1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
210 struct fixed_file_data {
211 struct fixed_file_table *table;
212 struct io_ring_ctx *ctx;
214 struct fixed_file_ref_node *node;
215 struct percpu_ref refs;
216 struct completion done;
217 struct list_head ref_list;
222 struct list_head list;
228 struct io_restriction {
229 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
230 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
231 u8 sqe_flags_allowed;
232 u8 sqe_flags_required;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list;
242 struct list_head ctx_new_list;
243 struct mutex ctx_lock;
245 struct task_struct *thread;
246 struct wait_queue_head wait;
251 struct percpu_ref refs;
252 } ____cacheline_aligned_in_smp;
256 unsigned int compat: 1;
257 unsigned int limit_mem: 1;
258 unsigned int cq_overflow_flushed: 1;
259 unsigned int drain_next: 1;
260 unsigned int eventfd_async: 1;
261 unsigned int restricted: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head;
278 unsigned sq_thread_idle;
279 unsigned cached_sq_dropped;
280 unsigned cached_cq_overflow;
281 unsigned long sq_check_overflow;
283 struct list_head defer_list;
284 struct list_head timeout_list;
285 struct list_head cq_overflow_list;
287 wait_queue_head_t inflight_wait;
288 struct io_uring_sqe *sq_sqes;
289 } ____cacheline_aligned_in_smp;
291 struct io_rings *rings;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct *sqo_task;
302 /* Only used for accounting purposes */
303 struct mm_struct *mm_account;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state *sqo_blkcg_css;
309 struct io_sq_data *sq_data; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait;
312 struct wait_queue_entry sqo_wait_entry;
313 struct list_head sqd_list;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data *file_data;
321 unsigned nr_user_files;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs;
325 struct io_mapped_ubuf *user_bufs;
327 struct user_struct *user;
329 const struct cred *creds;
333 unsigned int sessionid;
336 struct completion ref_comp;
337 struct completion sq_thread_comp;
339 /* if all else fails... */
340 struct io_kiocb *fallback_req;
342 #if defined(CONFIG_UNIX)
343 struct socket *ring_sock;
346 struct idr io_buffer_idr;
348 struct idr personality_idr;
351 unsigned cached_cq_tail;
354 atomic_t cq_timeouts;
355 unsigned long cq_check_overflow;
356 struct wait_queue_head cq_wait;
357 struct fasync_struct *cq_fasync;
358 struct eventfd_ctx *cq_ev_fd;
359 } ____cacheline_aligned_in_smp;
362 struct mutex uring_lock;
363 wait_queue_head_t wait;
364 } ____cacheline_aligned_in_smp;
367 spinlock_t completion_lock;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list;
376 struct hlist_head *cancel_hash;
377 unsigned cancel_hash_bits;
378 bool poll_multi_file;
380 spinlock_t inflight_lock;
381 struct list_head inflight_list;
382 } ____cacheline_aligned_in_smp;
384 struct delayed_work file_put_work;
385 struct llist_head file_put_llist;
387 struct work_struct exit_work;
388 struct io_restriction restrictions;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb {
398 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
409 struct file *put_file;
413 struct io_timeout_data {
414 struct io_kiocb *req;
415 struct hrtimer timer;
416 struct timespec64 ts;
417 enum hrtimer_mode mode;
422 struct sockaddr __user *addr;
423 int __user *addr_len;
425 unsigned long nofile;
445 struct list_head list;
448 struct io_timeout_rem {
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user *addr;
469 struct user_msghdr __user *umsg;
475 struct io_buffer *kbuf;
481 struct filename *filename;
483 unsigned long nofile;
486 struct io_files_update {
512 struct epoll_event event;
516 struct file *file_out;
517 struct file *file_in;
524 struct io_provide_buf {
538 const char __user *filename;
539 struct statx __user *buffer;
542 struct io_completion {
544 struct list_head list;
548 struct io_async_connect {
549 struct sockaddr_storage address;
552 struct io_async_msghdr {
553 struct iovec fast_iov[UIO_FASTIOV];
555 struct sockaddr __user *uaddr;
557 struct sockaddr_storage addr;
561 struct iovec fast_iov[UIO_FASTIOV];
562 const struct iovec *free_iovec;
563 struct iov_iter iter;
565 struct wait_page_queue wpq;
569 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
570 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
571 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
572 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
573 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
574 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
581 REQ_F_LINK_TIMEOUT_BIT,
583 REQ_F_NEED_CLEANUP_BIT,
585 REQ_F_BUFFER_SELECTED_BIT,
586 REQ_F_NO_FILE_TABLE_BIT,
587 REQ_F_WORK_INITIALIZED_BIT,
588 REQ_F_LTIMEOUT_ACTIVE_BIT,
590 /* not a real bit, just to check we're not overflowing the space */
596 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
597 /* drain existing IO first */
598 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
600 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
601 /* doesn't sever on completion < 0 */
602 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
604 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
605 /* IOSQE_BUFFER_SELECT */
606 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
609 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
610 /* fail rest of links */
611 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
612 /* on inflight list */
613 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
614 /* read/write uses file position */
615 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
616 /* must not punt to workers */
617 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
618 /* has or had linked timeout */
619 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
621 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
623 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
624 /* already went through poll handler */
625 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
626 /* buffer already selected */
627 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
628 /* doesn't need file table for this request */
629 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
630 /* io_wq_work is initialized */
631 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
632 /* linked timeout is active, i.e. prepared by link's head */
633 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
637 struct io_poll_iocb poll;
638 struct io_poll_iocb *double_poll;
642 * NOTE! Each of the iocb union members has the file pointer
643 * as the first entry in their struct definition. So you can
644 * access the file pointer through any of the sub-structs,
645 * or directly as just 'ki_filp' in this struct.
651 struct io_poll_iocb poll;
652 struct io_accept accept;
654 struct io_cancel cancel;
655 struct io_timeout timeout;
656 struct io_timeout_rem timeout_rem;
657 struct io_connect connect;
658 struct io_sr_msg sr_msg;
660 struct io_close close;
661 struct io_files_update files_update;
662 struct io_fadvise fadvise;
663 struct io_madvise madvise;
664 struct io_epoll epoll;
665 struct io_splice splice;
666 struct io_provide_buf pbuf;
667 struct io_statx statx;
668 /* use only after cleaning per-op data, see io_clean_op() */
669 struct io_completion compl;
672 /* opcode allocated if it needs to store data for async defer */
675 /* polled IO has completed */
681 struct io_ring_ctx *ctx;
684 struct task_struct *task;
687 struct list_head link_list;
690 * 1. used with ctx->iopoll_list with reads/writes
691 * 2. to track reqs with ->files (see io_op_def::file_table)
693 struct list_head inflight_entry;
695 struct percpu_ref *fixed_file_refs;
696 struct callback_head task_work;
697 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
698 struct hlist_node hash_node;
699 struct async_poll *apoll;
700 struct io_wq_work work;
703 struct io_defer_entry {
704 struct list_head list;
705 struct io_kiocb *req;
709 #define IO_IOPOLL_BATCH 8
711 struct io_comp_state {
713 struct list_head list;
714 struct io_ring_ctx *ctx;
717 struct io_submit_state {
718 struct blk_plug plug;
721 * io_kiocb alloc cache
723 void *reqs[IO_IOPOLL_BATCH];
724 unsigned int free_reqs;
727 * Batch completion logic
729 struct io_comp_state comp;
732 * File reference cache
736 unsigned int has_refs;
737 unsigned int ios_left;
741 /* needs req->file assigned */
742 unsigned needs_file : 1;
743 /* don't fail if file grab fails */
744 unsigned needs_file_no_error : 1;
745 /* hash wq insertion if file is a regular file */
746 unsigned hash_reg_file : 1;
747 /* unbound wq insertion if file is a non-regular file */
748 unsigned unbound_nonreg_file : 1;
749 /* opcode is not supported by this kernel */
750 unsigned not_supported : 1;
751 /* set if opcode supports polled "wait" */
753 unsigned pollout : 1;
754 /* op supports buffer selection */
755 unsigned buffer_select : 1;
756 /* must always have async data allocated */
757 unsigned needs_async_data : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
763 static const struct io_op_def io_op_defs[] = {
764 [IORING_OP_NOP] = {},
765 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
771 .async_size = sizeof(struct io_async_rw),
772 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
774 [IORING_OP_WRITEV] = {
777 .unbound_nonreg_file = 1,
779 .needs_async_data = 1,
780 .async_size = sizeof(struct io_async_rw),
781 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
784 [IORING_OP_FSYNC] = {
786 .work_flags = IO_WQ_WORK_BLKCG,
788 [IORING_OP_READ_FIXED] = {
790 .unbound_nonreg_file = 1,
792 .async_size = sizeof(struct io_async_rw),
793 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
795 [IORING_OP_WRITE_FIXED] = {
798 .unbound_nonreg_file = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
804 [IORING_OP_POLL_ADD] = {
806 .unbound_nonreg_file = 1,
808 [IORING_OP_POLL_REMOVE] = {},
809 [IORING_OP_SYNC_FILE_RANGE] = {
811 .work_flags = IO_WQ_WORK_BLKCG,
813 [IORING_OP_SENDMSG] = {
815 .unbound_nonreg_file = 1,
817 .needs_async_data = 1,
818 .async_size = sizeof(struct io_async_msghdr),
819 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
822 [IORING_OP_RECVMSG] = {
824 .unbound_nonreg_file = 1,
827 .needs_async_data = 1,
828 .async_size = sizeof(struct io_async_msghdr),
829 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
832 [IORING_OP_TIMEOUT] = {
833 .needs_async_data = 1,
834 .async_size = sizeof(struct io_timeout_data),
835 .work_flags = IO_WQ_WORK_MM,
837 [IORING_OP_TIMEOUT_REMOVE] = {},
838 [IORING_OP_ACCEPT] = {
840 .unbound_nonreg_file = 1,
842 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
844 [IORING_OP_ASYNC_CANCEL] = {},
845 [IORING_OP_LINK_TIMEOUT] = {
846 .needs_async_data = 1,
847 .async_size = sizeof(struct io_timeout_data),
848 .work_flags = IO_WQ_WORK_MM,
850 [IORING_OP_CONNECT] = {
852 .unbound_nonreg_file = 1,
854 .needs_async_data = 1,
855 .async_size = sizeof(struct io_async_connect),
856 .work_flags = IO_WQ_WORK_MM,
858 [IORING_OP_FALLOCATE] = {
860 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
862 [IORING_OP_OPENAT] = {
863 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
866 [IORING_OP_CLOSE] = {
868 .needs_file_no_error = 1,
869 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
871 [IORING_OP_FILES_UPDATE] = {
872 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
874 [IORING_OP_STATX] = {
875 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
876 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
880 .unbound_nonreg_file = 1,
883 .async_size = sizeof(struct io_async_rw),
884 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
886 [IORING_OP_WRITE] = {
888 .unbound_nonreg_file = 1,
890 .async_size = sizeof(struct io_async_rw),
891 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
894 [IORING_OP_FADVISE] = {
896 .work_flags = IO_WQ_WORK_BLKCG,
898 [IORING_OP_MADVISE] = {
899 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
903 .unbound_nonreg_file = 1,
905 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
909 .unbound_nonreg_file = 1,
912 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 [IORING_OP_OPENAT2] = {
915 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
918 [IORING_OP_EPOLL_CTL] = {
919 .unbound_nonreg_file = 1,
920 .work_flags = IO_WQ_WORK_FILES,
922 [IORING_OP_SPLICE] = {
925 .unbound_nonreg_file = 1,
926 .work_flags = IO_WQ_WORK_BLKCG,
928 [IORING_OP_PROVIDE_BUFFERS] = {},
929 [IORING_OP_REMOVE_BUFFERS] = {},
933 .unbound_nonreg_file = 1,
937 enum io_mem_account {
942 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
943 struct io_comp_state *cs);
944 static void io_cqring_fill_event(struct io_kiocb *req, long res);
945 static void io_put_req(struct io_kiocb *req);
946 static void io_put_req_deferred(struct io_kiocb *req, int nr);
947 static void io_double_put_req(struct io_kiocb *req);
948 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
949 static void __io_queue_linked_timeout(struct io_kiocb *req);
950 static void io_queue_linked_timeout(struct io_kiocb *req);
951 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
952 struct io_uring_files_update *ip,
954 static void __io_clean_op(struct io_kiocb *req);
955 static struct file *io_file_get(struct io_submit_state *state,
956 struct io_kiocb *req, int fd, bool fixed);
957 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
958 static void io_file_put_work(struct work_struct *work);
960 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
961 struct iovec **iovec, struct iov_iter *iter,
963 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
964 const struct iovec *fast_iov,
965 struct iov_iter *iter, bool force);
967 static struct kmem_cache *req_cachep;
969 static const struct file_operations io_uring_fops;
971 struct sock *io_uring_get_socket(struct file *file)
973 #if defined(CONFIG_UNIX)
974 if (file->f_op == &io_uring_fops) {
975 struct io_ring_ctx *ctx = file->private_data;
977 return ctx->ring_sock->sk;
982 EXPORT_SYMBOL(io_uring_get_socket);
984 static inline void io_clean_op(struct io_kiocb *req)
986 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
991 static void io_sq_thread_drop_mm(void)
993 struct mm_struct *mm = current->mm;
996 kthread_unuse_mm(mm);
1001 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1004 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1005 !ctx->sqo_task->mm ||
1006 !mmget_not_zero(ctx->sqo_task->mm)))
1008 kthread_use_mm(ctx->sqo_task->mm);
1014 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1015 struct io_kiocb *req)
1017 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1019 return __io_sq_thread_acquire_mm(ctx);
1022 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1023 struct cgroup_subsys_state **cur_css)
1026 #ifdef CONFIG_BLK_CGROUP
1027 /* puts the old one when swapping */
1028 if (*cur_css != ctx->sqo_blkcg_css) {
1029 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1030 *cur_css = ctx->sqo_blkcg_css;
1035 static void io_sq_thread_unassociate_blkcg(void)
1037 #ifdef CONFIG_BLK_CGROUP
1038 kthread_associate_blkcg(NULL);
1042 static inline void req_set_fail_links(struct io_kiocb *req)
1044 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1045 req->flags |= REQ_F_FAIL_LINK;
1049 * None of these are dereferenced, they are simply used to check if any of
1050 * them have changed. If we're under current and check they are still the
1051 * same, we're fine to grab references to them for actual out-of-line use.
1053 static void io_init_identity(struct io_identity *id)
1055 id->files = current->files;
1056 id->mm = current->mm;
1057 #ifdef CONFIG_BLK_CGROUP
1059 id->blkcg_css = blkcg_css();
1062 id->creds = current_cred();
1063 id->nsproxy = current->nsproxy;
1064 id->fs = current->fs;
1065 id->fsize = rlimit(RLIMIT_FSIZE);
1067 id->loginuid = current->loginuid;
1068 id->sessionid = current->sessionid;
1070 refcount_set(&id->count, 1);
1073 static inline void __io_req_init_async(struct io_kiocb *req)
1075 memset(&req->work, 0, sizeof(req->work));
1076 req->flags |= REQ_F_WORK_INITIALIZED;
1080 * Note: must call io_req_init_async() for the first time you
1081 * touch any members of io_wq_work.
1083 static inline void io_req_init_async(struct io_kiocb *req)
1085 struct io_uring_task *tctx = current->io_uring;
1087 if (req->flags & REQ_F_WORK_INITIALIZED)
1090 __io_req_init_async(req);
1092 /* Grab a ref if this isn't our static identity */
1093 req->work.identity = tctx->identity;
1094 if (tctx->identity != &tctx->__identity)
1095 refcount_inc(&req->work.identity->count);
1098 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1100 return ctx->flags & IORING_SETUP_SQPOLL;
1103 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1105 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1107 complete(&ctx->ref_comp);
1110 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1112 return !req->timeout.off;
1115 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1117 struct io_ring_ctx *ctx;
1120 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1124 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1125 if (!ctx->fallback_req)
1129 * Use 5 bits less than the max cq entries, that should give us around
1130 * 32 entries per hash list if totally full and uniformly spread.
1132 hash_bits = ilog2(p->cq_entries);
1136 ctx->cancel_hash_bits = hash_bits;
1137 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1139 if (!ctx->cancel_hash)
1141 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1143 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1144 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1147 ctx->flags = p->flags;
1148 init_waitqueue_head(&ctx->sqo_sq_wait);
1149 INIT_LIST_HEAD(&ctx->sqd_list);
1150 init_waitqueue_head(&ctx->cq_wait);
1151 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1152 init_completion(&ctx->ref_comp);
1153 init_completion(&ctx->sq_thread_comp);
1154 idr_init(&ctx->io_buffer_idr);
1155 idr_init(&ctx->personality_idr);
1156 mutex_init(&ctx->uring_lock);
1157 init_waitqueue_head(&ctx->wait);
1158 spin_lock_init(&ctx->completion_lock);
1159 INIT_LIST_HEAD(&ctx->iopoll_list);
1160 INIT_LIST_HEAD(&ctx->defer_list);
1161 INIT_LIST_HEAD(&ctx->timeout_list);
1162 init_waitqueue_head(&ctx->inflight_wait);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1166 init_llist_head(&ctx->file_put_llist);
1169 if (ctx->fallback_req)
1170 kmem_cache_free(req_cachep, ctx->fallback_req);
1171 kfree(ctx->cancel_hash);
1176 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1178 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1179 struct io_ring_ctx *ctx = req->ctx;
1181 return seq != ctx->cached_cq_tail
1182 + READ_ONCE(ctx->cached_cq_overflow);
1188 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1190 struct io_rings *rings = ctx->rings;
1192 /* order cqe stores with ring update */
1193 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1195 if (wq_has_sleeper(&ctx->cq_wait)) {
1196 wake_up_interruptible(&ctx->cq_wait);
1197 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1201 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1203 if (req->work.identity == &tctx->__identity)
1205 if (refcount_dec_and_test(&req->work.identity->count))
1206 kfree(req->work.identity);
1209 static void io_req_clean_work(struct io_kiocb *req)
1211 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1214 req->flags &= ~REQ_F_WORK_INITIALIZED;
1216 if (req->work.flags & IO_WQ_WORK_MM) {
1217 mmdrop(req->work.identity->mm);
1218 req->work.flags &= ~IO_WQ_WORK_MM;
1220 #ifdef CONFIG_BLK_CGROUP
1221 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1222 css_put(req->work.identity->blkcg_css);
1223 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1226 if (req->work.flags & IO_WQ_WORK_CREDS) {
1227 put_cred(req->work.identity->creds);
1228 req->work.flags &= ~IO_WQ_WORK_CREDS;
1230 if (req->work.flags & IO_WQ_WORK_FS) {
1231 struct fs_struct *fs = req->work.identity->fs;
1233 spin_lock(&req->work.identity->fs->lock);
1236 spin_unlock(&req->work.identity->fs->lock);
1239 req->work.flags &= ~IO_WQ_WORK_FS;
1242 io_put_identity(req->task->io_uring, req);
1246 * Create a private copy of io_identity, since some fields don't match
1247 * the current context.
1249 static bool io_identity_cow(struct io_kiocb *req)
1251 struct io_uring_task *tctx = current->io_uring;
1252 const struct cred *creds = NULL;
1253 struct io_identity *id;
1255 if (req->work.flags & IO_WQ_WORK_CREDS)
1256 creds = req->work.identity->creds;
1258 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1259 if (unlikely(!id)) {
1260 req->work.flags |= IO_WQ_WORK_CANCEL;
1265 * We can safely just re-init the creds we copied Either the field
1266 * matches the current one, or we haven't grabbed it yet. The only
1267 * exception is ->creds, through registered personalities, so handle
1268 * that one separately.
1270 io_init_identity(id);
1272 req->work.identity->creds = creds;
1274 /* add one for this request */
1275 refcount_inc(&id->count);
1277 /* drop old identity, assign new one. one ref for req, one for tctx */
1278 if (req->work.identity != tctx->identity &&
1279 refcount_sub_and_test(2, &req->work.identity->count))
1280 kfree(req->work.identity);
1282 req->work.identity = id;
1283 tctx->identity = id;
1287 static bool io_grab_identity(struct io_kiocb *req)
1289 const struct io_op_def *def = &io_op_defs[req->opcode];
1290 struct io_identity *id = req->work.identity;
1291 struct io_ring_ctx *ctx = req->ctx;
1293 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1294 if (id->fsize != rlimit(RLIMIT_FSIZE))
1296 req->work.flags |= IO_WQ_WORK_FSIZE;
1299 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1300 (def->work_flags & IO_WQ_WORK_FILES) &&
1301 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1302 if (id->files != current->files ||
1303 id->nsproxy != current->nsproxy)
1305 atomic_inc(&id->files->count);
1306 get_nsproxy(id->nsproxy);
1307 req->flags |= REQ_F_INFLIGHT;
1309 spin_lock_irq(&ctx->inflight_lock);
1310 list_add(&req->inflight_entry, &ctx->inflight_list);
1311 spin_unlock_irq(&ctx->inflight_lock);
1312 req->work.flags |= IO_WQ_WORK_FILES;
1314 #ifdef CONFIG_BLK_CGROUP
1315 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1316 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1318 if (id->blkcg_css != blkcg_css()) {
1323 * This should be rare, either the cgroup is dying or the task
1324 * is moving cgroups. Just punt to root for the handful of ios.
1326 if (css_tryget_online(id->blkcg_css))
1327 req->work.flags |= IO_WQ_WORK_BLKCG;
1331 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1332 if (id->creds != current_cred())
1334 get_cred(id->creds);
1335 req->work.flags |= IO_WQ_WORK_CREDS;
1338 if (!uid_eq(current->loginuid, id->loginuid) ||
1339 current->sessionid != id->sessionid)
1342 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1343 (def->work_flags & IO_WQ_WORK_FS)) {
1344 if (current->fs != id->fs)
1346 spin_lock(&id->fs->lock);
1347 if (!id->fs->in_exec) {
1349 req->work.flags |= IO_WQ_WORK_FS;
1351 req->work.flags |= IO_WQ_WORK_CANCEL;
1353 spin_unlock(¤t->fs->lock);
1359 static void io_prep_async_work(struct io_kiocb *req)
1361 const struct io_op_def *def = &io_op_defs[req->opcode];
1362 struct io_ring_ctx *ctx = req->ctx;
1363 struct io_identity *id;
1365 io_req_init_async(req);
1366 id = req->work.identity;
1368 if (req->flags & REQ_F_ISREG) {
1369 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1370 io_wq_hash_work(&req->work, file_inode(req->file));
1372 if (def->unbound_nonreg_file)
1373 req->work.flags |= IO_WQ_WORK_UNBOUND;
1376 /* ->mm can never change on us */
1377 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1378 (def->work_flags & IO_WQ_WORK_MM)) {
1380 req->work.flags |= IO_WQ_WORK_MM;
1383 /* if we fail grabbing identity, we must COW, regrab, and retry */
1384 if (io_grab_identity(req))
1387 if (!io_identity_cow(req))
1390 /* can't fail at this point */
1391 if (!io_grab_identity(req))
1395 static void io_prep_async_link(struct io_kiocb *req)
1397 struct io_kiocb *cur;
1399 io_prep_async_work(req);
1400 if (req->flags & REQ_F_LINK_HEAD)
1401 list_for_each_entry(cur, &req->link_list, link_list)
1402 io_prep_async_work(cur);
1405 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1407 struct io_ring_ctx *ctx = req->ctx;
1408 struct io_kiocb *link = io_prep_linked_timeout(req);
1410 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1411 &req->work, req->flags);
1412 io_wq_enqueue(ctx->io_wq, &req->work);
1416 static void io_queue_async_work(struct io_kiocb *req)
1418 struct io_kiocb *link;
1420 /* init ->work of the whole link before punting */
1421 io_prep_async_link(req);
1422 link = __io_queue_async_work(req);
1425 io_queue_linked_timeout(link);
1428 static void io_kill_timeout(struct io_kiocb *req)
1430 struct io_timeout_data *io = req->async_data;
1433 ret = hrtimer_try_to_cancel(&io->timer);
1435 atomic_set(&req->ctx->cq_timeouts,
1436 atomic_read(&req->ctx->cq_timeouts) + 1);
1437 list_del_init(&req->timeout.list);
1438 io_cqring_fill_event(req, 0);
1439 io_put_req_deferred(req, 1);
1443 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1445 struct io_ring_ctx *ctx = req->ctx;
1447 if (!tsk || req->task == tsk)
1449 if (ctx->flags & IORING_SETUP_SQPOLL) {
1450 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1457 * Returns true if we found and killed one or more timeouts
1459 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1461 struct io_kiocb *req, *tmp;
1464 spin_lock_irq(&ctx->completion_lock);
1465 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1466 if (io_task_match(req, tsk)) {
1467 io_kill_timeout(req);
1471 spin_unlock_irq(&ctx->completion_lock);
1472 return canceled != 0;
1475 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1478 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1479 struct io_defer_entry, list);
1480 struct io_kiocb *link;
1482 if (req_need_defer(de->req, de->seq))
1484 list_del_init(&de->list);
1485 /* punt-init is done before queueing for defer */
1486 link = __io_queue_async_work(de->req);
1488 __io_queue_linked_timeout(link);
1489 /* drop submission reference */
1490 io_put_req_deferred(link, 1);
1493 } while (!list_empty(&ctx->defer_list));
1496 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1498 while (!list_empty(&ctx->timeout_list)) {
1499 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1500 struct io_kiocb, timeout.list);
1502 if (io_is_timeout_noseq(req))
1504 if (req->timeout.target_seq != ctx->cached_cq_tail
1505 - atomic_read(&ctx->cq_timeouts))
1508 list_del_init(&req->timeout.list);
1509 io_kill_timeout(req);
1513 static void io_commit_cqring(struct io_ring_ctx *ctx)
1515 io_flush_timeouts(ctx);
1516 __io_commit_cqring(ctx);
1518 if (unlikely(!list_empty(&ctx->defer_list)))
1519 __io_queue_deferred(ctx);
1522 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1524 struct io_rings *r = ctx->rings;
1526 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1529 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1531 struct io_rings *rings = ctx->rings;
1534 tail = ctx->cached_cq_tail;
1536 * writes to the cq entry need to come after reading head; the
1537 * control dependency is enough as we're using WRITE_ONCE to
1540 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1543 ctx->cached_cq_tail++;
1544 return &rings->cqes[tail & ctx->cq_mask];
1547 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1551 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1553 if (!ctx->eventfd_async)
1555 return io_wq_current_is_worker();
1558 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1560 if (waitqueue_active(&ctx->wait))
1561 wake_up(&ctx->wait);
1562 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1563 wake_up(&ctx->sq_data->wait);
1564 if (io_should_trigger_evfd(ctx))
1565 eventfd_signal(ctx->cq_ev_fd, 1);
1568 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1570 if (list_empty(&ctx->cq_overflow_list)) {
1571 clear_bit(0, &ctx->sq_check_overflow);
1572 clear_bit(0, &ctx->cq_check_overflow);
1573 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1577 static inline bool io_match_files(struct io_kiocb *req,
1578 struct files_struct *files)
1582 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1583 (req->work.flags & IO_WQ_WORK_FILES))
1584 return req->work.identity->files == files;
1588 /* Returns true if there are no backlogged entries after the flush */
1589 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1590 struct task_struct *tsk,
1591 struct files_struct *files)
1593 struct io_rings *rings = ctx->rings;
1594 struct io_kiocb *req, *tmp;
1595 struct io_uring_cqe *cqe;
1596 unsigned long flags;
1600 if (list_empty_careful(&ctx->cq_overflow_list))
1602 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1603 rings->cq_ring_entries))
1607 spin_lock_irqsave(&ctx->completion_lock, flags);
1609 /* if force is set, the ring is going away. always drop after that */
1611 ctx->cq_overflow_flushed = 1;
1614 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1615 if (tsk && req->task != tsk)
1617 if (!io_match_files(req, files))
1620 cqe = io_get_cqring(ctx);
1624 list_move(&req->compl.list, &list);
1626 WRITE_ONCE(cqe->user_data, req->user_data);
1627 WRITE_ONCE(cqe->res, req->result);
1628 WRITE_ONCE(cqe->flags, req->compl.cflags);
1630 ctx->cached_cq_overflow++;
1631 WRITE_ONCE(ctx->rings->cq_overflow,
1632 ctx->cached_cq_overflow);
1636 io_commit_cqring(ctx);
1637 io_cqring_mark_overflow(ctx);
1639 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1640 io_cqring_ev_posted(ctx);
1642 while (!list_empty(&list)) {
1643 req = list_first_entry(&list, struct io_kiocb, compl.list);
1644 list_del(&req->compl.list);
1651 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1653 struct io_ring_ctx *ctx = req->ctx;
1654 struct io_uring_cqe *cqe;
1656 trace_io_uring_complete(ctx, req->user_data, res);
1659 * If we can't get a cq entry, userspace overflowed the
1660 * submission (by quite a lot). Increment the overflow count in
1663 cqe = io_get_cqring(ctx);
1665 WRITE_ONCE(cqe->user_data, req->user_data);
1666 WRITE_ONCE(cqe->res, res);
1667 WRITE_ONCE(cqe->flags, cflags);
1668 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1670 * If we're in ring overflow flush mode, or in task cancel mode,
1671 * then we cannot store the request for later flushing, we need
1672 * to drop it on the floor.
1674 ctx->cached_cq_overflow++;
1675 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1677 if (list_empty(&ctx->cq_overflow_list)) {
1678 set_bit(0, &ctx->sq_check_overflow);
1679 set_bit(0, &ctx->cq_check_overflow);
1680 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1684 req->compl.cflags = cflags;
1685 refcount_inc(&req->refs);
1686 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1690 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1692 __io_cqring_fill_event(req, res, 0);
1695 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1697 struct io_ring_ctx *ctx = req->ctx;
1698 unsigned long flags;
1700 spin_lock_irqsave(&ctx->completion_lock, flags);
1701 __io_cqring_fill_event(req, res, cflags);
1702 io_commit_cqring(ctx);
1703 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1705 io_cqring_ev_posted(ctx);
1708 static void io_submit_flush_completions(struct io_comp_state *cs)
1710 struct io_ring_ctx *ctx = cs->ctx;
1712 spin_lock_irq(&ctx->completion_lock);
1713 while (!list_empty(&cs->list)) {
1714 struct io_kiocb *req;
1716 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1717 list_del(&req->compl.list);
1718 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1721 * io_free_req() doesn't care about completion_lock unless one
1722 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1723 * because of a potential deadlock with req->work.fs->lock
1725 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1726 |REQ_F_WORK_INITIALIZED)) {
1727 spin_unlock_irq(&ctx->completion_lock);
1729 spin_lock_irq(&ctx->completion_lock);
1734 io_commit_cqring(ctx);
1735 spin_unlock_irq(&ctx->completion_lock);
1737 io_cqring_ev_posted(ctx);
1741 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1742 struct io_comp_state *cs)
1745 io_cqring_add_event(req, res, cflags);
1750 req->compl.cflags = cflags;
1751 list_add_tail(&req->compl.list, &cs->list);
1753 io_submit_flush_completions(cs);
1757 static void io_req_complete(struct io_kiocb *req, long res)
1759 __io_req_complete(req, res, 0, NULL);
1762 static inline bool io_is_fallback_req(struct io_kiocb *req)
1764 return req == (struct io_kiocb *)
1765 ((unsigned long) req->ctx->fallback_req & ~1UL);
1768 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1770 struct io_kiocb *req;
1772 req = ctx->fallback_req;
1773 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1779 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1780 struct io_submit_state *state)
1782 if (!state->free_reqs) {
1783 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1787 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1788 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1791 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1792 * retry single alloc to be on the safe side.
1794 if (unlikely(ret <= 0)) {
1795 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1796 if (!state->reqs[0])
1800 state->free_reqs = ret;
1804 return state->reqs[state->free_reqs];
1806 return io_get_fallback_req(ctx);
1809 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1813 percpu_ref_put(req->fixed_file_refs);
1818 static void io_dismantle_req(struct io_kiocb *req)
1822 if (req->async_data)
1823 kfree(req->async_data);
1825 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1827 io_req_clean_work(req);
1830 static void __io_free_req(struct io_kiocb *req)
1832 struct io_uring_task *tctx = req->task->io_uring;
1833 struct io_ring_ctx *ctx = req->ctx;
1835 io_dismantle_req(req);
1837 percpu_counter_dec(&tctx->inflight);
1839 wake_up(&tctx->wait);
1840 put_task_struct(req->task);
1842 if (likely(!io_is_fallback_req(req)))
1843 kmem_cache_free(req_cachep, req);
1845 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1846 percpu_ref_put(&ctx->refs);
1849 static void io_kill_linked_timeout(struct io_kiocb *req)
1851 struct io_ring_ctx *ctx = req->ctx;
1852 struct io_kiocb *link;
1853 bool cancelled = false;
1854 unsigned long flags;
1856 spin_lock_irqsave(&ctx->completion_lock, flags);
1857 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1860 * Can happen if a linked timeout fired and link had been like
1861 * req -> link t-out -> link t-out [-> ...]
1863 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1864 struct io_timeout_data *io = link->async_data;
1867 list_del_init(&link->link_list);
1868 ret = hrtimer_try_to_cancel(&io->timer);
1870 io_cqring_fill_event(link, -ECANCELED);
1871 io_commit_cqring(ctx);
1875 req->flags &= ~REQ_F_LINK_TIMEOUT;
1876 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1879 io_cqring_ev_posted(ctx);
1884 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1886 struct io_kiocb *nxt;
1889 * The list should never be empty when we are called here. But could
1890 * potentially happen if the chain is messed up, check to be on the
1893 if (unlikely(list_empty(&req->link_list)))
1896 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1897 list_del_init(&req->link_list);
1898 if (!list_empty(&nxt->link_list))
1899 nxt->flags |= REQ_F_LINK_HEAD;
1904 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1906 static void io_fail_links(struct io_kiocb *req)
1908 struct io_ring_ctx *ctx = req->ctx;
1909 unsigned long flags;
1911 spin_lock_irqsave(&ctx->completion_lock, flags);
1912 while (!list_empty(&req->link_list)) {
1913 struct io_kiocb *link = list_first_entry(&req->link_list,
1914 struct io_kiocb, link_list);
1916 list_del_init(&link->link_list);
1917 trace_io_uring_fail_link(req, link);
1919 io_cqring_fill_event(link, -ECANCELED);
1922 * It's ok to free under spinlock as they're not linked anymore,
1923 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1926 if (link->flags & REQ_F_WORK_INITIALIZED)
1927 io_put_req_deferred(link, 2);
1929 io_double_put_req(link);
1932 io_commit_cqring(ctx);
1933 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1935 io_cqring_ev_posted(ctx);
1938 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1940 req->flags &= ~REQ_F_LINK_HEAD;
1941 if (req->flags & REQ_F_LINK_TIMEOUT)
1942 io_kill_linked_timeout(req);
1945 * If LINK is set, we have dependent requests in this chain. If we
1946 * didn't fail this request, queue the first one up, moving any other
1947 * dependencies to the next request. In case of failure, fail the rest
1950 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1951 return io_req_link_next(req);
1956 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1958 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1960 return __io_req_find_next(req);
1963 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1965 struct task_struct *tsk = req->task;
1966 struct io_ring_ctx *ctx = req->ctx;
1969 if (tsk->flags & PF_EXITING)
1973 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1974 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1975 * processing task_work. There's no reliable way to tell if TWA_RESUME
1979 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1980 notify = TWA_SIGNAL;
1982 ret = task_work_add(tsk, &req->task_work, notify);
1984 wake_up_process(tsk);
1989 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1991 struct io_ring_ctx *ctx = req->ctx;
1993 spin_lock_irq(&ctx->completion_lock);
1994 io_cqring_fill_event(req, error);
1995 io_commit_cqring(ctx);
1996 spin_unlock_irq(&ctx->completion_lock);
1998 io_cqring_ev_posted(ctx);
1999 req_set_fail_links(req);
2000 io_double_put_req(req);
2003 static void io_req_task_cancel(struct callback_head *cb)
2005 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2006 struct io_ring_ctx *ctx = req->ctx;
2008 __io_req_task_cancel(req, -ECANCELED);
2009 percpu_ref_put(&ctx->refs);
2012 static void __io_req_task_submit(struct io_kiocb *req)
2014 struct io_ring_ctx *ctx = req->ctx;
2016 if (!__io_sq_thread_acquire_mm(ctx)) {
2017 mutex_lock(&ctx->uring_lock);
2018 __io_queue_sqe(req, NULL);
2019 mutex_unlock(&ctx->uring_lock);
2021 __io_req_task_cancel(req, -EFAULT);
2025 static void io_req_task_submit(struct callback_head *cb)
2027 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2028 struct io_ring_ctx *ctx = req->ctx;
2030 __io_req_task_submit(req);
2031 percpu_ref_put(&ctx->refs);
2034 static void io_req_task_queue(struct io_kiocb *req)
2038 init_task_work(&req->task_work, io_req_task_submit);
2039 percpu_ref_get(&req->ctx->refs);
2041 ret = io_req_task_work_add(req, true);
2042 if (unlikely(ret)) {
2043 struct task_struct *tsk;
2045 init_task_work(&req->task_work, io_req_task_cancel);
2046 tsk = io_wq_get_task(req->ctx->io_wq);
2047 task_work_add(tsk, &req->task_work, 0);
2048 wake_up_process(tsk);
2052 static void io_queue_next(struct io_kiocb *req)
2054 struct io_kiocb *nxt = io_req_find_next(req);
2057 io_req_task_queue(nxt);
2060 static void io_free_req(struct io_kiocb *req)
2067 void *reqs[IO_IOPOLL_BATCH];
2070 struct task_struct *task;
2074 static inline void io_init_req_batch(struct req_batch *rb)
2081 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2082 struct req_batch *rb)
2084 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2085 percpu_ref_put_many(&ctx->refs, rb->to_free);
2089 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2090 struct req_batch *rb)
2093 __io_req_free_batch_flush(ctx, rb);
2095 struct io_uring_task *tctx = rb->task->io_uring;
2097 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2098 put_task_struct_many(rb->task, rb->task_refs);
2103 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2105 if (unlikely(io_is_fallback_req(req))) {
2109 if (req->flags & REQ_F_LINK_HEAD)
2112 if (req->task != rb->task) {
2114 struct io_uring_task *tctx = rb->task->io_uring;
2116 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2117 put_task_struct_many(rb->task, rb->task_refs);
2119 rb->task = req->task;
2124 io_dismantle_req(req);
2125 rb->reqs[rb->to_free++] = req;
2126 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2127 __io_req_free_batch_flush(req->ctx, rb);
2131 * Drop reference to request, return next in chain (if there is one) if this
2132 * was the last reference to this request.
2134 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2136 struct io_kiocb *nxt = NULL;
2138 if (refcount_dec_and_test(&req->refs)) {
2139 nxt = io_req_find_next(req);
2145 static void io_put_req(struct io_kiocb *req)
2147 if (refcount_dec_and_test(&req->refs))
2151 static void io_put_req_deferred_cb(struct callback_head *cb)
2153 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2158 static void io_free_req_deferred(struct io_kiocb *req)
2162 init_task_work(&req->task_work, io_put_req_deferred_cb);
2163 ret = io_req_task_work_add(req, true);
2164 if (unlikely(ret)) {
2165 struct task_struct *tsk;
2167 tsk = io_wq_get_task(req->ctx->io_wq);
2168 task_work_add(tsk, &req->task_work, 0);
2169 wake_up_process(tsk);
2173 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2175 if (refcount_sub_and_test(refs, &req->refs))
2176 io_free_req_deferred(req);
2179 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2181 struct io_kiocb *nxt;
2184 * A ref is owned by io-wq in which context we're. So, if that's the
2185 * last one, it's safe to steal next work. False negatives are Ok,
2186 * it just will be re-punted async in io_put_work()
2188 if (refcount_read(&req->refs) != 1)
2191 nxt = io_req_find_next(req);
2192 return nxt ? &nxt->work : NULL;
2195 static void io_double_put_req(struct io_kiocb *req)
2197 /* drop both submit and complete references */
2198 if (refcount_sub_and_test(2, &req->refs))
2202 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2204 struct io_rings *rings = ctx->rings;
2206 if (test_bit(0, &ctx->cq_check_overflow)) {
2208 * noflush == true is from the waitqueue handler, just ensure
2209 * we wake up the task, and the next invocation will flush the
2210 * entries. We cannot safely to it from here.
2212 if (noflush && !list_empty(&ctx->cq_overflow_list))
2215 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2218 /* See comment at the top of this file */
2220 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2223 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2225 struct io_rings *rings = ctx->rings;
2227 /* make sure SQ entry isn't read before tail */
2228 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2231 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2233 unsigned int cflags;
2235 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2236 cflags |= IORING_CQE_F_BUFFER;
2237 req->flags &= ~REQ_F_BUFFER_SELECTED;
2242 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2244 struct io_buffer *kbuf;
2246 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2247 return io_put_kbuf(req, kbuf);
2250 static inline bool io_run_task_work(void)
2253 * Not safe to run on exiting task, and the task_work handling will
2254 * not add work to such a task.
2256 if (unlikely(current->flags & PF_EXITING))
2258 if (current->task_works) {
2259 __set_current_state(TASK_RUNNING);
2267 static void io_iopoll_queue(struct list_head *again)
2269 struct io_kiocb *req;
2272 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2273 list_del(&req->inflight_entry);
2274 __io_complete_rw(req, -EAGAIN, 0, NULL);
2275 } while (!list_empty(again));
2279 * Find and free completed poll iocbs
2281 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2282 struct list_head *done)
2284 struct req_batch rb;
2285 struct io_kiocb *req;
2288 /* order with ->result store in io_complete_rw_iopoll() */
2291 io_init_req_batch(&rb);
2292 while (!list_empty(done)) {
2295 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2296 if (READ_ONCE(req->result) == -EAGAIN) {
2298 req->iopoll_completed = 0;
2299 list_move_tail(&req->inflight_entry, &again);
2302 list_del(&req->inflight_entry);
2304 if (req->flags & REQ_F_BUFFER_SELECTED)
2305 cflags = io_put_rw_kbuf(req);
2307 __io_cqring_fill_event(req, req->result, cflags);
2310 if (refcount_dec_and_test(&req->refs))
2311 io_req_free_batch(&rb, req);
2314 io_commit_cqring(ctx);
2315 if (ctx->flags & IORING_SETUP_SQPOLL)
2316 io_cqring_ev_posted(ctx);
2317 io_req_free_batch_finish(ctx, &rb);
2319 if (!list_empty(&again))
2320 io_iopoll_queue(&again);
2323 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2326 struct io_kiocb *req, *tmp;
2332 * Only spin for completions if we don't have multiple devices hanging
2333 * off our complete list, and we're under the requested amount.
2335 spin = !ctx->poll_multi_file && *nr_events < min;
2338 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2339 struct kiocb *kiocb = &req->rw.kiocb;
2342 * Move completed and retryable entries to our local lists.
2343 * If we find a request that requires polling, break out
2344 * and complete those lists first, if we have entries there.
2346 if (READ_ONCE(req->iopoll_completed)) {
2347 list_move_tail(&req->inflight_entry, &done);
2350 if (!list_empty(&done))
2353 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2357 /* iopoll may have completed current req */
2358 if (READ_ONCE(req->iopoll_completed))
2359 list_move_tail(&req->inflight_entry, &done);
2366 if (!list_empty(&done))
2367 io_iopoll_complete(ctx, nr_events, &done);
2373 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2374 * non-spinning poll check - we'll still enter the driver poll loop, but only
2375 * as a non-spinning completion check.
2377 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2380 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2383 ret = io_do_iopoll(ctx, nr_events, min);
2386 if (*nr_events >= min)
2394 * We can't just wait for polled events to come to us, we have to actively
2395 * find and complete them.
2397 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2399 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2402 mutex_lock(&ctx->uring_lock);
2403 while (!list_empty(&ctx->iopoll_list)) {
2404 unsigned int nr_events = 0;
2406 io_do_iopoll(ctx, &nr_events, 0);
2408 /* let it sleep and repeat later if can't complete a request */
2412 * Ensure we allow local-to-the-cpu processing to take place,
2413 * in this case we need to ensure that we reap all events.
2414 * Also let task_work, etc. to progress by releasing the mutex
2416 if (need_resched()) {
2417 mutex_unlock(&ctx->uring_lock);
2419 mutex_lock(&ctx->uring_lock);
2422 mutex_unlock(&ctx->uring_lock);
2425 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2427 unsigned int nr_events = 0;
2428 int iters = 0, ret = 0;
2431 * We disallow the app entering submit/complete with polling, but we
2432 * still need to lock the ring to prevent racing with polled issue
2433 * that got punted to a workqueue.
2435 mutex_lock(&ctx->uring_lock);
2438 * Don't enter poll loop if we already have events pending.
2439 * If we do, we can potentially be spinning for commands that
2440 * already triggered a CQE (eg in error).
2442 if (io_cqring_events(ctx, false))
2446 * If a submit got punted to a workqueue, we can have the
2447 * application entering polling for a command before it gets
2448 * issued. That app will hold the uring_lock for the duration
2449 * of the poll right here, so we need to take a breather every
2450 * now and then to ensure that the issue has a chance to add
2451 * the poll to the issued list. Otherwise we can spin here
2452 * forever, while the workqueue is stuck trying to acquire the
2455 if (!(++iters & 7)) {
2456 mutex_unlock(&ctx->uring_lock);
2458 mutex_lock(&ctx->uring_lock);
2461 ret = io_iopoll_getevents(ctx, &nr_events, min);
2465 } while (min && !nr_events && !need_resched());
2467 mutex_unlock(&ctx->uring_lock);
2471 static void kiocb_end_write(struct io_kiocb *req)
2474 * Tell lockdep we inherited freeze protection from submission
2477 if (req->flags & REQ_F_ISREG) {
2478 struct inode *inode = file_inode(req->file);
2480 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2482 file_end_write(req->file);
2485 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2486 struct io_comp_state *cs)
2488 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2491 if (kiocb->ki_flags & IOCB_WRITE)
2492 kiocb_end_write(req);
2494 if (res != req->result)
2495 req_set_fail_links(req);
2496 if (req->flags & REQ_F_BUFFER_SELECTED)
2497 cflags = io_put_rw_kbuf(req);
2498 __io_req_complete(req, res, cflags, cs);
2502 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2504 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2505 ssize_t ret = -ECANCELED;
2506 struct iov_iter iter;
2514 switch (req->opcode) {
2515 case IORING_OP_READV:
2516 case IORING_OP_READ_FIXED:
2517 case IORING_OP_READ:
2520 case IORING_OP_WRITEV:
2521 case IORING_OP_WRITE_FIXED:
2522 case IORING_OP_WRITE:
2526 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2531 if (!req->async_data) {
2532 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2535 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2543 req_set_fail_links(req);
2544 io_req_complete(req, ret);
2549 static bool io_rw_reissue(struct io_kiocb *req, long res)
2552 umode_t mode = file_inode(req->file)->i_mode;
2555 if (!S_ISBLK(mode) && !S_ISREG(mode))
2557 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2560 ret = io_sq_thread_acquire_mm(req->ctx, req);
2562 if (io_resubmit_prep(req, ret)) {
2563 refcount_inc(&req->refs);
2564 io_queue_async_work(req);
2572 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2573 struct io_comp_state *cs)
2575 if (!io_rw_reissue(req, res))
2576 io_complete_rw_common(&req->rw.kiocb, res, cs);
2579 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2581 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2583 __io_complete_rw(req, res, res2, NULL);
2586 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2588 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2590 if (kiocb->ki_flags & IOCB_WRITE)
2591 kiocb_end_write(req);
2593 if (res != -EAGAIN && res != req->result)
2594 req_set_fail_links(req);
2596 WRITE_ONCE(req->result, res);
2597 /* order with io_poll_complete() checking ->result */
2599 WRITE_ONCE(req->iopoll_completed, 1);
2603 * After the iocb has been issued, it's safe to be found on the poll list.
2604 * Adding the kiocb to the list AFTER submission ensures that we don't
2605 * find it from a io_iopoll_getevents() thread before the issuer is done
2606 * accessing the kiocb cookie.
2608 static void io_iopoll_req_issued(struct io_kiocb *req)
2610 struct io_ring_ctx *ctx = req->ctx;
2613 * Track whether we have multiple files in our lists. This will impact
2614 * how we do polling eventually, not spinning if we're on potentially
2615 * different devices.
2617 if (list_empty(&ctx->iopoll_list)) {
2618 ctx->poll_multi_file = false;
2619 } else if (!ctx->poll_multi_file) {
2620 struct io_kiocb *list_req;
2622 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2624 if (list_req->file != req->file)
2625 ctx->poll_multi_file = true;
2629 * For fast devices, IO may have already completed. If it has, add
2630 * it to the front so we find it first.
2632 if (READ_ONCE(req->iopoll_completed))
2633 list_add(&req->inflight_entry, &ctx->iopoll_list);
2635 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2637 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2638 wq_has_sleeper(&ctx->sq_data->wait))
2639 wake_up(&ctx->sq_data->wait);
2642 static void __io_state_file_put(struct io_submit_state *state)
2644 if (state->has_refs)
2645 fput_many(state->file, state->has_refs);
2649 static inline void io_state_file_put(struct io_submit_state *state)
2652 __io_state_file_put(state);
2656 * Get as many references to a file as we have IOs left in this submission,
2657 * assuming most submissions are for one file, or at least that each file
2658 * has more than one submission.
2660 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2666 if (state->fd == fd) {
2670 __io_state_file_put(state);
2672 state->file = fget_many(fd, state->ios_left);
2677 state->has_refs = state->ios_left - 1;
2681 static bool io_bdev_nowait(struct block_device *bdev)
2684 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2691 * If we tracked the file through the SCM inflight mechanism, we could support
2692 * any file. For now, just ensure that anything potentially problematic is done
2695 static bool io_file_supports_async(struct file *file, int rw)
2697 umode_t mode = file_inode(file)->i_mode;
2699 if (S_ISBLK(mode)) {
2700 if (io_bdev_nowait(file->f_inode->i_bdev))
2704 if (S_ISCHR(mode) || S_ISSOCK(mode))
2706 if (S_ISREG(mode)) {
2707 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2708 file->f_op != &io_uring_fops)
2713 /* any ->read/write should understand O_NONBLOCK */
2714 if (file->f_flags & O_NONBLOCK)
2717 if (!(file->f_mode & FMODE_NOWAIT))
2721 return file->f_op->read_iter != NULL;
2723 return file->f_op->write_iter != NULL;
2726 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2728 struct io_ring_ctx *ctx = req->ctx;
2729 struct kiocb *kiocb = &req->rw.kiocb;
2733 if (S_ISREG(file_inode(req->file)->i_mode))
2734 req->flags |= REQ_F_ISREG;
2736 kiocb->ki_pos = READ_ONCE(sqe->off);
2737 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2738 req->flags |= REQ_F_CUR_POS;
2739 kiocb->ki_pos = req->file->f_pos;
2741 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2742 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2743 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2747 ioprio = READ_ONCE(sqe->ioprio);
2749 ret = ioprio_check_cap(ioprio);
2753 kiocb->ki_ioprio = ioprio;
2755 kiocb->ki_ioprio = get_current_ioprio();
2757 /* don't allow async punt if RWF_NOWAIT was requested */
2758 if (kiocb->ki_flags & IOCB_NOWAIT)
2759 req->flags |= REQ_F_NOWAIT;
2761 if (ctx->flags & IORING_SETUP_IOPOLL) {
2762 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2763 !kiocb->ki_filp->f_op->iopoll)
2766 kiocb->ki_flags |= IOCB_HIPRI;
2767 kiocb->ki_complete = io_complete_rw_iopoll;
2768 req->iopoll_completed = 0;
2770 if (kiocb->ki_flags & IOCB_HIPRI)
2772 kiocb->ki_complete = io_complete_rw;
2775 req->rw.addr = READ_ONCE(sqe->addr);
2776 req->rw.len = READ_ONCE(sqe->len);
2777 req->buf_index = READ_ONCE(sqe->buf_index);
2781 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2787 case -ERESTARTNOINTR:
2788 case -ERESTARTNOHAND:
2789 case -ERESTART_RESTARTBLOCK:
2791 * We can't just restart the syscall, since previously
2792 * submitted sqes may already be in progress. Just fail this
2798 kiocb->ki_complete(kiocb, ret, 0);
2802 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2803 struct io_comp_state *cs)
2805 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2806 struct io_async_rw *io = req->async_data;
2808 /* add previously done IO, if any */
2809 if (io && io->bytes_done > 0) {
2811 ret = io->bytes_done;
2813 ret += io->bytes_done;
2816 if (req->flags & REQ_F_CUR_POS)
2817 req->file->f_pos = kiocb->ki_pos;
2818 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2819 __io_complete_rw(req, ret, 0, cs);
2821 io_rw_done(kiocb, ret);
2824 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2825 struct iov_iter *iter)
2827 struct io_ring_ctx *ctx = req->ctx;
2828 size_t len = req->rw.len;
2829 struct io_mapped_ubuf *imu;
2830 u16 index, buf_index = req->buf_index;
2834 if (unlikely(buf_index >= ctx->nr_user_bufs))
2836 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2837 imu = &ctx->user_bufs[index];
2838 buf_addr = req->rw.addr;
2841 if (buf_addr + len < buf_addr)
2843 /* not inside the mapped region */
2844 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2848 * May not be a start of buffer, set size appropriately
2849 * and advance us to the beginning.
2851 offset = buf_addr - imu->ubuf;
2852 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2856 * Don't use iov_iter_advance() here, as it's really slow for
2857 * using the latter parts of a big fixed buffer - it iterates
2858 * over each segment manually. We can cheat a bit here, because
2861 * 1) it's a BVEC iter, we set it up
2862 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2863 * first and last bvec
2865 * So just find our index, and adjust the iterator afterwards.
2866 * If the offset is within the first bvec (or the whole first
2867 * bvec, just use iov_iter_advance(). This makes it easier
2868 * since we can just skip the first segment, which may not
2869 * be PAGE_SIZE aligned.
2871 const struct bio_vec *bvec = imu->bvec;
2873 if (offset <= bvec->bv_len) {
2874 iov_iter_advance(iter, offset);
2876 unsigned long seg_skip;
2878 /* skip first vec */
2879 offset -= bvec->bv_len;
2880 seg_skip = 1 + (offset >> PAGE_SHIFT);
2882 iter->bvec = bvec + seg_skip;
2883 iter->nr_segs -= seg_skip;
2884 iter->count -= bvec->bv_len + offset;
2885 iter->iov_offset = offset & ~PAGE_MASK;
2892 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2895 mutex_unlock(&ctx->uring_lock);
2898 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2901 * "Normal" inline submissions always hold the uring_lock, since we
2902 * grab it from the system call. Same is true for the SQPOLL offload.
2903 * The only exception is when we've detached the request and issue it
2904 * from an async worker thread, grab the lock for that case.
2907 mutex_lock(&ctx->uring_lock);
2910 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2911 int bgid, struct io_buffer *kbuf,
2914 struct io_buffer *head;
2916 if (req->flags & REQ_F_BUFFER_SELECTED)
2919 io_ring_submit_lock(req->ctx, needs_lock);
2921 lockdep_assert_held(&req->ctx->uring_lock);
2923 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2925 if (!list_empty(&head->list)) {
2926 kbuf = list_last_entry(&head->list, struct io_buffer,
2928 list_del(&kbuf->list);
2931 idr_remove(&req->ctx->io_buffer_idr, bgid);
2933 if (*len > kbuf->len)
2936 kbuf = ERR_PTR(-ENOBUFS);
2939 io_ring_submit_unlock(req->ctx, needs_lock);
2944 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2947 struct io_buffer *kbuf;
2950 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2951 bgid = req->buf_index;
2952 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2955 req->rw.addr = (u64) (unsigned long) kbuf;
2956 req->flags |= REQ_F_BUFFER_SELECTED;
2957 return u64_to_user_ptr(kbuf->addr);
2960 #ifdef CONFIG_COMPAT
2961 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2964 struct compat_iovec __user *uiov;
2965 compat_ssize_t clen;
2969 uiov = u64_to_user_ptr(req->rw.addr);
2970 if (!access_ok(uiov, sizeof(*uiov)))
2972 if (__get_user(clen, &uiov->iov_len))
2978 buf = io_rw_buffer_select(req, &len, needs_lock);
2980 return PTR_ERR(buf);
2981 iov[0].iov_base = buf;
2982 iov[0].iov_len = (compat_size_t) len;
2987 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2990 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2994 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2997 len = iov[0].iov_len;
3000 buf = io_rw_buffer_select(req, &len, needs_lock);
3002 return PTR_ERR(buf);
3003 iov[0].iov_base = buf;
3004 iov[0].iov_len = len;
3008 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3011 if (req->flags & REQ_F_BUFFER_SELECTED) {
3012 struct io_buffer *kbuf;
3014 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3015 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3016 iov[0].iov_len = kbuf->len;
3021 else if (req->rw.len > 1)
3024 #ifdef CONFIG_COMPAT
3025 if (req->ctx->compat)
3026 return io_compat_import(req, iov, needs_lock);
3029 return __io_iov_buffer_select(req, iov, needs_lock);
3032 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3033 struct iovec **iovec, struct iov_iter *iter,
3036 void __user *buf = u64_to_user_ptr(req->rw.addr);
3037 size_t sqe_len = req->rw.len;
3041 opcode = req->opcode;
3042 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3044 return io_import_fixed(req, rw, iter);
3047 /* buffer index only valid with fixed read/write, or buffer select */
3048 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3051 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3052 if (req->flags & REQ_F_BUFFER_SELECT) {
3053 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3055 return PTR_ERR(buf);
3056 req->rw.len = sqe_len;
3059 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3061 return ret < 0 ? ret : sqe_len;
3064 if (req->flags & REQ_F_BUFFER_SELECT) {
3065 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3067 ret = (*iovec)->iov_len;
3068 iov_iter_init(iter, rw, *iovec, 1, ret);
3074 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3078 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3079 struct iovec **iovec, struct iov_iter *iter,
3082 struct io_async_rw *iorw = req->async_data;
3085 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3087 return iov_iter_count(&iorw->iter);
3090 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3092 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3096 * For files that don't have ->read_iter() and ->write_iter(), handle them
3097 * by looping over ->read() or ->write() manually.
3099 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3101 struct kiocb *kiocb = &req->rw.kiocb;
3102 struct file *file = req->file;
3106 * Don't support polled IO through this interface, and we can't
3107 * support non-blocking either. For the latter, this just causes
3108 * the kiocb to be handled from an async context.
3110 if (kiocb->ki_flags & IOCB_HIPRI)
3112 if (kiocb->ki_flags & IOCB_NOWAIT)
3115 while (iov_iter_count(iter)) {
3119 if (!iov_iter_is_bvec(iter)) {
3120 iovec = iov_iter_iovec(iter);
3122 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3123 iovec.iov_len = req->rw.len;
3127 nr = file->f_op->read(file, iovec.iov_base,
3128 iovec.iov_len, io_kiocb_ppos(kiocb));
3130 nr = file->f_op->write(file, iovec.iov_base,
3131 iovec.iov_len, io_kiocb_ppos(kiocb));
3140 if (nr != iovec.iov_len)
3144 iov_iter_advance(iter, nr);
3150 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3151 const struct iovec *fast_iov, struct iov_iter *iter)
3153 struct io_async_rw *rw = req->async_data;
3155 memcpy(&rw->iter, iter, sizeof(*iter));
3156 rw->free_iovec = iovec;
3158 /* can only be fixed buffers, no need to do anything */
3159 if (iter->type == ITER_BVEC)
3162 unsigned iov_off = 0;
3164 rw->iter.iov = rw->fast_iov;
3165 if (iter->iov != fast_iov) {
3166 iov_off = iter->iov - fast_iov;
3167 rw->iter.iov += iov_off;
3169 if (rw->fast_iov != fast_iov)
3170 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3171 sizeof(struct iovec) * iter->nr_segs);
3173 req->flags |= REQ_F_NEED_CLEANUP;
3177 static inline int __io_alloc_async_data(struct io_kiocb *req)
3179 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3180 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3181 return req->async_data == NULL;
3184 static int io_alloc_async_data(struct io_kiocb *req)
3186 if (!io_op_defs[req->opcode].needs_async_data)
3189 return __io_alloc_async_data(req);
3192 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3193 const struct iovec *fast_iov,
3194 struct iov_iter *iter, bool force)
3196 if (!force && !io_op_defs[req->opcode].needs_async_data)
3198 if (!req->async_data) {
3199 if (__io_alloc_async_data(req))
3202 io_req_map_rw(req, iovec, fast_iov, iter);
3207 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3209 struct io_async_rw *iorw = req->async_data;
3210 struct iovec *iov = iorw->fast_iov;
3213 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3214 if (unlikely(ret < 0))
3217 iorw->bytes_done = 0;
3218 iorw->free_iovec = iov;
3220 req->flags |= REQ_F_NEED_CLEANUP;
3224 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3228 ret = io_prep_rw(req, sqe);
3232 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3235 /* either don't need iovec imported or already have it */
3236 if (!req->async_data)
3238 return io_rw_prep_async(req, READ);
3242 * This is our waitqueue callback handler, registered through lock_page_async()
3243 * when we initially tried to do the IO with the iocb armed our waitqueue.
3244 * This gets called when the page is unlocked, and we generally expect that to
3245 * happen when the page IO is completed and the page is now uptodate. This will
3246 * queue a task_work based retry of the operation, attempting to copy the data
3247 * again. If the latter fails because the page was NOT uptodate, then we will
3248 * do a thread based blocking retry of the operation. That's the unexpected
3251 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3252 int sync, void *arg)
3254 struct wait_page_queue *wpq;
3255 struct io_kiocb *req = wait->private;
3256 struct wait_page_key *key = arg;
3259 wpq = container_of(wait, struct wait_page_queue, wait);
3261 if (!wake_page_match(wpq, key))
3264 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3265 list_del_init(&wait->entry);
3267 init_task_work(&req->task_work, io_req_task_submit);
3268 percpu_ref_get(&req->ctx->refs);
3270 /* submit ref gets dropped, acquire a new one */
3271 refcount_inc(&req->refs);
3272 ret = io_req_task_work_add(req, true);
3273 if (unlikely(ret)) {
3274 struct task_struct *tsk;
3276 /* queue just for cancelation */
3277 init_task_work(&req->task_work, io_req_task_cancel);
3278 tsk = io_wq_get_task(req->ctx->io_wq);
3279 task_work_add(tsk, &req->task_work, 0);
3280 wake_up_process(tsk);
3286 * This controls whether a given IO request should be armed for async page
3287 * based retry. If we return false here, the request is handed to the async
3288 * worker threads for retry. If we're doing buffered reads on a regular file,
3289 * we prepare a private wait_page_queue entry and retry the operation. This
3290 * will either succeed because the page is now uptodate and unlocked, or it
3291 * will register a callback when the page is unlocked at IO completion. Through
3292 * that callback, io_uring uses task_work to setup a retry of the operation.
3293 * That retry will attempt the buffered read again. The retry will generally
3294 * succeed, or in rare cases where it fails, we then fall back to using the
3295 * async worker threads for a blocking retry.
3297 static bool io_rw_should_retry(struct io_kiocb *req)
3299 struct io_async_rw *rw = req->async_data;
3300 struct wait_page_queue *wait = &rw->wpq;
3301 struct kiocb *kiocb = &req->rw.kiocb;
3303 /* never retry for NOWAIT, we just complete with -EAGAIN */
3304 if (req->flags & REQ_F_NOWAIT)
3307 /* Only for buffered IO */
3308 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3312 * just use poll if we can, and don't attempt if the fs doesn't
3313 * support callback based unlocks
3315 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3318 wait->wait.func = io_async_buf_func;
3319 wait->wait.private = req;
3320 wait->wait.flags = 0;
3321 INIT_LIST_HEAD(&wait->wait.entry);
3322 kiocb->ki_flags |= IOCB_WAITQ;
3323 kiocb->ki_flags &= ~IOCB_NOWAIT;
3324 kiocb->ki_waitq = wait;
3328 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3330 if (req->file->f_op->read_iter)
3331 return call_read_iter(req->file, &req->rw.kiocb, iter);
3332 else if (req->file->f_op->read)
3333 return loop_rw_iter(READ, req, iter);
3338 static int io_read(struct io_kiocb *req, bool force_nonblock,
3339 struct io_comp_state *cs)
3341 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3342 struct kiocb *kiocb = &req->rw.kiocb;
3343 struct iov_iter __iter, *iter = &__iter;
3344 struct io_async_rw *rw = req->async_data;
3345 ssize_t io_size, ret, ret2;
3352 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3355 iov_count = iov_iter_count(iter);
3357 req->result = io_size;
3360 /* Ensure we clear previously set non-block flag */
3361 if (!force_nonblock)
3362 kiocb->ki_flags &= ~IOCB_NOWAIT;
3364 kiocb->ki_flags |= IOCB_NOWAIT;
3367 /* If the file doesn't support async, just async punt */
3368 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3372 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3376 ret = io_iter_do_read(req, iter);
3380 } else if (ret == -EIOCBQUEUED) {
3383 } else if (ret == -EAGAIN) {
3384 /* IOPOLL retry should happen for io-wq threads */
3385 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3387 /* no retry on NONBLOCK marked file */
3388 if (req->file->f_flags & O_NONBLOCK)
3390 /* some cases will consume bytes even on error returns */
3391 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3394 } else if (ret < 0) {
3395 /* make sure -ERESTARTSYS -> -EINTR is done */
3399 /* read it all, or we did blocking attempt. no retry. */
3400 if (!iov_iter_count(iter) || !force_nonblock ||
3401 (req->file->f_flags & O_NONBLOCK))
3406 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3413 rw = req->async_data;
3414 /* it's copied and will be cleaned with ->io */
3416 /* now use our persistent iterator, if we aren't already */
3419 rw->bytes_done += ret;
3420 /* if we can retry, do so with the callbacks armed */
3421 if (!io_rw_should_retry(req)) {
3422 kiocb->ki_flags &= ~IOCB_WAITQ;
3427 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3428 * get -EIOCBQUEUED, then we'll get a notification when the desired
3429 * page gets unlocked. We can also get a partial read here, and if we
3430 * do, then just retry at the new offset.
3432 ret = io_iter_do_read(req, iter);
3433 if (ret == -EIOCBQUEUED) {
3436 } else if (ret > 0 && ret < io_size) {
3437 /* we got some bytes, but not all. retry. */
3441 kiocb_done(kiocb, ret, cs);
3444 /* it's reportedly faster than delegating the null check to kfree() */
3450 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3454 ret = io_prep_rw(req, sqe);
3458 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3461 /* either don't need iovec imported or already have it */
3462 if (!req->async_data)
3464 return io_rw_prep_async(req, WRITE);
3467 static int io_write(struct io_kiocb *req, bool force_nonblock,
3468 struct io_comp_state *cs)
3470 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3471 struct kiocb *kiocb = &req->rw.kiocb;
3472 struct iov_iter __iter, *iter = &__iter;
3473 struct io_async_rw *rw = req->async_data;
3475 ssize_t ret, ret2, io_size;
3480 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3483 iov_count = iov_iter_count(iter);
3485 req->result = io_size;
3487 /* Ensure we clear previously set non-block flag */
3488 if (!force_nonblock)
3489 kiocb->ki_flags &= ~IOCB_NOWAIT;
3491 kiocb->ki_flags |= IOCB_NOWAIT;
3493 /* If the file doesn't support async, just async punt */
3494 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3497 /* file path doesn't support NOWAIT for non-direct_IO */
3498 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3499 (req->flags & REQ_F_ISREG))
3502 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3507 * Open-code file_start_write here to grab freeze protection,
3508 * which will be released by another thread in
3509 * io_complete_rw(). Fool lockdep by telling it the lock got
3510 * released so that it doesn't complain about the held lock when
3511 * we return to userspace.
3513 if (req->flags & REQ_F_ISREG) {
3514 __sb_start_write(file_inode(req->file)->i_sb,
3515 SB_FREEZE_WRITE, true);
3516 __sb_writers_release(file_inode(req->file)->i_sb,
3519 kiocb->ki_flags |= IOCB_WRITE;
3521 if (req->file->f_op->write_iter)
3522 ret2 = call_write_iter(req->file, kiocb, iter);
3523 else if (req->file->f_op->write)
3524 ret2 = loop_rw_iter(WRITE, req, iter);
3529 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3530 * retry them without IOCB_NOWAIT.
3532 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3534 /* no retry on NONBLOCK marked file */
3535 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3537 if (!force_nonblock || ret2 != -EAGAIN) {
3538 /* IOPOLL retry should happen for io-wq threads */
3539 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3542 kiocb_done(kiocb, ret2, cs);
3545 /* some cases will consume bytes even on error returns */
3546 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3547 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3552 /* it's reportedly faster than delegating the null check to kfree() */
3558 static int __io_splice_prep(struct io_kiocb *req,
3559 const struct io_uring_sqe *sqe)
3561 struct io_splice* sp = &req->splice;
3562 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3564 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3568 sp->len = READ_ONCE(sqe->len);
3569 sp->flags = READ_ONCE(sqe->splice_flags);
3571 if (unlikely(sp->flags & ~valid_flags))
3574 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3575 (sp->flags & SPLICE_F_FD_IN_FIXED));
3578 req->flags |= REQ_F_NEED_CLEANUP;
3580 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3582 * Splice operation will be punted aync, and here need to
3583 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3585 io_req_init_async(req);
3586 req->work.flags |= IO_WQ_WORK_UNBOUND;
3592 static int io_tee_prep(struct io_kiocb *req,
3593 const struct io_uring_sqe *sqe)
3595 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3597 return __io_splice_prep(req, sqe);
3600 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3602 struct io_splice *sp = &req->splice;
3603 struct file *in = sp->file_in;
3604 struct file *out = sp->file_out;
3605 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3611 ret = do_tee(in, out, sp->len, flags);
3613 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3614 req->flags &= ~REQ_F_NEED_CLEANUP;
3617 req_set_fail_links(req);
3618 io_req_complete(req, ret);
3622 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3624 struct io_splice* sp = &req->splice;
3626 sp->off_in = READ_ONCE(sqe->splice_off_in);
3627 sp->off_out = READ_ONCE(sqe->off);
3628 return __io_splice_prep(req, sqe);
3631 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3633 struct io_splice *sp = &req->splice;
3634 struct file *in = sp->file_in;
3635 struct file *out = sp->file_out;
3636 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3637 loff_t *poff_in, *poff_out;
3643 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3644 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3647 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3649 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3650 req->flags &= ~REQ_F_NEED_CLEANUP;
3653 req_set_fail_links(req);
3654 io_req_complete(req, ret);
3659 * IORING_OP_NOP just posts a completion event, nothing else.
3661 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3663 struct io_ring_ctx *ctx = req->ctx;
3665 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3668 __io_req_complete(req, 0, 0, cs);
3672 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3674 struct io_ring_ctx *ctx = req->ctx;
3679 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3681 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3684 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3685 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3688 req->sync.off = READ_ONCE(sqe->off);
3689 req->sync.len = READ_ONCE(sqe->len);
3693 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3695 loff_t end = req->sync.off + req->sync.len;
3698 /* fsync always requires a blocking context */
3702 ret = vfs_fsync_range(req->file, req->sync.off,
3703 end > 0 ? end : LLONG_MAX,
3704 req->sync.flags & IORING_FSYNC_DATASYNC);
3706 req_set_fail_links(req);
3707 io_req_complete(req, ret);
3711 static int io_fallocate_prep(struct io_kiocb *req,
3712 const struct io_uring_sqe *sqe)
3714 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3716 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3719 req->sync.off = READ_ONCE(sqe->off);
3720 req->sync.len = READ_ONCE(sqe->addr);
3721 req->sync.mode = READ_ONCE(sqe->len);
3725 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3729 /* fallocate always requiring blocking context */
3732 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3735 req_set_fail_links(req);
3736 io_req_complete(req, ret);
3740 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3742 const char __user *fname;
3745 if (unlikely(sqe->ioprio || sqe->buf_index))
3747 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3750 /* open.how should be already initialised */
3751 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3752 req->open.how.flags |= O_LARGEFILE;
3754 req->open.dfd = READ_ONCE(sqe->fd);
3755 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3756 req->open.filename = getname(fname);
3757 if (IS_ERR(req->open.filename)) {
3758 ret = PTR_ERR(req->open.filename);
3759 req->open.filename = NULL;
3762 req->open.nofile = rlimit(RLIMIT_NOFILE);
3763 req->flags |= REQ_F_NEED_CLEANUP;
3767 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3771 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3773 mode = READ_ONCE(sqe->len);
3774 flags = READ_ONCE(sqe->open_flags);
3775 req->open.how = build_open_how(flags, mode);
3776 return __io_openat_prep(req, sqe);
3779 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3781 struct open_how __user *how;
3785 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3787 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3788 len = READ_ONCE(sqe->len);
3789 if (len < OPEN_HOW_SIZE_VER0)
3792 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3797 return __io_openat_prep(req, sqe);
3800 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3802 struct open_flags op;
3809 ret = build_open_flags(&req->open.how, &op);
3813 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3817 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3820 ret = PTR_ERR(file);
3822 fsnotify_open(file);
3823 fd_install(ret, file);
3826 putname(req->open.filename);
3827 req->flags &= ~REQ_F_NEED_CLEANUP;
3829 req_set_fail_links(req);
3830 io_req_complete(req, ret);
3834 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3836 return io_openat2(req, force_nonblock);
3839 static int io_remove_buffers_prep(struct io_kiocb *req,
3840 const struct io_uring_sqe *sqe)
3842 struct io_provide_buf *p = &req->pbuf;
3845 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3848 tmp = READ_ONCE(sqe->fd);
3849 if (!tmp || tmp > USHRT_MAX)
3852 memset(p, 0, sizeof(*p));
3854 p->bgid = READ_ONCE(sqe->buf_group);
3858 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3859 int bgid, unsigned nbufs)
3863 /* shouldn't happen */
3867 /* the head kbuf is the list itself */
3868 while (!list_empty(&buf->list)) {
3869 struct io_buffer *nxt;
3871 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3872 list_del(&nxt->list);
3879 idr_remove(&ctx->io_buffer_idr, bgid);
3884 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3885 struct io_comp_state *cs)
3887 struct io_provide_buf *p = &req->pbuf;
3888 struct io_ring_ctx *ctx = req->ctx;
3889 struct io_buffer *head;
3892 io_ring_submit_lock(ctx, !force_nonblock);
3894 lockdep_assert_held(&ctx->uring_lock);
3897 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3899 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3901 io_ring_submit_lock(ctx, !force_nonblock);
3903 req_set_fail_links(req);
3904 __io_req_complete(req, ret, 0, cs);
3908 static int io_provide_buffers_prep(struct io_kiocb *req,
3909 const struct io_uring_sqe *sqe)
3911 struct io_provide_buf *p = &req->pbuf;
3914 if (sqe->ioprio || sqe->rw_flags)
3917 tmp = READ_ONCE(sqe->fd);
3918 if (!tmp || tmp > USHRT_MAX)
3921 p->addr = READ_ONCE(sqe->addr);
3922 p->len = READ_ONCE(sqe->len);
3924 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3927 p->bgid = READ_ONCE(sqe->buf_group);
3928 tmp = READ_ONCE(sqe->off);
3929 if (tmp > USHRT_MAX)
3935 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3937 struct io_buffer *buf;
3938 u64 addr = pbuf->addr;
3939 int i, bid = pbuf->bid;
3941 for (i = 0; i < pbuf->nbufs; i++) {
3942 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3947 buf->len = pbuf->len;
3952 INIT_LIST_HEAD(&buf->list);
3955 list_add_tail(&buf->list, &(*head)->list);
3959 return i ? i : -ENOMEM;
3962 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3963 struct io_comp_state *cs)
3965 struct io_provide_buf *p = &req->pbuf;
3966 struct io_ring_ctx *ctx = req->ctx;
3967 struct io_buffer *head, *list;
3970 io_ring_submit_lock(ctx, !force_nonblock);
3972 lockdep_assert_held(&ctx->uring_lock);
3974 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3976 ret = io_add_buffers(p, &head);
3981 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3984 __io_remove_buffers(ctx, head, p->bgid, -1U);
3989 io_ring_submit_unlock(ctx, !force_nonblock);
3991 req_set_fail_links(req);
3992 __io_req_complete(req, ret, 0, cs);
3996 static int io_epoll_ctl_prep(struct io_kiocb *req,
3997 const struct io_uring_sqe *sqe)
3999 #if defined(CONFIG_EPOLL)
4000 if (sqe->ioprio || sqe->buf_index)
4002 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4005 req->epoll.epfd = READ_ONCE(sqe->fd);
4006 req->epoll.op = READ_ONCE(sqe->len);
4007 req->epoll.fd = READ_ONCE(sqe->off);
4009 if (ep_op_has_event(req->epoll.op)) {
4010 struct epoll_event __user *ev;
4012 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4013 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4023 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4024 struct io_comp_state *cs)
4026 #if defined(CONFIG_EPOLL)
4027 struct io_epoll *ie = &req->epoll;
4030 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4031 if (force_nonblock && ret == -EAGAIN)
4035 req_set_fail_links(req);
4036 __io_req_complete(req, ret, 0, cs);
4043 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4045 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4046 if (sqe->ioprio || sqe->buf_index || sqe->off)
4048 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4051 req->madvise.addr = READ_ONCE(sqe->addr);
4052 req->madvise.len = READ_ONCE(sqe->len);
4053 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4060 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4062 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4063 struct io_madvise *ma = &req->madvise;
4069 ret = do_madvise(ma->addr, ma->len, ma->advice);
4071 req_set_fail_links(req);
4072 io_req_complete(req, ret);
4079 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4081 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4083 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4086 req->fadvise.offset = READ_ONCE(sqe->off);
4087 req->fadvise.len = READ_ONCE(sqe->len);
4088 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4092 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4094 struct io_fadvise *fa = &req->fadvise;
4097 if (force_nonblock) {
4098 switch (fa->advice) {
4099 case POSIX_FADV_NORMAL:
4100 case POSIX_FADV_RANDOM:
4101 case POSIX_FADV_SEQUENTIAL:
4108 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4110 req_set_fail_links(req);
4111 io_req_complete(req, ret);
4115 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4117 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4119 if (sqe->ioprio || sqe->buf_index)
4121 if (req->flags & REQ_F_FIXED_FILE)
4124 req->statx.dfd = READ_ONCE(sqe->fd);
4125 req->statx.mask = READ_ONCE(sqe->len);
4126 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4127 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4128 req->statx.flags = READ_ONCE(sqe->statx_flags);
4133 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4135 struct io_statx *ctx = &req->statx;
4138 if (force_nonblock) {
4139 /* only need file table for an actual valid fd */
4140 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4141 req->flags |= REQ_F_NO_FILE_TABLE;
4145 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4149 req_set_fail_links(req);
4150 io_req_complete(req, ret);
4154 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4157 * If we queue this for async, it must not be cancellable. That would
4158 * leave the 'file' in an undeterminate state, and here need to modify
4159 * io_wq_work.flags, so initialize io_wq_work firstly.
4161 io_req_init_async(req);
4162 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4164 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4166 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4167 sqe->rw_flags || sqe->buf_index)
4169 if (req->flags & REQ_F_FIXED_FILE)
4172 req->close.fd = READ_ONCE(sqe->fd);
4173 if ((req->file && req->file->f_op == &io_uring_fops))
4176 req->close.put_file = NULL;
4180 static int io_close(struct io_kiocb *req, bool force_nonblock,
4181 struct io_comp_state *cs)
4183 struct io_close *close = &req->close;
4186 /* might be already done during nonblock submission */
4187 if (!close->put_file) {
4188 ret = __close_fd_get_file(close->fd, &close->put_file);
4190 return (ret == -ENOENT) ? -EBADF : ret;
4193 /* if the file has a flush method, be safe and punt to async */
4194 if (close->put_file->f_op->flush && force_nonblock) {
4195 /* was never set, but play safe */
4196 req->flags &= ~REQ_F_NOWAIT;
4197 /* avoid grabbing files - we don't need the files */
4198 req->flags |= REQ_F_NO_FILE_TABLE;
4202 /* No ->flush() or already async, safely close from here */
4203 ret = filp_close(close->put_file, req->work.identity->files);
4205 req_set_fail_links(req);
4206 fput(close->put_file);
4207 close->put_file = NULL;
4208 __io_req_complete(req, ret, 0, cs);
4212 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4214 struct io_ring_ctx *ctx = req->ctx;
4219 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4221 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4224 req->sync.off = READ_ONCE(sqe->off);
4225 req->sync.len = READ_ONCE(sqe->len);
4226 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4230 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4234 /* sync_file_range always requires a blocking context */
4238 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4241 req_set_fail_links(req);
4242 io_req_complete(req, ret);
4246 #if defined(CONFIG_NET)
4247 static int io_setup_async_msg(struct io_kiocb *req,
4248 struct io_async_msghdr *kmsg)
4250 struct io_async_msghdr *async_msg = req->async_data;
4254 if (io_alloc_async_data(req)) {
4255 if (kmsg->iov != kmsg->fast_iov)
4259 async_msg = req->async_data;
4260 req->flags |= REQ_F_NEED_CLEANUP;
4261 memcpy(async_msg, kmsg, sizeof(*kmsg));
4265 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4266 struct io_async_msghdr *iomsg)
4268 iomsg->iov = iomsg->fast_iov;
4269 iomsg->msg.msg_name = &iomsg->addr;
4270 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4271 req->sr_msg.msg_flags, &iomsg->iov);
4274 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4276 struct io_async_msghdr *async_msg = req->async_data;
4277 struct io_sr_msg *sr = &req->sr_msg;
4280 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4283 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4284 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4285 sr->len = READ_ONCE(sqe->len);
4287 #ifdef CONFIG_COMPAT
4288 if (req->ctx->compat)
4289 sr->msg_flags |= MSG_CMSG_COMPAT;
4292 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4294 ret = io_sendmsg_copy_hdr(req, async_msg);
4296 req->flags |= REQ_F_NEED_CLEANUP;
4300 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4301 struct io_comp_state *cs)
4303 struct io_async_msghdr iomsg, *kmsg;
4304 struct socket *sock;
4308 sock = sock_from_file(req->file, &ret);
4309 if (unlikely(!sock))
4312 if (req->async_data) {
4313 kmsg = req->async_data;
4314 kmsg->msg.msg_name = &kmsg->addr;
4315 /* if iov is set, it's allocated already */
4317 kmsg->iov = kmsg->fast_iov;
4318 kmsg->msg.msg_iter.iov = kmsg->iov;
4320 ret = io_sendmsg_copy_hdr(req, &iomsg);
4326 flags = req->sr_msg.msg_flags;
4327 if (flags & MSG_DONTWAIT)
4328 req->flags |= REQ_F_NOWAIT;
4329 else if (force_nonblock)
4330 flags |= MSG_DONTWAIT;
4332 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4333 if (force_nonblock && ret == -EAGAIN)
4334 return io_setup_async_msg(req, kmsg);
4335 if (ret == -ERESTARTSYS)
4338 if (kmsg->iov != kmsg->fast_iov)
4340 req->flags &= ~REQ_F_NEED_CLEANUP;
4342 req_set_fail_links(req);
4343 __io_req_complete(req, ret, 0, cs);
4347 static int io_send(struct io_kiocb *req, bool force_nonblock,
4348 struct io_comp_state *cs)
4350 struct io_sr_msg *sr = &req->sr_msg;
4353 struct socket *sock;
4357 sock = sock_from_file(req->file, &ret);
4358 if (unlikely(!sock))
4361 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4365 msg.msg_name = NULL;
4366 msg.msg_control = NULL;
4367 msg.msg_controllen = 0;
4368 msg.msg_namelen = 0;
4370 flags = req->sr_msg.msg_flags;
4371 if (flags & MSG_DONTWAIT)
4372 req->flags |= REQ_F_NOWAIT;
4373 else if (force_nonblock)
4374 flags |= MSG_DONTWAIT;
4376 msg.msg_flags = flags;
4377 ret = sock_sendmsg(sock, &msg);
4378 if (force_nonblock && ret == -EAGAIN)
4380 if (ret == -ERESTARTSYS)
4384 req_set_fail_links(req);
4385 __io_req_complete(req, ret, 0, cs);
4389 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4390 struct io_async_msghdr *iomsg)
4392 struct io_sr_msg *sr = &req->sr_msg;
4393 struct iovec __user *uiov;
4397 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4398 &iomsg->uaddr, &uiov, &iov_len);
4402 if (req->flags & REQ_F_BUFFER_SELECT) {
4405 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4407 sr->len = iomsg->iov[0].iov_len;
4408 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4412 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4413 &iomsg->iov, &iomsg->msg.msg_iter,
4422 #ifdef CONFIG_COMPAT
4423 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4424 struct io_async_msghdr *iomsg)
4426 struct compat_msghdr __user *msg_compat;
4427 struct io_sr_msg *sr = &req->sr_msg;
4428 struct compat_iovec __user *uiov;
4433 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4434 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4439 uiov = compat_ptr(ptr);
4440 if (req->flags & REQ_F_BUFFER_SELECT) {
4441 compat_ssize_t clen;
4445 if (!access_ok(uiov, sizeof(*uiov)))
4447 if (__get_user(clen, &uiov->iov_len))
4451 sr->len = iomsg->iov[0].iov_len;
4454 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4455 UIO_FASTIOV, &iomsg->iov,
4456 &iomsg->msg.msg_iter, true);
4465 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4466 struct io_async_msghdr *iomsg)
4468 iomsg->msg.msg_name = &iomsg->addr;
4469 iomsg->iov = iomsg->fast_iov;
4471 #ifdef CONFIG_COMPAT
4472 if (req->ctx->compat)
4473 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4476 return __io_recvmsg_copy_hdr(req, iomsg);
4479 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4482 struct io_sr_msg *sr = &req->sr_msg;
4483 struct io_buffer *kbuf;
4485 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4490 req->flags |= REQ_F_BUFFER_SELECTED;
4494 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4496 return io_put_kbuf(req, req->sr_msg.kbuf);
4499 static int io_recvmsg_prep(struct io_kiocb *req,
4500 const struct io_uring_sqe *sqe)
4502 struct io_async_msghdr *async_msg = req->async_data;
4503 struct io_sr_msg *sr = &req->sr_msg;
4506 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4509 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4510 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4511 sr->len = READ_ONCE(sqe->len);
4512 sr->bgid = READ_ONCE(sqe->buf_group);
4514 #ifdef CONFIG_COMPAT
4515 if (req->ctx->compat)
4516 sr->msg_flags |= MSG_CMSG_COMPAT;
4519 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4521 ret = io_recvmsg_copy_hdr(req, async_msg);
4523 req->flags |= REQ_F_NEED_CLEANUP;
4527 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4528 struct io_comp_state *cs)
4530 struct io_async_msghdr iomsg, *kmsg;
4531 struct socket *sock;
4532 struct io_buffer *kbuf;
4534 int ret, cflags = 0;
4536 sock = sock_from_file(req->file, &ret);
4537 if (unlikely(!sock))
4540 if (req->async_data) {
4541 kmsg = req->async_data;
4542 kmsg->msg.msg_name = &kmsg->addr;
4543 /* if iov is set, it's allocated already */
4545 kmsg->iov = kmsg->fast_iov;
4546 kmsg->msg.msg_iter.iov = kmsg->iov;
4548 ret = io_recvmsg_copy_hdr(req, &iomsg);
4554 if (req->flags & REQ_F_BUFFER_SELECT) {
4555 kbuf = io_recv_buffer_select(req, !force_nonblock);
4557 return PTR_ERR(kbuf);
4558 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4559 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4560 1, req->sr_msg.len);
4563 flags = req->sr_msg.msg_flags;
4564 if (flags & MSG_DONTWAIT)
4565 req->flags |= REQ_F_NOWAIT;
4566 else if (force_nonblock)
4567 flags |= MSG_DONTWAIT;
4569 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4570 kmsg->uaddr, flags);
4571 if (force_nonblock && ret == -EAGAIN)
4572 return io_setup_async_msg(req, kmsg);
4573 if (ret == -ERESTARTSYS)
4576 if (req->flags & REQ_F_BUFFER_SELECTED)
4577 cflags = io_put_recv_kbuf(req);
4578 if (kmsg->iov != kmsg->fast_iov)
4580 req->flags &= ~REQ_F_NEED_CLEANUP;
4582 req_set_fail_links(req);
4583 __io_req_complete(req, ret, cflags, cs);
4587 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4588 struct io_comp_state *cs)
4590 struct io_buffer *kbuf;
4591 struct io_sr_msg *sr = &req->sr_msg;
4593 void __user *buf = sr->buf;
4594 struct socket *sock;
4597 int ret, cflags = 0;
4599 sock = sock_from_file(req->file, &ret);
4600 if (unlikely(!sock))
4603 if (req->flags & REQ_F_BUFFER_SELECT) {
4604 kbuf = io_recv_buffer_select(req, !force_nonblock);
4606 return PTR_ERR(kbuf);
4607 buf = u64_to_user_ptr(kbuf->addr);
4610 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4614 msg.msg_name = NULL;
4615 msg.msg_control = NULL;
4616 msg.msg_controllen = 0;
4617 msg.msg_namelen = 0;
4618 msg.msg_iocb = NULL;
4621 flags = req->sr_msg.msg_flags;
4622 if (flags & MSG_DONTWAIT)
4623 req->flags |= REQ_F_NOWAIT;
4624 else if (force_nonblock)
4625 flags |= MSG_DONTWAIT;
4627 ret = sock_recvmsg(sock, &msg, flags);
4628 if (force_nonblock && ret == -EAGAIN)
4630 if (ret == -ERESTARTSYS)
4633 if (req->flags & REQ_F_BUFFER_SELECTED)
4634 cflags = io_put_recv_kbuf(req);
4636 req_set_fail_links(req);
4637 __io_req_complete(req, ret, cflags, cs);
4641 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4643 struct io_accept *accept = &req->accept;
4645 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4647 if (sqe->ioprio || sqe->len || sqe->buf_index)
4650 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4651 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4652 accept->flags = READ_ONCE(sqe->accept_flags);
4653 accept->nofile = rlimit(RLIMIT_NOFILE);
4657 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4658 struct io_comp_state *cs)
4660 struct io_accept *accept = &req->accept;
4661 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4664 if (req->file->f_flags & O_NONBLOCK)
4665 req->flags |= REQ_F_NOWAIT;
4667 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4668 accept->addr_len, accept->flags,
4670 if (ret == -EAGAIN && force_nonblock)
4673 if (ret == -ERESTARTSYS)
4675 req_set_fail_links(req);
4677 __io_req_complete(req, ret, 0, cs);
4681 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4683 struct io_connect *conn = &req->connect;
4684 struct io_async_connect *io = req->async_data;
4686 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4688 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4691 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4692 conn->addr_len = READ_ONCE(sqe->addr2);
4697 return move_addr_to_kernel(conn->addr, conn->addr_len,
4701 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4702 struct io_comp_state *cs)
4704 struct io_async_connect __io, *io;
4705 unsigned file_flags;
4708 if (req->async_data) {
4709 io = req->async_data;
4711 ret = move_addr_to_kernel(req->connect.addr,
4712 req->connect.addr_len,
4719 file_flags = force_nonblock ? O_NONBLOCK : 0;
4721 ret = __sys_connect_file(req->file, &io->address,
4722 req->connect.addr_len, file_flags);
4723 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4724 if (req->async_data)
4726 if (io_alloc_async_data(req)) {
4730 io = req->async_data;
4731 memcpy(req->async_data, &__io, sizeof(__io));
4734 if (ret == -ERESTARTSYS)
4738 req_set_fail_links(req);
4739 __io_req_complete(req, ret, 0, cs);
4742 #else /* !CONFIG_NET */
4743 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4748 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4749 struct io_comp_state *cs)
4754 static int io_send(struct io_kiocb *req, bool force_nonblock,
4755 struct io_comp_state *cs)
4760 static int io_recvmsg_prep(struct io_kiocb *req,
4761 const struct io_uring_sqe *sqe)
4766 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4767 struct io_comp_state *cs)
4772 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4773 struct io_comp_state *cs)
4778 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4783 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4784 struct io_comp_state *cs)
4789 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4794 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4795 struct io_comp_state *cs)
4799 #endif /* CONFIG_NET */
4801 struct io_poll_table {
4802 struct poll_table_struct pt;
4803 struct io_kiocb *req;
4807 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4808 __poll_t mask, task_work_func_t func)
4813 /* for instances that support it check for an event match first: */
4814 if (mask && !(mask & poll->events))
4817 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4819 list_del_init(&poll->wait.entry);
4822 init_task_work(&req->task_work, func);
4823 percpu_ref_get(&req->ctx->refs);
4826 * If we using the signalfd wait_queue_head for this wakeup, then
4827 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4828 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4829 * either, as the normal wakeup will suffice.
4831 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4834 * If this fails, then the task is exiting. When a task exits, the
4835 * work gets canceled, so just cancel this request as well instead
4836 * of executing it. We can't safely execute it anyway, as we may not
4837 * have the needed state needed for it anyway.
4839 ret = io_req_task_work_add(req, twa_signal_ok);
4840 if (unlikely(ret)) {
4841 struct task_struct *tsk;
4843 WRITE_ONCE(poll->canceled, true);
4844 tsk = io_wq_get_task(req->ctx->io_wq);
4845 task_work_add(tsk, &req->task_work, 0);
4846 wake_up_process(tsk);
4851 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4852 __acquires(&req->ctx->completion_lock)
4854 struct io_ring_ctx *ctx = req->ctx;
4856 if (!req->result && !READ_ONCE(poll->canceled)) {
4857 struct poll_table_struct pt = { ._key = poll->events };
4859 req->result = vfs_poll(req->file, &pt) & poll->events;
4862 spin_lock_irq(&ctx->completion_lock);
4863 if (!req->result && !READ_ONCE(poll->canceled)) {
4864 add_wait_queue(poll->head, &poll->wait);
4871 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4873 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4874 if (req->opcode == IORING_OP_POLL_ADD)
4875 return req->async_data;
4876 return req->apoll->double_poll;
4879 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4881 if (req->opcode == IORING_OP_POLL_ADD)
4883 return &req->apoll->poll;
4886 static void io_poll_remove_double(struct io_kiocb *req)
4888 struct io_poll_iocb *poll = io_poll_get_double(req);
4890 lockdep_assert_held(&req->ctx->completion_lock);
4892 if (poll && poll->head) {
4893 struct wait_queue_head *head = poll->head;
4895 spin_lock(&head->lock);
4896 list_del_init(&poll->wait.entry);
4897 if (poll->wait.private)
4898 refcount_dec(&req->refs);
4900 spin_unlock(&head->lock);
4904 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4906 struct io_ring_ctx *ctx = req->ctx;
4908 io_poll_remove_double(req);
4909 req->poll.done = true;
4910 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4911 io_commit_cqring(ctx);
4914 static void io_poll_task_func(struct callback_head *cb)
4916 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4917 struct io_ring_ctx *ctx = req->ctx;
4918 struct io_kiocb *nxt;
4920 if (io_poll_rewait(req, &req->poll)) {
4921 spin_unlock_irq(&ctx->completion_lock);
4923 hash_del(&req->hash_node);
4924 io_poll_complete(req, req->result, 0);
4925 spin_unlock_irq(&ctx->completion_lock);
4927 nxt = io_put_req_find_next(req);
4928 io_cqring_ev_posted(ctx);
4930 __io_req_task_submit(nxt);
4933 percpu_ref_put(&ctx->refs);
4936 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4937 int sync, void *key)
4939 struct io_kiocb *req = wait->private;
4940 struct io_poll_iocb *poll = io_poll_get_single(req);
4941 __poll_t mask = key_to_poll(key);
4943 /* for instances that support it check for an event match first: */
4944 if (mask && !(mask & poll->events))
4947 list_del_init(&wait->entry);
4949 if (poll && poll->head) {
4952 spin_lock(&poll->head->lock);
4953 done = list_empty(&poll->wait.entry);
4955 list_del_init(&poll->wait.entry);
4956 /* make sure double remove sees this as being gone */
4957 wait->private = NULL;
4958 spin_unlock(&poll->head->lock);
4960 __io_async_wake(req, poll, mask, io_poll_task_func);
4962 refcount_dec(&req->refs);
4966 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4967 wait_queue_func_t wake_func)
4971 poll->canceled = false;
4972 poll->events = events;
4973 INIT_LIST_HEAD(&poll->wait.entry);
4974 init_waitqueue_func_entry(&poll->wait, wake_func);
4977 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4978 struct wait_queue_head *head,
4979 struct io_poll_iocb **poll_ptr)
4981 struct io_kiocb *req = pt->req;
4984 * If poll->head is already set, it's because the file being polled
4985 * uses multiple waitqueues for poll handling (eg one for read, one
4986 * for write). Setup a separate io_poll_iocb if this happens.
4988 if (unlikely(poll->head)) {
4989 struct io_poll_iocb *poll_one = poll;
4991 /* already have a 2nd entry, fail a third attempt */
4993 pt->error = -EINVAL;
4996 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4998 pt->error = -ENOMEM;
5001 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5002 refcount_inc(&req->refs);
5003 poll->wait.private = req;
5010 if (poll->events & EPOLLEXCLUSIVE)
5011 add_wait_queue_exclusive(head, &poll->wait);
5013 add_wait_queue(head, &poll->wait);
5016 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5017 struct poll_table_struct *p)
5019 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5020 struct async_poll *apoll = pt->req->apoll;
5022 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5025 static void io_async_task_func(struct callback_head *cb)
5027 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5028 struct async_poll *apoll = req->apoll;
5029 struct io_ring_ctx *ctx = req->ctx;
5031 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5033 if (io_poll_rewait(req, &apoll->poll)) {
5034 spin_unlock_irq(&ctx->completion_lock);
5035 percpu_ref_put(&ctx->refs);
5039 /* If req is still hashed, it cannot have been canceled. Don't check. */
5040 if (hash_hashed(&req->hash_node))
5041 hash_del(&req->hash_node);
5043 io_poll_remove_double(req);
5044 spin_unlock_irq(&ctx->completion_lock);
5046 if (!READ_ONCE(apoll->poll.canceled))
5047 __io_req_task_submit(req);
5049 __io_req_task_cancel(req, -ECANCELED);
5051 percpu_ref_put(&ctx->refs);
5052 kfree(apoll->double_poll);
5056 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5059 struct io_kiocb *req = wait->private;
5060 struct io_poll_iocb *poll = &req->apoll->poll;
5062 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5065 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5068 static void io_poll_req_insert(struct io_kiocb *req)
5070 struct io_ring_ctx *ctx = req->ctx;
5071 struct hlist_head *list;
5073 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5074 hlist_add_head(&req->hash_node, list);
5077 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5078 struct io_poll_iocb *poll,
5079 struct io_poll_table *ipt, __poll_t mask,
5080 wait_queue_func_t wake_func)
5081 __acquires(&ctx->completion_lock)
5083 struct io_ring_ctx *ctx = req->ctx;
5084 bool cancel = false;
5086 INIT_HLIST_NODE(&req->hash_node);
5087 io_init_poll_iocb(poll, mask, wake_func);
5088 poll->file = req->file;
5089 poll->wait.private = req;
5091 ipt->pt._key = mask;
5093 ipt->error = -EINVAL;
5095 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5097 spin_lock_irq(&ctx->completion_lock);
5098 if (likely(poll->head)) {
5099 spin_lock(&poll->head->lock);
5100 if (unlikely(list_empty(&poll->wait.entry))) {
5106 if (mask || ipt->error)
5107 list_del_init(&poll->wait.entry);
5109 WRITE_ONCE(poll->canceled, true);
5110 else if (!poll->done) /* actually waiting for an event */
5111 io_poll_req_insert(req);
5112 spin_unlock(&poll->head->lock);
5118 static bool io_arm_poll_handler(struct io_kiocb *req)
5120 const struct io_op_def *def = &io_op_defs[req->opcode];
5121 struct io_ring_ctx *ctx = req->ctx;
5122 struct async_poll *apoll;
5123 struct io_poll_table ipt;
5127 if (!req->file || !file_can_poll(req->file))
5129 if (req->flags & REQ_F_POLLED)
5133 else if (def->pollout)
5137 /* if we can't nonblock try, then no point in arming a poll handler */
5138 if (!io_file_supports_async(req->file, rw))
5141 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5142 if (unlikely(!apoll))
5144 apoll->double_poll = NULL;
5146 req->flags |= REQ_F_POLLED;
5151 mask |= POLLIN | POLLRDNORM;
5153 mask |= POLLOUT | POLLWRNORM;
5155 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5156 if ((req->opcode == IORING_OP_RECVMSG) &&
5157 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5160 mask |= POLLERR | POLLPRI;
5162 ipt.pt._qproc = io_async_queue_proc;
5164 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5166 if (ret || ipt.error) {
5167 io_poll_remove_double(req);
5168 spin_unlock_irq(&ctx->completion_lock);
5169 kfree(apoll->double_poll);
5173 spin_unlock_irq(&ctx->completion_lock);
5174 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5175 apoll->poll.events);
5179 static bool __io_poll_remove_one(struct io_kiocb *req,
5180 struct io_poll_iocb *poll)
5182 bool do_complete = false;
5184 spin_lock(&poll->head->lock);
5185 WRITE_ONCE(poll->canceled, true);
5186 if (!list_empty(&poll->wait.entry)) {
5187 list_del_init(&poll->wait.entry);
5190 spin_unlock(&poll->head->lock);
5191 hash_del(&req->hash_node);
5195 static bool io_poll_remove_one(struct io_kiocb *req)
5199 io_poll_remove_double(req);
5201 if (req->opcode == IORING_OP_POLL_ADD) {
5202 do_complete = __io_poll_remove_one(req, &req->poll);
5204 struct async_poll *apoll = req->apoll;
5206 /* non-poll requests have submit ref still */
5207 do_complete = __io_poll_remove_one(req, &apoll->poll);
5210 kfree(apoll->double_poll);
5216 io_cqring_fill_event(req, -ECANCELED);
5217 io_commit_cqring(req->ctx);
5218 req_set_fail_links(req);
5219 io_put_req_deferred(req, 1);
5226 * Returns true if we found and killed one or more poll requests
5228 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5230 struct hlist_node *tmp;
5231 struct io_kiocb *req;
5234 spin_lock_irq(&ctx->completion_lock);
5235 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5236 struct hlist_head *list;
5238 list = &ctx->cancel_hash[i];
5239 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5240 if (io_task_match(req, tsk))
5241 posted += io_poll_remove_one(req);
5244 spin_unlock_irq(&ctx->completion_lock);
5247 io_cqring_ev_posted(ctx);
5252 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5254 struct hlist_head *list;
5255 struct io_kiocb *req;
5257 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5258 hlist_for_each_entry(req, list, hash_node) {
5259 if (sqe_addr != req->user_data)
5261 if (io_poll_remove_one(req))
5269 static int io_poll_remove_prep(struct io_kiocb *req,
5270 const struct io_uring_sqe *sqe)
5272 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5274 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5278 req->poll.addr = READ_ONCE(sqe->addr);
5283 * Find a running poll command that matches one specified in sqe->addr,
5284 * and remove it if found.
5286 static int io_poll_remove(struct io_kiocb *req)
5288 struct io_ring_ctx *ctx = req->ctx;
5292 addr = req->poll.addr;
5293 spin_lock_irq(&ctx->completion_lock);
5294 ret = io_poll_cancel(ctx, addr);
5295 spin_unlock_irq(&ctx->completion_lock);
5298 req_set_fail_links(req);
5299 io_req_complete(req, ret);
5303 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5306 struct io_kiocb *req = wait->private;
5307 struct io_poll_iocb *poll = &req->poll;
5309 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5312 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5313 struct poll_table_struct *p)
5315 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5317 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5320 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5322 struct io_poll_iocb *poll = &req->poll;
5325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5327 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5330 events = READ_ONCE(sqe->poll32_events);
5332 events = swahw32(events);
5334 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5335 (events & EPOLLEXCLUSIVE);
5339 static int io_poll_add(struct io_kiocb *req)
5341 struct io_poll_iocb *poll = &req->poll;
5342 struct io_ring_ctx *ctx = req->ctx;
5343 struct io_poll_table ipt;
5346 ipt.pt._qproc = io_poll_queue_proc;
5348 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5351 if (mask) { /* no async, we'd stolen it */
5353 io_poll_complete(req, mask, 0);
5355 spin_unlock_irq(&ctx->completion_lock);
5358 io_cqring_ev_posted(ctx);
5364 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5366 struct io_timeout_data *data = container_of(timer,
5367 struct io_timeout_data, timer);
5368 struct io_kiocb *req = data->req;
5369 struct io_ring_ctx *ctx = req->ctx;
5370 unsigned long flags;
5372 spin_lock_irqsave(&ctx->completion_lock, flags);
5373 list_del_init(&req->timeout.list);
5374 atomic_set(&req->ctx->cq_timeouts,
5375 atomic_read(&req->ctx->cq_timeouts) + 1);
5377 io_cqring_fill_event(req, -ETIME);
5378 io_commit_cqring(ctx);
5379 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5381 io_cqring_ev_posted(ctx);
5382 req_set_fail_links(req);
5384 return HRTIMER_NORESTART;
5387 static int __io_timeout_cancel(struct io_kiocb *req)
5389 struct io_timeout_data *io = req->async_data;
5392 ret = hrtimer_try_to_cancel(&io->timer);
5395 list_del_init(&req->timeout.list);
5397 req_set_fail_links(req);
5398 io_cqring_fill_event(req, -ECANCELED);
5399 io_put_req_deferred(req, 1);
5403 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5405 struct io_kiocb *req;
5408 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5409 if (user_data == req->user_data) {
5418 return __io_timeout_cancel(req);
5421 static int io_timeout_remove_prep(struct io_kiocb *req,
5422 const struct io_uring_sqe *sqe)
5424 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5426 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5428 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5431 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5436 * Remove or update an existing timeout command
5438 static int io_timeout_remove(struct io_kiocb *req)
5440 struct io_ring_ctx *ctx = req->ctx;
5443 spin_lock_irq(&ctx->completion_lock);
5444 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5446 io_cqring_fill_event(req, ret);
5447 io_commit_cqring(ctx);
5448 spin_unlock_irq(&ctx->completion_lock);
5449 io_cqring_ev_posted(ctx);
5451 req_set_fail_links(req);
5456 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5457 bool is_timeout_link)
5459 struct io_timeout_data *data;
5461 u32 off = READ_ONCE(sqe->off);
5463 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5465 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5467 if (off && is_timeout_link)
5469 flags = READ_ONCE(sqe->timeout_flags);
5470 if (flags & ~IORING_TIMEOUT_ABS)
5473 req->timeout.off = off;
5475 if (!req->async_data && io_alloc_async_data(req))
5478 data = req->async_data;
5481 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5484 if (flags & IORING_TIMEOUT_ABS)
5485 data->mode = HRTIMER_MODE_ABS;
5487 data->mode = HRTIMER_MODE_REL;
5489 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5493 static int io_timeout(struct io_kiocb *req)
5495 struct io_ring_ctx *ctx = req->ctx;
5496 struct io_timeout_data *data = req->async_data;
5497 struct list_head *entry;
5498 u32 tail, off = req->timeout.off;
5500 spin_lock_irq(&ctx->completion_lock);
5503 * sqe->off holds how many events that need to occur for this
5504 * timeout event to be satisfied. If it isn't set, then this is
5505 * a pure timeout request, sequence isn't used.
5507 if (io_is_timeout_noseq(req)) {
5508 entry = ctx->timeout_list.prev;
5512 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5513 req->timeout.target_seq = tail + off;
5516 * Insertion sort, ensuring the first entry in the list is always
5517 * the one we need first.
5519 list_for_each_prev(entry, &ctx->timeout_list) {
5520 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5523 if (io_is_timeout_noseq(nxt))
5525 /* nxt.seq is behind @tail, otherwise would've been completed */
5526 if (off >= nxt->timeout.target_seq - tail)
5530 list_add(&req->timeout.list, entry);
5531 data->timer.function = io_timeout_fn;
5532 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5533 spin_unlock_irq(&ctx->completion_lock);
5537 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5539 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5541 return req->user_data == (unsigned long) data;
5544 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5546 enum io_wq_cancel cancel_ret;
5549 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5550 switch (cancel_ret) {
5551 case IO_WQ_CANCEL_OK:
5554 case IO_WQ_CANCEL_RUNNING:
5557 case IO_WQ_CANCEL_NOTFOUND:
5565 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5566 struct io_kiocb *req, __u64 sqe_addr,
5569 unsigned long flags;
5572 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5573 if (ret != -ENOENT) {
5574 spin_lock_irqsave(&ctx->completion_lock, flags);
5578 spin_lock_irqsave(&ctx->completion_lock, flags);
5579 ret = io_timeout_cancel(ctx, sqe_addr);
5582 ret = io_poll_cancel(ctx, sqe_addr);
5586 io_cqring_fill_event(req, ret);
5587 io_commit_cqring(ctx);
5588 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5589 io_cqring_ev_posted(ctx);
5592 req_set_fail_links(req);
5596 static int io_async_cancel_prep(struct io_kiocb *req,
5597 const struct io_uring_sqe *sqe)
5599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5601 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5603 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5606 req->cancel.addr = READ_ONCE(sqe->addr);
5610 static int io_async_cancel(struct io_kiocb *req)
5612 struct io_ring_ctx *ctx = req->ctx;
5614 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5618 static int io_files_update_prep(struct io_kiocb *req,
5619 const struct io_uring_sqe *sqe)
5621 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5623 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5625 if (sqe->ioprio || sqe->rw_flags)
5628 req->files_update.offset = READ_ONCE(sqe->off);
5629 req->files_update.nr_args = READ_ONCE(sqe->len);
5630 if (!req->files_update.nr_args)
5632 req->files_update.arg = READ_ONCE(sqe->addr);
5636 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5637 struct io_comp_state *cs)
5639 struct io_ring_ctx *ctx = req->ctx;
5640 struct io_uring_files_update up;
5646 up.offset = req->files_update.offset;
5647 up.fds = req->files_update.arg;
5649 mutex_lock(&ctx->uring_lock);
5650 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5651 mutex_unlock(&ctx->uring_lock);
5654 req_set_fail_links(req);
5655 __io_req_complete(req, ret, 0, cs);
5659 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5661 switch (req->opcode) {
5664 case IORING_OP_READV:
5665 case IORING_OP_READ_FIXED:
5666 case IORING_OP_READ:
5667 return io_read_prep(req, sqe);
5668 case IORING_OP_WRITEV:
5669 case IORING_OP_WRITE_FIXED:
5670 case IORING_OP_WRITE:
5671 return io_write_prep(req, sqe);
5672 case IORING_OP_POLL_ADD:
5673 return io_poll_add_prep(req, sqe);
5674 case IORING_OP_POLL_REMOVE:
5675 return io_poll_remove_prep(req, sqe);
5676 case IORING_OP_FSYNC:
5677 return io_prep_fsync(req, sqe);
5678 case IORING_OP_SYNC_FILE_RANGE:
5679 return io_prep_sfr(req, sqe);
5680 case IORING_OP_SENDMSG:
5681 case IORING_OP_SEND:
5682 return io_sendmsg_prep(req, sqe);
5683 case IORING_OP_RECVMSG:
5684 case IORING_OP_RECV:
5685 return io_recvmsg_prep(req, sqe);
5686 case IORING_OP_CONNECT:
5687 return io_connect_prep(req, sqe);
5688 case IORING_OP_TIMEOUT:
5689 return io_timeout_prep(req, sqe, false);
5690 case IORING_OP_TIMEOUT_REMOVE:
5691 return io_timeout_remove_prep(req, sqe);
5692 case IORING_OP_ASYNC_CANCEL:
5693 return io_async_cancel_prep(req, sqe);
5694 case IORING_OP_LINK_TIMEOUT:
5695 return io_timeout_prep(req, sqe, true);
5696 case IORING_OP_ACCEPT:
5697 return io_accept_prep(req, sqe);
5698 case IORING_OP_FALLOCATE:
5699 return io_fallocate_prep(req, sqe);
5700 case IORING_OP_OPENAT:
5701 return io_openat_prep(req, sqe);
5702 case IORING_OP_CLOSE:
5703 return io_close_prep(req, sqe);
5704 case IORING_OP_FILES_UPDATE:
5705 return io_files_update_prep(req, sqe);
5706 case IORING_OP_STATX:
5707 return io_statx_prep(req, sqe);
5708 case IORING_OP_FADVISE:
5709 return io_fadvise_prep(req, sqe);
5710 case IORING_OP_MADVISE:
5711 return io_madvise_prep(req, sqe);
5712 case IORING_OP_OPENAT2:
5713 return io_openat2_prep(req, sqe);
5714 case IORING_OP_EPOLL_CTL:
5715 return io_epoll_ctl_prep(req, sqe);
5716 case IORING_OP_SPLICE:
5717 return io_splice_prep(req, sqe);
5718 case IORING_OP_PROVIDE_BUFFERS:
5719 return io_provide_buffers_prep(req, sqe);
5720 case IORING_OP_REMOVE_BUFFERS:
5721 return io_remove_buffers_prep(req, sqe);
5723 return io_tee_prep(req, sqe);
5726 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5731 static int io_req_defer_prep(struct io_kiocb *req,
5732 const struct io_uring_sqe *sqe)
5736 if (io_alloc_async_data(req))
5738 return io_req_prep(req, sqe);
5741 static u32 io_get_sequence(struct io_kiocb *req)
5743 struct io_kiocb *pos;
5744 struct io_ring_ctx *ctx = req->ctx;
5745 u32 total_submitted, nr_reqs = 1;
5747 if (req->flags & REQ_F_LINK_HEAD)
5748 list_for_each_entry(pos, &req->link_list, link_list)
5751 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5752 return total_submitted - nr_reqs;
5755 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5757 struct io_ring_ctx *ctx = req->ctx;
5758 struct io_defer_entry *de;
5762 /* Still need defer if there is pending req in defer list. */
5763 if (likely(list_empty_careful(&ctx->defer_list) &&
5764 !(req->flags & REQ_F_IO_DRAIN)))
5767 seq = io_get_sequence(req);
5768 /* Still a chance to pass the sequence check */
5769 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5772 if (!req->async_data) {
5773 ret = io_req_defer_prep(req, sqe);
5777 io_prep_async_link(req);
5778 de = kmalloc(sizeof(*de), GFP_KERNEL);
5782 spin_lock_irq(&ctx->completion_lock);
5783 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5784 spin_unlock_irq(&ctx->completion_lock);
5786 io_queue_async_work(req);
5787 return -EIOCBQUEUED;
5790 trace_io_uring_defer(ctx, req, req->user_data);
5793 list_add_tail(&de->list, &ctx->defer_list);
5794 spin_unlock_irq(&ctx->completion_lock);
5795 return -EIOCBQUEUED;
5798 static void io_req_drop_files(struct io_kiocb *req)
5800 struct io_ring_ctx *ctx = req->ctx;
5801 unsigned long flags;
5803 spin_lock_irqsave(&ctx->inflight_lock, flags);
5804 list_del(&req->inflight_entry);
5805 if (waitqueue_active(&ctx->inflight_wait))
5806 wake_up(&ctx->inflight_wait);
5807 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5808 req->flags &= ~REQ_F_INFLIGHT;
5809 put_files_struct(req->work.identity->files);
5810 put_nsproxy(req->work.identity->nsproxy);
5811 req->work.flags &= ~IO_WQ_WORK_FILES;
5814 static void __io_clean_op(struct io_kiocb *req)
5816 if (req->flags & REQ_F_BUFFER_SELECTED) {
5817 switch (req->opcode) {
5818 case IORING_OP_READV:
5819 case IORING_OP_READ_FIXED:
5820 case IORING_OP_READ:
5821 kfree((void *)(unsigned long)req->rw.addr);
5823 case IORING_OP_RECVMSG:
5824 case IORING_OP_RECV:
5825 kfree(req->sr_msg.kbuf);
5828 req->flags &= ~REQ_F_BUFFER_SELECTED;
5831 if (req->flags & REQ_F_NEED_CLEANUP) {
5832 switch (req->opcode) {
5833 case IORING_OP_READV:
5834 case IORING_OP_READ_FIXED:
5835 case IORING_OP_READ:
5836 case IORING_OP_WRITEV:
5837 case IORING_OP_WRITE_FIXED:
5838 case IORING_OP_WRITE: {
5839 struct io_async_rw *io = req->async_data;
5841 kfree(io->free_iovec);
5844 case IORING_OP_RECVMSG:
5845 case IORING_OP_SENDMSG: {
5846 struct io_async_msghdr *io = req->async_data;
5847 if (io->iov != io->fast_iov)
5851 case IORING_OP_SPLICE:
5853 io_put_file(req, req->splice.file_in,
5854 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5856 case IORING_OP_OPENAT:
5857 case IORING_OP_OPENAT2:
5858 if (req->open.filename)
5859 putname(req->open.filename);
5862 req->flags &= ~REQ_F_NEED_CLEANUP;
5865 if (req->flags & REQ_F_INFLIGHT)
5866 io_req_drop_files(req);
5869 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5870 struct io_comp_state *cs)
5872 struct io_ring_ctx *ctx = req->ctx;
5875 switch (req->opcode) {
5877 ret = io_nop(req, cs);
5879 case IORING_OP_READV:
5880 case IORING_OP_READ_FIXED:
5881 case IORING_OP_READ:
5882 ret = io_read(req, force_nonblock, cs);
5884 case IORING_OP_WRITEV:
5885 case IORING_OP_WRITE_FIXED:
5886 case IORING_OP_WRITE:
5887 ret = io_write(req, force_nonblock, cs);
5889 case IORING_OP_FSYNC:
5890 ret = io_fsync(req, force_nonblock);
5892 case IORING_OP_POLL_ADD:
5893 ret = io_poll_add(req);
5895 case IORING_OP_POLL_REMOVE:
5896 ret = io_poll_remove(req);
5898 case IORING_OP_SYNC_FILE_RANGE:
5899 ret = io_sync_file_range(req, force_nonblock);
5901 case IORING_OP_SENDMSG:
5902 ret = io_sendmsg(req, force_nonblock, cs);
5904 case IORING_OP_SEND:
5905 ret = io_send(req, force_nonblock, cs);
5907 case IORING_OP_RECVMSG:
5908 ret = io_recvmsg(req, force_nonblock, cs);
5910 case IORING_OP_RECV:
5911 ret = io_recv(req, force_nonblock, cs);
5913 case IORING_OP_TIMEOUT:
5914 ret = io_timeout(req);
5916 case IORING_OP_TIMEOUT_REMOVE:
5917 ret = io_timeout_remove(req);
5919 case IORING_OP_ACCEPT:
5920 ret = io_accept(req, force_nonblock, cs);
5922 case IORING_OP_CONNECT:
5923 ret = io_connect(req, force_nonblock, cs);
5925 case IORING_OP_ASYNC_CANCEL:
5926 ret = io_async_cancel(req);
5928 case IORING_OP_FALLOCATE:
5929 ret = io_fallocate(req, force_nonblock);
5931 case IORING_OP_OPENAT:
5932 ret = io_openat(req, force_nonblock);
5934 case IORING_OP_CLOSE:
5935 ret = io_close(req, force_nonblock, cs);
5937 case IORING_OP_FILES_UPDATE:
5938 ret = io_files_update(req, force_nonblock, cs);
5940 case IORING_OP_STATX:
5941 ret = io_statx(req, force_nonblock);
5943 case IORING_OP_FADVISE:
5944 ret = io_fadvise(req, force_nonblock);
5946 case IORING_OP_MADVISE:
5947 ret = io_madvise(req, force_nonblock);
5949 case IORING_OP_OPENAT2:
5950 ret = io_openat2(req, force_nonblock);
5952 case IORING_OP_EPOLL_CTL:
5953 ret = io_epoll_ctl(req, force_nonblock, cs);
5955 case IORING_OP_SPLICE:
5956 ret = io_splice(req, force_nonblock);
5958 case IORING_OP_PROVIDE_BUFFERS:
5959 ret = io_provide_buffers(req, force_nonblock, cs);
5961 case IORING_OP_REMOVE_BUFFERS:
5962 ret = io_remove_buffers(req, force_nonblock, cs);
5965 ret = io_tee(req, force_nonblock);
5975 /* If the op doesn't have a file, we're not polling for it */
5976 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5977 const bool in_async = io_wq_current_is_worker();
5979 /* workqueue context doesn't hold uring_lock, grab it now */
5981 mutex_lock(&ctx->uring_lock);
5983 io_iopoll_req_issued(req);
5986 mutex_unlock(&ctx->uring_lock);
5992 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5994 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5995 struct io_kiocb *timeout;
5998 timeout = io_prep_linked_timeout(req);
6000 io_queue_linked_timeout(timeout);
6002 /* if NO_CANCEL is set, we must still run the work */
6003 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6004 IO_WQ_WORK_CANCEL) {
6010 ret = io_issue_sqe(req, false, NULL);
6012 * We can get EAGAIN for polled IO even though we're
6013 * forcing a sync submission from here, since we can't
6014 * wait for request slots on the block side.
6023 req_set_fail_links(req);
6024 io_req_complete(req, ret);
6027 return io_steal_work(req);
6030 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6033 struct fixed_file_table *table;
6035 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6036 return table->files[index & IORING_FILE_TABLE_MASK];
6039 static struct file *io_file_get(struct io_submit_state *state,
6040 struct io_kiocb *req, int fd, bool fixed)
6042 struct io_ring_ctx *ctx = req->ctx;
6046 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6048 fd = array_index_nospec(fd, ctx->nr_user_files);
6049 file = io_file_from_index(ctx, fd);
6051 req->fixed_file_refs = &ctx->file_data->node->refs;
6052 percpu_ref_get(req->fixed_file_refs);
6055 trace_io_uring_file_get(ctx, fd);
6056 file = __io_file_get(state, fd);
6062 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6067 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6068 if (unlikely(!fixed && io_async_submit(req->ctx)))
6071 req->file = io_file_get(state, req, fd, fixed);
6072 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6077 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6079 struct io_timeout_data *data = container_of(timer,
6080 struct io_timeout_data, timer);
6081 struct io_kiocb *req = data->req;
6082 struct io_ring_ctx *ctx = req->ctx;
6083 struct io_kiocb *prev = NULL;
6084 unsigned long flags;
6086 spin_lock_irqsave(&ctx->completion_lock, flags);
6089 * We don't expect the list to be empty, that will only happen if we
6090 * race with the completion of the linked work.
6092 if (!list_empty(&req->link_list)) {
6093 prev = list_entry(req->link_list.prev, struct io_kiocb,
6095 if (refcount_inc_not_zero(&prev->refs))
6096 list_del_init(&req->link_list);
6101 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6104 req_set_fail_links(prev);
6105 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6108 io_req_complete(req, -ETIME);
6110 return HRTIMER_NORESTART;
6113 static void __io_queue_linked_timeout(struct io_kiocb *req)
6116 * If the list is now empty, then our linked request finished before
6117 * we got a chance to setup the timer
6119 if (!list_empty(&req->link_list)) {
6120 struct io_timeout_data *data = req->async_data;
6122 data->timer.function = io_link_timeout_fn;
6123 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6128 static void io_queue_linked_timeout(struct io_kiocb *req)
6130 struct io_ring_ctx *ctx = req->ctx;
6132 spin_lock_irq(&ctx->completion_lock);
6133 __io_queue_linked_timeout(req);
6134 spin_unlock_irq(&ctx->completion_lock);
6136 /* drop submission reference */
6140 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6142 struct io_kiocb *nxt;
6144 if (!(req->flags & REQ_F_LINK_HEAD))
6146 if (req->flags & REQ_F_LINK_TIMEOUT)
6149 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6151 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6154 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6155 req->flags |= REQ_F_LINK_TIMEOUT;
6159 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6161 struct io_kiocb *linked_timeout;
6162 struct io_kiocb *nxt;
6163 const struct cred *old_creds = NULL;
6167 linked_timeout = io_prep_linked_timeout(req);
6169 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6170 (req->work.flags & IO_WQ_WORK_CREDS) &&
6171 req->work.identity->creds != current_cred()) {
6173 revert_creds(old_creds);
6174 if (old_creds == req->work.identity->creds)
6175 old_creds = NULL; /* restored original creds */
6177 old_creds = override_creds(req->work.identity->creds);
6180 ret = io_issue_sqe(req, true, cs);
6183 * We async punt it if the file wasn't marked NOWAIT, or if the file
6184 * doesn't support non-blocking read/write attempts
6186 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6187 if (!io_arm_poll_handler(req)) {
6190 * Queued up for async execution, worker will release
6191 * submit reference when the iocb is actually submitted.
6193 io_queue_async_work(req);
6197 io_queue_linked_timeout(linked_timeout);
6201 if (unlikely(ret)) {
6202 /* un-prep timeout, so it'll be killed as any other linked */
6203 req->flags &= ~REQ_F_LINK_TIMEOUT;
6204 req_set_fail_links(req);
6206 io_req_complete(req, ret);
6210 /* drop submission reference */
6211 nxt = io_put_req_find_next(req);
6213 io_queue_linked_timeout(linked_timeout);
6218 if (req->flags & REQ_F_FORCE_ASYNC) {
6219 linked_timeout = NULL;
6226 revert_creds(old_creds);
6229 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6230 struct io_comp_state *cs)
6234 ret = io_req_defer(req, sqe);
6236 if (ret != -EIOCBQUEUED) {
6238 req_set_fail_links(req);
6240 io_req_complete(req, ret);
6242 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6243 if (!req->async_data) {
6244 ret = io_req_defer_prep(req, sqe);
6250 * Never try inline submit of IOSQE_ASYNC is set, go straight
6251 * to async execution.
6253 io_req_init_async(req);
6254 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6255 io_queue_async_work(req);
6258 ret = io_req_prep(req, sqe);
6262 __io_queue_sqe(req, cs);
6266 static inline void io_queue_link_head(struct io_kiocb *req,
6267 struct io_comp_state *cs)
6269 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6271 io_req_complete(req, -ECANCELED);
6273 io_queue_sqe(req, NULL, cs);
6276 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6277 struct io_kiocb **link, struct io_comp_state *cs)
6279 struct io_ring_ctx *ctx = req->ctx;
6283 * If we already have a head request, queue this one for async
6284 * submittal once the head completes. If we don't have a head but
6285 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6286 * submitted sync once the chain is complete. If none of those
6287 * conditions are true (normal request), then just queue it.
6290 struct io_kiocb *head = *link;
6293 * Taking sequential execution of a link, draining both sides
6294 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6295 * requests in the link. So, it drains the head and the
6296 * next after the link request. The last one is done via
6297 * drain_next flag to persist the effect across calls.
6299 if (req->flags & REQ_F_IO_DRAIN) {
6300 head->flags |= REQ_F_IO_DRAIN;
6301 ctx->drain_next = 1;
6303 ret = io_req_defer_prep(req, sqe);
6304 if (unlikely(ret)) {
6305 /* fail even hard links since we don't submit */
6306 head->flags |= REQ_F_FAIL_LINK;
6309 trace_io_uring_link(ctx, req, head);
6310 list_add_tail(&req->link_list, &head->link_list);
6312 /* last request of a link, enqueue the link */
6313 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6314 io_queue_link_head(head, cs);
6318 if (unlikely(ctx->drain_next)) {
6319 req->flags |= REQ_F_IO_DRAIN;
6320 ctx->drain_next = 0;
6322 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6323 req->flags |= REQ_F_LINK_HEAD;
6324 INIT_LIST_HEAD(&req->link_list);
6326 ret = io_req_defer_prep(req, sqe);
6328 req->flags |= REQ_F_FAIL_LINK;
6331 io_queue_sqe(req, sqe, cs);
6339 * Batched submission is done, ensure local IO is flushed out.
6341 static void io_submit_state_end(struct io_submit_state *state)
6343 if (!list_empty(&state->comp.list))
6344 io_submit_flush_completions(&state->comp);
6345 blk_finish_plug(&state->plug);
6346 io_state_file_put(state);
6347 if (state->free_reqs)
6348 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6352 * Start submission side cache.
6354 static void io_submit_state_start(struct io_submit_state *state,
6355 struct io_ring_ctx *ctx, unsigned int max_ios)
6357 blk_start_plug(&state->plug);
6359 INIT_LIST_HEAD(&state->comp.list);
6360 state->comp.ctx = ctx;
6361 state->free_reqs = 0;
6363 state->ios_left = max_ios;
6366 static void io_commit_sqring(struct io_ring_ctx *ctx)
6368 struct io_rings *rings = ctx->rings;
6371 * Ensure any loads from the SQEs are done at this point,
6372 * since once we write the new head, the application could
6373 * write new data to them.
6375 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6379 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6380 * that is mapped by userspace. This means that care needs to be taken to
6381 * ensure that reads are stable, as we cannot rely on userspace always
6382 * being a good citizen. If members of the sqe are validated and then later
6383 * used, it's important that those reads are done through READ_ONCE() to
6384 * prevent a re-load down the line.
6386 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6388 u32 *sq_array = ctx->sq_array;
6392 * The cached sq head (or cq tail) serves two purposes:
6394 * 1) allows us to batch the cost of updating the user visible
6396 * 2) allows the kernel side to track the head on its own, even
6397 * though the application is the one updating it.
6399 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6400 if (likely(head < ctx->sq_entries))
6401 return &ctx->sq_sqes[head];
6403 /* drop invalid entries */
6404 ctx->cached_sq_dropped++;
6405 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6409 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6411 ctx->cached_sq_head++;
6415 * Check SQE restrictions (opcode and flags).
6417 * Returns 'true' if SQE is allowed, 'false' otherwise.
6419 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6420 struct io_kiocb *req,
6421 unsigned int sqe_flags)
6423 if (!ctx->restricted)
6426 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6429 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6430 ctx->restrictions.sqe_flags_required)
6433 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6434 ctx->restrictions.sqe_flags_required))
6440 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6441 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6442 IOSQE_BUFFER_SELECT)
6444 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6445 const struct io_uring_sqe *sqe,
6446 struct io_submit_state *state)
6448 unsigned int sqe_flags;
6451 req->opcode = READ_ONCE(sqe->opcode);
6452 req->user_data = READ_ONCE(sqe->user_data);
6453 req->async_data = NULL;
6457 /* one is dropped after submission, the other at completion */
6458 refcount_set(&req->refs, 2);
6459 req->task = current;
6462 if (unlikely(req->opcode >= IORING_OP_LAST))
6465 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6468 sqe_flags = READ_ONCE(sqe->flags);
6469 /* enforce forwards compatibility on users */
6470 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6473 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6476 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6477 !io_op_defs[req->opcode].buffer_select)
6480 id = READ_ONCE(sqe->personality);
6482 struct io_identity *iod;
6484 iod = idr_find(&ctx->personality_idr, id);
6487 refcount_inc(&iod->count);
6489 __io_req_init_async(req);
6490 get_cred(iod->creds);
6491 req->work.identity = iod;
6492 req->work.flags |= IO_WQ_WORK_CREDS;
6495 /* same numerical values with corresponding REQ_F_*, safe to copy */
6496 req->flags |= sqe_flags;
6498 if (!io_op_defs[req->opcode].needs_file)
6501 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6506 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6508 struct io_submit_state state;
6509 struct io_kiocb *link = NULL;
6510 int i, submitted = 0;
6512 /* if we have a backlog and couldn't flush it all, return BUSY */
6513 if (test_bit(0, &ctx->sq_check_overflow)) {
6514 if (!list_empty(&ctx->cq_overflow_list) &&
6515 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6519 /* make sure SQ entry isn't read before tail */
6520 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6522 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6525 percpu_counter_add(¤t->io_uring->inflight, nr);
6526 refcount_add(nr, ¤t->usage);
6528 io_submit_state_start(&state, ctx, nr);
6530 for (i = 0; i < nr; i++) {
6531 const struct io_uring_sqe *sqe;
6532 struct io_kiocb *req;
6535 sqe = io_get_sqe(ctx);
6536 if (unlikely(!sqe)) {
6537 io_consume_sqe(ctx);
6540 req = io_alloc_req(ctx, &state);
6541 if (unlikely(!req)) {
6543 submitted = -EAGAIN;
6546 io_consume_sqe(ctx);
6547 /* will complete beyond this point, count as submitted */
6550 err = io_init_req(ctx, req, sqe, &state);
6551 if (unlikely(err)) {
6554 io_req_complete(req, err);
6558 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6559 true, io_async_submit(ctx));
6560 err = io_submit_sqe(req, sqe, &link, &state.comp);
6565 if (unlikely(submitted != nr)) {
6566 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6567 struct io_uring_task *tctx = current->io_uring;
6568 int unused = nr - ref_used;
6570 percpu_ref_put_many(&ctx->refs, unused);
6571 percpu_counter_sub(&tctx->inflight, unused);
6572 put_task_struct_many(current, unused);
6575 io_queue_link_head(link, &state.comp);
6576 io_submit_state_end(&state);
6578 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6579 io_commit_sqring(ctx);
6584 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6586 /* Tell userspace we may need a wakeup call */
6587 spin_lock_irq(&ctx->completion_lock);
6588 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6589 spin_unlock_irq(&ctx->completion_lock);
6592 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6594 spin_lock_irq(&ctx->completion_lock);
6595 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6596 spin_unlock_irq(&ctx->completion_lock);
6599 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6600 int sync, void *key)
6602 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6605 ret = autoremove_wake_function(wqe, mode, sync, key);
6607 unsigned long flags;
6609 spin_lock_irqsave(&ctx->completion_lock, flags);
6610 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6611 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6622 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6623 unsigned long start_jiffies, bool cap_entries)
6625 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6626 struct io_sq_data *sqd = ctx->sq_data;
6627 unsigned int to_submit;
6631 if (!list_empty(&ctx->iopoll_list)) {
6632 unsigned nr_events = 0;
6634 mutex_lock(&ctx->uring_lock);
6635 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6636 io_do_iopoll(ctx, &nr_events, 0);
6637 mutex_unlock(&ctx->uring_lock);
6640 to_submit = io_sqring_entries(ctx);
6643 * If submit got -EBUSY, flag us as needing the application
6644 * to enter the kernel to reap and flush events.
6646 if (!to_submit || ret == -EBUSY || need_resched()) {
6648 * Drop cur_mm before scheduling, we can't hold it for
6649 * long periods (or over schedule()). Do this before
6650 * adding ourselves to the waitqueue, as the unuse/drop
6653 io_sq_thread_drop_mm();
6656 * We're polling. If we're within the defined idle
6657 * period, then let us spin without work before going
6658 * to sleep. The exception is if we got EBUSY doing
6659 * more IO, we should wait for the application to
6660 * reap events and wake us up.
6662 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6663 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6664 !percpu_ref_is_dying(&ctx->refs)))
6667 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6668 TASK_INTERRUPTIBLE);
6671 * While doing polled IO, before going to sleep, we need
6672 * to check if there are new reqs added to iopoll_list,
6673 * it is because reqs may have been punted to io worker
6674 * and will be added to iopoll_list later, hence check
6675 * the iopoll_list again.
6677 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6678 !list_empty_careful(&ctx->iopoll_list)) {
6679 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6683 to_submit = io_sqring_entries(ctx);
6684 if (!to_submit || ret == -EBUSY)
6688 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6689 io_ring_clear_wakeup_flag(ctx);
6691 /* if we're handling multiple rings, cap submit size for fairness */
6692 if (cap_entries && to_submit > 8)
6695 mutex_lock(&ctx->uring_lock);
6696 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6697 ret = io_submit_sqes(ctx, to_submit);
6698 mutex_unlock(&ctx->uring_lock);
6700 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6701 wake_up(&ctx->sqo_sq_wait);
6703 return SQT_DID_WORK;
6706 static void io_sqd_init_new(struct io_sq_data *sqd)
6708 struct io_ring_ctx *ctx;
6710 while (!list_empty(&sqd->ctx_new_list)) {
6711 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6712 init_wait(&ctx->sqo_wait_entry);
6713 ctx->sqo_wait_entry.func = io_sq_wake_function;
6714 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6715 complete(&ctx->sq_thread_comp);
6719 static int io_sq_thread(void *data)
6721 struct cgroup_subsys_state *cur_css = NULL;
6722 const struct cred *old_cred = NULL;
6723 struct io_sq_data *sqd = data;
6724 struct io_ring_ctx *ctx;
6725 unsigned long start_jiffies;
6727 start_jiffies = jiffies;
6728 while (!kthread_should_stop()) {
6729 enum sq_ret ret = 0;
6733 * Any changes to the sqd lists are synchronized through the
6734 * kthread parking. This synchronizes the thread vs users,
6735 * the users are synchronized on the sqd->ctx_lock.
6737 if (kthread_should_park())
6740 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6741 io_sqd_init_new(sqd);
6743 cap_entries = !list_is_singular(&sqd->ctx_list);
6745 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6746 if (current->cred != ctx->creds) {
6748 revert_creds(old_cred);
6749 old_cred = override_creds(ctx->creds);
6751 io_sq_thread_associate_blkcg(ctx, &cur_css);
6753 current->loginuid = ctx->loginuid;
6754 current->sessionid = ctx->sessionid;
6757 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6759 io_sq_thread_drop_mm();
6762 if (ret & SQT_SPIN) {
6765 } else if (ret == SQT_IDLE) {
6766 if (kthread_should_park())
6768 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6769 io_ring_set_wakeup_flag(ctx);
6771 start_jiffies = jiffies;
6772 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6773 io_ring_clear_wakeup_flag(ctx);
6780 io_sq_thread_unassociate_blkcg();
6782 revert_creds(old_cred);
6789 struct io_wait_queue {
6790 struct wait_queue_entry wq;
6791 struct io_ring_ctx *ctx;
6793 unsigned nr_timeouts;
6796 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6798 struct io_ring_ctx *ctx = iowq->ctx;
6801 * Wake up if we have enough events, or if a timeout occurred since we
6802 * started waiting. For timeouts, we always want to return to userspace,
6803 * regardless of event count.
6805 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6806 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6809 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6810 int wake_flags, void *key)
6812 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6815 /* use noflush == true, as we can't safely rely on locking context */
6816 if (!io_should_wake(iowq, true))
6819 return autoremove_wake_function(curr, mode, wake_flags, key);
6822 static int io_run_task_work_sig(void)
6824 if (io_run_task_work())
6826 if (!signal_pending(current))
6828 if (current->jobctl & JOBCTL_TASK_WORK) {
6829 spin_lock_irq(¤t->sighand->siglock);
6830 current->jobctl &= ~JOBCTL_TASK_WORK;
6831 recalc_sigpending();
6832 spin_unlock_irq(¤t->sighand->siglock);
6839 * Wait until events become available, if we don't already have some. The
6840 * application must reap them itself, as they reside on the shared cq ring.
6842 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6843 const sigset_t __user *sig, size_t sigsz)
6845 struct io_wait_queue iowq = {
6848 .func = io_wake_function,
6849 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6852 .to_wait = min_events,
6854 struct io_rings *rings = ctx->rings;
6858 if (io_cqring_events(ctx, false) >= min_events)
6860 if (!io_run_task_work())
6865 #ifdef CONFIG_COMPAT
6866 if (in_compat_syscall())
6867 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6871 ret = set_user_sigmask(sig, sigsz);
6877 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6878 trace_io_uring_cqring_wait(ctx, min_events);
6880 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6881 TASK_INTERRUPTIBLE);
6882 /* make sure we run task_work before checking for signals */
6883 ret = io_run_task_work_sig();
6888 if (io_should_wake(&iowq, false))
6892 finish_wait(&ctx->wait, &iowq.wq);
6894 restore_saved_sigmask_unless(ret == -EINTR);
6896 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6899 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6901 #if defined(CONFIG_UNIX)
6902 if (ctx->ring_sock) {
6903 struct sock *sock = ctx->ring_sock->sk;
6904 struct sk_buff *skb;
6906 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6912 for (i = 0; i < ctx->nr_user_files; i++) {
6915 file = io_file_from_index(ctx, i);
6922 static void io_file_ref_kill(struct percpu_ref *ref)
6924 struct fixed_file_data *data;
6926 data = container_of(ref, struct fixed_file_data, refs);
6927 complete(&data->done);
6930 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6932 struct fixed_file_data *data = ctx->file_data;
6933 struct fixed_file_ref_node *ref_node = NULL;
6934 unsigned nr_tables, i;
6939 spin_lock(&data->lock);
6940 if (!list_empty(&data->ref_list))
6941 ref_node = list_first_entry(&data->ref_list,
6942 struct fixed_file_ref_node, node);
6943 spin_unlock(&data->lock);
6945 percpu_ref_kill(&ref_node->refs);
6947 percpu_ref_kill(&data->refs);
6949 /* wait for all refs nodes to complete */
6950 flush_delayed_work(&ctx->file_put_work);
6951 wait_for_completion(&data->done);
6953 __io_sqe_files_unregister(ctx);
6954 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6955 for (i = 0; i < nr_tables; i++)
6956 kfree(data->table[i].files);
6958 percpu_ref_exit(&data->refs);
6960 ctx->file_data = NULL;
6961 ctx->nr_user_files = 0;
6965 static void io_put_sq_data(struct io_sq_data *sqd)
6967 if (refcount_dec_and_test(&sqd->refs)) {
6969 * The park is a bit of a work-around, without it we get
6970 * warning spews on shutdown with SQPOLL set and affinity
6971 * set to a single CPU.
6974 kthread_park(sqd->thread);
6975 kthread_stop(sqd->thread);
6982 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6984 struct io_ring_ctx *ctx_attach;
6985 struct io_sq_data *sqd;
6988 f = fdget(p->wq_fd);
6990 return ERR_PTR(-ENXIO);
6991 if (f.file->f_op != &io_uring_fops) {
6993 return ERR_PTR(-EINVAL);
6996 ctx_attach = f.file->private_data;
6997 sqd = ctx_attach->sq_data;
7000 return ERR_PTR(-EINVAL);
7003 refcount_inc(&sqd->refs);
7008 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7010 struct io_sq_data *sqd;
7012 if (p->flags & IORING_SETUP_ATTACH_WQ)
7013 return io_attach_sq_data(p);
7015 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7017 return ERR_PTR(-ENOMEM);
7019 refcount_set(&sqd->refs, 1);
7020 INIT_LIST_HEAD(&sqd->ctx_list);
7021 INIT_LIST_HEAD(&sqd->ctx_new_list);
7022 mutex_init(&sqd->ctx_lock);
7023 mutex_init(&sqd->lock);
7024 init_waitqueue_head(&sqd->wait);
7028 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7029 __releases(&sqd->lock)
7033 kthread_unpark(sqd->thread);
7034 mutex_unlock(&sqd->lock);
7037 static void io_sq_thread_park(struct io_sq_data *sqd)
7038 __acquires(&sqd->lock)
7042 mutex_lock(&sqd->lock);
7043 kthread_park(sqd->thread);
7046 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7048 struct io_sq_data *sqd = ctx->sq_data;
7053 * We may arrive here from the error branch in
7054 * io_sq_offload_create() where the kthread is created
7055 * without being waked up, thus wake it up now to make
7056 * sure the wait will complete.
7058 wake_up_process(sqd->thread);
7059 wait_for_completion(&ctx->sq_thread_comp);
7061 io_sq_thread_park(sqd);
7064 mutex_lock(&sqd->ctx_lock);
7065 list_del(&ctx->sqd_list);
7066 mutex_unlock(&sqd->ctx_lock);
7069 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7070 io_sq_thread_unpark(sqd);
7073 io_put_sq_data(sqd);
7074 ctx->sq_data = NULL;
7078 static void io_finish_async(struct io_ring_ctx *ctx)
7080 io_sq_thread_stop(ctx);
7083 io_wq_destroy(ctx->io_wq);
7088 #if defined(CONFIG_UNIX)
7090 * Ensure the UNIX gc is aware of our file set, so we are certain that
7091 * the io_uring can be safely unregistered on process exit, even if we have
7092 * loops in the file referencing.
7094 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7096 struct sock *sk = ctx->ring_sock->sk;
7097 struct scm_fp_list *fpl;
7098 struct sk_buff *skb;
7101 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7105 skb = alloc_skb(0, GFP_KERNEL);
7114 fpl->user = get_uid(ctx->user);
7115 for (i = 0; i < nr; i++) {
7116 struct file *file = io_file_from_index(ctx, i + offset);
7120 fpl->fp[nr_files] = get_file(file);
7121 unix_inflight(fpl->user, fpl->fp[nr_files]);
7126 fpl->max = SCM_MAX_FD;
7127 fpl->count = nr_files;
7128 UNIXCB(skb).fp = fpl;
7129 skb->destructor = unix_destruct_scm;
7130 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7131 skb_queue_head(&sk->sk_receive_queue, skb);
7133 for (i = 0; i < nr_files; i++)
7144 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7145 * causes regular reference counting to break down. We rely on the UNIX
7146 * garbage collection to take care of this problem for us.
7148 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7150 unsigned left, total;
7154 left = ctx->nr_user_files;
7156 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7158 ret = __io_sqe_files_scm(ctx, this_files, total);
7162 total += this_files;
7168 while (total < ctx->nr_user_files) {
7169 struct file *file = io_file_from_index(ctx, total);
7179 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7185 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7186 unsigned nr_tables, unsigned nr_files)
7190 for (i = 0; i < nr_tables; i++) {
7191 struct fixed_file_table *table = &file_data->table[i];
7192 unsigned this_files;
7194 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7195 table->files = kcalloc(this_files, sizeof(struct file *),
7199 nr_files -= this_files;
7205 for (i = 0; i < nr_tables; i++) {
7206 struct fixed_file_table *table = &file_data->table[i];
7207 kfree(table->files);
7212 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7214 #if defined(CONFIG_UNIX)
7215 struct sock *sock = ctx->ring_sock->sk;
7216 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7217 struct sk_buff *skb;
7220 __skb_queue_head_init(&list);
7223 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7224 * remove this entry and rearrange the file array.
7226 skb = skb_dequeue(head);
7228 struct scm_fp_list *fp;
7230 fp = UNIXCB(skb).fp;
7231 for (i = 0; i < fp->count; i++) {
7234 if (fp->fp[i] != file)
7237 unix_notinflight(fp->user, fp->fp[i]);
7238 left = fp->count - 1 - i;
7240 memmove(&fp->fp[i], &fp->fp[i + 1],
7241 left * sizeof(struct file *));
7248 __skb_queue_tail(&list, skb);
7258 __skb_queue_tail(&list, skb);
7260 skb = skb_dequeue(head);
7263 if (skb_peek(&list)) {
7264 spin_lock_irq(&head->lock);
7265 while ((skb = __skb_dequeue(&list)) != NULL)
7266 __skb_queue_tail(head, skb);
7267 spin_unlock_irq(&head->lock);
7274 struct io_file_put {
7275 struct list_head list;
7279 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7281 struct fixed_file_data *file_data = ref_node->file_data;
7282 struct io_ring_ctx *ctx = file_data->ctx;
7283 struct io_file_put *pfile, *tmp;
7285 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7286 list_del(&pfile->list);
7287 io_ring_file_put(ctx, pfile->file);
7291 spin_lock(&file_data->lock);
7292 list_del(&ref_node->node);
7293 spin_unlock(&file_data->lock);
7295 percpu_ref_exit(&ref_node->refs);
7297 percpu_ref_put(&file_data->refs);
7300 static void io_file_put_work(struct work_struct *work)
7302 struct io_ring_ctx *ctx;
7303 struct llist_node *node;
7305 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7306 node = llist_del_all(&ctx->file_put_llist);
7309 struct fixed_file_ref_node *ref_node;
7310 struct llist_node *next = node->next;
7312 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7313 __io_file_put_work(ref_node);
7318 static void io_file_data_ref_zero(struct percpu_ref *ref)
7320 struct fixed_file_ref_node *ref_node;
7321 struct io_ring_ctx *ctx;
7325 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7326 ctx = ref_node->file_data->ctx;
7328 if (percpu_ref_is_dying(&ctx->file_data->refs))
7331 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7333 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7335 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7338 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7339 struct io_ring_ctx *ctx)
7341 struct fixed_file_ref_node *ref_node;
7343 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7345 return ERR_PTR(-ENOMEM);
7347 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7350 return ERR_PTR(-ENOMEM);
7352 INIT_LIST_HEAD(&ref_node->node);
7353 INIT_LIST_HEAD(&ref_node->file_list);
7354 ref_node->file_data = ctx->file_data;
7358 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7360 percpu_ref_exit(&ref_node->refs);
7364 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7367 __s32 __user *fds = (__s32 __user *) arg;
7368 unsigned nr_tables, i;
7370 int fd, ret = -ENOMEM;
7371 struct fixed_file_ref_node *ref_node;
7372 struct fixed_file_data *file_data;
7378 if (nr_args > IORING_MAX_FIXED_FILES)
7381 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7384 file_data->ctx = ctx;
7385 init_completion(&file_data->done);
7386 INIT_LIST_HEAD(&file_data->ref_list);
7387 spin_lock_init(&file_data->lock);
7389 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7390 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7392 if (!file_data->table)
7395 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7396 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7399 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7401 ctx->file_data = file_data;
7403 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7404 struct fixed_file_table *table;
7407 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7411 /* allow sparse sets */
7421 * Don't allow io_uring instances to be registered. If UNIX
7422 * isn't enabled, then this causes a reference cycle and this
7423 * instance can never get freed. If UNIX is enabled we'll
7424 * handle it just fine, but there's still no point in allowing
7425 * a ring fd as it doesn't support regular read/write anyway.
7427 if (file->f_op == &io_uring_fops) {
7431 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7432 index = i & IORING_FILE_TABLE_MASK;
7433 table->files[index] = file;
7436 ret = io_sqe_files_scm(ctx);
7438 io_sqe_files_unregister(ctx);
7442 ref_node = alloc_fixed_file_ref_node(ctx);
7443 if (IS_ERR(ref_node)) {
7444 io_sqe_files_unregister(ctx);
7445 return PTR_ERR(ref_node);
7448 file_data->node = ref_node;
7449 spin_lock(&file_data->lock);
7450 list_add(&ref_node->node, &file_data->ref_list);
7451 spin_unlock(&file_data->lock);
7452 percpu_ref_get(&file_data->refs);
7455 for (i = 0; i < ctx->nr_user_files; i++) {
7456 file = io_file_from_index(ctx, i);
7460 for (i = 0; i < nr_tables; i++)
7461 kfree(file_data->table[i].files);
7462 ctx->nr_user_files = 0;
7464 percpu_ref_exit(&file_data->refs);
7466 kfree(file_data->table);
7468 ctx->file_data = NULL;
7472 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7475 #if defined(CONFIG_UNIX)
7476 struct sock *sock = ctx->ring_sock->sk;
7477 struct sk_buff_head *head = &sock->sk_receive_queue;
7478 struct sk_buff *skb;
7481 * See if we can merge this file into an existing skb SCM_RIGHTS
7482 * file set. If there's no room, fall back to allocating a new skb
7483 * and filling it in.
7485 spin_lock_irq(&head->lock);
7486 skb = skb_peek(head);
7488 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7490 if (fpl->count < SCM_MAX_FD) {
7491 __skb_unlink(skb, head);
7492 spin_unlock_irq(&head->lock);
7493 fpl->fp[fpl->count] = get_file(file);
7494 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7496 spin_lock_irq(&head->lock);
7497 __skb_queue_head(head, skb);
7502 spin_unlock_irq(&head->lock);
7509 return __io_sqe_files_scm(ctx, 1, index);
7515 static int io_queue_file_removal(struct fixed_file_data *data,
7518 struct io_file_put *pfile;
7519 struct fixed_file_ref_node *ref_node = data->node;
7521 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7526 list_add(&pfile->list, &ref_node->file_list);
7531 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7532 struct io_uring_files_update *up,
7535 struct fixed_file_data *data = ctx->file_data;
7536 struct fixed_file_ref_node *ref_node;
7541 bool needs_switch = false;
7543 if (check_add_overflow(up->offset, nr_args, &done))
7545 if (done > ctx->nr_user_files)
7548 ref_node = alloc_fixed_file_ref_node(ctx);
7549 if (IS_ERR(ref_node))
7550 return PTR_ERR(ref_node);
7553 fds = u64_to_user_ptr(up->fds);
7555 struct fixed_file_table *table;
7559 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7563 i = array_index_nospec(up->offset, ctx->nr_user_files);
7564 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7565 index = i & IORING_FILE_TABLE_MASK;
7566 if (table->files[index]) {
7567 file = table->files[index];
7568 err = io_queue_file_removal(data, file);
7571 table->files[index] = NULL;
7572 needs_switch = true;
7581 * Don't allow io_uring instances to be registered. If
7582 * UNIX isn't enabled, then this causes a reference
7583 * cycle and this instance can never get freed. If UNIX
7584 * is enabled we'll handle it just fine, but there's
7585 * still no point in allowing a ring fd as it doesn't
7586 * support regular read/write anyway.
7588 if (file->f_op == &io_uring_fops) {
7593 table->files[index] = file;
7594 err = io_sqe_file_register(ctx, file, i);
7596 table->files[index] = NULL;
7607 percpu_ref_kill(&data->node->refs);
7608 spin_lock(&data->lock);
7609 list_add(&ref_node->node, &data->ref_list);
7610 data->node = ref_node;
7611 spin_unlock(&data->lock);
7612 percpu_ref_get(&ctx->file_data->refs);
7614 destroy_fixed_file_ref_node(ref_node);
7616 return done ? done : err;
7619 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7622 struct io_uring_files_update up;
7624 if (!ctx->file_data)
7628 if (copy_from_user(&up, arg, sizeof(up)))
7633 return __io_sqe_files_update(ctx, &up, nr_args);
7636 static void io_free_work(struct io_wq_work *work)
7638 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7640 /* Consider that io_steal_work() relies on this ref */
7644 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7645 struct io_uring_params *p)
7647 struct io_wq_data data;
7649 struct io_ring_ctx *ctx_attach;
7650 unsigned int concurrency;
7653 data.user = ctx->user;
7654 data.free_work = io_free_work;
7655 data.do_work = io_wq_submit_work;
7657 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7658 /* Do QD, or 4 * CPUS, whatever is smallest */
7659 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7661 ctx->io_wq = io_wq_create(concurrency, &data);
7662 if (IS_ERR(ctx->io_wq)) {
7663 ret = PTR_ERR(ctx->io_wq);
7669 f = fdget(p->wq_fd);
7673 if (f.file->f_op != &io_uring_fops) {
7678 ctx_attach = f.file->private_data;
7679 /* @io_wq is protected by holding the fd */
7680 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7685 ctx->io_wq = ctx_attach->io_wq;
7691 static int io_uring_alloc_task_context(struct task_struct *task)
7693 struct io_uring_task *tctx;
7696 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7697 if (unlikely(!tctx))
7700 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7701 if (unlikely(ret)) {
7707 init_waitqueue_head(&tctx->wait);
7710 io_init_identity(&tctx->__identity);
7711 tctx->identity = &tctx->__identity;
7712 task->io_uring = tctx;
7716 void __io_uring_free(struct task_struct *tsk)
7718 struct io_uring_task *tctx = tsk->io_uring;
7720 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7721 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7722 if (tctx->identity != &tctx->__identity)
7723 kfree(tctx->identity);
7724 percpu_counter_destroy(&tctx->inflight);
7726 tsk->io_uring = NULL;
7729 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7730 struct io_uring_params *p)
7734 if (ctx->flags & IORING_SETUP_SQPOLL) {
7735 struct io_sq_data *sqd;
7738 if (!capable(CAP_SYS_ADMIN))
7741 sqd = io_get_sq_data(p);
7748 io_sq_thread_park(sqd);
7749 mutex_lock(&sqd->ctx_lock);
7750 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7751 mutex_unlock(&sqd->ctx_lock);
7752 io_sq_thread_unpark(sqd);
7754 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7755 if (!ctx->sq_thread_idle)
7756 ctx->sq_thread_idle = HZ;
7761 if (p->flags & IORING_SETUP_SQ_AFF) {
7762 int cpu = p->sq_thread_cpu;
7765 if (cpu >= nr_cpu_ids)
7767 if (!cpu_online(cpu))
7770 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7771 cpu, "io_uring-sq");
7773 sqd->thread = kthread_create(io_sq_thread, sqd,
7776 if (IS_ERR(sqd->thread)) {
7777 ret = PTR_ERR(sqd->thread);
7781 ret = io_uring_alloc_task_context(sqd->thread);
7784 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7785 /* Can't have SQ_AFF without SQPOLL */
7791 ret = io_init_wq_offload(ctx, p);
7797 io_finish_async(ctx);
7801 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7803 struct io_sq_data *sqd = ctx->sq_data;
7805 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7806 wake_up_process(sqd->thread);
7809 static inline void __io_unaccount_mem(struct user_struct *user,
7810 unsigned long nr_pages)
7812 atomic_long_sub(nr_pages, &user->locked_vm);
7815 static inline int __io_account_mem(struct user_struct *user,
7816 unsigned long nr_pages)
7818 unsigned long page_limit, cur_pages, new_pages;
7820 /* Don't allow more pages than we can safely lock */
7821 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7824 cur_pages = atomic_long_read(&user->locked_vm);
7825 new_pages = cur_pages + nr_pages;
7826 if (new_pages > page_limit)
7828 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7829 new_pages) != cur_pages);
7834 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7835 enum io_mem_account acct)
7838 __io_unaccount_mem(ctx->user, nr_pages);
7840 if (ctx->mm_account) {
7841 if (acct == ACCT_LOCKED)
7842 ctx->mm_account->locked_vm -= nr_pages;
7843 else if (acct == ACCT_PINNED)
7844 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7848 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7849 enum io_mem_account acct)
7853 if (ctx->limit_mem) {
7854 ret = __io_account_mem(ctx->user, nr_pages);
7859 if (ctx->mm_account) {
7860 if (acct == ACCT_LOCKED)
7861 ctx->mm_account->locked_vm += nr_pages;
7862 else if (acct == ACCT_PINNED)
7863 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7869 static void io_mem_free(void *ptr)
7876 page = virt_to_head_page(ptr);
7877 if (put_page_testzero(page))
7878 free_compound_page(page);
7881 static void *io_mem_alloc(size_t size)
7883 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7886 return (void *) __get_free_pages(gfp_flags, get_order(size));
7889 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7892 struct io_rings *rings;
7893 size_t off, sq_array_size;
7895 off = struct_size(rings, cqes, cq_entries);
7896 if (off == SIZE_MAX)
7900 off = ALIGN(off, SMP_CACHE_BYTES);
7908 sq_array_size = array_size(sizeof(u32), sq_entries);
7909 if (sq_array_size == SIZE_MAX)
7912 if (check_add_overflow(off, sq_array_size, &off))
7918 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7922 pages = (size_t)1 << get_order(
7923 rings_size(sq_entries, cq_entries, NULL));
7924 pages += (size_t)1 << get_order(
7925 array_size(sizeof(struct io_uring_sqe), sq_entries));
7930 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7934 if (!ctx->user_bufs)
7937 for (i = 0; i < ctx->nr_user_bufs; i++) {
7938 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7940 for (j = 0; j < imu->nr_bvecs; j++)
7941 unpin_user_page(imu->bvec[j].bv_page);
7943 if (imu->acct_pages)
7944 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7949 kfree(ctx->user_bufs);
7950 ctx->user_bufs = NULL;
7951 ctx->nr_user_bufs = 0;
7955 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7956 void __user *arg, unsigned index)
7958 struct iovec __user *src;
7960 #ifdef CONFIG_COMPAT
7962 struct compat_iovec __user *ciovs;
7963 struct compat_iovec ciov;
7965 ciovs = (struct compat_iovec __user *) arg;
7966 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7969 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7970 dst->iov_len = ciov.iov_len;
7974 src = (struct iovec __user *) arg;
7975 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7981 * Not super efficient, but this is just a registration time. And we do cache
7982 * the last compound head, so generally we'll only do a full search if we don't
7985 * We check if the given compound head page has already been accounted, to
7986 * avoid double accounting it. This allows us to account the full size of the
7987 * page, not just the constituent pages of a huge page.
7989 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7990 int nr_pages, struct page *hpage)
7994 /* check current page array */
7995 for (i = 0; i < nr_pages; i++) {
7996 if (!PageCompound(pages[i]))
7998 if (compound_head(pages[i]) == hpage)
8002 /* check previously registered pages */
8003 for (i = 0; i < ctx->nr_user_bufs; i++) {
8004 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8006 for (j = 0; j < imu->nr_bvecs; j++) {
8007 if (!PageCompound(imu->bvec[j].bv_page))
8009 if (compound_head(imu->bvec[j].bv_page) == hpage)
8017 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8018 int nr_pages, struct io_mapped_ubuf *imu,
8019 struct page **last_hpage)
8023 for (i = 0; i < nr_pages; i++) {
8024 if (!PageCompound(pages[i])) {
8029 hpage = compound_head(pages[i]);
8030 if (hpage == *last_hpage)
8032 *last_hpage = hpage;
8033 if (headpage_already_acct(ctx, pages, i, hpage))
8035 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8039 if (!imu->acct_pages)
8042 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8044 imu->acct_pages = 0;
8048 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8051 struct vm_area_struct **vmas = NULL;
8052 struct page **pages = NULL;
8053 struct page *last_hpage = NULL;
8054 int i, j, got_pages = 0;
8059 if (!nr_args || nr_args > UIO_MAXIOV)
8062 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8064 if (!ctx->user_bufs)
8067 for (i = 0; i < nr_args; i++) {
8068 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8069 unsigned long off, start, end, ubuf;
8074 ret = io_copy_iov(ctx, &iov, arg, i);
8079 * Don't impose further limits on the size and buffer
8080 * constraints here, we'll -EINVAL later when IO is
8081 * submitted if they are wrong.
8084 if (!iov.iov_base || !iov.iov_len)
8087 /* arbitrary limit, but we need something */
8088 if (iov.iov_len > SZ_1G)
8091 ubuf = (unsigned long) iov.iov_base;
8092 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8093 start = ubuf >> PAGE_SHIFT;
8094 nr_pages = end - start;
8097 if (!pages || nr_pages > got_pages) {
8100 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8102 vmas = kvmalloc_array(nr_pages,
8103 sizeof(struct vm_area_struct *),
8105 if (!pages || !vmas) {
8109 got_pages = nr_pages;
8112 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8119 mmap_read_lock(current->mm);
8120 pret = pin_user_pages(ubuf, nr_pages,
8121 FOLL_WRITE | FOLL_LONGTERM,
8123 if (pret == nr_pages) {
8124 /* don't support file backed memory */
8125 for (j = 0; j < nr_pages; j++) {
8126 struct vm_area_struct *vma = vmas[j];
8129 !is_file_hugepages(vma->vm_file)) {
8135 ret = pret < 0 ? pret : -EFAULT;
8137 mmap_read_unlock(current->mm);
8140 * if we did partial map, or found file backed vmas,
8141 * release any pages we did get
8144 unpin_user_pages(pages, pret);
8149 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8151 unpin_user_pages(pages, pret);
8156 off = ubuf & ~PAGE_MASK;
8158 for (j = 0; j < nr_pages; j++) {
8161 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8162 imu->bvec[j].bv_page = pages[j];
8163 imu->bvec[j].bv_len = vec_len;
8164 imu->bvec[j].bv_offset = off;
8168 /* store original address for later verification */
8170 imu->len = iov.iov_len;
8171 imu->nr_bvecs = nr_pages;
8173 ctx->nr_user_bufs++;
8181 io_sqe_buffer_unregister(ctx);
8185 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8187 __s32 __user *fds = arg;
8193 if (copy_from_user(&fd, fds, sizeof(*fds)))
8196 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8197 if (IS_ERR(ctx->cq_ev_fd)) {
8198 int ret = PTR_ERR(ctx->cq_ev_fd);
8199 ctx->cq_ev_fd = NULL;
8206 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8208 if (ctx->cq_ev_fd) {
8209 eventfd_ctx_put(ctx->cq_ev_fd);
8210 ctx->cq_ev_fd = NULL;
8217 static int __io_destroy_buffers(int id, void *p, void *data)
8219 struct io_ring_ctx *ctx = data;
8220 struct io_buffer *buf = p;
8222 __io_remove_buffers(ctx, buf, id, -1U);
8226 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8228 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8229 idr_destroy(&ctx->io_buffer_idr);
8232 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8234 io_finish_async(ctx);
8235 io_sqe_buffer_unregister(ctx);
8237 if (ctx->sqo_task) {
8238 put_task_struct(ctx->sqo_task);
8239 ctx->sqo_task = NULL;
8240 mmdrop(ctx->mm_account);
8241 ctx->mm_account = NULL;
8244 #ifdef CONFIG_BLK_CGROUP
8245 if (ctx->sqo_blkcg_css)
8246 css_put(ctx->sqo_blkcg_css);
8249 io_sqe_files_unregister(ctx);
8250 io_eventfd_unregister(ctx);
8251 io_destroy_buffers(ctx);
8252 idr_destroy(&ctx->personality_idr);
8254 #if defined(CONFIG_UNIX)
8255 if (ctx->ring_sock) {
8256 ctx->ring_sock->file = NULL; /* so that iput() is called */
8257 sock_release(ctx->ring_sock);
8261 io_mem_free(ctx->rings);
8262 io_mem_free(ctx->sq_sqes);
8264 percpu_ref_exit(&ctx->refs);
8265 free_uid(ctx->user);
8266 put_cred(ctx->creds);
8267 kfree(ctx->cancel_hash);
8268 kmem_cache_free(req_cachep, ctx->fallback_req);
8272 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8274 struct io_ring_ctx *ctx = file->private_data;
8277 poll_wait(file, &ctx->cq_wait, wait);
8279 * synchronizes with barrier from wq_has_sleeper call in
8283 if (!io_sqring_full(ctx))
8284 mask |= EPOLLOUT | EPOLLWRNORM;
8285 if (io_cqring_events(ctx, false))
8286 mask |= EPOLLIN | EPOLLRDNORM;
8291 static int io_uring_fasync(int fd, struct file *file, int on)
8293 struct io_ring_ctx *ctx = file->private_data;
8295 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8298 static int io_remove_personalities(int id, void *p, void *data)
8300 struct io_ring_ctx *ctx = data;
8301 struct io_identity *iod;
8303 iod = idr_remove(&ctx->personality_idr, id);
8305 put_cred(iod->creds);
8306 if (refcount_dec_and_test(&iod->count))
8312 static void io_ring_exit_work(struct work_struct *work)
8314 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8318 * If we're doing polled IO and end up having requests being
8319 * submitted async (out-of-line), then completions can come in while
8320 * we're waiting for refs to drop. We need to reap these manually,
8321 * as nobody else will be looking for them.
8325 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8326 io_iopoll_try_reap_events(ctx);
8327 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8328 io_ring_ctx_free(ctx);
8331 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8333 mutex_lock(&ctx->uring_lock);
8334 percpu_ref_kill(&ctx->refs);
8335 mutex_unlock(&ctx->uring_lock);
8337 io_kill_timeouts(ctx, NULL);
8338 io_poll_remove_all(ctx, NULL);
8341 io_wq_cancel_all(ctx->io_wq);
8343 /* if we failed setting up the ctx, we might not have any rings */
8345 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8346 io_iopoll_try_reap_events(ctx);
8347 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8350 * Do this upfront, so we won't have a grace period where the ring
8351 * is closed but resources aren't reaped yet. This can cause
8352 * spurious failure in setting up a new ring.
8354 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8357 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8359 * Use system_unbound_wq to avoid spawning tons of event kworkers
8360 * if we're exiting a ton of rings at the same time. It just adds
8361 * noise and overhead, there's no discernable change in runtime
8362 * over using system_wq.
8364 queue_work(system_unbound_wq, &ctx->exit_work);
8367 static int io_uring_release(struct inode *inode, struct file *file)
8369 struct io_ring_ctx *ctx = file->private_data;
8371 file->private_data = NULL;
8372 io_ring_ctx_wait_and_kill(ctx);
8376 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8378 struct files_struct *files = data;
8380 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8381 work->identity->files == files);
8385 * Returns true if 'preq' is the link parent of 'req'
8387 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8389 struct io_kiocb *link;
8391 if (!(preq->flags & REQ_F_LINK_HEAD))
8394 list_for_each_entry(link, &preq->link_list, link_list) {
8402 static bool io_match_link_files(struct io_kiocb *req,
8403 struct files_struct *files)
8405 struct io_kiocb *link;
8407 if (io_match_files(req, files))
8409 if (req->flags & REQ_F_LINK_HEAD) {
8410 list_for_each_entry(link, &req->link_list, link_list) {
8411 if (io_match_files(link, files))
8419 * We're looking to cancel 'req' because it's holding on to our files, but
8420 * 'req' could be a link to another request. See if it is, and cancel that
8421 * parent request if so.
8423 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8425 struct hlist_node *tmp;
8426 struct io_kiocb *preq;
8430 spin_lock_irq(&ctx->completion_lock);
8431 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8432 struct hlist_head *list;
8434 list = &ctx->cancel_hash[i];
8435 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8436 found = io_match_link(preq, req);
8438 io_poll_remove_one(preq);
8443 spin_unlock_irq(&ctx->completion_lock);
8447 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8448 struct io_kiocb *req)
8450 struct io_kiocb *preq;
8453 spin_lock_irq(&ctx->completion_lock);
8454 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8455 found = io_match_link(preq, req);
8457 __io_timeout_cancel(preq);
8461 spin_unlock_irq(&ctx->completion_lock);
8465 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8467 return io_match_link(container_of(work, struct io_kiocb, work), data);
8470 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8472 enum io_wq_cancel cret;
8474 /* cancel this particular work, if it's running */
8475 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8476 if (cret != IO_WQ_CANCEL_NOTFOUND)
8479 /* find links that hold this pending, cancel those */
8480 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8481 if (cret != IO_WQ_CANCEL_NOTFOUND)
8484 /* if we have a poll link holding this pending, cancel that */
8485 if (io_poll_remove_link(ctx, req))
8488 /* final option, timeout link is holding this req pending */
8489 io_timeout_remove_link(ctx, req);
8492 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8493 struct files_struct *files)
8495 struct io_defer_entry *de = NULL;
8498 spin_lock_irq(&ctx->completion_lock);
8499 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8500 if (io_match_link_files(de->req, files)) {
8501 list_cut_position(&list, &ctx->defer_list, &de->list);
8505 spin_unlock_irq(&ctx->completion_lock);
8507 while (!list_empty(&list)) {
8508 de = list_first_entry(&list, struct io_defer_entry, list);
8509 list_del_init(&de->list);
8510 req_set_fail_links(de->req);
8511 io_put_req(de->req);
8512 io_req_complete(de->req, -ECANCELED);
8518 * Returns true if we found and killed one or more files pinning requests
8520 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8521 struct files_struct *files)
8523 if (list_empty_careful(&ctx->inflight_list))
8526 io_cancel_defer_files(ctx, files);
8527 /* cancel all at once, should be faster than doing it one by one*/
8528 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8530 while (!list_empty_careful(&ctx->inflight_list)) {
8531 struct io_kiocb *cancel_req = NULL, *req;
8534 spin_lock_irq(&ctx->inflight_lock);
8535 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8536 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8537 req->work.identity->files != files)
8539 /* req is being completed, ignore */
8540 if (!refcount_inc_not_zero(&req->refs))
8546 prepare_to_wait(&ctx->inflight_wait, &wait,
8547 TASK_UNINTERRUPTIBLE);
8548 spin_unlock_irq(&ctx->inflight_lock);
8550 /* We need to keep going until we don't find a matching req */
8553 /* cancel this request, or head link requests */
8554 io_attempt_cancel(ctx, cancel_req);
8555 io_put_req(cancel_req);
8556 /* cancellations _may_ trigger task work */
8559 finish_wait(&ctx->inflight_wait, &wait);
8565 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8567 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8568 struct task_struct *task = data;
8570 return io_task_match(req, task);
8573 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8574 struct task_struct *task,
8575 struct files_struct *files)
8579 ret = io_uring_cancel_files(ctx, files);
8581 enum io_wq_cancel cret;
8583 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8584 if (cret != IO_WQ_CANCEL_NOTFOUND)
8587 /* SQPOLL thread does its own polling */
8588 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8589 while (!list_empty_careful(&ctx->iopoll_list)) {
8590 io_iopoll_try_reap_events(ctx);
8595 ret |= io_poll_remove_all(ctx, task);
8596 ret |= io_kill_timeouts(ctx, task);
8603 * We need to iteratively cancel requests, in case a request has dependent
8604 * hard links. These persist even for failure of cancelations, hence keep
8605 * looping until none are found.
8607 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8608 struct files_struct *files)
8610 struct task_struct *task = current;
8612 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8613 task = ctx->sq_data->thread;
8615 io_cqring_overflow_flush(ctx, true, task, files);
8617 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8624 * Note that this task has used io_uring. We use it for cancelation purposes.
8626 static int io_uring_add_task_file(struct file *file)
8628 struct io_uring_task *tctx = current->io_uring;
8630 if (unlikely(!tctx)) {
8633 ret = io_uring_alloc_task_context(current);
8636 tctx = current->io_uring;
8638 if (tctx->last != file) {
8639 void *old = xa_load(&tctx->xa, (unsigned long)file);
8643 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8652 * Remove this io_uring_file -> task mapping.
8654 static void io_uring_del_task_file(struct file *file)
8656 struct io_uring_task *tctx = current->io_uring;
8658 if (tctx->last == file)
8660 file = xa_erase(&tctx->xa, (unsigned long)file);
8666 * Drop task note for this file if we're the only ones that hold it after
8669 static void io_uring_attempt_task_drop(struct file *file)
8671 if (!current->io_uring)
8674 * fput() is pending, will be 2 if the only other ref is our potential
8675 * task file note. If the task is exiting, drop regardless of count.
8677 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8678 atomic_long_read(&file->f_count) == 2)
8679 io_uring_del_task_file(file);
8682 void __io_uring_files_cancel(struct files_struct *files)
8684 struct io_uring_task *tctx = current->io_uring;
8686 unsigned long index;
8688 /* make sure overflow events are dropped */
8689 tctx->in_idle = true;
8691 xa_for_each(&tctx->xa, index, file) {
8692 struct io_ring_ctx *ctx = file->private_data;
8694 io_uring_cancel_task_requests(ctx, files);
8696 io_uring_del_task_file(file);
8701 * Find any io_uring fd that this task has registered or done IO on, and cancel
8704 void __io_uring_task_cancel(void)
8706 struct io_uring_task *tctx = current->io_uring;
8710 /* make sure overflow events are dropped */
8711 tctx->in_idle = true;
8714 /* read completions before cancelations */
8715 inflight = percpu_counter_sum(&tctx->inflight);
8718 __io_uring_files_cancel(NULL);
8720 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8723 * If we've seen completions, retry. This avoids a race where
8724 * a completion comes in before we did prepare_to_wait().
8726 if (inflight != percpu_counter_sum(&tctx->inflight))
8731 finish_wait(&tctx->wait, &wait);
8732 tctx->in_idle = false;
8735 static int io_uring_flush(struct file *file, void *data)
8737 io_uring_attempt_task_drop(file);
8741 static void *io_uring_validate_mmap_request(struct file *file,
8742 loff_t pgoff, size_t sz)
8744 struct io_ring_ctx *ctx = file->private_data;
8745 loff_t offset = pgoff << PAGE_SHIFT;
8750 case IORING_OFF_SQ_RING:
8751 case IORING_OFF_CQ_RING:
8754 case IORING_OFF_SQES:
8758 return ERR_PTR(-EINVAL);
8761 page = virt_to_head_page(ptr);
8762 if (sz > page_size(page))
8763 return ERR_PTR(-EINVAL);
8770 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8772 size_t sz = vma->vm_end - vma->vm_start;
8776 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8778 return PTR_ERR(ptr);
8780 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8781 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8784 #else /* !CONFIG_MMU */
8786 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8788 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8791 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8793 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8796 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8797 unsigned long addr, unsigned long len,
8798 unsigned long pgoff, unsigned long flags)
8802 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8804 return PTR_ERR(ptr);
8806 return (unsigned long) ptr;
8809 #endif /* !CONFIG_MMU */
8811 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8816 if (!io_sqring_full(ctx))
8819 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8821 if (!io_sqring_full(ctx))
8825 } while (!signal_pending(current));
8827 finish_wait(&ctx->sqo_sq_wait, &wait);
8830 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8831 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8834 struct io_ring_ctx *ctx;
8841 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8842 IORING_ENTER_SQ_WAIT))
8850 if (f.file->f_op != &io_uring_fops)
8854 ctx = f.file->private_data;
8855 if (!percpu_ref_tryget(&ctx->refs))
8859 if (ctx->flags & IORING_SETUP_R_DISABLED)
8863 * For SQ polling, the thread will do all submissions and completions.
8864 * Just return the requested submit count, and wake the thread if
8868 if (ctx->flags & IORING_SETUP_SQPOLL) {
8869 if (!list_empty_careful(&ctx->cq_overflow_list))
8870 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8871 if (flags & IORING_ENTER_SQ_WAKEUP)
8872 wake_up(&ctx->sq_data->wait);
8873 if (flags & IORING_ENTER_SQ_WAIT)
8874 io_sqpoll_wait_sq(ctx);
8875 submitted = to_submit;
8876 } else if (to_submit) {
8877 ret = io_uring_add_task_file(f.file);
8880 mutex_lock(&ctx->uring_lock);
8881 submitted = io_submit_sqes(ctx, to_submit);
8882 mutex_unlock(&ctx->uring_lock);
8884 if (submitted != to_submit)
8887 if (flags & IORING_ENTER_GETEVENTS) {
8888 min_complete = min(min_complete, ctx->cq_entries);
8891 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8892 * space applications don't need to do io completion events
8893 * polling again, they can rely on io_sq_thread to do polling
8894 * work, which can reduce cpu usage and uring_lock contention.
8896 if (ctx->flags & IORING_SETUP_IOPOLL &&
8897 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8898 ret = io_iopoll_check(ctx, min_complete);
8900 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8905 percpu_ref_put(&ctx->refs);
8908 return submitted ? submitted : ret;
8911 #ifdef CONFIG_PROC_FS
8912 static int io_uring_show_cred(int id, void *p, void *data)
8914 const struct cred *cred = p;
8915 struct seq_file *m = data;
8916 struct user_namespace *uns = seq_user_ns(m);
8917 struct group_info *gi;
8922 seq_printf(m, "%5d\n", id);
8923 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8924 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8925 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8926 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8927 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8928 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8929 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8930 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8931 seq_puts(m, "\n\tGroups:\t");
8932 gi = cred->group_info;
8933 for (g = 0; g < gi->ngroups; g++) {
8934 seq_put_decimal_ull(m, g ? " " : "",
8935 from_kgid_munged(uns, gi->gid[g]));
8937 seq_puts(m, "\n\tCapEff:\t");
8938 cap = cred->cap_effective;
8939 CAP_FOR_EACH_U32(__capi)
8940 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8945 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8947 struct io_sq_data *sq = NULL;
8952 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8953 * since fdinfo case grabs it in the opposite direction of normal use
8954 * cases. If we fail to get the lock, we just don't iterate any
8955 * structures that could be going away outside the io_uring mutex.
8957 has_lock = mutex_trylock(&ctx->uring_lock);
8959 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8962 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8963 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8964 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8965 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8966 struct fixed_file_table *table;
8969 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8970 f = table->files[i & IORING_FILE_TABLE_MASK];
8972 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8974 seq_printf(m, "%5u: <none>\n", i);
8976 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8977 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8978 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8980 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8981 (unsigned int) buf->len);
8983 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8984 seq_printf(m, "Personalities:\n");
8985 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8987 seq_printf(m, "PollList:\n");
8988 spin_lock_irq(&ctx->completion_lock);
8989 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8990 struct hlist_head *list = &ctx->cancel_hash[i];
8991 struct io_kiocb *req;
8993 hlist_for_each_entry(req, list, hash_node)
8994 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8995 req->task->task_works != NULL);
8997 spin_unlock_irq(&ctx->completion_lock);
8999 mutex_unlock(&ctx->uring_lock);
9002 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9004 struct io_ring_ctx *ctx = f->private_data;
9006 if (percpu_ref_tryget(&ctx->refs)) {
9007 __io_uring_show_fdinfo(ctx, m);
9008 percpu_ref_put(&ctx->refs);
9013 static const struct file_operations io_uring_fops = {
9014 .release = io_uring_release,
9015 .flush = io_uring_flush,
9016 .mmap = io_uring_mmap,
9018 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9019 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9021 .poll = io_uring_poll,
9022 .fasync = io_uring_fasync,
9023 #ifdef CONFIG_PROC_FS
9024 .show_fdinfo = io_uring_show_fdinfo,
9028 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9029 struct io_uring_params *p)
9031 struct io_rings *rings;
9032 size_t size, sq_array_offset;
9034 /* make sure these are sane, as we already accounted them */
9035 ctx->sq_entries = p->sq_entries;
9036 ctx->cq_entries = p->cq_entries;
9038 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9039 if (size == SIZE_MAX)
9042 rings = io_mem_alloc(size);
9047 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9048 rings->sq_ring_mask = p->sq_entries - 1;
9049 rings->cq_ring_mask = p->cq_entries - 1;
9050 rings->sq_ring_entries = p->sq_entries;
9051 rings->cq_ring_entries = p->cq_entries;
9052 ctx->sq_mask = rings->sq_ring_mask;
9053 ctx->cq_mask = rings->cq_ring_mask;
9055 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9056 if (size == SIZE_MAX) {
9057 io_mem_free(ctx->rings);
9062 ctx->sq_sqes = io_mem_alloc(size);
9063 if (!ctx->sq_sqes) {
9064 io_mem_free(ctx->rings);
9073 * Allocate an anonymous fd, this is what constitutes the application
9074 * visible backing of an io_uring instance. The application mmaps this
9075 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9076 * we have to tie this fd to a socket for file garbage collection purposes.
9078 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9083 #if defined(CONFIG_UNIX)
9084 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9090 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9094 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9095 O_RDWR | O_CLOEXEC);
9099 ret = PTR_ERR(file);
9103 #if defined(CONFIG_UNIX)
9104 ctx->ring_sock->file = file;
9106 if (unlikely(io_uring_add_task_file(file))) {
9107 file = ERR_PTR(-ENOMEM);
9110 fd_install(ret, file);
9113 #if defined(CONFIG_UNIX)
9114 sock_release(ctx->ring_sock);
9115 ctx->ring_sock = NULL;
9120 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9121 struct io_uring_params __user *params)
9123 struct user_struct *user = NULL;
9124 struct io_ring_ctx *ctx;
9130 if (entries > IORING_MAX_ENTRIES) {
9131 if (!(p->flags & IORING_SETUP_CLAMP))
9133 entries = IORING_MAX_ENTRIES;
9137 * Use twice as many entries for the CQ ring. It's possible for the
9138 * application to drive a higher depth than the size of the SQ ring,
9139 * since the sqes are only used at submission time. This allows for
9140 * some flexibility in overcommitting a bit. If the application has
9141 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9142 * of CQ ring entries manually.
9144 p->sq_entries = roundup_pow_of_two(entries);
9145 if (p->flags & IORING_SETUP_CQSIZE) {
9147 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9148 * to a power-of-two, if it isn't already. We do NOT impose
9149 * any cq vs sq ring sizing.
9151 if (p->cq_entries < p->sq_entries)
9153 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9154 if (!(p->flags & IORING_SETUP_CLAMP))
9156 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9158 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9160 p->cq_entries = 2 * p->sq_entries;
9163 user = get_uid(current_user());
9164 limit_mem = !capable(CAP_IPC_LOCK);
9167 ret = __io_account_mem(user,
9168 ring_pages(p->sq_entries, p->cq_entries));
9175 ctx = io_ring_ctx_alloc(p);
9178 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9183 ctx->compat = in_compat_syscall();
9185 ctx->creds = get_current_cred();
9187 ctx->loginuid = current->loginuid;
9188 ctx->sessionid = current->sessionid;
9190 ctx->sqo_task = get_task_struct(current);
9193 * This is just grabbed for accounting purposes. When a process exits,
9194 * the mm is exited and dropped before the files, hence we need to hang
9195 * on to this mm purely for the purposes of being able to unaccount
9196 * memory (locked/pinned vm). It's not used for anything else.
9198 mmgrab(current->mm);
9199 ctx->mm_account = current->mm;
9201 #ifdef CONFIG_BLK_CGROUP
9203 * The sq thread will belong to the original cgroup it was inited in.
9204 * If the cgroup goes offline (e.g. disabling the io controller), then
9205 * issued bios will be associated with the closest cgroup later in the
9209 ctx->sqo_blkcg_css = blkcg_css();
9210 ret = css_tryget_online(ctx->sqo_blkcg_css);
9213 /* don't init against a dying cgroup, have the user try again */
9214 ctx->sqo_blkcg_css = NULL;
9221 * Account memory _before_ installing the file descriptor. Once
9222 * the descriptor is installed, it can get closed at any time. Also
9223 * do this before hitting the general error path, as ring freeing
9224 * will un-account as well.
9226 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9228 ctx->limit_mem = limit_mem;
9230 ret = io_allocate_scq_urings(ctx, p);
9234 ret = io_sq_offload_create(ctx, p);
9238 if (!(p->flags & IORING_SETUP_R_DISABLED))
9239 io_sq_offload_start(ctx);
9241 memset(&p->sq_off, 0, sizeof(p->sq_off));
9242 p->sq_off.head = offsetof(struct io_rings, sq.head);
9243 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9244 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9245 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9246 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9247 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9248 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9250 memset(&p->cq_off, 0, sizeof(p->cq_off));
9251 p->cq_off.head = offsetof(struct io_rings, cq.head);
9252 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9253 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9254 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9255 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9256 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9257 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9259 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9260 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9261 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9262 IORING_FEAT_POLL_32BITS;
9264 if (copy_to_user(params, p, sizeof(*p))) {
9270 * Install ring fd as the very last thing, so we don't risk someone
9271 * having closed it before we finish setup
9273 ret = io_uring_get_fd(ctx);
9277 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9280 io_ring_ctx_wait_and_kill(ctx);
9285 * Sets up an aio uring context, and returns the fd. Applications asks for a
9286 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9287 * params structure passed in.
9289 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9291 struct io_uring_params p;
9294 if (copy_from_user(&p, params, sizeof(p)))
9296 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9301 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9302 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9303 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9304 IORING_SETUP_R_DISABLED))
9307 return io_uring_create(entries, &p, params);
9310 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9311 struct io_uring_params __user *, params)
9313 return io_uring_setup(entries, params);
9316 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9318 struct io_uring_probe *p;
9322 size = struct_size(p, ops, nr_args);
9323 if (size == SIZE_MAX)
9325 p = kzalloc(size, GFP_KERNEL);
9330 if (copy_from_user(p, arg, size))
9333 if (memchr_inv(p, 0, size))
9336 p->last_op = IORING_OP_LAST - 1;
9337 if (nr_args > IORING_OP_LAST)
9338 nr_args = IORING_OP_LAST;
9340 for (i = 0; i < nr_args; i++) {
9342 if (!io_op_defs[i].not_supported)
9343 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9348 if (copy_to_user(arg, p, size))
9355 static int io_register_personality(struct io_ring_ctx *ctx)
9357 struct io_identity *id;
9360 id = kmalloc(sizeof(*id), GFP_KERNEL);
9364 io_init_identity(id);
9365 id->creds = get_current_cred();
9367 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9369 put_cred(id->creds);
9375 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9377 struct io_identity *iod;
9379 iod = idr_remove(&ctx->personality_idr, id);
9381 put_cred(iod->creds);
9382 if (refcount_dec_and_test(&iod->count))
9390 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9391 unsigned int nr_args)
9393 struct io_uring_restriction *res;
9397 /* Restrictions allowed only if rings started disabled */
9398 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9401 /* We allow only a single restrictions registration */
9402 if (ctx->restrictions.registered)
9405 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9408 size = array_size(nr_args, sizeof(*res));
9409 if (size == SIZE_MAX)
9412 res = memdup_user(arg, size);
9414 return PTR_ERR(res);
9418 for (i = 0; i < nr_args; i++) {
9419 switch (res[i].opcode) {
9420 case IORING_RESTRICTION_REGISTER_OP:
9421 if (res[i].register_op >= IORING_REGISTER_LAST) {
9426 __set_bit(res[i].register_op,
9427 ctx->restrictions.register_op);
9429 case IORING_RESTRICTION_SQE_OP:
9430 if (res[i].sqe_op >= IORING_OP_LAST) {
9435 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9437 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9438 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9440 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9441 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9450 /* Reset all restrictions if an error happened */
9452 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9454 ctx->restrictions.registered = true;
9460 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9462 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9465 if (ctx->restrictions.registered)
9466 ctx->restricted = 1;
9468 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9470 io_sq_offload_start(ctx);
9475 static bool io_register_op_must_quiesce(int op)
9478 case IORING_UNREGISTER_FILES:
9479 case IORING_REGISTER_FILES_UPDATE:
9480 case IORING_REGISTER_PROBE:
9481 case IORING_REGISTER_PERSONALITY:
9482 case IORING_UNREGISTER_PERSONALITY:
9489 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9490 void __user *arg, unsigned nr_args)
9491 __releases(ctx->uring_lock)
9492 __acquires(ctx->uring_lock)
9497 * We're inside the ring mutex, if the ref is already dying, then
9498 * someone else killed the ctx or is already going through
9499 * io_uring_register().
9501 if (percpu_ref_is_dying(&ctx->refs))
9504 if (io_register_op_must_quiesce(opcode)) {
9505 percpu_ref_kill(&ctx->refs);
9508 * Drop uring mutex before waiting for references to exit. If
9509 * another thread is currently inside io_uring_enter() it might
9510 * need to grab the uring_lock to make progress. If we hold it
9511 * here across the drain wait, then we can deadlock. It's safe
9512 * to drop the mutex here, since no new references will come in
9513 * after we've killed the percpu ref.
9515 mutex_unlock(&ctx->uring_lock);
9517 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9520 ret = io_run_task_work_sig();
9525 mutex_lock(&ctx->uring_lock);
9528 percpu_ref_resurrect(&ctx->refs);
9533 if (ctx->restricted) {
9534 if (opcode >= IORING_REGISTER_LAST) {
9539 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9546 case IORING_REGISTER_BUFFERS:
9547 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9549 case IORING_UNREGISTER_BUFFERS:
9553 ret = io_sqe_buffer_unregister(ctx);
9555 case IORING_REGISTER_FILES:
9556 ret = io_sqe_files_register(ctx, arg, nr_args);
9558 case IORING_UNREGISTER_FILES:
9562 ret = io_sqe_files_unregister(ctx);
9564 case IORING_REGISTER_FILES_UPDATE:
9565 ret = io_sqe_files_update(ctx, arg, nr_args);
9567 case IORING_REGISTER_EVENTFD:
9568 case IORING_REGISTER_EVENTFD_ASYNC:
9572 ret = io_eventfd_register(ctx, arg);
9575 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9576 ctx->eventfd_async = 1;
9578 ctx->eventfd_async = 0;
9580 case IORING_UNREGISTER_EVENTFD:
9584 ret = io_eventfd_unregister(ctx);
9586 case IORING_REGISTER_PROBE:
9588 if (!arg || nr_args > 256)
9590 ret = io_probe(ctx, arg, nr_args);
9592 case IORING_REGISTER_PERSONALITY:
9596 ret = io_register_personality(ctx);
9598 case IORING_UNREGISTER_PERSONALITY:
9602 ret = io_unregister_personality(ctx, nr_args);
9604 case IORING_REGISTER_ENABLE_RINGS:
9608 ret = io_register_enable_rings(ctx);
9610 case IORING_REGISTER_RESTRICTIONS:
9611 ret = io_register_restrictions(ctx, arg, nr_args);
9619 if (io_register_op_must_quiesce(opcode)) {
9620 /* bring the ctx back to life */
9621 percpu_ref_reinit(&ctx->refs);
9623 reinit_completion(&ctx->ref_comp);
9628 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9629 void __user *, arg, unsigned int, nr_args)
9631 struct io_ring_ctx *ctx;
9640 if (f.file->f_op != &io_uring_fops)
9643 ctx = f.file->private_data;
9645 mutex_lock(&ctx->uring_lock);
9646 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9647 mutex_unlock(&ctx->uring_lock);
9648 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9649 ctx->cq_ev_fd != NULL, ret);
9655 static int __init io_uring_init(void)
9657 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9658 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9659 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9662 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9663 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9664 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9665 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9666 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9667 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9668 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9669 BUILD_BUG_SQE_ELEM(8, __u64, off);
9670 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9671 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9672 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9673 BUILD_BUG_SQE_ELEM(24, __u32, len);
9674 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9675 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9676 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9677 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9678 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9679 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9680 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9681 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9682 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9683 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9684 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9685 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9686 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9687 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9688 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9689 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9690 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9691 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9692 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9694 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9695 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9696 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9699 __initcall(io_uring_init);