_op_defs[] = {
758 [IORING_OP_NOP] = {},
759 [IORING_OP_READV] = {
761 .unbound_nonreg_file = 1,
764 .needs_async_data = 1,
765 .async_size = sizeof(struct io_async_rw),
766 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
768 [IORING_OP_WRITEV] = {
771 .unbound_nonreg_file = 1,
774 .needs_async_data = 1,
775 .async_size = sizeof(struct io_async_rw),
776 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
778 [IORING_OP_FSYNC] = {
780 .work_flags = IO_WQ_WORK_BLKCG,
782 [IORING_OP_READ_FIXED] = {
784 .unbound_nonreg_file = 1,
786 .async_size = sizeof(struct io_async_rw),
787 .work_flags = IO_WQ_WORK_BLKCG,
789 [IORING_OP_WRITE_FIXED] = {
792 .unbound_nonreg_file = 1,
795 .async_size = sizeof(struct io_async_rw),
796 .work_flags = IO_WQ_WORK_BLKCG,
798 [IORING_OP_POLL_ADD] = {
800 .unbound_nonreg_file = 1,
802 [IORING_OP_POLL_REMOVE] = {},
803 [IORING_OP_SYNC_FILE_RANGE] = {
805 .work_flags = IO_WQ_WORK_BLKCG,
807 [IORING_OP_SENDMSG] = {
809 .unbound_nonreg_file = 1,
811 .needs_async_data = 1,
812 .async_size = sizeof(struct io_async_msghdr),
813 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
816 [IORING_OP_RECVMSG] = {
818 .unbound_nonreg_file = 1,
821 .needs_async_data = 1,
822 .async_size = sizeof(struct io_async_msghdr),
823 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
826 [IORING_OP_TIMEOUT] = {
827 .needs_async_data = 1,
828 .async_size = sizeof(struct io_timeout_data),
829 .work_flags = IO_WQ_WORK_MM,
831 [IORING_OP_TIMEOUT_REMOVE] = {},
832 [IORING_OP_ACCEPT] = {
834 .unbound_nonreg_file = 1,
836 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
838 [IORING_OP_ASYNC_CANCEL] = {},
839 [IORING_OP_LINK_TIMEOUT] = {
840 .needs_async_data = 1,
841 .async_size = sizeof(struct io_timeout_data),
842 .work_flags = IO_WQ_WORK_MM,
844 [IORING_OP_CONNECT] = {
846 .unbound_nonreg_file = 1,
848 .needs_async_data = 1,
849 .async_size = sizeof(struct io_async_connect),
850 .work_flags = IO_WQ_WORK_MM,
852 [IORING_OP_FALLOCATE] = {
855 .work_flags = IO_WQ_WORK_BLKCG,
857 [IORING_OP_OPENAT] = {
858 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
861 [IORING_OP_CLOSE] = {
863 .needs_file_no_error = 1,
864 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
866 [IORING_OP_FILES_UPDATE] = {
867 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
869 [IORING_OP_STATX] = {
870 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
871 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
875 .unbound_nonreg_file = 1,
878 .async_size = sizeof(struct io_async_rw),
879 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
881 [IORING_OP_WRITE] = {
883 .unbound_nonreg_file = 1,
886 .async_size = sizeof(struct io_async_rw),
887 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
889 [IORING_OP_FADVISE] = {
891 .work_flags = IO_WQ_WORK_BLKCG,
893 [IORING_OP_MADVISE] = {
894 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
898 .unbound_nonreg_file = 1,
900 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
904 .unbound_nonreg_file = 1,
907 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
909 [IORING_OP_OPENAT2] = {
910 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
913 [IORING_OP_EPOLL_CTL] = {
914 .unbound_nonreg_file = 1,
915 .work_flags = IO_WQ_WORK_FILES,
917 [IORING_OP_SPLICE] = {
920 .unbound_nonreg_file = 1,
921 .work_flags = IO_WQ_WORK_BLKCG,
923 [IORING_OP_PROVIDE_BUFFERS] = {},
924 [IORING_OP_REMOVE_BUFFERS] = {},
928 .unbound_nonreg_file = 1,
932 enum io_mem_account {
937 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
938 struct io_comp_state *cs);
939 static void io_cqring_fill_event(struct io_kiocb *req, long res);
940 static void io_put_req(struct io_kiocb *req);
941 static void io_put_req_deferred(struct io_kiocb *req, int nr);
942 static void io_double_put_req(struct io_kiocb *req);
943 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
944 static void __io_queue_linked_timeout(struct io_kiocb *req);
945 static void io_queue_linked_timeout(struct io_kiocb *req);
946 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
947 struct io_uring_files_update *ip,
949 static void __io_clean_op(struct io_kiocb *req);
950 static struct file *io_file_get(struct io_submit_state *state,
951 struct io_kiocb *req, int fd, bool fixed);
952 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
953 static void io_file_put_work(struct work_struct *work);
955 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
956 struct iovec **iovec, struct iov_iter *iter,
958 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
959 const struct iovec *fast_iov,
960 struct iov_iter *iter, bool force);
962 static struct kmem_cache *req_cachep;
964 static const struct file_operations io_uring_fops;
966 struct sock *io_uring_get_socket(struct file *file)
968 #if defined(CONFIG_UNIX)
969 if (file->f_op == &io_uring_fops) {
970 struct io_ring_ctx *ctx = file->private_data;
972 return ctx->ring_sock->sk;
977 EXPORT_SYMBOL(io_uring_get_socket);
979 static inline void io_clean_op(struct io_kiocb *req)
981 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
986 static void io_sq_thread_drop_mm(void)
988 struct mm_struct *mm = current->mm;
991 kthread_unuse_mm(mm);
996 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
999 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1000 !ctx->sqo_task->mm ||
1001 !mmget_not_zero(ctx->sqo_task->mm)))
1003 kthread_use_mm(ctx->sqo_task->mm);
1009 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1010 struct io_kiocb *req)
1012 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1014 return __io_sq_thread_acquire_mm(ctx);
1017 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1018 struct cgroup_subsys_state **cur_css)
1021 #ifdef CONFIG_BLK_CGROUP
1022 /* puts the old one when swapping */
1023 if (*cur_css != ctx->sqo_blkcg_css) {
1024 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1025 *cur_css = ctx->sqo_blkcg_css;
1030 static void io_sq_thread_unassociate_blkcg(void)
1032 #ifdef CONFIG_BLK_CGROUP
1033 kthread_associate_blkcg(NULL);
1037 static inline void req_set_fail_links(struct io_kiocb *req)
1039 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1040 req->flags |= REQ_F_FAIL_LINK;
1044 * None of these are dereferenced, they are simply used to check if any of
1045 * them have changed. If we're under current and check they are still the
1046 * same, we're fine to grab references to them for actual out-of-line use.
1048 static void io_init_identity(struct io_identity *id)
1050 id->files = current->files;
1051 id->mm = current->mm;
1052 #ifdef CONFIG_BLK_CGROUP
1054 id->blkcg_css = blkcg_css();
1057 id->creds = current_cred();
1058 id->nsproxy = current->nsproxy;
1059 id->fs = current->fs;
1060 id->fsize = rlimit(RLIMIT_FSIZE);
1061 refcount_set(&id->count, 1);
1065 * Note: must call io_req_init_async() for the first time you
1066 * touch any members of io_wq_work.
1068 static inline void io_req_init_async(struct io_kiocb *req)
1070 if (req->flags & REQ_F_WORK_INITIALIZED)
1073 memset(&req->work, 0, sizeof(req->work));
1074 req->flags |= REQ_F_WORK_INITIALIZED;
1075 io_init_identity(&req->identity);
1076 req->work.identity = &req->identity;
1079 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1081 return ctx->flags & IORING_SETUP_SQPOLL;
1084 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1086 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1088 complete(&ctx->ref_comp);
1091 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1093 return !req->timeout.off;
1096 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1098 struct io_ring_ctx *ctx;
1101 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1105 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1106 if (!ctx->fallback_req)
1110 * Use 5 bits less than the max cq entries, that should give us around
1111 * 32 entries per hash list if totally full and uniformly spread.
1113 hash_bits = ilog2(p->cq_entries);
1117 ctx->cancel_hash_bits = hash_bits;
1118 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1120 if (!ctx->cancel_hash)
1122 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1124 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1125 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1128 ctx->flags = p->flags;
1129 init_waitqueue_head(&ctx->sqo_sq_wait);
1130 INIT_LIST_HEAD(&ctx->sqd_list);
1131 init_waitqueue_head(&ctx->cq_wait);
1132 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1133 init_completion(&ctx->ref_comp);
1134 init_completion(&ctx->sq_thread_comp);
1135 idr_init(&ctx->io_buffer_idr);
1136 idr_init(&ctx->personality_idr);
1137 mutex_init(&ctx->uring_lock);
1138 init_waitqueue_head(&ctx->wait);
1139 spin_lock_init(&ctx->completion_lock);
1140 INIT_LIST_HEAD(&ctx->iopoll_list);
1141 INIT_LIST_HEAD(&ctx->defer_list);
1142 INIT_LIST_HEAD(&ctx->timeout_list);
1143 init_waitqueue_head(&ctx->inflight_wait);
1144 spin_lock_init(&ctx->inflight_lock);
1145 INIT_LIST_HEAD(&ctx->inflight_list);
1146 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1147 init_llist_head(&ctx->file_put_llist);
1150 if (ctx->fallback_req)
1151 kmem_cache_free(req_cachep, ctx->fallback_req);
1152 kfree(ctx->cancel_hash);
1157 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1159 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1160 struct io_ring_ctx *ctx = req->ctx;
1162 return seq != ctx->cached_cq_tail
1163 + atomic_read(&ctx->cached_cq_overflow);
1169 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1171 struct io_rings *rings = ctx->rings;
1173 /* order cqe stores with ring update */
1174 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1176 if (wq_has_sleeper(&ctx->cq_wait)) {
1177 wake_up_interruptible(&ctx->cq_wait);
1178 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1182 static void io_put_identity(struct io_kiocb *req)
1184 if (req->work.identity == &req->identity)
1186 if (refcount_dec_and_test(&req->work.identity->count))
1187 kfree(req->work.identity);
1190 static void io_req_clean_work(struct io_kiocb *req)
1192 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1195 req->flags &= ~REQ_F_WORK_INITIALIZED;
1197 if (req->work.flags & IO_WQ_WORK_MM) {
1198 mmdrop(req->work.identity->mm);
1199 req->work.flags &= ~IO_WQ_WORK_MM;
1201 #ifdef CONFIG_BLK_CGROUP
1202 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1203 css_put(req->work.identity->blkcg_css);
1204 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1207 if (req->work.flags & IO_WQ_WORK_CREDS) {
1208 put_cred(req->work.identity->creds);
1209 req->work.flags &= ~IO_WQ_WORK_CREDS;
1211 if (req->work.flags & IO_WQ_WORK_FS) {
1212 struct fs_struct *fs = req->work.identity->fs;
1214 spin_lock(&req->work.identity->fs->lock);
1217 spin_unlock(&req->work.identity->fs->lock);
1220 req->work.flags &= ~IO_WQ_WORK_FS;
1223 io_put_identity(req);
1227 * Create a private copy of io_identity, since some fields don't match
1228 * the current context.
1230 static bool io_identity_cow(struct io_kiocb *req)
1232 const struct cred *creds = NULL;
1233 struct io_identity *id;
1235 if (req->work.flags & IO_WQ_WORK_CREDS)
1236 creds = req->work.identity->creds;
1238 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1239 if (unlikely(!id)) {
1240 req->work.flags |= IO_WQ_WORK_CANCEL;
1245 * We can safely just re-init the creds we copied Either the field
1246 * matches the current one, or we haven't grabbed it yet. The only
1247 * exception is ->creds, through registered personalities, so handle
1248 * that one separately.
1250 io_init_identity(id);
1252 req->work.identity->creds = creds;
1254 /* add one for this request */
1255 refcount_inc(&id->count);
1257 /* drop old identity, assign new one. one ref for req, one for tctx */
1258 if (req->work.identity != &req->identity &&
1259 refcount_sub_and_test(2, &req->work.identity->count))
1260 kfree(req->work.identity);
1262 req->work.identity = id;
1266 static bool io_grab_identity(struct io_kiocb *req)
1268 const struct io_op_def *def = &io_op_defs[req->opcode];
1269 struct io_identity *id = &req->identity;
1270 struct io_ring_ctx *ctx = req->ctx;
1272 if (def->needs_fsize && id->fsize != rlimit(RLIMIT_FSIZE))
1275 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1276 (def->work_flags & IO_WQ_WORK_FILES) &&
1277 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1278 if (id->files != current->files ||
1279 id->nsproxy != current->nsproxy)
1281 atomic_inc(&id->files->count);
1282 get_nsproxy(id->nsproxy);
1283 req->flags |= REQ_F_INFLIGHT;
1285 spin_lock_irq(&ctx->inflight_lock);
1286 list_add(&req->inflight_entry, &ctx->inflight_list);
1287 spin_unlock_irq(&ctx->inflight_lock);
1288 req->work.flags |= IO_WQ_WORK_FILES;
1290 #ifdef CONFIG_BLK_CGROUP
1291 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1292 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1294 if (id->blkcg_css != blkcg_css()) {
1299 * This should be rare, either the cgroup is dying or the task
1300 * is moving cgroups. Just punt to root for the handful of ios.
1302 if (css_tryget_online(id->blkcg_css))
1303 req->work.flags |= IO_WQ_WORK_BLKCG;
1307 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1308 if (id->creds != current_cred())
1310 get_cred(id->creds);
1311 req->work.flags |= IO_WQ_WORK_CREDS;
1313 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1314 (def->work_flags & IO_WQ_WORK_FS)) {
1315 if (current->fs != id->fs)
1317 spin_lock(&id->fs->lock);
1318 if (!id->fs->in_exec) {
1320 req->work.flags |= IO_WQ_WORK_FS;
1322 req->work.flags |= IO_WQ_WORK_CANCEL;
1324 spin_unlock(¤t->fs->lock);
1330 static void io_prep_async_work(struct io_kiocb *req)
1332 const struct io_op_def *def = &io_op_defs[req->opcode];
1333 struct io_identity *id = &req->identity;
1334 struct io_ring_ctx *ctx = req->ctx;
1336 io_req_init_async(req);
1338 if (req->flags & REQ_F_ISREG) {
1339 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1340 io_wq_hash_work(&req->work, file_inode(req->file));
1342 if (def->unbound_nonreg_file)
1343 req->work.flags |= IO_WQ_WORK_UNBOUND;
1346 /* ->mm can never change on us */
1347 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1348 (def->work_flags & IO_WQ_WORK_MM)) {
1350 req->work.flags |= IO_WQ_WORK_MM;
1353 /* if we fail grabbing identity, we must COW, regrab, and retry */
1354 if (io_grab_identity(req))
1357 if (!io_identity_cow(req))
1360 /* can't fail at this point */
1361 if (!io_grab_identity(req))
1365 static void io_prep_async_link(struct io_kiocb *req)
1367 struct io_kiocb *cur;
1369 io_prep_async_work(req);
1370 if (req->flags & REQ_F_LINK_HEAD)
1371 list_for_each_entry(cur, &req->link_list, link_list)
1372 io_prep_async_work(cur);
1375 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1377 struct io_ring_ctx *ctx = req->ctx;
1378 struct io_kiocb *link = io_prep_linked_timeout(req);
1380 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1381 &req->work, req->flags);
1382 io_wq_enqueue(ctx->io_wq, &req->work);
1386 static void io_queue_async_work(struct io_kiocb *req)
1388 struct io_kiocb *link;
1390 /* init ->work of the whole link before punting */
1391 io_prep_async_link(req);
1392 link = __io_queue_async_work(req);
1395 io_queue_linked_timeout(link);
1398 static void io_kill_timeout(struct io_kiocb *req)
1400 struct io_timeout_data *io = req->async_data;
1403 ret = hrtimer_try_to_cancel(&io->timer);
1405 atomic_set(&req->ctx->cq_timeouts,
1406 atomic_read(&req->ctx->cq_timeouts) + 1);
1407 list_del_init(&req->timeout.list);
1408 io_cqring_fill_event(req, 0);
1409 io_put_req_deferred(req, 1);
1413 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1415 struct io_ring_ctx *ctx = req->ctx;
1417 if (!tsk || req->task == tsk)
1419 if (ctx->flags & IORING_SETUP_SQPOLL) {
1420 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1427 * Returns true if we found and killed one or more timeouts
1429 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1431 struct io_kiocb *req, *tmp;
1434 spin_lock_irq(&ctx->completion_lock);
1435 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1436 if (io_task_match(req, tsk)) {
1437 io_kill_timeout(req);
1441 spin_unlock_irq(&ctx->completion_lock);
1442 return canceled != 0;
1445 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1448 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1449 struct io_defer_entry, list);
1450 struct io_kiocb *link;
1452 if (req_need_defer(de->req, de->seq))
1454 list_del_init(&de->list);
1455 /* punt-init is done before queueing for defer */
1456 link = __io_queue_async_work(de->req);
1458 __io_queue_linked_timeout(link);
1459 /* drop submission reference */
1460 io_put_req_deferred(link, 1);
1463 } while (!list_empty(&ctx->defer_list));
1466 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1468 while (!list_empty(&ctx->timeout_list)) {
1469 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1470 struct io_kiocb, timeout.list);
1472 if (io_is_timeout_noseq(req))
1474 if (req->timeout.target_seq != ctx->cached_cq_tail
1475 - atomic_read(&ctx->cq_timeouts))
1478 list_del_init(&req->timeout.list);
1479 io_kill_timeout(req);
1483 static void io_commit_cqring(struct io_ring_ctx *ctx)
1485 io_flush_timeouts(ctx);
1486 __io_commit_cqring(ctx);
1488 if (unlikely(!list_empty(&ctx->defer_list)))
1489 __io_queue_deferred(ctx);
1492 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1494 struct io_rings *r = ctx->rings;
1496 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1499 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1501 struct io_rings *rings = ctx->rings;
1504 tail = ctx->cached_cq_tail;
1506 * writes to the cq entry need to come after reading head; the
1507 * control dependency is enough as we're using WRITE_ONCE to
1510 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1513 ctx->cached_cq_tail++;
1514 return &rings->cqes[tail & ctx->cq_mask];
1517 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1521 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1523 if (!ctx->eventfd_async)
1525 return io_wq_current_is_worker();
1528 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1530 if (waitqueue_active(&ctx->wait))
1531 wake_up(&ctx->wait);
1532 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1533 wake_up(&ctx->sq_data->wait);
1534 if (io_should_trigger_evfd(ctx))
1535 eventfd_signal(ctx->cq_ev_fd, 1);
1538 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1540 if (list_empty(&ctx->cq_overflow_list)) {
1541 clear_bit(0, &ctx->sq_check_overflow);
1542 clear_bit(0, &ctx->cq_check_overflow);
1543 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1547 static inline bool io_match_files(struct io_kiocb *req,
1548 struct files_struct *files)
1552 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1553 (req->work.flags & IO_WQ_WORK_FILES))
1554 return req->work.identity->files == files;
1558 /* Returns true if there are no backlogged entries after the flush */
1559 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1560 struct task_struct *tsk,
1561 struct files_struct *files)
1563 struct io_rings *rings = ctx->rings;
1564 struct io_kiocb *req, *tmp;
1565 struct io_uring_cqe *cqe;
1566 unsigned long flags;
1570 if (list_empty_careful(&ctx->cq_overflow_list))
1572 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1573 rings->cq_ring_entries))
1577 spin_lock_irqsave(&ctx->completion_lock, flags);
1579 /* if force is set, the ring is going away. always drop after that */
1581 ctx->cq_overflow_flushed = 1;
1584 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1585 if (tsk && req->task != tsk)
1587 if (!io_match_files(req, files))
1590 cqe = io_get_cqring(ctx);
1594 list_move(&req->compl.list, &list);
1596 WRITE_ONCE(cqe->user_data, req->user_data);
1597 WRITE_ONCE(cqe->res, req->result);
1598 WRITE_ONCE(cqe->flags, req->compl.cflags);
1600 WRITE_ONCE(ctx->rings->cq_overflow,
1601 atomic_inc_return(&ctx->cached_cq_overflow));
1605 io_commit_cqring(ctx);
1606 io_cqring_mark_overflow(ctx);
1608 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1609 io_cqring_ev_posted(ctx);
1611 while (!list_empty(&list)) {
1612 req = list_first_entry(&list, struct io_kiocb, compl.list);
1613 list_del(&req->compl.list);
1620 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1622 struct io_ring_ctx *ctx = req->ctx;
1623 struct io_uring_cqe *cqe;
1625 trace_io_uring_complete(ctx, req->user_data, res);
1628 * If we can't get a cq entry, userspace overflowed the
1629 * submission (by quite a lot). Increment the overflow count in
1632 cqe = io_get_cqring(ctx);
1634 WRITE_ONCE(cqe->user_data, req->user_data);
1635 WRITE_ONCE(cqe->res, res);
1636 WRITE_ONCE(cqe->flags, cflags);
1637 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1639 * If we're in ring overflow flush mode, or in task cancel mode,
1640 * then we cannot store the request for later flushing, we need
1641 * to drop it on the floor.
1643 WRITE_ONCE(ctx->rings->cq_overflow,
1644 atomic_inc_return(&ctx->cached_cq_overflow));
1646 if (list_empty(&ctx->cq_overflow_list)) {
1647 set_bit(0, &ctx->sq_check_overflow);
1648 set_bit(0, &ctx->cq_check_overflow);
1649 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1653 req->compl.cflags = cflags;
1654 refcount_inc(&req->refs);
1655 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1659 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1661 __io_cqring_fill_event(req, res, 0);
1664 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1666 struct io_ring_ctx *ctx = req->ctx;
1667 unsigned long flags;
1669 spin_lock_irqsave(&ctx->completion_lock, flags);
1670 __io_cqring_fill_event(req, res, cflags);
1671 io_commit_cqring(ctx);
1672 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1674 io_cqring_ev_posted(ctx);
1677 static void io_submit_flush_completions(struct io_comp_state *cs)
1679 struct io_ring_ctx *ctx = cs->ctx;
1681 spin_lock_irq(&ctx->completion_lock);
1682 while (!list_empty(&cs->list)) {
1683 struct io_kiocb *req;
1685 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1686 list_del(&req->compl.list);
1687 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1690 * io_free_req() doesn't care about completion_lock unless one
1691 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1692 * because of a potential deadlock with req->work.fs->lock
1694 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1695 |REQ_F_WORK_INITIALIZED)) {
1696 spin_unlock_irq(&ctx->completion_lock);
1698 spin_lock_irq(&ctx->completion_lock);
1703 io_commit_cqring(ctx);
1704 spin_unlock_irq(&ctx->completion_lock);
1706 io_cqring_ev_posted(ctx);
1710 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1711 struct io_comp_state *cs)
1714 io_cqring_add_event(req, res, cflags);
1719 req->compl.cflags = cflags;
1720 list_add_tail(&req->compl.list, &cs->list);
1722 io_submit_flush_completions(cs);
1726 static void io_req_complete(struct io_kiocb *req, long res)
1728 __io_req_complete(req, res, 0, NULL);
1731 static inline bool io_is_fallback_req(struct io_kiocb *req)
1733 return req == (struct io_kiocb *)
1734 ((unsigned long) req->ctx->fallback_req & ~1UL);
1737 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1739 struct io_kiocb *req;
1741 req = ctx->fallback_req;
1742 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1748 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1749 struct io_submit_state *state)
1751 if (!state->free_reqs) {
1752 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1756 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1757 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1760 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1761 * retry single alloc to be on the safe side.
1763 if (unlikely(ret <= 0)) {
1764 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1765 if (!state->reqs[0])
1769 state->free_reqs = ret;
1773 return state->reqs[state->free_reqs];
1775 return io_get_fallback_req(ctx);
1778 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1782 percpu_ref_put(req->fixed_file_refs);
1787 static void io_dismantle_req(struct io_kiocb *req)
1791 if (req->async_data)
1792 kfree(req->async_data);
1794 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1796 io_req_clean_work(req);
1799 static void __io_free_req(struct io_kiocb *req)
1801 struct io_uring_task *tctx = req->task->io_uring;
1802 struct io_ring_ctx *ctx = req->ctx;
1804 io_dismantle_req(req);
1806 atomic_long_inc(&tctx->req_complete);
1808 wake_up(&tctx->wait);
1809 put_task_struct(req->task);
1811 if (likely(!io_is_fallback_req(req)))
1812 kmem_cache_free(req_cachep, req);
1814 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1815 percpu_ref_put(&ctx->refs);
1818 static bool io_link_cancel_timeout(struct io_kiocb *req)
1820 struct io_timeout_data *io = req->async_data;
1821 struct io_ring_ctx *ctx = req->ctx;
1824 ret = hrtimer_try_to_cancel(&io->timer);
1826 io_cqring_fill_event(req, -ECANCELED);
1827 io_commit_cqring(ctx);
1828 req->flags &= ~REQ_F_LINK_HEAD;
1829 io_put_req_deferred(req, 1);
1836 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1838 struct io_kiocb *link;
1841 if (list_empty(&req->link_list))
1843 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1844 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1847 list_del_init(&link->link_list);
1848 wake_ev = io_link_cancel_timeout(link);
1849 req->flags &= ~REQ_F_LINK_TIMEOUT;
1853 static void io_kill_linked_timeout(struct io_kiocb *req)
1855 struct io_ring_ctx *ctx = req->ctx;
1856 unsigned long flags;
1859 spin_lock_irqsave(&ctx->completion_lock, flags);
1860 wake_ev = __io_kill_linked_timeout(req);
1861 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1864 io_cqring_ev_posted(ctx);
1867 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1869 struct io_kiocb *nxt;
1872 * The list should never be empty when we are called here. But could
1873 * potentially happen if the chain is messed up, check to be on the
1876 if (unlikely(list_empty(&req->link_list)))
1879 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1880 list_del_init(&req->link_list);
1881 if (!list_empty(&nxt->link_list))
1882 nxt->flags |= REQ_F_LINK_HEAD;
1887 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1889 static void __io_fail_links(struct io_kiocb *req)
1891 struct io_ring_ctx *ctx = req->ctx;
1893 while (!list_empty(&req->link_list)) {
1894 struct io_kiocb *link = list_first_entry(&req->link_list,
1895 struct io_kiocb, link_list);
1897 list_del_init(&link->link_list);
1898 trace_io_uring_fail_link(req, link);
1900 io_cqring_fill_event(link, -ECANCELED);
1903 * It's ok to free under spinlock as they're not linked anymore,
1904 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1907 if (link->flags & REQ_F_WORK_INITIALIZED)
1908 io_put_req_deferred(link, 2);
1910 io_double_put_req(link);
1913 io_commit_cqring(ctx);
1916 static void io_fail_links(struct io_kiocb *req)
1918 struct io_ring_ctx *ctx = req->ctx;
1919 unsigned long flags;
1921 spin_lock_irqsave(&ctx->completion_lock, flags);
1922 __io_fail_links(req);
1923 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1925 io_cqring_ev_posted(ctx);
1928 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1930 req->flags &= ~REQ_F_LINK_HEAD;
1931 if (req->flags & REQ_F_LINK_TIMEOUT)
1932 io_kill_linked_timeout(req);
1935 * If LINK is set, we have dependent requests in this chain. If we
1936 * didn't fail this request, queue the first one up, moving any other
1937 * dependencies to the next request. In case of failure, fail the rest
1940 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1941 return io_req_link_next(req);
1946 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1948 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1950 return __io_req_find_next(req);
1953 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1955 struct task_struct *tsk = req->task;
1956 struct io_ring_ctx *ctx = req->ctx;
1959 if (tsk->flags & PF_EXITING)
1963 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1964 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1965 * processing task_work. There's no reliable way to tell if TWA_RESUME
1969 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1970 notify = TWA_SIGNAL;
1972 ret = task_work_add(tsk, &req->task_work, notify);
1974 wake_up_process(tsk);
1979 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1981 struct io_ring_ctx *ctx = req->ctx;
1983 spin_lock_irq(&ctx->completion_lock);
1984 io_cqring_fill_event(req, error);
1985 io_commit_cqring(ctx);
1986 spin_unlock_irq(&ctx->completion_lock);
1988 io_cqring_ev_posted(ctx);
1989 req_set_fail_links(req);
1990 io_double_put_req(req);
1993 static void io_req_task_cancel(struct callback_head *cb)
1995 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1996 struct io_ring_ctx *ctx = req->ctx;
1998 __io_req_task_cancel(req, -ECANCELED);
1999 percpu_ref_put(&ctx->refs);
2002 static void __io_req_task_submit(struct io_kiocb *req)
2004 struct io_ring_ctx *ctx = req->ctx;
2006 if (!__io_sq_thread_acquire_mm(ctx)) {
2007 mutex_lock(&ctx->uring_lock);
2008 __io_queue_sqe(req, NULL);
2009 mutex_unlock(&ctx->uring_lock);
2011 __io_req_task_cancel(req, -EFAULT);
2015 static void io_req_task_submit(struct callback_head *cb)
2017 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2018 struct io_ring_ctx *ctx = req->ctx;
2020 __io_req_task_submit(req);
2021 percpu_ref_put(&ctx->refs);
2024 static void io_req_task_queue(struct io_kiocb *req)
2028 init_task_work(&req->task_work, io_req_task_submit);
2029 percpu_ref_get(&req->ctx->refs);
2031 ret = io_req_task_work_add(req, true);
2032 if (unlikely(ret)) {
2033 struct task_struct *tsk;
2035 init_task_work(&req->task_work, io_req_task_cancel);
2036 tsk = io_wq_get_task(req->ctx->io_wq);
2037 task_work_add(tsk, &req->task_work, 0);
2038 wake_up_process(tsk);
2042 static void io_queue_next(struct io_kiocb *req)
2044 struct io_kiocb *nxt = io_req_find_next(req);
2047 io_req_task_queue(nxt);
2050 static void io_free_req(struct io_kiocb *req)
2057 void *reqs[IO_IOPOLL_BATCH];
2060 struct task_struct *task;
2064 static inline void io_init_req_batch(struct req_batch *rb)
2071 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2072 struct req_batch *rb)
2074 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2075 percpu_ref_put_many(&ctx->refs, rb->to_free);
2079 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2080 struct req_batch *rb)
2083 __io_req_free_batch_flush(ctx, rb);
2085 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2086 put_task_struct_many(rb->task, rb->task_refs);
2091 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2093 if (unlikely(io_is_fallback_req(req))) {
2097 if (req->flags & REQ_F_LINK_HEAD)
2100 if (req->task != rb->task) {
2102 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2103 put_task_struct_many(rb->task, rb->task_refs);
2105 rb->task = req->task;
2110 io_dismantle_req(req);
2111 rb->reqs[rb->to_free++] = req;
2112 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2113 __io_req_free_batch_flush(req->ctx, rb);
2117 * Drop reference to request, return next in chain (if there is one) if this
2118 * was the last reference to this request.
2120 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2122 struct io_kiocb *nxt = NULL;
2124 if (refcount_dec_and_test(&req->refs)) {
2125 nxt = io_req_find_next(req);
2131 static void io_put_req(struct io_kiocb *req)
2133 if (refcount_dec_and_test(&req->refs))
2137 static void io_put_req_deferred_cb(struct callback_head *cb)
2139 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2144 static void io_free_req_deferred(struct io_kiocb *req)
2148 init_task_work(&req->task_work, io_put_req_deferred_cb);
2149 ret = io_req_task_work_add(req, true);
2150 if (unlikely(ret)) {
2151 struct task_struct *tsk;
2153 tsk = io_wq_get_task(req->ctx->io_wq);
2154 task_work_add(tsk, &req->task_work, 0);
2155 wake_up_process(tsk);
2159 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2161 if (refcount_sub_and_test(refs, &req->refs))
2162 io_free_req_deferred(req);
2165 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2167 struct io_kiocb *nxt;
2170 * A ref is owned by io-wq in which context we're. So, if that's the
2171 * last one, it's safe to steal next work. False negatives are Ok,
2172 * it just will be re-punted async in io_put_work()
2174 if (refcount_read(&req->refs) != 1)
2177 nxt = io_req_find_next(req);
2178 return nxt ? &nxt->work : NULL;
2181 static void io_double_put_req(struct io_kiocb *req)
2183 /* drop both submit and complete references */
2184 if (refcount_sub_and_test(2, &req->refs))
2188 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2190 struct io_rings *rings = ctx->rings;
2192 if (test_bit(0, &ctx->cq_check_overflow)) {
2194 * noflush == true is from the waitqueue handler, just ensure
2195 * we wake up the task, and the next invocation will flush the
2196 * entries. We cannot safely to it from here.
2198 if (noflush && !list_empty(&ctx->cq_overflow_list))
2201 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2204 /* See comment at the top of this file */
2206 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2209 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2211 struct io_rings *rings = ctx->rings;
2213 /* make sure SQ entry isn't read before tail */
2214 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2217 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2219 unsigned int cflags;
2221 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2222 cflags |= IORING_CQE_F_BUFFER;
2223 req->flags &= ~REQ_F_BUFFER_SELECTED;
2228 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2230 struct io_buffer *kbuf;
2232 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2233 return io_put_kbuf(req, kbuf);
2236 static inline bool io_run_task_work(void)
2239 * Not safe to run on exiting task, and the task_work handling will
2240 * not add work to such a task.
2242 if (unlikely(current->flags & PF_EXITING))
2244 if (current->task_works) {
2245 __set_current_state(TASK_RUNNING);
2253 static void io_iopoll_queue(struct list_head *again)
2255 struct io_kiocb *req;
2258 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2259 list_del(&req->inflight_entry);
2260 __io_complete_rw(req, -EAGAIN, 0, NULL);
2261 } while (!list_empty(again));
2265 * Find and free completed poll iocbs
2267 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2268 struct list_head *done)
2270 struct req_batch rb;
2271 struct io_kiocb *req;
2274 /* order with ->result store in io_complete_rw_iopoll() */
2277 io_init_req_batch(&rb);
2278 while (!list_empty(done)) {
2281 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2282 if (READ_ONCE(req->result) == -EAGAIN) {
2284 req->iopoll_completed = 0;
2285 list_move_tail(&req->inflight_entry, &again);
2288 list_del(&req->inflight_entry);
2290 if (req->flags & REQ_F_BUFFER_SELECTED)
2291 cflags = io_put_rw_kbuf(req);
2293 __io_cqring_fill_event(req, req->result, cflags);
2296 if (refcount_dec_and_test(&req->refs))
2297 io_req_free_batch(&rb, req);
2300 io_commit_cqring(ctx);
2301 if (ctx->flags & IORING_SETUP_SQPOLL)
2302 io_cqring_ev_posted(ctx);
2303 io_req_free_batch_finish(ctx, &rb);
2305 if (!list_empty(&again))
2306 io_iopoll_queue(&again);
2309 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2312 struct io_kiocb *req, *tmp;
2318 * Only spin for completions if we don't have multiple devices hanging
2319 * off our complete list, and we're under the requested amount.
2321 spin = !ctx->poll_multi_file && *nr_events < min;
2324 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2325 struct kiocb *kiocb = &req->rw.kiocb;
2328 * Move completed and retryable entries to our local lists.
2329 * If we find a request that requires polling, break out
2330 * and complete those lists first, if we have entries there.
2332 if (READ_ONCE(req->iopoll_completed)) {
2333 list_move_tail(&req->inflight_entry, &done);
2336 if (!list_empty(&done))
2339 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2343 /* iopoll may have completed current req */
2344 if (READ_ONCE(req->iopoll_completed))
2345 list_move_tail(&req->inflight_entry, &done);
2352 if (!list_empty(&done))
2353 io_iopoll_complete(ctx, nr_events, &done);
2359 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2360 * non-spinning poll check - we'll still enter the driver poll loop, but only
2361 * as a non-spinning completion check.
2363 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2366 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2369 ret = io_do_iopoll(ctx, nr_events, min);
2372 if (*nr_events >= min)
2380 * We can't just wait for polled events to come to us, we have to actively
2381 * find and complete them.
2383 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2385 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2388 mutex_lock(&ctx->uring_lock);
2389 while (!list_empty(&ctx->iopoll_list)) {
2390 unsigned int nr_events = 0;
2392 io_do_iopoll(ctx, &nr_events, 0);
2394 /* let it sleep and repeat later if can't complete a request */
2398 * Ensure we allow local-to-the-cpu processing to take place,
2399 * in this case we need to ensure that we reap all events.
2400 * Also let task_work, etc. to progress by releasing the mutex
2402 if (need_resched()) {
2403 mutex_unlock(&ctx->uring_lock);
2405 mutex_lock(&ctx->uring_lock);
2408 mutex_unlock(&ctx->uring_lock);
2411 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2413 unsigned int nr_events = 0;
2414 int iters = 0, ret = 0;
2417 * We disallow the app entering submit/complete with polling, but we
2418 * still need to lock the ring to prevent racing with polled issue
2419 * that got punted to a workqueue.
2421 mutex_lock(&ctx->uring_lock);
2424 * Don't enter poll loop if we already have events pending.
2425 * If we do, we can potentially be spinning for commands that
2426 * already triggered a CQE (eg in error).
2428 if (io_cqring_events(ctx, false))
2432 * If a submit got punted to a workqueue, we can have the
2433 * application entering polling for a command before it gets
2434 * issued. That app will hold the uring_lock for the duration
2435 * of the poll right here, so we need to take a breather every
2436 * now and then to ensure that the issue has a chance to add
2437 * the poll to the issued list. Otherwise we can spin here
2438 * forever, while the workqueue is stuck trying to acquire the
2441 if (!(++iters & 7)) {
2442 mutex_unlock(&ctx->uring_lock);
2444 mutex_lock(&ctx->uring_lock);
2447 ret = io_iopoll_getevents(ctx, &nr_events, min);
2451 } while (min && !nr_events && !need_resched());
2453 mutex_unlock(&ctx->uring_lock);
2457 static void kiocb_end_write(struct io_kiocb *req)
2460 * Tell lockdep we inherited freeze protection from submission
2463 if (req->flags & REQ_F_ISREG) {
2464 struct inode *inode = file_inode(req->file);
2466 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2468 file_end_write(req->file);
2471 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2472 struct io_comp_state *cs)
2474 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2477 if (kiocb->ki_flags & IOCB_WRITE)
2478 kiocb_end_write(req);
2480 if (res != req->result)
2481 req_set_fail_links(req);
2482 if (req->flags & REQ_F_BUFFER_SELECTED)
2483 cflags = io_put_rw_kbuf(req);
2484 __io_req_complete(req, res, cflags, cs);
2488 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2490 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2491 ssize_t ret = -ECANCELED;
2492 struct iov_iter iter;
2500 switch (req->opcode) {
2501 case IORING_OP_READV:
2502 case IORING_OP_READ_FIXED:
2503 case IORING_OP_READ:
2506 case IORING_OP_WRITEV:
2507 case IORING_OP_WRITE_FIXED:
2508 case IORING_OP_WRITE:
2512 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2517 if (!req->async_data) {
2518 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2521 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2529 req_set_fail_links(req);
2530 io_req_complete(req, ret);
2535 static bool io_rw_reissue(struct io_kiocb *req, long res)
2538 umode_t mode = file_inode(req->file)->i_mode;
2541 if (!S_ISBLK(mode) && !S_ISREG(mode))
2543 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2546 ret = io_sq_thread_acquire_mm(req->ctx, req);
2548 if (io_resubmit_prep(req, ret)) {
2549 refcount_inc(&req->refs);
2550 io_queue_async_work(req);
2558 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2559 struct io_comp_state *cs)
2561 if (!io_rw_reissue(req, res))
2562 io_complete_rw_common(&req->rw.kiocb, res, cs);
2565 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2567 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2569 __io_complete_rw(req, res, res2, NULL);
2572 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2574 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2576 if (kiocb->ki_flags & IOCB_WRITE)
2577 kiocb_end_write(req);
2579 if (res != -EAGAIN && res != req->result)
2580 req_set_fail_links(req);
2582 WRITE_ONCE(req->result, res);
2583 /* order with io_poll_complete() checking ->result */
2585 WRITE_ONCE(req->iopoll_completed, 1);
2589 * After the iocb has been issued, it's safe to be found on the poll list.
2590 * Adding the kiocb to the list AFTER submission ensures that we don't
2591 * find it from a io_iopoll_getevents() thread before the issuer is done
2592 * accessing the kiocb cookie.
2594 static void io_iopoll_req_issued(struct io_kiocb *req)
2596 struct io_ring_ctx *ctx = req->ctx;
2599 * Track whether we have multiple files in our lists. This will impact
2600 * how we do polling eventually, not spinning if we're on potentially
2601 * different devices.
2603 if (list_empty(&ctx->iopoll_list)) {
2604 ctx->poll_multi_file = false;
2605 } else if (!ctx->poll_multi_file) {
2606 struct io_kiocb *list_req;
2608 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2610 if (list_req->file != req->file)
2611 ctx->poll_multi_file = true;
2615 * For fast devices, IO may have already completed. If it has, add
2616 * it to the front so we find it first.
2618 if (READ_ONCE(req->iopoll_completed))
2619 list_add(&req->inflight_entry, &ctx->iopoll_list);
2621 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2623 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2624 wq_has_sleeper(&ctx->sq_data->wait))
2625 wake_up(&ctx->sq_data->wait);
2628 static void __io_state_file_put(struct io_submit_state *state)
2630 if (state->has_refs)
2631 fput_many(state->file, state->has_refs);
2635 static inline void io_state_file_put(struct io_submit_state *state)
2638 __io_state_file_put(state);
2642 * Get as many references to a file as we have IOs left in this submission,
2643 * assuming most submissions are for one file, or at least that each file
2644 * has more than one submission.
2646 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2652 if (state->fd == fd) {
2656 __io_state_file_put(state);
2658 state->file = fget_many(fd, state->ios_left);
2663 state->has_refs = state->ios_left - 1;
2667 static bool io_bdev_nowait(struct block_device *bdev)
2670 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2677 * If we tracked the file through the SCM inflight mechanism, we could support
2678 * any file. For now, just ensure that anything potentially problematic is done
2681 static bool io_file_supports_async(struct file *file, int rw)
2683 umode_t mode = file_inode(file)->i_mode;
2685 if (S_ISBLK(mode)) {
2686 if (io_bdev_nowait(file->f_inode->i_bdev))
2690 if (S_ISCHR(mode) || S_ISSOCK(mode))
2692 if (S_ISREG(mode)) {
2693 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2694 file->f_op != &io_uring_fops)
2699 /* any ->read/write should understand O_NONBLOCK */
2700 if (file->f_flags & O_NONBLOCK)
2703 if (!(file->f_mode & FMODE_NOWAIT))
2707 return file->f_op->read_iter != NULL;
2709 return file->f_op->write_iter != NULL;
2712 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2714 struct io_ring_ctx *ctx = req->ctx;
2715 struct kiocb *kiocb = &req->rw.kiocb;
2719 if (S_ISREG(file_inode(req->file)->i_mode))
2720 req->flags |= REQ_F_ISREG;
2722 kiocb->ki_pos = READ_ONCE(sqe->off);
2723 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2724 req->flags |= REQ_F_CUR_POS;
2725 kiocb->ki_pos = req->file->f_pos;
2727 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2728 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2729 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2733 ioprio = READ_ONCE(sqe->ioprio);
2735 ret = ioprio_check_cap(ioprio);
2739 kiocb->ki_ioprio = ioprio;
2741 kiocb->ki_ioprio = get_current_ioprio();
2743 /* don't allow async punt if RWF_NOWAIT was requested */
2744 if (kiocb->ki_flags & IOCB_NOWAIT)
2745 req->flags |= REQ_F_NOWAIT;
2747 if (ctx->flags & IORING_SETUP_IOPOLL) {
2748 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2749 !kiocb->ki_filp->f_op->iopoll)
2752 kiocb->ki_flags |= IOCB_HIPRI;
2753 kiocb->ki_complete = io_complete_rw_iopoll;
2754 req->iopoll_completed = 0;
2756 if (kiocb->ki_flags & IOCB_HIPRI)
2758 kiocb->ki_complete = io_complete_rw;
2761 req->rw.addr = READ_ONCE(sqe->addr);
2762 req->rw.len = READ_ONCE(sqe->len);
2763 req->buf_index = READ_ONCE(sqe->buf_index);
2767 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2773 case -ERESTARTNOINTR:
2774 case -ERESTARTNOHAND:
2775 case -ERESTART_RESTARTBLOCK:
2777 * We can't just restart the syscall, since previously
2778 * submitted sqes may already be in progress. Just fail this
2784 kiocb->ki_complete(kiocb, ret, 0);
2788 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2789 struct io_comp_state *cs)
2791 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2792 struct io_async_rw *io = req->async_data;
2794 /* add previously done IO, if any */
2795 if (io && io->bytes_done > 0) {
2797 ret = io->bytes_done;
2799 ret += io->bytes_done;
2802 if (req->flags & REQ_F_CUR_POS)
2803 req->file->f_pos = kiocb->ki_pos;
2804 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2805 __io_complete_rw(req, ret, 0, cs);
2807 io_rw_done(kiocb, ret);
2810 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2811 struct iov_iter *iter)
2813 struct io_ring_ctx *ctx = req->ctx;
2814 size_t len = req->rw.len;
2815 struct io_mapped_ubuf *imu;
2816 u16 index, buf_index = req->buf_index;
2820 if (unlikely(buf_index >= ctx->nr_user_bufs))
2822 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2823 imu = &ctx->user_bufs[index];
2824 buf_addr = req->rw.addr;
2827 if (buf_addr + len < buf_addr)
2829 /* not inside the mapped region */
2830 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2834 * May not be a start of buffer, set size appropriately
2835 * and advance us to the beginning.
2837 offset = buf_addr - imu->ubuf;
2838 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2842 * Don't use iov_iter_advance() here, as it's really slow for
2843 * using the latter parts of a big fixed buffer - it iterates
2844 * over each segment manually. We can cheat a bit here, because
2847 * 1) it's a BVEC iter, we set it up
2848 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2849 * first and last bvec
2851 * So just find our index, and adjust the iterator afterwards.
2852 * If the offset is within the first bvec (or the whole first
2853 * bvec, just use iov_iter_advance(). This makes it easier
2854 * since we can just skip the first segment, which may not
2855 * be PAGE_SIZE aligned.
2857 const struct bio_vec *bvec = imu->bvec;
2859 if (offset <= bvec->bv_len) {
2860 iov_iter_advance(iter, offset);
2862 unsigned long seg_skip;
2864 /* skip first vec */
2865 offset -= bvec->bv_len;
2866 seg_skip = 1 + (offset >> PAGE_SHIFT);
2868 iter->bvec = bvec + seg_skip;
2869 iter->nr_segs -= seg_skip;
2870 iter->count -= bvec->bv_len + offset;
2871 iter->iov_offset = offset & ~PAGE_MASK;
2878 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2881 mutex_unlock(&ctx->uring_lock);
2884 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2887 * "Normal" inline submissions always hold the uring_lock, since we
2888 * grab it from the system call. Same is true for the SQPOLL offload.
2889 * The only exception is when we've detached the request and issue it
2890 * from an async worker thread, grab the lock for that case.
2893 mutex_lock(&ctx->uring_lock);
2896 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2897 int bgid, struct io_buffer *kbuf,
2900 struct io_buffer *head;
2902 if (req->flags & REQ_F_BUFFER_SELECTED)
2905 io_ring_submit_lock(req->ctx, needs_lock);
2907 lockdep_assert_held(&req->ctx->uring_lock);
2909 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2911 if (!list_empty(&head->list)) {
2912 kbuf = list_last_entry(&head->list, struct io_buffer,
2914 list_del(&kbuf->list);
2917 idr_remove(&req->ctx->io_buffer_idr, bgid);
2919 if (*len > kbuf->len)
2922 kbuf = ERR_PTR(-ENOBUFS);
2925 io_ring_submit_unlock(req->ctx, needs_lock);
2930 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2933 struct io_buffer *kbuf;
2936 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2937 bgid = req->buf_index;
2938 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2941 req->rw.addr = (u64) (unsigned long) kbuf;
2942 req->flags |= REQ_F_BUFFER_SELECTED;
2943 return u64_to_user_ptr(kbuf->addr);
2946 #ifdef CONFIG_COMPAT
2947 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2950 struct compat_iovec __user *uiov;
2951 compat_ssize_t clen;
2955 uiov = u64_to_user_ptr(req->rw.addr);
2956 if (!access_ok(uiov, sizeof(*uiov)))
2958 if (__get_user(clen, &uiov->iov_len))
2964 buf = io_rw_buffer_select(req, &len, needs_lock);
2966 return PTR_ERR(buf);
2967 iov[0].iov_base = buf;
2968 iov[0].iov_len = (compat_size_t) len;
2973 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2976 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2980 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2983 len = iov[0].iov_len;
2986 buf = io_rw_buffer_select(req, &len, needs_lock);
2988 return PTR_ERR(buf);
2989 iov[0].iov_base = buf;
2990 iov[0].iov_len = len;
2994 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2997 if (req->flags & REQ_F_BUFFER_SELECTED) {
2998 struct io_buffer *kbuf;
3000 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3001 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3002 iov[0].iov_len = kbuf->len;
3007 else if (req->rw.len > 1)
3010 #ifdef CONFIG_COMPAT
3011 if (req->ctx->compat)
3012 return io_compat_import(req, iov, needs_lock);
3015 return __io_iov_buffer_select(req, iov, needs_lock);
3018 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3019 struct iovec **iovec, struct iov_iter *iter,
3022 void __user *buf = u64_to_user_ptr(req->rw.addr);
3023 size_t sqe_len = req->rw.len;
3027 opcode = req->opcode;
3028 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3030 return io_import_fixed(req, rw, iter);
3033 /* buffer index only valid with fixed read/write, or buffer select */
3034 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3037 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3038 if (req->flags & REQ_F_BUFFER_SELECT) {
3039 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3041 return PTR_ERR(buf);
3042 req->rw.len = sqe_len;
3045 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3047 return ret < 0 ? ret : sqe_len;
3050 if (req->flags & REQ_F_BUFFER_SELECT) {
3051 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3053 ret = (*iovec)->iov_len;
3054 iov_iter_init(iter, rw, *iovec, 1, ret);
3060 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3064 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3065 struct iovec **iovec, struct iov_iter *iter,
3068 struct io_async_rw *iorw = req->async_data;
3071 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3073 return iov_iter_count(&iorw->iter);
3076 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3078 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3082 * For files that don't have ->read_iter() and ->write_iter(), handle them
3083 * by looping over ->read() or ->write() manually.
3085 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3086 struct iov_iter *iter)
3091 * Don't support polled IO through this interface, and we can't
3092 * support non-blocking either. For the latter, this just causes
3093 * the kiocb to be handled from an async context.
3095 if (kiocb->ki_flags & IOCB_HIPRI)
3097 if (kiocb->ki_flags & IOCB_NOWAIT)
3100 while (iov_iter_count(iter)) {
3104 if (!iov_iter_is_bvec(iter)) {
3105 iovec = iov_iter_iovec(iter);
3107 /* fixed buffers import bvec */
3108 iovec.iov_base = kmap(iter->bvec->bv_page)
3110 iovec.iov_len = min(iter->count,
3111 iter->bvec->bv_len - iter->iov_offset);
3115 nr = file->f_op->read(file, iovec.iov_base,
3116 iovec.iov_len, io_kiocb_ppos(kiocb));
3118 nr = file->f_op->write(file, iovec.iov_base,
3119 iovec.iov_len, io_kiocb_ppos(kiocb));
3122 if (iov_iter_is_bvec(iter))
3123 kunmap(iter->bvec->bv_page);
3131 if (nr != iovec.iov_len)
3133 iov_iter_advance(iter, nr);
3139 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3140 const struct iovec *fast_iov, struct iov_iter *iter)
3142 struct io_async_rw *rw = req->async_data;
3144 memcpy(&rw->iter, iter, sizeof(*iter));
3145 rw->free_iovec = iovec;
3147 /* can only be fixed buffers, no need to do anything */
3148 if (iter->type == ITER_BVEC)
3151 unsigned iov_off = 0;
3153 rw->iter.iov = rw->fast_iov;
3154 if (iter->iov != fast_iov) {
3155 iov_off = iter->iov - fast_iov;
3156 rw->iter.iov += iov_off;
3158 if (rw->fast_iov != fast_iov)
3159 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3160 sizeof(struct iovec) * iter->nr_segs);
3162 req->flags |= REQ_F_NEED_CLEANUP;
3166 static inline int __io_alloc_async_data(struct io_kiocb *req)
3168 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3169 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3170 return req->async_data == NULL;
3173 static int io_alloc_async_data(struct io_kiocb *req)
3175 if (!io_op_defs[req->opcode].needs_async_data)
3178 return __io_alloc_async_data(req);
3181 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3182 const struct iovec *fast_iov,
3183 struct iov_iter *iter, bool force)
3185 if (!force && !io_op_defs[req->opcode].needs_async_data)
3187 if (!req->async_data) {
3188 if (__io_alloc_async_data(req))
3191 io_req_map_rw(req, iovec, fast_iov, iter);
3196 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3198 struct io_async_rw *iorw = req->async_data;
3199 struct iovec *iov = iorw->fast_iov;
3202 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3203 if (unlikely(ret < 0))
3206 iorw->bytes_done = 0;
3207 iorw->free_iovec = iov;
3209 req->flags |= REQ_F_NEED_CLEANUP;
3213 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3217 ret = io_prep_rw(req, sqe);
3221 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3224 /* either don't need iovec imported or already have it */
3225 if (!req->async_data)
3227 return io_rw_prep_async(req, READ);
3231 * This is our waitqueue callback handler, registered through lock_page_async()
3232 * when we initially tried to do the IO with the iocb armed our waitqueue.
3233 * This gets called when the page is unlocked, and we generally expect that to
3234 * happen when the page IO is completed and the page is now uptodate. This will
3235 * queue a task_work based retry of the operation, attempting to copy the data
3236 * again. If the latter fails because the page was NOT uptodate, then we will
3237 * do a thread based blocking retry of the operation. That's the unexpected
3240 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3241 int sync, void *arg)
3243 struct wait_page_queue *wpq;
3244 struct io_kiocb *req = wait->private;
3245 struct wait_page_key *key = arg;
3248 wpq = container_of(wait, struct wait_page_queue, wait);
3250 if (!wake_page_match(wpq, key))
3253 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3254 list_del_init(&wait->entry);
3256 init_task_work(&req->task_work, io_req_task_submit);
3257 percpu_ref_get(&req->ctx->refs);
3259 /* submit ref gets dropped, acquire a new one */
3260 refcount_inc(&req->refs);
3261 ret = io_req_task_work_add(req, true);
3262 if (unlikely(ret)) {
3263 struct task_struct *tsk;
3265 /* queue just for cancelation */
3266 init_task_work(&req->task_work, io_req_task_cancel);
3267 tsk = io_wq_get_task(req->ctx->io_wq);
3268 task_work_add(tsk, &req->task_work, 0);
3269 wake_up_process(tsk);
3275 * This controls whether a given IO request should be armed for async page
3276 * based retry. If we return false here, the request is handed to the async
3277 * worker threads for retry. If we're doing buffered reads on a regular file,
3278 * we prepare a private wait_page_queue entry and retry the operation. This
3279 * will either succeed because the page is now uptodate and unlocked, or it
3280 * will register a callback when the page is unlocked at IO completion. Through
3281 * that callback, io_uring uses task_work to setup a retry of the operation.
3282 * That retry will attempt the buffered read again. The retry will generally
3283 * succeed, or in rare cases where it fails, we then fall back to using the
3284 * async worker threads for a blocking retry.
3286 static bool io_rw_should_retry(struct io_kiocb *req)
3288 struct io_async_rw *rw = req->async_data;
3289 struct wait_page_queue *wait = &rw->wpq;
3290 struct kiocb *kiocb = &req->rw.kiocb;
3292 /* never retry for NOWAIT, we just complete with -EAGAIN */
3293 if (req->flags & REQ_F_NOWAIT)
3296 /* Only for buffered IO */
3297 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3301 * just use poll if we can, and don't attempt if the fs doesn't
3302 * support callback based unlocks
3304 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3307 wait->wait.func = io_async_buf_func;
3308 wait->wait.private = req;
3309 wait->wait.flags = 0;
3310 INIT_LIST_HEAD(&wait->wait.entry);
3311 kiocb->ki_flags |= IOCB_WAITQ;
3312 kiocb->ki_flags &= ~IOCB_NOWAIT;
3313 kiocb->ki_waitq = wait;
3317 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3319 if (req->file->f_op->read_iter)
3320 return call_read_iter(req->file, &req->rw.kiocb, iter);
3321 else if (req->file->f_op->read)
3322 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3327 static int io_read(struct io_kiocb *req, bool force_nonblock,
3328 struct io_comp_state *cs)
3330 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3331 struct kiocb *kiocb = &req->rw.kiocb;
3332 struct iov_iter __iter, *iter = &__iter;
3333 struct io_async_rw *rw = req->async_data;
3334 ssize_t io_size, ret, ret2;
3341 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3344 iov_count = iov_iter_count(iter);
3346 req->result = io_size;
3349 /* Ensure we clear previously set non-block flag */
3350 if (!force_nonblock)
3351 kiocb->ki_flags &= ~IOCB_NOWAIT;
3353 kiocb->ki_flags |= IOCB_NOWAIT;
3356 /* If the file doesn't support async, just async punt */
3357 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3361 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3365 ret = io_iter_do_read(req, iter);
3369 } else if (ret == -EIOCBQUEUED) {
3372 } else if (ret == -EAGAIN) {
3373 /* IOPOLL retry should happen for io-wq threads */
3374 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3376 /* no retry on NONBLOCK marked file */
3377 if (req->file->f_flags & O_NONBLOCK)
3379 /* some cases will consume bytes even on error returns */
3380 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3383 } else if (ret < 0) {
3384 /* make sure -ERESTARTSYS -> -EINTR is done */
3388 /* read it all, or we did blocking attempt. no retry. */
3389 if (!iov_iter_count(iter) || !force_nonblock ||
3390 (req->file->f_flags & O_NONBLOCK))
3395 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3402 rw = req->async_data;
3403 /* it's copied and will be cleaned with ->io */
3405 /* now use our persistent iterator, if we aren't already */
3408 rw->bytes_done += ret;
3409 /* if we can retry, do so with the callbacks armed */
3410 if (!io_rw_should_retry(req)) {
3411 kiocb->ki_flags &= ~IOCB_WAITQ;
3416 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3417 * get -EIOCBQUEUED, then we'll get a notification when the desired
3418 * page gets unlocked. We can also get a partial read here, and if we
3419 * do, then just retry at the new offset.
3421 ret = io_iter_do_read(req, iter);
3422 if (ret == -EIOCBQUEUED) {
3425 } else if (ret > 0 && ret < io_size) {
3426 /* we got some bytes, but not all. retry. */
3430 kiocb_done(kiocb, ret, cs);
3433 /* it's reportedly faster than delegating the null check to kfree() */
3439 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3443 ret = io_prep_rw(req, sqe);
3447 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3450 /* either don't need iovec imported or already have it */
3451 if (!req->async_data)
3453 return io_rw_prep_async(req, WRITE);
3456 static int io_write(struct io_kiocb *req, bool force_nonblock,
3457 struct io_comp_state *cs)
3459 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3460 struct kiocb *kiocb = &req->rw.kiocb;
3461 struct iov_iter __iter, *iter = &__iter;
3462 struct io_async_rw *rw = req->async_data;
3464 ssize_t ret, ret2, io_size;
3469 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3472 iov_count = iov_iter_count(iter);
3474 req->result = io_size;
3476 /* Ensure we clear previously set non-block flag */
3477 if (!force_nonblock)
3478 kiocb->ki_flags &= ~IOCB_NOWAIT;
3480 kiocb->ki_flags |= IOCB_NOWAIT;
3482 /* If the file doesn't support async, just async punt */
3483 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3486 /* file path doesn't support NOWAIT for non-direct_IO */
3487 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3488 (req->flags & REQ_F_ISREG))
3491 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3496 * Open-code file_start_write here to grab freeze protection,
3497 * which will be released by another thread in
3498 * io_complete_rw(). Fool lockdep by telling it the lock got
3499 * released so that it doesn't complain about the held lock when
3500 * we return to userspace.
3502 if (req->flags & REQ_F_ISREG) {
3503 __sb_start_write(file_inode(req->file)->i_sb,
3504 SB_FREEZE_WRITE, true);
3505 __sb_writers_release(file_inode(req->file)->i_sb,
3508 kiocb->ki_flags |= IOCB_WRITE;
3510 if (req->file->f_op->write_iter)
3511 ret2 = call_write_iter(req->file, kiocb, iter);
3512 else if (req->file->f_op->write)
3513 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3518 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3519 * retry them without IOCB_NOWAIT.
3521 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3523 /* no retry on NONBLOCK marked file */
3524 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3526 if (!force_nonblock || ret2 != -EAGAIN) {
3527 /* IOPOLL retry should happen for io-wq threads */
3528 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3531 kiocb_done(kiocb, ret2, cs);
3534 /* some cases will consume bytes even on error returns */
3535 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3536 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3541 /* it's reportedly faster than delegating the null check to kfree() */
3547 static int __io_splice_prep(struct io_kiocb *req,
3548 const struct io_uring_sqe *sqe)
3550 struct io_splice* sp = &req->splice;
3551 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3553 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3557 sp->len = READ_ONCE(sqe->len);
3558 sp->flags = READ_ONCE(sqe->splice_flags);
3560 if (unlikely(sp->flags & ~valid_flags))
3563 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3564 (sp->flags & SPLICE_F_FD_IN_FIXED));
3567 req->flags |= REQ_F_NEED_CLEANUP;
3569 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3571 * Splice operation will be punted aync, and here need to
3572 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3574 io_req_init_async(req);
3575 req->work.flags |= IO_WQ_WORK_UNBOUND;
3581 static int io_tee_prep(struct io_kiocb *req,
3582 const struct io_uring_sqe *sqe)
3584 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3586 return __io_splice_prep(req, sqe);
3589 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3591 struct io_splice *sp = &req->splice;
3592 struct file *in = sp->file_in;
3593 struct file *out = sp->file_out;
3594 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3600 ret = do_tee(in, out, sp->len, flags);
3602 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3603 req->flags &= ~REQ_F_NEED_CLEANUP;
3606 req_set_fail_links(req);
3607 io_req_complete(req, ret);
3611 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3613 struct io_splice* sp = &req->splice;
3615 sp->off_in = READ_ONCE(sqe->splice_off_in);
3616 sp->off_out = READ_ONCE(sqe->off);
3617 return __io_splice_prep(req, sqe);
3620 static int io_splice(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;
3626 loff_t *poff_in, *poff_out;
3632 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3633 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3636 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3638 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3639 req->flags &= ~REQ_F_NEED_CLEANUP;
3642 req_set_fail_links(req);
3643 io_req_complete(req, ret);
3648 * IORING_OP_NOP just posts a completion event, nothing else.
3650 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3652 struct io_ring_ctx *ctx = req->ctx;
3654 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3657 __io_req_complete(req, 0, 0, cs);
3661 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3663 struct io_ring_ctx *ctx = req->ctx;
3668 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3670 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3673 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3674 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3677 req->sync.off = READ_ONCE(sqe->off);
3678 req->sync.len = READ_ONCE(sqe->len);
3682 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3684 loff_t end = req->sync.off + req->sync.len;
3687 /* fsync always requires a blocking context */
3691 ret = vfs_fsync_range(req->file, req->sync.off,
3692 end > 0 ? end : LLONG_MAX,
3693 req->sync.flags & IORING_FSYNC_DATASYNC);
3695 req_set_fail_links(req);
3696 io_req_complete(req, ret);
3700 static int io_fallocate_prep(struct io_kiocb *req,
3701 const struct io_uring_sqe *sqe)
3703 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3705 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3708 req->sync.off = READ_ONCE(sqe->off);
3709 req->sync.len = READ_ONCE(sqe->addr);
3710 req->sync.mode = READ_ONCE(sqe->len);
3714 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3718 /* fallocate always requiring blocking context */
3721 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3724 req_set_fail_links(req);
3725 io_req_complete(req, ret);
3729 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3731 const char __user *fname;
3734 if (unlikely(sqe->ioprio || sqe->buf_index))
3736 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3739 /* open.how should be already initialised */
3740 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3741 req->open.how.flags |= O_LARGEFILE;
3743 req->open.dfd = READ_ONCE(sqe->fd);
3744 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3745 req->open.filename = getname(fname);
3746 if (IS_ERR(req->open.filename)) {
3747 ret = PTR_ERR(req->open.filename);
3748 req->open.filename = NULL;
3751 req->open.nofile = rlimit(RLIMIT_NOFILE);
3752 req->flags |= REQ_F_NEED_CLEANUP;
3756 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3760 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3762 mode = READ_ONCE(sqe->len);
3763 flags = READ_ONCE(sqe->open_flags);
3764 req->open.how = build_open_how(flags, mode);
3765 return __io_openat_prep(req, sqe);
3768 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3770 struct open_how __user *how;
3774 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3776 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3777 len = READ_ONCE(sqe->len);
3778 if (len < OPEN_HOW_SIZE_VER0)
3781 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3786 return __io_openat_prep(req, sqe);
3789 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3791 struct open_flags op;
3798 ret = build_open_flags(&req->open.how, &op);
3802 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3806 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3809 ret = PTR_ERR(file);
3811 fsnotify_open(file);
3812 fd_install(ret, file);
3815 putname(req->open.filename);
3816 req->flags &= ~REQ_F_NEED_CLEANUP;
3818 req_set_fail_links(req);
3819 io_req_complete(req, ret);
3823 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3825 return io_openat2(req, force_nonblock);
3828 static int io_remove_buffers_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 struct io_provide_buf *p = &req->pbuf;
3834 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3837 tmp = READ_ONCE(sqe->fd);
3838 if (!tmp || tmp > USHRT_MAX)
3841 memset(p, 0, sizeof(*p));
3843 p->bgid = READ_ONCE(sqe->buf_group);
3847 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3848 int bgid, unsigned nbufs)
3852 /* shouldn't happen */
3856 /* the head kbuf is the list itself */
3857 while (!list_empty(&buf->list)) {
3858 struct io_buffer *nxt;
3860 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3861 list_del(&nxt->list);
3868 idr_remove(&ctx->io_buffer_idr, bgid);
3873 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3874 struct io_comp_state *cs)
3876 struct io_provide_buf *p = &req->pbuf;
3877 struct io_ring_ctx *ctx = req->ctx;
3878 struct io_buffer *head;
3881 io_ring_submit_lock(ctx, !force_nonblock);
3883 lockdep_assert_held(&ctx->uring_lock);
3886 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3888 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3890 io_ring_submit_lock(ctx, !force_nonblock);
3892 req_set_fail_links(req);
3893 __io_req_complete(req, ret, 0, cs);
3897 static int io_provide_buffers_prep(struct io_kiocb *req,
3898 const struct io_uring_sqe *sqe)
3900 struct io_provide_buf *p = &req->pbuf;
3903 if (sqe->ioprio || sqe->rw_flags)
3906 tmp = READ_ONCE(sqe->fd);
3907 if (!tmp || tmp > USHRT_MAX)
3910 p->addr = READ_ONCE(sqe->addr);
3911 p->len = READ_ONCE(sqe->len);
3913 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3916 p->bgid = READ_ONCE(sqe->buf_group);
3917 tmp = READ_ONCE(sqe->off);
3918 if (tmp > USHRT_MAX)
3924 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3926 struct io_buffer *buf;
3927 u64 addr = pbuf->addr;
3928 int i, bid = pbuf->bid;
3930 for (i = 0; i < pbuf->nbufs; i++) {
3931 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3936 buf->len = pbuf->len;
3941 INIT_LIST_HEAD(&buf->list);
3944 list_add_tail(&buf->list, &(*head)->list);
3948 return i ? i : -ENOMEM;
3951 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3952 struct io_comp_state *cs)
3954 struct io_provide_buf *p = &req->pbuf;
3955 struct io_ring_ctx *ctx = req->ctx;
3956 struct io_buffer *head, *list;
3959 io_ring_submit_lock(ctx, !force_nonblock);
3961 lockdep_assert_held(&ctx->uring_lock);
3963 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3965 ret = io_add_buffers(p, &head);
3970 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3973 __io_remove_buffers(ctx, head, p->bgid, -1U);
3978 io_ring_submit_unlock(ctx, !force_nonblock);
3980 req_set_fail_links(req);
3981 __io_req_complete(req, ret, 0, cs);
3985 static int io_epoll_ctl_prep(struct io_kiocb *req,
3986 const struct io_uring_sqe *sqe)
3988 #if defined(CONFIG_EPOLL)
3989 if (sqe->ioprio || sqe->buf_index)
3991 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3994 req->epoll.epfd = READ_ONCE(sqe->fd);
3995 req->epoll.op = READ_ONCE(sqe->len);
3996 req->epoll.fd = READ_ONCE(sqe->off);
3998 if (ep_op_has_event(req->epoll.op)) {
3999 struct epoll_event __user *ev;
4001 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4002 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4012 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4013 struct io_comp_state *cs)
4015 #if defined(CONFIG_EPOLL)
4016 struct io_epoll *ie = &req->epoll;
4019 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4020 if (force_nonblock && ret == -EAGAIN)
4024 req_set_fail_links(req);
4025 __io_req_complete(req, ret, 0, cs);
4032 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4034 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4035 if (sqe->ioprio || sqe->buf_index || sqe->off)
4037 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4040 req->madvise.addr = READ_ONCE(sqe->addr);
4041 req->madvise.len = READ_ONCE(sqe->len);
4042 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4049 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4051 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4052 struct io_madvise *ma = &req->madvise;
4058 ret = do_madvise(ma->addr, ma->len, ma->advice);
4060 req_set_fail_links(req);
4061 io_req_complete(req, ret);
4068 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4070 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->fadvise.offset = READ_ONCE(sqe->off);
4076 req->fadvise.len = READ_ONCE(sqe->len);
4077 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4081 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4083 struct io_fadvise *fa = &req->fadvise;
4086 if (force_nonblock) {
4087 switch (fa->advice) {
4088 case POSIX_FADV_NORMAL:
4089 case POSIX_FADV_RANDOM:
4090 case POSIX_FADV_SEQUENTIAL:
4097 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4099 req_set_fail_links(req);
4100 io_req_complete(req, ret);
4104 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4106 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4108 if (sqe->ioprio || sqe->buf_index)
4110 if (req->flags & REQ_F_FIXED_FILE)
4113 req->statx.dfd = READ_ONCE(sqe->fd);
4114 req->statx.mask = READ_ONCE(sqe->len);
4115 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4116 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4117 req->statx.flags = READ_ONCE(sqe->statx_flags);
4122 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4124 struct io_statx *ctx = &req->statx;
4127 if (force_nonblock) {
4128 /* only need file table for an actual valid fd */
4129 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4130 req->flags |= REQ_F_NO_FILE_TABLE;
4134 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4138 req_set_fail_links(req);
4139 io_req_complete(req, ret);
4143 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4146 * If we queue this for async, it must not be cancellable. That would
4147 * leave the 'file' in an undeterminate state, and here need to modify
4148 * io_wq_work.flags, so initialize io_wq_work firstly.
4150 io_req_init_async(req);
4151 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4153 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4155 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4156 sqe->rw_flags || sqe->buf_index)
4158 if (req->flags & REQ_F_FIXED_FILE)
4161 req->close.fd = READ_ONCE(sqe->fd);
4162 if ((req->file && req->file->f_op == &io_uring_fops))
4165 req->close.put_file = NULL;
4169 static int io_close(struct io_kiocb *req, bool force_nonblock,
4170 struct io_comp_state *cs)
4172 struct io_close *close = &req->close;
4175 /* might be already done during nonblock submission */
4176 if (!close->put_file) {
4177 ret = __close_fd_get_file(close->fd, &close->put_file);
4179 return (ret == -ENOENT) ? -EBADF : ret;
4182 /* if the file has a flush method, be safe and punt to async */
4183 if (close->put_file->f_op->flush && force_nonblock) {
4184 /* was never set, but play safe */
4185 req->flags &= ~REQ_F_NOWAIT;
4186 /* avoid grabbing files - we don't need the files */
4187 req->flags |= REQ_F_NO_FILE_TABLE;
4191 /* No ->flush() or already async, safely close from here */
4192 ret = filp_close(close->put_file, req->work.identity->files);
4194 req_set_fail_links(req);
4195 fput(close->put_file);
4196 close->put_file = NULL;
4197 __io_req_complete(req, ret, 0, cs);
4201 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4203 struct io_ring_ctx *ctx = req->ctx;
4208 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4210 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4213 req->sync.off = READ_ONCE(sqe->off);
4214 req->sync.len = READ_ONCE(sqe->len);
4215 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4219 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4223 /* sync_file_range always requires a blocking context */
4227 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4230 req_set_fail_links(req);
4231 io_req_complete(req, ret);
4235 #if defined(CONFIG_NET)
4236 static int io_setup_async_msg(struct io_kiocb *req,
4237 struct io_async_msghdr *kmsg)
4239 struct io_async_msghdr *async_msg = req->async_data;
4243 if (io_alloc_async_data(req)) {
4244 if (kmsg->iov != kmsg->fast_iov)
4248 async_msg = req->async_data;
4249 req->flags |= REQ_F_NEED_CLEANUP;
4250 memcpy(async_msg, kmsg, sizeof(*kmsg));
4254 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4255 struct io_async_msghdr *iomsg)
4257 iomsg->iov = iomsg->fast_iov;
4258 iomsg->msg.msg_name = &iomsg->addr;
4259 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4260 req->sr_msg.msg_flags, &iomsg->iov);
4263 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4265 struct io_async_msghdr *async_msg = req->async_data;
4266 struct io_sr_msg *sr = &req->sr_msg;
4269 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4272 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4273 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4274 sr->len = READ_ONCE(sqe->len);
4276 #ifdef CONFIG_COMPAT
4277 if (req->ctx->compat)
4278 sr->msg_flags |= MSG_CMSG_COMPAT;
4281 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4283 ret = io_sendmsg_copy_hdr(req, async_msg);
4285 req->flags |= REQ_F_NEED_CLEANUP;
4289 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4290 struct io_comp_state *cs)
4292 struct io_async_msghdr iomsg, *kmsg;
4293 struct socket *sock;
4297 sock = sock_from_file(req->file, &ret);
4298 if (unlikely(!sock))
4301 if (req->async_data) {
4302 kmsg = req->async_data;
4303 kmsg->msg.msg_name = &kmsg->addr;
4304 /* if iov is set, it's allocated already */
4306 kmsg->iov = kmsg->fast_iov;
4307 kmsg->msg.msg_iter.iov = kmsg->iov;
4309 ret = io_sendmsg_copy_hdr(req, &iomsg);
4315 flags = req->sr_msg.msg_flags;
4316 if (flags & MSG_DONTWAIT)
4317 req->flags |= REQ_F_NOWAIT;
4318 else if (force_nonblock)
4319 flags |= MSG_DONTWAIT;
4321 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4322 if (force_nonblock && ret == -EAGAIN)
4323 return io_setup_async_msg(req, kmsg);
4324 if (ret == -ERESTARTSYS)
4327 if (kmsg->iov != kmsg->fast_iov)
4329 req->flags &= ~REQ_F_NEED_CLEANUP;
4331 req_set_fail_links(req);
4332 __io_req_complete(req, ret, 0, cs);
4336 static int io_send(struct io_kiocb *req, bool force_nonblock,
4337 struct io_comp_state *cs)
4339 struct io_sr_msg *sr = &req->sr_msg;
4342 struct socket *sock;
4346 sock = sock_from_file(req->file, &ret);
4347 if (unlikely(!sock))
4350 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4354 msg.msg_name = NULL;
4355 msg.msg_control = NULL;
4356 msg.msg_controllen = 0;
4357 msg.msg_namelen = 0;
4359 flags = req->sr_msg.msg_flags;
4360 if (flags & MSG_DONTWAIT)
4361 req->flags |= REQ_F_NOWAIT;
4362 else if (force_nonblock)
4363 flags |= MSG_DONTWAIT;
4365 msg.msg_flags = flags;
4366 ret = sock_sendmsg(sock, &msg);
4367 if (force_nonblock && ret == -EAGAIN)
4369 if (ret == -ERESTARTSYS)
4373 req_set_fail_links(req);
4374 __io_req_complete(req, ret, 0, cs);
4378 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4379 struct io_async_msghdr *iomsg)
4381 struct io_sr_msg *sr = &req->sr_msg;
4382 struct iovec __user *uiov;
4386 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4387 &iomsg->uaddr, &uiov, &iov_len);
4391 if (req->flags & REQ_F_BUFFER_SELECT) {
4394 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4396 sr->len = iomsg->iov[0].iov_len;
4397 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4401 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4402 &iomsg->iov, &iomsg->msg.msg_iter,
4411 #ifdef CONFIG_COMPAT
4412 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4413 struct io_async_msghdr *iomsg)
4415 struct compat_msghdr __user *msg_compat;
4416 struct io_sr_msg *sr = &req->sr_msg;
4417 struct compat_iovec __user *uiov;
4422 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4423 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4428 uiov = compat_ptr(ptr);
4429 if (req->flags & REQ_F_BUFFER_SELECT) {
4430 compat_ssize_t clen;
4434 if (!access_ok(uiov, sizeof(*uiov)))
4436 if (__get_user(clen, &uiov->iov_len))
4440 sr->len = iomsg->iov[0].iov_len;
4443 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4444 UIO_FASTIOV, &iomsg->iov,
4445 &iomsg->msg.msg_iter, true);
4454 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4455 struct io_async_msghdr *iomsg)
4457 iomsg->msg.msg_name = &iomsg->addr;
4458 iomsg->iov = iomsg->fast_iov;
4460 #ifdef CONFIG_COMPAT
4461 if (req->ctx->compat)
4462 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4465 return __io_recvmsg_copy_hdr(req, iomsg);
4468 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4471 struct io_sr_msg *sr = &req->sr_msg;
4472 struct io_buffer *kbuf;
4474 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4479 req->flags |= REQ_F_BUFFER_SELECTED;
4483 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4485 return io_put_kbuf(req, req->sr_msg.kbuf);
4488 static int io_recvmsg_prep(struct io_kiocb *req,
4489 const struct io_uring_sqe *sqe)
4491 struct io_async_msghdr *async_msg = req->async_data;
4492 struct io_sr_msg *sr = &req->sr_msg;
4495 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4498 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4499 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4500 sr->len = READ_ONCE(sqe->len);
4501 sr->bgid = READ_ONCE(sqe->buf_group);
4503 #ifdef CONFIG_COMPAT
4504 if (req->ctx->compat)
4505 sr->msg_flags |= MSG_CMSG_COMPAT;
4508 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4510 ret = io_recvmsg_copy_hdr(req, async_msg);
4512 req->flags |= REQ_F_NEED_CLEANUP;
4516 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4517 struct io_comp_state *cs)
4519 struct io_async_msghdr iomsg, *kmsg;
4520 struct socket *sock;
4521 struct io_buffer *kbuf;
4523 int ret, cflags = 0;
4525 sock = sock_from_file(req->file, &ret);
4526 if (unlikely(!sock))
4529 if (req->async_data) {
4530 kmsg = req->async_data;
4531 kmsg->msg.msg_name = &kmsg->addr;
4532 /* if iov is set, it's allocated already */
4534 kmsg->iov = kmsg->fast_iov;
4535 kmsg->msg.msg_iter.iov = kmsg->iov;
4537 ret = io_recvmsg_copy_hdr(req, &iomsg);
4543 if (req->flags & REQ_F_BUFFER_SELECT) {
4544 kbuf = io_recv_buffer_select(req, !force_nonblock);
4546 return PTR_ERR(kbuf);
4547 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4548 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4549 1, req->sr_msg.len);
4552 flags = req->sr_msg.msg_flags;
4553 if (flags & MSG_DONTWAIT)
4554 req->flags |= REQ_F_NOWAIT;
4555 else if (force_nonblock)
4556 flags |= MSG_DONTWAIT;
4558 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4559 kmsg->uaddr, flags);
4560 if (force_nonblock && ret == -EAGAIN)
4561 return io_setup_async_msg(req, kmsg);
4562 if (ret == -ERESTARTSYS)
4565 if (req->flags & REQ_F_BUFFER_SELECTED)
4566 cflags = io_put_recv_kbuf(req);
4567 if (kmsg->iov != kmsg->fast_iov)
4569 req->flags &= ~REQ_F_NEED_CLEANUP;
4571 req_set_fail_links(req);
4572 __io_req_complete(req, ret, cflags, cs);
4576 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4577 struct io_comp_state *cs)
4579 struct io_buffer *kbuf;
4580 struct io_sr_msg *sr = &req->sr_msg;
4582 void __user *buf = sr->buf;
4583 struct socket *sock;
4586 int ret, cflags = 0;
4588 sock = sock_from_file(req->file, &ret);
4589 if (unlikely(!sock))
4592 if (req->flags & REQ_F_BUFFER_SELECT) {
4593 kbuf = io_recv_buffer_select(req, !force_nonblock);
4595 return PTR_ERR(kbuf);
4596 buf = u64_to_user_ptr(kbuf->addr);
4599 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4603 msg.msg_name = NULL;
4604 msg.msg_control = NULL;
4605 msg.msg_controllen = 0;
4606 msg.msg_namelen = 0;
4607 msg.msg_iocb = NULL;
4610 flags = req->sr_msg.msg_flags;
4611 if (flags & MSG_DONTWAIT)
4612 req->flags |= REQ_F_NOWAIT;
4613 else if (force_nonblock)
4614 flags |= MSG_DONTWAIT;
4616 ret = sock_recvmsg(sock, &msg, flags);
4617 if (force_nonblock && ret == -EAGAIN)
4619 if (ret == -ERESTARTSYS)
4622 if (req->flags & REQ_F_BUFFER_SELECTED)
4623 cflags = io_put_recv_kbuf(req);
4625 req_set_fail_links(req);
4626 __io_req_complete(req, ret, cflags, cs);
4630 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4632 struct io_accept *accept = &req->accept;
4634 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4636 if (sqe->ioprio || sqe->len || sqe->buf_index)
4639 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4640 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4641 accept->flags = READ_ONCE(sqe->accept_flags);
4642 accept->nofile = rlimit(RLIMIT_NOFILE);
4646 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4647 struct io_comp_state *cs)
4649 struct io_accept *accept = &req->accept;
4650 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4653 if (req->file->f_flags & O_NONBLOCK)
4654 req->flags |= REQ_F_NOWAIT;
4656 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4657 accept->addr_len, accept->flags,
4659 if (ret == -EAGAIN && force_nonblock)
4662 if (ret == -ERESTARTSYS)
4664 req_set_fail_links(req);
4666 __io_req_complete(req, ret, 0, cs);
4670 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4672 struct io_connect *conn = &req->connect;
4673 struct io_async_connect *io = req->async_data;
4675 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4677 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4680 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4681 conn->addr_len = READ_ONCE(sqe->addr2);
4686 return move_addr_to_kernel(conn->addr, conn->addr_len,
4690 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4691 struct io_comp_state *cs)
4693 struct io_async_connect __io, *io;
4694 unsigned file_flags;
4697 if (req->async_data) {
4698 io = req->async_data;
4700 ret = move_addr_to_kernel(req->connect.addr,
4701 req->connect.addr_len,
4708 file_flags = force_nonblock ? O_NONBLOCK : 0;
4710 ret = __sys_connect_file(req->file, &io->address,
4711 req->connect.addr_len, file_flags);
4712 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4713 if (req->async_data)
4715 if (io_alloc_async_data(req)) {
4719 io = req->async_data;
4720 memcpy(req->async_data, &__io, sizeof(__io));
4723 if (ret == -ERESTARTSYS)
4727 req_set_fail_links(req);
4728 __io_req_complete(req, ret, 0, cs);
4731 #else /* !CONFIG_NET */
4732 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4737 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4738 struct io_comp_state *cs)
4743 static int io_send(struct io_kiocb *req, bool force_nonblock,
4744 struct io_comp_state *cs)
4749 static int io_recvmsg_prep(struct io_kiocb *req,
4750 const struct io_uring_sqe *sqe)
4755 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4756 struct io_comp_state *cs)
4761 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4762 struct io_comp_state *cs)
4767 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4772 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4773 struct io_comp_state *cs)
4778 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4783 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4784 struct io_comp_state *cs)
4788 #endif /* CONFIG_NET */
4790 struct io_poll_table {
4791 struct poll_table_struct pt;
4792 struct io_kiocb *req;
4796 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4797 __poll_t mask, task_work_func_t func)
4802 /* for instances that support it check for an event match first: */
4803 if (mask && !(mask & poll->events))
4806 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4808 list_del_init(&poll->wait.entry);
4811 init_task_work(&req->task_work, func);
4812 percpu_ref_get(&req->ctx->refs);
4815 * If we using the signalfd wait_queue_head for this wakeup, then
4816 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4817 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4818 * either, as the normal wakeup will suffice.
4820 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4823 * If this fails, then the task is exiting. When a task exits, the
4824 * work gets canceled, so just cancel this request as well instead
4825 * of executing it. We can't safely execute it anyway, as we may not
4826 * have the needed state needed for it anyway.
4828 ret = io_req_task_work_add(req, twa_signal_ok);
4829 if (unlikely(ret)) {
4830 struct task_struct *tsk;
4832 WRITE_ONCE(poll->canceled, true);
4833 tsk = io_wq_get_task(req->ctx->io_wq);
4834 task_work_add(tsk, &req->task_work, 0);
4835 wake_up_process(tsk);
4840 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4841 __acquires(&req->ctx->completion_lock)
4843 struct io_ring_ctx *ctx = req->ctx;
4845 if (!req->result && !READ_ONCE(poll->canceled)) {
4846 struct poll_table_struct pt = { ._key = poll->events };
4848 req->result = vfs_poll(req->file, &pt) & poll->events;
4851 spin_lock_irq(&ctx->completion_lock);
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 add_wait_queue(poll->head, &poll->wait);
4860 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4862 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4863 if (req->opcode == IORING_OP_POLL_ADD)
4864 return req->async_data;
4865 return req->apoll->double_poll;
4868 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4870 if (req->opcode == IORING_OP_POLL_ADD)
4872 return &req->apoll->poll;
4875 static void io_poll_remove_double(struct io_kiocb *req)
4877 struct io_poll_iocb *poll = io_poll_get_double(req);
4879 lockdep_assert_held(&req->ctx->completion_lock);
4881 if (poll && poll->head) {
4882 struct wait_queue_head *head = poll->head;
4884 spin_lock(&head->lock);
4885 list_del_init(&poll->wait.entry);
4886 if (poll->wait.private)
4887 refcount_dec(&req->refs);
4889 spin_unlock(&head->lock);
4893 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4895 struct io_ring_ctx *ctx = req->ctx;
4897 io_poll_remove_double(req);
4898 req->poll.done = true;
4899 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4900 io_commit_cqring(ctx);
4903 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4905 struct io_ring_ctx *ctx = req->ctx;
4907 if (io_poll_rewait(req, &req->poll)) {
4908 spin_unlock_irq(&ctx->completion_lock);
4912 hash_del(&req->hash_node);
4913 io_poll_complete(req, req->result, 0);
4914 spin_unlock_irq(&ctx->completion_lock);
4916 *nxt = io_put_req_find_next(req);
4917 io_cqring_ev_posted(ctx);
4920 static void io_poll_task_func(struct callback_head *cb)
4922 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4923 struct io_ring_ctx *ctx = req->ctx;
4924 struct io_kiocb *nxt = NULL;
4926 io_poll_task_handler(req, &nxt);
4928 __io_req_task_submit(nxt);
4929 percpu_ref_put(&ctx->refs);
4932 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4933 int sync, void *key)
4935 struct io_kiocb *req = wait->private;
4936 struct io_poll_iocb *poll = io_poll_get_single(req);
4937 __poll_t mask = key_to_poll(key);
4939 /* for instances that support it check for an event match first: */
4940 if (mask && !(mask & poll->events))
4943 list_del_init(&wait->entry);
4945 if (poll && poll->head) {
4948 spin_lock(&poll->head->lock);
4949 done = list_empty(&poll->wait.entry);
4951 list_del_init(&poll->wait.entry);
4952 /* make sure double remove sees this as being gone */
4953 wait->private = NULL;
4954 spin_unlock(&poll->head->lock);
4956 __io_async_wake(req, poll, mask, io_poll_task_func);
4958 refcount_dec(&req->refs);
4962 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4963 wait_queue_func_t wake_func)
4967 poll->canceled = false;
4968 poll->events = events;
4969 INIT_LIST_HEAD(&poll->wait.entry);
4970 init_waitqueue_func_entry(&poll->wait, wake_func);
4973 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4974 struct wait_queue_head *head,
4975 struct io_poll_iocb **poll_ptr)
4977 struct io_kiocb *req = pt->req;
4980 * If poll->head is already set, it's because the file being polled
4981 * uses multiple waitqueues for poll handling (eg one for read, one
4982 * for write). Setup a separate io_poll_iocb if this happens.
4984 if (unlikely(poll->head)) {
4985 /* already have a 2nd entry, fail a third attempt */
4987 pt->error = -EINVAL;
4990 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4992 pt->error = -ENOMEM;
4995 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4996 refcount_inc(&req->refs);
4997 poll->wait.private = req;
5004 if (poll->events & EPOLLEXCLUSIVE)
5005 add_wait_queue_exclusive(head, &poll->wait);
5007 add_wait_queue(head, &poll->wait);
5010 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5011 struct poll_table_struct *p)
5013 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5014 struct async_poll *apoll = pt->req->apoll;
5016 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5019 static void io_async_task_func(struct callback_head *cb)
5021 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5022 struct async_poll *apoll = req->apoll;
5023 struct io_ring_ctx *ctx = req->ctx;
5025 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5027 if (io_poll_rewait(req, &apoll->poll)) {
5028 spin_unlock_irq(&ctx->completion_lock);
5029 percpu_ref_put(&ctx->refs);
5033 /* If req is still hashed, it cannot have been canceled. Don't check. */
5034 if (hash_hashed(&req->hash_node))
5035 hash_del(&req->hash_node);
5037 io_poll_remove_double(req);
5038 spin_unlock_irq(&ctx->completion_lock);
5040 if (!READ_ONCE(apoll->poll.canceled))
5041 __io_req_task_submit(req);
5043 __io_req_task_cancel(req, -ECANCELED);
5045 percpu_ref_put(&ctx->refs);
5046 kfree(apoll->double_poll);
5050 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5053 struct io_kiocb *req = wait->private;
5054 struct io_poll_iocb *poll = &req->apoll->poll;
5056 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5059 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5062 static void io_poll_req_insert(struct io_kiocb *req)
5064 struct io_ring_ctx *ctx = req->ctx;
5065 struct hlist_head *list;
5067 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5068 hlist_add_head(&req->hash_node, list);
5071 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5072 struct io_poll_iocb *poll,
5073 struct io_poll_table *ipt, __poll_t mask,
5074 wait_queue_func_t wake_func)
5075 __acquires(&ctx->completion_lock)
5077 struct io_ring_ctx *ctx = req->ctx;
5078 bool cancel = false;
5080 io_init_poll_iocb(poll, mask, wake_func);
5081 poll->file = req->file;
5082 poll->wait.private = req;
5084 ipt->pt._key = mask;
5086 ipt->error = -EINVAL;
5088 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5090 spin_lock_irq(&ctx->completion_lock);
5091 if (likely(poll->head)) {
5092 spin_lock(&poll->head->lock);
5093 if (unlikely(list_empty(&poll->wait.entry))) {
5099 if (mask || ipt->error)
5100 list_del_init(&poll->wait.entry);
5102 WRITE_ONCE(poll->canceled, true);
5103 else if (!poll->done) /* actually waiting for an event */
5104 io_poll_req_insert(req);
5105 spin_unlock(&poll->head->lock);
5111 static bool io_arm_poll_handler(struct io_kiocb *req)
5113 const struct io_op_def *def = &io_op_defs[req->opcode];
5114 struct io_ring_ctx *ctx = req->ctx;
5115 struct async_poll *apoll;
5116 struct io_poll_table ipt;
5120 if (!req->file || !file_can_poll(req->file))
5122 if (req->flags & REQ_F_POLLED)
5126 else if (def->pollout)
5130 /* if we can't nonblock try, then no point in arming a poll handler */
5131 if (!io_file_supports_async(req->file, rw))
5134 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5135 if (unlikely(!apoll))
5137 apoll->double_poll = NULL;
5139 req->flags |= REQ_F_POLLED;
5141 INIT_HLIST_NODE(&req->hash_node);
5145 mask |= POLLIN | POLLRDNORM;
5147 mask |= POLLOUT | POLLWRNORM;
5149 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5150 if ((req->opcode == IORING_OP_RECVMSG) &&
5151 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5154 mask |= POLLERR | POLLPRI;
5156 ipt.pt._qproc = io_async_queue_proc;
5158 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5160 if (ret || ipt.error) {
5161 io_poll_remove_double(req);
5162 spin_unlock_irq(&ctx->completion_lock);
5163 kfree(apoll->double_poll);
5167 spin_unlock_irq(&ctx->completion_lock);
5168 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5169 apoll->poll.events);
5173 static bool __io_poll_remove_one(struct io_kiocb *req,
5174 struct io_poll_iocb *poll)
5176 bool do_complete = false;
5178 spin_lock(&poll->head->lock);
5179 WRITE_ONCE(poll->canceled, true);
5180 if (!list_empty(&poll->wait.entry)) {
5181 list_del_init(&poll->wait.entry);
5184 spin_unlock(&poll->head->lock);
5185 hash_del(&req->hash_node);
5189 static bool io_poll_remove_one(struct io_kiocb *req)
5193 io_poll_remove_double(req);
5195 if (req->opcode == IORING_OP_POLL_ADD) {
5196 do_complete = __io_poll_remove_one(req, &req->poll);
5198 struct async_poll *apoll = req->apoll;
5200 /* non-poll requests have submit ref still */
5201 do_complete = __io_poll_remove_one(req, &apoll->poll);
5204 kfree(apoll->double_poll);
5210 io_cqring_fill_event(req, -ECANCELED);
5211 io_commit_cqring(req->ctx);
5212 req_set_fail_links(req);
5213 io_put_req_deferred(req, 1);
5220 * Returns true if we found and killed one or more poll requests
5222 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5224 struct hlist_node *tmp;
5225 struct io_kiocb *req;
5228 spin_lock_irq(&ctx->completion_lock);
5229 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5230 struct hlist_head *list;
5232 list = &ctx->cancel_hash[i];
5233 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5234 if (io_task_match(req, tsk))
5235 posted += io_poll_remove_one(req);
5238 spin_unlock_irq(&ctx->completion_lock);
5241 io_cqring_ev_posted(ctx);
5246 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5248 struct hlist_head *list;
5249 struct io_kiocb *req;
5251 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5252 hlist_for_each_entry(req, list, hash_node) {
5253 if (sqe_addr != req->user_data)
5255 if (io_poll_remove_one(req))
5263 static int io_poll_remove_prep(struct io_kiocb *req,
5264 const struct io_uring_sqe *sqe)
5266 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5268 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5272 req->poll.addr = READ_ONCE(sqe->addr);
5277 * Find a running poll command that matches one specified in sqe->addr,
5278 * and remove it if found.
5280 static int io_poll_remove(struct io_kiocb *req)
5282 struct io_ring_ctx *ctx = req->ctx;
5286 addr = req->poll.addr;
5287 spin_lock_irq(&ctx->completion_lock);
5288 ret = io_poll_cancel(ctx, addr);
5289 spin_unlock_irq(&ctx->completion_lock);
5292 req_set_fail_links(req);
5293 io_req_complete(req, ret);
5297 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5300 struct io_kiocb *req = wait->private;
5301 struct io_poll_iocb *poll = &req->poll;
5303 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5306 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5307 struct poll_table_struct *p)
5309 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5311 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5314 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5316 struct io_poll_iocb *poll = &req->poll;
5319 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5321 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5326 events = READ_ONCE(sqe->poll32_events);
5328 events = swahw32(events);
5330 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5331 (events & EPOLLEXCLUSIVE);
5335 static int io_poll_add(struct io_kiocb *req)
5337 struct io_poll_iocb *poll = &req->poll;
5338 struct io_ring_ctx *ctx = req->ctx;
5339 struct io_poll_table ipt;
5342 INIT_HLIST_NODE(&req->hash_node);
5343 ipt.pt._qproc = io_poll_queue_proc;
5345 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5348 if (mask) { /* no async, we'd stolen it */
5350 io_poll_complete(req, mask, 0);
5352 spin_unlock_irq(&ctx->completion_lock);
5355 io_cqring_ev_posted(ctx);
5361 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5363 struct io_timeout_data *data = container_of(timer,
5364 struct io_timeout_data, timer);
5365 struct io_kiocb *req = data->req;
5366 struct io_ring_ctx *ctx = req->ctx;
5367 unsigned long flags;
5369 spin_lock_irqsave(&ctx->completion_lock, flags);
5370 list_del_init(&req->timeout.list);
5371 atomic_set(&req->ctx->cq_timeouts,
5372 atomic_read(&req->ctx->cq_timeouts) + 1);
5374 io_cqring_fill_event(req, -ETIME);
5375 io_commit_cqring(ctx);
5376 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5378 io_cqring_ev_posted(ctx);
5379 req_set_fail_links(req);
5381 return HRTIMER_NORESTART;
5384 static int __io_timeout_cancel(struct io_kiocb *req)
5386 struct io_timeout_data *io = req->async_data;
5389 ret = hrtimer_try_to_cancel(&io->timer);
5392 list_del_init(&req->timeout.list);
5394 req_set_fail_links(req);
5395 io_cqring_fill_event(req, -ECANCELED);
5396 io_put_req_deferred(req, 1);
5400 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5402 struct io_kiocb *req;
5405 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5406 if (user_data == req->user_data) {
5415 return __io_timeout_cancel(req);
5418 static int io_timeout_remove_prep(struct io_kiocb *req,
5419 const struct io_uring_sqe *sqe)
5421 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5423 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5425 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5428 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5433 * Remove or update an existing timeout command
5435 static int io_timeout_remove(struct io_kiocb *req)
5437 struct io_ring_ctx *ctx = req->ctx;
5440 spin_lock_irq(&ctx->completion_lock);
5441 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5443 io_cqring_fill_event(req, ret);
5444 io_commit_cqring(ctx);
5445 spin_unlock_irq(&ctx->completion_lock);
5446 io_cqring_ev_posted(ctx);
5448 req_set_fail_links(req);
5453 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5454 bool is_timeout_link)
5456 struct io_timeout_data *data;
5458 u32 off = READ_ONCE(sqe->off);
5460 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5462 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5464 if (off && is_timeout_link)
5466 flags = READ_ONCE(sqe->timeout_flags);
5467 if (flags & ~IORING_TIMEOUT_ABS)
5470 req->timeout.off = off;
5472 if (!req->async_data && io_alloc_async_data(req))
5475 data = req->async_data;
5478 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5481 if (flags & IORING_TIMEOUT_ABS)
5482 data->mode = HRTIMER_MODE_ABS;
5484 data->mode = HRTIMER_MODE_REL;
5486 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5490 static int io_timeout(struct io_kiocb *req)
5492 struct io_ring_ctx *ctx = req->ctx;
5493 struct io_timeout_data *data = req->async_data;
5494 struct list_head *entry;
5495 u32 tail, off = req->timeout.off;
5497 spin_lock_irq(&ctx->completion_lock);
5500 * sqe->off holds how many events that need to occur for this
5501 * timeout event to be satisfied. If it isn't set, then this is
5502 * a pure timeout request, sequence isn't used.
5504 if (io_is_timeout_noseq(req)) {
5505 entry = ctx->timeout_list.prev;
5509 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5510 req->timeout.target_seq = tail + off;
5513 * Insertion sort, ensuring the first entry in the list is always
5514 * the one we need first.
5516 list_for_each_prev(entry, &ctx->timeout_list) {
5517 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5520 if (io_is_timeout_noseq(nxt))
5522 /* nxt.seq is behind @tail, otherwise would've been completed */
5523 if (off >= nxt->timeout.target_seq - tail)
5527 list_add(&req->timeout.list, entry);
5528 data->timer.function = io_timeout_fn;
5529 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5530 spin_unlock_irq(&ctx->completion_lock);
5534 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5536 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5538 return req->user_data == (unsigned long) data;
5541 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5543 enum io_wq_cancel cancel_ret;
5546 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5547 switch (cancel_ret) {
5548 case IO_WQ_CANCEL_OK:
5551 case IO_WQ_CANCEL_RUNNING:
5554 case IO_WQ_CANCEL_NOTFOUND:
5562 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5563 struct io_kiocb *req, __u64 sqe_addr,
5566 unsigned long flags;
5569 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5570 if (ret != -ENOENT) {
5571 spin_lock_irqsave(&ctx->completion_lock, flags);
5575 spin_lock_irqsave(&ctx->completion_lock, flags);
5576 ret = io_timeout_cancel(ctx, sqe_addr);
5579 ret = io_poll_cancel(ctx, sqe_addr);
5583 io_cqring_fill_event(req, ret);
5584 io_commit_cqring(ctx);
5585 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5586 io_cqring_ev_posted(ctx);
5589 req_set_fail_links(req);
5593 static int io_async_cancel_prep(struct io_kiocb *req,
5594 const struct io_uring_sqe *sqe)
5596 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5598 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5600 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5603 req->cancel.addr = READ_ONCE(sqe->addr);
5607 static int io_async_cancel(struct io_kiocb *req)
5609 struct io_ring_ctx *ctx = req->ctx;
5611 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5615 static int io_files_update_prep(struct io_kiocb *req,
5616 const struct io_uring_sqe *sqe)
5618 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5620 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5622 if (sqe->ioprio || sqe->rw_flags)
5625 req->files_update.offset = READ_ONCE(sqe->off);
5626 req->files_update.nr_args = READ_ONCE(sqe->len);
5627 if (!req->files_update.nr_args)
5629 req->files_update.arg = READ_ONCE(sqe->addr);
5633 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5634 struct io_comp_state *cs)
5636 struct io_ring_ctx *ctx = req->ctx;
5637 struct io_uring_files_update up;
5643 up.offset = req->files_update.offset;
5644 up.fds = req->files_update.arg;
5646 mutex_lock(&ctx->uring_lock);
5647 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5648 mutex_unlock(&ctx->uring_lock);
5651 req_set_fail_links(req);
5652 __io_req_complete(req, ret, 0, cs);
5656 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5658 switch (req->opcode) {
5661 case IORING_OP_READV:
5662 case IORING_OP_READ_FIXED:
5663 case IORING_OP_READ:
5664 return io_read_prep(req, sqe);
5665 case IORING_OP_WRITEV:
5666 case IORING_OP_WRITE_FIXED:
5667 case IORING_OP_WRITE:
5668 return io_write_prep(req, sqe);
5669 case IORING_OP_POLL_ADD:
5670 return io_poll_add_prep(req, sqe);
5671 case IORING_OP_POLL_REMOVE:
5672 return io_poll_remove_prep(req, sqe);
5673 case IORING_OP_FSYNC:
5674 return io_prep_fsync(req, sqe);
5675 case IORING_OP_SYNC_FILE_RANGE:
5676 return io_prep_sfr(req, sqe);
5677 case IORING_OP_SENDMSG:
5678 case IORING_OP_SEND:
5679 return io_sendmsg_prep(req, sqe);
5680 case IORING_OP_RECVMSG:
5681 case IORING_OP_RECV:
5682 return io_recvmsg_prep(req, sqe);
5683 case IORING_OP_CONNECT:
5684 return io_connect_prep(req, sqe);
5685 case IORING_OP_TIMEOUT:
5686 return io_timeout_prep(req, sqe, false);
5687 case IORING_OP_TIMEOUT_REMOVE:
5688 return io_timeout_remove_prep(req, sqe);
5689 case IORING_OP_ASYNC_CANCEL:
5690 return io_async_cancel_prep(req, sqe);
5691 case IORING_OP_LINK_TIMEOUT:
5692 return io_timeout_prep(req, sqe, true);
5693 case IORING_OP_ACCEPT:
5694 return io_accept_prep(req, sqe);
5695 case IORING_OP_FALLOCATE:
5696 return io_fallocate_prep(req, sqe);
5697 case IORING_OP_OPENAT:
5698 return io_openat_prep(req, sqe);
5699 case IORING_OP_CLOSE:
5700 return io_close_prep(req, sqe);
5701 case IORING_OP_FILES_UPDATE:
5702 return io_files_update_prep(req, sqe);
5703 case IORING_OP_STATX:
5704 return io_statx_prep(req, sqe);
5705 case IORING_OP_FADVISE:
5706 return io_fadvise_prep(req, sqe);
5707 case IORING_OP_MADVISE:
5708 return io_madvise_prep(req, sqe);
5709 case IORING_OP_OPENAT2:
5710 return io_openat2_prep(req, sqe);
5711 case IORING_OP_EPOLL_CTL:
5712 return io_epoll_ctl_prep(req, sqe);
5713 case IORING_OP_SPLICE:
5714 return io_splice_prep(req, sqe);
5715 case IORING_OP_PROVIDE_BUFFERS:
5716 return io_provide_buffers_prep(req, sqe);
5717 case IORING_OP_REMOVE_BUFFERS:
5718 return io_remove_buffers_prep(req, sqe);
5720 return io_tee_prep(req, sqe);
5723 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5728 static int io_req_defer_prep(struct io_kiocb *req,
5729 const struct io_uring_sqe *sqe)
5733 if (io_alloc_async_data(req))
5735 return io_req_prep(req, sqe);
5738 static u32 io_get_sequence(struct io_kiocb *req)
5740 struct io_kiocb *pos;
5741 struct io_ring_ctx *ctx = req->ctx;
5742 u32 total_submitted, nr_reqs = 1;
5744 if (req->flags & REQ_F_LINK_HEAD)
5745 list_for_each_entry(pos, &req->link_list, link_list)
5748 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5749 return total_submitted - nr_reqs;
5752 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5754 struct io_ring_ctx *ctx = req->ctx;
5755 struct io_defer_entry *de;
5759 /* Still need defer if there is pending req in defer list. */
5760 if (likely(list_empty_careful(&ctx->defer_list) &&
5761 !(req->flags & REQ_F_IO_DRAIN)))
5764 seq = io_get_sequence(req);
5765 /* Still a chance to pass the sequence check */
5766 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5769 if (!req->async_data) {
5770 ret = io_req_defer_prep(req, sqe);
5774 io_prep_async_link(req);
5775 de = kmalloc(sizeof(*de), GFP_KERNEL);
5779 spin_lock_irq(&ctx->completion_lock);
5780 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5781 spin_unlock_irq(&ctx->completion_lock);
5783 io_queue_async_work(req);
5784 return -EIOCBQUEUED;
5787 trace_io_uring_defer(ctx, req, req->user_data);
5790 list_add_tail(&de->list, &ctx->defer_list);
5791 spin_unlock_irq(&ctx->completion_lock);
5792 return -EIOCBQUEUED;
5795 static void io_req_drop_files(struct io_kiocb *req)
5797 struct io_ring_ctx *ctx = req->ctx;
5798 unsigned long flags;
5800 spin_lock_irqsave(&ctx->inflight_lock, flags);
5801 list_del(&req->inflight_entry);
5802 if (waitqueue_active(&ctx->inflight_wait))
5803 wake_up(&ctx->inflight_wait);
5804 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5805 req->flags &= ~REQ_F_INFLIGHT;
5806 put_files_struct(req->work.identity->files);
5807 put_nsproxy(req->work.identity->nsproxy);
5808 req->work.flags &= ~IO_WQ_WORK_FILES;
5811 static void __io_clean_op(struct io_kiocb *req)
5813 if (req->flags & REQ_F_BUFFER_SELECTED) {
5814 switch (req->opcode) {
5815 case IORING_OP_READV:
5816 case IORING_OP_READ_FIXED:
5817 case IORING_OP_READ:
5818 kfree((void *)(unsigned long)req->rw.addr);
5820 case IORING_OP_RECVMSG:
5821 case IORING_OP_RECV:
5822 kfree(req->sr_msg.kbuf);
5825 req->flags &= ~REQ_F_BUFFER_SELECTED;
5828 if (req->flags & REQ_F_NEED_CLEANUP) {
5829 switch (req->opcode) {
5830 case IORING_OP_READV:
5831 case IORING_OP_READ_FIXED:
5832 case IORING_OP_READ:
5833 case IORING_OP_WRITEV:
5834 case IORING_OP_WRITE_FIXED:
5835 case IORING_OP_WRITE: {
5836 struct io_async_rw *io = req->async_data;
5838 kfree(io->free_iovec);
5841 case IORING_OP_RECVMSG:
5842 case IORING_OP_SENDMSG: {
5843 struct io_async_msghdr *io = req->async_data;
5844 if (io->iov != io->fast_iov)
5848 case IORING_OP_SPLICE:
5850 io_put_file(req, req->splice.file_in,
5851 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5853 case IORING_OP_OPENAT:
5854 case IORING_OP_OPENAT2:
5855 if (req->open.filename)
5856 putname(req->open.filename);
5859 req->flags &= ~REQ_F_NEED_CLEANUP;
5862 if (req->flags & REQ_F_INFLIGHT)
5863 io_req_drop_files(req);
5866 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5867 struct io_comp_state *cs)
5869 struct io_ring_ctx *ctx = req->ctx;
5872 switch (req->opcode) {
5874 ret = io_nop(req, cs);
5876 case IORING_OP_READV:
5877 case IORING_OP_READ_FIXED:
5878 case IORING_OP_READ:
5879 ret = io_read(req, force_nonblock, cs);
5881 case IORING_OP_WRITEV:
5882 case IORING_OP_WRITE_FIXED:
5883 case IORING_OP_WRITE:
5884 ret = io_write(req, force_nonblock, cs);
5886 case IORING_OP_FSYNC:
5887 ret = io_fsync(req, force_nonblock);
5889 case IORING_OP_POLL_ADD:
5890 ret = io_poll_add(req);
5892 case IORING_OP_POLL_REMOVE:
5893 ret = io_poll_remove(req);
5895 case IORING_OP_SYNC_FILE_RANGE:
5896 ret = io_sync_file_range(req, force_nonblock);
5898 case IORING_OP_SENDMSG:
5899 ret = io_sendmsg(req, force_nonblock, cs);
5901 case IORING_OP_SEND:
5902 ret = io_send(req, force_nonblock, cs);
5904 case IORING_OP_RECVMSG:
5905 ret = io_recvmsg(req, force_nonblock, cs);
5907 case IORING_OP_RECV:
5908 ret = io_recv(req, force_nonblock, cs);
5910 case IORING_OP_TIMEOUT:
5911 ret = io_timeout(req);
5913 case IORING_OP_TIMEOUT_REMOVE:
5914 ret = io_timeout_remove(req);
5916 case IORING_OP_ACCEPT:
5917 ret = io_accept(req, force_nonblock, cs);
5919 case IORING_OP_CONNECT:
5920 ret = io_connect(req, force_nonblock, cs);
5922 case IORING_OP_ASYNC_CANCEL:
5923 ret = io_async_cancel(req);
5925 case IORING_OP_FALLOCATE:
5926 ret = io_fallocate(req, force_nonblock);
5928 case IORING_OP_OPENAT:
5929 ret = io_openat(req, force_nonblock);
5931 case IORING_OP_CLOSE:
5932 ret = io_close(req, force_nonblock, cs);
5934 case IORING_OP_FILES_UPDATE:
5935 ret = io_files_update(req, force_nonblock, cs);
5937 case IORING_OP_STATX:
5938 ret = io_statx(req, force_nonblock);
5940 case IORING_OP_FADVISE:
5941 ret = io_fadvise(req, force_nonblock);
5943 case IORING_OP_MADVISE:
5944 ret = io_madvise(req, force_nonblock);
5946 case IORING_OP_OPENAT2:
5947 ret = io_openat2(req, force_nonblock);
5949 case IORING_OP_EPOLL_CTL:
5950 ret = io_epoll_ctl(req, force_nonblock, cs);
5952 case IORING_OP_SPLICE:
5953 ret = io_splice(req, force_nonblock);
5955 case IORING_OP_PROVIDE_BUFFERS:
5956 ret = io_provide_buffers(req, force_nonblock, cs);
5958 case IORING_OP_REMOVE_BUFFERS:
5959 ret = io_remove_buffers(req, force_nonblock, cs);
5962 ret = io_tee(req, force_nonblock);
5972 /* If the op doesn't have a file, we're not polling for it */
5973 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5974 const bool in_async = io_wq_current_is_worker();
5976 /* workqueue context doesn't hold uring_lock, grab it now */
5978 mutex_lock(&ctx->uring_lock);
5980 io_iopoll_req_issued(req);
5983 mutex_unlock(&ctx->uring_lock);
5989 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5991 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5992 struct io_kiocb *timeout;
5995 timeout = io_prep_linked_timeout(req);
5997 io_queue_linked_timeout(timeout);
5999 /* if NO_CANCEL is set, we must still run the work */
6000 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6001 IO_WQ_WORK_CANCEL) {
6007 ret = io_issue_sqe(req, false, NULL);
6009 * We can get EAGAIN for polled IO even though we're
6010 * forcing a sync submission from here, since we can't
6011 * wait for request slots on the block side.
6020 req_set_fail_links(req);
6021 io_req_complete(req, ret);
6024 return io_steal_work(req);
6027 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6030 struct fixed_file_table *table;
6032 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6033 return table->files[index & IORING_FILE_TABLE_MASK];
6036 static struct file *io_file_get(struct io_submit_state *state,
6037 struct io_kiocb *req, int fd, bool fixed)
6039 struct io_ring_ctx *ctx = req->ctx;
6043 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6045 fd = array_index_nospec(fd, ctx->nr_user_files);
6046 file = io_file_from_index(ctx, fd);
6048 req->fixed_file_refs = &ctx->file_data->node->refs;
6049 percpu_ref_get(req->fixed_file_refs);
6052 trace_io_uring_file_get(ctx, fd);
6053 file = __io_file_get(state, fd);
6059 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6064 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6065 if (unlikely(!fixed && io_async_submit(req->ctx)))
6068 req->file = io_file_get(state, req, fd, fixed);
6069 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6074 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6076 struct io_timeout_data *data = container_of(timer,
6077 struct io_timeout_data, timer);
6078 struct io_kiocb *req = data->req;
6079 struct io_ring_ctx *ctx = req->ctx;
6080 struct io_kiocb *prev = NULL;
6081 unsigned long flags;
6083 spin_lock_irqsave(&ctx->completion_lock, flags);
6086 * We don't expect the list to be empty, that will only happen if we
6087 * race with the completion of the linked work.
6089 if (!list_empty(&req->link_list)) {
6090 prev = list_entry(req->link_list.prev, struct io_kiocb,
6092 if (refcount_inc_not_zero(&prev->refs)) {
6093 list_del_init(&req->link_list);
6094 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6099 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6102 req_set_fail_links(prev);
6103 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6106 io_req_complete(req, -ETIME);
6108 return HRTIMER_NORESTART;
6111 static void __io_queue_linked_timeout(struct io_kiocb *req)
6114 * If the list is now empty, then our linked request finished before
6115 * we got a chance to setup the timer
6117 if (!list_empty(&req->link_list)) {
6118 struct io_timeout_data *data = req->async_data;
6120 data->timer.function = io_link_timeout_fn;
6121 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6126 static void io_queue_linked_timeout(struct io_kiocb *req)
6128 struct io_ring_ctx *ctx = req->ctx;
6130 spin_lock_irq(&ctx->completion_lock);
6131 __io_queue_linked_timeout(req);
6132 spin_unlock_irq(&ctx->completion_lock);
6134 /* drop submission reference */
6138 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6140 struct io_kiocb *nxt;
6142 if (!(req->flags & REQ_F_LINK_HEAD))
6144 if (req->flags & REQ_F_LINK_TIMEOUT)
6147 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6149 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6152 req->flags |= REQ_F_LINK_TIMEOUT;
6156 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6158 struct io_kiocb *linked_timeout;
6159 struct io_kiocb *nxt;
6160 const struct cred *old_creds = NULL;
6164 linked_timeout = io_prep_linked_timeout(req);
6166 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.identity->creds &&
6167 req->work.identity->creds != current_cred()) {
6169 revert_creds(old_creds);
6170 if (old_creds == req->work.identity->creds)
6171 old_creds = NULL; /* restored original creds */
6173 old_creds = override_creds(req->work.identity->creds);
6174 req->work.flags |= IO_WQ_WORK_CREDS;
6177 ret = io_issue_sqe(req, true, cs);
6180 * We async punt it if the file wasn't marked NOWAIT, or if the file
6181 * doesn't support non-blocking read/write attempts
6183 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6184 if (!io_arm_poll_handler(req)) {
6187 * Queued up for async execution, worker will release
6188 * submit reference when the iocb is actually submitted.
6190 io_queue_async_work(req);
6194 io_queue_linked_timeout(linked_timeout);
6198 if (unlikely(ret)) {
6199 /* un-prep timeout, so it'll be killed as any other linked */
6200 req->flags &= ~REQ_F_LINK_TIMEOUT;
6201 req_set_fail_links(req);
6203 io_req_complete(req, ret);
6207 /* drop submission reference */
6208 nxt = io_put_req_find_next(req);
6210 io_queue_linked_timeout(linked_timeout);
6215 if (req->flags & REQ_F_FORCE_ASYNC)
6221 revert_creds(old_creds);
6224 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6225 struct io_comp_state *cs)
6229 ret = io_req_defer(req, sqe);
6231 if (ret != -EIOCBQUEUED) {
6233 req_set_fail_links(req);
6235 io_req_complete(req, ret);
6237 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6238 if (!req->async_data) {
6239 ret = io_req_defer_prep(req, sqe);
6245 * Never try inline submit of IOSQE_ASYNC is set, go straight
6246 * to async execution.
6248 io_req_init_async(req);
6249 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6250 io_queue_async_work(req);
6253 ret = io_req_prep(req, sqe);
6257 __io_queue_sqe(req, cs);
6261 static inline void io_queue_link_head(struct io_kiocb *req,
6262 struct io_comp_state *cs)
6264 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6266 io_req_complete(req, -ECANCELED);
6268 io_queue_sqe(req, NULL, cs);
6271 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6272 struct io_kiocb **link, struct io_comp_state *cs)
6274 struct io_ring_ctx *ctx = req->ctx;
6278 * If we already have a head request, queue this one for async
6279 * submittal once the head completes. If we don't have a head but
6280 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6281 * submitted sync once the chain is complete. If none of those
6282 * conditions are true (normal request), then just queue it.
6285 struct io_kiocb *head = *link;
6288 * Taking sequential execution of a link, draining both sides
6289 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6290 * requests in the link. So, it drains the head and the
6291 * next after the link request. The last one is done via
6292 * drain_next flag to persist the effect across calls.
6294 if (req->flags & REQ_F_IO_DRAIN) {
6295 head->flags |= REQ_F_IO_DRAIN;
6296 ctx->drain_next = 1;
6298 ret = io_req_defer_prep(req, sqe);
6299 if (unlikely(ret)) {
6300 /* fail even hard links since we don't submit */
6301 head->flags |= REQ_F_FAIL_LINK;
6304 trace_io_uring_link(ctx, req, head);
6305 list_add_tail(&req->link_list, &head->link_list);
6307 /* last request of a link, enqueue the link */
6308 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6309 io_queue_link_head(head, cs);
6313 if (unlikely(ctx->drain_next)) {
6314 req->flags |= REQ_F_IO_DRAIN;
6315 ctx->drain_next = 0;
6317 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6318 req->flags |= REQ_F_LINK_HEAD;
6319 INIT_LIST_HEAD(&req->link_list);
6321 ret = io_req_defer_prep(req, sqe);
6323 req->flags |= REQ_F_FAIL_LINK;
6326 io_queue_sqe(req, sqe, cs);
6334 * Batched submission is done, ensure local IO is flushed out.
6336 static void io_submit_state_end(struct io_submit_state *state)
6338 if (!list_empty(&state->comp.list))
6339 io_submit_flush_completions(&state->comp);
6340 blk_finish_plug(&state->plug);
6341 io_state_file_put(state);
6342 if (state->free_reqs)
6343 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6347 * Start submission side cache.
6349 static void io_submit_state_start(struct io_submit_state *state,
6350 struct io_ring_ctx *ctx, unsigned int max_ios)
6352 blk_start_plug(&state->plug);
6354 INIT_LIST_HEAD(&state->comp.list);
6355 state->comp.ctx = ctx;
6356 state->free_reqs = 0;
6358 state->ios_left = max_ios;
6361 static void io_commit_sqring(struct io_ring_ctx *ctx)
6363 struct io_rings *rings = ctx->rings;
6366 * Ensure any loads from the SQEs are done at this point,
6367 * since once we write the new head, the application could
6368 * write new data to them.
6370 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6374 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6375 * that is mapped by userspace. This means that care needs to be taken to
6376 * ensure that reads are stable, as we cannot rely on userspace always
6377 * being a good citizen. If members of the sqe are validated and then later
6378 * used, it's important that those reads are done through READ_ONCE() to
6379 * prevent a re-load down the line.
6381 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6383 u32 *sq_array = ctx->sq_array;
6387 * The cached sq head (or cq tail) serves two purposes:
6389 * 1) allows us to batch the cost of updating the user visible
6391 * 2) allows the kernel side to track the head on its own, even
6392 * though the application is the one updating it.
6394 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6395 if (likely(head < ctx->sq_entries))
6396 return &ctx->sq_sqes[head];
6398 /* drop invalid entries */
6399 ctx->cached_sq_dropped++;
6400 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6404 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6406 ctx->cached_sq_head++;
6410 * Check SQE restrictions (opcode and flags).
6412 * Returns 'true' if SQE is allowed, 'false' otherwise.
6414 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6415 struct io_kiocb *req,
6416 unsigned int sqe_flags)
6418 if (!ctx->restricted)
6421 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6424 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6425 ctx->restrictions.sqe_flags_required)
6428 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6429 ctx->restrictions.sqe_flags_required))
6435 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6436 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6437 IOSQE_BUFFER_SELECT)
6439 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6440 const struct io_uring_sqe *sqe,
6441 struct io_submit_state *state)
6443 unsigned int sqe_flags;
6446 req->opcode = READ_ONCE(sqe->opcode);
6447 req->user_data = READ_ONCE(sqe->user_data);
6448 req->async_data = NULL;
6452 /* one is dropped after submission, the other at completion */
6453 refcount_set(&req->refs, 2);
6454 req->task = current;
6457 if (unlikely(req->opcode >= IORING_OP_LAST))
6460 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6463 sqe_flags = READ_ONCE(sqe->flags);
6464 /* enforce forwards compatibility on users */
6465 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6468 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6471 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6472 !io_op_defs[req->opcode].buffer_select)
6475 id = READ_ONCE(sqe->personality);
6477 struct io_identity *iod;
6479 io_req_init_async(req);
6480 iod = idr_find(&ctx->personality_idr, id);
6483 refcount_inc(&iod->count);
6484 io_put_identity(req);
6485 get_cred(iod->creds);
6486 req->work.identity = iod;
6487 req->work.flags |= IO_WQ_WORK_CREDS;
6490 /* same numerical values with corresponding REQ_F_*, safe to copy */
6491 req->flags |= sqe_flags;
6493 if (!io_op_defs[req->opcode].needs_file)
6496 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6501 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6503 struct io_submit_state state;
6504 struct io_kiocb *link = NULL;
6505 int i, submitted = 0;
6507 /* if we have a backlog and couldn't flush it all, return BUSY */
6508 if (test_bit(0, &ctx->sq_check_overflow)) {
6509 if (!list_empty(&ctx->cq_overflow_list) &&
6510 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6514 /* make sure SQ entry isn't read before tail */
6515 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6517 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6520 atomic_long_add(nr, ¤t->io_uring->req_issue);
6521 refcount_add(nr, ¤t->usage);
6523 io_submit_state_start(&state, ctx, nr);
6525 for (i = 0; i < nr; i++) {
6526 const struct io_uring_sqe *sqe;
6527 struct io_kiocb *req;
6530 sqe = io_get_sqe(ctx);
6531 if (unlikely(!sqe)) {
6532 io_consume_sqe(ctx);
6535 req = io_alloc_req(ctx, &state);
6536 if (unlikely(!req)) {
6538 submitted = -EAGAIN;
6541 io_consume_sqe(ctx);
6542 /* will complete beyond this point, count as submitted */
6545 err = io_init_req(ctx, req, sqe, &state);
6546 if (unlikely(err)) {
6549 io_req_complete(req, err);
6553 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6554 true, io_async_submit(ctx));
6555 err = io_submit_sqe(req, sqe, &link, &state.comp);
6560 if (unlikely(submitted != nr)) {
6561 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6563 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6564 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6565 put_task_struct_many(current, nr - ref_used);
6568 io_queue_link_head(link, &state.comp);
6569 io_submit_state_end(&state);
6571 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6572 io_commit_sqring(ctx);
6577 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6579 /* Tell userspace we may need a wakeup call */
6580 spin_lock_irq(&ctx->completion_lock);
6581 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6582 spin_unlock_irq(&ctx->completion_lock);
6585 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6587 spin_lock_irq(&ctx->completion_lock);
6588 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6589 spin_unlock_irq(&ctx->completion_lock);
6592 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6593 int sync, void *key)
6595 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6598 ret = autoremove_wake_function(wqe, mode, sync, key);
6600 unsigned long flags;
6602 spin_lock_irqsave(&ctx->completion_lock, flags);
6603 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6604 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6615 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6616 unsigned long start_jiffies, bool cap_entries)
6618 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6619 struct io_sq_data *sqd = ctx->sq_data;
6620 unsigned int to_submit;
6624 if (!list_empty(&ctx->iopoll_list)) {
6625 unsigned nr_events = 0;
6627 mutex_lock(&ctx->uring_lock);
6628 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6629 io_do_iopoll(ctx, &nr_events, 0);
6630 mutex_unlock(&ctx->uring_lock);
6633 to_submit = io_sqring_entries(ctx);
6636 * If submit got -EBUSY, flag us as needing the application
6637 * to enter the kernel to reap and flush events.
6639 if (!to_submit || ret == -EBUSY || need_resched()) {
6641 * Drop cur_mm before scheduling, we can't hold it for
6642 * long periods (or over schedule()). Do this before
6643 * adding ourselves to the waitqueue, as the unuse/drop
6646 io_sq_thread_drop_mm();
6649 * We're polling. If we're within the defined idle
6650 * period, then let us spin without work before going
6651 * to sleep. The exception is if we got EBUSY doing
6652 * more IO, we should wait for the application to
6653 * reap events and wake us up.
6655 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6656 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6657 !percpu_ref_is_dying(&ctx->refs)))
6660 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6661 TASK_INTERRUPTIBLE);
6664 * While doing polled IO, before going to sleep, we need
6665 * to check if there are new reqs added to iopoll_list,
6666 * it is because reqs may have been punted to io worker
6667 * and will be added to iopoll_list later, hence check
6668 * the iopoll_list again.
6670 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6671 !list_empty_careful(&ctx->iopoll_list)) {
6672 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6676 to_submit = io_sqring_entries(ctx);
6677 if (!to_submit || ret == -EBUSY)
6681 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6682 io_ring_clear_wakeup_flag(ctx);
6684 /* if we're handling multiple rings, cap submit size for fairness */
6685 if (cap_entries && to_submit > 8)
6688 mutex_lock(&ctx->uring_lock);
6689 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6690 ret = io_submit_sqes(ctx, to_submit);
6691 mutex_unlock(&ctx->uring_lock);
6693 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6694 wake_up(&ctx->sqo_sq_wait);
6696 return SQT_DID_WORK;
6699 static void io_sqd_init_new(struct io_sq_data *sqd)
6701 struct io_ring_ctx *ctx;
6703 while (!list_empty(&sqd->ctx_new_list)) {
6704 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6705 init_wait(&ctx->sqo_wait_entry);
6706 ctx->sqo_wait_entry.func = io_sq_wake_function;
6707 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6708 complete(&ctx->sq_thread_comp);
6712 static int io_sq_thread(void *data)
6714 struct cgroup_subsys_state *cur_css = NULL;
6715 const struct cred *old_cred = NULL;
6716 struct io_sq_data *sqd = data;
6717 struct io_ring_ctx *ctx;
6718 unsigned long start_jiffies;
6720 start_jiffies = jiffies;
6721 while (!kthread_should_stop()) {
6722 enum sq_ret ret = 0;
6726 * Any changes to the sqd lists are synchronized through the
6727 * kthread parking. This synchronizes the thread vs users,
6728 * the users are synchronized on the sqd->ctx_lock.
6730 if (kthread_should_park())
6733 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6734 io_sqd_init_new(sqd);
6736 cap_entries = !list_is_singular(&sqd->ctx_list);
6738 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6739 if (current->cred != ctx->creds) {
6741 revert_creds(old_cred);
6742 old_cred = override_creds(ctx->creds);
6744 io_sq_thread_associate_blkcg(ctx, &cur_css);
6746 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6748 io_sq_thread_drop_mm();
6751 if (ret & SQT_SPIN) {
6754 } else if (ret == SQT_IDLE) {
6755 if (kthread_should_park())
6757 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6758 io_ring_set_wakeup_flag(ctx);
6760 start_jiffies = jiffies;
6761 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6762 io_ring_clear_wakeup_flag(ctx);
6769 io_sq_thread_unassociate_blkcg();
6771 revert_creds(old_cred);
6778 struct io_wait_queue {
6779 struct wait_queue_entry wq;
6780 struct io_ring_ctx *ctx;
6782 unsigned nr_timeouts;
6785 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6787 struct io_ring_ctx *ctx = iowq->ctx;
6790 * Wake up if we have enough events, or if a timeout occurred since we
6791 * started waiting. For timeouts, we always want to return to userspace,
6792 * regardless of event count.
6794 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6795 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6798 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6799 int wake_flags, void *key)
6801 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6804 /* use noflush == true, as we can't safely rely on locking context */
6805 if (!io_should_wake(iowq, true))
6808 return autoremove_wake_function(curr, mode, wake_flags, key);
6811 static int io_run_task_work_sig(void)
6813 if (io_run_task_work())
6815 if (!signal_pending(current))
6817 if (current->jobctl & JOBCTL_TASK_WORK) {
6818 spin_lock_irq(¤t->sighand->siglock);
6819 current->jobctl &= ~JOBCTL_TASK_WORK;
6820 recalc_sigpending();
6821 spin_unlock_irq(¤t->sighand->siglock);
6828 * Wait until events become available, if we don't already have some. The
6829 * application must reap them itself, as they reside on the shared cq ring.
6831 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6832 const sigset_t __user *sig, size_t sigsz)
6834 struct io_wait_queue iowq = {
6837 .func = io_wake_function,
6838 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6841 .to_wait = min_events,
6843 struct io_rings *rings = ctx->rings;
6847 if (io_cqring_events(ctx, false) >= min_events)
6849 if (!io_run_task_work())
6854 #ifdef CONFIG_COMPAT
6855 if (in_compat_syscall())
6856 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6860 ret = set_user_sigmask(sig, sigsz);
6866 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6867 trace_io_uring_cqring_wait(ctx, min_events);
6869 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6870 TASK_INTERRUPTIBLE);
6871 /* make sure we run task_work before checking for signals */
6872 ret = io_run_task_work_sig();
6877 if (io_should_wake(&iowq, false))
6881 finish_wait(&ctx->wait, &iowq.wq);
6883 restore_saved_sigmask_unless(ret == -EINTR);
6885 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6888 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6890 #if defined(CONFIG_UNIX)
6891 if (ctx->ring_sock) {
6892 struct sock *sock = ctx->ring_sock->sk;
6893 struct sk_buff *skb;
6895 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6901 for (i = 0; i < ctx->nr_user_files; i++) {
6904 file = io_file_from_index(ctx, i);
6911 static void io_file_ref_kill(struct percpu_ref *ref)
6913 struct fixed_file_data *data;
6915 data = container_of(ref, struct fixed_file_data, refs);
6916 complete(&data->done);
6919 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6921 struct fixed_file_data *data = ctx->file_data;
6922 struct fixed_file_ref_node *ref_node = NULL;
6923 unsigned nr_tables, i;
6928 spin_lock(&data->lock);
6929 if (!list_empty(&data->ref_list))
6930 ref_node = list_first_entry(&data->ref_list,
6931 struct fixed_file_ref_node, node);
6932 spin_unlock(&data->lock);
6934 percpu_ref_kill(&ref_node->refs);
6936 percpu_ref_kill(&data->refs);
6938 /* wait for all refs nodes to complete */
6939 flush_delayed_work(&ctx->file_put_work);
6940 wait_for_completion(&data->done);
6942 __io_sqe_files_unregister(ctx);
6943 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6944 for (i = 0; i < nr_tables; i++)
6945 kfree(data->table[i].files);
6947 percpu_ref_exit(&data->refs);
6949 ctx->file_data = NULL;
6950 ctx->nr_user_files = 0;
6954 static void io_put_sq_data(struct io_sq_data *sqd)
6956 if (refcount_dec_and_test(&sqd->refs)) {
6958 * The park is a bit of a work-around, without it we get
6959 * warning spews on shutdown with SQPOLL set and affinity
6960 * set to a single CPU.
6963 kthread_park(sqd->thread);
6964 kthread_stop(sqd->thread);
6971 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6973 struct io_ring_ctx *ctx_attach;
6974 struct io_sq_data *sqd;
6977 f = fdget(p->wq_fd);
6979 return ERR_PTR(-ENXIO);
6980 if (f.file->f_op != &io_uring_fops) {
6982 return ERR_PTR(-EINVAL);
6985 ctx_attach = f.file->private_data;
6986 sqd = ctx_attach->sq_data;
6989 return ERR_PTR(-EINVAL);
6992 refcount_inc(&sqd->refs);
6997 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6999 struct io_sq_data *sqd;
7001 if (p->flags & IORING_SETUP_ATTACH_WQ)
7002 return io_attach_sq_data(p);
7004 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7006 return ERR_PTR(-ENOMEM);
7008 refcount_set(&sqd->refs, 1);
7009 INIT_LIST_HEAD(&sqd->ctx_list);
7010 INIT_LIST_HEAD(&sqd->ctx_new_list);
7011 mutex_init(&sqd->ctx_lock);
7012 mutex_init(&sqd->lock);
7013 init_waitqueue_head(&sqd->wait);
7017 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7018 __releases(&sqd->lock)
7022 kthread_unpark(sqd->thread);
7023 mutex_unlock(&sqd->lock);
7026 static void io_sq_thread_park(struct io_sq_data *sqd)
7027 __acquires(&sqd->lock)
7031 mutex_lock(&sqd->lock);
7032 kthread_park(sqd->thread);
7035 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7037 struct io_sq_data *sqd = ctx->sq_data;
7042 * We may arrive here from the error branch in
7043 * io_sq_offload_create() where the kthread is created
7044 * without being waked up, thus wake it up now to make
7045 * sure the wait will complete.
7047 wake_up_process(sqd->thread);
7048 wait_for_completion(&ctx->sq_thread_comp);
7050 io_sq_thread_park(sqd);
7053 mutex_lock(&sqd->ctx_lock);
7054 list_del(&ctx->sqd_list);
7055 mutex_unlock(&sqd->ctx_lock);
7058 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7059 io_sq_thread_unpark(sqd);
7062 io_put_sq_data(sqd);
7063 ctx->sq_data = NULL;
7067 static void io_finish_async(struct io_ring_ctx *ctx)
7069 io_sq_thread_stop(ctx);
7072 io_wq_destroy(ctx->io_wq);
7077 #if defined(CONFIG_UNIX)
7079 * Ensure the UNIX gc is aware of our file set, so we are certain that
7080 * the io_uring can be safely unregistered on process exit, even if we have
7081 * loops in the file referencing.
7083 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7085 struct sock *sk = ctx->ring_sock->sk;
7086 struct scm_fp_list *fpl;
7087 struct sk_buff *skb;
7090 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7094 skb = alloc_skb(0, GFP_KERNEL);
7103 fpl->user = get_uid(ctx->user);
7104 for (i = 0; i < nr; i++) {
7105 struct file *file = io_file_from_index(ctx, i + offset);
7109 fpl->fp[nr_files] = get_file(file);
7110 unix_inflight(fpl->user, fpl->fp[nr_files]);
7115 fpl->max = SCM_MAX_FD;
7116 fpl->count = nr_files;
7117 UNIXCB(skb).fp = fpl;
7118 skb->destructor = unix_destruct_scm;
7119 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7120 skb_queue_head(&sk->sk_receive_queue, skb);
7122 for (i = 0; i < nr_files; i++)
7133 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7134 * causes regular reference counting to break down. We rely on the UNIX
7135 * garbage collection to take care of this problem for us.
7137 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7139 unsigned left, total;
7143 left = ctx->nr_user_files;
7145 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7147 ret = __io_sqe_files_scm(ctx, this_files, total);
7151 total += this_files;
7157 while (total < ctx->nr_user_files) {
7158 struct file *file = io_file_from_index(ctx, total);
7168 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7174 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7175 unsigned nr_tables, unsigned nr_files)
7179 for (i = 0; i < nr_tables; i++) {
7180 struct fixed_file_table *table = &file_data->table[i];
7181 unsigned this_files;
7183 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7184 table->files = kcalloc(this_files, sizeof(struct file *),
7188 nr_files -= this_files;
7194 for (i = 0; i < nr_tables; i++) {
7195 struct fixed_file_table *table = &file_data->table[i];
7196 kfree(table->files);
7201 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7203 #if defined(CONFIG_UNIX)
7204 struct sock *sock = ctx->ring_sock->sk;
7205 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7206 struct sk_buff *skb;
7209 __skb_queue_head_init(&list);
7212 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7213 * remove this entry and rearrange the file array.
7215 skb = skb_dequeue(head);
7217 struct scm_fp_list *fp;
7219 fp = UNIXCB(skb).fp;
7220 for (i = 0; i < fp->count; i++) {
7223 if (fp->fp[i] != file)
7226 unix_notinflight(fp->user, fp->fp[i]);
7227 left = fp->count - 1 - i;
7229 memmove(&fp->fp[i], &fp->fp[i + 1],
7230 left * sizeof(struct file *));
7237 __skb_queue_tail(&list, skb);
7247 __skb_queue_tail(&list, skb);
7249 skb = skb_dequeue(head);
7252 if (skb_peek(&list)) {
7253 spin_lock_irq(&head->lock);
7254 while ((skb = __skb_dequeue(&list)) != NULL)
7255 __skb_queue_tail(head, skb);
7256 spin_unlock_irq(&head->lock);
7263 struct io_file_put {
7264 struct list_head list;
7268 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7270 struct fixed_file_data *file_data = ref_node->file_data;
7271 struct io_ring_ctx *ctx = file_data->ctx;
7272 struct io_file_put *pfile, *tmp;
7274 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7275 list_del(&pfile->list);
7276 io_ring_file_put(ctx, pfile->file);
7280 spin_lock(&file_data->lock);
7281 list_del(&ref_node->node);
7282 spin_unlock(&file_data->lock);
7284 percpu_ref_exit(&ref_node->refs);
7286 percpu_ref_put(&file_data->refs);
7289 static void io_file_put_work(struct work_struct *work)
7291 struct io_ring_ctx *ctx;
7292 struct llist_node *node;
7294 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7295 node = llist_del_all(&ctx->file_put_llist);
7298 struct fixed_file_ref_node *ref_node;
7299 struct llist_node *next = node->next;
7301 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7302 __io_file_put_work(ref_node);
7307 static void io_file_data_ref_zero(struct percpu_ref *ref)
7309 struct fixed_file_ref_node *ref_node;
7310 struct io_ring_ctx *ctx;
7314 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7315 ctx = ref_node->file_data->ctx;
7317 if (percpu_ref_is_dying(&ctx->file_data->refs))
7320 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7322 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7324 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7327 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7328 struct io_ring_ctx *ctx)
7330 struct fixed_file_ref_node *ref_node;
7332 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7334 return ERR_PTR(-ENOMEM);
7336 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7339 return ERR_PTR(-ENOMEM);
7341 INIT_LIST_HEAD(&ref_node->node);
7342 INIT_LIST_HEAD(&ref_node->file_list);
7343 ref_node->file_data = ctx->file_data;
7347 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7349 percpu_ref_exit(&ref_node->refs);
7353 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7356 __s32 __user *fds = (__s32 __user *) arg;
7357 unsigned nr_tables, i;
7359 int fd, ret = -ENOMEM;
7360 struct fixed_file_ref_node *ref_node;
7361 struct fixed_file_data *file_data;
7367 if (nr_args > IORING_MAX_FIXED_FILES)
7370 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7373 file_data->ctx = ctx;
7374 init_completion(&file_data->done);
7375 INIT_LIST_HEAD(&file_data->ref_list);
7376 spin_lock_init(&file_data->lock);
7378 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7379 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7381 if (!file_data->table)
7384 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7385 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7388 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7390 ctx->file_data = file_data;
7392 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7393 struct fixed_file_table *table;
7396 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7400 /* allow sparse sets */
7410 * Don't allow io_uring instances to be registered. If UNIX
7411 * isn't enabled, then this causes a reference cycle and this
7412 * instance can never get freed. If UNIX is enabled we'll
7413 * handle it just fine, but there's still no point in allowing
7414 * a ring fd as it doesn't support regular read/write anyway.
7416 if (file->f_op == &io_uring_fops) {
7420 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7421 index = i & IORING_FILE_TABLE_MASK;
7422 table->files[index] = file;
7425 ret = io_sqe_files_scm(ctx);
7427 io_sqe_files_unregister(ctx);
7431 ref_node = alloc_fixed_file_ref_node(ctx);
7432 if (IS_ERR(ref_node)) {
7433 io_sqe_files_unregister(ctx);
7434 return PTR_ERR(ref_node);
7437 file_data->node = ref_node;
7438 spin_lock(&file_data->lock);
7439 list_add(&ref_node->node, &file_data->ref_list);
7440 spin_unlock(&file_data->lock);
7441 percpu_ref_get(&file_data->refs);
7444 for (i = 0; i < ctx->nr_user_files; i++) {
7445 file = io_file_from_index(ctx, i);
7449 for (i = 0; i < nr_tables; i++)
7450 kfree(file_data->table[i].files);
7451 ctx->nr_user_files = 0;
7453 percpu_ref_exit(&file_data->refs);
7455 kfree(file_data->table);
7457 ctx->file_data = NULL;
7461 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7464 #if defined(CONFIG_UNIX)
7465 struct sock *sock = ctx->ring_sock->sk;
7466 struct sk_buff_head *head = &sock->sk_receive_queue;
7467 struct sk_buff *skb;
7470 * See if we can merge this file into an existing skb SCM_RIGHTS
7471 * file set. If there's no room, fall back to allocating a new skb
7472 * and filling it in.
7474 spin_lock_irq(&head->lock);
7475 skb = skb_peek(head);
7477 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7479 if (fpl->count < SCM_MAX_FD) {
7480 __skb_unlink(skb, head);
7481 spin_unlock_irq(&head->lock);
7482 fpl->fp[fpl->count] = get_file(file);
7483 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7485 spin_lock_irq(&head->lock);
7486 __skb_queue_head(head, skb);
7491 spin_unlock_irq(&head->lock);
7498 return __io_sqe_files_scm(ctx, 1, index);
7504 static int io_queue_file_removal(struct fixed_file_data *data,
7507 struct io_file_put *pfile;
7508 struct fixed_file_ref_node *ref_node = data->node;
7510 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7515 list_add(&pfile->list, &ref_node->file_list);
7520 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7521 struct io_uring_files_update *up,
7524 struct fixed_file_data *data = ctx->file_data;
7525 struct fixed_file_ref_node *ref_node;
7530 bool needs_switch = false;
7532 if (check_add_overflow(up->offset, nr_args, &done))
7534 if (done > ctx->nr_user_files)
7537 ref_node = alloc_fixed_file_ref_node(ctx);
7538 if (IS_ERR(ref_node))
7539 return PTR_ERR(ref_node);
7542 fds = u64_to_user_ptr(up->fds);
7544 struct fixed_file_table *table;
7548 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7552 i = array_index_nospec(up->offset, ctx->nr_user_files);
7553 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7554 index = i & IORING_FILE_TABLE_MASK;
7555 if (table->files[index]) {
7556 file = table->files[index];
7557 err = io_queue_file_removal(data, file);
7560 table->files[index] = NULL;
7561 needs_switch = true;
7570 * Don't allow io_uring instances to be registered. If
7571 * UNIX isn't enabled, then this causes a reference
7572 * cycle and this instance can never get freed. If UNIX
7573 * is enabled we'll handle it just fine, but there's
7574 * still no point in allowing a ring fd as it doesn't
7575 * support regular read/write anyway.
7577 if (file->f_op == &io_uring_fops) {
7582 table->files[index] = file;
7583 err = io_sqe_file_register(ctx, file, i);
7585 table->files[index] = NULL;
7596 percpu_ref_kill(&data->node->refs);
7597 spin_lock(&data->lock);
7598 list_add(&ref_node->node, &data->ref_list);
7599 data->node = ref_node;
7600 spin_unlock(&data->lock);
7601 percpu_ref_get(&ctx->file_data->refs);
7603 destroy_fixed_file_ref_node(ref_node);
7605 return done ? done : err;
7608 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7611 struct io_uring_files_update up;
7613 if (!ctx->file_data)
7617 if (copy_from_user(&up, arg, sizeof(up)))
7622 return __io_sqe_files_update(ctx, &up, nr_args);
7625 static void io_free_work(struct io_wq_work *work)
7627 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7629 /* Consider that io_steal_work() relies on this ref */
7633 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7634 struct io_uring_params *p)
7636 struct io_wq_data data;
7638 struct io_ring_ctx *ctx_attach;
7639 unsigned int concurrency;
7642 data.user = ctx->user;
7643 data.free_work = io_free_work;
7644 data.do_work = io_wq_submit_work;
7646 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7647 /* Do QD, or 4 * CPUS, whatever is smallest */
7648 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7650 ctx->io_wq = io_wq_create(concurrency, &data);
7651 if (IS_ERR(ctx->io_wq)) {
7652 ret = PTR_ERR(ctx->io_wq);
7658 f = fdget(p->wq_fd);
7662 if (f.file->f_op != &io_uring_fops) {
7667 ctx_attach = f.file->private_data;
7668 /* @io_wq is protected by holding the fd */
7669 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7674 ctx->io_wq = ctx_attach->io_wq;
7680 static int io_uring_alloc_task_context(struct task_struct *task)
7682 struct io_uring_task *tctx;
7684 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7685 if (unlikely(!tctx))
7689 init_waitqueue_head(&tctx->wait);
7692 atomic_long_set(&tctx->req_issue, 0);
7693 atomic_long_set(&tctx->req_complete, 0);
7694 task->io_uring = tctx;
7698 void __io_uring_free(struct task_struct *tsk)
7700 struct io_uring_task *tctx = tsk->io_uring;
7702 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7704 tsk->io_uring = NULL;
7707 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7708 struct io_uring_params *p)
7712 if (ctx->flags & IORING_SETUP_SQPOLL) {
7713 struct io_sq_data *sqd;
7716 if (!capable(CAP_SYS_ADMIN))
7719 sqd = io_get_sq_data(p);
7726 io_sq_thread_park(sqd);
7727 mutex_lock(&sqd->ctx_lock);
7728 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7729 mutex_unlock(&sqd->ctx_lock);
7730 io_sq_thread_unpark(sqd);
7732 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7733 if (!ctx->sq_thread_idle)
7734 ctx->sq_thread_idle = HZ;
7739 if (p->flags & IORING_SETUP_SQ_AFF) {
7740 int cpu = p->sq_thread_cpu;
7743 if (cpu >= nr_cpu_ids)
7745 if (!cpu_online(cpu))
7748 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7749 cpu, "io_uring-sq");
7751 sqd->thread = kthread_create(io_sq_thread, sqd,
7754 if (IS_ERR(sqd->thread)) {
7755 ret = PTR_ERR(sqd->thread);
7759 ret = io_uring_alloc_task_context(sqd->thread);
7762 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7763 /* Can't have SQ_AFF without SQPOLL */
7769 ret = io_init_wq_offload(ctx, p);
7775 io_finish_async(ctx);
7779 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7781 struct io_sq_data *sqd = ctx->sq_data;
7783 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7784 wake_up_process(sqd->thread);
7787 static inline void __io_unaccount_mem(struct user_struct *user,
7788 unsigned long nr_pages)
7790 atomic_long_sub(nr_pages, &user->locked_vm);
7793 static inline int __io_account_mem(struct user_struct *user,
7794 unsigned long nr_pages)
7796 unsigned long page_limit, cur_pages, new_pages;
7798 /* Don't allow more pages than we can safely lock */
7799 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7802 cur_pages = atomic_long_read(&user->locked_vm);
7803 new_pages = cur_pages + nr_pages;
7804 if (new_pages > page_limit)
7806 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7807 new_pages) != cur_pages);
7812 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7813 enum io_mem_account acct)
7816 __io_unaccount_mem(ctx->user, nr_pages);
7818 if (ctx->mm_account) {
7819 if (acct == ACCT_LOCKED)
7820 ctx->mm_account->locked_vm -= nr_pages;
7821 else if (acct == ACCT_PINNED)
7822 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7826 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7827 enum io_mem_account acct)
7831 if (ctx->limit_mem) {
7832 ret = __io_account_mem(ctx->user, nr_pages);
7837 if (ctx->mm_account) {
7838 if (acct == ACCT_LOCKED)
7839 ctx->mm_account->locked_vm += nr_pages;
7840 else if (acct == ACCT_PINNED)
7841 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7847 static void io_mem_free(void *ptr)
7854 page = virt_to_head_page(ptr);
7855 if (put_page_testzero(page))
7856 free_compound_page(page);
7859 static void *io_mem_alloc(size_t size)
7861 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7864 return (void *) __get_free_pages(gfp_flags, get_order(size));
7867 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7870 struct io_rings *rings;
7871 size_t off, sq_array_size;
7873 off = struct_size(rings, cqes, cq_entries);
7874 if (off == SIZE_MAX)
7878 off = ALIGN(off, SMP_CACHE_BYTES);
7886 sq_array_size = array_size(sizeof(u32), sq_entries);
7887 if (sq_array_size == SIZE_MAX)
7890 if (check_add_overflow(off, sq_array_size, &off))
7896 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7900 pages = (size_t)1 << get_order(
7901 rings_size(sq_entries, cq_entries, NULL));
7902 pages += (size_t)1 << get_order(
7903 array_size(sizeof(struct io_uring_sqe), sq_entries));
7908 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7912 if (!ctx->user_bufs)
7915 for (i = 0; i < ctx->nr_user_bufs; i++) {
7916 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7918 for (j = 0; j < imu->nr_bvecs; j++)
7919 unpin_user_page(imu->bvec[j].bv_page);
7921 if (imu->acct_pages)
7922 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7927 kfree(ctx->user_bufs);
7928 ctx->user_bufs = NULL;
7929 ctx->nr_user_bufs = 0;
7933 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7934 void __user *arg, unsigned index)
7936 struct iovec __user *src;
7938 #ifdef CONFIG_COMPAT
7940 struct compat_iovec __user *ciovs;
7941 struct compat_iovec ciov;
7943 ciovs = (struct compat_iovec __user *) arg;
7944 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7947 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7948 dst->iov_len = ciov.iov_len;
7952 src = (struct iovec __user *) arg;
7953 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7959 * Not super efficient, but this is just a registration time. And we do cache
7960 * the last compound head, so generally we'll only do a full search if we don't
7963 * We check if the given compound head page has already been accounted, to
7964 * avoid double accounting it. This allows us to account the full size of the
7965 * page, not just the constituent pages of a huge page.
7967 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7968 int nr_pages, struct page *hpage)
7972 /* check current page array */
7973 for (i = 0; i < nr_pages; i++) {
7974 if (!PageCompound(pages[i]))
7976 if (compound_head(pages[i]) == hpage)
7980 /* check previously registered pages */
7981 for (i = 0; i < ctx->nr_user_bufs; i++) {
7982 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7984 for (j = 0; j < imu->nr_bvecs; j++) {
7985 if (!PageCompound(imu->bvec[j].bv_page))
7987 if (compound_head(imu->bvec[j].bv_page) == hpage)
7995 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7996 int nr_pages, struct io_mapped_ubuf *imu,
7997 struct page **last_hpage)
8001 for (i = 0; i < nr_pages; i++) {
8002 if (!PageCompound(pages[i])) {
8007 hpage = compound_head(pages[i]);
8008 if (hpage == *last_hpage)
8010 *last_hpage = hpage;
8011 if (headpage_already_acct(ctx, pages, i, hpage))
8013 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8017 if (!imu->acct_pages)
8020 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8022 imu->acct_pages = 0;
8026 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8029 struct vm_area_struct **vmas = NULL;
8030 struct page **pages = NULL;
8031 struct page *last_hpage = NULL;
8032 int i, j, got_pages = 0;
8037 if (!nr_args || nr_args > UIO_MAXIOV)
8040 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8042 if (!ctx->user_bufs)
8045 for (i = 0; i < nr_args; i++) {
8046 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8047 unsigned long off, start, end, ubuf;
8052 ret = io_copy_iov(ctx, &iov, arg, i);
8057 * Don't impose further limits on the size and buffer
8058 * constraints here, we'll -EINVAL later when IO is
8059 * submitted if they are wrong.
8062 if (!iov.iov_base || !iov.iov_len)
8065 /* arbitrary limit, but we need something */
8066 if (iov.iov_len > SZ_1G)
8069 ubuf = (unsigned long) iov.iov_base;
8070 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8071 start = ubuf >> PAGE_SHIFT;
8072 nr_pages = end - start;
8075 if (!pages || nr_pages > got_pages) {
8078 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8080 vmas = kvmalloc_array(nr_pages,
8081 sizeof(struct vm_area_struct *),
8083 if (!pages || !vmas) {
8087 got_pages = nr_pages;
8090 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8097 mmap_read_lock(current->mm);
8098 pret = pin_user_pages(ubuf, nr_pages,
8099 FOLL_WRITE | FOLL_LONGTERM,
8101 if (pret == nr_pages) {
8102 /* don't support file backed memory */
8103 for (j = 0; j < nr_pages; j++) {
8104 struct vm_area_struct *vma = vmas[j];
8107 !is_file_hugepages(vma->vm_file)) {
8113 ret = pret < 0 ? pret : -EFAULT;
8115 mmap_read_unlock(current->mm);
8118 * if we did partial map, or found file backed vmas,
8119 * release any pages we did get
8122 unpin_user_pages(pages, pret);
8127 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8129 unpin_user_pages(pages, pret);
8134 off = ubuf & ~PAGE_MASK;
8136 for (j = 0; j < nr_pages; j++) {
8139 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8140 imu->bvec[j].bv_page = pages[j];
8141 imu->bvec[j].bv_len = vec_len;
8142 imu->bvec[j].bv_offset = off;
8146 /* store original address for later verification */
8148 imu->len = iov.iov_len;
8149 imu->nr_bvecs = nr_pages;
8151 ctx->nr_user_bufs++;
8159 io_sqe_buffer_unregister(ctx);
8163 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8165 __s32 __user *fds = arg;
8171 if (copy_from_user(&fd, fds, sizeof(*fds)))
8174 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8175 if (IS_ERR(ctx->cq_ev_fd)) {
8176 int ret = PTR_ERR(ctx->cq_ev_fd);
8177 ctx->cq_ev_fd = NULL;
8184 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8186 if (ctx->cq_ev_fd) {
8187 eventfd_ctx_put(ctx->cq_ev_fd);
8188 ctx->cq_ev_fd = NULL;
8195 static int __io_destroy_buffers(int id, void *p, void *data)
8197 struct io_ring_ctx *ctx = data;
8198 struct io_buffer *buf = p;
8200 __io_remove_buffers(ctx, buf, id, -1U);
8204 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8206 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8207 idr_destroy(&ctx->io_buffer_idr);
8210 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8212 io_finish_async(ctx);
8213 io_sqe_buffer_unregister(ctx);
8215 if (ctx->sqo_task) {
8216 put_task_struct(ctx->sqo_task);
8217 ctx->sqo_task = NULL;
8218 mmdrop(ctx->mm_account);
8219 ctx->mm_account = NULL;
8222 #ifdef CONFIG_BLK_CGROUP
8223 if (ctx->sqo_blkcg_css)
8224 css_put(ctx->sqo_blkcg_css);
8227 io_sqe_files_unregister(ctx);
8228 io_eventfd_unregister(ctx);
8229 io_destroy_buffers(ctx);
8230 idr_destroy(&ctx->personality_idr);
8232 #if defined(CONFIG_UNIX)
8233 if (ctx->ring_sock) {
8234 ctx->ring_sock->file = NULL; /* so that iput() is called */
8235 sock_release(ctx->ring_sock);
8239 io_mem_free(ctx->rings);
8240 io_mem_free(ctx->sq_sqes);
8242 percpu_ref_exit(&ctx->refs);
8243 free_uid(ctx->user);
8244 put_cred(ctx->creds);
8245 kfree(ctx->cancel_hash);
8246 kmem_cache_free(req_cachep, ctx->fallback_req);
8250 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8252 struct io_ring_ctx *ctx = file->private_data;
8255 poll_wait(file, &ctx->cq_wait, wait);
8257 * synchronizes with barrier from wq_has_sleeper call in
8261 if (!io_sqring_full(ctx))
8262 mask |= EPOLLOUT | EPOLLWRNORM;
8263 if (io_cqring_events(ctx, false))
8264 mask |= EPOLLIN | EPOLLRDNORM;
8269 static int io_uring_fasync(int fd, struct file *file, int on)
8271 struct io_ring_ctx *ctx = file->private_data;
8273 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8276 static int io_remove_personalities(int id, void *p, void *data)
8278 struct io_ring_ctx *ctx = data;
8279 struct io_identity *iod;
8281 iod = idr_remove(&ctx->personality_idr, id);
8283 put_cred(iod->creds);
8284 if (refcount_dec_and_test(&iod->count))
8290 static void io_ring_exit_work(struct work_struct *work)
8292 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8296 * If we're doing polled IO and end up having requests being
8297 * submitted async (out-of-line), then completions can come in while
8298 * we're waiting for refs to drop. We need to reap these manually,
8299 * as nobody else will be looking for them.
8303 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8304 io_iopoll_try_reap_events(ctx);
8305 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8306 io_ring_ctx_free(ctx);
8309 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8311 mutex_lock(&ctx->uring_lock);
8312 percpu_ref_kill(&ctx->refs);
8313 mutex_unlock(&ctx->uring_lock);
8315 io_kill_timeouts(ctx, NULL);
8316 io_poll_remove_all(ctx, NULL);
8319 io_wq_cancel_all(ctx->io_wq);
8321 /* if we failed setting up the ctx, we might not have any rings */
8323 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8324 io_iopoll_try_reap_events(ctx);
8325 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8328 * Do this upfront, so we won't have a grace period where the ring
8329 * is closed but resources aren't reaped yet. This can cause
8330 * spurious failure in setting up a new ring.
8332 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8335 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8337 * Use system_unbound_wq to avoid spawning tons of event kworkers
8338 * if we're exiting a ton of rings at the same time. It just adds
8339 * noise and overhead, there's no discernable change in runtime
8340 * over using system_wq.
8342 queue_work(system_unbound_wq, &ctx->exit_work);
8345 static int io_uring_release(struct inode *inode, struct file *file)
8347 struct io_ring_ctx *ctx = file->private_data;
8349 file->private_data = NULL;
8350 io_ring_ctx_wait_and_kill(ctx);
8354 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8356 struct files_struct *files = data;
8358 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8359 work->identity->files == files);
8363 * Returns true if 'preq' is the link parent of 'req'
8365 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8367 struct io_kiocb *link;
8369 if (!(preq->flags & REQ_F_LINK_HEAD))
8372 list_for_each_entry(link, &preq->link_list, link_list) {
8380 static bool io_match_link_files(struct io_kiocb *req,
8381 struct files_struct *files)
8383 struct io_kiocb *link;
8385 if (io_match_files(req, files))
8387 if (req->flags & REQ_F_LINK_HEAD) {
8388 list_for_each_entry(link, &req->link_list, link_list) {
8389 if (io_match_files(link, files))
8397 * We're looking to cancel 'req' because it's holding on to our files, but
8398 * 'req' could be a link to another request. See if it is, and cancel that
8399 * parent request if so.
8401 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8403 struct hlist_node *tmp;
8404 struct io_kiocb *preq;
8408 spin_lock_irq(&ctx->completion_lock);
8409 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8410 struct hlist_head *list;
8412 list = &ctx->cancel_hash[i];
8413 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8414 found = io_match_link(preq, req);
8416 io_poll_remove_one(preq);
8421 spin_unlock_irq(&ctx->completion_lock);
8425 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8426 struct io_kiocb *req)
8428 struct io_kiocb *preq;
8431 spin_lock_irq(&ctx->completion_lock);
8432 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8433 found = io_match_link(preq, req);
8435 __io_timeout_cancel(preq);
8439 spin_unlock_irq(&ctx->completion_lock);
8443 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8445 return io_match_link(container_of(work, struct io_kiocb, work), data);
8448 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8450 enum io_wq_cancel cret;
8452 /* cancel this particular work, if it's running */
8453 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8454 if (cret != IO_WQ_CANCEL_NOTFOUND)
8457 /* find links that hold this pending, cancel those */
8458 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8459 if (cret != IO_WQ_CANCEL_NOTFOUND)
8462 /* if we have a poll link holding this pending, cancel that */
8463 if (io_poll_remove_link(ctx, req))
8466 /* final option, timeout link is holding this req pending */
8467 io_timeout_remove_link(ctx, req);
8470 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8471 struct files_struct *files)
8473 struct io_defer_entry *de = NULL;
8476 spin_lock_irq(&ctx->completion_lock);
8477 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8478 if (io_match_link_files(de->req, files)) {
8479 list_cut_position(&list, &ctx->defer_list, &de->list);
8483 spin_unlock_irq(&ctx->completion_lock);
8485 while (!list_empty(&list)) {
8486 de = list_first_entry(&list, struct io_defer_entry, list);
8487 list_del_init(&de->list);
8488 req_set_fail_links(de->req);
8489 io_put_req(de->req);
8490 io_req_complete(de->req, -ECANCELED);
8496 * Returns true if we found and killed one or more files pinning requests
8498 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8499 struct files_struct *files)
8501 if (list_empty_careful(&ctx->inflight_list))
8504 io_cancel_defer_files(ctx, files);
8505 /* cancel all at once, should be faster than doing it one by one*/
8506 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8508 while (!list_empty_careful(&ctx->inflight_list)) {
8509 struct io_kiocb *cancel_req = NULL, *req;
8512 spin_lock_irq(&ctx->inflight_lock);
8513 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8514 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8515 req->work.identity->files != files)
8517 /* req is being completed, ignore */
8518 if (!refcount_inc_not_zero(&req->refs))
8524 prepare_to_wait(&ctx->inflight_wait, &wait,
8525 TASK_UNINTERRUPTIBLE);
8526 spin_unlock_irq(&ctx->inflight_lock);
8528 /* We need to keep going until we don't find a matching req */
8531 /* cancel this request, or head link requests */
8532 io_attempt_cancel(ctx, cancel_req);
8533 io_put_req(cancel_req);
8534 /* cancellations _may_ trigger task work */
8537 finish_wait(&ctx->inflight_wait, &wait);
8543 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8545 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8546 struct task_struct *task = data;
8548 return io_task_match(req, task);
8551 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8552 struct task_struct *task,
8553 struct files_struct *files)
8557 ret = io_uring_cancel_files(ctx, files);
8559 enum io_wq_cancel cret;
8561 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8562 if (cret != IO_WQ_CANCEL_NOTFOUND)
8565 /* SQPOLL thread does its own polling */
8566 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8567 while (!list_empty_careful(&ctx->iopoll_list)) {
8568 io_iopoll_try_reap_events(ctx);
8573 ret |= io_poll_remove_all(ctx, task);
8574 ret |= io_kill_timeouts(ctx, task);
8581 * We need to iteratively cancel requests, in case a request has dependent
8582 * hard links. These persist even for failure of cancelations, hence keep
8583 * looping until none are found.
8585 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8586 struct files_struct *files)
8588 struct task_struct *task = current;
8590 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8591 task = ctx->sq_data->thread;
8593 io_cqring_overflow_flush(ctx, true, task, files);
8595 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8602 * Note that this task has used io_uring. We use it for cancelation purposes.
8604 static int io_uring_add_task_file(struct file *file)
8606 struct io_uring_task *tctx = current->io_uring;
8608 if (unlikely(!tctx)) {
8611 ret = io_uring_alloc_task_context(current);
8614 tctx = current->io_uring;
8616 if (tctx->last != file) {
8617 void *old = xa_load(&tctx->xa, (unsigned long)file);
8621 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8630 * Remove this io_uring_file -> task mapping.
8632 static void io_uring_del_task_file(struct file *file)
8634 struct io_uring_task *tctx = current->io_uring;
8636 if (tctx->last == file)
8638 file = xa_erase(&tctx->xa, (unsigned long)file);
8643 static void __io_uring_attempt_task_drop(struct file *file)
8645 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8648 io_uring_del_task_file(file);
8652 * Drop task note for this file if we're the only ones that hold it after
8655 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8657 if (!current->io_uring)
8660 * fput() is pending, will be 2 if the only other ref is our potential
8661 * task file note. If the task is exiting, drop regardless of count.
8663 if (!exiting && atomic_long_read(&file->f_count) != 2)
8666 __io_uring_attempt_task_drop(file);
8669 void __io_uring_files_cancel(struct files_struct *files)
8671 struct io_uring_task *tctx = current->io_uring;
8673 unsigned long index;
8675 /* make sure overflow events are dropped */
8676 tctx->in_idle = true;
8678 xa_for_each(&tctx->xa, index, file) {
8679 struct io_ring_ctx *ctx = file->private_data;
8681 io_uring_cancel_task_requests(ctx, files);
8683 io_uring_del_task_file(file);
8687 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8689 return atomic_long_read(&tctx->req_issue) ==
8690 atomic_long_read(&tctx->req_complete);
8694 * Find any io_uring fd that this task has registered or done IO on, and cancel
8697 void __io_uring_task_cancel(void)
8699 struct io_uring_task *tctx = current->io_uring;
8703 /* make sure overflow events are dropped */
8704 tctx->in_idle = true;
8706 while (!io_uring_task_idle(tctx)) {
8707 /* read completions before cancelations */
8708 completions = atomic_long_read(&tctx->req_complete);
8709 __io_uring_files_cancel(NULL);
8711 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8714 * If we've seen completions, retry. This avoids a race where
8715 * a completion comes in before we did prepare_to_wait().
8717 if (completions != atomic_long_read(&tctx->req_complete))
8719 if (io_uring_task_idle(tctx))
8724 finish_wait(&tctx->wait, &wait);
8725 tctx->in_idle = false;
8728 static int io_uring_flush(struct file *file, void *data)
8730 struct io_ring_ctx *ctx = file->private_data;
8733 * If the task is going away, cancel work it may have pending
8735 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8738 io_uring_cancel_task_requests(ctx, data);
8739 io_uring_attempt_task_drop(file, !data);
8743 static void *io_uring_validate_mmap_request(struct file *file,
8744 loff_t pgoff, size_t sz)
8746 struct io_ring_ctx *ctx = file->private_data;
8747 loff_t offset = pgoff << PAGE_SHIFT;
8752 case IORING_OFF_SQ_RING:
8753 case IORING_OFF_CQ_RING:
8756 case IORING_OFF_SQES:
8760 return ERR_PTR(-EINVAL);
8763 page = virt_to_head_page(ptr);
8764 if (sz > page_size(page))
8765 return ERR_PTR(-EINVAL);
8772 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8774 size_t sz = vma->vm_end - vma->vm_start;
8778 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8780 return PTR_ERR(ptr);
8782 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8783 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8786 #else /* !CONFIG_MMU */
8788 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8790 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8793 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8795 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8798 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8799 unsigned long addr, unsigned long len,
8800 unsigned long pgoff, unsigned long flags)
8804 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8806 return PTR_ERR(ptr);
8808 return (unsigned long) ptr;
8811 #endif /* !CONFIG_MMU */
8813 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8818 if (!io_sqring_full(ctx))
8821 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8823 if (!io_sqring_full(ctx))
8827 } while (!signal_pending(current));
8829 finish_wait(&ctx->sqo_sq_wait, &wait);
8832 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8833 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8836 struct io_ring_ctx *ctx;
8843 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8844 IORING_ENTER_SQ_WAIT))
8852 if (f.file->f_op != &io_uring_fops)
8856 ctx = f.file->private_data;
8857 if (!percpu_ref_tryget(&ctx->refs))
8861 if (ctx->flags & IORING_SETUP_R_DISABLED)
8865 * For SQ polling, the thread will do all submissions and completions.
8866 * Just return the requested submit count, and wake the thread if
8870 if (ctx->flags & IORING_SETUP_SQPOLL) {
8871 if (!list_empty_careful(&ctx->cq_overflow_list))
8872 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8873 if (flags & IORING_ENTER_SQ_WAKEUP)
8874 wake_up(&ctx->sq_data->wait);
8875 if (flags & IORING_ENTER_SQ_WAIT)
8876 io_sqpoll_wait_sq(ctx);
8877 submitted = to_submit;
8878 } else if (to_submit) {
8879 ret = io_uring_add_task_file(f.file);
8882 mutex_lock(&ctx->uring_lock);
8883 submitted = io_submit_sqes(ctx, to_submit);
8884 mutex_unlock(&ctx->uring_lock);
8886 if (submitted != to_submit)
8889 if (flags & IORING_ENTER_GETEVENTS) {
8890 min_complete = min(min_complete, ctx->cq_entries);
8893 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8894 * space applications don't need to do io completion events
8895 * polling again, they can rely on io_sq_thread to do polling
8896 * work, which can reduce cpu usage and uring_lock contention.
8898 if (ctx->flags & IORING_SETUP_IOPOLL &&
8899 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8900 ret = io_iopoll_check(ctx, min_complete);
8902 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8907 percpu_ref_put(&ctx->refs);
8910 return submitted ? submitted : ret;
8913 #ifdef CONFIG_PROC_FS
8914 static int io_uring_show_cred(int id, void *p, void *data)
8916 const struct cred *cred = p;
8917 struct seq_file *m = data;
8918 struct user_namespace *uns = seq_user_ns(m);
8919 struct group_info *gi;
8924 seq_printf(m, "%5d\n", id);
8925 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8926 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8927 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8928 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8929 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8930 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8931 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8932 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8933 seq_puts(m, "\n\tGroups:\t");
8934 gi = cred->group_info;
8935 for (g = 0; g < gi->ngroups; g++) {
8936 seq_put_decimal_ull(m, g ? " " : "",
8937 from_kgid_munged(uns, gi->gid[g]));
8939 seq_puts(m, "\n\tCapEff:\t");
8940 cap = cred->cap_effective;
8941 CAP_FOR_EACH_U32(__capi)
8942 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8947 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8949 struct io_sq_data *sq = NULL;
8954 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8955 * since fdinfo case grabs it in the opposite direction of normal use
8956 * cases. If we fail to get the lock, we just don't iterate any
8957 * structures that could be going away outside the io_uring mutex.
8959 has_lock = mutex_trylock(&ctx->uring_lock);
8961 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8964 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8965 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8966 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8967 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8968 struct fixed_file_table *table;
8971 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8972 f = table->files[i & IORING_FILE_TABLE_MASK];
8974 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8976 seq_printf(m, "%5u: <none>\n", i);
8978 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8979 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8980 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8982 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8983 (unsigned int) buf->len);
8985 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8986 seq_printf(m, "Personalities:\n");
8987 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8989 seq_printf(m, "PollList:\n");
8990 spin_lock_irq(&ctx->completion_lock);
8991 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8992 struct hlist_head *list = &ctx->cancel_hash[i];
8993 struct io_kiocb *req;
8995 hlist_for_each_entry(req, list, hash_node)
8996 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8997 req->task->task_works != NULL);
8999 spin_unlock_irq(&ctx->completion_lock);
9001 mutex_unlock(&ctx->uring_lock);
9004 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9006 struct io_ring_ctx *ctx = f->private_data;
9008 if (percpu_ref_tryget(&ctx->refs)) {
9009 __io_uring_show_fdinfo(ctx, m);
9010 percpu_ref_put(&ctx->refs);
9015 static const struct file_operations io_uring_fops = {
9016 .release = io_uring_release,
9017 .flush = io_uring_flush,
9018 .mmap = io_uring_mmap,
9020 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9021 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9023 .poll = io_uring_poll,
9024 .fasync = io_uring_fasync,
9025 #ifdef CONFIG_PROC_FS
9026 .show_fdinfo = io_uring_show_fdinfo,
9030 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9031 struct io_uring_params *p)
9033 struct io_rings *rings;
9034 size_t size, sq_array_offset;
9036 /* make sure these are sane, as we already accounted them */
9037 ctx->sq_entries = p->sq_entries;
9038 ctx->cq_entries = p->cq_entries;
9040 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9041 if (size == SIZE_MAX)
9044 rings = io_mem_alloc(size);
9049 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9050 rings->sq_ring_mask = p->sq_entries - 1;
9051 rings->cq_ring_mask = p->cq_entries - 1;
9052 rings->sq_ring_entries = p->sq_entries;
9053 rings->cq_ring_entries = p->cq_entries;
9054 ctx->sq_mask = rings->sq_ring_mask;
9055 ctx->cq_mask = rings->cq_ring_mask;
9057 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9058 if (size == SIZE_MAX) {
9059 io_mem_free(ctx->rings);
9064 ctx->sq_sqes = io_mem_alloc(size);
9065 if (!ctx->sq_sqes) {
9066 io_mem_free(ctx->rings);
9075 * Allocate an anonymous fd, this is what constitutes the application
9076 * visible backing of an io_uring instance. The application mmaps this
9077 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9078 * we have to tie this fd to a socket for file garbage collection purposes.
9080 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9085 #if defined(CONFIG_UNIX)
9086 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9092 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9096 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9097 O_RDWR | O_CLOEXEC);
9101 ret = PTR_ERR(file);
9105 #if defined(CONFIG_UNIX)
9106 ctx->ring_sock->file = file;
9108 if (unlikely(io_uring_add_task_file(file))) {
9109 file = ERR_PTR(-ENOMEM);
9112 fd_install(ret, file);
9115 #if defined(CONFIG_UNIX)
9116 sock_release(ctx->ring_sock);
9117 ctx->ring_sock = NULL;
9122 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9123 struct io_uring_params __user *params)
9125 struct user_struct *user = NULL;
9126 struct io_ring_ctx *ctx;
9132 if (entries > IORING_MAX_ENTRIES) {
9133 if (!(p->flags & IORING_SETUP_CLAMP))
9135 entries = IORING_MAX_ENTRIES;
9139 * Use twice as many entries for the CQ ring. It's possible for the
9140 * application to drive a higher depth than the size of the SQ ring,
9141 * since the sqes are only used at submission time. This allows for
9142 * some flexibility in overcommitting a bit. If the application has
9143 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9144 * of CQ ring entries manually.
9146 p->sq_entries = roundup_pow_of_two(entries);
9147 if (p->flags & IORING_SETUP_CQSIZE) {
9149 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9150 * to a power-of-two, if it isn't already. We do NOT impose
9151 * any cq vs sq ring sizing.
9153 if (p->cq_entries < p->sq_entries)
9155 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9156 if (!(p->flags & IORING_SETUP_CLAMP))
9158 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9160 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9162 p->cq_entries = 2 * p->sq_entries;
9165 user = get_uid(current_user());
9166 limit_mem = !capable(CAP_IPC_LOCK);
9169 ret = __io_account_mem(user,
9170 ring_pages(p->sq_entries, p->cq_entries));
9177 ctx = io_ring_ctx_alloc(p);
9180 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9185 ctx->compat = in_compat_syscall();
9187 ctx->creds = get_current_cred();
9189 ctx->sqo_task = get_task_struct(current);
9192 * This is just grabbed for accounting purposes. When a process exits,
9193 * the mm is exited and dropped before the files, hence we need to hang
9194 * on to this mm purely for the purposes of being able to unaccount
9195 * memory (locked/pinned vm). It's not used for anything else.
9197 mmgrab(current->mm);
9198 ctx->mm_account = current->mm;
9200 #ifdef CONFIG_BLK_CGROUP
9202 * The sq thread will belong to the original cgroup it was inited in.
9203 * If the cgroup goes offline (e.g. disabling the io controller), then
9204 * issued bios will be associated with the closest cgroup later in the
9208 ctx->sqo_blkcg_css = blkcg_css();
9209 ret = css_tryget_online(ctx->sqo_blkcg_css);
9212 /* don't init against a dying cgroup, have the user try again */
9213 ctx->sqo_blkcg_css = NULL;
9220 * Account memory _before_ installing the file descriptor. Once
9221 * the descriptor is installed, it can get closed at any time. Also
9222 * do this before hitting the general error path, as ring freeing
9223 * will un-account as well.
9225 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9227 ctx->limit_mem = limit_mem;
9229 ret = io_allocate_scq_urings(ctx, p);
9233 ret = io_sq_offload_create(ctx, p);
9237 if (!(p->flags & IORING_SETUP_R_DISABLED))
9238 io_sq_offload_start(ctx);
9240 memset(&p->sq_off, 0, sizeof(p->sq_off));
9241 p->sq_off.head = offsetof(struct io_rings, sq.head);
9242 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9243 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9244 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9245 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9246 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9247 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9249 memset(&p->cq_off, 0, sizeof(p->cq_off));
9250 p->cq_off.head = offsetof(struct io_rings, cq.head);
9251 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9252 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9253 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9254 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9255 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9256 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9258 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9259 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9260 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9261 IORING_FEAT_POLL_32BITS;
9263 if (copy_to_user(params, p, sizeof(*p))) {
9269 * Install ring fd as the very last thing, so we don't risk someone
9270 * having closed it before we finish setup
9272 ret = io_uring_get_fd(ctx);
9276 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9279 io_ring_ctx_wait_and_kill(ctx);
9284 * Sets up an aio uring context, and returns the fd. Applications asks for a
9285 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9286 * params structure passed in.
9288 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9290 struct io_uring_params p;
9293 if (copy_from_user(&p, params, sizeof(p)))
9295 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9300 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9301 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9302 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9303 IORING_SETUP_R_DISABLED))
9306 return io_uring_create(entries, &p, params);
9309 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9310 struct io_uring_params __user *, params)
9312 return io_uring_setup(entries, params);
9315 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9317 struct io_uring_probe *p;
9321 size = struct_size(p, ops, nr_args);
9322 if (size == SIZE_MAX)
9324 p = kzalloc(size, GFP_KERNEL);
9329 if (copy_from_user(p, arg, size))
9332 if (memchr_inv(p, 0, size))
9335 p->last_op = IORING_OP_LAST - 1;
9336 if (nr_args > IORING_OP_LAST)
9337 nr_args = IORING_OP_LAST;
9339 for (i = 0; i < nr_args; i++) {
9341 if (!io_op_defs[i].not_supported)
9342 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9347 if (copy_to_user(arg, p, size))
9354 static int io_register_personality(struct io_ring_ctx *ctx)
9356 struct io_identity *id;
9359 id = kmalloc(sizeof(*id), GFP_KERNEL);
9363 io_init_identity(id);
9364 id->creds = get_current_cred();
9366 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9368 put_cred(id->creds);
9374 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9376 struct io_identity *iod;
9378 iod = idr_remove(&ctx->personality_idr, id);
9380 put_cred(iod->creds);
9381 if (refcount_dec_and_test(&iod->count))
9389 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9390 unsigned int nr_args)
9392 struct io_uring_restriction *res;
9396 /* Restrictions allowed only if rings started disabled */
9397 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9400 /* We allow only a single restrictions registration */
9401 if (ctx->restrictions.registered)
9404 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9407 size = array_size(nr_args, sizeof(*res));
9408 if (size == SIZE_MAX)
9411 res = memdup_user(arg, size);
9413 return PTR_ERR(res);
9417 for (i = 0; i < nr_args; i++) {
9418 switch (res[i].opcode) {
9419 case IORING_RESTRICTION_REGISTER_OP:
9420 if (res[i].register_op >= IORING_REGISTER_LAST) {
9425 __set_bit(res[i].register_op,
9426 ctx->restrictions.register_op);
9428 case IORING_RESTRICTION_SQE_OP:
9429 if (res[i].sqe_op >= IORING_OP_LAST) {
9434 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9436 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9437 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9439 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9440 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9449 /* Reset all restrictions if an error happened */
9451 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9453 ctx->restrictions.registered = true;
9459 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9461 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9464 if (ctx->restrictions.registered)
9465 ctx->restricted = 1;
9467 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9469 io_sq_offload_start(ctx);
9474 static bool io_register_op_must_quiesce(int op)
9477 case IORING_UNREGISTER_FILES:
9478 case IORING_REGISTER_FILES_UPDATE:
9479 case IORING_REGISTER_PROBE:
9480 case IORING_REGISTER_PERSONALITY:
9481 case IORING_UNREGISTER_PERSONALITY:
9488 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9489 void __user *arg, unsigned nr_args)
9490 __releases(ctx->uring_lock)
9491 __acquires(ctx->uring_lock)
9496 * We're inside the ring mutex, if the ref is already dying, then
9497 * someone else killed the ctx or is already going through
9498 * io_uring_register().
9500 if (percpu_ref_is_dying(&ctx->refs))
9503 if (io_register_op_must_quiesce(opcode)) {
9504 percpu_ref_kill(&ctx->refs);
9507 * Drop uring mutex before waiting for references to exit. If
9508 * another thread is currently inside io_uring_enter() it might
9509 * need to grab the uring_lock to make progress. If we hold it
9510 * here across the drain wait, then we can deadlock. It's safe
9511 * to drop the mutex here, since no new references will come in
9512 * after we've killed the percpu ref.
9514 mutex_unlock(&ctx->uring_lock);
9516 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9519 ret = io_run_task_work_sig();
9524 mutex_lock(&ctx->uring_lock);
9527 percpu_ref_resurrect(&ctx->refs);
9532 if (ctx->restricted) {
9533 if (opcode >= IORING_REGISTER_LAST) {
9538 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9545 case IORING_REGISTER_BUFFERS:
9546 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9548 case IORING_UNREGISTER_BUFFERS:
9552 ret = io_sqe_buffer_unregister(ctx);
9554 case IORING_REGISTER_FILES:
9555 ret = io_sqe_files_register(ctx, arg, nr_args);
9557 case IORING_UNREGISTER_FILES:
9561 ret = io_sqe_files_unregister(ctx);
9563 case IORING_REGISTER_FILES_UPDATE:
9564 ret = io_sqe_files_update(ctx, arg, nr_args);
9566 case IORING_REGISTER_EVENTFD:
9567 case IORING_REGISTER_EVENTFD_ASYNC:
9571 ret = io_eventfd_register(ctx, arg);
9574 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9575 ctx->eventfd_async = 1;
9577 ctx->eventfd_async = 0;
9579 case IORING_UNREGISTER_EVENTFD:
9583 ret = io_eventfd_unregister(ctx);
9585 case IORING_REGISTER_PROBE:
9587 if (!arg || nr_args > 256)
9589 ret = io_probe(ctx, arg, nr_args);
9591 case IORING_REGISTER_PERSONALITY:
9595 ret = io_register_personality(ctx);
9597 case IORING_UNREGISTER_PERSONALITY:
9601 ret = io_unregister_personality(ctx, nr_args);
9603 case IORING_REGISTER_ENABLE_RINGS:
9607 ret = io_register_enable_rings(ctx);
9609 case IORING_REGISTER_RESTRICTIONS:
9610 ret = io_register_restrictions(ctx, arg, nr_args);
9618 if (io_register_op_must_quiesce(opcode)) {
9619 /* bring the ctx back to life */
9620 percpu_ref_reinit(&ctx->refs);
9622 reinit_completion(&ctx->ref_comp);
9627 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9628 void __user *, arg, unsigned int, nr_args)
9630 struct io_ring_ctx *ctx;
9639 if (f.file->f_op != &io_uring_fops)
9642 ctx = f.file->private_data;
9644 mutex_lock(&ctx->uring_lock);
9645 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9646 mutex_unlock(&ctx->uring_lock);
9647 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9648 ctx->cq_ev_fd != NULL, ret);
9654 static int __init io_uring_init(void)
9656 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9657 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9658 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9661 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9662 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9663 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9664 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9665 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9666 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9667 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9668 BUILD_BUG_SQE_ELEM(8, __u64, off);
9669 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9670 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9671 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9672 BUILD_BUG_SQE_ELEM(24, __u32, len);
9673 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9674 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9675 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9676 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9677 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9678 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9679 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9680 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9681 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9682 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9683 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9684 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9685 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9686 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9687 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9688 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9689 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9690 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9691 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9693 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9694 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9695 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9698 __initcall(io_uring_init);