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);
1002 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1004 struct mm_struct *mm;
1009 /* Should never happen */
1010 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1013 task_lock(ctx->sqo_task);
1014 mm = ctx->sqo_task->mm;
1015 if (unlikely(!mm || !mmget_not_zero(mm)))
1017 task_unlock(ctx->sqo_task);
1027 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1028 struct io_kiocb *req)
1030 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1032 return __io_sq_thread_acquire_mm(ctx);
1035 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1036 struct cgroup_subsys_state **cur_css)
1039 #ifdef CONFIG_BLK_CGROUP
1040 /* puts the old one when swapping */
1041 if (*cur_css != ctx->sqo_blkcg_css) {
1042 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1043 *cur_css = ctx->sqo_blkcg_css;
1048 static void io_sq_thread_unassociate_blkcg(void)
1050 #ifdef CONFIG_BLK_CGROUP
1051 kthread_associate_blkcg(NULL);
1055 static inline void req_set_fail_links(struct io_kiocb *req)
1057 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1058 req->flags |= REQ_F_FAIL_LINK;
1062 * None of these are dereferenced, they are simply used to check if any of
1063 * them have changed. If we're under current and check they are still the
1064 * same, we're fine to grab references to them for actual out-of-line use.
1066 static void io_init_identity(struct io_identity *id)
1068 id->files = current->files;
1069 id->mm = current->mm;
1070 #ifdef CONFIG_BLK_CGROUP
1072 id->blkcg_css = blkcg_css();
1075 id->creds = current_cred();
1076 id->nsproxy = current->nsproxy;
1077 id->fs = current->fs;
1078 id->fsize = rlimit(RLIMIT_FSIZE);
1080 id->loginuid = current->loginuid;
1081 id->sessionid = current->sessionid;
1083 refcount_set(&id->count, 1);
1086 static inline void __io_req_init_async(struct io_kiocb *req)
1088 memset(&req->work, 0, sizeof(req->work));
1089 req->flags |= REQ_F_WORK_INITIALIZED;
1093 * Note: must call io_req_init_async() for the first time you
1094 * touch any members of io_wq_work.
1096 static inline void io_req_init_async(struct io_kiocb *req)
1098 struct io_uring_task *tctx = current->io_uring;
1100 if (req->flags & REQ_F_WORK_INITIALIZED)
1103 __io_req_init_async(req);
1105 /* Grab a ref if this isn't our static identity */
1106 req->work.identity = tctx->identity;
1107 if (tctx->identity != &tctx->__identity)
1108 refcount_inc(&req->work.identity->count);
1111 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1113 return ctx->flags & IORING_SETUP_SQPOLL;
1116 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1118 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1120 complete(&ctx->ref_comp);
1123 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1125 return !req->timeout.off;
1128 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1130 struct io_ring_ctx *ctx;
1133 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1137 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1138 if (!ctx->fallback_req)
1142 * Use 5 bits less than the max cq entries, that should give us around
1143 * 32 entries per hash list if totally full and uniformly spread.
1145 hash_bits = ilog2(p->cq_entries);
1149 ctx->cancel_hash_bits = hash_bits;
1150 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1152 if (!ctx->cancel_hash)
1154 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1156 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1157 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1160 ctx->flags = p->flags;
1161 init_waitqueue_head(&ctx->sqo_sq_wait);
1162 INIT_LIST_HEAD(&ctx->sqd_list);
1163 init_waitqueue_head(&ctx->cq_wait);
1164 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1165 init_completion(&ctx->ref_comp);
1166 init_completion(&ctx->sq_thread_comp);
1167 idr_init(&ctx->io_buffer_idr);
1168 idr_init(&ctx->personality_idr);
1169 mutex_init(&ctx->uring_lock);
1170 init_waitqueue_head(&ctx->wait);
1171 spin_lock_init(&ctx->completion_lock);
1172 INIT_LIST_HEAD(&ctx->iopoll_list);
1173 INIT_LIST_HEAD(&ctx->defer_list);
1174 INIT_LIST_HEAD(&ctx->timeout_list);
1175 init_waitqueue_head(&ctx->inflight_wait);
1176 spin_lock_init(&ctx->inflight_lock);
1177 INIT_LIST_HEAD(&ctx->inflight_list);
1178 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1179 init_llist_head(&ctx->file_put_llist);
1182 if (ctx->fallback_req)
1183 kmem_cache_free(req_cachep, ctx->fallback_req);
1184 kfree(ctx->cancel_hash);
1189 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1191 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1192 struct io_ring_ctx *ctx = req->ctx;
1194 return seq != ctx->cached_cq_tail
1195 + READ_ONCE(ctx->cached_cq_overflow);
1201 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1203 struct io_rings *rings = ctx->rings;
1205 /* order cqe stores with ring update */
1206 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1208 if (wq_has_sleeper(&ctx->cq_wait)) {
1209 wake_up_interruptible(&ctx->cq_wait);
1210 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1214 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1216 if (req->work.identity == &tctx->__identity)
1218 if (refcount_dec_and_test(&req->work.identity->count))
1219 kfree(req->work.identity);
1222 static void io_req_clean_work(struct io_kiocb *req)
1224 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1227 req->flags &= ~REQ_F_WORK_INITIALIZED;
1229 if (req->work.flags & IO_WQ_WORK_MM) {
1230 mmdrop(req->work.identity->mm);
1231 req->work.flags &= ~IO_WQ_WORK_MM;
1233 #ifdef CONFIG_BLK_CGROUP
1234 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1235 css_put(req->work.identity->blkcg_css);
1236 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1239 if (req->work.flags & IO_WQ_WORK_CREDS) {
1240 put_cred(req->work.identity->creds);
1241 req->work.flags &= ~IO_WQ_WORK_CREDS;
1243 if (req->work.flags & IO_WQ_WORK_FS) {
1244 struct fs_struct *fs = req->work.identity->fs;
1246 spin_lock(&req->work.identity->fs->lock);
1249 spin_unlock(&req->work.identity->fs->lock);
1252 req->work.flags &= ~IO_WQ_WORK_FS;
1255 io_put_identity(req->task->io_uring, req);
1259 * Create a private copy of io_identity, since some fields don't match
1260 * the current context.
1262 static bool io_identity_cow(struct io_kiocb *req)
1264 struct io_uring_task *tctx = current->io_uring;
1265 const struct cred *creds = NULL;
1266 struct io_identity *id;
1268 if (req->work.flags & IO_WQ_WORK_CREDS)
1269 creds = req->work.identity->creds;
1271 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1272 if (unlikely(!id)) {
1273 req->work.flags |= IO_WQ_WORK_CANCEL;
1278 * We can safely just re-init the creds we copied Either the field
1279 * matches the current one, or we haven't grabbed it yet. The only
1280 * exception is ->creds, through registered personalities, so handle
1281 * that one separately.
1283 io_init_identity(id);
1285 req->work.identity->creds = creds;
1287 /* add one for this request */
1288 refcount_inc(&id->count);
1290 /* drop tctx and req identity references, if needed */
1291 if (tctx->identity != &tctx->__identity &&
1292 refcount_dec_and_test(&tctx->identity->count))
1293 kfree(tctx->identity);
1294 if (req->work.identity != &tctx->__identity &&
1295 refcount_dec_and_test(&req->work.identity->count))
1296 kfree(req->work.identity);
1298 req->work.identity = id;
1299 tctx->identity = id;
1303 static bool io_grab_identity(struct io_kiocb *req)
1305 const struct io_op_def *def = &io_op_defs[req->opcode];
1306 struct io_identity *id = req->work.identity;
1307 struct io_ring_ctx *ctx = req->ctx;
1309 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1310 if (id->fsize != rlimit(RLIMIT_FSIZE))
1312 req->work.flags |= IO_WQ_WORK_FSIZE;
1315 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1316 (def->work_flags & IO_WQ_WORK_FILES) &&
1317 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1318 if (id->files != current->files ||
1319 id->nsproxy != current->nsproxy)
1321 atomic_inc(&id->files->count);
1322 get_nsproxy(id->nsproxy);
1323 req->flags |= REQ_F_INFLIGHT;
1325 spin_lock_irq(&ctx->inflight_lock);
1326 list_add(&req->inflight_entry, &ctx->inflight_list);
1327 spin_unlock_irq(&ctx->inflight_lock);
1328 req->work.flags |= IO_WQ_WORK_FILES;
1330 #ifdef CONFIG_BLK_CGROUP
1331 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1332 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1334 if (id->blkcg_css != blkcg_css()) {
1339 * This should be rare, either the cgroup is dying or the task
1340 * is moving cgroups. Just punt to root for the handful of ios.
1342 if (css_tryget_online(id->blkcg_css))
1343 req->work.flags |= IO_WQ_WORK_BLKCG;
1347 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1348 if (id->creds != current_cred())
1350 get_cred(id->creds);
1351 req->work.flags |= IO_WQ_WORK_CREDS;
1354 if (!uid_eq(current->loginuid, id->loginuid) ||
1355 current->sessionid != id->sessionid)
1358 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1359 (def->work_flags & IO_WQ_WORK_FS)) {
1360 if (current->fs != id->fs)
1362 spin_lock(&id->fs->lock);
1363 if (!id->fs->in_exec) {
1365 req->work.flags |= IO_WQ_WORK_FS;
1367 req->work.flags |= IO_WQ_WORK_CANCEL;
1369 spin_unlock(¤t->fs->lock);
1375 static void io_prep_async_work(struct io_kiocb *req)
1377 const struct io_op_def *def = &io_op_defs[req->opcode];
1378 struct io_ring_ctx *ctx = req->ctx;
1379 struct io_identity *id;
1381 io_req_init_async(req);
1382 id = req->work.identity;
1384 if (req->flags & REQ_F_FORCE_ASYNC)
1385 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1387 if (req->flags & REQ_F_ISREG) {
1388 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1389 io_wq_hash_work(&req->work, file_inode(req->file));
1391 if (def->unbound_nonreg_file)
1392 req->work.flags |= IO_WQ_WORK_UNBOUND;
1395 /* ->mm can never change on us */
1396 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1397 (def->work_flags & IO_WQ_WORK_MM)) {
1399 req->work.flags |= IO_WQ_WORK_MM;
1402 /* if we fail grabbing identity, we must COW, regrab, and retry */
1403 if (io_grab_identity(req))
1406 if (!io_identity_cow(req))
1409 /* can't fail at this point */
1410 if (!io_grab_identity(req))
1414 static void io_prep_async_link(struct io_kiocb *req)
1416 struct io_kiocb *cur;
1418 io_prep_async_work(req);
1419 if (req->flags & REQ_F_LINK_HEAD)
1420 list_for_each_entry(cur, &req->link_list, link_list)
1421 io_prep_async_work(cur);
1424 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1426 struct io_ring_ctx *ctx = req->ctx;
1427 struct io_kiocb *link = io_prep_linked_timeout(req);
1429 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1430 &req->work, req->flags);
1431 io_wq_enqueue(ctx->io_wq, &req->work);
1435 static void io_queue_async_work(struct io_kiocb *req)
1437 struct io_kiocb *link;
1439 /* init ->work of the whole link before punting */
1440 io_prep_async_link(req);
1441 link = __io_queue_async_work(req);
1444 io_queue_linked_timeout(link);
1447 static void io_kill_timeout(struct io_kiocb *req)
1449 struct io_timeout_data *io = req->async_data;
1452 ret = hrtimer_try_to_cancel(&io->timer);
1454 atomic_set(&req->ctx->cq_timeouts,
1455 atomic_read(&req->ctx->cq_timeouts) + 1);
1456 list_del_init(&req->timeout.list);
1457 io_cqring_fill_event(req, 0);
1458 io_put_req_deferred(req, 1);
1462 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1464 struct io_ring_ctx *ctx = req->ctx;
1466 if (!tsk || req->task == tsk)
1468 if (ctx->flags & IORING_SETUP_SQPOLL) {
1469 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1476 * Returns true if we found and killed one or more timeouts
1478 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1480 struct io_kiocb *req, *tmp;
1483 spin_lock_irq(&ctx->completion_lock);
1484 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1485 if (io_task_match(req, tsk)) {
1486 io_kill_timeout(req);
1490 spin_unlock_irq(&ctx->completion_lock);
1491 return canceled != 0;
1494 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1497 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1498 struct io_defer_entry, list);
1499 struct io_kiocb *link;
1501 if (req_need_defer(de->req, de->seq))
1503 list_del_init(&de->list);
1504 /* punt-init is done before queueing for defer */
1505 link = __io_queue_async_work(de->req);
1507 __io_queue_linked_timeout(link);
1508 /* drop submission reference */
1509 io_put_req_deferred(link, 1);
1512 } while (!list_empty(&ctx->defer_list));
1515 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1517 while (!list_empty(&ctx->timeout_list)) {
1518 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1519 struct io_kiocb, timeout.list);
1521 if (io_is_timeout_noseq(req))
1523 if (req->timeout.target_seq != ctx->cached_cq_tail
1524 - atomic_read(&ctx->cq_timeouts))
1527 list_del_init(&req->timeout.list);
1528 io_kill_timeout(req);
1532 static void io_commit_cqring(struct io_ring_ctx *ctx)
1534 io_flush_timeouts(ctx);
1535 __io_commit_cqring(ctx);
1537 if (unlikely(!list_empty(&ctx->defer_list)))
1538 __io_queue_deferred(ctx);
1541 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1543 struct io_rings *r = ctx->rings;
1545 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1548 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1550 struct io_rings *rings = ctx->rings;
1553 tail = ctx->cached_cq_tail;
1555 * writes to the cq entry need to come after reading head; the
1556 * control dependency is enough as we're using WRITE_ONCE to
1559 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1562 ctx->cached_cq_tail++;
1563 return &rings->cqes[tail & ctx->cq_mask];
1566 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1570 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1572 if (!ctx->eventfd_async)
1574 return io_wq_current_is_worker();
1577 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1579 if (waitqueue_active(&ctx->wait))
1580 wake_up(&ctx->wait);
1581 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1582 wake_up(&ctx->sq_data->wait);
1583 if (io_should_trigger_evfd(ctx))
1584 eventfd_signal(ctx->cq_ev_fd, 1);
1587 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1589 if (list_empty(&ctx->cq_overflow_list)) {
1590 clear_bit(0, &ctx->sq_check_overflow);
1591 clear_bit(0, &ctx->cq_check_overflow);
1592 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1596 static inline bool __io_match_files(struct io_kiocb *req,
1597 struct files_struct *files)
1599 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1600 (req->work.flags & IO_WQ_WORK_FILES)) &&
1601 req->work.identity->files == files;
1604 static bool io_match_files(struct io_kiocb *req,
1605 struct files_struct *files)
1607 struct io_kiocb *link;
1611 if (__io_match_files(req, files))
1613 if (req->flags & REQ_F_LINK_HEAD) {
1614 list_for_each_entry(link, &req->link_list, link_list) {
1615 if (__io_match_files(link, files))
1622 /* Returns true if there are no backlogged entries after the flush */
1623 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1624 struct task_struct *tsk,
1625 struct files_struct *files)
1627 struct io_rings *rings = ctx->rings;
1628 struct io_kiocb *req, *tmp;
1629 struct io_uring_cqe *cqe;
1630 unsigned long flags;
1634 if (list_empty_careful(&ctx->cq_overflow_list))
1636 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1637 rings->cq_ring_entries))
1641 spin_lock_irqsave(&ctx->completion_lock, flags);
1643 /* if force is set, the ring is going away. always drop after that */
1645 ctx->cq_overflow_flushed = 1;
1648 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1649 if (tsk && req->task != tsk)
1651 if (!io_match_files(req, files))
1654 cqe = io_get_cqring(ctx);
1658 list_move(&req->compl.list, &list);
1660 WRITE_ONCE(cqe->user_data, req->user_data);
1661 WRITE_ONCE(cqe->res, req->result);
1662 WRITE_ONCE(cqe->flags, req->compl.cflags);
1664 ctx->cached_cq_overflow++;
1665 WRITE_ONCE(ctx->rings->cq_overflow,
1666 ctx->cached_cq_overflow);
1670 io_commit_cqring(ctx);
1671 io_cqring_mark_overflow(ctx);
1673 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1674 io_cqring_ev_posted(ctx);
1676 while (!list_empty(&list)) {
1677 req = list_first_entry(&list, struct io_kiocb, compl.list);
1678 list_del(&req->compl.list);
1685 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1687 struct io_ring_ctx *ctx = req->ctx;
1688 struct io_uring_cqe *cqe;
1690 trace_io_uring_complete(ctx, req->user_data, res);
1693 * If we can't get a cq entry, userspace overflowed the
1694 * submission (by quite a lot). Increment the overflow count in
1697 cqe = io_get_cqring(ctx);
1699 WRITE_ONCE(cqe->user_data, req->user_data);
1700 WRITE_ONCE(cqe->res, res);
1701 WRITE_ONCE(cqe->flags, cflags);
1702 } else if (ctx->cq_overflow_flushed ||
1703 atomic_read(&req->task->io_uring->in_idle)) {
1705 * If we're in ring overflow flush mode, or in task cancel mode,
1706 * then we cannot store the request for later flushing, we need
1707 * to drop it on the floor.
1709 ctx->cached_cq_overflow++;
1710 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1712 if (list_empty(&ctx->cq_overflow_list)) {
1713 set_bit(0, &ctx->sq_check_overflow);
1714 set_bit(0, &ctx->cq_check_overflow);
1715 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1719 req->compl.cflags = cflags;
1720 refcount_inc(&req->refs);
1721 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1725 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1727 __io_cqring_fill_event(req, res, 0);
1730 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1732 struct io_ring_ctx *ctx = req->ctx;
1733 unsigned long flags;
1735 spin_lock_irqsave(&ctx->completion_lock, flags);
1736 __io_cqring_fill_event(req, res, cflags);
1737 io_commit_cqring(ctx);
1738 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1740 io_cqring_ev_posted(ctx);
1743 static void io_submit_flush_completions(struct io_comp_state *cs)
1745 struct io_ring_ctx *ctx = cs->ctx;
1747 spin_lock_irq(&ctx->completion_lock);
1748 while (!list_empty(&cs->list)) {
1749 struct io_kiocb *req;
1751 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1752 list_del(&req->compl.list);
1753 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1756 * io_free_req() doesn't care about completion_lock unless one
1757 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1758 * because of a potential deadlock with req->work.fs->lock
1760 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1761 |REQ_F_WORK_INITIALIZED)) {
1762 spin_unlock_irq(&ctx->completion_lock);
1764 spin_lock_irq(&ctx->completion_lock);
1769 io_commit_cqring(ctx);
1770 spin_unlock_irq(&ctx->completion_lock);
1772 io_cqring_ev_posted(ctx);
1776 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1777 struct io_comp_state *cs)
1780 io_cqring_add_event(req, res, cflags);
1785 req->compl.cflags = cflags;
1786 list_add_tail(&req->compl.list, &cs->list);
1788 io_submit_flush_completions(cs);
1792 static void io_req_complete(struct io_kiocb *req, long res)
1794 __io_req_complete(req, res, 0, NULL);
1797 static inline bool io_is_fallback_req(struct io_kiocb *req)
1799 return req == (struct io_kiocb *)
1800 ((unsigned long) req->ctx->fallback_req & ~1UL);
1803 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1805 struct io_kiocb *req;
1807 req = ctx->fallback_req;
1808 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1814 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1815 struct io_submit_state *state)
1817 if (!state->free_reqs) {
1818 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1822 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1823 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1826 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1827 * retry single alloc to be on the safe side.
1829 if (unlikely(ret <= 0)) {
1830 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1831 if (!state->reqs[0])
1835 state->free_reqs = ret;
1839 return state->reqs[state->free_reqs];
1841 return io_get_fallback_req(ctx);
1844 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1848 percpu_ref_put(req->fixed_file_refs);
1853 static void io_dismantle_req(struct io_kiocb *req)
1857 if (req->async_data)
1858 kfree(req->async_data);
1860 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1862 io_req_clean_work(req);
1865 static void __io_free_req(struct io_kiocb *req)
1867 struct io_uring_task *tctx = req->task->io_uring;
1868 struct io_ring_ctx *ctx = req->ctx;
1870 io_dismantle_req(req);
1872 percpu_counter_dec(&tctx->inflight);
1873 if (atomic_read(&tctx->in_idle))
1874 wake_up(&tctx->wait);
1875 put_task_struct(req->task);
1877 if (likely(!io_is_fallback_req(req)))
1878 kmem_cache_free(req_cachep, req);
1880 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1881 percpu_ref_put(&ctx->refs);
1884 static void io_kill_linked_timeout(struct io_kiocb *req)
1886 struct io_ring_ctx *ctx = req->ctx;
1887 struct io_kiocb *link;
1888 bool cancelled = false;
1889 unsigned long flags;
1891 spin_lock_irqsave(&ctx->completion_lock, flags);
1892 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1895 * Can happen if a linked timeout fired and link had been like
1896 * req -> link t-out -> link t-out [-> ...]
1898 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1899 struct io_timeout_data *io = link->async_data;
1902 list_del_init(&link->link_list);
1903 ret = hrtimer_try_to_cancel(&io->timer);
1905 io_cqring_fill_event(link, -ECANCELED);
1906 io_commit_cqring(ctx);
1910 req->flags &= ~REQ_F_LINK_TIMEOUT;
1911 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1914 io_cqring_ev_posted(ctx);
1919 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1921 struct io_kiocb *nxt;
1924 * The list should never be empty when we are called here. But could
1925 * potentially happen if the chain is messed up, check to be on the
1928 if (unlikely(list_empty(&req->link_list)))
1931 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1932 list_del_init(&req->link_list);
1933 if (!list_empty(&nxt->link_list))
1934 nxt->flags |= REQ_F_LINK_HEAD;
1939 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1941 static void io_fail_links(struct io_kiocb *req)
1943 struct io_ring_ctx *ctx = req->ctx;
1944 unsigned long flags;
1946 spin_lock_irqsave(&ctx->completion_lock, flags);
1947 while (!list_empty(&req->link_list)) {
1948 struct io_kiocb *link = list_first_entry(&req->link_list,
1949 struct io_kiocb, link_list);
1951 list_del_init(&link->link_list);
1952 trace_io_uring_fail_link(req, link);
1954 io_cqring_fill_event(link, -ECANCELED);
1957 * It's ok to free under spinlock as they're not linked anymore,
1958 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1961 if (link->flags & REQ_F_WORK_INITIALIZED)
1962 io_put_req_deferred(link, 2);
1964 io_double_put_req(link);
1967 io_commit_cqring(ctx);
1968 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1970 io_cqring_ev_posted(ctx);
1973 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1975 req->flags &= ~REQ_F_LINK_HEAD;
1976 if (req->flags & REQ_F_LINK_TIMEOUT)
1977 io_kill_linked_timeout(req);
1980 * If LINK is set, we have dependent requests in this chain. If we
1981 * didn't fail this request, queue the first one up, moving any other
1982 * dependencies to the next request. In case of failure, fail the rest
1985 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1986 return io_req_link_next(req);
1991 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1993 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1995 return __io_req_find_next(req);
1998 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2000 struct task_struct *tsk = req->task;
2001 struct io_ring_ctx *ctx = req->ctx;
2002 enum task_work_notify_mode notify;
2005 if (tsk->flags & PF_EXITING)
2009 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2010 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2011 * processing task_work. There's no reliable way to tell if TWA_RESUME
2015 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2016 notify = TWA_SIGNAL;
2018 ret = task_work_add(tsk, &req->task_work, notify);
2020 wake_up_process(tsk);
2025 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2027 struct io_ring_ctx *ctx = req->ctx;
2029 spin_lock_irq(&ctx->completion_lock);
2030 io_cqring_fill_event(req, error);
2031 io_commit_cqring(ctx);
2032 spin_unlock_irq(&ctx->completion_lock);
2034 io_cqring_ev_posted(ctx);
2035 req_set_fail_links(req);
2036 io_double_put_req(req);
2039 static void io_req_task_cancel(struct callback_head *cb)
2041 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2042 struct io_ring_ctx *ctx = req->ctx;
2044 __io_req_task_cancel(req, -ECANCELED);
2045 percpu_ref_put(&ctx->refs);
2048 static void __io_req_task_submit(struct io_kiocb *req)
2050 struct io_ring_ctx *ctx = req->ctx;
2052 if (!__io_sq_thread_acquire_mm(ctx)) {
2053 mutex_lock(&ctx->uring_lock);
2054 __io_queue_sqe(req, NULL);
2055 mutex_unlock(&ctx->uring_lock);
2057 __io_req_task_cancel(req, -EFAULT);
2061 static void io_req_task_submit(struct callback_head *cb)
2063 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2064 struct io_ring_ctx *ctx = req->ctx;
2066 __io_req_task_submit(req);
2067 percpu_ref_put(&ctx->refs);
2070 static void io_req_task_queue(struct io_kiocb *req)
2074 init_task_work(&req->task_work, io_req_task_submit);
2075 percpu_ref_get(&req->ctx->refs);
2077 ret = io_req_task_work_add(req, true);
2078 if (unlikely(ret)) {
2079 struct task_struct *tsk;
2081 init_task_work(&req->task_work, io_req_task_cancel);
2082 tsk = io_wq_get_task(req->ctx->io_wq);
2083 task_work_add(tsk, &req->task_work, TWA_NONE);
2084 wake_up_process(tsk);
2088 static void io_queue_next(struct io_kiocb *req)
2090 struct io_kiocb *nxt = io_req_find_next(req);
2093 io_req_task_queue(nxt);
2096 static void io_free_req(struct io_kiocb *req)
2103 void *reqs[IO_IOPOLL_BATCH];
2106 struct task_struct *task;
2110 static inline void io_init_req_batch(struct req_batch *rb)
2117 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2118 struct req_batch *rb)
2120 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2121 percpu_ref_put_many(&ctx->refs, rb->to_free);
2125 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2126 struct req_batch *rb)
2129 __io_req_free_batch_flush(ctx, rb);
2131 struct io_uring_task *tctx = rb->task->io_uring;
2133 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2134 put_task_struct_many(rb->task, rb->task_refs);
2139 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2141 if (unlikely(io_is_fallback_req(req))) {
2145 if (req->flags & REQ_F_LINK_HEAD)
2148 if (req->task != rb->task) {
2150 struct io_uring_task *tctx = rb->task->io_uring;
2152 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2153 put_task_struct_many(rb->task, rb->task_refs);
2155 rb->task = req->task;
2160 io_dismantle_req(req);
2161 rb->reqs[rb->to_free++] = req;
2162 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2163 __io_req_free_batch_flush(req->ctx, rb);
2167 * Drop reference to request, return next in chain (if there is one) if this
2168 * was the last reference to this request.
2170 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2172 struct io_kiocb *nxt = NULL;
2174 if (refcount_dec_and_test(&req->refs)) {
2175 nxt = io_req_find_next(req);
2181 static void io_put_req(struct io_kiocb *req)
2183 if (refcount_dec_and_test(&req->refs))
2187 static void io_put_req_deferred_cb(struct callback_head *cb)
2189 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2194 static void io_free_req_deferred(struct io_kiocb *req)
2198 init_task_work(&req->task_work, io_put_req_deferred_cb);
2199 ret = io_req_task_work_add(req, true);
2200 if (unlikely(ret)) {
2201 struct task_struct *tsk;
2203 tsk = io_wq_get_task(req->ctx->io_wq);
2204 task_work_add(tsk, &req->task_work, TWA_NONE);
2205 wake_up_process(tsk);
2209 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2211 if (refcount_sub_and_test(refs, &req->refs))
2212 io_free_req_deferred(req);
2215 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2217 struct io_kiocb *nxt;
2220 * A ref is owned by io-wq in which context we're. So, if that's the
2221 * last one, it's safe to steal next work. False negatives are Ok,
2222 * it just will be re-punted async in io_put_work()
2224 if (refcount_read(&req->refs) != 1)
2227 nxt = io_req_find_next(req);
2228 return nxt ? &nxt->work : NULL;
2231 static void io_double_put_req(struct io_kiocb *req)
2233 /* drop both submit and complete references */
2234 if (refcount_sub_and_test(2, &req->refs))
2238 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2240 struct io_rings *rings = ctx->rings;
2242 if (test_bit(0, &ctx->cq_check_overflow)) {
2244 * noflush == true is from the waitqueue handler, just ensure
2245 * we wake up the task, and the next invocation will flush the
2246 * entries. We cannot safely to it from here.
2248 if (noflush && !list_empty(&ctx->cq_overflow_list))
2251 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2254 /* See comment at the top of this file */
2256 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2259 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2261 struct io_rings *rings = ctx->rings;
2263 /* make sure SQ entry isn't read before tail */
2264 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2267 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2269 unsigned int cflags;
2271 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2272 cflags |= IORING_CQE_F_BUFFER;
2273 req->flags &= ~REQ_F_BUFFER_SELECTED;
2278 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2280 struct io_buffer *kbuf;
2282 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2283 return io_put_kbuf(req, kbuf);
2286 static inline bool io_run_task_work(void)
2289 * Not safe to run on exiting task, and the task_work handling will
2290 * not add work to such a task.
2292 if (unlikely(current->flags & PF_EXITING))
2294 if (current->task_works) {
2295 __set_current_state(TASK_RUNNING);
2303 static void io_iopoll_queue(struct list_head *again)
2305 struct io_kiocb *req;
2308 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2309 list_del(&req->inflight_entry);
2310 __io_complete_rw(req, -EAGAIN, 0, NULL);
2311 } while (!list_empty(again));
2315 * Find and free completed poll iocbs
2317 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2318 struct list_head *done)
2320 struct req_batch rb;
2321 struct io_kiocb *req;
2324 /* order with ->result store in io_complete_rw_iopoll() */
2327 io_init_req_batch(&rb);
2328 while (!list_empty(done)) {
2331 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2332 if (READ_ONCE(req->result) == -EAGAIN) {
2334 req->iopoll_completed = 0;
2335 list_move_tail(&req->inflight_entry, &again);
2338 list_del(&req->inflight_entry);
2340 if (req->flags & REQ_F_BUFFER_SELECTED)
2341 cflags = io_put_rw_kbuf(req);
2343 __io_cqring_fill_event(req, req->result, cflags);
2346 if (refcount_dec_and_test(&req->refs))
2347 io_req_free_batch(&rb, req);
2350 io_commit_cqring(ctx);
2351 if (ctx->flags & IORING_SETUP_SQPOLL)
2352 io_cqring_ev_posted(ctx);
2353 io_req_free_batch_finish(ctx, &rb);
2355 if (!list_empty(&again))
2356 io_iopoll_queue(&again);
2359 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2362 struct io_kiocb *req, *tmp;
2368 * Only spin for completions if we don't have multiple devices hanging
2369 * off our complete list, and we're under the requested amount.
2371 spin = !ctx->poll_multi_file && *nr_events < min;
2374 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2375 struct kiocb *kiocb = &req->rw.kiocb;
2378 * Move completed and retryable entries to our local lists.
2379 * If we find a request that requires polling, break out
2380 * and complete those lists first, if we have entries there.
2382 if (READ_ONCE(req->iopoll_completed)) {
2383 list_move_tail(&req->inflight_entry, &done);
2386 if (!list_empty(&done))
2389 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2393 /* iopoll may have completed current req */
2394 if (READ_ONCE(req->iopoll_completed))
2395 list_move_tail(&req->inflight_entry, &done);
2402 if (!list_empty(&done))
2403 io_iopoll_complete(ctx, nr_events, &done);
2409 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2410 * non-spinning poll check - we'll still enter the driver poll loop, but only
2411 * as a non-spinning completion check.
2413 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2416 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2419 ret = io_do_iopoll(ctx, nr_events, min);
2422 if (*nr_events >= min)
2430 * We can't just wait for polled events to come to us, we have to actively
2431 * find and complete them.
2433 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2435 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2438 mutex_lock(&ctx->uring_lock);
2439 while (!list_empty(&ctx->iopoll_list)) {
2440 unsigned int nr_events = 0;
2442 io_do_iopoll(ctx, &nr_events, 0);
2444 /* let it sleep and repeat later if can't complete a request */
2448 * Ensure we allow local-to-the-cpu processing to take place,
2449 * in this case we need to ensure that we reap all events.
2450 * Also let task_work, etc. to progress by releasing the mutex
2452 if (need_resched()) {
2453 mutex_unlock(&ctx->uring_lock);
2455 mutex_lock(&ctx->uring_lock);
2458 mutex_unlock(&ctx->uring_lock);
2461 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2463 unsigned int nr_events = 0;
2464 int iters = 0, ret = 0;
2467 * We disallow the app entering submit/complete with polling, but we
2468 * still need to lock the ring to prevent racing with polled issue
2469 * that got punted to a workqueue.
2471 mutex_lock(&ctx->uring_lock);
2474 * Don't enter poll loop if we already have events pending.
2475 * If we do, we can potentially be spinning for commands that
2476 * already triggered a CQE (eg in error).
2478 if (io_cqring_events(ctx, false))
2482 * If a submit got punted to a workqueue, we can have the
2483 * application entering polling for a command before it gets
2484 * issued. That app will hold the uring_lock for the duration
2485 * of the poll right here, so we need to take a breather every
2486 * now and then to ensure that the issue has a chance to add
2487 * the poll to the issued list. Otherwise we can spin here
2488 * forever, while the workqueue is stuck trying to acquire the
2491 if (!(++iters & 7)) {
2492 mutex_unlock(&ctx->uring_lock);
2494 mutex_lock(&ctx->uring_lock);
2497 ret = io_iopoll_getevents(ctx, &nr_events, min);
2501 } while (min && !nr_events && !need_resched());
2503 mutex_unlock(&ctx->uring_lock);
2507 static void kiocb_end_write(struct io_kiocb *req)
2510 * Tell lockdep we inherited freeze protection from submission
2513 if (req->flags & REQ_F_ISREG) {
2514 struct inode *inode = file_inode(req->file);
2516 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2518 file_end_write(req->file);
2521 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2522 struct io_comp_state *cs)
2524 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2527 if (kiocb->ki_flags & IOCB_WRITE)
2528 kiocb_end_write(req);
2530 if (res != req->result)
2531 req_set_fail_links(req);
2532 if (req->flags & REQ_F_BUFFER_SELECTED)
2533 cflags = io_put_rw_kbuf(req);
2534 __io_req_complete(req, res, cflags, cs);
2538 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2540 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2541 ssize_t ret = -ECANCELED;
2542 struct iov_iter iter;
2550 switch (req->opcode) {
2551 case IORING_OP_READV:
2552 case IORING_OP_READ_FIXED:
2553 case IORING_OP_READ:
2556 case IORING_OP_WRITEV:
2557 case IORING_OP_WRITE_FIXED:
2558 case IORING_OP_WRITE:
2562 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2567 if (!req->async_data) {
2568 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2571 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2579 req_set_fail_links(req);
2580 io_req_complete(req, ret);
2585 static bool io_rw_reissue(struct io_kiocb *req, long res)
2588 umode_t mode = file_inode(req->file)->i_mode;
2591 if (!S_ISBLK(mode) && !S_ISREG(mode))
2593 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2596 ret = io_sq_thread_acquire_mm(req->ctx, req);
2598 if (io_resubmit_prep(req, ret)) {
2599 refcount_inc(&req->refs);
2600 io_queue_async_work(req);
2608 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2609 struct io_comp_state *cs)
2611 if (!io_rw_reissue(req, res))
2612 io_complete_rw_common(&req->rw.kiocb, res, cs);
2615 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2617 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2619 __io_complete_rw(req, res, res2, NULL);
2622 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2624 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2626 if (kiocb->ki_flags & IOCB_WRITE)
2627 kiocb_end_write(req);
2629 if (res != -EAGAIN && res != req->result)
2630 req_set_fail_links(req);
2632 WRITE_ONCE(req->result, res);
2633 /* order with io_poll_complete() checking ->result */
2635 WRITE_ONCE(req->iopoll_completed, 1);
2639 * After the iocb has been issued, it's safe to be found on the poll list.
2640 * Adding the kiocb to the list AFTER submission ensures that we don't
2641 * find it from a io_iopoll_getevents() thread before the issuer is done
2642 * accessing the kiocb cookie.
2644 static void io_iopoll_req_issued(struct io_kiocb *req)
2646 struct io_ring_ctx *ctx = req->ctx;
2649 * Track whether we have multiple files in our lists. This will impact
2650 * how we do polling eventually, not spinning if we're on potentially
2651 * different devices.
2653 if (list_empty(&ctx->iopoll_list)) {
2654 ctx->poll_multi_file = false;
2655 } else if (!ctx->poll_multi_file) {
2656 struct io_kiocb *list_req;
2658 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2660 if (list_req->file != req->file)
2661 ctx->poll_multi_file = true;
2665 * For fast devices, IO may have already completed. If it has, add
2666 * it to the front so we find it first.
2668 if (READ_ONCE(req->iopoll_completed))
2669 list_add(&req->inflight_entry, &ctx->iopoll_list);
2671 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2673 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2674 wq_has_sleeper(&ctx->sq_data->wait))
2675 wake_up(&ctx->sq_data->wait);
2678 static void __io_state_file_put(struct io_submit_state *state)
2680 if (state->has_refs)
2681 fput_many(state->file, state->has_refs);
2685 static inline void io_state_file_put(struct io_submit_state *state)
2688 __io_state_file_put(state);
2692 * Get as many references to a file as we have IOs left in this submission,
2693 * assuming most submissions are for one file, or at least that each file
2694 * has more than one submission.
2696 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2702 if (state->fd == fd) {
2706 __io_state_file_put(state);
2708 state->file = fget_many(fd, state->ios_left);
2713 state->has_refs = state->ios_left - 1;
2717 static bool io_bdev_nowait(struct block_device *bdev)
2720 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2727 * If we tracked the file through the SCM inflight mechanism, we could support
2728 * any file. For now, just ensure that anything potentially problematic is done
2731 static bool io_file_supports_async(struct file *file, int rw)
2733 umode_t mode = file_inode(file)->i_mode;
2735 if (S_ISBLK(mode)) {
2736 if (io_bdev_nowait(file->f_inode->i_bdev))
2740 if (S_ISCHR(mode) || S_ISSOCK(mode))
2742 if (S_ISREG(mode)) {
2743 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2744 file->f_op != &io_uring_fops)
2749 /* any ->read/write should understand O_NONBLOCK */
2750 if (file->f_flags & O_NONBLOCK)
2753 if (!(file->f_mode & FMODE_NOWAIT))
2757 return file->f_op->read_iter != NULL;
2759 return file->f_op->write_iter != NULL;
2762 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2764 struct io_ring_ctx *ctx = req->ctx;
2765 struct kiocb *kiocb = &req->rw.kiocb;
2769 if (S_ISREG(file_inode(req->file)->i_mode))
2770 req->flags |= REQ_F_ISREG;
2772 kiocb->ki_pos = READ_ONCE(sqe->off);
2773 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2774 req->flags |= REQ_F_CUR_POS;
2775 kiocb->ki_pos = req->file->f_pos;
2777 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2778 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2779 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2783 ioprio = READ_ONCE(sqe->ioprio);
2785 ret = ioprio_check_cap(ioprio);
2789 kiocb->ki_ioprio = ioprio;
2791 kiocb->ki_ioprio = get_current_ioprio();
2793 /* don't allow async punt if RWF_NOWAIT was requested */
2794 if (kiocb->ki_flags & IOCB_NOWAIT)
2795 req->flags |= REQ_F_NOWAIT;
2797 if (ctx->flags & IORING_SETUP_IOPOLL) {
2798 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2799 !kiocb->ki_filp->f_op->iopoll)
2802 kiocb->ki_flags |= IOCB_HIPRI;
2803 kiocb->ki_complete = io_complete_rw_iopoll;
2804 req->iopoll_completed = 0;
2806 if (kiocb->ki_flags & IOCB_HIPRI)
2808 kiocb->ki_complete = io_complete_rw;
2811 req->rw.addr = READ_ONCE(sqe->addr);
2812 req->rw.len = READ_ONCE(sqe->len);
2813 req->buf_index = READ_ONCE(sqe->buf_index);
2817 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2823 case -ERESTARTNOINTR:
2824 case -ERESTARTNOHAND:
2825 case -ERESTART_RESTARTBLOCK:
2827 * We can't just restart the syscall, since previously
2828 * submitted sqes may already be in progress. Just fail this
2834 kiocb->ki_complete(kiocb, ret, 0);
2838 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2839 struct io_comp_state *cs)
2841 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2842 struct io_async_rw *io = req->async_data;
2844 /* add previously done IO, if any */
2845 if (io && io->bytes_done > 0) {
2847 ret = io->bytes_done;
2849 ret += io->bytes_done;
2852 if (req->flags & REQ_F_CUR_POS)
2853 req->file->f_pos = kiocb->ki_pos;
2854 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2855 __io_complete_rw(req, ret, 0, cs);
2857 io_rw_done(kiocb, ret);
2860 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2861 struct iov_iter *iter)
2863 struct io_ring_ctx *ctx = req->ctx;
2864 size_t len = req->rw.len;
2865 struct io_mapped_ubuf *imu;
2866 u16 index, buf_index = req->buf_index;
2870 if (unlikely(buf_index >= ctx->nr_user_bufs))
2872 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2873 imu = &ctx->user_bufs[index];
2874 buf_addr = req->rw.addr;
2877 if (buf_addr + len < buf_addr)
2879 /* not inside the mapped region */
2880 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2884 * May not be a start of buffer, set size appropriately
2885 * and advance us to the beginning.
2887 offset = buf_addr - imu->ubuf;
2888 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2892 * Don't use iov_iter_advance() here, as it's really slow for
2893 * using the latter parts of a big fixed buffer - it iterates
2894 * over each segment manually. We can cheat a bit here, because
2897 * 1) it's a BVEC iter, we set it up
2898 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2899 * first and last bvec
2901 * So just find our index, and adjust the iterator afterwards.
2902 * If the offset is within the first bvec (or the whole first
2903 * bvec, just use iov_iter_advance(). This makes it easier
2904 * since we can just skip the first segment, which may not
2905 * be PAGE_SIZE aligned.
2907 const struct bio_vec *bvec = imu->bvec;
2909 if (offset <= bvec->bv_len) {
2910 iov_iter_advance(iter, offset);
2912 unsigned long seg_skip;
2914 /* skip first vec */
2915 offset -= bvec->bv_len;
2916 seg_skip = 1 + (offset >> PAGE_SHIFT);
2918 iter->bvec = bvec + seg_skip;
2919 iter->nr_segs -= seg_skip;
2920 iter->count -= bvec->bv_len + offset;
2921 iter->iov_offset = offset & ~PAGE_MASK;
2928 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2931 mutex_unlock(&ctx->uring_lock);
2934 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2937 * "Normal" inline submissions always hold the uring_lock, since we
2938 * grab it from the system call. Same is true for the SQPOLL offload.
2939 * The only exception is when we've detached the request and issue it
2940 * from an async worker thread, grab the lock for that case.
2943 mutex_lock(&ctx->uring_lock);
2946 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2947 int bgid, struct io_buffer *kbuf,
2950 struct io_buffer *head;
2952 if (req->flags & REQ_F_BUFFER_SELECTED)
2955 io_ring_submit_lock(req->ctx, needs_lock);
2957 lockdep_assert_held(&req->ctx->uring_lock);
2959 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2961 if (!list_empty(&head->list)) {
2962 kbuf = list_last_entry(&head->list, struct io_buffer,
2964 list_del(&kbuf->list);
2967 idr_remove(&req->ctx->io_buffer_idr, bgid);
2969 if (*len > kbuf->len)
2972 kbuf = ERR_PTR(-ENOBUFS);
2975 io_ring_submit_unlock(req->ctx, needs_lock);
2980 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2983 struct io_buffer *kbuf;
2986 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2987 bgid = req->buf_index;
2988 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2991 req->rw.addr = (u64) (unsigned long) kbuf;
2992 req->flags |= REQ_F_BUFFER_SELECTED;
2993 return u64_to_user_ptr(kbuf->addr);
2996 #ifdef CONFIG_COMPAT
2997 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3000 struct compat_iovec __user *uiov;
3001 compat_ssize_t clen;
3005 uiov = u64_to_user_ptr(req->rw.addr);
3006 if (!access_ok(uiov, sizeof(*uiov)))
3008 if (__get_user(clen, &uiov->iov_len))
3014 buf = io_rw_buffer_select(req, &len, needs_lock);
3016 return PTR_ERR(buf);
3017 iov[0].iov_base = buf;
3018 iov[0].iov_len = (compat_size_t) len;
3023 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3026 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3030 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3033 len = iov[0].iov_len;
3036 buf = io_rw_buffer_select(req, &len, needs_lock);
3038 return PTR_ERR(buf);
3039 iov[0].iov_base = buf;
3040 iov[0].iov_len = len;
3044 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3047 if (req->flags & REQ_F_BUFFER_SELECTED) {
3048 struct io_buffer *kbuf;
3050 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3051 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3052 iov[0].iov_len = kbuf->len;
3057 else if (req->rw.len > 1)
3060 #ifdef CONFIG_COMPAT
3061 if (req->ctx->compat)
3062 return io_compat_import(req, iov, needs_lock);
3065 return __io_iov_buffer_select(req, iov, needs_lock);
3068 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3069 struct iovec **iovec, struct iov_iter *iter,
3072 void __user *buf = u64_to_user_ptr(req->rw.addr);
3073 size_t sqe_len = req->rw.len;
3077 opcode = req->opcode;
3078 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3080 return io_import_fixed(req, rw, iter);
3083 /* buffer index only valid with fixed read/write, or buffer select */
3084 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3087 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3088 if (req->flags & REQ_F_BUFFER_SELECT) {
3089 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3091 return PTR_ERR(buf);
3092 req->rw.len = sqe_len;
3095 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3097 return ret < 0 ? ret : sqe_len;
3100 if (req->flags & REQ_F_BUFFER_SELECT) {
3101 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3103 ret = (*iovec)->iov_len;
3104 iov_iter_init(iter, rw, *iovec, 1, ret);
3110 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3114 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3115 struct iovec **iovec, struct iov_iter *iter,
3118 struct io_async_rw *iorw = req->async_data;
3121 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3123 return iov_iter_count(&iorw->iter);
3126 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3128 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3132 * For files that don't have ->read_iter() and ->write_iter(), handle them
3133 * by looping over ->read() or ->write() manually.
3135 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3137 struct kiocb *kiocb = &req->rw.kiocb;
3138 struct file *file = req->file;
3142 * Don't support polled IO through this interface, and we can't
3143 * support non-blocking either. For the latter, this just causes
3144 * the kiocb to be handled from an async context.
3146 if (kiocb->ki_flags & IOCB_HIPRI)
3148 if (kiocb->ki_flags & IOCB_NOWAIT)
3151 while (iov_iter_count(iter)) {
3155 if (!iov_iter_is_bvec(iter)) {
3156 iovec = iov_iter_iovec(iter);
3158 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3159 iovec.iov_len = req->rw.len;
3163 nr = file->f_op->read(file, iovec.iov_base,
3164 iovec.iov_len, io_kiocb_ppos(kiocb));
3166 nr = file->f_op->write(file, iovec.iov_base,
3167 iovec.iov_len, io_kiocb_ppos(kiocb));
3176 if (nr != iovec.iov_len)
3180 iov_iter_advance(iter, nr);
3186 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3187 const struct iovec *fast_iov, struct iov_iter *iter)
3189 struct io_async_rw *rw = req->async_data;
3191 memcpy(&rw->iter, iter, sizeof(*iter));
3192 rw->free_iovec = iovec;
3194 /* can only be fixed buffers, no need to do anything */
3195 if (iter->type == ITER_BVEC)
3198 unsigned iov_off = 0;
3200 rw->iter.iov = rw->fast_iov;
3201 if (iter->iov != fast_iov) {
3202 iov_off = iter->iov - fast_iov;
3203 rw->iter.iov += iov_off;
3205 if (rw->fast_iov != fast_iov)
3206 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3207 sizeof(struct iovec) * iter->nr_segs);
3209 req->flags |= REQ_F_NEED_CLEANUP;
3213 static inline int __io_alloc_async_data(struct io_kiocb *req)
3215 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3216 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3217 return req->async_data == NULL;
3220 static int io_alloc_async_data(struct io_kiocb *req)
3222 if (!io_op_defs[req->opcode].needs_async_data)
3225 return __io_alloc_async_data(req);
3228 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3229 const struct iovec *fast_iov,
3230 struct iov_iter *iter, bool force)
3232 if (!force && !io_op_defs[req->opcode].needs_async_data)
3234 if (!req->async_data) {
3235 if (__io_alloc_async_data(req))
3238 io_req_map_rw(req, iovec, fast_iov, iter);
3243 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3245 struct io_async_rw *iorw = req->async_data;
3246 struct iovec *iov = iorw->fast_iov;
3249 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3250 if (unlikely(ret < 0))
3253 iorw->bytes_done = 0;
3254 iorw->free_iovec = iov;
3256 req->flags |= REQ_F_NEED_CLEANUP;
3260 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3264 ret = io_prep_rw(req, sqe);
3268 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3271 /* either don't need iovec imported or already have it */
3272 if (!req->async_data)
3274 return io_rw_prep_async(req, READ);
3278 * This is our waitqueue callback handler, registered through lock_page_async()
3279 * when we initially tried to do the IO with the iocb armed our waitqueue.
3280 * This gets called when the page is unlocked, and we generally expect that to
3281 * happen when the page IO is completed and the page is now uptodate. This will
3282 * queue a task_work based retry of the operation, attempting to copy the data
3283 * again. If the latter fails because the page was NOT uptodate, then we will
3284 * do a thread based blocking retry of the operation. That's the unexpected
3287 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3288 int sync, void *arg)
3290 struct wait_page_queue *wpq;
3291 struct io_kiocb *req = wait->private;
3292 struct wait_page_key *key = arg;
3295 wpq = container_of(wait, struct wait_page_queue, wait);
3297 if (!wake_page_match(wpq, key))
3300 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3301 list_del_init(&wait->entry);
3303 init_task_work(&req->task_work, io_req_task_submit);
3304 percpu_ref_get(&req->ctx->refs);
3306 /* submit ref gets dropped, acquire a new one */
3307 refcount_inc(&req->refs);
3308 ret = io_req_task_work_add(req, true);
3309 if (unlikely(ret)) {
3310 struct task_struct *tsk;
3312 /* queue just for cancelation */
3313 init_task_work(&req->task_work, io_req_task_cancel);
3314 tsk = io_wq_get_task(req->ctx->io_wq);
3315 task_work_add(tsk, &req->task_work, TWA_NONE);
3316 wake_up_process(tsk);
3322 * This controls whether a given IO request should be armed for async page
3323 * based retry. If we return false here, the request is handed to the async
3324 * worker threads for retry. If we're doing buffered reads on a regular file,
3325 * we prepare a private wait_page_queue entry and retry the operation. This
3326 * will either succeed because the page is now uptodate and unlocked, or it
3327 * will register a callback when the page is unlocked at IO completion. Through
3328 * that callback, io_uring uses task_work to setup a retry of the operation.
3329 * That retry will attempt the buffered read again. The retry will generally
3330 * succeed, or in rare cases where it fails, we then fall back to using the
3331 * async worker threads for a blocking retry.
3333 static bool io_rw_should_retry(struct io_kiocb *req)
3335 struct io_async_rw *rw = req->async_data;
3336 struct wait_page_queue *wait = &rw->wpq;
3337 struct kiocb *kiocb = &req->rw.kiocb;
3339 /* never retry for NOWAIT, we just complete with -EAGAIN */
3340 if (req->flags & REQ_F_NOWAIT)
3343 /* Only for buffered IO */
3344 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3348 * just use poll if we can, and don't attempt if the fs doesn't
3349 * support callback based unlocks
3351 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3354 wait->wait.func = io_async_buf_func;
3355 wait->wait.private = req;
3356 wait->wait.flags = 0;
3357 INIT_LIST_HEAD(&wait->wait.entry);
3358 kiocb->ki_flags |= IOCB_WAITQ;
3359 kiocb->ki_flags &= ~IOCB_NOWAIT;
3360 kiocb->ki_waitq = wait;
3364 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3366 if (req->file->f_op->read_iter)
3367 return call_read_iter(req->file, &req->rw.kiocb, iter);
3368 else if (req->file->f_op->read)
3369 return loop_rw_iter(READ, req, iter);
3374 static int io_read(struct io_kiocb *req, bool force_nonblock,
3375 struct io_comp_state *cs)
3377 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3378 struct kiocb *kiocb = &req->rw.kiocb;
3379 struct iov_iter __iter, *iter = &__iter;
3380 struct io_async_rw *rw = req->async_data;
3381 ssize_t io_size, ret, ret2;
3388 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3391 iov_count = iov_iter_count(iter);
3393 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3403 /* If the file doesn't support async, just async punt */
3404 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3408 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3412 ret = io_iter_do_read(req, iter);
3416 } else if (ret == -EIOCBQUEUED) {
3419 } else if (ret == -EAGAIN) {
3420 /* IOPOLL retry should happen for io-wq threads */
3421 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3423 /* no retry on NONBLOCK marked file */
3424 if (req->file->f_flags & O_NONBLOCK)
3426 /* some cases will consume bytes even on error returns */
3427 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3430 } else if (ret < 0) {
3431 /* make sure -ERESTARTSYS -> -EINTR is done */
3435 /* read it all, or we did blocking attempt. no retry. */
3436 if (!iov_iter_count(iter) || !force_nonblock ||
3437 (req->file->f_flags & O_NONBLOCK))
3442 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3449 rw = req->async_data;
3450 /* it's copied and will be cleaned with ->io */
3452 /* now use our persistent iterator, if we aren't already */
3455 rw->bytes_done += ret;
3456 /* if we can retry, do so with the callbacks armed */
3457 if (!io_rw_should_retry(req)) {
3458 kiocb->ki_flags &= ~IOCB_WAITQ;
3463 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3464 * get -EIOCBQUEUED, then we'll get a notification when the desired
3465 * page gets unlocked. We can also get a partial read here, and if we
3466 * do, then just retry at the new offset.
3468 ret = io_iter_do_read(req, iter);
3469 if (ret == -EIOCBQUEUED) {
3472 } else if (ret > 0 && ret < io_size) {
3473 /* we got some bytes, but not all. retry. */
3477 kiocb_done(kiocb, ret, cs);
3480 /* it's reportedly faster than delegating the null check to kfree() */
3486 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3490 ret = io_prep_rw(req, sqe);
3494 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3497 /* either don't need iovec imported or already have it */
3498 if (!req->async_data)
3500 return io_rw_prep_async(req, WRITE);
3503 static int io_write(struct io_kiocb *req, bool force_nonblock,
3504 struct io_comp_state *cs)
3506 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3507 struct kiocb *kiocb = &req->rw.kiocb;
3508 struct iov_iter __iter, *iter = &__iter;
3509 struct io_async_rw *rw = req->async_data;
3511 ssize_t ret, ret2, io_size;
3516 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3519 iov_count = iov_iter_count(iter);
3521 req->result = io_size;
3523 /* Ensure we clear previously set non-block flag */
3524 if (!force_nonblock)
3525 kiocb->ki_flags &= ~IOCB_NOWAIT;
3527 kiocb->ki_flags |= IOCB_NOWAIT;
3529 /* If the file doesn't support async, just async punt */
3530 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3533 /* file path doesn't support NOWAIT for non-direct_IO */
3534 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3535 (req->flags & REQ_F_ISREG))
3538 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3543 * Open-code file_start_write here to grab freeze protection,
3544 * which will be released by another thread in
3545 * io_complete_rw(). Fool lockdep by telling it the lock got
3546 * released so that it doesn't complain about the held lock when
3547 * we return to userspace.
3549 if (req->flags & REQ_F_ISREG) {
3550 sb_start_write(file_inode(req->file)->i_sb);
3551 __sb_writers_release(file_inode(req->file)->i_sb,
3554 kiocb->ki_flags |= IOCB_WRITE;
3556 if (req->file->f_op->write_iter)
3557 ret2 = call_write_iter(req->file, kiocb, iter);
3558 else if (req->file->f_op->write)
3559 ret2 = loop_rw_iter(WRITE, req, iter);
3564 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3565 * retry them without IOCB_NOWAIT.
3567 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3569 /* no retry on NONBLOCK marked file */
3570 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3572 if (!force_nonblock || ret2 != -EAGAIN) {
3573 /* IOPOLL retry should happen for io-wq threads */
3574 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3577 kiocb_done(kiocb, ret2, cs);
3580 /* some cases will consume bytes even on error returns */
3581 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3582 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3587 /* it's reportedly faster than delegating the null check to kfree() */
3593 static int __io_splice_prep(struct io_kiocb *req,
3594 const struct io_uring_sqe *sqe)
3596 struct io_splice* sp = &req->splice;
3597 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3603 sp->len = READ_ONCE(sqe->len);
3604 sp->flags = READ_ONCE(sqe->splice_flags);
3606 if (unlikely(sp->flags & ~valid_flags))
3609 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3610 (sp->flags & SPLICE_F_FD_IN_FIXED));
3613 req->flags |= REQ_F_NEED_CLEANUP;
3615 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3617 * Splice operation will be punted aync, and here need to
3618 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3620 io_req_init_async(req);
3621 req->work.flags |= IO_WQ_WORK_UNBOUND;
3627 static int io_tee_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3632 return __io_splice_prep(req, sqe);
3635 static int io_tee(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;
3646 ret = do_tee(in, out, sp->len, flags);
3648 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3649 req->flags &= ~REQ_F_NEED_CLEANUP;
3652 req_set_fail_links(req);
3653 io_req_complete(req, ret);
3657 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3659 struct io_splice* sp = &req->splice;
3661 sp->off_in = READ_ONCE(sqe->splice_off_in);
3662 sp->off_out = READ_ONCE(sqe->off);
3663 return __io_splice_prep(req, sqe);
3666 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3668 struct io_splice *sp = &req->splice;
3669 struct file *in = sp->file_in;
3670 struct file *out = sp->file_out;
3671 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3672 loff_t *poff_in, *poff_out;
3678 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3679 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3682 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3684 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3685 req->flags &= ~REQ_F_NEED_CLEANUP;
3688 req_set_fail_links(req);
3689 io_req_complete(req, ret);
3694 * IORING_OP_NOP just posts a completion event, nothing else.
3696 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3698 struct io_ring_ctx *ctx = req->ctx;
3700 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3703 __io_req_complete(req, 0, 0, cs);
3707 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3709 struct io_ring_ctx *ctx = req->ctx;
3714 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3716 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3719 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3720 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3723 req->sync.off = READ_ONCE(sqe->off);
3724 req->sync.len = READ_ONCE(sqe->len);
3728 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3730 loff_t end = req->sync.off + req->sync.len;
3733 /* fsync always requires a blocking context */
3737 ret = vfs_fsync_range(req->file, req->sync.off,
3738 end > 0 ? end : LLONG_MAX,
3739 req->sync.flags & IORING_FSYNC_DATASYNC);
3741 req_set_fail_links(req);
3742 io_req_complete(req, ret);
3746 static int io_fallocate_prep(struct io_kiocb *req,
3747 const struct io_uring_sqe *sqe)
3749 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3751 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3754 req->sync.off = READ_ONCE(sqe->off);
3755 req->sync.len = READ_ONCE(sqe->addr);
3756 req->sync.mode = READ_ONCE(sqe->len);
3760 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3764 /* fallocate always requiring blocking context */
3767 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3770 req_set_fail_links(req);
3771 io_req_complete(req, ret);
3775 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3777 const char __user *fname;
3780 if (unlikely(sqe->ioprio || sqe->buf_index))
3782 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3785 /* open.how should be already initialised */
3786 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3787 req->open.how.flags |= O_LARGEFILE;
3789 req->open.dfd = READ_ONCE(sqe->fd);
3790 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3791 req->open.filename = getname(fname);
3792 if (IS_ERR(req->open.filename)) {
3793 ret = PTR_ERR(req->open.filename);
3794 req->open.filename = NULL;
3797 req->open.nofile = rlimit(RLIMIT_NOFILE);
3798 req->flags |= REQ_F_NEED_CLEANUP;
3802 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3806 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3808 mode = READ_ONCE(sqe->len);
3809 flags = READ_ONCE(sqe->open_flags);
3810 req->open.how = build_open_how(flags, mode);
3811 return __io_openat_prep(req, sqe);
3814 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3816 struct open_how __user *how;
3820 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3822 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3823 len = READ_ONCE(sqe->len);
3824 if (len < OPEN_HOW_SIZE_VER0)
3827 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3832 return __io_openat_prep(req, sqe);
3835 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3837 struct open_flags op;
3844 ret = build_open_flags(&req->open.how, &op);
3848 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3852 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3855 ret = PTR_ERR(file);
3857 fsnotify_open(file);
3858 fd_install(ret, file);
3861 putname(req->open.filename);
3862 req->flags &= ~REQ_F_NEED_CLEANUP;
3864 req_set_fail_links(req);
3865 io_req_complete(req, ret);
3869 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3871 return io_openat2(req, force_nonblock);
3874 static int io_remove_buffers_prep(struct io_kiocb *req,
3875 const struct io_uring_sqe *sqe)
3877 struct io_provide_buf *p = &req->pbuf;
3880 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3883 tmp = READ_ONCE(sqe->fd);
3884 if (!tmp || tmp > USHRT_MAX)
3887 memset(p, 0, sizeof(*p));
3889 p->bgid = READ_ONCE(sqe->buf_group);
3893 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3894 int bgid, unsigned nbufs)
3898 /* shouldn't happen */
3902 /* the head kbuf is the list itself */
3903 while (!list_empty(&buf->list)) {
3904 struct io_buffer *nxt;
3906 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3907 list_del(&nxt->list);
3914 idr_remove(&ctx->io_buffer_idr, bgid);
3919 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3920 struct io_comp_state *cs)
3922 struct io_provide_buf *p = &req->pbuf;
3923 struct io_ring_ctx *ctx = req->ctx;
3924 struct io_buffer *head;
3927 io_ring_submit_lock(ctx, !force_nonblock);
3929 lockdep_assert_held(&ctx->uring_lock);
3932 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3934 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3936 io_ring_submit_lock(ctx, !force_nonblock);
3938 req_set_fail_links(req);
3939 __io_req_complete(req, ret, 0, cs);
3943 static int io_provide_buffers_prep(struct io_kiocb *req,
3944 const struct io_uring_sqe *sqe)
3946 struct io_provide_buf *p = &req->pbuf;
3949 if (sqe->ioprio || sqe->rw_flags)
3952 tmp = READ_ONCE(sqe->fd);
3953 if (!tmp || tmp > USHRT_MAX)
3956 p->addr = READ_ONCE(sqe->addr);
3957 p->len = READ_ONCE(sqe->len);
3959 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3962 p->bgid = READ_ONCE(sqe->buf_group);
3963 tmp = READ_ONCE(sqe->off);
3964 if (tmp > USHRT_MAX)
3970 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3972 struct io_buffer *buf;
3973 u64 addr = pbuf->addr;
3974 int i, bid = pbuf->bid;
3976 for (i = 0; i < pbuf->nbufs; i++) {
3977 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3982 buf->len = pbuf->len;
3987 INIT_LIST_HEAD(&buf->list);
3990 list_add_tail(&buf->list, &(*head)->list);
3994 return i ? i : -ENOMEM;
3997 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3998 struct io_comp_state *cs)
4000 struct io_provide_buf *p = &req->pbuf;
4001 struct io_ring_ctx *ctx = req->ctx;
4002 struct io_buffer *head, *list;
4005 io_ring_submit_lock(ctx, !force_nonblock);
4007 lockdep_assert_held(&ctx->uring_lock);
4009 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4011 ret = io_add_buffers(p, &head);
4016 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4019 __io_remove_buffers(ctx, head, p->bgid, -1U);
4024 io_ring_submit_unlock(ctx, !force_nonblock);
4026 req_set_fail_links(req);
4027 __io_req_complete(req, ret, 0, cs);
4031 static int io_epoll_ctl_prep(struct io_kiocb *req,
4032 const struct io_uring_sqe *sqe)
4034 #if defined(CONFIG_EPOLL)
4035 if (sqe->ioprio || sqe->buf_index)
4037 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4040 req->epoll.epfd = READ_ONCE(sqe->fd);
4041 req->epoll.op = READ_ONCE(sqe->len);
4042 req->epoll.fd = READ_ONCE(sqe->off);
4044 if (ep_op_has_event(req->epoll.op)) {
4045 struct epoll_event __user *ev;
4047 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4048 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4058 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4059 struct io_comp_state *cs)
4061 #if defined(CONFIG_EPOLL)
4062 struct io_epoll *ie = &req->epoll;
4065 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4066 if (force_nonblock && ret == -EAGAIN)
4070 req_set_fail_links(req);
4071 __io_req_complete(req, ret, 0, cs);
4078 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4080 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4081 if (sqe->ioprio || sqe->buf_index || sqe->off)
4083 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4086 req->madvise.addr = READ_ONCE(sqe->addr);
4087 req->madvise.len = READ_ONCE(sqe->len);
4088 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4095 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4097 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4098 struct io_madvise *ma = &req->madvise;
4104 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4106 req_set_fail_links(req);
4107 io_req_complete(req, ret);
4114 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4116 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->fadvise.offset = READ_ONCE(sqe->off);
4122 req->fadvise.len = READ_ONCE(sqe->len);
4123 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4127 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4129 struct io_fadvise *fa = &req->fadvise;
4132 if (force_nonblock) {
4133 switch (fa->advice) {
4134 case POSIX_FADV_NORMAL:
4135 case POSIX_FADV_RANDOM:
4136 case POSIX_FADV_SEQUENTIAL:
4143 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4145 req_set_fail_links(req);
4146 io_req_complete(req, ret);
4150 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4152 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4154 if (sqe->ioprio || sqe->buf_index)
4156 if (req->flags & REQ_F_FIXED_FILE)
4159 req->statx.dfd = READ_ONCE(sqe->fd);
4160 req->statx.mask = READ_ONCE(sqe->len);
4161 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4162 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4163 req->statx.flags = READ_ONCE(sqe->statx_flags);
4168 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4170 struct io_statx *ctx = &req->statx;
4173 if (force_nonblock) {
4174 /* only need file table for an actual valid fd */
4175 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4176 req->flags |= REQ_F_NO_FILE_TABLE;
4180 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4184 req_set_fail_links(req);
4185 io_req_complete(req, ret);
4189 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4192 * If we queue this for async, it must not be cancellable. That would
4193 * leave the 'file' in an undeterminate state, and here need to modify
4194 * io_wq_work.flags, so initialize io_wq_work firstly.
4196 io_req_init_async(req);
4197 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4199 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4201 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4202 sqe->rw_flags || sqe->buf_index)
4204 if (req->flags & REQ_F_FIXED_FILE)
4207 req->close.fd = READ_ONCE(sqe->fd);
4208 if ((req->file && req->file->f_op == &io_uring_fops))
4211 req->close.put_file = NULL;
4215 static int io_close(struct io_kiocb *req, bool force_nonblock,
4216 struct io_comp_state *cs)
4218 struct io_close *close = &req->close;
4221 /* might be already done during nonblock submission */
4222 if (!close->put_file) {
4223 ret = __close_fd_get_file(close->fd, &close->put_file);
4225 return (ret == -ENOENT) ? -EBADF : ret;
4228 /* if the file has a flush method, be safe and punt to async */
4229 if (close->put_file->f_op->flush && force_nonblock) {
4230 /* was never set, but play safe */
4231 req->flags &= ~REQ_F_NOWAIT;
4232 /* avoid grabbing files - we don't need the files */
4233 req->flags |= REQ_F_NO_FILE_TABLE;
4237 /* No ->flush() or already async, safely close from here */
4238 ret = filp_close(close->put_file, req->work.identity->files);
4240 req_set_fail_links(req);
4241 fput(close->put_file);
4242 close->put_file = NULL;
4243 __io_req_complete(req, ret, 0, cs);
4247 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4249 struct io_ring_ctx *ctx = req->ctx;
4254 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4256 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4259 req->sync.off = READ_ONCE(sqe->off);
4260 req->sync.len = READ_ONCE(sqe->len);
4261 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4265 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4269 /* sync_file_range always requires a blocking context */
4273 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4276 req_set_fail_links(req);
4277 io_req_complete(req, ret);
4281 #if defined(CONFIG_NET)
4282 static int io_setup_async_msg(struct io_kiocb *req,
4283 struct io_async_msghdr *kmsg)
4285 struct io_async_msghdr *async_msg = req->async_data;
4289 if (io_alloc_async_data(req)) {
4290 if (kmsg->iov != kmsg->fast_iov)
4294 async_msg = req->async_data;
4295 req->flags |= REQ_F_NEED_CLEANUP;
4296 memcpy(async_msg, kmsg, sizeof(*kmsg));
4300 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4301 struct io_async_msghdr *iomsg)
4303 iomsg->iov = iomsg->fast_iov;
4304 iomsg->msg.msg_name = &iomsg->addr;
4305 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4306 req->sr_msg.msg_flags, &iomsg->iov);
4309 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4311 struct io_async_msghdr *async_msg = req->async_data;
4312 struct io_sr_msg *sr = &req->sr_msg;
4315 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4318 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4319 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4320 sr->len = READ_ONCE(sqe->len);
4322 #ifdef CONFIG_COMPAT
4323 if (req->ctx->compat)
4324 sr->msg_flags |= MSG_CMSG_COMPAT;
4327 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4329 ret = io_sendmsg_copy_hdr(req, async_msg);
4331 req->flags |= REQ_F_NEED_CLEANUP;
4335 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4336 struct io_comp_state *cs)
4338 struct io_async_msghdr iomsg, *kmsg;
4339 struct socket *sock;
4343 sock = sock_from_file(req->file, &ret);
4344 if (unlikely(!sock))
4347 if (req->async_data) {
4348 kmsg = req->async_data;
4349 kmsg->msg.msg_name = &kmsg->addr;
4350 /* if iov is set, it's allocated already */
4352 kmsg->iov = kmsg->fast_iov;
4353 kmsg->msg.msg_iter.iov = kmsg->iov;
4355 ret = io_sendmsg_copy_hdr(req, &iomsg);
4361 flags = req->sr_msg.msg_flags;
4362 if (flags & MSG_DONTWAIT)
4363 req->flags |= REQ_F_NOWAIT;
4364 else if (force_nonblock)
4365 flags |= MSG_DONTWAIT;
4367 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4368 if (force_nonblock && ret == -EAGAIN)
4369 return io_setup_async_msg(req, kmsg);
4370 if (ret == -ERESTARTSYS)
4373 if (kmsg->iov != kmsg->fast_iov)
4375 req->flags &= ~REQ_F_NEED_CLEANUP;
4377 req_set_fail_links(req);
4378 __io_req_complete(req, ret, 0, cs);
4382 static int io_send(struct io_kiocb *req, bool force_nonblock,
4383 struct io_comp_state *cs)
4385 struct io_sr_msg *sr = &req->sr_msg;
4388 struct socket *sock;
4392 sock = sock_from_file(req->file, &ret);
4393 if (unlikely(!sock))
4396 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4400 msg.msg_name = NULL;
4401 msg.msg_control = NULL;
4402 msg.msg_controllen = 0;
4403 msg.msg_namelen = 0;
4405 flags = req->sr_msg.msg_flags;
4406 if (flags & MSG_DONTWAIT)
4407 req->flags |= REQ_F_NOWAIT;
4408 else if (force_nonblock)
4409 flags |= MSG_DONTWAIT;
4411 msg.msg_flags = flags;
4412 ret = sock_sendmsg(sock, &msg);
4413 if (force_nonblock && ret == -EAGAIN)
4415 if (ret == -ERESTARTSYS)
4419 req_set_fail_links(req);
4420 __io_req_complete(req, ret, 0, cs);
4424 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4425 struct io_async_msghdr *iomsg)
4427 struct io_sr_msg *sr = &req->sr_msg;
4428 struct iovec __user *uiov;
4432 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4433 &iomsg->uaddr, &uiov, &iov_len);
4437 if (req->flags & REQ_F_BUFFER_SELECT) {
4440 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4442 sr->len = iomsg->iov[0].iov_len;
4443 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4447 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4448 &iomsg->iov, &iomsg->msg.msg_iter,
4457 #ifdef CONFIG_COMPAT
4458 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4459 struct io_async_msghdr *iomsg)
4461 struct compat_msghdr __user *msg_compat;
4462 struct io_sr_msg *sr = &req->sr_msg;
4463 struct compat_iovec __user *uiov;
4468 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4469 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4474 uiov = compat_ptr(ptr);
4475 if (req->flags & REQ_F_BUFFER_SELECT) {
4476 compat_ssize_t clen;
4480 if (!access_ok(uiov, sizeof(*uiov)))
4482 if (__get_user(clen, &uiov->iov_len))
4486 sr->len = iomsg->iov[0].iov_len;
4489 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4490 UIO_FASTIOV, &iomsg->iov,
4491 &iomsg->msg.msg_iter, true);
4500 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4501 struct io_async_msghdr *iomsg)
4503 iomsg->msg.msg_name = &iomsg->addr;
4504 iomsg->iov = iomsg->fast_iov;
4506 #ifdef CONFIG_COMPAT
4507 if (req->ctx->compat)
4508 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4511 return __io_recvmsg_copy_hdr(req, iomsg);
4514 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4517 struct io_sr_msg *sr = &req->sr_msg;
4518 struct io_buffer *kbuf;
4520 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4525 req->flags |= REQ_F_BUFFER_SELECTED;
4529 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4531 return io_put_kbuf(req, req->sr_msg.kbuf);
4534 static int io_recvmsg_prep(struct io_kiocb *req,
4535 const struct io_uring_sqe *sqe)
4537 struct io_async_msghdr *async_msg = req->async_data;
4538 struct io_sr_msg *sr = &req->sr_msg;
4541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4544 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4545 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4546 sr->len = READ_ONCE(sqe->len);
4547 sr->bgid = READ_ONCE(sqe->buf_group);
4549 #ifdef CONFIG_COMPAT
4550 if (req->ctx->compat)
4551 sr->msg_flags |= MSG_CMSG_COMPAT;
4554 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4556 ret = io_recvmsg_copy_hdr(req, async_msg);
4558 req->flags |= REQ_F_NEED_CLEANUP;
4562 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4563 struct io_comp_state *cs)
4565 struct io_async_msghdr iomsg, *kmsg;
4566 struct socket *sock;
4567 struct io_buffer *kbuf;
4569 int ret, cflags = 0;
4571 sock = sock_from_file(req->file, &ret);
4572 if (unlikely(!sock))
4575 if (req->async_data) {
4576 kmsg = req->async_data;
4577 kmsg->msg.msg_name = &kmsg->addr;
4578 /* if iov is set, it's allocated already */
4580 kmsg->iov = kmsg->fast_iov;
4581 kmsg->msg.msg_iter.iov = kmsg->iov;
4583 ret = io_recvmsg_copy_hdr(req, &iomsg);
4589 if (req->flags & REQ_F_BUFFER_SELECT) {
4590 kbuf = io_recv_buffer_select(req, !force_nonblock);
4592 return PTR_ERR(kbuf);
4593 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4594 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4595 1, req->sr_msg.len);
4598 flags = req->sr_msg.msg_flags;
4599 if (flags & MSG_DONTWAIT)
4600 req->flags |= REQ_F_NOWAIT;
4601 else if (force_nonblock)
4602 flags |= MSG_DONTWAIT;
4604 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4605 kmsg->uaddr, flags);
4606 if (force_nonblock && ret == -EAGAIN)
4607 return io_setup_async_msg(req, kmsg);
4608 if (ret == -ERESTARTSYS)
4611 if (req->flags & REQ_F_BUFFER_SELECTED)
4612 cflags = io_put_recv_kbuf(req);
4613 if (kmsg->iov != kmsg->fast_iov)
4615 req->flags &= ~REQ_F_NEED_CLEANUP;
4617 req_set_fail_links(req);
4618 __io_req_complete(req, ret, cflags, cs);
4622 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4623 struct io_comp_state *cs)
4625 struct io_buffer *kbuf;
4626 struct io_sr_msg *sr = &req->sr_msg;
4628 void __user *buf = sr->buf;
4629 struct socket *sock;
4632 int ret, cflags = 0;
4634 sock = sock_from_file(req->file, &ret);
4635 if (unlikely(!sock))
4638 if (req->flags & REQ_F_BUFFER_SELECT) {
4639 kbuf = io_recv_buffer_select(req, !force_nonblock);
4641 return PTR_ERR(kbuf);
4642 buf = u64_to_user_ptr(kbuf->addr);
4645 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4649 msg.msg_name = NULL;
4650 msg.msg_control = NULL;
4651 msg.msg_controllen = 0;
4652 msg.msg_namelen = 0;
4653 msg.msg_iocb = NULL;
4656 flags = req->sr_msg.msg_flags;
4657 if (flags & MSG_DONTWAIT)
4658 req->flags |= REQ_F_NOWAIT;
4659 else if (force_nonblock)
4660 flags |= MSG_DONTWAIT;
4662 ret = sock_recvmsg(sock, &msg, flags);
4663 if (force_nonblock && ret == -EAGAIN)
4665 if (ret == -ERESTARTSYS)
4668 if (req->flags & REQ_F_BUFFER_SELECTED)
4669 cflags = io_put_recv_kbuf(req);
4671 req_set_fail_links(req);
4672 __io_req_complete(req, ret, cflags, cs);
4676 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4678 struct io_accept *accept = &req->accept;
4680 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4682 if (sqe->ioprio || sqe->len || sqe->buf_index)
4685 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4686 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4687 accept->flags = READ_ONCE(sqe->accept_flags);
4688 accept->nofile = rlimit(RLIMIT_NOFILE);
4692 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4693 struct io_comp_state *cs)
4695 struct io_accept *accept = &req->accept;
4696 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4699 if (req->file->f_flags & O_NONBLOCK)
4700 req->flags |= REQ_F_NOWAIT;
4702 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4703 accept->addr_len, accept->flags,
4705 if (ret == -EAGAIN && force_nonblock)
4708 if (ret == -ERESTARTSYS)
4710 req_set_fail_links(req);
4712 __io_req_complete(req, ret, 0, cs);
4716 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_connect *conn = &req->connect;
4719 struct io_async_connect *io = req->async_data;
4721 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4723 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4726 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4727 conn->addr_len = READ_ONCE(sqe->addr2);
4732 return move_addr_to_kernel(conn->addr, conn->addr_len,
4736 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4737 struct io_comp_state *cs)
4739 struct io_async_connect __io, *io;
4740 unsigned file_flags;
4743 if (req->async_data) {
4744 io = req->async_data;
4746 ret = move_addr_to_kernel(req->connect.addr,
4747 req->connect.addr_len,
4754 file_flags = force_nonblock ? O_NONBLOCK : 0;
4756 ret = __sys_connect_file(req->file, &io->address,
4757 req->connect.addr_len, file_flags);
4758 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4759 if (req->async_data)
4761 if (io_alloc_async_data(req)) {
4765 io = req->async_data;
4766 memcpy(req->async_data, &__io, sizeof(__io));
4769 if (ret == -ERESTARTSYS)
4773 req_set_fail_links(req);
4774 __io_req_complete(req, ret, 0, cs);
4777 #else /* !CONFIG_NET */
4778 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4783 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4784 struct io_comp_state *cs)
4789 static int io_send(struct io_kiocb *req, bool force_nonblock,
4790 struct io_comp_state *cs)
4795 static int io_recvmsg_prep(struct io_kiocb *req,
4796 const struct io_uring_sqe *sqe)
4801 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4802 struct io_comp_state *cs)
4807 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4808 struct io_comp_state *cs)
4813 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4818 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4819 struct io_comp_state *cs)
4824 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4829 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4830 struct io_comp_state *cs)
4834 #endif /* CONFIG_NET */
4836 struct io_poll_table {
4837 struct poll_table_struct pt;
4838 struct io_kiocb *req;
4842 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4843 __poll_t mask, task_work_func_t func)
4848 /* for instances that support it check for an event match first: */
4849 if (mask && !(mask & poll->events))
4852 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4854 list_del_init(&poll->wait.entry);
4857 init_task_work(&req->task_work, func);
4858 percpu_ref_get(&req->ctx->refs);
4861 * If we using the signalfd wait_queue_head for this wakeup, then
4862 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4863 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4864 * either, as the normal wakeup will suffice.
4866 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4869 * If this fails, then the task is exiting. When a task exits, the
4870 * work gets canceled, so just cancel this request as well instead
4871 * of executing it. We can't safely execute it anyway, as we may not
4872 * have the needed state needed for it anyway.
4874 ret = io_req_task_work_add(req, twa_signal_ok);
4875 if (unlikely(ret)) {
4876 struct task_struct *tsk;
4878 WRITE_ONCE(poll->canceled, true);
4879 tsk = io_wq_get_task(req->ctx->io_wq);
4880 task_work_add(tsk, &req->task_work, TWA_NONE);
4881 wake_up_process(tsk);
4886 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4887 __acquires(&req->ctx->completion_lock)
4889 struct io_ring_ctx *ctx = req->ctx;
4891 if (!req->result && !READ_ONCE(poll->canceled)) {
4892 struct poll_table_struct pt = { ._key = poll->events };
4894 req->result = vfs_poll(req->file, &pt) & poll->events;
4897 spin_lock_irq(&ctx->completion_lock);
4898 if (!req->result && !READ_ONCE(poll->canceled)) {
4899 add_wait_queue(poll->head, &poll->wait);
4906 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4908 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4909 if (req->opcode == IORING_OP_POLL_ADD)
4910 return req->async_data;
4911 return req->apoll->double_poll;
4914 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4916 if (req->opcode == IORING_OP_POLL_ADD)
4918 return &req->apoll->poll;
4921 static void io_poll_remove_double(struct io_kiocb *req)
4923 struct io_poll_iocb *poll = io_poll_get_double(req);
4925 lockdep_assert_held(&req->ctx->completion_lock);
4927 if (poll && poll->head) {
4928 struct wait_queue_head *head = poll->head;
4930 spin_lock(&head->lock);
4931 list_del_init(&poll->wait.entry);
4932 if (poll->wait.private)
4933 refcount_dec(&req->refs);
4935 spin_unlock(&head->lock);
4939 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4941 struct io_ring_ctx *ctx = req->ctx;
4943 io_poll_remove_double(req);
4944 req->poll.done = true;
4945 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4946 io_commit_cqring(ctx);
4949 static void io_poll_task_func(struct callback_head *cb)
4951 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4952 struct io_ring_ctx *ctx = req->ctx;
4953 struct io_kiocb *nxt;
4955 if (io_poll_rewait(req, &req->poll)) {
4956 spin_unlock_irq(&ctx->completion_lock);
4958 hash_del(&req->hash_node);
4959 io_poll_complete(req, req->result, 0);
4960 spin_unlock_irq(&ctx->completion_lock);
4962 nxt = io_put_req_find_next(req);
4963 io_cqring_ev_posted(ctx);
4965 __io_req_task_submit(nxt);
4968 percpu_ref_put(&ctx->refs);
4971 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4972 int sync, void *key)
4974 struct io_kiocb *req = wait->private;
4975 struct io_poll_iocb *poll = io_poll_get_single(req);
4976 __poll_t mask = key_to_poll(key);
4978 /* for instances that support it check for an event match first: */
4979 if (mask && !(mask & poll->events))
4982 list_del_init(&wait->entry);
4984 if (poll && poll->head) {
4987 spin_lock(&poll->head->lock);
4988 done = list_empty(&poll->wait.entry);
4990 list_del_init(&poll->wait.entry);
4991 /* make sure double remove sees this as being gone */
4992 wait->private = NULL;
4993 spin_unlock(&poll->head->lock);
4995 /* use wait func handler, so it matches the rq type */
4996 poll->wait.func(&poll->wait, mode, sync, key);
4999 refcount_dec(&req->refs);
5003 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5004 wait_queue_func_t wake_func)
5008 poll->canceled = false;
5009 poll->events = events;
5010 INIT_LIST_HEAD(&poll->wait.entry);
5011 init_waitqueue_func_entry(&poll->wait, wake_func);
5014 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5015 struct wait_queue_head *head,
5016 struct io_poll_iocb **poll_ptr)
5018 struct io_kiocb *req = pt->req;
5021 * If poll->head is already set, it's because the file being polled
5022 * uses multiple waitqueues for poll handling (eg one for read, one
5023 * for write). Setup a separate io_poll_iocb if this happens.
5025 if (unlikely(poll->head)) {
5026 struct io_poll_iocb *poll_one = poll;
5028 /* already have a 2nd entry, fail a third attempt */
5030 pt->error = -EINVAL;
5033 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5035 pt->error = -ENOMEM;
5038 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5039 refcount_inc(&req->refs);
5040 poll->wait.private = req;
5047 if (poll->events & EPOLLEXCLUSIVE)
5048 add_wait_queue_exclusive(head, &poll->wait);
5050 add_wait_queue(head, &poll->wait);
5053 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5054 struct poll_table_struct *p)
5056 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5057 struct async_poll *apoll = pt->req->apoll;
5059 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5062 static void io_async_task_func(struct callback_head *cb)
5064 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5065 struct async_poll *apoll = req->apoll;
5066 struct io_ring_ctx *ctx = req->ctx;
5068 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5070 if (io_poll_rewait(req, &apoll->poll)) {
5071 spin_unlock_irq(&ctx->completion_lock);
5072 percpu_ref_put(&ctx->refs);
5076 /* If req is still hashed, it cannot have been canceled. Don't check. */
5077 if (hash_hashed(&req->hash_node))
5078 hash_del(&req->hash_node);
5080 io_poll_remove_double(req);
5081 spin_unlock_irq(&ctx->completion_lock);
5083 if (!READ_ONCE(apoll->poll.canceled))
5084 __io_req_task_submit(req);
5086 __io_req_task_cancel(req, -ECANCELED);
5088 percpu_ref_put(&ctx->refs);
5089 kfree(apoll->double_poll);
5093 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5096 struct io_kiocb *req = wait->private;
5097 struct io_poll_iocb *poll = &req->apoll->poll;
5099 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5102 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5105 static void io_poll_req_insert(struct io_kiocb *req)
5107 struct io_ring_ctx *ctx = req->ctx;
5108 struct hlist_head *list;
5110 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5111 hlist_add_head(&req->hash_node, list);
5114 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5115 struct io_poll_iocb *poll,
5116 struct io_poll_table *ipt, __poll_t mask,
5117 wait_queue_func_t wake_func)
5118 __acquires(&ctx->completion_lock)
5120 struct io_ring_ctx *ctx = req->ctx;
5121 bool cancel = false;
5123 INIT_HLIST_NODE(&req->hash_node);
5124 io_init_poll_iocb(poll, mask, wake_func);
5125 poll->file = req->file;
5126 poll->wait.private = req;
5128 ipt->pt._key = mask;
5130 ipt->error = -EINVAL;
5132 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5134 spin_lock_irq(&ctx->completion_lock);
5135 if (likely(poll->head)) {
5136 spin_lock(&poll->head->lock);
5137 if (unlikely(list_empty(&poll->wait.entry))) {
5143 if (mask || ipt->error)
5144 list_del_init(&poll->wait.entry);
5146 WRITE_ONCE(poll->canceled, true);
5147 else if (!poll->done) /* actually waiting for an event */
5148 io_poll_req_insert(req);
5149 spin_unlock(&poll->head->lock);
5155 static bool io_arm_poll_handler(struct io_kiocb *req)
5157 const struct io_op_def *def = &io_op_defs[req->opcode];
5158 struct io_ring_ctx *ctx = req->ctx;
5159 struct async_poll *apoll;
5160 struct io_poll_table ipt;
5164 if (!req->file || !file_can_poll(req->file))
5166 if (req->flags & REQ_F_POLLED)
5170 else if (def->pollout)
5174 /* if we can't nonblock try, then no point in arming a poll handler */
5175 if (!io_file_supports_async(req->file, rw))
5178 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5179 if (unlikely(!apoll))
5181 apoll->double_poll = NULL;
5183 req->flags |= REQ_F_POLLED;
5188 mask |= POLLIN | POLLRDNORM;
5190 mask |= POLLOUT | POLLWRNORM;
5192 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5193 if ((req->opcode == IORING_OP_RECVMSG) &&
5194 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5197 mask |= POLLERR | POLLPRI;
5199 ipt.pt._qproc = io_async_queue_proc;
5201 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5203 if (ret || ipt.error) {
5204 io_poll_remove_double(req);
5205 spin_unlock_irq(&ctx->completion_lock);
5206 kfree(apoll->double_poll);
5210 spin_unlock_irq(&ctx->completion_lock);
5211 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5212 apoll->poll.events);
5216 static bool __io_poll_remove_one(struct io_kiocb *req,
5217 struct io_poll_iocb *poll)
5219 bool do_complete = false;
5221 spin_lock(&poll->head->lock);
5222 WRITE_ONCE(poll->canceled, true);
5223 if (!list_empty(&poll->wait.entry)) {
5224 list_del_init(&poll->wait.entry);
5227 spin_unlock(&poll->head->lock);
5228 hash_del(&req->hash_node);
5232 static bool io_poll_remove_one(struct io_kiocb *req)
5236 io_poll_remove_double(req);
5238 if (req->opcode == IORING_OP_POLL_ADD) {
5239 do_complete = __io_poll_remove_one(req, &req->poll);
5241 struct async_poll *apoll = req->apoll;
5243 /* non-poll requests have submit ref still */
5244 do_complete = __io_poll_remove_one(req, &apoll->poll);
5247 kfree(apoll->double_poll);
5253 io_cqring_fill_event(req, -ECANCELED);
5254 io_commit_cqring(req->ctx);
5255 req_set_fail_links(req);
5256 io_put_req_deferred(req, 1);
5263 * Returns true if we found and killed one or more poll requests
5265 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5267 struct hlist_node *tmp;
5268 struct io_kiocb *req;
5271 spin_lock_irq(&ctx->completion_lock);
5272 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5273 struct hlist_head *list;
5275 list = &ctx->cancel_hash[i];
5276 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5277 if (io_task_match(req, tsk))
5278 posted += io_poll_remove_one(req);
5281 spin_unlock_irq(&ctx->completion_lock);
5284 io_cqring_ev_posted(ctx);
5289 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5291 struct hlist_head *list;
5292 struct io_kiocb *req;
5294 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5295 hlist_for_each_entry(req, list, hash_node) {
5296 if (sqe_addr != req->user_data)
5298 if (io_poll_remove_one(req))
5306 static int io_poll_remove_prep(struct io_kiocb *req,
5307 const struct io_uring_sqe *sqe)
5309 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5311 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5315 req->poll.addr = READ_ONCE(sqe->addr);
5320 * Find a running poll command that matches one specified in sqe->addr,
5321 * and remove it if found.
5323 static int io_poll_remove(struct io_kiocb *req)
5325 struct io_ring_ctx *ctx = req->ctx;
5329 addr = req->poll.addr;
5330 spin_lock_irq(&ctx->completion_lock);
5331 ret = io_poll_cancel(ctx, addr);
5332 spin_unlock_irq(&ctx->completion_lock);
5335 req_set_fail_links(req);
5336 io_req_complete(req, ret);
5340 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5343 struct io_kiocb *req = wait->private;
5344 struct io_poll_iocb *poll = &req->poll;
5346 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5349 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5350 struct poll_table_struct *p)
5352 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5354 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5357 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5359 struct io_poll_iocb *poll = &req->poll;
5362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5364 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5367 events = READ_ONCE(sqe->poll32_events);
5369 events = swahw32(events);
5371 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5372 (events & EPOLLEXCLUSIVE);
5376 static int io_poll_add(struct io_kiocb *req)
5378 struct io_poll_iocb *poll = &req->poll;
5379 struct io_ring_ctx *ctx = req->ctx;
5380 struct io_poll_table ipt;
5383 ipt.pt._qproc = io_poll_queue_proc;
5385 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5388 if (mask) { /* no async, we'd stolen it */
5390 io_poll_complete(req, mask, 0);
5392 spin_unlock_irq(&ctx->completion_lock);
5395 io_cqring_ev_posted(ctx);
5401 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5403 struct io_timeout_data *data = container_of(timer,
5404 struct io_timeout_data, timer);
5405 struct io_kiocb *req = data->req;
5406 struct io_ring_ctx *ctx = req->ctx;
5407 unsigned long flags;
5409 spin_lock_irqsave(&ctx->completion_lock, flags);
5410 list_del_init(&req->timeout.list);
5411 atomic_set(&req->ctx->cq_timeouts,
5412 atomic_read(&req->ctx->cq_timeouts) + 1);
5414 io_cqring_fill_event(req, -ETIME);
5415 io_commit_cqring(ctx);
5416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5418 io_cqring_ev_posted(ctx);
5419 req_set_fail_links(req);
5421 return HRTIMER_NORESTART;
5424 static int __io_timeout_cancel(struct io_kiocb *req)
5426 struct io_timeout_data *io = req->async_data;
5429 ret = hrtimer_try_to_cancel(&io->timer);
5432 list_del_init(&req->timeout.list);
5434 req_set_fail_links(req);
5435 io_cqring_fill_event(req, -ECANCELED);
5436 io_put_req_deferred(req, 1);
5440 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5442 struct io_kiocb *req;
5445 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5446 if (user_data == req->user_data) {
5455 return __io_timeout_cancel(req);
5458 static int io_timeout_remove_prep(struct io_kiocb *req,
5459 const struct io_uring_sqe *sqe)
5461 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5463 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5465 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5468 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5473 * Remove or update an existing timeout command
5475 static int io_timeout_remove(struct io_kiocb *req)
5477 struct io_ring_ctx *ctx = req->ctx;
5480 spin_lock_irq(&ctx->completion_lock);
5481 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5483 io_cqring_fill_event(req, ret);
5484 io_commit_cqring(ctx);
5485 spin_unlock_irq(&ctx->completion_lock);
5486 io_cqring_ev_posted(ctx);
5488 req_set_fail_links(req);
5493 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5494 bool is_timeout_link)
5496 struct io_timeout_data *data;
5498 u32 off = READ_ONCE(sqe->off);
5500 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5502 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5504 if (off && is_timeout_link)
5506 flags = READ_ONCE(sqe->timeout_flags);
5507 if (flags & ~IORING_TIMEOUT_ABS)
5510 req->timeout.off = off;
5512 if (!req->async_data && io_alloc_async_data(req))
5515 data = req->async_data;
5518 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5521 if (flags & IORING_TIMEOUT_ABS)
5522 data->mode = HRTIMER_MODE_ABS;
5524 data->mode = HRTIMER_MODE_REL;
5526 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5530 static int io_timeout(struct io_kiocb *req)
5532 struct io_ring_ctx *ctx = req->ctx;
5533 struct io_timeout_data *data = req->async_data;
5534 struct list_head *entry;
5535 u32 tail, off = req->timeout.off;
5537 spin_lock_irq(&ctx->completion_lock);
5540 * sqe->off holds how many events that need to occur for this
5541 * timeout event to be satisfied. If it isn't set, then this is
5542 * a pure timeout request, sequence isn't used.
5544 if (io_is_timeout_noseq(req)) {
5545 entry = ctx->timeout_list.prev;
5549 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5550 req->timeout.target_seq = tail + off;
5553 * Insertion sort, ensuring the first entry in the list is always
5554 * the one we need first.
5556 list_for_each_prev(entry, &ctx->timeout_list) {
5557 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5560 if (io_is_timeout_noseq(nxt))
5562 /* nxt.seq is behind @tail, otherwise would've been completed */
5563 if (off >= nxt->timeout.target_seq - tail)
5567 list_add(&req->timeout.list, entry);
5568 data->timer.function = io_timeout_fn;
5569 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5570 spin_unlock_irq(&ctx->completion_lock);
5574 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5576 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5578 return req->user_data == (unsigned long) data;
5581 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5583 enum io_wq_cancel cancel_ret;
5586 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5587 switch (cancel_ret) {
5588 case IO_WQ_CANCEL_OK:
5591 case IO_WQ_CANCEL_RUNNING:
5594 case IO_WQ_CANCEL_NOTFOUND:
5602 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5603 struct io_kiocb *req, __u64 sqe_addr,
5606 unsigned long flags;
5609 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5610 if (ret != -ENOENT) {
5611 spin_lock_irqsave(&ctx->completion_lock, flags);
5615 spin_lock_irqsave(&ctx->completion_lock, flags);
5616 ret = io_timeout_cancel(ctx, sqe_addr);
5619 ret = io_poll_cancel(ctx, sqe_addr);
5623 io_cqring_fill_event(req, ret);
5624 io_commit_cqring(ctx);
5625 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5626 io_cqring_ev_posted(ctx);
5629 req_set_fail_links(req);
5633 static int io_async_cancel_prep(struct io_kiocb *req,
5634 const struct io_uring_sqe *sqe)
5636 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5638 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5640 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5643 req->cancel.addr = READ_ONCE(sqe->addr);
5647 static int io_async_cancel(struct io_kiocb *req)
5649 struct io_ring_ctx *ctx = req->ctx;
5651 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5655 static int io_files_update_prep(struct io_kiocb *req,
5656 const struct io_uring_sqe *sqe)
5658 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5660 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5662 if (sqe->ioprio || sqe->rw_flags)
5665 req->files_update.offset = READ_ONCE(sqe->off);
5666 req->files_update.nr_args = READ_ONCE(sqe->len);
5667 if (!req->files_update.nr_args)
5669 req->files_update.arg = READ_ONCE(sqe->addr);
5673 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5674 struct io_comp_state *cs)
5676 struct io_ring_ctx *ctx = req->ctx;
5677 struct io_uring_files_update up;
5683 up.offset = req->files_update.offset;
5684 up.fds = req->files_update.arg;
5686 mutex_lock(&ctx->uring_lock);
5687 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5688 mutex_unlock(&ctx->uring_lock);
5691 req_set_fail_links(req);
5692 __io_req_complete(req, ret, 0, cs);
5696 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5698 switch (req->opcode) {
5701 case IORING_OP_READV:
5702 case IORING_OP_READ_FIXED:
5703 case IORING_OP_READ:
5704 return io_read_prep(req, sqe);
5705 case IORING_OP_WRITEV:
5706 case IORING_OP_WRITE_FIXED:
5707 case IORING_OP_WRITE:
5708 return io_write_prep(req, sqe);
5709 case IORING_OP_POLL_ADD:
5710 return io_poll_add_prep(req, sqe);
5711 case IORING_OP_POLL_REMOVE:
5712 return io_poll_remove_prep(req, sqe);
5713 case IORING_OP_FSYNC:
5714 return io_prep_fsync(req, sqe);
5715 case IORING_OP_SYNC_FILE_RANGE:
5716 return io_prep_sfr(req, sqe);
5717 case IORING_OP_SENDMSG:
5718 case IORING_OP_SEND:
5719 return io_sendmsg_prep(req, sqe);
5720 case IORING_OP_RECVMSG:
5721 case IORING_OP_RECV:
5722 return io_recvmsg_prep(req, sqe);
5723 case IORING_OP_CONNECT:
5724 return io_connect_prep(req, sqe);
5725 case IORING_OP_TIMEOUT:
5726 return io_timeout_prep(req, sqe, false);
5727 case IORING_OP_TIMEOUT_REMOVE:
5728 return io_timeout_remove_prep(req, sqe);
5729 case IORING_OP_ASYNC_CANCEL:
5730 return io_async_cancel_prep(req, sqe);
5731 case IORING_OP_LINK_TIMEOUT:
5732 return io_timeout_prep(req, sqe, true);
5733 case IORING_OP_ACCEPT:
5734 return io_accept_prep(req, sqe);
5735 case IORING_OP_FALLOCATE:
5736 return io_fallocate_prep(req, sqe);
5737 case IORING_OP_OPENAT:
5738 return io_openat_prep(req, sqe);
5739 case IORING_OP_CLOSE:
5740 return io_close_prep(req, sqe);
5741 case IORING_OP_FILES_UPDATE:
5742 return io_files_update_prep(req, sqe);
5743 case IORING_OP_STATX:
5744 return io_statx_prep(req, sqe);
5745 case IORING_OP_FADVISE:
5746 return io_fadvise_prep(req, sqe);
5747 case IORING_OP_MADVISE:
5748 return io_madvise_prep(req, sqe);
5749 case IORING_OP_OPENAT2:
5750 return io_openat2_prep(req, sqe);
5751 case IORING_OP_EPOLL_CTL:
5752 return io_epoll_ctl_prep(req, sqe);
5753 case IORING_OP_SPLICE:
5754 return io_splice_prep(req, sqe);
5755 case IORING_OP_PROVIDE_BUFFERS:
5756 return io_provide_buffers_prep(req, sqe);
5757 case IORING_OP_REMOVE_BUFFERS:
5758 return io_remove_buffers_prep(req, sqe);
5760 return io_tee_prep(req, sqe);
5763 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5768 static int io_req_defer_prep(struct io_kiocb *req,
5769 const struct io_uring_sqe *sqe)
5773 if (io_alloc_async_data(req))
5775 return io_req_prep(req, sqe);
5778 static u32 io_get_sequence(struct io_kiocb *req)
5780 struct io_kiocb *pos;
5781 struct io_ring_ctx *ctx = req->ctx;
5782 u32 total_submitted, nr_reqs = 1;
5784 if (req->flags & REQ_F_LINK_HEAD)
5785 list_for_each_entry(pos, &req->link_list, link_list)
5788 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5789 return total_submitted - nr_reqs;
5792 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5794 struct io_ring_ctx *ctx = req->ctx;
5795 struct io_defer_entry *de;
5799 /* Still need defer if there is pending req in defer list. */
5800 if (likely(list_empty_careful(&ctx->defer_list) &&
5801 !(req->flags & REQ_F_IO_DRAIN)))
5804 seq = io_get_sequence(req);
5805 /* Still a chance to pass the sequence check */
5806 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5809 if (!req->async_data) {
5810 ret = io_req_defer_prep(req, sqe);
5814 io_prep_async_link(req);
5815 de = kmalloc(sizeof(*de), GFP_KERNEL);
5819 spin_lock_irq(&ctx->completion_lock);
5820 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5821 spin_unlock_irq(&ctx->completion_lock);
5823 io_queue_async_work(req);
5824 return -EIOCBQUEUED;
5827 trace_io_uring_defer(ctx, req, req->user_data);
5830 list_add_tail(&de->list, &ctx->defer_list);
5831 spin_unlock_irq(&ctx->completion_lock);
5832 return -EIOCBQUEUED;
5835 static void io_req_drop_files(struct io_kiocb *req)
5837 struct io_ring_ctx *ctx = req->ctx;
5838 unsigned long flags;
5840 spin_lock_irqsave(&ctx->inflight_lock, flags);
5841 list_del(&req->inflight_entry);
5842 if (waitqueue_active(&ctx->inflight_wait))
5843 wake_up(&ctx->inflight_wait);
5844 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5845 req->flags &= ~REQ_F_INFLIGHT;
5846 put_files_struct(req->work.identity->files);
5847 put_nsproxy(req->work.identity->nsproxy);
5848 req->work.flags &= ~IO_WQ_WORK_FILES;
5851 static void __io_clean_op(struct io_kiocb *req)
5853 if (req->flags & REQ_F_BUFFER_SELECTED) {
5854 switch (req->opcode) {
5855 case IORING_OP_READV:
5856 case IORING_OP_READ_FIXED:
5857 case IORING_OP_READ:
5858 kfree((void *)(unsigned long)req->rw.addr);
5860 case IORING_OP_RECVMSG:
5861 case IORING_OP_RECV:
5862 kfree(req->sr_msg.kbuf);
5865 req->flags &= ~REQ_F_BUFFER_SELECTED;
5868 if (req->flags & REQ_F_NEED_CLEANUP) {
5869 switch (req->opcode) {
5870 case IORING_OP_READV:
5871 case IORING_OP_READ_FIXED:
5872 case IORING_OP_READ:
5873 case IORING_OP_WRITEV:
5874 case IORING_OP_WRITE_FIXED:
5875 case IORING_OP_WRITE: {
5876 struct io_async_rw *io = req->async_data;
5878 kfree(io->free_iovec);
5881 case IORING_OP_RECVMSG:
5882 case IORING_OP_SENDMSG: {
5883 struct io_async_msghdr *io = req->async_data;
5884 if (io->iov != io->fast_iov)
5888 case IORING_OP_SPLICE:
5890 io_put_file(req, req->splice.file_in,
5891 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5893 case IORING_OP_OPENAT:
5894 case IORING_OP_OPENAT2:
5895 if (req->open.filename)
5896 putname(req->open.filename);
5899 req->flags &= ~REQ_F_NEED_CLEANUP;
5902 if (req->flags & REQ_F_INFLIGHT)
5903 io_req_drop_files(req);
5906 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5907 struct io_comp_state *cs)
5909 struct io_ring_ctx *ctx = req->ctx;
5912 switch (req->opcode) {
5914 ret = io_nop(req, cs);
5916 case IORING_OP_READV:
5917 case IORING_OP_READ_FIXED:
5918 case IORING_OP_READ:
5919 ret = io_read(req, force_nonblock, cs);
5921 case IORING_OP_WRITEV:
5922 case IORING_OP_WRITE_FIXED:
5923 case IORING_OP_WRITE:
5924 ret = io_write(req, force_nonblock, cs);
5926 case IORING_OP_FSYNC:
5927 ret = io_fsync(req, force_nonblock);
5929 case IORING_OP_POLL_ADD:
5930 ret = io_poll_add(req);
5932 case IORING_OP_POLL_REMOVE:
5933 ret = io_poll_remove(req);
5935 case IORING_OP_SYNC_FILE_RANGE:
5936 ret = io_sync_file_range(req, force_nonblock);
5938 case IORING_OP_SENDMSG:
5939 ret = io_sendmsg(req, force_nonblock, cs);
5941 case IORING_OP_SEND:
5942 ret = io_send(req, force_nonblock, cs);
5944 case IORING_OP_RECVMSG:
5945 ret = io_recvmsg(req, force_nonblock, cs);
5947 case IORING_OP_RECV:
5948 ret = io_recv(req, force_nonblock, cs);
5950 case IORING_OP_TIMEOUT:
5951 ret = io_timeout(req);
5953 case IORING_OP_TIMEOUT_REMOVE:
5954 ret = io_timeout_remove(req);
5956 case IORING_OP_ACCEPT:
5957 ret = io_accept(req, force_nonblock, cs);
5959 case IORING_OP_CONNECT:
5960 ret = io_connect(req, force_nonblock, cs);
5962 case IORING_OP_ASYNC_CANCEL:
5963 ret = io_async_cancel(req);
5965 case IORING_OP_FALLOCATE:
5966 ret = io_fallocate(req, force_nonblock);
5968 case IORING_OP_OPENAT:
5969 ret = io_openat(req, force_nonblock);
5971 case IORING_OP_CLOSE:
5972 ret = io_close(req, force_nonblock, cs);
5974 case IORING_OP_FILES_UPDATE:
5975 ret = io_files_update(req, force_nonblock, cs);
5977 case IORING_OP_STATX:
5978 ret = io_statx(req, force_nonblock);
5980 case IORING_OP_FADVISE:
5981 ret = io_fadvise(req, force_nonblock);
5983 case IORING_OP_MADVISE:
5984 ret = io_madvise(req, force_nonblock);
5986 case IORING_OP_OPENAT2:
5987 ret = io_openat2(req, force_nonblock);
5989 case IORING_OP_EPOLL_CTL:
5990 ret = io_epoll_ctl(req, force_nonblock, cs);
5992 case IORING_OP_SPLICE:
5993 ret = io_splice(req, force_nonblock);
5995 case IORING_OP_PROVIDE_BUFFERS:
5996 ret = io_provide_buffers(req, force_nonblock, cs);
5998 case IORING_OP_REMOVE_BUFFERS:
5999 ret = io_remove_buffers(req, force_nonblock, cs);
6002 ret = io_tee(req, force_nonblock);
6012 /* If the op doesn't have a file, we're not polling for it */
6013 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6014 const bool in_async = io_wq_current_is_worker();
6016 /* workqueue context doesn't hold uring_lock, grab it now */
6018 mutex_lock(&ctx->uring_lock);
6020 io_iopoll_req_issued(req);
6023 mutex_unlock(&ctx->uring_lock);
6029 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6031 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6032 struct io_kiocb *timeout;
6035 timeout = io_prep_linked_timeout(req);
6037 io_queue_linked_timeout(timeout);
6039 /* if NO_CANCEL is set, we must still run the work */
6040 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6041 IO_WQ_WORK_CANCEL) {
6047 ret = io_issue_sqe(req, false, NULL);
6049 * We can get EAGAIN for polled IO even though we're
6050 * forcing a sync submission from here, since we can't
6051 * wait for request slots on the block side.
6060 req_set_fail_links(req);
6061 io_req_complete(req, ret);
6064 return io_steal_work(req);
6067 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6070 struct fixed_file_table *table;
6072 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6073 return table->files[index & IORING_FILE_TABLE_MASK];
6076 static struct file *io_file_get(struct io_submit_state *state,
6077 struct io_kiocb *req, int fd, bool fixed)
6079 struct io_ring_ctx *ctx = req->ctx;
6083 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6085 fd = array_index_nospec(fd, ctx->nr_user_files);
6086 file = io_file_from_index(ctx, fd);
6088 req->fixed_file_refs = &ctx->file_data->node->refs;
6089 percpu_ref_get(req->fixed_file_refs);
6092 trace_io_uring_file_get(ctx, fd);
6093 file = __io_file_get(state, fd);
6099 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6104 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6105 if (unlikely(!fixed && io_async_submit(req->ctx)))
6108 req->file = io_file_get(state, req, fd, fixed);
6109 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6114 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6116 struct io_timeout_data *data = container_of(timer,
6117 struct io_timeout_data, timer);
6118 struct io_kiocb *req = data->req;
6119 struct io_ring_ctx *ctx = req->ctx;
6120 struct io_kiocb *prev = NULL;
6121 unsigned long flags;
6123 spin_lock_irqsave(&ctx->completion_lock, flags);
6126 * We don't expect the list to be empty, that will only happen if we
6127 * race with the completion of the linked work.
6129 if (!list_empty(&req->link_list)) {
6130 prev = list_entry(req->link_list.prev, struct io_kiocb,
6132 if (refcount_inc_not_zero(&prev->refs))
6133 list_del_init(&req->link_list);
6138 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6141 req_set_fail_links(prev);
6142 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6145 io_req_complete(req, -ETIME);
6147 return HRTIMER_NORESTART;
6150 static void __io_queue_linked_timeout(struct io_kiocb *req)
6153 * If the list is now empty, then our linked request finished before
6154 * we got a chance to setup the timer
6156 if (!list_empty(&req->link_list)) {
6157 struct io_timeout_data *data = req->async_data;
6159 data->timer.function = io_link_timeout_fn;
6160 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6165 static void io_queue_linked_timeout(struct io_kiocb *req)
6167 struct io_ring_ctx *ctx = req->ctx;
6169 spin_lock_irq(&ctx->completion_lock);
6170 __io_queue_linked_timeout(req);
6171 spin_unlock_irq(&ctx->completion_lock);
6173 /* drop submission reference */
6177 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6179 struct io_kiocb *nxt;
6181 if (!(req->flags & REQ_F_LINK_HEAD))
6183 if (req->flags & REQ_F_LINK_TIMEOUT)
6186 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6188 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6191 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6192 req->flags |= REQ_F_LINK_TIMEOUT;
6196 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6198 struct io_kiocb *linked_timeout;
6199 const struct cred *old_creds = NULL;
6203 linked_timeout = io_prep_linked_timeout(req);
6205 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6206 (req->work.flags & IO_WQ_WORK_CREDS) &&
6207 req->work.identity->creds != current_cred()) {
6209 revert_creds(old_creds);
6210 if (old_creds == req->work.identity->creds)
6211 old_creds = NULL; /* restored original creds */
6213 old_creds = override_creds(req->work.identity->creds);
6216 ret = io_issue_sqe(req, true, cs);
6219 * We async punt it if the file wasn't marked NOWAIT, or if the file
6220 * doesn't support non-blocking read/write attempts
6222 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6223 if (!io_arm_poll_handler(req)) {
6225 * Queued up for async execution, worker will release
6226 * submit reference when the iocb is actually submitted.
6228 io_queue_async_work(req);
6232 io_queue_linked_timeout(linked_timeout);
6233 } else if (likely(!ret)) {
6234 /* drop submission reference */
6235 req = io_put_req_find_next(req);
6237 io_queue_linked_timeout(linked_timeout);
6240 if (!(req->flags & REQ_F_FORCE_ASYNC))
6242 io_queue_async_work(req);
6245 /* un-prep timeout, so it'll be killed as any other linked */
6246 req->flags &= ~REQ_F_LINK_TIMEOUT;
6247 req_set_fail_links(req);
6249 io_req_complete(req, ret);
6253 revert_creds(old_creds);
6256 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6257 struct io_comp_state *cs)
6261 ret = io_req_defer(req, sqe);
6263 if (ret != -EIOCBQUEUED) {
6265 req_set_fail_links(req);
6267 io_req_complete(req, ret);
6269 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6270 if (!req->async_data) {
6271 ret = io_req_defer_prep(req, sqe);
6275 io_queue_async_work(req);
6278 ret = io_req_prep(req, sqe);
6282 __io_queue_sqe(req, cs);
6286 static inline void io_queue_link_head(struct io_kiocb *req,
6287 struct io_comp_state *cs)
6289 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6291 io_req_complete(req, -ECANCELED);
6293 io_queue_sqe(req, NULL, cs);
6296 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6297 struct io_kiocb **link, struct io_comp_state *cs)
6299 struct io_ring_ctx *ctx = req->ctx;
6303 * If we already have a head request, queue this one for async
6304 * submittal once the head completes. If we don't have a head but
6305 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6306 * submitted sync once the chain is complete. If none of those
6307 * conditions are true (normal request), then just queue it.
6310 struct io_kiocb *head = *link;
6313 * Taking sequential execution of a link, draining both sides
6314 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6315 * requests in the link. So, it drains the head and the
6316 * next after the link request. The last one is done via
6317 * drain_next flag to persist the effect across calls.
6319 if (req->flags & REQ_F_IO_DRAIN) {
6320 head->flags |= REQ_F_IO_DRAIN;
6321 ctx->drain_next = 1;
6323 ret = io_req_defer_prep(req, sqe);
6324 if (unlikely(ret)) {
6325 /* fail even hard links since we don't submit */
6326 head->flags |= REQ_F_FAIL_LINK;
6329 trace_io_uring_link(ctx, req, head);
6330 list_add_tail(&req->link_list, &head->link_list);
6332 /* last request of a link, enqueue the link */
6333 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6334 io_queue_link_head(head, cs);
6338 if (unlikely(ctx->drain_next)) {
6339 req->flags |= REQ_F_IO_DRAIN;
6340 ctx->drain_next = 0;
6342 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6343 req->flags |= REQ_F_LINK_HEAD;
6344 INIT_LIST_HEAD(&req->link_list);
6346 ret = io_req_defer_prep(req, sqe);
6348 req->flags |= REQ_F_FAIL_LINK;
6351 io_queue_sqe(req, sqe, cs);
6359 * Batched submission is done, ensure local IO is flushed out.
6361 static void io_submit_state_end(struct io_submit_state *state)
6363 if (!list_empty(&state->comp.list))
6364 io_submit_flush_completions(&state->comp);
6365 blk_finish_plug(&state->plug);
6366 io_state_file_put(state);
6367 if (state->free_reqs)
6368 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6372 * Start submission side cache.
6374 static void io_submit_state_start(struct io_submit_state *state,
6375 struct io_ring_ctx *ctx, unsigned int max_ios)
6377 blk_start_plug(&state->plug);
6379 INIT_LIST_HEAD(&state->comp.list);
6380 state->comp.ctx = ctx;
6381 state->free_reqs = 0;
6383 state->ios_left = max_ios;
6386 static void io_commit_sqring(struct io_ring_ctx *ctx)
6388 struct io_rings *rings = ctx->rings;
6391 * Ensure any loads from the SQEs are done at this point,
6392 * since once we write the new head, the application could
6393 * write new data to them.
6395 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6399 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6400 * that is mapped by userspace. This means that care needs to be taken to
6401 * ensure that reads are stable, as we cannot rely on userspace always
6402 * being a good citizen. If members of the sqe are validated and then later
6403 * used, it's important that those reads are done through READ_ONCE() to
6404 * prevent a re-load down the line.
6406 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6408 u32 *sq_array = ctx->sq_array;
6412 * The cached sq head (or cq tail) serves two purposes:
6414 * 1) allows us to batch the cost of updating the user visible
6416 * 2) allows the kernel side to track the head on its own, even
6417 * though the application is the one updating it.
6419 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6420 if (likely(head < ctx->sq_entries))
6421 return &ctx->sq_sqes[head];
6423 /* drop invalid entries */
6424 ctx->cached_sq_dropped++;
6425 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6429 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6431 ctx->cached_sq_head++;
6435 * Check SQE restrictions (opcode and flags).
6437 * Returns 'true' if SQE is allowed, 'false' otherwise.
6439 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6440 struct io_kiocb *req,
6441 unsigned int sqe_flags)
6443 if (!ctx->restricted)
6446 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6449 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6450 ctx->restrictions.sqe_flags_required)
6453 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6454 ctx->restrictions.sqe_flags_required))
6460 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6461 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6462 IOSQE_BUFFER_SELECT)
6464 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6465 const struct io_uring_sqe *sqe,
6466 struct io_submit_state *state)
6468 unsigned int sqe_flags;
6471 req->opcode = READ_ONCE(sqe->opcode);
6472 req->user_data = READ_ONCE(sqe->user_data);
6473 req->async_data = NULL;
6477 /* one is dropped after submission, the other at completion */
6478 refcount_set(&req->refs, 2);
6479 req->task = current;
6482 if (unlikely(req->opcode >= IORING_OP_LAST))
6485 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6488 sqe_flags = READ_ONCE(sqe->flags);
6489 /* enforce forwards compatibility on users */
6490 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6493 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6496 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6497 !io_op_defs[req->opcode].buffer_select)
6500 id = READ_ONCE(sqe->personality);
6502 struct io_identity *iod;
6504 iod = idr_find(&ctx->personality_idr, id);
6507 refcount_inc(&iod->count);
6509 __io_req_init_async(req);
6510 get_cred(iod->creds);
6511 req->work.identity = iod;
6512 req->work.flags |= IO_WQ_WORK_CREDS;
6515 /* same numerical values with corresponding REQ_F_*, safe to copy */
6516 req->flags |= sqe_flags;
6518 if (!io_op_defs[req->opcode].needs_file)
6521 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6526 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6528 struct io_submit_state state;
6529 struct io_kiocb *link = NULL;
6530 int i, submitted = 0;
6532 /* if we have a backlog and couldn't flush it all, return BUSY */
6533 if (test_bit(0, &ctx->sq_check_overflow)) {
6534 if (!list_empty(&ctx->cq_overflow_list) &&
6535 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6539 /* make sure SQ entry isn't read before tail */
6540 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6542 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6545 percpu_counter_add(¤t->io_uring->inflight, nr);
6546 refcount_add(nr, ¤t->usage);
6548 io_submit_state_start(&state, ctx, nr);
6550 for (i = 0; i < nr; i++) {
6551 const struct io_uring_sqe *sqe;
6552 struct io_kiocb *req;
6555 sqe = io_get_sqe(ctx);
6556 if (unlikely(!sqe)) {
6557 io_consume_sqe(ctx);
6560 req = io_alloc_req(ctx, &state);
6561 if (unlikely(!req)) {
6563 submitted = -EAGAIN;
6566 io_consume_sqe(ctx);
6567 /* will complete beyond this point, count as submitted */
6570 err = io_init_req(ctx, req, sqe, &state);
6571 if (unlikely(err)) {
6574 io_req_complete(req, err);
6578 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6579 true, io_async_submit(ctx));
6580 err = io_submit_sqe(req, sqe, &link, &state.comp);
6585 if (unlikely(submitted != nr)) {
6586 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6587 struct io_uring_task *tctx = current->io_uring;
6588 int unused = nr - ref_used;
6590 percpu_ref_put_many(&ctx->refs, unused);
6591 percpu_counter_sub(&tctx->inflight, unused);
6592 put_task_struct_many(current, unused);
6595 io_queue_link_head(link, &state.comp);
6596 io_submit_state_end(&state);
6598 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6599 io_commit_sqring(ctx);
6604 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6606 /* Tell userspace we may need a wakeup call */
6607 spin_lock_irq(&ctx->completion_lock);
6608 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6609 spin_unlock_irq(&ctx->completion_lock);
6612 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6614 spin_lock_irq(&ctx->completion_lock);
6615 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6616 spin_unlock_irq(&ctx->completion_lock);
6619 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6620 int sync, void *key)
6622 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6625 ret = autoremove_wake_function(wqe, mode, sync, key);
6627 unsigned long flags;
6629 spin_lock_irqsave(&ctx->completion_lock, flags);
6630 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6631 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6642 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6643 unsigned long start_jiffies, bool cap_entries)
6645 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6646 struct io_sq_data *sqd = ctx->sq_data;
6647 unsigned int to_submit;
6651 if (!list_empty(&ctx->iopoll_list)) {
6652 unsigned nr_events = 0;
6654 mutex_lock(&ctx->uring_lock);
6655 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6656 io_do_iopoll(ctx, &nr_events, 0);
6657 mutex_unlock(&ctx->uring_lock);
6660 to_submit = io_sqring_entries(ctx);
6663 * If submit got -EBUSY, flag us as needing the application
6664 * to enter the kernel to reap and flush events.
6666 if (!to_submit || ret == -EBUSY || need_resched()) {
6668 * Drop cur_mm before scheduling, we can't hold it for
6669 * long periods (or over schedule()). Do this before
6670 * adding ourselves to the waitqueue, as the unuse/drop
6673 io_sq_thread_drop_mm();
6676 * We're polling. If we're within the defined idle
6677 * period, then let us spin without work before going
6678 * to sleep. The exception is if we got EBUSY doing
6679 * more IO, we should wait for the application to
6680 * reap events and wake us up.
6682 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6683 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6684 !percpu_ref_is_dying(&ctx->refs)))
6687 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6688 TASK_INTERRUPTIBLE);
6691 * While doing polled IO, before going to sleep, we need
6692 * to check if there are new reqs added to iopoll_list,
6693 * it is because reqs may have been punted to io worker
6694 * and will be added to iopoll_list later, hence check
6695 * the iopoll_list again.
6697 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6698 !list_empty_careful(&ctx->iopoll_list)) {
6699 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6703 to_submit = io_sqring_entries(ctx);
6704 if (!to_submit || ret == -EBUSY)
6708 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6709 io_ring_clear_wakeup_flag(ctx);
6711 /* if we're handling multiple rings, cap submit size for fairness */
6712 if (cap_entries && to_submit > 8)
6715 mutex_lock(&ctx->uring_lock);
6716 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6717 ret = io_submit_sqes(ctx, to_submit);
6718 mutex_unlock(&ctx->uring_lock);
6720 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6721 wake_up(&ctx->sqo_sq_wait);
6723 return SQT_DID_WORK;
6726 static void io_sqd_init_new(struct io_sq_data *sqd)
6728 struct io_ring_ctx *ctx;
6730 while (!list_empty(&sqd->ctx_new_list)) {
6731 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6732 init_wait(&ctx->sqo_wait_entry);
6733 ctx->sqo_wait_entry.func = io_sq_wake_function;
6734 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6735 complete(&ctx->sq_thread_comp);
6739 static int io_sq_thread(void *data)
6741 struct cgroup_subsys_state *cur_css = NULL;
6742 const struct cred *old_cred = NULL;
6743 struct io_sq_data *sqd = data;
6744 struct io_ring_ctx *ctx;
6745 unsigned long start_jiffies;
6747 start_jiffies = jiffies;
6748 while (!kthread_should_stop()) {
6749 enum sq_ret ret = 0;
6753 * Any changes to the sqd lists are synchronized through the
6754 * kthread parking. This synchronizes the thread vs users,
6755 * the users are synchronized on the sqd->ctx_lock.
6757 if (kthread_should_park())
6760 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6761 io_sqd_init_new(sqd);
6763 cap_entries = !list_is_singular(&sqd->ctx_list);
6765 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6766 if (current->cred != ctx->creds) {
6768 revert_creds(old_cred);
6769 old_cred = override_creds(ctx->creds);
6771 io_sq_thread_associate_blkcg(ctx, &cur_css);
6773 current->loginuid = ctx->loginuid;
6774 current->sessionid = ctx->sessionid;
6777 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6779 io_sq_thread_drop_mm();
6782 if (ret & SQT_SPIN) {
6785 } else if (ret == SQT_IDLE) {
6786 if (kthread_should_park())
6788 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6789 io_ring_set_wakeup_flag(ctx);
6791 start_jiffies = jiffies;
6792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6793 io_ring_clear_wakeup_flag(ctx);
6800 io_sq_thread_unassociate_blkcg();
6802 revert_creds(old_cred);
6809 struct io_wait_queue {
6810 struct wait_queue_entry wq;
6811 struct io_ring_ctx *ctx;
6813 unsigned nr_timeouts;
6816 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6818 struct io_ring_ctx *ctx = iowq->ctx;
6821 * Wake up if we have enough events, or if a timeout occurred since we
6822 * started waiting. For timeouts, we always want to return to userspace,
6823 * regardless of event count.
6825 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6826 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6829 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6830 int wake_flags, void *key)
6832 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6835 /* use noflush == true, as we can't safely rely on locking context */
6836 if (!io_should_wake(iowq, true))
6839 return autoremove_wake_function(curr, mode, wake_flags, key);
6842 static int io_run_task_work_sig(void)
6844 if (io_run_task_work())
6846 if (!signal_pending(current))
6848 if (current->jobctl & JOBCTL_TASK_WORK) {
6849 spin_lock_irq(¤t->sighand->siglock);
6850 current->jobctl &= ~JOBCTL_TASK_WORK;
6851 recalc_sigpending();
6852 spin_unlock_irq(¤t->sighand->siglock);
6859 * Wait until events become available, if we don't already have some. The
6860 * application must reap them itself, as they reside on the shared cq ring.
6862 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6863 const sigset_t __user *sig, size_t sigsz)
6865 struct io_wait_queue iowq = {
6868 .func = io_wake_function,
6869 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6872 .to_wait = min_events,
6874 struct io_rings *rings = ctx->rings;
6878 if (io_cqring_events(ctx, false) >= min_events)
6880 if (!io_run_task_work())
6885 #ifdef CONFIG_COMPAT
6886 if (in_compat_syscall())
6887 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6891 ret = set_user_sigmask(sig, sigsz);
6897 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6898 trace_io_uring_cqring_wait(ctx, min_events);
6900 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6901 TASK_INTERRUPTIBLE);
6902 /* make sure we run task_work before checking for signals */
6903 ret = io_run_task_work_sig();
6908 if (io_should_wake(&iowq, false))
6912 finish_wait(&ctx->wait, &iowq.wq);
6914 restore_saved_sigmask_unless(ret == -EINTR);
6916 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6919 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6921 #if defined(CONFIG_UNIX)
6922 if (ctx->ring_sock) {
6923 struct sock *sock = ctx->ring_sock->sk;
6924 struct sk_buff *skb;
6926 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6932 for (i = 0; i < ctx->nr_user_files; i++) {
6935 file = io_file_from_index(ctx, i);
6942 static void io_file_ref_kill(struct percpu_ref *ref)
6944 struct fixed_file_data *data;
6946 data = container_of(ref, struct fixed_file_data, refs);
6947 complete(&data->done);
6950 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6952 struct fixed_file_data *data = ctx->file_data;
6953 struct fixed_file_ref_node *ref_node = NULL;
6954 unsigned nr_tables, i;
6959 spin_lock(&data->lock);
6960 if (!list_empty(&data->ref_list))
6961 ref_node = list_first_entry(&data->ref_list,
6962 struct fixed_file_ref_node, node);
6963 spin_unlock(&data->lock);
6965 percpu_ref_kill(&ref_node->refs);
6967 percpu_ref_kill(&data->refs);
6969 /* wait for all refs nodes to complete */
6970 flush_delayed_work(&ctx->file_put_work);
6971 wait_for_completion(&data->done);
6973 __io_sqe_files_unregister(ctx);
6974 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6975 for (i = 0; i < nr_tables; i++)
6976 kfree(data->table[i].files);
6978 percpu_ref_exit(&data->refs);
6980 ctx->file_data = NULL;
6981 ctx->nr_user_files = 0;
6985 static void io_put_sq_data(struct io_sq_data *sqd)
6987 if (refcount_dec_and_test(&sqd->refs)) {
6989 * The park is a bit of a work-around, without it we get
6990 * warning spews on shutdown with SQPOLL set and affinity
6991 * set to a single CPU.
6994 kthread_park(sqd->thread);
6995 kthread_stop(sqd->thread);
7002 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7004 struct io_ring_ctx *ctx_attach;
7005 struct io_sq_data *sqd;
7008 f = fdget(p->wq_fd);
7010 return ERR_PTR(-ENXIO);
7011 if (f.file->f_op != &io_uring_fops) {
7013 return ERR_PTR(-EINVAL);
7016 ctx_attach = f.file->private_data;
7017 sqd = ctx_attach->sq_data;
7020 return ERR_PTR(-EINVAL);
7023 refcount_inc(&sqd->refs);
7028 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7030 struct io_sq_data *sqd;
7032 if (p->flags & IORING_SETUP_ATTACH_WQ)
7033 return io_attach_sq_data(p);
7035 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7037 return ERR_PTR(-ENOMEM);
7039 refcount_set(&sqd->refs, 1);
7040 INIT_LIST_HEAD(&sqd->ctx_list);
7041 INIT_LIST_HEAD(&sqd->ctx_new_list);
7042 mutex_init(&sqd->ctx_lock);
7043 mutex_init(&sqd->lock);
7044 init_waitqueue_head(&sqd->wait);
7048 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7049 __releases(&sqd->lock)
7053 kthread_unpark(sqd->thread);
7054 mutex_unlock(&sqd->lock);
7057 static void io_sq_thread_park(struct io_sq_data *sqd)
7058 __acquires(&sqd->lock)
7062 mutex_lock(&sqd->lock);
7063 kthread_park(sqd->thread);
7066 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7068 struct io_sq_data *sqd = ctx->sq_data;
7073 * We may arrive here from the error branch in
7074 * io_sq_offload_create() where the kthread is created
7075 * without being waked up, thus wake it up now to make
7076 * sure the wait will complete.
7078 wake_up_process(sqd->thread);
7079 wait_for_completion(&ctx->sq_thread_comp);
7081 io_sq_thread_park(sqd);
7084 mutex_lock(&sqd->ctx_lock);
7085 list_del(&ctx->sqd_list);
7086 mutex_unlock(&sqd->ctx_lock);
7089 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7090 io_sq_thread_unpark(sqd);
7093 io_put_sq_data(sqd);
7094 ctx->sq_data = NULL;
7098 static void io_finish_async(struct io_ring_ctx *ctx)
7100 io_sq_thread_stop(ctx);
7103 io_wq_destroy(ctx->io_wq);
7108 #if defined(CONFIG_UNIX)
7110 * Ensure the UNIX gc is aware of our file set, so we are certain that
7111 * the io_uring can be safely unregistered on process exit, even if we have
7112 * loops in the file referencing.
7114 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7116 struct sock *sk = ctx->ring_sock->sk;
7117 struct scm_fp_list *fpl;
7118 struct sk_buff *skb;
7121 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7125 skb = alloc_skb(0, GFP_KERNEL);
7134 fpl->user = get_uid(ctx->user);
7135 for (i = 0; i < nr; i++) {
7136 struct file *file = io_file_from_index(ctx, i + offset);
7140 fpl->fp[nr_files] = get_file(file);
7141 unix_inflight(fpl->user, fpl->fp[nr_files]);
7146 fpl->max = SCM_MAX_FD;
7147 fpl->count = nr_files;
7148 UNIXCB(skb).fp = fpl;
7149 skb->destructor = unix_destruct_scm;
7150 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7151 skb_queue_head(&sk->sk_receive_queue, skb);
7153 for (i = 0; i < nr_files; i++)
7164 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7165 * causes regular reference counting to break down. We rely on the UNIX
7166 * garbage collection to take care of this problem for us.
7168 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7170 unsigned left, total;
7174 left = ctx->nr_user_files;
7176 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7178 ret = __io_sqe_files_scm(ctx, this_files, total);
7182 total += this_files;
7188 while (total < ctx->nr_user_files) {
7189 struct file *file = io_file_from_index(ctx, total);
7199 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7205 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7206 unsigned nr_tables, unsigned nr_files)
7210 for (i = 0; i < nr_tables; i++) {
7211 struct fixed_file_table *table = &file_data->table[i];
7212 unsigned this_files;
7214 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7215 table->files = kcalloc(this_files, sizeof(struct file *),
7219 nr_files -= this_files;
7225 for (i = 0; i < nr_tables; i++) {
7226 struct fixed_file_table *table = &file_data->table[i];
7227 kfree(table->files);
7232 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7234 #if defined(CONFIG_UNIX)
7235 struct sock *sock = ctx->ring_sock->sk;
7236 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7237 struct sk_buff *skb;
7240 __skb_queue_head_init(&list);
7243 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7244 * remove this entry and rearrange the file array.
7246 skb = skb_dequeue(head);
7248 struct scm_fp_list *fp;
7250 fp = UNIXCB(skb).fp;
7251 for (i = 0; i < fp->count; i++) {
7254 if (fp->fp[i] != file)
7257 unix_notinflight(fp->user, fp->fp[i]);
7258 left = fp->count - 1 - i;
7260 memmove(&fp->fp[i], &fp->fp[i + 1],
7261 left * sizeof(struct file *));
7268 __skb_queue_tail(&list, skb);
7278 __skb_queue_tail(&list, skb);
7280 skb = skb_dequeue(head);
7283 if (skb_peek(&list)) {
7284 spin_lock_irq(&head->lock);
7285 while ((skb = __skb_dequeue(&list)) != NULL)
7286 __skb_queue_tail(head, skb);
7287 spin_unlock_irq(&head->lock);
7294 struct io_file_put {
7295 struct list_head list;
7299 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7301 struct fixed_file_data *file_data = ref_node->file_data;
7302 struct io_ring_ctx *ctx = file_data->ctx;
7303 struct io_file_put *pfile, *tmp;
7305 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7306 list_del(&pfile->list);
7307 io_ring_file_put(ctx, pfile->file);
7311 spin_lock(&file_data->lock);
7312 list_del(&ref_node->node);
7313 spin_unlock(&file_data->lock);
7315 percpu_ref_exit(&ref_node->refs);
7317 percpu_ref_put(&file_data->refs);
7320 static void io_file_put_work(struct work_struct *work)
7322 struct io_ring_ctx *ctx;
7323 struct llist_node *node;
7325 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7326 node = llist_del_all(&ctx->file_put_llist);
7329 struct fixed_file_ref_node *ref_node;
7330 struct llist_node *next = node->next;
7332 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7333 __io_file_put_work(ref_node);
7338 static void io_file_data_ref_zero(struct percpu_ref *ref)
7340 struct fixed_file_ref_node *ref_node;
7341 struct io_ring_ctx *ctx;
7345 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7346 ctx = ref_node->file_data->ctx;
7348 if (percpu_ref_is_dying(&ctx->file_data->refs))
7351 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7353 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7355 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7358 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7359 struct io_ring_ctx *ctx)
7361 struct fixed_file_ref_node *ref_node;
7363 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7365 return ERR_PTR(-ENOMEM);
7367 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7370 return ERR_PTR(-ENOMEM);
7372 INIT_LIST_HEAD(&ref_node->node);
7373 INIT_LIST_HEAD(&ref_node->file_list);
7374 ref_node->file_data = ctx->file_data;
7378 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7380 percpu_ref_exit(&ref_node->refs);
7384 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7387 __s32 __user *fds = (__s32 __user *) arg;
7388 unsigned nr_tables, i;
7390 int fd, ret = -ENOMEM;
7391 struct fixed_file_ref_node *ref_node;
7392 struct fixed_file_data *file_data;
7398 if (nr_args > IORING_MAX_FIXED_FILES)
7401 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7404 file_data->ctx = ctx;
7405 init_completion(&file_data->done);
7406 INIT_LIST_HEAD(&file_data->ref_list);
7407 spin_lock_init(&file_data->lock);
7409 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7410 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7412 if (!file_data->table)
7415 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7416 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7419 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7421 ctx->file_data = file_data;
7423 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7424 struct fixed_file_table *table;
7427 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7431 /* allow sparse sets */
7441 * Don't allow io_uring instances to be registered. If UNIX
7442 * isn't enabled, then this causes a reference cycle and this
7443 * instance can never get freed. If UNIX is enabled we'll
7444 * handle it just fine, but there's still no point in allowing
7445 * a ring fd as it doesn't support regular read/write anyway.
7447 if (file->f_op == &io_uring_fops) {
7451 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7452 index = i & IORING_FILE_TABLE_MASK;
7453 table->files[index] = file;
7456 ret = io_sqe_files_scm(ctx);
7458 io_sqe_files_unregister(ctx);
7462 ref_node = alloc_fixed_file_ref_node(ctx);
7463 if (IS_ERR(ref_node)) {
7464 io_sqe_files_unregister(ctx);
7465 return PTR_ERR(ref_node);
7468 file_data->node = ref_node;
7469 spin_lock(&file_data->lock);
7470 list_add(&ref_node->node, &file_data->ref_list);
7471 spin_unlock(&file_data->lock);
7472 percpu_ref_get(&file_data->refs);
7475 for (i = 0; i < ctx->nr_user_files; i++) {
7476 file = io_file_from_index(ctx, i);
7480 for (i = 0; i < nr_tables; i++)
7481 kfree(file_data->table[i].files);
7482 ctx->nr_user_files = 0;
7484 percpu_ref_exit(&file_data->refs);
7486 kfree(file_data->table);
7488 ctx->file_data = NULL;
7492 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7495 #if defined(CONFIG_UNIX)
7496 struct sock *sock = ctx->ring_sock->sk;
7497 struct sk_buff_head *head = &sock->sk_receive_queue;
7498 struct sk_buff *skb;
7501 * See if we can merge this file into an existing skb SCM_RIGHTS
7502 * file set. If there's no room, fall back to allocating a new skb
7503 * and filling it in.
7505 spin_lock_irq(&head->lock);
7506 skb = skb_peek(head);
7508 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7510 if (fpl->count < SCM_MAX_FD) {
7511 __skb_unlink(skb, head);
7512 spin_unlock_irq(&head->lock);
7513 fpl->fp[fpl->count] = get_file(file);
7514 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7516 spin_lock_irq(&head->lock);
7517 __skb_queue_head(head, skb);
7522 spin_unlock_irq(&head->lock);
7529 return __io_sqe_files_scm(ctx, 1, index);
7535 static int io_queue_file_removal(struct fixed_file_data *data,
7538 struct io_file_put *pfile;
7539 struct fixed_file_ref_node *ref_node = data->node;
7541 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7546 list_add(&pfile->list, &ref_node->file_list);
7551 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7552 struct io_uring_files_update *up,
7555 struct fixed_file_data *data = ctx->file_data;
7556 struct fixed_file_ref_node *ref_node;
7561 bool needs_switch = false;
7563 if (check_add_overflow(up->offset, nr_args, &done))
7565 if (done > ctx->nr_user_files)
7568 ref_node = alloc_fixed_file_ref_node(ctx);
7569 if (IS_ERR(ref_node))
7570 return PTR_ERR(ref_node);
7573 fds = u64_to_user_ptr(up->fds);
7575 struct fixed_file_table *table;
7579 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7583 i = array_index_nospec(up->offset, ctx->nr_user_files);
7584 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7585 index = i & IORING_FILE_TABLE_MASK;
7586 if (table->files[index]) {
7587 file = table->files[index];
7588 err = io_queue_file_removal(data, file);
7591 table->files[index] = NULL;
7592 needs_switch = true;
7601 * Don't allow io_uring instances to be registered. If
7602 * UNIX isn't enabled, then this causes a reference
7603 * cycle and this instance can never get freed. If UNIX
7604 * is enabled we'll handle it just fine, but there's
7605 * still no point in allowing a ring fd as it doesn't
7606 * support regular read/write anyway.
7608 if (file->f_op == &io_uring_fops) {
7613 table->files[index] = file;
7614 err = io_sqe_file_register(ctx, file, i);
7616 table->files[index] = NULL;
7627 percpu_ref_kill(&data->node->refs);
7628 spin_lock(&data->lock);
7629 list_add(&ref_node->node, &data->ref_list);
7630 data->node = ref_node;
7631 spin_unlock(&data->lock);
7632 percpu_ref_get(&ctx->file_data->refs);
7634 destroy_fixed_file_ref_node(ref_node);
7636 return done ? done : err;
7639 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7642 struct io_uring_files_update up;
7644 if (!ctx->file_data)
7648 if (copy_from_user(&up, arg, sizeof(up)))
7653 return __io_sqe_files_update(ctx, &up, nr_args);
7656 static void io_free_work(struct io_wq_work *work)
7658 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7660 /* Consider that io_steal_work() relies on this ref */
7664 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7665 struct io_uring_params *p)
7667 struct io_wq_data data;
7669 struct io_ring_ctx *ctx_attach;
7670 unsigned int concurrency;
7673 data.user = ctx->user;
7674 data.free_work = io_free_work;
7675 data.do_work = io_wq_submit_work;
7677 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7678 /* Do QD, or 4 * CPUS, whatever is smallest */
7679 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7681 ctx->io_wq = io_wq_create(concurrency, &data);
7682 if (IS_ERR(ctx->io_wq)) {
7683 ret = PTR_ERR(ctx->io_wq);
7689 f = fdget(p->wq_fd);
7693 if (f.file->f_op != &io_uring_fops) {
7698 ctx_attach = f.file->private_data;
7699 /* @io_wq is protected by holding the fd */
7700 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7705 ctx->io_wq = ctx_attach->io_wq;
7711 static int io_uring_alloc_task_context(struct task_struct *task)
7713 struct io_uring_task *tctx;
7716 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7717 if (unlikely(!tctx))
7720 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7721 if (unlikely(ret)) {
7727 init_waitqueue_head(&tctx->wait);
7729 atomic_set(&tctx->in_idle, 0);
7730 tctx->sqpoll = false;
7731 io_init_identity(&tctx->__identity);
7732 tctx->identity = &tctx->__identity;
7733 task->io_uring = tctx;
7737 void __io_uring_free(struct task_struct *tsk)
7739 struct io_uring_task *tctx = tsk->io_uring;
7741 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7742 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7743 if (tctx->identity != &tctx->__identity)
7744 kfree(tctx->identity);
7745 percpu_counter_destroy(&tctx->inflight);
7747 tsk->io_uring = NULL;
7750 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7751 struct io_uring_params *p)
7755 if (ctx->flags & IORING_SETUP_SQPOLL) {
7756 struct io_sq_data *sqd;
7759 if (!capable(CAP_SYS_ADMIN))
7762 sqd = io_get_sq_data(p);
7769 io_sq_thread_park(sqd);
7770 mutex_lock(&sqd->ctx_lock);
7771 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7772 mutex_unlock(&sqd->ctx_lock);
7773 io_sq_thread_unpark(sqd);
7775 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7776 if (!ctx->sq_thread_idle)
7777 ctx->sq_thread_idle = HZ;
7782 if (p->flags & IORING_SETUP_SQ_AFF) {
7783 int cpu = p->sq_thread_cpu;
7786 if (cpu >= nr_cpu_ids)
7788 if (!cpu_online(cpu))
7791 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7792 cpu, "io_uring-sq");
7794 sqd->thread = kthread_create(io_sq_thread, sqd,
7797 if (IS_ERR(sqd->thread)) {
7798 ret = PTR_ERR(sqd->thread);
7802 ret = io_uring_alloc_task_context(sqd->thread);
7805 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7806 /* Can't have SQ_AFF without SQPOLL */
7812 ret = io_init_wq_offload(ctx, p);
7818 io_finish_async(ctx);
7822 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7824 struct io_sq_data *sqd = ctx->sq_data;
7826 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7827 wake_up_process(sqd->thread);
7830 static inline void __io_unaccount_mem(struct user_struct *user,
7831 unsigned long nr_pages)
7833 atomic_long_sub(nr_pages, &user->locked_vm);
7836 static inline int __io_account_mem(struct user_struct *user,
7837 unsigned long nr_pages)
7839 unsigned long page_limit, cur_pages, new_pages;
7841 /* Don't allow more pages than we can safely lock */
7842 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7845 cur_pages = atomic_long_read(&user->locked_vm);
7846 new_pages = cur_pages + nr_pages;
7847 if (new_pages > page_limit)
7849 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7850 new_pages) != cur_pages);
7855 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7856 enum io_mem_account acct)
7859 __io_unaccount_mem(ctx->user, nr_pages);
7861 if (ctx->mm_account) {
7862 if (acct == ACCT_LOCKED)
7863 ctx->mm_account->locked_vm -= nr_pages;
7864 else if (acct == ACCT_PINNED)
7865 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7869 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7870 enum io_mem_account acct)
7874 if (ctx->limit_mem) {
7875 ret = __io_account_mem(ctx->user, nr_pages);
7880 if (ctx->mm_account) {
7881 if (acct == ACCT_LOCKED)
7882 ctx->mm_account->locked_vm += nr_pages;
7883 else if (acct == ACCT_PINNED)
7884 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7890 static void io_mem_free(void *ptr)
7897 page = virt_to_head_page(ptr);
7898 if (put_page_testzero(page))
7899 free_compound_page(page);
7902 static void *io_mem_alloc(size_t size)
7904 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7907 return (void *) __get_free_pages(gfp_flags, get_order(size));
7910 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7913 struct io_rings *rings;
7914 size_t off, sq_array_size;
7916 off = struct_size(rings, cqes, cq_entries);
7917 if (off == SIZE_MAX)
7921 off = ALIGN(off, SMP_CACHE_BYTES);
7929 sq_array_size = array_size(sizeof(u32), sq_entries);
7930 if (sq_array_size == SIZE_MAX)
7933 if (check_add_overflow(off, sq_array_size, &off))
7939 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7943 pages = (size_t)1 << get_order(
7944 rings_size(sq_entries, cq_entries, NULL));
7945 pages += (size_t)1 << get_order(
7946 array_size(sizeof(struct io_uring_sqe), sq_entries));
7951 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7955 if (!ctx->user_bufs)
7958 for (i = 0; i < ctx->nr_user_bufs; i++) {
7959 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7961 for (j = 0; j < imu->nr_bvecs; j++)
7962 unpin_user_page(imu->bvec[j].bv_page);
7964 if (imu->acct_pages)
7965 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7970 kfree(ctx->user_bufs);
7971 ctx->user_bufs = NULL;
7972 ctx->nr_user_bufs = 0;
7976 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7977 void __user *arg, unsigned index)
7979 struct iovec __user *src;
7981 #ifdef CONFIG_COMPAT
7983 struct compat_iovec __user *ciovs;
7984 struct compat_iovec ciov;
7986 ciovs = (struct compat_iovec __user *) arg;
7987 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7990 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7991 dst->iov_len = ciov.iov_len;
7995 src = (struct iovec __user *) arg;
7996 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8002 * Not super efficient, but this is just a registration time. And we do cache
8003 * the last compound head, so generally we'll only do a full search if we don't
8006 * We check if the given compound head page has already been accounted, to
8007 * avoid double accounting it. This allows us to account the full size of the
8008 * page, not just the constituent pages of a huge page.
8010 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8011 int nr_pages, struct page *hpage)
8015 /* check current page array */
8016 for (i = 0; i < nr_pages; i++) {
8017 if (!PageCompound(pages[i]))
8019 if (compound_head(pages[i]) == hpage)
8023 /* check previously registered pages */
8024 for (i = 0; i < ctx->nr_user_bufs; i++) {
8025 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8027 for (j = 0; j < imu->nr_bvecs; j++) {
8028 if (!PageCompound(imu->bvec[j].bv_page))
8030 if (compound_head(imu->bvec[j].bv_page) == hpage)
8038 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8039 int nr_pages, struct io_mapped_ubuf *imu,
8040 struct page **last_hpage)
8044 for (i = 0; i < nr_pages; i++) {
8045 if (!PageCompound(pages[i])) {
8050 hpage = compound_head(pages[i]);
8051 if (hpage == *last_hpage)
8053 *last_hpage = hpage;
8054 if (headpage_already_acct(ctx, pages, i, hpage))
8056 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8060 if (!imu->acct_pages)
8063 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8065 imu->acct_pages = 0;
8069 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8072 struct vm_area_struct **vmas = NULL;
8073 struct page **pages = NULL;
8074 struct page *last_hpage = NULL;
8075 int i, j, got_pages = 0;
8080 if (!nr_args || nr_args > UIO_MAXIOV)
8083 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8085 if (!ctx->user_bufs)
8088 for (i = 0; i < nr_args; i++) {
8089 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8090 unsigned long off, start, end, ubuf;
8095 ret = io_copy_iov(ctx, &iov, arg, i);
8100 * Don't impose further limits on the size and buffer
8101 * constraints here, we'll -EINVAL later when IO is
8102 * submitted if they are wrong.
8105 if (!iov.iov_base || !iov.iov_len)
8108 /* arbitrary limit, but we need something */
8109 if (iov.iov_len > SZ_1G)
8112 ubuf = (unsigned long) iov.iov_base;
8113 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8114 start = ubuf >> PAGE_SHIFT;
8115 nr_pages = end - start;
8118 if (!pages || nr_pages > got_pages) {
8121 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8123 vmas = kvmalloc_array(nr_pages,
8124 sizeof(struct vm_area_struct *),
8126 if (!pages || !vmas) {
8130 got_pages = nr_pages;
8133 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8140 mmap_read_lock(current->mm);
8141 pret = pin_user_pages(ubuf, nr_pages,
8142 FOLL_WRITE | FOLL_LONGTERM,
8144 if (pret == nr_pages) {
8145 /* don't support file backed memory */
8146 for (j = 0; j < nr_pages; j++) {
8147 struct vm_area_struct *vma = vmas[j];
8150 !is_file_hugepages(vma->vm_file)) {
8156 ret = pret < 0 ? pret : -EFAULT;
8158 mmap_read_unlock(current->mm);
8161 * if we did partial map, or found file backed vmas,
8162 * release any pages we did get
8165 unpin_user_pages(pages, pret);
8170 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8172 unpin_user_pages(pages, pret);
8177 off = ubuf & ~PAGE_MASK;
8179 for (j = 0; j < nr_pages; j++) {
8182 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8183 imu->bvec[j].bv_page = pages[j];
8184 imu->bvec[j].bv_len = vec_len;
8185 imu->bvec[j].bv_offset = off;
8189 /* store original address for later verification */
8191 imu->len = iov.iov_len;
8192 imu->nr_bvecs = nr_pages;
8194 ctx->nr_user_bufs++;
8202 io_sqe_buffer_unregister(ctx);
8206 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8208 __s32 __user *fds = arg;
8214 if (copy_from_user(&fd, fds, sizeof(*fds)))
8217 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8218 if (IS_ERR(ctx->cq_ev_fd)) {
8219 int ret = PTR_ERR(ctx->cq_ev_fd);
8220 ctx->cq_ev_fd = NULL;
8227 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8229 if (ctx->cq_ev_fd) {
8230 eventfd_ctx_put(ctx->cq_ev_fd);
8231 ctx->cq_ev_fd = NULL;
8238 static int __io_destroy_buffers(int id, void *p, void *data)
8240 struct io_ring_ctx *ctx = data;
8241 struct io_buffer *buf = p;
8243 __io_remove_buffers(ctx, buf, id, -1U);
8247 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8249 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8250 idr_destroy(&ctx->io_buffer_idr);
8253 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8255 io_finish_async(ctx);
8256 io_sqe_buffer_unregister(ctx);
8258 if (ctx->sqo_task) {
8259 put_task_struct(ctx->sqo_task);
8260 ctx->sqo_task = NULL;
8261 mmdrop(ctx->mm_account);
8262 ctx->mm_account = NULL;
8265 #ifdef CONFIG_BLK_CGROUP
8266 if (ctx->sqo_blkcg_css)
8267 css_put(ctx->sqo_blkcg_css);
8270 io_sqe_files_unregister(ctx);
8271 io_eventfd_unregister(ctx);
8272 io_destroy_buffers(ctx);
8273 idr_destroy(&ctx->personality_idr);
8275 #if defined(CONFIG_UNIX)
8276 if (ctx->ring_sock) {
8277 ctx->ring_sock->file = NULL; /* so that iput() is called */
8278 sock_release(ctx->ring_sock);
8282 io_mem_free(ctx->rings);
8283 io_mem_free(ctx->sq_sqes);
8285 percpu_ref_exit(&ctx->refs);
8286 free_uid(ctx->user);
8287 put_cred(ctx->creds);
8288 kfree(ctx->cancel_hash);
8289 kmem_cache_free(req_cachep, ctx->fallback_req);
8293 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8295 struct io_ring_ctx *ctx = file->private_data;
8298 poll_wait(file, &ctx->cq_wait, wait);
8300 * synchronizes with barrier from wq_has_sleeper call in
8304 if (!io_sqring_full(ctx))
8305 mask |= EPOLLOUT | EPOLLWRNORM;
8306 if (io_cqring_events(ctx, false))
8307 mask |= EPOLLIN | EPOLLRDNORM;
8312 static int io_uring_fasync(int fd, struct file *file, int on)
8314 struct io_ring_ctx *ctx = file->private_data;
8316 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8319 static int io_remove_personalities(int id, void *p, void *data)
8321 struct io_ring_ctx *ctx = data;
8322 struct io_identity *iod;
8324 iod = idr_remove(&ctx->personality_idr, id);
8326 put_cred(iod->creds);
8327 if (refcount_dec_and_test(&iod->count))
8333 static void io_ring_exit_work(struct work_struct *work)
8335 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8339 * If we're doing polled IO and end up having requests being
8340 * submitted async (out-of-line), then completions can come in while
8341 * we're waiting for refs to drop. We need to reap these manually,
8342 * as nobody else will be looking for them.
8346 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8347 io_iopoll_try_reap_events(ctx);
8348 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8349 io_ring_ctx_free(ctx);
8352 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8354 mutex_lock(&ctx->uring_lock);
8355 percpu_ref_kill(&ctx->refs);
8356 mutex_unlock(&ctx->uring_lock);
8358 io_kill_timeouts(ctx, NULL);
8359 io_poll_remove_all(ctx, NULL);
8362 io_wq_cancel_all(ctx->io_wq);
8364 /* if we failed setting up the ctx, we might not have any rings */
8366 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8367 io_iopoll_try_reap_events(ctx);
8368 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8371 * Do this upfront, so we won't have a grace period where the ring
8372 * is closed but resources aren't reaped yet. This can cause
8373 * spurious failure in setting up a new ring.
8375 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8378 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8380 * Use system_unbound_wq to avoid spawning tons of event kworkers
8381 * if we're exiting a ton of rings at the same time. It just adds
8382 * noise and overhead, there's no discernable change in runtime
8383 * over using system_wq.
8385 queue_work(system_unbound_wq, &ctx->exit_work);
8388 static int io_uring_release(struct inode *inode, struct file *file)
8390 struct io_ring_ctx *ctx = file->private_data;
8392 file->private_data = NULL;
8393 io_ring_ctx_wait_and_kill(ctx);
8397 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8399 struct files_struct *files = data;
8401 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8402 work->identity->files == files);
8406 * Returns true if 'preq' is the link parent of 'req'
8408 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8410 struct io_kiocb *link;
8412 if (!(preq->flags & REQ_F_LINK_HEAD))
8415 list_for_each_entry(link, &preq->link_list, link_list) {
8424 * We're looking to cancel 'req' because it's holding on to our files, but
8425 * 'req' could be a link to another request. See if it is, and cancel that
8426 * parent request if so.
8428 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8430 struct hlist_node *tmp;
8431 struct io_kiocb *preq;
8435 spin_lock_irq(&ctx->completion_lock);
8436 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8437 struct hlist_head *list;
8439 list = &ctx->cancel_hash[i];
8440 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8441 found = io_match_link(preq, req);
8443 io_poll_remove_one(preq);
8448 spin_unlock_irq(&ctx->completion_lock);
8452 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8453 struct io_kiocb *req)
8455 struct io_kiocb *preq;
8458 spin_lock_irq(&ctx->completion_lock);
8459 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8460 found = io_match_link(preq, req);
8462 __io_timeout_cancel(preq);
8466 spin_unlock_irq(&ctx->completion_lock);
8470 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8472 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8475 if (req->flags & REQ_F_LINK_TIMEOUT) {
8476 unsigned long flags;
8477 struct io_ring_ctx *ctx = req->ctx;
8479 /* protect against races with linked timeouts */
8480 spin_lock_irqsave(&ctx->completion_lock, flags);
8481 ret = io_match_link(req, data);
8482 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8484 ret = io_match_link(req, data);
8489 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8491 enum io_wq_cancel cret;
8493 /* cancel this particular work, if it's running */
8494 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8495 if (cret != IO_WQ_CANCEL_NOTFOUND)
8498 /* find links that hold this pending, cancel those */
8499 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8500 if (cret != IO_WQ_CANCEL_NOTFOUND)
8503 /* if we have a poll link holding this pending, cancel that */
8504 if (io_poll_remove_link(ctx, req))
8507 /* final option, timeout link is holding this req pending */
8508 io_timeout_remove_link(ctx, req);
8511 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8512 struct task_struct *task,
8513 struct files_struct *files)
8515 struct io_defer_entry *de = NULL;
8518 spin_lock_irq(&ctx->completion_lock);
8519 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8520 if (io_task_match(de->req, task) &&
8521 io_match_files(de->req, files)) {
8522 list_cut_position(&list, &ctx->defer_list, &de->list);
8526 spin_unlock_irq(&ctx->completion_lock);
8528 while (!list_empty(&list)) {
8529 de = list_first_entry(&list, struct io_defer_entry, list);
8530 list_del_init(&de->list);
8531 req_set_fail_links(de->req);
8532 io_put_req(de->req);
8533 io_req_complete(de->req, -ECANCELED);
8539 * Returns true if we found and killed one or more files pinning requests
8541 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8542 struct files_struct *files)
8544 if (list_empty_careful(&ctx->inflight_list))
8547 /* cancel all at once, should be faster than doing it one by one*/
8548 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8550 while (!list_empty_careful(&ctx->inflight_list)) {
8551 struct io_kiocb *cancel_req = NULL, *req;
8554 spin_lock_irq(&ctx->inflight_lock);
8555 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8556 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8557 req->work.identity->files != files)
8559 /* req is being completed, ignore */
8560 if (!refcount_inc_not_zero(&req->refs))
8566 prepare_to_wait(&ctx->inflight_wait, &wait,
8567 TASK_UNINTERRUPTIBLE);
8568 spin_unlock_irq(&ctx->inflight_lock);
8570 /* We need to keep going until we don't find a matching req */
8573 /* cancel this request, or head link requests */
8574 io_attempt_cancel(ctx, cancel_req);
8575 io_put_req(cancel_req);
8576 /* cancellations _may_ trigger task work */
8579 finish_wait(&ctx->inflight_wait, &wait);
8585 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8587 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8588 struct task_struct *task = data;
8590 return io_task_match(req, task);
8593 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8594 struct task_struct *task,
8595 struct files_struct *files)
8599 ret = io_uring_cancel_files(ctx, files);
8601 enum io_wq_cancel cret;
8603 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8604 if (cret != IO_WQ_CANCEL_NOTFOUND)
8607 /* SQPOLL thread does its own polling */
8608 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8609 while (!list_empty_careful(&ctx->iopoll_list)) {
8610 io_iopoll_try_reap_events(ctx);
8615 ret |= io_poll_remove_all(ctx, task);
8616 ret |= io_kill_timeouts(ctx, task);
8623 * We need to iteratively cancel requests, in case a request has dependent
8624 * hard links. These persist even for failure of cancelations, hence keep
8625 * looping until none are found.
8627 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8628 struct files_struct *files)
8630 struct task_struct *task = current;
8632 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8633 task = ctx->sq_data->thread;
8634 atomic_inc(&task->io_uring->in_idle);
8635 io_sq_thread_park(ctx->sq_data);
8639 io_cancel_defer_files(ctx, NULL, files);
8641 io_cancel_defer_files(ctx, task, NULL);
8643 io_cqring_overflow_flush(ctx, true, task, files);
8645 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8650 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8651 atomic_dec(&task->io_uring->in_idle);
8653 * If the files that are going away are the ones in the thread
8654 * identity, clear them out.
8656 if (task->io_uring->identity->files == files)
8657 task->io_uring->identity->files = NULL;
8658 io_sq_thread_unpark(ctx->sq_data);
8663 * Note that this task has used io_uring. We use it for cancelation purposes.
8665 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8667 struct io_uring_task *tctx = current->io_uring;
8669 if (unlikely(!tctx)) {
8672 ret = io_uring_alloc_task_context(current);
8675 tctx = current->io_uring;
8677 if (tctx->last != file) {
8678 void *old = xa_load(&tctx->xa, (unsigned long)file);
8682 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8688 * This is race safe in that the task itself is doing this, hence it
8689 * cannot be going through the exit/cancel paths at the same time.
8690 * This cannot be modified while exit/cancel is running.
8692 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8693 tctx->sqpoll = true;
8699 * Remove this io_uring_file -> task mapping.
8701 static void io_uring_del_task_file(struct file *file)
8703 struct io_uring_task *tctx = current->io_uring;
8705 if (tctx->last == file)
8707 file = xa_erase(&tctx->xa, (unsigned long)file);
8713 * Drop task note for this file if we're the only ones that hold it after
8716 static void io_uring_attempt_task_drop(struct file *file)
8718 if (!current->io_uring)
8721 * fput() is pending, will be 2 if the only other ref is our potential
8722 * task file note. If the task is exiting, drop regardless of count.
8724 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8725 atomic_long_read(&file->f_count) == 2)
8726 io_uring_del_task_file(file);
8729 void __io_uring_files_cancel(struct files_struct *files)
8731 struct io_uring_task *tctx = current->io_uring;
8733 unsigned long index;
8735 /* make sure overflow events are dropped */
8736 atomic_inc(&tctx->in_idle);
8738 xa_for_each(&tctx->xa, index, file) {
8739 struct io_ring_ctx *ctx = file->private_data;
8741 io_uring_cancel_task_requests(ctx, files);
8743 io_uring_del_task_file(file);
8746 atomic_dec(&tctx->in_idle);
8749 static s64 tctx_inflight(struct io_uring_task *tctx)
8751 unsigned long index;
8755 inflight = percpu_counter_sum(&tctx->inflight);
8760 * If we have SQPOLL rings, then we need to iterate and find them, and
8761 * add the pending count for those.
8763 xa_for_each(&tctx->xa, index, file) {
8764 struct io_ring_ctx *ctx = file->private_data;
8766 if (ctx->flags & IORING_SETUP_SQPOLL) {
8767 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8769 inflight += percpu_counter_sum(&__tctx->inflight);
8777 * Find any io_uring fd that this task has registered or done IO on, and cancel
8780 void __io_uring_task_cancel(void)
8782 struct io_uring_task *tctx = current->io_uring;
8786 /* make sure overflow events are dropped */
8787 atomic_inc(&tctx->in_idle);
8790 /* read completions before cancelations */
8791 inflight = tctx_inflight(tctx);
8794 __io_uring_files_cancel(NULL);
8796 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8799 * If we've seen completions, retry. This avoids a race where
8800 * a completion comes in before we did prepare_to_wait().
8802 if (inflight != tctx_inflight(tctx))
8807 finish_wait(&tctx->wait, &wait);
8808 atomic_dec(&tctx->in_idle);
8811 static int io_uring_flush(struct file *file, void *data)
8813 io_uring_attempt_task_drop(file);
8817 static void *io_uring_validate_mmap_request(struct file *file,
8818 loff_t pgoff, size_t sz)
8820 struct io_ring_ctx *ctx = file->private_data;
8821 loff_t offset = pgoff << PAGE_SHIFT;
8826 case IORING_OFF_SQ_RING:
8827 case IORING_OFF_CQ_RING:
8830 case IORING_OFF_SQES:
8834 return ERR_PTR(-EINVAL);
8837 page = virt_to_head_page(ptr);
8838 if (sz > page_size(page))
8839 return ERR_PTR(-EINVAL);
8846 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8848 size_t sz = vma->vm_end - vma->vm_start;
8852 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8854 return PTR_ERR(ptr);
8856 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8857 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8860 #else /* !CONFIG_MMU */
8862 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8864 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8867 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8869 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8872 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8873 unsigned long addr, unsigned long len,
8874 unsigned long pgoff, unsigned long flags)
8878 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8880 return PTR_ERR(ptr);
8882 return (unsigned long) ptr;
8885 #endif /* !CONFIG_MMU */
8887 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8892 if (!io_sqring_full(ctx))
8895 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8897 if (!io_sqring_full(ctx))
8901 } while (!signal_pending(current));
8903 finish_wait(&ctx->sqo_sq_wait, &wait);
8906 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8907 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8910 struct io_ring_ctx *ctx;
8917 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8918 IORING_ENTER_SQ_WAIT))
8926 if (f.file->f_op != &io_uring_fops)
8930 ctx = f.file->private_data;
8931 if (!percpu_ref_tryget(&ctx->refs))
8935 if (ctx->flags & IORING_SETUP_R_DISABLED)
8939 * For SQ polling, the thread will do all submissions and completions.
8940 * Just return the requested submit count, and wake the thread if
8944 if (ctx->flags & IORING_SETUP_SQPOLL) {
8945 if (!list_empty_careful(&ctx->cq_overflow_list))
8946 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8947 if (flags & IORING_ENTER_SQ_WAKEUP)
8948 wake_up(&ctx->sq_data->wait);
8949 if (flags & IORING_ENTER_SQ_WAIT)
8950 io_sqpoll_wait_sq(ctx);
8951 submitted = to_submit;
8952 } else if (to_submit) {
8953 ret = io_uring_add_task_file(ctx, f.file);
8956 mutex_lock(&ctx->uring_lock);
8957 submitted = io_submit_sqes(ctx, to_submit);
8958 mutex_unlock(&ctx->uring_lock);
8960 if (submitted != to_submit)
8963 if (flags & IORING_ENTER_GETEVENTS) {
8964 min_complete = min(min_complete, ctx->cq_entries);
8967 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8968 * space applications don't need to do io completion events
8969 * polling again, they can rely on io_sq_thread to do polling
8970 * work, which can reduce cpu usage and uring_lock contention.
8972 if (ctx->flags & IORING_SETUP_IOPOLL &&
8973 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8974 ret = io_iopoll_check(ctx, min_complete);
8976 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8981 percpu_ref_put(&ctx->refs);
8984 return submitted ? submitted : ret;
8987 #ifdef CONFIG_PROC_FS
8988 static int io_uring_show_cred(int id, void *p, void *data)
8990 struct io_identity *iod = p;
8991 const struct cred *cred = iod->creds;
8992 struct seq_file *m = data;
8993 struct user_namespace *uns = seq_user_ns(m);
8994 struct group_info *gi;
8999 seq_printf(m, "%5d\n", id);
9000 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9001 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9002 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9003 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9004 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9005 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9006 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9007 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9008 seq_puts(m, "\n\tGroups:\t");
9009 gi = cred->group_info;
9010 for (g = 0; g < gi->ngroups; g++) {
9011 seq_put_decimal_ull(m, g ? " " : "",
9012 from_kgid_munged(uns, gi->gid[g]));
9014 seq_puts(m, "\n\tCapEff:\t");
9015 cap = cred->cap_effective;
9016 CAP_FOR_EACH_U32(__capi)
9017 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9022 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9024 struct io_sq_data *sq = NULL;
9029 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9030 * since fdinfo case grabs it in the opposite direction of normal use
9031 * cases. If we fail to get the lock, we just don't iterate any
9032 * structures that could be going away outside the io_uring mutex.
9034 has_lock = mutex_trylock(&ctx->uring_lock);
9036 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9039 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9040 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9041 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9042 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9043 struct fixed_file_table *table;
9046 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9047 f = table->files[i & IORING_FILE_TABLE_MASK];
9049 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9051 seq_printf(m, "%5u: <none>\n", i);
9053 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9054 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9055 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9057 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9058 (unsigned int) buf->len);
9060 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9061 seq_printf(m, "Personalities:\n");
9062 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9064 seq_printf(m, "PollList:\n");
9065 spin_lock_irq(&ctx->completion_lock);
9066 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9067 struct hlist_head *list = &ctx->cancel_hash[i];
9068 struct io_kiocb *req;
9070 hlist_for_each_entry(req, list, hash_node)
9071 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9072 req->task->task_works != NULL);
9074 spin_unlock_irq(&ctx->completion_lock);
9076 mutex_unlock(&ctx->uring_lock);
9079 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9081 struct io_ring_ctx *ctx = f->private_data;
9083 if (percpu_ref_tryget(&ctx->refs)) {
9084 __io_uring_show_fdinfo(ctx, m);
9085 percpu_ref_put(&ctx->refs);
9090 static const struct file_operations io_uring_fops = {
9091 .release = io_uring_release,
9092 .flush = io_uring_flush,
9093 .mmap = io_uring_mmap,
9095 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9096 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9098 .poll = io_uring_poll,
9099 .fasync = io_uring_fasync,
9100 #ifdef CONFIG_PROC_FS
9101 .show_fdinfo = io_uring_show_fdinfo,
9105 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9106 struct io_uring_params *p)
9108 struct io_rings *rings;
9109 size_t size, sq_array_offset;
9111 /* make sure these are sane, as we already accounted them */
9112 ctx->sq_entries = p->sq_entries;
9113 ctx->cq_entries = p->cq_entries;
9115 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9116 if (size == SIZE_MAX)
9119 rings = io_mem_alloc(size);
9124 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9125 rings->sq_ring_mask = p->sq_entries - 1;
9126 rings->cq_ring_mask = p->cq_entries - 1;
9127 rings->sq_ring_entries = p->sq_entries;
9128 rings->cq_ring_entries = p->cq_entries;
9129 ctx->sq_mask = rings->sq_ring_mask;
9130 ctx->cq_mask = rings->cq_ring_mask;
9132 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9133 if (size == SIZE_MAX) {
9134 io_mem_free(ctx->rings);
9139 ctx->sq_sqes = io_mem_alloc(size);
9140 if (!ctx->sq_sqes) {
9141 io_mem_free(ctx->rings);
9150 * Allocate an anonymous fd, this is what constitutes the application
9151 * visible backing of an io_uring instance. The application mmaps this
9152 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9153 * we have to tie this fd to a socket for file garbage collection purposes.
9155 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9160 #if defined(CONFIG_UNIX)
9161 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9167 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9171 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9172 O_RDWR | O_CLOEXEC);
9176 ret = PTR_ERR(file);
9180 #if defined(CONFIG_UNIX)
9181 ctx->ring_sock->file = file;
9183 if (unlikely(io_uring_add_task_file(ctx, file))) {
9184 file = ERR_PTR(-ENOMEM);
9187 fd_install(ret, file);
9190 #if defined(CONFIG_UNIX)
9191 sock_release(ctx->ring_sock);
9192 ctx->ring_sock = NULL;
9197 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9198 struct io_uring_params __user *params)
9200 struct user_struct *user = NULL;
9201 struct io_ring_ctx *ctx;
9207 if (entries > IORING_MAX_ENTRIES) {
9208 if (!(p->flags & IORING_SETUP_CLAMP))
9210 entries = IORING_MAX_ENTRIES;
9214 * Use twice as many entries for the CQ ring. It's possible for the
9215 * application to drive a higher depth than the size of the SQ ring,
9216 * since the sqes are only used at submission time. This allows for
9217 * some flexibility in overcommitting a bit. If the application has
9218 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9219 * of CQ ring entries manually.
9221 p->sq_entries = roundup_pow_of_two(entries);
9222 if (p->flags & IORING_SETUP_CQSIZE) {
9224 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9225 * to a power-of-two, if it isn't already. We do NOT impose
9226 * any cq vs sq ring sizing.
9228 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9229 if (p->cq_entries < p->sq_entries)
9231 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9232 if (!(p->flags & IORING_SETUP_CLAMP))
9234 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9237 p->cq_entries = 2 * p->sq_entries;
9240 user = get_uid(current_user());
9241 limit_mem = !capable(CAP_IPC_LOCK);
9244 ret = __io_account_mem(user,
9245 ring_pages(p->sq_entries, p->cq_entries));
9252 ctx = io_ring_ctx_alloc(p);
9255 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9260 ctx->compat = in_compat_syscall();
9262 ctx->creds = get_current_cred();
9264 ctx->loginuid = current->loginuid;
9265 ctx->sessionid = current->sessionid;
9267 ctx->sqo_task = get_task_struct(current);
9270 * This is just grabbed for accounting purposes. When a process exits,
9271 * the mm is exited and dropped before the files, hence we need to hang
9272 * on to this mm purely for the purposes of being able to unaccount
9273 * memory (locked/pinned vm). It's not used for anything else.
9275 mmgrab(current->mm);
9276 ctx->mm_account = current->mm;
9278 #ifdef CONFIG_BLK_CGROUP
9280 * The sq thread will belong to the original cgroup it was inited in.
9281 * If the cgroup goes offline (e.g. disabling the io controller), then
9282 * issued bios will be associated with the closest cgroup later in the
9286 ctx->sqo_blkcg_css = blkcg_css();
9287 ret = css_tryget_online(ctx->sqo_blkcg_css);
9290 /* don't init against a dying cgroup, have the user try again */
9291 ctx->sqo_blkcg_css = NULL;
9298 * Account memory _before_ installing the file descriptor. Once
9299 * the descriptor is installed, it can get closed at any time. Also
9300 * do this before hitting the general error path, as ring freeing
9301 * will un-account as well.
9303 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9305 ctx->limit_mem = limit_mem;
9307 ret = io_allocate_scq_urings(ctx, p);
9311 ret = io_sq_offload_create(ctx, p);
9315 if (!(p->flags & IORING_SETUP_R_DISABLED))
9316 io_sq_offload_start(ctx);
9318 memset(&p->sq_off, 0, sizeof(p->sq_off));
9319 p->sq_off.head = offsetof(struct io_rings, sq.head);
9320 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9321 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9322 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9323 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9324 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9325 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9327 memset(&p->cq_off, 0, sizeof(p->cq_off));
9328 p->cq_off.head = offsetof(struct io_rings, cq.head);
9329 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9330 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9331 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9332 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9333 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9334 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9336 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9337 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9338 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9339 IORING_FEAT_POLL_32BITS;
9341 if (copy_to_user(params, p, sizeof(*p))) {
9347 * Install ring fd as the very last thing, so we don't risk someone
9348 * having closed it before we finish setup
9350 ret = io_uring_get_fd(ctx);
9354 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9357 io_ring_ctx_wait_and_kill(ctx);
9362 * Sets up an aio uring context, and returns the fd. Applications asks for a
9363 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9364 * params structure passed in.
9366 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9368 struct io_uring_params p;
9371 if (copy_from_user(&p, params, sizeof(p)))
9373 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9378 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9379 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9380 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9381 IORING_SETUP_R_DISABLED))
9384 return io_uring_create(entries, &p, params);
9387 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9388 struct io_uring_params __user *, params)
9390 return io_uring_setup(entries, params);
9393 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9395 struct io_uring_probe *p;
9399 size = struct_size(p, ops, nr_args);
9400 if (size == SIZE_MAX)
9402 p = kzalloc(size, GFP_KERNEL);
9407 if (copy_from_user(p, arg, size))
9410 if (memchr_inv(p, 0, size))
9413 p->last_op = IORING_OP_LAST - 1;
9414 if (nr_args > IORING_OP_LAST)
9415 nr_args = IORING_OP_LAST;
9417 for (i = 0; i < nr_args; i++) {
9419 if (!io_op_defs[i].not_supported)
9420 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9425 if (copy_to_user(arg, p, size))
9432 static int io_register_personality(struct io_ring_ctx *ctx)
9434 struct io_identity *id;
9437 id = kmalloc(sizeof(*id), GFP_KERNEL);
9441 io_init_identity(id);
9442 id->creds = get_current_cred();
9444 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9446 put_cred(id->creds);
9452 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9454 struct io_identity *iod;
9456 iod = idr_remove(&ctx->personality_idr, id);
9458 put_cred(iod->creds);
9459 if (refcount_dec_and_test(&iod->count))
9467 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9468 unsigned int nr_args)
9470 struct io_uring_restriction *res;
9474 /* Restrictions allowed only if rings started disabled */
9475 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9478 /* We allow only a single restrictions registration */
9479 if (ctx->restrictions.registered)
9482 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9485 size = array_size(nr_args, sizeof(*res));
9486 if (size == SIZE_MAX)
9489 res = memdup_user(arg, size);
9491 return PTR_ERR(res);
9495 for (i = 0; i < nr_args; i++) {
9496 switch (res[i].opcode) {
9497 case IORING_RESTRICTION_REGISTER_OP:
9498 if (res[i].register_op >= IORING_REGISTER_LAST) {
9503 __set_bit(res[i].register_op,
9504 ctx->restrictions.register_op);
9506 case IORING_RESTRICTION_SQE_OP:
9507 if (res[i].sqe_op >= IORING_OP_LAST) {
9512 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9514 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9515 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9517 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9518 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9527 /* Reset all restrictions if an error happened */
9529 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9531 ctx->restrictions.registered = true;
9537 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9539 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9542 if (ctx->restrictions.registered)
9543 ctx->restricted = 1;
9545 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9547 io_sq_offload_start(ctx);
9552 static bool io_register_op_must_quiesce(int op)
9555 case IORING_UNREGISTER_FILES:
9556 case IORING_REGISTER_FILES_UPDATE:
9557 case IORING_REGISTER_PROBE:
9558 case IORING_REGISTER_PERSONALITY:
9559 case IORING_UNREGISTER_PERSONALITY:
9566 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9567 void __user *arg, unsigned nr_args)
9568 __releases(ctx->uring_lock)
9569 __acquires(ctx->uring_lock)
9574 * We're inside the ring mutex, if the ref is already dying, then
9575 * someone else killed the ctx or is already going through
9576 * io_uring_register().
9578 if (percpu_ref_is_dying(&ctx->refs))
9581 if (io_register_op_must_quiesce(opcode)) {
9582 percpu_ref_kill(&ctx->refs);
9585 * Drop uring mutex before waiting for references to exit. If
9586 * another thread is currently inside io_uring_enter() it might
9587 * need to grab the uring_lock to make progress. If we hold it
9588 * here across the drain wait, then we can deadlock. It's safe
9589 * to drop the mutex here, since no new references will come in
9590 * after we've killed the percpu ref.
9592 mutex_unlock(&ctx->uring_lock);
9594 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9597 ret = io_run_task_work_sig();
9602 mutex_lock(&ctx->uring_lock);
9605 percpu_ref_resurrect(&ctx->refs);
9610 if (ctx->restricted) {
9611 if (opcode >= IORING_REGISTER_LAST) {
9616 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9623 case IORING_REGISTER_BUFFERS:
9624 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9626 case IORING_UNREGISTER_BUFFERS:
9630 ret = io_sqe_buffer_unregister(ctx);
9632 case IORING_REGISTER_FILES:
9633 ret = io_sqe_files_register(ctx, arg, nr_args);
9635 case IORING_UNREGISTER_FILES:
9639 ret = io_sqe_files_unregister(ctx);
9641 case IORING_REGISTER_FILES_UPDATE:
9642 ret = io_sqe_files_update(ctx, arg, nr_args);
9644 case IORING_REGISTER_EVENTFD:
9645 case IORING_REGISTER_EVENTFD_ASYNC:
9649 ret = io_eventfd_register(ctx, arg);
9652 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9653 ctx->eventfd_async = 1;
9655 ctx->eventfd_async = 0;
9657 case IORING_UNREGISTER_EVENTFD:
9661 ret = io_eventfd_unregister(ctx);
9663 case IORING_REGISTER_PROBE:
9665 if (!arg || nr_args > 256)
9667 ret = io_probe(ctx, arg, nr_args);
9669 case IORING_REGISTER_PERSONALITY:
9673 ret = io_register_personality(ctx);
9675 case IORING_UNREGISTER_PERSONALITY:
9679 ret = io_unregister_personality(ctx, nr_args);
9681 case IORING_REGISTER_ENABLE_RINGS:
9685 ret = io_register_enable_rings(ctx);
9687 case IORING_REGISTER_RESTRICTIONS:
9688 ret = io_register_restrictions(ctx, arg, nr_args);
9696 if (io_register_op_must_quiesce(opcode)) {
9697 /* bring the ctx back to life */
9698 percpu_ref_reinit(&ctx->refs);
9700 reinit_completion(&ctx->ref_comp);
9705 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9706 void __user *, arg, unsigned int, nr_args)
9708 struct io_ring_ctx *ctx;
9717 if (f.file->f_op != &io_uring_fops)
9720 ctx = f.file->private_data;
9722 mutex_lock(&ctx->uring_lock);
9723 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9724 mutex_unlock(&ctx->uring_lock);
9725 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9726 ctx->cq_ev_fd != NULL, ret);
9732 static int __init io_uring_init(void)
9734 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9735 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9736 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9739 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9740 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9741 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9742 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9743 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9744 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9745 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9746 BUILD_BUG_SQE_ELEM(8, __u64, off);
9747 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9748 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9749 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9750 BUILD_BUG_SQE_ELEM(24, __u32, len);
9751 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9752 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9753 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9754 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9755 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9756 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9757 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9758 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9759 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9760 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9761 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9762 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9763 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9764 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9765 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9766 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9767 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9768 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9769 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9771 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9772 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9773 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9776 __initcall(io_uring_init);