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,
588 REQ_F_LTIMEOUT_ACTIVE_BIT,
590 /* not a real bit, just to check we're not overflowing the space */
596 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
597 /* drain existing IO first */
598 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
600 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
601 /* doesn't sever on completion < 0 */
602 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
604 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
605 /* IOSQE_BUFFER_SELECT */
606 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
609 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
610 /* fail rest of links */
611 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
612 /* on inflight list */
613 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
614 /* read/write uses file position */
615 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
616 /* must not punt to workers */
617 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
618 /* has or had linked timeout */
619 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
621 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
623 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
624 /* already went through poll handler */
625 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
626 /* buffer already selected */
627 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
628 /* doesn't need file table for this request */
629 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
630 /* io_wq_work is initialized */
631 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
632 /* linked timeout is active, i.e. prepared by link's head */
633 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
637 struct io_poll_iocb poll;
638 struct io_poll_iocb *double_poll;
642 * NOTE! Each of the iocb union members has the file pointer
643 * as the first entry in their struct definition. So you can
644 * access the file pointer through any of the sub-structs,
645 * or directly as just 'ki_filp' in this struct.
651 struct io_poll_iocb poll;
652 struct io_accept accept;
654 struct io_cancel cancel;
655 struct io_timeout timeout;
656 struct io_timeout_rem timeout_rem;
657 struct io_connect connect;
658 struct io_sr_msg sr_msg;
660 struct io_close close;
661 struct io_files_update files_update;
662 struct io_fadvise fadvise;
663 struct io_madvise madvise;
664 struct io_epoll epoll;
665 struct io_splice splice;
666 struct io_provide_buf pbuf;
667 struct io_statx statx;
668 /* use only after cleaning per-op data, see io_clean_op() */
669 struct io_completion compl;
672 /* opcode allocated if it needs to store data for async defer */
675 /* polled IO has completed */
681 struct io_ring_ctx *ctx;
684 struct task_struct *task;
687 struct list_head link_list;
690 * 1. used with ctx->iopoll_list with reads/writes
691 * 2. to track reqs with ->files (see io_op_def::file_table)
693 struct list_head inflight_entry;
695 struct percpu_ref *fixed_file_refs;
696 struct callback_head task_work;
697 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
698 struct hlist_node hash_node;
699 struct async_poll *apoll;
700 struct io_wq_work work;
703 struct io_defer_entry {
704 struct list_head list;
705 struct io_kiocb *req;
709 #define IO_IOPOLL_BATCH 8
711 struct io_comp_state {
713 struct list_head list;
714 struct io_ring_ctx *ctx;
717 struct io_submit_state {
718 struct blk_plug plug;
721 * io_kiocb alloc cache
723 void *reqs[IO_IOPOLL_BATCH];
724 unsigned int free_reqs;
727 * Batch completion logic
729 struct io_comp_state comp;
732 * File reference cache
736 unsigned int has_refs;
737 unsigned int ios_left;
741 /* needs req->file assigned */
742 unsigned needs_file : 1;
743 /* don't fail if file grab fails */
744 unsigned needs_file_no_error : 1;
745 /* hash wq insertion if file is a regular file */
746 unsigned hash_reg_file : 1;
747 /* unbound wq insertion if file is a non-regular file */
748 unsigned unbound_nonreg_file : 1;
749 /* opcode is not supported by this kernel */
750 unsigned not_supported : 1;
751 /* set if opcode supports polled "wait" */
753 unsigned pollout : 1;
754 /* op supports buffer selection */
755 unsigned buffer_select : 1;
756 /* must always have async data allocated */
757 unsigned needs_async_data : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
763 static const struct io_op_def io_op_defs[] = {
764 [IORING_OP_NOP] = {},
765 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
771 .async_size = sizeof(struct io_async_rw),
772 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
774 [IORING_OP_WRITEV] = {
777 .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 |
784 [IORING_OP_FSYNC] = {
786 .work_flags = IO_WQ_WORK_BLKCG,
788 [IORING_OP_READ_FIXED] = {
790 .unbound_nonreg_file = 1,
792 .async_size = sizeof(struct io_async_rw),
793 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
795 [IORING_OP_WRITE_FIXED] = {
798 .unbound_nonreg_file = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
804 [IORING_OP_POLL_ADD] = {
806 .unbound_nonreg_file = 1,
808 [IORING_OP_POLL_REMOVE] = {},
809 [IORING_OP_SYNC_FILE_RANGE] = {
811 .work_flags = IO_WQ_WORK_BLKCG,
813 [IORING_OP_SENDMSG] = {
815 .unbound_nonreg_file = 1,
817 .needs_async_data = 1,
818 .async_size = sizeof(struct io_async_msghdr),
819 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
822 [IORING_OP_RECVMSG] = {
824 .unbound_nonreg_file = 1,
827 .needs_async_data = 1,
828 .async_size = sizeof(struct io_async_msghdr),
829 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
832 [IORING_OP_TIMEOUT] = {
833 .needs_async_data = 1,
834 .async_size = sizeof(struct io_timeout_data),
835 .work_flags = IO_WQ_WORK_MM,
837 [IORING_OP_TIMEOUT_REMOVE] = {},
838 [IORING_OP_ACCEPT] = {
840 .unbound_nonreg_file = 1,
842 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
844 [IORING_OP_ASYNC_CANCEL] = {},
845 [IORING_OP_LINK_TIMEOUT] = {
846 .needs_async_data = 1,
847 .async_size = sizeof(struct io_timeout_data),
848 .work_flags = IO_WQ_WORK_MM,
850 [IORING_OP_CONNECT] = {
852 .unbound_nonreg_file = 1,
854 .needs_async_data = 1,
855 .async_size = sizeof(struct io_async_connect),
856 .work_flags = IO_WQ_WORK_MM,
858 [IORING_OP_FALLOCATE] = {
860 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
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,
890 .async_size = sizeof(struct io_async_rw),
891 .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->work_flags & IO_WQ_WORK_FSIZE) {
1294 if (id->fsize != rlimit(RLIMIT_FSIZE))
1296 req->work.flags |= IO_WQ_WORK_FSIZE;
1299 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1300 (def->work_flags & IO_WQ_WORK_FILES) &&
1301 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1302 if (id->files != current->files ||
1303 id->nsproxy != current->nsproxy)
1305 atomic_inc(&id->files->count);
1306 get_nsproxy(id->nsproxy);
1307 req->flags |= REQ_F_INFLIGHT;
1309 spin_lock_irq(&ctx->inflight_lock);
1310 list_add(&req->inflight_entry, &ctx->inflight_list);
1311 spin_unlock_irq(&ctx->inflight_lock);
1312 req->work.flags |= IO_WQ_WORK_FILES;
1314 #ifdef CONFIG_BLK_CGROUP
1315 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1316 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1318 if (id->blkcg_css != blkcg_css()) {
1323 * This should be rare, either the cgroup is dying or the task
1324 * is moving cgroups. Just punt to root for the handful of ios.
1326 if (css_tryget_online(id->blkcg_css))
1327 req->work.flags |= IO_WQ_WORK_BLKCG;
1331 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1332 if (id->creds != current_cred())
1334 get_cred(id->creds);
1335 req->work.flags |= IO_WQ_WORK_CREDS;
1338 if (!uid_eq(current->loginuid, id->loginuid) ||
1339 current->sessionid != id->sessionid)
1342 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1343 (def->work_flags & IO_WQ_WORK_FS)) {
1344 if (current->fs != id->fs)
1346 spin_lock(&id->fs->lock);
1347 if (!id->fs->in_exec) {
1349 req->work.flags |= IO_WQ_WORK_FS;
1351 req->work.flags |= IO_WQ_WORK_CANCEL;
1353 spin_unlock(¤t->fs->lock);
1359 static void io_prep_async_work(struct io_kiocb *req)
1361 const struct io_op_def *def = &io_op_defs[req->opcode];
1362 struct io_ring_ctx *ctx = req->ctx;
1363 struct io_identity *id;
1365 io_req_init_async(req);
1366 id = req->work.identity;
1368 if (req->flags & REQ_F_ISREG) {
1369 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1370 io_wq_hash_work(&req->work, file_inode(req->file));
1372 if (def->unbound_nonreg_file)
1373 req->work.flags |= IO_WQ_WORK_UNBOUND;
1376 /* ->mm can never change on us */
1377 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1378 (def->work_flags & IO_WQ_WORK_MM)) {
1380 req->work.flags |= IO_WQ_WORK_MM;
1383 /* if we fail grabbing identity, we must COW, regrab, and retry */
1384 if (io_grab_identity(req))
1387 if (!io_identity_cow(req))
1390 /* can't fail at this point */
1391 if (!io_grab_identity(req))
1395 static void io_prep_async_link(struct io_kiocb *req)
1397 struct io_kiocb *cur;
1399 io_prep_async_work(req);
1400 if (req->flags & REQ_F_LINK_HEAD)
1401 list_for_each_entry(cur, &req->link_list, link_list)
1402 io_prep_async_work(cur);
1405 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1407 struct io_ring_ctx *ctx = req->ctx;
1408 struct io_kiocb *link = io_prep_linked_timeout(req);
1410 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1411 &req->work, req->flags);
1412 io_wq_enqueue(ctx->io_wq, &req->work);
1416 static void io_queue_async_work(struct io_kiocb *req)
1418 struct io_kiocb *link;
1420 /* init ->work of the whole link before punting */
1421 io_prep_async_link(req);
1422 link = __io_queue_async_work(req);
1425 io_queue_linked_timeout(link);
1428 static void io_kill_timeout(struct io_kiocb *req)
1430 struct io_timeout_data *io = req->async_data;
1433 ret = hrtimer_try_to_cancel(&io->timer);
1435 atomic_set(&req->ctx->cq_timeouts,
1436 atomic_read(&req->ctx->cq_timeouts) + 1);
1437 list_del_init(&req->timeout.list);
1438 io_cqring_fill_event(req, 0);
1439 io_put_req_deferred(req, 1);
1443 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1445 struct io_ring_ctx *ctx = req->ctx;
1447 if (!tsk || req->task == tsk)
1449 if (ctx->flags & IORING_SETUP_SQPOLL) {
1450 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1457 * Returns true if we found and killed one or more timeouts
1459 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1461 struct io_kiocb *req, *tmp;
1464 spin_lock_irq(&ctx->completion_lock);
1465 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1466 if (io_task_match(req, tsk)) {
1467 io_kill_timeout(req);
1471 spin_unlock_irq(&ctx->completion_lock);
1472 return canceled != 0;
1475 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1478 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1479 struct io_defer_entry, list);
1480 struct io_kiocb *link;
1482 if (req_need_defer(de->req, de->seq))
1484 list_del_init(&de->list);
1485 /* punt-init is done before queueing for defer */
1486 link = __io_queue_async_work(de->req);
1488 __io_queue_linked_timeout(link);
1489 /* drop submission reference */
1490 io_put_req_deferred(link, 1);
1493 } while (!list_empty(&ctx->defer_list));
1496 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1498 while (!list_empty(&ctx->timeout_list)) {
1499 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1500 struct io_kiocb, timeout.list);
1502 if (io_is_timeout_noseq(req))
1504 if (req->timeout.target_seq != ctx->cached_cq_tail
1505 - atomic_read(&ctx->cq_timeouts))
1508 list_del_init(&req->timeout.list);
1509 io_kill_timeout(req);
1513 static void io_commit_cqring(struct io_ring_ctx *ctx)
1515 io_flush_timeouts(ctx);
1516 __io_commit_cqring(ctx);
1518 if (unlikely(!list_empty(&ctx->defer_list)))
1519 __io_queue_deferred(ctx);
1522 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1524 struct io_rings *r = ctx->rings;
1526 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1529 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1531 struct io_rings *rings = ctx->rings;
1534 tail = ctx->cached_cq_tail;
1536 * writes to the cq entry need to come after reading head; the
1537 * control dependency is enough as we're using WRITE_ONCE to
1540 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1543 ctx->cached_cq_tail++;
1544 return &rings->cqes[tail & ctx->cq_mask];
1547 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1551 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1553 if (!ctx->eventfd_async)
1555 return io_wq_current_is_worker();
1558 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1560 if (waitqueue_active(&ctx->wait))
1561 wake_up(&ctx->wait);
1562 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1563 wake_up(&ctx->sq_data->wait);
1564 if (io_should_trigger_evfd(ctx))
1565 eventfd_signal(ctx->cq_ev_fd, 1);
1568 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1570 if (list_empty(&ctx->cq_overflow_list)) {
1571 clear_bit(0, &ctx->sq_check_overflow);
1572 clear_bit(0, &ctx->cq_check_overflow);
1573 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1577 static inline bool io_match_files(struct io_kiocb *req,
1578 struct files_struct *files)
1582 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1583 (req->work.flags & IO_WQ_WORK_FILES))
1584 return req->work.identity->files == files;
1588 /* Returns true if there are no backlogged entries after the flush */
1589 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1590 struct task_struct *tsk,
1591 struct files_struct *files)
1593 struct io_rings *rings = ctx->rings;
1594 struct io_kiocb *req, *tmp;
1595 struct io_uring_cqe *cqe;
1596 unsigned long flags;
1600 if (list_empty_careful(&ctx->cq_overflow_list))
1602 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1603 rings->cq_ring_entries))
1607 spin_lock_irqsave(&ctx->completion_lock, flags);
1609 /* if force is set, the ring is going away. always drop after that */
1611 ctx->cq_overflow_flushed = 1;
1614 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1615 if (tsk && req->task != tsk)
1617 if (!io_match_files(req, files))
1620 cqe = io_get_cqring(ctx);
1624 list_move(&req->compl.list, &list);
1626 WRITE_ONCE(cqe->user_data, req->user_data);
1627 WRITE_ONCE(cqe->res, req->result);
1628 WRITE_ONCE(cqe->flags, req->compl.cflags);
1630 ctx->cached_cq_overflow++;
1631 WRITE_ONCE(ctx->rings->cq_overflow,
1632 ctx->cached_cq_overflow);
1636 io_commit_cqring(ctx);
1637 io_cqring_mark_overflow(ctx);
1639 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1640 io_cqring_ev_posted(ctx);
1642 while (!list_empty(&list)) {
1643 req = list_first_entry(&list, struct io_kiocb, compl.list);
1644 list_del(&req->compl.list);
1651 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1653 struct io_ring_ctx *ctx = req->ctx;
1654 struct io_uring_cqe *cqe;
1656 trace_io_uring_complete(ctx, req->user_data, res);
1659 * If we can't get a cq entry, userspace overflowed the
1660 * submission (by quite a lot). Increment the overflow count in
1663 cqe = io_get_cqring(ctx);
1665 WRITE_ONCE(cqe->user_data, req->user_data);
1666 WRITE_ONCE(cqe->res, res);
1667 WRITE_ONCE(cqe->flags, cflags);
1668 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1670 * If we're in ring overflow flush mode, or in task cancel mode,
1671 * then we cannot store the request for later flushing, we need
1672 * to drop it on the floor.
1674 ctx->cached_cq_overflow++;
1675 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1677 if (list_empty(&ctx->cq_overflow_list)) {
1678 set_bit(0, &ctx->sq_check_overflow);
1679 set_bit(0, &ctx->cq_check_overflow);
1680 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1684 req->compl.cflags = cflags;
1685 refcount_inc(&req->refs);
1686 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1690 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1692 __io_cqring_fill_event(req, res, 0);
1695 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1697 struct io_ring_ctx *ctx = req->ctx;
1698 unsigned long flags;
1700 spin_lock_irqsave(&ctx->completion_lock, flags);
1701 __io_cqring_fill_event(req, res, cflags);
1702 io_commit_cqring(ctx);
1703 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1705 io_cqring_ev_posted(ctx);
1708 static void io_submit_flush_completions(struct io_comp_state *cs)
1710 struct io_ring_ctx *ctx = cs->ctx;
1712 spin_lock_irq(&ctx->completion_lock);
1713 while (!list_empty(&cs->list)) {
1714 struct io_kiocb *req;
1716 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1717 list_del(&req->compl.list);
1718 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1721 * io_free_req() doesn't care about completion_lock unless one
1722 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1723 * because of a potential deadlock with req->work.fs->lock
1725 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1726 |REQ_F_WORK_INITIALIZED)) {
1727 spin_unlock_irq(&ctx->completion_lock);
1729 spin_lock_irq(&ctx->completion_lock);
1734 io_commit_cqring(ctx);
1735 spin_unlock_irq(&ctx->completion_lock);
1737 io_cqring_ev_posted(ctx);
1741 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1742 struct io_comp_state *cs)
1745 io_cqring_add_event(req, res, cflags);
1750 req->compl.cflags = cflags;
1751 list_add_tail(&req->compl.list, &cs->list);
1753 io_submit_flush_completions(cs);
1757 static void io_req_complete(struct io_kiocb *req, long res)
1759 __io_req_complete(req, res, 0, NULL);
1762 static inline bool io_is_fallback_req(struct io_kiocb *req)
1764 return req == (struct io_kiocb *)
1765 ((unsigned long) req->ctx->fallback_req & ~1UL);
1768 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1770 struct io_kiocb *req;
1772 req = ctx->fallback_req;
1773 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1779 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1780 struct io_submit_state *state)
1782 if (!state->free_reqs) {
1783 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1787 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1788 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1791 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1792 * retry single alloc to be on the safe side.
1794 if (unlikely(ret <= 0)) {
1795 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1796 if (!state->reqs[0])
1800 state->free_reqs = ret;
1804 return state->reqs[state->free_reqs];
1806 return io_get_fallback_req(ctx);
1809 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1813 percpu_ref_put(req->fixed_file_refs);
1818 static void io_dismantle_req(struct io_kiocb *req)
1822 if (req->async_data)
1823 kfree(req->async_data);
1825 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1827 io_req_clean_work(req);
1830 static void __io_free_req(struct io_kiocb *req)
1832 struct io_uring_task *tctx = req->task->io_uring;
1833 struct io_ring_ctx *ctx = req->ctx;
1835 io_dismantle_req(req);
1837 percpu_counter_dec(&tctx->inflight);
1839 wake_up(&tctx->wait);
1840 put_task_struct(req->task);
1842 if (likely(!io_is_fallback_req(req)))
1843 kmem_cache_free(req_cachep, req);
1845 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1846 percpu_ref_put(&ctx->refs);
1849 static bool io_link_cancel_timeout(struct io_kiocb *req)
1851 struct io_timeout_data *io = req->async_data;
1852 struct io_ring_ctx *ctx = req->ctx;
1855 ret = hrtimer_try_to_cancel(&io->timer);
1857 io_cqring_fill_event(req, -ECANCELED);
1858 io_commit_cqring(ctx);
1859 req->flags &= ~REQ_F_LINK_HEAD;
1860 io_put_req_deferred(req, 1);
1867 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1869 struct io_kiocb *link;
1872 if (list_empty(&req->link_list))
1874 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1875 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1878 * Can happen if a linked timeout fired and link had been like
1879 * req -> link t-out -> link t-out [-> ...]
1881 if (!(link->flags & REQ_F_LTIMEOUT_ACTIVE))
1884 list_del_init(&link->link_list);
1885 wake_ev = io_link_cancel_timeout(link);
1886 req->flags &= ~REQ_F_LINK_TIMEOUT;
1890 static void io_kill_linked_timeout(struct io_kiocb *req)
1892 struct io_ring_ctx *ctx = req->ctx;
1893 unsigned long flags;
1896 spin_lock_irqsave(&ctx->completion_lock, flags);
1897 wake_ev = __io_kill_linked_timeout(req);
1898 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1901 io_cqring_ev_posted(ctx);
1904 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1906 struct io_kiocb *nxt;
1909 * The list should never be empty when we are called here. But could
1910 * potentially happen if the chain is messed up, check to be on the
1913 if (unlikely(list_empty(&req->link_list)))
1916 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1917 list_del_init(&req->link_list);
1918 if (!list_empty(&nxt->link_list))
1919 nxt->flags |= REQ_F_LINK_HEAD;
1924 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1926 static void io_fail_links(struct io_kiocb *req)
1928 struct io_ring_ctx *ctx = req->ctx;
1929 unsigned long flags;
1931 spin_lock_irqsave(&ctx->completion_lock, flags);
1932 while (!list_empty(&req->link_list)) {
1933 struct io_kiocb *link = list_first_entry(&req->link_list,
1934 struct io_kiocb, link_list);
1936 list_del_init(&link->link_list);
1937 trace_io_uring_fail_link(req, link);
1939 io_cqring_fill_event(link, -ECANCELED);
1942 * It's ok to free under spinlock as they're not linked anymore,
1943 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1946 if (link->flags & REQ_F_WORK_INITIALIZED)
1947 io_put_req_deferred(link, 2);
1949 io_double_put_req(link);
1952 io_commit_cqring(ctx);
1953 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1955 io_cqring_ev_posted(ctx);
1958 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1960 req->flags &= ~REQ_F_LINK_HEAD;
1961 if (req->flags & REQ_F_LINK_TIMEOUT)
1962 io_kill_linked_timeout(req);
1965 * If LINK is set, we have dependent requests in this chain. If we
1966 * didn't fail this request, queue the first one up, moving any other
1967 * dependencies to the next request. In case of failure, fail the rest
1970 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1971 return io_req_link_next(req);
1976 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1978 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1980 return __io_req_find_next(req);
1983 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1985 struct task_struct *tsk = req->task;
1986 struct io_ring_ctx *ctx = req->ctx;
1987 enum task_work_notify_mode notify;
1990 if (tsk->flags & PF_EXITING)
1994 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1995 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1996 * processing task_work. There's no reliable way to tell if TWA_RESUME
2000 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2001 notify = TWA_SIGNAL;
2003 ret = task_work_add(tsk, &req->task_work, notify);
2005 wake_up_process(tsk);
2010 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2012 struct io_ring_ctx *ctx = req->ctx;
2014 spin_lock_irq(&ctx->completion_lock);
2015 io_cqring_fill_event(req, error);
2016 io_commit_cqring(ctx);
2017 spin_unlock_irq(&ctx->completion_lock);
2019 io_cqring_ev_posted(ctx);
2020 req_set_fail_links(req);
2021 io_double_put_req(req);
2024 static void io_req_task_cancel(struct callback_head *cb)
2026 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2027 struct io_ring_ctx *ctx = req->ctx;
2029 __io_req_task_cancel(req, -ECANCELED);
2030 percpu_ref_put(&ctx->refs);
2033 static void __io_req_task_submit(struct io_kiocb *req)
2035 struct io_ring_ctx *ctx = req->ctx;
2037 if (!__io_sq_thread_acquire_mm(ctx)) {
2038 mutex_lock(&ctx->uring_lock);
2039 __io_queue_sqe(req, NULL);
2040 mutex_unlock(&ctx->uring_lock);
2042 __io_req_task_cancel(req, -EFAULT);
2046 static void io_req_task_submit(struct callback_head *cb)
2048 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2049 struct io_ring_ctx *ctx = req->ctx;
2051 __io_req_task_submit(req);
2052 percpu_ref_put(&ctx->refs);
2055 static void io_req_task_queue(struct io_kiocb *req)
2059 init_task_work(&req->task_work, io_req_task_submit);
2060 percpu_ref_get(&req->ctx->refs);
2062 ret = io_req_task_work_add(req, true);
2063 if (unlikely(ret)) {
2064 struct task_struct *tsk;
2066 init_task_work(&req->task_work, io_req_task_cancel);
2067 tsk = io_wq_get_task(req->ctx->io_wq);
2068 task_work_add(tsk, &req->task_work, TWA_NONE);
2069 wake_up_process(tsk);
2073 static void io_queue_next(struct io_kiocb *req)
2075 struct io_kiocb *nxt = io_req_find_next(req);
2078 io_req_task_queue(nxt);
2081 static void io_free_req(struct io_kiocb *req)
2088 void *reqs[IO_IOPOLL_BATCH];
2091 struct task_struct *task;
2095 static inline void io_init_req_batch(struct req_batch *rb)
2102 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2103 struct req_batch *rb)
2105 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2106 percpu_ref_put_many(&ctx->refs, rb->to_free);
2110 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2111 struct req_batch *rb)
2114 __io_req_free_batch_flush(ctx, rb);
2116 struct io_uring_task *tctx = rb->task->io_uring;
2118 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2119 put_task_struct_many(rb->task, rb->task_refs);
2124 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2126 if (unlikely(io_is_fallback_req(req))) {
2130 if (req->flags & REQ_F_LINK_HEAD)
2133 if (req->task != rb->task) {
2135 struct io_uring_task *tctx = rb->task->io_uring;
2137 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2138 put_task_struct_many(rb->task, rb->task_refs);
2140 rb->task = req->task;
2145 io_dismantle_req(req);
2146 rb->reqs[rb->to_free++] = req;
2147 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2148 __io_req_free_batch_flush(req->ctx, rb);
2152 * Drop reference to request, return next in chain (if there is one) if this
2153 * was the last reference to this request.
2155 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2157 struct io_kiocb *nxt = NULL;
2159 if (refcount_dec_and_test(&req->refs)) {
2160 nxt = io_req_find_next(req);
2166 static void io_put_req(struct io_kiocb *req)
2168 if (refcount_dec_and_test(&req->refs))
2172 static void io_put_req_deferred_cb(struct callback_head *cb)
2174 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2179 static void io_free_req_deferred(struct io_kiocb *req)
2183 init_task_work(&req->task_work, io_put_req_deferred_cb);
2184 ret = io_req_task_work_add(req, true);
2185 if (unlikely(ret)) {
2186 struct task_struct *tsk;
2188 tsk = io_wq_get_task(req->ctx->io_wq);
2189 task_work_add(tsk, &req->task_work, TWA_NONE);
2190 wake_up_process(tsk);
2194 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2196 if (refcount_sub_and_test(refs, &req->refs))
2197 io_free_req_deferred(req);
2200 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2202 struct io_kiocb *nxt;
2205 * A ref is owned by io-wq in which context we're. So, if that's the
2206 * last one, it's safe to steal next work. False negatives are Ok,
2207 * it just will be re-punted async in io_put_work()
2209 if (refcount_read(&req->refs) != 1)
2212 nxt = io_req_find_next(req);
2213 return nxt ? &nxt->work : NULL;
2216 static void io_double_put_req(struct io_kiocb *req)
2218 /* drop both submit and complete references */
2219 if (refcount_sub_and_test(2, &req->refs))
2223 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2225 struct io_rings *rings = ctx->rings;
2227 if (test_bit(0, &ctx->cq_check_overflow)) {
2229 * noflush == true is from the waitqueue handler, just ensure
2230 * we wake up the task, and the next invocation will flush the
2231 * entries. We cannot safely to it from here.
2233 if (noflush && !list_empty(&ctx->cq_overflow_list))
2236 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2239 /* See comment at the top of this file */
2241 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2244 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2246 struct io_rings *rings = ctx->rings;
2248 /* make sure SQ entry isn't read before tail */
2249 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2252 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2254 unsigned int cflags;
2256 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2257 cflags |= IORING_CQE_F_BUFFER;
2258 req->flags &= ~REQ_F_BUFFER_SELECTED;
2263 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2265 struct io_buffer *kbuf;
2267 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2268 return io_put_kbuf(req, kbuf);
2271 static inline bool io_run_task_work(void)
2274 * Not safe to run on exiting task, and the task_work handling will
2275 * not add work to such a task.
2277 if (unlikely(current->flags & PF_EXITING))
2279 if (current->task_works) {
2280 __set_current_state(TASK_RUNNING);
2288 static void io_iopoll_queue(struct list_head *again)
2290 struct io_kiocb *req;
2293 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2294 list_del(&req->inflight_entry);
2295 __io_complete_rw(req, -EAGAIN, 0, NULL);
2296 } while (!list_empty(again));
2300 * Find and free completed poll iocbs
2302 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2303 struct list_head *done)
2305 struct req_batch rb;
2306 struct io_kiocb *req;
2309 /* order with ->result store in io_complete_rw_iopoll() */
2312 io_init_req_batch(&rb);
2313 while (!list_empty(done)) {
2316 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2317 if (READ_ONCE(req->result) == -EAGAIN) {
2319 req->iopoll_completed = 0;
2320 list_move_tail(&req->inflight_entry, &again);
2323 list_del(&req->inflight_entry);
2325 if (req->flags & REQ_F_BUFFER_SELECTED)
2326 cflags = io_put_rw_kbuf(req);
2328 __io_cqring_fill_event(req, req->result, cflags);
2331 if (refcount_dec_and_test(&req->refs))
2332 io_req_free_batch(&rb, req);
2335 io_commit_cqring(ctx);
2336 if (ctx->flags & IORING_SETUP_SQPOLL)
2337 io_cqring_ev_posted(ctx);
2338 io_req_free_batch_finish(ctx, &rb);
2340 if (!list_empty(&again))
2341 io_iopoll_queue(&again);
2344 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2347 struct io_kiocb *req, *tmp;
2353 * Only spin for completions if we don't have multiple devices hanging
2354 * off our complete list, and we're under the requested amount.
2356 spin = !ctx->poll_multi_file && *nr_events < min;
2359 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2360 struct kiocb *kiocb = &req->rw.kiocb;
2363 * Move completed and retryable entries to our local lists.
2364 * If we find a request that requires polling, break out
2365 * and complete those lists first, if we have entries there.
2367 if (READ_ONCE(req->iopoll_completed)) {
2368 list_move_tail(&req->inflight_entry, &done);
2371 if (!list_empty(&done))
2374 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2378 /* iopoll may have completed current req */
2379 if (READ_ONCE(req->iopoll_completed))
2380 list_move_tail(&req->inflight_entry, &done);
2387 if (!list_empty(&done))
2388 io_iopoll_complete(ctx, nr_events, &done);
2394 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2395 * non-spinning poll check - we'll still enter the driver poll loop, but only
2396 * as a non-spinning completion check.
2398 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2401 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2404 ret = io_do_iopoll(ctx, nr_events, min);
2407 if (*nr_events >= min)
2415 * We can't just wait for polled events to come to us, we have to actively
2416 * find and complete them.
2418 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2420 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2423 mutex_lock(&ctx->uring_lock);
2424 while (!list_empty(&ctx->iopoll_list)) {
2425 unsigned int nr_events = 0;
2427 io_do_iopoll(ctx, &nr_events, 0);
2429 /* let it sleep and repeat later if can't complete a request */
2433 * Ensure we allow local-to-the-cpu processing to take place,
2434 * in this case we need to ensure that we reap all events.
2435 * Also let task_work, etc. to progress by releasing the mutex
2437 if (need_resched()) {
2438 mutex_unlock(&ctx->uring_lock);
2440 mutex_lock(&ctx->uring_lock);
2443 mutex_unlock(&ctx->uring_lock);
2446 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2448 unsigned int nr_events = 0;
2449 int iters = 0, ret = 0;
2452 * We disallow the app entering submit/complete with polling, but we
2453 * still need to lock the ring to prevent racing with polled issue
2454 * that got punted to a workqueue.
2456 mutex_lock(&ctx->uring_lock);
2459 * Don't enter poll loop if we already have events pending.
2460 * If we do, we can potentially be spinning for commands that
2461 * already triggered a CQE (eg in error).
2463 if (io_cqring_events(ctx, false))
2467 * If a submit got punted to a workqueue, we can have the
2468 * application entering polling for a command before it gets
2469 * issued. That app will hold the uring_lock for the duration
2470 * of the poll right here, so we need to take a breather every
2471 * now and then to ensure that the issue has a chance to add
2472 * the poll to the issued list. Otherwise we can spin here
2473 * forever, while the workqueue is stuck trying to acquire the
2476 if (!(++iters & 7)) {
2477 mutex_unlock(&ctx->uring_lock);
2479 mutex_lock(&ctx->uring_lock);
2482 ret = io_iopoll_getevents(ctx, &nr_events, min);
2486 } while (min && !nr_events && !need_resched());
2488 mutex_unlock(&ctx->uring_lock);
2492 static void kiocb_end_write(struct io_kiocb *req)
2495 * Tell lockdep we inherited freeze protection from submission
2498 if (req->flags & REQ_F_ISREG) {
2499 struct inode *inode = file_inode(req->file);
2501 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2503 file_end_write(req->file);
2506 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2507 struct io_comp_state *cs)
2509 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2512 if (kiocb->ki_flags & IOCB_WRITE)
2513 kiocb_end_write(req);
2515 if (res != req->result)
2516 req_set_fail_links(req);
2517 if (req->flags & REQ_F_BUFFER_SELECTED)
2518 cflags = io_put_rw_kbuf(req);
2519 __io_req_complete(req, res, cflags, cs);
2523 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2525 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2526 ssize_t ret = -ECANCELED;
2527 struct iov_iter iter;
2535 switch (req->opcode) {
2536 case IORING_OP_READV:
2537 case IORING_OP_READ_FIXED:
2538 case IORING_OP_READ:
2541 case IORING_OP_WRITEV:
2542 case IORING_OP_WRITE_FIXED:
2543 case IORING_OP_WRITE:
2547 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2552 if (!req->async_data) {
2553 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2556 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2564 req_set_fail_links(req);
2565 io_req_complete(req, ret);
2570 static bool io_rw_reissue(struct io_kiocb *req, long res)
2573 umode_t mode = file_inode(req->file)->i_mode;
2576 if (!S_ISBLK(mode) && !S_ISREG(mode))
2578 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2581 ret = io_sq_thread_acquire_mm(req->ctx, req);
2583 if (io_resubmit_prep(req, ret)) {
2584 refcount_inc(&req->refs);
2585 io_queue_async_work(req);
2593 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2594 struct io_comp_state *cs)
2596 if (!io_rw_reissue(req, res))
2597 io_complete_rw_common(&req->rw.kiocb, res, cs);
2600 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2602 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2604 __io_complete_rw(req, res, res2, NULL);
2607 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2609 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2611 if (kiocb->ki_flags & IOCB_WRITE)
2612 kiocb_end_write(req);
2614 if (res != -EAGAIN && res != req->result)
2615 req_set_fail_links(req);
2617 WRITE_ONCE(req->result, res);
2618 /* order with io_poll_complete() checking ->result */
2620 WRITE_ONCE(req->iopoll_completed, 1);
2624 * After the iocb has been issued, it's safe to be found on the poll list.
2625 * Adding the kiocb to the list AFTER submission ensures that we don't
2626 * find it from a io_iopoll_getevents() thread before the issuer is done
2627 * accessing the kiocb cookie.
2629 static void io_iopoll_req_issued(struct io_kiocb *req)
2631 struct io_ring_ctx *ctx = req->ctx;
2634 * Track whether we have multiple files in our lists. This will impact
2635 * how we do polling eventually, not spinning if we're on potentially
2636 * different devices.
2638 if (list_empty(&ctx->iopoll_list)) {
2639 ctx->poll_multi_file = false;
2640 } else if (!ctx->poll_multi_file) {
2641 struct io_kiocb *list_req;
2643 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2645 if (list_req->file != req->file)
2646 ctx->poll_multi_file = true;
2650 * For fast devices, IO may have already completed. If it has, add
2651 * it to the front so we find it first.
2653 if (READ_ONCE(req->iopoll_completed))
2654 list_add(&req->inflight_entry, &ctx->iopoll_list);
2656 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2658 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2659 wq_has_sleeper(&ctx->sq_data->wait))
2660 wake_up(&ctx->sq_data->wait);
2663 static void __io_state_file_put(struct io_submit_state *state)
2665 if (state->has_refs)
2666 fput_many(state->file, state->has_refs);
2670 static inline void io_state_file_put(struct io_submit_state *state)
2673 __io_state_file_put(state);
2677 * Get as many references to a file as we have IOs left in this submission,
2678 * assuming most submissions are for one file, or at least that each file
2679 * has more than one submission.
2681 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2687 if (state->fd == fd) {
2691 __io_state_file_put(state);
2693 state->file = fget_many(fd, state->ios_left);
2698 state->has_refs = state->ios_left - 1;
2702 static bool io_bdev_nowait(struct block_device *bdev)
2705 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2712 * If we tracked the file through the SCM inflight mechanism, we could support
2713 * any file. For now, just ensure that anything potentially problematic is done
2716 static bool io_file_supports_async(struct file *file, int rw)
2718 umode_t mode = file_inode(file)->i_mode;
2720 if (S_ISBLK(mode)) {
2721 if (io_bdev_nowait(file->f_inode->i_bdev))
2725 if (S_ISCHR(mode) || S_ISSOCK(mode))
2727 if (S_ISREG(mode)) {
2728 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2729 file->f_op != &io_uring_fops)
2734 /* any ->read/write should understand O_NONBLOCK */
2735 if (file->f_flags & O_NONBLOCK)
2738 if (!(file->f_mode & FMODE_NOWAIT))
2742 return file->f_op->read_iter != NULL;
2744 return file->f_op->write_iter != NULL;
2747 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2749 struct io_ring_ctx *ctx = req->ctx;
2750 struct kiocb *kiocb = &req->rw.kiocb;
2754 if (S_ISREG(file_inode(req->file)->i_mode))
2755 req->flags |= REQ_F_ISREG;
2757 kiocb->ki_pos = READ_ONCE(sqe->off);
2758 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2759 req->flags |= REQ_F_CUR_POS;
2760 kiocb->ki_pos = req->file->f_pos;
2762 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2763 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2764 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2768 ioprio = READ_ONCE(sqe->ioprio);
2770 ret = ioprio_check_cap(ioprio);
2774 kiocb->ki_ioprio = ioprio;
2776 kiocb->ki_ioprio = get_current_ioprio();
2778 /* don't allow async punt if RWF_NOWAIT was requested */
2779 if (kiocb->ki_flags & IOCB_NOWAIT)
2780 req->flags |= REQ_F_NOWAIT;
2782 if (ctx->flags & IORING_SETUP_IOPOLL) {
2783 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2784 !kiocb->ki_filp->f_op->iopoll)
2787 kiocb->ki_flags |= IOCB_HIPRI;
2788 kiocb->ki_complete = io_complete_rw_iopoll;
2789 req->iopoll_completed = 0;
2791 if (kiocb->ki_flags & IOCB_HIPRI)
2793 kiocb->ki_complete = io_complete_rw;
2796 req->rw.addr = READ_ONCE(sqe->addr);
2797 req->rw.len = READ_ONCE(sqe->len);
2798 req->buf_index = READ_ONCE(sqe->buf_index);
2802 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2808 case -ERESTARTNOINTR:
2809 case -ERESTARTNOHAND:
2810 case -ERESTART_RESTARTBLOCK:
2812 * We can't just restart the syscall, since previously
2813 * submitted sqes may already be in progress. Just fail this
2819 kiocb->ki_complete(kiocb, ret, 0);
2823 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2824 struct io_comp_state *cs)
2826 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2827 struct io_async_rw *io = req->async_data;
2829 /* add previously done IO, if any */
2830 if (io && io->bytes_done > 0) {
2832 ret = io->bytes_done;
2834 ret += io->bytes_done;
2837 if (req->flags & REQ_F_CUR_POS)
2838 req->file->f_pos = kiocb->ki_pos;
2839 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2840 __io_complete_rw(req, ret, 0, cs);
2842 io_rw_done(kiocb, ret);
2845 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2846 struct iov_iter *iter)
2848 struct io_ring_ctx *ctx = req->ctx;
2849 size_t len = req->rw.len;
2850 struct io_mapped_ubuf *imu;
2851 u16 index, buf_index = req->buf_index;
2855 if (unlikely(buf_index >= ctx->nr_user_bufs))
2857 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2858 imu = &ctx->user_bufs[index];
2859 buf_addr = req->rw.addr;
2862 if (buf_addr + len < buf_addr)
2864 /* not inside the mapped region */
2865 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2869 * May not be a start of buffer, set size appropriately
2870 * and advance us to the beginning.
2872 offset = buf_addr - imu->ubuf;
2873 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2877 * Don't use iov_iter_advance() here, as it's really slow for
2878 * using the latter parts of a big fixed buffer - it iterates
2879 * over each segment manually. We can cheat a bit here, because
2882 * 1) it's a BVEC iter, we set it up
2883 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2884 * first and last bvec
2886 * So just find our index, and adjust the iterator afterwards.
2887 * If the offset is within the first bvec (or the whole first
2888 * bvec, just use iov_iter_advance(). This makes it easier
2889 * since we can just skip the first segment, which may not
2890 * be PAGE_SIZE aligned.
2892 const struct bio_vec *bvec = imu->bvec;
2894 if (offset <= bvec->bv_len) {
2895 iov_iter_advance(iter, offset);
2897 unsigned long seg_skip;
2899 /* skip first vec */
2900 offset -= bvec->bv_len;
2901 seg_skip = 1 + (offset >> PAGE_SHIFT);
2903 iter->bvec = bvec + seg_skip;
2904 iter->nr_segs -= seg_skip;
2905 iter->count -= bvec->bv_len + offset;
2906 iter->iov_offset = offset & ~PAGE_MASK;
2913 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2916 mutex_unlock(&ctx->uring_lock);
2919 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2922 * "Normal" inline submissions always hold the uring_lock, since we
2923 * grab it from the system call. Same is true for the SQPOLL offload.
2924 * The only exception is when we've detached the request and issue it
2925 * from an async worker thread, grab the lock for that case.
2928 mutex_lock(&ctx->uring_lock);
2931 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2932 int bgid, struct io_buffer *kbuf,
2935 struct io_buffer *head;
2937 if (req->flags & REQ_F_BUFFER_SELECTED)
2940 io_ring_submit_lock(req->ctx, needs_lock);
2942 lockdep_assert_held(&req->ctx->uring_lock);
2944 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2946 if (!list_empty(&head->list)) {
2947 kbuf = list_last_entry(&head->list, struct io_buffer,
2949 list_del(&kbuf->list);
2952 idr_remove(&req->ctx->io_buffer_idr, bgid);
2954 if (*len > kbuf->len)
2957 kbuf = ERR_PTR(-ENOBUFS);
2960 io_ring_submit_unlock(req->ctx, needs_lock);
2965 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2968 struct io_buffer *kbuf;
2971 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2972 bgid = req->buf_index;
2973 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2976 req->rw.addr = (u64) (unsigned long) kbuf;
2977 req->flags |= REQ_F_BUFFER_SELECTED;
2978 return u64_to_user_ptr(kbuf->addr);
2981 #ifdef CONFIG_COMPAT
2982 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2985 struct compat_iovec __user *uiov;
2986 compat_ssize_t clen;
2990 uiov = u64_to_user_ptr(req->rw.addr);
2991 if (!access_ok(uiov, sizeof(*uiov)))
2993 if (__get_user(clen, &uiov->iov_len))
2999 buf = io_rw_buffer_select(req, &len, needs_lock);
3001 return PTR_ERR(buf);
3002 iov[0].iov_base = buf;
3003 iov[0].iov_len = (compat_size_t) len;
3008 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3011 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3015 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3018 len = iov[0].iov_len;
3021 buf = io_rw_buffer_select(req, &len, needs_lock);
3023 return PTR_ERR(buf);
3024 iov[0].iov_base = buf;
3025 iov[0].iov_len = len;
3029 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3032 if (req->flags & REQ_F_BUFFER_SELECTED) {
3033 struct io_buffer *kbuf;
3035 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3036 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3037 iov[0].iov_len = kbuf->len;
3042 else if (req->rw.len > 1)
3045 #ifdef CONFIG_COMPAT
3046 if (req->ctx->compat)
3047 return io_compat_import(req, iov, needs_lock);
3050 return __io_iov_buffer_select(req, iov, needs_lock);
3053 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3054 struct iovec **iovec, struct iov_iter *iter,
3057 void __user *buf = u64_to_user_ptr(req->rw.addr);
3058 size_t sqe_len = req->rw.len;
3062 opcode = req->opcode;
3063 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3065 return io_import_fixed(req, rw, iter);
3068 /* buffer index only valid with fixed read/write, or buffer select */
3069 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3072 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3073 if (req->flags & REQ_F_BUFFER_SELECT) {
3074 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3076 return PTR_ERR(buf);
3077 req->rw.len = sqe_len;
3080 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3082 return ret < 0 ? ret : sqe_len;
3085 if (req->flags & REQ_F_BUFFER_SELECT) {
3086 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3088 ret = (*iovec)->iov_len;
3089 iov_iter_init(iter, rw, *iovec, 1, ret);
3095 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3099 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3100 struct iovec **iovec, struct iov_iter *iter,
3103 struct io_async_rw *iorw = req->async_data;
3106 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3108 return iov_iter_count(&iorw->iter);
3111 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3113 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3117 * For files that don't have ->read_iter() and ->write_iter(), handle them
3118 * by looping over ->read() or ->write() manually.
3120 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3122 struct kiocb *kiocb = &req->rw.kiocb;
3123 struct file *file = req->file;
3127 * Don't support polled IO through this interface, and we can't
3128 * support non-blocking either. For the latter, this just causes
3129 * the kiocb to be handled from an async context.
3131 if (kiocb->ki_flags & IOCB_HIPRI)
3133 if (kiocb->ki_flags & IOCB_NOWAIT)
3136 while (iov_iter_count(iter)) {
3140 if (!iov_iter_is_bvec(iter)) {
3141 iovec = iov_iter_iovec(iter);
3143 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3144 iovec.iov_len = req->rw.len;
3148 nr = file->f_op->read(file, iovec.iov_base,
3149 iovec.iov_len, io_kiocb_ppos(kiocb));
3151 nr = file->f_op->write(file, iovec.iov_base,
3152 iovec.iov_len, io_kiocb_ppos(kiocb));
3161 if (nr != iovec.iov_len)
3165 iov_iter_advance(iter, nr);
3171 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3172 const struct iovec *fast_iov, struct iov_iter *iter)
3174 struct io_async_rw *rw = req->async_data;
3176 memcpy(&rw->iter, iter, sizeof(*iter));
3177 rw->free_iovec = iovec;
3179 /* can only be fixed buffers, no need to do anything */
3180 if (iter->type == ITER_BVEC)
3183 unsigned iov_off = 0;
3185 rw->iter.iov = rw->fast_iov;
3186 if (iter->iov != fast_iov) {
3187 iov_off = iter->iov - fast_iov;
3188 rw->iter.iov += iov_off;
3190 if (rw->fast_iov != fast_iov)
3191 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3192 sizeof(struct iovec) * iter->nr_segs);
3194 req->flags |= REQ_F_NEED_CLEANUP;
3198 static inline int __io_alloc_async_data(struct io_kiocb *req)
3200 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3201 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3202 return req->async_data == NULL;
3205 static int io_alloc_async_data(struct io_kiocb *req)
3207 if (!io_op_defs[req->opcode].needs_async_data)
3210 return __io_alloc_async_data(req);
3213 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3214 const struct iovec *fast_iov,
3215 struct iov_iter *iter, bool force)
3217 if (!force && !io_op_defs[req->opcode].needs_async_data)
3219 if (!req->async_data) {
3220 if (__io_alloc_async_data(req))
3223 io_req_map_rw(req, iovec, fast_iov, iter);
3228 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3230 struct io_async_rw *iorw = req->async_data;
3231 struct iovec *iov = iorw->fast_iov;
3234 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3235 if (unlikely(ret < 0))
3238 iorw->bytes_done = 0;
3239 iorw->free_iovec = iov;
3241 req->flags |= REQ_F_NEED_CLEANUP;
3245 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3249 ret = io_prep_rw(req, sqe);
3253 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3256 /* either don't need iovec imported or already have it */
3257 if (!req->async_data)
3259 return io_rw_prep_async(req, READ);
3263 * This is our waitqueue callback handler, registered through lock_page_async()
3264 * when we initially tried to do the IO with the iocb armed our waitqueue.
3265 * This gets called when the page is unlocked, and we generally expect that to
3266 * happen when the page IO is completed and the page is now uptodate. This will
3267 * queue a task_work based retry of the operation, attempting to copy the data
3268 * again. If the latter fails because the page was NOT uptodate, then we will
3269 * do a thread based blocking retry of the operation. That's the unexpected
3272 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3273 int sync, void *arg)
3275 struct wait_page_queue *wpq;
3276 struct io_kiocb *req = wait->private;
3277 struct wait_page_key *key = arg;
3280 wpq = container_of(wait, struct wait_page_queue, wait);
3282 if (!wake_page_match(wpq, key))
3285 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3286 list_del_init(&wait->entry);
3288 init_task_work(&req->task_work, io_req_task_submit);
3289 percpu_ref_get(&req->ctx->refs);
3291 /* submit ref gets dropped, acquire a new one */
3292 refcount_inc(&req->refs);
3293 ret = io_req_task_work_add(req, true);
3294 if (unlikely(ret)) {
3295 struct task_struct *tsk;
3297 /* queue just for cancelation */
3298 init_task_work(&req->task_work, io_req_task_cancel);
3299 tsk = io_wq_get_task(req->ctx->io_wq);
3300 task_work_add(tsk, &req->task_work, TWA_NONE);
3301 wake_up_process(tsk);
3307 * This controls whether a given IO request should be armed for async page
3308 * based retry. If we return false here, the request is handed to the async
3309 * worker threads for retry. If we're doing buffered reads on a regular file,
3310 * we prepare a private wait_page_queue entry and retry the operation. This
3311 * will either succeed because the page is now uptodate and unlocked, or it
3312 * will register a callback when the page is unlocked at IO completion. Through
3313 * that callback, io_uring uses task_work to setup a retry of the operation.
3314 * That retry will attempt the buffered read again. The retry will generally
3315 * succeed, or in rare cases where it fails, we then fall back to using the
3316 * async worker threads for a blocking retry.
3318 static bool io_rw_should_retry(struct io_kiocb *req)
3320 struct io_async_rw *rw = req->async_data;
3321 struct wait_page_queue *wait = &rw->wpq;
3322 struct kiocb *kiocb = &req->rw.kiocb;
3324 /* never retry for NOWAIT, we just complete with -EAGAIN */
3325 if (req->flags & REQ_F_NOWAIT)
3328 /* Only for buffered IO */
3329 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3333 * just use poll if we can, and don't attempt if the fs doesn't
3334 * support callback based unlocks
3336 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3339 wait->wait.func = io_async_buf_func;
3340 wait->wait.private = req;
3341 wait->wait.flags = 0;
3342 INIT_LIST_HEAD(&wait->wait.entry);
3343 kiocb->ki_flags |= IOCB_WAITQ;
3344 kiocb->ki_flags &= ~IOCB_NOWAIT;
3345 kiocb->ki_waitq = wait;
3349 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3351 if (req->file->f_op->read_iter)
3352 return call_read_iter(req->file, &req->rw.kiocb, iter);
3353 else if (req->file->f_op->read)
3354 return loop_rw_iter(READ, req, iter);
3359 static int io_read(struct io_kiocb *req, bool force_nonblock,
3360 struct io_comp_state *cs)
3362 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3363 struct kiocb *kiocb = &req->rw.kiocb;
3364 struct iov_iter __iter, *iter = &__iter;
3365 struct io_async_rw *rw = req->async_data;
3366 ssize_t io_size, ret, ret2;
3373 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3376 iov_count = iov_iter_count(iter);
3378 req->result = io_size;
3381 /* Ensure we clear previously set non-block flag */
3382 if (!force_nonblock)
3383 kiocb->ki_flags &= ~IOCB_NOWAIT;
3385 kiocb->ki_flags |= IOCB_NOWAIT;
3388 /* If the file doesn't support async, just async punt */
3389 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3393 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3397 ret = io_iter_do_read(req, iter);
3401 } else if (ret == -EIOCBQUEUED) {
3404 } else if (ret == -EAGAIN) {
3405 /* IOPOLL retry should happen for io-wq threads */
3406 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3408 /* no retry on NONBLOCK marked file */
3409 if (req->file->f_flags & O_NONBLOCK)
3411 /* some cases will consume bytes even on error returns */
3412 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3415 } else if (ret < 0) {
3416 /* make sure -ERESTARTSYS -> -EINTR is done */
3420 /* read it all, or we did blocking attempt. no retry. */
3421 if (!iov_iter_count(iter) || !force_nonblock ||
3422 (req->file->f_flags & O_NONBLOCK))
3427 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3434 rw = req->async_data;
3435 /* it's copied and will be cleaned with ->io */
3437 /* now use our persistent iterator, if we aren't already */
3440 rw->bytes_done += ret;
3441 /* if we can retry, do so with the callbacks armed */
3442 if (!io_rw_should_retry(req)) {
3443 kiocb->ki_flags &= ~IOCB_WAITQ;
3448 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3449 * get -EIOCBQUEUED, then we'll get a notification when the desired
3450 * page gets unlocked. We can also get a partial read here, and if we
3451 * do, then just retry at the new offset.
3453 ret = io_iter_do_read(req, iter);
3454 if (ret == -EIOCBQUEUED) {
3457 } else if (ret > 0 && ret < io_size) {
3458 /* we got some bytes, but not all. retry. */
3462 kiocb_done(kiocb, ret, cs);
3465 /* it's reportedly faster than delegating the null check to kfree() */
3471 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3475 ret = io_prep_rw(req, sqe);
3479 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3482 /* either don't need iovec imported or already have it */
3483 if (!req->async_data)
3485 return io_rw_prep_async(req, WRITE);
3488 static int io_write(struct io_kiocb *req, bool force_nonblock,
3489 struct io_comp_state *cs)
3491 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3492 struct kiocb *kiocb = &req->rw.kiocb;
3493 struct iov_iter __iter, *iter = &__iter;
3494 struct io_async_rw *rw = req->async_data;
3496 ssize_t ret, ret2, io_size;
3501 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3504 iov_count = iov_iter_count(iter);
3506 req->result = io_size;
3508 /* Ensure we clear previously set non-block flag */
3509 if (!force_nonblock)
3510 kiocb->ki_flags &= ~IOCB_NOWAIT;
3512 kiocb->ki_flags |= IOCB_NOWAIT;
3514 /* If the file doesn't support async, just async punt */
3515 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3518 /* file path doesn't support NOWAIT for non-direct_IO */
3519 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3520 (req->flags & REQ_F_ISREG))
3523 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3528 * Open-code file_start_write here to grab freeze protection,
3529 * which will be released by another thread in
3530 * io_complete_rw(). Fool lockdep by telling it the lock got
3531 * released so that it doesn't complain about the held lock when
3532 * we return to userspace.
3534 if (req->flags & REQ_F_ISREG) {
3535 __sb_start_write(file_inode(req->file)->i_sb,
3536 SB_FREEZE_WRITE, true);
3537 __sb_writers_release(file_inode(req->file)->i_sb,
3540 kiocb->ki_flags |= IOCB_WRITE;
3542 if (req->file->f_op->write_iter)
3543 ret2 = call_write_iter(req->file, kiocb, iter);
3544 else if (req->file->f_op->write)
3545 ret2 = loop_rw_iter(WRITE, req, iter);
3550 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3551 * retry them without IOCB_NOWAIT.
3553 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3555 /* no retry on NONBLOCK marked file */
3556 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3558 if (!force_nonblock || ret2 != -EAGAIN) {
3559 /* IOPOLL retry should happen for io-wq threads */
3560 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3563 kiocb_done(kiocb, ret2, cs);
3566 /* some cases will consume bytes even on error returns */
3567 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3568 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3573 /* it's reportedly faster than delegating the null check to kfree() */
3579 static int __io_splice_prep(struct io_kiocb *req,
3580 const struct io_uring_sqe *sqe)
3582 struct io_splice* sp = &req->splice;
3583 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3589 sp->len = READ_ONCE(sqe->len);
3590 sp->flags = READ_ONCE(sqe->splice_flags);
3592 if (unlikely(sp->flags & ~valid_flags))
3595 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3596 (sp->flags & SPLICE_F_FD_IN_FIXED));
3599 req->flags |= REQ_F_NEED_CLEANUP;
3601 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3603 * Splice operation will be punted aync, and here need to
3604 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3606 io_req_init_async(req);
3607 req->work.flags |= IO_WQ_WORK_UNBOUND;
3613 static int io_tee_prep(struct io_kiocb *req,
3614 const struct io_uring_sqe *sqe)
3616 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3618 return __io_splice_prep(req, sqe);
3621 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3623 struct io_splice *sp = &req->splice;
3624 struct file *in = sp->file_in;
3625 struct file *out = sp->file_out;
3626 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3632 ret = do_tee(in, out, sp->len, flags);
3634 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3635 req->flags &= ~REQ_F_NEED_CLEANUP;
3638 req_set_fail_links(req);
3639 io_req_complete(req, ret);
3643 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3645 struct io_splice* sp = &req->splice;
3647 sp->off_in = READ_ONCE(sqe->splice_off_in);
3648 sp->off_out = READ_ONCE(sqe->off);
3649 return __io_splice_prep(req, sqe);
3652 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3654 struct io_splice *sp = &req->splice;
3655 struct file *in = sp->file_in;
3656 struct file *out = sp->file_out;
3657 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3658 loff_t *poff_in, *poff_out;
3664 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3665 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3668 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3670 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3671 req->flags &= ~REQ_F_NEED_CLEANUP;
3674 req_set_fail_links(req);
3675 io_req_complete(req, ret);
3680 * IORING_OP_NOP just posts a completion event, nothing else.
3682 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3684 struct io_ring_ctx *ctx = req->ctx;
3686 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3689 __io_req_complete(req, 0, 0, cs);
3693 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3695 struct io_ring_ctx *ctx = req->ctx;
3700 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3702 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3705 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3706 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3709 req->sync.off = READ_ONCE(sqe->off);
3710 req->sync.len = READ_ONCE(sqe->len);
3714 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3716 loff_t end = req->sync.off + req->sync.len;
3719 /* fsync always requires a blocking context */
3723 ret = vfs_fsync_range(req->file, req->sync.off,
3724 end > 0 ? end : LLONG_MAX,
3725 req->sync.flags & IORING_FSYNC_DATASYNC);
3727 req_set_fail_links(req);
3728 io_req_complete(req, ret);
3732 static int io_fallocate_prep(struct io_kiocb *req,
3733 const struct io_uring_sqe *sqe)
3735 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3737 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3740 req->sync.off = READ_ONCE(sqe->off);
3741 req->sync.len = READ_ONCE(sqe->addr);
3742 req->sync.mode = READ_ONCE(sqe->len);
3746 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3750 /* fallocate always requiring blocking context */
3753 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3756 req_set_fail_links(req);
3757 io_req_complete(req, ret);
3761 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3763 const char __user *fname;
3766 if (unlikely(sqe->ioprio || sqe->buf_index))
3768 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3771 /* open.how should be already initialised */
3772 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3773 req->open.how.flags |= O_LARGEFILE;
3775 req->open.dfd = READ_ONCE(sqe->fd);
3776 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3777 req->open.filename = getname(fname);
3778 if (IS_ERR(req->open.filename)) {
3779 ret = PTR_ERR(req->open.filename);
3780 req->open.filename = NULL;
3783 req->open.nofile = rlimit(RLIMIT_NOFILE);
3784 req->flags |= REQ_F_NEED_CLEANUP;
3788 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3792 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3794 mode = READ_ONCE(sqe->len);
3795 flags = READ_ONCE(sqe->open_flags);
3796 req->open.how = build_open_how(flags, mode);
3797 return __io_openat_prep(req, sqe);
3800 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3802 struct open_how __user *how;
3806 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3808 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3809 len = READ_ONCE(sqe->len);
3810 if (len < OPEN_HOW_SIZE_VER0)
3813 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3818 return __io_openat_prep(req, sqe);
3821 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3823 struct open_flags op;
3830 ret = build_open_flags(&req->open.how, &op);
3834 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3838 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3841 ret = PTR_ERR(file);
3843 fsnotify_open(file);
3844 fd_install(ret, file);
3847 putname(req->open.filename);
3848 req->flags &= ~REQ_F_NEED_CLEANUP;
3850 req_set_fail_links(req);
3851 io_req_complete(req, ret);
3855 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3857 return io_openat2(req, force_nonblock);
3860 static int io_remove_buffers_prep(struct io_kiocb *req,
3861 const struct io_uring_sqe *sqe)
3863 struct io_provide_buf *p = &req->pbuf;
3866 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3869 tmp = READ_ONCE(sqe->fd);
3870 if (!tmp || tmp > USHRT_MAX)
3873 memset(p, 0, sizeof(*p));
3875 p->bgid = READ_ONCE(sqe->buf_group);
3879 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3880 int bgid, unsigned nbufs)
3884 /* shouldn't happen */
3888 /* the head kbuf is the list itself */
3889 while (!list_empty(&buf->list)) {
3890 struct io_buffer *nxt;
3892 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3893 list_del(&nxt->list);
3900 idr_remove(&ctx->io_buffer_idr, bgid);
3905 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3906 struct io_comp_state *cs)
3908 struct io_provide_buf *p = &req->pbuf;
3909 struct io_ring_ctx *ctx = req->ctx;
3910 struct io_buffer *head;
3913 io_ring_submit_lock(ctx, !force_nonblock);
3915 lockdep_assert_held(&ctx->uring_lock);
3918 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3920 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3922 io_ring_submit_lock(ctx, !force_nonblock);
3924 req_set_fail_links(req);
3925 __io_req_complete(req, ret, 0, cs);
3929 static int io_provide_buffers_prep(struct io_kiocb *req,
3930 const struct io_uring_sqe *sqe)
3932 struct io_provide_buf *p = &req->pbuf;
3935 if (sqe->ioprio || sqe->rw_flags)
3938 tmp = READ_ONCE(sqe->fd);
3939 if (!tmp || tmp > USHRT_MAX)
3942 p->addr = READ_ONCE(sqe->addr);
3943 p->len = READ_ONCE(sqe->len);
3945 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3948 p->bgid = READ_ONCE(sqe->buf_group);
3949 tmp = READ_ONCE(sqe->off);
3950 if (tmp > USHRT_MAX)
3956 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3958 struct io_buffer *buf;
3959 u64 addr = pbuf->addr;
3960 int i, bid = pbuf->bid;
3962 for (i = 0; i < pbuf->nbufs; i++) {
3963 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3968 buf->len = pbuf->len;
3973 INIT_LIST_HEAD(&buf->list);
3976 list_add_tail(&buf->list, &(*head)->list);
3980 return i ? i : -ENOMEM;
3983 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3984 struct io_comp_state *cs)
3986 struct io_provide_buf *p = &req->pbuf;
3987 struct io_ring_ctx *ctx = req->ctx;
3988 struct io_buffer *head, *list;
3991 io_ring_submit_lock(ctx, !force_nonblock);
3993 lockdep_assert_held(&ctx->uring_lock);
3995 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3997 ret = io_add_buffers(p, &head);
4002 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4005 __io_remove_buffers(ctx, head, p->bgid, -1U);
4010 io_ring_submit_unlock(ctx, !force_nonblock);
4012 req_set_fail_links(req);
4013 __io_req_complete(req, ret, 0, cs);
4017 static int io_epoll_ctl_prep(struct io_kiocb *req,
4018 const struct io_uring_sqe *sqe)
4020 #if defined(CONFIG_EPOLL)
4021 if (sqe->ioprio || sqe->buf_index)
4023 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4026 req->epoll.epfd = READ_ONCE(sqe->fd);
4027 req->epoll.op = READ_ONCE(sqe->len);
4028 req->epoll.fd = READ_ONCE(sqe->off);
4030 if (ep_op_has_event(req->epoll.op)) {
4031 struct epoll_event __user *ev;
4033 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4034 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4044 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4045 struct io_comp_state *cs)
4047 #if defined(CONFIG_EPOLL)
4048 struct io_epoll *ie = &req->epoll;
4051 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4052 if (force_nonblock && ret == -EAGAIN)
4056 req_set_fail_links(req);
4057 __io_req_complete(req, ret, 0, cs);
4064 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4066 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4067 if (sqe->ioprio || sqe->buf_index || sqe->off)
4069 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4072 req->madvise.addr = READ_ONCE(sqe->addr);
4073 req->madvise.len = READ_ONCE(sqe->len);
4074 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4081 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4083 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4084 struct io_madvise *ma = &req->madvise;
4090 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4092 req_set_fail_links(req);
4093 io_req_complete(req, ret);
4100 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4102 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4104 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4107 req->fadvise.offset = READ_ONCE(sqe->off);
4108 req->fadvise.len = READ_ONCE(sqe->len);
4109 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4113 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4115 struct io_fadvise *fa = &req->fadvise;
4118 if (force_nonblock) {
4119 switch (fa->advice) {
4120 case POSIX_FADV_NORMAL:
4121 case POSIX_FADV_RANDOM:
4122 case POSIX_FADV_SEQUENTIAL:
4129 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4131 req_set_fail_links(req);
4132 io_req_complete(req, ret);
4136 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4138 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4140 if (sqe->ioprio || sqe->buf_index)
4142 if (req->flags & REQ_F_FIXED_FILE)
4145 req->statx.dfd = READ_ONCE(sqe->fd);
4146 req->statx.mask = READ_ONCE(sqe->len);
4147 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4148 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4149 req->statx.flags = READ_ONCE(sqe->statx_flags);
4154 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4156 struct io_statx *ctx = &req->statx;
4159 if (force_nonblock) {
4160 /* only need file table for an actual valid fd */
4161 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4162 req->flags |= REQ_F_NO_FILE_TABLE;
4166 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4170 req_set_fail_links(req);
4171 io_req_complete(req, ret);
4175 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4178 * If we queue this for async, it must not be cancellable. That would
4179 * leave the 'file' in an undeterminate state, and here need to modify
4180 * io_wq_work.flags, so initialize io_wq_work firstly.
4182 io_req_init_async(req);
4183 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4185 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4187 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4188 sqe->rw_flags || sqe->buf_index)
4190 if (req->flags & REQ_F_FIXED_FILE)
4193 req->close.fd = READ_ONCE(sqe->fd);
4194 if ((req->file && req->file->f_op == &io_uring_fops))
4197 req->close.put_file = NULL;
4201 static int io_close(struct io_kiocb *req, bool force_nonblock,
4202 struct io_comp_state *cs)
4204 struct io_close *close = &req->close;
4207 /* might be already done during nonblock submission */
4208 if (!close->put_file) {
4209 ret = __close_fd_get_file(close->fd, &close->put_file);
4211 return (ret == -ENOENT) ? -EBADF : ret;
4214 /* if the file has a flush method, be safe and punt to async */
4215 if (close->put_file->f_op->flush && force_nonblock) {
4216 /* was never set, but play safe */
4217 req->flags &= ~REQ_F_NOWAIT;
4218 /* avoid grabbing files - we don't need the files */
4219 req->flags |= REQ_F_NO_FILE_TABLE;
4223 /* No ->flush() or already async, safely close from here */
4224 ret = filp_close(close->put_file, req->work.identity->files);
4226 req_set_fail_links(req);
4227 fput(close->put_file);
4228 close->put_file = NULL;
4229 __io_req_complete(req, ret, 0, cs);
4233 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4235 struct io_ring_ctx *ctx = req->ctx;
4240 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4242 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4245 req->sync.off = READ_ONCE(sqe->off);
4246 req->sync.len = READ_ONCE(sqe->len);
4247 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4251 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4255 /* sync_file_range always requires a blocking context */
4259 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4262 req_set_fail_links(req);
4263 io_req_complete(req, ret);
4267 #if defined(CONFIG_NET)
4268 static int io_setup_async_msg(struct io_kiocb *req,
4269 struct io_async_msghdr *kmsg)
4271 struct io_async_msghdr *async_msg = req->async_data;
4275 if (io_alloc_async_data(req)) {
4276 if (kmsg->iov != kmsg->fast_iov)
4280 async_msg = req->async_data;
4281 req->flags |= REQ_F_NEED_CLEANUP;
4282 memcpy(async_msg, kmsg, sizeof(*kmsg));
4286 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4287 struct io_async_msghdr *iomsg)
4289 iomsg->iov = iomsg->fast_iov;
4290 iomsg->msg.msg_name = &iomsg->addr;
4291 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4292 req->sr_msg.msg_flags, &iomsg->iov);
4295 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4297 struct io_async_msghdr *async_msg = req->async_data;
4298 struct io_sr_msg *sr = &req->sr_msg;
4301 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4304 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4305 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4306 sr->len = READ_ONCE(sqe->len);
4308 #ifdef CONFIG_COMPAT
4309 if (req->ctx->compat)
4310 sr->msg_flags |= MSG_CMSG_COMPAT;
4313 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4315 ret = io_sendmsg_copy_hdr(req, async_msg);
4317 req->flags |= REQ_F_NEED_CLEANUP;
4321 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4322 struct io_comp_state *cs)
4324 struct io_async_msghdr iomsg, *kmsg;
4325 struct socket *sock;
4329 sock = sock_from_file(req->file, &ret);
4330 if (unlikely(!sock))
4333 if (req->async_data) {
4334 kmsg = req->async_data;
4335 kmsg->msg.msg_name = &kmsg->addr;
4336 /* if iov is set, it's allocated already */
4338 kmsg->iov = kmsg->fast_iov;
4339 kmsg->msg.msg_iter.iov = kmsg->iov;
4341 ret = io_sendmsg_copy_hdr(req, &iomsg);
4347 flags = req->sr_msg.msg_flags;
4348 if (flags & MSG_DONTWAIT)
4349 req->flags |= REQ_F_NOWAIT;
4350 else if (force_nonblock)
4351 flags |= MSG_DONTWAIT;
4353 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4354 if (force_nonblock && ret == -EAGAIN)
4355 return io_setup_async_msg(req, kmsg);
4356 if (ret == -ERESTARTSYS)
4359 if (kmsg->iov != kmsg->fast_iov)
4361 req->flags &= ~REQ_F_NEED_CLEANUP;
4363 req_set_fail_links(req);
4364 __io_req_complete(req, ret, 0, cs);
4368 static int io_send(struct io_kiocb *req, bool force_nonblock,
4369 struct io_comp_state *cs)
4371 struct io_sr_msg *sr = &req->sr_msg;
4374 struct socket *sock;
4378 sock = sock_from_file(req->file, &ret);
4379 if (unlikely(!sock))
4382 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4386 msg.msg_name = NULL;
4387 msg.msg_control = NULL;
4388 msg.msg_controllen = 0;
4389 msg.msg_namelen = 0;
4391 flags = req->sr_msg.msg_flags;
4392 if (flags & MSG_DONTWAIT)
4393 req->flags |= REQ_F_NOWAIT;
4394 else if (force_nonblock)
4395 flags |= MSG_DONTWAIT;
4397 msg.msg_flags = flags;
4398 ret = sock_sendmsg(sock, &msg);
4399 if (force_nonblock && ret == -EAGAIN)
4401 if (ret == -ERESTARTSYS)
4405 req_set_fail_links(req);
4406 __io_req_complete(req, ret, 0, cs);
4410 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4411 struct io_async_msghdr *iomsg)
4413 struct io_sr_msg *sr = &req->sr_msg;
4414 struct iovec __user *uiov;
4418 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4419 &iomsg->uaddr, &uiov, &iov_len);
4423 if (req->flags & REQ_F_BUFFER_SELECT) {
4426 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4428 sr->len = iomsg->iov[0].iov_len;
4429 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4433 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4434 &iomsg->iov, &iomsg->msg.msg_iter,
4443 #ifdef CONFIG_COMPAT
4444 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4445 struct io_async_msghdr *iomsg)
4447 struct compat_msghdr __user *msg_compat;
4448 struct io_sr_msg *sr = &req->sr_msg;
4449 struct compat_iovec __user *uiov;
4454 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4455 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4460 uiov = compat_ptr(ptr);
4461 if (req->flags & REQ_F_BUFFER_SELECT) {
4462 compat_ssize_t clen;
4466 if (!access_ok(uiov, sizeof(*uiov)))
4468 if (__get_user(clen, &uiov->iov_len))
4472 sr->len = iomsg->iov[0].iov_len;
4475 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4476 UIO_FASTIOV, &iomsg->iov,
4477 &iomsg->msg.msg_iter, true);
4486 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4487 struct io_async_msghdr *iomsg)
4489 iomsg->msg.msg_name = &iomsg->addr;
4490 iomsg->iov = iomsg->fast_iov;
4492 #ifdef CONFIG_COMPAT
4493 if (req->ctx->compat)
4494 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4497 return __io_recvmsg_copy_hdr(req, iomsg);
4500 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4503 struct io_sr_msg *sr = &req->sr_msg;
4504 struct io_buffer *kbuf;
4506 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4511 req->flags |= REQ_F_BUFFER_SELECTED;
4515 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4517 return io_put_kbuf(req, req->sr_msg.kbuf);
4520 static int io_recvmsg_prep(struct io_kiocb *req,
4521 const struct io_uring_sqe *sqe)
4523 struct io_async_msghdr *async_msg = req->async_data;
4524 struct io_sr_msg *sr = &req->sr_msg;
4527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4530 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4531 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4532 sr->len = READ_ONCE(sqe->len);
4533 sr->bgid = READ_ONCE(sqe->buf_group);
4535 #ifdef CONFIG_COMPAT
4536 if (req->ctx->compat)
4537 sr->msg_flags |= MSG_CMSG_COMPAT;
4540 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4542 ret = io_recvmsg_copy_hdr(req, async_msg);
4544 req->flags |= REQ_F_NEED_CLEANUP;
4548 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4549 struct io_comp_state *cs)
4551 struct io_async_msghdr iomsg, *kmsg;
4552 struct socket *sock;
4553 struct io_buffer *kbuf;
4555 int ret, cflags = 0;
4557 sock = sock_from_file(req->file, &ret);
4558 if (unlikely(!sock))
4561 if (req->async_data) {
4562 kmsg = req->async_data;
4563 kmsg->msg.msg_name = &kmsg->addr;
4564 /* if iov is set, it's allocated already */
4566 kmsg->iov = kmsg->fast_iov;
4567 kmsg->msg.msg_iter.iov = kmsg->iov;
4569 ret = io_recvmsg_copy_hdr(req, &iomsg);
4575 if (req->flags & REQ_F_BUFFER_SELECT) {
4576 kbuf = io_recv_buffer_select(req, !force_nonblock);
4578 return PTR_ERR(kbuf);
4579 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4580 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4581 1, req->sr_msg.len);
4584 flags = req->sr_msg.msg_flags;
4585 if (flags & MSG_DONTWAIT)
4586 req->flags |= REQ_F_NOWAIT;
4587 else if (force_nonblock)
4588 flags |= MSG_DONTWAIT;
4590 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4591 kmsg->uaddr, flags);
4592 if (force_nonblock && ret == -EAGAIN)
4593 return io_setup_async_msg(req, kmsg);
4594 if (ret == -ERESTARTSYS)
4597 if (req->flags & REQ_F_BUFFER_SELECTED)
4598 cflags = io_put_recv_kbuf(req);
4599 if (kmsg->iov != kmsg->fast_iov)
4601 req->flags &= ~REQ_F_NEED_CLEANUP;
4603 req_set_fail_links(req);
4604 __io_req_complete(req, ret, cflags, cs);
4608 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4609 struct io_comp_state *cs)
4611 struct io_buffer *kbuf;
4612 struct io_sr_msg *sr = &req->sr_msg;
4614 void __user *buf = sr->buf;
4615 struct socket *sock;
4618 int ret, cflags = 0;
4620 sock = sock_from_file(req->file, &ret);
4621 if (unlikely(!sock))
4624 if (req->flags & REQ_F_BUFFER_SELECT) {
4625 kbuf = io_recv_buffer_select(req, !force_nonblock);
4627 return PTR_ERR(kbuf);
4628 buf = u64_to_user_ptr(kbuf->addr);
4631 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4635 msg.msg_name = NULL;
4636 msg.msg_control = NULL;
4637 msg.msg_controllen = 0;
4638 msg.msg_namelen = 0;
4639 msg.msg_iocb = NULL;
4642 flags = req->sr_msg.msg_flags;
4643 if (flags & MSG_DONTWAIT)
4644 req->flags |= REQ_F_NOWAIT;
4645 else if (force_nonblock)
4646 flags |= MSG_DONTWAIT;
4648 ret = sock_recvmsg(sock, &msg, flags);
4649 if (force_nonblock && ret == -EAGAIN)
4651 if (ret == -ERESTARTSYS)
4654 if (req->flags & REQ_F_BUFFER_SELECTED)
4655 cflags = io_put_recv_kbuf(req);
4657 req_set_fail_links(req);
4658 __io_req_complete(req, ret, cflags, cs);
4662 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4664 struct io_accept *accept = &req->accept;
4666 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4668 if (sqe->ioprio || sqe->len || sqe->buf_index)
4671 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4672 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4673 accept->flags = READ_ONCE(sqe->accept_flags);
4674 accept->nofile = rlimit(RLIMIT_NOFILE);
4678 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4679 struct io_comp_state *cs)
4681 struct io_accept *accept = &req->accept;
4682 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4685 if (req->file->f_flags & O_NONBLOCK)
4686 req->flags |= REQ_F_NOWAIT;
4688 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4689 accept->addr_len, accept->flags,
4691 if (ret == -EAGAIN && force_nonblock)
4694 if (ret == -ERESTARTSYS)
4696 req_set_fail_links(req);
4698 __io_req_complete(req, ret, 0, cs);
4702 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4704 struct io_connect *conn = &req->connect;
4705 struct io_async_connect *io = req->async_data;
4707 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4709 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4712 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4713 conn->addr_len = READ_ONCE(sqe->addr2);
4718 return move_addr_to_kernel(conn->addr, conn->addr_len,
4722 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4723 struct io_comp_state *cs)
4725 struct io_async_connect __io, *io;
4726 unsigned file_flags;
4729 if (req->async_data) {
4730 io = req->async_data;
4732 ret = move_addr_to_kernel(req->connect.addr,
4733 req->connect.addr_len,
4740 file_flags = force_nonblock ? O_NONBLOCK : 0;
4742 ret = __sys_connect_file(req->file, &io->address,
4743 req->connect.addr_len, file_flags);
4744 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4745 if (req->async_data)
4747 if (io_alloc_async_data(req)) {
4751 io = req->async_data;
4752 memcpy(req->async_data, &__io, sizeof(__io));
4755 if (ret == -ERESTARTSYS)
4759 req_set_fail_links(req);
4760 __io_req_complete(req, ret, 0, cs);
4763 #else /* !CONFIG_NET */
4764 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4769 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4770 struct io_comp_state *cs)
4775 static int io_send(struct io_kiocb *req, bool force_nonblock,
4776 struct io_comp_state *cs)
4781 static int io_recvmsg_prep(struct io_kiocb *req,
4782 const struct io_uring_sqe *sqe)
4787 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4788 struct io_comp_state *cs)
4793 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4794 struct io_comp_state *cs)
4799 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4804 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4805 struct io_comp_state *cs)
4810 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4815 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4816 struct io_comp_state *cs)
4820 #endif /* CONFIG_NET */
4822 struct io_poll_table {
4823 struct poll_table_struct pt;
4824 struct io_kiocb *req;
4828 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4829 __poll_t mask, task_work_func_t func)
4834 /* for instances that support it check for an event match first: */
4835 if (mask && !(mask & poll->events))
4838 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4840 list_del_init(&poll->wait.entry);
4843 init_task_work(&req->task_work, func);
4844 percpu_ref_get(&req->ctx->refs);
4847 * If we using the signalfd wait_queue_head for this wakeup, then
4848 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4849 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4850 * either, as the normal wakeup will suffice.
4852 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4855 * If this fails, then the task is exiting. When a task exits, the
4856 * work gets canceled, so just cancel this request as well instead
4857 * of executing it. We can't safely execute it anyway, as we may not
4858 * have the needed state needed for it anyway.
4860 ret = io_req_task_work_add(req, twa_signal_ok);
4861 if (unlikely(ret)) {
4862 struct task_struct *tsk;
4864 WRITE_ONCE(poll->canceled, true);
4865 tsk = io_wq_get_task(req->ctx->io_wq);
4866 task_work_add(tsk, &req->task_work, TWA_NONE);
4867 wake_up_process(tsk);
4872 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4873 __acquires(&req->ctx->completion_lock)
4875 struct io_ring_ctx *ctx = req->ctx;
4877 if (!req->result && !READ_ONCE(poll->canceled)) {
4878 struct poll_table_struct pt = { ._key = poll->events };
4880 req->result = vfs_poll(req->file, &pt) & poll->events;
4883 spin_lock_irq(&ctx->completion_lock);
4884 if (!req->result && !READ_ONCE(poll->canceled)) {
4885 add_wait_queue(poll->head, &poll->wait);
4892 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4894 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4895 if (req->opcode == IORING_OP_POLL_ADD)
4896 return req->async_data;
4897 return req->apoll->double_poll;
4900 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4902 if (req->opcode == IORING_OP_POLL_ADD)
4904 return &req->apoll->poll;
4907 static void io_poll_remove_double(struct io_kiocb *req)
4909 struct io_poll_iocb *poll = io_poll_get_double(req);
4911 lockdep_assert_held(&req->ctx->completion_lock);
4913 if (poll && poll->head) {
4914 struct wait_queue_head *head = poll->head;
4916 spin_lock(&head->lock);
4917 list_del_init(&poll->wait.entry);
4918 if (poll->wait.private)
4919 refcount_dec(&req->refs);
4921 spin_unlock(&head->lock);
4925 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4927 struct io_ring_ctx *ctx = req->ctx;
4929 io_poll_remove_double(req);
4930 req->poll.done = true;
4931 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4932 io_commit_cqring(ctx);
4935 static void io_poll_task_func(struct callback_head *cb)
4937 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4938 struct io_ring_ctx *ctx = req->ctx;
4939 struct io_kiocb *nxt;
4941 if (io_poll_rewait(req, &req->poll)) {
4942 spin_unlock_irq(&ctx->completion_lock);
4944 hash_del(&req->hash_node);
4945 io_poll_complete(req, req->result, 0);
4946 spin_unlock_irq(&ctx->completion_lock);
4948 nxt = io_put_req_find_next(req);
4949 io_cqring_ev_posted(ctx);
4951 __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 INIT_HLIST_NODE(&req->hash_node);
5108 io_init_poll_iocb(poll, mask, wake_func);
5109 poll->file = req->file;
5110 poll->wait.private = req;
5112 ipt->pt._key = mask;
5114 ipt->error = -EINVAL;
5116 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5118 spin_lock_irq(&ctx->completion_lock);
5119 if (likely(poll->head)) {
5120 spin_lock(&poll->head->lock);
5121 if (unlikely(list_empty(&poll->wait.entry))) {
5127 if (mask || ipt->error)
5128 list_del_init(&poll->wait.entry);
5130 WRITE_ONCE(poll->canceled, true);
5131 else if (!poll->done) /* actually waiting for an event */
5132 io_poll_req_insert(req);
5133 spin_unlock(&poll->head->lock);
5139 static bool io_arm_poll_handler(struct io_kiocb *req)
5141 const struct io_op_def *def = &io_op_defs[req->opcode];
5142 struct io_ring_ctx *ctx = req->ctx;
5143 struct async_poll *apoll;
5144 struct io_poll_table ipt;
5148 if (!req->file || !file_can_poll(req->file))
5150 if (req->flags & REQ_F_POLLED)
5154 else if (def->pollout)
5158 /* if we can't nonblock try, then no point in arming a poll handler */
5159 if (!io_file_supports_async(req->file, rw))
5162 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5163 if (unlikely(!apoll))
5165 apoll->double_poll = NULL;
5167 req->flags |= REQ_F_POLLED;
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 ipt.pt._qproc = io_poll_queue_proc;
5369 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5372 if (mask) { /* no async, we'd stolen it */
5374 io_poll_complete(req, mask, 0);
5376 spin_unlock_irq(&ctx->completion_lock);
5379 io_cqring_ev_posted(ctx);
5385 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5387 struct io_timeout_data *data = container_of(timer,
5388 struct io_timeout_data, timer);
5389 struct io_kiocb *req = data->req;
5390 struct io_ring_ctx *ctx = req->ctx;
5391 unsigned long flags;
5393 spin_lock_irqsave(&ctx->completion_lock, flags);
5394 list_del_init(&req->timeout.list);
5395 atomic_set(&req->ctx->cq_timeouts,
5396 atomic_read(&req->ctx->cq_timeouts) + 1);
5398 io_cqring_fill_event(req, -ETIME);
5399 io_commit_cqring(ctx);
5400 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5402 io_cqring_ev_posted(ctx);
5403 req_set_fail_links(req);
5405 return HRTIMER_NORESTART;
5408 static int __io_timeout_cancel(struct io_kiocb *req)
5410 struct io_timeout_data *io = req->async_data;
5413 ret = hrtimer_try_to_cancel(&io->timer);
5416 list_del_init(&req->timeout.list);
5418 req_set_fail_links(req);
5419 io_cqring_fill_event(req, -ECANCELED);
5420 io_put_req_deferred(req, 1);
5424 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5426 struct io_kiocb *req;
5429 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5430 if (user_data == req->user_data) {
5439 return __io_timeout_cancel(req);
5442 static int io_timeout_remove_prep(struct io_kiocb *req,
5443 const struct io_uring_sqe *sqe)
5445 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5447 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5449 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5452 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5457 * Remove or update an existing timeout command
5459 static int io_timeout_remove(struct io_kiocb *req)
5461 struct io_ring_ctx *ctx = req->ctx;
5464 spin_lock_irq(&ctx->completion_lock);
5465 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5467 io_cqring_fill_event(req, ret);
5468 io_commit_cqring(ctx);
5469 spin_unlock_irq(&ctx->completion_lock);
5470 io_cqring_ev_posted(ctx);
5472 req_set_fail_links(req);
5477 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5478 bool is_timeout_link)
5480 struct io_timeout_data *data;
5482 u32 off = READ_ONCE(sqe->off);
5484 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5486 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5488 if (off && is_timeout_link)
5490 flags = READ_ONCE(sqe->timeout_flags);
5491 if (flags & ~IORING_TIMEOUT_ABS)
5494 req->timeout.off = off;
5496 if (!req->async_data && io_alloc_async_data(req))
5499 data = req->async_data;
5502 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5505 if (flags & IORING_TIMEOUT_ABS)
5506 data->mode = HRTIMER_MODE_ABS;
5508 data->mode = HRTIMER_MODE_REL;
5510 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5514 static int io_timeout(struct io_kiocb *req)
5516 struct io_ring_ctx *ctx = req->ctx;
5517 struct io_timeout_data *data = req->async_data;
5518 struct list_head *entry;
5519 u32 tail, off = req->timeout.off;
5521 spin_lock_irq(&ctx->completion_lock);
5524 * sqe->off holds how many events that need to occur for this
5525 * timeout event to be satisfied. If it isn't set, then this is
5526 * a pure timeout request, sequence isn't used.
5528 if (io_is_timeout_noseq(req)) {
5529 entry = ctx->timeout_list.prev;
5533 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5534 req->timeout.target_seq = tail + off;
5537 * Insertion sort, ensuring the first entry in the list is always
5538 * the one we need first.
5540 list_for_each_prev(entry, &ctx->timeout_list) {
5541 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5544 if (io_is_timeout_noseq(nxt))
5546 /* nxt.seq is behind @tail, otherwise would've been completed */
5547 if (off >= nxt->timeout.target_seq - tail)
5551 list_add(&req->timeout.list, entry);
5552 data->timer.function = io_timeout_fn;
5553 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5554 spin_unlock_irq(&ctx->completion_lock);
5558 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5560 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5562 return req->user_data == (unsigned long) data;
5565 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5567 enum io_wq_cancel cancel_ret;
5570 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5571 switch (cancel_ret) {
5572 case IO_WQ_CANCEL_OK:
5575 case IO_WQ_CANCEL_RUNNING:
5578 case IO_WQ_CANCEL_NOTFOUND:
5586 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5587 struct io_kiocb *req, __u64 sqe_addr,
5590 unsigned long flags;
5593 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5594 if (ret != -ENOENT) {
5595 spin_lock_irqsave(&ctx->completion_lock, flags);
5599 spin_lock_irqsave(&ctx->completion_lock, flags);
5600 ret = io_timeout_cancel(ctx, sqe_addr);
5603 ret = io_poll_cancel(ctx, sqe_addr);
5607 io_cqring_fill_event(req, ret);
5608 io_commit_cqring(ctx);
5609 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5610 io_cqring_ev_posted(ctx);
5613 req_set_fail_links(req);
5617 static int io_async_cancel_prep(struct io_kiocb *req,
5618 const struct io_uring_sqe *sqe)
5620 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5622 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5624 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5627 req->cancel.addr = READ_ONCE(sqe->addr);
5631 static int io_async_cancel(struct io_kiocb *req)
5633 struct io_ring_ctx *ctx = req->ctx;
5635 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5639 static int io_files_update_prep(struct io_kiocb *req,
5640 const struct io_uring_sqe *sqe)
5642 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5644 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5646 if (sqe->ioprio || sqe->rw_flags)
5649 req->files_update.offset = READ_ONCE(sqe->off);
5650 req->files_update.nr_args = READ_ONCE(sqe->len);
5651 if (!req->files_update.nr_args)
5653 req->files_update.arg = READ_ONCE(sqe->addr);
5657 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5658 struct io_comp_state *cs)
5660 struct io_ring_ctx *ctx = req->ctx;
5661 struct io_uring_files_update up;
5667 up.offset = req->files_update.offset;
5668 up.fds = req->files_update.arg;
5670 mutex_lock(&ctx->uring_lock);
5671 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5672 mutex_unlock(&ctx->uring_lock);
5675 req_set_fail_links(req);
5676 __io_req_complete(req, ret, 0, cs);
5680 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5682 switch (req->opcode) {
5685 case IORING_OP_READV:
5686 case IORING_OP_READ_FIXED:
5687 case IORING_OP_READ:
5688 return io_read_prep(req, sqe);
5689 case IORING_OP_WRITEV:
5690 case IORING_OP_WRITE_FIXED:
5691 case IORING_OP_WRITE:
5692 return io_write_prep(req, sqe);
5693 case IORING_OP_POLL_ADD:
5694 return io_poll_add_prep(req, sqe);
5695 case IORING_OP_POLL_REMOVE:
5696 return io_poll_remove_prep(req, sqe);
5697 case IORING_OP_FSYNC:
5698 return io_prep_fsync(req, sqe);
5699 case IORING_OP_SYNC_FILE_RANGE:
5700 return io_prep_sfr(req, sqe);
5701 case IORING_OP_SENDMSG:
5702 case IORING_OP_SEND:
5703 return io_sendmsg_prep(req, sqe);
5704 case IORING_OP_RECVMSG:
5705 case IORING_OP_RECV:
5706 return io_recvmsg_prep(req, sqe);
5707 case IORING_OP_CONNECT:
5708 return io_connect_prep(req, sqe);
5709 case IORING_OP_TIMEOUT:
5710 return io_timeout_prep(req, sqe, false);
5711 case IORING_OP_TIMEOUT_REMOVE:
5712 return io_timeout_remove_prep(req, sqe);
5713 case IORING_OP_ASYNC_CANCEL:
5714 return io_async_cancel_prep(req, sqe);
5715 case IORING_OP_LINK_TIMEOUT:
5716 return io_timeout_prep(req, sqe, true);
5717 case IORING_OP_ACCEPT:
5718 return io_accept_prep(req, sqe);
5719 case IORING_OP_FALLOCATE:
5720 return io_fallocate_prep(req, sqe);
5721 case IORING_OP_OPENAT:
5722 return io_openat_prep(req, sqe);
5723 case IORING_OP_CLOSE:
5724 return io_close_prep(req, sqe);
5725 case IORING_OP_FILES_UPDATE:
5726 return io_files_update_prep(req, sqe);
5727 case IORING_OP_STATX:
5728 return io_statx_prep(req, sqe);
5729 case IORING_OP_FADVISE:
5730 return io_fadvise_prep(req, sqe);
5731 case IORING_OP_MADVISE:
5732 return io_madvise_prep(req, sqe);
5733 case IORING_OP_OPENAT2:
5734 return io_openat2_prep(req, sqe);
5735 case IORING_OP_EPOLL_CTL:
5736 return io_epoll_ctl_prep(req, sqe);
5737 case IORING_OP_SPLICE:
5738 return io_splice_prep(req, sqe);
5739 case IORING_OP_PROVIDE_BUFFERS:
5740 return io_provide_buffers_prep(req, sqe);
5741 case IORING_OP_REMOVE_BUFFERS:
5742 return io_remove_buffers_prep(req, sqe);
5744 return io_tee_prep(req, sqe);
5747 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5752 static int io_req_defer_prep(struct io_kiocb *req,
5753 const struct io_uring_sqe *sqe)
5757 if (io_alloc_async_data(req))
5759 return io_req_prep(req, sqe);
5762 static u32 io_get_sequence(struct io_kiocb *req)
5764 struct io_kiocb *pos;
5765 struct io_ring_ctx *ctx = req->ctx;
5766 u32 total_submitted, nr_reqs = 1;
5768 if (req->flags & REQ_F_LINK_HEAD)
5769 list_for_each_entry(pos, &req->link_list, link_list)
5772 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5773 return total_submitted - nr_reqs;
5776 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5778 struct io_ring_ctx *ctx = req->ctx;
5779 struct io_defer_entry *de;
5783 /* Still need defer if there is pending req in defer list. */
5784 if (likely(list_empty_careful(&ctx->defer_list) &&
5785 !(req->flags & REQ_F_IO_DRAIN)))
5788 seq = io_get_sequence(req);
5789 /* Still a chance to pass the sequence check */
5790 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5793 if (!req->async_data) {
5794 ret = io_req_defer_prep(req, sqe);
5798 io_prep_async_link(req);
5799 de = kmalloc(sizeof(*de), GFP_KERNEL);
5803 spin_lock_irq(&ctx->completion_lock);
5804 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5805 spin_unlock_irq(&ctx->completion_lock);
5807 io_queue_async_work(req);
5808 return -EIOCBQUEUED;
5811 trace_io_uring_defer(ctx, req, req->user_data);
5814 list_add_tail(&de->list, &ctx->defer_list);
5815 spin_unlock_irq(&ctx->completion_lock);
5816 return -EIOCBQUEUED;
5819 static void io_req_drop_files(struct io_kiocb *req)
5821 struct io_ring_ctx *ctx = req->ctx;
5822 unsigned long flags;
5824 spin_lock_irqsave(&ctx->inflight_lock, flags);
5825 list_del(&req->inflight_entry);
5826 if (waitqueue_active(&ctx->inflight_wait))
5827 wake_up(&ctx->inflight_wait);
5828 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5829 req->flags &= ~REQ_F_INFLIGHT;
5830 put_files_struct(req->work.identity->files);
5831 put_nsproxy(req->work.identity->nsproxy);
5832 req->work.flags &= ~IO_WQ_WORK_FILES;
5835 static void __io_clean_op(struct io_kiocb *req)
5837 if (req->flags & REQ_F_BUFFER_SELECTED) {
5838 switch (req->opcode) {
5839 case IORING_OP_READV:
5840 case IORING_OP_READ_FIXED:
5841 case IORING_OP_READ:
5842 kfree((void *)(unsigned long)req->rw.addr);
5844 case IORING_OP_RECVMSG:
5845 case IORING_OP_RECV:
5846 kfree(req->sr_msg.kbuf);
5849 req->flags &= ~REQ_F_BUFFER_SELECTED;
5852 if (req->flags & REQ_F_NEED_CLEANUP) {
5853 switch (req->opcode) {
5854 case IORING_OP_READV:
5855 case IORING_OP_READ_FIXED:
5856 case IORING_OP_READ:
5857 case IORING_OP_WRITEV:
5858 case IORING_OP_WRITE_FIXED:
5859 case IORING_OP_WRITE: {
5860 struct io_async_rw *io = req->async_data;
5862 kfree(io->free_iovec);
5865 case IORING_OP_RECVMSG:
5866 case IORING_OP_SENDMSG: {
5867 struct io_async_msghdr *io = req->async_data;
5868 if (io->iov != io->fast_iov)
5872 case IORING_OP_SPLICE:
5874 io_put_file(req, req->splice.file_in,
5875 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5877 case IORING_OP_OPENAT:
5878 case IORING_OP_OPENAT2:
5879 if (req->open.filename)
5880 putname(req->open.filename);
5883 req->flags &= ~REQ_F_NEED_CLEANUP;
5886 if (req->flags & REQ_F_INFLIGHT)
5887 io_req_drop_files(req);
5890 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5891 struct io_comp_state *cs)
5893 struct io_ring_ctx *ctx = req->ctx;
5896 switch (req->opcode) {
5898 ret = io_nop(req, cs);
5900 case IORING_OP_READV:
5901 case IORING_OP_READ_FIXED:
5902 case IORING_OP_READ:
5903 ret = io_read(req, force_nonblock, cs);
5905 case IORING_OP_WRITEV:
5906 case IORING_OP_WRITE_FIXED:
5907 case IORING_OP_WRITE:
5908 ret = io_write(req, force_nonblock, cs);
5910 case IORING_OP_FSYNC:
5911 ret = io_fsync(req, force_nonblock);
5913 case IORING_OP_POLL_ADD:
5914 ret = io_poll_add(req);
5916 case IORING_OP_POLL_REMOVE:
5917 ret = io_poll_remove(req);
5919 case IORING_OP_SYNC_FILE_RANGE:
5920 ret = io_sync_file_range(req, force_nonblock);
5922 case IORING_OP_SENDMSG:
5923 ret = io_sendmsg(req, force_nonblock, cs);
5925 case IORING_OP_SEND:
5926 ret = io_send(req, force_nonblock, cs);
5928 case IORING_OP_RECVMSG:
5929 ret = io_recvmsg(req, force_nonblock, cs);
5931 case IORING_OP_RECV:
5932 ret = io_recv(req, force_nonblock, cs);
5934 case IORING_OP_TIMEOUT:
5935 ret = io_timeout(req);
5937 case IORING_OP_TIMEOUT_REMOVE:
5938 ret = io_timeout_remove(req);
5940 case IORING_OP_ACCEPT:
5941 ret = io_accept(req, force_nonblock, cs);
5943 case IORING_OP_CONNECT:
5944 ret = io_connect(req, force_nonblock, cs);
5946 case IORING_OP_ASYNC_CANCEL:
5947 ret = io_async_cancel(req);
5949 case IORING_OP_FALLOCATE:
5950 ret = io_fallocate(req, force_nonblock);
5952 case IORING_OP_OPENAT:
5953 ret = io_openat(req, force_nonblock);
5955 case IORING_OP_CLOSE:
5956 ret = io_close(req, force_nonblock, cs);
5958 case IORING_OP_FILES_UPDATE:
5959 ret = io_files_update(req, force_nonblock, cs);
5961 case IORING_OP_STATX:
5962 ret = io_statx(req, force_nonblock);
5964 case IORING_OP_FADVISE:
5965 ret = io_fadvise(req, force_nonblock);
5967 case IORING_OP_MADVISE:
5968 ret = io_madvise(req, force_nonblock);
5970 case IORING_OP_OPENAT2:
5971 ret = io_openat2(req, force_nonblock);
5973 case IORING_OP_EPOLL_CTL:
5974 ret = io_epoll_ctl(req, force_nonblock, cs);
5976 case IORING_OP_SPLICE:
5977 ret = io_splice(req, force_nonblock);
5979 case IORING_OP_PROVIDE_BUFFERS:
5980 ret = io_provide_buffers(req, force_nonblock, cs);
5982 case IORING_OP_REMOVE_BUFFERS:
5983 ret = io_remove_buffers(req, force_nonblock, cs);
5986 ret = io_tee(req, force_nonblock);
5996 /* If the op doesn't have a file, we're not polling for it */
5997 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5998 const bool in_async = io_wq_current_is_worker();
6000 /* workqueue context doesn't hold uring_lock, grab it now */
6002 mutex_lock(&ctx->uring_lock);
6004 io_iopoll_req_issued(req);
6007 mutex_unlock(&ctx->uring_lock);
6013 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6015 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6016 struct io_kiocb *timeout;
6019 timeout = io_prep_linked_timeout(req);
6021 io_queue_linked_timeout(timeout);
6023 /* if NO_CANCEL is set, we must still run the work */
6024 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6025 IO_WQ_WORK_CANCEL) {
6031 ret = io_issue_sqe(req, false, NULL);
6033 * We can get EAGAIN for polled IO even though we're
6034 * forcing a sync submission from here, since we can't
6035 * wait for request slots on the block side.
6044 req_set_fail_links(req);
6045 io_req_complete(req, ret);
6048 return io_steal_work(req);
6051 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6054 struct fixed_file_table *table;
6056 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6057 return table->files[index & IORING_FILE_TABLE_MASK];
6060 static struct file *io_file_get(struct io_submit_state *state,
6061 struct io_kiocb *req, int fd, bool fixed)
6063 struct io_ring_ctx *ctx = req->ctx;
6067 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6069 fd = array_index_nospec(fd, ctx->nr_user_files);
6070 file = io_file_from_index(ctx, fd);
6072 req->fixed_file_refs = &ctx->file_data->node->refs;
6073 percpu_ref_get(req->fixed_file_refs);
6076 trace_io_uring_file_get(ctx, fd);
6077 file = __io_file_get(state, fd);
6083 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6088 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6089 if (unlikely(!fixed && io_async_submit(req->ctx)))
6092 req->file = io_file_get(state, req, fd, fixed);
6093 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6098 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6100 struct io_timeout_data *data = container_of(timer,
6101 struct io_timeout_data, timer);
6102 struct io_kiocb *req = data->req;
6103 struct io_ring_ctx *ctx = req->ctx;
6104 struct io_kiocb *prev = NULL;
6105 unsigned long flags;
6107 spin_lock_irqsave(&ctx->completion_lock, flags);
6110 * We don't expect the list to be empty, that will only happen if we
6111 * race with the completion of the linked work.
6113 if (!list_empty(&req->link_list)) {
6114 prev = list_entry(req->link_list.prev, struct io_kiocb,
6116 if (refcount_inc_not_zero(&prev->refs))
6117 list_del_init(&req->link_list);
6122 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6125 req_set_fail_links(prev);
6126 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6129 io_req_complete(req, -ETIME);
6131 return HRTIMER_NORESTART;
6134 static void __io_queue_linked_timeout(struct io_kiocb *req)
6137 * If the list is now empty, then our linked request finished before
6138 * we got a chance to setup the timer
6140 if (!list_empty(&req->link_list)) {
6141 struct io_timeout_data *data = req->async_data;
6143 data->timer.function = io_link_timeout_fn;
6144 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6149 static void io_queue_linked_timeout(struct io_kiocb *req)
6151 struct io_ring_ctx *ctx = req->ctx;
6153 spin_lock_irq(&ctx->completion_lock);
6154 __io_queue_linked_timeout(req);
6155 spin_unlock_irq(&ctx->completion_lock);
6157 /* drop submission reference */
6161 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6163 struct io_kiocb *nxt;
6165 if (!(req->flags & REQ_F_LINK_HEAD))
6167 if (req->flags & REQ_F_LINK_TIMEOUT)
6170 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6172 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6175 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6176 req->flags |= REQ_F_LINK_TIMEOUT;
6180 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6182 struct io_kiocb *linked_timeout;
6183 struct io_kiocb *nxt;
6184 const struct cred *old_creds = NULL;
6188 linked_timeout = io_prep_linked_timeout(req);
6190 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6191 (req->work.flags & IO_WQ_WORK_CREDS) &&
6192 req->work.identity->creds != current_cred()) {
6194 revert_creds(old_creds);
6195 if (old_creds == req->work.identity->creds)
6196 old_creds = NULL; /* restored original creds */
6198 old_creds = override_creds(req->work.identity->creds);
6201 ret = io_issue_sqe(req, true, cs);
6204 * We async punt it if the file wasn't marked NOWAIT, or if the file
6205 * doesn't support non-blocking read/write attempts
6207 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6208 if (!io_arm_poll_handler(req)) {
6211 * Queued up for async execution, worker will release
6212 * submit reference when the iocb is actually submitted.
6214 io_queue_async_work(req);
6218 io_queue_linked_timeout(linked_timeout);
6222 if (unlikely(ret)) {
6223 /* un-prep timeout, so it'll be killed as any other linked */
6224 req->flags &= ~REQ_F_LINK_TIMEOUT;
6225 req_set_fail_links(req);
6227 io_req_complete(req, ret);
6231 /* drop submission reference */
6232 nxt = io_put_req_find_next(req);
6234 io_queue_linked_timeout(linked_timeout);
6239 if (req->flags & REQ_F_FORCE_ASYNC) {
6240 linked_timeout = NULL;
6247 revert_creds(old_creds);
6250 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6251 struct io_comp_state *cs)
6255 ret = io_req_defer(req, sqe);
6257 if (ret != -EIOCBQUEUED) {
6259 req_set_fail_links(req);
6261 io_req_complete(req, ret);
6263 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6264 if (!req->async_data) {
6265 ret = io_req_defer_prep(req, sqe);
6271 * Never try inline submit of IOSQE_ASYNC is set, go straight
6272 * to async execution.
6274 io_req_init_async(req);
6275 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6276 io_queue_async_work(req);
6279 ret = io_req_prep(req, sqe);
6283 __io_queue_sqe(req, cs);
6287 static inline void io_queue_link_head(struct io_kiocb *req,
6288 struct io_comp_state *cs)
6290 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6292 io_req_complete(req, -ECANCELED);
6294 io_queue_sqe(req, NULL, cs);
6297 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6298 struct io_kiocb **link, struct io_comp_state *cs)
6300 struct io_ring_ctx *ctx = req->ctx;
6304 * If we already have a head request, queue this one for async
6305 * submittal once the head completes. If we don't have a head but
6306 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6307 * submitted sync once the chain is complete. If none of those
6308 * conditions are true (normal request), then just queue it.
6311 struct io_kiocb *head = *link;
6314 * Taking sequential execution of a link, draining both sides
6315 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6316 * requests in the link. So, it drains the head and the
6317 * next after the link request. The last one is done via
6318 * drain_next flag to persist the effect across calls.
6320 if (req->flags & REQ_F_IO_DRAIN) {
6321 head->flags |= REQ_F_IO_DRAIN;
6322 ctx->drain_next = 1;
6324 ret = io_req_defer_prep(req, sqe);
6325 if (unlikely(ret)) {
6326 /* fail even hard links since we don't submit */
6327 head->flags |= REQ_F_FAIL_LINK;
6330 trace_io_uring_link(ctx, req, head);
6331 list_add_tail(&req->link_list, &head->link_list);
6333 /* last request of a link, enqueue the link */
6334 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6335 io_queue_link_head(head, cs);
6339 if (unlikely(ctx->drain_next)) {
6340 req->flags |= REQ_F_IO_DRAIN;
6341 ctx->drain_next = 0;
6343 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6344 req->flags |= REQ_F_LINK_HEAD;
6345 INIT_LIST_HEAD(&req->link_list);
6347 ret = io_req_defer_prep(req, sqe);
6349 req->flags |= REQ_F_FAIL_LINK;
6352 io_queue_sqe(req, sqe, cs);
6360 * Batched submission is done, ensure local IO is flushed out.
6362 static void io_submit_state_end(struct io_submit_state *state)
6364 if (!list_empty(&state->comp.list))
6365 io_submit_flush_completions(&state->comp);
6366 blk_finish_plug(&state->plug);
6367 io_state_file_put(state);
6368 if (state->free_reqs)
6369 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6373 * Start submission side cache.
6375 static void io_submit_state_start(struct io_submit_state *state,
6376 struct io_ring_ctx *ctx, unsigned int max_ios)
6378 blk_start_plug(&state->plug);
6380 INIT_LIST_HEAD(&state->comp.list);
6381 state->comp.ctx = ctx;
6382 state->free_reqs = 0;
6384 state->ios_left = max_ios;
6387 static void io_commit_sqring(struct io_ring_ctx *ctx)
6389 struct io_rings *rings = ctx->rings;
6392 * Ensure any loads from the SQEs are done at this point,
6393 * since once we write the new head, the application could
6394 * write new data to them.
6396 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6400 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6401 * that is mapped by userspace. This means that care needs to be taken to
6402 * ensure that reads are stable, as we cannot rely on userspace always
6403 * being a good citizen. If members of the sqe are validated and then later
6404 * used, it's important that those reads are done through READ_ONCE() to
6405 * prevent a re-load down the line.
6407 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6409 u32 *sq_array = ctx->sq_array;
6413 * The cached sq head (or cq tail) serves two purposes:
6415 * 1) allows us to batch the cost of updating the user visible
6417 * 2) allows the kernel side to track the head on its own, even
6418 * though the application is the one updating it.
6420 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6421 if (likely(head < ctx->sq_entries))
6422 return &ctx->sq_sqes[head];
6424 /* drop invalid entries */
6425 ctx->cached_sq_dropped++;
6426 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6430 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6432 ctx->cached_sq_head++;
6436 * Check SQE restrictions (opcode and flags).
6438 * Returns 'true' if SQE is allowed, 'false' otherwise.
6440 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6441 struct io_kiocb *req,
6442 unsigned int sqe_flags)
6444 if (!ctx->restricted)
6447 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6450 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6451 ctx->restrictions.sqe_flags_required)
6454 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6455 ctx->restrictions.sqe_flags_required))
6461 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6462 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6463 IOSQE_BUFFER_SELECT)
6465 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6466 const struct io_uring_sqe *sqe,
6467 struct io_submit_state *state)
6469 unsigned int sqe_flags;
6472 req->opcode = READ_ONCE(sqe->opcode);
6473 req->user_data = READ_ONCE(sqe->user_data);
6474 req->async_data = NULL;
6478 /* one is dropped after submission, the other at completion */
6479 refcount_set(&req->refs, 2);
6480 req->task = current;
6483 if (unlikely(req->opcode >= IORING_OP_LAST))
6486 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6489 sqe_flags = READ_ONCE(sqe->flags);
6490 /* enforce forwards compatibility on users */
6491 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6494 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6497 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6498 !io_op_defs[req->opcode].buffer_select)
6501 id = READ_ONCE(sqe->personality);
6503 struct io_identity *iod;
6505 iod = idr_find(&ctx->personality_idr, id);
6508 refcount_inc(&iod->count);
6510 __io_req_init_async(req);
6511 get_cred(iod->creds);
6512 req->work.identity = iod;
6513 req->work.flags |= IO_WQ_WORK_CREDS;
6516 /* same numerical values with corresponding REQ_F_*, safe to copy */
6517 req->flags |= sqe_flags;
6519 if (!io_op_defs[req->opcode].needs_file)
6522 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6527 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6529 struct io_submit_state state;
6530 struct io_kiocb *link = NULL;
6531 int i, submitted = 0;
6533 /* if we have a backlog and couldn't flush it all, return BUSY */
6534 if (test_bit(0, &ctx->sq_check_overflow)) {
6535 if (!list_empty(&ctx->cq_overflow_list) &&
6536 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6540 /* make sure SQ entry isn't read before tail */
6541 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6543 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6546 percpu_counter_add(¤t->io_uring->inflight, nr);
6547 refcount_add(nr, ¤t->usage);
6549 io_submit_state_start(&state, ctx, nr);
6551 for (i = 0; i < nr; i++) {
6552 const struct io_uring_sqe *sqe;
6553 struct io_kiocb *req;
6556 sqe = io_get_sqe(ctx);
6557 if (unlikely(!sqe)) {
6558 io_consume_sqe(ctx);
6561 req = io_alloc_req(ctx, &state);
6562 if (unlikely(!req)) {
6564 submitted = -EAGAIN;
6567 io_consume_sqe(ctx);
6568 /* will complete beyond this point, count as submitted */
6571 err = io_init_req(ctx, req, sqe, &state);
6572 if (unlikely(err)) {
6575 io_req_complete(req, err);
6579 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6580 true, io_async_submit(ctx));
6581 err = io_submit_sqe(req, sqe, &link, &state.comp);
6586 if (unlikely(submitted != nr)) {
6587 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6588 struct io_uring_task *tctx = current->io_uring;
6589 int unused = nr - ref_used;
6591 percpu_ref_put_many(&ctx->refs, unused);
6592 percpu_counter_sub(&tctx->inflight, unused);
6593 put_task_struct_many(current, unused);
6596 io_queue_link_head(link, &state.comp);
6597 io_submit_state_end(&state);
6599 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6600 io_commit_sqring(ctx);
6605 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6607 /* Tell userspace we may need a wakeup call */
6608 spin_lock_irq(&ctx->completion_lock);
6609 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6610 spin_unlock_irq(&ctx->completion_lock);
6613 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6615 spin_lock_irq(&ctx->completion_lock);
6616 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6617 spin_unlock_irq(&ctx->completion_lock);
6620 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6621 int sync, void *key)
6623 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6626 ret = autoremove_wake_function(wqe, mode, sync, key);
6628 unsigned long flags;
6630 spin_lock_irqsave(&ctx->completion_lock, flags);
6631 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6632 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6643 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6644 unsigned long start_jiffies, bool cap_entries)
6646 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6647 struct io_sq_data *sqd = ctx->sq_data;
6648 unsigned int to_submit;
6652 if (!list_empty(&ctx->iopoll_list)) {
6653 unsigned nr_events = 0;
6655 mutex_lock(&ctx->uring_lock);
6656 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6657 io_do_iopoll(ctx, &nr_events, 0);
6658 mutex_unlock(&ctx->uring_lock);
6661 to_submit = io_sqring_entries(ctx);
6664 * If submit got -EBUSY, flag us as needing the application
6665 * to enter the kernel to reap and flush events.
6667 if (!to_submit || ret == -EBUSY || need_resched()) {
6669 * Drop cur_mm before scheduling, we can't hold it for
6670 * long periods (or over schedule()). Do this before
6671 * adding ourselves to the waitqueue, as the unuse/drop
6674 io_sq_thread_drop_mm();
6677 * We're polling. If we're within the defined idle
6678 * period, then let us spin without work before going
6679 * to sleep. The exception is if we got EBUSY doing
6680 * more IO, we should wait for the application to
6681 * reap events and wake us up.
6683 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6684 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6685 !percpu_ref_is_dying(&ctx->refs)))
6688 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6689 TASK_INTERRUPTIBLE);
6692 * While doing polled IO, before going to sleep, we need
6693 * to check if there are new reqs added to iopoll_list,
6694 * it is because reqs may have been punted to io worker
6695 * and will be added to iopoll_list later, hence check
6696 * the iopoll_list again.
6698 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6699 !list_empty_careful(&ctx->iopoll_list)) {
6700 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6704 to_submit = io_sqring_entries(ctx);
6705 if (!to_submit || ret == -EBUSY)
6709 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6710 io_ring_clear_wakeup_flag(ctx);
6712 /* if we're handling multiple rings, cap submit size for fairness */
6713 if (cap_entries && to_submit > 8)
6716 mutex_lock(&ctx->uring_lock);
6717 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6718 ret = io_submit_sqes(ctx, to_submit);
6719 mutex_unlock(&ctx->uring_lock);
6721 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6722 wake_up(&ctx->sqo_sq_wait);
6724 return SQT_DID_WORK;
6727 static void io_sqd_init_new(struct io_sq_data *sqd)
6729 struct io_ring_ctx *ctx;
6731 while (!list_empty(&sqd->ctx_new_list)) {
6732 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6733 init_wait(&ctx->sqo_wait_entry);
6734 ctx->sqo_wait_entry.func = io_sq_wake_function;
6735 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6736 complete(&ctx->sq_thread_comp);
6740 static int io_sq_thread(void *data)
6742 struct cgroup_subsys_state *cur_css = NULL;
6743 const struct cred *old_cred = NULL;
6744 struct io_sq_data *sqd = data;
6745 struct io_ring_ctx *ctx;
6746 unsigned long start_jiffies;
6748 start_jiffies = jiffies;
6749 while (!kthread_should_stop()) {
6750 enum sq_ret ret = 0;
6754 * Any changes to the sqd lists are synchronized through the
6755 * kthread parking. This synchronizes the thread vs users,
6756 * the users are synchronized on the sqd->ctx_lock.
6758 if (kthread_should_park())
6761 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6762 io_sqd_init_new(sqd);
6764 cap_entries = !list_is_singular(&sqd->ctx_list);
6766 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6767 if (current->cred != ctx->creds) {
6769 revert_creds(old_cred);
6770 old_cred = override_creds(ctx->creds);
6772 io_sq_thread_associate_blkcg(ctx, &cur_css);
6774 current->loginuid = ctx->loginuid;
6775 current->sessionid = ctx->sessionid;
6778 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6780 io_sq_thread_drop_mm();
6783 if (ret & SQT_SPIN) {
6786 } else if (ret == SQT_IDLE) {
6787 if (kthread_should_park())
6789 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6790 io_ring_set_wakeup_flag(ctx);
6792 start_jiffies = jiffies;
6793 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6794 io_ring_clear_wakeup_flag(ctx);
6801 io_sq_thread_unassociate_blkcg();
6803 revert_creds(old_cred);
6810 struct io_wait_queue {
6811 struct wait_queue_entry wq;
6812 struct io_ring_ctx *ctx;
6814 unsigned nr_timeouts;
6817 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6819 struct io_ring_ctx *ctx = iowq->ctx;
6822 * Wake up if we have enough events, or if a timeout occurred since we
6823 * started waiting. For timeouts, we always want to return to userspace,
6824 * regardless of event count.
6826 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6827 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6830 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6831 int wake_flags, void *key)
6833 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6836 /* use noflush == true, as we can't safely rely on locking context */
6837 if (!io_should_wake(iowq, true))
6840 return autoremove_wake_function(curr, mode, wake_flags, key);
6843 static int io_run_task_work_sig(void)
6845 if (io_run_task_work())
6847 if (!signal_pending(current))
6849 if (current->jobctl & JOBCTL_TASK_WORK) {
6850 spin_lock_irq(¤t->sighand->siglock);
6851 current->jobctl &= ~JOBCTL_TASK_WORK;
6852 recalc_sigpending();
6853 spin_unlock_irq(¤t->sighand->siglock);
6860 * Wait until events become available, if we don't already have some. The
6861 * application must reap them itself, as they reside on the shared cq ring.
6863 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6864 const sigset_t __user *sig, size_t sigsz)
6866 struct io_wait_queue iowq = {
6869 .func = io_wake_function,
6870 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6873 .to_wait = min_events,
6875 struct io_rings *rings = ctx->rings;
6879 if (io_cqring_events(ctx, false) >= min_events)
6881 if (!io_run_task_work())
6886 #ifdef CONFIG_COMPAT
6887 if (in_compat_syscall())
6888 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6892 ret = set_user_sigmask(sig, sigsz);
6898 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6899 trace_io_uring_cqring_wait(ctx, min_events);
6901 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6902 TASK_INTERRUPTIBLE);
6903 /* make sure we run task_work before checking for signals */
6904 ret = io_run_task_work_sig();
6909 if (io_should_wake(&iowq, false))
6913 finish_wait(&ctx->wait, &iowq.wq);
6915 restore_saved_sigmask_unless(ret == -EINTR);
6917 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6920 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6922 #if defined(CONFIG_UNIX)
6923 if (ctx->ring_sock) {
6924 struct sock *sock = ctx->ring_sock->sk;
6925 struct sk_buff *skb;
6927 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6933 for (i = 0; i < ctx->nr_user_files; i++) {
6936 file = io_file_from_index(ctx, i);
6943 static void io_file_ref_kill(struct percpu_ref *ref)
6945 struct fixed_file_data *data;
6947 data = container_of(ref, struct fixed_file_data, refs);
6948 complete(&data->done);
6951 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6953 struct fixed_file_data *data = ctx->file_data;
6954 struct fixed_file_ref_node *ref_node = NULL;
6955 unsigned nr_tables, i;
6960 spin_lock(&data->lock);
6961 if (!list_empty(&data->ref_list))
6962 ref_node = list_first_entry(&data->ref_list,
6963 struct fixed_file_ref_node, node);
6964 spin_unlock(&data->lock);
6966 percpu_ref_kill(&ref_node->refs);
6968 percpu_ref_kill(&data->refs);
6970 /* wait for all refs nodes to complete */
6971 flush_delayed_work(&ctx->file_put_work);
6972 wait_for_completion(&data->done);
6974 __io_sqe_files_unregister(ctx);
6975 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6976 for (i = 0; i < nr_tables; i++)
6977 kfree(data->table[i].files);
6979 percpu_ref_exit(&data->refs);
6981 ctx->file_data = NULL;
6982 ctx->nr_user_files = 0;
6986 static void io_put_sq_data(struct io_sq_data *sqd)
6988 if (refcount_dec_and_test(&sqd->refs)) {
6990 * The park is a bit of a work-around, without it we get
6991 * warning spews on shutdown with SQPOLL set and affinity
6992 * set to a single CPU.
6995 kthread_park(sqd->thread);
6996 kthread_stop(sqd->thread);
7003 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7005 struct io_ring_ctx *ctx_attach;
7006 struct io_sq_data *sqd;
7009 f = fdget(p->wq_fd);
7011 return ERR_PTR(-ENXIO);
7012 if (f.file->f_op != &io_uring_fops) {
7014 return ERR_PTR(-EINVAL);
7017 ctx_attach = f.file->private_data;
7018 sqd = ctx_attach->sq_data;
7021 return ERR_PTR(-EINVAL);
7024 refcount_inc(&sqd->refs);
7029 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7031 struct io_sq_data *sqd;
7033 if (p->flags & IORING_SETUP_ATTACH_WQ)
7034 return io_attach_sq_data(p);
7036 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7038 return ERR_PTR(-ENOMEM);
7040 refcount_set(&sqd->refs, 1);
7041 INIT_LIST_HEAD(&sqd->ctx_list);
7042 INIT_LIST_HEAD(&sqd->ctx_new_list);
7043 mutex_init(&sqd->ctx_lock);
7044 mutex_init(&sqd->lock);
7045 init_waitqueue_head(&sqd->wait);
7049 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7050 __releases(&sqd->lock)
7054 kthread_unpark(sqd->thread);
7055 mutex_unlock(&sqd->lock);
7058 static void io_sq_thread_park(struct io_sq_data *sqd)
7059 __acquires(&sqd->lock)
7063 mutex_lock(&sqd->lock);
7064 kthread_park(sqd->thread);
7067 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7069 struct io_sq_data *sqd = ctx->sq_data;
7074 * We may arrive here from the error branch in
7075 * io_sq_offload_create() where the kthread is created
7076 * without being waked up, thus wake it up now to make
7077 * sure the wait will complete.
7079 wake_up_process(sqd->thread);
7080 wait_for_completion(&ctx->sq_thread_comp);
7082 io_sq_thread_park(sqd);
7085 mutex_lock(&sqd->ctx_lock);
7086 list_del(&ctx->sqd_list);
7087 mutex_unlock(&sqd->ctx_lock);
7090 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7091 io_sq_thread_unpark(sqd);
7094 io_put_sq_data(sqd);
7095 ctx->sq_data = NULL;
7099 static void io_finish_async(struct io_ring_ctx *ctx)
7101 io_sq_thread_stop(ctx);
7104 io_wq_destroy(ctx->io_wq);
7109 #if defined(CONFIG_UNIX)
7111 * Ensure the UNIX gc is aware of our file set, so we are certain that
7112 * the io_uring can be safely unregistered on process exit, even if we have
7113 * loops in the file referencing.
7115 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7117 struct sock *sk = ctx->ring_sock->sk;
7118 struct scm_fp_list *fpl;
7119 struct sk_buff *skb;
7122 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7126 skb = alloc_skb(0, GFP_KERNEL);
7135 fpl->user = get_uid(ctx->user);
7136 for (i = 0; i < nr; i++) {
7137 struct file *file = io_file_from_index(ctx, i + offset);
7141 fpl->fp[nr_files] = get_file(file);
7142 unix_inflight(fpl->user, fpl->fp[nr_files]);
7147 fpl->max = SCM_MAX_FD;
7148 fpl->count = nr_files;
7149 UNIXCB(skb).fp = fpl;
7150 skb->destructor = unix_destruct_scm;
7151 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7152 skb_queue_head(&sk->sk_receive_queue, skb);
7154 for (i = 0; i < nr_files; i++)
7165 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7166 * causes regular reference counting to break down. We rely on the UNIX
7167 * garbage collection to take care of this problem for us.
7169 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7171 unsigned left, total;
7175 left = ctx->nr_user_files;
7177 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7179 ret = __io_sqe_files_scm(ctx, this_files, total);
7183 total += this_files;
7189 while (total < ctx->nr_user_files) {
7190 struct file *file = io_file_from_index(ctx, total);
7200 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7206 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7207 unsigned nr_tables, unsigned nr_files)
7211 for (i = 0; i < nr_tables; i++) {
7212 struct fixed_file_table *table = &file_data->table[i];
7213 unsigned this_files;
7215 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7216 table->files = kcalloc(this_files, sizeof(struct file *),
7220 nr_files -= this_files;
7226 for (i = 0; i < nr_tables; i++) {
7227 struct fixed_file_table *table = &file_data->table[i];
7228 kfree(table->files);
7233 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7235 #if defined(CONFIG_UNIX)
7236 struct sock *sock = ctx->ring_sock->sk;
7237 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7238 struct sk_buff *skb;
7241 __skb_queue_head_init(&list);
7244 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7245 * remove this entry and rearrange the file array.
7247 skb = skb_dequeue(head);
7249 struct scm_fp_list *fp;
7251 fp = UNIXCB(skb).fp;
7252 for (i = 0; i < fp->count; i++) {
7255 if (fp->fp[i] != file)
7258 unix_notinflight(fp->user, fp->fp[i]);
7259 left = fp->count - 1 - i;
7261 memmove(&fp->fp[i], &fp->fp[i + 1],
7262 left * sizeof(struct file *));
7269 __skb_queue_tail(&list, skb);
7279 __skb_queue_tail(&list, skb);
7281 skb = skb_dequeue(head);
7284 if (skb_peek(&list)) {
7285 spin_lock_irq(&head->lock);
7286 while ((skb = __skb_dequeue(&list)) != NULL)
7287 __skb_queue_tail(head, skb);
7288 spin_unlock_irq(&head->lock);
7295 struct io_file_put {
7296 struct list_head list;
7300 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7302 struct fixed_file_data *file_data = ref_node->file_data;
7303 struct io_ring_ctx *ctx = file_data->ctx;
7304 struct io_file_put *pfile, *tmp;
7306 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7307 list_del(&pfile->list);
7308 io_ring_file_put(ctx, pfile->file);
7312 spin_lock(&file_data->lock);
7313 list_del(&ref_node->node);
7314 spin_unlock(&file_data->lock);
7316 percpu_ref_exit(&ref_node->refs);
7318 percpu_ref_put(&file_data->refs);
7321 static void io_file_put_work(struct work_struct *work)
7323 struct io_ring_ctx *ctx;
7324 struct llist_node *node;
7326 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7327 node = llist_del_all(&ctx->file_put_llist);
7330 struct fixed_file_ref_node *ref_node;
7331 struct llist_node *next = node->next;
7333 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7334 __io_file_put_work(ref_node);
7339 static void io_file_data_ref_zero(struct percpu_ref *ref)
7341 struct fixed_file_ref_node *ref_node;
7342 struct io_ring_ctx *ctx;
7346 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7347 ctx = ref_node->file_data->ctx;
7349 if (percpu_ref_is_dying(&ctx->file_data->refs))
7352 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7354 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7356 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7359 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7360 struct io_ring_ctx *ctx)
7362 struct fixed_file_ref_node *ref_node;
7364 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7366 return ERR_PTR(-ENOMEM);
7368 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7371 return ERR_PTR(-ENOMEM);
7373 INIT_LIST_HEAD(&ref_node->node);
7374 INIT_LIST_HEAD(&ref_node->file_list);
7375 ref_node->file_data = ctx->file_data;
7379 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7381 percpu_ref_exit(&ref_node->refs);
7385 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7388 __s32 __user *fds = (__s32 __user *) arg;
7389 unsigned nr_tables, i;
7391 int fd, ret = -ENOMEM;
7392 struct fixed_file_ref_node *ref_node;
7393 struct fixed_file_data *file_data;
7399 if (nr_args > IORING_MAX_FIXED_FILES)
7402 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7405 file_data->ctx = ctx;
7406 init_completion(&file_data->done);
7407 INIT_LIST_HEAD(&file_data->ref_list);
7408 spin_lock_init(&file_data->lock);
7410 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7411 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7413 if (!file_data->table)
7416 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7417 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7420 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7422 ctx->file_data = file_data;
7424 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7425 struct fixed_file_table *table;
7428 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7432 /* allow sparse sets */
7442 * Don't allow io_uring instances to be registered. If UNIX
7443 * isn't enabled, then this causes a reference cycle and this
7444 * instance can never get freed. If UNIX is enabled we'll
7445 * handle it just fine, but there's still no point in allowing
7446 * a ring fd as it doesn't support regular read/write anyway.
7448 if (file->f_op == &io_uring_fops) {
7452 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7453 index = i & IORING_FILE_TABLE_MASK;
7454 table->files[index] = file;
7457 ret = io_sqe_files_scm(ctx);
7459 io_sqe_files_unregister(ctx);
7463 ref_node = alloc_fixed_file_ref_node(ctx);
7464 if (IS_ERR(ref_node)) {
7465 io_sqe_files_unregister(ctx);
7466 return PTR_ERR(ref_node);
7469 file_data->node = ref_node;
7470 spin_lock(&file_data->lock);
7471 list_add(&ref_node->node, &file_data->ref_list);
7472 spin_unlock(&file_data->lock);
7473 percpu_ref_get(&file_data->refs);
7476 for (i = 0; i < ctx->nr_user_files; i++) {
7477 file = io_file_from_index(ctx, i);
7481 for (i = 0; i < nr_tables; i++)
7482 kfree(file_data->table[i].files);
7483 ctx->nr_user_files = 0;
7485 percpu_ref_exit(&file_data->refs);
7487 kfree(file_data->table);
7489 ctx->file_data = NULL;
7493 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7496 #if defined(CONFIG_UNIX)
7497 struct sock *sock = ctx->ring_sock->sk;
7498 struct sk_buff_head *head = &sock->sk_receive_queue;
7499 struct sk_buff *skb;
7502 * See if we can merge this file into an existing skb SCM_RIGHTS
7503 * file set. If there's no room, fall back to allocating a new skb
7504 * and filling it in.
7506 spin_lock_irq(&head->lock);
7507 skb = skb_peek(head);
7509 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7511 if (fpl->count < SCM_MAX_FD) {
7512 __skb_unlink(skb, head);
7513 spin_unlock_irq(&head->lock);
7514 fpl->fp[fpl->count] = get_file(file);
7515 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7517 spin_lock_irq(&head->lock);
7518 __skb_queue_head(head, skb);
7523 spin_unlock_irq(&head->lock);
7530 return __io_sqe_files_scm(ctx, 1, index);
7536 static int io_queue_file_removal(struct fixed_file_data *data,
7539 struct io_file_put *pfile;
7540 struct fixed_file_ref_node *ref_node = data->node;
7542 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7547 list_add(&pfile->list, &ref_node->file_list);
7552 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7553 struct io_uring_files_update *up,
7556 struct fixed_file_data *data = ctx->file_data;
7557 struct fixed_file_ref_node *ref_node;
7562 bool needs_switch = false;
7564 if (check_add_overflow(up->offset, nr_args, &done))
7566 if (done > ctx->nr_user_files)
7569 ref_node = alloc_fixed_file_ref_node(ctx);
7570 if (IS_ERR(ref_node))
7571 return PTR_ERR(ref_node);
7574 fds = u64_to_user_ptr(up->fds);
7576 struct fixed_file_table *table;
7580 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7584 i = array_index_nospec(up->offset, ctx->nr_user_files);
7585 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7586 index = i & IORING_FILE_TABLE_MASK;
7587 if (table->files[index]) {
7588 file = table->files[index];
7589 err = io_queue_file_removal(data, file);
7592 table->files[index] = NULL;
7593 needs_switch = true;
7602 * Don't allow io_uring instances to be registered. If
7603 * UNIX isn't enabled, then this causes a reference
7604 * cycle and this instance can never get freed. If UNIX
7605 * is enabled we'll handle it just fine, but there's
7606 * still no point in allowing a ring fd as it doesn't
7607 * support regular read/write anyway.
7609 if (file->f_op == &io_uring_fops) {
7614 table->files[index] = file;
7615 err = io_sqe_file_register(ctx, file, i);
7617 table->files[index] = NULL;
7628 percpu_ref_kill(&data->node->refs);
7629 spin_lock(&data->lock);
7630 list_add(&ref_node->node, &data->ref_list);
7631 data->node = ref_node;
7632 spin_unlock(&data->lock);
7633 percpu_ref_get(&ctx->file_data->refs);
7635 destroy_fixed_file_ref_node(ref_node);
7637 return done ? done : err;
7640 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7643 struct io_uring_files_update up;
7645 if (!ctx->file_data)
7649 if (copy_from_user(&up, arg, sizeof(up)))
7654 return __io_sqe_files_update(ctx, &up, nr_args);
7657 static void io_free_work(struct io_wq_work *work)
7659 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7661 /* Consider that io_steal_work() relies on this ref */
7665 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7666 struct io_uring_params *p)
7668 struct io_wq_data data;
7670 struct io_ring_ctx *ctx_attach;
7671 unsigned int concurrency;
7674 data.user = ctx->user;
7675 data.free_work = io_free_work;
7676 data.do_work = io_wq_submit_work;
7678 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7679 /* Do QD, or 4 * CPUS, whatever is smallest */
7680 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7682 ctx->io_wq = io_wq_create(concurrency, &data);
7683 if (IS_ERR(ctx->io_wq)) {
7684 ret = PTR_ERR(ctx->io_wq);
7690 f = fdget(p->wq_fd);
7694 if (f.file->f_op != &io_uring_fops) {
7699 ctx_attach = f.file->private_data;
7700 /* @io_wq is protected by holding the fd */
7701 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7706 ctx->io_wq = ctx_attach->io_wq;
7712 static int io_uring_alloc_task_context(struct task_struct *task)
7714 struct io_uring_task *tctx;
7717 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7718 if (unlikely(!tctx))
7721 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7722 if (unlikely(ret)) {
7728 init_waitqueue_head(&tctx->wait);
7731 io_init_identity(&tctx->__identity);
7732 tctx->identity = &tctx->__identity;
7733 task->io_uring = tctx;
7737 void __io_uring_free(struct task_struct *tsk)
7739 struct io_uring_task *tctx = tsk->io_uring;
7741 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7742 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7743 if (tctx->identity != &tctx->__identity)
7744 kfree(tctx->identity);
7745 percpu_counter_destroy(&tctx->inflight);
7747 tsk->io_uring = NULL;
7750 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7751 struct io_uring_params *p)
7755 if (ctx->flags & IORING_SETUP_SQPOLL) {
7756 struct io_sq_data *sqd;
7759 if (!capable(CAP_SYS_ADMIN))
7762 sqd = io_get_sq_data(p);
7769 io_sq_thread_park(sqd);
7770 mutex_lock(&sqd->ctx_lock);
7771 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7772 mutex_unlock(&sqd->ctx_lock);
7773 io_sq_thread_unpark(sqd);
7775 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7776 if (!ctx->sq_thread_idle)
7777 ctx->sq_thread_idle = HZ;
7782 if (p->flags & IORING_SETUP_SQ_AFF) {
7783 int cpu = p->sq_thread_cpu;
7786 if (cpu >= nr_cpu_ids)
7788 if (!cpu_online(cpu))
7791 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7792 cpu, "io_uring-sq");
7794 sqd->thread = kthread_create(io_sq_thread, sqd,
7797 if (IS_ERR(sqd->thread)) {
7798 ret = PTR_ERR(sqd->thread);
7802 ret = io_uring_alloc_task_context(sqd->thread);
7805 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7806 /* Can't have SQ_AFF without SQPOLL */
7812 ret = io_init_wq_offload(ctx, p);
7818 io_finish_async(ctx);
7822 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7824 struct io_sq_data *sqd = ctx->sq_data;
7826 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7827 wake_up_process(sqd->thread);
7830 static inline void __io_unaccount_mem(struct user_struct *user,
7831 unsigned long nr_pages)
7833 atomic_long_sub(nr_pages, &user->locked_vm);
7836 static inline int __io_account_mem(struct user_struct *user,
7837 unsigned long nr_pages)
7839 unsigned long page_limit, cur_pages, new_pages;
7841 /* Don't allow more pages than we can safely lock */
7842 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7845 cur_pages = atomic_long_read(&user->locked_vm);
7846 new_pages = cur_pages + nr_pages;
7847 if (new_pages > page_limit)
7849 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7850 new_pages) != cur_pages);
7855 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7856 enum io_mem_account acct)
7859 __io_unaccount_mem(ctx->user, nr_pages);
7861 if (ctx->mm_account) {
7862 if (acct == ACCT_LOCKED)
7863 ctx->mm_account->locked_vm -= nr_pages;
7864 else if (acct == ACCT_PINNED)
7865 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7869 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7870 enum io_mem_account acct)
7874 if (ctx->limit_mem) {
7875 ret = __io_account_mem(ctx->user, nr_pages);
7880 if (ctx->mm_account) {
7881 if (acct == ACCT_LOCKED)
7882 ctx->mm_account->locked_vm += nr_pages;
7883 else if (acct == ACCT_PINNED)
7884 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7890 static void io_mem_free(void *ptr)
7897 page = virt_to_head_page(ptr);
7898 if (put_page_testzero(page))
7899 free_compound_page(page);
7902 static void *io_mem_alloc(size_t size)
7904 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7907 return (void *) __get_free_pages(gfp_flags, get_order(size));
7910 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7913 struct io_rings *rings;
7914 size_t off, sq_array_size;
7916 off = struct_size(rings, cqes, cq_entries);
7917 if (off == SIZE_MAX)
7921 off = ALIGN(off, SMP_CACHE_BYTES);
7929 sq_array_size = array_size(sizeof(u32), sq_entries);
7930 if (sq_array_size == SIZE_MAX)
7933 if (check_add_overflow(off, sq_array_size, &off))
7939 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7943 pages = (size_t)1 << get_order(
7944 rings_size(sq_entries, cq_entries, NULL));
7945 pages += (size_t)1 << get_order(
7946 array_size(sizeof(struct io_uring_sqe), sq_entries));
7951 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7955 if (!ctx->user_bufs)
7958 for (i = 0; i < ctx->nr_user_bufs; i++) {
7959 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7961 for (j = 0; j < imu->nr_bvecs; j++)
7962 unpin_user_page(imu->bvec[j].bv_page);
7964 if (imu->acct_pages)
7965 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7970 kfree(ctx->user_bufs);
7971 ctx->user_bufs = NULL;
7972 ctx->nr_user_bufs = 0;
7976 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7977 void __user *arg, unsigned index)
7979 struct iovec __user *src;
7981 #ifdef CONFIG_COMPAT
7983 struct compat_iovec __user *ciovs;
7984 struct compat_iovec ciov;
7986 ciovs = (struct compat_iovec __user *) arg;
7987 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7990 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7991 dst->iov_len = ciov.iov_len;
7995 src = (struct iovec __user *) arg;
7996 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8002 * Not super efficient, but this is just a registration time. And we do cache
8003 * the last compound head, so generally we'll only do a full search if we don't
8006 * We check if the given compound head page has already been accounted, to
8007 * avoid double accounting it. This allows us to account the full size of the
8008 * page, not just the constituent pages of a huge page.
8010 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8011 int nr_pages, struct page *hpage)
8015 /* check current page array */
8016 for (i = 0; i < nr_pages; i++) {
8017 if (!PageCompound(pages[i]))
8019 if (compound_head(pages[i]) == hpage)
8023 /* check previously registered pages */
8024 for (i = 0; i < ctx->nr_user_bufs; i++) {
8025 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8027 for (j = 0; j < imu->nr_bvecs; j++) {
8028 if (!PageCompound(imu->bvec[j].bv_page))
8030 if (compound_head(imu->bvec[j].bv_page) == hpage)
8038 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8039 int nr_pages, struct io_mapped_ubuf *imu,
8040 struct page **last_hpage)
8044 for (i = 0; i < nr_pages; i++) {
8045 if (!PageCompound(pages[i])) {
8050 hpage = compound_head(pages[i]);
8051 if (hpage == *last_hpage)
8053 *last_hpage = hpage;
8054 if (headpage_already_acct(ctx, pages, i, hpage))
8056 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8060 if (!imu->acct_pages)
8063 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8065 imu->acct_pages = 0;
8069 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8072 struct vm_area_struct **vmas = NULL;
8073 struct page **pages = NULL;
8074 struct page *last_hpage = NULL;
8075 int i, j, got_pages = 0;
8080 if (!nr_args || nr_args > UIO_MAXIOV)
8083 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8085 if (!ctx->user_bufs)
8088 for (i = 0; i < nr_args; i++) {
8089 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8090 unsigned long off, start, end, ubuf;
8095 ret = io_copy_iov(ctx, &iov, arg, i);
8100 * Don't impose further limits on the size and buffer
8101 * constraints here, we'll -EINVAL later when IO is
8102 * submitted if they are wrong.
8105 if (!iov.iov_base || !iov.iov_len)
8108 /* arbitrary limit, but we need something */
8109 if (iov.iov_len > SZ_1G)
8112 ubuf = (unsigned long) iov.iov_base;
8113 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8114 start = ubuf >> PAGE_SHIFT;
8115 nr_pages = end - start;
8118 if (!pages || nr_pages > got_pages) {
8121 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8123 vmas = kvmalloc_array(nr_pages,
8124 sizeof(struct vm_area_struct *),
8126 if (!pages || !vmas) {
8130 got_pages = nr_pages;
8133 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8140 mmap_read_lock(current->mm);
8141 pret = pin_user_pages(ubuf, nr_pages,
8142 FOLL_WRITE | FOLL_LONGTERM,
8144 if (pret == nr_pages) {
8145 /* don't support file backed memory */
8146 for (j = 0; j < nr_pages; j++) {
8147 struct vm_area_struct *vma = vmas[j];
8150 !is_file_hugepages(vma->vm_file)) {
8156 ret = pret < 0 ? pret : -EFAULT;
8158 mmap_read_unlock(current->mm);
8161 * if we did partial map, or found file backed vmas,
8162 * release any pages we did get
8165 unpin_user_pages(pages, pret);
8170 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8172 unpin_user_pages(pages, pret);
8177 off = ubuf & ~PAGE_MASK;
8179 for (j = 0; j < nr_pages; j++) {
8182 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8183 imu->bvec[j].bv_page = pages[j];
8184 imu->bvec[j].bv_len = vec_len;
8185 imu->bvec[j].bv_offset = off;
8189 /* store original address for later verification */
8191 imu->len = iov.iov_len;
8192 imu->nr_bvecs = nr_pages;
8194 ctx->nr_user_bufs++;
8202 io_sqe_buffer_unregister(ctx);
8206 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8208 __s32 __user *fds = arg;
8214 if (copy_from_user(&fd, fds, sizeof(*fds)))
8217 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8218 if (IS_ERR(ctx->cq_ev_fd)) {
8219 int ret = PTR_ERR(ctx->cq_ev_fd);
8220 ctx->cq_ev_fd = NULL;
8227 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8229 if (ctx->cq_ev_fd) {
8230 eventfd_ctx_put(ctx->cq_ev_fd);
8231 ctx->cq_ev_fd = NULL;
8238 static int __io_destroy_buffers(int id, void *p, void *data)
8240 struct io_ring_ctx *ctx = data;
8241 struct io_buffer *buf = p;
8243 __io_remove_buffers(ctx, buf, id, -1U);
8247 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8249 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8250 idr_destroy(&ctx->io_buffer_idr);
8253 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8255 io_finish_async(ctx);
8256 io_sqe_buffer_unregister(ctx);
8258 if (ctx->sqo_task) {
8259 put_task_struct(ctx->sqo_task);
8260 ctx->sqo_task = NULL;
8261 mmdrop(ctx->mm_account);
8262 ctx->mm_account = NULL;
8265 #ifdef CONFIG_BLK_CGROUP
8266 if (ctx->sqo_blkcg_css)
8267 css_put(ctx->sqo_blkcg_css);
8270 io_sqe_files_unregister(ctx);
8271 io_eventfd_unregister(ctx);
8272 io_destroy_buffers(ctx);
8273 idr_destroy(&ctx->personality_idr);
8275 #if defined(CONFIG_UNIX)
8276 if (ctx->ring_sock) {
8277 ctx->ring_sock->file = NULL; /* so that iput() is called */
8278 sock_release(ctx->ring_sock);
8282 io_mem_free(ctx->rings);
8283 io_mem_free(ctx->sq_sqes);
8285 percpu_ref_exit(&ctx->refs);
8286 free_uid(ctx->user);
8287 put_cred(ctx->creds);
8288 kfree(ctx->cancel_hash);
8289 kmem_cache_free(req_cachep, ctx->fallback_req);
8293 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8295 struct io_ring_ctx *ctx = file->private_data;
8298 poll_wait(file, &ctx->cq_wait, wait);
8300 * synchronizes with barrier from wq_has_sleeper call in
8304 if (!io_sqring_full(ctx))
8305 mask |= EPOLLOUT | EPOLLWRNORM;
8306 if (io_cqring_events(ctx, false))
8307 mask |= EPOLLIN | EPOLLRDNORM;
8312 static int io_uring_fasync(int fd, struct file *file, int on)
8314 struct io_ring_ctx *ctx = file->private_data;
8316 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8319 static int io_remove_personalities(int id, void *p, void *data)
8321 struct io_ring_ctx *ctx = data;
8322 struct io_identity *iod;
8324 iod = idr_remove(&ctx->personality_idr, id);
8326 put_cred(iod->creds);
8327 if (refcount_dec_and_test(&iod->count))
8333 static void io_ring_exit_work(struct work_struct *work)
8335 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8339 * If we're doing polled IO and end up having requests being
8340 * submitted async (out-of-line), then completions can come in while
8341 * we're waiting for refs to drop. We need to reap these manually,
8342 * as nobody else will be looking for them.
8346 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8347 io_iopoll_try_reap_events(ctx);
8348 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8349 io_ring_ctx_free(ctx);
8352 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8354 mutex_lock(&ctx->uring_lock);
8355 percpu_ref_kill(&ctx->refs);
8356 mutex_unlock(&ctx->uring_lock);
8358 io_kill_timeouts(ctx, NULL);
8359 io_poll_remove_all(ctx, NULL);
8362 io_wq_cancel_all(ctx->io_wq);
8364 /* if we failed setting up the ctx, we might not have any rings */
8366 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8367 io_iopoll_try_reap_events(ctx);
8368 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8371 * Do this upfront, so we won't have a grace period where the ring
8372 * is closed but resources aren't reaped yet. This can cause
8373 * spurious failure in setting up a new ring.
8375 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8378 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8380 * Use system_unbound_wq to avoid spawning tons of event kworkers
8381 * if we're exiting a ton of rings at the same time. It just adds
8382 * noise and overhead, there's no discernable change in runtime
8383 * over using system_wq.
8385 queue_work(system_unbound_wq, &ctx->exit_work);
8388 static int io_uring_release(struct inode *inode, struct file *file)
8390 struct io_ring_ctx *ctx = file->private_data;
8392 file->private_data = NULL;
8393 io_ring_ctx_wait_and_kill(ctx);
8397 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8399 struct files_struct *files = data;
8401 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8402 work->identity->files == files);
8406 * Returns true if 'preq' is the link parent of 'req'
8408 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8410 struct io_kiocb *link;
8412 if (!(preq->flags & REQ_F_LINK_HEAD))
8415 list_for_each_entry(link, &preq->link_list, link_list) {
8423 static bool io_match_link_files(struct io_kiocb *req,
8424 struct files_struct *files)
8426 struct io_kiocb *link;
8428 if (io_match_files(req, files))
8430 if (req->flags & REQ_F_LINK_HEAD) {
8431 list_for_each_entry(link, &req->link_list, link_list) {
8432 if (io_match_files(link, files))
8440 * We're looking to cancel 'req' because it's holding on to our files, but
8441 * 'req' could be a link to another request. See if it is, and cancel that
8442 * parent request if so.
8444 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8446 struct hlist_node *tmp;
8447 struct io_kiocb *preq;
8451 spin_lock_irq(&ctx->completion_lock);
8452 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8453 struct hlist_head *list;
8455 list = &ctx->cancel_hash[i];
8456 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8457 found = io_match_link(preq, req);
8459 io_poll_remove_one(preq);
8464 spin_unlock_irq(&ctx->completion_lock);
8468 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8469 struct io_kiocb *req)
8471 struct io_kiocb *preq;
8474 spin_lock_irq(&ctx->completion_lock);
8475 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8476 found = io_match_link(preq, req);
8478 __io_timeout_cancel(preq);
8482 spin_unlock_irq(&ctx->completion_lock);
8486 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8488 return io_match_link(container_of(work, struct io_kiocb, work), data);
8491 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8493 enum io_wq_cancel cret;
8495 /* cancel this particular work, if it's running */
8496 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8497 if (cret != IO_WQ_CANCEL_NOTFOUND)
8500 /* find links that hold this pending, cancel those */
8501 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8502 if (cret != IO_WQ_CANCEL_NOTFOUND)
8505 /* if we have a poll link holding this pending, cancel that */
8506 if (io_poll_remove_link(ctx, req))
8509 /* final option, timeout link is holding this req pending */
8510 io_timeout_remove_link(ctx, req);
8513 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8514 struct files_struct *files)
8516 struct io_defer_entry *de = NULL;
8519 spin_lock_irq(&ctx->completion_lock);
8520 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8521 if (io_match_link_files(de->req, files)) {
8522 list_cut_position(&list, &ctx->defer_list, &de->list);
8526 spin_unlock_irq(&ctx->completion_lock);
8528 while (!list_empty(&list)) {
8529 de = list_first_entry(&list, struct io_defer_entry, list);
8530 list_del_init(&de->list);
8531 req_set_fail_links(de->req);
8532 io_put_req(de->req);
8533 io_req_complete(de->req, -ECANCELED);
8539 * Returns true if we found and killed one or more files pinning requests
8541 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8542 struct files_struct *files)
8544 if (list_empty_careful(&ctx->inflight_list))
8547 io_cancel_defer_files(ctx, files);
8548 /* cancel all at once, should be faster than doing it one by one*/
8549 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8551 while (!list_empty_careful(&ctx->inflight_list)) {
8552 struct io_kiocb *cancel_req = NULL, *req;
8555 spin_lock_irq(&ctx->inflight_lock);
8556 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8557 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8558 req->work.identity->files != files)
8560 /* req is being completed, ignore */
8561 if (!refcount_inc_not_zero(&req->refs))
8567 prepare_to_wait(&ctx->inflight_wait, &wait,
8568 TASK_UNINTERRUPTIBLE);
8569 spin_unlock_irq(&ctx->inflight_lock);
8571 /* We need to keep going until we don't find a matching req */
8574 /* cancel this request, or head link requests */
8575 io_attempt_cancel(ctx, cancel_req);
8576 io_put_req(cancel_req);
8577 /* cancellations _may_ trigger task work */
8580 finish_wait(&ctx->inflight_wait, &wait);
8586 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8588 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8589 struct task_struct *task = data;
8591 return io_task_match(req, task);
8594 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8595 struct task_struct *task,
8596 struct files_struct *files)
8600 ret = io_uring_cancel_files(ctx, files);
8602 enum io_wq_cancel cret;
8604 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8605 if (cret != IO_WQ_CANCEL_NOTFOUND)
8608 /* SQPOLL thread does its own polling */
8609 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8610 while (!list_empty_careful(&ctx->iopoll_list)) {
8611 io_iopoll_try_reap_events(ctx);
8616 ret |= io_poll_remove_all(ctx, task);
8617 ret |= io_kill_timeouts(ctx, task);
8624 * We need to iteratively cancel requests, in case a request has dependent
8625 * hard links. These persist even for failure of cancelations, hence keep
8626 * looping until none are found.
8628 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8629 struct files_struct *files)
8631 struct task_struct *task = current;
8633 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8634 task = ctx->sq_data->thread;
8636 io_cqring_overflow_flush(ctx, true, task, files);
8638 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8645 * Note that this task has used io_uring. We use it for cancelation purposes.
8647 static int io_uring_add_task_file(struct file *file)
8649 struct io_uring_task *tctx = current->io_uring;
8651 if (unlikely(!tctx)) {
8654 ret = io_uring_alloc_task_context(current);
8657 tctx = current->io_uring;
8659 if (tctx->last != file) {
8660 void *old = xa_load(&tctx->xa, (unsigned long)file);
8664 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8673 * Remove this io_uring_file -> task mapping.
8675 static void io_uring_del_task_file(struct file *file)
8677 struct io_uring_task *tctx = current->io_uring;
8679 if (tctx->last == file)
8681 file = xa_erase(&tctx->xa, (unsigned long)file);
8687 * Drop task note for this file if we're the only ones that hold it after
8690 static void io_uring_attempt_task_drop(struct file *file)
8692 if (!current->io_uring)
8695 * fput() is pending, will be 2 if the only other ref is our potential
8696 * task file note. If the task is exiting, drop regardless of count.
8698 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8699 atomic_long_read(&file->f_count) == 2)
8700 io_uring_del_task_file(file);
8703 void __io_uring_files_cancel(struct files_struct *files)
8705 struct io_uring_task *tctx = current->io_uring;
8707 unsigned long index;
8709 /* make sure overflow events are dropped */
8710 tctx->in_idle = true;
8712 xa_for_each(&tctx->xa, index, file) {
8713 struct io_ring_ctx *ctx = file->private_data;
8715 io_uring_cancel_task_requests(ctx, files);
8717 io_uring_del_task_file(file);
8722 * Find any io_uring fd that this task has registered or done IO on, and cancel
8725 void __io_uring_task_cancel(void)
8727 struct io_uring_task *tctx = current->io_uring;
8731 /* make sure overflow events are dropped */
8732 tctx->in_idle = true;
8735 /* read completions before cancelations */
8736 inflight = percpu_counter_sum(&tctx->inflight);
8739 __io_uring_files_cancel(NULL);
8741 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8744 * If we've seen completions, retry. This avoids a race where
8745 * a completion comes in before we did prepare_to_wait().
8747 if (inflight != percpu_counter_sum(&tctx->inflight))
8752 finish_wait(&tctx->wait, &wait);
8753 tctx->in_idle = false;
8756 static int io_uring_flush(struct file *file, void *data)
8758 io_uring_attempt_task_drop(file);
8762 static void *io_uring_validate_mmap_request(struct file *file,
8763 loff_t pgoff, size_t sz)
8765 struct io_ring_ctx *ctx = file->private_data;
8766 loff_t offset = pgoff << PAGE_SHIFT;
8771 case IORING_OFF_SQ_RING:
8772 case IORING_OFF_CQ_RING:
8775 case IORING_OFF_SQES:
8779 return ERR_PTR(-EINVAL);
8782 page = virt_to_head_page(ptr);
8783 if (sz > page_size(page))
8784 return ERR_PTR(-EINVAL);
8791 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8793 size_t sz = vma->vm_end - vma->vm_start;
8797 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8799 return PTR_ERR(ptr);
8801 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8802 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8805 #else /* !CONFIG_MMU */
8807 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8809 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8812 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8814 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8817 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8818 unsigned long addr, unsigned long len,
8819 unsigned long pgoff, unsigned long flags)
8823 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8825 return PTR_ERR(ptr);
8827 return (unsigned long) ptr;
8830 #endif /* !CONFIG_MMU */
8832 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8837 if (!io_sqring_full(ctx))
8840 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8842 if (!io_sqring_full(ctx))
8846 } while (!signal_pending(current));
8848 finish_wait(&ctx->sqo_sq_wait, &wait);
8851 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8852 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8855 struct io_ring_ctx *ctx;
8862 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8863 IORING_ENTER_SQ_WAIT))
8871 if (f.file->f_op != &io_uring_fops)
8875 ctx = f.file->private_data;
8876 if (!percpu_ref_tryget(&ctx->refs))
8880 if (ctx->flags & IORING_SETUP_R_DISABLED)
8884 * For SQ polling, the thread will do all submissions and completions.
8885 * Just return the requested submit count, and wake the thread if
8889 if (ctx->flags & IORING_SETUP_SQPOLL) {
8890 if (!list_empty_careful(&ctx->cq_overflow_list))
8891 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8892 if (flags & IORING_ENTER_SQ_WAKEUP)
8893 wake_up(&ctx->sq_data->wait);
8894 if (flags & IORING_ENTER_SQ_WAIT)
8895 io_sqpoll_wait_sq(ctx);
8896 submitted = to_submit;
8897 } else if (to_submit) {
8898 ret = io_uring_add_task_file(f.file);
8901 mutex_lock(&ctx->uring_lock);
8902 submitted = io_submit_sqes(ctx, to_submit);
8903 mutex_unlock(&ctx->uring_lock);
8905 if (submitted != to_submit)
8908 if (flags & IORING_ENTER_GETEVENTS) {
8909 min_complete = min(min_complete, ctx->cq_entries);
8912 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8913 * space applications don't need to do io completion events
8914 * polling again, they can rely on io_sq_thread to do polling
8915 * work, which can reduce cpu usage and uring_lock contention.
8917 if (ctx->flags & IORING_SETUP_IOPOLL &&
8918 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8919 ret = io_iopoll_check(ctx, min_complete);
8921 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8926 percpu_ref_put(&ctx->refs);
8929 return submitted ? submitted : ret;
8932 #ifdef CONFIG_PROC_FS
8933 static int io_uring_show_cred(int id, void *p, void *data)
8935 const struct cred *cred = p;
8936 struct seq_file *m = data;
8937 struct user_namespace *uns = seq_user_ns(m);
8938 struct group_info *gi;
8943 seq_printf(m, "%5d\n", id);
8944 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8945 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8946 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8947 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8948 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8949 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8950 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8951 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8952 seq_puts(m, "\n\tGroups:\t");
8953 gi = cred->group_info;
8954 for (g = 0; g < gi->ngroups; g++) {
8955 seq_put_decimal_ull(m, g ? " " : "",
8956 from_kgid_munged(uns, gi->gid[g]));
8958 seq_puts(m, "\n\tCapEff:\t");
8959 cap = cred->cap_effective;
8960 CAP_FOR_EACH_U32(__capi)
8961 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8966 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8968 struct io_sq_data *sq = NULL;
8973 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8974 * since fdinfo case grabs it in the opposite direction of normal use
8975 * cases. If we fail to get the lock, we just don't iterate any
8976 * structures that could be going away outside the io_uring mutex.
8978 has_lock = mutex_trylock(&ctx->uring_lock);
8980 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8983 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8984 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8985 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8986 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8987 struct fixed_file_table *table;
8990 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8991 f = table->files[i & IORING_FILE_TABLE_MASK];
8993 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8995 seq_printf(m, "%5u: <none>\n", i);
8997 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8998 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8999 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9001 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9002 (unsigned int) buf->len);
9004 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9005 seq_printf(m, "Personalities:\n");
9006 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9008 seq_printf(m, "PollList:\n");
9009 spin_lock_irq(&ctx->completion_lock);
9010 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9011 struct hlist_head *list = &ctx->cancel_hash[i];
9012 struct io_kiocb *req;
9014 hlist_for_each_entry(req, list, hash_node)
9015 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9016 req->task->task_works != NULL);
9018 spin_unlock_irq(&ctx->completion_lock);
9020 mutex_unlock(&ctx->uring_lock);
9023 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9025 struct io_ring_ctx *ctx = f->private_data;
9027 if (percpu_ref_tryget(&ctx->refs)) {
9028 __io_uring_show_fdinfo(ctx, m);
9029 percpu_ref_put(&ctx->refs);
9034 static const struct file_operations io_uring_fops = {
9035 .release = io_uring_release,
9036 .flush = io_uring_flush,
9037 .mmap = io_uring_mmap,
9039 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9040 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9042 .poll = io_uring_poll,
9043 .fasync = io_uring_fasync,
9044 #ifdef CONFIG_PROC_FS
9045 .show_fdinfo = io_uring_show_fdinfo,
9049 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9050 struct io_uring_params *p)
9052 struct io_rings *rings;
9053 size_t size, sq_array_offset;
9055 /* make sure these are sane, as we already accounted them */
9056 ctx->sq_entries = p->sq_entries;
9057 ctx->cq_entries = p->cq_entries;
9059 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9060 if (size == SIZE_MAX)
9063 rings = io_mem_alloc(size);
9068 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9069 rings->sq_ring_mask = p->sq_entries - 1;
9070 rings->cq_ring_mask = p->cq_entries - 1;
9071 rings->sq_ring_entries = p->sq_entries;
9072 rings->cq_ring_entries = p->cq_entries;
9073 ctx->sq_mask = rings->sq_ring_mask;
9074 ctx->cq_mask = rings->cq_ring_mask;
9076 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9077 if (size == SIZE_MAX) {
9078 io_mem_free(ctx->rings);
9083 ctx->sq_sqes = io_mem_alloc(size);
9084 if (!ctx->sq_sqes) {
9085 io_mem_free(ctx->rings);
9094 * Allocate an anonymous fd, this is what constitutes the application
9095 * visible backing of an io_uring instance. The application mmaps this
9096 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9097 * we have to tie this fd to a socket for file garbage collection purposes.
9099 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9104 #if defined(CONFIG_UNIX)
9105 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9111 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9115 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9116 O_RDWR | O_CLOEXEC);
9120 ret = PTR_ERR(file);
9124 #if defined(CONFIG_UNIX)
9125 ctx->ring_sock->file = file;
9127 if (unlikely(io_uring_add_task_file(file))) {
9128 file = ERR_PTR(-ENOMEM);
9131 fd_install(ret, file);
9134 #if defined(CONFIG_UNIX)
9135 sock_release(ctx->ring_sock);
9136 ctx->ring_sock = NULL;
9141 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9142 struct io_uring_params __user *params)
9144 struct user_struct *user = NULL;
9145 struct io_ring_ctx *ctx;
9151 if (entries > IORING_MAX_ENTRIES) {
9152 if (!(p->flags & IORING_SETUP_CLAMP))
9154 entries = IORING_MAX_ENTRIES;
9158 * Use twice as many entries for the CQ ring. It's possible for the
9159 * application to drive a higher depth than the size of the SQ ring,
9160 * since the sqes are only used at submission time. This allows for
9161 * some flexibility in overcommitting a bit. If the application has
9162 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9163 * of CQ ring entries manually.
9165 p->sq_entries = roundup_pow_of_two(entries);
9166 if (p->flags & IORING_SETUP_CQSIZE) {
9168 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9169 * to a power-of-two, if it isn't already. We do NOT impose
9170 * any cq vs sq ring sizing.
9172 if (p->cq_entries < p->sq_entries)
9174 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9175 if (!(p->flags & IORING_SETUP_CLAMP))
9177 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9179 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9181 p->cq_entries = 2 * p->sq_entries;
9184 user = get_uid(current_user());
9185 limit_mem = !capable(CAP_IPC_LOCK);
9188 ret = __io_account_mem(user,
9189 ring_pages(p->sq_entries, p->cq_entries));
9196 ctx = io_ring_ctx_alloc(p);
9199 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9204 ctx->compat = in_compat_syscall();
9206 ctx->creds = get_current_cred();
9208 ctx->loginuid = current->loginuid;
9209 ctx->sessionid = current->sessionid;
9211 ctx->sqo_task = get_task_struct(current);
9214 * This is just grabbed for accounting purposes. When a process exits,
9215 * the mm is exited and dropped before the files, hence we need to hang
9216 * on to this mm purely for the purposes of being able to unaccount
9217 * memory (locked/pinned vm). It's not used for anything else.
9219 mmgrab(current->mm);
9220 ctx->mm_account = current->mm;
9222 #ifdef CONFIG_BLK_CGROUP
9224 * The sq thread will belong to the original cgroup it was inited in.
9225 * If the cgroup goes offline (e.g. disabling the io controller), then
9226 * issued bios will be associated with the closest cgroup later in the
9230 ctx->sqo_blkcg_css = blkcg_css();
9231 ret = css_tryget_online(ctx->sqo_blkcg_css);
9234 /* don't init against a dying cgroup, have the user try again */
9235 ctx->sqo_blkcg_css = NULL;
9242 * Account memory _before_ installing the file descriptor. Once
9243 * the descriptor is installed, it can get closed at any time. Also
9244 * do this before hitting the general error path, as ring freeing
9245 * will un-account as well.
9247 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9249 ctx->limit_mem = limit_mem;
9251 ret = io_allocate_scq_urings(ctx, p);
9255 ret = io_sq_offload_create(ctx, p);
9259 if (!(p->flags & IORING_SETUP_R_DISABLED))
9260 io_sq_offload_start(ctx);
9262 memset(&p->sq_off, 0, sizeof(p->sq_off));
9263 p->sq_off.head = offsetof(struct io_rings, sq.head);
9264 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9265 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9266 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9267 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9268 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9269 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9271 memset(&p->cq_off, 0, sizeof(p->cq_off));
9272 p->cq_off.head = offsetof(struct io_rings, cq.head);
9273 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9274 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9275 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9276 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9277 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9278 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9280 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9281 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9282 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9283 IORING_FEAT_POLL_32BITS;
9285 if (copy_to_user(params, p, sizeof(*p))) {
9291 * Install ring fd as the very last thing, so we don't risk someone
9292 * having closed it before we finish setup
9294 ret = io_uring_get_fd(ctx);
9298 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9301 io_ring_ctx_wait_and_kill(ctx);
9306 * Sets up an aio uring context, and returns the fd. Applications asks for a
9307 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9308 * params structure passed in.
9310 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9312 struct io_uring_params p;
9315 if (copy_from_user(&p, params, sizeof(p)))
9317 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9322 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9323 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9324 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9325 IORING_SETUP_R_DISABLED))
9328 return io_uring_create(entries, &p, params);
9331 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9332 struct io_uring_params __user *, params)
9334 return io_uring_setup(entries, params);
9337 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9339 struct io_uring_probe *p;
9343 size = struct_size(p, ops, nr_args);
9344 if (size == SIZE_MAX)
9346 p = kzalloc(size, GFP_KERNEL);
9351 if (copy_from_user(p, arg, size))
9354 if (memchr_inv(p, 0, size))
9357 p->last_op = IORING_OP_LAST - 1;
9358 if (nr_args > IORING_OP_LAST)
9359 nr_args = IORING_OP_LAST;
9361 for (i = 0; i < nr_args; i++) {
9363 if (!io_op_defs[i].not_supported)
9364 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9369 if (copy_to_user(arg, p, size))
9376 static int io_register_personality(struct io_ring_ctx *ctx)
9378 struct io_identity *id;
9381 id = kmalloc(sizeof(*id), GFP_KERNEL);
9385 io_init_identity(id);
9386 id->creds = get_current_cred();
9388 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9390 put_cred(id->creds);
9396 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9398 struct io_identity *iod;
9400 iod = idr_remove(&ctx->personality_idr, id);
9402 put_cred(iod->creds);
9403 if (refcount_dec_and_test(&iod->count))
9411 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9412 unsigned int nr_args)
9414 struct io_uring_restriction *res;
9418 /* Restrictions allowed only if rings started disabled */
9419 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9422 /* We allow only a single restrictions registration */
9423 if (ctx->restrictions.registered)
9426 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9429 size = array_size(nr_args, sizeof(*res));
9430 if (size == SIZE_MAX)
9433 res = memdup_user(arg, size);
9435 return PTR_ERR(res);
9439 for (i = 0; i < nr_args; i++) {
9440 switch (res[i].opcode) {
9441 case IORING_RESTRICTION_REGISTER_OP:
9442 if (res[i].register_op >= IORING_REGISTER_LAST) {
9447 __set_bit(res[i].register_op,
9448 ctx->restrictions.register_op);
9450 case IORING_RESTRICTION_SQE_OP:
9451 if (res[i].sqe_op >= IORING_OP_LAST) {
9456 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9458 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9459 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9461 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9462 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9471 /* Reset all restrictions if an error happened */
9473 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9475 ctx->restrictions.registered = true;
9481 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9483 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9486 if (ctx->restrictions.registered)
9487 ctx->restricted = 1;
9489 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9491 io_sq_offload_start(ctx);
9496 static bool io_register_op_must_quiesce(int op)
9499 case IORING_UNREGISTER_FILES:
9500 case IORING_REGISTER_FILES_UPDATE:
9501 case IORING_REGISTER_PROBE:
9502 case IORING_REGISTER_PERSONALITY:
9503 case IORING_UNREGISTER_PERSONALITY:
9510 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9511 void __user *arg, unsigned nr_args)
9512 __releases(ctx->uring_lock)
9513 __acquires(ctx->uring_lock)
9518 * We're inside the ring mutex, if the ref is already dying, then
9519 * someone else killed the ctx or is already going through
9520 * io_uring_register().
9522 if (percpu_ref_is_dying(&ctx->refs))
9525 if (io_register_op_must_quiesce(opcode)) {
9526 percpu_ref_kill(&ctx->refs);
9529 * Drop uring mutex before waiting for references to exit. If
9530 * another thread is currently inside io_uring_enter() it might
9531 * need to grab the uring_lock to make progress. If we hold it
9532 * here across the drain wait, then we can deadlock. It's safe
9533 * to drop the mutex here, since no new references will come in
9534 * after we've killed the percpu ref.
9536 mutex_unlock(&ctx->uring_lock);
9538 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9541 ret = io_run_task_work_sig();
9546 mutex_lock(&ctx->uring_lock);
9549 percpu_ref_resurrect(&ctx->refs);
9554 if (ctx->restricted) {
9555 if (opcode >= IORING_REGISTER_LAST) {
9560 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9567 case IORING_REGISTER_BUFFERS:
9568 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9570 case IORING_UNREGISTER_BUFFERS:
9574 ret = io_sqe_buffer_unregister(ctx);
9576 case IORING_REGISTER_FILES:
9577 ret = io_sqe_files_register(ctx, arg, nr_args);
9579 case IORING_UNREGISTER_FILES:
9583 ret = io_sqe_files_unregister(ctx);
9585 case IORING_REGISTER_FILES_UPDATE:
9586 ret = io_sqe_files_update(ctx, arg, nr_args);
9588 case IORING_REGISTER_EVENTFD:
9589 case IORING_REGISTER_EVENTFD_ASYNC:
9593 ret = io_eventfd_register(ctx, arg);
9596 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9597 ctx->eventfd_async = 1;
9599 ctx->eventfd_async = 0;
9601 case IORING_UNREGISTER_EVENTFD:
9605 ret = io_eventfd_unregister(ctx);
9607 case IORING_REGISTER_PROBE:
9609 if (!arg || nr_args > 256)
9611 ret = io_probe(ctx, arg, nr_args);
9613 case IORING_REGISTER_PERSONALITY:
9617 ret = io_register_personality(ctx);
9619 case IORING_UNREGISTER_PERSONALITY:
9623 ret = io_unregister_personality(ctx, nr_args);
9625 case IORING_REGISTER_ENABLE_RINGS:
9629 ret = io_register_enable_rings(ctx);
9631 case IORING_REGISTER_RESTRICTIONS:
9632 ret = io_register_restrictions(ctx, arg, nr_args);
9640 if (io_register_op_must_quiesce(opcode)) {
9641 /* bring the ctx back to life */
9642 percpu_ref_reinit(&ctx->refs);
9644 reinit_completion(&ctx->ref_comp);
9649 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9650 void __user *, arg, unsigned int, nr_args)
9652 struct io_ring_ctx *ctx;
9661 if (f.file->f_op != &io_uring_fops)
9664 ctx = f.file->private_data;
9666 mutex_lock(&ctx->uring_lock);
9667 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9668 mutex_unlock(&ctx->uring_lock);
9669 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9670 ctx->cq_ev_fd != NULL, ret);
9676 static int __init io_uring_init(void)
9678 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9679 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9680 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9683 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9684 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9685 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9686 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9687 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9688 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9689 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9690 BUILD_BUG_SQE_ELEM(8, __u64, off);
9691 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9692 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9693 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9694 BUILD_BUG_SQE_ELEM(24, __u32, len);
9695 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9696 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9697 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9698 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9699 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9700 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9701 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9702 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9703 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9704 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9705 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9706 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9707 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9708 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9709 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9710 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9711 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9712 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9713 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9715 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9716 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9717 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9720 __initcall(io_uring_init);