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_FORCE_ASYNC)
1369 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1371 if (req->flags & REQ_F_ISREG) {
1372 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1373 io_wq_hash_work(&req->work, file_inode(req->file));
1375 if (def->unbound_nonreg_file)
1376 req->work.flags |= IO_WQ_WORK_UNBOUND;
1379 /* ->mm can never change on us */
1380 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1381 (def->work_flags & IO_WQ_WORK_MM)) {
1383 req->work.flags |= IO_WQ_WORK_MM;
1386 /* if we fail grabbing identity, we must COW, regrab, and retry */
1387 if (io_grab_identity(req))
1390 if (!io_identity_cow(req))
1393 /* can't fail at this point */
1394 if (!io_grab_identity(req))
1398 static void io_prep_async_link(struct io_kiocb *req)
1400 struct io_kiocb *cur;
1402 io_prep_async_work(req);
1403 if (req->flags & REQ_F_LINK_HEAD)
1404 list_for_each_entry(cur, &req->link_list, link_list)
1405 io_prep_async_work(cur);
1408 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1410 struct io_ring_ctx *ctx = req->ctx;
1411 struct io_kiocb *link = io_prep_linked_timeout(req);
1413 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1414 &req->work, req->flags);
1415 io_wq_enqueue(ctx->io_wq, &req->work);
1419 static void io_queue_async_work(struct io_kiocb *req)
1421 struct io_kiocb *link;
1423 /* init ->work of the whole link before punting */
1424 io_prep_async_link(req);
1425 link = __io_queue_async_work(req);
1428 io_queue_linked_timeout(link);
1431 static void io_kill_timeout(struct io_kiocb *req)
1433 struct io_timeout_data *io = req->async_data;
1436 ret = hrtimer_try_to_cancel(&io->timer);
1438 atomic_set(&req->ctx->cq_timeouts,
1439 atomic_read(&req->ctx->cq_timeouts) + 1);
1440 list_del_init(&req->timeout.list);
1441 io_cqring_fill_event(req, 0);
1442 io_put_req_deferred(req, 1);
1446 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1448 struct io_ring_ctx *ctx = req->ctx;
1450 if (!tsk || req->task == tsk)
1452 if (ctx->flags & IORING_SETUP_SQPOLL) {
1453 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1460 * Returns true if we found and killed one or more timeouts
1462 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1464 struct io_kiocb *req, *tmp;
1467 spin_lock_irq(&ctx->completion_lock);
1468 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1469 if (io_task_match(req, tsk)) {
1470 io_kill_timeout(req);
1474 spin_unlock_irq(&ctx->completion_lock);
1475 return canceled != 0;
1478 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1481 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1482 struct io_defer_entry, list);
1483 struct io_kiocb *link;
1485 if (req_need_defer(de->req, de->seq))
1487 list_del_init(&de->list);
1488 /* punt-init is done before queueing for defer */
1489 link = __io_queue_async_work(de->req);
1491 __io_queue_linked_timeout(link);
1492 /* drop submission reference */
1493 io_put_req_deferred(link, 1);
1496 } while (!list_empty(&ctx->defer_list));
1499 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1501 while (!list_empty(&ctx->timeout_list)) {
1502 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1503 struct io_kiocb, timeout.list);
1505 if (io_is_timeout_noseq(req))
1507 if (req->timeout.target_seq != ctx->cached_cq_tail
1508 - atomic_read(&ctx->cq_timeouts))
1511 list_del_init(&req->timeout.list);
1512 io_kill_timeout(req);
1516 static void io_commit_cqring(struct io_ring_ctx *ctx)
1518 io_flush_timeouts(ctx);
1519 __io_commit_cqring(ctx);
1521 if (unlikely(!list_empty(&ctx->defer_list)))
1522 __io_queue_deferred(ctx);
1525 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1527 struct io_rings *r = ctx->rings;
1529 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1532 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1534 struct io_rings *rings = ctx->rings;
1537 tail = ctx->cached_cq_tail;
1539 * writes to the cq entry need to come after reading head; the
1540 * control dependency is enough as we're using WRITE_ONCE to
1543 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1546 ctx->cached_cq_tail++;
1547 return &rings->cqes[tail & ctx->cq_mask];
1550 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1554 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1556 if (!ctx->eventfd_async)
1558 return io_wq_current_is_worker();
1561 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1563 if (waitqueue_active(&ctx->wait))
1564 wake_up(&ctx->wait);
1565 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1566 wake_up(&ctx->sq_data->wait);
1567 if (io_should_trigger_evfd(ctx))
1568 eventfd_signal(ctx->cq_ev_fd, 1);
1571 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1573 if (list_empty(&ctx->cq_overflow_list)) {
1574 clear_bit(0, &ctx->sq_check_overflow);
1575 clear_bit(0, &ctx->cq_check_overflow);
1576 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1580 static inline bool io_match_files(struct io_kiocb *req,
1581 struct files_struct *files)
1585 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1586 (req->work.flags & IO_WQ_WORK_FILES))
1587 return req->work.identity->files == files;
1591 /* Returns true if there are no backlogged entries after the flush */
1592 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1593 struct task_struct *tsk,
1594 struct files_struct *files)
1596 struct io_rings *rings = ctx->rings;
1597 struct io_kiocb *req, *tmp;
1598 struct io_uring_cqe *cqe;
1599 unsigned long flags;
1603 if (list_empty_careful(&ctx->cq_overflow_list))
1605 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1606 rings->cq_ring_entries))
1610 spin_lock_irqsave(&ctx->completion_lock, flags);
1612 /* if force is set, the ring is going away. always drop after that */
1614 ctx->cq_overflow_flushed = 1;
1617 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1618 if (tsk && req->task != tsk)
1620 if (!io_match_files(req, files))
1623 cqe = io_get_cqring(ctx);
1627 list_move(&req->compl.list, &list);
1629 WRITE_ONCE(cqe->user_data, req->user_data);
1630 WRITE_ONCE(cqe->res, req->result);
1631 WRITE_ONCE(cqe->flags, req->compl.cflags);
1633 ctx->cached_cq_overflow++;
1634 WRITE_ONCE(ctx->rings->cq_overflow,
1635 ctx->cached_cq_overflow);
1639 io_commit_cqring(ctx);
1640 io_cqring_mark_overflow(ctx);
1642 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1643 io_cqring_ev_posted(ctx);
1645 while (!list_empty(&list)) {
1646 req = list_first_entry(&list, struct io_kiocb, compl.list);
1647 list_del(&req->compl.list);
1654 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1656 struct io_ring_ctx *ctx = req->ctx;
1657 struct io_uring_cqe *cqe;
1659 trace_io_uring_complete(ctx, req->user_data, res);
1662 * If we can't get a cq entry, userspace overflowed the
1663 * submission (by quite a lot). Increment the overflow count in
1666 cqe = io_get_cqring(ctx);
1668 WRITE_ONCE(cqe->user_data, req->user_data);
1669 WRITE_ONCE(cqe->res, res);
1670 WRITE_ONCE(cqe->flags, cflags);
1671 } else if (ctx->cq_overflow_flushed ||
1672 atomic_read(&req->task->io_uring->in_idle)) {
1674 * If we're in ring overflow flush mode, or in task cancel mode,
1675 * then we cannot store the request for later flushing, we need
1676 * to drop it on the floor.
1678 ctx->cached_cq_overflow++;
1679 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1681 if (list_empty(&ctx->cq_overflow_list)) {
1682 set_bit(0, &ctx->sq_check_overflow);
1683 set_bit(0, &ctx->cq_check_overflow);
1684 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1688 req->compl.cflags = cflags;
1689 refcount_inc(&req->refs);
1690 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1694 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1696 __io_cqring_fill_event(req, res, 0);
1699 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1701 struct io_ring_ctx *ctx = req->ctx;
1702 unsigned long flags;
1704 spin_lock_irqsave(&ctx->completion_lock, flags);
1705 __io_cqring_fill_event(req, res, cflags);
1706 io_commit_cqring(ctx);
1707 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1709 io_cqring_ev_posted(ctx);
1712 static void io_submit_flush_completions(struct io_comp_state *cs)
1714 struct io_ring_ctx *ctx = cs->ctx;
1716 spin_lock_irq(&ctx->completion_lock);
1717 while (!list_empty(&cs->list)) {
1718 struct io_kiocb *req;
1720 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1721 list_del(&req->compl.list);
1722 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1725 * io_free_req() doesn't care about completion_lock unless one
1726 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1727 * because of a potential deadlock with req->work.fs->lock
1729 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1730 |REQ_F_WORK_INITIALIZED)) {
1731 spin_unlock_irq(&ctx->completion_lock);
1733 spin_lock_irq(&ctx->completion_lock);
1738 io_commit_cqring(ctx);
1739 spin_unlock_irq(&ctx->completion_lock);
1741 io_cqring_ev_posted(ctx);
1745 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1746 struct io_comp_state *cs)
1749 io_cqring_add_event(req, res, cflags);
1754 req->compl.cflags = cflags;
1755 list_add_tail(&req->compl.list, &cs->list);
1757 io_submit_flush_completions(cs);
1761 static void io_req_complete(struct io_kiocb *req, long res)
1763 __io_req_complete(req, res, 0, NULL);
1766 static inline bool io_is_fallback_req(struct io_kiocb *req)
1768 return req == (struct io_kiocb *)
1769 ((unsigned long) req->ctx->fallback_req & ~1UL);
1772 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1774 struct io_kiocb *req;
1776 req = ctx->fallback_req;
1777 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1783 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1784 struct io_submit_state *state)
1786 if (!state->free_reqs) {
1787 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1791 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1792 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1795 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1796 * retry single alloc to be on the safe side.
1798 if (unlikely(ret <= 0)) {
1799 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1800 if (!state->reqs[0])
1804 state->free_reqs = ret;
1808 return state->reqs[state->free_reqs];
1810 return io_get_fallback_req(ctx);
1813 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1817 percpu_ref_put(req->fixed_file_refs);
1822 static void io_dismantle_req(struct io_kiocb *req)
1826 if (req->async_data)
1827 kfree(req->async_data);
1829 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1831 io_req_clean_work(req);
1834 static void __io_free_req(struct io_kiocb *req)
1836 struct io_uring_task *tctx = req->task->io_uring;
1837 struct io_ring_ctx *ctx = req->ctx;
1839 io_dismantle_req(req);
1841 percpu_counter_dec(&tctx->inflight);
1842 if (atomic_read(&tctx->in_idle))
1843 wake_up(&tctx->wait);
1844 put_task_struct(req->task);
1846 if (likely(!io_is_fallback_req(req)))
1847 kmem_cache_free(req_cachep, req);
1849 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1850 percpu_ref_put(&ctx->refs);
1853 static void io_kill_linked_timeout(struct io_kiocb *req)
1855 struct io_ring_ctx *ctx = req->ctx;
1856 struct io_kiocb *link;
1857 bool cancelled = false;
1858 unsigned long flags;
1860 spin_lock_irqsave(&ctx->completion_lock, flags);
1861 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1864 * Can happen if a linked timeout fired and link had been like
1865 * req -> link t-out -> link t-out [-> ...]
1867 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1868 struct io_timeout_data *io = link->async_data;
1871 list_del_init(&link->link_list);
1872 ret = hrtimer_try_to_cancel(&io->timer);
1874 io_cqring_fill_event(link, -ECANCELED);
1875 io_commit_cqring(ctx);
1879 req->flags &= ~REQ_F_LINK_TIMEOUT;
1880 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1883 io_cqring_ev_posted(ctx);
1888 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1890 struct io_kiocb *nxt;
1893 * The list should never be empty when we are called here. But could
1894 * potentially happen if the chain is messed up, check to be on the
1897 if (unlikely(list_empty(&req->link_list)))
1900 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1901 list_del_init(&req->link_list);
1902 if (!list_empty(&nxt->link_list))
1903 nxt->flags |= REQ_F_LINK_HEAD;
1908 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1910 static void io_fail_links(struct io_kiocb *req)
1912 struct io_ring_ctx *ctx = req->ctx;
1913 unsigned long flags;
1915 spin_lock_irqsave(&ctx->completion_lock, flags);
1916 while (!list_empty(&req->link_list)) {
1917 struct io_kiocb *link = list_first_entry(&req->link_list,
1918 struct io_kiocb, link_list);
1920 list_del_init(&link->link_list);
1921 trace_io_uring_fail_link(req, link);
1923 io_cqring_fill_event(link, -ECANCELED);
1926 * It's ok to free under spinlock as they're not linked anymore,
1927 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1930 if (link->flags & REQ_F_WORK_INITIALIZED)
1931 io_put_req_deferred(link, 2);
1933 io_double_put_req(link);
1936 io_commit_cqring(ctx);
1937 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1939 io_cqring_ev_posted(ctx);
1942 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1944 req->flags &= ~REQ_F_LINK_HEAD;
1945 if (req->flags & REQ_F_LINK_TIMEOUT)
1946 io_kill_linked_timeout(req);
1949 * If LINK is set, we have dependent requests in this chain. If we
1950 * didn't fail this request, queue the first one up, moving any other
1951 * dependencies to the next request. In case of failure, fail the rest
1954 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1955 return io_req_link_next(req);
1960 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1962 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1964 return __io_req_find_next(req);
1967 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1969 struct task_struct *tsk = req->task;
1970 struct io_ring_ctx *ctx = req->ctx;
1971 enum task_work_notify_mode notify;
1974 if (tsk->flags & PF_EXITING)
1978 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1979 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1980 * processing task_work. There's no reliable way to tell if TWA_RESUME
1984 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1985 notify = TWA_SIGNAL;
1987 ret = task_work_add(tsk, &req->task_work, notify);
1989 wake_up_process(tsk);
1994 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1996 struct io_ring_ctx *ctx = req->ctx;
1998 spin_lock_irq(&ctx->completion_lock);
1999 io_cqring_fill_event(req, error);
2000 io_commit_cqring(ctx);
2001 spin_unlock_irq(&ctx->completion_lock);
2003 io_cqring_ev_posted(ctx);
2004 req_set_fail_links(req);
2005 io_double_put_req(req);
2008 static void io_req_task_cancel(struct callback_head *cb)
2010 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2011 struct io_ring_ctx *ctx = req->ctx;
2013 __io_req_task_cancel(req, -ECANCELED);
2014 percpu_ref_put(&ctx->refs);
2017 static void __io_req_task_submit(struct io_kiocb *req)
2019 struct io_ring_ctx *ctx = req->ctx;
2021 if (!__io_sq_thread_acquire_mm(ctx)) {
2022 mutex_lock(&ctx->uring_lock);
2023 __io_queue_sqe(req, NULL);
2024 mutex_unlock(&ctx->uring_lock);
2026 __io_req_task_cancel(req, -EFAULT);
2030 static void io_req_task_submit(struct callback_head *cb)
2032 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2033 struct io_ring_ctx *ctx = req->ctx;
2035 __io_req_task_submit(req);
2036 percpu_ref_put(&ctx->refs);
2039 static void io_req_task_queue(struct io_kiocb *req)
2043 init_task_work(&req->task_work, io_req_task_submit);
2044 percpu_ref_get(&req->ctx->refs);
2046 ret = io_req_task_work_add(req, true);
2047 if (unlikely(ret)) {
2048 struct task_struct *tsk;
2050 init_task_work(&req->task_work, io_req_task_cancel);
2051 tsk = io_wq_get_task(req->ctx->io_wq);
2052 task_work_add(tsk, &req->task_work, TWA_NONE);
2053 wake_up_process(tsk);
2057 static void io_queue_next(struct io_kiocb *req)
2059 struct io_kiocb *nxt = io_req_find_next(req);
2062 io_req_task_queue(nxt);
2065 static void io_free_req(struct io_kiocb *req)
2072 void *reqs[IO_IOPOLL_BATCH];
2075 struct task_struct *task;
2079 static inline void io_init_req_batch(struct req_batch *rb)
2086 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2087 struct req_batch *rb)
2089 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2090 percpu_ref_put_many(&ctx->refs, rb->to_free);
2094 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2095 struct req_batch *rb)
2098 __io_req_free_batch_flush(ctx, rb);
2100 struct io_uring_task *tctx = rb->task->io_uring;
2102 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2103 put_task_struct_many(rb->task, rb->task_refs);
2108 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2110 if (unlikely(io_is_fallback_req(req))) {
2114 if (req->flags & REQ_F_LINK_HEAD)
2117 if (req->task != rb->task) {
2119 struct io_uring_task *tctx = rb->task->io_uring;
2121 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2122 put_task_struct_many(rb->task, rb->task_refs);
2124 rb->task = req->task;
2129 io_dismantle_req(req);
2130 rb->reqs[rb->to_free++] = req;
2131 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2132 __io_req_free_batch_flush(req->ctx, rb);
2136 * Drop reference to request, return next in chain (if there is one) if this
2137 * was the last reference to this request.
2139 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2141 struct io_kiocb *nxt = NULL;
2143 if (refcount_dec_and_test(&req->refs)) {
2144 nxt = io_req_find_next(req);
2150 static void io_put_req(struct io_kiocb *req)
2152 if (refcount_dec_and_test(&req->refs))
2156 static void io_put_req_deferred_cb(struct callback_head *cb)
2158 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2163 static void io_free_req_deferred(struct io_kiocb *req)
2167 init_task_work(&req->task_work, io_put_req_deferred_cb);
2168 ret = io_req_task_work_add(req, true);
2169 if (unlikely(ret)) {
2170 struct task_struct *tsk;
2172 tsk = io_wq_get_task(req->ctx->io_wq);
2173 task_work_add(tsk, &req->task_work, TWA_NONE);
2174 wake_up_process(tsk);
2178 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2180 if (refcount_sub_and_test(refs, &req->refs))
2181 io_free_req_deferred(req);
2184 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2186 struct io_kiocb *nxt;
2189 * A ref is owned by io-wq in which context we're. So, if that's the
2190 * last one, it's safe to steal next work. False negatives are Ok,
2191 * it just will be re-punted async in io_put_work()
2193 if (refcount_read(&req->refs) != 1)
2196 nxt = io_req_find_next(req);
2197 return nxt ? &nxt->work : NULL;
2200 static void io_double_put_req(struct io_kiocb *req)
2202 /* drop both submit and complete references */
2203 if (refcount_sub_and_test(2, &req->refs))
2207 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2209 struct io_rings *rings = ctx->rings;
2211 if (test_bit(0, &ctx->cq_check_overflow)) {
2213 * noflush == true is from the waitqueue handler, just ensure
2214 * we wake up the task, and the next invocation will flush the
2215 * entries. We cannot safely to it from here.
2217 if (noflush && !list_empty(&ctx->cq_overflow_list))
2220 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2223 /* See comment at the top of this file */
2225 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2228 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2230 struct io_rings *rings = ctx->rings;
2232 /* make sure SQ entry isn't read before tail */
2233 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2236 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2238 unsigned int cflags;
2240 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2241 cflags |= IORING_CQE_F_BUFFER;
2242 req->flags &= ~REQ_F_BUFFER_SELECTED;
2247 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2249 struct io_buffer *kbuf;
2251 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2252 return io_put_kbuf(req, kbuf);
2255 static inline bool io_run_task_work(void)
2258 * Not safe to run on exiting task, and the task_work handling will
2259 * not add work to such a task.
2261 if (unlikely(current->flags & PF_EXITING))
2263 if (current->task_works) {
2264 __set_current_state(TASK_RUNNING);
2272 static void io_iopoll_queue(struct list_head *again)
2274 struct io_kiocb *req;
2277 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2278 list_del(&req->inflight_entry);
2279 __io_complete_rw(req, -EAGAIN, 0, NULL);
2280 } while (!list_empty(again));
2284 * Find and free completed poll iocbs
2286 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2287 struct list_head *done)
2289 struct req_batch rb;
2290 struct io_kiocb *req;
2293 /* order with ->result store in io_complete_rw_iopoll() */
2296 io_init_req_batch(&rb);
2297 while (!list_empty(done)) {
2300 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2301 if (READ_ONCE(req->result) == -EAGAIN) {
2303 req->iopoll_completed = 0;
2304 list_move_tail(&req->inflight_entry, &again);
2307 list_del(&req->inflight_entry);
2309 if (req->flags & REQ_F_BUFFER_SELECTED)
2310 cflags = io_put_rw_kbuf(req);
2312 __io_cqring_fill_event(req, req->result, cflags);
2315 if (refcount_dec_and_test(&req->refs))
2316 io_req_free_batch(&rb, req);
2319 io_commit_cqring(ctx);
2320 if (ctx->flags & IORING_SETUP_SQPOLL)
2321 io_cqring_ev_posted(ctx);
2322 io_req_free_batch_finish(ctx, &rb);
2324 if (!list_empty(&again))
2325 io_iopoll_queue(&again);
2328 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2331 struct io_kiocb *req, *tmp;
2337 * Only spin for completions if we don't have multiple devices hanging
2338 * off our complete list, and we're under the requested amount.
2340 spin = !ctx->poll_multi_file && *nr_events < min;
2343 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2344 struct kiocb *kiocb = &req->rw.kiocb;
2347 * Move completed and retryable entries to our local lists.
2348 * If we find a request that requires polling, break out
2349 * and complete those lists first, if we have entries there.
2351 if (READ_ONCE(req->iopoll_completed)) {
2352 list_move_tail(&req->inflight_entry, &done);
2355 if (!list_empty(&done))
2358 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2362 /* iopoll may have completed current req */
2363 if (READ_ONCE(req->iopoll_completed))
2364 list_move_tail(&req->inflight_entry, &done);
2371 if (!list_empty(&done))
2372 io_iopoll_complete(ctx, nr_events, &done);
2378 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2379 * non-spinning poll check - we'll still enter the driver poll loop, but only
2380 * as a non-spinning completion check.
2382 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2385 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2388 ret = io_do_iopoll(ctx, nr_events, min);
2391 if (*nr_events >= min)
2399 * We can't just wait for polled events to come to us, we have to actively
2400 * find and complete them.
2402 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2404 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2407 mutex_lock(&ctx->uring_lock);
2408 while (!list_empty(&ctx->iopoll_list)) {
2409 unsigned int nr_events = 0;
2411 io_do_iopoll(ctx, &nr_events, 0);
2413 /* let it sleep and repeat later if can't complete a request */
2417 * Ensure we allow local-to-the-cpu processing to take place,
2418 * in this case we need to ensure that we reap all events.
2419 * Also let task_work, etc. to progress by releasing the mutex
2421 if (need_resched()) {
2422 mutex_unlock(&ctx->uring_lock);
2424 mutex_lock(&ctx->uring_lock);
2427 mutex_unlock(&ctx->uring_lock);
2430 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2432 unsigned int nr_events = 0;
2433 int iters = 0, ret = 0;
2436 * We disallow the app entering submit/complete with polling, but we
2437 * still need to lock the ring to prevent racing with polled issue
2438 * that got punted to a workqueue.
2440 mutex_lock(&ctx->uring_lock);
2443 * Don't enter poll loop if we already have events pending.
2444 * If we do, we can potentially be spinning for commands that
2445 * already triggered a CQE (eg in error).
2447 if (io_cqring_events(ctx, false))
2451 * If a submit got punted to a workqueue, we can have the
2452 * application entering polling for a command before it gets
2453 * issued. That app will hold the uring_lock for the duration
2454 * of the poll right here, so we need to take a breather every
2455 * now and then to ensure that the issue has a chance to add
2456 * the poll to the issued list. Otherwise we can spin here
2457 * forever, while the workqueue is stuck trying to acquire the
2460 if (!(++iters & 7)) {
2461 mutex_unlock(&ctx->uring_lock);
2463 mutex_lock(&ctx->uring_lock);
2466 ret = io_iopoll_getevents(ctx, &nr_events, min);
2470 } while (min && !nr_events && !need_resched());
2472 mutex_unlock(&ctx->uring_lock);
2476 static void kiocb_end_write(struct io_kiocb *req)
2479 * Tell lockdep we inherited freeze protection from submission
2482 if (req->flags & REQ_F_ISREG) {
2483 struct inode *inode = file_inode(req->file);
2485 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2487 file_end_write(req->file);
2490 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2491 struct io_comp_state *cs)
2493 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2496 if (kiocb->ki_flags & IOCB_WRITE)
2497 kiocb_end_write(req);
2499 if (res != req->result)
2500 req_set_fail_links(req);
2501 if (req->flags & REQ_F_BUFFER_SELECTED)
2502 cflags = io_put_rw_kbuf(req);
2503 __io_req_complete(req, res, cflags, cs);
2507 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2509 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2510 ssize_t ret = -ECANCELED;
2511 struct iov_iter iter;
2519 switch (req->opcode) {
2520 case IORING_OP_READV:
2521 case IORING_OP_READ_FIXED:
2522 case IORING_OP_READ:
2525 case IORING_OP_WRITEV:
2526 case IORING_OP_WRITE_FIXED:
2527 case IORING_OP_WRITE:
2531 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2536 if (!req->async_data) {
2537 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2540 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2548 req_set_fail_links(req);
2549 io_req_complete(req, ret);
2554 static bool io_rw_reissue(struct io_kiocb *req, long res)
2557 umode_t mode = file_inode(req->file)->i_mode;
2560 if (!S_ISBLK(mode) && !S_ISREG(mode))
2562 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2565 ret = io_sq_thread_acquire_mm(req->ctx, req);
2567 if (io_resubmit_prep(req, ret)) {
2568 refcount_inc(&req->refs);
2569 io_queue_async_work(req);
2577 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2578 struct io_comp_state *cs)
2580 if (!io_rw_reissue(req, res))
2581 io_complete_rw_common(&req->rw.kiocb, res, cs);
2584 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2586 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2588 __io_complete_rw(req, res, res2, NULL);
2591 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2593 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2595 if (kiocb->ki_flags & IOCB_WRITE)
2596 kiocb_end_write(req);
2598 if (res != -EAGAIN && res != req->result)
2599 req_set_fail_links(req);
2601 WRITE_ONCE(req->result, res);
2602 /* order with io_poll_complete() checking ->result */
2604 WRITE_ONCE(req->iopoll_completed, 1);
2608 * After the iocb has been issued, it's safe to be found on the poll list.
2609 * Adding the kiocb to the list AFTER submission ensures that we don't
2610 * find it from a io_iopoll_getevents() thread before the issuer is done
2611 * accessing the kiocb cookie.
2613 static void io_iopoll_req_issued(struct io_kiocb *req)
2615 struct io_ring_ctx *ctx = req->ctx;
2618 * Track whether we have multiple files in our lists. This will impact
2619 * how we do polling eventually, not spinning if we're on potentially
2620 * different devices.
2622 if (list_empty(&ctx->iopoll_list)) {
2623 ctx->poll_multi_file = false;
2624 } else if (!ctx->poll_multi_file) {
2625 struct io_kiocb *list_req;
2627 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2629 if (list_req->file != req->file)
2630 ctx->poll_multi_file = true;
2634 * For fast devices, IO may have already completed. If it has, add
2635 * it to the front so we find it first.
2637 if (READ_ONCE(req->iopoll_completed))
2638 list_add(&req->inflight_entry, &ctx->iopoll_list);
2640 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2642 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2643 wq_has_sleeper(&ctx->sq_data->wait))
2644 wake_up(&ctx->sq_data->wait);
2647 static void __io_state_file_put(struct io_submit_state *state)
2649 if (state->has_refs)
2650 fput_many(state->file, state->has_refs);
2654 static inline void io_state_file_put(struct io_submit_state *state)
2657 __io_state_file_put(state);
2661 * Get as many references to a file as we have IOs left in this submission,
2662 * assuming most submissions are for one file, or at least that each file
2663 * has more than one submission.
2665 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2671 if (state->fd == fd) {
2675 __io_state_file_put(state);
2677 state->file = fget_many(fd, state->ios_left);
2682 state->has_refs = state->ios_left - 1;
2686 static bool io_bdev_nowait(struct block_device *bdev)
2689 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2696 * If we tracked the file through the SCM inflight mechanism, we could support
2697 * any file. For now, just ensure that anything potentially problematic is done
2700 static bool io_file_supports_async(struct file *file, int rw)
2702 umode_t mode = file_inode(file)->i_mode;
2704 if (S_ISBLK(mode)) {
2705 if (io_bdev_nowait(file->f_inode->i_bdev))
2709 if (S_ISCHR(mode) || S_ISSOCK(mode))
2711 if (S_ISREG(mode)) {
2712 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2713 file->f_op != &io_uring_fops)
2718 /* any ->read/write should understand O_NONBLOCK */
2719 if (file->f_flags & O_NONBLOCK)
2722 if (!(file->f_mode & FMODE_NOWAIT))
2726 return file->f_op->read_iter != NULL;
2728 return file->f_op->write_iter != NULL;
2731 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2733 struct io_ring_ctx *ctx = req->ctx;
2734 struct kiocb *kiocb = &req->rw.kiocb;
2738 if (S_ISREG(file_inode(req->file)->i_mode))
2739 req->flags |= REQ_F_ISREG;
2741 kiocb->ki_pos = READ_ONCE(sqe->off);
2742 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2743 req->flags |= REQ_F_CUR_POS;
2744 kiocb->ki_pos = req->file->f_pos;
2746 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2747 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2748 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2752 ioprio = READ_ONCE(sqe->ioprio);
2754 ret = ioprio_check_cap(ioprio);
2758 kiocb->ki_ioprio = ioprio;
2760 kiocb->ki_ioprio = get_current_ioprio();
2762 /* don't allow async punt if RWF_NOWAIT was requested */
2763 if (kiocb->ki_flags & IOCB_NOWAIT)
2764 req->flags |= REQ_F_NOWAIT;
2766 if (ctx->flags & IORING_SETUP_IOPOLL) {
2767 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2768 !kiocb->ki_filp->f_op->iopoll)
2771 kiocb->ki_flags |= IOCB_HIPRI;
2772 kiocb->ki_complete = io_complete_rw_iopoll;
2773 req->iopoll_completed = 0;
2775 if (kiocb->ki_flags & IOCB_HIPRI)
2777 kiocb->ki_complete = io_complete_rw;
2780 req->rw.addr = READ_ONCE(sqe->addr);
2781 req->rw.len = READ_ONCE(sqe->len);
2782 req->buf_index = READ_ONCE(sqe->buf_index);
2786 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2792 case -ERESTARTNOINTR:
2793 case -ERESTARTNOHAND:
2794 case -ERESTART_RESTARTBLOCK:
2796 * We can't just restart the syscall, since previously
2797 * submitted sqes may already be in progress. Just fail this
2803 kiocb->ki_complete(kiocb, ret, 0);
2807 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2808 struct io_comp_state *cs)
2810 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2811 struct io_async_rw *io = req->async_data;
2813 /* add previously done IO, if any */
2814 if (io && io->bytes_done > 0) {
2816 ret = io->bytes_done;
2818 ret += io->bytes_done;
2821 if (req->flags & REQ_F_CUR_POS)
2822 req->file->f_pos = kiocb->ki_pos;
2823 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2824 __io_complete_rw(req, ret, 0, cs);
2826 io_rw_done(kiocb, ret);
2829 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2830 struct iov_iter *iter)
2832 struct io_ring_ctx *ctx = req->ctx;
2833 size_t len = req->rw.len;
2834 struct io_mapped_ubuf *imu;
2835 u16 index, buf_index = req->buf_index;
2839 if (unlikely(buf_index >= ctx->nr_user_bufs))
2841 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2842 imu = &ctx->user_bufs[index];
2843 buf_addr = req->rw.addr;
2846 if (buf_addr + len < buf_addr)
2848 /* not inside the mapped region */
2849 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2853 * May not be a start of buffer, set size appropriately
2854 * and advance us to the beginning.
2856 offset = buf_addr - imu->ubuf;
2857 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2861 * Don't use iov_iter_advance() here, as it's really slow for
2862 * using the latter parts of a big fixed buffer - it iterates
2863 * over each segment manually. We can cheat a bit here, because
2866 * 1) it's a BVEC iter, we set it up
2867 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2868 * first and last bvec
2870 * So just find our index, and adjust the iterator afterwards.
2871 * If the offset is within the first bvec (or the whole first
2872 * bvec, just use iov_iter_advance(). This makes it easier
2873 * since we can just skip the first segment, which may not
2874 * be PAGE_SIZE aligned.
2876 const struct bio_vec *bvec = imu->bvec;
2878 if (offset <= bvec->bv_len) {
2879 iov_iter_advance(iter, offset);
2881 unsigned long seg_skip;
2883 /* skip first vec */
2884 offset -= bvec->bv_len;
2885 seg_skip = 1 + (offset >> PAGE_SHIFT);
2887 iter->bvec = bvec + seg_skip;
2888 iter->nr_segs -= seg_skip;
2889 iter->count -= bvec->bv_len + offset;
2890 iter->iov_offset = offset & ~PAGE_MASK;
2897 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2900 mutex_unlock(&ctx->uring_lock);
2903 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2906 * "Normal" inline submissions always hold the uring_lock, since we
2907 * grab it from the system call. Same is true for the SQPOLL offload.
2908 * The only exception is when we've detached the request and issue it
2909 * from an async worker thread, grab the lock for that case.
2912 mutex_lock(&ctx->uring_lock);
2915 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2916 int bgid, struct io_buffer *kbuf,
2919 struct io_buffer *head;
2921 if (req->flags & REQ_F_BUFFER_SELECTED)
2924 io_ring_submit_lock(req->ctx, needs_lock);
2926 lockdep_assert_held(&req->ctx->uring_lock);
2928 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2930 if (!list_empty(&head->list)) {
2931 kbuf = list_last_entry(&head->list, struct io_buffer,
2933 list_del(&kbuf->list);
2936 idr_remove(&req->ctx->io_buffer_idr, bgid);
2938 if (*len > kbuf->len)
2941 kbuf = ERR_PTR(-ENOBUFS);
2944 io_ring_submit_unlock(req->ctx, needs_lock);
2949 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2952 struct io_buffer *kbuf;
2955 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2956 bgid = req->buf_index;
2957 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2960 req->rw.addr = (u64) (unsigned long) kbuf;
2961 req->flags |= REQ_F_BUFFER_SELECTED;
2962 return u64_to_user_ptr(kbuf->addr);
2965 #ifdef CONFIG_COMPAT
2966 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2969 struct compat_iovec __user *uiov;
2970 compat_ssize_t clen;
2974 uiov = u64_to_user_ptr(req->rw.addr);
2975 if (!access_ok(uiov, sizeof(*uiov)))
2977 if (__get_user(clen, &uiov->iov_len))
2983 buf = io_rw_buffer_select(req, &len, needs_lock);
2985 return PTR_ERR(buf);
2986 iov[0].iov_base = buf;
2987 iov[0].iov_len = (compat_size_t) len;
2992 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2995 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2999 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3002 len = iov[0].iov_len;
3005 buf = io_rw_buffer_select(req, &len, needs_lock);
3007 return PTR_ERR(buf);
3008 iov[0].iov_base = buf;
3009 iov[0].iov_len = len;
3013 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3016 if (req->flags & REQ_F_BUFFER_SELECTED) {
3017 struct io_buffer *kbuf;
3019 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3020 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3021 iov[0].iov_len = kbuf->len;
3026 else if (req->rw.len > 1)
3029 #ifdef CONFIG_COMPAT
3030 if (req->ctx->compat)
3031 return io_compat_import(req, iov, needs_lock);
3034 return __io_iov_buffer_select(req, iov, needs_lock);
3037 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3038 struct iovec **iovec, struct iov_iter *iter,
3041 void __user *buf = u64_to_user_ptr(req->rw.addr);
3042 size_t sqe_len = req->rw.len;
3046 opcode = req->opcode;
3047 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3049 return io_import_fixed(req, rw, iter);
3052 /* buffer index only valid with fixed read/write, or buffer select */
3053 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3056 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3057 if (req->flags & REQ_F_BUFFER_SELECT) {
3058 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3060 return PTR_ERR(buf);
3061 req->rw.len = sqe_len;
3064 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3066 return ret < 0 ? ret : sqe_len;
3069 if (req->flags & REQ_F_BUFFER_SELECT) {
3070 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3072 ret = (*iovec)->iov_len;
3073 iov_iter_init(iter, rw, *iovec, 1, ret);
3079 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3083 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3084 struct iovec **iovec, struct iov_iter *iter,
3087 struct io_async_rw *iorw = req->async_data;
3090 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3092 return iov_iter_count(&iorw->iter);
3095 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3097 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3101 * For files that don't have ->read_iter() and ->write_iter(), handle them
3102 * by looping over ->read() or ->write() manually.
3104 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3106 struct kiocb *kiocb = &req->rw.kiocb;
3107 struct file *file = req->file;
3111 * Don't support polled IO through this interface, and we can't
3112 * support non-blocking either. For the latter, this just causes
3113 * the kiocb to be handled from an async context.
3115 if (kiocb->ki_flags & IOCB_HIPRI)
3117 if (kiocb->ki_flags & IOCB_NOWAIT)
3120 while (iov_iter_count(iter)) {
3124 if (!iov_iter_is_bvec(iter)) {
3125 iovec = iov_iter_iovec(iter);
3127 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3128 iovec.iov_len = req->rw.len;
3132 nr = file->f_op->read(file, iovec.iov_base,
3133 iovec.iov_len, io_kiocb_ppos(kiocb));
3135 nr = file->f_op->write(file, iovec.iov_base,
3136 iovec.iov_len, io_kiocb_ppos(kiocb));
3145 if (nr != iovec.iov_len)
3149 iov_iter_advance(iter, nr);
3155 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3156 const struct iovec *fast_iov, struct iov_iter *iter)
3158 struct io_async_rw *rw = req->async_data;
3160 memcpy(&rw->iter, iter, sizeof(*iter));
3161 rw->free_iovec = iovec;
3163 /* can only be fixed buffers, no need to do anything */
3164 if (iter->type == ITER_BVEC)
3167 unsigned iov_off = 0;
3169 rw->iter.iov = rw->fast_iov;
3170 if (iter->iov != fast_iov) {
3171 iov_off = iter->iov - fast_iov;
3172 rw->iter.iov += iov_off;
3174 if (rw->fast_iov != fast_iov)
3175 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3176 sizeof(struct iovec) * iter->nr_segs);
3178 req->flags |= REQ_F_NEED_CLEANUP;
3182 static inline int __io_alloc_async_data(struct io_kiocb *req)
3184 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3185 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3186 return req->async_data == NULL;
3189 static int io_alloc_async_data(struct io_kiocb *req)
3191 if (!io_op_defs[req->opcode].needs_async_data)
3194 return __io_alloc_async_data(req);
3197 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3198 const struct iovec *fast_iov,
3199 struct iov_iter *iter, bool force)
3201 if (!force && !io_op_defs[req->opcode].needs_async_data)
3203 if (!req->async_data) {
3204 if (__io_alloc_async_data(req))
3207 io_req_map_rw(req, iovec, fast_iov, iter);
3212 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3214 struct io_async_rw *iorw = req->async_data;
3215 struct iovec *iov = iorw->fast_iov;
3218 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3219 if (unlikely(ret < 0))
3222 iorw->bytes_done = 0;
3223 iorw->free_iovec = iov;
3225 req->flags |= REQ_F_NEED_CLEANUP;
3229 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3233 ret = io_prep_rw(req, sqe);
3237 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3240 /* either don't need iovec imported or already have it */
3241 if (!req->async_data)
3243 return io_rw_prep_async(req, READ);
3247 * This is our waitqueue callback handler, registered through lock_page_async()
3248 * when we initially tried to do the IO with the iocb armed our waitqueue.
3249 * This gets called when the page is unlocked, and we generally expect that to
3250 * happen when the page IO is completed and the page is now uptodate. This will
3251 * queue a task_work based retry of the operation, attempting to copy the data
3252 * again. If the latter fails because the page was NOT uptodate, then we will
3253 * do a thread based blocking retry of the operation. That's the unexpected
3256 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3257 int sync, void *arg)
3259 struct wait_page_queue *wpq;
3260 struct io_kiocb *req = wait->private;
3261 struct wait_page_key *key = arg;
3264 wpq = container_of(wait, struct wait_page_queue, wait);
3266 if (!wake_page_match(wpq, key))
3269 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3270 list_del_init(&wait->entry);
3272 init_task_work(&req->task_work, io_req_task_submit);
3273 percpu_ref_get(&req->ctx->refs);
3275 /* submit ref gets dropped, acquire a new one */
3276 refcount_inc(&req->refs);
3277 ret = io_req_task_work_add(req, true);
3278 if (unlikely(ret)) {
3279 struct task_struct *tsk;
3281 /* queue just for cancelation */
3282 init_task_work(&req->task_work, io_req_task_cancel);
3283 tsk = io_wq_get_task(req->ctx->io_wq);
3284 task_work_add(tsk, &req->task_work, TWA_NONE);
3285 wake_up_process(tsk);
3291 * This controls whether a given IO request should be armed for async page
3292 * based retry. If we return false here, the request is handed to the async
3293 * worker threads for retry. If we're doing buffered reads on a regular file,
3294 * we prepare a private wait_page_queue entry and retry the operation. This
3295 * will either succeed because the page is now uptodate and unlocked, or it
3296 * will register a callback when the page is unlocked at IO completion. Through
3297 * that callback, io_uring uses task_work to setup a retry of the operation.
3298 * That retry will attempt the buffered read again. The retry will generally
3299 * succeed, or in rare cases where it fails, we then fall back to using the
3300 * async worker threads for a blocking retry.
3302 static bool io_rw_should_retry(struct io_kiocb *req)
3304 struct io_async_rw *rw = req->async_data;
3305 struct wait_page_queue *wait = &rw->wpq;
3306 struct kiocb *kiocb = &req->rw.kiocb;
3308 /* never retry for NOWAIT, we just complete with -EAGAIN */
3309 if (req->flags & REQ_F_NOWAIT)
3312 /* Only for buffered IO */
3313 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3317 * just use poll if we can, and don't attempt if the fs doesn't
3318 * support callback based unlocks
3320 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3323 wait->wait.func = io_async_buf_func;
3324 wait->wait.private = req;
3325 wait->wait.flags = 0;
3326 INIT_LIST_HEAD(&wait->wait.entry);
3327 kiocb->ki_flags |= IOCB_WAITQ;
3328 kiocb->ki_flags &= ~IOCB_NOWAIT;
3329 kiocb->ki_waitq = wait;
3333 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3335 if (req->file->f_op->read_iter)
3336 return call_read_iter(req->file, &req->rw.kiocb, iter);
3337 else if (req->file->f_op->read)
3338 return loop_rw_iter(READ, req, iter);
3343 static int io_read(struct io_kiocb *req, bool force_nonblock,
3344 struct io_comp_state *cs)
3346 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3347 struct kiocb *kiocb = &req->rw.kiocb;
3348 struct iov_iter __iter, *iter = &__iter;
3349 struct io_async_rw *rw = req->async_data;
3350 ssize_t io_size, ret, ret2;
3357 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3360 iov_count = iov_iter_count(iter);
3362 req->result = io_size;
3365 /* Ensure we clear previously set non-block flag */
3366 if (!force_nonblock)
3367 kiocb->ki_flags &= ~IOCB_NOWAIT;
3369 kiocb->ki_flags |= IOCB_NOWAIT;
3372 /* If the file doesn't support async, just async punt */
3373 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3377 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3381 ret = io_iter_do_read(req, iter);
3385 } else if (ret == -EIOCBQUEUED) {
3388 } else if (ret == -EAGAIN) {
3389 /* IOPOLL retry should happen for io-wq threads */
3390 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3392 /* no retry on NONBLOCK marked file */
3393 if (req->file->f_flags & O_NONBLOCK)
3395 /* some cases will consume bytes even on error returns */
3396 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3399 } else if (ret < 0) {
3400 /* make sure -ERESTARTSYS -> -EINTR is done */
3404 /* read it all, or we did blocking attempt. no retry. */
3405 if (!iov_iter_count(iter) || !force_nonblock ||
3406 (req->file->f_flags & O_NONBLOCK))
3411 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3418 rw = req->async_data;
3419 /* it's copied and will be cleaned with ->io */
3421 /* now use our persistent iterator, if we aren't already */
3424 rw->bytes_done += ret;
3425 /* if we can retry, do so with the callbacks armed */
3426 if (!io_rw_should_retry(req)) {
3427 kiocb->ki_flags &= ~IOCB_WAITQ;
3432 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3433 * get -EIOCBQUEUED, then we'll get a notification when the desired
3434 * page gets unlocked. We can also get a partial read here, and if we
3435 * do, then just retry at the new offset.
3437 ret = io_iter_do_read(req, iter);
3438 if (ret == -EIOCBQUEUED) {
3441 } else if (ret > 0 && ret < io_size) {
3442 /* we got some bytes, but not all. retry. */
3446 kiocb_done(kiocb, ret, cs);
3449 /* it's reportedly faster than delegating the null check to kfree() */
3455 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3459 ret = io_prep_rw(req, sqe);
3463 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3466 /* either don't need iovec imported or already have it */
3467 if (!req->async_data)
3469 return io_rw_prep_async(req, WRITE);
3472 static int io_write(struct io_kiocb *req, bool force_nonblock,
3473 struct io_comp_state *cs)
3475 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3476 struct kiocb *kiocb = &req->rw.kiocb;
3477 struct iov_iter __iter, *iter = &__iter;
3478 struct io_async_rw *rw = req->async_data;
3480 ssize_t ret, ret2, io_size;
3485 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3488 iov_count = iov_iter_count(iter);
3490 req->result = io_size;
3492 /* Ensure we clear previously set non-block flag */
3493 if (!force_nonblock)
3494 kiocb->ki_flags &= ~IOCB_NOWAIT;
3496 kiocb->ki_flags |= IOCB_NOWAIT;
3498 /* If the file doesn't support async, just async punt */
3499 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3502 /* file path doesn't support NOWAIT for non-direct_IO */
3503 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3504 (req->flags & REQ_F_ISREG))
3507 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3512 * Open-code file_start_write here to grab freeze protection,
3513 * which will be released by another thread in
3514 * io_complete_rw(). Fool lockdep by telling it the lock got
3515 * released so that it doesn't complain about the held lock when
3516 * we return to userspace.
3518 if (req->flags & REQ_F_ISREG) {
3519 __sb_start_write(file_inode(req->file)->i_sb,
3520 SB_FREEZE_WRITE, true);
3521 __sb_writers_release(file_inode(req->file)->i_sb,
3524 kiocb->ki_flags |= IOCB_WRITE;
3526 if (req->file->f_op->write_iter)
3527 ret2 = call_write_iter(req->file, kiocb, iter);
3528 else if (req->file->f_op->write)
3529 ret2 = loop_rw_iter(WRITE, req, iter);
3534 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3535 * retry them without IOCB_NOWAIT.
3537 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3539 /* no retry on NONBLOCK marked file */
3540 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3542 if (!force_nonblock || ret2 != -EAGAIN) {
3543 /* IOPOLL retry should happen for io-wq threads */
3544 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3547 kiocb_done(kiocb, ret2, cs);
3550 /* some cases will consume bytes even on error returns */
3551 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3552 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3557 /* it's reportedly faster than delegating the null check to kfree() */
3563 static int __io_splice_prep(struct io_kiocb *req,
3564 const struct io_uring_sqe *sqe)
3566 struct io_splice* sp = &req->splice;
3567 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3569 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3573 sp->len = READ_ONCE(sqe->len);
3574 sp->flags = READ_ONCE(sqe->splice_flags);
3576 if (unlikely(sp->flags & ~valid_flags))
3579 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3580 (sp->flags & SPLICE_F_FD_IN_FIXED));
3583 req->flags |= REQ_F_NEED_CLEANUP;
3585 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3587 * Splice operation will be punted aync, and here need to
3588 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3590 io_req_init_async(req);
3591 req->work.flags |= IO_WQ_WORK_UNBOUND;
3597 static int io_tee_prep(struct io_kiocb *req,
3598 const struct io_uring_sqe *sqe)
3600 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3602 return __io_splice_prep(req, sqe);
3605 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3607 struct io_splice *sp = &req->splice;
3608 struct file *in = sp->file_in;
3609 struct file *out = sp->file_out;
3610 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3616 ret = do_tee(in, out, sp->len, flags);
3618 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3619 req->flags &= ~REQ_F_NEED_CLEANUP;
3622 req_set_fail_links(req);
3623 io_req_complete(req, ret);
3627 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3629 struct io_splice* sp = &req->splice;
3631 sp->off_in = READ_ONCE(sqe->splice_off_in);
3632 sp->off_out = READ_ONCE(sqe->off);
3633 return __io_splice_prep(req, sqe);
3636 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3638 struct io_splice *sp = &req->splice;
3639 struct file *in = sp->file_in;
3640 struct file *out = sp->file_out;
3641 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3642 loff_t *poff_in, *poff_out;
3648 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3649 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3652 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3654 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3655 req->flags &= ~REQ_F_NEED_CLEANUP;
3658 req_set_fail_links(req);
3659 io_req_complete(req, ret);
3664 * IORING_OP_NOP just posts a completion event, nothing else.
3666 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3668 struct io_ring_ctx *ctx = req->ctx;
3670 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3673 __io_req_complete(req, 0, 0, cs);
3677 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3679 struct io_ring_ctx *ctx = req->ctx;
3684 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3686 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3689 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3690 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3693 req->sync.off = READ_ONCE(sqe->off);
3694 req->sync.len = READ_ONCE(sqe->len);
3698 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3700 loff_t end = req->sync.off + req->sync.len;
3703 /* fsync always requires a blocking context */
3707 ret = vfs_fsync_range(req->file, req->sync.off,
3708 end > 0 ? end : LLONG_MAX,
3709 req->sync.flags & IORING_FSYNC_DATASYNC);
3711 req_set_fail_links(req);
3712 io_req_complete(req, ret);
3716 static int io_fallocate_prep(struct io_kiocb *req,
3717 const struct io_uring_sqe *sqe)
3719 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3721 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3724 req->sync.off = READ_ONCE(sqe->off);
3725 req->sync.len = READ_ONCE(sqe->addr);
3726 req->sync.mode = READ_ONCE(sqe->len);
3730 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3734 /* fallocate always requiring blocking context */
3737 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3740 req_set_fail_links(req);
3741 io_req_complete(req, ret);
3745 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3747 const char __user *fname;
3750 if (unlikely(sqe->ioprio || sqe->buf_index))
3752 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3755 /* open.how should be already initialised */
3756 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3757 req->open.how.flags |= O_LARGEFILE;
3759 req->open.dfd = READ_ONCE(sqe->fd);
3760 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3761 req->open.filename = getname(fname);
3762 if (IS_ERR(req->open.filename)) {
3763 ret = PTR_ERR(req->open.filename);
3764 req->open.filename = NULL;
3767 req->open.nofile = rlimit(RLIMIT_NOFILE);
3768 req->flags |= REQ_F_NEED_CLEANUP;
3772 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3776 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3778 mode = READ_ONCE(sqe->len);
3779 flags = READ_ONCE(sqe->open_flags);
3780 req->open.how = build_open_how(flags, mode);
3781 return __io_openat_prep(req, sqe);
3784 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3786 struct open_how __user *how;
3790 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3792 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3793 len = READ_ONCE(sqe->len);
3794 if (len < OPEN_HOW_SIZE_VER0)
3797 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3802 return __io_openat_prep(req, sqe);
3805 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3807 struct open_flags op;
3814 ret = build_open_flags(&req->open.how, &op);
3818 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3822 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3825 ret = PTR_ERR(file);
3827 fsnotify_open(file);
3828 fd_install(ret, file);
3831 putname(req->open.filename);
3832 req->flags &= ~REQ_F_NEED_CLEANUP;
3834 req_set_fail_links(req);
3835 io_req_complete(req, ret);
3839 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3841 return io_openat2(req, force_nonblock);
3844 static int io_remove_buffers_prep(struct io_kiocb *req,
3845 const struct io_uring_sqe *sqe)
3847 struct io_provide_buf *p = &req->pbuf;
3850 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3853 tmp = READ_ONCE(sqe->fd);
3854 if (!tmp || tmp > USHRT_MAX)
3857 memset(p, 0, sizeof(*p));
3859 p->bgid = READ_ONCE(sqe->buf_group);
3863 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3864 int bgid, unsigned nbufs)
3868 /* shouldn't happen */
3872 /* the head kbuf is the list itself */
3873 while (!list_empty(&buf->list)) {
3874 struct io_buffer *nxt;
3876 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3877 list_del(&nxt->list);
3884 idr_remove(&ctx->io_buffer_idr, bgid);
3889 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3890 struct io_comp_state *cs)
3892 struct io_provide_buf *p = &req->pbuf;
3893 struct io_ring_ctx *ctx = req->ctx;
3894 struct io_buffer *head;
3897 io_ring_submit_lock(ctx, !force_nonblock);
3899 lockdep_assert_held(&ctx->uring_lock);
3902 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3904 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3906 io_ring_submit_lock(ctx, !force_nonblock);
3908 req_set_fail_links(req);
3909 __io_req_complete(req, ret, 0, cs);
3913 static int io_provide_buffers_prep(struct io_kiocb *req,
3914 const struct io_uring_sqe *sqe)
3916 struct io_provide_buf *p = &req->pbuf;
3919 if (sqe->ioprio || sqe->rw_flags)
3922 tmp = READ_ONCE(sqe->fd);
3923 if (!tmp || tmp > USHRT_MAX)
3926 p->addr = READ_ONCE(sqe->addr);
3927 p->len = READ_ONCE(sqe->len);
3929 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3932 p->bgid = READ_ONCE(sqe->buf_group);
3933 tmp = READ_ONCE(sqe->off);
3934 if (tmp > USHRT_MAX)
3940 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3942 struct io_buffer *buf;
3943 u64 addr = pbuf->addr;
3944 int i, bid = pbuf->bid;
3946 for (i = 0; i < pbuf->nbufs; i++) {
3947 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3952 buf->len = pbuf->len;
3957 INIT_LIST_HEAD(&buf->list);
3960 list_add_tail(&buf->list, &(*head)->list);
3964 return i ? i : -ENOMEM;
3967 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3968 struct io_comp_state *cs)
3970 struct io_provide_buf *p = &req->pbuf;
3971 struct io_ring_ctx *ctx = req->ctx;
3972 struct io_buffer *head, *list;
3975 io_ring_submit_lock(ctx, !force_nonblock);
3977 lockdep_assert_held(&ctx->uring_lock);
3979 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3981 ret = io_add_buffers(p, &head);
3986 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3989 __io_remove_buffers(ctx, head, p->bgid, -1U);
3994 io_ring_submit_unlock(ctx, !force_nonblock);
3996 req_set_fail_links(req);
3997 __io_req_complete(req, ret, 0, cs);
4001 static int io_epoll_ctl_prep(struct io_kiocb *req,
4002 const struct io_uring_sqe *sqe)
4004 #if defined(CONFIG_EPOLL)
4005 if (sqe->ioprio || sqe->buf_index)
4007 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4010 req->epoll.epfd = READ_ONCE(sqe->fd);
4011 req->epoll.op = READ_ONCE(sqe->len);
4012 req->epoll.fd = READ_ONCE(sqe->off);
4014 if (ep_op_has_event(req->epoll.op)) {
4015 struct epoll_event __user *ev;
4017 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4018 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4028 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4029 struct io_comp_state *cs)
4031 #if defined(CONFIG_EPOLL)
4032 struct io_epoll *ie = &req->epoll;
4035 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4036 if (force_nonblock && ret == -EAGAIN)
4040 req_set_fail_links(req);
4041 __io_req_complete(req, ret, 0, cs);
4048 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4050 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4051 if (sqe->ioprio || sqe->buf_index || sqe->off)
4053 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4056 req->madvise.addr = READ_ONCE(sqe->addr);
4057 req->madvise.len = READ_ONCE(sqe->len);
4058 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4065 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4067 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4068 struct io_madvise *ma = &req->madvise;
4074 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4076 req_set_fail_links(req);
4077 io_req_complete(req, ret);
4084 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4086 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4088 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4091 req->fadvise.offset = READ_ONCE(sqe->off);
4092 req->fadvise.len = READ_ONCE(sqe->len);
4093 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4097 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4099 struct io_fadvise *fa = &req->fadvise;
4102 if (force_nonblock) {
4103 switch (fa->advice) {
4104 case POSIX_FADV_NORMAL:
4105 case POSIX_FADV_RANDOM:
4106 case POSIX_FADV_SEQUENTIAL:
4113 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4115 req_set_fail_links(req);
4116 io_req_complete(req, ret);
4120 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4122 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4124 if (sqe->ioprio || sqe->buf_index)
4126 if (req->flags & REQ_F_FIXED_FILE)
4129 req->statx.dfd = READ_ONCE(sqe->fd);
4130 req->statx.mask = READ_ONCE(sqe->len);
4131 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4132 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4133 req->statx.flags = READ_ONCE(sqe->statx_flags);
4138 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4140 struct io_statx *ctx = &req->statx;
4143 if (force_nonblock) {
4144 /* only need file table for an actual valid fd */
4145 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4146 req->flags |= REQ_F_NO_FILE_TABLE;
4150 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4154 req_set_fail_links(req);
4155 io_req_complete(req, ret);
4159 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4162 * If we queue this for async, it must not be cancellable. That would
4163 * leave the 'file' in an undeterminate state, and here need to modify
4164 * io_wq_work.flags, so initialize io_wq_work firstly.
4166 io_req_init_async(req);
4167 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4169 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4171 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4172 sqe->rw_flags || sqe->buf_index)
4174 if (req->flags & REQ_F_FIXED_FILE)
4177 req->close.fd = READ_ONCE(sqe->fd);
4178 if ((req->file && req->file->f_op == &io_uring_fops))
4181 req->close.put_file = NULL;
4185 static int io_close(struct io_kiocb *req, bool force_nonblock,
4186 struct io_comp_state *cs)
4188 struct io_close *close = &req->close;
4191 /* might be already done during nonblock submission */
4192 if (!close->put_file) {
4193 ret = __close_fd_get_file(close->fd, &close->put_file);
4195 return (ret == -ENOENT) ? -EBADF : ret;
4198 /* if the file has a flush method, be safe and punt to async */
4199 if (close->put_file->f_op->flush && force_nonblock) {
4200 /* was never set, but play safe */
4201 req->flags &= ~REQ_F_NOWAIT;
4202 /* avoid grabbing files - we don't need the files */
4203 req->flags |= REQ_F_NO_FILE_TABLE;
4207 /* No ->flush() or already async, safely close from here */
4208 ret = filp_close(close->put_file, req->work.identity->files);
4210 req_set_fail_links(req);
4211 fput(close->put_file);
4212 close->put_file = NULL;
4213 __io_req_complete(req, ret, 0, cs);
4217 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4219 struct io_ring_ctx *ctx = req->ctx;
4224 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4226 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4229 req->sync.off = READ_ONCE(sqe->off);
4230 req->sync.len = READ_ONCE(sqe->len);
4231 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4235 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4239 /* sync_file_range always requires a blocking context */
4243 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4246 req_set_fail_links(req);
4247 io_req_complete(req, ret);
4251 #if defined(CONFIG_NET)
4252 static int io_setup_async_msg(struct io_kiocb *req,
4253 struct io_async_msghdr *kmsg)
4255 struct io_async_msghdr *async_msg = req->async_data;
4259 if (io_alloc_async_data(req)) {
4260 if (kmsg->iov != kmsg->fast_iov)
4264 async_msg = req->async_data;
4265 req->flags |= REQ_F_NEED_CLEANUP;
4266 memcpy(async_msg, kmsg, sizeof(*kmsg));
4270 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4271 struct io_async_msghdr *iomsg)
4273 iomsg->iov = iomsg->fast_iov;
4274 iomsg->msg.msg_name = &iomsg->addr;
4275 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4276 req->sr_msg.msg_flags, &iomsg->iov);
4279 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4281 struct io_async_msghdr *async_msg = req->async_data;
4282 struct io_sr_msg *sr = &req->sr_msg;
4285 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4288 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4289 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4290 sr->len = READ_ONCE(sqe->len);
4292 #ifdef CONFIG_COMPAT
4293 if (req->ctx->compat)
4294 sr->msg_flags |= MSG_CMSG_COMPAT;
4297 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4299 ret = io_sendmsg_copy_hdr(req, async_msg);
4301 req->flags |= REQ_F_NEED_CLEANUP;
4305 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4306 struct io_comp_state *cs)
4308 struct io_async_msghdr iomsg, *kmsg;
4309 struct socket *sock;
4313 sock = sock_from_file(req->file, &ret);
4314 if (unlikely(!sock))
4317 if (req->async_data) {
4318 kmsg = req->async_data;
4319 kmsg->msg.msg_name = &kmsg->addr;
4320 /* if iov is set, it's allocated already */
4322 kmsg->iov = kmsg->fast_iov;
4323 kmsg->msg.msg_iter.iov = kmsg->iov;
4325 ret = io_sendmsg_copy_hdr(req, &iomsg);
4331 flags = req->sr_msg.msg_flags;
4332 if (flags & MSG_DONTWAIT)
4333 req->flags |= REQ_F_NOWAIT;
4334 else if (force_nonblock)
4335 flags |= MSG_DONTWAIT;
4337 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4338 if (force_nonblock && ret == -EAGAIN)
4339 return io_setup_async_msg(req, kmsg);
4340 if (ret == -ERESTARTSYS)
4343 if (kmsg->iov != kmsg->fast_iov)
4345 req->flags &= ~REQ_F_NEED_CLEANUP;
4347 req_set_fail_links(req);
4348 __io_req_complete(req, ret, 0, cs);
4352 static int io_send(struct io_kiocb *req, bool force_nonblock,
4353 struct io_comp_state *cs)
4355 struct io_sr_msg *sr = &req->sr_msg;
4358 struct socket *sock;
4362 sock = sock_from_file(req->file, &ret);
4363 if (unlikely(!sock))
4366 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4370 msg.msg_name = NULL;
4371 msg.msg_control = NULL;
4372 msg.msg_controllen = 0;
4373 msg.msg_namelen = 0;
4375 flags = req->sr_msg.msg_flags;
4376 if (flags & MSG_DONTWAIT)
4377 req->flags |= REQ_F_NOWAIT;
4378 else if (force_nonblock)
4379 flags |= MSG_DONTWAIT;
4381 msg.msg_flags = flags;
4382 ret = sock_sendmsg(sock, &msg);
4383 if (force_nonblock && ret == -EAGAIN)
4385 if (ret == -ERESTARTSYS)
4389 req_set_fail_links(req);
4390 __io_req_complete(req, ret, 0, cs);
4394 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4395 struct io_async_msghdr *iomsg)
4397 struct io_sr_msg *sr = &req->sr_msg;
4398 struct iovec __user *uiov;
4402 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4403 &iomsg->uaddr, &uiov, &iov_len);
4407 if (req->flags & REQ_F_BUFFER_SELECT) {
4410 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4412 sr->len = iomsg->iov[0].iov_len;
4413 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4417 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4418 &iomsg->iov, &iomsg->msg.msg_iter,
4427 #ifdef CONFIG_COMPAT
4428 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4429 struct io_async_msghdr *iomsg)
4431 struct compat_msghdr __user *msg_compat;
4432 struct io_sr_msg *sr = &req->sr_msg;
4433 struct compat_iovec __user *uiov;
4438 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4439 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4444 uiov = compat_ptr(ptr);
4445 if (req->flags & REQ_F_BUFFER_SELECT) {
4446 compat_ssize_t clen;
4450 if (!access_ok(uiov, sizeof(*uiov)))
4452 if (__get_user(clen, &uiov->iov_len))
4456 sr->len = iomsg->iov[0].iov_len;
4459 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4460 UIO_FASTIOV, &iomsg->iov,
4461 &iomsg->msg.msg_iter, true);
4470 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4471 struct io_async_msghdr *iomsg)
4473 iomsg->msg.msg_name = &iomsg->addr;
4474 iomsg->iov = iomsg->fast_iov;
4476 #ifdef CONFIG_COMPAT
4477 if (req->ctx->compat)
4478 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4481 return __io_recvmsg_copy_hdr(req, iomsg);
4484 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4487 struct io_sr_msg *sr = &req->sr_msg;
4488 struct io_buffer *kbuf;
4490 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4495 req->flags |= REQ_F_BUFFER_SELECTED;
4499 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4501 return io_put_kbuf(req, req->sr_msg.kbuf);
4504 static int io_recvmsg_prep(struct io_kiocb *req,
4505 const struct io_uring_sqe *sqe)
4507 struct io_async_msghdr *async_msg = req->async_data;
4508 struct io_sr_msg *sr = &req->sr_msg;
4511 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4514 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4515 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4516 sr->len = READ_ONCE(sqe->len);
4517 sr->bgid = READ_ONCE(sqe->buf_group);
4519 #ifdef CONFIG_COMPAT
4520 if (req->ctx->compat)
4521 sr->msg_flags |= MSG_CMSG_COMPAT;
4524 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4526 ret = io_recvmsg_copy_hdr(req, async_msg);
4528 req->flags |= REQ_F_NEED_CLEANUP;
4532 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4533 struct io_comp_state *cs)
4535 struct io_async_msghdr iomsg, *kmsg;
4536 struct socket *sock;
4537 struct io_buffer *kbuf;
4539 int ret, cflags = 0;
4541 sock = sock_from_file(req->file, &ret);
4542 if (unlikely(!sock))
4545 if (req->async_data) {
4546 kmsg = req->async_data;
4547 kmsg->msg.msg_name = &kmsg->addr;
4548 /* if iov is set, it's allocated already */
4550 kmsg->iov = kmsg->fast_iov;
4551 kmsg->msg.msg_iter.iov = kmsg->iov;
4553 ret = io_recvmsg_copy_hdr(req, &iomsg);
4559 if (req->flags & REQ_F_BUFFER_SELECT) {
4560 kbuf = io_recv_buffer_select(req, !force_nonblock);
4562 return PTR_ERR(kbuf);
4563 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4564 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4565 1, req->sr_msg.len);
4568 flags = req->sr_msg.msg_flags;
4569 if (flags & MSG_DONTWAIT)
4570 req->flags |= REQ_F_NOWAIT;
4571 else if (force_nonblock)
4572 flags |= MSG_DONTWAIT;
4574 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4575 kmsg->uaddr, flags);
4576 if (force_nonblock && ret == -EAGAIN)
4577 return io_setup_async_msg(req, kmsg);
4578 if (ret == -ERESTARTSYS)
4581 if (req->flags & REQ_F_BUFFER_SELECTED)
4582 cflags = io_put_recv_kbuf(req);
4583 if (kmsg->iov != kmsg->fast_iov)
4585 req->flags &= ~REQ_F_NEED_CLEANUP;
4587 req_set_fail_links(req);
4588 __io_req_complete(req, ret, cflags, cs);
4592 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4593 struct io_comp_state *cs)
4595 struct io_buffer *kbuf;
4596 struct io_sr_msg *sr = &req->sr_msg;
4598 void __user *buf = sr->buf;
4599 struct socket *sock;
4602 int ret, cflags = 0;
4604 sock = sock_from_file(req->file, &ret);
4605 if (unlikely(!sock))
4608 if (req->flags & REQ_F_BUFFER_SELECT) {
4609 kbuf = io_recv_buffer_select(req, !force_nonblock);
4611 return PTR_ERR(kbuf);
4612 buf = u64_to_user_ptr(kbuf->addr);
4615 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4619 msg.msg_name = NULL;
4620 msg.msg_control = NULL;
4621 msg.msg_controllen = 0;
4622 msg.msg_namelen = 0;
4623 msg.msg_iocb = NULL;
4626 flags = req->sr_msg.msg_flags;
4627 if (flags & MSG_DONTWAIT)
4628 req->flags |= REQ_F_NOWAIT;
4629 else if (force_nonblock)
4630 flags |= MSG_DONTWAIT;
4632 ret = sock_recvmsg(sock, &msg, flags);
4633 if (force_nonblock && ret == -EAGAIN)
4635 if (ret == -ERESTARTSYS)
4638 if (req->flags & REQ_F_BUFFER_SELECTED)
4639 cflags = io_put_recv_kbuf(req);
4641 req_set_fail_links(req);
4642 __io_req_complete(req, ret, cflags, cs);
4646 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4648 struct io_accept *accept = &req->accept;
4650 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4652 if (sqe->ioprio || sqe->len || sqe->buf_index)
4655 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4656 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4657 accept->flags = READ_ONCE(sqe->accept_flags);
4658 accept->nofile = rlimit(RLIMIT_NOFILE);
4662 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4663 struct io_comp_state *cs)
4665 struct io_accept *accept = &req->accept;
4666 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4669 if (req->file->f_flags & O_NONBLOCK)
4670 req->flags |= REQ_F_NOWAIT;
4672 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4673 accept->addr_len, accept->flags,
4675 if (ret == -EAGAIN && force_nonblock)
4678 if (ret == -ERESTARTSYS)
4680 req_set_fail_links(req);
4682 __io_req_complete(req, ret, 0, cs);
4686 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4688 struct io_connect *conn = &req->connect;
4689 struct io_async_connect *io = req->async_data;
4691 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4693 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4696 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4697 conn->addr_len = READ_ONCE(sqe->addr2);
4702 return move_addr_to_kernel(conn->addr, conn->addr_len,
4706 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4707 struct io_comp_state *cs)
4709 struct io_async_connect __io, *io;
4710 unsigned file_flags;
4713 if (req->async_data) {
4714 io = req->async_data;
4716 ret = move_addr_to_kernel(req->connect.addr,
4717 req->connect.addr_len,
4724 file_flags = force_nonblock ? O_NONBLOCK : 0;
4726 ret = __sys_connect_file(req->file, &io->address,
4727 req->connect.addr_len, file_flags);
4728 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4729 if (req->async_data)
4731 if (io_alloc_async_data(req)) {
4735 io = req->async_data;
4736 memcpy(req->async_data, &__io, sizeof(__io));
4739 if (ret == -ERESTARTSYS)
4743 req_set_fail_links(req);
4744 __io_req_complete(req, ret, 0, cs);
4747 #else /* !CONFIG_NET */
4748 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4753 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4754 struct io_comp_state *cs)
4759 static int io_send(struct io_kiocb *req, bool force_nonblock,
4760 struct io_comp_state *cs)
4765 static int io_recvmsg_prep(struct io_kiocb *req,
4766 const struct io_uring_sqe *sqe)
4771 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4772 struct io_comp_state *cs)
4777 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4778 struct io_comp_state *cs)
4783 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4788 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4789 struct io_comp_state *cs)
4794 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4799 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4800 struct io_comp_state *cs)
4804 #endif /* CONFIG_NET */
4806 struct io_poll_table {
4807 struct poll_table_struct pt;
4808 struct io_kiocb *req;
4812 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4813 __poll_t mask, task_work_func_t func)
4818 /* for instances that support it check for an event match first: */
4819 if (mask && !(mask & poll->events))
4822 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4824 list_del_init(&poll->wait.entry);
4827 init_task_work(&req->task_work, func);
4828 percpu_ref_get(&req->ctx->refs);
4831 * If we using the signalfd wait_queue_head for this wakeup, then
4832 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4833 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4834 * either, as the normal wakeup will suffice.
4836 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4839 * If this fails, then the task is exiting. When a task exits, the
4840 * work gets canceled, so just cancel this request as well instead
4841 * of executing it. We can't safely execute it anyway, as we may not
4842 * have the needed state needed for it anyway.
4844 ret = io_req_task_work_add(req, twa_signal_ok);
4845 if (unlikely(ret)) {
4846 struct task_struct *tsk;
4848 WRITE_ONCE(poll->canceled, true);
4849 tsk = io_wq_get_task(req->ctx->io_wq);
4850 task_work_add(tsk, &req->task_work, TWA_NONE);
4851 wake_up_process(tsk);
4856 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4857 __acquires(&req->ctx->completion_lock)
4859 struct io_ring_ctx *ctx = req->ctx;
4861 if (!req->result && !READ_ONCE(poll->canceled)) {
4862 struct poll_table_struct pt = { ._key = poll->events };
4864 req->result = vfs_poll(req->file, &pt) & poll->events;
4867 spin_lock_irq(&ctx->completion_lock);
4868 if (!req->result && !READ_ONCE(poll->canceled)) {
4869 add_wait_queue(poll->head, &poll->wait);
4876 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4878 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4879 if (req->opcode == IORING_OP_POLL_ADD)
4880 return req->async_data;
4881 return req->apoll->double_poll;
4884 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4886 if (req->opcode == IORING_OP_POLL_ADD)
4888 return &req->apoll->poll;
4891 static void io_poll_remove_double(struct io_kiocb *req)
4893 struct io_poll_iocb *poll = io_poll_get_double(req);
4895 lockdep_assert_held(&req->ctx->completion_lock);
4897 if (poll && poll->head) {
4898 struct wait_queue_head *head = poll->head;
4900 spin_lock(&head->lock);
4901 list_del_init(&poll->wait.entry);
4902 if (poll->wait.private)
4903 refcount_dec(&req->refs);
4905 spin_unlock(&head->lock);
4909 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4911 struct io_ring_ctx *ctx = req->ctx;
4913 io_poll_remove_double(req);
4914 req->poll.done = true;
4915 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4916 io_commit_cqring(ctx);
4919 static void io_poll_task_func(struct callback_head *cb)
4921 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4922 struct io_ring_ctx *ctx = req->ctx;
4923 struct io_kiocb *nxt;
4925 if (io_poll_rewait(req, &req->poll)) {
4926 spin_unlock_irq(&ctx->completion_lock);
4928 hash_del(&req->hash_node);
4929 io_poll_complete(req, req->result, 0);
4930 spin_unlock_irq(&ctx->completion_lock);
4932 nxt = io_put_req_find_next(req);
4933 io_cqring_ev_posted(ctx);
4935 __io_req_task_submit(nxt);
4938 percpu_ref_put(&ctx->refs);
4941 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4942 int sync, void *key)
4944 struct io_kiocb *req = wait->private;
4945 struct io_poll_iocb *poll = io_poll_get_single(req);
4946 __poll_t mask = key_to_poll(key);
4948 /* for instances that support it check for an event match first: */
4949 if (mask && !(mask & poll->events))
4952 list_del_init(&wait->entry);
4954 if (poll && poll->head) {
4957 spin_lock(&poll->head->lock);
4958 done = list_empty(&poll->wait.entry);
4960 list_del_init(&poll->wait.entry);
4961 /* make sure double remove sees this as being gone */
4962 wait->private = NULL;
4963 spin_unlock(&poll->head->lock);
4965 /* use wait func handler, so it matches the rq type */
4966 poll->wait.func(&poll->wait, mode, sync, key);
4969 refcount_dec(&req->refs);
4973 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4974 wait_queue_func_t wake_func)
4978 poll->canceled = false;
4979 poll->events = events;
4980 INIT_LIST_HEAD(&poll->wait.entry);
4981 init_waitqueue_func_entry(&poll->wait, wake_func);
4984 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4985 struct wait_queue_head *head,
4986 struct io_poll_iocb **poll_ptr)
4988 struct io_kiocb *req = pt->req;
4991 * If poll->head is already set, it's because the file being polled
4992 * uses multiple waitqueues for poll handling (eg one for read, one
4993 * for write). Setup a separate io_poll_iocb if this happens.
4995 if (unlikely(poll->head)) {
4996 struct io_poll_iocb *poll_one = poll;
4998 /* already have a 2nd entry, fail a third attempt */
5000 pt->error = -EINVAL;
5003 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5005 pt->error = -ENOMEM;
5008 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5009 refcount_inc(&req->refs);
5010 poll->wait.private = req;
5017 if (poll->events & EPOLLEXCLUSIVE)
5018 add_wait_queue_exclusive(head, &poll->wait);
5020 add_wait_queue(head, &poll->wait);
5023 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5024 struct poll_table_struct *p)
5026 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5027 struct async_poll *apoll = pt->req->apoll;
5029 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5032 static void io_async_task_func(struct callback_head *cb)
5034 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5035 struct async_poll *apoll = req->apoll;
5036 struct io_ring_ctx *ctx = req->ctx;
5038 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5040 if (io_poll_rewait(req, &apoll->poll)) {
5041 spin_unlock_irq(&ctx->completion_lock);
5042 percpu_ref_put(&ctx->refs);
5046 /* If req is still hashed, it cannot have been canceled. Don't check. */
5047 if (hash_hashed(&req->hash_node))
5048 hash_del(&req->hash_node);
5050 io_poll_remove_double(req);
5051 spin_unlock_irq(&ctx->completion_lock);
5053 if (!READ_ONCE(apoll->poll.canceled))
5054 __io_req_task_submit(req);
5056 __io_req_task_cancel(req, -ECANCELED);
5058 percpu_ref_put(&ctx->refs);
5059 kfree(apoll->double_poll);
5063 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5066 struct io_kiocb *req = wait->private;
5067 struct io_poll_iocb *poll = &req->apoll->poll;
5069 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5072 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5075 static void io_poll_req_insert(struct io_kiocb *req)
5077 struct io_ring_ctx *ctx = req->ctx;
5078 struct hlist_head *list;
5080 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5081 hlist_add_head(&req->hash_node, list);
5084 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5085 struct io_poll_iocb *poll,
5086 struct io_poll_table *ipt, __poll_t mask,
5087 wait_queue_func_t wake_func)
5088 __acquires(&ctx->completion_lock)
5090 struct io_ring_ctx *ctx = req->ctx;
5091 bool cancel = false;
5093 INIT_HLIST_NODE(&req->hash_node);
5094 io_init_poll_iocb(poll, mask, wake_func);
5095 poll->file = req->file;
5096 poll->wait.private = req;
5098 ipt->pt._key = mask;
5100 ipt->error = -EINVAL;
5102 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5104 spin_lock_irq(&ctx->completion_lock);
5105 if (likely(poll->head)) {
5106 spin_lock(&poll->head->lock);
5107 if (unlikely(list_empty(&poll->wait.entry))) {
5113 if (mask || ipt->error)
5114 list_del_init(&poll->wait.entry);
5116 WRITE_ONCE(poll->canceled, true);
5117 else if (!poll->done) /* actually waiting for an event */
5118 io_poll_req_insert(req);
5119 spin_unlock(&poll->head->lock);
5125 static bool io_arm_poll_handler(struct io_kiocb *req)
5127 const struct io_op_def *def = &io_op_defs[req->opcode];
5128 struct io_ring_ctx *ctx = req->ctx;
5129 struct async_poll *apoll;
5130 struct io_poll_table ipt;
5134 if (!req->file || !file_can_poll(req->file))
5136 if (req->flags & REQ_F_POLLED)
5140 else if (def->pollout)
5144 /* if we can't nonblock try, then no point in arming a poll handler */
5145 if (!io_file_supports_async(req->file, rw))
5148 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5149 if (unlikely(!apoll))
5151 apoll->double_poll = NULL;
5153 req->flags |= REQ_F_POLLED;
5158 mask |= POLLIN | POLLRDNORM;
5160 mask |= POLLOUT | POLLWRNORM;
5162 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5163 if ((req->opcode == IORING_OP_RECVMSG) &&
5164 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5167 mask |= POLLERR | POLLPRI;
5169 ipt.pt._qproc = io_async_queue_proc;
5171 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5173 if (ret || ipt.error) {
5174 io_poll_remove_double(req);
5175 spin_unlock_irq(&ctx->completion_lock);
5176 kfree(apoll->double_poll);
5180 spin_unlock_irq(&ctx->completion_lock);
5181 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5182 apoll->poll.events);
5186 static bool __io_poll_remove_one(struct io_kiocb *req,
5187 struct io_poll_iocb *poll)
5189 bool do_complete = false;
5191 spin_lock(&poll->head->lock);
5192 WRITE_ONCE(poll->canceled, true);
5193 if (!list_empty(&poll->wait.entry)) {
5194 list_del_init(&poll->wait.entry);
5197 spin_unlock(&poll->head->lock);
5198 hash_del(&req->hash_node);
5202 static bool io_poll_remove_one(struct io_kiocb *req)
5206 io_poll_remove_double(req);
5208 if (req->opcode == IORING_OP_POLL_ADD) {
5209 do_complete = __io_poll_remove_one(req, &req->poll);
5211 struct async_poll *apoll = req->apoll;
5213 /* non-poll requests have submit ref still */
5214 do_complete = __io_poll_remove_one(req, &apoll->poll);
5217 kfree(apoll->double_poll);
5223 io_cqring_fill_event(req, -ECANCELED);
5224 io_commit_cqring(req->ctx);
5225 req_set_fail_links(req);
5226 io_put_req_deferred(req, 1);
5233 * Returns true if we found and killed one or more poll requests
5235 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5237 struct hlist_node *tmp;
5238 struct io_kiocb *req;
5241 spin_lock_irq(&ctx->completion_lock);
5242 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5243 struct hlist_head *list;
5245 list = &ctx->cancel_hash[i];
5246 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5247 if (io_task_match(req, tsk))
5248 posted += io_poll_remove_one(req);
5251 spin_unlock_irq(&ctx->completion_lock);
5254 io_cqring_ev_posted(ctx);
5259 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5261 struct hlist_head *list;
5262 struct io_kiocb *req;
5264 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5265 hlist_for_each_entry(req, list, hash_node) {
5266 if (sqe_addr != req->user_data)
5268 if (io_poll_remove_one(req))
5276 static int io_poll_remove_prep(struct io_kiocb *req,
5277 const struct io_uring_sqe *sqe)
5279 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5281 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5285 req->poll.addr = READ_ONCE(sqe->addr);
5290 * Find a running poll command that matches one specified in sqe->addr,
5291 * and remove it if found.
5293 static int io_poll_remove(struct io_kiocb *req)
5295 struct io_ring_ctx *ctx = req->ctx;
5299 addr = req->poll.addr;
5300 spin_lock_irq(&ctx->completion_lock);
5301 ret = io_poll_cancel(ctx, addr);
5302 spin_unlock_irq(&ctx->completion_lock);
5305 req_set_fail_links(req);
5306 io_req_complete(req, ret);
5310 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5313 struct io_kiocb *req = wait->private;
5314 struct io_poll_iocb *poll = &req->poll;
5316 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5319 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5320 struct poll_table_struct *p)
5322 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5324 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5327 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5329 struct io_poll_iocb *poll = &req->poll;
5332 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5334 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5337 events = READ_ONCE(sqe->poll32_events);
5339 events = swahw32(events);
5341 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5342 (events & EPOLLEXCLUSIVE);
5346 static int io_poll_add(struct io_kiocb *req)
5348 struct io_poll_iocb *poll = &req->poll;
5349 struct io_ring_ctx *ctx = req->ctx;
5350 struct io_poll_table ipt;
5353 ipt.pt._qproc = io_poll_queue_proc;
5355 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5358 if (mask) { /* no async, we'd stolen it */
5360 io_poll_complete(req, mask, 0);
5362 spin_unlock_irq(&ctx->completion_lock);
5365 io_cqring_ev_posted(ctx);
5371 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5373 struct io_timeout_data *data = container_of(timer,
5374 struct io_timeout_data, timer);
5375 struct io_kiocb *req = data->req;
5376 struct io_ring_ctx *ctx = req->ctx;
5377 unsigned long flags;
5379 spin_lock_irqsave(&ctx->completion_lock, flags);
5380 list_del_init(&req->timeout.list);
5381 atomic_set(&req->ctx->cq_timeouts,
5382 atomic_read(&req->ctx->cq_timeouts) + 1);
5384 io_cqring_fill_event(req, -ETIME);
5385 io_commit_cqring(ctx);
5386 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5388 io_cqring_ev_posted(ctx);
5389 req_set_fail_links(req);
5391 return HRTIMER_NORESTART;
5394 static int __io_timeout_cancel(struct io_kiocb *req)
5396 struct io_timeout_data *io = req->async_data;
5399 ret = hrtimer_try_to_cancel(&io->timer);
5402 list_del_init(&req->timeout.list);
5404 req_set_fail_links(req);
5405 io_cqring_fill_event(req, -ECANCELED);
5406 io_put_req_deferred(req, 1);
5410 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5412 struct io_kiocb *req;
5415 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5416 if (user_data == req->user_data) {
5425 return __io_timeout_cancel(req);
5428 static int io_timeout_remove_prep(struct io_kiocb *req,
5429 const struct io_uring_sqe *sqe)
5431 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5433 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5435 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5438 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5443 * Remove or update an existing timeout command
5445 static int io_timeout_remove(struct io_kiocb *req)
5447 struct io_ring_ctx *ctx = req->ctx;
5450 spin_lock_irq(&ctx->completion_lock);
5451 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5453 io_cqring_fill_event(req, ret);
5454 io_commit_cqring(ctx);
5455 spin_unlock_irq(&ctx->completion_lock);
5456 io_cqring_ev_posted(ctx);
5458 req_set_fail_links(req);
5463 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5464 bool is_timeout_link)
5466 struct io_timeout_data *data;
5468 u32 off = READ_ONCE(sqe->off);
5470 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5472 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5474 if (off && is_timeout_link)
5476 flags = READ_ONCE(sqe->timeout_flags);
5477 if (flags & ~IORING_TIMEOUT_ABS)
5480 req->timeout.off = off;
5482 if (!req->async_data && io_alloc_async_data(req))
5485 data = req->async_data;
5488 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5491 if (flags & IORING_TIMEOUT_ABS)
5492 data->mode = HRTIMER_MODE_ABS;
5494 data->mode = HRTIMER_MODE_REL;
5496 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5500 static int io_timeout(struct io_kiocb *req)
5502 struct io_ring_ctx *ctx = req->ctx;
5503 struct io_timeout_data *data = req->async_data;
5504 struct list_head *entry;
5505 u32 tail, off = req->timeout.off;
5507 spin_lock_irq(&ctx->completion_lock);
5510 * sqe->off holds how many events that need to occur for this
5511 * timeout event to be satisfied. If it isn't set, then this is
5512 * a pure timeout request, sequence isn't used.
5514 if (io_is_timeout_noseq(req)) {
5515 entry = ctx->timeout_list.prev;
5519 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5520 req->timeout.target_seq = tail + off;
5523 * Insertion sort, ensuring the first entry in the list is always
5524 * the one we need first.
5526 list_for_each_prev(entry, &ctx->timeout_list) {
5527 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5530 if (io_is_timeout_noseq(nxt))
5532 /* nxt.seq is behind @tail, otherwise would've been completed */
5533 if (off >= nxt->timeout.target_seq - tail)
5537 list_add(&req->timeout.list, entry);
5538 data->timer.function = io_timeout_fn;
5539 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5540 spin_unlock_irq(&ctx->completion_lock);
5544 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5546 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5548 return req->user_data == (unsigned long) data;
5551 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5553 enum io_wq_cancel cancel_ret;
5556 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5557 switch (cancel_ret) {
5558 case IO_WQ_CANCEL_OK:
5561 case IO_WQ_CANCEL_RUNNING:
5564 case IO_WQ_CANCEL_NOTFOUND:
5572 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5573 struct io_kiocb *req, __u64 sqe_addr,
5576 unsigned long flags;
5579 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5580 if (ret != -ENOENT) {
5581 spin_lock_irqsave(&ctx->completion_lock, flags);
5585 spin_lock_irqsave(&ctx->completion_lock, flags);
5586 ret = io_timeout_cancel(ctx, sqe_addr);
5589 ret = io_poll_cancel(ctx, sqe_addr);
5593 io_cqring_fill_event(req, ret);
5594 io_commit_cqring(ctx);
5595 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5596 io_cqring_ev_posted(ctx);
5599 req_set_fail_links(req);
5603 static int io_async_cancel_prep(struct io_kiocb *req,
5604 const struct io_uring_sqe *sqe)
5606 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5608 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5610 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5613 req->cancel.addr = READ_ONCE(sqe->addr);
5617 static int io_async_cancel(struct io_kiocb *req)
5619 struct io_ring_ctx *ctx = req->ctx;
5621 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5625 static int io_files_update_prep(struct io_kiocb *req,
5626 const struct io_uring_sqe *sqe)
5628 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5630 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5632 if (sqe->ioprio || sqe->rw_flags)
5635 req->files_update.offset = READ_ONCE(sqe->off);
5636 req->files_update.nr_args = READ_ONCE(sqe->len);
5637 if (!req->files_update.nr_args)
5639 req->files_update.arg = READ_ONCE(sqe->addr);
5643 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5644 struct io_comp_state *cs)
5646 struct io_ring_ctx *ctx = req->ctx;
5647 struct io_uring_files_update up;
5653 up.offset = req->files_update.offset;
5654 up.fds = req->files_update.arg;
5656 mutex_lock(&ctx->uring_lock);
5657 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5658 mutex_unlock(&ctx->uring_lock);
5661 req_set_fail_links(req);
5662 __io_req_complete(req, ret, 0, cs);
5666 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5668 switch (req->opcode) {
5671 case IORING_OP_READV:
5672 case IORING_OP_READ_FIXED:
5673 case IORING_OP_READ:
5674 return io_read_prep(req, sqe);
5675 case IORING_OP_WRITEV:
5676 case IORING_OP_WRITE_FIXED:
5677 case IORING_OP_WRITE:
5678 return io_write_prep(req, sqe);
5679 case IORING_OP_POLL_ADD:
5680 return io_poll_add_prep(req, sqe);
5681 case IORING_OP_POLL_REMOVE:
5682 return io_poll_remove_prep(req, sqe);
5683 case IORING_OP_FSYNC:
5684 return io_prep_fsync(req, sqe);
5685 case IORING_OP_SYNC_FILE_RANGE:
5686 return io_prep_sfr(req, sqe);
5687 case IORING_OP_SENDMSG:
5688 case IORING_OP_SEND:
5689 return io_sendmsg_prep(req, sqe);
5690 case IORING_OP_RECVMSG:
5691 case IORING_OP_RECV:
5692 return io_recvmsg_prep(req, sqe);
5693 case IORING_OP_CONNECT:
5694 return io_connect_prep(req, sqe);
5695 case IORING_OP_TIMEOUT:
5696 return io_timeout_prep(req, sqe, false);
5697 case IORING_OP_TIMEOUT_REMOVE:
5698 return io_timeout_remove_prep(req, sqe);
5699 case IORING_OP_ASYNC_CANCEL:
5700 return io_async_cancel_prep(req, sqe);
5701 case IORING_OP_LINK_TIMEOUT:
5702 return io_timeout_prep(req, sqe, true);
5703 case IORING_OP_ACCEPT:
5704 return io_accept_prep(req, sqe);
5705 case IORING_OP_FALLOCATE:
5706 return io_fallocate_prep(req, sqe);
5707 case IORING_OP_OPENAT:
5708 return io_openat_prep(req, sqe);
5709 case IORING_OP_CLOSE:
5710 return io_close_prep(req, sqe);
5711 case IORING_OP_FILES_UPDATE:
5712 return io_files_update_prep(req, sqe);
5713 case IORING_OP_STATX:
5714 return io_statx_prep(req, sqe);
5715 case IORING_OP_FADVISE:
5716 return io_fadvise_prep(req, sqe);
5717 case IORING_OP_MADVISE:
5718 return io_madvise_prep(req, sqe);
5719 case IORING_OP_OPENAT2:
5720 return io_openat2_prep(req, sqe);
5721 case IORING_OP_EPOLL_CTL:
5722 return io_epoll_ctl_prep(req, sqe);
5723 case IORING_OP_SPLICE:
5724 return io_splice_prep(req, sqe);
5725 case IORING_OP_PROVIDE_BUFFERS:
5726 return io_provide_buffers_prep(req, sqe);
5727 case IORING_OP_REMOVE_BUFFERS:
5728 return io_remove_buffers_prep(req, sqe);
5730 return io_tee_prep(req, sqe);
5733 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5738 static int io_req_defer_prep(struct io_kiocb *req,
5739 const struct io_uring_sqe *sqe)
5743 if (io_alloc_async_data(req))
5745 return io_req_prep(req, sqe);
5748 static u32 io_get_sequence(struct io_kiocb *req)
5750 struct io_kiocb *pos;
5751 struct io_ring_ctx *ctx = req->ctx;
5752 u32 total_submitted, nr_reqs = 1;
5754 if (req->flags & REQ_F_LINK_HEAD)
5755 list_for_each_entry(pos, &req->link_list, link_list)
5758 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5759 return total_submitted - nr_reqs;
5762 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5764 struct io_ring_ctx *ctx = req->ctx;
5765 struct io_defer_entry *de;
5769 /* Still need defer if there is pending req in defer list. */
5770 if (likely(list_empty_careful(&ctx->defer_list) &&
5771 !(req->flags & REQ_F_IO_DRAIN)))
5774 seq = io_get_sequence(req);
5775 /* Still a chance to pass the sequence check */
5776 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5779 if (!req->async_data) {
5780 ret = io_req_defer_prep(req, sqe);
5784 io_prep_async_link(req);
5785 de = kmalloc(sizeof(*de), GFP_KERNEL);
5789 spin_lock_irq(&ctx->completion_lock);
5790 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5791 spin_unlock_irq(&ctx->completion_lock);
5793 io_queue_async_work(req);
5794 return -EIOCBQUEUED;
5797 trace_io_uring_defer(ctx, req, req->user_data);
5800 list_add_tail(&de->list, &ctx->defer_list);
5801 spin_unlock_irq(&ctx->completion_lock);
5802 return -EIOCBQUEUED;
5805 static void io_req_drop_files(struct io_kiocb *req)
5807 struct io_ring_ctx *ctx = req->ctx;
5808 unsigned long flags;
5810 spin_lock_irqsave(&ctx->inflight_lock, flags);
5811 list_del(&req->inflight_entry);
5812 if (waitqueue_active(&ctx->inflight_wait))
5813 wake_up(&ctx->inflight_wait);
5814 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5815 req->flags &= ~REQ_F_INFLIGHT;
5816 put_files_struct(req->work.identity->files);
5817 put_nsproxy(req->work.identity->nsproxy);
5818 req->work.flags &= ~IO_WQ_WORK_FILES;
5821 static void __io_clean_op(struct io_kiocb *req)
5823 if (req->flags & REQ_F_BUFFER_SELECTED) {
5824 switch (req->opcode) {
5825 case IORING_OP_READV:
5826 case IORING_OP_READ_FIXED:
5827 case IORING_OP_READ:
5828 kfree((void *)(unsigned long)req->rw.addr);
5830 case IORING_OP_RECVMSG:
5831 case IORING_OP_RECV:
5832 kfree(req->sr_msg.kbuf);
5835 req->flags &= ~REQ_F_BUFFER_SELECTED;
5838 if (req->flags & REQ_F_NEED_CLEANUP) {
5839 switch (req->opcode) {
5840 case IORING_OP_READV:
5841 case IORING_OP_READ_FIXED:
5842 case IORING_OP_READ:
5843 case IORING_OP_WRITEV:
5844 case IORING_OP_WRITE_FIXED:
5845 case IORING_OP_WRITE: {
5846 struct io_async_rw *io = req->async_data;
5848 kfree(io->free_iovec);
5851 case IORING_OP_RECVMSG:
5852 case IORING_OP_SENDMSG: {
5853 struct io_async_msghdr *io = req->async_data;
5854 if (io->iov != io->fast_iov)
5858 case IORING_OP_SPLICE:
5860 io_put_file(req, req->splice.file_in,
5861 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5863 case IORING_OP_OPENAT:
5864 case IORING_OP_OPENAT2:
5865 if (req->open.filename)
5866 putname(req->open.filename);
5869 req->flags &= ~REQ_F_NEED_CLEANUP;
5872 if (req->flags & REQ_F_INFLIGHT)
5873 io_req_drop_files(req);
5876 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5877 struct io_comp_state *cs)
5879 struct io_ring_ctx *ctx = req->ctx;
5882 switch (req->opcode) {
5884 ret = io_nop(req, cs);
5886 case IORING_OP_READV:
5887 case IORING_OP_READ_FIXED:
5888 case IORING_OP_READ:
5889 ret = io_read(req, force_nonblock, cs);
5891 case IORING_OP_WRITEV:
5892 case IORING_OP_WRITE_FIXED:
5893 case IORING_OP_WRITE:
5894 ret = io_write(req, force_nonblock, cs);
5896 case IORING_OP_FSYNC:
5897 ret = io_fsync(req, force_nonblock);
5899 case IORING_OP_POLL_ADD:
5900 ret = io_poll_add(req);
5902 case IORING_OP_POLL_REMOVE:
5903 ret = io_poll_remove(req);
5905 case IORING_OP_SYNC_FILE_RANGE:
5906 ret = io_sync_file_range(req, force_nonblock);
5908 case IORING_OP_SENDMSG:
5909 ret = io_sendmsg(req, force_nonblock, cs);
5911 case IORING_OP_SEND:
5912 ret = io_send(req, force_nonblock, cs);
5914 case IORING_OP_RECVMSG:
5915 ret = io_recvmsg(req, force_nonblock, cs);
5917 case IORING_OP_RECV:
5918 ret = io_recv(req, force_nonblock, cs);
5920 case IORING_OP_TIMEOUT:
5921 ret = io_timeout(req);
5923 case IORING_OP_TIMEOUT_REMOVE:
5924 ret = io_timeout_remove(req);
5926 case IORING_OP_ACCEPT:
5927 ret = io_accept(req, force_nonblock, cs);
5929 case IORING_OP_CONNECT:
5930 ret = io_connect(req, force_nonblock, cs);
5932 case IORING_OP_ASYNC_CANCEL:
5933 ret = io_async_cancel(req);
5935 case IORING_OP_FALLOCATE:
5936 ret = io_fallocate(req, force_nonblock);
5938 case IORING_OP_OPENAT:
5939 ret = io_openat(req, force_nonblock);
5941 case IORING_OP_CLOSE:
5942 ret = io_close(req, force_nonblock, cs);
5944 case IORING_OP_FILES_UPDATE:
5945 ret = io_files_update(req, force_nonblock, cs);
5947 case IORING_OP_STATX:
5948 ret = io_statx(req, force_nonblock);
5950 case IORING_OP_FADVISE:
5951 ret = io_fadvise(req, force_nonblock);
5953 case IORING_OP_MADVISE:
5954 ret = io_madvise(req, force_nonblock);
5956 case IORING_OP_OPENAT2:
5957 ret = io_openat2(req, force_nonblock);
5959 case IORING_OP_EPOLL_CTL:
5960 ret = io_epoll_ctl(req, force_nonblock, cs);
5962 case IORING_OP_SPLICE:
5963 ret = io_splice(req, force_nonblock);
5965 case IORING_OP_PROVIDE_BUFFERS:
5966 ret = io_provide_buffers(req, force_nonblock, cs);
5968 case IORING_OP_REMOVE_BUFFERS:
5969 ret = io_remove_buffers(req, force_nonblock, cs);
5972 ret = io_tee(req, force_nonblock);
5982 /* If the op doesn't have a file, we're not polling for it */
5983 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5984 const bool in_async = io_wq_current_is_worker();
5986 /* workqueue context doesn't hold uring_lock, grab it now */
5988 mutex_lock(&ctx->uring_lock);
5990 io_iopoll_req_issued(req);
5993 mutex_unlock(&ctx->uring_lock);
5999 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6001 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6002 struct io_kiocb *timeout;
6005 timeout = io_prep_linked_timeout(req);
6007 io_queue_linked_timeout(timeout);
6009 /* if NO_CANCEL is set, we must still run the work */
6010 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6011 IO_WQ_WORK_CANCEL) {
6017 ret = io_issue_sqe(req, false, NULL);
6019 * We can get EAGAIN for polled IO even though we're
6020 * forcing a sync submission from here, since we can't
6021 * wait for request slots on the block side.
6030 req_set_fail_links(req);
6031 io_req_complete(req, ret);
6034 return io_steal_work(req);
6037 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6040 struct fixed_file_table *table;
6042 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6043 return table->files[index & IORING_FILE_TABLE_MASK];
6046 static struct file *io_file_get(struct io_submit_state *state,
6047 struct io_kiocb *req, int fd, bool fixed)
6049 struct io_ring_ctx *ctx = req->ctx;
6053 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6055 fd = array_index_nospec(fd, ctx->nr_user_files);
6056 file = io_file_from_index(ctx, fd);
6058 req->fixed_file_refs = &ctx->file_data->node->refs;
6059 percpu_ref_get(req->fixed_file_refs);
6062 trace_io_uring_file_get(ctx, fd);
6063 file = __io_file_get(state, fd);
6069 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6074 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6075 if (unlikely(!fixed && io_async_submit(req->ctx)))
6078 req->file = io_file_get(state, req, fd, fixed);
6079 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6084 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6086 struct io_timeout_data *data = container_of(timer,
6087 struct io_timeout_data, timer);
6088 struct io_kiocb *req = data->req;
6089 struct io_ring_ctx *ctx = req->ctx;
6090 struct io_kiocb *prev = NULL;
6091 unsigned long flags;
6093 spin_lock_irqsave(&ctx->completion_lock, flags);
6096 * We don't expect the list to be empty, that will only happen if we
6097 * race with the completion of the linked work.
6099 if (!list_empty(&req->link_list)) {
6100 prev = list_entry(req->link_list.prev, struct io_kiocb,
6102 if (refcount_inc_not_zero(&prev->refs))
6103 list_del_init(&req->link_list);
6108 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6111 req_set_fail_links(prev);
6112 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6115 io_req_complete(req, -ETIME);
6117 return HRTIMER_NORESTART;
6120 static void __io_queue_linked_timeout(struct io_kiocb *req)
6123 * If the list is now empty, then our linked request finished before
6124 * we got a chance to setup the timer
6126 if (!list_empty(&req->link_list)) {
6127 struct io_timeout_data *data = req->async_data;
6129 data->timer.function = io_link_timeout_fn;
6130 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6135 static void io_queue_linked_timeout(struct io_kiocb *req)
6137 struct io_ring_ctx *ctx = req->ctx;
6139 spin_lock_irq(&ctx->completion_lock);
6140 __io_queue_linked_timeout(req);
6141 spin_unlock_irq(&ctx->completion_lock);
6143 /* drop submission reference */
6147 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6149 struct io_kiocb *nxt;
6151 if (!(req->flags & REQ_F_LINK_HEAD))
6153 if (req->flags & REQ_F_LINK_TIMEOUT)
6156 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6158 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6161 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6162 req->flags |= REQ_F_LINK_TIMEOUT;
6166 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6168 struct io_kiocb *linked_timeout;
6169 const struct cred *old_creds = NULL;
6173 linked_timeout = io_prep_linked_timeout(req);
6175 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6176 (req->work.flags & IO_WQ_WORK_CREDS) &&
6177 req->work.identity->creds != current_cred()) {
6179 revert_creds(old_creds);
6180 if (old_creds == req->work.identity->creds)
6181 old_creds = NULL; /* restored original creds */
6183 old_creds = override_creds(req->work.identity->creds);
6186 ret = io_issue_sqe(req, true, cs);
6189 * We async punt it if the file wasn't marked NOWAIT, or if the file
6190 * doesn't support non-blocking read/write attempts
6192 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6193 if (!io_arm_poll_handler(req)) {
6195 * Queued up for async execution, worker will release
6196 * submit reference when the iocb is actually submitted.
6198 io_queue_async_work(req);
6202 io_queue_linked_timeout(linked_timeout);
6203 } else if (likely(!ret)) {
6204 /* drop submission reference */
6205 req = io_put_req_find_next(req);
6207 io_queue_linked_timeout(linked_timeout);
6210 if (!(req->flags & REQ_F_FORCE_ASYNC))
6212 io_queue_async_work(req);
6215 /* un-prep timeout, so it'll be killed as any other linked */
6216 req->flags &= ~REQ_F_LINK_TIMEOUT;
6217 req_set_fail_links(req);
6219 io_req_complete(req, ret);
6223 revert_creds(old_creds);
6226 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6227 struct io_comp_state *cs)
6231 ret = io_req_defer(req, sqe);
6233 if (ret != -EIOCBQUEUED) {
6235 req_set_fail_links(req);
6237 io_req_complete(req, ret);
6239 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6240 if (!req->async_data) {
6241 ret = io_req_defer_prep(req, sqe);
6245 io_queue_async_work(req);
6248 ret = io_req_prep(req, sqe);
6252 __io_queue_sqe(req, cs);
6256 static inline void io_queue_link_head(struct io_kiocb *req,
6257 struct io_comp_state *cs)
6259 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6261 io_req_complete(req, -ECANCELED);
6263 io_queue_sqe(req, NULL, cs);
6266 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6267 struct io_kiocb **link, struct io_comp_state *cs)
6269 struct io_ring_ctx *ctx = req->ctx;
6273 * If we already have a head request, queue this one for async
6274 * submittal once the head completes. If we don't have a head but
6275 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6276 * submitted sync once the chain is complete. If none of those
6277 * conditions are true (normal request), then just queue it.
6280 struct io_kiocb *head = *link;
6283 * Taking sequential execution of a link, draining both sides
6284 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6285 * requests in the link. So, it drains the head and the
6286 * next after the link request. The last one is done via
6287 * drain_next flag to persist the effect across calls.
6289 if (req->flags & REQ_F_IO_DRAIN) {
6290 head->flags |= REQ_F_IO_DRAIN;
6291 ctx->drain_next = 1;
6293 ret = io_req_defer_prep(req, sqe);
6294 if (unlikely(ret)) {
6295 /* fail even hard links since we don't submit */
6296 head->flags |= REQ_F_FAIL_LINK;
6299 trace_io_uring_link(ctx, req, head);
6300 list_add_tail(&req->link_list, &head->link_list);
6302 /* last request of a link, enqueue the link */
6303 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6304 io_queue_link_head(head, cs);
6308 if (unlikely(ctx->drain_next)) {
6309 req->flags |= REQ_F_IO_DRAIN;
6310 ctx->drain_next = 0;
6312 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6313 req->flags |= REQ_F_LINK_HEAD;
6314 INIT_LIST_HEAD(&req->link_list);
6316 ret = io_req_defer_prep(req, sqe);
6318 req->flags |= REQ_F_FAIL_LINK;
6321 io_queue_sqe(req, sqe, cs);
6329 * Batched submission is done, ensure local IO is flushed out.
6331 static void io_submit_state_end(struct io_submit_state *state)
6333 if (!list_empty(&state->comp.list))
6334 io_submit_flush_completions(&state->comp);
6335 blk_finish_plug(&state->plug);
6336 io_state_file_put(state);
6337 if (state->free_reqs)
6338 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6342 * Start submission side cache.
6344 static void io_submit_state_start(struct io_submit_state *state,
6345 struct io_ring_ctx *ctx, unsigned int max_ios)
6347 blk_start_plug(&state->plug);
6349 INIT_LIST_HEAD(&state->comp.list);
6350 state->comp.ctx = ctx;
6351 state->free_reqs = 0;
6353 state->ios_left = max_ios;
6356 static void io_commit_sqring(struct io_ring_ctx *ctx)
6358 struct io_rings *rings = ctx->rings;
6361 * Ensure any loads from the SQEs are done at this point,
6362 * since once we write the new head, the application could
6363 * write new data to them.
6365 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6369 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6370 * that is mapped by userspace. This means that care needs to be taken to
6371 * ensure that reads are stable, as we cannot rely on userspace always
6372 * being a good citizen. If members of the sqe are validated and then later
6373 * used, it's important that those reads are done through READ_ONCE() to
6374 * prevent a re-load down the line.
6376 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6378 u32 *sq_array = ctx->sq_array;
6382 * The cached sq head (or cq tail) serves two purposes:
6384 * 1) allows us to batch the cost of updating the user visible
6386 * 2) allows the kernel side to track the head on its own, even
6387 * though the application is the one updating it.
6389 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6390 if (likely(head < ctx->sq_entries))
6391 return &ctx->sq_sqes[head];
6393 /* drop invalid entries */
6394 ctx->cached_sq_dropped++;
6395 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6399 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6401 ctx->cached_sq_head++;
6405 * Check SQE restrictions (opcode and flags).
6407 * Returns 'true' if SQE is allowed, 'false' otherwise.
6409 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6410 struct io_kiocb *req,
6411 unsigned int sqe_flags)
6413 if (!ctx->restricted)
6416 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6419 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6420 ctx->restrictions.sqe_flags_required)
6423 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6424 ctx->restrictions.sqe_flags_required))
6430 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6431 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6432 IOSQE_BUFFER_SELECT)
6434 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6435 const struct io_uring_sqe *sqe,
6436 struct io_submit_state *state)
6438 unsigned int sqe_flags;
6441 req->opcode = READ_ONCE(sqe->opcode);
6442 req->user_data = READ_ONCE(sqe->user_data);
6443 req->async_data = NULL;
6447 /* one is dropped after submission, the other at completion */
6448 refcount_set(&req->refs, 2);
6449 req->task = current;
6452 if (unlikely(req->opcode >= IORING_OP_LAST))
6455 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6458 sqe_flags = READ_ONCE(sqe->flags);
6459 /* enforce forwards compatibility on users */
6460 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6463 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6466 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6467 !io_op_defs[req->opcode].buffer_select)
6470 id = READ_ONCE(sqe->personality);
6472 struct io_identity *iod;
6474 iod = idr_find(&ctx->personality_idr, id);
6477 refcount_inc(&iod->count);
6479 __io_req_init_async(req);
6480 get_cred(iod->creds);
6481 req->work.identity = iod;
6482 req->work.flags |= IO_WQ_WORK_CREDS;
6485 /* same numerical values with corresponding REQ_F_*, safe to copy */
6486 req->flags |= sqe_flags;
6488 if (!io_op_defs[req->opcode].needs_file)
6491 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6496 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6498 struct io_submit_state state;
6499 struct io_kiocb *link = NULL;
6500 int i, submitted = 0;
6502 /* if we have a backlog and couldn't flush it all, return BUSY */
6503 if (test_bit(0, &ctx->sq_check_overflow)) {
6504 if (!list_empty(&ctx->cq_overflow_list) &&
6505 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6509 /* make sure SQ entry isn't read before tail */
6510 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6512 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6515 percpu_counter_add(¤t->io_uring->inflight, nr);
6516 refcount_add(nr, ¤t->usage);
6518 io_submit_state_start(&state, ctx, nr);
6520 for (i = 0; i < nr; i++) {
6521 const struct io_uring_sqe *sqe;
6522 struct io_kiocb *req;
6525 sqe = io_get_sqe(ctx);
6526 if (unlikely(!sqe)) {
6527 io_consume_sqe(ctx);
6530 req = io_alloc_req(ctx, &state);
6531 if (unlikely(!req)) {
6533 submitted = -EAGAIN;
6536 io_consume_sqe(ctx);
6537 /* will complete beyond this point, count as submitted */
6540 err = io_init_req(ctx, req, sqe, &state);
6541 if (unlikely(err)) {
6544 io_req_complete(req, err);
6548 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6549 true, io_async_submit(ctx));
6550 err = io_submit_sqe(req, sqe, &link, &state.comp);
6555 if (unlikely(submitted != nr)) {
6556 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6557 struct io_uring_task *tctx = current->io_uring;
6558 int unused = nr - ref_used;
6560 percpu_ref_put_many(&ctx->refs, unused);
6561 percpu_counter_sub(&tctx->inflight, unused);
6562 put_task_struct_many(current, unused);
6565 io_queue_link_head(link, &state.comp);
6566 io_submit_state_end(&state);
6568 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6569 io_commit_sqring(ctx);
6574 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6576 /* Tell userspace we may need a wakeup call */
6577 spin_lock_irq(&ctx->completion_lock);
6578 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6579 spin_unlock_irq(&ctx->completion_lock);
6582 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6584 spin_lock_irq(&ctx->completion_lock);
6585 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6586 spin_unlock_irq(&ctx->completion_lock);
6589 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6590 int sync, void *key)
6592 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6595 ret = autoremove_wake_function(wqe, mode, sync, key);
6597 unsigned long flags;
6599 spin_lock_irqsave(&ctx->completion_lock, flags);
6600 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6601 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6612 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6613 unsigned long start_jiffies, bool cap_entries)
6615 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6616 struct io_sq_data *sqd = ctx->sq_data;
6617 unsigned int to_submit;
6621 if (!list_empty(&ctx->iopoll_list)) {
6622 unsigned nr_events = 0;
6624 mutex_lock(&ctx->uring_lock);
6625 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6626 io_do_iopoll(ctx, &nr_events, 0);
6627 mutex_unlock(&ctx->uring_lock);
6630 to_submit = io_sqring_entries(ctx);
6633 * If submit got -EBUSY, flag us as needing the application
6634 * to enter the kernel to reap and flush events.
6636 if (!to_submit || ret == -EBUSY || need_resched()) {
6638 * Drop cur_mm before scheduling, we can't hold it for
6639 * long periods (or over schedule()). Do this before
6640 * adding ourselves to the waitqueue, as the unuse/drop
6643 io_sq_thread_drop_mm();
6646 * We're polling. If we're within the defined idle
6647 * period, then let us spin without work before going
6648 * to sleep. The exception is if we got EBUSY doing
6649 * more IO, we should wait for the application to
6650 * reap events and wake us up.
6652 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6653 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6654 !percpu_ref_is_dying(&ctx->refs)))
6657 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6658 TASK_INTERRUPTIBLE);
6661 * While doing polled IO, before going to sleep, we need
6662 * to check if there are new reqs added to iopoll_list,
6663 * it is because reqs may have been punted to io worker
6664 * and will be added to iopoll_list later, hence check
6665 * the iopoll_list again.
6667 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6668 !list_empty_careful(&ctx->iopoll_list)) {
6669 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6673 to_submit = io_sqring_entries(ctx);
6674 if (!to_submit || ret == -EBUSY)
6678 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6679 io_ring_clear_wakeup_flag(ctx);
6681 /* if we're handling multiple rings, cap submit size for fairness */
6682 if (cap_entries && to_submit > 8)
6685 mutex_lock(&ctx->uring_lock);
6686 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6687 ret = io_submit_sqes(ctx, to_submit);
6688 mutex_unlock(&ctx->uring_lock);
6690 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6691 wake_up(&ctx->sqo_sq_wait);
6693 return SQT_DID_WORK;
6696 static void io_sqd_init_new(struct io_sq_data *sqd)
6698 struct io_ring_ctx *ctx;
6700 while (!list_empty(&sqd->ctx_new_list)) {
6701 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6702 init_wait(&ctx->sqo_wait_entry);
6703 ctx->sqo_wait_entry.func = io_sq_wake_function;
6704 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6705 complete(&ctx->sq_thread_comp);
6709 static int io_sq_thread(void *data)
6711 struct cgroup_subsys_state *cur_css = NULL;
6712 const struct cred *old_cred = NULL;
6713 struct io_sq_data *sqd = data;
6714 struct io_ring_ctx *ctx;
6715 unsigned long start_jiffies;
6717 start_jiffies = jiffies;
6718 while (!kthread_should_stop()) {
6719 enum sq_ret ret = 0;
6723 * Any changes to the sqd lists are synchronized through the
6724 * kthread parking. This synchronizes the thread vs users,
6725 * the users are synchronized on the sqd->ctx_lock.
6727 if (kthread_should_park())
6730 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6731 io_sqd_init_new(sqd);
6733 cap_entries = !list_is_singular(&sqd->ctx_list);
6735 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6736 if (current->cred != ctx->creds) {
6738 revert_creds(old_cred);
6739 old_cred = override_creds(ctx->creds);
6741 io_sq_thread_associate_blkcg(ctx, &cur_css);
6743 current->loginuid = ctx->loginuid;
6744 current->sessionid = ctx->sessionid;
6747 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6749 io_sq_thread_drop_mm();
6752 if (ret & SQT_SPIN) {
6755 } else if (ret == SQT_IDLE) {
6756 if (kthread_should_park())
6758 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6759 io_ring_set_wakeup_flag(ctx);
6761 start_jiffies = jiffies;
6762 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6763 io_ring_clear_wakeup_flag(ctx);
6770 io_sq_thread_unassociate_blkcg();
6772 revert_creds(old_cred);
6779 struct io_wait_queue {
6780 struct wait_queue_entry wq;
6781 struct io_ring_ctx *ctx;
6783 unsigned nr_timeouts;
6786 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6788 struct io_ring_ctx *ctx = iowq->ctx;
6791 * Wake up if we have enough events, or if a timeout occurred since we
6792 * started waiting. For timeouts, we always want to return to userspace,
6793 * regardless of event count.
6795 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6796 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6799 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6800 int wake_flags, void *key)
6802 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6805 /* use noflush == true, as we can't safely rely on locking context */
6806 if (!io_should_wake(iowq, true))
6809 return autoremove_wake_function(curr, mode, wake_flags, key);
6812 static int io_run_task_work_sig(void)
6814 if (io_run_task_work())
6816 if (!signal_pending(current))
6818 if (current->jobctl & JOBCTL_TASK_WORK) {
6819 spin_lock_irq(¤t->sighand->siglock);
6820 current->jobctl &= ~JOBCTL_TASK_WORK;
6821 recalc_sigpending();
6822 spin_unlock_irq(¤t->sighand->siglock);
6829 * Wait until events become available, if we don't already have some. The
6830 * application must reap them itself, as they reside on the shared cq ring.
6832 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6833 const sigset_t __user *sig, size_t sigsz)
6835 struct io_wait_queue iowq = {
6838 .func = io_wake_function,
6839 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6842 .to_wait = min_events,
6844 struct io_rings *rings = ctx->rings;
6848 if (io_cqring_events(ctx, false) >= min_events)
6850 if (!io_run_task_work())
6855 #ifdef CONFIG_COMPAT
6856 if (in_compat_syscall())
6857 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6861 ret = set_user_sigmask(sig, sigsz);
6867 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6868 trace_io_uring_cqring_wait(ctx, min_events);
6870 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6871 TASK_INTERRUPTIBLE);
6872 /* make sure we run task_work before checking for signals */
6873 ret = io_run_task_work_sig();
6878 if (io_should_wake(&iowq, false))
6882 finish_wait(&ctx->wait, &iowq.wq);
6884 restore_saved_sigmask_unless(ret == -EINTR);
6886 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6889 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6891 #if defined(CONFIG_UNIX)
6892 if (ctx->ring_sock) {
6893 struct sock *sock = ctx->ring_sock->sk;
6894 struct sk_buff *skb;
6896 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6902 for (i = 0; i < ctx->nr_user_files; i++) {
6905 file = io_file_from_index(ctx, i);
6912 static void io_file_ref_kill(struct percpu_ref *ref)
6914 struct fixed_file_data *data;
6916 data = container_of(ref, struct fixed_file_data, refs);
6917 complete(&data->done);
6920 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6922 struct fixed_file_data *data = ctx->file_data;
6923 struct fixed_file_ref_node *ref_node = NULL;
6924 unsigned nr_tables, i;
6929 spin_lock(&data->lock);
6930 if (!list_empty(&data->ref_list))
6931 ref_node = list_first_entry(&data->ref_list,
6932 struct fixed_file_ref_node, node);
6933 spin_unlock(&data->lock);
6935 percpu_ref_kill(&ref_node->refs);
6937 percpu_ref_kill(&data->refs);
6939 /* wait for all refs nodes to complete */
6940 flush_delayed_work(&ctx->file_put_work);
6941 wait_for_completion(&data->done);
6943 __io_sqe_files_unregister(ctx);
6944 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6945 for (i = 0; i < nr_tables; i++)
6946 kfree(data->table[i].files);
6948 percpu_ref_exit(&data->refs);
6950 ctx->file_data = NULL;
6951 ctx->nr_user_files = 0;
6955 static void io_put_sq_data(struct io_sq_data *sqd)
6957 if (refcount_dec_and_test(&sqd->refs)) {
6959 * The park is a bit of a work-around, without it we get
6960 * warning spews on shutdown with SQPOLL set and affinity
6961 * set to a single CPU.
6964 kthread_park(sqd->thread);
6965 kthread_stop(sqd->thread);
6972 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6974 struct io_ring_ctx *ctx_attach;
6975 struct io_sq_data *sqd;
6978 f = fdget(p->wq_fd);
6980 return ERR_PTR(-ENXIO);
6981 if (f.file->f_op != &io_uring_fops) {
6983 return ERR_PTR(-EINVAL);
6986 ctx_attach = f.file->private_data;
6987 sqd = ctx_attach->sq_data;
6990 return ERR_PTR(-EINVAL);
6993 refcount_inc(&sqd->refs);
6998 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7000 struct io_sq_data *sqd;
7002 if (p->flags & IORING_SETUP_ATTACH_WQ)
7003 return io_attach_sq_data(p);
7005 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7007 return ERR_PTR(-ENOMEM);
7009 refcount_set(&sqd->refs, 1);
7010 INIT_LIST_HEAD(&sqd->ctx_list);
7011 INIT_LIST_HEAD(&sqd->ctx_new_list);
7012 mutex_init(&sqd->ctx_lock);
7013 mutex_init(&sqd->lock);
7014 init_waitqueue_head(&sqd->wait);
7018 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7019 __releases(&sqd->lock)
7023 kthread_unpark(sqd->thread);
7024 mutex_unlock(&sqd->lock);
7027 static void io_sq_thread_park(struct io_sq_data *sqd)
7028 __acquires(&sqd->lock)
7032 mutex_lock(&sqd->lock);
7033 kthread_park(sqd->thread);
7036 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7038 struct io_sq_data *sqd = ctx->sq_data;
7043 * We may arrive here from the error branch in
7044 * io_sq_offload_create() where the kthread is created
7045 * without being waked up, thus wake it up now to make
7046 * sure the wait will complete.
7048 wake_up_process(sqd->thread);
7049 wait_for_completion(&ctx->sq_thread_comp);
7051 io_sq_thread_park(sqd);
7054 mutex_lock(&sqd->ctx_lock);
7055 list_del(&ctx->sqd_list);
7056 mutex_unlock(&sqd->ctx_lock);
7059 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7060 io_sq_thread_unpark(sqd);
7063 io_put_sq_data(sqd);
7064 ctx->sq_data = NULL;
7068 static void io_finish_async(struct io_ring_ctx *ctx)
7070 io_sq_thread_stop(ctx);
7073 io_wq_destroy(ctx->io_wq);
7078 #if defined(CONFIG_UNIX)
7080 * Ensure the UNIX gc is aware of our file set, so we are certain that
7081 * the io_uring can be safely unregistered on process exit, even if we have
7082 * loops in the file referencing.
7084 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7086 struct sock *sk = ctx->ring_sock->sk;
7087 struct scm_fp_list *fpl;
7088 struct sk_buff *skb;
7091 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7095 skb = alloc_skb(0, GFP_KERNEL);
7104 fpl->user = get_uid(ctx->user);
7105 for (i = 0; i < nr; i++) {
7106 struct file *file = io_file_from_index(ctx, i + offset);
7110 fpl->fp[nr_files] = get_file(file);
7111 unix_inflight(fpl->user, fpl->fp[nr_files]);
7116 fpl->max = SCM_MAX_FD;
7117 fpl->count = nr_files;
7118 UNIXCB(skb).fp = fpl;
7119 skb->destructor = unix_destruct_scm;
7120 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7121 skb_queue_head(&sk->sk_receive_queue, skb);
7123 for (i = 0; i < nr_files; i++)
7134 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7135 * causes regular reference counting to break down. We rely on the UNIX
7136 * garbage collection to take care of this problem for us.
7138 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7140 unsigned left, total;
7144 left = ctx->nr_user_files;
7146 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7148 ret = __io_sqe_files_scm(ctx, this_files, total);
7152 total += this_files;
7158 while (total < ctx->nr_user_files) {
7159 struct file *file = io_file_from_index(ctx, total);
7169 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7175 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7176 unsigned nr_tables, unsigned nr_files)
7180 for (i = 0; i < nr_tables; i++) {
7181 struct fixed_file_table *table = &file_data->table[i];
7182 unsigned this_files;
7184 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7185 table->files = kcalloc(this_files, sizeof(struct file *),
7189 nr_files -= this_files;
7195 for (i = 0; i < nr_tables; i++) {
7196 struct fixed_file_table *table = &file_data->table[i];
7197 kfree(table->files);
7202 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7204 #if defined(CONFIG_UNIX)
7205 struct sock *sock = ctx->ring_sock->sk;
7206 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7207 struct sk_buff *skb;
7210 __skb_queue_head_init(&list);
7213 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7214 * remove this entry and rearrange the file array.
7216 skb = skb_dequeue(head);
7218 struct scm_fp_list *fp;
7220 fp = UNIXCB(skb).fp;
7221 for (i = 0; i < fp->count; i++) {
7224 if (fp->fp[i] != file)
7227 unix_notinflight(fp->user, fp->fp[i]);
7228 left = fp->count - 1 - i;
7230 memmove(&fp->fp[i], &fp->fp[i + 1],
7231 left * sizeof(struct file *));
7238 __skb_queue_tail(&list, skb);
7248 __skb_queue_tail(&list, skb);
7250 skb = skb_dequeue(head);
7253 if (skb_peek(&list)) {
7254 spin_lock_irq(&head->lock);
7255 while ((skb = __skb_dequeue(&list)) != NULL)
7256 __skb_queue_tail(head, skb);
7257 spin_unlock_irq(&head->lock);
7264 struct io_file_put {
7265 struct list_head list;
7269 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7271 struct fixed_file_data *file_data = ref_node->file_data;
7272 struct io_ring_ctx *ctx = file_data->ctx;
7273 struct io_file_put *pfile, *tmp;
7275 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7276 list_del(&pfile->list);
7277 io_ring_file_put(ctx, pfile->file);
7281 spin_lock(&file_data->lock);
7282 list_del(&ref_node->node);
7283 spin_unlock(&file_data->lock);
7285 percpu_ref_exit(&ref_node->refs);
7287 percpu_ref_put(&file_data->refs);
7290 static void io_file_put_work(struct work_struct *work)
7292 struct io_ring_ctx *ctx;
7293 struct llist_node *node;
7295 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7296 node = llist_del_all(&ctx->file_put_llist);
7299 struct fixed_file_ref_node *ref_node;
7300 struct llist_node *next = node->next;
7302 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7303 __io_file_put_work(ref_node);
7308 static void io_file_data_ref_zero(struct percpu_ref *ref)
7310 struct fixed_file_ref_node *ref_node;
7311 struct io_ring_ctx *ctx;
7315 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7316 ctx = ref_node->file_data->ctx;
7318 if (percpu_ref_is_dying(&ctx->file_data->refs))
7321 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7323 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7325 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7328 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7329 struct io_ring_ctx *ctx)
7331 struct fixed_file_ref_node *ref_node;
7333 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7335 return ERR_PTR(-ENOMEM);
7337 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7340 return ERR_PTR(-ENOMEM);
7342 INIT_LIST_HEAD(&ref_node->node);
7343 INIT_LIST_HEAD(&ref_node->file_list);
7344 ref_node->file_data = ctx->file_data;
7348 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7350 percpu_ref_exit(&ref_node->refs);
7354 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7357 __s32 __user *fds = (__s32 __user *) arg;
7358 unsigned nr_tables, i;
7360 int fd, ret = -ENOMEM;
7361 struct fixed_file_ref_node *ref_node;
7362 struct fixed_file_data *file_data;
7368 if (nr_args > IORING_MAX_FIXED_FILES)
7371 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7374 file_data->ctx = ctx;
7375 init_completion(&file_data->done);
7376 INIT_LIST_HEAD(&file_data->ref_list);
7377 spin_lock_init(&file_data->lock);
7379 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7380 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7382 if (!file_data->table)
7385 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7386 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7389 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7391 ctx->file_data = file_data;
7393 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7394 struct fixed_file_table *table;
7397 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7401 /* allow sparse sets */
7411 * Don't allow io_uring instances to be registered. If UNIX
7412 * isn't enabled, then this causes a reference cycle and this
7413 * instance can never get freed. If UNIX is enabled we'll
7414 * handle it just fine, but there's still no point in allowing
7415 * a ring fd as it doesn't support regular read/write anyway.
7417 if (file->f_op == &io_uring_fops) {
7421 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7422 index = i & IORING_FILE_TABLE_MASK;
7423 table->files[index] = file;
7426 ret = io_sqe_files_scm(ctx);
7428 io_sqe_files_unregister(ctx);
7432 ref_node = alloc_fixed_file_ref_node(ctx);
7433 if (IS_ERR(ref_node)) {
7434 io_sqe_files_unregister(ctx);
7435 return PTR_ERR(ref_node);
7438 file_data->node = ref_node;
7439 spin_lock(&file_data->lock);
7440 list_add(&ref_node->node, &file_data->ref_list);
7441 spin_unlock(&file_data->lock);
7442 percpu_ref_get(&file_data->refs);
7445 for (i = 0; i < ctx->nr_user_files; i++) {
7446 file = io_file_from_index(ctx, i);
7450 for (i = 0; i < nr_tables; i++)
7451 kfree(file_data->table[i].files);
7452 ctx->nr_user_files = 0;
7454 percpu_ref_exit(&file_data->refs);
7456 kfree(file_data->table);
7458 ctx->file_data = NULL;
7462 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7465 #if defined(CONFIG_UNIX)
7466 struct sock *sock = ctx->ring_sock->sk;
7467 struct sk_buff_head *head = &sock->sk_receive_queue;
7468 struct sk_buff *skb;
7471 * See if we can merge this file into an existing skb SCM_RIGHTS
7472 * file set. If there's no room, fall back to allocating a new skb
7473 * and filling it in.
7475 spin_lock_irq(&head->lock);
7476 skb = skb_peek(head);
7478 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7480 if (fpl->count < SCM_MAX_FD) {
7481 __skb_unlink(skb, head);
7482 spin_unlock_irq(&head->lock);
7483 fpl->fp[fpl->count] = get_file(file);
7484 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7486 spin_lock_irq(&head->lock);
7487 __skb_queue_head(head, skb);
7492 spin_unlock_irq(&head->lock);
7499 return __io_sqe_files_scm(ctx, 1, index);
7505 static int io_queue_file_removal(struct fixed_file_data *data,
7508 struct io_file_put *pfile;
7509 struct fixed_file_ref_node *ref_node = data->node;
7511 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7516 list_add(&pfile->list, &ref_node->file_list);
7521 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7522 struct io_uring_files_update *up,
7525 struct fixed_file_data *data = ctx->file_data;
7526 struct fixed_file_ref_node *ref_node;
7531 bool needs_switch = false;
7533 if (check_add_overflow(up->offset, nr_args, &done))
7535 if (done > ctx->nr_user_files)
7538 ref_node = alloc_fixed_file_ref_node(ctx);
7539 if (IS_ERR(ref_node))
7540 return PTR_ERR(ref_node);
7543 fds = u64_to_user_ptr(up->fds);
7545 struct fixed_file_table *table;
7549 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7553 i = array_index_nospec(up->offset, ctx->nr_user_files);
7554 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7555 index = i & IORING_FILE_TABLE_MASK;
7556 if (table->files[index]) {
7557 file = table->files[index];
7558 err = io_queue_file_removal(data, file);
7561 table->files[index] = NULL;
7562 needs_switch = true;
7571 * Don't allow io_uring instances to be registered. If
7572 * UNIX isn't enabled, then this causes a reference
7573 * cycle and this instance can never get freed. If UNIX
7574 * is enabled we'll handle it just fine, but there's
7575 * still no point in allowing a ring fd as it doesn't
7576 * support regular read/write anyway.
7578 if (file->f_op == &io_uring_fops) {
7583 table->files[index] = file;
7584 err = io_sqe_file_register(ctx, file, i);
7586 table->files[index] = NULL;
7597 percpu_ref_kill(&data->node->refs);
7598 spin_lock(&data->lock);
7599 list_add(&ref_node->node, &data->ref_list);
7600 data->node = ref_node;
7601 spin_unlock(&data->lock);
7602 percpu_ref_get(&ctx->file_data->refs);
7604 destroy_fixed_file_ref_node(ref_node);
7606 return done ? done : err;
7609 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7612 struct io_uring_files_update up;
7614 if (!ctx->file_data)
7618 if (copy_from_user(&up, arg, sizeof(up)))
7623 return __io_sqe_files_update(ctx, &up, nr_args);
7626 static void io_free_work(struct io_wq_work *work)
7628 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7630 /* Consider that io_steal_work() relies on this ref */
7634 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7635 struct io_uring_params *p)
7637 struct io_wq_data data;
7639 struct io_ring_ctx *ctx_attach;
7640 unsigned int concurrency;
7643 data.user = ctx->user;
7644 data.free_work = io_free_work;
7645 data.do_work = io_wq_submit_work;
7647 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7648 /* Do QD, or 4 * CPUS, whatever is smallest */
7649 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7651 ctx->io_wq = io_wq_create(concurrency, &data);
7652 if (IS_ERR(ctx->io_wq)) {
7653 ret = PTR_ERR(ctx->io_wq);
7659 f = fdget(p->wq_fd);
7663 if (f.file->f_op != &io_uring_fops) {
7668 ctx_attach = f.file->private_data;
7669 /* @io_wq is protected by holding the fd */
7670 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7675 ctx->io_wq = ctx_attach->io_wq;
7681 static int io_uring_alloc_task_context(struct task_struct *task)
7683 struct io_uring_task *tctx;
7686 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7687 if (unlikely(!tctx))
7690 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7691 if (unlikely(ret)) {
7697 init_waitqueue_head(&tctx->wait);
7699 atomic_set(&tctx->in_idle, 0);
7700 tctx->sqpoll = false;
7701 io_init_identity(&tctx->__identity);
7702 tctx->identity = &tctx->__identity;
7703 task->io_uring = tctx;
7707 void __io_uring_free(struct task_struct *tsk)
7709 struct io_uring_task *tctx = tsk->io_uring;
7711 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7712 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7713 if (tctx->identity != &tctx->__identity)
7714 kfree(tctx->identity);
7715 percpu_counter_destroy(&tctx->inflight);
7717 tsk->io_uring = NULL;
7720 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7721 struct io_uring_params *p)
7725 if (ctx->flags & IORING_SETUP_SQPOLL) {
7726 struct io_sq_data *sqd;
7729 if (!capable(CAP_SYS_ADMIN))
7732 sqd = io_get_sq_data(p);
7739 io_sq_thread_park(sqd);
7740 mutex_lock(&sqd->ctx_lock);
7741 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7742 mutex_unlock(&sqd->ctx_lock);
7743 io_sq_thread_unpark(sqd);
7745 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7746 if (!ctx->sq_thread_idle)
7747 ctx->sq_thread_idle = HZ;
7752 if (p->flags & IORING_SETUP_SQ_AFF) {
7753 int cpu = p->sq_thread_cpu;
7756 if (cpu >= nr_cpu_ids)
7758 if (!cpu_online(cpu))
7761 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7762 cpu, "io_uring-sq");
7764 sqd->thread = kthread_create(io_sq_thread, sqd,
7767 if (IS_ERR(sqd->thread)) {
7768 ret = PTR_ERR(sqd->thread);
7772 ret = io_uring_alloc_task_context(sqd->thread);
7775 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7776 /* Can't have SQ_AFF without SQPOLL */
7782 ret = io_init_wq_offload(ctx, p);
7788 io_finish_async(ctx);
7792 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7794 struct io_sq_data *sqd = ctx->sq_data;
7796 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7797 wake_up_process(sqd->thread);
7800 static inline void __io_unaccount_mem(struct user_struct *user,
7801 unsigned long nr_pages)
7803 atomic_long_sub(nr_pages, &user->locked_vm);
7806 static inline int __io_account_mem(struct user_struct *user,
7807 unsigned long nr_pages)
7809 unsigned long page_limit, cur_pages, new_pages;
7811 /* Don't allow more pages than we can safely lock */
7812 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7815 cur_pages = atomic_long_read(&user->locked_vm);
7816 new_pages = cur_pages + nr_pages;
7817 if (new_pages > page_limit)
7819 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7820 new_pages) != cur_pages);
7825 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7826 enum io_mem_account acct)
7829 __io_unaccount_mem(ctx->user, nr_pages);
7831 if (ctx->mm_account) {
7832 if (acct == ACCT_LOCKED)
7833 ctx->mm_account->locked_vm -= nr_pages;
7834 else if (acct == ACCT_PINNED)
7835 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7839 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7840 enum io_mem_account acct)
7844 if (ctx->limit_mem) {
7845 ret = __io_account_mem(ctx->user, nr_pages);
7850 if (ctx->mm_account) {
7851 if (acct == ACCT_LOCKED)
7852 ctx->mm_account->locked_vm += nr_pages;
7853 else if (acct == ACCT_PINNED)
7854 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7860 static void io_mem_free(void *ptr)
7867 page = virt_to_head_page(ptr);
7868 if (put_page_testzero(page))
7869 free_compound_page(page);
7872 static void *io_mem_alloc(size_t size)
7874 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7877 return (void *) __get_free_pages(gfp_flags, get_order(size));
7880 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7883 struct io_rings *rings;
7884 size_t off, sq_array_size;
7886 off = struct_size(rings, cqes, cq_entries);
7887 if (off == SIZE_MAX)
7891 off = ALIGN(off, SMP_CACHE_BYTES);
7899 sq_array_size = array_size(sizeof(u32), sq_entries);
7900 if (sq_array_size == SIZE_MAX)
7903 if (check_add_overflow(off, sq_array_size, &off))
7909 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7913 pages = (size_t)1 << get_order(
7914 rings_size(sq_entries, cq_entries, NULL));
7915 pages += (size_t)1 << get_order(
7916 array_size(sizeof(struct io_uring_sqe), sq_entries));
7921 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7925 if (!ctx->user_bufs)
7928 for (i = 0; i < ctx->nr_user_bufs; i++) {
7929 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7931 for (j = 0; j < imu->nr_bvecs; j++)
7932 unpin_user_page(imu->bvec[j].bv_page);
7934 if (imu->acct_pages)
7935 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7940 kfree(ctx->user_bufs);
7941 ctx->user_bufs = NULL;
7942 ctx->nr_user_bufs = 0;
7946 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7947 void __user *arg, unsigned index)
7949 struct iovec __user *src;
7951 #ifdef CONFIG_COMPAT
7953 struct compat_iovec __user *ciovs;
7954 struct compat_iovec ciov;
7956 ciovs = (struct compat_iovec __user *) arg;
7957 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7960 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7961 dst->iov_len = ciov.iov_len;
7965 src = (struct iovec __user *) arg;
7966 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7972 * Not super efficient, but this is just a registration time. And we do cache
7973 * the last compound head, so generally we'll only do a full search if we don't
7976 * We check if the given compound head page has already been accounted, to
7977 * avoid double accounting it. This allows us to account the full size of the
7978 * page, not just the constituent pages of a huge page.
7980 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7981 int nr_pages, struct page *hpage)
7985 /* check current page array */
7986 for (i = 0; i < nr_pages; i++) {
7987 if (!PageCompound(pages[i]))
7989 if (compound_head(pages[i]) == hpage)
7993 /* check previously registered pages */
7994 for (i = 0; i < ctx->nr_user_bufs; i++) {
7995 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7997 for (j = 0; j < imu->nr_bvecs; j++) {
7998 if (!PageCompound(imu->bvec[j].bv_page))
8000 if (compound_head(imu->bvec[j].bv_page) == hpage)
8008 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8009 int nr_pages, struct io_mapped_ubuf *imu,
8010 struct page **last_hpage)
8014 for (i = 0; i < nr_pages; i++) {
8015 if (!PageCompound(pages[i])) {
8020 hpage = compound_head(pages[i]);
8021 if (hpage == *last_hpage)
8023 *last_hpage = hpage;
8024 if (headpage_already_acct(ctx, pages, i, hpage))
8026 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8030 if (!imu->acct_pages)
8033 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8035 imu->acct_pages = 0;
8039 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8042 struct vm_area_struct **vmas = NULL;
8043 struct page **pages = NULL;
8044 struct page *last_hpage = NULL;
8045 int i, j, got_pages = 0;
8050 if (!nr_args || nr_args > UIO_MAXIOV)
8053 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8055 if (!ctx->user_bufs)
8058 for (i = 0; i < nr_args; i++) {
8059 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8060 unsigned long off, start, end, ubuf;
8065 ret = io_copy_iov(ctx, &iov, arg, i);
8070 * Don't impose further limits on the size and buffer
8071 * constraints here, we'll -EINVAL later when IO is
8072 * submitted if they are wrong.
8075 if (!iov.iov_base || !iov.iov_len)
8078 /* arbitrary limit, but we need something */
8079 if (iov.iov_len > SZ_1G)
8082 ubuf = (unsigned long) iov.iov_base;
8083 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8084 start = ubuf >> PAGE_SHIFT;
8085 nr_pages = end - start;
8088 if (!pages || nr_pages > got_pages) {
8091 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8093 vmas = kvmalloc_array(nr_pages,
8094 sizeof(struct vm_area_struct *),
8096 if (!pages || !vmas) {
8100 got_pages = nr_pages;
8103 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8110 mmap_read_lock(current->mm);
8111 pret = pin_user_pages(ubuf, nr_pages,
8112 FOLL_WRITE | FOLL_LONGTERM,
8114 if (pret == nr_pages) {
8115 /* don't support file backed memory */
8116 for (j = 0; j < nr_pages; j++) {
8117 struct vm_area_struct *vma = vmas[j];
8120 !is_file_hugepages(vma->vm_file)) {
8126 ret = pret < 0 ? pret : -EFAULT;
8128 mmap_read_unlock(current->mm);
8131 * if we did partial map, or found file backed vmas,
8132 * release any pages we did get
8135 unpin_user_pages(pages, pret);
8140 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8142 unpin_user_pages(pages, pret);
8147 off = ubuf & ~PAGE_MASK;
8149 for (j = 0; j < nr_pages; j++) {
8152 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8153 imu->bvec[j].bv_page = pages[j];
8154 imu->bvec[j].bv_len = vec_len;
8155 imu->bvec[j].bv_offset = off;
8159 /* store original address for later verification */
8161 imu->len = iov.iov_len;
8162 imu->nr_bvecs = nr_pages;
8164 ctx->nr_user_bufs++;
8172 io_sqe_buffer_unregister(ctx);
8176 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8178 __s32 __user *fds = arg;
8184 if (copy_from_user(&fd, fds, sizeof(*fds)))
8187 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8188 if (IS_ERR(ctx->cq_ev_fd)) {
8189 int ret = PTR_ERR(ctx->cq_ev_fd);
8190 ctx->cq_ev_fd = NULL;
8197 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8199 if (ctx->cq_ev_fd) {
8200 eventfd_ctx_put(ctx->cq_ev_fd);
8201 ctx->cq_ev_fd = NULL;
8208 static int __io_destroy_buffers(int id, void *p, void *data)
8210 struct io_ring_ctx *ctx = data;
8211 struct io_buffer *buf = p;
8213 __io_remove_buffers(ctx, buf, id, -1U);
8217 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8219 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8220 idr_destroy(&ctx->io_buffer_idr);
8223 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8225 io_finish_async(ctx);
8226 io_sqe_buffer_unregister(ctx);
8228 if (ctx->sqo_task) {
8229 put_task_struct(ctx->sqo_task);
8230 ctx->sqo_task = NULL;
8231 mmdrop(ctx->mm_account);
8232 ctx->mm_account = NULL;
8235 #ifdef CONFIG_BLK_CGROUP
8236 if (ctx->sqo_blkcg_css)
8237 css_put(ctx->sqo_blkcg_css);
8240 io_sqe_files_unregister(ctx);
8241 io_eventfd_unregister(ctx);
8242 io_destroy_buffers(ctx);
8243 idr_destroy(&ctx->personality_idr);
8245 #if defined(CONFIG_UNIX)
8246 if (ctx->ring_sock) {
8247 ctx->ring_sock->file = NULL; /* so that iput() is called */
8248 sock_release(ctx->ring_sock);
8252 io_mem_free(ctx->rings);
8253 io_mem_free(ctx->sq_sqes);
8255 percpu_ref_exit(&ctx->refs);
8256 free_uid(ctx->user);
8257 put_cred(ctx->creds);
8258 kfree(ctx->cancel_hash);
8259 kmem_cache_free(req_cachep, ctx->fallback_req);
8263 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8265 struct io_ring_ctx *ctx = file->private_data;
8268 poll_wait(file, &ctx->cq_wait, wait);
8270 * synchronizes with barrier from wq_has_sleeper call in
8274 if (!io_sqring_full(ctx))
8275 mask |= EPOLLOUT | EPOLLWRNORM;
8276 if (io_cqring_events(ctx, false))
8277 mask |= EPOLLIN | EPOLLRDNORM;
8282 static int io_uring_fasync(int fd, struct file *file, int on)
8284 struct io_ring_ctx *ctx = file->private_data;
8286 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8289 static int io_remove_personalities(int id, void *p, void *data)
8291 struct io_ring_ctx *ctx = data;
8292 struct io_identity *iod;
8294 iod = idr_remove(&ctx->personality_idr, id);
8296 put_cred(iod->creds);
8297 if (refcount_dec_and_test(&iod->count))
8303 static void io_ring_exit_work(struct work_struct *work)
8305 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8309 * If we're doing polled IO and end up having requests being
8310 * submitted async (out-of-line), then completions can come in while
8311 * we're waiting for refs to drop. We need to reap these manually,
8312 * as nobody else will be looking for them.
8316 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8317 io_iopoll_try_reap_events(ctx);
8318 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8319 io_ring_ctx_free(ctx);
8322 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8324 mutex_lock(&ctx->uring_lock);
8325 percpu_ref_kill(&ctx->refs);
8326 mutex_unlock(&ctx->uring_lock);
8328 io_kill_timeouts(ctx, NULL);
8329 io_poll_remove_all(ctx, NULL);
8332 io_wq_cancel_all(ctx->io_wq);
8334 /* if we failed setting up the ctx, we might not have any rings */
8336 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8337 io_iopoll_try_reap_events(ctx);
8338 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8341 * Do this upfront, so we won't have a grace period where the ring
8342 * is closed but resources aren't reaped yet. This can cause
8343 * spurious failure in setting up a new ring.
8345 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8348 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8350 * Use system_unbound_wq to avoid spawning tons of event kworkers
8351 * if we're exiting a ton of rings at the same time. It just adds
8352 * noise and overhead, there's no discernable change in runtime
8353 * over using system_wq.
8355 queue_work(system_unbound_wq, &ctx->exit_work);
8358 static int io_uring_release(struct inode *inode, struct file *file)
8360 struct io_ring_ctx *ctx = file->private_data;
8362 file->private_data = NULL;
8363 io_ring_ctx_wait_and_kill(ctx);
8367 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8369 struct files_struct *files = data;
8371 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8372 work->identity->files == files);
8376 * Returns true if 'preq' is the link parent of 'req'
8378 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8380 struct io_kiocb *link;
8382 if (!(preq->flags & REQ_F_LINK_HEAD))
8385 list_for_each_entry(link, &preq->link_list, link_list) {
8393 static bool io_match_link_files(struct io_kiocb *req,
8394 struct files_struct *files)
8396 struct io_kiocb *link;
8398 if (io_match_files(req, files))
8400 if (req->flags & REQ_F_LINK_HEAD) {
8401 list_for_each_entry(link, &req->link_list, link_list) {
8402 if (io_match_files(link, files))
8410 * We're looking to cancel 'req' because it's holding on to our files, but
8411 * 'req' could be a link to another request. See if it is, and cancel that
8412 * parent request if so.
8414 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8416 struct hlist_node *tmp;
8417 struct io_kiocb *preq;
8421 spin_lock_irq(&ctx->completion_lock);
8422 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8423 struct hlist_head *list;
8425 list = &ctx->cancel_hash[i];
8426 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8427 found = io_match_link(preq, req);
8429 io_poll_remove_one(preq);
8434 spin_unlock_irq(&ctx->completion_lock);
8438 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8439 struct io_kiocb *req)
8441 struct io_kiocb *preq;
8444 spin_lock_irq(&ctx->completion_lock);
8445 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8446 found = io_match_link(preq, req);
8448 __io_timeout_cancel(preq);
8452 spin_unlock_irq(&ctx->completion_lock);
8456 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8458 return io_match_link(container_of(work, struct io_kiocb, work), data);
8461 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8463 enum io_wq_cancel cret;
8465 /* cancel this particular work, if it's running */
8466 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8467 if (cret != IO_WQ_CANCEL_NOTFOUND)
8470 /* find links that hold this pending, cancel those */
8471 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8472 if (cret != IO_WQ_CANCEL_NOTFOUND)
8475 /* if we have a poll link holding this pending, cancel that */
8476 if (io_poll_remove_link(ctx, req))
8479 /* final option, timeout link is holding this req pending */
8480 io_timeout_remove_link(ctx, req);
8483 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8484 struct files_struct *files)
8486 struct io_defer_entry *de = NULL;
8489 spin_lock_irq(&ctx->completion_lock);
8490 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8491 if (io_match_link_files(de->req, files)) {
8492 list_cut_position(&list, &ctx->defer_list, &de->list);
8496 spin_unlock_irq(&ctx->completion_lock);
8498 while (!list_empty(&list)) {
8499 de = list_first_entry(&list, struct io_defer_entry, list);
8500 list_del_init(&de->list);
8501 req_set_fail_links(de->req);
8502 io_put_req(de->req);
8503 io_req_complete(de->req, -ECANCELED);
8509 * Returns true if we found and killed one or more files pinning requests
8511 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8512 struct files_struct *files)
8514 if (list_empty_careful(&ctx->inflight_list))
8517 io_cancel_defer_files(ctx, files);
8518 /* cancel all at once, should be faster than doing it one by one*/
8519 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8521 while (!list_empty_careful(&ctx->inflight_list)) {
8522 struct io_kiocb *cancel_req = NULL, *req;
8525 spin_lock_irq(&ctx->inflight_lock);
8526 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8527 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8528 req->work.identity->files != files)
8530 /* req is being completed, ignore */
8531 if (!refcount_inc_not_zero(&req->refs))
8537 prepare_to_wait(&ctx->inflight_wait, &wait,
8538 TASK_UNINTERRUPTIBLE);
8539 spin_unlock_irq(&ctx->inflight_lock);
8541 /* We need to keep going until we don't find a matching req */
8544 /* cancel this request, or head link requests */
8545 io_attempt_cancel(ctx, cancel_req);
8546 io_put_req(cancel_req);
8547 /* cancellations _may_ trigger task work */
8550 finish_wait(&ctx->inflight_wait, &wait);
8556 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8558 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8559 struct task_struct *task = data;
8561 return io_task_match(req, task);
8564 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8565 struct task_struct *task,
8566 struct files_struct *files)
8570 ret = io_uring_cancel_files(ctx, files);
8572 enum io_wq_cancel cret;
8574 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8575 if (cret != IO_WQ_CANCEL_NOTFOUND)
8578 /* SQPOLL thread does its own polling */
8579 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8580 while (!list_empty_careful(&ctx->iopoll_list)) {
8581 io_iopoll_try_reap_events(ctx);
8586 ret |= io_poll_remove_all(ctx, task);
8587 ret |= io_kill_timeouts(ctx, task);
8594 * We need to iteratively cancel requests, in case a request has dependent
8595 * hard links. These persist even for failure of cancelations, hence keep
8596 * looping until none are found.
8598 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8599 struct files_struct *files)
8601 struct task_struct *task = current;
8603 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8604 task = ctx->sq_data->thread;
8605 atomic_inc(&task->io_uring->in_idle);
8606 io_sq_thread_park(ctx->sq_data);
8609 io_cqring_overflow_flush(ctx, true, task, files);
8611 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8616 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8617 atomic_dec(&task->io_uring->in_idle);
8619 * If the files that are going away are the ones in the thread
8620 * identity, clear them out.
8622 if (task->io_uring->identity->files == files)
8623 task->io_uring->identity->files = NULL;
8624 io_sq_thread_unpark(ctx->sq_data);
8629 * Note that this task has used io_uring. We use it for cancelation purposes.
8631 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8633 struct io_uring_task *tctx = current->io_uring;
8635 if (unlikely(!tctx)) {
8638 ret = io_uring_alloc_task_context(current);
8641 tctx = current->io_uring;
8643 if (tctx->last != file) {
8644 void *old = xa_load(&tctx->xa, (unsigned long)file);
8648 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8654 * This is race safe in that the task itself is doing this, hence it
8655 * cannot be going through the exit/cancel paths at the same time.
8656 * This cannot be modified while exit/cancel is running.
8658 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8659 tctx->sqpoll = true;
8665 * Remove this io_uring_file -> task mapping.
8667 static void io_uring_del_task_file(struct file *file)
8669 struct io_uring_task *tctx = current->io_uring;
8671 if (tctx->last == file)
8673 file = xa_erase(&tctx->xa, (unsigned long)file);
8679 * Drop task note for this file if we're the only ones that hold it after
8682 static void io_uring_attempt_task_drop(struct file *file)
8684 if (!current->io_uring)
8687 * fput() is pending, will be 2 if the only other ref is our potential
8688 * task file note. If the task is exiting, drop regardless of count.
8690 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8691 atomic_long_read(&file->f_count) == 2)
8692 io_uring_del_task_file(file);
8695 void __io_uring_files_cancel(struct files_struct *files)
8697 struct io_uring_task *tctx = current->io_uring;
8699 unsigned long index;
8701 /* make sure overflow events are dropped */
8702 atomic_inc(&tctx->in_idle);
8704 xa_for_each(&tctx->xa, index, file) {
8705 struct io_ring_ctx *ctx = file->private_data;
8707 io_uring_cancel_task_requests(ctx, files);
8709 io_uring_del_task_file(file);
8712 atomic_dec(&tctx->in_idle);
8715 static s64 tctx_inflight(struct io_uring_task *tctx)
8717 unsigned long index;
8721 inflight = percpu_counter_sum(&tctx->inflight);
8726 * If we have SQPOLL rings, then we need to iterate and find them, and
8727 * add the pending count for those.
8729 xa_for_each(&tctx->xa, index, file) {
8730 struct io_ring_ctx *ctx = file->private_data;
8732 if (ctx->flags & IORING_SETUP_SQPOLL) {
8733 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8735 inflight += percpu_counter_sum(&__tctx->inflight);
8743 * Find any io_uring fd that this task has registered or done IO on, and cancel
8746 void __io_uring_task_cancel(void)
8748 struct io_uring_task *tctx = current->io_uring;
8752 /* make sure overflow events are dropped */
8753 atomic_inc(&tctx->in_idle);
8756 /* read completions before cancelations */
8757 inflight = tctx_inflight(tctx);
8760 __io_uring_files_cancel(NULL);
8762 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8765 * If we've seen completions, retry. This avoids a race where
8766 * a completion comes in before we did prepare_to_wait().
8768 if (inflight != tctx_inflight(tctx))
8773 finish_wait(&tctx->wait, &wait);
8774 atomic_dec(&tctx->in_idle);
8777 static int io_uring_flush(struct file *file, void *data)
8779 io_uring_attempt_task_drop(file);
8783 static void *io_uring_validate_mmap_request(struct file *file,
8784 loff_t pgoff, size_t sz)
8786 struct io_ring_ctx *ctx = file->private_data;
8787 loff_t offset = pgoff << PAGE_SHIFT;
8792 case IORING_OFF_SQ_RING:
8793 case IORING_OFF_CQ_RING:
8796 case IORING_OFF_SQES:
8800 return ERR_PTR(-EINVAL);
8803 page = virt_to_head_page(ptr);
8804 if (sz > page_size(page))
8805 return ERR_PTR(-EINVAL);
8812 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8814 size_t sz = vma->vm_end - vma->vm_start;
8818 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8820 return PTR_ERR(ptr);
8822 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8823 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8826 #else /* !CONFIG_MMU */
8828 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8830 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8833 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8835 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8838 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8839 unsigned long addr, unsigned long len,
8840 unsigned long pgoff, unsigned long flags)
8844 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8846 return PTR_ERR(ptr);
8848 return (unsigned long) ptr;
8851 #endif /* !CONFIG_MMU */
8853 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8858 if (!io_sqring_full(ctx))
8861 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8863 if (!io_sqring_full(ctx))
8867 } while (!signal_pending(current));
8869 finish_wait(&ctx->sqo_sq_wait, &wait);
8872 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8873 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8876 struct io_ring_ctx *ctx;
8883 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8884 IORING_ENTER_SQ_WAIT))
8892 if (f.file->f_op != &io_uring_fops)
8896 ctx = f.file->private_data;
8897 if (!percpu_ref_tryget(&ctx->refs))
8901 if (ctx->flags & IORING_SETUP_R_DISABLED)
8905 * For SQ polling, the thread will do all submissions and completions.
8906 * Just return the requested submit count, and wake the thread if
8910 if (ctx->flags & IORING_SETUP_SQPOLL) {
8911 if (!list_empty_careful(&ctx->cq_overflow_list))
8912 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8913 if (flags & IORING_ENTER_SQ_WAKEUP)
8914 wake_up(&ctx->sq_data->wait);
8915 if (flags & IORING_ENTER_SQ_WAIT)
8916 io_sqpoll_wait_sq(ctx);
8917 submitted = to_submit;
8918 } else if (to_submit) {
8919 ret = io_uring_add_task_file(ctx, f.file);
8922 mutex_lock(&ctx->uring_lock);
8923 submitted = io_submit_sqes(ctx, to_submit);
8924 mutex_unlock(&ctx->uring_lock);
8926 if (submitted != to_submit)
8929 if (flags & IORING_ENTER_GETEVENTS) {
8930 min_complete = min(min_complete, ctx->cq_entries);
8933 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8934 * space applications don't need to do io completion events
8935 * polling again, they can rely on io_sq_thread to do polling
8936 * work, which can reduce cpu usage and uring_lock contention.
8938 if (ctx->flags & IORING_SETUP_IOPOLL &&
8939 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8940 ret = io_iopoll_check(ctx, min_complete);
8942 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8947 percpu_ref_put(&ctx->refs);
8950 return submitted ? submitted : ret;
8953 #ifdef CONFIG_PROC_FS
8954 static int io_uring_show_cred(int id, void *p, void *data)
8956 const struct cred *cred = p;
8957 struct seq_file *m = data;
8958 struct user_namespace *uns = seq_user_ns(m);
8959 struct group_info *gi;
8964 seq_printf(m, "%5d\n", id);
8965 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8966 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8967 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8968 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8969 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8970 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8971 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8972 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8973 seq_puts(m, "\n\tGroups:\t");
8974 gi = cred->group_info;
8975 for (g = 0; g < gi->ngroups; g++) {
8976 seq_put_decimal_ull(m, g ? " " : "",
8977 from_kgid_munged(uns, gi->gid[g]));
8979 seq_puts(m, "\n\tCapEff:\t");
8980 cap = cred->cap_effective;
8981 CAP_FOR_EACH_U32(__capi)
8982 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8987 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8989 struct io_sq_data *sq = NULL;
8994 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8995 * since fdinfo case grabs it in the opposite direction of normal use
8996 * cases. If we fail to get the lock, we just don't iterate any
8997 * structures that could be going away outside the io_uring mutex.
8999 has_lock = mutex_trylock(&ctx->uring_lock);
9001 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9004 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9005 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9006 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9007 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9008 struct fixed_file_table *table;
9011 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9012 f = table->files[i & IORING_FILE_TABLE_MASK];
9014 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9016 seq_printf(m, "%5u: <none>\n", i);
9018 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9019 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9020 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9022 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9023 (unsigned int) buf->len);
9025 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9026 seq_printf(m, "Personalities:\n");
9027 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9029 seq_printf(m, "PollList:\n");
9030 spin_lock_irq(&ctx->completion_lock);
9031 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9032 struct hlist_head *list = &ctx->cancel_hash[i];
9033 struct io_kiocb *req;
9035 hlist_for_each_entry(req, list, hash_node)
9036 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9037 req->task->task_works != NULL);
9039 spin_unlock_irq(&ctx->completion_lock);
9041 mutex_unlock(&ctx->uring_lock);
9044 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9046 struct io_ring_ctx *ctx = f->private_data;
9048 if (percpu_ref_tryget(&ctx->refs)) {
9049 __io_uring_show_fdinfo(ctx, m);
9050 percpu_ref_put(&ctx->refs);
9055 static const struct file_operations io_uring_fops = {
9056 .release = io_uring_release,
9057 .flush = io_uring_flush,
9058 .mmap = io_uring_mmap,
9060 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9061 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9063 .poll = io_uring_poll,
9064 .fasync = io_uring_fasync,
9065 #ifdef CONFIG_PROC_FS
9066 .show_fdinfo = io_uring_show_fdinfo,
9070 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9071 struct io_uring_params *p)
9073 struct io_rings *rings;
9074 size_t size, sq_array_offset;
9076 /* make sure these are sane, as we already accounted them */
9077 ctx->sq_entries = p->sq_entries;
9078 ctx->cq_entries = p->cq_entries;
9080 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9081 if (size == SIZE_MAX)
9084 rings = io_mem_alloc(size);
9089 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9090 rings->sq_ring_mask = p->sq_entries - 1;
9091 rings->cq_ring_mask = p->cq_entries - 1;
9092 rings->sq_ring_entries = p->sq_entries;
9093 rings->cq_ring_entries = p->cq_entries;
9094 ctx->sq_mask = rings->sq_ring_mask;
9095 ctx->cq_mask = rings->cq_ring_mask;
9097 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9098 if (size == SIZE_MAX) {
9099 io_mem_free(ctx->rings);
9104 ctx->sq_sqes = io_mem_alloc(size);
9105 if (!ctx->sq_sqes) {
9106 io_mem_free(ctx->rings);
9115 * Allocate an anonymous fd, this is what constitutes the application
9116 * visible backing of an io_uring instance. The application mmaps this
9117 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9118 * we have to tie this fd to a socket for file garbage collection purposes.
9120 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9125 #if defined(CONFIG_UNIX)
9126 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9132 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9136 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9137 O_RDWR | O_CLOEXEC);
9141 ret = PTR_ERR(file);
9145 #if defined(CONFIG_UNIX)
9146 ctx->ring_sock->file = file;
9148 if (unlikely(io_uring_add_task_file(ctx, file))) {
9149 file = ERR_PTR(-ENOMEM);
9152 fd_install(ret, file);
9155 #if defined(CONFIG_UNIX)
9156 sock_release(ctx->ring_sock);
9157 ctx->ring_sock = NULL;
9162 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9163 struct io_uring_params __user *params)
9165 struct user_struct *user = NULL;
9166 struct io_ring_ctx *ctx;
9172 if (entries > IORING_MAX_ENTRIES) {
9173 if (!(p->flags & IORING_SETUP_CLAMP))
9175 entries = IORING_MAX_ENTRIES;
9179 * Use twice as many entries for the CQ ring. It's possible for the
9180 * application to drive a higher depth than the size of the SQ ring,
9181 * since the sqes are only used at submission time. This allows for
9182 * some flexibility in overcommitting a bit. If the application has
9183 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9184 * of CQ ring entries manually.
9186 p->sq_entries = roundup_pow_of_two(entries);
9187 if (p->flags & IORING_SETUP_CQSIZE) {
9189 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9190 * to a power-of-two, if it isn't already. We do NOT impose
9191 * any cq vs sq ring sizing.
9193 if (p->cq_entries < p->sq_entries)
9195 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9196 if (!(p->flags & IORING_SETUP_CLAMP))
9198 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9200 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9202 p->cq_entries = 2 * p->sq_entries;
9205 user = get_uid(current_user());
9206 limit_mem = !capable(CAP_IPC_LOCK);
9209 ret = __io_account_mem(user,
9210 ring_pages(p->sq_entries, p->cq_entries));
9217 ctx = io_ring_ctx_alloc(p);
9220 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9225 ctx->compat = in_compat_syscall();
9227 ctx->creds = get_current_cred();
9229 ctx->loginuid = current->loginuid;
9230 ctx->sessionid = current->sessionid;
9232 ctx->sqo_task = get_task_struct(current);
9235 * This is just grabbed for accounting purposes. When a process exits,
9236 * the mm is exited and dropped before the files, hence we need to hang
9237 * on to this mm purely for the purposes of being able to unaccount
9238 * memory (locked/pinned vm). It's not used for anything else.
9240 mmgrab(current->mm);
9241 ctx->mm_account = current->mm;
9243 #ifdef CONFIG_BLK_CGROUP
9245 * The sq thread will belong to the original cgroup it was inited in.
9246 * If the cgroup goes offline (e.g. disabling the io controller), then
9247 * issued bios will be associated with the closest cgroup later in the
9251 ctx->sqo_blkcg_css = blkcg_css();
9252 ret = css_tryget_online(ctx->sqo_blkcg_css);
9255 /* don't init against a dying cgroup, have the user try again */
9256 ctx->sqo_blkcg_css = NULL;
9263 * Account memory _before_ installing the file descriptor. Once
9264 * the descriptor is installed, it can get closed at any time. Also
9265 * do this before hitting the general error path, as ring freeing
9266 * will un-account as well.
9268 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9270 ctx->limit_mem = limit_mem;
9272 ret = io_allocate_scq_urings(ctx, p);
9276 ret = io_sq_offload_create(ctx, p);
9280 if (!(p->flags & IORING_SETUP_R_DISABLED))
9281 io_sq_offload_start(ctx);
9283 memset(&p->sq_off, 0, sizeof(p->sq_off));
9284 p->sq_off.head = offsetof(struct io_rings, sq.head);
9285 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9286 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9287 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9288 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9289 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9290 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9292 memset(&p->cq_off, 0, sizeof(p->cq_off));
9293 p->cq_off.head = offsetof(struct io_rings, cq.head);
9294 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9295 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9296 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9297 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9298 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9299 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9301 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9302 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9303 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9304 IORING_FEAT_POLL_32BITS;
9306 if (copy_to_user(params, p, sizeof(*p))) {
9312 * Install ring fd as the very last thing, so we don't risk someone
9313 * having closed it before we finish setup
9315 ret = io_uring_get_fd(ctx);
9319 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9322 io_ring_ctx_wait_and_kill(ctx);
9327 * Sets up an aio uring context, and returns the fd. Applications asks for a
9328 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9329 * params structure passed in.
9331 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9333 struct io_uring_params p;
9336 if (copy_from_user(&p, params, sizeof(p)))
9338 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9343 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9344 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9345 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9346 IORING_SETUP_R_DISABLED))
9349 return io_uring_create(entries, &p, params);
9352 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9353 struct io_uring_params __user *, params)
9355 return io_uring_setup(entries, params);
9358 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9360 struct io_uring_probe *p;
9364 size = struct_size(p, ops, nr_args);
9365 if (size == SIZE_MAX)
9367 p = kzalloc(size, GFP_KERNEL);
9372 if (copy_from_user(p, arg, size))
9375 if (memchr_inv(p, 0, size))
9378 p->last_op = IORING_OP_LAST - 1;
9379 if (nr_args > IORING_OP_LAST)
9380 nr_args = IORING_OP_LAST;
9382 for (i = 0; i < nr_args; i++) {
9384 if (!io_op_defs[i].not_supported)
9385 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9390 if (copy_to_user(arg, p, size))
9397 static int io_register_personality(struct io_ring_ctx *ctx)
9399 struct io_identity *id;
9402 id = kmalloc(sizeof(*id), GFP_KERNEL);
9406 io_init_identity(id);
9407 id->creds = get_current_cred();
9409 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9411 put_cred(id->creds);
9417 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9419 struct io_identity *iod;
9421 iod = idr_remove(&ctx->personality_idr, id);
9423 put_cred(iod->creds);
9424 if (refcount_dec_and_test(&iod->count))
9432 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9433 unsigned int nr_args)
9435 struct io_uring_restriction *res;
9439 /* Restrictions allowed only if rings started disabled */
9440 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9443 /* We allow only a single restrictions registration */
9444 if (ctx->restrictions.registered)
9447 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9450 size = array_size(nr_args, sizeof(*res));
9451 if (size == SIZE_MAX)
9454 res = memdup_user(arg, size);
9456 return PTR_ERR(res);
9460 for (i = 0; i < nr_args; i++) {
9461 switch (res[i].opcode) {
9462 case IORING_RESTRICTION_REGISTER_OP:
9463 if (res[i].register_op >= IORING_REGISTER_LAST) {
9468 __set_bit(res[i].register_op,
9469 ctx->restrictions.register_op);
9471 case IORING_RESTRICTION_SQE_OP:
9472 if (res[i].sqe_op >= IORING_OP_LAST) {
9477 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9479 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9480 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9482 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9483 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9492 /* Reset all restrictions if an error happened */
9494 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9496 ctx->restrictions.registered = true;
9502 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9504 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9507 if (ctx->restrictions.registered)
9508 ctx->restricted = 1;
9510 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9512 io_sq_offload_start(ctx);
9517 static bool io_register_op_must_quiesce(int op)
9520 case IORING_UNREGISTER_FILES:
9521 case IORING_REGISTER_FILES_UPDATE:
9522 case IORING_REGISTER_PROBE:
9523 case IORING_REGISTER_PERSONALITY:
9524 case IORING_UNREGISTER_PERSONALITY:
9531 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9532 void __user *arg, unsigned nr_args)
9533 __releases(ctx->uring_lock)
9534 __acquires(ctx->uring_lock)
9539 * We're inside the ring mutex, if the ref is already dying, then
9540 * someone else killed the ctx or is already going through
9541 * io_uring_register().
9543 if (percpu_ref_is_dying(&ctx->refs))
9546 if (io_register_op_must_quiesce(opcode)) {
9547 percpu_ref_kill(&ctx->refs);
9550 * Drop uring mutex before waiting for references to exit. If
9551 * another thread is currently inside io_uring_enter() it might
9552 * need to grab the uring_lock to make progress. If we hold it
9553 * here across the drain wait, then we can deadlock. It's safe
9554 * to drop the mutex here, since no new references will come in
9555 * after we've killed the percpu ref.
9557 mutex_unlock(&ctx->uring_lock);
9559 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9562 ret = io_run_task_work_sig();
9567 mutex_lock(&ctx->uring_lock);
9570 percpu_ref_resurrect(&ctx->refs);
9575 if (ctx->restricted) {
9576 if (opcode >= IORING_REGISTER_LAST) {
9581 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9588 case IORING_REGISTER_BUFFERS:
9589 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9591 case IORING_UNREGISTER_BUFFERS:
9595 ret = io_sqe_buffer_unregister(ctx);
9597 case IORING_REGISTER_FILES:
9598 ret = io_sqe_files_register(ctx, arg, nr_args);
9600 case IORING_UNREGISTER_FILES:
9604 ret = io_sqe_files_unregister(ctx);
9606 case IORING_REGISTER_FILES_UPDATE:
9607 ret = io_sqe_files_update(ctx, arg, nr_args);
9609 case IORING_REGISTER_EVENTFD:
9610 case IORING_REGISTER_EVENTFD_ASYNC:
9614 ret = io_eventfd_register(ctx, arg);
9617 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9618 ctx->eventfd_async = 1;
9620 ctx->eventfd_async = 0;
9622 case IORING_UNREGISTER_EVENTFD:
9626 ret = io_eventfd_unregister(ctx);
9628 case IORING_REGISTER_PROBE:
9630 if (!arg || nr_args > 256)
9632 ret = io_probe(ctx, arg, nr_args);
9634 case IORING_REGISTER_PERSONALITY:
9638 ret = io_register_personality(ctx);
9640 case IORING_UNREGISTER_PERSONALITY:
9644 ret = io_unregister_personality(ctx, nr_args);
9646 case IORING_REGISTER_ENABLE_RINGS:
9650 ret = io_register_enable_rings(ctx);
9652 case IORING_REGISTER_RESTRICTIONS:
9653 ret = io_register_restrictions(ctx, arg, nr_args);
9661 if (io_register_op_must_quiesce(opcode)) {
9662 /* bring the ctx back to life */
9663 percpu_ref_reinit(&ctx->refs);
9665 reinit_completion(&ctx->ref_comp);
9670 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9671 void __user *, arg, unsigned int, nr_args)
9673 struct io_ring_ctx *ctx;
9682 if (f.file->f_op != &io_uring_fops)
9685 ctx = f.file->private_data;
9687 mutex_lock(&ctx->uring_lock);
9688 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9689 mutex_unlock(&ctx->uring_lock);
9690 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9691 ctx->cq_ev_fd != NULL, ret);
9697 static int __init io_uring_init(void)
9699 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9700 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9701 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9704 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9705 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9706 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9707 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9708 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9709 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9710 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9711 BUILD_BUG_SQE_ELEM(8, __u64, off);
9712 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9713 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9714 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9715 BUILD_BUG_SQE_ELEM(24, __u32, len);
9716 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9717 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9718 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9719 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9720 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9721 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9722 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9723 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9724 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9725 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9726 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9727 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9728 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9729 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9730 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9731 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9732 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9733 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9734 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9736 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9737 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9738 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9741 __initcall(io_uring_init);