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;
1970 enum task_work_notify_mode notify;
1973 if (tsk->flags & PF_EXITING)
1977 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1978 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1979 * processing task_work. There's no reliable way to tell if TWA_RESUME
1983 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1984 notify = TWA_SIGNAL;
1986 ret = task_work_add(tsk, &req->task_work, notify);
1988 wake_up_process(tsk);
1993 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1995 struct io_ring_ctx *ctx = req->ctx;
1997 spin_lock_irq(&ctx->completion_lock);
1998 io_cqring_fill_event(req, error);
1999 io_commit_cqring(ctx);
2000 spin_unlock_irq(&ctx->completion_lock);
2002 io_cqring_ev_posted(ctx);
2003 req_set_fail_links(req);
2004 io_double_put_req(req);
2007 static void io_req_task_cancel(struct callback_head *cb)
2009 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2010 struct io_ring_ctx *ctx = req->ctx;
2012 __io_req_task_cancel(req, -ECANCELED);
2013 percpu_ref_put(&ctx->refs);
2016 static void __io_req_task_submit(struct io_kiocb *req)
2018 struct io_ring_ctx *ctx = req->ctx;
2020 if (!__io_sq_thread_acquire_mm(ctx)) {
2021 mutex_lock(&ctx->uring_lock);
2022 __io_queue_sqe(req, NULL);
2023 mutex_unlock(&ctx->uring_lock);
2025 __io_req_task_cancel(req, -EFAULT);
2029 static void io_req_task_submit(struct callback_head *cb)
2031 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2032 struct io_ring_ctx *ctx = req->ctx;
2034 __io_req_task_submit(req);
2035 percpu_ref_put(&ctx->refs);
2038 static void io_req_task_queue(struct io_kiocb *req)
2042 init_task_work(&req->task_work, io_req_task_submit);
2043 percpu_ref_get(&req->ctx->refs);
2045 ret = io_req_task_work_add(req, true);
2046 if (unlikely(ret)) {
2047 struct task_struct *tsk;
2049 init_task_work(&req->task_work, io_req_task_cancel);
2050 tsk = io_wq_get_task(req->ctx->io_wq);
2051 task_work_add(tsk, &req->task_work, TWA_NONE);
2052 wake_up_process(tsk);
2056 static void io_queue_next(struct io_kiocb *req)
2058 struct io_kiocb *nxt = io_req_find_next(req);
2061 io_req_task_queue(nxt);
2064 static void io_free_req(struct io_kiocb *req)
2071 void *reqs[IO_IOPOLL_BATCH];
2074 struct task_struct *task;
2078 static inline void io_init_req_batch(struct req_batch *rb)
2085 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2086 struct req_batch *rb)
2088 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2089 percpu_ref_put_many(&ctx->refs, rb->to_free);
2093 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2094 struct req_batch *rb)
2097 __io_req_free_batch_flush(ctx, rb);
2099 struct io_uring_task *tctx = rb->task->io_uring;
2101 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2102 put_task_struct_many(rb->task, rb->task_refs);
2107 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2109 if (unlikely(io_is_fallback_req(req))) {
2113 if (req->flags & REQ_F_LINK_HEAD)
2116 if (req->task != rb->task) {
2118 struct io_uring_task *tctx = rb->task->io_uring;
2120 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2121 put_task_struct_many(rb->task, rb->task_refs);
2123 rb->task = req->task;
2128 io_dismantle_req(req);
2129 rb->reqs[rb->to_free++] = req;
2130 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2131 __io_req_free_batch_flush(req->ctx, rb);
2135 * Drop reference to request, return next in chain (if there is one) if this
2136 * was the last reference to this request.
2138 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2140 struct io_kiocb *nxt = NULL;
2142 if (refcount_dec_and_test(&req->refs)) {
2143 nxt = io_req_find_next(req);
2149 static void io_put_req(struct io_kiocb *req)
2151 if (refcount_dec_and_test(&req->refs))
2155 static void io_put_req_deferred_cb(struct callback_head *cb)
2157 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2162 static void io_free_req_deferred(struct io_kiocb *req)
2166 init_task_work(&req->task_work, io_put_req_deferred_cb);
2167 ret = io_req_task_work_add(req, true);
2168 if (unlikely(ret)) {
2169 struct task_struct *tsk;
2171 tsk = io_wq_get_task(req->ctx->io_wq);
2172 task_work_add(tsk, &req->task_work, TWA_NONE);
2173 wake_up_process(tsk);
2177 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2179 if (refcount_sub_and_test(refs, &req->refs))
2180 io_free_req_deferred(req);
2183 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2185 struct io_kiocb *nxt;
2188 * A ref is owned by io-wq in which context we're. So, if that's the
2189 * last one, it's safe to steal next work. False negatives are Ok,
2190 * it just will be re-punted async in io_put_work()
2192 if (refcount_read(&req->refs) != 1)
2195 nxt = io_req_find_next(req);
2196 return nxt ? &nxt->work : NULL;
2199 static void io_double_put_req(struct io_kiocb *req)
2201 /* drop both submit and complete references */
2202 if (refcount_sub_and_test(2, &req->refs))
2206 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2208 struct io_rings *rings = ctx->rings;
2210 if (test_bit(0, &ctx->cq_check_overflow)) {
2212 * noflush == true is from the waitqueue handler, just ensure
2213 * we wake up the task, and the next invocation will flush the
2214 * entries. We cannot safely to it from here.
2216 if (noflush && !list_empty(&ctx->cq_overflow_list))
2219 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2222 /* See comment at the top of this file */
2224 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2227 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2229 struct io_rings *rings = ctx->rings;
2231 /* make sure SQ entry isn't read before tail */
2232 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2235 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2237 unsigned int cflags;
2239 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2240 cflags |= IORING_CQE_F_BUFFER;
2241 req->flags &= ~REQ_F_BUFFER_SELECTED;
2246 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2248 struct io_buffer *kbuf;
2250 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2251 return io_put_kbuf(req, kbuf);
2254 static inline bool io_run_task_work(void)
2257 * Not safe to run on exiting task, and the task_work handling will
2258 * not add work to such a task.
2260 if (unlikely(current->flags & PF_EXITING))
2262 if (current->task_works) {
2263 __set_current_state(TASK_RUNNING);
2271 static void io_iopoll_queue(struct list_head *again)
2273 struct io_kiocb *req;
2276 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2277 list_del(&req->inflight_entry);
2278 __io_complete_rw(req, -EAGAIN, 0, NULL);
2279 } while (!list_empty(again));
2283 * Find and free completed poll iocbs
2285 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2286 struct list_head *done)
2288 struct req_batch rb;
2289 struct io_kiocb *req;
2292 /* order with ->result store in io_complete_rw_iopoll() */
2295 io_init_req_batch(&rb);
2296 while (!list_empty(done)) {
2299 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2300 if (READ_ONCE(req->result) == -EAGAIN) {
2302 req->iopoll_completed = 0;
2303 list_move_tail(&req->inflight_entry, &again);
2306 list_del(&req->inflight_entry);
2308 if (req->flags & REQ_F_BUFFER_SELECTED)
2309 cflags = io_put_rw_kbuf(req);
2311 __io_cqring_fill_event(req, req->result, cflags);
2314 if (refcount_dec_and_test(&req->refs))
2315 io_req_free_batch(&rb, req);
2318 io_commit_cqring(ctx);
2319 if (ctx->flags & IORING_SETUP_SQPOLL)
2320 io_cqring_ev_posted(ctx);
2321 io_req_free_batch_finish(ctx, &rb);
2323 if (!list_empty(&again))
2324 io_iopoll_queue(&again);
2327 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2330 struct io_kiocb *req, *tmp;
2336 * Only spin for completions if we don't have multiple devices hanging
2337 * off our complete list, and we're under the requested amount.
2339 spin = !ctx->poll_multi_file && *nr_events < min;
2342 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2343 struct kiocb *kiocb = &req->rw.kiocb;
2346 * Move completed and retryable entries to our local lists.
2347 * If we find a request that requires polling, break out
2348 * and complete those lists first, if we have entries there.
2350 if (READ_ONCE(req->iopoll_completed)) {
2351 list_move_tail(&req->inflight_entry, &done);
2354 if (!list_empty(&done))
2357 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2361 /* iopoll may have completed current req */
2362 if (READ_ONCE(req->iopoll_completed))
2363 list_move_tail(&req->inflight_entry, &done);
2370 if (!list_empty(&done))
2371 io_iopoll_complete(ctx, nr_events, &done);
2377 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2378 * non-spinning poll check - we'll still enter the driver poll loop, but only
2379 * as a non-spinning completion check.
2381 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2384 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2387 ret = io_do_iopoll(ctx, nr_events, min);
2390 if (*nr_events >= min)
2398 * We can't just wait for polled events to come to us, we have to actively
2399 * find and complete them.
2401 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2403 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2406 mutex_lock(&ctx->uring_lock);
2407 while (!list_empty(&ctx->iopoll_list)) {
2408 unsigned int nr_events = 0;
2410 io_do_iopoll(ctx, &nr_events, 0);
2412 /* let it sleep and repeat later if can't complete a request */
2416 * Ensure we allow local-to-the-cpu processing to take place,
2417 * in this case we need to ensure that we reap all events.
2418 * Also let task_work, etc. to progress by releasing the mutex
2420 if (need_resched()) {
2421 mutex_unlock(&ctx->uring_lock);
2423 mutex_lock(&ctx->uring_lock);
2426 mutex_unlock(&ctx->uring_lock);
2429 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2431 unsigned int nr_events = 0;
2432 int iters = 0, ret = 0;
2435 * We disallow the app entering submit/complete with polling, but we
2436 * still need to lock the ring to prevent racing with polled issue
2437 * that got punted to a workqueue.
2439 mutex_lock(&ctx->uring_lock);
2442 * Don't enter poll loop if we already have events pending.
2443 * If we do, we can potentially be spinning for commands that
2444 * already triggered a CQE (eg in error).
2446 if (io_cqring_events(ctx, false))
2450 * If a submit got punted to a workqueue, we can have the
2451 * application entering polling for a command before it gets
2452 * issued. That app will hold the uring_lock for the duration
2453 * of the poll right here, so we need to take a breather every
2454 * now and then to ensure that the issue has a chance to add
2455 * the poll to the issued list. Otherwise we can spin here
2456 * forever, while the workqueue is stuck trying to acquire the
2459 if (!(++iters & 7)) {
2460 mutex_unlock(&ctx->uring_lock);
2462 mutex_lock(&ctx->uring_lock);
2465 ret = io_iopoll_getevents(ctx, &nr_events, min);
2469 } while (min && !nr_events && !need_resched());
2471 mutex_unlock(&ctx->uring_lock);
2475 static void kiocb_end_write(struct io_kiocb *req)
2478 * Tell lockdep we inherited freeze protection from submission
2481 if (req->flags & REQ_F_ISREG) {
2482 struct inode *inode = file_inode(req->file);
2484 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2486 file_end_write(req->file);
2489 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2490 struct io_comp_state *cs)
2492 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2495 if (kiocb->ki_flags & IOCB_WRITE)
2496 kiocb_end_write(req);
2498 if (res != req->result)
2499 req_set_fail_links(req);
2500 if (req->flags & REQ_F_BUFFER_SELECTED)
2501 cflags = io_put_rw_kbuf(req);
2502 __io_req_complete(req, res, cflags, cs);
2506 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2508 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2509 ssize_t ret = -ECANCELED;
2510 struct iov_iter iter;
2518 switch (req->opcode) {
2519 case IORING_OP_READV:
2520 case IORING_OP_READ_FIXED:
2521 case IORING_OP_READ:
2524 case IORING_OP_WRITEV:
2525 case IORING_OP_WRITE_FIXED:
2526 case IORING_OP_WRITE:
2530 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2535 if (!req->async_data) {
2536 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2539 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2547 req_set_fail_links(req);
2548 io_req_complete(req, ret);
2553 static bool io_rw_reissue(struct io_kiocb *req, long res)
2556 umode_t mode = file_inode(req->file)->i_mode;
2559 if (!S_ISBLK(mode) && !S_ISREG(mode))
2561 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2564 ret = io_sq_thread_acquire_mm(req->ctx, req);
2566 if (io_resubmit_prep(req, ret)) {
2567 refcount_inc(&req->refs);
2568 io_queue_async_work(req);
2576 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2577 struct io_comp_state *cs)
2579 if (!io_rw_reissue(req, res))
2580 io_complete_rw_common(&req->rw.kiocb, res, cs);
2583 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2585 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2587 __io_complete_rw(req, res, res2, NULL);
2590 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2592 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2594 if (kiocb->ki_flags & IOCB_WRITE)
2595 kiocb_end_write(req);
2597 if (res != -EAGAIN && res != req->result)
2598 req_set_fail_links(req);
2600 WRITE_ONCE(req->result, res);
2601 /* order with io_poll_complete() checking ->result */
2603 WRITE_ONCE(req->iopoll_completed, 1);
2607 * After the iocb has been issued, it's safe to be found on the poll list.
2608 * Adding the kiocb to the list AFTER submission ensures that we don't
2609 * find it from a io_iopoll_getevents() thread before the issuer is done
2610 * accessing the kiocb cookie.
2612 static void io_iopoll_req_issued(struct io_kiocb *req)
2614 struct io_ring_ctx *ctx = req->ctx;
2617 * Track whether we have multiple files in our lists. This will impact
2618 * how we do polling eventually, not spinning if we're on potentially
2619 * different devices.
2621 if (list_empty(&ctx->iopoll_list)) {
2622 ctx->poll_multi_file = false;
2623 } else if (!ctx->poll_multi_file) {
2624 struct io_kiocb *list_req;
2626 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2628 if (list_req->file != req->file)
2629 ctx->poll_multi_file = true;
2633 * For fast devices, IO may have already completed. If it has, add
2634 * it to the front so we find it first.
2636 if (READ_ONCE(req->iopoll_completed))
2637 list_add(&req->inflight_entry, &ctx->iopoll_list);
2639 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2641 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2642 wq_has_sleeper(&ctx->sq_data->wait))
2643 wake_up(&ctx->sq_data->wait);
2646 static void __io_state_file_put(struct io_submit_state *state)
2648 if (state->has_refs)
2649 fput_many(state->file, state->has_refs);
2653 static inline void io_state_file_put(struct io_submit_state *state)
2656 __io_state_file_put(state);
2660 * Get as many references to a file as we have IOs left in this submission,
2661 * assuming most submissions are for one file, or at least that each file
2662 * has more than one submission.
2664 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2670 if (state->fd == fd) {
2674 __io_state_file_put(state);
2676 state->file = fget_many(fd, state->ios_left);
2681 state->has_refs = state->ios_left - 1;
2685 static bool io_bdev_nowait(struct block_device *bdev)
2688 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2695 * If we tracked the file through the SCM inflight mechanism, we could support
2696 * any file. For now, just ensure that anything potentially problematic is done
2699 static bool io_file_supports_async(struct file *file, int rw)
2701 umode_t mode = file_inode(file)->i_mode;
2703 if (S_ISBLK(mode)) {
2704 if (io_bdev_nowait(file->f_inode->i_bdev))
2708 if (S_ISCHR(mode) || S_ISSOCK(mode))
2710 if (S_ISREG(mode)) {
2711 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2712 file->f_op != &io_uring_fops)
2717 /* any ->read/write should understand O_NONBLOCK */
2718 if (file->f_flags & O_NONBLOCK)
2721 if (!(file->f_mode & FMODE_NOWAIT))
2725 return file->f_op->read_iter != NULL;
2727 return file->f_op->write_iter != NULL;
2730 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2732 struct io_ring_ctx *ctx = req->ctx;
2733 struct kiocb *kiocb = &req->rw.kiocb;
2737 if (S_ISREG(file_inode(req->file)->i_mode))
2738 req->flags |= REQ_F_ISREG;
2740 kiocb->ki_pos = READ_ONCE(sqe->off);
2741 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2742 req->flags |= REQ_F_CUR_POS;
2743 kiocb->ki_pos = req->file->f_pos;
2745 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2746 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2747 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2751 ioprio = READ_ONCE(sqe->ioprio);
2753 ret = ioprio_check_cap(ioprio);
2757 kiocb->ki_ioprio = ioprio;
2759 kiocb->ki_ioprio = get_current_ioprio();
2761 /* don't allow async punt if RWF_NOWAIT was requested */
2762 if (kiocb->ki_flags & IOCB_NOWAIT)
2763 req->flags |= REQ_F_NOWAIT;
2765 if (ctx->flags & IORING_SETUP_IOPOLL) {
2766 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2767 !kiocb->ki_filp->f_op->iopoll)
2770 kiocb->ki_flags |= IOCB_HIPRI;
2771 kiocb->ki_complete = io_complete_rw_iopoll;
2772 req->iopoll_completed = 0;
2774 if (kiocb->ki_flags & IOCB_HIPRI)
2776 kiocb->ki_complete = io_complete_rw;
2779 req->rw.addr = READ_ONCE(sqe->addr);
2780 req->rw.len = READ_ONCE(sqe->len);
2781 req->buf_index = READ_ONCE(sqe->buf_index);
2785 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2791 case -ERESTARTNOINTR:
2792 case -ERESTARTNOHAND:
2793 case -ERESTART_RESTARTBLOCK:
2795 * We can't just restart the syscall, since previously
2796 * submitted sqes may already be in progress. Just fail this
2802 kiocb->ki_complete(kiocb, ret, 0);
2806 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2807 struct io_comp_state *cs)
2809 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2810 struct io_async_rw *io = req->async_data;
2812 /* add previously done IO, if any */
2813 if (io && io->bytes_done > 0) {
2815 ret = io->bytes_done;
2817 ret += io->bytes_done;
2820 if (req->flags & REQ_F_CUR_POS)
2821 req->file->f_pos = kiocb->ki_pos;
2822 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2823 __io_complete_rw(req, ret, 0, cs);
2825 io_rw_done(kiocb, ret);
2828 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2829 struct iov_iter *iter)
2831 struct io_ring_ctx *ctx = req->ctx;
2832 size_t len = req->rw.len;
2833 struct io_mapped_ubuf *imu;
2834 u16 index, buf_index = req->buf_index;
2838 if (unlikely(buf_index >= ctx->nr_user_bufs))
2840 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2841 imu = &ctx->user_bufs[index];
2842 buf_addr = req->rw.addr;
2845 if (buf_addr + len < buf_addr)
2847 /* not inside the mapped region */
2848 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2852 * May not be a start of buffer, set size appropriately
2853 * and advance us to the beginning.
2855 offset = buf_addr - imu->ubuf;
2856 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2860 * Don't use iov_iter_advance() here, as it's really slow for
2861 * using the latter parts of a big fixed buffer - it iterates
2862 * over each segment manually. We can cheat a bit here, because
2865 * 1) it's a BVEC iter, we set it up
2866 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2867 * first and last bvec
2869 * So just find our index, and adjust the iterator afterwards.
2870 * If the offset is within the first bvec (or the whole first
2871 * bvec, just use iov_iter_advance(). This makes it easier
2872 * since we can just skip the first segment, which may not
2873 * be PAGE_SIZE aligned.
2875 const struct bio_vec *bvec = imu->bvec;
2877 if (offset <= bvec->bv_len) {
2878 iov_iter_advance(iter, offset);
2880 unsigned long seg_skip;
2882 /* skip first vec */
2883 offset -= bvec->bv_len;
2884 seg_skip = 1 + (offset >> PAGE_SHIFT);
2886 iter->bvec = bvec + seg_skip;
2887 iter->nr_segs -= seg_skip;
2888 iter->count -= bvec->bv_len + offset;
2889 iter->iov_offset = offset & ~PAGE_MASK;
2896 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2899 mutex_unlock(&ctx->uring_lock);
2902 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2905 * "Normal" inline submissions always hold the uring_lock, since we
2906 * grab it from the system call. Same is true for the SQPOLL offload.
2907 * The only exception is when we've detached the request and issue it
2908 * from an async worker thread, grab the lock for that case.
2911 mutex_lock(&ctx->uring_lock);
2914 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2915 int bgid, struct io_buffer *kbuf,
2918 struct io_buffer *head;
2920 if (req->flags & REQ_F_BUFFER_SELECTED)
2923 io_ring_submit_lock(req->ctx, needs_lock);
2925 lockdep_assert_held(&req->ctx->uring_lock);
2927 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2929 if (!list_empty(&head->list)) {
2930 kbuf = list_last_entry(&head->list, struct io_buffer,
2932 list_del(&kbuf->list);
2935 idr_remove(&req->ctx->io_buffer_idr, bgid);
2937 if (*len > kbuf->len)
2940 kbuf = ERR_PTR(-ENOBUFS);
2943 io_ring_submit_unlock(req->ctx, needs_lock);
2948 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2951 struct io_buffer *kbuf;
2954 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2955 bgid = req->buf_index;
2956 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2959 req->rw.addr = (u64) (unsigned long) kbuf;
2960 req->flags |= REQ_F_BUFFER_SELECTED;
2961 return u64_to_user_ptr(kbuf->addr);
2964 #ifdef CONFIG_COMPAT
2965 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2968 struct compat_iovec __user *uiov;
2969 compat_ssize_t clen;
2973 uiov = u64_to_user_ptr(req->rw.addr);
2974 if (!access_ok(uiov, sizeof(*uiov)))
2976 if (__get_user(clen, &uiov->iov_len))
2982 buf = io_rw_buffer_select(req, &len, needs_lock);
2984 return PTR_ERR(buf);
2985 iov[0].iov_base = buf;
2986 iov[0].iov_len = (compat_size_t) len;
2991 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2994 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2998 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3001 len = iov[0].iov_len;
3004 buf = io_rw_buffer_select(req, &len, needs_lock);
3006 return PTR_ERR(buf);
3007 iov[0].iov_base = buf;
3008 iov[0].iov_len = len;
3012 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3015 if (req->flags & REQ_F_BUFFER_SELECTED) {
3016 struct io_buffer *kbuf;
3018 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3019 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3020 iov[0].iov_len = kbuf->len;
3025 else if (req->rw.len > 1)
3028 #ifdef CONFIG_COMPAT
3029 if (req->ctx->compat)
3030 return io_compat_import(req, iov, needs_lock);
3033 return __io_iov_buffer_select(req, iov, needs_lock);
3036 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3037 struct iovec **iovec, struct iov_iter *iter,
3040 void __user *buf = u64_to_user_ptr(req->rw.addr);
3041 size_t sqe_len = req->rw.len;
3045 opcode = req->opcode;
3046 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3048 return io_import_fixed(req, rw, iter);
3051 /* buffer index only valid with fixed read/write, or buffer select */
3052 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3055 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3056 if (req->flags & REQ_F_BUFFER_SELECT) {
3057 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3059 return PTR_ERR(buf);
3060 req->rw.len = sqe_len;
3063 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3065 return ret < 0 ? ret : sqe_len;
3068 if (req->flags & REQ_F_BUFFER_SELECT) {
3069 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3071 ret = (*iovec)->iov_len;
3072 iov_iter_init(iter, rw, *iovec, 1, ret);
3078 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3082 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3083 struct iovec **iovec, struct iov_iter *iter,
3086 struct io_async_rw *iorw = req->async_data;
3089 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3091 return iov_iter_count(&iorw->iter);
3094 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3096 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3100 * For files that don't have ->read_iter() and ->write_iter(), handle them
3101 * by looping over ->read() or ->write() manually.
3103 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3105 struct kiocb *kiocb = &req->rw.kiocb;
3106 struct file *file = req->file;
3110 * Don't support polled IO through this interface, and we can't
3111 * support non-blocking either. For the latter, this just causes
3112 * the kiocb to be handled from an async context.
3114 if (kiocb->ki_flags & IOCB_HIPRI)
3116 if (kiocb->ki_flags & IOCB_NOWAIT)
3119 while (iov_iter_count(iter)) {
3123 if (!iov_iter_is_bvec(iter)) {
3124 iovec = iov_iter_iovec(iter);
3126 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3127 iovec.iov_len = req->rw.len;
3131 nr = file->f_op->read(file, iovec.iov_base,
3132 iovec.iov_len, io_kiocb_ppos(kiocb));
3134 nr = file->f_op->write(file, iovec.iov_base,
3135 iovec.iov_len, io_kiocb_ppos(kiocb));
3144 if (nr != iovec.iov_len)
3148 iov_iter_advance(iter, nr);
3154 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3155 const struct iovec *fast_iov, struct iov_iter *iter)
3157 struct io_async_rw *rw = req->async_data;
3159 memcpy(&rw->iter, iter, sizeof(*iter));
3160 rw->free_iovec = iovec;
3162 /* can only be fixed buffers, no need to do anything */
3163 if (iter->type == ITER_BVEC)
3166 unsigned iov_off = 0;
3168 rw->iter.iov = rw->fast_iov;
3169 if (iter->iov != fast_iov) {
3170 iov_off = iter->iov - fast_iov;
3171 rw->iter.iov += iov_off;
3173 if (rw->fast_iov != fast_iov)
3174 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3175 sizeof(struct iovec) * iter->nr_segs);
3177 req->flags |= REQ_F_NEED_CLEANUP;
3181 static inline int __io_alloc_async_data(struct io_kiocb *req)
3183 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3184 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3185 return req->async_data == NULL;
3188 static int io_alloc_async_data(struct io_kiocb *req)
3190 if (!io_op_defs[req->opcode].needs_async_data)
3193 return __io_alloc_async_data(req);
3196 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3197 const struct iovec *fast_iov,
3198 struct iov_iter *iter, bool force)
3200 if (!force && !io_op_defs[req->opcode].needs_async_data)
3202 if (!req->async_data) {
3203 if (__io_alloc_async_data(req))
3206 io_req_map_rw(req, iovec, fast_iov, iter);
3211 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3213 struct io_async_rw *iorw = req->async_data;
3214 struct iovec *iov = iorw->fast_iov;
3217 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3218 if (unlikely(ret < 0))
3221 iorw->bytes_done = 0;
3222 iorw->free_iovec = iov;
3224 req->flags |= REQ_F_NEED_CLEANUP;
3228 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3232 ret = io_prep_rw(req, sqe);
3236 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3239 /* either don't need iovec imported or already have it */
3240 if (!req->async_data)
3242 return io_rw_prep_async(req, READ);
3246 * This is our waitqueue callback handler, registered through lock_page_async()
3247 * when we initially tried to do the IO with the iocb armed our waitqueue.
3248 * This gets called when the page is unlocked, and we generally expect that to
3249 * happen when the page IO is completed and the page is now uptodate. This will
3250 * queue a task_work based retry of the operation, attempting to copy the data
3251 * again. If the latter fails because the page was NOT uptodate, then we will
3252 * do a thread based blocking retry of the operation. That's the unexpected
3255 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3256 int sync, void *arg)
3258 struct wait_page_queue *wpq;
3259 struct io_kiocb *req = wait->private;
3260 struct wait_page_key *key = arg;
3263 wpq = container_of(wait, struct wait_page_queue, wait);
3265 if (!wake_page_match(wpq, key))
3268 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3269 list_del_init(&wait->entry);
3271 init_task_work(&req->task_work, io_req_task_submit);
3272 percpu_ref_get(&req->ctx->refs);
3274 /* submit ref gets dropped, acquire a new one */
3275 refcount_inc(&req->refs);
3276 ret = io_req_task_work_add(req, true);
3277 if (unlikely(ret)) {
3278 struct task_struct *tsk;
3280 /* queue just for cancelation */
3281 init_task_work(&req->task_work, io_req_task_cancel);
3282 tsk = io_wq_get_task(req->ctx->io_wq);
3283 task_work_add(tsk, &req->task_work, TWA_NONE);
3284 wake_up_process(tsk);
3290 * This controls whether a given IO request should be armed for async page
3291 * based retry. If we return false here, the request is handed to the async
3292 * worker threads for retry. If we're doing buffered reads on a regular file,
3293 * we prepare a private wait_page_queue entry and retry the operation. This
3294 * will either succeed because the page is now uptodate and unlocked, or it
3295 * will register a callback when the page is unlocked at IO completion. Through
3296 * that callback, io_uring uses task_work to setup a retry of the operation.
3297 * That retry will attempt the buffered read again. The retry will generally
3298 * succeed, or in rare cases where it fails, we then fall back to using the
3299 * async worker threads for a blocking retry.
3301 static bool io_rw_should_retry(struct io_kiocb *req)
3303 struct io_async_rw *rw = req->async_data;
3304 struct wait_page_queue *wait = &rw->wpq;
3305 struct kiocb *kiocb = &req->rw.kiocb;
3307 /* never retry for NOWAIT, we just complete with -EAGAIN */
3308 if (req->flags & REQ_F_NOWAIT)
3311 /* Only for buffered IO */
3312 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3316 * just use poll if we can, and don't attempt if the fs doesn't
3317 * support callback based unlocks
3319 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3322 wait->wait.func = io_async_buf_func;
3323 wait->wait.private = req;
3324 wait->wait.flags = 0;
3325 INIT_LIST_HEAD(&wait->wait.entry);
3326 kiocb->ki_flags |= IOCB_WAITQ;
3327 kiocb->ki_flags &= ~IOCB_NOWAIT;
3328 kiocb->ki_waitq = wait;
3332 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3334 if (req->file->f_op->read_iter)
3335 return call_read_iter(req->file, &req->rw.kiocb, iter);
3336 else if (req->file->f_op->read)
3337 return loop_rw_iter(READ, req, iter);
3342 static int io_read(struct io_kiocb *req, bool force_nonblock,
3343 struct io_comp_state *cs)
3345 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3346 struct kiocb *kiocb = &req->rw.kiocb;
3347 struct iov_iter __iter, *iter = &__iter;
3348 struct io_async_rw *rw = req->async_data;
3349 ssize_t io_size, ret, ret2;
3356 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3359 iov_count = iov_iter_count(iter);
3361 req->result = io_size;
3364 /* Ensure we clear previously set non-block flag */
3365 if (!force_nonblock)
3366 kiocb->ki_flags &= ~IOCB_NOWAIT;
3368 kiocb->ki_flags |= IOCB_NOWAIT;
3371 /* If the file doesn't support async, just async punt */
3372 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3376 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3380 ret = io_iter_do_read(req, iter);
3384 } else if (ret == -EIOCBQUEUED) {
3387 } else if (ret == -EAGAIN) {
3388 /* IOPOLL retry should happen for io-wq threads */
3389 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3391 /* no retry on NONBLOCK marked file */
3392 if (req->file->f_flags & O_NONBLOCK)
3394 /* some cases will consume bytes even on error returns */
3395 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3398 } else if (ret < 0) {
3399 /* make sure -ERESTARTSYS -> -EINTR is done */
3403 /* read it all, or we did blocking attempt. no retry. */
3404 if (!iov_iter_count(iter) || !force_nonblock ||
3405 (req->file->f_flags & O_NONBLOCK))
3410 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3417 rw = req->async_data;
3418 /* it's copied and will be cleaned with ->io */
3420 /* now use our persistent iterator, if we aren't already */
3423 rw->bytes_done += ret;
3424 /* if we can retry, do so with the callbacks armed */
3425 if (!io_rw_should_retry(req)) {
3426 kiocb->ki_flags &= ~IOCB_WAITQ;
3431 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3432 * get -EIOCBQUEUED, then we'll get a notification when the desired
3433 * page gets unlocked. We can also get a partial read here, and if we
3434 * do, then just retry at the new offset.
3436 ret = io_iter_do_read(req, iter);
3437 if (ret == -EIOCBQUEUED) {
3440 } else if (ret > 0 && ret < io_size) {
3441 /* we got some bytes, but not all. retry. */
3445 kiocb_done(kiocb, ret, cs);
3448 /* it's reportedly faster than delegating the null check to kfree() */
3454 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3458 ret = io_prep_rw(req, sqe);
3462 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3465 /* either don't need iovec imported or already have it */
3466 if (!req->async_data)
3468 return io_rw_prep_async(req, WRITE);
3471 static int io_write(struct io_kiocb *req, bool force_nonblock,
3472 struct io_comp_state *cs)
3474 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3475 struct kiocb *kiocb = &req->rw.kiocb;
3476 struct iov_iter __iter, *iter = &__iter;
3477 struct io_async_rw *rw = req->async_data;
3479 ssize_t ret, ret2, io_size;
3484 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3487 iov_count = iov_iter_count(iter);
3489 req->result = io_size;
3491 /* Ensure we clear previously set non-block flag */
3492 if (!force_nonblock)
3493 kiocb->ki_flags &= ~IOCB_NOWAIT;
3495 kiocb->ki_flags |= IOCB_NOWAIT;
3497 /* If the file doesn't support async, just async punt */
3498 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3501 /* file path doesn't support NOWAIT for non-direct_IO */
3502 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3503 (req->flags & REQ_F_ISREG))
3506 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3511 * Open-code file_start_write here to grab freeze protection,
3512 * which will be released by another thread in
3513 * io_complete_rw(). Fool lockdep by telling it the lock got
3514 * released so that it doesn't complain about the held lock when
3515 * we return to userspace.
3517 if (req->flags & REQ_F_ISREG) {
3518 __sb_start_write(file_inode(req->file)->i_sb,
3519 SB_FREEZE_WRITE, true);
3520 __sb_writers_release(file_inode(req->file)->i_sb,
3523 kiocb->ki_flags |= IOCB_WRITE;
3525 if (req->file->f_op->write_iter)
3526 ret2 = call_write_iter(req->file, kiocb, iter);
3527 else if (req->file->f_op->write)
3528 ret2 = loop_rw_iter(WRITE, req, iter);
3533 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3534 * retry them without IOCB_NOWAIT.
3536 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3538 /* no retry on NONBLOCK marked file */
3539 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3541 if (!force_nonblock || ret2 != -EAGAIN) {
3542 /* IOPOLL retry should happen for io-wq threads */
3543 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3546 kiocb_done(kiocb, ret2, cs);
3549 /* some cases will consume bytes even on error returns */
3550 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3551 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3556 /* it's reportedly faster than delegating the null check to kfree() */
3562 static int __io_splice_prep(struct io_kiocb *req,
3563 const struct io_uring_sqe *sqe)
3565 struct io_splice* sp = &req->splice;
3566 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3568 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3572 sp->len = READ_ONCE(sqe->len);
3573 sp->flags = READ_ONCE(sqe->splice_flags);
3575 if (unlikely(sp->flags & ~valid_flags))
3578 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3579 (sp->flags & SPLICE_F_FD_IN_FIXED));
3582 req->flags |= REQ_F_NEED_CLEANUP;
3584 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3586 * Splice operation will be punted aync, and here need to
3587 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3589 io_req_init_async(req);
3590 req->work.flags |= IO_WQ_WORK_UNBOUND;
3596 static int io_tee_prep(struct io_kiocb *req,
3597 const struct io_uring_sqe *sqe)
3599 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3601 return __io_splice_prep(req, sqe);
3604 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3606 struct io_splice *sp = &req->splice;
3607 struct file *in = sp->file_in;
3608 struct file *out = sp->file_out;
3609 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3615 ret = do_tee(in, out, sp->len, flags);
3617 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3618 req->flags &= ~REQ_F_NEED_CLEANUP;
3621 req_set_fail_links(req);
3622 io_req_complete(req, ret);
3626 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3628 struct io_splice* sp = &req->splice;
3630 sp->off_in = READ_ONCE(sqe->splice_off_in);
3631 sp->off_out = READ_ONCE(sqe->off);
3632 return __io_splice_prep(req, sqe);
3635 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3637 struct io_splice *sp = &req->splice;
3638 struct file *in = sp->file_in;
3639 struct file *out = sp->file_out;
3640 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3641 loff_t *poff_in, *poff_out;
3647 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3648 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3651 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3653 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3654 req->flags &= ~REQ_F_NEED_CLEANUP;
3657 req_set_fail_links(req);
3658 io_req_complete(req, ret);
3663 * IORING_OP_NOP just posts a completion event, nothing else.
3665 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3667 struct io_ring_ctx *ctx = req->ctx;
3669 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3672 __io_req_complete(req, 0, 0, cs);
3676 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3678 struct io_ring_ctx *ctx = req->ctx;
3683 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3685 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3688 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3689 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3692 req->sync.off = READ_ONCE(sqe->off);
3693 req->sync.len = READ_ONCE(sqe->len);
3697 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3699 loff_t end = req->sync.off + req->sync.len;
3702 /* fsync always requires a blocking context */
3706 ret = vfs_fsync_range(req->file, req->sync.off,
3707 end > 0 ? end : LLONG_MAX,
3708 req->sync.flags & IORING_FSYNC_DATASYNC);
3710 req_set_fail_links(req);
3711 io_req_complete(req, ret);
3715 static int io_fallocate_prep(struct io_kiocb *req,
3716 const struct io_uring_sqe *sqe)
3718 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3723 req->sync.off = READ_ONCE(sqe->off);
3724 req->sync.len = READ_ONCE(sqe->addr);
3725 req->sync.mode = READ_ONCE(sqe->len);
3729 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3733 /* fallocate always requiring blocking context */
3736 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3739 req_set_fail_links(req);
3740 io_req_complete(req, ret);
3744 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3746 const char __user *fname;
3749 if (unlikely(sqe->ioprio || sqe->buf_index))
3751 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3754 /* open.how should be already initialised */
3755 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3756 req->open.how.flags |= O_LARGEFILE;
3758 req->open.dfd = READ_ONCE(sqe->fd);
3759 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3760 req->open.filename = getname(fname);
3761 if (IS_ERR(req->open.filename)) {
3762 ret = PTR_ERR(req->open.filename);
3763 req->open.filename = NULL;
3766 req->open.nofile = rlimit(RLIMIT_NOFILE);
3767 req->flags |= REQ_F_NEED_CLEANUP;
3771 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3775 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3777 mode = READ_ONCE(sqe->len);
3778 flags = READ_ONCE(sqe->open_flags);
3779 req->open.how = build_open_how(flags, mode);
3780 return __io_openat_prep(req, sqe);
3783 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3785 struct open_how __user *how;
3789 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3791 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3792 len = READ_ONCE(sqe->len);
3793 if (len < OPEN_HOW_SIZE_VER0)
3796 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3801 return __io_openat_prep(req, sqe);
3804 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3806 struct open_flags op;
3813 ret = build_open_flags(&req->open.how, &op);
3817 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3821 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3824 ret = PTR_ERR(file);
3826 fsnotify_open(file);
3827 fd_install(ret, file);
3830 putname(req->open.filename);
3831 req->flags &= ~REQ_F_NEED_CLEANUP;
3833 req_set_fail_links(req);
3834 io_req_complete(req, ret);
3838 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3840 return io_openat2(req, force_nonblock);
3843 static int io_remove_buffers_prep(struct io_kiocb *req,
3844 const struct io_uring_sqe *sqe)
3846 struct io_provide_buf *p = &req->pbuf;
3849 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3852 tmp = READ_ONCE(sqe->fd);
3853 if (!tmp || tmp > USHRT_MAX)
3856 memset(p, 0, sizeof(*p));
3858 p->bgid = READ_ONCE(sqe->buf_group);
3862 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3863 int bgid, unsigned nbufs)
3867 /* shouldn't happen */
3871 /* the head kbuf is the list itself */
3872 while (!list_empty(&buf->list)) {
3873 struct io_buffer *nxt;
3875 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3876 list_del(&nxt->list);
3883 idr_remove(&ctx->io_buffer_idr, bgid);
3888 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3889 struct io_comp_state *cs)
3891 struct io_provide_buf *p = &req->pbuf;
3892 struct io_ring_ctx *ctx = req->ctx;
3893 struct io_buffer *head;
3896 io_ring_submit_lock(ctx, !force_nonblock);
3898 lockdep_assert_held(&ctx->uring_lock);
3901 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3903 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3905 io_ring_submit_lock(ctx, !force_nonblock);
3907 req_set_fail_links(req);
3908 __io_req_complete(req, ret, 0, cs);
3912 static int io_provide_buffers_prep(struct io_kiocb *req,
3913 const struct io_uring_sqe *sqe)
3915 struct io_provide_buf *p = &req->pbuf;
3918 if (sqe->ioprio || sqe->rw_flags)
3921 tmp = READ_ONCE(sqe->fd);
3922 if (!tmp || tmp > USHRT_MAX)
3925 p->addr = READ_ONCE(sqe->addr);
3926 p->len = READ_ONCE(sqe->len);
3928 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3931 p->bgid = READ_ONCE(sqe->buf_group);
3932 tmp = READ_ONCE(sqe->off);
3933 if (tmp > USHRT_MAX)
3939 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3941 struct io_buffer *buf;
3942 u64 addr = pbuf->addr;
3943 int i, bid = pbuf->bid;
3945 for (i = 0; i < pbuf->nbufs; i++) {
3946 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3951 buf->len = pbuf->len;
3956 INIT_LIST_HEAD(&buf->list);
3959 list_add_tail(&buf->list, &(*head)->list);
3963 return i ? i : -ENOMEM;
3966 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3967 struct io_comp_state *cs)
3969 struct io_provide_buf *p = &req->pbuf;
3970 struct io_ring_ctx *ctx = req->ctx;
3971 struct io_buffer *head, *list;
3974 io_ring_submit_lock(ctx, !force_nonblock);
3976 lockdep_assert_held(&ctx->uring_lock);
3978 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3980 ret = io_add_buffers(p, &head);
3985 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3988 __io_remove_buffers(ctx, head, p->bgid, -1U);
3993 io_ring_submit_unlock(ctx, !force_nonblock);
3995 req_set_fail_links(req);
3996 __io_req_complete(req, ret, 0, cs);
4000 static int io_epoll_ctl_prep(struct io_kiocb *req,
4001 const struct io_uring_sqe *sqe)
4003 #if defined(CONFIG_EPOLL)
4004 if (sqe->ioprio || sqe->buf_index)
4006 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4009 req->epoll.epfd = READ_ONCE(sqe->fd);
4010 req->epoll.op = READ_ONCE(sqe->len);
4011 req->epoll.fd = READ_ONCE(sqe->off);
4013 if (ep_op_has_event(req->epoll.op)) {
4014 struct epoll_event __user *ev;
4016 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4017 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4027 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4028 struct io_comp_state *cs)
4030 #if defined(CONFIG_EPOLL)
4031 struct io_epoll *ie = &req->epoll;
4034 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4035 if (force_nonblock && ret == -EAGAIN)
4039 req_set_fail_links(req);
4040 __io_req_complete(req, ret, 0, cs);
4047 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4049 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4050 if (sqe->ioprio || sqe->buf_index || sqe->off)
4052 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4055 req->madvise.addr = READ_ONCE(sqe->addr);
4056 req->madvise.len = READ_ONCE(sqe->len);
4057 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4064 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4066 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4067 struct io_madvise *ma = &req->madvise;
4073 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4075 req_set_fail_links(req);
4076 io_req_complete(req, ret);
4083 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4085 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4087 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4090 req->fadvise.offset = READ_ONCE(sqe->off);
4091 req->fadvise.len = READ_ONCE(sqe->len);
4092 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4096 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4098 struct io_fadvise *fa = &req->fadvise;
4101 if (force_nonblock) {
4102 switch (fa->advice) {
4103 case POSIX_FADV_NORMAL:
4104 case POSIX_FADV_RANDOM:
4105 case POSIX_FADV_SEQUENTIAL:
4112 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4114 req_set_fail_links(req);
4115 io_req_complete(req, ret);
4119 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4121 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4123 if (sqe->ioprio || sqe->buf_index)
4125 if (req->flags & REQ_F_FIXED_FILE)
4128 req->statx.dfd = READ_ONCE(sqe->fd);
4129 req->statx.mask = READ_ONCE(sqe->len);
4130 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4131 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4132 req->statx.flags = READ_ONCE(sqe->statx_flags);
4137 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4139 struct io_statx *ctx = &req->statx;
4142 if (force_nonblock) {
4143 /* only need file table for an actual valid fd */
4144 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4145 req->flags |= REQ_F_NO_FILE_TABLE;
4149 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4153 req_set_fail_links(req);
4154 io_req_complete(req, ret);
4158 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4161 * If we queue this for async, it must not be cancellable. That would
4162 * leave the 'file' in an undeterminate state, and here need to modify
4163 * io_wq_work.flags, so initialize io_wq_work firstly.
4165 io_req_init_async(req);
4166 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4168 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4170 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4171 sqe->rw_flags || sqe->buf_index)
4173 if (req->flags & REQ_F_FIXED_FILE)
4176 req->close.fd = READ_ONCE(sqe->fd);
4177 if ((req->file && req->file->f_op == &io_uring_fops))
4180 req->close.put_file = NULL;
4184 static int io_close(struct io_kiocb *req, bool force_nonblock,
4185 struct io_comp_state *cs)
4187 struct io_close *close = &req->close;
4190 /* might be already done during nonblock submission */
4191 if (!close->put_file) {
4192 ret = __close_fd_get_file(close->fd, &close->put_file);
4194 return (ret == -ENOENT) ? -EBADF : ret;
4197 /* if the file has a flush method, be safe and punt to async */
4198 if (close->put_file->f_op->flush && force_nonblock) {
4199 /* was never set, but play safe */
4200 req->flags &= ~REQ_F_NOWAIT;
4201 /* avoid grabbing files - we don't need the files */
4202 req->flags |= REQ_F_NO_FILE_TABLE;
4206 /* No ->flush() or already async, safely close from here */
4207 ret = filp_close(close->put_file, req->work.identity->files);
4209 req_set_fail_links(req);
4210 fput(close->put_file);
4211 close->put_file = NULL;
4212 __io_req_complete(req, ret, 0, cs);
4216 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4218 struct io_ring_ctx *ctx = req->ctx;
4223 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4225 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4228 req->sync.off = READ_ONCE(sqe->off);
4229 req->sync.len = READ_ONCE(sqe->len);
4230 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4234 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4238 /* sync_file_range always requires a blocking context */
4242 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4245 req_set_fail_links(req);
4246 io_req_complete(req, ret);
4250 #if defined(CONFIG_NET)
4251 static int io_setup_async_msg(struct io_kiocb *req,
4252 struct io_async_msghdr *kmsg)
4254 struct io_async_msghdr *async_msg = req->async_data;
4258 if (io_alloc_async_data(req)) {
4259 if (kmsg->iov != kmsg->fast_iov)
4263 async_msg = req->async_data;
4264 req->flags |= REQ_F_NEED_CLEANUP;
4265 memcpy(async_msg, kmsg, sizeof(*kmsg));
4269 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4270 struct io_async_msghdr *iomsg)
4272 iomsg->iov = iomsg->fast_iov;
4273 iomsg->msg.msg_name = &iomsg->addr;
4274 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4275 req->sr_msg.msg_flags, &iomsg->iov);
4278 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4280 struct io_async_msghdr *async_msg = req->async_data;
4281 struct io_sr_msg *sr = &req->sr_msg;
4284 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4287 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4288 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4289 sr->len = READ_ONCE(sqe->len);
4291 #ifdef CONFIG_COMPAT
4292 if (req->ctx->compat)
4293 sr->msg_flags |= MSG_CMSG_COMPAT;
4296 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4298 ret = io_sendmsg_copy_hdr(req, async_msg);
4300 req->flags |= REQ_F_NEED_CLEANUP;
4304 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4305 struct io_comp_state *cs)
4307 struct io_async_msghdr iomsg, *kmsg;
4308 struct socket *sock;
4312 sock = sock_from_file(req->file, &ret);
4313 if (unlikely(!sock))
4316 if (req->async_data) {
4317 kmsg = req->async_data;
4318 kmsg->msg.msg_name = &kmsg->addr;
4319 /* if iov is set, it's allocated already */
4321 kmsg->iov = kmsg->fast_iov;
4322 kmsg->msg.msg_iter.iov = kmsg->iov;
4324 ret = io_sendmsg_copy_hdr(req, &iomsg);
4330 flags = req->sr_msg.msg_flags;
4331 if (flags & MSG_DONTWAIT)
4332 req->flags |= REQ_F_NOWAIT;
4333 else if (force_nonblock)
4334 flags |= MSG_DONTWAIT;
4336 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4337 if (force_nonblock && ret == -EAGAIN)
4338 return io_setup_async_msg(req, kmsg);
4339 if (ret == -ERESTARTSYS)
4342 if (kmsg->iov != kmsg->fast_iov)
4344 req->flags &= ~REQ_F_NEED_CLEANUP;
4346 req_set_fail_links(req);
4347 __io_req_complete(req, ret, 0, cs);
4351 static int io_send(struct io_kiocb *req, bool force_nonblock,
4352 struct io_comp_state *cs)
4354 struct io_sr_msg *sr = &req->sr_msg;
4357 struct socket *sock;
4361 sock = sock_from_file(req->file, &ret);
4362 if (unlikely(!sock))
4365 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4369 msg.msg_name = NULL;
4370 msg.msg_control = NULL;
4371 msg.msg_controllen = 0;
4372 msg.msg_namelen = 0;
4374 flags = req->sr_msg.msg_flags;
4375 if (flags & MSG_DONTWAIT)
4376 req->flags |= REQ_F_NOWAIT;
4377 else if (force_nonblock)
4378 flags |= MSG_DONTWAIT;
4380 msg.msg_flags = flags;
4381 ret = sock_sendmsg(sock, &msg);
4382 if (force_nonblock && ret == -EAGAIN)
4384 if (ret == -ERESTARTSYS)
4388 req_set_fail_links(req);
4389 __io_req_complete(req, ret, 0, cs);
4393 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4394 struct io_async_msghdr *iomsg)
4396 struct io_sr_msg *sr = &req->sr_msg;
4397 struct iovec __user *uiov;
4401 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4402 &iomsg->uaddr, &uiov, &iov_len);
4406 if (req->flags & REQ_F_BUFFER_SELECT) {
4409 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4411 sr->len = iomsg->iov[0].iov_len;
4412 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4416 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4417 &iomsg->iov, &iomsg->msg.msg_iter,
4426 #ifdef CONFIG_COMPAT
4427 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4428 struct io_async_msghdr *iomsg)
4430 struct compat_msghdr __user *msg_compat;
4431 struct io_sr_msg *sr = &req->sr_msg;
4432 struct compat_iovec __user *uiov;
4437 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4438 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4443 uiov = compat_ptr(ptr);
4444 if (req->flags & REQ_F_BUFFER_SELECT) {
4445 compat_ssize_t clen;
4449 if (!access_ok(uiov, sizeof(*uiov)))
4451 if (__get_user(clen, &uiov->iov_len))
4455 sr->len = iomsg->iov[0].iov_len;
4458 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4459 UIO_FASTIOV, &iomsg->iov,
4460 &iomsg->msg.msg_iter, true);
4469 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4470 struct io_async_msghdr *iomsg)
4472 iomsg->msg.msg_name = &iomsg->addr;
4473 iomsg->iov = iomsg->fast_iov;
4475 #ifdef CONFIG_COMPAT
4476 if (req->ctx->compat)
4477 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4480 return __io_recvmsg_copy_hdr(req, iomsg);
4483 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4486 struct io_sr_msg *sr = &req->sr_msg;
4487 struct io_buffer *kbuf;
4489 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4494 req->flags |= REQ_F_BUFFER_SELECTED;
4498 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4500 return io_put_kbuf(req, req->sr_msg.kbuf);
4503 static int io_recvmsg_prep(struct io_kiocb *req,
4504 const struct io_uring_sqe *sqe)
4506 struct io_async_msghdr *async_msg = req->async_data;
4507 struct io_sr_msg *sr = &req->sr_msg;
4510 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4513 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4514 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4515 sr->len = READ_ONCE(sqe->len);
4516 sr->bgid = READ_ONCE(sqe->buf_group);
4518 #ifdef CONFIG_COMPAT
4519 if (req->ctx->compat)
4520 sr->msg_flags |= MSG_CMSG_COMPAT;
4523 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4525 ret = io_recvmsg_copy_hdr(req, async_msg);
4527 req->flags |= REQ_F_NEED_CLEANUP;
4531 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4532 struct io_comp_state *cs)
4534 struct io_async_msghdr iomsg, *kmsg;
4535 struct socket *sock;
4536 struct io_buffer *kbuf;
4538 int ret, cflags = 0;
4540 sock = sock_from_file(req->file, &ret);
4541 if (unlikely(!sock))
4544 if (req->async_data) {
4545 kmsg = req->async_data;
4546 kmsg->msg.msg_name = &kmsg->addr;
4547 /* if iov is set, it's allocated already */
4549 kmsg->iov = kmsg->fast_iov;
4550 kmsg->msg.msg_iter.iov = kmsg->iov;
4552 ret = io_recvmsg_copy_hdr(req, &iomsg);
4558 if (req->flags & REQ_F_BUFFER_SELECT) {
4559 kbuf = io_recv_buffer_select(req, !force_nonblock);
4561 return PTR_ERR(kbuf);
4562 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4563 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4564 1, req->sr_msg.len);
4567 flags = req->sr_msg.msg_flags;
4568 if (flags & MSG_DONTWAIT)
4569 req->flags |= REQ_F_NOWAIT;
4570 else if (force_nonblock)
4571 flags |= MSG_DONTWAIT;
4573 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4574 kmsg->uaddr, flags);
4575 if (force_nonblock && ret == -EAGAIN)
4576 return io_setup_async_msg(req, kmsg);
4577 if (ret == -ERESTARTSYS)
4580 if (req->flags & REQ_F_BUFFER_SELECTED)
4581 cflags = io_put_recv_kbuf(req);
4582 if (kmsg->iov != kmsg->fast_iov)
4584 req->flags &= ~REQ_F_NEED_CLEANUP;
4586 req_set_fail_links(req);
4587 __io_req_complete(req, ret, cflags, cs);
4591 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4592 struct io_comp_state *cs)
4594 struct io_buffer *kbuf;
4595 struct io_sr_msg *sr = &req->sr_msg;
4597 void __user *buf = sr->buf;
4598 struct socket *sock;
4601 int ret, cflags = 0;
4603 sock = sock_from_file(req->file, &ret);
4604 if (unlikely(!sock))
4607 if (req->flags & REQ_F_BUFFER_SELECT) {
4608 kbuf = io_recv_buffer_select(req, !force_nonblock);
4610 return PTR_ERR(kbuf);
4611 buf = u64_to_user_ptr(kbuf->addr);
4614 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4618 msg.msg_name = NULL;
4619 msg.msg_control = NULL;
4620 msg.msg_controllen = 0;
4621 msg.msg_namelen = 0;
4622 msg.msg_iocb = NULL;
4625 flags = req->sr_msg.msg_flags;
4626 if (flags & MSG_DONTWAIT)
4627 req->flags |= REQ_F_NOWAIT;
4628 else if (force_nonblock)
4629 flags |= MSG_DONTWAIT;
4631 ret = sock_recvmsg(sock, &msg, flags);
4632 if (force_nonblock && ret == -EAGAIN)
4634 if (ret == -ERESTARTSYS)
4637 if (req->flags & REQ_F_BUFFER_SELECTED)
4638 cflags = io_put_recv_kbuf(req);
4640 req_set_fail_links(req);
4641 __io_req_complete(req, ret, cflags, cs);
4645 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4647 struct io_accept *accept = &req->accept;
4649 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4651 if (sqe->ioprio || sqe->len || sqe->buf_index)
4654 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4655 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4656 accept->flags = READ_ONCE(sqe->accept_flags);
4657 accept->nofile = rlimit(RLIMIT_NOFILE);
4661 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4662 struct io_comp_state *cs)
4664 struct io_accept *accept = &req->accept;
4665 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4668 if (req->file->f_flags & O_NONBLOCK)
4669 req->flags |= REQ_F_NOWAIT;
4671 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4672 accept->addr_len, accept->flags,
4674 if (ret == -EAGAIN && force_nonblock)
4677 if (ret == -ERESTARTSYS)
4679 req_set_fail_links(req);
4681 __io_req_complete(req, ret, 0, cs);
4685 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4687 struct io_connect *conn = &req->connect;
4688 struct io_async_connect *io = req->async_data;
4690 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4692 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4695 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4696 conn->addr_len = READ_ONCE(sqe->addr2);
4701 return move_addr_to_kernel(conn->addr, conn->addr_len,
4705 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4706 struct io_comp_state *cs)
4708 struct io_async_connect __io, *io;
4709 unsigned file_flags;
4712 if (req->async_data) {
4713 io = req->async_data;
4715 ret = move_addr_to_kernel(req->connect.addr,
4716 req->connect.addr_len,
4723 file_flags = force_nonblock ? O_NONBLOCK : 0;
4725 ret = __sys_connect_file(req->file, &io->address,
4726 req->connect.addr_len, file_flags);
4727 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4728 if (req->async_data)
4730 if (io_alloc_async_data(req)) {
4734 io = req->async_data;
4735 memcpy(req->async_data, &__io, sizeof(__io));
4738 if (ret == -ERESTARTSYS)
4742 req_set_fail_links(req);
4743 __io_req_complete(req, ret, 0, cs);
4746 #else /* !CONFIG_NET */
4747 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4752 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4753 struct io_comp_state *cs)
4758 static int io_send(struct io_kiocb *req, bool force_nonblock,
4759 struct io_comp_state *cs)
4764 static int io_recvmsg_prep(struct io_kiocb *req,
4765 const struct io_uring_sqe *sqe)
4770 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4771 struct io_comp_state *cs)
4776 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4777 struct io_comp_state *cs)
4782 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4787 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4788 struct io_comp_state *cs)
4793 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4798 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4799 struct io_comp_state *cs)
4803 #endif /* CONFIG_NET */
4805 struct io_poll_table {
4806 struct poll_table_struct pt;
4807 struct io_kiocb *req;
4811 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4812 __poll_t mask, task_work_func_t func)
4817 /* for instances that support it check for an event match first: */
4818 if (mask && !(mask & poll->events))
4821 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4823 list_del_init(&poll->wait.entry);
4826 init_task_work(&req->task_work, func);
4827 percpu_ref_get(&req->ctx->refs);
4830 * If we using the signalfd wait_queue_head for this wakeup, then
4831 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4832 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4833 * either, as the normal wakeup will suffice.
4835 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4838 * If this fails, then the task is exiting. When a task exits, the
4839 * work gets canceled, so just cancel this request as well instead
4840 * of executing it. We can't safely execute it anyway, as we may not
4841 * have the needed state needed for it anyway.
4843 ret = io_req_task_work_add(req, twa_signal_ok);
4844 if (unlikely(ret)) {
4845 struct task_struct *tsk;
4847 WRITE_ONCE(poll->canceled, true);
4848 tsk = io_wq_get_task(req->ctx->io_wq);
4849 task_work_add(tsk, &req->task_work, TWA_NONE);
4850 wake_up_process(tsk);
4855 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4856 __acquires(&req->ctx->completion_lock)
4858 struct io_ring_ctx *ctx = req->ctx;
4860 if (!req->result && !READ_ONCE(poll->canceled)) {
4861 struct poll_table_struct pt = { ._key = poll->events };
4863 req->result = vfs_poll(req->file, &pt) & poll->events;
4866 spin_lock_irq(&ctx->completion_lock);
4867 if (!req->result && !READ_ONCE(poll->canceled)) {
4868 add_wait_queue(poll->head, &poll->wait);
4875 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4877 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4878 if (req->opcode == IORING_OP_POLL_ADD)
4879 return req->async_data;
4880 return req->apoll->double_poll;
4883 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4885 if (req->opcode == IORING_OP_POLL_ADD)
4887 return &req->apoll->poll;
4890 static void io_poll_remove_double(struct io_kiocb *req)
4892 struct io_poll_iocb *poll = io_poll_get_double(req);
4894 lockdep_assert_held(&req->ctx->completion_lock);
4896 if (poll && poll->head) {
4897 struct wait_queue_head *head = poll->head;
4899 spin_lock(&head->lock);
4900 list_del_init(&poll->wait.entry);
4901 if (poll->wait.private)
4902 refcount_dec(&req->refs);
4904 spin_unlock(&head->lock);
4908 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4910 struct io_ring_ctx *ctx = req->ctx;
4912 io_poll_remove_double(req);
4913 req->poll.done = true;
4914 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4915 io_commit_cqring(ctx);
4918 static void io_poll_task_func(struct callback_head *cb)
4920 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4921 struct io_ring_ctx *ctx = req->ctx;
4922 struct io_kiocb *nxt;
4924 if (io_poll_rewait(req, &req->poll)) {
4925 spin_unlock_irq(&ctx->completion_lock);
4927 hash_del(&req->hash_node);
4928 io_poll_complete(req, req->result, 0);
4929 spin_unlock_irq(&ctx->completion_lock);
4931 nxt = io_put_req_find_next(req);
4932 io_cqring_ev_posted(ctx);
4934 __io_req_task_submit(nxt);
4937 percpu_ref_put(&ctx->refs);
4940 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4941 int sync, void *key)
4943 struct io_kiocb *req = wait->private;
4944 struct io_poll_iocb *poll = io_poll_get_single(req);
4945 __poll_t mask = key_to_poll(key);
4947 /* for instances that support it check for an event match first: */
4948 if (mask && !(mask & poll->events))
4951 list_del_init(&wait->entry);
4953 if (poll && poll->head) {
4956 spin_lock(&poll->head->lock);
4957 done = list_empty(&poll->wait.entry);
4959 list_del_init(&poll->wait.entry);
4960 /* make sure double remove sees this as being gone */
4961 wait->private = NULL;
4962 spin_unlock(&poll->head->lock);
4964 /* use wait func handler, so it matches the rq type */
4965 poll->wait.func(&poll->wait, mode, sync, key);
4968 refcount_dec(&req->refs);
4972 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4973 wait_queue_func_t wake_func)
4977 poll->canceled = false;
4978 poll->events = events;
4979 INIT_LIST_HEAD(&poll->wait.entry);
4980 init_waitqueue_func_entry(&poll->wait, wake_func);
4983 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4984 struct wait_queue_head *head,
4985 struct io_poll_iocb **poll_ptr)
4987 struct io_kiocb *req = pt->req;
4990 * If poll->head is already set, it's because the file being polled
4991 * uses multiple waitqueues for poll handling (eg one for read, one
4992 * for write). Setup a separate io_poll_iocb if this happens.
4994 if (unlikely(poll->head)) {
4995 struct io_poll_iocb *poll_one = poll;
4997 /* already have a 2nd entry, fail a third attempt */
4999 pt->error = -EINVAL;
5002 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5004 pt->error = -ENOMEM;
5007 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5008 refcount_inc(&req->refs);
5009 poll->wait.private = req;
5016 if (poll->events & EPOLLEXCLUSIVE)
5017 add_wait_queue_exclusive(head, &poll->wait);
5019 add_wait_queue(head, &poll->wait);
5022 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5023 struct poll_table_struct *p)
5025 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5026 struct async_poll *apoll = pt->req->apoll;
5028 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5031 static void io_async_task_func(struct callback_head *cb)
5033 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5034 struct async_poll *apoll = req->apoll;
5035 struct io_ring_ctx *ctx = req->ctx;
5037 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5039 if (io_poll_rewait(req, &apoll->poll)) {
5040 spin_unlock_irq(&ctx->completion_lock);
5041 percpu_ref_put(&ctx->refs);
5045 /* If req is still hashed, it cannot have been canceled. Don't check. */
5046 if (hash_hashed(&req->hash_node))
5047 hash_del(&req->hash_node);
5049 io_poll_remove_double(req);
5050 spin_unlock_irq(&ctx->completion_lock);
5052 if (!READ_ONCE(apoll->poll.canceled))
5053 __io_req_task_submit(req);
5055 __io_req_task_cancel(req, -ECANCELED);
5057 percpu_ref_put(&ctx->refs);
5058 kfree(apoll->double_poll);
5062 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5065 struct io_kiocb *req = wait->private;
5066 struct io_poll_iocb *poll = &req->apoll->poll;
5068 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5071 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5074 static void io_poll_req_insert(struct io_kiocb *req)
5076 struct io_ring_ctx *ctx = req->ctx;
5077 struct hlist_head *list;
5079 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5080 hlist_add_head(&req->hash_node, list);
5083 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5084 struct io_poll_iocb *poll,
5085 struct io_poll_table *ipt, __poll_t mask,
5086 wait_queue_func_t wake_func)
5087 __acquires(&ctx->completion_lock)
5089 struct io_ring_ctx *ctx = req->ctx;
5090 bool cancel = false;
5092 INIT_HLIST_NODE(&req->hash_node);
5093 io_init_poll_iocb(poll, mask, wake_func);
5094 poll->file = req->file;
5095 poll->wait.private = req;
5097 ipt->pt._key = mask;
5099 ipt->error = -EINVAL;
5101 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5103 spin_lock_irq(&ctx->completion_lock);
5104 if (likely(poll->head)) {
5105 spin_lock(&poll->head->lock);
5106 if (unlikely(list_empty(&poll->wait.entry))) {
5112 if (mask || ipt->error)
5113 list_del_init(&poll->wait.entry);
5115 WRITE_ONCE(poll->canceled, true);
5116 else if (!poll->done) /* actually waiting for an event */
5117 io_poll_req_insert(req);
5118 spin_unlock(&poll->head->lock);
5124 static bool io_arm_poll_handler(struct io_kiocb *req)
5126 const struct io_op_def *def = &io_op_defs[req->opcode];
5127 struct io_ring_ctx *ctx = req->ctx;
5128 struct async_poll *apoll;
5129 struct io_poll_table ipt;
5133 if (!req->file || !file_can_poll(req->file))
5135 if (req->flags & REQ_F_POLLED)
5139 else if (def->pollout)
5143 /* if we can't nonblock try, then no point in arming a poll handler */
5144 if (!io_file_supports_async(req->file, rw))
5147 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5148 if (unlikely(!apoll))
5150 apoll->double_poll = NULL;
5152 req->flags |= REQ_F_POLLED;
5157 mask |= POLLIN | POLLRDNORM;
5159 mask |= POLLOUT | POLLWRNORM;
5161 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5162 if ((req->opcode == IORING_OP_RECVMSG) &&
5163 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5166 mask |= POLLERR | POLLPRI;
5168 ipt.pt._qproc = io_async_queue_proc;
5170 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5172 if (ret || ipt.error) {
5173 io_poll_remove_double(req);
5174 spin_unlock_irq(&ctx->completion_lock);
5175 kfree(apoll->double_poll);
5179 spin_unlock_irq(&ctx->completion_lock);
5180 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5181 apoll->poll.events);
5185 static bool __io_poll_remove_one(struct io_kiocb *req,
5186 struct io_poll_iocb *poll)
5188 bool do_complete = false;
5190 spin_lock(&poll->head->lock);
5191 WRITE_ONCE(poll->canceled, true);
5192 if (!list_empty(&poll->wait.entry)) {
5193 list_del_init(&poll->wait.entry);
5196 spin_unlock(&poll->head->lock);
5197 hash_del(&req->hash_node);
5201 static bool io_poll_remove_one(struct io_kiocb *req)
5205 io_poll_remove_double(req);
5207 if (req->opcode == IORING_OP_POLL_ADD) {
5208 do_complete = __io_poll_remove_one(req, &req->poll);
5210 struct async_poll *apoll = req->apoll;
5212 /* non-poll requests have submit ref still */
5213 do_complete = __io_poll_remove_one(req, &apoll->poll);
5216 kfree(apoll->double_poll);
5222 io_cqring_fill_event(req, -ECANCELED);
5223 io_commit_cqring(req->ctx);
5224 req_set_fail_links(req);
5225 io_put_req_deferred(req, 1);
5232 * Returns true if we found and killed one or more poll requests
5234 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5236 struct hlist_node *tmp;
5237 struct io_kiocb *req;
5240 spin_lock_irq(&ctx->completion_lock);
5241 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5242 struct hlist_head *list;
5244 list = &ctx->cancel_hash[i];
5245 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5246 if (io_task_match(req, tsk))
5247 posted += io_poll_remove_one(req);
5250 spin_unlock_irq(&ctx->completion_lock);
5253 io_cqring_ev_posted(ctx);
5258 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5260 struct hlist_head *list;
5261 struct io_kiocb *req;
5263 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5264 hlist_for_each_entry(req, list, hash_node) {
5265 if (sqe_addr != req->user_data)
5267 if (io_poll_remove_one(req))
5275 static int io_poll_remove_prep(struct io_kiocb *req,
5276 const struct io_uring_sqe *sqe)
5278 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5280 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5284 req->poll.addr = READ_ONCE(sqe->addr);
5289 * Find a running poll command that matches one specified in sqe->addr,
5290 * and remove it if found.
5292 static int io_poll_remove(struct io_kiocb *req)
5294 struct io_ring_ctx *ctx = req->ctx;
5298 addr = req->poll.addr;
5299 spin_lock_irq(&ctx->completion_lock);
5300 ret = io_poll_cancel(ctx, addr);
5301 spin_unlock_irq(&ctx->completion_lock);
5304 req_set_fail_links(req);
5305 io_req_complete(req, ret);
5309 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5312 struct io_kiocb *req = wait->private;
5313 struct io_poll_iocb *poll = &req->poll;
5315 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5318 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5319 struct poll_table_struct *p)
5321 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5323 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5326 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5328 struct io_poll_iocb *poll = &req->poll;
5331 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5333 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5336 events = READ_ONCE(sqe->poll32_events);
5338 events = swahw32(events);
5340 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5341 (events & EPOLLEXCLUSIVE);
5345 static int io_poll_add(struct io_kiocb *req)
5347 struct io_poll_iocb *poll = &req->poll;
5348 struct io_ring_ctx *ctx = req->ctx;
5349 struct io_poll_table ipt;
5352 ipt.pt._qproc = io_poll_queue_proc;
5354 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5357 if (mask) { /* no async, we'd stolen it */
5359 io_poll_complete(req, mask, 0);
5361 spin_unlock_irq(&ctx->completion_lock);
5364 io_cqring_ev_posted(ctx);
5370 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5372 struct io_timeout_data *data = container_of(timer,
5373 struct io_timeout_data, timer);
5374 struct io_kiocb *req = data->req;
5375 struct io_ring_ctx *ctx = req->ctx;
5376 unsigned long flags;
5378 spin_lock_irqsave(&ctx->completion_lock, flags);
5379 list_del_init(&req->timeout.list);
5380 atomic_set(&req->ctx->cq_timeouts,
5381 atomic_read(&req->ctx->cq_timeouts) + 1);
5383 io_cqring_fill_event(req, -ETIME);
5384 io_commit_cqring(ctx);
5385 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5387 io_cqring_ev_posted(ctx);
5388 req_set_fail_links(req);
5390 return HRTIMER_NORESTART;
5393 static int __io_timeout_cancel(struct io_kiocb *req)
5395 struct io_timeout_data *io = req->async_data;
5398 ret = hrtimer_try_to_cancel(&io->timer);
5401 list_del_init(&req->timeout.list);
5403 req_set_fail_links(req);
5404 io_cqring_fill_event(req, -ECANCELED);
5405 io_put_req_deferred(req, 1);
5409 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5411 struct io_kiocb *req;
5414 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5415 if (user_data == req->user_data) {
5424 return __io_timeout_cancel(req);
5427 static int io_timeout_remove_prep(struct io_kiocb *req,
5428 const struct io_uring_sqe *sqe)
5430 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5432 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5434 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5437 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5442 * Remove or update an existing timeout command
5444 static int io_timeout_remove(struct io_kiocb *req)
5446 struct io_ring_ctx *ctx = req->ctx;
5449 spin_lock_irq(&ctx->completion_lock);
5450 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5452 io_cqring_fill_event(req, ret);
5453 io_commit_cqring(ctx);
5454 spin_unlock_irq(&ctx->completion_lock);
5455 io_cqring_ev_posted(ctx);
5457 req_set_fail_links(req);
5462 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5463 bool is_timeout_link)
5465 struct io_timeout_data *data;
5467 u32 off = READ_ONCE(sqe->off);
5469 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5471 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5473 if (off && is_timeout_link)
5475 flags = READ_ONCE(sqe->timeout_flags);
5476 if (flags & ~IORING_TIMEOUT_ABS)
5479 req->timeout.off = off;
5481 if (!req->async_data && io_alloc_async_data(req))
5484 data = req->async_data;
5487 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5490 if (flags & IORING_TIMEOUT_ABS)
5491 data->mode = HRTIMER_MODE_ABS;
5493 data->mode = HRTIMER_MODE_REL;
5495 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5499 static int io_timeout(struct io_kiocb *req)
5501 struct io_ring_ctx *ctx = req->ctx;
5502 struct io_timeout_data *data = req->async_data;
5503 struct list_head *entry;
5504 u32 tail, off = req->timeout.off;
5506 spin_lock_irq(&ctx->completion_lock);
5509 * sqe->off holds how many events that need to occur for this
5510 * timeout event to be satisfied. If it isn't set, then this is
5511 * a pure timeout request, sequence isn't used.
5513 if (io_is_timeout_noseq(req)) {
5514 entry = ctx->timeout_list.prev;
5518 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5519 req->timeout.target_seq = tail + off;
5522 * Insertion sort, ensuring the first entry in the list is always
5523 * the one we need first.
5525 list_for_each_prev(entry, &ctx->timeout_list) {
5526 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5529 if (io_is_timeout_noseq(nxt))
5531 /* nxt.seq is behind @tail, otherwise would've been completed */
5532 if (off >= nxt->timeout.target_seq - tail)
5536 list_add(&req->timeout.list, entry);
5537 data->timer.function = io_timeout_fn;
5538 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5539 spin_unlock_irq(&ctx->completion_lock);
5543 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5545 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5547 return req->user_data == (unsigned long) data;
5550 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5552 enum io_wq_cancel cancel_ret;
5555 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5556 switch (cancel_ret) {
5557 case IO_WQ_CANCEL_OK:
5560 case IO_WQ_CANCEL_RUNNING:
5563 case IO_WQ_CANCEL_NOTFOUND:
5571 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5572 struct io_kiocb *req, __u64 sqe_addr,
5575 unsigned long flags;
5578 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5579 if (ret != -ENOENT) {
5580 spin_lock_irqsave(&ctx->completion_lock, flags);
5584 spin_lock_irqsave(&ctx->completion_lock, flags);
5585 ret = io_timeout_cancel(ctx, sqe_addr);
5588 ret = io_poll_cancel(ctx, sqe_addr);
5592 io_cqring_fill_event(req, ret);
5593 io_commit_cqring(ctx);
5594 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5595 io_cqring_ev_posted(ctx);
5598 req_set_fail_links(req);
5602 static int io_async_cancel_prep(struct io_kiocb *req,
5603 const struct io_uring_sqe *sqe)
5605 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5607 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5609 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5612 req->cancel.addr = READ_ONCE(sqe->addr);
5616 static int io_async_cancel(struct io_kiocb *req)
5618 struct io_ring_ctx *ctx = req->ctx;
5620 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5624 static int io_files_update_prep(struct io_kiocb *req,
5625 const struct io_uring_sqe *sqe)
5627 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5629 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5631 if (sqe->ioprio || sqe->rw_flags)
5634 req->files_update.offset = READ_ONCE(sqe->off);
5635 req->files_update.nr_args = READ_ONCE(sqe->len);
5636 if (!req->files_update.nr_args)
5638 req->files_update.arg = READ_ONCE(sqe->addr);
5642 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5643 struct io_comp_state *cs)
5645 struct io_ring_ctx *ctx = req->ctx;
5646 struct io_uring_files_update up;
5652 up.offset = req->files_update.offset;
5653 up.fds = req->files_update.arg;
5655 mutex_lock(&ctx->uring_lock);
5656 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5657 mutex_unlock(&ctx->uring_lock);
5660 req_set_fail_links(req);
5661 __io_req_complete(req, ret, 0, cs);
5665 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5667 switch (req->opcode) {
5670 case IORING_OP_READV:
5671 case IORING_OP_READ_FIXED:
5672 case IORING_OP_READ:
5673 return io_read_prep(req, sqe);
5674 case IORING_OP_WRITEV:
5675 case IORING_OP_WRITE_FIXED:
5676 case IORING_OP_WRITE:
5677 return io_write_prep(req, sqe);
5678 case IORING_OP_POLL_ADD:
5679 return io_poll_add_prep(req, sqe);
5680 case IORING_OP_POLL_REMOVE:
5681 return io_poll_remove_prep(req, sqe);
5682 case IORING_OP_FSYNC:
5683 return io_prep_fsync(req, sqe);
5684 case IORING_OP_SYNC_FILE_RANGE:
5685 return io_prep_sfr(req, sqe);
5686 case IORING_OP_SENDMSG:
5687 case IORING_OP_SEND:
5688 return io_sendmsg_prep(req, sqe);
5689 case IORING_OP_RECVMSG:
5690 case IORING_OP_RECV:
5691 return io_recvmsg_prep(req, sqe);
5692 case IORING_OP_CONNECT:
5693 return io_connect_prep(req, sqe);
5694 case IORING_OP_TIMEOUT:
5695 return io_timeout_prep(req, sqe, false);
5696 case IORING_OP_TIMEOUT_REMOVE:
5697 return io_timeout_remove_prep(req, sqe);
5698 case IORING_OP_ASYNC_CANCEL:
5699 return io_async_cancel_prep(req, sqe);
5700 case IORING_OP_LINK_TIMEOUT:
5701 return io_timeout_prep(req, sqe, true);
5702 case IORING_OP_ACCEPT:
5703 return io_accept_prep(req, sqe);
5704 case IORING_OP_FALLOCATE:
5705 return io_fallocate_prep(req, sqe);
5706 case IORING_OP_OPENAT:
5707 return io_openat_prep(req, sqe);
5708 case IORING_OP_CLOSE:
5709 return io_close_prep(req, sqe);
5710 case IORING_OP_FILES_UPDATE:
5711 return io_files_update_prep(req, sqe);
5712 case IORING_OP_STATX:
5713 return io_statx_prep(req, sqe);
5714 case IORING_OP_FADVISE:
5715 return io_fadvise_prep(req, sqe);
5716 case IORING_OP_MADVISE:
5717 return io_madvise_prep(req, sqe);
5718 case IORING_OP_OPENAT2:
5719 return io_openat2_prep(req, sqe);
5720 case IORING_OP_EPOLL_CTL:
5721 return io_epoll_ctl_prep(req, sqe);
5722 case IORING_OP_SPLICE:
5723 return io_splice_prep(req, sqe);
5724 case IORING_OP_PROVIDE_BUFFERS:
5725 return io_provide_buffers_prep(req, sqe);
5726 case IORING_OP_REMOVE_BUFFERS:
5727 return io_remove_buffers_prep(req, sqe);
5729 return io_tee_prep(req, sqe);
5732 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5737 static int io_req_defer_prep(struct io_kiocb *req,
5738 const struct io_uring_sqe *sqe)
5742 if (io_alloc_async_data(req))
5744 return io_req_prep(req, sqe);
5747 static u32 io_get_sequence(struct io_kiocb *req)
5749 struct io_kiocb *pos;
5750 struct io_ring_ctx *ctx = req->ctx;
5751 u32 total_submitted, nr_reqs = 1;
5753 if (req->flags & REQ_F_LINK_HEAD)
5754 list_for_each_entry(pos, &req->link_list, link_list)
5757 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5758 return total_submitted - nr_reqs;
5761 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5763 struct io_ring_ctx *ctx = req->ctx;
5764 struct io_defer_entry *de;
5768 /* Still need defer if there is pending req in defer list. */
5769 if (likely(list_empty_careful(&ctx->defer_list) &&
5770 !(req->flags & REQ_F_IO_DRAIN)))
5773 seq = io_get_sequence(req);
5774 /* Still a chance to pass the sequence check */
5775 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5778 if (!req->async_data) {
5779 ret = io_req_defer_prep(req, sqe);
5783 io_prep_async_link(req);
5784 de = kmalloc(sizeof(*de), GFP_KERNEL);
5788 spin_lock_irq(&ctx->completion_lock);
5789 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5790 spin_unlock_irq(&ctx->completion_lock);
5792 io_queue_async_work(req);
5793 return -EIOCBQUEUED;
5796 trace_io_uring_defer(ctx, req, req->user_data);
5799 list_add_tail(&de->list, &ctx->defer_list);
5800 spin_unlock_irq(&ctx->completion_lock);
5801 return -EIOCBQUEUED;
5804 static void io_req_drop_files(struct io_kiocb *req)
5806 struct io_ring_ctx *ctx = req->ctx;
5807 unsigned long flags;
5809 spin_lock_irqsave(&ctx->inflight_lock, flags);
5810 list_del(&req->inflight_entry);
5811 if (waitqueue_active(&ctx->inflight_wait))
5812 wake_up(&ctx->inflight_wait);
5813 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5814 req->flags &= ~REQ_F_INFLIGHT;
5815 put_files_struct(req->work.identity->files);
5816 put_nsproxy(req->work.identity->nsproxy);
5817 req->work.flags &= ~IO_WQ_WORK_FILES;
5820 static void __io_clean_op(struct io_kiocb *req)
5822 if (req->flags & REQ_F_BUFFER_SELECTED) {
5823 switch (req->opcode) {
5824 case IORING_OP_READV:
5825 case IORING_OP_READ_FIXED:
5826 case IORING_OP_READ:
5827 kfree((void *)(unsigned long)req->rw.addr);
5829 case IORING_OP_RECVMSG:
5830 case IORING_OP_RECV:
5831 kfree(req->sr_msg.kbuf);
5834 req->flags &= ~REQ_F_BUFFER_SELECTED;
5837 if (req->flags & REQ_F_NEED_CLEANUP) {
5838 switch (req->opcode) {
5839 case IORING_OP_READV:
5840 case IORING_OP_READ_FIXED:
5841 case IORING_OP_READ:
5842 case IORING_OP_WRITEV:
5843 case IORING_OP_WRITE_FIXED:
5844 case IORING_OP_WRITE: {
5845 struct io_async_rw *io = req->async_data;
5847 kfree(io->free_iovec);
5850 case IORING_OP_RECVMSG:
5851 case IORING_OP_SENDMSG: {
5852 struct io_async_msghdr *io = req->async_data;
5853 if (io->iov != io->fast_iov)
5857 case IORING_OP_SPLICE:
5859 io_put_file(req, req->splice.file_in,
5860 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5862 case IORING_OP_OPENAT:
5863 case IORING_OP_OPENAT2:
5864 if (req->open.filename)
5865 putname(req->open.filename);
5868 req->flags &= ~REQ_F_NEED_CLEANUP;
5871 if (req->flags & REQ_F_INFLIGHT)
5872 io_req_drop_files(req);
5875 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5876 struct io_comp_state *cs)
5878 struct io_ring_ctx *ctx = req->ctx;
5881 switch (req->opcode) {
5883 ret = io_nop(req, cs);
5885 case IORING_OP_READV:
5886 case IORING_OP_READ_FIXED:
5887 case IORING_OP_READ:
5888 ret = io_read(req, force_nonblock, cs);
5890 case IORING_OP_WRITEV:
5891 case IORING_OP_WRITE_FIXED:
5892 case IORING_OP_WRITE:
5893 ret = io_write(req, force_nonblock, cs);
5895 case IORING_OP_FSYNC:
5896 ret = io_fsync(req, force_nonblock);
5898 case IORING_OP_POLL_ADD:
5899 ret = io_poll_add(req);
5901 case IORING_OP_POLL_REMOVE:
5902 ret = io_poll_remove(req);
5904 case IORING_OP_SYNC_FILE_RANGE:
5905 ret = io_sync_file_range(req, force_nonblock);
5907 case IORING_OP_SENDMSG:
5908 ret = io_sendmsg(req, force_nonblock, cs);
5910 case IORING_OP_SEND:
5911 ret = io_send(req, force_nonblock, cs);
5913 case IORING_OP_RECVMSG:
5914 ret = io_recvmsg(req, force_nonblock, cs);
5916 case IORING_OP_RECV:
5917 ret = io_recv(req, force_nonblock, cs);
5919 case IORING_OP_TIMEOUT:
5920 ret = io_timeout(req);
5922 case IORING_OP_TIMEOUT_REMOVE:
5923 ret = io_timeout_remove(req);
5925 case IORING_OP_ACCEPT:
5926 ret = io_accept(req, force_nonblock, cs);
5928 case IORING_OP_CONNECT:
5929 ret = io_connect(req, force_nonblock, cs);
5931 case IORING_OP_ASYNC_CANCEL:
5932 ret = io_async_cancel(req);
5934 case IORING_OP_FALLOCATE:
5935 ret = io_fallocate(req, force_nonblock);
5937 case IORING_OP_OPENAT:
5938 ret = io_openat(req, force_nonblock);
5940 case IORING_OP_CLOSE:
5941 ret = io_close(req, force_nonblock, cs);
5943 case IORING_OP_FILES_UPDATE:
5944 ret = io_files_update(req, force_nonblock, cs);
5946 case IORING_OP_STATX:
5947 ret = io_statx(req, force_nonblock);
5949 case IORING_OP_FADVISE:
5950 ret = io_fadvise(req, force_nonblock);
5952 case IORING_OP_MADVISE:
5953 ret = io_madvise(req, force_nonblock);
5955 case IORING_OP_OPENAT2:
5956 ret = io_openat2(req, force_nonblock);
5958 case IORING_OP_EPOLL_CTL:
5959 ret = io_epoll_ctl(req, force_nonblock, cs);
5961 case IORING_OP_SPLICE:
5962 ret = io_splice(req, force_nonblock);
5964 case IORING_OP_PROVIDE_BUFFERS:
5965 ret = io_provide_buffers(req, force_nonblock, cs);
5967 case IORING_OP_REMOVE_BUFFERS:
5968 ret = io_remove_buffers(req, force_nonblock, cs);
5971 ret = io_tee(req, force_nonblock);
5981 /* If the op doesn't have a file, we're not polling for it */
5982 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5983 const bool in_async = io_wq_current_is_worker();
5985 /* workqueue context doesn't hold uring_lock, grab it now */
5987 mutex_lock(&ctx->uring_lock);
5989 io_iopoll_req_issued(req);
5992 mutex_unlock(&ctx->uring_lock);
5998 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6000 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6001 struct io_kiocb *timeout;
6004 timeout = io_prep_linked_timeout(req);
6006 io_queue_linked_timeout(timeout);
6008 /* if NO_CANCEL is set, we must still run the work */
6009 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6010 IO_WQ_WORK_CANCEL) {
6016 ret = io_issue_sqe(req, false, NULL);
6018 * We can get EAGAIN for polled IO even though we're
6019 * forcing a sync submission from here, since we can't
6020 * wait for request slots on the block side.
6029 req_set_fail_links(req);
6030 io_req_complete(req, ret);
6033 return io_steal_work(req);
6036 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6039 struct fixed_file_table *table;
6041 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6042 return table->files[index & IORING_FILE_TABLE_MASK];
6045 static struct file *io_file_get(struct io_submit_state *state,
6046 struct io_kiocb *req, int fd, bool fixed)
6048 struct io_ring_ctx *ctx = req->ctx;
6052 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6054 fd = array_index_nospec(fd, ctx->nr_user_files);
6055 file = io_file_from_index(ctx, fd);
6057 req->fixed_file_refs = &ctx->file_data->node->refs;
6058 percpu_ref_get(req->fixed_file_refs);
6061 trace_io_uring_file_get(ctx, fd);
6062 file = __io_file_get(state, fd);
6068 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6073 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6074 if (unlikely(!fixed && io_async_submit(req->ctx)))
6077 req->file = io_file_get(state, req, fd, fixed);
6078 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6083 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6085 struct io_timeout_data *data = container_of(timer,
6086 struct io_timeout_data, timer);
6087 struct io_kiocb *req = data->req;
6088 struct io_ring_ctx *ctx = req->ctx;
6089 struct io_kiocb *prev = NULL;
6090 unsigned long flags;
6092 spin_lock_irqsave(&ctx->completion_lock, flags);
6095 * We don't expect the list to be empty, that will only happen if we
6096 * race with the completion of the linked work.
6098 if (!list_empty(&req->link_list)) {
6099 prev = list_entry(req->link_list.prev, struct io_kiocb,
6101 if (refcount_inc_not_zero(&prev->refs))
6102 list_del_init(&req->link_list);
6107 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6110 req_set_fail_links(prev);
6111 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6114 io_req_complete(req, -ETIME);
6116 return HRTIMER_NORESTART;
6119 static void __io_queue_linked_timeout(struct io_kiocb *req)
6122 * If the list is now empty, then our linked request finished before
6123 * we got a chance to setup the timer
6125 if (!list_empty(&req->link_list)) {
6126 struct io_timeout_data *data = req->async_data;
6128 data->timer.function = io_link_timeout_fn;
6129 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6134 static void io_queue_linked_timeout(struct io_kiocb *req)
6136 struct io_ring_ctx *ctx = req->ctx;
6138 spin_lock_irq(&ctx->completion_lock);
6139 __io_queue_linked_timeout(req);
6140 spin_unlock_irq(&ctx->completion_lock);
6142 /* drop submission reference */
6146 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6148 struct io_kiocb *nxt;
6150 if (!(req->flags & REQ_F_LINK_HEAD))
6152 if (req->flags & REQ_F_LINK_TIMEOUT)
6155 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6157 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6160 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6161 req->flags |= REQ_F_LINK_TIMEOUT;
6165 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6167 struct io_kiocb *linked_timeout;
6168 const struct cred *old_creds = NULL;
6172 linked_timeout = io_prep_linked_timeout(req);
6174 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6175 (req->work.flags & IO_WQ_WORK_CREDS) &&
6176 req->work.identity->creds != current_cred()) {
6178 revert_creds(old_creds);
6179 if (old_creds == req->work.identity->creds)
6180 old_creds = NULL; /* restored original creds */
6182 old_creds = override_creds(req->work.identity->creds);
6185 ret = io_issue_sqe(req, true, cs);
6188 * We async punt it if the file wasn't marked NOWAIT, or if the file
6189 * doesn't support non-blocking read/write attempts
6191 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6192 if (!io_arm_poll_handler(req)) {
6194 * Queued up for async execution, worker will release
6195 * submit reference when the iocb is actually submitted.
6197 io_queue_async_work(req);
6201 io_queue_linked_timeout(linked_timeout);
6202 } else if (likely(!ret)) {
6203 /* drop submission reference */
6204 req = io_put_req_find_next(req);
6206 io_queue_linked_timeout(linked_timeout);
6209 if (!(req->flags & REQ_F_FORCE_ASYNC))
6211 io_queue_async_work(req);
6214 /* un-prep timeout, so it'll be killed as any other linked */
6215 req->flags &= ~REQ_F_LINK_TIMEOUT;
6216 req_set_fail_links(req);
6218 io_req_complete(req, ret);
6222 revert_creds(old_creds);
6225 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6226 struct io_comp_state *cs)
6230 ret = io_req_defer(req, sqe);
6232 if (ret != -EIOCBQUEUED) {
6234 req_set_fail_links(req);
6236 io_req_complete(req, ret);
6238 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6239 if (!req->async_data) {
6240 ret = io_req_defer_prep(req, sqe);
6244 io_queue_async_work(req);
6247 ret = io_req_prep(req, sqe);
6251 __io_queue_sqe(req, cs);
6255 static inline void io_queue_link_head(struct io_kiocb *req,
6256 struct io_comp_state *cs)
6258 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6260 io_req_complete(req, -ECANCELED);
6262 io_queue_sqe(req, NULL, cs);
6265 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6266 struct io_kiocb **link, struct io_comp_state *cs)
6268 struct io_ring_ctx *ctx = req->ctx;
6272 * If we already have a head request, queue this one for async
6273 * submittal once the head completes. If we don't have a head but
6274 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6275 * submitted sync once the chain is complete. If none of those
6276 * conditions are true (normal request), then just queue it.
6279 struct io_kiocb *head = *link;
6282 * Taking sequential execution of a link, draining both sides
6283 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6284 * requests in the link. So, it drains the head and the
6285 * next after the link request. The last one is done via
6286 * drain_next flag to persist the effect across calls.
6288 if (req->flags & REQ_F_IO_DRAIN) {
6289 head->flags |= REQ_F_IO_DRAIN;
6290 ctx->drain_next = 1;
6292 ret = io_req_defer_prep(req, sqe);
6293 if (unlikely(ret)) {
6294 /* fail even hard links since we don't submit */
6295 head->flags |= REQ_F_FAIL_LINK;
6298 trace_io_uring_link(ctx, req, head);
6299 list_add_tail(&req->link_list, &head->link_list);
6301 /* last request of a link, enqueue the link */
6302 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6303 io_queue_link_head(head, cs);
6307 if (unlikely(ctx->drain_next)) {
6308 req->flags |= REQ_F_IO_DRAIN;
6309 ctx->drain_next = 0;
6311 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6312 req->flags |= REQ_F_LINK_HEAD;
6313 INIT_LIST_HEAD(&req->link_list);
6315 ret = io_req_defer_prep(req, sqe);
6317 req->flags |= REQ_F_FAIL_LINK;
6320 io_queue_sqe(req, sqe, cs);
6328 * Batched submission is done, ensure local IO is flushed out.
6330 static void io_submit_state_end(struct io_submit_state *state)
6332 if (!list_empty(&state->comp.list))
6333 io_submit_flush_completions(&state->comp);
6334 blk_finish_plug(&state->plug);
6335 io_state_file_put(state);
6336 if (state->free_reqs)
6337 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6341 * Start submission side cache.
6343 static void io_submit_state_start(struct io_submit_state *state,
6344 struct io_ring_ctx *ctx, unsigned int max_ios)
6346 blk_start_plug(&state->plug);
6348 INIT_LIST_HEAD(&state->comp.list);
6349 state->comp.ctx = ctx;
6350 state->free_reqs = 0;
6352 state->ios_left = max_ios;
6355 static void io_commit_sqring(struct io_ring_ctx *ctx)
6357 struct io_rings *rings = ctx->rings;
6360 * Ensure any loads from the SQEs are done at this point,
6361 * since once we write the new head, the application could
6362 * write new data to them.
6364 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6368 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6369 * that is mapped by userspace. This means that care needs to be taken to
6370 * ensure that reads are stable, as we cannot rely on userspace always
6371 * being a good citizen. If members of the sqe are validated and then later
6372 * used, it's important that those reads are done through READ_ONCE() to
6373 * prevent a re-load down the line.
6375 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6377 u32 *sq_array = ctx->sq_array;
6381 * The cached sq head (or cq tail) serves two purposes:
6383 * 1) allows us to batch the cost of updating the user visible
6385 * 2) allows the kernel side to track the head on its own, even
6386 * though the application is the one updating it.
6388 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6389 if (likely(head < ctx->sq_entries))
6390 return &ctx->sq_sqes[head];
6392 /* drop invalid entries */
6393 ctx->cached_sq_dropped++;
6394 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6398 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6400 ctx->cached_sq_head++;
6404 * Check SQE restrictions (opcode and flags).
6406 * Returns 'true' if SQE is allowed, 'false' otherwise.
6408 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6409 struct io_kiocb *req,
6410 unsigned int sqe_flags)
6412 if (!ctx->restricted)
6415 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6418 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6419 ctx->restrictions.sqe_flags_required)
6422 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6423 ctx->restrictions.sqe_flags_required))
6429 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6430 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6431 IOSQE_BUFFER_SELECT)
6433 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6434 const struct io_uring_sqe *sqe,
6435 struct io_submit_state *state)
6437 unsigned int sqe_flags;
6440 req->opcode = READ_ONCE(sqe->opcode);
6441 req->user_data = READ_ONCE(sqe->user_data);
6442 req->async_data = NULL;
6446 /* one is dropped after submission, the other at completion */
6447 refcount_set(&req->refs, 2);
6448 req->task = current;
6451 if (unlikely(req->opcode >= IORING_OP_LAST))
6454 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6457 sqe_flags = READ_ONCE(sqe->flags);
6458 /* enforce forwards compatibility on users */
6459 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6462 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6465 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6466 !io_op_defs[req->opcode].buffer_select)
6469 id = READ_ONCE(sqe->personality);
6471 struct io_identity *iod;
6473 iod = idr_find(&ctx->personality_idr, id);
6476 refcount_inc(&iod->count);
6478 __io_req_init_async(req);
6479 get_cred(iod->creds);
6480 req->work.identity = iod;
6481 req->work.flags |= IO_WQ_WORK_CREDS;
6484 /* same numerical values with corresponding REQ_F_*, safe to copy */
6485 req->flags |= sqe_flags;
6487 if (!io_op_defs[req->opcode].needs_file)
6490 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6495 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6497 struct io_submit_state state;
6498 struct io_kiocb *link = NULL;
6499 int i, submitted = 0;
6501 /* if we have a backlog and couldn't flush it all, return BUSY */
6502 if (test_bit(0, &ctx->sq_check_overflow)) {
6503 if (!list_empty(&ctx->cq_overflow_list) &&
6504 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6508 /* make sure SQ entry isn't read before tail */
6509 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6511 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6514 percpu_counter_add(¤t->io_uring->inflight, nr);
6515 refcount_add(nr, ¤t->usage);
6517 io_submit_state_start(&state, ctx, nr);
6519 for (i = 0; i < nr; i++) {
6520 const struct io_uring_sqe *sqe;
6521 struct io_kiocb *req;
6524 sqe = io_get_sqe(ctx);
6525 if (unlikely(!sqe)) {
6526 io_consume_sqe(ctx);
6529 req = io_alloc_req(ctx, &state);
6530 if (unlikely(!req)) {
6532 submitted = -EAGAIN;
6535 io_consume_sqe(ctx);
6536 /* will complete beyond this point, count as submitted */
6539 err = io_init_req(ctx, req, sqe, &state);
6540 if (unlikely(err)) {
6543 io_req_complete(req, err);
6547 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6548 true, io_async_submit(ctx));
6549 err = io_submit_sqe(req, sqe, &link, &state.comp);
6554 if (unlikely(submitted != nr)) {
6555 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6556 struct io_uring_task *tctx = current->io_uring;
6557 int unused = nr - ref_used;
6559 percpu_ref_put_many(&ctx->refs, unused);
6560 percpu_counter_sub(&tctx->inflight, unused);
6561 put_task_struct_many(current, unused);
6564 io_queue_link_head(link, &state.comp);
6565 io_submit_state_end(&state);
6567 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6568 io_commit_sqring(ctx);
6573 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6575 /* Tell userspace we may need a wakeup call */
6576 spin_lock_irq(&ctx->completion_lock);
6577 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6578 spin_unlock_irq(&ctx->completion_lock);
6581 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
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 int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6589 int sync, void *key)
6591 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6594 ret = autoremove_wake_function(wqe, mode, sync, key);
6596 unsigned long flags;
6598 spin_lock_irqsave(&ctx->completion_lock, flags);
6599 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6600 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6611 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6612 unsigned long start_jiffies, bool cap_entries)
6614 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6615 struct io_sq_data *sqd = ctx->sq_data;
6616 unsigned int to_submit;
6620 if (!list_empty(&ctx->iopoll_list)) {
6621 unsigned nr_events = 0;
6623 mutex_lock(&ctx->uring_lock);
6624 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6625 io_do_iopoll(ctx, &nr_events, 0);
6626 mutex_unlock(&ctx->uring_lock);
6629 to_submit = io_sqring_entries(ctx);
6632 * If submit got -EBUSY, flag us as needing the application
6633 * to enter the kernel to reap and flush events.
6635 if (!to_submit || ret == -EBUSY || need_resched()) {
6637 * Drop cur_mm before scheduling, we can't hold it for
6638 * long periods (or over schedule()). Do this before
6639 * adding ourselves to the waitqueue, as the unuse/drop
6642 io_sq_thread_drop_mm();
6645 * We're polling. If we're within the defined idle
6646 * period, then let us spin without work before going
6647 * to sleep. The exception is if we got EBUSY doing
6648 * more IO, we should wait for the application to
6649 * reap events and wake us up.
6651 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6652 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6653 !percpu_ref_is_dying(&ctx->refs)))
6656 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6657 TASK_INTERRUPTIBLE);
6660 * While doing polled IO, before going to sleep, we need
6661 * to check if there are new reqs added to iopoll_list,
6662 * it is because reqs may have been punted to io worker
6663 * and will be added to iopoll_list later, hence check
6664 * the iopoll_list again.
6666 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6667 !list_empty_careful(&ctx->iopoll_list)) {
6668 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6672 to_submit = io_sqring_entries(ctx);
6673 if (!to_submit || ret == -EBUSY)
6677 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6678 io_ring_clear_wakeup_flag(ctx);
6680 /* if we're handling multiple rings, cap submit size for fairness */
6681 if (cap_entries && to_submit > 8)
6684 mutex_lock(&ctx->uring_lock);
6685 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6686 ret = io_submit_sqes(ctx, to_submit);
6687 mutex_unlock(&ctx->uring_lock);
6689 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6690 wake_up(&ctx->sqo_sq_wait);
6692 return SQT_DID_WORK;
6695 static void io_sqd_init_new(struct io_sq_data *sqd)
6697 struct io_ring_ctx *ctx;
6699 while (!list_empty(&sqd->ctx_new_list)) {
6700 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6701 init_wait(&ctx->sqo_wait_entry);
6702 ctx->sqo_wait_entry.func = io_sq_wake_function;
6703 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6704 complete(&ctx->sq_thread_comp);
6708 static int io_sq_thread(void *data)
6710 struct cgroup_subsys_state *cur_css = NULL;
6711 const struct cred *old_cred = NULL;
6712 struct io_sq_data *sqd = data;
6713 struct io_ring_ctx *ctx;
6714 unsigned long start_jiffies;
6716 start_jiffies = jiffies;
6717 while (!kthread_should_stop()) {
6718 enum sq_ret ret = 0;
6722 * Any changes to the sqd lists are synchronized through the
6723 * kthread parking. This synchronizes the thread vs users,
6724 * the users are synchronized on the sqd->ctx_lock.
6726 if (kthread_should_park())
6729 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6730 io_sqd_init_new(sqd);
6732 cap_entries = !list_is_singular(&sqd->ctx_list);
6734 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6735 if (current->cred != ctx->creds) {
6737 revert_creds(old_cred);
6738 old_cred = override_creds(ctx->creds);
6740 io_sq_thread_associate_blkcg(ctx, &cur_css);
6742 current->loginuid = ctx->loginuid;
6743 current->sessionid = ctx->sessionid;
6746 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6748 io_sq_thread_drop_mm();
6751 if (ret & SQT_SPIN) {
6754 } else if (ret == SQT_IDLE) {
6755 if (kthread_should_park())
6757 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6758 io_ring_set_wakeup_flag(ctx);
6760 start_jiffies = jiffies;
6761 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6762 io_ring_clear_wakeup_flag(ctx);
6769 io_sq_thread_unassociate_blkcg();
6771 revert_creds(old_cred);
6778 struct io_wait_queue {
6779 struct wait_queue_entry wq;
6780 struct io_ring_ctx *ctx;
6782 unsigned nr_timeouts;
6785 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6787 struct io_ring_ctx *ctx = iowq->ctx;
6790 * Wake up if we have enough events, or if a timeout occurred since we
6791 * started waiting. For timeouts, we always want to return to userspace,
6792 * regardless of event count.
6794 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6795 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6798 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6799 int wake_flags, void *key)
6801 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6804 /* use noflush == true, as we can't safely rely on locking context */
6805 if (!io_should_wake(iowq, true))
6808 return autoremove_wake_function(curr, mode, wake_flags, key);
6811 static int io_run_task_work_sig(void)
6813 if (io_run_task_work())
6815 if (!signal_pending(current))
6817 if (current->jobctl & JOBCTL_TASK_WORK) {
6818 spin_lock_irq(¤t->sighand->siglock);
6819 current->jobctl &= ~JOBCTL_TASK_WORK;
6820 recalc_sigpending();
6821 spin_unlock_irq(¤t->sighand->siglock);
6828 * Wait until events become available, if we don't already have some. The
6829 * application must reap them itself, as they reside on the shared cq ring.
6831 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6832 const sigset_t __user *sig, size_t sigsz)
6834 struct io_wait_queue iowq = {
6837 .func = io_wake_function,
6838 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6841 .to_wait = min_events,
6843 struct io_rings *rings = ctx->rings;
6847 if (io_cqring_events(ctx, false) >= min_events)
6849 if (!io_run_task_work())
6854 #ifdef CONFIG_COMPAT
6855 if (in_compat_syscall())
6856 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6860 ret = set_user_sigmask(sig, sigsz);
6866 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6867 trace_io_uring_cqring_wait(ctx, min_events);
6869 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6870 TASK_INTERRUPTIBLE);
6871 /* make sure we run task_work before checking for signals */
6872 ret = io_run_task_work_sig();
6877 if (io_should_wake(&iowq, false))
6881 finish_wait(&ctx->wait, &iowq.wq);
6883 restore_saved_sigmask_unless(ret == -EINTR);
6885 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6888 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6890 #if defined(CONFIG_UNIX)
6891 if (ctx->ring_sock) {
6892 struct sock *sock = ctx->ring_sock->sk;
6893 struct sk_buff *skb;
6895 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6901 for (i = 0; i < ctx->nr_user_files; i++) {
6904 file = io_file_from_index(ctx, i);
6911 static void io_file_ref_kill(struct percpu_ref *ref)
6913 struct fixed_file_data *data;
6915 data = container_of(ref, struct fixed_file_data, refs);
6916 complete(&data->done);
6919 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6921 struct fixed_file_data *data = ctx->file_data;
6922 struct fixed_file_ref_node *ref_node = NULL;
6923 unsigned nr_tables, i;
6928 spin_lock(&data->lock);
6929 if (!list_empty(&data->ref_list))
6930 ref_node = list_first_entry(&data->ref_list,
6931 struct fixed_file_ref_node, node);
6932 spin_unlock(&data->lock);
6934 percpu_ref_kill(&ref_node->refs);
6936 percpu_ref_kill(&data->refs);
6938 /* wait for all refs nodes to complete */
6939 flush_delayed_work(&ctx->file_put_work);
6940 wait_for_completion(&data->done);
6942 __io_sqe_files_unregister(ctx);
6943 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6944 for (i = 0; i < nr_tables; i++)
6945 kfree(data->table[i].files);
6947 percpu_ref_exit(&data->refs);
6949 ctx->file_data = NULL;
6950 ctx->nr_user_files = 0;
6954 static void io_put_sq_data(struct io_sq_data *sqd)
6956 if (refcount_dec_and_test(&sqd->refs)) {
6958 * The park is a bit of a work-around, without it we get
6959 * warning spews on shutdown with SQPOLL set and affinity
6960 * set to a single CPU.
6963 kthread_park(sqd->thread);
6964 kthread_stop(sqd->thread);
6971 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6973 struct io_ring_ctx *ctx_attach;
6974 struct io_sq_data *sqd;
6977 f = fdget(p->wq_fd);
6979 return ERR_PTR(-ENXIO);
6980 if (f.file->f_op != &io_uring_fops) {
6982 return ERR_PTR(-EINVAL);
6985 ctx_attach = f.file->private_data;
6986 sqd = ctx_attach->sq_data;
6989 return ERR_PTR(-EINVAL);
6992 refcount_inc(&sqd->refs);
6997 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6999 struct io_sq_data *sqd;
7001 if (p->flags & IORING_SETUP_ATTACH_WQ)
7002 return io_attach_sq_data(p);
7004 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7006 return ERR_PTR(-ENOMEM);
7008 refcount_set(&sqd->refs, 1);
7009 INIT_LIST_HEAD(&sqd->ctx_list);
7010 INIT_LIST_HEAD(&sqd->ctx_new_list);
7011 mutex_init(&sqd->ctx_lock);
7012 mutex_init(&sqd->lock);
7013 init_waitqueue_head(&sqd->wait);
7017 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7018 __releases(&sqd->lock)
7022 kthread_unpark(sqd->thread);
7023 mutex_unlock(&sqd->lock);
7026 static void io_sq_thread_park(struct io_sq_data *sqd)
7027 __acquires(&sqd->lock)
7031 mutex_lock(&sqd->lock);
7032 kthread_park(sqd->thread);
7035 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7037 struct io_sq_data *sqd = ctx->sq_data;
7042 * We may arrive here from the error branch in
7043 * io_sq_offload_create() where the kthread is created
7044 * without being waked up, thus wake it up now to make
7045 * sure the wait will complete.
7047 wake_up_process(sqd->thread);
7048 wait_for_completion(&ctx->sq_thread_comp);
7050 io_sq_thread_park(sqd);
7053 mutex_lock(&sqd->ctx_lock);
7054 list_del(&ctx->sqd_list);
7055 mutex_unlock(&sqd->ctx_lock);
7058 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7059 io_sq_thread_unpark(sqd);
7062 io_put_sq_data(sqd);
7063 ctx->sq_data = NULL;
7067 static void io_finish_async(struct io_ring_ctx *ctx)
7069 io_sq_thread_stop(ctx);
7072 io_wq_destroy(ctx->io_wq);
7077 #if defined(CONFIG_UNIX)
7079 * Ensure the UNIX gc is aware of our file set, so we are certain that
7080 * the io_uring can be safely unregistered on process exit, even if we have
7081 * loops in the file referencing.
7083 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7085 struct sock *sk = ctx->ring_sock->sk;
7086 struct scm_fp_list *fpl;
7087 struct sk_buff *skb;
7090 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7094 skb = alloc_skb(0, GFP_KERNEL);
7103 fpl->user = get_uid(ctx->user);
7104 for (i = 0; i < nr; i++) {
7105 struct file *file = io_file_from_index(ctx, i + offset);
7109 fpl->fp[nr_files] = get_file(file);
7110 unix_inflight(fpl->user, fpl->fp[nr_files]);
7115 fpl->max = SCM_MAX_FD;
7116 fpl->count = nr_files;
7117 UNIXCB(skb).fp = fpl;
7118 skb->destructor = unix_destruct_scm;
7119 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7120 skb_queue_head(&sk->sk_receive_queue, skb);
7122 for (i = 0; i < nr_files; i++)
7133 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7134 * causes regular reference counting to break down. We rely on the UNIX
7135 * garbage collection to take care of this problem for us.
7137 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7139 unsigned left, total;
7143 left = ctx->nr_user_files;
7145 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7147 ret = __io_sqe_files_scm(ctx, this_files, total);
7151 total += this_files;
7157 while (total < ctx->nr_user_files) {
7158 struct file *file = io_file_from_index(ctx, total);
7168 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7174 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7175 unsigned nr_tables, unsigned nr_files)
7179 for (i = 0; i < nr_tables; i++) {
7180 struct fixed_file_table *table = &file_data->table[i];
7181 unsigned this_files;
7183 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7184 table->files = kcalloc(this_files, sizeof(struct file *),
7188 nr_files -= this_files;
7194 for (i = 0; i < nr_tables; i++) {
7195 struct fixed_file_table *table = &file_data->table[i];
7196 kfree(table->files);
7201 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7203 #if defined(CONFIG_UNIX)
7204 struct sock *sock = ctx->ring_sock->sk;
7205 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7206 struct sk_buff *skb;
7209 __skb_queue_head_init(&list);
7212 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7213 * remove this entry and rearrange the file array.
7215 skb = skb_dequeue(head);
7217 struct scm_fp_list *fp;
7219 fp = UNIXCB(skb).fp;
7220 for (i = 0; i < fp->count; i++) {
7223 if (fp->fp[i] != file)
7226 unix_notinflight(fp->user, fp->fp[i]);
7227 left = fp->count - 1 - i;
7229 memmove(&fp->fp[i], &fp->fp[i + 1],
7230 left * sizeof(struct file *));
7237 __skb_queue_tail(&list, skb);
7247 __skb_queue_tail(&list, skb);
7249 skb = skb_dequeue(head);
7252 if (skb_peek(&list)) {
7253 spin_lock_irq(&head->lock);
7254 while ((skb = __skb_dequeue(&list)) != NULL)
7255 __skb_queue_tail(head, skb);
7256 spin_unlock_irq(&head->lock);
7263 struct io_file_put {
7264 struct list_head list;
7268 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7270 struct fixed_file_data *file_data = ref_node->file_data;
7271 struct io_ring_ctx *ctx = file_data->ctx;
7272 struct io_file_put *pfile, *tmp;
7274 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7275 list_del(&pfile->list);
7276 io_ring_file_put(ctx, pfile->file);
7280 spin_lock(&file_data->lock);
7281 list_del(&ref_node->node);
7282 spin_unlock(&file_data->lock);
7284 percpu_ref_exit(&ref_node->refs);
7286 percpu_ref_put(&file_data->refs);
7289 static void io_file_put_work(struct work_struct *work)
7291 struct io_ring_ctx *ctx;
7292 struct llist_node *node;
7294 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7295 node = llist_del_all(&ctx->file_put_llist);
7298 struct fixed_file_ref_node *ref_node;
7299 struct llist_node *next = node->next;
7301 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7302 __io_file_put_work(ref_node);
7307 static void io_file_data_ref_zero(struct percpu_ref *ref)
7309 struct fixed_file_ref_node *ref_node;
7310 struct io_ring_ctx *ctx;
7314 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7315 ctx = ref_node->file_data->ctx;
7317 if (percpu_ref_is_dying(&ctx->file_data->refs))
7320 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7322 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7324 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7327 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7328 struct io_ring_ctx *ctx)
7330 struct fixed_file_ref_node *ref_node;
7332 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7334 return ERR_PTR(-ENOMEM);
7336 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7339 return ERR_PTR(-ENOMEM);
7341 INIT_LIST_HEAD(&ref_node->node);
7342 INIT_LIST_HEAD(&ref_node->file_list);
7343 ref_node->file_data = ctx->file_data;
7347 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7349 percpu_ref_exit(&ref_node->refs);
7353 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7356 __s32 __user *fds = (__s32 __user *) arg;
7357 unsigned nr_tables, i;
7359 int fd, ret = -ENOMEM;
7360 struct fixed_file_ref_node *ref_node;
7361 struct fixed_file_data *file_data;
7367 if (nr_args > IORING_MAX_FIXED_FILES)
7370 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7373 file_data->ctx = ctx;
7374 init_completion(&file_data->done);
7375 INIT_LIST_HEAD(&file_data->ref_list);
7376 spin_lock_init(&file_data->lock);
7378 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7379 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7381 if (!file_data->table)
7384 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7385 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7388 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7390 ctx->file_data = file_data;
7392 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7393 struct fixed_file_table *table;
7396 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7400 /* allow sparse sets */
7410 * Don't allow io_uring instances to be registered. If UNIX
7411 * isn't enabled, then this causes a reference cycle and this
7412 * instance can never get freed. If UNIX is enabled we'll
7413 * handle it just fine, but there's still no point in allowing
7414 * a ring fd as it doesn't support regular read/write anyway.
7416 if (file->f_op == &io_uring_fops) {
7420 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7421 index = i & IORING_FILE_TABLE_MASK;
7422 table->files[index] = file;
7425 ret = io_sqe_files_scm(ctx);
7427 io_sqe_files_unregister(ctx);
7431 ref_node = alloc_fixed_file_ref_node(ctx);
7432 if (IS_ERR(ref_node)) {
7433 io_sqe_files_unregister(ctx);
7434 return PTR_ERR(ref_node);
7437 file_data->node = ref_node;
7438 spin_lock(&file_data->lock);
7439 list_add(&ref_node->node, &file_data->ref_list);
7440 spin_unlock(&file_data->lock);
7441 percpu_ref_get(&file_data->refs);
7444 for (i = 0; i < ctx->nr_user_files; i++) {
7445 file = io_file_from_index(ctx, i);
7449 for (i = 0; i < nr_tables; i++)
7450 kfree(file_data->table[i].files);
7451 ctx->nr_user_files = 0;
7453 percpu_ref_exit(&file_data->refs);
7455 kfree(file_data->table);
7457 ctx->file_data = NULL;
7461 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7464 #if defined(CONFIG_UNIX)
7465 struct sock *sock = ctx->ring_sock->sk;
7466 struct sk_buff_head *head = &sock->sk_receive_queue;
7467 struct sk_buff *skb;
7470 * See if we can merge this file into an existing skb SCM_RIGHTS
7471 * file set. If there's no room, fall back to allocating a new skb
7472 * and filling it in.
7474 spin_lock_irq(&head->lock);
7475 skb = skb_peek(head);
7477 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7479 if (fpl->count < SCM_MAX_FD) {
7480 __skb_unlink(skb, head);
7481 spin_unlock_irq(&head->lock);
7482 fpl->fp[fpl->count] = get_file(file);
7483 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7485 spin_lock_irq(&head->lock);
7486 __skb_queue_head(head, skb);
7491 spin_unlock_irq(&head->lock);
7498 return __io_sqe_files_scm(ctx, 1, index);
7504 static int io_queue_file_removal(struct fixed_file_data *data,
7507 struct io_file_put *pfile;
7508 struct fixed_file_ref_node *ref_node = data->node;
7510 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7515 list_add(&pfile->list, &ref_node->file_list);
7520 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7521 struct io_uring_files_update *up,
7524 struct fixed_file_data *data = ctx->file_data;
7525 struct fixed_file_ref_node *ref_node;
7530 bool needs_switch = false;
7532 if (check_add_overflow(up->offset, nr_args, &done))
7534 if (done > ctx->nr_user_files)
7537 ref_node = alloc_fixed_file_ref_node(ctx);
7538 if (IS_ERR(ref_node))
7539 return PTR_ERR(ref_node);
7542 fds = u64_to_user_ptr(up->fds);
7544 struct fixed_file_table *table;
7548 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7552 i = array_index_nospec(up->offset, ctx->nr_user_files);
7553 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7554 index = i & IORING_FILE_TABLE_MASK;
7555 if (table->files[index]) {
7556 file = table->files[index];
7557 err = io_queue_file_removal(data, file);
7560 table->files[index] = NULL;
7561 needs_switch = true;
7570 * Don't allow io_uring instances to be registered. If
7571 * UNIX isn't enabled, then this causes a reference
7572 * cycle and this instance can never get freed. If UNIX
7573 * is enabled we'll handle it just fine, but there's
7574 * still no point in allowing a ring fd as it doesn't
7575 * support regular read/write anyway.
7577 if (file->f_op == &io_uring_fops) {
7582 table->files[index] = file;
7583 err = io_sqe_file_register(ctx, file, i);
7585 table->files[index] = NULL;
7596 percpu_ref_kill(&data->node->refs);
7597 spin_lock(&data->lock);
7598 list_add(&ref_node->node, &data->ref_list);
7599 data->node = ref_node;
7600 spin_unlock(&data->lock);
7601 percpu_ref_get(&ctx->file_data->refs);
7603 destroy_fixed_file_ref_node(ref_node);
7605 return done ? done : err;
7608 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7611 struct io_uring_files_update up;
7613 if (!ctx->file_data)
7617 if (copy_from_user(&up, arg, sizeof(up)))
7622 return __io_sqe_files_update(ctx, &up, nr_args);
7625 static void io_free_work(struct io_wq_work *work)
7627 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7629 /* Consider that io_steal_work() relies on this ref */
7633 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7634 struct io_uring_params *p)
7636 struct io_wq_data data;
7638 struct io_ring_ctx *ctx_attach;
7639 unsigned int concurrency;
7642 data.user = ctx->user;
7643 data.free_work = io_free_work;
7644 data.do_work = io_wq_submit_work;
7646 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7647 /* Do QD, or 4 * CPUS, whatever is smallest */
7648 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7650 ctx->io_wq = io_wq_create(concurrency, &data);
7651 if (IS_ERR(ctx->io_wq)) {
7652 ret = PTR_ERR(ctx->io_wq);
7658 f = fdget(p->wq_fd);
7662 if (f.file->f_op != &io_uring_fops) {
7667 ctx_attach = f.file->private_data;
7668 /* @io_wq is protected by holding the fd */
7669 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7674 ctx->io_wq = ctx_attach->io_wq;
7680 static int io_uring_alloc_task_context(struct task_struct *task)
7682 struct io_uring_task *tctx;
7685 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7686 if (unlikely(!tctx))
7689 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7690 if (unlikely(ret)) {
7696 init_waitqueue_head(&tctx->wait);
7699 io_init_identity(&tctx->__identity);
7700 tctx->identity = &tctx->__identity;
7701 task->io_uring = tctx;
7705 void __io_uring_free(struct task_struct *tsk)
7707 struct io_uring_task *tctx = tsk->io_uring;
7709 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7710 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7711 if (tctx->identity != &tctx->__identity)
7712 kfree(tctx->identity);
7713 percpu_counter_destroy(&tctx->inflight);
7715 tsk->io_uring = NULL;
7718 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7719 struct io_uring_params *p)
7723 if (ctx->flags & IORING_SETUP_SQPOLL) {
7724 struct io_sq_data *sqd;
7727 if (!capable(CAP_SYS_ADMIN))
7730 sqd = io_get_sq_data(p);
7737 io_sq_thread_park(sqd);
7738 mutex_lock(&sqd->ctx_lock);
7739 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7740 mutex_unlock(&sqd->ctx_lock);
7741 io_sq_thread_unpark(sqd);
7743 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7744 if (!ctx->sq_thread_idle)
7745 ctx->sq_thread_idle = HZ;
7750 if (p->flags & IORING_SETUP_SQ_AFF) {
7751 int cpu = p->sq_thread_cpu;
7754 if (cpu >= nr_cpu_ids)
7756 if (!cpu_online(cpu))
7759 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7760 cpu, "io_uring-sq");
7762 sqd->thread = kthread_create(io_sq_thread, sqd,
7765 if (IS_ERR(sqd->thread)) {
7766 ret = PTR_ERR(sqd->thread);
7770 ret = io_uring_alloc_task_context(sqd->thread);
7773 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7774 /* Can't have SQ_AFF without SQPOLL */
7780 ret = io_init_wq_offload(ctx, p);
7786 io_finish_async(ctx);
7790 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7792 struct io_sq_data *sqd = ctx->sq_data;
7794 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7795 wake_up_process(sqd->thread);
7798 static inline void __io_unaccount_mem(struct user_struct *user,
7799 unsigned long nr_pages)
7801 atomic_long_sub(nr_pages, &user->locked_vm);
7804 static inline int __io_account_mem(struct user_struct *user,
7805 unsigned long nr_pages)
7807 unsigned long page_limit, cur_pages, new_pages;
7809 /* Don't allow more pages than we can safely lock */
7810 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7813 cur_pages = atomic_long_read(&user->locked_vm);
7814 new_pages = cur_pages + nr_pages;
7815 if (new_pages > page_limit)
7817 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7818 new_pages) != cur_pages);
7823 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7824 enum io_mem_account acct)
7827 __io_unaccount_mem(ctx->user, nr_pages);
7829 if (ctx->mm_account) {
7830 if (acct == ACCT_LOCKED)
7831 ctx->mm_account->locked_vm -= nr_pages;
7832 else if (acct == ACCT_PINNED)
7833 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7837 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7838 enum io_mem_account acct)
7842 if (ctx->limit_mem) {
7843 ret = __io_account_mem(ctx->user, nr_pages);
7848 if (ctx->mm_account) {
7849 if (acct == ACCT_LOCKED)
7850 ctx->mm_account->locked_vm += nr_pages;
7851 else if (acct == ACCT_PINNED)
7852 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7858 static void io_mem_free(void *ptr)
7865 page = virt_to_head_page(ptr);
7866 if (put_page_testzero(page))
7867 free_compound_page(page);
7870 static void *io_mem_alloc(size_t size)
7872 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7875 return (void *) __get_free_pages(gfp_flags, get_order(size));
7878 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7881 struct io_rings *rings;
7882 size_t off, sq_array_size;
7884 off = struct_size(rings, cqes, cq_entries);
7885 if (off == SIZE_MAX)
7889 off = ALIGN(off, SMP_CACHE_BYTES);
7897 sq_array_size = array_size(sizeof(u32), sq_entries);
7898 if (sq_array_size == SIZE_MAX)
7901 if (check_add_overflow(off, sq_array_size, &off))
7907 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7911 pages = (size_t)1 << get_order(
7912 rings_size(sq_entries, cq_entries, NULL));
7913 pages += (size_t)1 << get_order(
7914 array_size(sizeof(struct io_uring_sqe), sq_entries));
7919 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7923 if (!ctx->user_bufs)
7926 for (i = 0; i < ctx->nr_user_bufs; i++) {
7927 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7929 for (j = 0; j < imu->nr_bvecs; j++)
7930 unpin_user_page(imu->bvec[j].bv_page);
7932 if (imu->acct_pages)
7933 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7938 kfree(ctx->user_bufs);
7939 ctx->user_bufs = NULL;
7940 ctx->nr_user_bufs = 0;
7944 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7945 void __user *arg, unsigned index)
7947 struct iovec __user *src;
7949 #ifdef CONFIG_COMPAT
7951 struct compat_iovec __user *ciovs;
7952 struct compat_iovec ciov;
7954 ciovs = (struct compat_iovec __user *) arg;
7955 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7958 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7959 dst->iov_len = ciov.iov_len;
7963 src = (struct iovec __user *) arg;
7964 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7970 * Not super efficient, but this is just a registration time. And we do cache
7971 * the last compound head, so generally we'll only do a full search if we don't
7974 * We check if the given compound head page has already been accounted, to
7975 * avoid double accounting it. This allows us to account the full size of the
7976 * page, not just the constituent pages of a huge page.
7978 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7979 int nr_pages, struct page *hpage)
7983 /* check current page array */
7984 for (i = 0; i < nr_pages; i++) {
7985 if (!PageCompound(pages[i]))
7987 if (compound_head(pages[i]) == hpage)
7991 /* check previously registered pages */
7992 for (i = 0; i < ctx->nr_user_bufs; i++) {
7993 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7995 for (j = 0; j < imu->nr_bvecs; j++) {
7996 if (!PageCompound(imu->bvec[j].bv_page))
7998 if (compound_head(imu->bvec[j].bv_page) == hpage)
8006 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8007 int nr_pages, struct io_mapped_ubuf *imu,
8008 struct page **last_hpage)
8012 for (i = 0; i < nr_pages; i++) {
8013 if (!PageCompound(pages[i])) {
8018 hpage = compound_head(pages[i]);
8019 if (hpage == *last_hpage)
8021 *last_hpage = hpage;
8022 if (headpage_already_acct(ctx, pages, i, hpage))
8024 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8028 if (!imu->acct_pages)
8031 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8033 imu->acct_pages = 0;
8037 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8040 struct vm_area_struct **vmas = NULL;
8041 struct page **pages = NULL;
8042 struct page *last_hpage = NULL;
8043 int i, j, got_pages = 0;
8048 if (!nr_args || nr_args > UIO_MAXIOV)
8051 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8053 if (!ctx->user_bufs)
8056 for (i = 0; i < nr_args; i++) {
8057 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8058 unsigned long off, start, end, ubuf;
8063 ret = io_copy_iov(ctx, &iov, arg, i);
8068 * Don't impose further limits on the size and buffer
8069 * constraints here, we'll -EINVAL later when IO is
8070 * submitted if they are wrong.
8073 if (!iov.iov_base || !iov.iov_len)
8076 /* arbitrary limit, but we need something */
8077 if (iov.iov_len > SZ_1G)
8080 ubuf = (unsigned long) iov.iov_base;
8081 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8082 start = ubuf >> PAGE_SHIFT;
8083 nr_pages = end - start;
8086 if (!pages || nr_pages > got_pages) {
8089 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8091 vmas = kvmalloc_array(nr_pages,
8092 sizeof(struct vm_area_struct *),
8094 if (!pages || !vmas) {
8098 got_pages = nr_pages;
8101 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8108 mmap_read_lock(current->mm);
8109 pret = pin_user_pages(ubuf, nr_pages,
8110 FOLL_WRITE | FOLL_LONGTERM,
8112 if (pret == nr_pages) {
8113 /* don't support file backed memory */
8114 for (j = 0; j < nr_pages; j++) {
8115 struct vm_area_struct *vma = vmas[j];
8118 !is_file_hugepages(vma->vm_file)) {
8124 ret = pret < 0 ? pret : -EFAULT;
8126 mmap_read_unlock(current->mm);
8129 * if we did partial map, or found file backed vmas,
8130 * release any pages we did get
8133 unpin_user_pages(pages, pret);
8138 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8140 unpin_user_pages(pages, pret);
8145 off = ubuf & ~PAGE_MASK;
8147 for (j = 0; j < nr_pages; j++) {
8150 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8151 imu->bvec[j].bv_page = pages[j];
8152 imu->bvec[j].bv_len = vec_len;
8153 imu->bvec[j].bv_offset = off;
8157 /* store original address for later verification */
8159 imu->len = iov.iov_len;
8160 imu->nr_bvecs = nr_pages;
8162 ctx->nr_user_bufs++;
8170 io_sqe_buffer_unregister(ctx);
8174 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8176 __s32 __user *fds = arg;
8182 if (copy_from_user(&fd, fds, sizeof(*fds)))
8185 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8186 if (IS_ERR(ctx->cq_ev_fd)) {
8187 int ret = PTR_ERR(ctx->cq_ev_fd);
8188 ctx->cq_ev_fd = NULL;
8195 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8197 if (ctx->cq_ev_fd) {
8198 eventfd_ctx_put(ctx->cq_ev_fd);
8199 ctx->cq_ev_fd = NULL;
8206 static int __io_destroy_buffers(int id, void *p, void *data)
8208 struct io_ring_ctx *ctx = data;
8209 struct io_buffer *buf = p;
8211 __io_remove_buffers(ctx, buf, id, -1U);
8215 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8217 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8218 idr_destroy(&ctx->io_buffer_idr);
8221 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8223 io_finish_async(ctx);
8224 io_sqe_buffer_unregister(ctx);
8226 if (ctx->sqo_task) {
8227 put_task_struct(ctx->sqo_task);
8228 ctx->sqo_task = NULL;
8229 mmdrop(ctx->mm_account);
8230 ctx->mm_account = NULL;
8233 #ifdef CONFIG_BLK_CGROUP
8234 if (ctx->sqo_blkcg_css)
8235 css_put(ctx->sqo_blkcg_css);
8238 io_sqe_files_unregister(ctx);
8239 io_eventfd_unregister(ctx);
8240 io_destroy_buffers(ctx);
8241 idr_destroy(&ctx->personality_idr);
8243 #if defined(CONFIG_UNIX)
8244 if (ctx->ring_sock) {
8245 ctx->ring_sock->file = NULL; /* so that iput() is called */
8246 sock_release(ctx->ring_sock);
8250 io_mem_free(ctx->rings);
8251 io_mem_free(ctx->sq_sqes);
8253 percpu_ref_exit(&ctx->refs);
8254 free_uid(ctx->user);
8255 put_cred(ctx->creds);
8256 kfree(ctx->cancel_hash);
8257 kmem_cache_free(req_cachep, ctx->fallback_req);
8261 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8263 struct io_ring_ctx *ctx = file->private_data;
8266 poll_wait(file, &ctx->cq_wait, wait);
8268 * synchronizes with barrier from wq_has_sleeper call in
8272 if (!io_sqring_full(ctx))
8273 mask |= EPOLLOUT | EPOLLWRNORM;
8274 if (io_cqring_events(ctx, false))
8275 mask |= EPOLLIN | EPOLLRDNORM;
8280 static int io_uring_fasync(int fd, struct file *file, int on)
8282 struct io_ring_ctx *ctx = file->private_data;
8284 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8287 static int io_remove_personalities(int id, void *p, void *data)
8289 struct io_ring_ctx *ctx = data;
8290 struct io_identity *iod;
8292 iod = idr_remove(&ctx->personality_idr, id);
8294 put_cred(iod->creds);
8295 if (refcount_dec_and_test(&iod->count))
8301 static void io_ring_exit_work(struct work_struct *work)
8303 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8307 * If we're doing polled IO and end up having requests being
8308 * submitted async (out-of-line), then completions can come in while
8309 * we're waiting for refs to drop. We need to reap these manually,
8310 * as nobody else will be looking for them.
8314 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8315 io_iopoll_try_reap_events(ctx);
8316 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8317 io_ring_ctx_free(ctx);
8320 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8322 mutex_lock(&ctx->uring_lock);
8323 percpu_ref_kill(&ctx->refs);
8324 mutex_unlock(&ctx->uring_lock);
8326 io_kill_timeouts(ctx, NULL);
8327 io_poll_remove_all(ctx, NULL);
8330 io_wq_cancel_all(ctx->io_wq);
8332 /* if we failed setting up the ctx, we might not have any rings */
8334 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8335 io_iopoll_try_reap_events(ctx);
8336 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8339 * Do this upfront, so we won't have a grace period where the ring
8340 * is closed but resources aren't reaped yet. This can cause
8341 * spurious failure in setting up a new ring.
8343 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8346 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8348 * Use system_unbound_wq to avoid spawning tons of event kworkers
8349 * if we're exiting a ton of rings at the same time. It just adds
8350 * noise and overhead, there's no discernable change in runtime
8351 * over using system_wq.
8353 queue_work(system_unbound_wq, &ctx->exit_work);
8356 static int io_uring_release(struct inode *inode, struct file *file)
8358 struct io_ring_ctx *ctx = file->private_data;
8360 file->private_data = NULL;
8361 io_ring_ctx_wait_and_kill(ctx);
8365 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8367 struct files_struct *files = data;
8369 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8370 work->identity->files == files);
8374 * Returns true if 'preq' is the link parent of 'req'
8376 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8378 struct io_kiocb *link;
8380 if (!(preq->flags & REQ_F_LINK_HEAD))
8383 list_for_each_entry(link, &preq->link_list, link_list) {
8391 static bool io_match_link_files(struct io_kiocb *req,
8392 struct files_struct *files)
8394 struct io_kiocb *link;
8396 if (io_match_files(req, files))
8398 if (req->flags & REQ_F_LINK_HEAD) {
8399 list_for_each_entry(link, &req->link_list, link_list) {
8400 if (io_match_files(link, files))
8408 * We're looking to cancel 'req' because it's holding on to our files, but
8409 * 'req' could be a link to another request. See if it is, and cancel that
8410 * parent request if so.
8412 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8414 struct hlist_node *tmp;
8415 struct io_kiocb *preq;
8419 spin_lock_irq(&ctx->completion_lock);
8420 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8421 struct hlist_head *list;
8423 list = &ctx->cancel_hash[i];
8424 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8425 found = io_match_link(preq, req);
8427 io_poll_remove_one(preq);
8432 spin_unlock_irq(&ctx->completion_lock);
8436 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8437 struct io_kiocb *req)
8439 struct io_kiocb *preq;
8442 spin_lock_irq(&ctx->completion_lock);
8443 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8444 found = io_match_link(preq, req);
8446 __io_timeout_cancel(preq);
8450 spin_unlock_irq(&ctx->completion_lock);
8454 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8456 return io_match_link(container_of(work, struct io_kiocb, work), data);
8459 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8461 enum io_wq_cancel cret;
8463 /* cancel this particular work, if it's running */
8464 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8465 if (cret != IO_WQ_CANCEL_NOTFOUND)
8468 /* find links that hold this pending, cancel those */
8469 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8470 if (cret != IO_WQ_CANCEL_NOTFOUND)
8473 /* if we have a poll link holding this pending, cancel that */
8474 if (io_poll_remove_link(ctx, req))
8477 /* final option, timeout link is holding this req pending */
8478 io_timeout_remove_link(ctx, req);
8481 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8482 struct files_struct *files)
8484 struct io_defer_entry *de = NULL;
8487 spin_lock_irq(&ctx->completion_lock);
8488 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8489 if (io_match_link_files(de->req, files)) {
8490 list_cut_position(&list, &ctx->defer_list, &de->list);
8494 spin_unlock_irq(&ctx->completion_lock);
8496 while (!list_empty(&list)) {
8497 de = list_first_entry(&list, struct io_defer_entry, list);
8498 list_del_init(&de->list);
8499 req_set_fail_links(de->req);
8500 io_put_req(de->req);
8501 io_req_complete(de->req, -ECANCELED);
8507 * Returns true if we found and killed one or more files pinning requests
8509 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8510 struct files_struct *files)
8512 if (list_empty_careful(&ctx->inflight_list))
8515 io_cancel_defer_files(ctx, files);
8516 /* cancel all at once, should be faster than doing it one by one*/
8517 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8519 while (!list_empty_careful(&ctx->inflight_list)) {
8520 struct io_kiocb *cancel_req = NULL, *req;
8523 spin_lock_irq(&ctx->inflight_lock);
8524 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8525 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8526 req->work.identity->files != files)
8528 /* req is being completed, ignore */
8529 if (!refcount_inc_not_zero(&req->refs))
8535 prepare_to_wait(&ctx->inflight_wait, &wait,
8536 TASK_UNINTERRUPTIBLE);
8537 spin_unlock_irq(&ctx->inflight_lock);
8539 /* We need to keep going until we don't find a matching req */
8542 /* cancel this request, or head link requests */
8543 io_attempt_cancel(ctx, cancel_req);
8544 io_put_req(cancel_req);
8545 /* cancellations _may_ trigger task work */
8548 finish_wait(&ctx->inflight_wait, &wait);
8554 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8556 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8557 struct task_struct *task = data;
8559 return io_task_match(req, task);
8562 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8563 struct task_struct *task,
8564 struct files_struct *files)
8568 ret = io_uring_cancel_files(ctx, files);
8570 enum io_wq_cancel cret;
8572 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8573 if (cret != IO_WQ_CANCEL_NOTFOUND)
8576 /* SQPOLL thread does its own polling */
8577 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8578 while (!list_empty_careful(&ctx->iopoll_list)) {
8579 io_iopoll_try_reap_events(ctx);
8584 ret |= io_poll_remove_all(ctx, task);
8585 ret |= io_kill_timeouts(ctx, task);
8592 * We need to iteratively cancel requests, in case a request has dependent
8593 * hard links. These persist even for failure of cancelations, hence keep
8594 * looping until none are found.
8596 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8597 struct files_struct *files)
8599 struct task_struct *task = current;
8601 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8602 task = ctx->sq_data->thread;
8604 io_cqring_overflow_flush(ctx, true, task, files);
8606 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8613 * Note that this task has used io_uring. We use it for cancelation purposes.
8615 static int io_uring_add_task_file(struct file *file)
8617 struct io_uring_task *tctx = current->io_uring;
8619 if (unlikely(!tctx)) {
8622 ret = io_uring_alloc_task_context(current);
8625 tctx = current->io_uring;
8627 if (tctx->last != file) {
8628 void *old = xa_load(&tctx->xa, (unsigned long)file);
8632 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8641 * Remove this io_uring_file -> task mapping.
8643 static void io_uring_del_task_file(struct file *file)
8645 struct io_uring_task *tctx = current->io_uring;
8647 if (tctx->last == file)
8649 file = xa_erase(&tctx->xa, (unsigned long)file);
8655 * Drop task note for this file if we're the only ones that hold it after
8658 static void io_uring_attempt_task_drop(struct file *file)
8660 if (!current->io_uring)
8663 * fput() is pending, will be 2 if the only other ref is our potential
8664 * task file note. If the task is exiting, drop regardless of count.
8666 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8667 atomic_long_read(&file->f_count) == 2)
8668 io_uring_del_task_file(file);
8671 void __io_uring_files_cancel(struct files_struct *files)
8673 struct io_uring_task *tctx = current->io_uring;
8675 unsigned long index;
8677 /* make sure overflow events are dropped */
8678 tctx->in_idle = true;
8680 xa_for_each(&tctx->xa, index, file) {
8681 struct io_ring_ctx *ctx = file->private_data;
8683 io_uring_cancel_task_requests(ctx, files);
8685 io_uring_del_task_file(file);
8690 * Find any io_uring fd that this task has registered or done IO on, and cancel
8693 void __io_uring_task_cancel(void)
8695 struct io_uring_task *tctx = current->io_uring;
8699 /* make sure overflow events are dropped */
8700 tctx->in_idle = true;
8703 /* read completions before cancelations */
8704 inflight = percpu_counter_sum(&tctx->inflight);
8707 __io_uring_files_cancel(NULL);
8709 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8712 * If we've seen completions, retry. This avoids a race where
8713 * a completion comes in before we did prepare_to_wait().
8715 if (inflight != percpu_counter_sum(&tctx->inflight))
8720 finish_wait(&tctx->wait, &wait);
8721 tctx->in_idle = false;
8724 static int io_uring_flush(struct file *file, void *data)
8726 io_uring_attempt_task_drop(file);
8730 static void *io_uring_validate_mmap_request(struct file *file,
8731 loff_t pgoff, size_t sz)
8733 struct io_ring_ctx *ctx = file->private_data;
8734 loff_t offset = pgoff << PAGE_SHIFT;
8739 case IORING_OFF_SQ_RING:
8740 case IORING_OFF_CQ_RING:
8743 case IORING_OFF_SQES:
8747 return ERR_PTR(-EINVAL);
8750 page = virt_to_head_page(ptr);
8751 if (sz > page_size(page))
8752 return ERR_PTR(-EINVAL);
8759 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8761 size_t sz = vma->vm_end - vma->vm_start;
8765 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8767 return PTR_ERR(ptr);
8769 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8770 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8773 #else /* !CONFIG_MMU */
8775 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8777 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8780 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8782 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8785 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8786 unsigned long addr, unsigned long len,
8787 unsigned long pgoff, unsigned long flags)
8791 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8793 return PTR_ERR(ptr);
8795 return (unsigned long) ptr;
8798 #endif /* !CONFIG_MMU */
8800 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8805 if (!io_sqring_full(ctx))
8808 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8810 if (!io_sqring_full(ctx))
8814 } while (!signal_pending(current));
8816 finish_wait(&ctx->sqo_sq_wait, &wait);
8819 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8820 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8823 struct io_ring_ctx *ctx;
8830 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8831 IORING_ENTER_SQ_WAIT))
8839 if (f.file->f_op != &io_uring_fops)
8843 ctx = f.file->private_data;
8844 if (!percpu_ref_tryget(&ctx->refs))
8848 if (ctx->flags & IORING_SETUP_R_DISABLED)
8852 * For SQ polling, the thread will do all submissions and completions.
8853 * Just return the requested submit count, and wake the thread if
8857 if (ctx->flags & IORING_SETUP_SQPOLL) {
8858 if (!list_empty_careful(&ctx->cq_overflow_list))
8859 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8860 if (flags & IORING_ENTER_SQ_WAKEUP)
8861 wake_up(&ctx->sq_data->wait);
8862 if (flags & IORING_ENTER_SQ_WAIT)
8863 io_sqpoll_wait_sq(ctx);
8864 submitted = to_submit;
8865 } else if (to_submit) {
8866 ret = io_uring_add_task_file(f.file);
8869 mutex_lock(&ctx->uring_lock);
8870 submitted = io_submit_sqes(ctx, to_submit);
8871 mutex_unlock(&ctx->uring_lock);
8873 if (submitted != to_submit)
8876 if (flags & IORING_ENTER_GETEVENTS) {
8877 min_complete = min(min_complete, ctx->cq_entries);
8880 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8881 * space applications don't need to do io completion events
8882 * polling again, they can rely on io_sq_thread to do polling
8883 * work, which can reduce cpu usage and uring_lock contention.
8885 if (ctx->flags & IORING_SETUP_IOPOLL &&
8886 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8887 ret = io_iopoll_check(ctx, min_complete);
8889 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8894 percpu_ref_put(&ctx->refs);
8897 return submitted ? submitted : ret;
8900 #ifdef CONFIG_PROC_FS
8901 static int io_uring_show_cred(int id, void *p, void *data)
8903 const struct cred *cred = p;
8904 struct seq_file *m = data;
8905 struct user_namespace *uns = seq_user_ns(m);
8906 struct group_info *gi;
8911 seq_printf(m, "%5d\n", id);
8912 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8913 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8914 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8915 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8916 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8917 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8918 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8919 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8920 seq_puts(m, "\n\tGroups:\t");
8921 gi = cred->group_info;
8922 for (g = 0; g < gi->ngroups; g++) {
8923 seq_put_decimal_ull(m, g ? " " : "",
8924 from_kgid_munged(uns, gi->gid[g]));
8926 seq_puts(m, "\n\tCapEff:\t");
8927 cap = cred->cap_effective;
8928 CAP_FOR_EACH_U32(__capi)
8929 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8934 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8936 struct io_sq_data *sq = NULL;
8941 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8942 * since fdinfo case grabs it in the opposite direction of normal use
8943 * cases. If we fail to get the lock, we just don't iterate any
8944 * structures that could be going away outside the io_uring mutex.
8946 has_lock = mutex_trylock(&ctx->uring_lock);
8948 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8951 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8952 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8953 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8954 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8955 struct fixed_file_table *table;
8958 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8959 f = table->files[i & IORING_FILE_TABLE_MASK];
8961 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8963 seq_printf(m, "%5u: <none>\n", i);
8965 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8966 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8967 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8969 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8970 (unsigned int) buf->len);
8972 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8973 seq_printf(m, "Personalities:\n");
8974 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8976 seq_printf(m, "PollList:\n");
8977 spin_lock_irq(&ctx->completion_lock);
8978 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8979 struct hlist_head *list = &ctx->cancel_hash[i];
8980 struct io_kiocb *req;
8982 hlist_for_each_entry(req, list, hash_node)
8983 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8984 req->task->task_works != NULL);
8986 spin_unlock_irq(&ctx->completion_lock);
8988 mutex_unlock(&ctx->uring_lock);
8991 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8993 struct io_ring_ctx *ctx = f->private_data;
8995 if (percpu_ref_tryget(&ctx->refs)) {
8996 __io_uring_show_fdinfo(ctx, m);
8997 percpu_ref_put(&ctx->refs);
9002 static const struct file_operations io_uring_fops = {
9003 .release = io_uring_release,
9004 .flush = io_uring_flush,
9005 .mmap = io_uring_mmap,
9007 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9008 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9010 .poll = io_uring_poll,
9011 .fasync = io_uring_fasync,
9012 #ifdef CONFIG_PROC_FS
9013 .show_fdinfo = io_uring_show_fdinfo,
9017 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9018 struct io_uring_params *p)
9020 struct io_rings *rings;
9021 size_t size, sq_array_offset;
9023 /* make sure these are sane, as we already accounted them */
9024 ctx->sq_entries = p->sq_entries;
9025 ctx->cq_entries = p->cq_entries;
9027 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9028 if (size == SIZE_MAX)
9031 rings = io_mem_alloc(size);
9036 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9037 rings->sq_ring_mask = p->sq_entries - 1;
9038 rings->cq_ring_mask = p->cq_entries - 1;
9039 rings->sq_ring_entries = p->sq_entries;
9040 rings->cq_ring_entries = p->cq_entries;
9041 ctx->sq_mask = rings->sq_ring_mask;
9042 ctx->cq_mask = rings->cq_ring_mask;
9044 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9045 if (size == SIZE_MAX) {
9046 io_mem_free(ctx->rings);
9051 ctx->sq_sqes = io_mem_alloc(size);
9052 if (!ctx->sq_sqes) {
9053 io_mem_free(ctx->rings);
9062 * Allocate an anonymous fd, this is what constitutes the application
9063 * visible backing of an io_uring instance. The application mmaps this
9064 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9065 * we have to tie this fd to a socket for file garbage collection purposes.
9067 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9072 #if defined(CONFIG_UNIX)
9073 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9079 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9083 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9084 O_RDWR | O_CLOEXEC);
9088 ret = PTR_ERR(file);
9092 #if defined(CONFIG_UNIX)
9093 ctx->ring_sock->file = file;
9095 if (unlikely(io_uring_add_task_file(file))) {
9096 file = ERR_PTR(-ENOMEM);
9099 fd_install(ret, file);
9102 #if defined(CONFIG_UNIX)
9103 sock_release(ctx->ring_sock);
9104 ctx->ring_sock = NULL;
9109 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9110 struct io_uring_params __user *params)
9112 struct user_struct *user = NULL;
9113 struct io_ring_ctx *ctx;
9119 if (entries > IORING_MAX_ENTRIES) {
9120 if (!(p->flags & IORING_SETUP_CLAMP))
9122 entries = IORING_MAX_ENTRIES;
9126 * Use twice as many entries for the CQ ring. It's possible for the
9127 * application to drive a higher depth than the size of the SQ ring,
9128 * since the sqes are only used at submission time. This allows for
9129 * some flexibility in overcommitting a bit. If the application has
9130 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9131 * of CQ ring entries manually.
9133 p->sq_entries = roundup_pow_of_two(entries);
9134 if (p->flags & IORING_SETUP_CQSIZE) {
9136 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9137 * to a power-of-two, if it isn't already. We do NOT impose
9138 * any cq vs sq ring sizing.
9140 if (p->cq_entries < p->sq_entries)
9142 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9143 if (!(p->flags & IORING_SETUP_CLAMP))
9145 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9147 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9149 p->cq_entries = 2 * p->sq_entries;
9152 user = get_uid(current_user());
9153 limit_mem = !capable(CAP_IPC_LOCK);
9156 ret = __io_account_mem(user,
9157 ring_pages(p->sq_entries, p->cq_entries));
9164 ctx = io_ring_ctx_alloc(p);
9167 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9172 ctx->compat = in_compat_syscall();
9174 ctx->creds = get_current_cred();
9176 ctx->loginuid = current->loginuid;
9177 ctx->sessionid = current->sessionid;
9179 ctx->sqo_task = get_task_struct(current);
9182 * This is just grabbed for accounting purposes. When a process exits,
9183 * the mm is exited and dropped before the files, hence we need to hang
9184 * on to this mm purely for the purposes of being able to unaccount
9185 * memory (locked/pinned vm). It's not used for anything else.
9187 mmgrab(current->mm);
9188 ctx->mm_account = current->mm;
9190 #ifdef CONFIG_BLK_CGROUP
9192 * The sq thread will belong to the original cgroup it was inited in.
9193 * If the cgroup goes offline (e.g. disabling the io controller), then
9194 * issued bios will be associated with the closest cgroup later in the
9198 ctx->sqo_blkcg_css = blkcg_css();
9199 ret = css_tryget_online(ctx->sqo_blkcg_css);
9202 /* don't init against a dying cgroup, have the user try again */
9203 ctx->sqo_blkcg_css = NULL;
9210 * Account memory _before_ installing the file descriptor. Once
9211 * the descriptor is installed, it can get closed at any time. Also
9212 * do this before hitting the general error path, as ring freeing
9213 * will un-account as well.
9215 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9217 ctx->limit_mem = limit_mem;
9219 ret = io_allocate_scq_urings(ctx, p);
9223 ret = io_sq_offload_create(ctx, p);
9227 if (!(p->flags & IORING_SETUP_R_DISABLED))
9228 io_sq_offload_start(ctx);
9230 memset(&p->sq_off, 0, sizeof(p->sq_off));
9231 p->sq_off.head = offsetof(struct io_rings, sq.head);
9232 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9233 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9234 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9235 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9236 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9237 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9239 memset(&p->cq_off, 0, sizeof(p->cq_off));
9240 p->cq_off.head = offsetof(struct io_rings, cq.head);
9241 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9242 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9243 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9244 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9245 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9246 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9248 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9249 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9250 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9251 IORING_FEAT_POLL_32BITS;
9253 if (copy_to_user(params, p, sizeof(*p))) {
9259 * Install ring fd as the very last thing, so we don't risk someone
9260 * having closed it before we finish setup
9262 ret = io_uring_get_fd(ctx);
9266 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9269 io_ring_ctx_wait_and_kill(ctx);
9274 * Sets up an aio uring context, and returns the fd. Applications asks for a
9275 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9276 * params structure passed in.
9278 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9280 struct io_uring_params p;
9283 if (copy_from_user(&p, params, sizeof(p)))
9285 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9290 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9291 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9292 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9293 IORING_SETUP_R_DISABLED))
9296 return io_uring_create(entries, &p, params);
9299 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9300 struct io_uring_params __user *, params)
9302 return io_uring_setup(entries, params);
9305 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9307 struct io_uring_probe *p;
9311 size = struct_size(p, ops, nr_args);
9312 if (size == SIZE_MAX)
9314 p = kzalloc(size, GFP_KERNEL);
9319 if (copy_from_user(p, arg, size))
9322 if (memchr_inv(p, 0, size))
9325 p->last_op = IORING_OP_LAST - 1;
9326 if (nr_args > IORING_OP_LAST)
9327 nr_args = IORING_OP_LAST;
9329 for (i = 0; i < nr_args; i++) {
9331 if (!io_op_defs[i].not_supported)
9332 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9337 if (copy_to_user(arg, p, size))
9344 static int io_register_personality(struct io_ring_ctx *ctx)
9346 struct io_identity *id;
9349 id = kmalloc(sizeof(*id), GFP_KERNEL);
9353 io_init_identity(id);
9354 id->creds = get_current_cred();
9356 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9358 put_cred(id->creds);
9364 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9366 struct io_identity *iod;
9368 iod = idr_remove(&ctx->personality_idr, id);
9370 put_cred(iod->creds);
9371 if (refcount_dec_and_test(&iod->count))
9379 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9380 unsigned int nr_args)
9382 struct io_uring_restriction *res;
9386 /* Restrictions allowed only if rings started disabled */
9387 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9390 /* We allow only a single restrictions registration */
9391 if (ctx->restrictions.registered)
9394 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9397 size = array_size(nr_args, sizeof(*res));
9398 if (size == SIZE_MAX)
9401 res = memdup_user(arg, size);
9403 return PTR_ERR(res);
9407 for (i = 0; i < nr_args; i++) {
9408 switch (res[i].opcode) {
9409 case IORING_RESTRICTION_REGISTER_OP:
9410 if (res[i].register_op >= IORING_REGISTER_LAST) {
9415 __set_bit(res[i].register_op,
9416 ctx->restrictions.register_op);
9418 case IORING_RESTRICTION_SQE_OP:
9419 if (res[i].sqe_op >= IORING_OP_LAST) {
9424 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9426 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9427 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9429 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9430 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9439 /* Reset all restrictions if an error happened */
9441 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9443 ctx->restrictions.registered = true;
9449 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9451 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9454 if (ctx->restrictions.registered)
9455 ctx->restricted = 1;
9457 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9459 io_sq_offload_start(ctx);
9464 static bool io_register_op_must_quiesce(int op)
9467 case IORING_UNREGISTER_FILES:
9468 case IORING_REGISTER_FILES_UPDATE:
9469 case IORING_REGISTER_PROBE:
9470 case IORING_REGISTER_PERSONALITY:
9471 case IORING_UNREGISTER_PERSONALITY:
9478 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9479 void __user *arg, unsigned nr_args)
9480 __releases(ctx->uring_lock)
9481 __acquires(ctx->uring_lock)
9486 * We're inside the ring mutex, if the ref is already dying, then
9487 * someone else killed the ctx or is already going through
9488 * io_uring_register().
9490 if (percpu_ref_is_dying(&ctx->refs))
9493 if (io_register_op_must_quiesce(opcode)) {
9494 percpu_ref_kill(&ctx->refs);
9497 * Drop uring mutex before waiting for references to exit. If
9498 * another thread is currently inside io_uring_enter() it might
9499 * need to grab the uring_lock to make progress. If we hold it
9500 * here across the drain wait, then we can deadlock. It's safe
9501 * to drop the mutex here, since no new references will come in
9502 * after we've killed the percpu ref.
9504 mutex_unlock(&ctx->uring_lock);
9506 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9509 ret = io_run_task_work_sig();
9514 mutex_lock(&ctx->uring_lock);
9517 percpu_ref_resurrect(&ctx->refs);
9522 if (ctx->restricted) {
9523 if (opcode >= IORING_REGISTER_LAST) {
9528 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9535 case IORING_REGISTER_BUFFERS:
9536 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9538 case IORING_UNREGISTER_BUFFERS:
9542 ret = io_sqe_buffer_unregister(ctx);
9544 case IORING_REGISTER_FILES:
9545 ret = io_sqe_files_register(ctx, arg, nr_args);
9547 case IORING_UNREGISTER_FILES:
9551 ret = io_sqe_files_unregister(ctx);
9553 case IORING_REGISTER_FILES_UPDATE:
9554 ret = io_sqe_files_update(ctx, arg, nr_args);
9556 case IORING_REGISTER_EVENTFD:
9557 case IORING_REGISTER_EVENTFD_ASYNC:
9561 ret = io_eventfd_register(ctx, arg);
9564 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9565 ctx->eventfd_async = 1;
9567 ctx->eventfd_async = 0;
9569 case IORING_UNREGISTER_EVENTFD:
9573 ret = io_eventfd_unregister(ctx);
9575 case IORING_REGISTER_PROBE:
9577 if (!arg || nr_args > 256)
9579 ret = io_probe(ctx, arg, nr_args);
9581 case IORING_REGISTER_PERSONALITY:
9585 ret = io_register_personality(ctx);
9587 case IORING_UNREGISTER_PERSONALITY:
9591 ret = io_unregister_personality(ctx, nr_args);
9593 case IORING_REGISTER_ENABLE_RINGS:
9597 ret = io_register_enable_rings(ctx);
9599 case IORING_REGISTER_RESTRICTIONS:
9600 ret = io_register_restrictions(ctx, arg, nr_args);
9608 if (io_register_op_must_quiesce(opcode)) {
9609 /* bring the ctx back to life */
9610 percpu_ref_reinit(&ctx->refs);
9612 reinit_completion(&ctx->ref_comp);
9617 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9618 void __user *, arg, unsigned int, nr_args)
9620 struct io_ring_ctx *ctx;
9629 if (f.file->f_op != &io_uring_fops)
9632 ctx = f.file->private_data;
9634 mutex_lock(&ctx->uring_lock);
9635 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9636 mutex_unlock(&ctx->uring_lock);
9637 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9638 ctx->cq_ev_fd != NULL, ret);
9644 static int __init io_uring_init(void)
9646 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9647 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9648 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9651 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9652 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9653 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9654 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9655 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9656 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9657 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9658 BUILD_BUG_SQE_ELEM(8, __u64, off);
9659 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9660 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9661 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9662 BUILD_BUG_SQE_ELEM(24, __u32, len);
9663 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9664 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9665 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9666 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9667 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9668 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9669 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9670 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9671 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9672 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9673 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9674 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9675 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9676 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9677 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9678 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9679 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9680 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9681 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9683 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9684 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9685 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9688 __initcall(io_uring_init);