1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
210 struct fixed_file_data {
211 struct fixed_file_table *table;
212 struct io_ring_ctx *ctx;
214 struct fixed_file_ref_node *node;
215 struct percpu_ref refs;
216 struct completion done;
217 struct list_head ref_list;
222 struct list_head list;
228 struct io_restriction {
229 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
230 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
231 u8 sqe_flags_allowed;
232 u8 sqe_flags_required;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list;
242 struct list_head ctx_new_list;
243 struct mutex ctx_lock;
245 struct task_struct *thread;
246 struct wait_queue_head wait;
251 struct percpu_ref refs;
252 } ____cacheline_aligned_in_smp;
256 unsigned int compat: 1;
257 unsigned int limit_mem: 1;
258 unsigned int cq_overflow_flushed: 1;
259 unsigned int drain_next: 1;
260 unsigned int eventfd_async: 1;
261 unsigned int restricted: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head;
278 unsigned sq_thread_idle;
279 unsigned cached_sq_dropped;
280 unsigned cached_cq_overflow;
281 unsigned long sq_check_overflow;
283 struct list_head defer_list;
284 struct list_head timeout_list;
285 struct list_head cq_overflow_list;
287 wait_queue_head_t inflight_wait;
288 struct io_uring_sqe *sq_sqes;
289 } ____cacheline_aligned_in_smp;
291 struct io_rings *rings;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct *sqo_task;
302 /* Only used for accounting purposes */
303 struct mm_struct *mm_account;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state *sqo_blkcg_css;
309 struct io_sq_data *sq_data; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait;
312 struct wait_queue_entry sqo_wait_entry;
313 struct list_head sqd_list;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data *file_data;
321 unsigned nr_user_files;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs;
325 struct io_mapped_ubuf *user_bufs;
327 struct user_struct *user;
329 const struct cred *creds;
333 unsigned int sessionid;
336 struct completion ref_comp;
337 struct completion sq_thread_comp;
339 /* if all else fails... */
340 struct io_kiocb *fallback_req;
342 #if defined(CONFIG_UNIX)
343 struct socket *ring_sock;
346 struct idr io_buffer_idr;
348 struct idr personality_idr;
351 unsigned cached_cq_tail;
354 atomic_t cq_timeouts;
355 unsigned long cq_check_overflow;
356 struct wait_queue_head cq_wait;
357 struct fasync_struct *cq_fasync;
358 struct eventfd_ctx *cq_ev_fd;
359 } ____cacheline_aligned_in_smp;
362 struct mutex uring_lock;
363 wait_queue_head_t wait;
364 } ____cacheline_aligned_in_smp;
367 spinlock_t completion_lock;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list;
376 struct hlist_head *cancel_hash;
377 unsigned cancel_hash_bits;
378 bool poll_multi_file;
380 spinlock_t inflight_lock;
381 struct list_head inflight_list;
382 } ____cacheline_aligned_in_smp;
384 struct delayed_work file_put_work;
385 struct llist_head file_put_llist;
387 struct work_struct exit_work;
388 struct io_restriction restrictions;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb {
398 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
409 struct file *put_file;
413 struct io_timeout_data {
414 struct io_kiocb *req;
415 struct hrtimer timer;
416 struct timespec64 ts;
417 enum hrtimer_mode mode;
422 struct sockaddr __user *addr;
423 int __user *addr_len;
425 unsigned long nofile;
445 struct list_head list;
448 struct io_timeout_rem {
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user *addr;
469 struct user_msghdr __user *umsg;
475 struct io_buffer *kbuf;
481 struct filename *filename;
483 unsigned long nofile;
486 struct io_files_update {
512 struct epoll_event event;
516 struct file *file_out;
517 struct file *file_in;
524 struct io_provide_buf {
538 const char __user *filename;
539 struct statx __user *buffer;
542 struct io_completion {
544 struct list_head list;
548 struct io_async_connect {
549 struct sockaddr_storage address;
552 struct io_async_msghdr {
553 struct iovec fast_iov[UIO_FASTIOV];
555 struct sockaddr __user *uaddr;
557 struct sockaddr_storage addr;
561 struct iovec fast_iov[UIO_FASTIOV];
562 const struct iovec *free_iovec;
563 struct iov_iter iter;
565 struct wait_page_queue wpq;
569 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
570 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
571 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
572 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
573 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
574 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
581 REQ_F_LINK_TIMEOUT_BIT,
583 REQ_F_NEED_CLEANUP_BIT,
585 REQ_F_BUFFER_SELECTED_BIT,
586 REQ_F_NO_FILE_TABLE_BIT,
587 REQ_F_WORK_INITIALIZED_BIT,
589 /* not a real bit, just to check we're not overflowing the space */
595 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
596 /* drain existing IO first */
597 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
599 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
600 /* doesn't sever on completion < 0 */
601 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
603 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
604 /* IOSQE_BUFFER_SELECT */
605 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
608 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
609 /* fail rest of links */
610 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
611 /* on inflight list */
612 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
613 /* read/write uses file position */
614 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
615 /* must not punt to workers */
616 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
617 /* has linked timeout */
618 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
620 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
622 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
623 /* already went through poll handler */
624 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
625 /* buffer already selected */
626 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
627 /* doesn't need file table for this request */
628 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
629 /* io_wq_work is initialized */
630 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
634 struct io_poll_iocb poll;
635 struct io_poll_iocb *double_poll;
639 * NOTE! Each of the iocb union members has the file pointer
640 * as the first entry in their struct definition. So you can
641 * access the file pointer through any of the sub-structs,
642 * or directly as just 'ki_filp' in this struct.
648 struct io_poll_iocb poll;
649 struct io_accept accept;
651 struct io_cancel cancel;
652 struct io_timeout timeout;
653 struct io_timeout_rem timeout_rem;
654 struct io_connect connect;
655 struct io_sr_msg sr_msg;
657 struct io_close close;
658 struct io_files_update files_update;
659 struct io_fadvise fadvise;
660 struct io_madvise madvise;
661 struct io_epoll epoll;
662 struct io_splice splice;
663 struct io_provide_buf pbuf;
664 struct io_statx statx;
665 /* use only after cleaning per-op data, see io_clean_op() */
666 struct io_completion compl;
669 /* opcode allocated if it needs to store data for async defer */
672 /* polled IO has completed */
678 struct io_ring_ctx *ctx;
681 struct task_struct *task;
684 struct list_head link_list;
687 * 1. used with ctx->iopoll_list with reads/writes
688 * 2. to track reqs with ->files (see io_op_def::file_table)
690 struct list_head inflight_entry;
692 struct percpu_ref *fixed_file_refs;
693 struct callback_head task_work;
694 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
695 struct hlist_node hash_node;
696 struct async_poll *apoll;
697 struct io_wq_work work;
700 struct io_defer_entry {
701 struct list_head list;
702 struct io_kiocb *req;
706 #define IO_IOPOLL_BATCH 8
708 struct io_comp_state {
710 struct list_head list;
711 struct io_ring_ctx *ctx;
714 struct io_submit_state {
715 struct blk_plug plug;
718 * io_kiocb alloc cache
720 void *reqs[IO_IOPOLL_BATCH];
721 unsigned int free_reqs;
724 * Batch completion logic
726 struct io_comp_state comp;
729 * File reference cache
733 unsigned int has_refs;
734 unsigned int ios_left;
738 /* needs req->file assigned */
739 unsigned needs_file : 1;
740 /* don't fail if file grab fails */
741 unsigned needs_file_no_error : 1;
742 /* hash wq insertion if file is a regular file */
743 unsigned hash_reg_file : 1;
744 /* unbound wq insertion if file is a non-regular file */
745 unsigned unbound_nonreg_file : 1;
746 /* opcode is not supported by this kernel */
747 unsigned not_supported : 1;
748 /* set if opcode supports polled "wait" */
750 unsigned pollout : 1;
751 /* op supports buffer selection */
752 unsigned buffer_select : 1;
753 /* needs rlimit(RLIMIT_FSIZE) assigned */
754 unsigned needs_fsize : 1;
755 /* must always have async data allocated */
756 unsigned needs_async_data : 1;
757 /* size of async data needed, if any */
758 unsigned short async_size;
762 static const struct io_op_def io_op_defs[] = {
763 [IORING_OP_NOP] = {},
764 [IORING_OP_READV] = {
766 .unbound_nonreg_file = 1,
769 .needs_async_data = 1,
770 .async_size = sizeof(struct io_async_rw),
771 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
773 [IORING_OP_WRITEV] = {
776 .unbound_nonreg_file = 1,
779 .needs_async_data = 1,
780 .async_size = sizeof(struct io_async_rw),
781 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
783 [IORING_OP_FSYNC] = {
785 .work_flags = IO_WQ_WORK_BLKCG,
787 [IORING_OP_READ_FIXED] = {
789 .unbound_nonreg_file = 1,
791 .async_size = sizeof(struct io_async_rw),
792 .work_flags = IO_WQ_WORK_BLKCG,
794 [IORING_OP_WRITE_FIXED] = {
797 .unbound_nonreg_file = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_BLKCG,
803 [IORING_OP_POLL_ADD] = {
805 .unbound_nonreg_file = 1,
807 [IORING_OP_POLL_REMOVE] = {},
808 [IORING_OP_SYNC_FILE_RANGE] = {
810 .work_flags = IO_WQ_WORK_BLKCG,
812 [IORING_OP_SENDMSG] = {
814 .unbound_nonreg_file = 1,
816 .needs_async_data = 1,
817 .async_size = sizeof(struct io_async_msghdr),
818 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
821 [IORING_OP_RECVMSG] = {
823 .unbound_nonreg_file = 1,
826 .needs_async_data = 1,
827 .async_size = sizeof(struct io_async_msghdr),
828 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
831 [IORING_OP_TIMEOUT] = {
832 .needs_async_data = 1,
833 .async_size = sizeof(struct io_timeout_data),
834 .work_flags = IO_WQ_WORK_MM,
836 [IORING_OP_TIMEOUT_REMOVE] = {},
837 [IORING_OP_ACCEPT] = {
839 .unbound_nonreg_file = 1,
841 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
843 [IORING_OP_ASYNC_CANCEL] = {},
844 [IORING_OP_LINK_TIMEOUT] = {
845 .needs_async_data = 1,
846 .async_size = sizeof(struct io_timeout_data),
847 .work_flags = IO_WQ_WORK_MM,
849 [IORING_OP_CONNECT] = {
851 .unbound_nonreg_file = 1,
853 .needs_async_data = 1,
854 .async_size = sizeof(struct io_async_connect),
855 .work_flags = IO_WQ_WORK_MM,
857 [IORING_OP_FALLOCATE] = {
860 .work_flags = IO_WQ_WORK_BLKCG,
862 [IORING_OP_OPENAT] = {
863 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
866 [IORING_OP_CLOSE] = {
868 .needs_file_no_error = 1,
869 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
871 [IORING_OP_FILES_UPDATE] = {
872 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
874 [IORING_OP_STATX] = {
875 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
876 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
880 .unbound_nonreg_file = 1,
883 .async_size = sizeof(struct io_async_rw),
884 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
886 [IORING_OP_WRITE] = {
888 .unbound_nonreg_file = 1,
891 .async_size = sizeof(struct io_async_rw),
892 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
894 [IORING_OP_FADVISE] = {
896 .work_flags = IO_WQ_WORK_BLKCG,
898 [IORING_OP_MADVISE] = {
899 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
903 .unbound_nonreg_file = 1,
905 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
909 .unbound_nonreg_file = 1,
912 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 [IORING_OP_OPENAT2] = {
915 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
918 [IORING_OP_EPOLL_CTL] = {
919 .unbound_nonreg_file = 1,
920 .work_flags = IO_WQ_WORK_FILES,
922 [IORING_OP_SPLICE] = {
925 .unbound_nonreg_file = 1,
926 .work_flags = IO_WQ_WORK_BLKCG,
928 [IORING_OP_PROVIDE_BUFFERS] = {},
929 [IORING_OP_REMOVE_BUFFERS] = {},
933 .unbound_nonreg_file = 1,
937 enum io_mem_account {
942 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
943 struct io_comp_state *cs);
944 static void io_cqring_fill_event(struct io_kiocb *req, long res);
945 static void io_put_req(struct io_kiocb *req);
946 static void io_put_req_deferred(struct io_kiocb *req, int nr);
947 static void io_double_put_req(struct io_kiocb *req);
948 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
949 static void __io_queue_linked_timeout(struct io_kiocb *req);
950 static void io_queue_linked_timeout(struct io_kiocb *req);
951 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
952 struct io_uring_files_update *ip,
954 static void __io_clean_op(struct io_kiocb *req);
955 static struct file *io_file_get(struct io_submit_state *state,
956 struct io_kiocb *req, int fd, bool fixed);
957 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
958 static void io_file_put_work(struct work_struct *work);
960 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
961 struct iovec **iovec, struct iov_iter *iter,
963 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
964 const struct iovec *fast_iov,
965 struct iov_iter *iter, bool force);
967 static struct kmem_cache *req_cachep;
969 static const struct file_operations io_uring_fops;
971 struct sock *io_uring_get_socket(struct file *file)
973 #if defined(CONFIG_UNIX)
974 if (file->f_op == &io_uring_fops) {
975 struct io_ring_ctx *ctx = file->private_data;
977 return ctx->ring_sock->sk;
982 EXPORT_SYMBOL(io_uring_get_socket);
984 static inline void io_clean_op(struct io_kiocb *req)
986 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
991 static void io_sq_thread_drop_mm(void)
993 struct mm_struct *mm = current->mm;
996 kthread_unuse_mm(mm);
1001 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1004 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1005 !ctx->sqo_task->mm ||
1006 !mmget_not_zero(ctx->sqo_task->mm)))
1008 kthread_use_mm(ctx->sqo_task->mm);
1014 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1015 struct io_kiocb *req)
1017 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1019 return __io_sq_thread_acquire_mm(ctx);
1022 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1023 struct cgroup_subsys_state **cur_css)
1026 #ifdef CONFIG_BLK_CGROUP
1027 /* puts the old one when swapping */
1028 if (*cur_css != ctx->sqo_blkcg_css) {
1029 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1030 *cur_css = ctx->sqo_blkcg_css;
1035 static void io_sq_thread_unassociate_blkcg(void)
1037 #ifdef CONFIG_BLK_CGROUP
1038 kthread_associate_blkcg(NULL);
1042 static inline void req_set_fail_links(struct io_kiocb *req)
1044 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1045 req->flags |= REQ_F_FAIL_LINK;
1049 * None of these are dereferenced, they are simply used to check if any of
1050 * them have changed. If we're under current and check they are still the
1051 * same, we're fine to grab references to them for actual out-of-line use.
1053 static void io_init_identity(struct io_identity *id)
1055 id->files = current->files;
1056 id->mm = current->mm;
1057 #ifdef CONFIG_BLK_CGROUP
1059 id->blkcg_css = blkcg_css();
1062 id->creds = current_cred();
1063 id->nsproxy = current->nsproxy;
1064 id->fs = current->fs;
1065 id->fsize = rlimit(RLIMIT_FSIZE);
1067 id->loginuid = current->loginuid;
1068 id->sessionid = current->sessionid;
1070 refcount_set(&id->count, 1);
1073 static inline void __io_req_init_async(struct io_kiocb *req)
1075 memset(&req->work, 0, sizeof(req->work));
1076 req->flags |= REQ_F_WORK_INITIALIZED;
1080 * Note: must call io_req_init_async() for the first time you
1081 * touch any members of io_wq_work.
1083 static inline void io_req_init_async(struct io_kiocb *req)
1085 struct io_uring_task *tctx = current->io_uring;
1087 if (req->flags & REQ_F_WORK_INITIALIZED)
1090 __io_req_init_async(req);
1092 /* Grab a ref if this isn't our static identity */
1093 req->work.identity = tctx->identity;
1094 if (tctx->identity != &tctx->__identity)
1095 refcount_inc(&req->work.identity->count);
1098 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1100 return ctx->flags & IORING_SETUP_SQPOLL;
1103 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1105 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1107 complete(&ctx->ref_comp);
1110 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1112 return !req->timeout.off;
1115 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1117 struct io_ring_ctx *ctx;
1120 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1124 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1125 if (!ctx->fallback_req)
1129 * Use 5 bits less than the max cq entries, that should give us around
1130 * 32 entries per hash list if totally full and uniformly spread.
1132 hash_bits = ilog2(p->cq_entries);
1136 ctx->cancel_hash_bits = hash_bits;
1137 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1139 if (!ctx->cancel_hash)
1141 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1143 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1144 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1147 ctx->flags = p->flags;
1148 init_waitqueue_head(&ctx->sqo_sq_wait);
1149 INIT_LIST_HEAD(&ctx->sqd_list);
1150 init_waitqueue_head(&ctx->cq_wait);
1151 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1152 init_completion(&ctx->ref_comp);
1153 init_completion(&ctx->sq_thread_comp);
1154 idr_init(&ctx->io_buffer_idr);
1155 idr_init(&ctx->personality_idr);
1156 mutex_init(&ctx->uring_lock);
1157 init_waitqueue_head(&ctx->wait);
1158 spin_lock_init(&ctx->completion_lock);
1159 INIT_LIST_HEAD(&ctx->iopoll_list);
1160 INIT_LIST_HEAD(&ctx->defer_list);
1161 INIT_LIST_HEAD(&ctx->timeout_list);
1162 init_waitqueue_head(&ctx->inflight_wait);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1166 init_llist_head(&ctx->file_put_llist);
1169 if (ctx->fallback_req)
1170 kmem_cache_free(req_cachep, ctx->fallback_req);
1171 kfree(ctx->cancel_hash);
1176 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1178 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1179 struct io_ring_ctx *ctx = req->ctx;
1181 return seq != ctx->cached_cq_tail
1182 + READ_ONCE(ctx->cached_cq_overflow);
1188 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1190 struct io_rings *rings = ctx->rings;
1192 /* order cqe stores with ring update */
1193 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1195 if (wq_has_sleeper(&ctx->cq_wait)) {
1196 wake_up_interruptible(&ctx->cq_wait);
1197 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1201 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1203 if (req->work.identity == &tctx->__identity)
1205 if (refcount_dec_and_test(&req->work.identity->count))
1206 kfree(req->work.identity);
1209 static void io_req_clean_work(struct io_kiocb *req)
1211 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1214 req->flags &= ~REQ_F_WORK_INITIALIZED;
1216 if (req->work.flags & IO_WQ_WORK_MM) {
1217 mmdrop(req->work.identity->mm);
1218 req->work.flags &= ~IO_WQ_WORK_MM;
1220 #ifdef CONFIG_BLK_CGROUP
1221 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1222 css_put(req->work.identity->blkcg_css);
1223 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1226 if (req->work.flags & IO_WQ_WORK_CREDS) {
1227 put_cred(req->work.identity->creds);
1228 req->work.flags &= ~IO_WQ_WORK_CREDS;
1230 if (req->work.flags & IO_WQ_WORK_FS) {
1231 struct fs_struct *fs = req->work.identity->fs;
1233 spin_lock(&req->work.identity->fs->lock);
1236 spin_unlock(&req->work.identity->fs->lock);
1239 req->work.flags &= ~IO_WQ_WORK_FS;
1242 io_put_identity(req->task->io_uring, req);
1246 * Create a private copy of io_identity, since some fields don't match
1247 * the current context.
1249 static bool io_identity_cow(struct io_kiocb *req)
1251 struct io_uring_task *tctx = current->io_uring;
1252 const struct cred *creds = NULL;
1253 struct io_identity *id;
1255 if (req->work.flags & IO_WQ_WORK_CREDS)
1256 creds = req->work.identity->creds;
1258 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1259 if (unlikely(!id)) {
1260 req->work.flags |= IO_WQ_WORK_CANCEL;
1265 * We can safely just re-init the creds we copied Either the field
1266 * matches the current one, or we haven't grabbed it yet. The only
1267 * exception is ->creds, through registered personalities, so handle
1268 * that one separately.
1270 io_init_identity(id);
1272 req->work.identity->creds = creds;
1274 /* add one for this request */
1275 refcount_inc(&id->count);
1277 /* drop old identity, assign new one. one ref for req, one for tctx */
1278 if (req->work.identity != tctx->identity &&
1279 refcount_sub_and_test(2, &req->work.identity->count))
1280 kfree(req->work.identity);
1282 req->work.identity = id;
1283 tctx->identity = id;
1287 static bool io_grab_identity(struct io_kiocb *req)
1289 const struct io_op_def *def = &io_op_defs[req->opcode];
1290 struct io_identity *id = req->work.identity;
1291 struct io_ring_ctx *ctx = req->ctx;
1293 if (def->needs_fsize && id->fsize != rlimit(RLIMIT_FSIZE))
1296 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1297 (def->work_flags & IO_WQ_WORK_FILES) &&
1298 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1299 if (id->files != current->files ||
1300 id->nsproxy != current->nsproxy)
1302 atomic_inc(&id->files->count);
1303 get_nsproxy(id->nsproxy);
1304 req->flags |= REQ_F_INFLIGHT;
1306 spin_lock_irq(&ctx->inflight_lock);
1307 list_add(&req->inflight_entry, &ctx->inflight_list);
1308 spin_unlock_irq(&ctx->inflight_lock);
1309 req->work.flags |= IO_WQ_WORK_FILES;
1311 #ifdef CONFIG_BLK_CGROUP
1312 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1313 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1315 if (id->blkcg_css != blkcg_css()) {
1320 * This should be rare, either the cgroup is dying or the task
1321 * is moving cgroups. Just punt to root for the handful of ios.
1323 if (css_tryget_online(id->blkcg_css))
1324 req->work.flags |= IO_WQ_WORK_BLKCG;
1328 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1329 if (id->creds != current_cred())
1331 get_cred(id->creds);
1332 req->work.flags |= IO_WQ_WORK_CREDS;
1335 if (!uid_eq(current->loginuid, id->loginuid) ||
1336 current->sessionid != id->sessionid)
1339 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1340 (def->work_flags & IO_WQ_WORK_FS)) {
1341 if (current->fs != id->fs)
1343 spin_lock(&id->fs->lock);
1344 if (!id->fs->in_exec) {
1346 req->work.flags |= IO_WQ_WORK_FS;
1348 req->work.flags |= IO_WQ_WORK_CANCEL;
1350 spin_unlock(¤t->fs->lock);
1356 static void io_prep_async_work(struct io_kiocb *req)
1358 const struct io_op_def *def = &io_op_defs[req->opcode];
1359 struct io_ring_ctx *ctx = req->ctx;
1360 struct io_identity *id;
1362 io_req_init_async(req);
1363 id = req->work.identity;
1365 if (req->flags & REQ_F_ISREG) {
1366 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1367 io_wq_hash_work(&req->work, file_inode(req->file));
1369 if (def->unbound_nonreg_file)
1370 req->work.flags |= IO_WQ_WORK_UNBOUND;
1373 /* ->mm can never change on us */
1374 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1375 (def->work_flags & IO_WQ_WORK_MM)) {
1377 req->work.flags |= IO_WQ_WORK_MM;
1380 /* if we fail grabbing identity, we must COW, regrab, and retry */
1381 if (io_grab_identity(req))
1384 if (!io_identity_cow(req))
1387 /* can't fail at this point */
1388 if (!io_grab_identity(req))
1392 static void io_prep_async_link(struct io_kiocb *req)
1394 struct io_kiocb *cur;
1396 io_prep_async_work(req);
1397 if (req->flags & REQ_F_LINK_HEAD)
1398 list_for_each_entry(cur, &req->link_list, link_list)
1399 io_prep_async_work(cur);
1402 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1404 struct io_ring_ctx *ctx = req->ctx;
1405 struct io_kiocb *link = io_prep_linked_timeout(req);
1407 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1408 &req->work, req->flags);
1409 io_wq_enqueue(ctx->io_wq, &req->work);
1413 static void io_queue_async_work(struct io_kiocb *req)
1415 struct io_kiocb *link;
1417 /* init ->work of the whole link before punting */
1418 io_prep_async_link(req);
1419 link = __io_queue_async_work(req);
1422 io_queue_linked_timeout(link);
1425 static void io_kill_timeout(struct io_kiocb *req)
1427 struct io_timeout_data *io = req->async_data;
1430 ret = hrtimer_try_to_cancel(&io->timer);
1432 atomic_set(&req->ctx->cq_timeouts,
1433 atomic_read(&req->ctx->cq_timeouts) + 1);
1434 list_del_init(&req->timeout.list);
1435 io_cqring_fill_event(req, 0);
1436 io_put_req_deferred(req, 1);
1440 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1442 struct io_ring_ctx *ctx = req->ctx;
1444 if (!tsk || req->task == tsk)
1446 if (ctx->flags & IORING_SETUP_SQPOLL) {
1447 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1454 * Returns true if we found and killed one or more timeouts
1456 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1458 struct io_kiocb *req, *tmp;
1461 spin_lock_irq(&ctx->completion_lock);
1462 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1463 if (io_task_match(req, tsk)) {
1464 io_kill_timeout(req);
1468 spin_unlock_irq(&ctx->completion_lock);
1469 return canceled != 0;
1472 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1475 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1476 struct io_defer_entry, list);
1477 struct io_kiocb *link;
1479 if (req_need_defer(de->req, de->seq))
1481 list_del_init(&de->list);
1482 /* punt-init is done before queueing for defer */
1483 link = __io_queue_async_work(de->req);
1485 __io_queue_linked_timeout(link);
1486 /* drop submission reference */
1487 io_put_req_deferred(link, 1);
1490 } while (!list_empty(&ctx->defer_list));
1493 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1495 while (!list_empty(&ctx->timeout_list)) {
1496 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1497 struct io_kiocb, timeout.list);
1499 if (io_is_timeout_noseq(req))
1501 if (req->timeout.target_seq != ctx->cached_cq_tail
1502 - atomic_read(&ctx->cq_timeouts))
1505 list_del_init(&req->timeout.list);
1506 io_kill_timeout(req);
1510 static void io_commit_cqring(struct io_ring_ctx *ctx)
1512 io_flush_timeouts(ctx);
1513 __io_commit_cqring(ctx);
1515 if (unlikely(!list_empty(&ctx->defer_list)))
1516 __io_queue_deferred(ctx);
1519 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1521 struct io_rings *r = ctx->rings;
1523 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1526 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1528 struct io_rings *rings = ctx->rings;
1531 tail = ctx->cached_cq_tail;
1533 * writes to the cq entry need to come after reading head; the
1534 * control dependency is enough as we're using WRITE_ONCE to
1537 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1540 ctx->cached_cq_tail++;
1541 return &rings->cqes[tail & ctx->cq_mask];
1544 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1548 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1550 if (!ctx->eventfd_async)
1552 return io_wq_current_is_worker();
1555 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1557 if (waitqueue_active(&ctx->wait))
1558 wake_up(&ctx->wait);
1559 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1560 wake_up(&ctx->sq_data->wait);
1561 if (io_should_trigger_evfd(ctx))
1562 eventfd_signal(ctx->cq_ev_fd, 1);
1565 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1567 if (list_empty(&ctx->cq_overflow_list)) {
1568 clear_bit(0, &ctx->sq_check_overflow);
1569 clear_bit(0, &ctx->cq_check_overflow);
1570 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1574 static inline bool io_match_files(struct io_kiocb *req,
1575 struct files_struct *files)
1579 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1580 (req->work.flags & IO_WQ_WORK_FILES))
1581 return req->work.identity->files == files;
1585 /* Returns true if there are no backlogged entries after the flush */
1586 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1587 struct task_struct *tsk,
1588 struct files_struct *files)
1590 struct io_rings *rings = ctx->rings;
1591 struct io_kiocb *req, *tmp;
1592 struct io_uring_cqe *cqe;
1593 unsigned long flags;
1597 if (list_empty_careful(&ctx->cq_overflow_list))
1599 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1600 rings->cq_ring_entries))
1604 spin_lock_irqsave(&ctx->completion_lock, flags);
1606 /* if force is set, the ring is going away. always drop after that */
1608 ctx->cq_overflow_flushed = 1;
1611 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1612 if (tsk && req->task != tsk)
1614 if (!io_match_files(req, files))
1617 cqe = io_get_cqring(ctx);
1621 list_move(&req->compl.list, &list);
1623 WRITE_ONCE(cqe->user_data, req->user_data);
1624 WRITE_ONCE(cqe->res, req->result);
1625 WRITE_ONCE(cqe->flags, req->compl.cflags);
1627 ctx->cached_cq_overflow++;
1628 WRITE_ONCE(ctx->rings->cq_overflow,
1629 ctx->cached_cq_overflow);
1633 io_commit_cqring(ctx);
1634 io_cqring_mark_overflow(ctx);
1636 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1637 io_cqring_ev_posted(ctx);
1639 while (!list_empty(&list)) {
1640 req = list_first_entry(&list, struct io_kiocb, compl.list);
1641 list_del(&req->compl.list);
1648 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1650 struct io_ring_ctx *ctx = req->ctx;
1651 struct io_uring_cqe *cqe;
1653 trace_io_uring_complete(ctx, req->user_data, res);
1656 * If we can't get a cq entry, userspace overflowed the
1657 * submission (by quite a lot). Increment the overflow count in
1660 cqe = io_get_cqring(ctx);
1662 WRITE_ONCE(cqe->user_data, req->user_data);
1663 WRITE_ONCE(cqe->res, res);
1664 WRITE_ONCE(cqe->flags, cflags);
1665 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1667 * If we're in ring overflow flush mode, or in task cancel mode,
1668 * then we cannot store the request for later flushing, we need
1669 * to drop it on the floor.
1671 ctx->cached_cq_overflow++;
1672 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1674 if (list_empty(&ctx->cq_overflow_list)) {
1675 set_bit(0, &ctx->sq_check_overflow);
1676 set_bit(0, &ctx->cq_check_overflow);
1677 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1681 req->compl.cflags = cflags;
1682 refcount_inc(&req->refs);
1683 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1687 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1689 __io_cqring_fill_event(req, res, 0);
1692 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1694 struct io_ring_ctx *ctx = req->ctx;
1695 unsigned long flags;
1697 spin_lock_irqsave(&ctx->completion_lock, flags);
1698 __io_cqring_fill_event(req, res, cflags);
1699 io_commit_cqring(ctx);
1700 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1702 io_cqring_ev_posted(ctx);
1705 static void io_submit_flush_completions(struct io_comp_state *cs)
1707 struct io_ring_ctx *ctx = cs->ctx;
1709 spin_lock_irq(&ctx->completion_lock);
1710 while (!list_empty(&cs->list)) {
1711 struct io_kiocb *req;
1713 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1714 list_del(&req->compl.list);
1715 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1718 * io_free_req() doesn't care about completion_lock unless one
1719 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1720 * because of a potential deadlock with req->work.fs->lock
1722 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1723 |REQ_F_WORK_INITIALIZED)) {
1724 spin_unlock_irq(&ctx->completion_lock);
1726 spin_lock_irq(&ctx->completion_lock);
1731 io_commit_cqring(ctx);
1732 spin_unlock_irq(&ctx->completion_lock);
1734 io_cqring_ev_posted(ctx);
1738 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1739 struct io_comp_state *cs)
1742 io_cqring_add_event(req, res, cflags);
1747 req->compl.cflags = cflags;
1748 list_add_tail(&req->compl.list, &cs->list);
1750 io_submit_flush_completions(cs);
1754 static void io_req_complete(struct io_kiocb *req, long res)
1756 __io_req_complete(req, res, 0, NULL);
1759 static inline bool io_is_fallback_req(struct io_kiocb *req)
1761 return req == (struct io_kiocb *)
1762 ((unsigned long) req->ctx->fallback_req & ~1UL);
1765 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1767 struct io_kiocb *req;
1769 req = ctx->fallback_req;
1770 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1776 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1777 struct io_submit_state *state)
1779 if (!state->free_reqs) {
1780 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1784 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1785 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1788 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1789 * retry single alloc to be on the safe side.
1791 if (unlikely(ret <= 0)) {
1792 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1793 if (!state->reqs[0])
1797 state->free_reqs = ret;
1801 return state->reqs[state->free_reqs];
1803 return io_get_fallback_req(ctx);
1806 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1810 percpu_ref_put(req->fixed_file_refs);
1815 static void io_dismantle_req(struct io_kiocb *req)
1819 if (req->async_data)
1820 kfree(req->async_data);
1822 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1824 io_req_clean_work(req);
1827 static void __io_free_req(struct io_kiocb *req)
1829 struct io_uring_task *tctx = req->task->io_uring;
1830 struct io_ring_ctx *ctx = req->ctx;
1832 io_dismantle_req(req);
1834 percpu_counter_dec(&tctx->inflight);
1836 wake_up(&tctx->wait);
1837 put_task_struct(req->task);
1839 if (likely(!io_is_fallback_req(req)))
1840 kmem_cache_free(req_cachep, req);
1842 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1843 percpu_ref_put(&ctx->refs);
1846 static bool io_link_cancel_timeout(struct io_kiocb *req)
1848 struct io_timeout_data *io = req->async_data;
1849 struct io_ring_ctx *ctx = req->ctx;
1852 ret = hrtimer_try_to_cancel(&io->timer);
1854 io_cqring_fill_event(req, -ECANCELED);
1855 io_commit_cqring(ctx);
1856 req->flags &= ~REQ_F_LINK_HEAD;
1857 io_put_req_deferred(req, 1);
1864 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1866 struct io_kiocb *link;
1869 if (list_empty(&req->link_list))
1871 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1872 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1875 list_del_init(&link->link_list);
1876 wake_ev = io_link_cancel_timeout(link);
1877 req->flags &= ~REQ_F_LINK_TIMEOUT;
1881 static void io_kill_linked_timeout(struct io_kiocb *req)
1883 struct io_ring_ctx *ctx = req->ctx;
1884 unsigned long flags;
1887 spin_lock_irqsave(&ctx->completion_lock, flags);
1888 wake_ev = __io_kill_linked_timeout(req);
1889 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1892 io_cqring_ev_posted(ctx);
1895 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1897 struct io_kiocb *nxt;
1900 * The list should never be empty when we are called here. But could
1901 * potentially happen if the chain is messed up, check to be on the
1904 if (unlikely(list_empty(&req->link_list)))
1907 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1908 list_del_init(&req->link_list);
1909 if (!list_empty(&nxt->link_list))
1910 nxt->flags |= REQ_F_LINK_HEAD;
1915 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1917 static void io_fail_links(struct io_kiocb *req)
1919 struct io_ring_ctx *ctx = req->ctx;
1920 unsigned long flags;
1922 spin_lock_irqsave(&ctx->completion_lock, flags);
1923 while (!list_empty(&req->link_list)) {
1924 struct io_kiocb *link = list_first_entry(&req->link_list,
1925 struct io_kiocb, link_list);
1927 list_del_init(&link->link_list);
1928 trace_io_uring_fail_link(req, link);
1930 io_cqring_fill_event(link, -ECANCELED);
1933 * It's ok to free under spinlock as they're not linked anymore,
1934 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1937 if (link->flags & REQ_F_WORK_INITIALIZED)
1938 io_put_req_deferred(link, 2);
1940 io_double_put_req(link);
1943 io_commit_cqring(ctx);
1944 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1946 io_cqring_ev_posted(ctx);
1949 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1951 req->flags &= ~REQ_F_LINK_HEAD;
1952 if (req->flags & REQ_F_LINK_TIMEOUT)
1953 io_kill_linked_timeout(req);
1956 * If LINK is set, we have dependent requests in this chain. If we
1957 * didn't fail this request, queue the first one up, moving any other
1958 * dependencies to the next request. In case of failure, fail the rest
1961 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1962 return io_req_link_next(req);
1967 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1969 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1971 return __io_req_find_next(req);
1974 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1976 struct task_struct *tsk = req->task;
1977 struct io_ring_ctx *ctx = req->ctx;
1980 if (tsk->flags & PF_EXITING)
1984 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1985 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1986 * processing task_work. There's no reliable way to tell if TWA_RESUME
1990 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1991 notify = TWA_SIGNAL;
1993 ret = task_work_add(tsk, &req->task_work, notify);
1995 wake_up_process(tsk);
2000 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2002 struct io_ring_ctx *ctx = req->ctx;
2004 spin_lock_irq(&ctx->completion_lock);
2005 io_cqring_fill_event(req, error);
2006 io_commit_cqring(ctx);
2007 spin_unlock_irq(&ctx->completion_lock);
2009 io_cqring_ev_posted(ctx);
2010 req_set_fail_links(req);
2011 io_double_put_req(req);
2014 static void io_req_task_cancel(struct callback_head *cb)
2016 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2017 struct io_ring_ctx *ctx = req->ctx;
2019 __io_req_task_cancel(req, -ECANCELED);
2020 percpu_ref_put(&ctx->refs);
2023 static void __io_req_task_submit(struct io_kiocb *req)
2025 struct io_ring_ctx *ctx = req->ctx;
2027 if (!__io_sq_thread_acquire_mm(ctx)) {
2028 mutex_lock(&ctx->uring_lock);
2029 __io_queue_sqe(req, NULL);
2030 mutex_unlock(&ctx->uring_lock);
2032 __io_req_task_cancel(req, -EFAULT);
2036 static void io_req_task_submit(struct callback_head *cb)
2038 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2039 struct io_ring_ctx *ctx = req->ctx;
2041 __io_req_task_submit(req);
2042 percpu_ref_put(&ctx->refs);
2045 static void io_req_task_queue(struct io_kiocb *req)
2049 init_task_work(&req->task_work, io_req_task_submit);
2050 percpu_ref_get(&req->ctx->refs);
2052 ret = io_req_task_work_add(req, true);
2053 if (unlikely(ret)) {
2054 struct task_struct *tsk;
2056 init_task_work(&req->task_work, io_req_task_cancel);
2057 tsk = io_wq_get_task(req->ctx->io_wq);
2058 task_work_add(tsk, &req->task_work, 0);
2059 wake_up_process(tsk);
2063 static void io_queue_next(struct io_kiocb *req)
2065 struct io_kiocb *nxt = io_req_find_next(req);
2068 io_req_task_queue(nxt);
2071 static void io_free_req(struct io_kiocb *req)
2078 void *reqs[IO_IOPOLL_BATCH];
2081 struct task_struct *task;
2085 static inline void io_init_req_batch(struct req_batch *rb)
2092 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2093 struct req_batch *rb)
2095 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2096 percpu_ref_put_many(&ctx->refs, rb->to_free);
2100 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2101 struct req_batch *rb)
2104 __io_req_free_batch_flush(ctx, rb);
2106 struct io_uring_task *tctx = rb->task->io_uring;
2108 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2109 put_task_struct_many(rb->task, rb->task_refs);
2114 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2116 if (unlikely(io_is_fallback_req(req))) {
2120 if (req->flags & REQ_F_LINK_HEAD)
2123 if (req->task != rb->task) {
2125 struct io_uring_task *tctx = rb->task->io_uring;
2127 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2128 put_task_struct_many(rb->task, rb->task_refs);
2130 rb->task = req->task;
2135 io_dismantle_req(req);
2136 rb->reqs[rb->to_free++] = req;
2137 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2138 __io_req_free_batch_flush(req->ctx, rb);
2142 * Drop reference to request, return next in chain (if there is one) if this
2143 * was the last reference to this request.
2145 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2147 struct io_kiocb *nxt = NULL;
2149 if (refcount_dec_and_test(&req->refs)) {
2150 nxt = io_req_find_next(req);
2156 static void io_put_req(struct io_kiocb *req)
2158 if (refcount_dec_and_test(&req->refs))
2162 static void io_put_req_deferred_cb(struct callback_head *cb)
2164 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2169 static void io_free_req_deferred(struct io_kiocb *req)
2173 init_task_work(&req->task_work, io_put_req_deferred_cb);
2174 ret = io_req_task_work_add(req, true);
2175 if (unlikely(ret)) {
2176 struct task_struct *tsk;
2178 tsk = io_wq_get_task(req->ctx->io_wq);
2179 task_work_add(tsk, &req->task_work, 0);
2180 wake_up_process(tsk);
2184 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2186 if (refcount_sub_and_test(refs, &req->refs))
2187 io_free_req_deferred(req);
2190 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2192 struct io_kiocb *nxt;
2195 * A ref is owned by io-wq in which context we're. So, if that's the
2196 * last one, it's safe to steal next work. False negatives are Ok,
2197 * it just will be re-punted async in io_put_work()
2199 if (refcount_read(&req->refs) != 1)
2202 nxt = io_req_find_next(req);
2203 return nxt ? &nxt->work : NULL;
2206 static void io_double_put_req(struct io_kiocb *req)
2208 /* drop both submit and complete references */
2209 if (refcount_sub_and_test(2, &req->refs))
2213 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2215 struct io_rings *rings = ctx->rings;
2217 if (test_bit(0, &ctx->cq_check_overflow)) {
2219 * noflush == true is from the waitqueue handler, just ensure
2220 * we wake up the task, and the next invocation will flush the
2221 * entries. We cannot safely to it from here.
2223 if (noflush && !list_empty(&ctx->cq_overflow_list))
2226 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2229 /* See comment at the top of this file */
2231 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2234 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2236 struct io_rings *rings = ctx->rings;
2238 /* make sure SQ entry isn't read before tail */
2239 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2242 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2244 unsigned int cflags;
2246 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2247 cflags |= IORING_CQE_F_BUFFER;
2248 req->flags &= ~REQ_F_BUFFER_SELECTED;
2253 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2255 struct io_buffer *kbuf;
2257 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2258 return io_put_kbuf(req, kbuf);
2261 static inline bool io_run_task_work(void)
2264 * Not safe to run on exiting task, and the task_work handling will
2265 * not add work to such a task.
2267 if (unlikely(current->flags & PF_EXITING))
2269 if (current->task_works) {
2270 __set_current_state(TASK_RUNNING);
2278 static void io_iopoll_queue(struct list_head *again)
2280 struct io_kiocb *req;
2283 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2284 list_del(&req->inflight_entry);
2285 __io_complete_rw(req, -EAGAIN, 0, NULL);
2286 } while (!list_empty(again));
2290 * Find and free completed poll iocbs
2292 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2293 struct list_head *done)
2295 struct req_batch rb;
2296 struct io_kiocb *req;
2299 /* order with ->result store in io_complete_rw_iopoll() */
2302 io_init_req_batch(&rb);
2303 while (!list_empty(done)) {
2306 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2307 if (READ_ONCE(req->result) == -EAGAIN) {
2309 req->iopoll_completed = 0;
2310 list_move_tail(&req->inflight_entry, &again);
2313 list_del(&req->inflight_entry);
2315 if (req->flags & REQ_F_BUFFER_SELECTED)
2316 cflags = io_put_rw_kbuf(req);
2318 __io_cqring_fill_event(req, req->result, cflags);
2321 if (refcount_dec_and_test(&req->refs))
2322 io_req_free_batch(&rb, req);
2325 io_commit_cqring(ctx);
2326 if (ctx->flags & IORING_SETUP_SQPOLL)
2327 io_cqring_ev_posted(ctx);
2328 io_req_free_batch_finish(ctx, &rb);
2330 if (!list_empty(&again))
2331 io_iopoll_queue(&again);
2334 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2337 struct io_kiocb *req, *tmp;
2343 * Only spin for completions if we don't have multiple devices hanging
2344 * off our complete list, and we're under the requested amount.
2346 spin = !ctx->poll_multi_file && *nr_events < min;
2349 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2350 struct kiocb *kiocb = &req->rw.kiocb;
2353 * Move completed and retryable entries to our local lists.
2354 * If we find a request that requires polling, break out
2355 * and complete those lists first, if we have entries there.
2357 if (READ_ONCE(req->iopoll_completed)) {
2358 list_move_tail(&req->inflight_entry, &done);
2361 if (!list_empty(&done))
2364 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2368 /* iopoll may have completed current req */
2369 if (READ_ONCE(req->iopoll_completed))
2370 list_move_tail(&req->inflight_entry, &done);
2377 if (!list_empty(&done))
2378 io_iopoll_complete(ctx, nr_events, &done);
2384 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2385 * non-spinning poll check - we'll still enter the driver poll loop, but only
2386 * as a non-spinning completion check.
2388 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2391 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2394 ret = io_do_iopoll(ctx, nr_events, min);
2397 if (*nr_events >= min)
2405 * We can't just wait for polled events to come to us, we have to actively
2406 * find and complete them.
2408 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2410 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2413 mutex_lock(&ctx->uring_lock);
2414 while (!list_empty(&ctx->iopoll_list)) {
2415 unsigned int nr_events = 0;
2417 io_do_iopoll(ctx, &nr_events, 0);
2419 /* let it sleep and repeat later if can't complete a request */
2423 * Ensure we allow local-to-the-cpu processing to take place,
2424 * in this case we need to ensure that we reap all events.
2425 * Also let task_work, etc. to progress by releasing the mutex
2427 if (need_resched()) {
2428 mutex_unlock(&ctx->uring_lock);
2430 mutex_lock(&ctx->uring_lock);
2433 mutex_unlock(&ctx->uring_lock);
2436 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2438 unsigned int nr_events = 0;
2439 int iters = 0, ret = 0;
2442 * We disallow the app entering submit/complete with polling, but we
2443 * still need to lock the ring to prevent racing with polled issue
2444 * that got punted to a workqueue.
2446 mutex_lock(&ctx->uring_lock);
2449 * Don't enter poll loop if we already have events pending.
2450 * If we do, we can potentially be spinning for commands that
2451 * already triggered a CQE (eg in error).
2453 if (io_cqring_events(ctx, false))
2457 * If a submit got punted to a workqueue, we can have the
2458 * application entering polling for a command before it gets
2459 * issued. That app will hold the uring_lock for the duration
2460 * of the poll right here, so we need to take a breather every
2461 * now and then to ensure that the issue has a chance to add
2462 * the poll to the issued list. Otherwise we can spin here
2463 * forever, while the workqueue is stuck trying to acquire the
2466 if (!(++iters & 7)) {
2467 mutex_unlock(&ctx->uring_lock);
2469 mutex_lock(&ctx->uring_lock);
2472 ret = io_iopoll_getevents(ctx, &nr_events, min);
2476 } while (min && !nr_events && !need_resched());
2478 mutex_unlock(&ctx->uring_lock);
2482 static void kiocb_end_write(struct io_kiocb *req)
2485 * Tell lockdep we inherited freeze protection from submission
2488 if (req->flags & REQ_F_ISREG) {
2489 struct inode *inode = file_inode(req->file);
2491 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2493 file_end_write(req->file);
2496 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2497 struct io_comp_state *cs)
2499 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 if (kiocb->ki_flags & IOCB_WRITE)
2503 kiocb_end_write(req);
2505 if (res != req->result)
2506 req_set_fail_links(req);
2507 if (req->flags & REQ_F_BUFFER_SELECTED)
2508 cflags = io_put_rw_kbuf(req);
2509 __io_req_complete(req, res, cflags, cs);
2513 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2515 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2516 ssize_t ret = -ECANCELED;
2517 struct iov_iter iter;
2525 switch (req->opcode) {
2526 case IORING_OP_READV:
2527 case IORING_OP_READ_FIXED:
2528 case IORING_OP_READ:
2531 case IORING_OP_WRITEV:
2532 case IORING_OP_WRITE_FIXED:
2533 case IORING_OP_WRITE:
2537 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2542 if (!req->async_data) {
2543 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2546 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2554 req_set_fail_links(req);
2555 io_req_complete(req, ret);
2560 static bool io_rw_reissue(struct io_kiocb *req, long res)
2563 umode_t mode = file_inode(req->file)->i_mode;
2566 if (!S_ISBLK(mode) && !S_ISREG(mode))
2568 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2571 ret = io_sq_thread_acquire_mm(req->ctx, req);
2573 if (io_resubmit_prep(req, ret)) {
2574 refcount_inc(&req->refs);
2575 io_queue_async_work(req);
2583 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2584 struct io_comp_state *cs)
2586 if (!io_rw_reissue(req, res))
2587 io_complete_rw_common(&req->rw.kiocb, res, cs);
2590 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2592 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2594 __io_complete_rw(req, res, res2, NULL);
2597 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2599 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2601 if (kiocb->ki_flags & IOCB_WRITE)
2602 kiocb_end_write(req);
2604 if (res != -EAGAIN && res != req->result)
2605 req_set_fail_links(req);
2607 WRITE_ONCE(req->result, res);
2608 /* order with io_poll_complete() checking ->result */
2610 WRITE_ONCE(req->iopoll_completed, 1);
2614 * After the iocb has been issued, it's safe to be found on the poll list.
2615 * Adding the kiocb to the list AFTER submission ensures that we don't
2616 * find it from a io_iopoll_getevents() thread before the issuer is done
2617 * accessing the kiocb cookie.
2619 static void io_iopoll_req_issued(struct io_kiocb *req)
2621 struct io_ring_ctx *ctx = req->ctx;
2624 * Track whether we have multiple files in our lists. This will impact
2625 * how we do polling eventually, not spinning if we're on potentially
2626 * different devices.
2628 if (list_empty(&ctx->iopoll_list)) {
2629 ctx->poll_multi_file = false;
2630 } else if (!ctx->poll_multi_file) {
2631 struct io_kiocb *list_req;
2633 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2635 if (list_req->file != req->file)
2636 ctx->poll_multi_file = true;
2640 * For fast devices, IO may have already completed. If it has, add
2641 * it to the front so we find it first.
2643 if (READ_ONCE(req->iopoll_completed))
2644 list_add(&req->inflight_entry, &ctx->iopoll_list);
2646 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2648 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2649 wq_has_sleeper(&ctx->sq_data->wait))
2650 wake_up(&ctx->sq_data->wait);
2653 static void __io_state_file_put(struct io_submit_state *state)
2655 if (state->has_refs)
2656 fput_many(state->file, state->has_refs);
2660 static inline void io_state_file_put(struct io_submit_state *state)
2663 __io_state_file_put(state);
2667 * Get as many references to a file as we have IOs left in this submission,
2668 * assuming most submissions are for one file, or at least that each file
2669 * has more than one submission.
2671 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2677 if (state->fd == fd) {
2681 __io_state_file_put(state);
2683 state->file = fget_many(fd, state->ios_left);
2688 state->has_refs = state->ios_left - 1;
2692 static bool io_bdev_nowait(struct block_device *bdev)
2695 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2702 * If we tracked the file through the SCM inflight mechanism, we could support
2703 * any file. For now, just ensure that anything potentially problematic is done
2706 static bool io_file_supports_async(struct file *file, int rw)
2708 umode_t mode = file_inode(file)->i_mode;
2710 if (S_ISBLK(mode)) {
2711 if (io_bdev_nowait(file->f_inode->i_bdev))
2715 if (S_ISCHR(mode) || S_ISSOCK(mode))
2717 if (S_ISREG(mode)) {
2718 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2719 file->f_op != &io_uring_fops)
2724 /* any ->read/write should understand O_NONBLOCK */
2725 if (file->f_flags & O_NONBLOCK)
2728 if (!(file->f_mode & FMODE_NOWAIT))
2732 return file->f_op->read_iter != NULL;
2734 return file->f_op->write_iter != NULL;
2737 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2739 struct io_ring_ctx *ctx = req->ctx;
2740 struct kiocb *kiocb = &req->rw.kiocb;
2744 if (S_ISREG(file_inode(req->file)->i_mode))
2745 req->flags |= REQ_F_ISREG;
2747 kiocb->ki_pos = READ_ONCE(sqe->off);
2748 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2749 req->flags |= REQ_F_CUR_POS;
2750 kiocb->ki_pos = req->file->f_pos;
2752 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2753 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2754 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2758 ioprio = READ_ONCE(sqe->ioprio);
2760 ret = ioprio_check_cap(ioprio);
2764 kiocb->ki_ioprio = ioprio;
2766 kiocb->ki_ioprio = get_current_ioprio();
2768 /* don't allow async punt if RWF_NOWAIT was requested */
2769 if (kiocb->ki_flags & IOCB_NOWAIT)
2770 req->flags |= REQ_F_NOWAIT;
2772 if (ctx->flags & IORING_SETUP_IOPOLL) {
2773 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2774 !kiocb->ki_filp->f_op->iopoll)
2777 kiocb->ki_flags |= IOCB_HIPRI;
2778 kiocb->ki_complete = io_complete_rw_iopoll;
2779 req->iopoll_completed = 0;
2781 if (kiocb->ki_flags & IOCB_HIPRI)
2783 kiocb->ki_complete = io_complete_rw;
2786 req->rw.addr = READ_ONCE(sqe->addr);
2787 req->rw.len = READ_ONCE(sqe->len);
2788 req->buf_index = READ_ONCE(sqe->buf_index);
2792 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2798 case -ERESTARTNOINTR:
2799 case -ERESTARTNOHAND:
2800 case -ERESTART_RESTARTBLOCK:
2802 * We can't just restart the syscall, since previously
2803 * submitted sqes may already be in progress. Just fail this
2809 kiocb->ki_complete(kiocb, ret, 0);
2813 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2814 struct io_comp_state *cs)
2816 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2817 struct io_async_rw *io = req->async_data;
2819 /* add previously done IO, if any */
2820 if (io && io->bytes_done > 0) {
2822 ret = io->bytes_done;
2824 ret += io->bytes_done;
2827 if (req->flags & REQ_F_CUR_POS)
2828 req->file->f_pos = kiocb->ki_pos;
2829 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2830 __io_complete_rw(req, ret, 0, cs);
2832 io_rw_done(kiocb, ret);
2835 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2836 struct iov_iter *iter)
2838 struct io_ring_ctx *ctx = req->ctx;
2839 size_t len = req->rw.len;
2840 struct io_mapped_ubuf *imu;
2841 u16 index, buf_index = req->buf_index;
2845 if (unlikely(buf_index >= ctx->nr_user_bufs))
2847 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2848 imu = &ctx->user_bufs[index];
2849 buf_addr = req->rw.addr;
2852 if (buf_addr + len < buf_addr)
2854 /* not inside the mapped region */
2855 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2859 * May not be a start of buffer, set size appropriately
2860 * and advance us to the beginning.
2862 offset = buf_addr - imu->ubuf;
2863 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2867 * Don't use iov_iter_advance() here, as it's really slow for
2868 * using the latter parts of a big fixed buffer - it iterates
2869 * over each segment manually. We can cheat a bit here, because
2872 * 1) it's a BVEC iter, we set it up
2873 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2874 * first and last bvec
2876 * So just find our index, and adjust the iterator afterwards.
2877 * If the offset is within the first bvec (or the whole first
2878 * bvec, just use iov_iter_advance(). This makes it easier
2879 * since we can just skip the first segment, which may not
2880 * be PAGE_SIZE aligned.
2882 const struct bio_vec *bvec = imu->bvec;
2884 if (offset <= bvec->bv_len) {
2885 iov_iter_advance(iter, offset);
2887 unsigned long seg_skip;
2889 /* skip first vec */
2890 offset -= bvec->bv_len;
2891 seg_skip = 1 + (offset >> PAGE_SHIFT);
2893 iter->bvec = bvec + seg_skip;
2894 iter->nr_segs -= seg_skip;
2895 iter->count -= bvec->bv_len + offset;
2896 iter->iov_offset = offset & ~PAGE_MASK;
2903 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2906 mutex_unlock(&ctx->uring_lock);
2909 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2912 * "Normal" inline submissions always hold the uring_lock, since we
2913 * grab it from the system call. Same is true for the SQPOLL offload.
2914 * The only exception is when we've detached the request and issue it
2915 * from an async worker thread, grab the lock for that case.
2918 mutex_lock(&ctx->uring_lock);
2921 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2922 int bgid, struct io_buffer *kbuf,
2925 struct io_buffer *head;
2927 if (req->flags & REQ_F_BUFFER_SELECTED)
2930 io_ring_submit_lock(req->ctx, needs_lock);
2932 lockdep_assert_held(&req->ctx->uring_lock);
2934 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2936 if (!list_empty(&head->list)) {
2937 kbuf = list_last_entry(&head->list, struct io_buffer,
2939 list_del(&kbuf->list);
2942 idr_remove(&req->ctx->io_buffer_idr, bgid);
2944 if (*len > kbuf->len)
2947 kbuf = ERR_PTR(-ENOBUFS);
2950 io_ring_submit_unlock(req->ctx, needs_lock);
2955 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2958 struct io_buffer *kbuf;
2961 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2962 bgid = req->buf_index;
2963 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2966 req->rw.addr = (u64) (unsigned long) kbuf;
2967 req->flags |= REQ_F_BUFFER_SELECTED;
2968 return u64_to_user_ptr(kbuf->addr);
2971 #ifdef CONFIG_COMPAT
2972 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2975 struct compat_iovec __user *uiov;
2976 compat_ssize_t clen;
2980 uiov = u64_to_user_ptr(req->rw.addr);
2981 if (!access_ok(uiov, sizeof(*uiov)))
2983 if (__get_user(clen, &uiov->iov_len))
2989 buf = io_rw_buffer_select(req, &len, needs_lock);
2991 return PTR_ERR(buf);
2992 iov[0].iov_base = buf;
2993 iov[0].iov_len = (compat_size_t) len;
2998 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3001 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3005 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3008 len = iov[0].iov_len;
3011 buf = io_rw_buffer_select(req, &len, needs_lock);
3013 return PTR_ERR(buf);
3014 iov[0].iov_base = buf;
3015 iov[0].iov_len = len;
3019 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3022 if (req->flags & REQ_F_BUFFER_SELECTED) {
3023 struct io_buffer *kbuf;
3025 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3026 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3027 iov[0].iov_len = kbuf->len;
3032 else if (req->rw.len > 1)
3035 #ifdef CONFIG_COMPAT
3036 if (req->ctx->compat)
3037 return io_compat_import(req, iov, needs_lock);
3040 return __io_iov_buffer_select(req, iov, needs_lock);
3043 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3044 struct iovec **iovec, struct iov_iter *iter,
3047 void __user *buf = u64_to_user_ptr(req->rw.addr);
3048 size_t sqe_len = req->rw.len;
3052 opcode = req->opcode;
3053 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3055 return io_import_fixed(req, rw, iter);
3058 /* buffer index only valid with fixed read/write, or buffer select */
3059 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3062 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3063 if (req->flags & REQ_F_BUFFER_SELECT) {
3064 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3066 return PTR_ERR(buf);
3067 req->rw.len = sqe_len;
3070 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3072 return ret < 0 ? ret : sqe_len;
3075 if (req->flags & REQ_F_BUFFER_SELECT) {
3076 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3078 ret = (*iovec)->iov_len;
3079 iov_iter_init(iter, rw, *iovec, 1, ret);
3085 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3089 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3090 struct iovec **iovec, struct iov_iter *iter,
3093 struct io_async_rw *iorw = req->async_data;
3096 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3098 return iov_iter_count(&iorw->iter);
3101 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3103 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3107 * For files that don't have ->read_iter() and ->write_iter(), handle them
3108 * by looping over ->read() or ->write() manually.
3110 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3111 struct iov_iter *iter)
3116 * Don't support polled IO through this interface, and we can't
3117 * support non-blocking either. For the latter, this just causes
3118 * the kiocb to be handled from an async context.
3120 if (kiocb->ki_flags & IOCB_HIPRI)
3122 if (kiocb->ki_flags & IOCB_NOWAIT)
3125 while (iov_iter_count(iter)) {
3129 if (!iov_iter_is_bvec(iter)) {
3130 iovec = iov_iter_iovec(iter);
3132 /* fixed buffers import bvec */
3133 iovec.iov_base = kmap(iter->bvec->bv_page)
3135 iovec.iov_len = min(iter->count,
3136 iter->bvec->bv_len - iter->iov_offset);
3140 nr = file->f_op->read(file, iovec.iov_base,
3141 iovec.iov_len, io_kiocb_ppos(kiocb));
3143 nr = file->f_op->write(file, iovec.iov_base,
3144 iovec.iov_len, io_kiocb_ppos(kiocb));
3147 if (iov_iter_is_bvec(iter))
3148 kunmap(iter->bvec->bv_page);
3156 if (nr != iovec.iov_len)
3158 iov_iter_advance(iter, nr);
3164 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3165 const struct iovec *fast_iov, struct iov_iter *iter)
3167 struct io_async_rw *rw = req->async_data;
3169 memcpy(&rw->iter, iter, sizeof(*iter));
3170 rw->free_iovec = iovec;
3172 /* can only be fixed buffers, no need to do anything */
3173 if (iter->type == ITER_BVEC)
3176 unsigned iov_off = 0;
3178 rw->iter.iov = rw->fast_iov;
3179 if (iter->iov != fast_iov) {
3180 iov_off = iter->iov - fast_iov;
3181 rw->iter.iov += iov_off;
3183 if (rw->fast_iov != fast_iov)
3184 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3185 sizeof(struct iovec) * iter->nr_segs);
3187 req->flags |= REQ_F_NEED_CLEANUP;
3191 static inline int __io_alloc_async_data(struct io_kiocb *req)
3193 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3194 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3195 return req->async_data == NULL;
3198 static int io_alloc_async_data(struct io_kiocb *req)
3200 if (!io_op_defs[req->opcode].needs_async_data)
3203 return __io_alloc_async_data(req);
3206 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3207 const struct iovec *fast_iov,
3208 struct iov_iter *iter, bool force)
3210 if (!force && !io_op_defs[req->opcode].needs_async_data)
3212 if (!req->async_data) {
3213 if (__io_alloc_async_data(req))
3216 io_req_map_rw(req, iovec, fast_iov, iter);
3221 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3223 struct io_async_rw *iorw = req->async_data;
3224 struct iovec *iov = iorw->fast_iov;
3227 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3228 if (unlikely(ret < 0))
3231 iorw->bytes_done = 0;
3232 iorw->free_iovec = iov;
3234 req->flags |= REQ_F_NEED_CLEANUP;
3238 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3242 ret = io_prep_rw(req, sqe);
3246 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3249 /* either don't need iovec imported or already have it */
3250 if (!req->async_data)
3252 return io_rw_prep_async(req, READ);
3256 * This is our waitqueue callback handler, registered through lock_page_async()
3257 * when we initially tried to do the IO with the iocb armed our waitqueue.
3258 * This gets called when the page is unlocked, and we generally expect that to
3259 * happen when the page IO is completed and the page is now uptodate. This will
3260 * queue a task_work based retry of the operation, attempting to copy the data
3261 * again. If the latter fails because the page was NOT uptodate, then we will
3262 * do a thread based blocking retry of the operation. That's the unexpected
3265 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3266 int sync, void *arg)
3268 struct wait_page_queue *wpq;
3269 struct io_kiocb *req = wait->private;
3270 struct wait_page_key *key = arg;
3273 wpq = container_of(wait, struct wait_page_queue, wait);
3275 if (!wake_page_match(wpq, key))
3278 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3279 list_del_init(&wait->entry);
3281 init_task_work(&req->task_work, io_req_task_submit);
3282 percpu_ref_get(&req->ctx->refs);
3284 /* submit ref gets dropped, acquire a new one */
3285 refcount_inc(&req->refs);
3286 ret = io_req_task_work_add(req, true);
3287 if (unlikely(ret)) {
3288 struct task_struct *tsk;
3290 /* queue just for cancelation */
3291 init_task_work(&req->task_work, io_req_task_cancel);
3292 tsk = io_wq_get_task(req->ctx->io_wq);
3293 task_work_add(tsk, &req->task_work, 0);
3294 wake_up_process(tsk);
3300 * This controls whether a given IO request should be armed for async page
3301 * based retry. If we return false here, the request is handed to the async
3302 * worker threads for retry. If we're doing buffered reads on a regular file,
3303 * we prepare a private wait_page_queue entry and retry the operation. This
3304 * will either succeed because the page is now uptodate and unlocked, or it
3305 * will register a callback when the page is unlocked at IO completion. Through
3306 * that callback, io_uring uses task_work to setup a retry of the operation.
3307 * That retry will attempt the buffered read again. The retry will generally
3308 * succeed, or in rare cases where it fails, we then fall back to using the
3309 * async worker threads for a blocking retry.
3311 static bool io_rw_should_retry(struct io_kiocb *req)
3313 struct io_async_rw *rw = req->async_data;
3314 struct wait_page_queue *wait = &rw->wpq;
3315 struct kiocb *kiocb = &req->rw.kiocb;
3317 /* never retry for NOWAIT, we just complete with -EAGAIN */
3318 if (req->flags & REQ_F_NOWAIT)
3321 /* Only for buffered IO */
3322 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3326 * just use poll if we can, and don't attempt if the fs doesn't
3327 * support callback based unlocks
3329 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3332 wait->wait.func = io_async_buf_func;
3333 wait->wait.private = req;
3334 wait->wait.flags = 0;
3335 INIT_LIST_HEAD(&wait->wait.entry);
3336 kiocb->ki_flags |= IOCB_WAITQ;
3337 kiocb->ki_flags &= ~IOCB_NOWAIT;
3338 kiocb->ki_waitq = wait;
3342 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3344 if (req->file->f_op->read_iter)
3345 return call_read_iter(req->file, &req->rw.kiocb, iter);
3346 else if (req->file->f_op->read)
3347 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3352 static int io_read(struct io_kiocb *req, bool force_nonblock,
3353 struct io_comp_state *cs)
3355 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3356 struct kiocb *kiocb = &req->rw.kiocb;
3357 struct iov_iter __iter, *iter = &__iter;
3358 struct io_async_rw *rw = req->async_data;
3359 ssize_t io_size, ret, ret2;
3366 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3369 iov_count = iov_iter_count(iter);
3371 req->result = io_size;
3374 /* Ensure we clear previously set non-block flag */
3375 if (!force_nonblock)
3376 kiocb->ki_flags &= ~IOCB_NOWAIT;
3378 kiocb->ki_flags |= IOCB_NOWAIT;
3381 /* If the file doesn't support async, just async punt */
3382 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3386 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3390 ret = io_iter_do_read(req, iter);
3394 } else if (ret == -EIOCBQUEUED) {
3397 } else if (ret == -EAGAIN) {
3398 /* IOPOLL retry should happen for io-wq threads */
3399 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3401 /* no retry on NONBLOCK marked file */
3402 if (req->file->f_flags & O_NONBLOCK)
3404 /* some cases will consume bytes even on error returns */
3405 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3408 } else if (ret < 0) {
3409 /* make sure -ERESTARTSYS -> -EINTR is done */
3413 /* read it all, or we did blocking attempt. no retry. */
3414 if (!iov_iter_count(iter) || !force_nonblock ||
3415 (req->file->f_flags & O_NONBLOCK))
3420 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3427 rw = req->async_data;
3428 /* it's copied and will be cleaned with ->io */
3430 /* now use our persistent iterator, if we aren't already */
3433 rw->bytes_done += ret;
3434 /* if we can retry, do so with the callbacks armed */
3435 if (!io_rw_should_retry(req)) {
3436 kiocb->ki_flags &= ~IOCB_WAITQ;
3441 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3442 * get -EIOCBQUEUED, then we'll get a notification when the desired
3443 * page gets unlocked. We can also get a partial read here, and if we
3444 * do, then just retry at the new offset.
3446 ret = io_iter_do_read(req, iter);
3447 if (ret == -EIOCBQUEUED) {
3450 } else if (ret > 0 && ret < io_size) {
3451 /* we got some bytes, but not all. retry. */
3455 kiocb_done(kiocb, ret, cs);
3458 /* it's reportedly faster than delegating the null check to kfree() */
3464 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3468 ret = io_prep_rw(req, sqe);
3472 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3475 /* either don't need iovec imported or already have it */
3476 if (!req->async_data)
3478 return io_rw_prep_async(req, WRITE);
3481 static int io_write(struct io_kiocb *req, bool force_nonblock,
3482 struct io_comp_state *cs)
3484 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3485 struct kiocb *kiocb = &req->rw.kiocb;
3486 struct iov_iter __iter, *iter = &__iter;
3487 struct io_async_rw *rw = req->async_data;
3489 ssize_t ret, ret2, io_size;
3494 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3497 iov_count = iov_iter_count(iter);
3499 req->result = io_size;
3501 /* Ensure we clear previously set non-block flag */
3502 if (!force_nonblock)
3503 kiocb->ki_flags &= ~IOCB_NOWAIT;
3505 kiocb->ki_flags |= IOCB_NOWAIT;
3507 /* If the file doesn't support async, just async punt */
3508 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3511 /* file path doesn't support NOWAIT for non-direct_IO */
3512 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3513 (req->flags & REQ_F_ISREG))
3516 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3521 * Open-code file_start_write here to grab freeze protection,
3522 * which will be released by another thread in
3523 * io_complete_rw(). Fool lockdep by telling it the lock got
3524 * released so that it doesn't complain about the held lock when
3525 * we return to userspace.
3527 if (req->flags & REQ_F_ISREG) {
3528 __sb_start_write(file_inode(req->file)->i_sb,
3529 SB_FREEZE_WRITE, true);
3530 __sb_writers_release(file_inode(req->file)->i_sb,
3533 kiocb->ki_flags |= IOCB_WRITE;
3535 if (req->file->f_op->write_iter)
3536 ret2 = call_write_iter(req->file, kiocb, iter);
3537 else if (req->file->f_op->write)
3538 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3543 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3544 * retry them without IOCB_NOWAIT.
3546 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3548 /* no retry on NONBLOCK marked file */
3549 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3551 if (!force_nonblock || ret2 != -EAGAIN) {
3552 /* IOPOLL retry should happen for io-wq threads */
3553 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3556 kiocb_done(kiocb, ret2, cs);
3559 /* some cases will consume bytes even on error returns */
3560 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3561 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3566 /* it's reportedly faster than delegating the null check to kfree() */
3572 static int __io_splice_prep(struct io_kiocb *req,
3573 const struct io_uring_sqe *sqe)
3575 struct io_splice* sp = &req->splice;
3576 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3582 sp->len = READ_ONCE(sqe->len);
3583 sp->flags = READ_ONCE(sqe->splice_flags);
3585 if (unlikely(sp->flags & ~valid_flags))
3588 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3589 (sp->flags & SPLICE_F_FD_IN_FIXED));
3592 req->flags |= REQ_F_NEED_CLEANUP;
3594 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3596 * Splice operation will be punted aync, and here need to
3597 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3599 io_req_init_async(req);
3600 req->work.flags |= IO_WQ_WORK_UNBOUND;
3606 static int io_tee_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3611 return __io_splice_prep(req, sqe);
3614 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3616 struct io_splice *sp = &req->splice;
3617 struct file *in = sp->file_in;
3618 struct file *out = sp->file_out;
3619 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3625 ret = do_tee(in, out, sp->len, flags);
3627 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3628 req->flags &= ~REQ_F_NEED_CLEANUP;
3631 req_set_fail_links(req);
3632 io_req_complete(req, ret);
3636 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3638 struct io_splice* sp = &req->splice;
3640 sp->off_in = READ_ONCE(sqe->splice_off_in);
3641 sp->off_out = READ_ONCE(sqe->off);
3642 return __io_splice_prep(req, sqe);
3645 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3647 struct io_splice *sp = &req->splice;
3648 struct file *in = sp->file_in;
3649 struct file *out = sp->file_out;
3650 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3651 loff_t *poff_in, *poff_out;
3657 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3658 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3661 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3663 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3664 req->flags &= ~REQ_F_NEED_CLEANUP;
3667 req_set_fail_links(req);
3668 io_req_complete(req, ret);
3673 * IORING_OP_NOP just posts a completion event, nothing else.
3675 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3677 struct io_ring_ctx *ctx = req->ctx;
3679 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3682 __io_req_complete(req, 0, 0, cs);
3686 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3688 struct io_ring_ctx *ctx = req->ctx;
3693 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3695 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3698 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3699 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3702 req->sync.off = READ_ONCE(sqe->off);
3703 req->sync.len = READ_ONCE(sqe->len);
3707 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3709 loff_t end = req->sync.off + req->sync.len;
3712 /* fsync always requires a blocking context */
3716 ret = vfs_fsync_range(req->file, req->sync.off,
3717 end > 0 ? end : LLONG_MAX,
3718 req->sync.flags & IORING_FSYNC_DATASYNC);
3720 req_set_fail_links(req);
3721 io_req_complete(req, ret);
3725 static int io_fallocate_prep(struct io_kiocb *req,
3726 const struct io_uring_sqe *sqe)
3728 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3730 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3733 req->sync.off = READ_ONCE(sqe->off);
3734 req->sync.len = READ_ONCE(sqe->addr);
3735 req->sync.mode = READ_ONCE(sqe->len);
3739 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3743 /* fallocate always requiring blocking context */
3746 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3749 req_set_fail_links(req);
3750 io_req_complete(req, ret);
3754 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3756 const char __user *fname;
3759 if (unlikely(sqe->ioprio || sqe->buf_index))
3761 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3764 /* open.how should be already initialised */
3765 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3766 req->open.how.flags |= O_LARGEFILE;
3768 req->open.dfd = READ_ONCE(sqe->fd);
3769 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3770 req->open.filename = getname(fname);
3771 if (IS_ERR(req->open.filename)) {
3772 ret = PTR_ERR(req->open.filename);
3773 req->open.filename = NULL;
3776 req->open.nofile = rlimit(RLIMIT_NOFILE);
3777 req->flags |= REQ_F_NEED_CLEANUP;
3781 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3785 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3787 mode = READ_ONCE(sqe->len);
3788 flags = READ_ONCE(sqe->open_flags);
3789 req->open.how = build_open_how(flags, mode);
3790 return __io_openat_prep(req, sqe);
3793 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3795 struct open_how __user *how;
3799 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3801 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3802 len = READ_ONCE(sqe->len);
3803 if (len < OPEN_HOW_SIZE_VER0)
3806 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3811 return __io_openat_prep(req, sqe);
3814 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3816 struct open_flags op;
3823 ret = build_open_flags(&req->open.how, &op);
3827 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3831 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3834 ret = PTR_ERR(file);
3836 fsnotify_open(file);
3837 fd_install(ret, file);
3840 putname(req->open.filename);
3841 req->flags &= ~REQ_F_NEED_CLEANUP;
3843 req_set_fail_links(req);
3844 io_req_complete(req, ret);
3848 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3850 return io_openat2(req, force_nonblock);
3853 static int io_remove_buffers_prep(struct io_kiocb *req,
3854 const struct io_uring_sqe *sqe)
3856 struct io_provide_buf *p = &req->pbuf;
3859 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3862 tmp = READ_ONCE(sqe->fd);
3863 if (!tmp || tmp > USHRT_MAX)
3866 memset(p, 0, sizeof(*p));
3868 p->bgid = READ_ONCE(sqe->buf_group);
3872 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3873 int bgid, unsigned nbufs)
3877 /* shouldn't happen */
3881 /* the head kbuf is the list itself */
3882 while (!list_empty(&buf->list)) {
3883 struct io_buffer *nxt;
3885 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3886 list_del(&nxt->list);
3893 idr_remove(&ctx->io_buffer_idr, bgid);
3898 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3899 struct io_comp_state *cs)
3901 struct io_provide_buf *p = &req->pbuf;
3902 struct io_ring_ctx *ctx = req->ctx;
3903 struct io_buffer *head;
3906 io_ring_submit_lock(ctx, !force_nonblock);
3908 lockdep_assert_held(&ctx->uring_lock);
3911 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3913 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3915 io_ring_submit_lock(ctx, !force_nonblock);
3917 req_set_fail_links(req);
3918 __io_req_complete(req, ret, 0, cs);
3922 static int io_provide_buffers_prep(struct io_kiocb *req,
3923 const struct io_uring_sqe *sqe)
3925 struct io_provide_buf *p = &req->pbuf;
3928 if (sqe->ioprio || sqe->rw_flags)
3931 tmp = READ_ONCE(sqe->fd);
3932 if (!tmp || tmp > USHRT_MAX)
3935 p->addr = READ_ONCE(sqe->addr);
3936 p->len = READ_ONCE(sqe->len);
3938 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3941 p->bgid = READ_ONCE(sqe->buf_group);
3942 tmp = READ_ONCE(sqe->off);
3943 if (tmp > USHRT_MAX)
3949 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3951 struct io_buffer *buf;
3952 u64 addr = pbuf->addr;
3953 int i, bid = pbuf->bid;
3955 for (i = 0; i < pbuf->nbufs; i++) {
3956 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3961 buf->len = pbuf->len;
3966 INIT_LIST_HEAD(&buf->list);
3969 list_add_tail(&buf->list, &(*head)->list);
3973 return i ? i : -ENOMEM;
3976 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3977 struct io_comp_state *cs)
3979 struct io_provide_buf *p = &req->pbuf;
3980 struct io_ring_ctx *ctx = req->ctx;
3981 struct io_buffer *head, *list;
3984 io_ring_submit_lock(ctx, !force_nonblock);
3986 lockdep_assert_held(&ctx->uring_lock);
3988 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3990 ret = io_add_buffers(p, &head);
3995 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3998 __io_remove_buffers(ctx, head, p->bgid, -1U);
4003 io_ring_submit_unlock(ctx, !force_nonblock);
4005 req_set_fail_links(req);
4006 __io_req_complete(req, ret, 0, cs);
4010 static int io_epoll_ctl_prep(struct io_kiocb *req,
4011 const struct io_uring_sqe *sqe)
4013 #if defined(CONFIG_EPOLL)
4014 if (sqe->ioprio || sqe->buf_index)
4016 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4019 req->epoll.epfd = READ_ONCE(sqe->fd);
4020 req->epoll.op = READ_ONCE(sqe->len);
4021 req->epoll.fd = READ_ONCE(sqe->off);
4023 if (ep_op_has_event(req->epoll.op)) {
4024 struct epoll_event __user *ev;
4026 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4027 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4037 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4038 struct io_comp_state *cs)
4040 #if defined(CONFIG_EPOLL)
4041 struct io_epoll *ie = &req->epoll;
4044 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4045 if (force_nonblock && ret == -EAGAIN)
4049 req_set_fail_links(req);
4050 __io_req_complete(req, ret, 0, cs);
4057 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4059 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4060 if (sqe->ioprio || sqe->buf_index || sqe->off)
4062 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4065 req->madvise.addr = READ_ONCE(sqe->addr);
4066 req->madvise.len = READ_ONCE(sqe->len);
4067 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4074 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4076 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4077 struct io_madvise *ma = &req->madvise;
4083 ret = do_madvise(ma->addr, ma->len, ma->advice);
4085 req_set_fail_links(req);
4086 io_req_complete(req, ret);
4093 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4095 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4097 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4100 req->fadvise.offset = READ_ONCE(sqe->off);
4101 req->fadvise.len = READ_ONCE(sqe->len);
4102 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4106 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4108 struct io_fadvise *fa = &req->fadvise;
4111 if (force_nonblock) {
4112 switch (fa->advice) {
4113 case POSIX_FADV_NORMAL:
4114 case POSIX_FADV_RANDOM:
4115 case POSIX_FADV_SEQUENTIAL:
4122 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4124 req_set_fail_links(req);
4125 io_req_complete(req, ret);
4129 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4133 if (sqe->ioprio || sqe->buf_index)
4135 if (req->flags & REQ_F_FIXED_FILE)
4138 req->statx.dfd = READ_ONCE(sqe->fd);
4139 req->statx.mask = READ_ONCE(sqe->len);
4140 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4141 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4142 req->statx.flags = READ_ONCE(sqe->statx_flags);
4147 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4149 struct io_statx *ctx = &req->statx;
4152 if (force_nonblock) {
4153 /* only need file table for an actual valid fd */
4154 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4155 req->flags |= REQ_F_NO_FILE_TABLE;
4159 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4163 req_set_fail_links(req);
4164 io_req_complete(req, ret);
4168 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4171 * If we queue this for async, it must not be cancellable. That would
4172 * leave the 'file' in an undeterminate state, and here need to modify
4173 * io_wq_work.flags, so initialize io_wq_work firstly.
4175 io_req_init_async(req);
4176 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4178 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4180 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4181 sqe->rw_flags || sqe->buf_index)
4183 if (req->flags & REQ_F_FIXED_FILE)
4186 req->close.fd = READ_ONCE(sqe->fd);
4187 if ((req->file && req->file->f_op == &io_uring_fops))
4190 req->close.put_file = NULL;
4194 static int io_close(struct io_kiocb *req, bool force_nonblock,
4195 struct io_comp_state *cs)
4197 struct io_close *close = &req->close;
4200 /* might be already done during nonblock submission */
4201 if (!close->put_file) {
4202 ret = __close_fd_get_file(close->fd, &close->put_file);
4204 return (ret == -ENOENT) ? -EBADF : ret;
4207 /* if the file has a flush method, be safe and punt to async */
4208 if (close->put_file->f_op->flush && force_nonblock) {
4209 /* was never set, but play safe */
4210 req->flags &= ~REQ_F_NOWAIT;
4211 /* avoid grabbing files - we don't need the files */
4212 req->flags |= REQ_F_NO_FILE_TABLE;
4216 /* No ->flush() or already async, safely close from here */
4217 ret = filp_close(close->put_file, req->work.identity->files);
4219 req_set_fail_links(req);
4220 fput(close->put_file);
4221 close->put_file = NULL;
4222 __io_req_complete(req, ret, 0, cs);
4226 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4228 struct io_ring_ctx *ctx = req->ctx;
4233 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4235 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4238 req->sync.off = READ_ONCE(sqe->off);
4239 req->sync.len = READ_ONCE(sqe->len);
4240 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4244 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4248 /* sync_file_range always requires a blocking context */
4252 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4255 req_set_fail_links(req);
4256 io_req_complete(req, ret);
4260 #if defined(CONFIG_NET)
4261 static int io_setup_async_msg(struct io_kiocb *req,
4262 struct io_async_msghdr *kmsg)
4264 struct io_async_msghdr *async_msg = req->async_data;
4268 if (io_alloc_async_data(req)) {
4269 if (kmsg->iov != kmsg->fast_iov)
4273 async_msg = req->async_data;
4274 req->flags |= REQ_F_NEED_CLEANUP;
4275 memcpy(async_msg, kmsg, sizeof(*kmsg));
4279 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4280 struct io_async_msghdr *iomsg)
4282 iomsg->iov = iomsg->fast_iov;
4283 iomsg->msg.msg_name = &iomsg->addr;
4284 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4285 req->sr_msg.msg_flags, &iomsg->iov);
4288 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4290 struct io_async_msghdr *async_msg = req->async_data;
4291 struct io_sr_msg *sr = &req->sr_msg;
4294 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4297 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4298 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4299 sr->len = READ_ONCE(sqe->len);
4301 #ifdef CONFIG_COMPAT
4302 if (req->ctx->compat)
4303 sr->msg_flags |= MSG_CMSG_COMPAT;
4306 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4308 ret = io_sendmsg_copy_hdr(req, async_msg);
4310 req->flags |= REQ_F_NEED_CLEANUP;
4314 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4315 struct io_comp_state *cs)
4317 struct io_async_msghdr iomsg, *kmsg;
4318 struct socket *sock;
4322 sock = sock_from_file(req->file, &ret);
4323 if (unlikely(!sock))
4326 if (req->async_data) {
4327 kmsg = req->async_data;
4328 kmsg->msg.msg_name = &kmsg->addr;
4329 /* if iov is set, it's allocated already */
4331 kmsg->iov = kmsg->fast_iov;
4332 kmsg->msg.msg_iter.iov = kmsg->iov;
4334 ret = io_sendmsg_copy_hdr(req, &iomsg);
4340 flags = req->sr_msg.msg_flags;
4341 if (flags & MSG_DONTWAIT)
4342 req->flags |= REQ_F_NOWAIT;
4343 else if (force_nonblock)
4344 flags |= MSG_DONTWAIT;
4346 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4347 if (force_nonblock && ret == -EAGAIN)
4348 return io_setup_async_msg(req, kmsg);
4349 if (ret == -ERESTARTSYS)
4352 if (kmsg->iov != kmsg->fast_iov)
4354 req->flags &= ~REQ_F_NEED_CLEANUP;
4356 req_set_fail_links(req);
4357 __io_req_complete(req, ret, 0, cs);
4361 static int io_send(struct io_kiocb *req, bool force_nonblock,
4362 struct io_comp_state *cs)
4364 struct io_sr_msg *sr = &req->sr_msg;
4367 struct socket *sock;
4371 sock = sock_from_file(req->file, &ret);
4372 if (unlikely(!sock))
4375 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4379 msg.msg_name = NULL;
4380 msg.msg_control = NULL;
4381 msg.msg_controllen = 0;
4382 msg.msg_namelen = 0;
4384 flags = req->sr_msg.msg_flags;
4385 if (flags & MSG_DONTWAIT)
4386 req->flags |= REQ_F_NOWAIT;
4387 else if (force_nonblock)
4388 flags |= MSG_DONTWAIT;
4390 msg.msg_flags = flags;
4391 ret = sock_sendmsg(sock, &msg);
4392 if (force_nonblock && ret == -EAGAIN)
4394 if (ret == -ERESTARTSYS)
4398 req_set_fail_links(req);
4399 __io_req_complete(req, ret, 0, cs);
4403 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4404 struct io_async_msghdr *iomsg)
4406 struct io_sr_msg *sr = &req->sr_msg;
4407 struct iovec __user *uiov;
4411 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4412 &iomsg->uaddr, &uiov, &iov_len);
4416 if (req->flags & REQ_F_BUFFER_SELECT) {
4419 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4421 sr->len = iomsg->iov[0].iov_len;
4422 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4426 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4427 &iomsg->iov, &iomsg->msg.msg_iter,
4436 #ifdef CONFIG_COMPAT
4437 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4438 struct io_async_msghdr *iomsg)
4440 struct compat_msghdr __user *msg_compat;
4441 struct io_sr_msg *sr = &req->sr_msg;
4442 struct compat_iovec __user *uiov;
4447 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4448 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4453 uiov = compat_ptr(ptr);
4454 if (req->flags & REQ_F_BUFFER_SELECT) {
4455 compat_ssize_t clen;
4459 if (!access_ok(uiov, sizeof(*uiov)))
4461 if (__get_user(clen, &uiov->iov_len))
4465 sr->len = iomsg->iov[0].iov_len;
4468 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4469 UIO_FASTIOV, &iomsg->iov,
4470 &iomsg->msg.msg_iter, true);
4479 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4480 struct io_async_msghdr *iomsg)
4482 iomsg->msg.msg_name = &iomsg->addr;
4483 iomsg->iov = iomsg->fast_iov;
4485 #ifdef CONFIG_COMPAT
4486 if (req->ctx->compat)
4487 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4490 return __io_recvmsg_copy_hdr(req, iomsg);
4493 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4496 struct io_sr_msg *sr = &req->sr_msg;
4497 struct io_buffer *kbuf;
4499 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4504 req->flags |= REQ_F_BUFFER_SELECTED;
4508 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4510 return io_put_kbuf(req, req->sr_msg.kbuf);
4513 static int io_recvmsg_prep(struct io_kiocb *req,
4514 const struct io_uring_sqe *sqe)
4516 struct io_async_msghdr *async_msg = req->async_data;
4517 struct io_sr_msg *sr = &req->sr_msg;
4520 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4523 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4524 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4525 sr->len = READ_ONCE(sqe->len);
4526 sr->bgid = READ_ONCE(sqe->buf_group);
4528 #ifdef CONFIG_COMPAT
4529 if (req->ctx->compat)
4530 sr->msg_flags |= MSG_CMSG_COMPAT;
4533 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4535 ret = io_recvmsg_copy_hdr(req, async_msg);
4537 req->flags |= REQ_F_NEED_CLEANUP;
4541 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4542 struct io_comp_state *cs)
4544 struct io_async_msghdr iomsg, *kmsg;
4545 struct socket *sock;
4546 struct io_buffer *kbuf;
4548 int ret, cflags = 0;
4550 sock = sock_from_file(req->file, &ret);
4551 if (unlikely(!sock))
4554 if (req->async_data) {
4555 kmsg = req->async_data;
4556 kmsg->msg.msg_name = &kmsg->addr;
4557 /* if iov is set, it's allocated already */
4559 kmsg->iov = kmsg->fast_iov;
4560 kmsg->msg.msg_iter.iov = kmsg->iov;
4562 ret = io_recvmsg_copy_hdr(req, &iomsg);
4568 if (req->flags & REQ_F_BUFFER_SELECT) {
4569 kbuf = io_recv_buffer_select(req, !force_nonblock);
4571 return PTR_ERR(kbuf);
4572 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4573 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4574 1, req->sr_msg.len);
4577 flags = req->sr_msg.msg_flags;
4578 if (flags & MSG_DONTWAIT)
4579 req->flags |= REQ_F_NOWAIT;
4580 else if (force_nonblock)
4581 flags |= MSG_DONTWAIT;
4583 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4584 kmsg->uaddr, flags);
4585 if (force_nonblock && ret == -EAGAIN)
4586 return io_setup_async_msg(req, kmsg);
4587 if (ret == -ERESTARTSYS)
4590 if (req->flags & REQ_F_BUFFER_SELECTED)
4591 cflags = io_put_recv_kbuf(req);
4592 if (kmsg->iov != kmsg->fast_iov)
4594 req->flags &= ~REQ_F_NEED_CLEANUP;
4596 req_set_fail_links(req);
4597 __io_req_complete(req, ret, cflags, cs);
4601 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4602 struct io_comp_state *cs)
4604 struct io_buffer *kbuf;
4605 struct io_sr_msg *sr = &req->sr_msg;
4607 void __user *buf = sr->buf;
4608 struct socket *sock;
4611 int ret, cflags = 0;
4613 sock = sock_from_file(req->file, &ret);
4614 if (unlikely(!sock))
4617 if (req->flags & REQ_F_BUFFER_SELECT) {
4618 kbuf = io_recv_buffer_select(req, !force_nonblock);
4620 return PTR_ERR(kbuf);
4621 buf = u64_to_user_ptr(kbuf->addr);
4624 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4628 msg.msg_name = NULL;
4629 msg.msg_control = NULL;
4630 msg.msg_controllen = 0;
4631 msg.msg_namelen = 0;
4632 msg.msg_iocb = NULL;
4635 flags = req->sr_msg.msg_flags;
4636 if (flags & MSG_DONTWAIT)
4637 req->flags |= REQ_F_NOWAIT;
4638 else if (force_nonblock)
4639 flags |= MSG_DONTWAIT;
4641 ret = sock_recvmsg(sock, &msg, flags);
4642 if (force_nonblock && ret == -EAGAIN)
4644 if (ret == -ERESTARTSYS)
4647 if (req->flags & REQ_F_BUFFER_SELECTED)
4648 cflags = io_put_recv_kbuf(req);
4650 req_set_fail_links(req);
4651 __io_req_complete(req, ret, cflags, cs);
4655 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4657 struct io_accept *accept = &req->accept;
4659 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4661 if (sqe->ioprio || sqe->len || sqe->buf_index)
4664 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4665 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4666 accept->flags = READ_ONCE(sqe->accept_flags);
4667 accept->nofile = rlimit(RLIMIT_NOFILE);
4671 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4672 struct io_comp_state *cs)
4674 struct io_accept *accept = &req->accept;
4675 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4678 if (req->file->f_flags & O_NONBLOCK)
4679 req->flags |= REQ_F_NOWAIT;
4681 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4682 accept->addr_len, accept->flags,
4684 if (ret == -EAGAIN && force_nonblock)
4687 if (ret == -ERESTARTSYS)
4689 req_set_fail_links(req);
4691 __io_req_complete(req, ret, 0, cs);
4695 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4697 struct io_connect *conn = &req->connect;
4698 struct io_async_connect *io = req->async_data;
4700 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4702 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4705 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4706 conn->addr_len = READ_ONCE(sqe->addr2);
4711 return move_addr_to_kernel(conn->addr, conn->addr_len,
4715 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4716 struct io_comp_state *cs)
4718 struct io_async_connect __io, *io;
4719 unsigned file_flags;
4722 if (req->async_data) {
4723 io = req->async_data;
4725 ret = move_addr_to_kernel(req->connect.addr,
4726 req->connect.addr_len,
4733 file_flags = force_nonblock ? O_NONBLOCK : 0;
4735 ret = __sys_connect_file(req->file, &io->address,
4736 req->connect.addr_len, file_flags);
4737 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4738 if (req->async_data)
4740 if (io_alloc_async_data(req)) {
4744 io = req->async_data;
4745 memcpy(req->async_data, &__io, sizeof(__io));
4748 if (ret == -ERESTARTSYS)
4752 req_set_fail_links(req);
4753 __io_req_complete(req, ret, 0, cs);
4756 #else /* !CONFIG_NET */
4757 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4762 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4763 struct io_comp_state *cs)
4768 static int io_send(struct io_kiocb *req, bool force_nonblock,
4769 struct io_comp_state *cs)
4774 static int io_recvmsg_prep(struct io_kiocb *req,
4775 const struct io_uring_sqe *sqe)
4780 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4781 struct io_comp_state *cs)
4786 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4787 struct io_comp_state *cs)
4792 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4797 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4798 struct io_comp_state *cs)
4803 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4808 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4809 struct io_comp_state *cs)
4813 #endif /* CONFIG_NET */
4815 struct io_poll_table {
4816 struct poll_table_struct pt;
4817 struct io_kiocb *req;
4821 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4822 __poll_t mask, task_work_func_t func)
4827 /* for instances that support it check for an event match first: */
4828 if (mask && !(mask & poll->events))
4831 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4833 list_del_init(&poll->wait.entry);
4836 init_task_work(&req->task_work, func);
4837 percpu_ref_get(&req->ctx->refs);
4840 * If we using the signalfd wait_queue_head for this wakeup, then
4841 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4842 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4843 * either, as the normal wakeup will suffice.
4845 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4848 * If this fails, then the task is exiting. When a task exits, the
4849 * work gets canceled, so just cancel this request as well instead
4850 * of executing it. We can't safely execute it anyway, as we may not
4851 * have the needed state needed for it anyway.
4853 ret = io_req_task_work_add(req, twa_signal_ok);
4854 if (unlikely(ret)) {
4855 struct task_struct *tsk;
4857 WRITE_ONCE(poll->canceled, true);
4858 tsk = io_wq_get_task(req->ctx->io_wq);
4859 task_work_add(tsk, &req->task_work, 0);
4860 wake_up_process(tsk);
4865 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4866 __acquires(&req->ctx->completion_lock)
4868 struct io_ring_ctx *ctx = req->ctx;
4870 if (!req->result && !READ_ONCE(poll->canceled)) {
4871 struct poll_table_struct pt = { ._key = poll->events };
4873 req->result = vfs_poll(req->file, &pt) & poll->events;
4876 spin_lock_irq(&ctx->completion_lock);
4877 if (!req->result && !READ_ONCE(poll->canceled)) {
4878 add_wait_queue(poll->head, &poll->wait);
4885 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4887 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4888 if (req->opcode == IORING_OP_POLL_ADD)
4889 return req->async_data;
4890 return req->apoll->double_poll;
4893 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4895 if (req->opcode == IORING_OP_POLL_ADD)
4897 return &req->apoll->poll;
4900 static void io_poll_remove_double(struct io_kiocb *req)
4902 struct io_poll_iocb *poll = io_poll_get_double(req);
4904 lockdep_assert_held(&req->ctx->completion_lock);
4906 if (poll && poll->head) {
4907 struct wait_queue_head *head = poll->head;
4909 spin_lock(&head->lock);
4910 list_del_init(&poll->wait.entry);
4911 if (poll->wait.private)
4912 refcount_dec(&req->refs);
4914 spin_unlock(&head->lock);
4918 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4920 struct io_ring_ctx *ctx = req->ctx;
4922 io_poll_remove_double(req);
4923 req->poll.done = true;
4924 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4925 io_commit_cqring(ctx);
4928 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4930 struct io_ring_ctx *ctx = req->ctx;
4932 if (io_poll_rewait(req, &req->poll)) {
4933 spin_unlock_irq(&ctx->completion_lock);
4937 hash_del(&req->hash_node);
4938 io_poll_complete(req, req->result, 0);
4939 spin_unlock_irq(&ctx->completion_lock);
4941 *nxt = io_put_req_find_next(req);
4942 io_cqring_ev_posted(ctx);
4945 static void io_poll_task_func(struct callback_head *cb)
4947 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4948 struct io_ring_ctx *ctx = req->ctx;
4949 struct io_kiocb *nxt = NULL;
4951 io_poll_task_handler(req, &nxt);
4953 __io_req_task_submit(nxt);
4954 percpu_ref_put(&ctx->refs);
4957 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4958 int sync, void *key)
4960 struct io_kiocb *req = wait->private;
4961 struct io_poll_iocb *poll = io_poll_get_single(req);
4962 __poll_t mask = key_to_poll(key);
4964 /* for instances that support it check for an event match first: */
4965 if (mask && !(mask & poll->events))
4968 list_del_init(&wait->entry);
4970 if (poll && poll->head) {
4973 spin_lock(&poll->head->lock);
4974 done = list_empty(&poll->wait.entry);
4976 list_del_init(&poll->wait.entry);
4977 /* make sure double remove sees this as being gone */
4978 wait->private = NULL;
4979 spin_unlock(&poll->head->lock);
4981 __io_async_wake(req, poll, mask, io_poll_task_func);
4983 refcount_dec(&req->refs);
4987 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4988 wait_queue_func_t wake_func)
4992 poll->canceled = false;
4993 poll->events = events;
4994 INIT_LIST_HEAD(&poll->wait.entry);
4995 init_waitqueue_func_entry(&poll->wait, wake_func);
4998 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4999 struct wait_queue_head *head,
5000 struct io_poll_iocb **poll_ptr)
5002 struct io_kiocb *req = pt->req;
5005 * If poll->head is already set, it's because the file being polled
5006 * uses multiple waitqueues for poll handling (eg one for read, one
5007 * for write). Setup a separate io_poll_iocb if this happens.
5009 if (unlikely(poll->head)) {
5010 struct io_poll_iocb *poll_one = poll;
5012 /* already have a 2nd entry, fail a third attempt */
5014 pt->error = -EINVAL;
5017 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5019 pt->error = -ENOMEM;
5022 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5023 refcount_inc(&req->refs);
5024 poll->wait.private = req;
5031 if (poll->events & EPOLLEXCLUSIVE)
5032 add_wait_queue_exclusive(head, &poll->wait);
5034 add_wait_queue(head, &poll->wait);
5037 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5038 struct poll_table_struct *p)
5040 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5041 struct async_poll *apoll = pt->req->apoll;
5043 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5046 static void io_async_task_func(struct callback_head *cb)
5048 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5049 struct async_poll *apoll = req->apoll;
5050 struct io_ring_ctx *ctx = req->ctx;
5052 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5054 if (io_poll_rewait(req, &apoll->poll)) {
5055 spin_unlock_irq(&ctx->completion_lock);
5056 percpu_ref_put(&ctx->refs);
5060 /* If req is still hashed, it cannot have been canceled. Don't check. */
5061 if (hash_hashed(&req->hash_node))
5062 hash_del(&req->hash_node);
5064 io_poll_remove_double(req);
5065 spin_unlock_irq(&ctx->completion_lock);
5067 if (!READ_ONCE(apoll->poll.canceled))
5068 __io_req_task_submit(req);
5070 __io_req_task_cancel(req, -ECANCELED);
5072 percpu_ref_put(&ctx->refs);
5073 kfree(apoll->double_poll);
5077 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5080 struct io_kiocb *req = wait->private;
5081 struct io_poll_iocb *poll = &req->apoll->poll;
5083 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5086 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5089 static void io_poll_req_insert(struct io_kiocb *req)
5091 struct io_ring_ctx *ctx = req->ctx;
5092 struct hlist_head *list;
5094 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5095 hlist_add_head(&req->hash_node, list);
5098 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5099 struct io_poll_iocb *poll,
5100 struct io_poll_table *ipt, __poll_t mask,
5101 wait_queue_func_t wake_func)
5102 __acquires(&ctx->completion_lock)
5104 struct io_ring_ctx *ctx = req->ctx;
5105 bool cancel = false;
5107 io_init_poll_iocb(poll, mask, wake_func);
5108 poll->file = req->file;
5109 poll->wait.private = req;
5111 ipt->pt._key = mask;
5113 ipt->error = -EINVAL;
5115 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5117 spin_lock_irq(&ctx->completion_lock);
5118 if (likely(poll->head)) {
5119 spin_lock(&poll->head->lock);
5120 if (unlikely(list_empty(&poll->wait.entry))) {
5126 if (mask || ipt->error)
5127 list_del_init(&poll->wait.entry);
5129 WRITE_ONCE(poll->canceled, true);
5130 else if (!poll->done) /* actually waiting for an event */
5131 io_poll_req_insert(req);
5132 spin_unlock(&poll->head->lock);
5138 static bool io_arm_poll_handler(struct io_kiocb *req)
5140 const struct io_op_def *def = &io_op_defs[req->opcode];
5141 struct io_ring_ctx *ctx = req->ctx;
5142 struct async_poll *apoll;
5143 struct io_poll_table ipt;
5147 if (!req->file || !file_can_poll(req->file))
5149 if (req->flags & REQ_F_POLLED)
5153 else if (def->pollout)
5157 /* if we can't nonblock try, then no point in arming a poll handler */
5158 if (!io_file_supports_async(req->file, rw))
5161 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5162 if (unlikely(!apoll))
5164 apoll->double_poll = NULL;
5166 req->flags |= REQ_F_POLLED;
5168 INIT_HLIST_NODE(&req->hash_node);
5172 mask |= POLLIN | POLLRDNORM;
5174 mask |= POLLOUT | POLLWRNORM;
5176 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5177 if ((req->opcode == IORING_OP_RECVMSG) &&
5178 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5181 mask |= POLLERR | POLLPRI;
5183 ipt.pt._qproc = io_async_queue_proc;
5185 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5187 if (ret || ipt.error) {
5188 io_poll_remove_double(req);
5189 spin_unlock_irq(&ctx->completion_lock);
5190 kfree(apoll->double_poll);
5194 spin_unlock_irq(&ctx->completion_lock);
5195 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5196 apoll->poll.events);
5200 static bool __io_poll_remove_one(struct io_kiocb *req,
5201 struct io_poll_iocb *poll)
5203 bool do_complete = false;
5205 spin_lock(&poll->head->lock);
5206 WRITE_ONCE(poll->canceled, true);
5207 if (!list_empty(&poll->wait.entry)) {
5208 list_del_init(&poll->wait.entry);
5211 spin_unlock(&poll->head->lock);
5212 hash_del(&req->hash_node);
5216 static bool io_poll_remove_one(struct io_kiocb *req)
5220 io_poll_remove_double(req);
5222 if (req->opcode == IORING_OP_POLL_ADD) {
5223 do_complete = __io_poll_remove_one(req, &req->poll);
5225 struct async_poll *apoll = req->apoll;
5227 /* non-poll requests have submit ref still */
5228 do_complete = __io_poll_remove_one(req, &apoll->poll);
5231 kfree(apoll->double_poll);
5237 io_cqring_fill_event(req, -ECANCELED);
5238 io_commit_cqring(req->ctx);
5239 req_set_fail_links(req);
5240 io_put_req_deferred(req, 1);
5247 * Returns true if we found and killed one or more poll requests
5249 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5251 struct hlist_node *tmp;
5252 struct io_kiocb *req;
5255 spin_lock_irq(&ctx->completion_lock);
5256 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5257 struct hlist_head *list;
5259 list = &ctx->cancel_hash[i];
5260 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5261 if (io_task_match(req, tsk))
5262 posted += io_poll_remove_one(req);
5265 spin_unlock_irq(&ctx->completion_lock);
5268 io_cqring_ev_posted(ctx);
5273 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5275 struct hlist_head *list;
5276 struct io_kiocb *req;
5278 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5279 hlist_for_each_entry(req, list, hash_node) {
5280 if (sqe_addr != req->user_data)
5282 if (io_poll_remove_one(req))
5290 static int io_poll_remove_prep(struct io_kiocb *req,
5291 const struct io_uring_sqe *sqe)
5293 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5295 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5299 req->poll.addr = READ_ONCE(sqe->addr);
5304 * Find a running poll command that matches one specified in sqe->addr,
5305 * and remove it if found.
5307 static int io_poll_remove(struct io_kiocb *req)
5309 struct io_ring_ctx *ctx = req->ctx;
5313 addr = req->poll.addr;
5314 spin_lock_irq(&ctx->completion_lock);
5315 ret = io_poll_cancel(ctx, addr);
5316 spin_unlock_irq(&ctx->completion_lock);
5319 req_set_fail_links(req);
5320 io_req_complete(req, ret);
5324 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5327 struct io_kiocb *req = wait->private;
5328 struct io_poll_iocb *poll = &req->poll;
5330 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5333 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5334 struct poll_table_struct *p)
5336 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5338 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5341 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5343 struct io_poll_iocb *poll = &req->poll;
5346 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5348 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5351 events = READ_ONCE(sqe->poll32_events);
5353 events = swahw32(events);
5355 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5356 (events & EPOLLEXCLUSIVE);
5360 static int io_poll_add(struct io_kiocb *req)
5362 struct io_poll_iocb *poll = &req->poll;
5363 struct io_ring_ctx *ctx = req->ctx;
5364 struct io_poll_table ipt;
5367 INIT_HLIST_NODE(&req->hash_node);
5368 ipt.pt._qproc = io_poll_queue_proc;
5370 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5373 if (mask) { /* no async, we'd stolen it */
5375 io_poll_complete(req, mask, 0);
5377 spin_unlock_irq(&ctx->completion_lock);
5380 io_cqring_ev_posted(ctx);
5386 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5388 struct io_timeout_data *data = container_of(timer,
5389 struct io_timeout_data, timer);
5390 struct io_kiocb *req = data->req;
5391 struct io_ring_ctx *ctx = req->ctx;
5392 unsigned long flags;
5394 spin_lock_irqsave(&ctx->completion_lock, flags);
5395 list_del_init(&req->timeout.list);
5396 atomic_set(&req->ctx->cq_timeouts,
5397 atomic_read(&req->ctx->cq_timeouts) + 1);
5399 io_cqring_fill_event(req, -ETIME);
5400 io_commit_cqring(ctx);
5401 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5403 io_cqring_ev_posted(ctx);
5404 req_set_fail_links(req);
5406 return HRTIMER_NORESTART;
5409 static int __io_timeout_cancel(struct io_kiocb *req)
5411 struct io_timeout_data *io = req->async_data;
5414 ret = hrtimer_try_to_cancel(&io->timer);
5417 list_del_init(&req->timeout.list);
5419 req_set_fail_links(req);
5420 io_cqring_fill_event(req, -ECANCELED);
5421 io_put_req_deferred(req, 1);
5425 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5427 struct io_kiocb *req;
5430 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5431 if (user_data == req->user_data) {
5440 return __io_timeout_cancel(req);
5443 static int io_timeout_remove_prep(struct io_kiocb *req,
5444 const struct io_uring_sqe *sqe)
5446 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5448 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5450 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5453 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5458 * Remove or update an existing timeout command
5460 static int io_timeout_remove(struct io_kiocb *req)
5462 struct io_ring_ctx *ctx = req->ctx;
5465 spin_lock_irq(&ctx->completion_lock);
5466 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5468 io_cqring_fill_event(req, ret);
5469 io_commit_cqring(ctx);
5470 spin_unlock_irq(&ctx->completion_lock);
5471 io_cqring_ev_posted(ctx);
5473 req_set_fail_links(req);
5478 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5479 bool is_timeout_link)
5481 struct io_timeout_data *data;
5483 u32 off = READ_ONCE(sqe->off);
5485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5487 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5489 if (off && is_timeout_link)
5491 flags = READ_ONCE(sqe->timeout_flags);
5492 if (flags & ~IORING_TIMEOUT_ABS)
5495 req->timeout.off = off;
5497 if (!req->async_data && io_alloc_async_data(req))
5500 data = req->async_data;
5503 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5506 if (flags & IORING_TIMEOUT_ABS)
5507 data->mode = HRTIMER_MODE_ABS;
5509 data->mode = HRTIMER_MODE_REL;
5511 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5515 static int io_timeout(struct io_kiocb *req)
5517 struct io_ring_ctx *ctx = req->ctx;
5518 struct io_timeout_data *data = req->async_data;
5519 struct list_head *entry;
5520 u32 tail, off = req->timeout.off;
5522 spin_lock_irq(&ctx->completion_lock);
5525 * sqe->off holds how many events that need to occur for this
5526 * timeout event to be satisfied. If it isn't set, then this is
5527 * a pure timeout request, sequence isn't used.
5529 if (io_is_timeout_noseq(req)) {
5530 entry = ctx->timeout_list.prev;
5534 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5535 req->timeout.target_seq = tail + off;
5538 * Insertion sort, ensuring the first entry in the list is always
5539 * the one we need first.
5541 list_for_each_prev(entry, &ctx->timeout_list) {
5542 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5545 if (io_is_timeout_noseq(nxt))
5547 /* nxt.seq is behind @tail, otherwise would've been completed */
5548 if (off >= nxt->timeout.target_seq - tail)
5552 list_add(&req->timeout.list, entry);
5553 data->timer.function = io_timeout_fn;
5554 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5555 spin_unlock_irq(&ctx->completion_lock);
5559 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5561 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5563 return req->user_data == (unsigned long) data;
5566 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5568 enum io_wq_cancel cancel_ret;
5571 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5572 switch (cancel_ret) {
5573 case IO_WQ_CANCEL_OK:
5576 case IO_WQ_CANCEL_RUNNING:
5579 case IO_WQ_CANCEL_NOTFOUND:
5587 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5588 struct io_kiocb *req, __u64 sqe_addr,
5591 unsigned long flags;
5594 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5595 if (ret != -ENOENT) {
5596 spin_lock_irqsave(&ctx->completion_lock, flags);
5600 spin_lock_irqsave(&ctx->completion_lock, flags);
5601 ret = io_timeout_cancel(ctx, sqe_addr);
5604 ret = io_poll_cancel(ctx, sqe_addr);
5608 io_cqring_fill_event(req, ret);
5609 io_commit_cqring(ctx);
5610 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5611 io_cqring_ev_posted(ctx);
5614 req_set_fail_links(req);
5618 static int io_async_cancel_prep(struct io_kiocb *req,
5619 const struct io_uring_sqe *sqe)
5621 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5623 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5625 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5628 req->cancel.addr = READ_ONCE(sqe->addr);
5632 static int io_async_cancel(struct io_kiocb *req)
5634 struct io_ring_ctx *ctx = req->ctx;
5636 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5640 static int io_files_update_prep(struct io_kiocb *req,
5641 const struct io_uring_sqe *sqe)
5643 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5645 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5647 if (sqe->ioprio || sqe->rw_flags)
5650 req->files_update.offset = READ_ONCE(sqe->off);
5651 req->files_update.nr_args = READ_ONCE(sqe->len);
5652 if (!req->files_update.nr_args)
5654 req->files_update.arg = READ_ONCE(sqe->addr);
5658 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5659 struct io_comp_state *cs)
5661 struct io_ring_ctx *ctx = req->ctx;
5662 struct io_uring_files_update up;
5668 up.offset = req->files_update.offset;
5669 up.fds = req->files_update.arg;
5671 mutex_lock(&ctx->uring_lock);
5672 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5673 mutex_unlock(&ctx->uring_lock);
5676 req_set_fail_links(req);
5677 __io_req_complete(req, ret, 0, cs);
5681 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5683 switch (req->opcode) {
5686 case IORING_OP_READV:
5687 case IORING_OP_READ_FIXED:
5688 case IORING_OP_READ:
5689 return io_read_prep(req, sqe);
5690 case IORING_OP_WRITEV:
5691 case IORING_OP_WRITE_FIXED:
5692 case IORING_OP_WRITE:
5693 return io_write_prep(req, sqe);
5694 case IORING_OP_POLL_ADD:
5695 return io_poll_add_prep(req, sqe);
5696 case IORING_OP_POLL_REMOVE:
5697 return io_poll_remove_prep(req, sqe);
5698 case IORING_OP_FSYNC:
5699 return io_prep_fsync(req, sqe);
5700 case IORING_OP_SYNC_FILE_RANGE:
5701 return io_prep_sfr(req, sqe);
5702 case IORING_OP_SENDMSG:
5703 case IORING_OP_SEND:
5704 return io_sendmsg_prep(req, sqe);
5705 case IORING_OP_RECVMSG:
5706 case IORING_OP_RECV:
5707 return io_recvmsg_prep(req, sqe);
5708 case IORING_OP_CONNECT:
5709 return io_connect_prep(req, sqe);
5710 case IORING_OP_TIMEOUT:
5711 return io_timeout_prep(req, sqe, false);
5712 case IORING_OP_TIMEOUT_REMOVE:
5713 return io_timeout_remove_prep(req, sqe);
5714 case IORING_OP_ASYNC_CANCEL:
5715 return io_async_cancel_prep(req, sqe);
5716 case IORING_OP_LINK_TIMEOUT:
5717 return io_timeout_prep(req, sqe, true);
5718 case IORING_OP_ACCEPT:
5719 return io_accept_prep(req, sqe);
5720 case IORING_OP_FALLOCATE:
5721 return io_fallocate_prep(req, sqe);
5722 case IORING_OP_OPENAT:
5723 return io_openat_prep(req, sqe);
5724 case IORING_OP_CLOSE:
5725 return io_close_prep(req, sqe);
5726 case IORING_OP_FILES_UPDATE:
5727 return io_files_update_prep(req, sqe);
5728 case IORING_OP_STATX:
5729 return io_statx_prep(req, sqe);
5730 case IORING_OP_FADVISE:
5731 return io_fadvise_prep(req, sqe);
5732 case IORING_OP_MADVISE:
5733 return io_madvise_prep(req, sqe);
5734 case IORING_OP_OPENAT2:
5735 return io_openat2_prep(req, sqe);
5736 case IORING_OP_EPOLL_CTL:
5737 return io_epoll_ctl_prep(req, sqe);
5738 case IORING_OP_SPLICE:
5739 return io_splice_prep(req, sqe);
5740 case IORING_OP_PROVIDE_BUFFERS:
5741 return io_provide_buffers_prep(req, sqe);
5742 case IORING_OP_REMOVE_BUFFERS:
5743 return io_remove_buffers_prep(req, sqe);
5745 return io_tee_prep(req, sqe);
5748 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5753 static int io_req_defer_prep(struct io_kiocb *req,
5754 const struct io_uring_sqe *sqe)
5758 if (io_alloc_async_data(req))
5760 return io_req_prep(req, sqe);
5763 static u32 io_get_sequence(struct io_kiocb *req)
5765 struct io_kiocb *pos;
5766 struct io_ring_ctx *ctx = req->ctx;
5767 u32 total_submitted, nr_reqs = 1;
5769 if (req->flags & REQ_F_LINK_HEAD)
5770 list_for_each_entry(pos, &req->link_list, link_list)
5773 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5774 return total_submitted - nr_reqs;
5777 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5779 struct io_ring_ctx *ctx = req->ctx;
5780 struct io_defer_entry *de;
5784 /* Still need defer if there is pending req in defer list. */
5785 if (likely(list_empty_careful(&ctx->defer_list) &&
5786 !(req->flags & REQ_F_IO_DRAIN)))
5789 seq = io_get_sequence(req);
5790 /* Still a chance to pass the sequence check */
5791 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5794 if (!req->async_data) {
5795 ret = io_req_defer_prep(req, sqe);
5799 io_prep_async_link(req);
5800 de = kmalloc(sizeof(*de), GFP_KERNEL);
5804 spin_lock_irq(&ctx->completion_lock);
5805 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5806 spin_unlock_irq(&ctx->completion_lock);
5808 io_queue_async_work(req);
5809 return -EIOCBQUEUED;
5812 trace_io_uring_defer(ctx, req, req->user_data);
5815 list_add_tail(&de->list, &ctx->defer_list);
5816 spin_unlock_irq(&ctx->completion_lock);
5817 return -EIOCBQUEUED;
5820 static void io_req_drop_files(struct io_kiocb *req)
5822 struct io_ring_ctx *ctx = req->ctx;
5823 unsigned long flags;
5825 spin_lock_irqsave(&ctx->inflight_lock, flags);
5826 list_del(&req->inflight_entry);
5827 if (waitqueue_active(&ctx->inflight_wait))
5828 wake_up(&ctx->inflight_wait);
5829 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5830 req->flags &= ~REQ_F_INFLIGHT;
5831 put_files_struct(req->work.identity->files);
5832 put_nsproxy(req->work.identity->nsproxy);
5833 req->work.flags &= ~IO_WQ_WORK_FILES;
5836 static void __io_clean_op(struct io_kiocb *req)
5838 if (req->flags & REQ_F_BUFFER_SELECTED) {
5839 switch (req->opcode) {
5840 case IORING_OP_READV:
5841 case IORING_OP_READ_FIXED:
5842 case IORING_OP_READ:
5843 kfree((void *)(unsigned long)req->rw.addr);
5845 case IORING_OP_RECVMSG:
5846 case IORING_OP_RECV:
5847 kfree(req->sr_msg.kbuf);
5850 req->flags &= ~REQ_F_BUFFER_SELECTED;
5853 if (req->flags & REQ_F_NEED_CLEANUP) {
5854 switch (req->opcode) {
5855 case IORING_OP_READV:
5856 case IORING_OP_READ_FIXED:
5857 case IORING_OP_READ:
5858 case IORING_OP_WRITEV:
5859 case IORING_OP_WRITE_FIXED:
5860 case IORING_OP_WRITE: {
5861 struct io_async_rw *io = req->async_data;
5863 kfree(io->free_iovec);
5866 case IORING_OP_RECVMSG:
5867 case IORING_OP_SENDMSG: {
5868 struct io_async_msghdr *io = req->async_data;
5869 if (io->iov != io->fast_iov)
5873 case IORING_OP_SPLICE:
5875 io_put_file(req, req->splice.file_in,
5876 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5878 case IORING_OP_OPENAT:
5879 case IORING_OP_OPENAT2:
5880 if (req->open.filename)
5881 putname(req->open.filename);
5884 req->flags &= ~REQ_F_NEED_CLEANUP;
5887 if (req->flags & REQ_F_INFLIGHT)
5888 io_req_drop_files(req);
5891 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5892 struct io_comp_state *cs)
5894 struct io_ring_ctx *ctx = req->ctx;
5897 switch (req->opcode) {
5899 ret = io_nop(req, cs);
5901 case IORING_OP_READV:
5902 case IORING_OP_READ_FIXED:
5903 case IORING_OP_READ:
5904 ret = io_read(req, force_nonblock, cs);
5906 case IORING_OP_WRITEV:
5907 case IORING_OP_WRITE_FIXED:
5908 case IORING_OP_WRITE:
5909 ret = io_write(req, force_nonblock, cs);
5911 case IORING_OP_FSYNC:
5912 ret = io_fsync(req, force_nonblock);
5914 case IORING_OP_POLL_ADD:
5915 ret = io_poll_add(req);
5917 case IORING_OP_POLL_REMOVE:
5918 ret = io_poll_remove(req);
5920 case IORING_OP_SYNC_FILE_RANGE:
5921 ret = io_sync_file_range(req, force_nonblock);
5923 case IORING_OP_SENDMSG:
5924 ret = io_sendmsg(req, force_nonblock, cs);
5926 case IORING_OP_SEND:
5927 ret = io_send(req, force_nonblock, cs);
5929 case IORING_OP_RECVMSG:
5930 ret = io_recvmsg(req, force_nonblock, cs);
5932 case IORING_OP_RECV:
5933 ret = io_recv(req, force_nonblock, cs);
5935 case IORING_OP_TIMEOUT:
5936 ret = io_timeout(req);
5938 case IORING_OP_TIMEOUT_REMOVE:
5939 ret = io_timeout_remove(req);
5941 case IORING_OP_ACCEPT:
5942 ret = io_accept(req, force_nonblock, cs);
5944 case IORING_OP_CONNECT:
5945 ret = io_connect(req, force_nonblock, cs);
5947 case IORING_OP_ASYNC_CANCEL:
5948 ret = io_async_cancel(req);
5950 case IORING_OP_FALLOCATE:
5951 ret = io_fallocate(req, force_nonblock);
5953 case IORING_OP_OPENAT:
5954 ret = io_openat(req, force_nonblock);
5956 case IORING_OP_CLOSE:
5957 ret = io_close(req, force_nonblock, cs);
5959 case IORING_OP_FILES_UPDATE:
5960 ret = io_files_update(req, force_nonblock, cs);
5962 case IORING_OP_STATX:
5963 ret = io_statx(req, force_nonblock);
5965 case IORING_OP_FADVISE:
5966 ret = io_fadvise(req, force_nonblock);
5968 case IORING_OP_MADVISE:
5969 ret = io_madvise(req, force_nonblock);
5971 case IORING_OP_OPENAT2:
5972 ret = io_openat2(req, force_nonblock);
5974 case IORING_OP_EPOLL_CTL:
5975 ret = io_epoll_ctl(req, force_nonblock, cs);
5977 case IORING_OP_SPLICE:
5978 ret = io_splice(req, force_nonblock);
5980 case IORING_OP_PROVIDE_BUFFERS:
5981 ret = io_provide_buffers(req, force_nonblock, cs);
5983 case IORING_OP_REMOVE_BUFFERS:
5984 ret = io_remove_buffers(req, force_nonblock, cs);
5987 ret = io_tee(req, force_nonblock);
5997 /* If the op doesn't have a file, we're not polling for it */
5998 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5999 const bool in_async = io_wq_current_is_worker();
6001 /* workqueue context doesn't hold uring_lock, grab it now */
6003 mutex_lock(&ctx->uring_lock);
6005 io_iopoll_req_issued(req);
6008 mutex_unlock(&ctx->uring_lock);
6014 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6016 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6017 struct io_kiocb *timeout;
6020 timeout = io_prep_linked_timeout(req);
6022 io_queue_linked_timeout(timeout);
6024 /* if NO_CANCEL is set, we must still run the work */
6025 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6026 IO_WQ_WORK_CANCEL) {
6032 ret = io_issue_sqe(req, false, NULL);
6034 * We can get EAGAIN for polled IO even though we're
6035 * forcing a sync submission from here, since we can't
6036 * wait for request slots on the block side.
6045 req_set_fail_links(req);
6046 io_req_complete(req, ret);
6049 return io_steal_work(req);
6052 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6055 struct fixed_file_table *table;
6057 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6058 return table->files[index & IORING_FILE_TABLE_MASK];
6061 static struct file *io_file_get(struct io_submit_state *state,
6062 struct io_kiocb *req, int fd, bool fixed)
6064 struct io_ring_ctx *ctx = req->ctx;
6068 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6070 fd = array_index_nospec(fd, ctx->nr_user_files);
6071 file = io_file_from_index(ctx, fd);
6073 req->fixed_file_refs = &ctx->file_data->node->refs;
6074 percpu_ref_get(req->fixed_file_refs);
6077 trace_io_uring_file_get(ctx, fd);
6078 file = __io_file_get(state, fd);
6084 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6089 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6090 if (unlikely(!fixed && io_async_submit(req->ctx)))
6093 req->file = io_file_get(state, req, fd, fixed);
6094 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6099 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6101 struct io_timeout_data *data = container_of(timer,
6102 struct io_timeout_data, timer);
6103 struct io_kiocb *req = data->req;
6104 struct io_ring_ctx *ctx = req->ctx;
6105 struct io_kiocb *prev = NULL;
6106 unsigned long flags;
6108 spin_lock_irqsave(&ctx->completion_lock, flags);
6111 * We don't expect the list to be empty, that will only happen if we
6112 * race with the completion of the linked work.
6114 if (!list_empty(&req->link_list)) {
6115 prev = list_entry(req->link_list.prev, struct io_kiocb,
6117 if (refcount_inc_not_zero(&prev->refs)) {
6118 list_del_init(&req->link_list);
6119 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6124 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6127 req_set_fail_links(prev);
6128 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6131 io_req_complete(req, -ETIME);
6133 return HRTIMER_NORESTART;
6136 static void __io_queue_linked_timeout(struct io_kiocb *req)
6139 * If the list is now empty, then our linked request finished before
6140 * we got a chance to setup the timer
6142 if (!list_empty(&req->link_list)) {
6143 struct io_timeout_data *data = req->async_data;
6145 data->timer.function = io_link_timeout_fn;
6146 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6151 static void io_queue_linked_timeout(struct io_kiocb *req)
6153 struct io_ring_ctx *ctx = req->ctx;
6155 spin_lock_irq(&ctx->completion_lock);
6156 __io_queue_linked_timeout(req);
6157 spin_unlock_irq(&ctx->completion_lock);
6159 /* drop submission reference */
6163 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6165 struct io_kiocb *nxt;
6167 if (!(req->flags & REQ_F_LINK_HEAD))
6169 if (req->flags & REQ_F_LINK_TIMEOUT)
6172 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6174 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6177 req->flags |= REQ_F_LINK_TIMEOUT;
6181 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6183 struct io_kiocb *linked_timeout;
6184 struct io_kiocb *nxt;
6185 const struct cred *old_creds = NULL;
6189 linked_timeout = io_prep_linked_timeout(req);
6191 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6192 (req->work.flags & IO_WQ_WORK_CREDS) &&
6193 req->work.identity->creds != current_cred()) {
6195 revert_creds(old_creds);
6196 if (old_creds == req->work.identity->creds)
6197 old_creds = NULL; /* restored original creds */
6199 old_creds = override_creds(req->work.identity->creds);
6202 ret = io_issue_sqe(req, true, cs);
6205 * We async punt it if the file wasn't marked NOWAIT, or if the file
6206 * doesn't support non-blocking read/write attempts
6208 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6209 if (!io_arm_poll_handler(req)) {
6212 * Queued up for async execution, worker will release
6213 * submit reference when the iocb is actually submitted.
6215 io_queue_async_work(req);
6219 io_queue_linked_timeout(linked_timeout);
6223 if (unlikely(ret)) {
6224 /* un-prep timeout, so it'll be killed as any other linked */
6225 req->flags &= ~REQ_F_LINK_TIMEOUT;
6226 req_set_fail_links(req);
6228 io_req_complete(req, ret);
6232 /* drop submission reference */
6233 nxt = io_put_req_find_next(req);
6235 io_queue_linked_timeout(linked_timeout);
6240 if (req->flags & REQ_F_FORCE_ASYNC)
6246 revert_creds(old_creds);
6249 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6250 struct io_comp_state *cs)
6254 ret = io_req_defer(req, sqe);
6256 if (ret != -EIOCBQUEUED) {
6258 req_set_fail_links(req);
6260 io_req_complete(req, ret);
6262 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6263 if (!req->async_data) {
6264 ret = io_req_defer_prep(req, sqe);
6270 * Never try inline submit of IOSQE_ASYNC is set, go straight
6271 * to async execution.
6273 io_req_init_async(req);
6274 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6275 io_queue_async_work(req);
6278 ret = io_req_prep(req, sqe);
6282 __io_queue_sqe(req, cs);
6286 static inline void io_queue_link_head(struct io_kiocb *req,
6287 struct io_comp_state *cs)
6289 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6291 io_req_complete(req, -ECANCELED);
6293 io_queue_sqe(req, NULL, cs);
6296 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6297 struct io_kiocb **link, struct io_comp_state *cs)
6299 struct io_ring_ctx *ctx = req->ctx;
6303 * If we already have a head request, queue this one for async
6304 * submittal once the head completes. If we don't have a head but
6305 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6306 * submitted sync once the chain is complete. If none of those
6307 * conditions are true (normal request), then just queue it.
6310 struct io_kiocb *head = *link;
6313 * Taking sequential execution of a link, draining both sides
6314 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6315 * requests in the link. So, it drains the head and the
6316 * next after the link request. The last one is done via
6317 * drain_next flag to persist the effect across calls.
6319 if (req->flags & REQ_F_IO_DRAIN) {
6320 head->flags |= REQ_F_IO_DRAIN;
6321 ctx->drain_next = 1;
6323 ret = io_req_defer_prep(req, sqe);
6324 if (unlikely(ret)) {
6325 /* fail even hard links since we don't submit */
6326 head->flags |= REQ_F_FAIL_LINK;
6329 trace_io_uring_link(ctx, req, head);
6330 list_add_tail(&req->link_list, &head->link_list);
6332 /* last request of a link, enqueue the link */
6333 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6334 io_queue_link_head(head, cs);
6338 if (unlikely(ctx->drain_next)) {
6339 req->flags |= REQ_F_IO_DRAIN;
6340 ctx->drain_next = 0;
6342 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6343 req->flags |= REQ_F_LINK_HEAD;
6344 INIT_LIST_HEAD(&req->link_list);
6346 ret = io_req_defer_prep(req, sqe);
6348 req->flags |= REQ_F_FAIL_LINK;
6351 io_queue_sqe(req, sqe, cs);
6359 * Batched submission is done, ensure local IO is flushed out.
6361 static void io_submit_state_end(struct io_submit_state *state)
6363 if (!list_empty(&state->comp.list))
6364 io_submit_flush_completions(&state->comp);
6365 blk_finish_plug(&state->plug);
6366 io_state_file_put(state);
6367 if (state->free_reqs)
6368 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6372 * Start submission side cache.
6374 static void io_submit_state_start(struct io_submit_state *state,
6375 struct io_ring_ctx *ctx, unsigned int max_ios)
6377 blk_start_plug(&state->plug);
6379 INIT_LIST_HEAD(&state->comp.list);
6380 state->comp.ctx = ctx;
6381 state->free_reqs = 0;
6383 state->ios_left = max_ios;
6386 static void io_commit_sqring(struct io_ring_ctx *ctx)
6388 struct io_rings *rings = ctx->rings;
6391 * Ensure any loads from the SQEs are done at this point,
6392 * since once we write the new head, the application could
6393 * write new data to them.
6395 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6399 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6400 * that is mapped by userspace. This means that care needs to be taken to
6401 * ensure that reads are stable, as we cannot rely on userspace always
6402 * being a good citizen. If members of the sqe are validated and then later
6403 * used, it's important that those reads are done through READ_ONCE() to
6404 * prevent a re-load down the line.
6406 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6408 u32 *sq_array = ctx->sq_array;
6412 * The cached sq head (or cq tail) serves two purposes:
6414 * 1) allows us to batch the cost of updating the user visible
6416 * 2) allows the kernel side to track the head on its own, even
6417 * though the application is the one updating it.
6419 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6420 if (likely(head < ctx->sq_entries))
6421 return &ctx->sq_sqes[head];
6423 /* drop invalid entries */
6424 ctx->cached_sq_dropped++;
6425 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6429 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6431 ctx->cached_sq_head++;
6435 * Check SQE restrictions (opcode and flags).
6437 * Returns 'true' if SQE is allowed, 'false' otherwise.
6439 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6440 struct io_kiocb *req,
6441 unsigned int sqe_flags)
6443 if (!ctx->restricted)
6446 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6449 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6450 ctx->restrictions.sqe_flags_required)
6453 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6454 ctx->restrictions.sqe_flags_required))
6460 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6461 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6462 IOSQE_BUFFER_SELECT)
6464 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6465 const struct io_uring_sqe *sqe,
6466 struct io_submit_state *state)
6468 unsigned int sqe_flags;
6471 req->opcode = READ_ONCE(sqe->opcode);
6472 req->user_data = READ_ONCE(sqe->user_data);
6473 req->async_data = NULL;
6477 /* one is dropped after submission, the other at completion */
6478 refcount_set(&req->refs, 2);
6479 req->task = current;
6482 if (unlikely(req->opcode >= IORING_OP_LAST))
6485 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6488 sqe_flags = READ_ONCE(sqe->flags);
6489 /* enforce forwards compatibility on users */
6490 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6493 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6496 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6497 !io_op_defs[req->opcode].buffer_select)
6500 id = READ_ONCE(sqe->personality);
6502 struct io_identity *iod;
6504 iod = idr_find(&ctx->personality_idr, id);
6507 refcount_inc(&iod->count);
6509 __io_req_init_async(req);
6510 get_cred(iod->creds);
6511 req->work.identity = iod;
6512 req->work.flags |= IO_WQ_WORK_CREDS;
6515 /* same numerical values with corresponding REQ_F_*, safe to copy */
6516 req->flags |= sqe_flags;
6518 if (!io_op_defs[req->opcode].needs_file)
6521 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6526 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6528 struct io_submit_state state;
6529 struct io_kiocb *link = NULL;
6530 int i, submitted = 0;
6532 /* if we have a backlog and couldn't flush it all, return BUSY */
6533 if (test_bit(0, &ctx->sq_check_overflow)) {
6534 if (!list_empty(&ctx->cq_overflow_list) &&
6535 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6539 /* make sure SQ entry isn't read before tail */
6540 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6542 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6545 percpu_counter_add(¤t->io_uring->inflight, nr);
6546 refcount_add(nr, ¤t->usage);
6548 io_submit_state_start(&state, ctx, nr);
6550 for (i = 0; i < nr; i++) {
6551 const struct io_uring_sqe *sqe;
6552 struct io_kiocb *req;
6555 sqe = io_get_sqe(ctx);
6556 if (unlikely(!sqe)) {
6557 io_consume_sqe(ctx);
6560 req = io_alloc_req(ctx, &state);
6561 if (unlikely(!req)) {
6563 submitted = -EAGAIN;
6566 io_consume_sqe(ctx);
6567 /* will complete beyond this point, count as submitted */
6570 err = io_init_req(ctx, req, sqe, &state);
6571 if (unlikely(err)) {
6574 io_req_complete(req, err);
6578 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6579 true, io_async_submit(ctx));
6580 err = io_submit_sqe(req, sqe, &link, &state.comp);
6585 if (unlikely(submitted != nr)) {
6586 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6587 struct io_uring_task *tctx = current->io_uring;
6588 int unused = nr - ref_used;
6590 percpu_ref_put_many(&ctx->refs, unused);
6591 percpu_counter_sub(&tctx->inflight, unused);
6592 put_task_struct_many(current, unused);
6595 io_queue_link_head(link, &state.comp);
6596 io_submit_state_end(&state);
6598 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6599 io_commit_sqring(ctx);
6604 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6606 /* Tell userspace we may need a wakeup call */
6607 spin_lock_irq(&ctx->completion_lock);
6608 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6609 spin_unlock_irq(&ctx->completion_lock);
6612 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6614 spin_lock_irq(&ctx->completion_lock);
6615 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6616 spin_unlock_irq(&ctx->completion_lock);
6619 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6620 int sync, void *key)
6622 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6625 ret = autoremove_wake_function(wqe, mode, sync, key);
6627 unsigned long flags;
6629 spin_lock_irqsave(&ctx->completion_lock, flags);
6630 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6631 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6642 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6643 unsigned long start_jiffies, bool cap_entries)
6645 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6646 struct io_sq_data *sqd = ctx->sq_data;
6647 unsigned int to_submit;
6651 if (!list_empty(&ctx->iopoll_list)) {
6652 unsigned nr_events = 0;
6654 mutex_lock(&ctx->uring_lock);
6655 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6656 io_do_iopoll(ctx, &nr_events, 0);
6657 mutex_unlock(&ctx->uring_lock);
6660 to_submit = io_sqring_entries(ctx);
6663 * If submit got -EBUSY, flag us as needing the application
6664 * to enter the kernel to reap and flush events.
6666 if (!to_submit || ret == -EBUSY || need_resched()) {
6668 * Drop cur_mm before scheduling, we can't hold it for
6669 * long periods (or over schedule()). Do this before
6670 * adding ourselves to the waitqueue, as the unuse/drop
6673 io_sq_thread_drop_mm();
6676 * We're polling. If we're within the defined idle
6677 * period, then let us spin without work before going
6678 * to sleep. The exception is if we got EBUSY doing
6679 * more IO, we should wait for the application to
6680 * reap events and wake us up.
6682 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6683 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6684 !percpu_ref_is_dying(&ctx->refs)))
6687 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6688 TASK_INTERRUPTIBLE);
6691 * While doing polled IO, before going to sleep, we need
6692 * to check if there are new reqs added to iopoll_list,
6693 * it is because reqs may have been punted to io worker
6694 * and will be added to iopoll_list later, hence check
6695 * the iopoll_list again.
6697 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6698 !list_empty_careful(&ctx->iopoll_list)) {
6699 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6703 to_submit = io_sqring_entries(ctx);
6704 if (!to_submit || ret == -EBUSY)
6708 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6709 io_ring_clear_wakeup_flag(ctx);
6711 /* if we're handling multiple rings, cap submit size for fairness */
6712 if (cap_entries && to_submit > 8)
6715 mutex_lock(&ctx->uring_lock);
6716 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6717 ret = io_submit_sqes(ctx, to_submit);
6718 mutex_unlock(&ctx->uring_lock);
6720 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6721 wake_up(&ctx->sqo_sq_wait);
6723 return SQT_DID_WORK;
6726 static void io_sqd_init_new(struct io_sq_data *sqd)
6728 struct io_ring_ctx *ctx;
6730 while (!list_empty(&sqd->ctx_new_list)) {
6731 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6732 init_wait(&ctx->sqo_wait_entry);
6733 ctx->sqo_wait_entry.func = io_sq_wake_function;
6734 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6735 complete(&ctx->sq_thread_comp);
6739 static int io_sq_thread(void *data)
6741 struct cgroup_subsys_state *cur_css = NULL;
6742 const struct cred *old_cred = NULL;
6743 struct io_sq_data *sqd = data;
6744 struct io_ring_ctx *ctx;
6745 unsigned long start_jiffies;
6747 start_jiffies = jiffies;
6748 while (!kthread_should_stop()) {
6749 enum sq_ret ret = 0;
6753 * Any changes to the sqd lists are synchronized through the
6754 * kthread parking. This synchronizes the thread vs users,
6755 * the users are synchronized on the sqd->ctx_lock.
6757 if (kthread_should_park())
6760 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6761 io_sqd_init_new(sqd);
6763 cap_entries = !list_is_singular(&sqd->ctx_list);
6765 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6766 if (current->cred != ctx->creds) {
6768 revert_creds(old_cred);
6769 old_cred = override_creds(ctx->creds);
6771 io_sq_thread_associate_blkcg(ctx, &cur_css);
6773 current->loginuid = ctx->loginuid;
6774 current->sessionid = ctx->sessionid;
6777 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6779 io_sq_thread_drop_mm();
6782 if (ret & SQT_SPIN) {
6785 } else if (ret == SQT_IDLE) {
6786 if (kthread_should_park())
6788 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6789 io_ring_set_wakeup_flag(ctx);
6791 start_jiffies = jiffies;
6792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6793 io_ring_clear_wakeup_flag(ctx);
6800 io_sq_thread_unassociate_blkcg();
6802 revert_creds(old_cred);
6809 struct io_wait_queue {
6810 struct wait_queue_entry wq;
6811 struct io_ring_ctx *ctx;
6813 unsigned nr_timeouts;
6816 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6818 struct io_ring_ctx *ctx = iowq->ctx;
6821 * Wake up if we have enough events, or if a timeout occurred since we
6822 * started waiting. For timeouts, we always want to return to userspace,
6823 * regardless of event count.
6825 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6826 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6829 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6830 int wake_flags, void *key)
6832 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6835 /* use noflush == true, as we can't safely rely on locking context */
6836 if (!io_should_wake(iowq, true))
6839 return autoremove_wake_function(curr, mode, wake_flags, key);
6842 static int io_run_task_work_sig(void)
6844 if (io_run_task_work())
6846 if (!signal_pending(current))
6848 if (current->jobctl & JOBCTL_TASK_WORK) {
6849 spin_lock_irq(¤t->sighand->siglock);
6850 current->jobctl &= ~JOBCTL_TASK_WORK;
6851 recalc_sigpending();
6852 spin_unlock_irq(¤t->sighand->siglock);
6859 * Wait until events become available, if we don't already have some. The
6860 * application must reap them itself, as they reside on the shared cq ring.
6862 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6863 const sigset_t __user *sig, size_t sigsz)
6865 struct io_wait_queue iowq = {
6868 .func = io_wake_function,
6869 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6872 .to_wait = min_events,
6874 struct io_rings *rings = ctx->rings;
6878 if (io_cqring_events(ctx, false) >= min_events)
6880 if (!io_run_task_work())
6885 #ifdef CONFIG_COMPAT
6886 if (in_compat_syscall())
6887 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6891 ret = set_user_sigmask(sig, sigsz);
6897 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6898 trace_io_uring_cqring_wait(ctx, min_events);
6900 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6901 TASK_INTERRUPTIBLE);
6902 /* make sure we run task_work before checking for signals */
6903 ret = io_run_task_work_sig();
6908 if (io_should_wake(&iowq, false))
6912 finish_wait(&ctx->wait, &iowq.wq);
6914 restore_saved_sigmask_unless(ret == -EINTR);
6916 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6919 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6921 #if defined(CONFIG_UNIX)
6922 if (ctx->ring_sock) {
6923 struct sock *sock = ctx->ring_sock->sk;
6924 struct sk_buff *skb;
6926 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6932 for (i = 0; i < ctx->nr_user_files; i++) {
6935 file = io_file_from_index(ctx, i);
6942 static void io_file_ref_kill(struct percpu_ref *ref)
6944 struct fixed_file_data *data;
6946 data = container_of(ref, struct fixed_file_data, refs);
6947 complete(&data->done);
6950 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6952 struct fixed_file_data *data = ctx->file_data;
6953 struct fixed_file_ref_node *ref_node = NULL;
6954 unsigned nr_tables, i;
6959 spin_lock(&data->lock);
6960 if (!list_empty(&data->ref_list))
6961 ref_node = list_first_entry(&data->ref_list,
6962 struct fixed_file_ref_node, node);
6963 spin_unlock(&data->lock);
6965 percpu_ref_kill(&ref_node->refs);
6967 percpu_ref_kill(&data->refs);
6969 /* wait for all refs nodes to complete */
6970 flush_delayed_work(&ctx->file_put_work);
6971 wait_for_completion(&data->done);
6973 __io_sqe_files_unregister(ctx);
6974 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6975 for (i = 0; i < nr_tables; i++)
6976 kfree(data->table[i].files);
6978 percpu_ref_exit(&data->refs);
6980 ctx->file_data = NULL;
6981 ctx->nr_user_files = 0;
6985 static void io_put_sq_data(struct io_sq_data *sqd)
6987 if (refcount_dec_and_test(&sqd->refs)) {
6989 * The park is a bit of a work-around, without it we get
6990 * warning spews on shutdown with SQPOLL set and affinity
6991 * set to a single CPU.
6994 kthread_park(sqd->thread);
6995 kthread_stop(sqd->thread);
7002 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7004 struct io_ring_ctx *ctx_attach;
7005 struct io_sq_data *sqd;
7008 f = fdget(p->wq_fd);
7010 return ERR_PTR(-ENXIO);
7011 if (f.file->f_op != &io_uring_fops) {
7013 return ERR_PTR(-EINVAL);
7016 ctx_attach = f.file->private_data;
7017 sqd = ctx_attach->sq_data;
7020 return ERR_PTR(-EINVAL);
7023 refcount_inc(&sqd->refs);
7028 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7030 struct io_sq_data *sqd;
7032 if (p->flags & IORING_SETUP_ATTACH_WQ)
7033 return io_attach_sq_data(p);
7035 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7037 return ERR_PTR(-ENOMEM);
7039 refcount_set(&sqd->refs, 1);
7040 INIT_LIST_HEAD(&sqd->ctx_list);
7041 INIT_LIST_HEAD(&sqd->ctx_new_list);
7042 mutex_init(&sqd->ctx_lock);
7043 mutex_init(&sqd->lock);
7044 init_waitqueue_head(&sqd->wait);
7048 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7049 __releases(&sqd->lock)
7053 kthread_unpark(sqd->thread);
7054 mutex_unlock(&sqd->lock);
7057 static void io_sq_thread_park(struct io_sq_data *sqd)
7058 __acquires(&sqd->lock)
7062 mutex_lock(&sqd->lock);
7063 kthread_park(sqd->thread);
7066 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7068 struct io_sq_data *sqd = ctx->sq_data;
7073 * We may arrive here from the error branch in
7074 * io_sq_offload_create() where the kthread is created
7075 * without being waked up, thus wake it up now to make
7076 * sure the wait will complete.
7078 wake_up_process(sqd->thread);
7079 wait_for_completion(&ctx->sq_thread_comp);
7081 io_sq_thread_park(sqd);
7084 mutex_lock(&sqd->ctx_lock);
7085 list_del(&ctx->sqd_list);
7086 mutex_unlock(&sqd->ctx_lock);
7089 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7090 io_sq_thread_unpark(sqd);
7093 io_put_sq_data(sqd);
7094 ctx->sq_data = NULL;
7098 static void io_finish_async(struct io_ring_ctx *ctx)
7100 io_sq_thread_stop(ctx);
7103 io_wq_destroy(ctx->io_wq);
7108 #if defined(CONFIG_UNIX)
7110 * Ensure the UNIX gc is aware of our file set, so we are certain that
7111 * the io_uring can be safely unregistered on process exit, even if we have
7112 * loops in the file referencing.
7114 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7116 struct sock *sk = ctx->ring_sock->sk;
7117 struct scm_fp_list *fpl;
7118 struct sk_buff *skb;
7121 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7125 skb = alloc_skb(0, GFP_KERNEL);
7134 fpl->user = get_uid(ctx->user);
7135 for (i = 0; i < nr; i++) {
7136 struct file *file = io_file_from_index(ctx, i + offset);
7140 fpl->fp[nr_files] = get_file(file);
7141 unix_inflight(fpl->user, fpl->fp[nr_files]);
7146 fpl->max = SCM_MAX_FD;
7147 fpl->count = nr_files;
7148 UNIXCB(skb).fp = fpl;
7149 skb->destructor = unix_destruct_scm;
7150 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7151 skb_queue_head(&sk->sk_receive_queue, skb);
7153 for (i = 0; i < nr_files; i++)
7164 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7165 * causes regular reference counting to break down. We rely on the UNIX
7166 * garbage collection to take care of this problem for us.
7168 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7170 unsigned left, total;
7174 left = ctx->nr_user_files;
7176 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7178 ret = __io_sqe_files_scm(ctx, this_files, total);
7182 total += this_files;
7188 while (total < ctx->nr_user_files) {
7189 struct file *file = io_file_from_index(ctx, total);
7199 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7205 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7206 unsigned nr_tables, unsigned nr_files)
7210 for (i = 0; i < nr_tables; i++) {
7211 struct fixed_file_table *table = &file_data->table[i];
7212 unsigned this_files;
7214 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7215 table->files = kcalloc(this_files, sizeof(struct file *),
7219 nr_files -= this_files;
7225 for (i = 0; i < nr_tables; i++) {
7226 struct fixed_file_table *table = &file_data->table[i];
7227 kfree(table->files);
7232 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7234 #if defined(CONFIG_UNIX)
7235 struct sock *sock = ctx->ring_sock->sk;
7236 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7237 struct sk_buff *skb;
7240 __skb_queue_head_init(&list);
7243 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7244 * remove this entry and rearrange the file array.
7246 skb = skb_dequeue(head);
7248 struct scm_fp_list *fp;
7250 fp = UNIXCB(skb).fp;
7251 for (i = 0; i < fp->count; i++) {
7254 if (fp->fp[i] != file)
7257 unix_notinflight(fp->user, fp->fp[i]);
7258 left = fp->count - 1 - i;
7260 memmove(&fp->fp[i], &fp->fp[i + 1],
7261 left * sizeof(struct file *));
7268 __skb_queue_tail(&list, skb);
7278 __skb_queue_tail(&list, skb);
7280 skb = skb_dequeue(head);
7283 if (skb_peek(&list)) {
7284 spin_lock_irq(&head->lock);
7285 while ((skb = __skb_dequeue(&list)) != NULL)
7286 __skb_queue_tail(head, skb);
7287 spin_unlock_irq(&head->lock);
7294 struct io_file_put {
7295 struct list_head list;
7299 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7301 struct fixed_file_data *file_data = ref_node->file_data;
7302 struct io_ring_ctx *ctx = file_data->ctx;
7303 struct io_file_put *pfile, *tmp;
7305 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7306 list_del(&pfile->list);
7307 io_ring_file_put(ctx, pfile->file);
7311 spin_lock(&file_data->lock);
7312 list_del(&ref_node->node);
7313 spin_unlock(&file_data->lock);
7315 percpu_ref_exit(&ref_node->refs);
7317 percpu_ref_put(&file_data->refs);
7320 static void io_file_put_work(struct work_struct *work)
7322 struct io_ring_ctx *ctx;
7323 struct llist_node *node;
7325 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7326 node = llist_del_all(&ctx->file_put_llist);
7329 struct fixed_file_ref_node *ref_node;
7330 struct llist_node *next = node->next;
7332 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7333 __io_file_put_work(ref_node);
7338 static void io_file_data_ref_zero(struct percpu_ref *ref)
7340 struct fixed_file_ref_node *ref_node;
7341 struct io_ring_ctx *ctx;
7345 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7346 ctx = ref_node->file_data->ctx;
7348 if (percpu_ref_is_dying(&ctx->file_data->refs))
7351 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7353 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7355 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7358 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7359 struct io_ring_ctx *ctx)
7361 struct fixed_file_ref_node *ref_node;
7363 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7365 return ERR_PTR(-ENOMEM);
7367 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7370 return ERR_PTR(-ENOMEM);
7372 INIT_LIST_HEAD(&ref_node->node);
7373 INIT_LIST_HEAD(&ref_node->file_list);
7374 ref_node->file_data = ctx->file_data;
7378 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7380 percpu_ref_exit(&ref_node->refs);
7384 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7387 __s32 __user *fds = (__s32 __user *) arg;
7388 unsigned nr_tables, i;
7390 int fd, ret = -ENOMEM;
7391 struct fixed_file_ref_node *ref_node;
7392 struct fixed_file_data *file_data;
7398 if (nr_args > IORING_MAX_FIXED_FILES)
7401 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7404 file_data->ctx = ctx;
7405 init_completion(&file_data->done);
7406 INIT_LIST_HEAD(&file_data->ref_list);
7407 spin_lock_init(&file_data->lock);
7409 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7410 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7412 if (!file_data->table)
7415 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7416 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7419 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7421 ctx->file_data = file_data;
7423 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7424 struct fixed_file_table *table;
7427 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7431 /* allow sparse sets */
7441 * Don't allow io_uring instances to be registered. If UNIX
7442 * isn't enabled, then this causes a reference cycle and this
7443 * instance can never get freed. If UNIX is enabled we'll
7444 * handle it just fine, but there's still no point in allowing
7445 * a ring fd as it doesn't support regular read/write anyway.
7447 if (file->f_op == &io_uring_fops) {
7451 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7452 index = i & IORING_FILE_TABLE_MASK;
7453 table->files[index] = file;
7456 ret = io_sqe_files_scm(ctx);
7458 io_sqe_files_unregister(ctx);
7462 ref_node = alloc_fixed_file_ref_node(ctx);
7463 if (IS_ERR(ref_node)) {
7464 io_sqe_files_unregister(ctx);
7465 return PTR_ERR(ref_node);
7468 file_data->node = ref_node;
7469 spin_lock(&file_data->lock);
7470 list_add(&ref_node->node, &file_data->ref_list);
7471 spin_unlock(&file_data->lock);
7472 percpu_ref_get(&file_data->refs);
7475 for (i = 0; i < ctx->nr_user_files; i++) {
7476 file = io_file_from_index(ctx, i);
7480 for (i = 0; i < nr_tables; i++)
7481 kfree(file_data->table[i].files);
7482 ctx->nr_user_files = 0;
7484 percpu_ref_exit(&file_data->refs);
7486 kfree(file_data->table);
7488 ctx->file_data = NULL;
7492 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7495 #if defined(CONFIG_UNIX)
7496 struct sock *sock = ctx->ring_sock->sk;
7497 struct sk_buff_head *head = &sock->sk_receive_queue;
7498 struct sk_buff *skb;
7501 * See if we can merge this file into an existing skb SCM_RIGHTS
7502 * file set. If there's no room, fall back to allocating a new skb
7503 * and filling it in.
7505 spin_lock_irq(&head->lock);
7506 skb = skb_peek(head);
7508 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7510 if (fpl->count < SCM_MAX_FD) {
7511 __skb_unlink(skb, head);
7512 spin_unlock_irq(&head->lock);
7513 fpl->fp[fpl->count] = get_file(file);
7514 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7516 spin_lock_irq(&head->lock);
7517 __skb_queue_head(head, skb);
7522 spin_unlock_irq(&head->lock);
7529 return __io_sqe_files_scm(ctx, 1, index);
7535 static int io_queue_file_removal(struct fixed_file_data *data,
7538 struct io_file_put *pfile;
7539 struct fixed_file_ref_node *ref_node = data->node;
7541 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7546 list_add(&pfile->list, &ref_node->file_list);
7551 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7552 struct io_uring_files_update *up,
7555 struct fixed_file_data *data = ctx->file_data;
7556 struct fixed_file_ref_node *ref_node;
7561 bool needs_switch = false;
7563 if (check_add_overflow(up->offset, nr_args, &done))
7565 if (done > ctx->nr_user_files)
7568 ref_node = alloc_fixed_file_ref_node(ctx);
7569 if (IS_ERR(ref_node))
7570 return PTR_ERR(ref_node);
7573 fds = u64_to_user_ptr(up->fds);
7575 struct fixed_file_table *table;
7579 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7583 i = array_index_nospec(up->offset, ctx->nr_user_files);
7584 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7585 index = i & IORING_FILE_TABLE_MASK;
7586 if (table->files[index]) {
7587 file = table->files[index];
7588 err = io_queue_file_removal(data, file);
7591 table->files[index] = NULL;
7592 needs_switch = true;
7601 * Don't allow io_uring instances to be registered. If
7602 * UNIX isn't enabled, then this causes a reference
7603 * cycle and this instance can never get freed. If UNIX
7604 * is enabled we'll handle it just fine, but there's
7605 * still no point in allowing a ring fd as it doesn't
7606 * support regular read/write anyway.
7608 if (file->f_op == &io_uring_fops) {
7613 table->files[index] = file;
7614 err = io_sqe_file_register(ctx, file, i);
7616 table->files[index] = NULL;
7627 percpu_ref_kill(&data->node->refs);
7628 spin_lock(&data->lock);
7629 list_add(&ref_node->node, &data->ref_list);
7630 data->node = ref_node;
7631 spin_unlock(&data->lock);
7632 percpu_ref_get(&ctx->file_data->refs);
7634 destroy_fixed_file_ref_node(ref_node);
7636 return done ? done : err;
7639 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7642 struct io_uring_files_update up;
7644 if (!ctx->file_data)
7648 if (copy_from_user(&up, arg, sizeof(up)))
7653 return __io_sqe_files_update(ctx, &up, nr_args);
7656 static void io_free_work(struct io_wq_work *work)
7658 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7660 /* Consider that io_steal_work() relies on this ref */
7664 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7665 struct io_uring_params *p)
7667 struct io_wq_data data;
7669 struct io_ring_ctx *ctx_attach;
7670 unsigned int concurrency;
7673 data.user = ctx->user;
7674 data.free_work = io_free_work;
7675 data.do_work = io_wq_submit_work;
7677 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7678 /* Do QD, or 4 * CPUS, whatever is smallest */
7679 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7681 ctx->io_wq = io_wq_create(concurrency, &data);
7682 if (IS_ERR(ctx->io_wq)) {
7683 ret = PTR_ERR(ctx->io_wq);
7689 f = fdget(p->wq_fd);
7693 if (f.file->f_op != &io_uring_fops) {
7698 ctx_attach = f.file->private_data;
7699 /* @io_wq is protected by holding the fd */
7700 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7705 ctx->io_wq = ctx_attach->io_wq;
7711 static int io_uring_alloc_task_context(struct task_struct *task)
7713 struct io_uring_task *tctx;
7716 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7717 if (unlikely(!tctx))
7720 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7721 if (unlikely(ret)) {
7727 init_waitqueue_head(&tctx->wait);
7730 io_init_identity(&tctx->__identity);
7731 tctx->identity = &tctx->__identity;
7732 task->io_uring = tctx;
7736 void __io_uring_free(struct task_struct *tsk)
7738 struct io_uring_task *tctx = tsk->io_uring;
7740 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7741 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7742 if (tctx->identity != &tctx->__identity)
7743 kfree(tctx->identity);
7744 percpu_counter_destroy(&tctx->inflight);
7746 tsk->io_uring = NULL;
7749 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7750 struct io_uring_params *p)
7754 if (ctx->flags & IORING_SETUP_SQPOLL) {
7755 struct io_sq_data *sqd;
7758 if (!capable(CAP_SYS_ADMIN))
7761 sqd = io_get_sq_data(p);
7768 io_sq_thread_park(sqd);
7769 mutex_lock(&sqd->ctx_lock);
7770 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7771 mutex_unlock(&sqd->ctx_lock);
7772 io_sq_thread_unpark(sqd);
7774 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7775 if (!ctx->sq_thread_idle)
7776 ctx->sq_thread_idle = HZ;
7781 if (p->flags & IORING_SETUP_SQ_AFF) {
7782 int cpu = p->sq_thread_cpu;
7785 if (cpu >= nr_cpu_ids)
7787 if (!cpu_online(cpu))
7790 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7791 cpu, "io_uring-sq");
7793 sqd->thread = kthread_create(io_sq_thread, sqd,
7796 if (IS_ERR(sqd->thread)) {
7797 ret = PTR_ERR(sqd->thread);
7801 ret = io_uring_alloc_task_context(sqd->thread);
7804 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7805 /* Can't have SQ_AFF without SQPOLL */
7811 ret = io_init_wq_offload(ctx, p);
7817 io_finish_async(ctx);
7821 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7823 struct io_sq_data *sqd = ctx->sq_data;
7825 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7826 wake_up_process(sqd->thread);
7829 static inline void __io_unaccount_mem(struct user_struct *user,
7830 unsigned long nr_pages)
7832 atomic_long_sub(nr_pages, &user->locked_vm);
7835 static inline int __io_account_mem(struct user_struct *user,
7836 unsigned long nr_pages)
7838 unsigned long page_limit, cur_pages, new_pages;
7840 /* Don't allow more pages than we can safely lock */
7841 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7844 cur_pages = atomic_long_read(&user->locked_vm);
7845 new_pages = cur_pages + nr_pages;
7846 if (new_pages > page_limit)
7848 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7849 new_pages) != cur_pages);
7854 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7855 enum io_mem_account acct)
7858 __io_unaccount_mem(ctx->user, nr_pages);
7860 if (ctx->mm_account) {
7861 if (acct == ACCT_LOCKED)
7862 ctx->mm_account->locked_vm -= nr_pages;
7863 else if (acct == ACCT_PINNED)
7864 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7868 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7869 enum io_mem_account acct)
7873 if (ctx->limit_mem) {
7874 ret = __io_account_mem(ctx->user, nr_pages);
7879 if (ctx->mm_account) {
7880 if (acct == ACCT_LOCKED)
7881 ctx->mm_account->locked_vm += nr_pages;
7882 else if (acct == ACCT_PINNED)
7883 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7889 static void io_mem_free(void *ptr)
7896 page = virt_to_head_page(ptr);
7897 if (put_page_testzero(page))
7898 free_compound_page(page);
7901 static void *io_mem_alloc(size_t size)
7903 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7906 return (void *) __get_free_pages(gfp_flags, get_order(size));
7909 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7912 struct io_rings *rings;
7913 size_t off, sq_array_size;
7915 off = struct_size(rings, cqes, cq_entries);
7916 if (off == SIZE_MAX)
7920 off = ALIGN(off, SMP_CACHE_BYTES);
7928 sq_array_size = array_size(sizeof(u32), sq_entries);
7929 if (sq_array_size == SIZE_MAX)
7932 if (check_add_overflow(off, sq_array_size, &off))
7938 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7942 pages = (size_t)1 << get_order(
7943 rings_size(sq_entries, cq_entries, NULL));
7944 pages += (size_t)1 << get_order(
7945 array_size(sizeof(struct io_uring_sqe), sq_entries));
7950 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7954 if (!ctx->user_bufs)
7957 for (i = 0; i < ctx->nr_user_bufs; i++) {
7958 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7960 for (j = 0; j < imu->nr_bvecs; j++)
7961 unpin_user_page(imu->bvec[j].bv_page);
7963 if (imu->acct_pages)
7964 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7969 kfree(ctx->user_bufs);
7970 ctx->user_bufs = NULL;
7971 ctx->nr_user_bufs = 0;
7975 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7976 void __user *arg, unsigned index)
7978 struct iovec __user *src;
7980 #ifdef CONFIG_COMPAT
7982 struct compat_iovec __user *ciovs;
7983 struct compat_iovec ciov;
7985 ciovs = (struct compat_iovec __user *) arg;
7986 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7989 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7990 dst->iov_len = ciov.iov_len;
7994 src = (struct iovec __user *) arg;
7995 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8001 * Not super efficient, but this is just a registration time. And we do cache
8002 * the last compound head, so generally we'll only do a full search if we don't
8005 * We check if the given compound head page has already been accounted, to
8006 * avoid double accounting it. This allows us to account the full size of the
8007 * page, not just the constituent pages of a huge page.
8009 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8010 int nr_pages, struct page *hpage)
8014 /* check current page array */
8015 for (i = 0; i < nr_pages; i++) {
8016 if (!PageCompound(pages[i]))
8018 if (compound_head(pages[i]) == hpage)
8022 /* check previously registered pages */
8023 for (i = 0; i < ctx->nr_user_bufs; i++) {
8024 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8026 for (j = 0; j < imu->nr_bvecs; j++) {
8027 if (!PageCompound(imu->bvec[j].bv_page))
8029 if (compound_head(imu->bvec[j].bv_page) == hpage)
8037 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8038 int nr_pages, struct io_mapped_ubuf *imu,
8039 struct page **last_hpage)
8043 for (i = 0; i < nr_pages; i++) {
8044 if (!PageCompound(pages[i])) {
8049 hpage = compound_head(pages[i]);
8050 if (hpage == *last_hpage)
8052 *last_hpage = hpage;
8053 if (headpage_already_acct(ctx, pages, i, hpage))
8055 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8059 if (!imu->acct_pages)
8062 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8064 imu->acct_pages = 0;
8068 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8071 struct vm_area_struct **vmas = NULL;
8072 struct page **pages = NULL;
8073 struct page *last_hpage = NULL;
8074 int i, j, got_pages = 0;
8079 if (!nr_args || nr_args > UIO_MAXIOV)
8082 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8084 if (!ctx->user_bufs)
8087 for (i = 0; i < nr_args; i++) {
8088 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8089 unsigned long off, start, end, ubuf;
8094 ret = io_copy_iov(ctx, &iov, arg, i);
8099 * Don't impose further limits on the size and buffer
8100 * constraints here, we'll -EINVAL later when IO is
8101 * submitted if they are wrong.
8104 if (!iov.iov_base || !iov.iov_len)
8107 /* arbitrary limit, but we need something */
8108 if (iov.iov_len > SZ_1G)
8111 ubuf = (unsigned long) iov.iov_base;
8112 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8113 start = ubuf >> PAGE_SHIFT;
8114 nr_pages = end - start;
8117 if (!pages || nr_pages > got_pages) {
8120 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8122 vmas = kvmalloc_array(nr_pages,
8123 sizeof(struct vm_area_struct *),
8125 if (!pages || !vmas) {
8129 got_pages = nr_pages;
8132 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8139 mmap_read_lock(current->mm);
8140 pret = pin_user_pages(ubuf, nr_pages,
8141 FOLL_WRITE | FOLL_LONGTERM,
8143 if (pret == nr_pages) {
8144 /* don't support file backed memory */
8145 for (j = 0; j < nr_pages; j++) {
8146 struct vm_area_struct *vma = vmas[j];
8149 !is_file_hugepages(vma->vm_file)) {
8155 ret = pret < 0 ? pret : -EFAULT;
8157 mmap_read_unlock(current->mm);
8160 * if we did partial map, or found file backed vmas,
8161 * release any pages we did get
8164 unpin_user_pages(pages, pret);
8169 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8171 unpin_user_pages(pages, pret);
8176 off = ubuf & ~PAGE_MASK;
8178 for (j = 0; j < nr_pages; j++) {
8181 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8182 imu->bvec[j].bv_page = pages[j];
8183 imu->bvec[j].bv_len = vec_len;
8184 imu->bvec[j].bv_offset = off;
8188 /* store original address for later verification */
8190 imu->len = iov.iov_len;
8191 imu->nr_bvecs = nr_pages;
8193 ctx->nr_user_bufs++;
8201 io_sqe_buffer_unregister(ctx);
8205 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8207 __s32 __user *fds = arg;
8213 if (copy_from_user(&fd, fds, sizeof(*fds)))
8216 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8217 if (IS_ERR(ctx->cq_ev_fd)) {
8218 int ret = PTR_ERR(ctx->cq_ev_fd);
8219 ctx->cq_ev_fd = NULL;
8226 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8228 if (ctx->cq_ev_fd) {
8229 eventfd_ctx_put(ctx->cq_ev_fd);
8230 ctx->cq_ev_fd = NULL;
8237 static int __io_destroy_buffers(int id, void *p, void *data)
8239 struct io_ring_ctx *ctx = data;
8240 struct io_buffer *buf = p;
8242 __io_remove_buffers(ctx, buf, id, -1U);
8246 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8248 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8249 idr_destroy(&ctx->io_buffer_idr);
8252 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8254 io_finish_async(ctx);
8255 io_sqe_buffer_unregister(ctx);
8257 if (ctx->sqo_task) {
8258 put_task_struct(ctx->sqo_task);
8259 ctx->sqo_task = NULL;
8260 mmdrop(ctx->mm_account);
8261 ctx->mm_account = NULL;
8264 #ifdef CONFIG_BLK_CGROUP
8265 if (ctx->sqo_blkcg_css)
8266 css_put(ctx->sqo_blkcg_css);
8269 io_sqe_files_unregister(ctx);
8270 io_eventfd_unregister(ctx);
8271 io_destroy_buffers(ctx);
8272 idr_destroy(&ctx->personality_idr);
8274 #if defined(CONFIG_UNIX)
8275 if (ctx->ring_sock) {
8276 ctx->ring_sock->file = NULL; /* so that iput() is called */
8277 sock_release(ctx->ring_sock);
8281 io_mem_free(ctx->rings);
8282 io_mem_free(ctx->sq_sqes);
8284 percpu_ref_exit(&ctx->refs);
8285 free_uid(ctx->user);
8286 put_cred(ctx->creds);
8287 kfree(ctx->cancel_hash);
8288 kmem_cache_free(req_cachep, ctx->fallback_req);
8292 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8294 struct io_ring_ctx *ctx = file->private_data;
8297 poll_wait(file, &ctx->cq_wait, wait);
8299 * synchronizes with barrier from wq_has_sleeper call in
8303 if (!io_sqring_full(ctx))
8304 mask |= EPOLLOUT | EPOLLWRNORM;
8305 if (io_cqring_events(ctx, false))
8306 mask |= EPOLLIN | EPOLLRDNORM;
8311 static int io_uring_fasync(int fd, struct file *file, int on)
8313 struct io_ring_ctx *ctx = file->private_data;
8315 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8318 static int io_remove_personalities(int id, void *p, void *data)
8320 struct io_ring_ctx *ctx = data;
8321 struct io_identity *iod;
8323 iod = idr_remove(&ctx->personality_idr, id);
8325 put_cred(iod->creds);
8326 if (refcount_dec_and_test(&iod->count))
8332 static void io_ring_exit_work(struct work_struct *work)
8334 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8338 * If we're doing polled IO and end up having requests being
8339 * submitted async (out-of-line), then completions can come in while
8340 * we're waiting for refs to drop. We need to reap these manually,
8341 * as nobody else will be looking for them.
8345 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8346 io_iopoll_try_reap_events(ctx);
8347 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8348 io_ring_ctx_free(ctx);
8351 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8353 mutex_lock(&ctx->uring_lock);
8354 percpu_ref_kill(&ctx->refs);
8355 mutex_unlock(&ctx->uring_lock);
8357 io_kill_timeouts(ctx, NULL);
8358 io_poll_remove_all(ctx, NULL);
8361 io_wq_cancel_all(ctx->io_wq);
8363 /* if we failed setting up the ctx, we might not have any rings */
8365 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8366 io_iopoll_try_reap_events(ctx);
8367 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8370 * Do this upfront, so we won't have a grace period where the ring
8371 * is closed but resources aren't reaped yet. This can cause
8372 * spurious failure in setting up a new ring.
8374 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8377 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8379 * Use system_unbound_wq to avoid spawning tons of event kworkers
8380 * if we're exiting a ton of rings at the same time. It just adds
8381 * noise and overhead, there's no discernable change in runtime
8382 * over using system_wq.
8384 queue_work(system_unbound_wq, &ctx->exit_work);
8387 static int io_uring_release(struct inode *inode, struct file *file)
8389 struct io_ring_ctx *ctx = file->private_data;
8391 file->private_data = NULL;
8392 io_ring_ctx_wait_and_kill(ctx);
8396 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8398 struct files_struct *files = data;
8400 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8401 work->identity->files == files);
8405 * Returns true if 'preq' is the link parent of 'req'
8407 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8409 struct io_kiocb *link;
8411 if (!(preq->flags & REQ_F_LINK_HEAD))
8414 list_for_each_entry(link, &preq->link_list, link_list) {
8422 static bool io_match_link_files(struct io_kiocb *req,
8423 struct files_struct *files)
8425 struct io_kiocb *link;
8427 if (io_match_files(req, files))
8429 if (req->flags & REQ_F_LINK_HEAD) {
8430 list_for_each_entry(link, &req->link_list, link_list) {
8431 if (io_match_files(link, files))
8439 * We're looking to cancel 'req' because it's holding on to our files, but
8440 * 'req' could be a link to another request. See if it is, and cancel that
8441 * parent request if so.
8443 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8445 struct hlist_node *tmp;
8446 struct io_kiocb *preq;
8450 spin_lock_irq(&ctx->completion_lock);
8451 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8452 struct hlist_head *list;
8454 list = &ctx->cancel_hash[i];
8455 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8456 found = io_match_link(preq, req);
8458 io_poll_remove_one(preq);
8463 spin_unlock_irq(&ctx->completion_lock);
8467 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8468 struct io_kiocb *req)
8470 struct io_kiocb *preq;
8473 spin_lock_irq(&ctx->completion_lock);
8474 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8475 found = io_match_link(preq, req);
8477 __io_timeout_cancel(preq);
8481 spin_unlock_irq(&ctx->completion_lock);
8485 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8487 return io_match_link(container_of(work, struct io_kiocb, work), data);
8490 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8492 enum io_wq_cancel cret;
8494 /* cancel this particular work, if it's running */
8495 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8496 if (cret != IO_WQ_CANCEL_NOTFOUND)
8499 /* find links that hold this pending, cancel those */
8500 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8501 if (cret != IO_WQ_CANCEL_NOTFOUND)
8504 /* if we have a poll link holding this pending, cancel that */
8505 if (io_poll_remove_link(ctx, req))
8508 /* final option, timeout link is holding this req pending */
8509 io_timeout_remove_link(ctx, req);
8512 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8513 struct files_struct *files)
8515 struct io_defer_entry *de = NULL;
8518 spin_lock_irq(&ctx->completion_lock);
8519 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8520 if (io_match_link_files(de->req, files)) {
8521 list_cut_position(&list, &ctx->defer_list, &de->list);
8525 spin_unlock_irq(&ctx->completion_lock);
8527 while (!list_empty(&list)) {
8528 de = list_first_entry(&list, struct io_defer_entry, list);
8529 list_del_init(&de->list);
8530 req_set_fail_links(de->req);
8531 io_put_req(de->req);
8532 io_req_complete(de->req, -ECANCELED);
8538 * Returns true if we found and killed one or more files pinning requests
8540 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8541 struct files_struct *files)
8543 if (list_empty_careful(&ctx->inflight_list))
8546 io_cancel_defer_files(ctx, files);
8547 /* cancel all at once, should be faster than doing it one by one*/
8548 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8550 while (!list_empty_careful(&ctx->inflight_list)) {
8551 struct io_kiocb *cancel_req = NULL, *req;
8554 spin_lock_irq(&ctx->inflight_lock);
8555 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8556 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8557 req->work.identity->files != files)
8559 /* req is being completed, ignore */
8560 if (!refcount_inc_not_zero(&req->refs))
8566 prepare_to_wait(&ctx->inflight_wait, &wait,
8567 TASK_UNINTERRUPTIBLE);
8568 spin_unlock_irq(&ctx->inflight_lock);
8570 /* We need to keep going until we don't find a matching req */
8573 /* cancel this request, or head link requests */
8574 io_attempt_cancel(ctx, cancel_req);
8575 io_put_req(cancel_req);
8576 /* cancellations _may_ trigger task work */
8579 finish_wait(&ctx->inflight_wait, &wait);
8585 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8587 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8588 struct task_struct *task = data;
8590 return io_task_match(req, task);
8593 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8594 struct task_struct *task,
8595 struct files_struct *files)
8599 ret = io_uring_cancel_files(ctx, files);
8601 enum io_wq_cancel cret;
8603 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8604 if (cret != IO_WQ_CANCEL_NOTFOUND)
8607 /* SQPOLL thread does its own polling */
8608 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8609 while (!list_empty_careful(&ctx->iopoll_list)) {
8610 io_iopoll_try_reap_events(ctx);
8615 ret |= io_poll_remove_all(ctx, task);
8616 ret |= io_kill_timeouts(ctx, task);
8623 * We need to iteratively cancel requests, in case a request has dependent
8624 * hard links. These persist even for failure of cancelations, hence keep
8625 * looping until none are found.
8627 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8628 struct files_struct *files)
8630 struct task_struct *task = current;
8632 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8633 task = ctx->sq_data->thread;
8635 io_cqring_overflow_flush(ctx, true, task, files);
8637 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8644 * Note that this task has used io_uring. We use it for cancelation purposes.
8646 static int io_uring_add_task_file(struct file *file)
8648 struct io_uring_task *tctx = current->io_uring;
8650 if (unlikely(!tctx)) {
8653 ret = io_uring_alloc_task_context(current);
8656 tctx = current->io_uring;
8658 if (tctx->last != file) {
8659 void *old = xa_load(&tctx->xa, (unsigned long)file);
8663 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8672 * Remove this io_uring_file -> task mapping.
8674 static void io_uring_del_task_file(struct file *file)
8676 struct io_uring_task *tctx = current->io_uring;
8678 if (tctx->last == file)
8680 file = xa_erase(&tctx->xa, (unsigned long)file);
8685 static void __io_uring_attempt_task_drop(struct file *file)
8687 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8690 io_uring_del_task_file(file);
8694 * Drop task note for this file if we're the only ones that hold it after
8697 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8699 if (!current->io_uring)
8702 * fput() is pending, will be 2 if the only other ref is our potential
8703 * task file note. If the task is exiting, drop regardless of count.
8705 if (!exiting && atomic_long_read(&file->f_count) != 2)
8708 __io_uring_attempt_task_drop(file);
8711 void __io_uring_files_cancel(struct files_struct *files)
8713 struct io_uring_task *tctx = current->io_uring;
8715 unsigned long index;
8717 /* make sure overflow events are dropped */
8718 tctx->in_idle = true;
8720 xa_for_each(&tctx->xa, index, file) {
8721 struct io_ring_ctx *ctx = file->private_data;
8723 io_uring_cancel_task_requests(ctx, files);
8725 io_uring_del_task_file(file);
8730 * Find any io_uring fd that this task has registered or done IO on, and cancel
8733 void __io_uring_task_cancel(void)
8735 struct io_uring_task *tctx = current->io_uring;
8739 /* make sure overflow events are dropped */
8740 tctx->in_idle = true;
8743 /* read completions before cancelations */
8744 inflight = percpu_counter_sum(&tctx->inflight);
8747 __io_uring_files_cancel(NULL);
8749 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8752 * If we've seen completions, retry. This avoids a race where
8753 * a completion comes in before we did prepare_to_wait().
8755 if (inflight != percpu_counter_sum(&tctx->inflight))
8760 finish_wait(&tctx->wait, &wait);
8761 tctx->in_idle = false;
8764 static int io_uring_flush(struct file *file, void *data)
8766 struct io_ring_ctx *ctx = file->private_data;
8769 * If the task is going away, cancel work it may have pending
8771 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8774 io_uring_cancel_task_requests(ctx, data);
8775 io_uring_attempt_task_drop(file, !data);
8779 static void *io_uring_validate_mmap_request(struct file *file,
8780 loff_t pgoff, size_t sz)
8782 struct io_ring_ctx *ctx = file->private_data;
8783 loff_t offset = pgoff << PAGE_SHIFT;
8788 case IORING_OFF_SQ_RING:
8789 case IORING_OFF_CQ_RING:
8792 case IORING_OFF_SQES:
8796 return ERR_PTR(-EINVAL);
8799 page = virt_to_head_page(ptr);
8800 if (sz > page_size(page))
8801 return ERR_PTR(-EINVAL);
8808 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8810 size_t sz = vma->vm_end - vma->vm_start;
8814 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8816 return PTR_ERR(ptr);
8818 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8819 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8822 #else /* !CONFIG_MMU */
8824 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8826 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8829 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8831 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8834 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8835 unsigned long addr, unsigned long len,
8836 unsigned long pgoff, unsigned long flags)
8840 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8842 return PTR_ERR(ptr);
8844 return (unsigned long) ptr;
8847 #endif /* !CONFIG_MMU */
8849 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8854 if (!io_sqring_full(ctx))
8857 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8859 if (!io_sqring_full(ctx))
8863 } while (!signal_pending(current));
8865 finish_wait(&ctx->sqo_sq_wait, &wait);
8868 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8869 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8872 struct io_ring_ctx *ctx;
8879 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8880 IORING_ENTER_SQ_WAIT))
8888 if (f.file->f_op != &io_uring_fops)
8892 ctx = f.file->private_data;
8893 if (!percpu_ref_tryget(&ctx->refs))
8897 if (ctx->flags & IORING_SETUP_R_DISABLED)
8901 * For SQ polling, the thread will do all submissions and completions.
8902 * Just return the requested submit count, and wake the thread if
8906 if (ctx->flags & IORING_SETUP_SQPOLL) {
8907 if (!list_empty_careful(&ctx->cq_overflow_list))
8908 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8909 if (flags & IORING_ENTER_SQ_WAKEUP)
8910 wake_up(&ctx->sq_data->wait);
8911 if (flags & IORING_ENTER_SQ_WAIT)
8912 io_sqpoll_wait_sq(ctx);
8913 submitted = to_submit;
8914 } else if (to_submit) {
8915 ret = io_uring_add_task_file(f.file);
8918 mutex_lock(&ctx->uring_lock);
8919 submitted = io_submit_sqes(ctx, to_submit);
8920 mutex_unlock(&ctx->uring_lock);
8922 if (submitted != to_submit)
8925 if (flags & IORING_ENTER_GETEVENTS) {
8926 min_complete = min(min_complete, ctx->cq_entries);
8929 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8930 * space applications don't need to do io completion events
8931 * polling again, they can rely on io_sq_thread to do polling
8932 * work, which can reduce cpu usage and uring_lock contention.
8934 if (ctx->flags & IORING_SETUP_IOPOLL &&
8935 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8936 ret = io_iopoll_check(ctx, min_complete);
8938 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8943 percpu_ref_put(&ctx->refs);
8946 return submitted ? submitted : ret;
8949 #ifdef CONFIG_PROC_FS
8950 static int io_uring_show_cred(int id, void *p, void *data)
8952 const struct cred *cred = p;
8953 struct seq_file *m = data;
8954 struct user_namespace *uns = seq_user_ns(m);
8955 struct group_info *gi;
8960 seq_printf(m, "%5d\n", id);
8961 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8962 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8963 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8964 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8965 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8966 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8967 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8968 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8969 seq_puts(m, "\n\tGroups:\t");
8970 gi = cred->group_info;
8971 for (g = 0; g < gi->ngroups; g++) {
8972 seq_put_decimal_ull(m, g ? " " : "",
8973 from_kgid_munged(uns, gi->gid[g]));
8975 seq_puts(m, "\n\tCapEff:\t");
8976 cap = cred->cap_effective;
8977 CAP_FOR_EACH_U32(__capi)
8978 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8983 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8985 struct io_sq_data *sq = NULL;
8990 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8991 * since fdinfo case grabs it in the opposite direction of normal use
8992 * cases. If we fail to get the lock, we just don't iterate any
8993 * structures that could be going away outside the io_uring mutex.
8995 has_lock = mutex_trylock(&ctx->uring_lock);
8997 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9000 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9001 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9002 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9003 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9004 struct fixed_file_table *table;
9007 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9008 f = table->files[i & IORING_FILE_TABLE_MASK];
9010 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9012 seq_printf(m, "%5u: <none>\n", i);
9014 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9015 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9016 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9018 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9019 (unsigned int) buf->len);
9021 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9022 seq_printf(m, "Personalities:\n");
9023 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9025 seq_printf(m, "PollList:\n");
9026 spin_lock_irq(&ctx->completion_lock);
9027 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9028 struct hlist_head *list = &ctx->cancel_hash[i];
9029 struct io_kiocb *req;
9031 hlist_for_each_entry(req, list, hash_node)
9032 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9033 req->task->task_works != NULL);
9035 spin_unlock_irq(&ctx->completion_lock);
9037 mutex_unlock(&ctx->uring_lock);
9040 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9042 struct io_ring_ctx *ctx = f->private_data;
9044 if (percpu_ref_tryget(&ctx->refs)) {
9045 __io_uring_show_fdinfo(ctx, m);
9046 percpu_ref_put(&ctx->refs);
9051 static const struct file_operations io_uring_fops = {
9052 .release = io_uring_release,
9053 .flush = io_uring_flush,
9054 .mmap = io_uring_mmap,
9056 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9057 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9059 .poll = io_uring_poll,
9060 .fasync = io_uring_fasync,
9061 #ifdef CONFIG_PROC_FS
9062 .show_fdinfo = io_uring_show_fdinfo,
9066 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9067 struct io_uring_params *p)
9069 struct io_rings *rings;
9070 size_t size, sq_array_offset;
9072 /* make sure these are sane, as we already accounted them */
9073 ctx->sq_entries = p->sq_entries;
9074 ctx->cq_entries = p->cq_entries;
9076 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9077 if (size == SIZE_MAX)
9080 rings = io_mem_alloc(size);
9085 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9086 rings->sq_ring_mask = p->sq_entries - 1;
9087 rings->cq_ring_mask = p->cq_entries - 1;
9088 rings->sq_ring_entries = p->sq_entries;
9089 rings->cq_ring_entries = p->cq_entries;
9090 ctx->sq_mask = rings->sq_ring_mask;
9091 ctx->cq_mask = rings->cq_ring_mask;
9093 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9094 if (size == SIZE_MAX) {
9095 io_mem_free(ctx->rings);
9100 ctx->sq_sqes = io_mem_alloc(size);
9101 if (!ctx->sq_sqes) {
9102 io_mem_free(ctx->rings);
9111 * Allocate an anonymous fd, this is what constitutes the application
9112 * visible backing of an io_uring instance. The application mmaps this
9113 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9114 * we have to tie this fd to a socket for file garbage collection purposes.
9116 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9121 #if defined(CONFIG_UNIX)
9122 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9128 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9132 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9133 O_RDWR | O_CLOEXEC);
9137 ret = PTR_ERR(file);
9141 #if defined(CONFIG_UNIX)
9142 ctx->ring_sock->file = file;
9144 if (unlikely(io_uring_add_task_file(file))) {
9145 file = ERR_PTR(-ENOMEM);
9148 fd_install(ret, file);
9151 #if defined(CONFIG_UNIX)
9152 sock_release(ctx->ring_sock);
9153 ctx->ring_sock = NULL;
9158 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9159 struct io_uring_params __user *params)
9161 struct user_struct *user = NULL;
9162 struct io_ring_ctx *ctx;
9168 if (entries > IORING_MAX_ENTRIES) {
9169 if (!(p->flags & IORING_SETUP_CLAMP))
9171 entries = IORING_MAX_ENTRIES;
9175 * Use twice as many entries for the CQ ring. It's possible for the
9176 * application to drive a higher depth than the size of the SQ ring,
9177 * since the sqes are only used at submission time. This allows for
9178 * some flexibility in overcommitting a bit. If the application has
9179 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9180 * of CQ ring entries manually.
9182 p->sq_entries = roundup_pow_of_two(entries);
9183 if (p->flags & IORING_SETUP_CQSIZE) {
9185 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9186 * to a power-of-two, if it isn't already. We do NOT impose
9187 * any cq vs sq ring sizing.
9189 if (p->cq_entries < p->sq_entries)
9191 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9192 if (!(p->flags & IORING_SETUP_CLAMP))
9194 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9196 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9198 p->cq_entries = 2 * p->sq_entries;
9201 user = get_uid(current_user());
9202 limit_mem = !capable(CAP_IPC_LOCK);
9205 ret = __io_account_mem(user,
9206 ring_pages(p->sq_entries, p->cq_entries));
9213 ctx = io_ring_ctx_alloc(p);
9216 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9221 ctx->compat = in_compat_syscall();
9223 ctx->creds = get_current_cred();
9225 ctx->loginuid = current->loginuid;
9226 ctx->sessionid = current->sessionid;
9228 ctx->sqo_task = get_task_struct(current);
9231 * This is just grabbed for accounting purposes. When a process exits,
9232 * the mm is exited and dropped before the files, hence we need to hang
9233 * on to this mm purely for the purposes of being able to unaccount
9234 * memory (locked/pinned vm). It's not used for anything else.
9236 mmgrab(current->mm);
9237 ctx->mm_account = current->mm;
9239 #ifdef CONFIG_BLK_CGROUP
9241 * The sq thread will belong to the original cgroup it was inited in.
9242 * If the cgroup goes offline (e.g. disabling the io controller), then
9243 * issued bios will be associated with the closest cgroup later in the
9247 ctx->sqo_blkcg_css = blkcg_css();
9248 ret = css_tryget_online(ctx->sqo_blkcg_css);
9251 /* don't init against a dying cgroup, have the user try again */
9252 ctx->sqo_blkcg_css = NULL;
9259 * Account memory _before_ installing the file descriptor. Once
9260 * the descriptor is installed, it can get closed at any time. Also
9261 * do this before hitting the general error path, as ring freeing
9262 * will un-account as well.
9264 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9266 ctx->limit_mem = limit_mem;
9268 ret = io_allocate_scq_urings(ctx, p);
9272 ret = io_sq_offload_create(ctx, p);
9276 if (!(p->flags & IORING_SETUP_R_DISABLED))
9277 io_sq_offload_start(ctx);
9279 memset(&p->sq_off, 0, sizeof(p->sq_off));
9280 p->sq_off.head = offsetof(struct io_rings, sq.head);
9281 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9282 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9283 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9284 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9285 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9286 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9288 memset(&p->cq_off, 0, sizeof(p->cq_off));
9289 p->cq_off.head = offsetof(struct io_rings, cq.head);
9290 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9291 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9292 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9293 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9294 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9295 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9297 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9298 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9299 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9300 IORING_FEAT_POLL_32BITS;
9302 if (copy_to_user(params, p, sizeof(*p))) {
9308 * Install ring fd as the very last thing, so we don't risk someone
9309 * having closed it before we finish setup
9311 ret = io_uring_get_fd(ctx);
9315 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9318 io_ring_ctx_wait_and_kill(ctx);
9323 * Sets up an aio uring context, and returns the fd. Applications asks for a
9324 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9325 * params structure passed in.
9327 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9329 struct io_uring_params p;
9332 if (copy_from_user(&p, params, sizeof(p)))
9334 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9339 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9340 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9341 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9342 IORING_SETUP_R_DISABLED))
9345 return io_uring_create(entries, &p, params);
9348 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9349 struct io_uring_params __user *, params)
9351 return io_uring_setup(entries, params);
9354 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9356 struct io_uring_probe *p;
9360 size = struct_size(p, ops, nr_args);
9361 if (size == SIZE_MAX)
9363 p = kzalloc(size, GFP_KERNEL);
9368 if (copy_from_user(p, arg, size))
9371 if (memchr_inv(p, 0, size))
9374 p->last_op = IORING_OP_LAST - 1;
9375 if (nr_args > IORING_OP_LAST)
9376 nr_args = IORING_OP_LAST;
9378 for (i = 0; i < nr_args; i++) {
9380 if (!io_op_defs[i].not_supported)
9381 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9386 if (copy_to_user(arg, p, size))
9393 static int io_register_personality(struct io_ring_ctx *ctx)
9395 struct io_identity *id;
9398 id = kmalloc(sizeof(*id), GFP_KERNEL);
9402 io_init_identity(id);
9403 id->creds = get_current_cred();
9405 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9407 put_cred(id->creds);
9413 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9415 struct io_identity *iod;
9417 iod = idr_remove(&ctx->personality_idr, id);
9419 put_cred(iod->creds);
9420 if (refcount_dec_and_test(&iod->count))
9428 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9429 unsigned int nr_args)
9431 struct io_uring_restriction *res;
9435 /* Restrictions allowed only if rings started disabled */
9436 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9439 /* We allow only a single restrictions registration */
9440 if (ctx->restrictions.registered)
9443 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9446 size = array_size(nr_args, sizeof(*res));
9447 if (size == SIZE_MAX)
9450 res = memdup_user(arg, size);
9452 return PTR_ERR(res);
9456 for (i = 0; i < nr_args; i++) {
9457 switch (res[i].opcode) {
9458 case IORING_RESTRICTION_REGISTER_OP:
9459 if (res[i].register_op >= IORING_REGISTER_LAST) {
9464 __set_bit(res[i].register_op,
9465 ctx->restrictions.register_op);
9467 case IORING_RESTRICTION_SQE_OP:
9468 if (res[i].sqe_op >= IORING_OP_LAST) {
9473 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9475 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9476 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9478 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9479 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9488 /* Reset all restrictions if an error happened */
9490 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9492 ctx->restrictions.registered = true;
9498 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9500 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9503 if (ctx->restrictions.registered)
9504 ctx->restricted = 1;
9506 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9508 io_sq_offload_start(ctx);
9513 static bool io_register_op_must_quiesce(int op)
9516 case IORING_UNREGISTER_FILES:
9517 case IORING_REGISTER_FILES_UPDATE:
9518 case IORING_REGISTER_PROBE:
9519 case IORING_REGISTER_PERSONALITY:
9520 case IORING_UNREGISTER_PERSONALITY:
9527 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9528 void __user *arg, unsigned nr_args)
9529 __releases(ctx->uring_lock)
9530 __acquires(ctx->uring_lock)
9535 * We're inside the ring mutex, if the ref is already dying, then
9536 * someone else killed the ctx or is already going through
9537 * io_uring_register().
9539 if (percpu_ref_is_dying(&ctx->refs))
9542 if (io_register_op_must_quiesce(opcode)) {
9543 percpu_ref_kill(&ctx->refs);
9546 * Drop uring mutex before waiting for references to exit. If
9547 * another thread is currently inside io_uring_enter() it might
9548 * need to grab the uring_lock to make progress. If we hold it
9549 * here across the drain wait, then we can deadlock. It's safe
9550 * to drop the mutex here, since no new references will come in
9551 * after we've killed the percpu ref.
9553 mutex_unlock(&ctx->uring_lock);
9555 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9558 ret = io_run_task_work_sig();
9563 mutex_lock(&ctx->uring_lock);
9566 percpu_ref_resurrect(&ctx->refs);
9571 if (ctx->restricted) {
9572 if (opcode >= IORING_REGISTER_LAST) {
9577 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9584 case IORING_REGISTER_BUFFERS:
9585 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9587 case IORING_UNREGISTER_BUFFERS:
9591 ret = io_sqe_buffer_unregister(ctx);
9593 case IORING_REGISTER_FILES:
9594 ret = io_sqe_files_register(ctx, arg, nr_args);
9596 case IORING_UNREGISTER_FILES:
9600 ret = io_sqe_files_unregister(ctx);
9602 case IORING_REGISTER_FILES_UPDATE:
9603 ret = io_sqe_files_update(ctx, arg, nr_args);
9605 case IORING_REGISTER_EVENTFD:
9606 case IORING_REGISTER_EVENTFD_ASYNC:
9610 ret = io_eventfd_register(ctx, arg);
9613 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9614 ctx->eventfd_async = 1;
9616 ctx->eventfd_async = 0;
9618 case IORING_UNREGISTER_EVENTFD:
9622 ret = io_eventfd_unregister(ctx);
9624 case IORING_REGISTER_PROBE:
9626 if (!arg || nr_args > 256)
9628 ret = io_probe(ctx, arg, nr_args);
9630 case IORING_REGISTER_PERSONALITY:
9634 ret = io_register_personality(ctx);
9636 case IORING_UNREGISTER_PERSONALITY:
9640 ret = io_unregister_personality(ctx, nr_args);
9642 case IORING_REGISTER_ENABLE_RINGS:
9646 ret = io_register_enable_rings(ctx);
9648 case IORING_REGISTER_RESTRICTIONS:
9649 ret = io_register_restrictions(ctx, arg, nr_args);
9657 if (io_register_op_must_quiesce(opcode)) {
9658 /* bring the ctx back to life */
9659 percpu_ref_reinit(&ctx->refs);
9661 reinit_completion(&ctx->ref_comp);
9666 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9667 void __user *, arg, unsigned int, nr_args)
9669 struct io_ring_ctx *ctx;
9678 if (f.file->f_op != &io_uring_fops)
9681 ctx = f.file->private_data;
9683 mutex_lock(&ctx->uring_lock);
9684 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9685 mutex_unlock(&ctx->uring_lock);
9686 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9687 ctx->cq_ev_fd != NULL, ret);
9693 static int __init io_uring_init(void)
9695 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9696 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9697 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9700 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9701 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9702 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9703 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9704 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9705 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9706 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9707 BUILD_BUG_SQE_ELEM(8, __u64, off);
9708 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9709 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9710 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9711 BUILD_BUG_SQE_ELEM(24, __u32, len);
9712 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9713 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9714 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9715 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9716 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9717 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9718 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9719 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9720 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9721 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9722 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9723 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9724 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9725 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9726 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9727 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9728 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9729 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9730 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9732 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9733 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9734 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9737 __initcall(io_uring_init);