1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
210 struct fixed_file_data {
211 struct fixed_file_table *table;
212 struct io_ring_ctx *ctx;
214 struct fixed_file_ref_node *node;
215 struct percpu_ref refs;
216 struct completion done;
217 struct list_head ref_list;
222 struct list_head list;
228 struct io_restriction {
229 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
230 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
231 u8 sqe_flags_allowed;
232 u8 sqe_flags_required;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list;
242 struct list_head ctx_new_list;
243 struct mutex ctx_lock;
245 struct task_struct *thread;
246 struct wait_queue_head wait;
251 struct percpu_ref refs;
252 } ____cacheline_aligned_in_smp;
256 unsigned int compat: 1;
257 unsigned int limit_mem: 1;
258 unsigned int cq_overflow_flushed: 1;
259 unsigned int drain_next: 1;
260 unsigned int eventfd_async: 1;
261 unsigned int restricted: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head;
278 unsigned sq_thread_idle;
279 unsigned cached_sq_dropped;
280 unsigned cached_cq_overflow;
281 unsigned long sq_check_overflow;
283 struct list_head defer_list;
284 struct list_head timeout_list;
285 struct list_head cq_overflow_list;
287 wait_queue_head_t inflight_wait;
288 struct io_uring_sqe *sq_sqes;
289 } ____cacheline_aligned_in_smp;
291 struct io_rings *rings;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct *sqo_task;
302 /* Only used for accounting purposes */
303 struct mm_struct *mm_account;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state *sqo_blkcg_css;
309 struct io_sq_data *sq_data; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait;
312 struct wait_queue_entry sqo_wait_entry;
313 struct list_head sqd_list;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data *file_data;
321 unsigned nr_user_files;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs;
325 struct io_mapped_ubuf *user_bufs;
327 struct user_struct *user;
329 const struct cred *creds;
333 unsigned int sessionid;
336 struct completion ref_comp;
337 struct completion sq_thread_comp;
339 /* if all else fails... */
340 struct io_kiocb *fallback_req;
342 #if defined(CONFIG_UNIX)
343 struct socket *ring_sock;
346 struct idr io_buffer_idr;
348 struct idr personality_idr;
351 unsigned cached_cq_tail;
354 atomic_t cq_timeouts;
355 unsigned long cq_check_overflow;
356 struct wait_queue_head cq_wait;
357 struct fasync_struct *cq_fasync;
358 struct eventfd_ctx *cq_ev_fd;
359 } ____cacheline_aligned_in_smp;
362 struct mutex uring_lock;
363 wait_queue_head_t wait;
364 } ____cacheline_aligned_in_smp;
367 spinlock_t completion_lock;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list;
376 struct hlist_head *cancel_hash;
377 unsigned cancel_hash_bits;
378 bool poll_multi_file;
380 spinlock_t inflight_lock;
381 struct list_head inflight_list;
382 } ____cacheline_aligned_in_smp;
384 struct delayed_work file_put_work;
385 struct llist_head file_put_llist;
387 struct work_struct exit_work;
388 struct io_restriction restrictions;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb {
398 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
409 struct file *put_file;
413 struct io_timeout_data {
414 struct io_kiocb *req;
415 struct hrtimer timer;
416 struct timespec64 ts;
417 enum hrtimer_mode mode;
422 struct sockaddr __user *addr;
423 int __user *addr_len;
425 unsigned long nofile;
445 struct list_head list;
448 struct io_timeout_rem {
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user *addr;
469 struct user_msghdr __user *umsg;
475 struct io_buffer *kbuf;
481 struct filename *filename;
483 unsigned long nofile;
486 struct io_files_update {
512 struct epoll_event event;
516 struct file *file_out;
517 struct file *file_in;
524 struct io_provide_buf {
538 const char __user *filename;
539 struct statx __user *buffer;
542 struct io_completion {
544 struct list_head list;
548 struct io_async_connect {
549 struct sockaddr_storage address;
552 struct io_async_msghdr {
553 struct iovec fast_iov[UIO_FASTIOV];
555 struct sockaddr __user *uaddr;
557 struct sockaddr_storage addr;
561 struct iovec fast_iov[UIO_FASTIOV];
562 const struct iovec *free_iovec;
563 struct iov_iter iter;
565 struct wait_page_queue wpq;
569 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
570 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
571 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
572 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
573 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
574 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
581 REQ_F_LINK_TIMEOUT_BIT,
583 REQ_F_NEED_CLEANUP_BIT,
585 REQ_F_BUFFER_SELECTED_BIT,
586 REQ_F_NO_FILE_TABLE_BIT,
587 REQ_F_WORK_INITIALIZED_BIT,
588 REQ_F_LTIMEOUT_ACTIVE_BIT,
590 /* not a real bit, just to check we're not overflowing the space */
596 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
597 /* drain existing IO first */
598 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
600 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
601 /* doesn't sever on completion < 0 */
602 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
604 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
605 /* IOSQE_BUFFER_SELECT */
606 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
609 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
610 /* fail rest of links */
611 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
612 /* on inflight list */
613 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
614 /* read/write uses file position */
615 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
616 /* must not punt to workers */
617 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
618 /* has or had linked timeout */
619 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
621 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
623 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
624 /* already went through poll handler */
625 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
626 /* buffer already selected */
627 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
628 /* doesn't need file table for this request */
629 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
630 /* io_wq_work is initialized */
631 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
632 /* linked timeout is active, i.e. prepared by link's head */
633 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
637 struct io_poll_iocb poll;
638 struct io_poll_iocb *double_poll;
642 * NOTE! Each of the iocb union members has the file pointer
643 * as the first entry in their struct definition. So you can
644 * access the file pointer through any of the sub-structs,
645 * or directly as just 'ki_filp' in this struct.
651 struct io_poll_iocb poll;
652 struct io_accept accept;
654 struct io_cancel cancel;
655 struct io_timeout timeout;
656 struct io_timeout_rem timeout_rem;
657 struct io_connect connect;
658 struct io_sr_msg sr_msg;
660 struct io_close close;
661 struct io_files_update files_update;
662 struct io_fadvise fadvise;
663 struct io_madvise madvise;
664 struct io_epoll epoll;
665 struct io_splice splice;
666 struct io_provide_buf pbuf;
667 struct io_statx statx;
668 /* use only after cleaning per-op data, see io_clean_op() */
669 struct io_completion compl;
672 /* opcode allocated if it needs to store data for async defer */
675 /* polled IO has completed */
681 struct io_ring_ctx *ctx;
684 struct task_struct *task;
687 struct list_head link_list;
690 * 1. used with ctx->iopoll_list with reads/writes
691 * 2. to track reqs with ->files (see io_op_def::file_table)
693 struct list_head inflight_entry;
695 struct percpu_ref *fixed_file_refs;
696 struct callback_head task_work;
697 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
698 struct hlist_node hash_node;
699 struct async_poll *apoll;
700 struct io_wq_work work;
703 struct io_defer_entry {
704 struct list_head list;
705 struct io_kiocb *req;
709 #define IO_IOPOLL_BATCH 8
711 struct io_comp_state {
713 struct list_head list;
714 struct io_ring_ctx *ctx;
717 struct io_submit_state {
718 struct blk_plug plug;
721 * io_kiocb alloc cache
723 void *reqs[IO_IOPOLL_BATCH];
724 unsigned int free_reqs;
727 * Batch completion logic
729 struct io_comp_state comp;
732 * File reference cache
736 unsigned int has_refs;
737 unsigned int ios_left;
741 /* needs req->file assigned */
742 unsigned needs_file : 1;
743 /* don't fail if file grab fails */
744 unsigned needs_file_no_error : 1;
745 /* hash wq insertion if file is a regular file */
746 unsigned hash_reg_file : 1;
747 /* unbound wq insertion if file is a non-regular file */
748 unsigned unbound_nonreg_file : 1;
749 /* opcode is not supported by this kernel */
750 unsigned not_supported : 1;
751 /* set if opcode supports polled "wait" */
753 unsigned pollout : 1;
754 /* op supports buffer selection */
755 unsigned buffer_select : 1;
756 /* must always have async data allocated */
757 unsigned needs_async_data : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
763 static const struct io_op_def io_op_defs[] = {
764 [IORING_OP_NOP] = {},
765 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
771 .async_size = sizeof(struct io_async_rw),
772 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
774 [IORING_OP_WRITEV] = {
777 .unbound_nonreg_file = 1,
779 .needs_async_data = 1,
780 .async_size = sizeof(struct io_async_rw),
781 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
784 [IORING_OP_FSYNC] = {
786 .work_flags = IO_WQ_WORK_BLKCG,
788 [IORING_OP_READ_FIXED] = {
790 .unbound_nonreg_file = 1,
792 .async_size = sizeof(struct io_async_rw),
793 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
795 [IORING_OP_WRITE_FIXED] = {
798 .unbound_nonreg_file = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
804 [IORING_OP_POLL_ADD] = {
806 .unbound_nonreg_file = 1,
808 [IORING_OP_POLL_REMOVE] = {},
809 [IORING_OP_SYNC_FILE_RANGE] = {
811 .work_flags = IO_WQ_WORK_BLKCG,
813 [IORING_OP_SENDMSG] = {
815 .unbound_nonreg_file = 1,
817 .needs_async_data = 1,
818 .async_size = sizeof(struct io_async_msghdr),
819 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
822 [IORING_OP_RECVMSG] = {
824 .unbound_nonreg_file = 1,
827 .needs_async_data = 1,
828 .async_size = sizeof(struct io_async_msghdr),
829 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
832 [IORING_OP_TIMEOUT] = {
833 .needs_async_data = 1,
834 .async_size = sizeof(struct io_timeout_data),
835 .work_flags = IO_WQ_WORK_MM,
837 [IORING_OP_TIMEOUT_REMOVE] = {},
838 [IORING_OP_ACCEPT] = {
840 .unbound_nonreg_file = 1,
842 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
844 [IORING_OP_ASYNC_CANCEL] = {},
845 [IORING_OP_LINK_TIMEOUT] = {
846 .needs_async_data = 1,
847 .async_size = sizeof(struct io_timeout_data),
848 .work_flags = IO_WQ_WORK_MM,
850 [IORING_OP_CONNECT] = {
852 .unbound_nonreg_file = 1,
854 .needs_async_data = 1,
855 .async_size = sizeof(struct io_async_connect),
856 .work_flags = IO_WQ_WORK_MM,
858 [IORING_OP_FALLOCATE] = {
860 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
862 [IORING_OP_OPENAT] = {
863 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
866 [IORING_OP_CLOSE] = {
868 .needs_file_no_error = 1,
869 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
871 [IORING_OP_FILES_UPDATE] = {
872 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
874 [IORING_OP_STATX] = {
875 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
876 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
880 .unbound_nonreg_file = 1,
883 .async_size = sizeof(struct io_async_rw),
884 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
886 [IORING_OP_WRITE] = {
888 .unbound_nonreg_file = 1,
890 .async_size = sizeof(struct io_async_rw),
891 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
894 [IORING_OP_FADVISE] = {
896 .work_flags = IO_WQ_WORK_BLKCG,
898 [IORING_OP_MADVISE] = {
899 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
903 .unbound_nonreg_file = 1,
905 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
909 .unbound_nonreg_file = 1,
912 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 [IORING_OP_OPENAT2] = {
915 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
918 [IORING_OP_EPOLL_CTL] = {
919 .unbound_nonreg_file = 1,
920 .work_flags = IO_WQ_WORK_FILES,
922 [IORING_OP_SPLICE] = {
925 .unbound_nonreg_file = 1,
926 .work_flags = IO_WQ_WORK_BLKCG,
928 [IORING_OP_PROVIDE_BUFFERS] = {},
929 [IORING_OP_REMOVE_BUFFERS] = {},
933 .unbound_nonreg_file = 1,
937 enum io_mem_account {
942 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
943 struct io_comp_state *cs);
944 static void io_cqring_fill_event(struct io_kiocb *req, long res);
945 static void io_put_req(struct io_kiocb *req);
946 static void io_put_req_deferred(struct io_kiocb *req, int nr);
947 static void io_double_put_req(struct io_kiocb *req);
948 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
949 static void __io_queue_linked_timeout(struct io_kiocb *req);
950 static void io_queue_linked_timeout(struct io_kiocb *req);
951 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
952 struct io_uring_files_update *ip,
954 static void __io_clean_op(struct io_kiocb *req);
955 static struct file *io_file_get(struct io_submit_state *state,
956 struct io_kiocb *req, int fd, bool fixed);
957 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
958 static void io_file_put_work(struct work_struct *work);
960 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
961 struct iovec **iovec, struct iov_iter *iter,
963 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
964 const struct iovec *fast_iov,
965 struct iov_iter *iter, bool force);
967 static struct kmem_cache *req_cachep;
969 static const struct file_operations io_uring_fops;
971 struct sock *io_uring_get_socket(struct file *file)
973 #if defined(CONFIG_UNIX)
974 if (file->f_op == &io_uring_fops) {
975 struct io_ring_ctx *ctx = file->private_data;
977 return ctx->ring_sock->sk;
982 EXPORT_SYMBOL(io_uring_get_socket);
984 static inline void io_clean_op(struct io_kiocb *req)
986 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
991 static void io_sq_thread_drop_mm(void)
993 struct mm_struct *mm = current->mm;
996 kthread_unuse_mm(mm);
1001 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1004 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1005 !ctx->sqo_task->mm ||
1006 !mmget_not_zero(ctx->sqo_task->mm)))
1008 kthread_use_mm(ctx->sqo_task->mm);
1014 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1015 struct io_kiocb *req)
1017 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1019 return __io_sq_thread_acquire_mm(ctx);
1022 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1023 struct cgroup_subsys_state **cur_css)
1026 #ifdef CONFIG_BLK_CGROUP
1027 /* puts the old one when swapping */
1028 if (*cur_css != ctx->sqo_blkcg_css) {
1029 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1030 *cur_css = ctx->sqo_blkcg_css;
1035 static void io_sq_thread_unassociate_blkcg(void)
1037 #ifdef CONFIG_BLK_CGROUP
1038 kthread_associate_blkcg(NULL);
1042 static inline void req_set_fail_links(struct io_kiocb *req)
1044 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1045 req->flags |= REQ_F_FAIL_LINK;
1049 * None of these are dereferenced, they are simply used to check if any of
1050 * them have changed. If we're under current and check they are still the
1051 * same, we're fine to grab references to them for actual out-of-line use.
1053 static void io_init_identity(struct io_identity *id)
1055 id->files = current->files;
1056 id->mm = current->mm;
1057 #ifdef CONFIG_BLK_CGROUP
1059 id->blkcg_css = blkcg_css();
1062 id->creds = current_cred();
1063 id->nsproxy = current->nsproxy;
1064 id->fs = current->fs;
1065 id->fsize = rlimit(RLIMIT_FSIZE);
1067 id->loginuid = current->loginuid;
1068 id->sessionid = current->sessionid;
1070 refcount_set(&id->count, 1);
1073 static inline void __io_req_init_async(struct io_kiocb *req)
1075 memset(&req->work, 0, sizeof(req->work));
1076 req->flags |= REQ_F_WORK_INITIALIZED;
1080 * Note: must call io_req_init_async() for the first time you
1081 * touch any members of io_wq_work.
1083 static inline void io_req_init_async(struct io_kiocb *req)
1085 struct io_uring_task *tctx = current->io_uring;
1087 if (req->flags & REQ_F_WORK_INITIALIZED)
1090 __io_req_init_async(req);
1092 /* Grab a ref if this isn't our static identity */
1093 req->work.identity = tctx->identity;
1094 if (tctx->identity != &tctx->__identity)
1095 refcount_inc(&req->work.identity->count);
1098 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1100 return ctx->flags & IORING_SETUP_SQPOLL;
1103 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1105 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1107 complete(&ctx->ref_comp);
1110 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1112 return !req->timeout.off;
1115 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1117 struct io_ring_ctx *ctx;
1120 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1124 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1125 if (!ctx->fallback_req)
1129 * Use 5 bits less than the max cq entries, that should give us around
1130 * 32 entries per hash list if totally full and uniformly spread.
1132 hash_bits = ilog2(p->cq_entries);
1136 ctx->cancel_hash_bits = hash_bits;
1137 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1139 if (!ctx->cancel_hash)
1141 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1143 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1144 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1147 ctx->flags = p->flags;
1148 init_waitqueue_head(&ctx->sqo_sq_wait);
1149 INIT_LIST_HEAD(&ctx->sqd_list);
1150 init_waitqueue_head(&ctx->cq_wait);
1151 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1152 init_completion(&ctx->ref_comp);
1153 init_completion(&ctx->sq_thread_comp);
1154 idr_init(&ctx->io_buffer_idr);
1155 idr_init(&ctx->personality_idr);
1156 mutex_init(&ctx->uring_lock);
1157 init_waitqueue_head(&ctx->wait);
1158 spin_lock_init(&ctx->completion_lock);
1159 INIT_LIST_HEAD(&ctx->iopoll_list);
1160 INIT_LIST_HEAD(&ctx->defer_list);
1161 INIT_LIST_HEAD(&ctx->timeout_list);
1162 init_waitqueue_head(&ctx->inflight_wait);
1163 spin_lock_init(&ctx->inflight_lock);
1164 INIT_LIST_HEAD(&ctx->inflight_list);
1165 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1166 init_llist_head(&ctx->file_put_llist);
1169 if (ctx->fallback_req)
1170 kmem_cache_free(req_cachep, ctx->fallback_req);
1171 kfree(ctx->cancel_hash);
1176 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1178 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1179 struct io_ring_ctx *ctx = req->ctx;
1181 return seq != ctx->cached_cq_tail
1182 + READ_ONCE(ctx->cached_cq_overflow);
1188 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1190 struct io_rings *rings = ctx->rings;
1192 /* order cqe stores with ring update */
1193 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1195 if (wq_has_sleeper(&ctx->cq_wait)) {
1196 wake_up_interruptible(&ctx->cq_wait);
1197 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1201 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1203 if (req->work.identity == &tctx->__identity)
1205 if (refcount_dec_and_test(&req->work.identity->count))
1206 kfree(req->work.identity);
1209 static void io_req_clean_work(struct io_kiocb *req)
1211 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1214 req->flags &= ~REQ_F_WORK_INITIALIZED;
1216 if (req->work.flags & IO_WQ_WORK_MM) {
1217 mmdrop(req->work.identity->mm);
1218 req->work.flags &= ~IO_WQ_WORK_MM;
1220 #ifdef CONFIG_BLK_CGROUP
1221 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1222 css_put(req->work.identity->blkcg_css);
1223 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1226 if (req->work.flags & IO_WQ_WORK_CREDS) {
1227 put_cred(req->work.identity->creds);
1228 req->work.flags &= ~IO_WQ_WORK_CREDS;
1230 if (req->work.flags & IO_WQ_WORK_FS) {
1231 struct fs_struct *fs = req->work.identity->fs;
1233 spin_lock(&req->work.identity->fs->lock);
1236 spin_unlock(&req->work.identity->fs->lock);
1239 req->work.flags &= ~IO_WQ_WORK_FS;
1242 io_put_identity(req->task->io_uring, req);
1246 * Create a private copy of io_identity, since some fields don't match
1247 * the current context.
1249 static bool io_identity_cow(struct io_kiocb *req)
1251 struct io_uring_task *tctx = current->io_uring;
1252 const struct cred *creds = NULL;
1253 struct io_identity *id;
1255 if (req->work.flags & IO_WQ_WORK_CREDS)
1256 creds = req->work.identity->creds;
1258 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1259 if (unlikely(!id)) {
1260 req->work.flags |= IO_WQ_WORK_CANCEL;
1265 * We can safely just re-init the creds we copied Either the field
1266 * matches the current one, or we haven't grabbed it yet. The only
1267 * exception is ->creds, through registered personalities, so handle
1268 * that one separately.
1270 io_init_identity(id);
1272 req->work.identity->creds = creds;
1274 /* add one for this request */
1275 refcount_inc(&id->count);
1277 /* drop old identity, assign new one. one ref for req, one for tctx */
1278 if (req->work.identity != tctx->identity &&
1279 refcount_sub_and_test(2, &req->work.identity->count))
1280 kfree(req->work.identity);
1282 req->work.identity = id;
1283 tctx->identity = id;
1287 static bool io_grab_identity(struct io_kiocb *req)
1289 const struct io_op_def *def = &io_op_defs[req->opcode];
1290 struct io_identity *id = req->work.identity;
1291 struct io_ring_ctx *ctx = req->ctx;
1293 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1294 if (id->fsize != rlimit(RLIMIT_FSIZE))
1296 req->work.flags |= IO_WQ_WORK_FSIZE;
1299 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1300 (def->work_flags & IO_WQ_WORK_FILES) &&
1301 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1302 if (id->files != current->files ||
1303 id->nsproxy != current->nsproxy)
1305 atomic_inc(&id->files->count);
1306 get_nsproxy(id->nsproxy);
1307 req->flags |= REQ_F_INFLIGHT;
1309 spin_lock_irq(&ctx->inflight_lock);
1310 list_add(&req->inflight_entry, &ctx->inflight_list);
1311 spin_unlock_irq(&ctx->inflight_lock);
1312 req->work.flags |= IO_WQ_WORK_FILES;
1314 #ifdef CONFIG_BLK_CGROUP
1315 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1316 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1318 if (id->blkcg_css != blkcg_css()) {
1323 * This should be rare, either the cgroup is dying or the task
1324 * is moving cgroups. Just punt to root for the handful of ios.
1326 if (css_tryget_online(id->blkcg_css))
1327 req->work.flags |= IO_WQ_WORK_BLKCG;
1331 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1332 if (id->creds != current_cred())
1334 get_cred(id->creds);
1335 req->work.flags |= IO_WQ_WORK_CREDS;
1338 if (!uid_eq(current->loginuid, id->loginuid) ||
1339 current->sessionid != id->sessionid)
1342 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1343 (def->work_flags & IO_WQ_WORK_FS)) {
1344 if (current->fs != id->fs)
1346 spin_lock(&id->fs->lock);
1347 if (!id->fs->in_exec) {
1349 req->work.flags |= IO_WQ_WORK_FS;
1351 req->work.flags |= IO_WQ_WORK_CANCEL;
1353 spin_unlock(¤t->fs->lock);
1359 static void io_prep_async_work(struct io_kiocb *req)
1361 const struct io_op_def *def = &io_op_defs[req->opcode];
1362 struct io_ring_ctx *ctx = req->ctx;
1363 struct io_identity *id;
1365 io_req_init_async(req);
1366 id = req->work.identity;
1368 if (req->flags & REQ_F_FORCE_ASYNC)
1369 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1371 if (req->flags & REQ_F_ISREG) {
1372 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1373 io_wq_hash_work(&req->work, file_inode(req->file));
1375 if (def->unbound_nonreg_file)
1376 req->work.flags |= IO_WQ_WORK_UNBOUND;
1379 /* ->mm can never change on us */
1380 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1381 (def->work_flags & IO_WQ_WORK_MM)) {
1383 req->work.flags |= IO_WQ_WORK_MM;
1386 /* if we fail grabbing identity, we must COW, regrab, and retry */
1387 if (io_grab_identity(req))
1390 if (!io_identity_cow(req))
1393 /* can't fail at this point */
1394 if (!io_grab_identity(req))
1398 static void io_prep_async_link(struct io_kiocb *req)
1400 struct io_kiocb *cur;
1402 io_prep_async_work(req);
1403 if (req->flags & REQ_F_LINK_HEAD)
1404 list_for_each_entry(cur, &req->link_list, link_list)
1405 io_prep_async_work(cur);
1408 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1410 struct io_ring_ctx *ctx = req->ctx;
1411 struct io_kiocb *link = io_prep_linked_timeout(req);
1413 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1414 &req->work, req->flags);
1415 io_wq_enqueue(ctx->io_wq, &req->work);
1419 static void io_queue_async_work(struct io_kiocb *req)
1421 struct io_kiocb *link;
1423 /* init ->work of the whole link before punting */
1424 io_prep_async_link(req);
1425 link = __io_queue_async_work(req);
1428 io_queue_linked_timeout(link);
1431 static void io_kill_timeout(struct io_kiocb *req)
1433 struct io_timeout_data *io = req->async_data;
1436 ret = hrtimer_try_to_cancel(&io->timer);
1438 atomic_set(&req->ctx->cq_timeouts,
1439 atomic_read(&req->ctx->cq_timeouts) + 1);
1440 list_del_init(&req->timeout.list);
1441 io_cqring_fill_event(req, 0);
1442 io_put_req_deferred(req, 1);
1446 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1448 struct io_ring_ctx *ctx = req->ctx;
1450 if (!tsk || req->task == tsk)
1452 if (ctx->flags & IORING_SETUP_SQPOLL) {
1453 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1460 * Returns true if we found and killed one or more timeouts
1462 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1464 struct io_kiocb *req, *tmp;
1467 spin_lock_irq(&ctx->completion_lock);
1468 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1469 if (io_task_match(req, tsk)) {
1470 io_kill_timeout(req);
1474 spin_unlock_irq(&ctx->completion_lock);
1475 return canceled != 0;
1478 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1481 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1482 struct io_defer_entry, list);
1483 struct io_kiocb *link;
1485 if (req_need_defer(de->req, de->seq))
1487 list_del_init(&de->list);
1488 /* punt-init is done before queueing for defer */
1489 link = __io_queue_async_work(de->req);
1491 __io_queue_linked_timeout(link);
1492 /* drop submission reference */
1493 io_put_req_deferred(link, 1);
1496 } while (!list_empty(&ctx->defer_list));
1499 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1501 while (!list_empty(&ctx->timeout_list)) {
1502 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1503 struct io_kiocb, timeout.list);
1505 if (io_is_timeout_noseq(req))
1507 if (req->timeout.target_seq != ctx->cached_cq_tail
1508 - atomic_read(&ctx->cq_timeouts))
1511 list_del_init(&req->timeout.list);
1512 io_kill_timeout(req);
1516 static void io_commit_cqring(struct io_ring_ctx *ctx)
1518 io_flush_timeouts(ctx);
1519 __io_commit_cqring(ctx);
1521 if (unlikely(!list_empty(&ctx->defer_list)))
1522 __io_queue_deferred(ctx);
1525 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1527 struct io_rings *r = ctx->rings;
1529 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1532 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1534 struct io_rings *rings = ctx->rings;
1537 tail = ctx->cached_cq_tail;
1539 * writes to the cq entry need to come after reading head; the
1540 * control dependency is enough as we're using WRITE_ONCE to
1543 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1546 ctx->cached_cq_tail++;
1547 return &rings->cqes[tail & ctx->cq_mask];
1550 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1554 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1556 if (!ctx->eventfd_async)
1558 return io_wq_current_is_worker();
1561 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1563 if (waitqueue_active(&ctx->wait))
1564 wake_up(&ctx->wait);
1565 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1566 wake_up(&ctx->sq_data->wait);
1567 if (io_should_trigger_evfd(ctx))
1568 eventfd_signal(ctx->cq_ev_fd, 1);
1571 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1573 if (list_empty(&ctx->cq_overflow_list)) {
1574 clear_bit(0, &ctx->sq_check_overflow);
1575 clear_bit(0, &ctx->cq_check_overflow);
1576 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1580 static inline bool io_match_files(struct io_kiocb *req,
1581 struct files_struct *files)
1585 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1586 (req->work.flags & IO_WQ_WORK_FILES))
1587 return req->work.identity->files == files;
1591 /* Returns true if there are no backlogged entries after the flush */
1592 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1593 struct task_struct *tsk,
1594 struct files_struct *files)
1596 struct io_rings *rings = ctx->rings;
1597 struct io_kiocb *req, *tmp;
1598 struct io_uring_cqe *cqe;
1599 unsigned long flags;
1603 if (list_empty_careful(&ctx->cq_overflow_list))
1605 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1606 rings->cq_ring_entries))
1610 spin_lock_irqsave(&ctx->completion_lock, flags);
1612 /* if force is set, the ring is going away. always drop after that */
1614 ctx->cq_overflow_flushed = 1;
1617 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1618 if (tsk && req->task != tsk)
1620 if (!io_match_files(req, files))
1623 cqe = io_get_cqring(ctx);
1627 list_move(&req->compl.list, &list);
1629 WRITE_ONCE(cqe->user_data, req->user_data);
1630 WRITE_ONCE(cqe->res, req->result);
1631 WRITE_ONCE(cqe->flags, req->compl.cflags);
1633 ctx->cached_cq_overflow++;
1634 WRITE_ONCE(ctx->rings->cq_overflow,
1635 ctx->cached_cq_overflow);
1639 io_commit_cqring(ctx);
1640 io_cqring_mark_overflow(ctx);
1642 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1643 io_cqring_ev_posted(ctx);
1645 while (!list_empty(&list)) {
1646 req = list_first_entry(&list, struct io_kiocb, compl.list);
1647 list_del(&req->compl.list);
1654 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1656 struct io_ring_ctx *ctx = req->ctx;
1657 struct io_uring_cqe *cqe;
1659 trace_io_uring_complete(ctx, req->user_data, res);
1662 * If we can't get a cq entry, userspace overflowed the
1663 * submission (by quite a lot). Increment the overflow count in
1666 cqe = io_get_cqring(ctx);
1668 WRITE_ONCE(cqe->user_data, req->user_data);
1669 WRITE_ONCE(cqe->res, res);
1670 WRITE_ONCE(cqe->flags, cflags);
1671 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1673 * If we're in ring overflow flush mode, or in task cancel mode,
1674 * then we cannot store the request for later flushing, we need
1675 * to drop it on the floor.
1677 ctx->cached_cq_overflow++;
1678 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1680 if (list_empty(&ctx->cq_overflow_list)) {
1681 set_bit(0, &ctx->sq_check_overflow);
1682 set_bit(0, &ctx->cq_check_overflow);
1683 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1687 req->compl.cflags = cflags;
1688 refcount_inc(&req->refs);
1689 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1693 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1695 __io_cqring_fill_event(req, res, 0);
1698 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1700 struct io_ring_ctx *ctx = req->ctx;
1701 unsigned long flags;
1703 spin_lock_irqsave(&ctx->completion_lock, flags);
1704 __io_cqring_fill_event(req, res, cflags);
1705 io_commit_cqring(ctx);
1706 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1708 io_cqring_ev_posted(ctx);
1711 static void io_submit_flush_completions(struct io_comp_state *cs)
1713 struct io_ring_ctx *ctx = cs->ctx;
1715 spin_lock_irq(&ctx->completion_lock);
1716 while (!list_empty(&cs->list)) {
1717 struct io_kiocb *req;
1719 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1720 list_del(&req->compl.list);
1721 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1724 * io_free_req() doesn't care about completion_lock unless one
1725 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1726 * because of a potential deadlock with req->work.fs->lock
1728 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1729 |REQ_F_WORK_INITIALIZED)) {
1730 spin_unlock_irq(&ctx->completion_lock);
1732 spin_lock_irq(&ctx->completion_lock);
1737 io_commit_cqring(ctx);
1738 spin_unlock_irq(&ctx->completion_lock);
1740 io_cqring_ev_posted(ctx);
1744 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1745 struct io_comp_state *cs)
1748 io_cqring_add_event(req, res, cflags);
1753 req->compl.cflags = cflags;
1754 list_add_tail(&req->compl.list, &cs->list);
1756 io_submit_flush_completions(cs);
1760 static void io_req_complete(struct io_kiocb *req, long res)
1762 __io_req_complete(req, res, 0, NULL);
1765 static inline bool io_is_fallback_req(struct io_kiocb *req)
1767 return req == (struct io_kiocb *)
1768 ((unsigned long) req->ctx->fallback_req & ~1UL);
1771 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1773 struct io_kiocb *req;
1775 req = ctx->fallback_req;
1776 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1782 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1783 struct io_submit_state *state)
1785 if (!state->free_reqs) {
1786 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1790 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1791 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1794 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1795 * retry single alloc to be on the safe side.
1797 if (unlikely(ret <= 0)) {
1798 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1799 if (!state->reqs[0])
1803 state->free_reqs = ret;
1807 return state->reqs[state->free_reqs];
1809 return io_get_fallback_req(ctx);
1812 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1816 percpu_ref_put(req->fixed_file_refs);
1821 static void io_dismantle_req(struct io_kiocb *req)
1825 if (req->async_data)
1826 kfree(req->async_data);
1828 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1830 io_req_clean_work(req);
1833 static void __io_free_req(struct io_kiocb *req)
1835 struct io_uring_task *tctx = req->task->io_uring;
1836 struct io_ring_ctx *ctx = req->ctx;
1838 io_dismantle_req(req);
1840 percpu_counter_dec(&tctx->inflight);
1842 wake_up(&tctx->wait);
1843 put_task_struct(req->task);
1845 if (likely(!io_is_fallback_req(req)))
1846 kmem_cache_free(req_cachep, req);
1848 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1849 percpu_ref_put(&ctx->refs);
1852 static void io_kill_linked_timeout(struct io_kiocb *req)
1854 struct io_ring_ctx *ctx = req->ctx;
1855 struct io_kiocb *link;
1856 bool cancelled = false;
1857 unsigned long flags;
1859 spin_lock_irqsave(&ctx->completion_lock, flags);
1860 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1863 * Can happen if a linked timeout fired and link had been like
1864 * req -> link t-out -> link t-out [-> ...]
1866 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1867 struct io_timeout_data *io = link->async_data;
1870 list_del_init(&link->link_list);
1871 ret = hrtimer_try_to_cancel(&io->timer);
1873 io_cqring_fill_event(link, -ECANCELED);
1874 io_commit_cqring(ctx);
1878 req->flags &= ~REQ_F_LINK_TIMEOUT;
1879 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1882 io_cqring_ev_posted(ctx);
1887 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1889 struct io_kiocb *nxt;
1892 * The list should never be empty when we are called here. But could
1893 * potentially happen if the chain is messed up, check to be on the
1896 if (unlikely(list_empty(&req->link_list)))
1899 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1900 list_del_init(&req->link_list);
1901 if (!list_empty(&nxt->link_list))
1902 nxt->flags |= REQ_F_LINK_HEAD;
1907 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1909 static void io_fail_links(struct io_kiocb *req)
1911 struct io_ring_ctx *ctx = req->ctx;
1912 unsigned long flags;
1914 spin_lock_irqsave(&ctx->completion_lock, flags);
1915 while (!list_empty(&req->link_list)) {
1916 struct io_kiocb *link = list_first_entry(&req->link_list,
1917 struct io_kiocb, link_list);
1919 list_del_init(&link->link_list);
1920 trace_io_uring_fail_link(req, link);
1922 io_cqring_fill_event(link, -ECANCELED);
1925 * It's ok to free under spinlock as they're not linked anymore,
1926 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1929 if (link->flags & REQ_F_WORK_INITIALIZED)
1930 io_put_req_deferred(link, 2);
1932 io_double_put_req(link);
1935 io_commit_cqring(ctx);
1936 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1938 io_cqring_ev_posted(ctx);
1941 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1943 req->flags &= ~REQ_F_LINK_HEAD;
1944 if (req->flags & REQ_F_LINK_TIMEOUT)
1945 io_kill_linked_timeout(req);
1948 * If LINK is set, we have dependent requests in this chain. If we
1949 * didn't fail this request, queue the first one up, moving any other
1950 * dependencies to the next request. In case of failure, fail the rest
1953 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1954 return io_req_link_next(req);
1959 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1961 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1963 return __io_req_find_next(req);
1966 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1968 struct task_struct *tsk = req->task;
1969 struct io_ring_ctx *ctx = req->ctx;
1972 if (tsk->flags & PF_EXITING)
1976 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1977 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1978 * processing task_work. There's no reliable way to tell if TWA_RESUME
1982 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1983 notify = TWA_SIGNAL;
1985 ret = task_work_add(tsk, &req->task_work, notify);
1987 wake_up_process(tsk);
1992 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1994 struct io_ring_ctx *ctx = req->ctx;
1996 spin_lock_irq(&ctx->completion_lock);
1997 io_cqring_fill_event(req, error);
1998 io_commit_cqring(ctx);
1999 spin_unlock_irq(&ctx->completion_lock);
2001 io_cqring_ev_posted(ctx);
2002 req_set_fail_links(req);
2003 io_double_put_req(req);
2006 static void io_req_task_cancel(struct callback_head *cb)
2008 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2009 struct io_ring_ctx *ctx = req->ctx;
2011 __io_req_task_cancel(req, -ECANCELED);
2012 percpu_ref_put(&ctx->refs);
2015 static void __io_req_task_submit(struct io_kiocb *req)
2017 struct io_ring_ctx *ctx = req->ctx;
2019 if (!__io_sq_thread_acquire_mm(ctx)) {
2020 mutex_lock(&ctx->uring_lock);
2021 __io_queue_sqe(req, NULL);
2022 mutex_unlock(&ctx->uring_lock);
2024 __io_req_task_cancel(req, -EFAULT);
2028 static void io_req_task_submit(struct callback_head *cb)
2030 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2031 struct io_ring_ctx *ctx = req->ctx;
2033 __io_req_task_submit(req);
2034 percpu_ref_put(&ctx->refs);
2037 static void io_req_task_queue(struct io_kiocb *req)
2041 init_task_work(&req->task_work, io_req_task_submit);
2042 percpu_ref_get(&req->ctx->refs);
2044 ret = io_req_task_work_add(req, true);
2045 if (unlikely(ret)) {
2046 struct task_struct *tsk;
2048 init_task_work(&req->task_work, io_req_task_cancel);
2049 tsk = io_wq_get_task(req->ctx->io_wq);
2050 task_work_add(tsk, &req->task_work, 0);
2051 wake_up_process(tsk);
2055 static void io_queue_next(struct io_kiocb *req)
2057 struct io_kiocb *nxt = io_req_find_next(req);
2060 io_req_task_queue(nxt);
2063 static void io_free_req(struct io_kiocb *req)
2070 void *reqs[IO_IOPOLL_BATCH];
2073 struct task_struct *task;
2077 static inline void io_init_req_batch(struct req_batch *rb)
2084 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2085 struct req_batch *rb)
2087 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2088 percpu_ref_put_many(&ctx->refs, rb->to_free);
2092 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2093 struct req_batch *rb)
2096 __io_req_free_batch_flush(ctx, rb);
2098 struct io_uring_task *tctx = rb->task->io_uring;
2100 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2101 put_task_struct_many(rb->task, rb->task_refs);
2106 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2108 if (unlikely(io_is_fallback_req(req))) {
2112 if (req->flags & REQ_F_LINK_HEAD)
2115 if (req->task != rb->task) {
2117 struct io_uring_task *tctx = rb->task->io_uring;
2119 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2120 put_task_struct_many(rb->task, rb->task_refs);
2122 rb->task = req->task;
2127 io_dismantle_req(req);
2128 rb->reqs[rb->to_free++] = req;
2129 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2130 __io_req_free_batch_flush(req->ctx, rb);
2134 * Drop reference to request, return next in chain (if there is one) if this
2135 * was the last reference to this request.
2137 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2139 struct io_kiocb *nxt = NULL;
2141 if (refcount_dec_and_test(&req->refs)) {
2142 nxt = io_req_find_next(req);
2148 static void io_put_req(struct io_kiocb *req)
2150 if (refcount_dec_and_test(&req->refs))
2154 static void io_put_req_deferred_cb(struct callback_head *cb)
2156 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2161 static void io_free_req_deferred(struct io_kiocb *req)
2165 init_task_work(&req->task_work, io_put_req_deferred_cb);
2166 ret = io_req_task_work_add(req, true);
2167 if (unlikely(ret)) {
2168 struct task_struct *tsk;
2170 tsk = io_wq_get_task(req->ctx->io_wq);
2171 task_work_add(tsk, &req->task_work, 0);
2172 wake_up_process(tsk);
2176 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2178 if (refcount_sub_and_test(refs, &req->refs))
2179 io_free_req_deferred(req);
2182 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2184 struct io_kiocb *nxt;
2187 * A ref is owned by io-wq in which context we're. So, if that's the
2188 * last one, it's safe to steal next work. False negatives are Ok,
2189 * it just will be re-punted async in io_put_work()
2191 if (refcount_read(&req->refs) != 1)
2194 nxt = io_req_find_next(req);
2195 return nxt ? &nxt->work : NULL;
2198 static void io_double_put_req(struct io_kiocb *req)
2200 /* drop both submit and complete references */
2201 if (refcount_sub_and_test(2, &req->refs))
2205 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2207 struct io_rings *rings = ctx->rings;
2209 if (test_bit(0, &ctx->cq_check_overflow)) {
2211 * noflush == true is from the waitqueue handler, just ensure
2212 * we wake up the task, and the next invocation will flush the
2213 * entries. We cannot safely to it from here.
2215 if (noflush && !list_empty(&ctx->cq_overflow_list))
2218 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2221 /* See comment at the top of this file */
2223 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2226 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2228 struct io_rings *rings = ctx->rings;
2230 /* make sure SQ entry isn't read before tail */
2231 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2234 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2236 unsigned int cflags;
2238 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2239 cflags |= IORING_CQE_F_BUFFER;
2240 req->flags &= ~REQ_F_BUFFER_SELECTED;
2245 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2247 struct io_buffer *kbuf;
2249 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2250 return io_put_kbuf(req, kbuf);
2253 static inline bool io_run_task_work(void)
2256 * Not safe to run on exiting task, and the task_work handling will
2257 * not add work to such a task.
2259 if (unlikely(current->flags & PF_EXITING))
2261 if (current->task_works) {
2262 __set_current_state(TASK_RUNNING);
2270 static void io_iopoll_queue(struct list_head *again)
2272 struct io_kiocb *req;
2275 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2276 list_del(&req->inflight_entry);
2277 __io_complete_rw(req, -EAGAIN, 0, NULL);
2278 } while (!list_empty(again));
2282 * Find and free completed poll iocbs
2284 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2285 struct list_head *done)
2287 struct req_batch rb;
2288 struct io_kiocb *req;
2291 /* order with ->result store in io_complete_rw_iopoll() */
2294 io_init_req_batch(&rb);
2295 while (!list_empty(done)) {
2298 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2299 if (READ_ONCE(req->result) == -EAGAIN) {
2301 req->iopoll_completed = 0;
2302 list_move_tail(&req->inflight_entry, &again);
2305 list_del(&req->inflight_entry);
2307 if (req->flags & REQ_F_BUFFER_SELECTED)
2308 cflags = io_put_rw_kbuf(req);
2310 __io_cqring_fill_event(req, req->result, cflags);
2313 if (refcount_dec_and_test(&req->refs))
2314 io_req_free_batch(&rb, req);
2317 io_commit_cqring(ctx);
2318 if (ctx->flags & IORING_SETUP_SQPOLL)
2319 io_cqring_ev_posted(ctx);
2320 io_req_free_batch_finish(ctx, &rb);
2322 if (!list_empty(&again))
2323 io_iopoll_queue(&again);
2326 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2329 struct io_kiocb *req, *tmp;
2335 * Only spin for completions if we don't have multiple devices hanging
2336 * off our complete list, and we're under the requested amount.
2338 spin = !ctx->poll_multi_file && *nr_events < min;
2341 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2342 struct kiocb *kiocb = &req->rw.kiocb;
2345 * Move completed and retryable entries to our local lists.
2346 * If we find a request that requires polling, break out
2347 * and complete those lists first, if we have entries there.
2349 if (READ_ONCE(req->iopoll_completed)) {
2350 list_move_tail(&req->inflight_entry, &done);
2353 if (!list_empty(&done))
2356 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2360 /* iopoll may have completed current req */
2361 if (READ_ONCE(req->iopoll_completed))
2362 list_move_tail(&req->inflight_entry, &done);
2369 if (!list_empty(&done))
2370 io_iopoll_complete(ctx, nr_events, &done);
2376 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2377 * non-spinning poll check - we'll still enter the driver poll loop, but only
2378 * as a non-spinning completion check.
2380 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2383 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2386 ret = io_do_iopoll(ctx, nr_events, min);
2389 if (*nr_events >= min)
2397 * We can't just wait for polled events to come to us, we have to actively
2398 * find and complete them.
2400 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2402 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2405 mutex_lock(&ctx->uring_lock);
2406 while (!list_empty(&ctx->iopoll_list)) {
2407 unsigned int nr_events = 0;
2409 io_do_iopoll(ctx, &nr_events, 0);
2411 /* let it sleep and repeat later if can't complete a request */
2415 * Ensure we allow local-to-the-cpu processing to take place,
2416 * in this case we need to ensure that we reap all events.
2417 * Also let task_work, etc. to progress by releasing the mutex
2419 if (need_resched()) {
2420 mutex_unlock(&ctx->uring_lock);
2422 mutex_lock(&ctx->uring_lock);
2425 mutex_unlock(&ctx->uring_lock);
2428 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2430 unsigned int nr_events = 0;
2431 int iters = 0, ret = 0;
2434 * We disallow the app entering submit/complete with polling, but we
2435 * still need to lock the ring to prevent racing with polled issue
2436 * that got punted to a workqueue.
2438 mutex_lock(&ctx->uring_lock);
2441 * Don't enter poll loop if we already have events pending.
2442 * If we do, we can potentially be spinning for commands that
2443 * already triggered a CQE (eg in error).
2445 if (io_cqring_events(ctx, false))
2449 * If a submit got punted to a workqueue, we can have the
2450 * application entering polling for a command before it gets
2451 * issued. That app will hold the uring_lock for the duration
2452 * of the poll right here, so we need to take a breather every
2453 * now and then to ensure that the issue has a chance to add
2454 * the poll to the issued list. Otherwise we can spin here
2455 * forever, while the workqueue is stuck trying to acquire the
2458 if (!(++iters & 7)) {
2459 mutex_unlock(&ctx->uring_lock);
2461 mutex_lock(&ctx->uring_lock);
2464 ret = io_iopoll_getevents(ctx, &nr_events, min);
2468 } while (min && !nr_events && !need_resched());
2470 mutex_unlock(&ctx->uring_lock);
2474 static void kiocb_end_write(struct io_kiocb *req)
2477 * Tell lockdep we inherited freeze protection from submission
2480 if (req->flags & REQ_F_ISREG) {
2481 struct inode *inode = file_inode(req->file);
2483 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2485 file_end_write(req->file);
2488 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2489 struct io_comp_state *cs)
2491 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2494 if (kiocb->ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
2497 if (res != req->result)
2498 req_set_fail_links(req);
2499 if (req->flags & REQ_F_BUFFER_SELECTED)
2500 cflags = io_put_rw_kbuf(req);
2501 __io_req_complete(req, res, cflags, cs);
2505 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2507 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2508 ssize_t ret = -ECANCELED;
2509 struct iov_iter iter;
2517 switch (req->opcode) {
2518 case IORING_OP_READV:
2519 case IORING_OP_READ_FIXED:
2520 case IORING_OP_READ:
2523 case IORING_OP_WRITEV:
2524 case IORING_OP_WRITE_FIXED:
2525 case IORING_OP_WRITE:
2529 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2534 if (!req->async_data) {
2535 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2538 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2546 req_set_fail_links(req);
2547 io_req_complete(req, ret);
2552 static bool io_rw_reissue(struct io_kiocb *req, long res)
2555 umode_t mode = file_inode(req->file)->i_mode;
2558 if (!S_ISBLK(mode) && !S_ISREG(mode))
2560 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2563 ret = io_sq_thread_acquire_mm(req->ctx, req);
2565 if (io_resubmit_prep(req, ret)) {
2566 refcount_inc(&req->refs);
2567 io_queue_async_work(req);
2575 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2576 struct io_comp_state *cs)
2578 if (!io_rw_reissue(req, res))
2579 io_complete_rw_common(&req->rw.kiocb, res, cs);
2582 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2584 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2586 __io_complete_rw(req, res, res2, NULL);
2589 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2591 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2593 if (kiocb->ki_flags & IOCB_WRITE)
2594 kiocb_end_write(req);
2596 if (res != -EAGAIN && res != req->result)
2597 req_set_fail_links(req);
2599 WRITE_ONCE(req->result, res);
2600 /* order with io_poll_complete() checking ->result */
2602 WRITE_ONCE(req->iopoll_completed, 1);
2606 * After the iocb has been issued, it's safe to be found on the poll list.
2607 * Adding the kiocb to the list AFTER submission ensures that we don't
2608 * find it from a io_iopoll_getevents() thread before the issuer is done
2609 * accessing the kiocb cookie.
2611 static void io_iopoll_req_issued(struct io_kiocb *req)
2613 struct io_ring_ctx *ctx = req->ctx;
2616 * Track whether we have multiple files in our lists. This will impact
2617 * how we do polling eventually, not spinning if we're on potentially
2618 * different devices.
2620 if (list_empty(&ctx->iopoll_list)) {
2621 ctx->poll_multi_file = false;
2622 } else if (!ctx->poll_multi_file) {
2623 struct io_kiocb *list_req;
2625 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2627 if (list_req->file != req->file)
2628 ctx->poll_multi_file = true;
2632 * For fast devices, IO may have already completed. If it has, add
2633 * it to the front so we find it first.
2635 if (READ_ONCE(req->iopoll_completed))
2636 list_add(&req->inflight_entry, &ctx->iopoll_list);
2638 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2640 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2641 wq_has_sleeper(&ctx->sq_data->wait))
2642 wake_up(&ctx->sq_data->wait);
2645 static void __io_state_file_put(struct io_submit_state *state)
2647 if (state->has_refs)
2648 fput_many(state->file, state->has_refs);
2652 static inline void io_state_file_put(struct io_submit_state *state)
2655 __io_state_file_put(state);
2659 * Get as many references to a file as we have IOs left in this submission,
2660 * assuming most submissions are for one file, or at least that each file
2661 * has more than one submission.
2663 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2669 if (state->fd == fd) {
2673 __io_state_file_put(state);
2675 state->file = fget_many(fd, state->ios_left);
2680 state->has_refs = state->ios_left - 1;
2684 static bool io_bdev_nowait(struct block_device *bdev)
2687 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2694 * If we tracked the file through the SCM inflight mechanism, we could support
2695 * any file. For now, just ensure that anything potentially problematic is done
2698 static bool io_file_supports_async(struct file *file, int rw)
2700 umode_t mode = file_inode(file)->i_mode;
2702 if (S_ISBLK(mode)) {
2703 if (io_bdev_nowait(file->f_inode->i_bdev))
2707 if (S_ISCHR(mode) || S_ISSOCK(mode))
2709 if (S_ISREG(mode)) {
2710 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2711 file->f_op != &io_uring_fops)
2716 /* any ->read/write should understand O_NONBLOCK */
2717 if (file->f_flags & O_NONBLOCK)
2720 if (!(file->f_mode & FMODE_NOWAIT))
2724 return file->f_op->read_iter != NULL;
2726 return file->f_op->write_iter != NULL;
2729 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2731 struct io_ring_ctx *ctx = req->ctx;
2732 struct kiocb *kiocb = &req->rw.kiocb;
2736 if (S_ISREG(file_inode(req->file)->i_mode))
2737 req->flags |= REQ_F_ISREG;
2739 kiocb->ki_pos = READ_ONCE(sqe->off);
2740 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2741 req->flags |= REQ_F_CUR_POS;
2742 kiocb->ki_pos = req->file->f_pos;
2744 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2745 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2746 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2750 ioprio = READ_ONCE(sqe->ioprio);
2752 ret = ioprio_check_cap(ioprio);
2756 kiocb->ki_ioprio = ioprio;
2758 kiocb->ki_ioprio = get_current_ioprio();
2760 /* don't allow async punt if RWF_NOWAIT was requested */
2761 if (kiocb->ki_flags & IOCB_NOWAIT)
2762 req->flags |= REQ_F_NOWAIT;
2764 if (ctx->flags & IORING_SETUP_IOPOLL) {
2765 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2766 !kiocb->ki_filp->f_op->iopoll)
2769 kiocb->ki_flags |= IOCB_HIPRI;
2770 kiocb->ki_complete = io_complete_rw_iopoll;
2771 req->iopoll_completed = 0;
2773 if (kiocb->ki_flags & IOCB_HIPRI)
2775 kiocb->ki_complete = io_complete_rw;
2778 req->rw.addr = READ_ONCE(sqe->addr);
2779 req->rw.len = READ_ONCE(sqe->len);
2780 req->buf_index = READ_ONCE(sqe->buf_index);
2784 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2790 case -ERESTARTNOINTR:
2791 case -ERESTARTNOHAND:
2792 case -ERESTART_RESTARTBLOCK:
2794 * We can't just restart the syscall, since previously
2795 * submitted sqes may already be in progress. Just fail this
2801 kiocb->ki_complete(kiocb, ret, 0);
2805 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2806 struct io_comp_state *cs)
2808 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2809 struct io_async_rw *io = req->async_data;
2811 /* add previously done IO, if any */
2812 if (io && io->bytes_done > 0) {
2814 ret = io->bytes_done;
2816 ret += io->bytes_done;
2819 if (req->flags & REQ_F_CUR_POS)
2820 req->file->f_pos = kiocb->ki_pos;
2821 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2822 __io_complete_rw(req, ret, 0, cs);
2824 io_rw_done(kiocb, ret);
2827 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2828 struct iov_iter *iter)
2830 struct io_ring_ctx *ctx = req->ctx;
2831 size_t len = req->rw.len;
2832 struct io_mapped_ubuf *imu;
2833 u16 index, buf_index = req->buf_index;
2837 if (unlikely(buf_index >= ctx->nr_user_bufs))
2839 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2840 imu = &ctx->user_bufs[index];
2841 buf_addr = req->rw.addr;
2844 if (buf_addr + len < buf_addr)
2846 /* not inside the mapped region */
2847 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2851 * May not be a start of buffer, set size appropriately
2852 * and advance us to the beginning.
2854 offset = buf_addr - imu->ubuf;
2855 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2859 * Don't use iov_iter_advance() here, as it's really slow for
2860 * using the latter parts of a big fixed buffer - it iterates
2861 * over each segment manually. We can cheat a bit here, because
2864 * 1) it's a BVEC iter, we set it up
2865 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2866 * first and last bvec
2868 * So just find our index, and adjust the iterator afterwards.
2869 * If the offset is within the first bvec (or the whole first
2870 * bvec, just use iov_iter_advance(). This makes it easier
2871 * since we can just skip the first segment, which may not
2872 * be PAGE_SIZE aligned.
2874 const struct bio_vec *bvec = imu->bvec;
2876 if (offset <= bvec->bv_len) {
2877 iov_iter_advance(iter, offset);
2879 unsigned long seg_skip;
2881 /* skip first vec */
2882 offset -= bvec->bv_len;
2883 seg_skip = 1 + (offset >> PAGE_SHIFT);
2885 iter->bvec = bvec + seg_skip;
2886 iter->nr_segs -= seg_skip;
2887 iter->count -= bvec->bv_len + offset;
2888 iter->iov_offset = offset & ~PAGE_MASK;
2895 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2898 mutex_unlock(&ctx->uring_lock);
2901 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2904 * "Normal" inline submissions always hold the uring_lock, since we
2905 * grab it from the system call. Same is true for the SQPOLL offload.
2906 * The only exception is when we've detached the request and issue it
2907 * from an async worker thread, grab the lock for that case.
2910 mutex_lock(&ctx->uring_lock);
2913 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2914 int bgid, struct io_buffer *kbuf,
2917 struct io_buffer *head;
2919 if (req->flags & REQ_F_BUFFER_SELECTED)
2922 io_ring_submit_lock(req->ctx, needs_lock);
2924 lockdep_assert_held(&req->ctx->uring_lock);
2926 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2928 if (!list_empty(&head->list)) {
2929 kbuf = list_last_entry(&head->list, struct io_buffer,
2931 list_del(&kbuf->list);
2934 idr_remove(&req->ctx->io_buffer_idr, bgid);
2936 if (*len > kbuf->len)
2939 kbuf = ERR_PTR(-ENOBUFS);
2942 io_ring_submit_unlock(req->ctx, needs_lock);
2947 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2950 struct io_buffer *kbuf;
2953 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2954 bgid = req->buf_index;
2955 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2958 req->rw.addr = (u64) (unsigned long) kbuf;
2959 req->flags |= REQ_F_BUFFER_SELECTED;
2960 return u64_to_user_ptr(kbuf->addr);
2963 #ifdef CONFIG_COMPAT
2964 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2967 struct compat_iovec __user *uiov;
2968 compat_ssize_t clen;
2972 uiov = u64_to_user_ptr(req->rw.addr);
2973 if (!access_ok(uiov, sizeof(*uiov)))
2975 if (__get_user(clen, &uiov->iov_len))
2981 buf = io_rw_buffer_select(req, &len, needs_lock);
2983 return PTR_ERR(buf);
2984 iov[0].iov_base = buf;
2985 iov[0].iov_len = (compat_size_t) len;
2990 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2993 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2997 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3000 len = iov[0].iov_len;
3003 buf = io_rw_buffer_select(req, &len, needs_lock);
3005 return PTR_ERR(buf);
3006 iov[0].iov_base = buf;
3007 iov[0].iov_len = len;
3011 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3014 if (req->flags & REQ_F_BUFFER_SELECTED) {
3015 struct io_buffer *kbuf;
3017 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3018 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3019 iov[0].iov_len = kbuf->len;
3024 else if (req->rw.len > 1)
3027 #ifdef CONFIG_COMPAT
3028 if (req->ctx->compat)
3029 return io_compat_import(req, iov, needs_lock);
3032 return __io_iov_buffer_select(req, iov, needs_lock);
3035 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3036 struct iovec **iovec, struct iov_iter *iter,
3039 void __user *buf = u64_to_user_ptr(req->rw.addr);
3040 size_t sqe_len = req->rw.len;
3044 opcode = req->opcode;
3045 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3047 return io_import_fixed(req, rw, iter);
3050 /* buffer index only valid with fixed read/write, or buffer select */
3051 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3054 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3055 if (req->flags & REQ_F_BUFFER_SELECT) {
3056 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3058 return PTR_ERR(buf);
3059 req->rw.len = sqe_len;
3062 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3064 return ret < 0 ? ret : sqe_len;
3067 if (req->flags & REQ_F_BUFFER_SELECT) {
3068 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3070 ret = (*iovec)->iov_len;
3071 iov_iter_init(iter, rw, *iovec, 1, ret);
3077 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3081 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3082 struct iovec **iovec, struct iov_iter *iter,
3085 struct io_async_rw *iorw = req->async_data;
3088 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3090 return iov_iter_count(&iorw->iter);
3093 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3095 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3099 * For files that don't have ->read_iter() and ->write_iter(), handle them
3100 * by looping over ->read() or ->write() manually.
3102 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3104 struct kiocb *kiocb = &req->rw.kiocb;
3105 struct file *file = req->file;
3109 * Don't support polled IO through this interface, and we can't
3110 * support non-blocking either. For the latter, this just causes
3111 * the kiocb to be handled from an async context.
3113 if (kiocb->ki_flags & IOCB_HIPRI)
3115 if (kiocb->ki_flags & IOCB_NOWAIT)
3118 while (iov_iter_count(iter)) {
3122 if (!iov_iter_is_bvec(iter)) {
3123 iovec = iov_iter_iovec(iter);
3125 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3126 iovec.iov_len = req->rw.len;
3130 nr = file->f_op->read(file, iovec.iov_base,
3131 iovec.iov_len, io_kiocb_ppos(kiocb));
3133 nr = file->f_op->write(file, iovec.iov_base,
3134 iovec.iov_len, io_kiocb_ppos(kiocb));
3143 if (nr != iovec.iov_len)
3147 iov_iter_advance(iter, nr);
3153 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3154 const struct iovec *fast_iov, struct iov_iter *iter)
3156 struct io_async_rw *rw = req->async_data;
3158 memcpy(&rw->iter, iter, sizeof(*iter));
3159 rw->free_iovec = iovec;
3161 /* can only be fixed buffers, no need to do anything */
3162 if (iter->type == ITER_BVEC)
3165 unsigned iov_off = 0;
3167 rw->iter.iov = rw->fast_iov;
3168 if (iter->iov != fast_iov) {
3169 iov_off = iter->iov - fast_iov;
3170 rw->iter.iov += iov_off;
3172 if (rw->fast_iov != fast_iov)
3173 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3174 sizeof(struct iovec) * iter->nr_segs);
3176 req->flags |= REQ_F_NEED_CLEANUP;
3180 static inline int __io_alloc_async_data(struct io_kiocb *req)
3182 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3183 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3184 return req->async_data == NULL;
3187 static int io_alloc_async_data(struct io_kiocb *req)
3189 if (!io_op_defs[req->opcode].needs_async_data)
3192 return __io_alloc_async_data(req);
3195 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3196 const struct iovec *fast_iov,
3197 struct iov_iter *iter, bool force)
3199 if (!force && !io_op_defs[req->opcode].needs_async_data)
3201 if (!req->async_data) {
3202 if (__io_alloc_async_data(req))
3205 io_req_map_rw(req, iovec, fast_iov, iter);
3210 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3212 struct io_async_rw *iorw = req->async_data;
3213 struct iovec *iov = iorw->fast_iov;
3216 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3217 if (unlikely(ret < 0))
3220 iorw->bytes_done = 0;
3221 iorw->free_iovec = iov;
3223 req->flags |= REQ_F_NEED_CLEANUP;
3227 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3231 ret = io_prep_rw(req, sqe);
3235 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3238 /* either don't need iovec imported or already have it */
3239 if (!req->async_data)
3241 return io_rw_prep_async(req, READ);
3245 * This is our waitqueue callback handler, registered through lock_page_async()
3246 * when we initially tried to do the IO with the iocb armed our waitqueue.
3247 * This gets called when the page is unlocked, and we generally expect that to
3248 * happen when the page IO is completed and the page is now uptodate. This will
3249 * queue a task_work based retry of the operation, attempting to copy the data
3250 * again. If the latter fails because the page was NOT uptodate, then we will
3251 * do a thread based blocking retry of the operation. That's the unexpected
3254 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3255 int sync, void *arg)
3257 struct wait_page_queue *wpq;
3258 struct io_kiocb *req = wait->private;
3259 struct wait_page_key *key = arg;
3262 wpq = container_of(wait, struct wait_page_queue, wait);
3264 if (!wake_page_match(wpq, key))
3267 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3268 list_del_init(&wait->entry);
3270 init_task_work(&req->task_work, io_req_task_submit);
3271 percpu_ref_get(&req->ctx->refs);
3273 /* submit ref gets dropped, acquire a new one */
3274 refcount_inc(&req->refs);
3275 ret = io_req_task_work_add(req, true);
3276 if (unlikely(ret)) {
3277 struct task_struct *tsk;
3279 /* queue just for cancelation */
3280 init_task_work(&req->task_work, io_req_task_cancel);
3281 tsk = io_wq_get_task(req->ctx->io_wq);
3282 task_work_add(tsk, &req->task_work, 0);
3283 wake_up_process(tsk);
3289 * This controls whether a given IO request should be armed for async page
3290 * based retry. If we return false here, the request is handed to the async
3291 * worker threads for retry. If we're doing buffered reads on a regular file,
3292 * we prepare a private wait_page_queue entry and retry the operation. This
3293 * will either succeed because the page is now uptodate and unlocked, or it
3294 * will register a callback when the page is unlocked at IO completion. Through
3295 * that callback, io_uring uses task_work to setup a retry of the operation.
3296 * That retry will attempt the buffered read again. The retry will generally
3297 * succeed, or in rare cases where it fails, we then fall back to using the
3298 * async worker threads for a blocking retry.
3300 static bool io_rw_should_retry(struct io_kiocb *req)
3302 struct io_async_rw *rw = req->async_data;
3303 struct wait_page_queue *wait = &rw->wpq;
3304 struct kiocb *kiocb = &req->rw.kiocb;
3306 /* never retry for NOWAIT, we just complete with -EAGAIN */
3307 if (req->flags & REQ_F_NOWAIT)
3310 /* Only for buffered IO */
3311 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3315 * just use poll if we can, and don't attempt if the fs doesn't
3316 * support callback based unlocks
3318 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3321 wait->wait.func = io_async_buf_func;
3322 wait->wait.private = req;
3323 wait->wait.flags = 0;
3324 INIT_LIST_HEAD(&wait->wait.entry);
3325 kiocb->ki_flags |= IOCB_WAITQ;
3326 kiocb->ki_flags &= ~IOCB_NOWAIT;
3327 kiocb->ki_waitq = wait;
3331 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3333 if (req->file->f_op->read_iter)
3334 return call_read_iter(req->file, &req->rw.kiocb, iter);
3335 else if (req->file->f_op->read)
3336 return loop_rw_iter(READ, req, iter);
3341 static int io_read(struct io_kiocb *req, bool force_nonblock,
3342 struct io_comp_state *cs)
3344 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3345 struct kiocb *kiocb = &req->rw.kiocb;
3346 struct iov_iter __iter, *iter = &__iter;
3347 struct io_async_rw *rw = req->async_data;
3348 ssize_t io_size, ret, ret2;
3355 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3358 iov_count = iov_iter_count(iter);
3360 req->result = io_size;
3363 /* Ensure we clear previously set non-block flag */
3364 if (!force_nonblock)
3365 kiocb->ki_flags &= ~IOCB_NOWAIT;
3367 kiocb->ki_flags |= IOCB_NOWAIT;
3370 /* If the file doesn't support async, just async punt */
3371 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3375 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3379 ret = io_iter_do_read(req, iter);
3383 } else if (ret == -EIOCBQUEUED) {
3386 } else if (ret == -EAGAIN) {
3387 /* IOPOLL retry should happen for io-wq threads */
3388 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3390 /* no retry on NONBLOCK marked file */
3391 if (req->file->f_flags & O_NONBLOCK)
3393 /* some cases will consume bytes even on error returns */
3394 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3397 } else if (ret < 0) {
3398 /* make sure -ERESTARTSYS -> -EINTR is done */
3402 /* read it all, or we did blocking attempt. no retry. */
3403 if (!iov_iter_count(iter) || !force_nonblock ||
3404 (req->file->f_flags & O_NONBLOCK))
3409 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3416 rw = req->async_data;
3417 /* it's copied and will be cleaned with ->io */
3419 /* now use our persistent iterator, if we aren't already */
3422 rw->bytes_done += ret;
3423 /* if we can retry, do so with the callbacks armed */
3424 if (!io_rw_should_retry(req)) {
3425 kiocb->ki_flags &= ~IOCB_WAITQ;
3430 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3431 * get -EIOCBQUEUED, then we'll get a notification when the desired
3432 * page gets unlocked. We can also get a partial read here, and if we
3433 * do, then just retry at the new offset.
3435 ret = io_iter_do_read(req, iter);
3436 if (ret == -EIOCBQUEUED) {
3439 } else if (ret > 0 && ret < io_size) {
3440 /* we got some bytes, but not all. retry. */
3444 kiocb_done(kiocb, ret, cs);
3447 /* it's reportedly faster than delegating the null check to kfree() */
3453 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3457 ret = io_prep_rw(req, sqe);
3461 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3464 /* either don't need iovec imported or already have it */
3465 if (!req->async_data)
3467 return io_rw_prep_async(req, WRITE);
3470 static int io_write(struct io_kiocb *req, bool force_nonblock,
3471 struct io_comp_state *cs)
3473 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3474 struct kiocb *kiocb = &req->rw.kiocb;
3475 struct iov_iter __iter, *iter = &__iter;
3476 struct io_async_rw *rw = req->async_data;
3478 ssize_t ret, ret2, io_size;
3483 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3486 iov_count = iov_iter_count(iter);
3488 req->result = io_size;
3490 /* Ensure we clear previously set non-block flag */
3491 if (!force_nonblock)
3492 kiocb->ki_flags &= ~IOCB_NOWAIT;
3494 kiocb->ki_flags |= IOCB_NOWAIT;
3496 /* If the file doesn't support async, just async punt */
3497 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3500 /* file path doesn't support NOWAIT for non-direct_IO */
3501 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3502 (req->flags & REQ_F_ISREG))
3505 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3510 * Open-code file_start_write here to grab freeze protection,
3511 * which will be released by another thread in
3512 * io_complete_rw(). Fool lockdep by telling it the lock got
3513 * released so that it doesn't complain about the held lock when
3514 * we return to userspace.
3516 if (req->flags & REQ_F_ISREG) {
3517 __sb_start_write(file_inode(req->file)->i_sb,
3518 SB_FREEZE_WRITE, true);
3519 __sb_writers_release(file_inode(req->file)->i_sb,
3522 kiocb->ki_flags |= IOCB_WRITE;
3524 if (req->file->f_op->write_iter)
3525 ret2 = call_write_iter(req->file, kiocb, iter);
3526 else if (req->file->f_op->write)
3527 ret2 = loop_rw_iter(WRITE, req, iter);
3532 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3533 * retry them without IOCB_NOWAIT.
3535 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3537 /* no retry on NONBLOCK marked file */
3538 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3540 if (!force_nonblock || ret2 != -EAGAIN) {
3541 /* IOPOLL retry should happen for io-wq threads */
3542 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3545 kiocb_done(kiocb, ret2, cs);
3548 /* some cases will consume bytes even on error returns */
3549 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3550 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3555 /* it's reportedly faster than delegating the null check to kfree() */
3561 static int __io_splice_prep(struct io_kiocb *req,
3562 const struct io_uring_sqe *sqe)
3564 struct io_splice* sp = &req->splice;
3565 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3567 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3571 sp->len = READ_ONCE(sqe->len);
3572 sp->flags = READ_ONCE(sqe->splice_flags);
3574 if (unlikely(sp->flags & ~valid_flags))
3577 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3578 (sp->flags & SPLICE_F_FD_IN_FIXED));
3581 req->flags |= REQ_F_NEED_CLEANUP;
3583 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3585 * Splice operation will be punted aync, and here need to
3586 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3588 io_req_init_async(req);
3589 req->work.flags |= IO_WQ_WORK_UNBOUND;
3595 static int io_tee_prep(struct io_kiocb *req,
3596 const struct io_uring_sqe *sqe)
3598 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3600 return __io_splice_prep(req, sqe);
3603 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3605 struct io_splice *sp = &req->splice;
3606 struct file *in = sp->file_in;
3607 struct file *out = sp->file_out;
3608 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3614 ret = do_tee(in, out, sp->len, flags);
3616 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3617 req->flags &= ~REQ_F_NEED_CLEANUP;
3620 req_set_fail_links(req);
3621 io_req_complete(req, ret);
3625 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3627 struct io_splice* sp = &req->splice;
3629 sp->off_in = READ_ONCE(sqe->splice_off_in);
3630 sp->off_out = READ_ONCE(sqe->off);
3631 return __io_splice_prep(req, sqe);
3634 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3636 struct io_splice *sp = &req->splice;
3637 struct file *in = sp->file_in;
3638 struct file *out = sp->file_out;
3639 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3640 loff_t *poff_in, *poff_out;
3646 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3647 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3650 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3652 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3653 req->flags &= ~REQ_F_NEED_CLEANUP;
3656 req_set_fail_links(req);
3657 io_req_complete(req, ret);
3662 * IORING_OP_NOP just posts a completion event, nothing else.
3664 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3666 struct io_ring_ctx *ctx = req->ctx;
3668 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3671 __io_req_complete(req, 0, 0, cs);
3675 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3677 struct io_ring_ctx *ctx = req->ctx;
3682 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3684 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3687 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3688 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3691 req->sync.off = READ_ONCE(sqe->off);
3692 req->sync.len = READ_ONCE(sqe->len);
3696 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3698 loff_t end = req->sync.off + req->sync.len;
3701 /* fsync always requires a blocking context */
3705 ret = vfs_fsync_range(req->file, req->sync.off,
3706 end > 0 ? end : LLONG_MAX,
3707 req->sync.flags & IORING_FSYNC_DATASYNC);
3709 req_set_fail_links(req);
3710 io_req_complete(req, ret);
3714 static int io_fallocate_prep(struct io_kiocb *req,
3715 const struct io_uring_sqe *sqe)
3717 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3719 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3722 req->sync.off = READ_ONCE(sqe->off);
3723 req->sync.len = READ_ONCE(sqe->addr);
3724 req->sync.mode = READ_ONCE(sqe->len);
3728 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3732 /* fallocate always requiring blocking context */
3735 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3738 req_set_fail_links(req);
3739 io_req_complete(req, ret);
3743 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3745 const char __user *fname;
3748 if (unlikely(sqe->ioprio || sqe->buf_index))
3750 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3753 /* open.how should be already initialised */
3754 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3755 req->open.how.flags |= O_LARGEFILE;
3757 req->open.dfd = READ_ONCE(sqe->fd);
3758 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3759 req->open.filename = getname(fname);
3760 if (IS_ERR(req->open.filename)) {
3761 ret = PTR_ERR(req->open.filename);
3762 req->open.filename = NULL;
3765 req->open.nofile = rlimit(RLIMIT_NOFILE);
3766 req->flags |= REQ_F_NEED_CLEANUP;
3770 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3774 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3776 mode = READ_ONCE(sqe->len);
3777 flags = READ_ONCE(sqe->open_flags);
3778 req->open.how = build_open_how(flags, mode);
3779 return __io_openat_prep(req, sqe);
3782 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3784 struct open_how __user *how;
3788 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3790 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3791 len = READ_ONCE(sqe->len);
3792 if (len < OPEN_HOW_SIZE_VER0)
3795 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3800 return __io_openat_prep(req, sqe);
3803 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3805 struct open_flags op;
3812 ret = build_open_flags(&req->open.how, &op);
3816 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3820 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3823 ret = PTR_ERR(file);
3825 fsnotify_open(file);
3826 fd_install(ret, file);
3829 putname(req->open.filename);
3830 req->flags &= ~REQ_F_NEED_CLEANUP;
3832 req_set_fail_links(req);
3833 io_req_complete(req, ret);
3837 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3839 return io_openat2(req, force_nonblock);
3842 static int io_remove_buffers_prep(struct io_kiocb *req,
3843 const struct io_uring_sqe *sqe)
3845 struct io_provide_buf *p = &req->pbuf;
3848 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3851 tmp = READ_ONCE(sqe->fd);
3852 if (!tmp || tmp > USHRT_MAX)
3855 memset(p, 0, sizeof(*p));
3857 p->bgid = READ_ONCE(sqe->buf_group);
3861 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3862 int bgid, unsigned nbufs)
3866 /* shouldn't happen */
3870 /* the head kbuf is the list itself */
3871 while (!list_empty(&buf->list)) {
3872 struct io_buffer *nxt;
3874 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3875 list_del(&nxt->list);
3882 idr_remove(&ctx->io_buffer_idr, bgid);
3887 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3888 struct io_comp_state *cs)
3890 struct io_provide_buf *p = &req->pbuf;
3891 struct io_ring_ctx *ctx = req->ctx;
3892 struct io_buffer *head;
3895 io_ring_submit_lock(ctx, !force_nonblock);
3897 lockdep_assert_held(&ctx->uring_lock);
3900 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3902 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3904 io_ring_submit_lock(ctx, !force_nonblock);
3906 req_set_fail_links(req);
3907 __io_req_complete(req, ret, 0, cs);
3911 static int io_provide_buffers_prep(struct io_kiocb *req,
3912 const struct io_uring_sqe *sqe)
3914 struct io_provide_buf *p = &req->pbuf;
3917 if (sqe->ioprio || sqe->rw_flags)
3920 tmp = READ_ONCE(sqe->fd);
3921 if (!tmp || tmp > USHRT_MAX)
3924 p->addr = READ_ONCE(sqe->addr);
3925 p->len = READ_ONCE(sqe->len);
3927 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3930 p->bgid = READ_ONCE(sqe->buf_group);
3931 tmp = READ_ONCE(sqe->off);
3932 if (tmp > USHRT_MAX)
3938 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3940 struct io_buffer *buf;
3941 u64 addr = pbuf->addr;
3942 int i, bid = pbuf->bid;
3944 for (i = 0; i < pbuf->nbufs; i++) {
3945 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3950 buf->len = pbuf->len;
3955 INIT_LIST_HEAD(&buf->list);
3958 list_add_tail(&buf->list, &(*head)->list);
3962 return i ? i : -ENOMEM;
3965 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3966 struct io_comp_state *cs)
3968 struct io_provide_buf *p = &req->pbuf;
3969 struct io_ring_ctx *ctx = req->ctx;
3970 struct io_buffer *head, *list;
3973 io_ring_submit_lock(ctx, !force_nonblock);
3975 lockdep_assert_held(&ctx->uring_lock);
3977 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3979 ret = io_add_buffers(p, &head);
3984 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3987 __io_remove_buffers(ctx, head, p->bgid, -1U);
3992 io_ring_submit_unlock(ctx, !force_nonblock);
3994 req_set_fail_links(req);
3995 __io_req_complete(req, ret, 0, cs);
3999 static int io_epoll_ctl_prep(struct io_kiocb *req,
4000 const struct io_uring_sqe *sqe)
4002 #if defined(CONFIG_EPOLL)
4003 if (sqe->ioprio || sqe->buf_index)
4005 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4008 req->epoll.epfd = READ_ONCE(sqe->fd);
4009 req->epoll.op = READ_ONCE(sqe->len);
4010 req->epoll.fd = READ_ONCE(sqe->off);
4012 if (ep_op_has_event(req->epoll.op)) {
4013 struct epoll_event __user *ev;
4015 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4016 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4026 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4027 struct io_comp_state *cs)
4029 #if defined(CONFIG_EPOLL)
4030 struct io_epoll *ie = &req->epoll;
4033 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4034 if (force_nonblock && ret == -EAGAIN)
4038 req_set_fail_links(req);
4039 __io_req_complete(req, ret, 0, cs);
4046 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4048 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4049 if (sqe->ioprio || sqe->buf_index || sqe->off)
4051 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4054 req->madvise.addr = READ_ONCE(sqe->addr);
4055 req->madvise.len = READ_ONCE(sqe->len);
4056 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4063 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4065 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4066 struct io_madvise *ma = &req->madvise;
4072 ret = do_madvise(ma->addr, ma->len, ma->advice);
4074 req_set_fail_links(req);
4075 io_req_complete(req, ret);
4082 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4084 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4086 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4089 req->fadvise.offset = READ_ONCE(sqe->off);
4090 req->fadvise.len = READ_ONCE(sqe->len);
4091 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4095 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4097 struct io_fadvise *fa = &req->fadvise;
4100 if (force_nonblock) {
4101 switch (fa->advice) {
4102 case POSIX_FADV_NORMAL:
4103 case POSIX_FADV_RANDOM:
4104 case POSIX_FADV_SEQUENTIAL:
4111 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4113 req_set_fail_links(req);
4114 io_req_complete(req, ret);
4118 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4120 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4122 if (sqe->ioprio || sqe->buf_index)
4124 if (req->flags & REQ_F_FIXED_FILE)
4127 req->statx.dfd = READ_ONCE(sqe->fd);
4128 req->statx.mask = READ_ONCE(sqe->len);
4129 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4130 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4131 req->statx.flags = READ_ONCE(sqe->statx_flags);
4136 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4138 struct io_statx *ctx = &req->statx;
4141 if (force_nonblock) {
4142 /* only need file table for an actual valid fd */
4143 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4144 req->flags |= REQ_F_NO_FILE_TABLE;
4148 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4152 req_set_fail_links(req);
4153 io_req_complete(req, ret);
4157 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4160 * If we queue this for async, it must not be cancellable. That would
4161 * leave the 'file' in an undeterminate state, and here need to modify
4162 * io_wq_work.flags, so initialize io_wq_work firstly.
4164 io_req_init_async(req);
4165 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4167 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4169 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4170 sqe->rw_flags || sqe->buf_index)
4172 if (req->flags & REQ_F_FIXED_FILE)
4175 req->close.fd = READ_ONCE(sqe->fd);
4176 if ((req->file && req->file->f_op == &io_uring_fops))
4179 req->close.put_file = NULL;
4183 static int io_close(struct io_kiocb *req, bool force_nonblock,
4184 struct io_comp_state *cs)
4186 struct io_close *close = &req->close;
4189 /* might be already done during nonblock submission */
4190 if (!close->put_file) {
4191 ret = __close_fd_get_file(close->fd, &close->put_file);
4193 return (ret == -ENOENT) ? -EBADF : ret;
4196 /* if the file has a flush method, be safe and punt to async */
4197 if (close->put_file->f_op->flush && force_nonblock) {
4198 /* was never set, but play safe */
4199 req->flags &= ~REQ_F_NOWAIT;
4200 /* avoid grabbing files - we don't need the files */
4201 req->flags |= REQ_F_NO_FILE_TABLE;
4205 /* No ->flush() or already async, safely close from here */
4206 ret = filp_close(close->put_file, req->work.identity->files);
4208 req_set_fail_links(req);
4209 fput(close->put_file);
4210 close->put_file = NULL;
4211 __io_req_complete(req, ret, 0, cs);
4215 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4217 struct io_ring_ctx *ctx = req->ctx;
4222 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4224 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4227 req->sync.off = READ_ONCE(sqe->off);
4228 req->sync.len = READ_ONCE(sqe->len);
4229 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4233 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4237 /* sync_file_range always requires a blocking context */
4241 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4244 req_set_fail_links(req);
4245 io_req_complete(req, ret);
4249 #if defined(CONFIG_NET)
4250 static int io_setup_async_msg(struct io_kiocb *req,
4251 struct io_async_msghdr *kmsg)
4253 struct io_async_msghdr *async_msg = req->async_data;
4257 if (io_alloc_async_data(req)) {
4258 if (kmsg->iov != kmsg->fast_iov)
4262 async_msg = req->async_data;
4263 req->flags |= REQ_F_NEED_CLEANUP;
4264 memcpy(async_msg, kmsg, sizeof(*kmsg));
4268 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4269 struct io_async_msghdr *iomsg)
4271 iomsg->iov = iomsg->fast_iov;
4272 iomsg->msg.msg_name = &iomsg->addr;
4273 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4274 req->sr_msg.msg_flags, &iomsg->iov);
4277 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4279 struct io_async_msghdr *async_msg = req->async_data;
4280 struct io_sr_msg *sr = &req->sr_msg;
4283 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4286 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4287 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4288 sr->len = READ_ONCE(sqe->len);
4290 #ifdef CONFIG_COMPAT
4291 if (req->ctx->compat)
4292 sr->msg_flags |= MSG_CMSG_COMPAT;
4295 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4297 ret = io_sendmsg_copy_hdr(req, async_msg);
4299 req->flags |= REQ_F_NEED_CLEANUP;
4303 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4304 struct io_comp_state *cs)
4306 struct io_async_msghdr iomsg, *kmsg;
4307 struct socket *sock;
4311 sock = sock_from_file(req->file, &ret);
4312 if (unlikely(!sock))
4315 if (req->async_data) {
4316 kmsg = req->async_data;
4317 kmsg->msg.msg_name = &kmsg->addr;
4318 /* if iov is set, it's allocated already */
4320 kmsg->iov = kmsg->fast_iov;
4321 kmsg->msg.msg_iter.iov = kmsg->iov;
4323 ret = io_sendmsg_copy_hdr(req, &iomsg);
4329 flags = req->sr_msg.msg_flags;
4330 if (flags & MSG_DONTWAIT)
4331 req->flags |= REQ_F_NOWAIT;
4332 else if (force_nonblock)
4333 flags |= MSG_DONTWAIT;
4335 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4336 if (force_nonblock && ret == -EAGAIN)
4337 return io_setup_async_msg(req, kmsg);
4338 if (ret == -ERESTARTSYS)
4341 if (kmsg->iov != kmsg->fast_iov)
4343 req->flags &= ~REQ_F_NEED_CLEANUP;
4345 req_set_fail_links(req);
4346 __io_req_complete(req, ret, 0, cs);
4350 static int io_send(struct io_kiocb *req, bool force_nonblock,
4351 struct io_comp_state *cs)
4353 struct io_sr_msg *sr = &req->sr_msg;
4356 struct socket *sock;
4360 sock = sock_from_file(req->file, &ret);
4361 if (unlikely(!sock))
4364 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4368 msg.msg_name = NULL;
4369 msg.msg_control = NULL;
4370 msg.msg_controllen = 0;
4371 msg.msg_namelen = 0;
4373 flags = req->sr_msg.msg_flags;
4374 if (flags & MSG_DONTWAIT)
4375 req->flags |= REQ_F_NOWAIT;
4376 else if (force_nonblock)
4377 flags |= MSG_DONTWAIT;
4379 msg.msg_flags = flags;
4380 ret = sock_sendmsg(sock, &msg);
4381 if (force_nonblock && ret == -EAGAIN)
4383 if (ret == -ERESTARTSYS)
4387 req_set_fail_links(req);
4388 __io_req_complete(req, ret, 0, cs);
4392 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4393 struct io_async_msghdr *iomsg)
4395 struct io_sr_msg *sr = &req->sr_msg;
4396 struct iovec __user *uiov;
4400 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4401 &iomsg->uaddr, &uiov, &iov_len);
4405 if (req->flags & REQ_F_BUFFER_SELECT) {
4408 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4410 sr->len = iomsg->iov[0].iov_len;
4411 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4415 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4416 &iomsg->iov, &iomsg->msg.msg_iter,
4425 #ifdef CONFIG_COMPAT
4426 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4427 struct io_async_msghdr *iomsg)
4429 struct compat_msghdr __user *msg_compat;
4430 struct io_sr_msg *sr = &req->sr_msg;
4431 struct compat_iovec __user *uiov;
4436 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4437 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4442 uiov = compat_ptr(ptr);
4443 if (req->flags & REQ_F_BUFFER_SELECT) {
4444 compat_ssize_t clen;
4448 if (!access_ok(uiov, sizeof(*uiov)))
4450 if (__get_user(clen, &uiov->iov_len))
4454 sr->len = iomsg->iov[0].iov_len;
4457 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4458 UIO_FASTIOV, &iomsg->iov,
4459 &iomsg->msg.msg_iter, true);
4468 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4469 struct io_async_msghdr *iomsg)
4471 iomsg->msg.msg_name = &iomsg->addr;
4472 iomsg->iov = iomsg->fast_iov;
4474 #ifdef CONFIG_COMPAT
4475 if (req->ctx->compat)
4476 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4479 return __io_recvmsg_copy_hdr(req, iomsg);
4482 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4485 struct io_sr_msg *sr = &req->sr_msg;
4486 struct io_buffer *kbuf;
4488 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4493 req->flags |= REQ_F_BUFFER_SELECTED;
4497 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4499 return io_put_kbuf(req, req->sr_msg.kbuf);
4502 static int io_recvmsg_prep(struct io_kiocb *req,
4503 const struct io_uring_sqe *sqe)
4505 struct io_async_msghdr *async_msg = req->async_data;
4506 struct io_sr_msg *sr = &req->sr_msg;
4509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4512 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4513 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4514 sr->len = READ_ONCE(sqe->len);
4515 sr->bgid = READ_ONCE(sqe->buf_group);
4517 #ifdef CONFIG_COMPAT
4518 if (req->ctx->compat)
4519 sr->msg_flags |= MSG_CMSG_COMPAT;
4522 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4524 ret = io_recvmsg_copy_hdr(req, async_msg);
4526 req->flags |= REQ_F_NEED_CLEANUP;
4530 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4531 struct io_comp_state *cs)
4533 struct io_async_msghdr iomsg, *kmsg;
4534 struct socket *sock;
4535 struct io_buffer *kbuf;
4537 int ret, cflags = 0;
4539 sock = sock_from_file(req->file, &ret);
4540 if (unlikely(!sock))
4543 if (req->async_data) {
4544 kmsg = req->async_data;
4545 kmsg->msg.msg_name = &kmsg->addr;
4546 /* if iov is set, it's allocated already */
4548 kmsg->iov = kmsg->fast_iov;
4549 kmsg->msg.msg_iter.iov = kmsg->iov;
4551 ret = io_recvmsg_copy_hdr(req, &iomsg);
4557 if (req->flags & REQ_F_BUFFER_SELECT) {
4558 kbuf = io_recv_buffer_select(req, !force_nonblock);
4560 return PTR_ERR(kbuf);
4561 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4562 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4563 1, req->sr_msg.len);
4566 flags = req->sr_msg.msg_flags;
4567 if (flags & MSG_DONTWAIT)
4568 req->flags |= REQ_F_NOWAIT;
4569 else if (force_nonblock)
4570 flags |= MSG_DONTWAIT;
4572 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4573 kmsg->uaddr, flags);
4574 if (force_nonblock && ret == -EAGAIN)
4575 return io_setup_async_msg(req, kmsg);
4576 if (ret == -ERESTARTSYS)
4579 if (req->flags & REQ_F_BUFFER_SELECTED)
4580 cflags = io_put_recv_kbuf(req);
4581 if (kmsg->iov != kmsg->fast_iov)
4583 req->flags &= ~REQ_F_NEED_CLEANUP;
4585 req_set_fail_links(req);
4586 __io_req_complete(req, ret, cflags, cs);
4590 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4591 struct io_comp_state *cs)
4593 struct io_buffer *kbuf;
4594 struct io_sr_msg *sr = &req->sr_msg;
4596 void __user *buf = sr->buf;
4597 struct socket *sock;
4600 int ret, cflags = 0;
4602 sock = sock_from_file(req->file, &ret);
4603 if (unlikely(!sock))
4606 if (req->flags & REQ_F_BUFFER_SELECT) {
4607 kbuf = io_recv_buffer_select(req, !force_nonblock);
4609 return PTR_ERR(kbuf);
4610 buf = u64_to_user_ptr(kbuf->addr);
4613 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4617 msg.msg_name = NULL;
4618 msg.msg_control = NULL;
4619 msg.msg_controllen = 0;
4620 msg.msg_namelen = 0;
4621 msg.msg_iocb = NULL;
4624 flags = req->sr_msg.msg_flags;
4625 if (flags & MSG_DONTWAIT)
4626 req->flags |= REQ_F_NOWAIT;
4627 else if (force_nonblock)
4628 flags |= MSG_DONTWAIT;
4630 ret = sock_recvmsg(sock, &msg, flags);
4631 if (force_nonblock && ret == -EAGAIN)
4633 if (ret == -ERESTARTSYS)
4636 if (req->flags & REQ_F_BUFFER_SELECTED)
4637 cflags = io_put_recv_kbuf(req);
4639 req_set_fail_links(req);
4640 __io_req_complete(req, ret, cflags, cs);
4644 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4646 struct io_accept *accept = &req->accept;
4648 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4650 if (sqe->ioprio || sqe->len || sqe->buf_index)
4653 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4654 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4655 accept->flags = READ_ONCE(sqe->accept_flags);
4656 accept->nofile = rlimit(RLIMIT_NOFILE);
4660 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4661 struct io_comp_state *cs)
4663 struct io_accept *accept = &req->accept;
4664 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4667 if (req->file->f_flags & O_NONBLOCK)
4668 req->flags |= REQ_F_NOWAIT;
4670 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4671 accept->addr_len, accept->flags,
4673 if (ret == -EAGAIN && force_nonblock)
4676 if (ret == -ERESTARTSYS)
4678 req_set_fail_links(req);
4680 __io_req_complete(req, ret, 0, cs);
4684 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4686 struct io_connect *conn = &req->connect;
4687 struct io_async_connect *io = req->async_data;
4689 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4691 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4694 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4695 conn->addr_len = READ_ONCE(sqe->addr2);
4700 return move_addr_to_kernel(conn->addr, conn->addr_len,
4704 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4705 struct io_comp_state *cs)
4707 struct io_async_connect __io, *io;
4708 unsigned file_flags;
4711 if (req->async_data) {
4712 io = req->async_data;
4714 ret = move_addr_to_kernel(req->connect.addr,
4715 req->connect.addr_len,
4722 file_flags = force_nonblock ? O_NONBLOCK : 0;
4724 ret = __sys_connect_file(req->file, &io->address,
4725 req->connect.addr_len, file_flags);
4726 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4727 if (req->async_data)
4729 if (io_alloc_async_data(req)) {
4733 io = req->async_data;
4734 memcpy(req->async_data, &__io, sizeof(__io));
4737 if (ret == -ERESTARTSYS)
4741 req_set_fail_links(req);
4742 __io_req_complete(req, ret, 0, cs);
4745 #else /* !CONFIG_NET */
4746 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4751 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4752 struct io_comp_state *cs)
4757 static int io_send(struct io_kiocb *req, bool force_nonblock,
4758 struct io_comp_state *cs)
4763 static int io_recvmsg_prep(struct io_kiocb *req,
4764 const struct io_uring_sqe *sqe)
4769 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4770 struct io_comp_state *cs)
4775 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4776 struct io_comp_state *cs)
4781 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4786 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4787 struct io_comp_state *cs)
4792 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4797 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4798 struct io_comp_state *cs)
4802 #endif /* CONFIG_NET */
4804 struct io_poll_table {
4805 struct poll_table_struct pt;
4806 struct io_kiocb *req;
4810 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4811 __poll_t mask, task_work_func_t func)
4816 /* for instances that support it check for an event match first: */
4817 if (mask && !(mask & poll->events))
4820 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4822 list_del_init(&poll->wait.entry);
4825 init_task_work(&req->task_work, func);
4826 percpu_ref_get(&req->ctx->refs);
4829 * If we using the signalfd wait_queue_head for this wakeup, then
4830 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4831 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4832 * either, as the normal wakeup will suffice.
4834 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4837 * If this fails, then the task is exiting. When a task exits, the
4838 * work gets canceled, so just cancel this request as well instead
4839 * of executing it. We can't safely execute it anyway, as we may not
4840 * have the needed state needed for it anyway.
4842 ret = io_req_task_work_add(req, twa_signal_ok);
4843 if (unlikely(ret)) {
4844 struct task_struct *tsk;
4846 WRITE_ONCE(poll->canceled, true);
4847 tsk = io_wq_get_task(req->ctx->io_wq);
4848 task_work_add(tsk, &req->task_work, 0);
4849 wake_up_process(tsk);
4854 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4855 __acquires(&req->ctx->completion_lock)
4857 struct io_ring_ctx *ctx = req->ctx;
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 struct poll_table_struct pt = { ._key = poll->events };
4862 req->result = vfs_poll(req->file, &pt) & poll->events;
4865 spin_lock_irq(&ctx->completion_lock);
4866 if (!req->result && !READ_ONCE(poll->canceled)) {
4867 add_wait_queue(poll->head, &poll->wait);
4874 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4876 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4877 if (req->opcode == IORING_OP_POLL_ADD)
4878 return req->async_data;
4879 return req->apoll->double_poll;
4882 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4884 if (req->opcode == IORING_OP_POLL_ADD)
4886 return &req->apoll->poll;
4889 static void io_poll_remove_double(struct io_kiocb *req)
4891 struct io_poll_iocb *poll = io_poll_get_double(req);
4893 lockdep_assert_held(&req->ctx->completion_lock);
4895 if (poll && poll->head) {
4896 struct wait_queue_head *head = poll->head;
4898 spin_lock(&head->lock);
4899 list_del_init(&poll->wait.entry);
4900 if (poll->wait.private)
4901 refcount_dec(&req->refs);
4903 spin_unlock(&head->lock);
4907 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4909 struct io_ring_ctx *ctx = req->ctx;
4911 io_poll_remove_double(req);
4912 req->poll.done = true;
4913 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4914 io_commit_cqring(ctx);
4917 static void io_poll_task_func(struct callback_head *cb)
4919 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4920 struct io_ring_ctx *ctx = req->ctx;
4921 struct io_kiocb *nxt;
4923 if (io_poll_rewait(req, &req->poll)) {
4924 spin_unlock_irq(&ctx->completion_lock);
4926 hash_del(&req->hash_node);
4927 io_poll_complete(req, req->result, 0);
4928 spin_unlock_irq(&ctx->completion_lock);
4930 nxt = io_put_req_find_next(req);
4931 io_cqring_ev_posted(ctx);
4933 __io_req_task_submit(nxt);
4936 percpu_ref_put(&ctx->refs);
4939 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4940 int sync, void *key)
4942 struct io_kiocb *req = wait->private;
4943 struct io_poll_iocb *poll = io_poll_get_single(req);
4944 __poll_t mask = key_to_poll(key);
4946 /* for instances that support it check for an event match first: */
4947 if (mask && !(mask & poll->events))
4950 list_del_init(&wait->entry);
4952 if (poll && poll->head) {
4955 spin_lock(&poll->head->lock);
4956 done = list_empty(&poll->wait.entry);
4958 list_del_init(&poll->wait.entry);
4959 /* make sure double remove sees this as being gone */
4960 wait->private = NULL;
4961 spin_unlock(&poll->head->lock);
4963 __io_async_wake(req, poll, mask, io_poll_task_func);
4965 refcount_dec(&req->refs);
4969 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4970 wait_queue_func_t wake_func)
4974 poll->canceled = false;
4975 poll->events = events;
4976 INIT_LIST_HEAD(&poll->wait.entry);
4977 init_waitqueue_func_entry(&poll->wait, wake_func);
4980 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4981 struct wait_queue_head *head,
4982 struct io_poll_iocb **poll_ptr)
4984 struct io_kiocb *req = pt->req;
4987 * If poll->head is already set, it's because the file being polled
4988 * uses multiple waitqueues for poll handling (eg one for read, one
4989 * for write). Setup a separate io_poll_iocb if this happens.
4991 if (unlikely(poll->head)) {
4992 struct io_poll_iocb *poll_one = poll;
4994 /* already have a 2nd entry, fail a third attempt */
4996 pt->error = -EINVAL;
4999 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5001 pt->error = -ENOMEM;
5004 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5005 refcount_inc(&req->refs);
5006 poll->wait.private = req;
5013 if (poll->events & EPOLLEXCLUSIVE)
5014 add_wait_queue_exclusive(head, &poll->wait);
5016 add_wait_queue(head, &poll->wait);
5019 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5020 struct poll_table_struct *p)
5022 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5023 struct async_poll *apoll = pt->req->apoll;
5025 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5028 static void io_async_task_func(struct callback_head *cb)
5030 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5031 struct async_poll *apoll = req->apoll;
5032 struct io_ring_ctx *ctx = req->ctx;
5034 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5036 if (io_poll_rewait(req, &apoll->poll)) {
5037 spin_unlock_irq(&ctx->completion_lock);
5038 percpu_ref_put(&ctx->refs);
5042 /* If req is still hashed, it cannot have been canceled. Don't check. */
5043 if (hash_hashed(&req->hash_node))
5044 hash_del(&req->hash_node);
5046 io_poll_remove_double(req);
5047 spin_unlock_irq(&ctx->completion_lock);
5049 if (!READ_ONCE(apoll->poll.canceled))
5050 __io_req_task_submit(req);
5052 __io_req_task_cancel(req, -ECANCELED);
5054 percpu_ref_put(&ctx->refs);
5055 kfree(apoll->double_poll);
5059 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5062 struct io_kiocb *req = wait->private;
5063 struct io_poll_iocb *poll = &req->apoll->poll;
5065 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5068 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5071 static void io_poll_req_insert(struct io_kiocb *req)
5073 struct io_ring_ctx *ctx = req->ctx;
5074 struct hlist_head *list;
5076 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5077 hlist_add_head(&req->hash_node, list);
5080 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5081 struct io_poll_iocb *poll,
5082 struct io_poll_table *ipt, __poll_t mask,
5083 wait_queue_func_t wake_func)
5084 __acquires(&ctx->completion_lock)
5086 struct io_ring_ctx *ctx = req->ctx;
5087 bool cancel = false;
5089 INIT_HLIST_NODE(&req->hash_node);
5090 io_init_poll_iocb(poll, mask, wake_func);
5091 poll->file = req->file;
5092 poll->wait.private = req;
5094 ipt->pt._key = mask;
5096 ipt->error = -EINVAL;
5098 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5100 spin_lock_irq(&ctx->completion_lock);
5101 if (likely(poll->head)) {
5102 spin_lock(&poll->head->lock);
5103 if (unlikely(list_empty(&poll->wait.entry))) {
5109 if (mask || ipt->error)
5110 list_del_init(&poll->wait.entry);
5112 WRITE_ONCE(poll->canceled, true);
5113 else if (!poll->done) /* actually waiting for an event */
5114 io_poll_req_insert(req);
5115 spin_unlock(&poll->head->lock);
5121 static bool io_arm_poll_handler(struct io_kiocb *req)
5123 const struct io_op_def *def = &io_op_defs[req->opcode];
5124 struct io_ring_ctx *ctx = req->ctx;
5125 struct async_poll *apoll;
5126 struct io_poll_table ipt;
5130 if (!req->file || !file_can_poll(req->file))
5132 if (req->flags & REQ_F_POLLED)
5136 else if (def->pollout)
5140 /* if we can't nonblock try, then no point in arming a poll handler */
5141 if (!io_file_supports_async(req->file, rw))
5144 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5145 if (unlikely(!apoll))
5147 apoll->double_poll = NULL;
5149 req->flags |= REQ_F_POLLED;
5154 mask |= POLLIN | POLLRDNORM;
5156 mask |= POLLOUT | POLLWRNORM;
5158 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5159 if ((req->opcode == IORING_OP_RECVMSG) &&
5160 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5163 mask |= POLLERR | POLLPRI;
5165 ipt.pt._qproc = io_async_queue_proc;
5167 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5169 if (ret || ipt.error) {
5170 io_poll_remove_double(req);
5171 spin_unlock_irq(&ctx->completion_lock);
5172 kfree(apoll->double_poll);
5176 spin_unlock_irq(&ctx->completion_lock);
5177 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5178 apoll->poll.events);
5182 static bool __io_poll_remove_one(struct io_kiocb *req,
5183 struct io_poll_iocb *poll)
5185 bool do_complete = false;
5187 spin_lock(&poll->head->lock);
5188 WRITE_ONCE(poll->canceled, true);
5189 if (!list_empty(&poll->wait.entry)) {
5190 list_del_init(&poll->wait.entry);
5193 spin_unlock(&poll->head->lock);
5194 hash_del(&req->hash_node);
5198 static bool io_poll_remove_one(struct io_kiocb *req)
5202 io_poll_remove_double(req);
5204 if (req->opcode == IORING_OP_POLL_ADD) {
5205 do_complete = __io_poll_remove_one(req, &req->poll);
5207 struct async_poll *apoll = req->apoll;
5209 /* non-poll requests have submit ref still */
5210 do_complete = __io_poll_remove_one(req, &apoll->poll);
5213 kfree(apoll->double_poll);
5219 io_cqring_fill_event(req, -ECANCELED);
5220 io_commit_cqring(req->ctx);
5221 req_set_fail_links(req);
5222 io_put_req_deferred(req, 1);
5229 * Returns true if we found and killed one or more poll requests
5231 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5233 struct hlist_node *tmp;
5234 struct io_kiocb *req;
5237 spin_lock_irq(&ctx->completion_lock);
5238 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5239 struct hlist_head *list;
5241 list = &ctx->cancel_hash[i];
5242 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5243 if (io_task_match(req, tsk))
5244 posted += io_poll_remove_one(req);
5247 spin_unlock_irq(&ctx->completion_lock);
5250 io_cqring_ev_posted(ctx);
5255 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5257 struct hlist_head *list;
5258 struct io_kiocb *req;
5260 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5261 hlist_for_each_entry(req, list, hash_node) {
5262 if (sqe_addr != req->user_data)
5264 if (io_poll_remove_one(req))
5272 static int io_poll_remove_prep(struct io_kiocb *req,
5273 const struct io_uring_sqe *sqe)
5275 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5277 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5281 req->poll.addr = READ_ONCE(sqe->addr);
5286 * Find a running poll command that matches one specified in sqe->addr,
5287 * and remove it if found.
5289 static int io_poll_remove(struct io_kiocb *req)
5291 struct io_ring_ctx *ctx = req->ctx;
5295 addr = req->poll.addr;
5296 spin_lock_irq(&ctx->completion_lock);
5297 ret = io_poll_cancel(ctx, addr);
5298 spin_unlock_irq(&ctx->completion_lock);
5301 req_set_fail_links(req);
5302 io_req_complete(req, ret);
5306 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5309 struct io_kiocb *req = wait->private;
5310 struct io_poll_iocb *poll = &req->poll;
5312 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5315 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5316 struct poll_table_struct *p)
5318 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5320 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5323 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5325 struct io_poll_iocb *poll = &req->poll;
5328 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5330 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5333 events = READ_ONCE(sqe->poll32_events);
5335 events = swahw32(events);
5337 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5338 (events & EPOLLEXCLUSIVE);
5342 static int io_poll_add(struct io_kiocb *req)
5344 struct io_poll_iocb *poll = &req->poll;
5345 struct io_ring_ctx *ctx = req->ctx;
5346 struct io_poll_table ipt;
5349 ipt.pt._qproc = io_poll_queue_proc;
5351 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5354 if (mask) { /* no async, we'd stolen it */
5356 io_poll_complete(req, mask, 0);
5358 spin_unlock_irq(&ctx->completion_lock);
5361 io_cqring_ev_posted(ctx);
5367 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5369 struct io_timeout_data *data = container_of(timer,
5370 struct io_timeout_data, timer);
5371 struct io_kiocb *req = data->req;
5372 struct io_ring_ctx *ctx = req->ctx;
5373 unsigned long flags;
5375 spin_lock_irqsave(&ctx->completion_lock, flags);
5376 list_del_init(&req->timeout.list);
5377 atomic_set(&req->ctx->cq_timeouts,
5378 atomic_read(&req->ctx->cq_timeouts) + 1);
5380 io_cqring_fill_event(req, -ETIME);
5381 io_commit_cqring(ctx);
5382 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5384 io_cqring_ev_posted(ctx);
5385 req_set_fail_links(req);
5387 return HRTIMER_NORESTART;
5390 static int __io_timeout_cancel(struct io_kiocb *req)
5392 struct io_timeout_data *io = req->async_data;
5395 ret = hrtimer_try_to_cancel(&io->timer);
5398 list_del_init(&req->timeout.list);
5400 req_set_fail_links(req);
5401 io_cqring_fill_event(req, -ECANCELED);
5402 io_put_req_deferred(req, 1);
5406 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5408 struct io_kiocb *req;
5411 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5412 if (user_data == req->user_data) {
5421 return __io_timeout_cancel(req);
5424 static int io_timeout_remove_prep(struct io_kiocb *req,
5425 const struct io_uring_sqe *sqe)
5427 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5429 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5431 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5434 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5439 * Remove or update an existing timeout command
5441 static int io_timeout_remove(struct io_kiocb *req)
5443 struct io_ring_ctx *ctx = req->ctx;
5446 spin_lock_irq(&ctx->completion_lock);
5447 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5449 io_cqring_fill_event(req, ret);
5450 io_commit_cqring(ctx);
5451 spin_unlock_irq(&ctx->completion_lock);
5452 io_cqring_ev_posted(ctx);
5454 req_set_fail_links(req);
5459 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5460 bool is_timeout_link)
5462 struct io_timeout_data *data;
5464 u32 off = READ_ONCE(sqe->off);
5466 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5468 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5470 if (off && is_timeout_link)
5472 flags = READ_ONCE(sqe->timeout_flags);
5473 if (flags & ~IORING_TIMEOUT_ABS)
5476 req->timeout.off = off;
5478 if (!req->async_data && io_alloc_async_data(req))
5481 data = req->async_data;
5484 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5487 if (flags & IORING_TIMEOUT_ABS)
5488 data->mode = HRTIMER_MODE_ABS;
5490 data->mode = HRTIMER_MODE_REL;
5492 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5496 static int io_timeout(struct io_kiocb *req)
5498 struct io_ring_ctx *ctx = req->ctx;
5499 struct io_timeout_data *data = req->async_data;
5500 struct list_head *entry;
5501 u32 tail, off = req->timeout.off;
5503 spin_lock_irq(&ctx->completion_lock);
5506 * sqe->off holds how many events that need to occur for this
5507 * timeout event to be satisfied. If it isn't set, then this is
5508 * a pure timeout request, sequence isn't used.
5510 if (io_is_timeout_noseq(req)) {
5511 entry = ctx->timeout_list.prev;
5515 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5516 req->timeout.target_seq = tail + off;
5519 * Insertion sort, ensuring the first entry in the list is always
5520 * the one we need first.
5522 list_for_each_prev(entry, &ctx->timeout_list) {
5523 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5526 if (io_is_timeout_noseq(nxt))
5528 /* nxt.seq is behind @tail, otherwise would've been completed */
5529 if (off >= nxt->timeout.target_seq - tail)
5533 list_add(&req->timeout.list, entry);
5534 data->timer.function = io_timeout_fn;
5535 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5536 spin_unlock_irq(&ctx->completion_lock);
5540 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5542 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5544 return req->user_data == (unsigned long) data;
5547 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5549 enum io_wq_cancel cancel_ret;
5552 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5553 switch (cancel_ret) {
5554 case IO_WQ_CANCEL_OK:
5557 case IO_WQ_CANCEL_RUNNING:
5560 case IO_WQ_CANCEL_NOTFOUND:
5568 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5569 struct io_kiocb *req, __u64 sqe_addr,
5572 unsigned long flags;
5575 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5576 if (ret != -ENOENT) {
5577 spin_lock_irqsave(&ctx->completion_lock, flags);
5581 spin_lock_irqsave(&ctx->completion_lock, flags);
5582 ret = io_timeout_cancel(ctx, sqe_addr);
5585 ret = io_poll_cancel(ctx, sqe_addr);
5589 io_cqring_fill_event(req, ret);
5590 io_commit_cqring(ctx);
5591 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5592 io_cqring_ev_posted(ctx);
5595 req_set_fail_links(req);
5599 static int io_async_cancel_prep(struct io_kiocb *req,
5600 const struct io_uring_sqe *sqe)
5602 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5604 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5606 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5609 req->cancel.addr = READ_ONCE(sqe->addr);
5613 static int io_async_cancel(struct io_kiocb *req)
5615 struct io_ring_ctx *ctx = req->ctx;
5617 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5621 static int io_files_update_prep(struct io_kiocb *req,
5622 const struct io_uring_sqe *sqe)
5624 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5626 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5628 if (sqe->ioprio || sqe->rw_flags)
5631 req->files_update.offset = READ_ONCE(sqe->off);
5632 req->files_update.nr_args = READ_ONCE(sqe->len);
5633 if (!req->files_update.nr_args)
5635 req->files_update.arg = READ_ONCE(sqe->addr);
5639 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5640 struct io_comp_state *cs)
5642 struct io_ring_ctx *ctx = req->ctx;
5643 struct io_uring_files_update up;
5649 up.offset = req->files_update.offset;
5650 up.fds = req->files_update.arg;
5652 mutex_lock(&ctx->uring_lock);
5653 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5654 mutex_unlock(&ctx->uring_lock);
5657 req_set_fail_links(req);
5658 __io_req_complete(req, ret, 0, cs);
5662 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5664 switch (req->opcode) {
5667 case IORING_OP_READV:
5668 case IORING_OP_READ_FIXED:
5669 case IORING_OP_READ:
5670 return io_read_prep(req, sqe);
5671 case IORING_OP_WRITEV:
5672 case IORING_OP_WRITE_FIXED:
5673 case IORING_OP_WRITE:
5674 return io_write_prep(req, sqe);
5675 case IORING_OP_POLL_ADD:
5676 return io_poll_add_prep(req, sqe);
5677 case IORING_OP_POLL_REMOVE:
5678 return io_poll_remove_prep(req, sqe);
5679 case IORING_OP_FSYNC:
5680 return io_prep_fsync(req, sqe);
5681 case IORING_OP_SYNC_FILE_RANGE:
5682 return io_prep_sfr(req, sqe);
5683 case IORING_OP_SENDMSG:
5684 case IORING_OP_SEND:
5685 return io_sendmsg_prep(req, sqe);
5686 case IORING_OP_RECVMSG:
5687 case IORING_OP_RECV:
5688 return io_recvmsg_prep(req, sqe);
5689 case IORING_OP_CONNECT:
5690 return io_connect_prep(req, sqe);
5691 case IORING_OP_TIMEOUT:
5692 return io_timeout_prep(req, sqe, false);
5693 case IORING_OP_TIMEOUT_REMOVE:
5694 return io_timeout_remove_prep(req, sqe);
5695 case IORING_OP_ASYNC_CANCEL:
5696 return io_async_cancel_prep(req, sqe);
5697 case IORING_OP_LINK_TIMEOUT:
5698 return io_timeout_prep(req, sqe, true);
5699 case IORING_OP_ACCEPT:
5700 return io_accept_prep(req, sqe);
5701 case IORING_OP_FALLOCATE:
5702 return io_fallocate_prep(req, sqe);
5703 case IORING_OP_OPENAT:
5704 return io_openat_prep(req, sqe);
5705 case IORING_OP_CLOSE:
5706 return io_close_prep(req, sqe);
5707 case IORING_OP_FILES_UPDATE:
5708 return io_files_update_prep(req, sqe);
5709 case IORING_OP_STATX:
5710 return io_statx_prep(req, sqe);
5711 case IORING_OP_FADVISE:
5712 return io_fadvise_prep(req, sqe);
5713 case IORING_OP_MADVISE:
5714 return io_madvise_prep(req, sqe);
5715 case IORING_OP_OPENAT2:
5716 return io_openat2_prep(req, sqe);
5717 case IORING_OP_EPOLL_CTL:
5718 return io_epoll_ctl_prep(req, sqe);
5719 case IORING_OP_SPLICE:
5720 return io_splice_prep(req, sqe);
5721 case IORING_OP_PROVIDE_BUFFERS:
5722 return io_provide_buffers_prep(req, sqe);
5723 case IORING_OP_REMOVE_BUFFERS:
5724 return io_remove_buffers_prep(req, sqe);
5726 return io_tee_prep(req, sqe);
5729 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5734 static int io_req_defer_prep(struct io_kiocb *req,
5735 const struct io_uring_sqe *sqe)
5739 if (io_alloc_async_data(req))
5741 return io_req_prep(req, sqe);
5744 static u32 io_get_sequence(struct io_kiocb *req)
5746 struct io_kiocb *pos;
5747 struct io_ring_ctx *ctx = req->ctx;
5748 u32 total_submitted, nr_reqs = 1;
5750 if (req->flags & REQ_F_LINK_HEAD)
5751 list_for_each_entry(pos, &req->link_list, link_list)
5754 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5755 return total_submitted - nr_reqs;
5758 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5760 struct io_ring_ctx *ctx = req->ctx;
5761 struct io_defer_entry *de;
5765 /* Still need defer if there is pending req in defer list. */
5766 if (likely(list_empty_careful(&ctx->defer_list) &&
5767 !(req->flags & REQ_F_IO_DRAIN)))
5770 seq = io_get_sequence(req);
5771 /* Still a chance to pass the sequence check */
5772 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5775 if (!req->async_data) {
5776 ret = io_req_defer_prep(req, sqe);
5780 io_prep_async_link(req);
5781 de = kmalloc(sizeof(*de), GFP_KERNEL);
5785 spin_lock_irq(&ctx->completion_lock);
5786 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5787 spin_unlock_irq(&ctx->completion_lock);
5789 io_queue_async_work(req);
5790 return -EIOCBQUEUED;
5793 trace_io_uring_defer(ctx, req, req->user_data);
5796 list_add_tail(&de->list, &ctx->defer_list);
5797 spin_unlock_irq(&ctx->completion_lock);
5798 return -EIOCBQUEUED;
5801 static void io_req_drop_files(struct io_kiocb *req)
5803 struct io_ring_ctx *ctx = req->ctx;
5804 unsigned long flags;
5806 spin_lock_irqsave(&ctx->inflight_lock, flags);
5807 list_del(&req->inflight_entry);
5808 if (waitqueue_active(&ctx->inflight_wait))
5809 wake_up(&ctx->inflight_wait);
5810 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5811 req->flags &= ~REQ_F_INFLIGHT;
5812 put_files_struct(req->work.identity->files);
5813 put_nsproxy(req->work.identity->nsproxy);
5814 req->work.flags &= ~IO_WQ_WORK_FILES;
5817 static void __io_clean_op(struct io_kiocb *req)
5819 if (req->flags & REQ_F_BUFFER_SELECTED) {
5820 switch (req->opcode) {
5821 case IORING_OP_READV:
5822 case IORING_OP_READ_FIXED:
5823 case IORING_OP_READ:
5824 kfree((void *)(unsigned long)req->rw.addr);
5826 case IORING_OP_RECVMSG:
5827 case IORING_OP_RECV:
5828 kfree(req->sr_msg.kbuf);
5831 req->flags &= ~REQ_F_BUFFER_SELECTED;
5834 if (req->flags & REQ_F_NEED_CLEANUP) {
5835 switch (req->opcode) {
5836 case IORING_OP_READV:
5837 case IORING_OP_READ_FIXED:
5838 case IORING_OP_READ:
5839 case IORING_OP_WRITEV:
5840 case IORING_OP_WRITE_FIXED:
5841 case IORING_OP_WRITE: {
5842 struct io_async_rw *io = req->async_data;
5844 kfree(io->free_iovec);
5847 case IORING_OP_RECVMSG:
5848 case IORING_OP_SENDMSG: {
5849 struct io_async_msghdr *io = req->async_data;
5850 if (io->iov != io->fast_iov)
5854 case IORING_OP_SPLICE:
5856 io_put_file(req, req->splice.file_in,
5857 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5859 case IORING_OP_OPENAT:
5860 case IORING_OP_OPENAT2:
5861 if (req->open.filename)
5862 putname(req->open.filename);
5865 req->flags &= ~REQ_F_NEED_CLEANUP;
5868 if (req->flags & REQ_F_INFLIGHT)
5869 io_req_drop_files(req);
5872 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5873 struct io_comp_state *cs)
5875 struct io_ring_ctx *ctx = req->ctx;
5878 switch (req->opcode) {
5880 ret = io_nop(req, cs);
5882 case IORING_OP_READV:
5883 case IORING_OP_READ_FIXED:
5884 case IORING_OP_READ:
5885 ret = io_read(req, force_nonblock, cs);
5887 case IORING_OP_WRITEV:
5888 case IORING_OP_WRITE_FIXED:
5889 case IORING_OP_WRITE:
5890 ret = io_write(req, force_nonblock, cs);
5892 case IORING_OP_FSYNC:
5893 ret = io_fsync(req, force_nonblock);
5895 case IORING_OP_POLL_ADD:
5896 ret = io_poll_add(req);
5898 case IORING_OP_POLL_REMOVE:
5899 ret = io_poll_remove(req);
5901 case IORING_OP_SYNC_FILE_RANGE:
5902 ret = io_sync_file_range(req, force_nonblock);
5904 case IORING_OP_SENDMSG:
5905 ret = io_sendmsg(req, force_nonblock, cs);
5907 case IORING_OP_SEND:
5908 ret = io_send(req, force_nonblock, cs);
5910 case IORING_OP_RECVMSG:
5911 ret = io_recvmsg(req, force_nonblock, cs);
5913 case IORING_OP_RECV:
5914 ret = io_recv(req, force_nonblock, cs);
5916 case IORING_OP_TIMEOUT:
5917 ret = io_timeout(req);
5919 case IORING_OP_TIMEOUT_REMOVE:
5920 ret = io_timeout_remove(req);
5922 case IORING_OP_ACCEPT:
5923 ret = io_accept(req, force_nonblock, cs);
5925 case IORING_OP_CONNECT:
5926 ret = io_connect(req, force_nonblock, cs);
5928 case IORING_OP_ASYNC_CANCEL:
5929 ret = io_async_cancel(req);
5931 case IORING_OP_FALLOCATE:
5932 ret = io_fallocate(req, force_nonblock);
5934 case IORING_OP_OPENAT:
5935 ret = io_openat(req, force_nonblock);
5937 case IORING_OP_CLOSE:
5938 ret = io_close(req, force_nonblock, cs);
5940 case IORING_OP_FILES_UPDATE:
5941 ret = io_files_update(req, force_nonblock, cs);
5943 case IORING_OP_STATX:
5944 ret = io_statx(req, force_nonblock);
5946 case IORING_OP_FADVISE:
5947 ret = io_fadvise(req, force_nonblock);
5949 case IORING_OP_MADVISE:
5950 ret = io_madvise(req, force_nonblock);
5952 case IORING_OP_OPENAT2:
5953 ret = io_openat2(req, force_nonblock);
5955 case IORING_OP_EPOLL_CTL:
5956 ret = io_epoll_ctl(req, force_nonblock, cs);
5958 case IORING_OP_SPLICE:
5959 ret = io_splice(req, force_nonblock);
5961 case IORING_OP_PROVIDE_BUFFERS:
5962 ret = io_provide_buffers(req, force_nonblock, cs);
5964 case IORING_OP_REMOVE_BUFFERS:
5965 ret = io_remove_buffers(req, force_nonblock, cs);
5968 ret = io_tee(req, force_nonblock);
5978 /* If the op doesn't have a file, we're not polling for it */
5979 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5980 const bool in_async = io_wq_current_is_worker();
5982 /* workqueue context doesn't hold uring_lock, grab it now */
5984 mutex_lock(&ctx->uring_lock);
5986 io_iopoll_req_issued(req);
5989 mutex_unlock(&ctx->uring_lock);
5995 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5997 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5998 struct io_kiocb *timeout;
6001 timeout = io_prep_linked_timeout(req);
6003 io_queue_linked_timeout(timeout);
6005 /* if NO_CANCEL is set, we must still run the work */
6006 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6007 IO_WQ_WORK_CANCEL) {
6013 ret = io_issue_sqe(req, false, NULL);
6015 * We can get EAGAIN for polled IO even though we're
6016 * forcing a sync submission from here, since we can't
6017 * wait for request slots on the block side.
6026 req_set_fail_links(req);
6027 io_req_complete(req, ret);
6030 return io_steal_work(req);
6033 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6036 struct fixed_file_table *table;
6038 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6039 return table->files[index & IORING_FILE_TABLE_MASK];
6042 static struct file *io_file_get(struct io_submit_state *state,
6043 struct io_kiocb *req, int fd, bool fixed)
6045 struct io_ring_ctx *ctx = req->ctx;
6049 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6051 fd = array_index_nospec(fd, ctx->nr_user_files);
6052 file = io_file_from_index(ctx, fd);
6054 req->fixed_file_refs = &ctx->file_data->node->refs;
6055 percpu_ref_get(req->fixed_file_refs);
6058 trace_io_uring_file_get(ctx, fd);
6059 file = __io_file_get(state, fd);
6065 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6070 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6071 if (unlikely(!fixed && io_async_submit(req->ctx)))
6074 req->file = io_file_get(state, req, fd, fixed);
6075 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6080 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6082 struct io_timeout_data *data = container_of(timer,
6083 struct io_timeout_data, timer);
6084 struct io_kiocb *req = data->req;
6085 struct io_ring_ctx *ctx = req->ctx;
6086 struct io_kiocb *prev = NULL;
6087 unsigned long flags;
6089 spin_lock_irqsave(&ctx->completion_lock, flags);
6092 * We don't expect the list to be empty, that will only happen if we
6093 * race with the completion of the linked work.
6095 if (!list_empty(&req->link_list)) {
6096 prev = list_entry(req->link_list.prev, struct io_kiocb,
6098 if (refcount_inc_not_zero(&prev->refs))
6099 list_del_init(&req->link_list);
6104 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6107 req_set_fail_links(prev);
6108 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6111 io_req_complete(req, -ETIME);
6113 return HRTIMER_NORESTART;
6116 static void __io_queue_linked_timeout(struct io_kiocb *req)
6119 * If the list is now empty, then our linked request finished before
6120 * we got a chance to setup the timer
6122 if (!list_empty(&req->link_list)) {
6123 struct io_timeout_data *data = req->async_data;
6125 data->timer.function = io_link_timeout_fn;
6126 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6131 static void io_queue_linked_timeout(struct io_kiocb *req)
6133 struct io_ring_ctx *ctx = req->ctx;
6135 spin_lock_irq(&ctx->completion_lock);
6136 __io_queue_linked_timeout(req);
6137 spin_unlock_irq(&ctx->completion_lock);
6139 /* drop submission reference */
6143 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6145 struct io_kiocb *nxt;
6147 if (!(req->flags & REQ_F_LINK_HEAD))
6149 if (req->flags & REQ_F_LINK_TIMEOUT)
6152 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6154 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6157 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6158 req->flags |= REQ_F_LINK_TIMEOUT;
6162 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6164 struct io_kiocb *linked_timeout;
6165 struct io_kiocb *nxt;
6166 const struct cred *old_creds = NULL;
6170 linked_timeout = io_prep_linked_timeout(req);
6172 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6173 (req->work.flags & IO_WQ_WORK_CREDS) &&
6174 req->work.identity->creds != current_cred()) {
6176 revert_creds(old_creds);
6177 if (old_creds == req->work.identity->creds)
6178 old_creds = NULL; /* restored original creds */
6180 old_creds = override_creds(req->work.identity->creds);
6183 ret = io_issue_sqe(req, true, cs);
6186 * We async punt it if the file wasn't marked NOWAIT, or if the file
6187 * doesn't support non-blocking read/write attempts
6189 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6190 if (!io_arm_poll_handler(req)) {
6193 * Queued up for async execution, worker will release
6194 * submit reference when the iocb is actually submitted.
6196 io_queue_async_work(req);
6200 io_queue_linked_timeout(linked_timeout);
6204 if (unlikely(ret)) {
6205 /* un-prep timeout, so it'll be killed as any other linked */
6206 req->flags &= ~REQ_F_LINK_TIMEOUT;
6207 req_set_fail_links(req);
6209 io_req_complete(req, ret);
6213 /* drop submission reference */
6214 nxt = io_put_req_find_next(req);
6216 io_queue_linked_timeout(linked_timeout);
6221 if (req->flags & REQ_F_FORCE_ASYNC) {
6222 linked_timeout = NULL;
6229 revert_creds(old_creds);
6232 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6233 struct io_comp_state *cs)
6237 ret = io_req_defer(req, sqe);
6239 if (ret != -EIOCBQUEUED) {
6241 req_set_fail_links(req);
6243 io_req_complete(req, ret);
6245 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6246 if (!req->async_data) {
6247 ret = io_req_defer_prep(req, sqe);
6251 io_queue_async_work(req);
6254 ret = io_req_prep(req, sqe);
6258 __io_queue_sqe(req, cs);
6262 static inline void io_queue_link_head(struct io_kiocb *req,
6263 struct io_comp_state *cs)
6265 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6267 io_req_complete(req, -ECANCELED);
6269 io_queue_sqe(req, NULL, cs);
6272 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6273 struct io_kiocb **link, struct io_comp_state *cs)
6275 struct io_ring_ctx *ctx = req->ctx;
6279 * If we already have a head request, queue this one for async
6280 * submittal once the head completes. If we don't have a head but
6281 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6282 * submitted sync once the chain is complete. If none of those
6283 * conditions are true (normal request), then just queue it.
6286 struct io_kiocb *head = *link;
6289 * Taking sequential execution of a link, draining both sides
6290 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6291 * requests in the link. So, it drains the head and the
6292 * next after the link request. The last one is done via
6293 * drain_next flag to persist the effect across calls.
6295 if (req->flags & REQ_F_IO_DRAIN) {
6296 head->flags |= REQ_F_IO_DRAIN;
6297 ctx->drain_next = 1;
6299 ret = io_req_defer_prep(req, sqe);
6300 if (unlikely(ret)) {
6301 /* fail even hard links since we don't submit */
6302 head->flags |= REQ_F_FAIL_LINK;
6305 trace_io_uring_link(ctx, req, head);
6306 list_add_tail(&req->link_list, &head->link_list);
6308 /* last request of a link, enqueue the link */
6309 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6310 io_queue_link_head(head, cs);
6314 if (unlikely(ctx->drain_next)) {
6315 req->flags |= REQ_F_IO_DRAIN;
6316 ctx->drain_next = 0;
6318 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6319 req->flags |= REQ_F_LINK_HEAD;
6320 INIT_LIST_HEAD(&req->link_list);
6322 ret = io_req_defer_prep(req, sqe);
6324 req->flags |= REQ_F_FAIL_LINK;
6327 io_queue_sqe(req, sqe, cs);
6335 * Batched submission is done, ensure local IO is flushed out.
6337 static void io_submit_state_end(struct io_submit_state *state)
6339 if (!list_empty(&state->comp.list))
6340 io_submit_flush_completions(&state->comp);
6341 blk_finish_plug(&state->plug);
6342 io_state_file_put(state);
6343 if (state->free_reqs)
6344 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6348 * Start submission side cache.
6350 static void io_submit_state_start(struct io_submit_state *state,
6351 struct io_ring_ctx *ctx, unsigned int max_ios)
6353 blk_start_plug(&state->plug);
6355 INIT_LIST_HEAD(&state->comp.list);
6356 state->comp.ctx = ctx;
6357 state->free_reqs = 0;
6359 state->ios_left = max_ios;
6362 static void io_commit_sqring(struct io_ring_ctx *ctx)
6364 struct io_rings *rings = ctx->rings;
6367 * Ensure any loads from the SQEs are done at this point,
6368 * since once we write the new head, the application could
6369 * write new data to them.
6371 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6375 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6376 * that is mapped by userspace. This means that care needs to be taken to
6377 * ensure that reads are stable, as we cannot rely on userspace always
6378 * being a good citizen. If members of the sqe are validated and then later
6379 * used, it's important that those reads are done through READ_ONCE() to
6380 * prevent a re-load down the line.
6382 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6384 u32 *sq_array = ctx->sq_array;
6388 * The cached sq head (or cq tail) serves two purposes:
6390 * 1) allows us to batch the cost of updating the user visible
6392 * 2) allows the kernel side to track the head on its own, even
6393 * though the application is the one updating it.
6395 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6396 if (likely(head < ctx->sq_entries))
6397 return &ctx->sq_sqes[head];
6399 /* drop invalid entries */
6400 ctx->cached_sq_dropped++;
6401 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6405 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6407 ctx->cached_sq_head++;
6411 * Check SQE restrictions (opcode and flags).
6413 * Returns 'true' if SQE is allowed, 'false' otherwise.
6415 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6416 struct io_kiocb *req,
6417 unsigned int sqe_flags)
6419 if (!ctx->restricted)
6422 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6425 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6426 ctx->restrictions.sqe_flags_required)
6429 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6430 ctx->restrictions.sqe_flags_required))
6436 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6437 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6438 IOSQE_BUFFER_SELECT)
6440 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6441 const struct io_uring_sqe *sqe,
6442 struct io_submit_state *state)
6444 unsigned int sqe_flags;
6447 req->opcode = READ_ONCE(sqe->opcode);
6448 req->user_data = READ_ONCE(sqe->user_data);
6449 req->async_data = NULL;
6453 /* one is dropped after submission, the other at completion */
6454 refcount_set(&req->refs, 2);
6455 req->task = current;
6458 if (unlikely(req->opcode >= IORING_OP_LAST))
6461 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6464 sqe_flags = READ_ONCE(sqe->flags);
6465 /* enforce forwards compatibility on users */
6466 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6469 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6472 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6473 !io_op_defs[req->opcode].buffer_select)
6476 id = READ_ONCE(sqe->personality);
6478 struct io_identity *iod;
6480 iod = idr_find(&ctx->personality_idr, id);
6483 refcount_inc(&iod->count);
6485 __io_req_init_async(req);
6486 get_cred(iod->creds);
6487 req->work.identity = iod;
6488 req->work.flags |= IO_WQ_WORK_CREDS;
6491 /* same numerical values with corresponding REQ_F_*, safe to copy */
6492 req->flags |= sqe_flags;
6494 if (!io_op_defs[req->opcode].needs_file)
6497 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6502 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6504 struct io_submit_state state;
6505 struct io_kiocb *link = NULL;
6506 int i, submitted = 0;
6508 /* if we have a backlog and couldn't flush it all, return BUSY */
6509 if (test_bit(0, &ctx->sq_check_overflow)) {
6510 if (!list_empty(&ctx->cq_overflow_list) &&
6511 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6515 /* make sure SQ entry isn't read before tail */
6516 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6518 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6521 percpu_counter_add(¤t->io_uring->inflight, nr);
6522 refcount_add(nr, ¤t->usage);
6524 io_submit_state_start(&state, ctx, nr);
6526 for (i = 0; i < nr; i++) {
6527 const struct io_uring_sqe *sqe;
6528 struct io_kiocb *req;
6531 sqe = io_get_sqe(ctx);
6532 if (unlikely(!sqe)) {
6533 io_consume_sqe(ctx);
6536 req = io_alloc_req(ctx, &state);
6537 if (unlikely(!req)) {
6539 submitted = -EAGAIN;
6542 io_consume_sqe(ctx);
6543 /* will complete beyond this point, count as submitted */
6546 err = io_init_req(ctx, req, sqe, &state);
6547 if (unlikely(err)) {
6550 io_req_complete(req, err);
6554 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6555 true, io_async_submit(ctx));
6556 err = io_submit_sqe(req, sqe, &link, &state.comp);
6561 if (unlikely(submitted != nr)) {
6562 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6563 struct io_uring_task *tctx = current->io_uring;
6564 int unused = nr - ref_used;
6566 percpu_ref_put_many(&ctx->refs, unused);
6567 percpu_counter_sub(&tctx->inflight, unused);
6568 put_task_struct_many(current, unused);
6571 io_queue_link_head(link, &state.comp);
6572 io_submit_state_end(&state);
6574 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6575 io_commit_sqring(ctx);
6580 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6582 /* Tell userspace we may need a wakeup call */
6583 spin_lock_irq(&ctx->completion_lock);
6584 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6585 spin_unlock_irq(&ctx->completion_lock);
6588 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6590 spin_lock_irq(&ctx->completion_lock);
6591 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6592 spin_unlock_irq(&ctx->completion_lock);
6595 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6596 int sync, void *key)
6598 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6601 ret = autoremove_wake_function(wqe, mode, sync, key);
6603 unsigned long flags;
6605 spin_lock_irqsave(&ctx->completion_lock, flags);
6606 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6607 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6618 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6619 unsigned long start_jiffies, bool cap_entries)
6621 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6622 struct io_sq_data *sqd = ctx->sq_data;
6623 unsigned int to_submit;
6627 if (!list_empty(&ctx->iopoll_list)) {
6628 unsigned nr_events = 0;
6630 mutex_lock(&ctx->uring_lock);
6631 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6632 io_do_iopoll(ctx, &nr_events, 0);
6633 mutex_unlock(&ctx->uring_lock);
6636 to_submit = io_sqring_entries(ctx);
6639 * If submit got -EBUSY, flag us as needing the application
6640 * to enter the kernel to reap and flush events.
6642 if (!to_submit || ret == -EBUSY || need_resched()) {
6644 * Drop cur_mm before scheduling, we can't hold it for
6645 * long periods (or over schedule()). Do this before
6646 * adding ourselves to the waitqueue, as the unuse/drop
6649 io_sq_thread_drop_mm();
6652 * We're polling. If we're within the defined idle
6653 * period, then let us spin without work before going
6654 * to sleep. The exception is if we got EBUSY doing
6655 * more IO, we should wait for the application to
6656 * reap events and wake us up.
6658 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6659 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6660 !percpu_ref_is_dying(&ctx->refs)))
6663 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6664 TASK_INTERRUPTIBLE);
6667 * While doing polled IO, before going to sleep, we need
6668 * to check if there are new reqs added to iopoll_list,
6669 * it is because reqs may have been punted to io worker
6670 * and will be added to iopoll_list later, hence check
6671 * the iopoll_list again.
6673 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6674 !list_empty_careful(&ctx->iopoll_list)) {
6675 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6679 to_submit = io_sqring_entries(ctx);
6680 if (!to_submit || ret == -EBUSY)
6684 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6685 io_ring_clear_wakeup_flag(ctx);
6687 /* if we're handling multiple rings, cap submit size for fairness */
6688 if (cap_entries && to_submit > 8)
6691 mutex_lock(&ctx->uring_lock);
6692 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6693 ret = io_submit_sqes(ctx, to_submit);
6694 mutex_unlock(&ctx->uring_lock);
6696 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6697 wake_up(&ctx->sqo_sq_wait);
6699 return SQT_DID_WORK;
6702 static void io_sqd_init_new(struct io_sq_data *sqd)
6704 struct io_ring_ctx *ctx;
6706 while (!list_empty(&sqd->ctx_new_list)) {
6707 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6708 init_wait(&ctx->sqo_wait_entry);
6709 ctx->sqo_wait_entry.func = io_sq_wake_function;
6710 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6711 complete(&ctx->sq_thread_comp);
6715 static int io_sq_thread(void *data)
6717 struct cgroup_subsys_state *cur_css = NULL;
6718 const struct cred *old_cred = NULL;
6719 struct io_sq_data *sqd = data;
6720 struct io_ring_ctx *ctx;
6721 unsigned long start_jiffies;
6723 start_jiffies = jiffies;
6724 while (!kthread_should_stop()) {
6725 enum sq_ret ret = 0;
6729 * Any changes to the sqd lists are synchronized through the
6730 * kthread parking. This synchronizes the thread vs users,
6731 * the users are synchronized on the sqd->ctx_lock.
6733 if (kthread_should_park())
6736 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6737 io_sqd_init_new(sqd);
6739 cap_entries = !list_is_singular(&sqd->ctx_list);
6741 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6742 if (current->cred != ctx->creds) {
6744 revert_creds(old_cred);
6745 old_cred = override_creds(ctx->creds);
6747 io_sq_thread_associate_blkcg(ctx, &cur_css);
6749 current->loginuid = ctx->loginuid;
6750 current->sessionid = ctx->sessionid;
6753 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6755 io_sq_thread_drop_mm();
6758 if (ret & SQT_SPIN) {
6761 } else if (ret == SQT_IDLE) {
6762 if (kthread_should_park())
6764 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6765 io_ring_set_wakeup_flag(ctx);
6767 start_jiffies = jiffies;
6768 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6769 io_ring_clear_wakeup_flag(ctx);
6776 io_sq_thread_unassociate_blkcg();
6778 revert_creds(old_cred);
6785 struct io_wait_queue {
6786 struct wait_queue_entry wq;
6787 struct io_ring_ctx *ctx;
6789 unsigned nr_timeouts;
6792 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6794 struct io_ring_ctx *ctx = iowq->ctx;
6797 * Wake up if we have enough events, or if a timeout occurred since we
6798 * started waiting. For timeouts, we always want to return to userspace,
6799 * regardless of event count.
6801 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6802 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6805 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6806 int wake_flags, void *key)
6808 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6811 /* use noflush == true, as we can't safely rely on locking context */
6812 if (!io_should_wake(iowq, true))
6815 return autoremove_wake_function(curr, mode, wake_flags, key);
6818 static int io_run_task_work_sig(void)
6820 if (io_run_task_work())
6822 if (!signal_pending(current))
6824 if (current->jobctl & JOBCTL_TASK_WORK) {
6825 spin_lock_irq(¤t->sighand->siglock);
6826 current->jobctl &= ~JOBCTL_TASK_WORK;
6827 recalc_sigpending();
6828 spin_unlock_irq(¤t->sighand->siglock);
6835 * Wait until events become available, if we don't already have some. The
6836 * application must reap them itself, as they reside on the shared cq ring.
6838 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6839 const sigset_t __user *sig, size_t sigsz)
6841 struct io_wait_queue iowq = {
6844 .func = io_wake_function,
6845 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6848 .to_wait = min_events,
6850 struct io_rings *rings = ctx->rings;
6854 if (io_cqring_events(ctx, false) >= min_events)
6856 if (!io_run_task_work())
6861 #ifdef CONFIG_COMPAT
6862 if (in_compat_syscall())
6863 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6867 ret = set_user_sigmask(sig, sigsz);
6873 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6874 trace_io_uring_cqring_wait(ctx, min_events);
6876 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6877 TASK_INTERRUPTIBLE);
6878 /* make sure we run task_work before checking for signals */
6879 ret = io_run_task_work_sig();
6884 if (io_should_wake(&iowq, false))
6888 finish_wait(&ctx->wait, &iowq.wq);
6890 restore_saved_sigmask_unless(ret == -EINTR);
6892 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6895 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6897 #if defined(CONFIG_UNIX)
6898 if (ctx->ring_sock) {
6899 struct sock *sock = ctx->ring_sock->sk;
6900 struct sk_buff *skb;
6902 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6908 for (i = 0; i < ctx->nr_user_files; i++) {
6911 file = io_file_from_index(ctx, i);
6918 static void io_file_ref_kill(struct percpu_ref *ref)
6920 struct fixed_file_data *data;
6922 data = container_of(ref, struct fixed_file_data, refs);
6923 complete(&data->done);
6926 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6928 struct fixed_file_data *data = ctx->file_data;
6929 struct fixed_file_ref_node *ref_node = NULL;
6930 unsigned nr_tables, i;
6935 spin_lock(&data->lock);
6936 if (!list_empty(&data->ref_list))
6937 ref_node = list_first_entry(&data->ref_list,
6938 struct fixed_file_ref_node, node);
6939 spin_unlock(&data->lock);
6941 percpu_ref_kill(&ref_node->refs);
6943 percpu_ref_kill(&data->refs);
6945 /* wait for all refs nodes to complete */
6946 flush_delayed_work(&ctx->file_put_work);
6947 wait_for_completion(&data->done);
6949 __io_sqe_files_unregister(ctx);
6950 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6951 for (i = 0; i < nr_tables; i++)
6952 kfree(data->table[i].files);
6954 percpu_ref_exit(&data->refs);
6956 ctx->file_data = NULL;
6957 ctx->nr_user_files = 0;
6961 static void io_put_sq_data(struct io_sq_data *sqd)
6963 if (refcount_dec_and_test(&sqd->refs)) {
6965 * The park is a bit of a work-around, without it we get
6966 * warning spews on shutdown with SQPOLL set and affinity
6967 * set to a single CPU.
6970 kthread_park(sqd->thread);
6971 kthread_stop(sqd->thread);
6978 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6980 struct io_ring_ctx *ctx_attach;
6981 struct io_sq_data *sqd;
6984 f = fdget(p->wq_fd);
6986 return ERR_PTR(-ENXIO);
6987 if (f.file->f_op != &io_uring_fops) {
6989 return ERR_PTR(-EINVAL);
6992 ctx_attach = f.file->private_data;
6993 sqd = ctx_attach->sq_data;
6996 return ERR_PTR(-EINVAL);
6999 refcount_inc(&sqd->refs);
7004 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7006 struct io_sq_data *sqd;
7008 if (p->flags & IORING_SETUP_ATTACH_WQ)
7009 return io_attach_sq_data(p);
7011 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7013 return ERR_PTR(-ENOMEM);
7015 refcount_set(&sqd->refs, 1);
7016 INIT_LIST_HEAD(&sqd->ctx_list);
7017 INIT_LIST_HEAD(&sqd->ctx_new_list);
7018 mutex_init(&sqd->ctx_lock);
7019 mutex_init(&sqd->lock);
7020 init_waitqueue_head(&sqd->wait);
7024 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7025 __releases(&sqd->lock)
7029 kthread_unpark(sqd->thread);
7030 mutex_unlock(&sqd->lock);
7033 static void io_sq_thread_park(struct io_sq_data *sqd)
7034 __acquires(&sqd->lock)
7038 mutex_lock(&sqd->lock);
7039 kthread_park(sqd->thread);
7042 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7044 struct io_sq_data *sqd = ctx->sq_data;
7049 * We may arrive here from the error branch in
7050 * io_sq_offload_create() where the kthread is created
7051 * without being waked up, thus wake it up now to make
7052 * sure the wait will complete.
7054 wake_up_process(sqd->thread);
7055 wait_for_completion(&ctx->sq_thread_comp);
7057 io_sq_thread_park(sqd);
7060 mutex_lock(&sqd->ctx_lock);
7061 list_del(&ctx->sqd_list);
7062 mutex_unlock(&sqd->ctx_lock);
7065 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7066 io_sq_thread_unpark(sqd);
7069 io_put_sq_data(sqd);
7070 ctx->sq_data = NULL;
7074 static void io_finish_async(struct io_ring_ctx *ctx)
7076 io_sq_thread_stop(ctx);
7079 io_wq_destroy(ctx->io_wq);
7084 #if defined(CONFIG_UNIX)
7086 * Ensure the UNIX gc is aware of our file set, so we are certain that
7087 * the io_uring can be safely unregistered on process exit, even if we have
7088 * loops in the file referencing.
7090 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7092 struct sock *sk = ctx->ring_sock->sk;
7093 struct scm_fp_list *fpl;
7094 struct sk_buff *skb;
7097 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7101 skb = alloc_skb(0, GFP_KERNEL);
7110 fpl->user = get_uid(ctx->user);
7111 for (i = 0; i < nr; i++) {
7112 struct file *file = io_file_from_index(ctx, i + offset);
7116 fpl->fp[nr_files] = get_file(file);
7117 unix_inflight(fpl->user, fpl->fp[nr_files]);
7122 fpl->max = SCM_MAX_FD;
7123 fpl->count = nr_files;
7124 UNIXCB(skb).fp = fpl;
7125 skb->destructor = unix_destruct_scm;
7126 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7127 skb_queue_head(&sk->sk_receive_queue, skb);
7129 for (i = 0; i < nr_files; i++)
7140 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7141 * causes regular reference counting to break down. We rely on the UNIX
7142 * garbage collection to take care of this problem for us.
7144 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7146 unsigned left, total;
7150 left = ctx->nr_user_files;
7152 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7154 ret = __io_sqe_files_scm(ctx, this_files, total);
7158 total += this_files;
7164 while (total < ctx->nr_user_files) {
7165 struct file *file = io_file_from_index(ctx, total);
7175 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7181 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7182 unsigned nr_tables, unsigned nr_files)
7186 for (i = 0; i < nr_tables; i++) {
7187 struct fixed_file_table *table = &file_data->table[i];
7188 unsigned this_files;
7190 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7191 table->files = kcalloc(this_files, sizeof(struct file *),
7195 nr_files -= this_files;
7201 for (i = 0; i < nr_tables; i++) {
7202 struct fixed_file_table *table = &file_data->table[i];
7203 kfree(table->files);
7208 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7210 #if defined(CONFIG_UNIX)
7211 struct sock *sock = ctx->ring_sock->sk;
7212 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7213 struct sk_buff *skb;
7216 __skb_queue_head_init(&list);
7219 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7220 * remove this entry and rearrange the file array.
7222 skb = skb_dequeue(head);
7224 struct scm_fp_list *fp;
7226 fp = UNIXCB(skb).fp;
7227 for (i = 0; i < fp->count; i++) {
7230 if (fp->fp[i] != file)
7233 unix_notinflight(fp->user, fp->fp[i]);
7234 left = fp->count - 1 - i;
7236 memmove(&fp->fp[i], &fp->fp[i + 1],
7237 left * sizeof(struct file *));
7244 __skb_queue_tail(&list, skb);
7254 __skb_queue_tail(&list, skb);
7256 skb = skb_dequeue(head);
7259 if (skb_peek(&list)) {
7260 spin_lock_irq(&head->lock);
7261 while ((skb = __skb_dequeue(&list)) != NULL)
7262 __skb_queue_tail(head, skb);
7263 spin_unlock_irq(&head->lock);
7270 struct io_file_put {
7271 struct list_head list;
7275 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7277 struct fixed_file_data *file_data = ref_node->file_data;
7278 struct io_ring_ctx *ctx = file_data->ctx;
7279 struct io_file_put *pfile, *tmp;
7281 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7282 list_del(&pfile->list);
7283 io_ring_file_put(ctx, pfile->file);
7287 spin_lock(&file_data->lock);
7288 list_del(&ref_node->node);
7289 spin_unlock(&file_data->lock);
7291 percpu_ref_exit(&ref_node->refs);
7293 percpu_ref_put(&file_data->refs);
7296 static void io_file_put_work(struct work_struct *work)
7298 struct io_ring_ctx *ctx;
7299 struct llist_node *node;
7301 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7302 node = llist_del_all(&ctx->file_put_llist);
7305 struct fixed_file_ref_node *ref_node;
7306 struct llist_node *next = node->next;
7308 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7309 __io_file_put_work(ref_node);
7314 static void io_file_data_ref_zero(struct percpu_ref *ref)
7316 struct fixed_file_ref_node *ref_node;
7317 struct io_ring_ctx *ctx;
7321 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7322 ctx = ref_node->file_data->ctx;
7324 if (percpu_ref_is_dying(&ctx->file_data->refs))
7327 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7329 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7331 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7334 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7335 struct io_ring_ctx *ctx)
7337 struct fixed_file_ref_node *ref_node;
7339 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7341 return ERR_PTR(-ENOMEM);
7343 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7346 return ERR_PTR(-ENOMEM);
7348 INIT_LIST_HEAD(&ref_node->node);
7349 INIT_LIST_HEAD(&ref_node->file_list);
7350 ref_node->file_data = ctx->file_data;
7354 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7356 percpu_ref_exit(&ref_node->refs);
7360 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7363 __s32 __user *fds = (__s32 __user *) arg;
7364 unsigned nr_tables, i;
7366 int fd, ret = -ENOMEM;
7367 struct fixed_file_ref_node *ref_node;
7368 struct fixed_file_data *file_data;
7374 if (nr_args > IORING_MAX_FIXED_FILES)
7377 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7380 file_data->ctx = ctx;
7381 init_completion(&file_data->done);
7382 INIT_LIST_HEAD(&file_data->ref_list);
7383 spin_lock_init(&file_data->lock);
7385 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7386 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7388 if (!file_data->table)
7391 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7392 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7395 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7397 ctx->file_data = file_data;
7399 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7400 struct fixed_file_table *table;
7403 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7407 /* allow sparse sets */
7417 * Don't allow io_uring instances to be registered. If UNIX
7418 * isn't enabled, then this causes a reference cycle and this
7419 * instance can never get freed. If UNIX is enabled we'll
7420 * handle it just fine, but there's still no point in allowing
7421 * a ring fd as it doesn't support regular read/write anyway.
7423 if (file->f_op == &io_uring_fops) {
7427 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7428 index = i & IORING_FILE_TABLE_MASK;
7429 table->files[index] = file;
7432 ret = io_sqe_files_scm(ctx);
7434 io_sqe_files_unregister(ctx);
7438 ref_node = alloc_fixed_file_ref_node(ctx);
7439 if (IS_ERR(ref_node)) {
7440 io_sqe_files_unregister(ctx);
7441 return PTR_ERR(ref_node);
7444 file_data->node = ref_node;
7445 spin_lock(&file_data->lock);
7446 list_add(&ref_node->node, &file_data->ref_list);
7447 spin_unlock(&file_data->lock);
7448 percpu_ref_get(&file_data->refs);
7451 for (i = 0; i < ctx->nr_user_files; i++) {
7452 file = io_file_from_index(ctx, i);
7456 for (i = 0; i < nr_tables; i++)
7457 kfree(file_data->table[i].files);
7458 ctx->nr_user_files = 0;
7460 percpu_ref_exit(&file_data->refs);
7462 kfree(file_data->table);
7464 ctx->file_data = NULL;
7468 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7471 #if defined(CONFIG_UNIX)
7472 struct sock *sock = ctx->ring_sock->sk;
7473 struct sk_buff_head *head = &sock->sk_receive_queue;
7474 struct sk_buff *skb;
7477 * See if we can merge this file into an existing skb SCM_RIGHTS
7478 * file set. If there's no room, fall back to allocating a new skb
7479 * and filling it in.
7481 spin_lock_irq(&head->lock);
7482 skb = skb_peek(head);
7484 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7486 if (fpl->count < SCM_MAX_FD) {
7487 __skb_unlink(skb, head);
7488 spin_unlock_irq(&head->lock);
7489 fpl->fp[fpl->count] = get_file(file);
7490 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7492 spin_lock_irq(&head->lock);
7493 __skb_queue_head(head, skb);
7498 spin_unlock_irq(&head->lock);
7505 return __io_sqe_files_scm(ctx, 1, index);
7511 static int io_queue_file_removal(struct fixed_file_data *data,
7514 struct io_file_put *pfile;
7515 struct fixed_file_ref_node *ref_node = data->node;
7517 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7522 list_add(&pfile->list, &ref_node->file_list);
7527 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7528 struct io_uring_files_update *up,
7531 struct fixed_file_data *data = ctx->file_data;
7532 struct fixed_file_ref_node *ref_node;
7537 bool needs_switch = false;
7539 if (check_add_overflow(up->offset, nr_args, &done))
7541 if (done > ctx->nr_user_files)
7544 ref_node = alloc_fixed_file_ref_node(ctx);
7545 if (IS_ERR(ref_node))
7546 return PTR_ERR(ref_node);
7549 fds = u64_to_user_ptr(up->fds);
7551 struct fixed_file_table *table;
7555 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7559 i = array_index_nospec(up->offset, ctx->nr_user_files);
7560 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7561 index = i & IORING_FILE_TABLE_MASK;
7562 if (table->files[index]) {
7563 file = table->files[index];
7564 err = io_queue_file_removal(data, file);
7567 table->files[index] = NULL;
7568 needs_switch = true;
7577 * Don't allow io_uring instances to be registered. If
7578 * UNIX isn't enabled, then this causes a reference
7579 * cycle and this instance can never get freed. If UNIX
7580 * is enabled we'll handle it just fine, but there's
7581 * still no point in allowing a ring fd as it doesn't
7582 * support regular read/write anyway.
7584 if (file->f_op == &io_uring_fops) {
7589 table->files[index] = file;
7590 err = io_sqe_file_register(ctx, file, i);
7592 table->files[index] = NULL;
7603 percpu_ref_kill(&data->node->refs);
7604 spin_lock(&data->lock);
7605 list_add(&ref_node->node, &data->ref_list);
7606 data->node = ref_node;
7607 spin_unlock(&data->lock);
7608 percpu_ref_get(&ctx->file_data->refs);
7610 destroy_fixed_file_ref_node(ref_node);
7612 return done ? done : err;
7615 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7618 struct io_uring_files_update up;
7620 if (!ctx->file_data)
7624 if (copy_from_user(&up, arg, sizeof(up)))
7629 return __io_sqe_files_update(ctx, &up, nr_args);
7632 static void io_free_work(struct io_wq_work *work)
7634 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7636 /* Consider that io_steal_work() relies on this ref */
7640 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7641 struct io_uring_params *p)
7643 struct io_wq_data data;
7645 struct io_ring_ctx *ctx_attach;
7646 unsigned int concurrency;
7649 data.user = ctx->user;
7650 data.free_work = io_free_work;
7651 data.do_work = io_wq_submit_work;
7653 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7654 /* Do QD, or 4 * CPUS, whatever is smallest */
7655 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7657 ctx->io_wq = io_wq_create(concurrency, &data);
7658 if (IS_ERR(ctx->io_wq)) {
7659 ret = PTR_ERR(ctx->io_wq);
7665 f = fdget(p->wq_fd);
7669 if (f.file->f_op != &io_uring_fops) {
7674 ctx_attach = f.file->private_data;
7675 /* @io_wq is protected by holding the fd */
7676 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7681 ctx->io_wq = ctx_attach->io_wq;
7687 static int io_uring_alloc_task_context(struct task_struct *task)
7689 struct io_uring_task *tctx;
7692 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7693 if (unlikely(!tctx))
7696 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7697 if (unlikely(ret)) {
7703 init_waitqueue_head(&tctx->wait);
7706 io_init_identity(&tctx->__identity);
7707 tctx->identity = &tctx->__identity;
7708 task->io_uring = tctx;
7712 void __io_uring_free(struct task_struct *tsk)
7714 struct io_uring_task *tctx = tsk->io_uring;
7716 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7717 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7718 if (tctx->identity != &tctx->__identity)
7719 kfree(tctx->identity);
7720 percpu_counter_destroy(&tctx->inflight);
7722 tsk->io_uring = NULL;
7725 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7726 struct io_uring_params *p)
7730 if (ctx->flags & IORING_SETUP_SQPOLL) {
7731 struct io_sq_data *sqd;
7734 if (!capable(CAP_SYS_ADMIN))
7737 sqd = io_get_sq_data(p);
7744 io_sq_thread_park(sqd);
7745 mutex_lock(&sqd->ctx_lock);
7746 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7747 mutex_unlock(&sqd->ctx_lock);
7748 io_sq_thread_unpark(sqd);
7750 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7751 if (!ctx->sq_thread_idle)
7752 ctx->sq_thread_idle = HZ;
7757 if (p->flags & IORING_SETUP_SQ_AFF) {
7758 int cpu = p->sq_thread_cpu;
7761 if (cpu >= nr_cpu_ids)
7763 if (!cpu_online(cpu))
7766 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7767 cpu, "io_uring-sq");
7769 sqd->thread = kthread_create(io_sq_thread, sqd,
7772 if (IS_ERR(sqd->thread)) {
7773 ret = PTR_ERR(sqd->thread);
7777 ret = io_uring_alloc_task_context(sqd->thread);
7780 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7781 /* Can't have SQ_AFF without SQPOLL */
7787 ret = io_init_wq_offload(ctx, p);
7793 io_finish_async(ctx);
7797 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7799 struct io_sq_data *sqd = ctx->sq_data;
7801 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7802 wake_up_process(sqd->thread);
7805 static inline void __io_unaccount_mem(struct user_struct *user,
7806 unsigned long nr_pages)
7808 atomic_long_sub(nr_pages, &user->locked_vm);
7811 static inline int __io_account_mem(struct user_struct *user,
7812 unsigned long nr_pages)
7814 unsigned long page_limit, cur_pages, new_pages;
7816 /* Don't allow more pages than we can safely lock */
7817 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7820 cur_pages = atomic_long_read(&user->locked_vm);
7821 new_pages = cur_pages + nr_pages;
7822 if (new_pages > page_limit)
7824 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7825 new_pages) != cur_pages);
7830 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7831 enum io_mem_account acct)
7834 __io_unaccount_mem(ctx->user, nr_pages);
7836 if (ctx->mm_account) {
7837 if (acct == ACCT_LOCKED)
7838 ctx->mm_account->locked_vm -= nr_pages;
7839 else if (acct == ACCT_PINNED)
7840 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7844 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7845 enum io_mem_account acct)
7849 if (ctx->limit_mem) {
7850 ret = __io_account_mem(ctx->user, nr_pages);
7855 if (ctx->mm_account) {
7856 if (acct == ACCT_LOCKED)
7857 ctx->mm_account->locked_vm += nr_pages;
7858 else if (acct == ACCT_PINNED)
7859 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7865 static void io_mem_free(void *ptr)
7872 page = virt_to_head_page(ptr);
7873 if (put_page_testzero(page))
7874 free_compound_page(page);
7877 static void *io_mem_alloc(size_t size)
7879 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7882 return (void *) __get_free_pages(gfp_flags, get_order(size));
7885 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7888 struct io_rings *rings;
7889 size_t off, sq_array_size;
7891 off = struct_size(rings, cqes, cq_entries);
7892 if (off == SIZE_MAX)
7896 off = ALIGN(off, SMP_CACHE_BYTES);
7904 sq_array_size = array_size(sizeof(u32), sq_entries);
7905 if (sq_array_size == SIZE_MAX)
7908 if (check_add_overflow(off, sq_array_size, &off))
7914 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7918 pages = (size_t)1 << get_order(
7919 rings_size(sq_entries, cq_entries, NULL));
7920 pages += (size_t)1 << get_order(
7921 array_size(sizeof(struct io_uring_sqe), sq_entries));
7926 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7930 if (!ctx->user_bufs)
7933 for (i = 0; i < ctx->nr_user_bufs; i++) {
7934 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7936 for (j = 0; j < imu->nr_bvecs; j++)
7937 unpin_user_page(imu->bvec[j].bv_page);
7939 if (imu->acct_pages)
7940 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7945 kfree(ctx->user_bufs);
7946 ctx->user_bufs = NULL;
7947 ctx->nr_user_bufs = 0;
7951 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7952 void __user *arg, unsigned index)
7954 struct iovec __user *src;
7956 #ifdef CONFIG_COMPAT
7958 struct compat_iovec __user *ciovs;
7959 struct compat_iovec ciov;
7961 ciovs = (struct compat_iovec __user *) arg;
7962 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7965 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7966 dst->iov_len = ciov.iov_len;
7970 src = (struct iovec __user *) arg;
7971 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7977 * Not super efficient, but this is just a registration time. And we do cache
7978 * the last compound head, so generally we'll only do a full search if we don't
7981 * We check if the given compound head page has already been accounted, to
7982 * avoid double accounting it. This allows us to account the full size of the
7983 * page, not just the constituent pages of a huge page.
7985 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7986 int nr_pages, struct page *hpage)
7990 /* check current page array */
7991 for (i = 0; i < nr_pages; i++) {
7992 if (!PageCompound(pages[i]))
7994 if (compound_head(pages[i]) == hpage)
7998 /* check previously registered pages */
7999 for (i = 0; i < ctx->nr_user_bufs; i++) {
8000 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8002 for (j = 0; j < imu->nr_bvecs; j++) {
8003 if (!PageCompound(imu->bvec[j].bv_page))
8005 if (compound_head(imu->bvec[j].bv_page) == hpage)
8013 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8014 int nr_pages, struct io_mapped_ubuf *imu,
8015 struct page **last_hpage)
8019 for (i = 0; i < nr_pages; i++) {
8020 if (!PageCompound(pages[i])) {
8025 hpage = compound_head(pages[i]);
8026 if (hpage == *last_hpage)
8028 *last_hpage = hpage;
8029 if (headpage_already_acct(ctx, pages, i, hpage))
8031 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8035 if (!imu->acct_pages)
8038 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8040 imu->acct_pages = 0;
8044 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8047 struct vm_area_struct **vmas = NULL;
8048 struct page **pages = NULL;
8049 struct page *last_hpage = NULL;
8050 int i, j, got_pages = 0;
8055 if (!nr_args || nr_args > UIO_MAXIOV)
8058 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8060 if (!ctx->user_bufs)
8063 for (i = 0; i < nr_args; i++) {
8064 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8065 unsigned long off, start, end, ubuf;
8070 ret = io_copy_iov(ctx, &iov, arg, i);
8075 * Don't impose further limits on the size and buffer
8076 * constraints here, we'll -EINVAL later when IO is
8077 * submitted if they are wrong.
8080 if (!iov.iov_base || !iov.iov_len)
8083 /* arbitrary limit, but we need something */
8084 if (iov.iov_len > SZ_1G)
8087 ubuf = (unsigned long) iov.iov_base;
8088 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8089 start = ubuf >> PAGE_SHIFT;
8090 nr_pages = end - start;
8093 if (!pages || nr_pages > got_pages) {
8096 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8098 vmas = kvmalloc_array(nr_pages,
8099 sizeof(struct vm_area_struct *),
8101 if (!pages || !vmas) {
8105 got_pages = nr_pages;
8108 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8115 mmap_read_lock(current->mm);
8116 pret = pin_user_pages(ubuf, nr_pages,
8117 FOLL_WRITE | FOLL_LONGTERM,
8119 if (pret == nr_pages) {
8120 /* don't support file backed memory */
8121 for (j = 0; j < nr_pages; j++) {
8122 struct vm_area_struct *vma = vmas[j];
8125 !is_file_hugepages(vma->vm_file)) {
8131 ret = pret < 0 ? pret : -EFAULT;
8133 mmap_read_unlock(current->mm);
8136 * if we did partial map, or found file backed vmas,
8137 * release any pages we did get
8140 unpin_user_pages(pages, pret);
8145 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8147 unpin_user_pages(pages, pret);
8152 off = ubuf & ~PAGE_MASK;
8154 for (j = 0; j < nr_pages; j++) {
8157 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8158 imu->bvec[j].bv_page = pages[j];
8159 imu->bvec[j].bv_len = vec_len;
8160 imu->bvec[j].bv_offset = off;
8164 /* store original address for later verification */
8166 imu->len = iov.iov_len;
8167 imu->nr_bvecs = nr_pages;
8169 ctx->nr_user_bufs++;
8177 io_sqe_buffer_unregister(ctx);
8181 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8183 __s32 __user *fds = arg;
8189 if (copy_from_user(&fd, fds, sizeof(*fds)))
8192 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8193 if (IS_ERR(ctx->cq_ev_fd)) {
8194 int ret = PTR_ERR(ctx->cq_ev_fd);
8195 ctx->cq_ev_fd = NULL;
8202 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8204 if (ctx->cq_ev_fd) {
8205 eventfd_ctx_put(ctx->cq_ev_fd);
8206 ctx->cq_ev_fd = NULL;
8213 static int __io_destroy_buffers(int id, void *p, void *data)
8215 struct io_ring_ctx *ctx = data;
8216 struct io_buffer *buf = p;
8218 __io_remove_buffers(ctx, buf, id, -1U);
8222 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8224 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8225 idr_destroy(&ctx->io_buffer_idr);
8228 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8230 io_finish_async(ctx);
8231 io_sqe_buffer_unregister(ctx);
8233 if (ctx->sqo_task) {
8234 put_task_struct(ctx->sqo_task);
8235 ctx->sqo_task = NULL;
8236 mmdrop(ctx->mm_account);
8237 ctx->mm_account = NULL;
8240 #ifdef CONFIG_BLK_CGROUP
8241 if (ctx->sqo_blkcg_css)
8242 css_put(ctx->sqo_blkcg_css);
8245 io_sqe_files_unregister(ctx);
8246 io_eventfd_unregister(ctx);
8247 io_destroy_buffers(ctx);
8248 idr_destroy(&ctx->personality_idr);
8250 #if defined(CONFIG_UNIX)
8251 if (ctx->ring_sock) {
8252 ctx->ring_sock->file = NULL; /* so that iput() is called */
8253 sock_release(ctx->ring_sock);
8257 io_mem_free(ctx->rings);
8258 io_mem_free(ctx->sq_sqes);
8260 percpu_ref_exit(&ctx->refs);
8261 free_uid(ctx->user);
8262 put_cred(ctx->creds);
8263 kfree(ctx->cancel_hash);
8264 kmem_cache_free(req_cachep, ctx->fallback_req);
8268 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8270 struct io_ring_ctx *ctx = file->private_data;
8273 poll_wait(file, &ctx->cq_wait, wait);
8275 * synchronizes with barrier from wq_has_sleeper call in
8279 if (!io_sqring_full(ctx))
8280 mask |= EPOLLOUT | EPOLLWRNORM;
8281 if (io_cqring_events(ctx, false))
8282 mask |= EPOLLIN | EPOLLRDNORM;
8287 static int io_uring_fasync(int fd, struct file *file, int on)
8289 struct io_ring_ctx *ctx = file->private_data;
8291 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8294 static int io_remove_personalities(int id, void *p, void *data)
8296 struct io_ring_ctx *ctx = data;
8297 struct io_identity *iod;
8299 iod = idr_remove(&ctx->personality_idr, id);
8301 put_cred(iod->creds);
8302 if (refcount_dec_and_test(&iod->count))
8308 static void io_ring_exit_work(struct work_struct *work)
8310 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8314 * If we're doing polled IO and end up having requests being
8315 * submitted async (out-of-line), then completions can come in while
8316 * we're waiting for refs to drop. We need to reap these manually,
8317 * as nobody else will be looking for them.
8321 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8322 io_iopoll_try_reap_events(ctx);
8323 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8324 io_ring_ctx_free(ctx);
8327 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8329 mutex_lock(&ctx->uring_lock);
8330 percpu_ref_kill(&ctx->refs);
8331 mutex_unlock(&ctx->uring_lock);
8333 io_kill_timeouts(ctx, NULL);
8334 io_poll_remove_all(ctx, NULL);
8337 io_wq_cancel_all(ctx->io_wq);
8339 /* if we failed setting up the ctx, we might not have any rings */
8341 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8342 io_iopoll_try_reap_events(ctx);
8343 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8346 * Do this upfront, so we won't have a grace period where the ring
8347 * is closed but resources aren't reaped yet. This can cause
8348 * spurious failure in setting up a new ring.
8350 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8353 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8355 * Use system_unbound_wq to avoid spawning tons of event kworkers
8356 * if we're exiting a ton of rings at the same time. It just adds
8357 * noise and overhead, there's no discernable change in runtime
8358 * over using system_wq.
8360 queue_work(system_unbound_wq, &ctx->exit_work);
8363 static int io_uring_release(struct inode *inode, struct file *file)
8365 struct io_ring_ctx *ctx = file->private_data;
8367 file->private_data = NULL;
8368 io_ring_ctx_wait_and_kill(ctx);
8372 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8374 struct files_struct *files = data;
8376 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8377 work->identity->files == files);
8381 * Returns true if 'preq' is the link parent of 'req'
8383 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8385 struct io_kiocb *link;
8387 if (!(preq->flags & REQ_F_LINK_HEAD))
8390 list_for_each_entry(link, &preq->link_list, link_list) {
8398 static bool io_match_link_files(struct io_kiocb *req,
8399 struct files_struct *files)
8401 struct io_kiocb *link;
8403 if (io_match_files(req, files))
8405 if (req->flags & REQ_F_LINK_HEAD) {
8406 list_for_each_entry(link, &req->link_list, link_list) {
8407 if (io_match_files(link, files))
8415 * We're looking to cancel 'req' because it's holding on to our files, but
8416 * 'req' could be a link to another request. See if it is, and cancel that
8417 * parent request if so.
8419 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8421 struct hlist_node *tmp;
8422 struct io_kiocb *preq;
8426 spin_lock_irq(&ctx->completion_lock);
8427 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8428 struct hlist_head *list;
8430 list = &ctx->cancel_hash[i];
8431 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8432 found = io_match_link(preq, req);
8434 io_poll_remove_one(preq);
8439 spin_unlock_irq(&ctx->completion_lock);
8443 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8444 struct io_kiocb *req)
8446 struct io_kiocb *preq;
8449 spin_lock_irq(&ctx->completion_lock);
8450 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8451 found = io_match_link(preq, req);
8453 __io_timeout_cancel(preq);
8457 spin_unlock_irq(&ctx->completion_lock);
8461 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8463 return io_match_link(container_of(work, struct io_kiocb, work), data);
8466 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8468 enum io_wq_cancel cret;
8470 /* cancel this particular work, if it's running */
8471 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8472 if (cret != IO_WQ_CANCEL_NOTFOUND)
8475 /* find links that hold this pending, cancel those */
8476 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8477 if (cret != IO_WQ_CANCEL_NOTFOUND)
8480 /* if we have a poll link holding this pending, cancel that */
8481 if (io_poll_remove_link(ctx, req))
8484 /* final option, timeout link is holding this req pending */
8485 io_timeout_remove_link(ctx, req);
8488 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8489 struct files_struct *files)
8491 struct io_defer_entry *de = NULL;
8494 spin_lock_irq(&ctx->completion_lock);
8495 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8496 if (io_match_link_files(de->req, files)) {
8497 list_cut_position(&list, &ctx->defer_list, &de->list);
8501 spin_unlock_irq(&ctx->completion_lock);
8503 while (!list_empty(&list)) {
8504 de = list_first_entry(&list, struct io_defer_entry, list);
8505 list_del_init(&de->list);
8506 req_set_fail_links(de->req);
8507 io_put_req(de->req);
8508 io_req_complete(de->req, -ECANCELED);
8514 * Returns true if we found and killed one or more files pinning requests
8516 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8517 struct files_struct *files)
8519 if (list_empty_careful(&ctx->inflight_list))
8522 io_cancel_defer_files(ctx, files);
8523 /* cancel all at once, should be faster than doing it one by one*/
8524 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8526 while (!list_empty_careful(&ctx->inflight_list)) {
8527 struct io_kiocb *cancel_req = NULL, *req;
8530 spin_lock_irq(&ctx->inflight_lock);
8531 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8532 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8533 req->work.identity->files != files)
8535 /* req is being completed, ignore */
8536 if (!refcount_inc_not_zero(&req->refs))
8542 prepare_to_wait(&ctx->inflight_wait, &wait,
8543 TASK_UNINTERRUPTIBLE);
8544 spin_unlock_irq(&ctx->inflight_lock);
8546 /* We need to keep going until we don't find a matching req */
8549 /* cancel this request, or head link requests */
8550 io_attempt_cancel(ctx, cancel_req);
8551 io_put_req(cancel_req);
8552 /* cancellations _may_ trigger task work */
8555 finish_wait(&ctx->inflight_wait, &wait);
8561 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8563 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8564 struct task_struct *task = data;
8566 return io_task_match(req, task);
8569 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8570 struct task_struct *task,
8571 struct files_struct *files)
8575 ret = io_uring_cancel_files(ctx, files);
8577 enum io_wq_cancel cret;
8579 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8580 if (cret != IO_WQ_CANCEL_NOTFOUND)
8583 /* SQPOLL thread does its own polling */
8584 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8585 while (!list_empty_careful(&ctx->iopoll_list)) {
8586 io_iopoll_try_reap_events(ctx);
8591 ret |= io_poll_remove_all(ctx, task);
8592 ret |= io_kill_timeouts(ctx, task);
8599 * We need to iteratively cancel requests, in case a request has dependent
8600 * hard links. These persist even for failure of cancelations, hence keep
8601 * looping until none are found.
8603 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8604 struct files_struct *files)
8606 struct task_struct *task = current;
8608 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8609 task = ctx->sq_data->thread;
8611 io_cqring_overflow_flush(ctx, true, task, files);
8613 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8620 * Note that this task has used io_uring. We use it for cancelation purposes.
8622 static int io_uring_add_task_file(struct file *file)
8624 struct io_uring_task *tctx = current->io_uring;
8626 if (unlikely(!tctx)) {
8629 ret = io_uring_alloc_task_context(current);
8632 tctx = current->io_uring;
8634 if (tctx->last != file) {
8635 void *old = xa_load(&tctx->xa, (unsigned long)file);
8639 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8648 * Remove this io_uring_file -> task mapping.
8650 static void io_uring_del_task_file(struct file *file)
8652 struct io_uring_task *tctx = current->io_uring;
8654 if (tctx->last == file)
8656 file = xa_erase(&tctx->xa, (unsigned long)file);
8662 * Drop task note for this file if we're the only ones that hold it after
8665 static void io_uring_attempt_task_drop(struct file *file)
8667 if (!current->io_uring)
8670 * fput() is pending, will be 2 if the only other ref is our potential
8671 * task file note. If the task is exiting, drop regardless of count.
8673 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8674 atomic_long_read(&file->f_count) == 2)
8675 io_uring_del_task_file(file);
8678 void __io_uring_files_cancel(struct files_struct *files)
8680 struct io_uring_task *tctx = current->io_uring;
8682 unsigned long index;
8684 /* make sure overflow events are dropped */
8685 tctx->in_idle = true;
8687 xa_for_each(&tctx->xa, index, file) {
8688 struct io_ring_ctx *ctx = file->private_data;
8690 io_uring_cancel_task_requests(ctx, files);
8692 io_uring_del_task_file(file);
8697 * Find any io_uring fd that this task has registered or done IO on, and cancel
8700 void __io_uring_task_cancel(void)
8702 struct io_uring_task *tctx = current->io_uring;
8706 /* make sure overflow events are dropped */
8707 tctx->in_idle = true;
8710 /* read completions before cancelations */
8711 inflight = percpu_counter_sum(&tctx->inflight);
8714 __io_uring_files_cancel(NULL);
8716 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8719 * If we've seen completions, retry. This avoids a race where
8720 * a completion comes in before we did prepare_to_wait().
8722 if (inflight != percpu_counter_sum(&tctx->inflight))
8727 finish_wait(&tctx->wait, &wait);
8728 tctx->in_idle = false;
8731 static int io_uring_flush(struct file *file, void *data)
8733 io_uring_attempt_task_drop(file);
8737 static void *io_uring_validate_mmap_request(struct file *file,
8738 loff_t pgoff, size_t sz)
8740 struct io_ring_ctx *ctx = file->private_data;
8741 loff_t offset = pgoff << PAGE_SHIFT;
8746 case IORING_OFF_SQ_RING:
8747 case IORING_OFF_CQ_RING:
8750 case IORING_OFF_SQES:
8754 return ERR_PTR(-EINVAL);
8757 page = virt_to_head_page(ptr);
8758 if (sz > page_size(page))
8759 return ERR_PTR(-EINVAL);
8766 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8768 size_t sz = vma->vm_end - vma->vm_start;
8772 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8774 return PTR_ERR(ptr);
8776 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8777 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8780 #else /* !CONFIG_MMU */
8782 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8784 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8787 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8789 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8792 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8793 unsigned long addr, unsigned long len,
8794 unsigned long pgoff, unsigned long flags)
8798 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8800 return PTR_ERR(ptr);
8802 return (unsigned long) ptr;
8805 #endif /* !CONFIG_MMU */
8807 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8812 if (!io_sqring_full(ctx))
8815 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8817 if (!io_sqring_full(ctx))
8821 } while (!signal_pending(current));
8823 finish_wait(&ctx->sqo_sq_wait, &wait);
8826 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8827 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8830 struct io_ring_ctx *ctx;
8837 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8838 IORING_ENTER_SQ_WAIT))
8846 if (f.file->f_op != &io_uring_fops)
8850 ctx = f.file->private_data;
8851 if (!percpu_ref_tryget(&ctx->refs))
8855 if (ctx->flags & IORING_SETUP_R_DISABLED)
8859 * For SQ polling, the thread will do all submissions and completions.
8860 * Just return the requested submit count, and wake the thread if
8864 if (ctx->flags & IORING_SETUP_SQPOLL) {
8865 if (!list_empty_careful(&ctx->cq_overflow_list))
8866 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8867 if (flags & IORING_ENTER_SQ_WAKEUP)
8868 wake_up(&ctx->sq_data->wait);
8869 if (flags & IORING_ENTER_SQ_WAIT)
8870 io_sqpoll_wait_sq(ctx);
8871 submitted = to_submit;
8872 } else if (to_submit) {
8873 ret = io_uring_add_task_file(f.file);
8876 mutex_lock(&ctx->uring_lock);
8877 submitted = io_submit_sqes(ctx, to_submit);
8878 mutex_unlock(&ctx->uring_lock);
8880 if (submitted != to_submit)
8883 if (flags & IORING_ENTER_GETEVENTS) {
8884 min_complete = min(min_complete, ctx->cq_entries);
8887 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8888 * space applications don't need to do io completion events
8889 * polling again, they can rely on io_sq_thread to do polling
8890 * work, which can reduce cpu usage and uring_lock contention.
8892 if (ctx->flags & IORING_SETUP_IOPOLL &&
8893 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8894 ret = io_iopoll_check(ctx, min_complete);
8896 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8901 percpu_ref_put(&ctx->refs);
8904 return submitted ? submitted : ret;
8907 #ifdef CONFIG_PROC_FS
8908 static int io_uring_show_cred(int id, void *p, void *data)
8910 const struct cred *cred = p;
8911 struct seq_file *m = data;
8912 struct user_namespace *uns = seq_user_ns(m);
8913 struct group_info *gi;
8918 seq_printf(m, "%5d\n", id);
8919 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8920 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8921 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8922 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8923 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8924 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8925 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8926 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8927 seq_puts(m, "\n\tGroups:\t");
8928 gi = cred->group_info;
8929 for (g = 0; g < gi->ngroups; g++) {
8930 seq_put_decimal_ull(m, g ? " " : "",
8931 from_kgid_munged(uns, gi->gid[g]));
8933 seq_puts(m, "\n\tCapEff:\t");
8934 cap = cred->cap_effective;
8935 CAP_FOR_EACH_U32(__capi)
8936 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8941 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8943 struct io_sq_data *sq = NULL;
8948 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8949 * since fdinfo case grabs it in the opposite direction of normal use
8950 * cases. If we fail to get the lock, we just don't iterate any
8951 * structures that could be going away outside the io_uring mutex.
8953 has_lock = mutex_trylock(&ctx->uring_lock);
8955 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8958 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8959 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8960 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8961 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8962 struct fixed_file_table *table;
8965 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8966 f = table->files[i & IORING_FILE_TABLE_MASK];
8968 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8970 seq_printf(m, "%5u: <none>\n", i);
8972 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8973 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8974 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8976 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8977 (unsigned int) buf->len);
8979 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8980 seq_printf(m, "Personalities:\n");
8981 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8983 seq_printf(m, "PollList:\n");
8984 spin_lock_irq(&ctx->completion_lock);
8985 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8986 struct hlist_head *list = &ctx->cancel_hash[i];
8987 struct io_kiocb *req;
8989 hlist_for_each_entry(req, list, hash_node)
8990 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8991 req->task->task_works != NULL);
8993 spin_unlock_irq(&ctx->completion_lock);
8995 mutex_unlock(&ctx->uring_lock);
8998 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9000 struct io_ring_ctx *ctx = f->private_data;
9002 if (percpu_ref_tryget(&ctx->refs)) {
9003 __io_uring_show_fdinfo(ctx, m);
9004 percpu_ref_put(&ctx->refs);
9009 static const struct file_operations io_uring_fops = {
9010 .release = io_uring_release,
9011 .flush = io_uring_flush,
9012 .mmap = io_uring_mmap,
9014 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9015 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9017 .poll = io_uring_poll,
9018 .fasync = io_uring_fasync,
9019 #ifdef CONFIG_PROC_FS
9020 .show_fdinfo = io_uring_show_fdinfo,
9024 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9025 struct io_uring_params *p)
9027 struct io_rings *rings;
9028 size_t size, sq_array_offset;
9030 /* make sure these are sane, as we already accounted them */
9031 ctx->sq_entries = p->sq_entries;
9032 ctx->cq_entries = p->cq_entries;
9034 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9035 if (size == SIZE_MAX)
9038 rings = io_mem_alloc(size);
9043 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9044 rings->sq_ring_mask = p->sq_entries - 1;
9045 rings->cq_ring_mask = p->cq_entries - 1;
9046 rings->sq_ring_entries = p->sq_entries;
9047 rings->cq_ring_entries = p->cq_entries;
9048 ctx->sq_mask = rings->sq_ring_mask;
9049 ctx->cq_mask = rings->cq_ring_mask;
9051 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9052 if (size == SIZE_MAX) {
9053 io_mem_free(ctx->rings);
9058 ctx->sq_sqes = io_mem_alloc(size);
9059 if (!ctx->sq_sqes) {
9060 io_mem_free(ctx->rings);
9069 * Allocate an anonymous fd, this is what constitutes the application
9070 * visible backing of an io_uring instance. The application mmaps this
9071 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9072 * we have to tie this fd to a socket for file garbage collection purposes.
9074 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9079 #if defined(CONFIG_UNIX)
9080 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9086 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9090 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9091 O_RDWR | O_CLOEXEC);
9095 ret = PTR_ERR(file);
9099 #if defined(CONFIG_UNIX)
9100 ctx->ring_sock->file = file;
9102 if (unlikely(io_uring_add_task_file(file))) {
9103 file = ERR_PTR(-ENOMEM);
9106 fd_install(ret, file);
9109 #if defined(CONFIG_UNIX)
9110 sock_release(ctx->ring_sock);
9111 ctx->ring_sock = NULL;
9116 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9117 struct io_uring_params __user *params)
9119 struct user_struct *user = NULL;
9120 struct io_ring_ctx *ctx;
9126 if (entries > IORING_MAX_ENTRIES) {
9127 if (!(p->flags & IORING_SETUP_CLAMP))
9129 entries = IORING_MAX_ENTRIES;
9133 * Use twice as many entries for the CQ ring. It's possible for the
9134 * application to drive a higher depth than the size of the SQ ring,
9135 * since the sqes are only used at submission time. This allows for
9136 * some flexibility in overcommitting a bit. If the application has
9137 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9138 * of CQ ring entries manually.
9140 p->sq_entries = roundup_pow_of_two(entries);
9141 if (p->flags & IORING_SETUP_CQSIZE) {
9143 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9144 * to a power-of-two, if it isn't already. We do NOT impose
9145 * any cq vs sq ring sizing.
9147 if (p->cq_entries < p->sq_entries)
9149 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9150 if (!(p->flags & IORING_SETUP_CLAMP))
9152 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9154 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9156 p->cq_entries = 2 * p->sq_entries;
9159 user = get_uid(current_user());
9160 limit_mem = !capable(CAP_IPC_LOCK);
9163 ret = __io_account_mem(user,
9164 ring_pages(p->sq_entries, p->cq_entries));
9171 ctx = io_ring_ctx_alloc(p);
9174 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9179 ctx->compat = in_compat_syscall();
9181 ctx->creds = get_current_cred();
9183 ctx->loginuid = current->loginuid;
9184 ctx->sessionid = current->sessionid;
9186 ctx->sqo_task = get_task_struct(current);
9189 * This is just grabbed for accounting purposes. When a process exits,
9190 * the mm is exited and dropped before the files, hence we need to hang
9191 * on to this mm purely for the purposes of being able to unaccount
9192 * memory (locked/pinned vm). It's not used for anything else.
9194 mmgrab(current->mm);
9195 ctx->mm_account = current->mm;
9197 #ifdef CONFIG_BLK_CGROUP
9199 * The sq thread will belong to the original cgroup it was inited in.
9200 * If the cgroup goes offline (e.g. disabling the io controller), then
9201 * issued bios will be associated with the closest cgroup later in the
9205 ctx->sqo_blkcg_css = blkcg_css();
9206 ret = css_tryget_online(ctx->sqo_blkcg_css);
9209 /* don't init against a dying cgroup, have the user try again */
9210 ctx->sqo_blkcg_css = NULL;
9217 * Account memory _before_ installing the file descriptor. Once
9218 * the descriptor is installed, it can get closed at any time. Also
9219 * do this before hitting the general error path, as ring freeing
9220 * will un-account as well.
9222 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9224 ctx->limit_mem = limit_mem;
9226 ret = io_allocate_scq_urings(ctx, p);
9230 ret = io_sq_offload_create(ctx, p);
9234 if (!(p->flags & IORING_SETUP_R_DISABLED))
9235 io_sq_offload_start(ctx);
9237 memset(&p->sq_off, 0, sizeof(p->sq_off));
9238 p->sq_off.head = offsetof(struct io_rings, sq.head);
9239 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9240 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9241 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9242 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9243 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9244 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9246 memset(&p->cq_off, 0, sizeof(p->cq_off));
9247 p->cq_off.head = offsetof(struct io_rings, cq.head);
9248 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9249 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9250 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9251 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9252 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9253 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9255 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9256 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9257 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9258 IORING_FEAT_POLL_32BITS;
9260 if (copy_to_user(params, p, sizeof(*p))) {
9266 * Install ring fd as the very last thing, so we don't risk someone
9267 * having closed it before we finish setup
9269 ret = io_uring_get_fd(ctx);
9273 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9276 io_ring_ctx_wait_and_kill(ctx);
9281 * Sets up an aio uring context, and returns the fd. Applications asks for a
9282 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9283 * params structure passed in.
9285 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9287 struct io_uring_params p;
9290 if (copy_from_user(&p, params, sizeof(p)))
9292 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9297 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9298 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9299 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9300 IORING_SETUP_R_DISABLED))
9303 return io_uring_create(entries, &p, params);
9306 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9307 struct io_uring_params __user *, params)
9309 return io_uring_setup(entries, params);
9312 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9314 struct io_uring_probe *p;
9318 size = struct_size(p, ops, nr_args);
9319 if (size == SIZE_MAX)
9321 p = kzalloc(size, GFP_KERNEL);
9326 if (copy_from_user(p, arg, size))
9329 if (memchr_inv(p, 0, size))
9332 p->last_op = IORING_OP_LAST - 1;
9333 if (nr_args > IORING_OP_LAST)
9334 nr_args = IORING_OP_LAST;
9336 for (i = 0; i < nr_args; i++) {
9338 if (!io_op_defs[i].not_supported)
9339 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9344 if (copy_to_user(arg, p, size))
9351 static int io_register_personality(struct io_ring_ctx *ctx)
9353 struct io_identity *id;
9356 id = kmalloc(sizeof(*id), GFP_KERNEL);
9360 io_init_identity(id);
9361 id->creds = get_current_cred();
9363 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9365 put_cred(id->creds);
9371 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9373 struct io_identity *iod;
9375 iod = idr_remove(&ctx->personality_idr, id);
9377 put_cred(iod->creds);
9378 if (refcount_dec_and_test(&iod->count))
9386 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9387 unsigned int nr_args)
9389 struct io_uring_restriction *res;
9393 /* Restrictions allowed only if rings started disabled */
9394 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9397 /* We allow only a single restrictions registration */
9398 if (ctx->restrictions.registered)
9401 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9404 size = array_size(nr_args, sizeof(*res));
9405 if (size == SIZE_MAX)
9408 res = memdup_user(arg, size);
9410 return PTR_ERR(res);
9414 for (i = 0; i < nr_args; i++) {
9415 switch (res[i].opcode) {
9416 case IORING_RESTRICTION_REGISTER_OP:
9417 if (res[i].register_op >= IORING_REGISTER_LAST) {
9422 __set_bit(res[i].register_op,
9423 ctx->restrictions.register_op);
9425 case IORING_RESTRICTION_SQE_OP:
9426 if (res[i].sqe_op >= IORING_OP_LAST) {
9431 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9433 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9434 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9436 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9437 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9446 /* Reset all restrictions if an error happened */
9448 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9450 ctx->restrictions.registered = true;
9456 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9458 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9461 if (ctx->restrictions.registered)
9462 ctx->restricted = 1;
9464 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9466 io_sq_offload_start(ctx);
9471 static bool io_register_op_must_quiesce(int op)
9474 case IORING_UNREGISTER_FILES:
9475 case IORING_REGISTER_FILES_UPDATE:
9476 case IORING_REGISTER_PROBE:
9477 case IORING_REGISTER_PERSONALITY:
9478 case IORING_UNREGISTER_PERSONALITY:
9485 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9486 void __user *arg, unsigned nr_args)
9487 __releases(ctx->uring_lock)
9488 __acquires(ctx->uring_lock)
9493 * We're inside the ring mutex, if the ref is already dying, then
9494 * someone else killed the ctx or is already going through
9495 * io_uring_register().
9497 if (percpu_ref_is_dying(&ctx->refs))
9500 if (io_register_op_must_quiesce(opcode)) {
9501 percpu_ref_kill(&ctx->refs);
9504 * Drop uring mutex before waiting for references to exit. If
9505 * another thread is currently inside io_uring_enter() it might
9506 * need to grab the uring_lock to make progress. If we hold it
9507 * here across the drain wait, then we can deadlock. It's safe
9508 * to drop the mutex here, since no new references will come in
9509 * after we've killed the percpu ref.
9511 mutex_unlock(&ctx->uring_lock);
9513 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9516 ret = io_run_task_work_sig();
9521 mutex_lock(&ctx->uring_lock);
9524 percpu_ref_resurrect(&ctx->refs);
9529 if (ctx->restricted) {
9530 if (opcode >= IORING_REGISTER_LAST) {
9535 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9542 case IORING_REGISTER_BUFFERS:
9543 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9545 case IORING_UNREGISTER_BUFFERS:
9549 ret = io_sqe_buffer_unregister(ctx);
9551 case IORING_REGISTER_FILES:
9552 ret = io_sqe_files_register(ctx, arg, nr_args);
9554 case IORING_UNREGISTER_FILES:
9558 ret = io_sqe_files_unregister(ctx);
9560 case IORING_REGISTER_FILES_UPDATE:
9561 ret = io_sqe_files_update(ctx, arg, nr_args);
9563 case IORING_REGISTER_EVENTFD:
9564 case IORING_REGISTER_EVENTFD_ASYNC:
9568 ret = io_eventfd_register(ctx, arg);
9571 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9572 ctx->eventfd_async = 1;
9574 ctx->eventfd_async = 0;
9576 case IORING_UNREGISTER_EVENTFD:
9580 ret = io_eventfd_unregister(ctx);
9582 case IORING_REGISTER_PROBE:
9584 if (!arg || nr_args > 256)
9586 ret = io_probe(ctx, arg, nr_args);
9588 case IORING_REGISTER_PERSONALITY:
9592 ret = io_register_personality(ctx);
9594 case IORING_UNREGISTER_PERSONALITY:
9598 ret = io_unregister_personality(ctx, nr_args);
9600 case IORING_REGISTER_ENABLE_RINGS:
9604 ret = io_register_enable_rings(ctx);
9606 case IORING_REGISTER_RESTRICTIONS:
9607 ret = io_register_restrictions(ctx, arg, nr_args);
9615 if (io_register_op_must_quiesce(opcode)) {
9616 /* bring the ctx back to life */
9617 percpu_ref_reinit(&ctx->refs);
9619 reinit_completion(&ctx->ref_comp);
9624 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9625 void __user *, arg, unsigned int, nr_args)
9627 struct io_ring_ctx *ctx;
9636 if (f.file->f_op != &io_uring_fops)
9639 ctx = f.file->private_data;
9641 mutex_lock(&ctx->uring_lock);
9642 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9643 mutex_unlock(&ctx->uring_lock);
9644 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9645 ctx->cq_ev_fd != NULL, ret);
9651 static int __init io_uring_init(void)
9653 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9654 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9655 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9658 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9659 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9660 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9661 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9662 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9663 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9664 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9665 BUILD_BUG_SQE_ELEM(8, __u64, off);
9666 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9667 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9668 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9669 BUILD_BUG_SQE_ELEM(24, __u32, len);
9670 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9671 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9672 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9673 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9674 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9675 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9676 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9677 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9678 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9679 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9680 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9681 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9682 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9683 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9684 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9685 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9686 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9687 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9688 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9690 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9691 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9692 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9695 __initcall(io_uring_init);