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>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
107 u32 head ____cacheline_aligned_in_smp;
108 u32 tail ____cacheline_aligned_in_smp;
112 * This data is shared with the application through the mmap at offsets
113 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
115 * The offsets to the member fields are published through struct
116 * io_sqring_offsets when calling io_uring_setup.
120 * Head and tail offsets into the ring; the offsets need to be
121 * masked to get valid indices.
123 * The kernel controls head of the sq ring and the tail of the cq ring,
124 * and the application controls tail of the sq ring and the head of the
127 struct io_uring sq, cq;
129 * Bitmasks to apply to head and tail offsets (constant, equals
132 u32 sq_ring_mask, cq_ring_mask;
133 /* Ring sizes (constant, power of 2) */
134 u32 sq_ring_entries, cq_ring_entries;
136 * Number of invalid entries dropped by the kernel due to
137 * invalid index stored in array
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * After a new SQ head value was read by the application this
144 * counter includes all submissions that were dropped reaching
145 * the new SQ head (and possibly more).
151 * Written by the kernel, shouldn't be modified by the
154 * The application needs a full memory barrier before checking
155 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
161 * Written by the application, shouldn't be modified by the
166 * Number of completion events lost because the queue was full;
167 * this should be avoided by the application by making sure
168 * there are not more requests pending than there is space in
169 * the completion queue.
171 * Written by the kernel, shouldn't be modified by the
172 * application (i.e. get number of "new events" by comparing to
175 * As completion events come in out of order this counter is not
176 * ordered with any other data.
180 * Ring buffer of completion events.
182 * The kernel writes completion events fresh every time they are
183 * produced, so the application is allowed to modify pending
186 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
189 struct io_mapped_ubuf {
192 struct bio_vec *bvec;
193 unsigned int nr_bvecs;
194 unsigned long acct_pages;
197 struct fixed_file_table {
201 struct fixed_file_ref_node {
202 struct percpu_ref refs;
203 struct list_head node;
204 struct list_head file_list;
205 struct fixed_file_data *file_data;
206 struct llist_node llist;
209 struct fixed_file_data {
210 struct fixed_file_table *table;
211 struct io_ring_ctx *ctx;
213 struct fixed_file_ref_node *node;
214 struct percpu_ref refs;
215 struct completion done;
216 struct list_head ref_list;
221 struct list_head list;
227 struct io_restriction {
228 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
229 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
230 u8 sqe_flags_allowed;
231 u8 sqe_flags_required;
239 /* ctx's that are using this sqd */
240 struct list_head ctx_list;
241 struct list_head ctx_new_list;
242 struct mutex ctx_lock;
244 struct task_struct *thread;
245 struct wait_queue_head wait;
250 struct percpu_ref refs;
251 } ____cacheline_aligned_in_smp;
255 unsigned int compat: 1;
256 unsigned int limit_mem: 1;
257 unsigned int cq_overflow_flushed: 1;
258 unsigned int drain_next: 1;
259 unsigned int eventfd_async: 1;
260 unsigned int restricted: 1;
263 * Ring buffer of indices into array of io_uring_sqe, which is
264 * mmapped by the application using the IORING_OFF_SQES offset.
266 * This indirection could e.g. be used to assign fixed
267 * io_uring_sqe entries to operations and only submit them to
268 * the queue when needed.
270 * The kernel modifies neither the indices array nor the entries
274 unsigned cached_sq_head;
277 unsigned sq_thread_idle;
278 unsigned cached_sq_dropped;
279 atomic_t cached_cq_overflow;
280 unsigned long sq_check_overflow;
282 struct list_head defer_list;
283 struct list_head timeout_list;
284 struct list_head cq_overflow_list;
286 wait_queue_head_t inflight_wait;
287 struct io_uring_sqe *sq_sqes;
288 } ____cacheline_aligned_in_smp;
290 struct io_rings *rings;
296 * For SQPOLL usage - we hold a reference to the parent task, so we
297 * have access to the ->files
299 struct task_struct *sqo_task;
301 /* Only used for accounting purposes */
302 struct mm_struct *mm_account;
304 #ifdef CONFIG_BLK_CGROUP
305 struct cgroup_subsys_state *sqo_blkcg_css;
308 struct io_sq_data *sq_data; /* if using sq thread polling */
310 struct wait_queue_head sqo_sq_wait;
311 struct wait_queue_entry sqo_wait_entry;
312 struct list_head sqd_list;
315 * If used, fixed file set. Writers must ensure that ->refs is dead,
316 * readers must ensure that ->refs is alive as long as the file* is
317 * used. Only updated through io_uring_register(2).
319 struct fixed_file_data *file_data;
320 unsigned nr_user_files;
322 /* if used, fixed mapped user buffers */
323 unsigned nr_user_bufs;
324 struct io_mapped_ubuf *user_bufs;
326 struct user_struct *user;
328 const struct cred *creds;
330 struct completion ref_comp;
331 struct completion sq_thread_comp;
333 /* if all else fails... */
334 struct io_kiocb *fallback_req;
336 #if defined(CONFIG_UNIX)
337 struct socket *ring_sock;
340 struct idr io_buffer_idr;
342 struct idr personality_idr;
345 unsigned cached_cq_tail;
348 atomic_t cq_timeouts;
349 unsigned long cq_check_overflow;
350 struct wait_queue_head cq_wait;
351 struct fasync_struct *cq_fasync;
352 struct eventfd_ctx *cq_ev_fd;
353 } ____cacheline_aligned_in_smp;
356 struct mutex uring_lock;
357 wait_queue_head_t wait;
358 } ____cacheline_aligned_in_smp;
361 spinlock_t completion_lock;
364 * ->iopoll_list is protected by the ctx->uring_lock for
365 * io_uring instances that don't use IORING_SETUP_SQPOLL.
366 * For SQPOLL, only the single threaded io_sq_thread() will
367 * manipulate the list, hence no extra locking is needed there.
369 struct list_head iopoll_list;
370 struct hlist_head *cancel_hash;
371 unsigned cancel_hash_bits;
372 bool poll_multi_file;
374 spinlock_t inflight_lock;
375 struct list_head inflight_list;
376 } ____cacheline_aligned_in_smp;
378 struct delayed_work file_put_work;
379 struct llist_head file_put_llist;
381 struct work_struct exit_work;
382 struct io_restriction restrictions;
386 * First field must be the file pointer in all the
387 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
389 struct io_poll_iocb {
392 struct wait_queue_head *head;
398 struct wait_queue_entry wait;
403 struct file *put_file;
407 struct io_timeout_data {
408 struct io_kiocb *req;
409 struct hrtimer timer;
410 struct timespec64 ts;
411 enum hrtimer_mode mode;
416 struct sockaddr __user *addr;
417 int __user *addr_len;
419 unsigned long nofile;
439 struct list_head list;
442 struct io_timeout_rem {
448 /* NOTE: kiocb has the file as the first member, so don't do it here */
456 struct sockaddr __user *addr;
463 struct user_msghdr __user *umsg;
469 struct io_buffer *kbuf;
475 struct filename *filename;
477 unsigned long nofile;
480 struct io_files_update {
506 struct epoll_event event;
510 struct file *file_out;
511 struct file *file_in;
518 struct io_provide_buf {
532 const char __user *filename;
533 struct statx __user *buffer;
536 struct io_completion {
538 struct list_head list;
542 struct io_async_connect {
543 struct sockaddr_storage address;
546 struct io_async_msghdr {
547 struct iovec fast_iov[UIO_FASTIOV];
549 struct sockaddr __user *uaddr;
551 struct sockaddr_storage addr;
555 struct iovec fast_iov[UIO_FASTIOV];
556 const struct iovec *free_iovec;
557 struct iov_iter iter;
559 struct wait_page_queue wpq;
563 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
564 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
565 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
566 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
567 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
568 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
575 REQ_F_LINK_TIMEOUT_BIT,
577 REQ_F_NEED_CLEANUP_BIT,
579 REQ_F_BUFFER_SELECTED_BIT,
580 REQ_F_NO_FILE_TABLE_BIT,
581 REQ_F_WORK_INITIALIZED_BIT,
583 /* not a real bit, just to check we're not overflowing the space */
589 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
590 /* drain existing IO first */
591 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
593 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
594 /* doesn't sever on completion < 0 */
595 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
597 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
598 /* IOSQE_BUFFER_SELECT */
599 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
602 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
603 /* fail rest of links */
604 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
605 /* on inflight list */
606 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
607 /* read/write uses file position */
608 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
609 /* must not punt to workers */
610 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
611 /* has linked timeout */
612 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
614 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
616 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
617 /* already went through poll handler */
618 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
619 /* buffer already selected */
620 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
621 /* doesn't need file table for this request */
622 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
623 /* io_wq_work is initialized */
624 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
628 struct io_poll_iocb poll;
629 struct io_poll_iocb *double_poll;
633 * NOTE! Each of the iocb union members has the file pointer
634 * as the first entry in their struct definition. So you can
635 * access the file pointer through any of the sub-structs,
636 * or directly as just 'ki_filp' in this struct.
642 struct io_poll_iocb poll;
643 struct io_accept accept;
645 struct io_cancel cancel;
646 struct io_timeout timeout;
647 struct io_timeout_rem timeout_rem;
648 struct io_connect connect;
649 struct io_sr_msg sr_msg;
651 struct io_close close;
652 struct io_files_update files_update;
653 struct io_fadvise fadvise;
654 struct io_madvise madvise;
655 struct io_epoll epoll;
656 struct io_splice splice;
657 struct io_provide_buf pbuf;
658 struct io_statx statx;
659 /* use only after cleaning per-op data, see io_clean_op() */
660 struct io_completion compl;
663 /* opcode allocated if it needs to store data for async defer */
666 /* polled IO has completed */
672 struct io_ring_ctx *ctx;
675 struct task_struct *task;
678 struct list_head link_list;
681 * 1. used with ctx->iopoll_list with reads/writes
682 * 2. to track reqs with ->files (see io_op_def::file_table)
684 struct list_head inflight_entry;
686 struct percpu_ref *fixed_file_refs;
687 struct callback_head task_work;
688 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
689 struct hlist_node hash_node;
690 struct async_poll *apoll;
691 struct io_wq_work work;
694 struct io_defer_entry {
695 struct list_head list;
696 struct io_kiocb *req;
700 #define IO_IOPOLL_BATCH 8
702 struct io_comp_state {
704 struct list_head list;
705 struct io_ring_ctx *ctx;
708 struct io_submit_state {
709 struct blk_plug plug;
712 * io_kiocb alloc cache
714 void *reqs[IO_IOPOLL_BATCH];
715 unsigned int free_reqs;
718 * Batch completion logic
720 struct io_comp_state comp;
723 * File reference cache
727 unsigned int has_refs;
728 unsigned int ios_left;
732 /* needs req->file assigned */
733 unsigned needs_file : 1;
734 /* don't fail if file grab fails */
735 unsigned needs_file_no_error : 1;
736 /* hash wq insertion if file is a regular file */
737 unsigned hash_reg_file : 1;
738 /* unbound wq insertion if file is a non-regular file */
739 unsigned unbound_nonreg_file : 1;
740 /* opcode is not supported by this kernel */
741 unsigned not_supported : 1;
742 /* set if opcode supports polled "wait" */
744 unsigned pollout : 1;
745 /* op supports buffer selection */
746 unsigned buffer_select : 1;
747 /* needs rlimit(RLIMIT_FSIZE) assigned */
748 unsigned needs_fsize : 1;
749 /* must always have async data allocated */
750 unsigned needs_async_data : 1;
751 /* size of async data needed, if any */
752 unsigned short async_size;
756 static const struct io_op_def io_op_defs[] = {
757 [IORING_OP_NOP] = {},
758 [IORING_OP_READV] = {
760 .unbound_nonreg_file = 1,
763 .needs_async_data = 1,
764 .async_size = sizeof(struct io_async_rw),
765 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
767 [IORING_OP_WRITEV] = {
770 .unbound_nonreg_file = 1,
773 .needs_async_data = 1,
774 .async_size = sizeof(struct io_async_rw),
775 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
777 [IORING_OP_FSYNC] = {
779 .work_flags = IO_WQ_WORK_BLKCG,
781 [IORING_OP_READ_FIXED] = {
783 .unbound_nonreg_file = 1,
785 .async_size = sizeof(struct io_async_rw),
786 .work_flags = IO_WQ_WORK_BLKCG,
788 [IORING_OP_WRITE_FIXED] = {
791 .unbound_nonreg_file = 1,
794 .async_size = sizeof(struct io_async_rw),
795 .work_flags = IO_WQ_WORK_BLKCG,
797 [IORING_OP_POLL_ADD] = {
799 .unbound_nonreg_file = 1,
801 [IORING_OP_POLL_REMOVE] = {},
802 [IORING_OP_SYNC_FILE_RANGE] = {
804 .work_flags = IO_WQ_WORK_BLKCG,
806 [IORING_OP_SENDMSG] = {
808 .unbound_nonreg_file = 1,
810 .needs_async_data = 1,
811 .async_size = sizeof(struct io_async_msghdr),
812 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
815 [IORING_OP_RECVMSG] = {
817 .unbound_nonreg_file = 1,
820 .needs_async_data = 1,
821 .async_size = sizeof(struct io_async_msghdr),
822 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
825 [IORING_OP_TIMEOUT] = {
826 .needs_async_data = 1,
827 .async_size = sizeof(struct io_timeout_data),
828 .work_flags = IO_WQ_WORK_MM,
830 [IORING_OP_TIMEOUT_REMOVE] = {},
831 [IORING_OP_ACCEPT] = {
833 .unbound_nonreg_file = 1,
835 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
837 [IORING_OP_ASYNC_CANCEL] = {},
838 [IORING_OP_LINK_TIMEOUT] = {
839 .needs_async_data = 1,
840 .async_size = sizeof(struct io_timeout_data),
841 .work_flags = IO_WQ_WORK_MM,
843 [IORING_OP_CONNECT] = {
845 .unbound_nonreg_file = 1,
847 .needs_async_data = 1,
848 .async_size = sizeof(struct io_async_connect),
849 .work_flags = IO_WQ_WORK_MM,
851 [IORING_OP_FALLOCATE] = {
854 .work_flags = IO_WQ_WORK_BLKCG,
856 [IORING_OP_OPENAT] = {
857 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
860 [IORING_OP_CLOSE] = {
862 .needs_file_no_error = 1,
863 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
865 [IORING_OP_FILES_UPDATE] = {
866 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
868 [IORING_OP_STATX] = {
869 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
870 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
874 .unbound_nonreg_file = 1,
877 .async_size = sizeof(struct io_async_rw),
878 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
880 [IORING_OP_WRITE] = {
882 .unbound_nonreg_file = 1,
885 .async_size = sizeof(struct io_async_rw),
886 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
888 [IORING_OP_FADVISE] = {
890 .work_flags = IO_WQ_WORK_BLKCG,
892 [IORING_OP_MADVISE] = {
893 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
897 .unbound_nonreg_file = 1,
899 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
903 .unbound_nonreg_file = 1,
906 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
908 [IORING_OP_OPENAT2] = {
909 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
912 [IORING_OP_EPOLL_CTL] = {
913 .unbound_nonreg_file = 1,
914 .work_flags = IO_WQ_WORK_FILES,
916 [IORING_OP_SPLICE] = {
919 .unbound_nonreg_file = 1,
920 .work_flags = IO_WQ_WORK_BLKCG,
922 [IORING_OP_PROVIDE_BUFFERS] = {},
923 [IORING_OP_REMOVE_BUFFERS] = {},
927 .unbound_nonreg_file = 1,
931 enum io_mem_account {
936 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
937 struct io_comp_state *cs);
938 static void io_cqring_fill_event(struct io_kiocb *req, long res);
939 static void io_put_req(struct io_kiocb *req);
940 static void io_put_req_deferred(struct io_kiocb *req, int nr);
941 static void io_double_put_req(struct io_kiocb *req);
942 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
943 static void __io_queue_linked_timeout(struct io_kiocb *req);
944 static void io_queue_linked_timeout(struct io_kiocb *req);
945 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
946 struct io_uring_files_update *ip,
948 static void __io_clean_op(struct io_kiocb *req);
949 static struct file *io_file_get(struct io_submit_state *state,
950 struct io_kiocb *req, int fd, bool fixed);
951 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
952 static void io_file_put_work(struct work_struct *work);
954 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
955 struct iovec **iovec, struct iov_iter *iter,
957 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
958 const struct iovec *fast_iov,
959 struct iov_iter *iter, bool force);
961 static struct kmem_cache *req_cachep;
963 static const struct file_operations io_uring_fops;
965 struct sock *io_uring_get_socket(struct file *file)
967 #if defined(CONFIG_UNIX)
968 if (file->f_op == &io_uring_fops) {
969 struct io_ring_ctx *ctx = file->private_data;
971 return ctx->ring_sock->sk;
976 EXPORT_SYMBOL(io_uring_get_socket);
978 static inline void io_clean_op(struct io_kiocb *req)
980 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
985 static void io_sq_thread_drop_mm(void)
987 struct mm_struct *mm = current->mm;
990 kthread_unuse_mm(mm);
995 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
998 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
999 !ctx->sqo_task->mm ||
1000 !mmget_not_zero(ctx->sqo_task->mm)))
1002 kthread_use_mm(ctx->sqo_task->mm);
1008 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1009 struct io_kiocb *req)
1011 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1013 return __io_sq_thread_acquire_mm(ctx);
1016 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1017 struct cgroup_subsys_state **cur_css)
1020 #ifdef CONFIG_BLK_CGROUP
1021 /* puts the old one when swapping */
1022 if (*cur_css != ctx->sqo_blkcg_css) {
1023 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1024 *cur_css = ctx->sqo_blkcg_css;
1029 static void io_sq_thread_unassociate_blkcg(void)
1031 #ifdef CONFIG_BLK_CGROUP
1032 kthread_associate_blkcg(NULL);
1036 static inline void req_set_fail_links(struct io_kiocb *req)
1038 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1039 req->flags |= REQ_F_FAIL_LINK;
1043 * None of these are dereferenced, they are simply used to check if any of
1044 * them have changed. If we're under current and check they are still the
1045 * same, we're fine to grab references to them for actual out-of-line use.
1047 static void io_init_identity(struct io_identity *id)
1049 id->files = current->files;
1050 id->mm = current->mm;
1051 #ifdef CONFIG_BLK_CGROUP
1053 id->blkcg_css = blkcg_css();
1056 id->creds = current_cred();
1057 id->nsproxy = current->nsproxy;
1058 id->fs = current->fs;
1059 id->fsize = rlimit(RLIMIT_FSIZE);
1060 refcount_set(&id->count, 1);
1064 * Note: must call io_req_init_async() for the first time you
1065 * touch any members of io_wq_work.
1067 static inline void io_req_init_async(struct io_kiocb *req)
1069 struct io_uring_task *tctx = current->io_uring;
1071 if (req->flags & REQ_F_WORK_INITIALIZED)
1074 memset(&req->work, 0, sizeof(req->work));
1075 req->flags |= REQ_F_WORK_INITIALIZED;
1077 /* Grab a ref if this isn't our static identity */
1078 req->work.identity = tctx->identity;
1079 if (tctx->identity != &tctx->__identity)
1080 refcount_inc(&req->work.identity->count);
1083 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1085 return ctx->flags & IORING_SETUP_SQPOLL;
1088 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1090 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1092 complete(&ctx->ref_comp);
1095 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1097 return !req->timeout.off;
1100 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1102 struct io_ring_ctx *ctx;
1105 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1109 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1110 if (!ctx->fallback_req)
1114 * Use 5 bits less than the max cq entries, that should give us around
1115 * 32 entries per hash list if totally full and uniformly spread.
1117 hash_bits = ilog2(p->cq_entries);
1121 ctx->cancel_hash_bits = hash_bits;
1122 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1124 if (!ctx->cancel_hash)
1126 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1128 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1129 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1132 ctx->flags = p->flags;
1133 init_waitqueue_head(&ctx->sqo_sq_wait);
1134 INIT_LIST_HEAD(&ctx->sqd_list);
1135 init_waitqueue_head(&ctx->cq_wait);
1136 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1137 init_completion(&ctx->ref_comp);
1138 init_completion(&ctx->sq_thread_comp);
1139 idr_init(&ctx->io_buffer_idr);
1140 idr_init(&ctx->personality_idr);
1141 mutex_init(&ctx->uring_lock);
1142 init_waitqueue_head(&ctx->wait);
1143 spin_lock_init(&ctx->completion_lock);
1144 INIT_LIST_HEAD(&ctx->iopoll_list);
1145 INIT_LIST_HEAD(&ctx->defer_list);
1146 INIT_LIST_HEAD(&ctx->timeout_list);
1147 init_waitqueue_head(&ctx->inflight_wait);
1148 spin_lock_init(&ctx->inflight_lock);
1149 INIT_LIST_HEAD(&ctx->inflight_list);
1150 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1151 init_llist_head(&ctx->file_put_llist);
1154 if (ctx->fallback_req)
1155 kmem_cache_free(req_cachep, ctx->fallback_req);
1156 kfree(ctx->cancel_hash);
1161 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1163 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1164 struct io_ring_ctx *ctx = req->ctx;
1166 return seq != ctx->cached_cq_tail
1167 + atomic_read(&ctx->cached_cq_overflow);
1173 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1175 struct io_rings *rings = ctx->rings;
1177 /* order cqe stores with ring update */
1178 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1180 if (wq_has_sleeper(&ctx->cq_wait)) {
1181 wake_up_interruptible(&ctx->cq_wait);
1182 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1186 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1188 if (req->work.identity == &tctx->__identity)
1190 if (refcount_dec_and_test(&req->work.identity->count))
1191 kfree(req->work.identity);
1194 static void io_req_clean_work(struct io_kiocb *req)
1196 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1199 req->flags &= ~REQ_F_WORK_INITIALIZED;
1201 if (req->work.flags & IO_WQ_WORK_MM) {
1202 mmdrop(req->work.identity->mm);
1203 req->work.flags &= ~IO_WQ_WORK_MM;
1205 #ifdef CONFIG_BLK_CGROUP
1206 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1207 css_put(req->work.identity->blkcg_css);
1208 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1211 if (req->work.flags & IO_WQ_WORK_CREDS) {
1212 put_cred(req->work.identity->creds);
1213 req->work.flags &= ~IO_WQ_WORK_CREDS;
1215 if (req->work.flags & IO_WQ_WORK_FS) {
1216 struct fs_struct *fs = req->work.identity->fs;
1218 spin_lock(&req->work.identity->fs->lock);
1221 spin_unlock(&req->work.identity->fs->lock);
1224 req->work.flags &= ~IO_WQ_WORK_FS;
1227 io_put_identity(req->task->io_uring, req);
1231 * Create a private copy of io_identity, since some fields don't match
1232 * the current context.
1234 static bool io_identity_cow(struct io_kiocb *req)
1236 struct io_uring_task *tctx = current->io_uring;
1237 const struct cred *creds = NULL;
1238 struct io_identity *id;
1240 if (req->work.flags & IO_WQ_WORK_CREDS)
1241 creds = req->work.identity->creds;
1243 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1244 if (unlikely(!id)) {
1245 req->work.flags |= IO_WQ_WORK_CANCEL;
1250 * We can safely just re-init the creds we copied Either the field
1251 * matches the current one, or we haven't grabbed it yet. The only
1252 * exception is ->creds, through registered personalities, so handle
1253 * that one separately.
1255 io_init_identity(id);
1257 req->work.identity->creds = creds;
1259 /* add one for this request */
1260 refcount_inc(&id->count);
1262 /* drop old identity, assign new one. one ref for req, one for tctx */
1263 if (req->work.identity != tctx->identity &&
1264 refcount_sub_and_test(2, &req->work.identity->count))
1265 kfree(req->work.identity);
1267 req->work.identity = id;
1268 tctx->identity = id;
1272 static bool io_grab_identity(struct io_kiocb *req)
1274 const struct io_op_def *def = &io_op_defs[req->opcode];
1275 struct io_identity *id = req->work.identity;
1276 struct io_ring_ctx *ctx = req->ctx;
1278 if (def->needs_fsize && id->fsize != rlimit(RLIMIT_FSIZE))
1281 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1282 (def->work_flags & IO_WQ_WORK_FILES) &&
1283 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1284 if (id->files != current->files ||
1285 id->nsproxy != current->nsproxy)
1287 atomic_inc(&id->files->count);
1288 get_nsproxy(id->nsproxy);
1289 req->flags |= REQ_F_INFLIGHT;
1291 spin_lock_irq(&ctx->inflight_lock);
1292 list_add(&req->inflight_entry, &ctx->inflight_list);
1293 spin_unlock_irq(&ctx->inflight_lock);
1294 req->work.flags |= IO_WQ_WORK_FILES;
1296 #ifdef CONFIG_BLK_CGROUP
1297 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1298 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1300 if (id->blkcg_css != blkcg_css()) {
1305 * This should be rare, either the cgroup is dying or the task
1306 * is moving cgroups. Just punt to root for the handful of ios.
1308 if (css_tryget_online(id->blkcg_css))
1309 req->work.flags |= IO_WQ_WORK_BLKCG;
1313 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1314 if (id->creds != current_cred())
1316 get_cred(id->creds);
1317 req->work.flags |= IO_WQ_WORK_CREDS;
1319 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1320 (def->work_flags & IO_WQ_WORK_FS)) {
1321 if (current->fs != id->fs)
1323 spin_lock(&id->fs->lock);
1324 if (!id->fs->in_exec) {
1326 req->work.flags |= IO_WQ_WORK_FS;
1328 req->work.flags |= IO_WQ_WORK_CANCEL;
1330 spin_unlock(¤t->fs->lock);
1336 static void io_prep_async_work(struct io_kiocb *req)
1338 const struct io_op_def *def = &io_op_defs[req->opcode];
1339 struct io_ring_ctx *ctx = req->ctx;
1340 struct io_identity *id;
1342 io_req_init_async(req);
1343 id = req->work.identity;
1345 if (req->flags & REQ_F_ISREG) {
1346 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1347 io_wq_hash_work(&req->work, file_inode(req->file));
1349 if (def->unbound_nonreg_file)
1350 req->work.flags |= IO_WQ_WORK_UNBOUND;
1353 /* ->mm can never change on us */
1354 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1355 (def->work_flags & IO_WQ_WORK_MM)) {
1357 req->work.flags |= IO_WQ_WORK_MM;
1360 /* if we fail grabbing identity, we must COW, regrab, and retry */
1361 if (io_grab_identity(req))
1364 if (!io_identity_cow(req))
1367 /* can't fail at this point */
1368 if (!io_grab_identity(req))
1372 static void io_prep_async_link(struct io_kiocb *req)
1374 struct io_kiocb *cur;
1376 io_prep_async_work(req);
1377 if (req->flags & REQ_F_LINK_HEAD)
1378 list_for_each_entry(cur, &req->link_list, link_list)
1379 io_prep_async_work(cur);
1382 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1384 struct io_ring_ctx *ctx = req->ctx;
1385 struct io_kiocb *link = io_prep_linked_timeout(req);
1387 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1388 &req->work, req->flags);
1389 io_wq_enqueue(ctx->io_wq, &req->work);
1393 static void io_queue_async_work(struct io_kiocb *req)
1395 struct io_kiocb *link;
1397 /* init ->work of the whole link before punting */
1398 io_prep_async_link(req);
1399 link = __io_queue_async_work(req);
1402 io_queue_linked_timeout(link);
1405 static void io_kill_timeout(struct io_kiocb *req)
1407 struct io_timeout_data *io = req->async_data;
1410 ret = hrtimer_try_to_cancel(&io->timer);
1412 atomic_set(&req->ctx->cq_timeouts,
1413 atomic_read(&req->ctx->cq_timeouts) + 1);
1414 list_del_init(&req->timeout.list);
1415 io_cqring_fill_event(req, 0);
1416 io_put_req_deferred(req, 1);
1420 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1422 struct io_ring_ctx *ctx = req->ctx;
1424 if (!tsk || req->task == tsk)
1426 if (ctx->flags & IORING_SETUP_SQPOLL) {
1427 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1434 * Returns true if we found and killed one or more timeouts
1436 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1438 struct io_kiocb *req, *tmp;
1441 spin_lock_irq(&ctx->completion_lock);
1442 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1443 if (io_task_match(req, tsk)) {
1444 io_kill_timeout(req);
1448 spin_unlock_irq(&ctx->completion_lock);
1449 return canceled != 0;
1452 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1455 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1456 struct io_defer_entry, list);
1457 struct io_kiocb *link;
1459 if (req_need_defer(de->req, de->seq))
1461 list_del_init(&de->list);
1462 /* punt-init is done before queueing for defer */
1463 link = __io_queue_async_work(de->req);
1465 __io_queue_linked_timeout(link);
1466 /* drop submission reference */
1467 io_put_req_deferred(link, 1);
1470 } while (!list_empty(&ctx->defer_list));
1473 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1475 while (!list_empty(&ctx->timeout_list)) {
1476 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1477 struct io_kiocb, timeout.list);
1479 if (io_is_timeout_noseq(req))
1481 if (req->timeout.target_seq != ctx->cached_cq_tail
1482 - atomic_read(&ctx->cq_timeouts))
1485 list_del_init(&req->timeout.list);
1486 io_kill_timeout(req);
1490 static void io_commit_cqring(struct io_ring_ctx *ctx)
1492 io_flush_timeouts(ctx);
1493 __io_commit_cqring(ctx);
1495 if (unlikely(!list_empty(&ctx->defer_list)))
1496 __io_queue_deferred(ctx);
1499 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1501 struct io_rings *r = ctx->rings;
1503 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1506 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1508 struct io_rings *rings = ctx->rings;
1511 tail = ctx->cached_cq_tail;
1513 * writes to the cq entry need to come after reading head; the
1514 * control dependency is enough as we're using WRITE_ONCE to
1517 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1520 ctx->cached_cq_tail++;
1521 return &rings->cqes[tail & ctx->cq_mask];
1524 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1528 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1530 if (!ctx->eventfd_async)
1532 return io_wq_current_is_worker();
1535 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1537 if (waitqueue_active(&ctx->wait))
1538 wake_up(&ctx->wait);
1539 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1540 wake_up(&ctx->sq_data->wait);
1541 if (io_should_trigger_evfd(ctx))
1542 eventfd_signal(ctx->cq_ev_fd, 1);
1545 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1547 if (list_empty(&ctx->cq_overflow_list)) {
1548 clear_bit(0, &ctx->sq_check_overflow);
1549 clear_bit(0, &ctx->cq_check_overflow);
1550 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1554 static inline bool io_match_files(struct io_kiocb *req,
1555 struct files_struct *files)
1559 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1560 (req->work.flags & IO_WQ_WORK_FILES))
1561 return req->work.identity->files == files;
1565 /* Returns true if there are no backlogged entries after the flush */
1566 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1567 struct task_struct *tsk,
1568 struct files_struct *files)
1570 struct io_rings *rings = ctx->rings;
1571 struct io_kiocb *req, *tmp;
1572 struct io_uring_cqe *cqe;
1573 unsigned long flags;
1577 if (list_empty_careful(&ctx->cq_overflow_list))
1579 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1580 rings->cq_ring_entries))
1584 spin_lock_irqsave(&ctx->completion_lock, flags);
1586 /* if force is set, the ring is going away. always drop after that */
1588 ctx->cq_overflow_flushed = 1;
1591 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1592 if (tsk && req->task != tsk)
1594 if (!io_match_files(req, files))
1597 cqe = io_get_cqring(ctx);
1601 list_move(&req->compl.list, &list);
1603 WRITE_ONCE(cqe->user_data, req->user_data);
1604 WRITE_ONCE(cqe->res, req->result);
1605 WRITE_ONCE(cqe->flags, req->compl.cflags);
1607 WRITE_ONCE(ctx->rings->cq_overflow,
1608 atomic_inc_return(&ctx->cached_cq_overflow));
1612 io_commit_cqring(ctx);
1613 io_cqring_mark_overflow(ctx);
1615 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1616 io_cqring_ev_posted(ctx);
1618 while (!list_empty(&list)) {
1619 req = list_first_entry(&list, struct io_kiocb, compl.list);
1620 list_del(&req->compl.list);
1627 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1629 struct io_ring_ctx *ctx = req->ctx;
1630 struct io_uring_cqe *cqe;
1632 trace_io_uring_complete(ctx, req->user_data, res);
1635 * If we can't get a cq entry, userspace overflowed the
1636 * submission (by quite a lot). Increment the overflow count in
1639 cqe = io_get_cqring(ctx);
1641 WRITE_ONCE(cqe->user_data, req->user_data);
1642 WRITE_ONCE(cqe->res, res);
1643 WRITE_ONCE(cqe->flags, cflags);
1644 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1646 * If we're in ring overflow flush mode, or in task cancel mode,
1647 * then we cannot store the request for later flushing, we need
1648 * to drop it on the floor.
1650 WRITE_ONCE(ctx->rings->cq_overflow,
1651 atomic_inc_return(&ctx->cached_cq_overflow));
1653 if (list_empty(&ctx->cq_overflow_list)) {
1654 set_bit(0, &ctx->sq_check_overflow);
1655 set_bit(0, &ctx->cq_check_overflow);
1656 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1660 req->compl.cflags = cflags;
1661 refcount_inc(&req->refs);
1662 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1666 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1668 __io_cqring_fill_event(req, res, 0);
1671 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1673 struct io_ring_ctx *ctx = req->ctx;
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->completion_lock, flags);
1677 __io_cqring_fill_event(req, res, cflags);
1678 io_commit_cqring(ctx);
1679 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1681 io_cqring_ev_posted(ctx);
1684 static void io_submit_flush_completions(struct io_comp_state *cs)
1686 struct io_ring_ctx *ctx = cs->ctx;
1688 spin_lock_irq(&ctx->completion_lock);
1689 while (!list_empty(&cs->list)) {
1690 struct io_kiocb *req;
1692 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1693 list_del(&req->compl.list);
1694 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1697 * io_free_req() doesn't care about completion_lock unless one
1698 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1699 * because of a potential deadlock with req->work.fs->lock
1701 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1702 |REQ_F_WORK_INITIALIZED)) {
1703 spin_unlock_irq(&ctx->completion_lock);
1705 spin_lock_irq(&ctx->completion_lock);
1710 io_commit_cqring(ctx);
1711 spin_unlock_irq(&ctx->completion_lock);
1713 io_cqring_ev_posted(ctx);
1717 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1718 struct io_comp_state *cs)
1721 io_cqring_add_event(req, res, cflags);
1726 req->compl.cflags = cflags;
1727 list_add_tail(&req->compl.list, &cs->list);
1729 io_submit_flush_completions(cs);
1733 static void io_req_complete(struct io_kiocb *req, long res)
1735 __io_req_complete(req, res, 0, NULL);
1738 static inline bool io_is_fallback_req(struct io_kiocb *req)
1740 return req == (struct io_kiocb *)
1741 ((unsigned long) req->ctx->fallback_req & ~1UL);
1744 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1746 struct io_kiocb *req;
1748 req = ctx->fallback_req;
1749 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1755 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1756 struct io_submit_state *state)
1758 if (!state->free_reqs) {
1759 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1763 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1764 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1767 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1768 * retry single alloc to be on the safe side.
1770 if (unlikely(ret <= 0)) {
1771 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1772 if (!state->reqs[0])
1776 state->free_reqs = ret;
1780 return state->reqs[state->free_reqs];
1782 return io_get_fallback_req(ctx);
1785 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1789 percpu_ref_put(req->fixed_file_refs);
1794 static void io_dismantle_req(struct io_kiocb *req)
1798 if (req->async_data)
1799 kfree(req->async_data);
1801 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1803 io_req_clean_work(req);
1806 static void __io_free_req(struct io_kiocb *req)
1808 struct io_uring_task *tctx = req->task->io_uring;
1809 struct io_ring_ctx *ctx = req->ctx;
1811 io_dismantle_req(req);
1813 atomic_long_inc(&tctx->req_complete);
1815 wake_up(&tctx->wait);
1816 put_task_struct(req->task);
1818 if (likely(!io_is_fallback_req(req)))
1819 kmem_cache_free(req_cachep, req);
1821 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1822 percpu_ref_put(&ctx->refs);
1825 static bool io_link_cancel_timeout(struct io_kiocb *req)
1827 struct io_timeout_data *io = req->async_data;
1828 struct io_ring_ctx *ctx = req->ctx;
1831 ret = hrtimer_try_to_cancel(&io->timer);
1833 io_cqring_fill_event(req, -ECANCELED);
1834 io_commit_cqring(ctx);
1835 req->flags &= ~REQ_F_LINK_HEAD;
1836 io_put_req_deferred(req, 1);
1843 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1845 struct io_kiocb *link;
1848 if (list_empty(&req->link_list))
1850 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1851 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1854 list_del_init(&link->link_list);
1855 wake_ev = io_link_cancel_timeout(link);
1856 req->flags &= ~REQ_F_LINK_TIMEOUT;
1860 static void io_kill_linked_timeout(struct io_kiocb *req)
1862 struct io_ring_ctx *ctx = req->ctx;
1863 unsigned long flags;
1866 spin_lock_irqsave(&ctx->completion_lock, flags);
1867 wake_ev = __io_kill_linked_timeout(req);
1868 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1871 io_cqring_ev_posted(ctx);
1874 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1876 struct io_kiocb *nxt;
1879 * The list should never be empty when we are called here. But could
1880 * potentially happen if the chain is messed up, check to be on the
1883 if (unlikely(list_empty(&req->link_list)))
1886 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1887 list_del_init(&req->link_list);
1888 if (!list_empty(&nxt->link_list))
1889 nxt->flags |= REQ_F_LINK_HEAD;
1894 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1896 static void __io_fail_links(struct io_kiocb *req)
1898 struct io_ring_ctx *ctx = req->ctx;
1900 while (!list_empty(&req->link_list)) {
1901 struct io_kiocb *link = list_first_entry(&req->link_list,
1902 struct io_kiocb, link_list);
1904 list_del_init(&link->link_list);
1905 trace_io_uring_fail_link(req, link);
1907 io_cqring_fill_event(link, -ECANCELED);
1910 * It's ok to free under spinlock as they're not linked anymore,
1911 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1914 if (link->flags & REQ_F_WORK_INITIALIZED)
1915 io_put_req_deferred(link, 2);
1917 io_double_put_req(link);
1920 io_commit_cqring(ctx);
1923 static void io_fail_links(struct io_kiocb *req)
1925 struct io_ring_ctx *ctx = req->ctx;
1926 unsigned long flags;
1928 spin_lock_irqsave(&ctx->completion_lock, flags);
1929 __io_fail_links(req);
1930 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1932 io_cqring_ev_posted(ctx);
1935 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1937 req->flags &= ~REQ_F_LINK_HEAD;
1938 if (req->flags & REQ_F_LINK_TIMEOUT)
1939 io_kill_linked_timeout(req);
1942 * If LINK is set, we have dependent requests in this chain. If we
1943 * didn't fail this request, queue the first one up, moving any other
1944 * dependencies to the next request. In case of failure, fail the rest
1947 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1948 return io_req_link_next(req);
1953 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1955 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1957 return __io_req_find_next(req);
1960 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1962 struct task_struct *tsk = req->task;
1963 struct io_ring_ctx *ctx = req->ctx;
1966 if (tsk->flags & PF_EXITING)
1970 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1971 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1972 * processing task_work. There's no reliable way to tell if TWA_RESUME
1976 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1977 notify = TWA_SIGNAL;
1979 ret = task_work_add(tsk, &req->task_work, notify);
1981 wake_up_process(tsk);
1986 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1988 struct io_ring_ctx *ctx = req->ctx;
1990 spin_lock_irq(&ctx->completion_lock);
1991 io_cqring_fill_event(req, error);
1992 io_commit_cqring(ctx);
1993 spin_unlock_irq(&ctx->completion_lock);
1995 io_cqring_ev_posted(ctx);
1996 req_set_fail_links(req);
1997 io_double_put_req(req);
2000 static void io_req_task_cancel(struct callback_head *cb)
2002 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2003 struct io_ring_ctx *ctx = req->ctx;
2005 __io_req_task_cancel(req, -ECANCELED);
2006 percpu_ref_put(&ctx->refs);
2009 static void __io_req_task_submit(struct io_kiocb *req)
2011 struct io_ring_ctx *ctx = req->ctx;
2013 if (!__io_sq_thread_acquire_mm(ctx)) {
2014 mutex_lock(&ctx->uring_lock);
2015 __io_queue_sqe(req, NULL);
2016 mutex_unlock(&ctx->uring_lock);
2018 __io_req_task_cancel(req, -EFAULT);
2022 static void io_req_task_submit(struct callback_head *cb)
2024 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2025 struct io_ring_ctx *ctx = req->ctx;
2027 __io_req_task_submit(req);
2028 percpu_ref_put(&ctx->refs);
2031 static void io_req_task_queue(struct io_kiocb *req)
2035 init_task_work(&req->task_work, io_req_task_submit);
2036 percpu_ref_get(&req->ctx->refs);
2038 ret = io_req_task_work_add(req, true);
2039 if (unlikely(ret)) {
2040 struct task_struct *tsk;
2042 init_task_work(&req->task_work, io_req_task_cancel);
2043 tsk = io_wq_get_task(req->ctx->io_wq);
2044 task_work_add(tsk, &req->task_work, 0);
2045 wake_up_process(tsk);
2049 static void io_queue_next(struct io_kiocb *req)
2051 struct io_kiocb *nxt = io_req_find_next(req);
2054 io_req_task_queue(nxt);
2057 static void io_free_req(struct io_kiocb *req)
2064 void *reqs[IO_IOPOLL_BATCH];
2067 struct task_struct *task;
2071 static inline void io_init_req_batch(struct req_batch *rb)
2078 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2079 struct req_batch *rb)
2081 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2082 percpu_ref_put_many(&ctx->refs, rb->to_free);
2086 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2087 struct req_batch *rb)
2090 __io_req_free_batch_flush(ctx, rb);
2092 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2093 put_task_struct_many(rb->task, rb->task_refs);
2098 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2100 if (unlikely(io_is_fallback_req(req))) {
2104 if (req->flags & REQ_F_LINK_HEAD)
2107 if (req->task != rb->task) {
2109 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2110 put_task_struct_many(rb->task, rb->task_refs);
2112 rb->task = req->task;
2117 io_dismantle_req(req);
2118 rb->reqs[rb->to_free++] = req;
2119 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2120 __io_req_free_batch_flush(req->ctx, rb);
2124 * Drop reference to request, return next in chain (if there is one) if this
2125 * was the last reference to this request.
2127 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2129 struct io_kiocb *nxt = NULL;
2131 if (refcount_dec_and_test(&req->refs)) {
2132 nxt = io_req_find_next(req);
2138 static void io_put_req(struct io_kiocb *req)
2140 if (refcount_dec_and_test(&req->refs))
2144 static void io_put_req_deferred_cb(struct callback_head *cb)
2146 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2151 static void io_free_req_deferred(struct io_kiocb *req)
2155 init_task_work(&req->task_work, io_put_req_deferred_cb);
2156 ret = io_req_task_work_add(req, true);
2157 if (unlikely(ret)) {
2158 struct task_struct *tsk;
2160 tsk = io_wq_get_task(req->ctx->io_wq);
2161 task_work_add(tsk, &req->task_work, 0);
2162 wake_up_process(tsk);
2166 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2168 if (refcount_sub_and_test(refs, &req->refs))
2169 io_free_req_deferred(req);
2172 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2174 struct io_kiocb *nxt;
2177 * A ref is owned by io-wq in which context we're. So, if that's the
2178 * last one, it's safe to steal next work. False negatives are Ok,
2179 * it just will be re-punted async in io_put_work()
2181 if (refcount_read(&req->refs) != 1)
2184 nxt = io_req_find_next(req);
2185 return nxt ? &nxt->work : NULL;
2188 static void io_double_put_req(struct io_kiocb *req)
2190 /* drop both submit and complete references */
2191 if (refcount_sub_and_test(2, &req->refs))
2195 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2197 struct io_rings *rings = ctx->rings;
2199 if (test_bit(0, &ctx->cq_check_overflow)) {
2201 * noflush == true is from the waitqueue handler, just ensure
2202 * we wake up the task, and the next invocation will flush the
2203 * entries. We cannot safely to it from here.
2205 if (noflush && !list_empty(&ctx->cq_overflow_list))
2208 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2211 /* See comment at the top of this file */
2213 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2216 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2218 struct io_rings *rings = ctx->rings;
2220 /* make sure SQ entry isn't read before tail */
2221 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2224 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2226 unsigned int cflags;
2228 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2229 cflags |= IORING_CQE_F_BUFFER;
2230 req->flags &= ~REQ_F_BUFFER_SELECTED;
2235 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2237 struct io_buffer *kbuf;
2239 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2240 return io_put_kbuf(req, kbuf);
2243 static inline bool io_run_task_work(void)
2246 * Not safe to run on exiting task, and the task_work handling will
2247 * not add work to such a task.
2249 if (unlikely(current->flags & PF_EXITING))
2251 if (current->task_works) {
2252 __set_current_state(TASK_RUNNING);
2260 static void io_iopoll_queue(struct list_head *again)
2262 struct io_kiocb *req;
2265 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2266 list_del(&req->inflight_entry);
2267 __io_complete_rw(req, -EAGAIN, 0, NULL);
2268 } while (!list_empty(again));
2272 * Find and free completed poll iocbs
2274 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2275 struct list_head *done)
2277 struct req_batch rb;
2278 struct io_kiocb *req;
2281 /* order with ->result store in io_complete_rw_iopoll() */
2284 io_init_req_batch(&rb);
2285 while (!list_empty(done)) {
2288 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2289 if (READ_ONCE(req->result) == -EAGAIN) {
2291 req->iopoll_completed = 0;
2292 list_move_tail(&req->inflight_entry, &again);
2295 list_del(&req->inflight_entry);
2297 if (req->flags & REQ_F_BUFFER_SELECTED)
2298 cflags = io_put_rw_kbuf(req);
2300 __io_cqring_fill_event(req, req->result, cflags);
2303 if (refcount_dec_and_test(&req->refs))
2304 io_req_free_batch(&rb, req);
2307 io_commit_cqring(ctx);
2308 if (ctx->flags & IORING_SETUP_SQPOLL)
2309 io_cqring_ev_posted(ctx);
2310 io_req_free_batch_finish(ctx, &rb);
2312 if (!list_empty(&again))
2313 io_iopoll_queue(&again);
2316 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2319 struct io_kiocb *req, *tmp;
2325 * Only spin for completions if we don't have multiple devices hanging
2326 * off our complete list, and we're under the requested amount.
2328 spin = !ctx->poll_multi_file && *nr_events < min;
2331 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2332 struct kiocb *kiocb = &req->rw.kiocb;
2335 * Move completed and retryable entries to our local lists.
2336 * If we find a request that requires polling, break out
2337 * and complete those lists first, if we have entries there.
2339 if (READ_ONCE(req->iopoll_completed)) {
2340 list_move_tail(&req->inflight_entry, &done);
2343 if (!list_empty(&done))
2346 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2350 /* iopoll may have completed current req */
2351 if (READ_ONCE(req->iopoll_completed))
2352 list_move_tail(&req->inflight_entry, &done);
2359 if (!list_empty(&done))
2360 io_iopoll_complete(ctx, nr_events, &done);
2366 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2367 * non-spinning poll check - we'll still enter the driver poll loop, but only
2368 * as a non-spinning completion check.
2370 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2373 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2376 ret = io_do_iopoll(ctx, nr_events, min);
2379 if (*nr_events >= min)
2387 * We can't just wait for polled events to come to us, we have to actively
2388 * find and complete them.
2390 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2392 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2395 mutex_lock(&ctx->uring_lock);
2396 while (!list_empty(&ctx->iopoll_list)) {
2397 unsigned int nr_events = 0;
2399 io_do_iopoll(ctx, &nr_events, 0);
2401 /* let it sleep and repeat later if can't complete a request */
2405 * Ensure we allow local-to-the-cpu processing to take place,
2406 * in this case we need to ensure that we reap all events.
2407 * Also let task_work, etc. to progress by releasing the mutex
2409 if (need_resched()) {
2410 mutex_unlock(&ctx->uring_lock);
2412 mutex_lock(&ctx->uring_lock);
2415 mutex_unlock(&ctx->uring_lock);
2418 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2420 unsigned int nr_events = 0;
2421 int iters = 0, ret = 0;
2424 * We disallow the app entering submit/complete with polling, but we
2425 * still need to lock the ring to prevent racing with polled issue
2426 * that got punted to a workqueue.
2428 mutex_lock(&ctx->uring_lock);
2431 * Don't enter poll loop if we already have events pending.
2432 * If we do, we can potentially be spinning for commands that
2433 * already triggered a CQE (eg in error).
2435 if (io_cqring_events(ctx, false))
2439 * If a submit got punted to a workqueue, we can have the
2440 * application entering polling for a command before it gets
2441 * issued. That app will hold the uring_lock for the duration
2442 * of the poll right here, so we need to take a breather every
2443 * now and then to ensure that the issue has a chance to add
2444 * the poll to the issued list. Otherwise we can spin here
2445 * forever, while the workqueue is stuck trying to acquire the
2448 if (!(++iters & 7)) {
2449 mutex_unlock(&ctx->uring_lock);
2451 mutex_lock(&ctx->uring_lock);
2454 ret = io_iopoll_getevents(ctx, &nr_events, min);
2458 } while (min && !nr_events && !need_resched());
2460 mutex_unlock(&ctx->uring_lock);
2464 static void kiocb_end_write(struct io_kiocb *req)
2467 * Tell lockdep we inherited freeze protection from submission
2470 if (req->flags & REQ_F_ISREG) {
2471 struct inode *inode = file_inode(req->file);
2473 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2475 file_end_write(req->file);
2478 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2479 struct io_comp_state *cs)
2481 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2484 if (kiocb->ki_flags & IOCB_WRITE)
2485 kiocb_end_write(req);
2487 if (res != req->result)
2488 req_set_fail_links(req);
2489 if (req->flags & REQ_F_BUFFER_SELECTED)
2490 cflags = io_put_rw_kbuf(req);
2491 __io_req_complete(req, res, cflags, cs);
2495 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2497 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2498 ssize_t ret = -ECANCELED;
2499 struct iov_iter iter;
2507 switch (req->opcode) {
2508 case IORING_OP_READV:
2509 case IORING_OP_READ_FIXED:
2510 case IORING_OP_READ:
2513 case IORING_OP_WRITEV:
2514 case IORING_OP_WRITE_FIXED:
2515 case IORING_OP_WRITE:
2519 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2524 if (!req->async_data) {
2525 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2528 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2536 req_set_fail_links(req);
2537 io_req_complete(req, ret);
2542 static bool io_rw_reissue(struct io_kiocb *req, long res)
2545 umode_t mode = file_inode(req->file)->i_mode;
2548 if (!S_ISBLK(mode) && !S_ISREG(mode))
2550 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2553 ret = io_sq_thread_acquire_mm(req->ctx, req);
2555 if (io_resubmit_prep(req, ret)) {
2556 refcount_inc(&req->refs);
2557 io_queue_async_work(req);
2565 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2566 struct io_comp_state *cs)
2568 if (!io_rw_reissue(req, res))
2569 io_complete_rw_common(&req->rw.kiocb, res, cs);
2572 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2574 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2576 __io_complete_rw(req, res, res2, NULL);
2579 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2581 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2583 if (kiocb->ki_flags & IOCB_WRITE)
2584 kiocb_end_write(req);
2586 if (res != -EAGAIN && res != req->result)
2587 req_set_fail_links(req);
2589 WRITE_ONCE(req->result, res);
2590 /* order with io_poll_complete() checking ->result */
2592 WRITE_ONCE(req->iopoll_completed, 1);
2596 * After the iocb has been issued, it's safe to be found on the poll list.
2597 * Adding the kiocb to the list AFTER submission ensures that we don't
2598 * find it from a io_iopoll_getevents() thread before the issuer is done
2599 * accessing the kiocb cookie.
2601 static void io_iopoll_req_issued(struct io_kiocb *req)
2603 struct io_ring_ctx *ctx = req->ctx;
2606 * Track whether we have multiple files in our lists. This will impact
2607 * how we do polling eventually, not spinning if we're on potentially
2608 * different devices.
2610 if (list_empty(&ctx->iopoll_list)) {
2611 ctx->poll_multi_file = false;
2612 } else if (!ctx->poll_multi_file) {
2613 struct io_kiocb *list_req;
2615 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2617 if (list_req->file != req->file)
2618 ctx->poll_multi_file = true;
2622 * For fast devices, IO may have already completed. If it has, add
2623 * it to the front so we find it first.
2625 if (READ_ONCE(req->iopoll_completed))
2626 list_add(&req->inflight_entry, &ctx->iopoll_list);
2628 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2630 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2631 wq_has_sleeper(&ctx->sq_data->wait))
2632 wake_up(&ctx->sq_data->wait);
2635 static void __io_state_file_put(struct io_submit_state *state)
2637 if (state->has_refs)
2638 fput_many(state->file, state->has_refs);
2642 static inline void io_state_file_put(struct io_submit_state *state)
2645 __io_state_file_put(state);
2649 * Get as many references to a file as we have IOs left in this submission,
2650 * assuming most submissions are for one file, or at least that each file
2651 * has more than one submission.
2653 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2659 if (state->fd == fd) {
2663 __io_state_file_put(state);
2665 state->file = fget_many(fd, state->ios_left);
2670 state->has_refs = state->ios_left - 1;
2674 static bool io_bdev_nowait(struct block_device *bdev)
2677 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2684 * If we tracked the file through the SCM inflight mechanism, we could support
2685 * any file. For now, just ensure that anything potentially problematic is done
2688 static bool io_file_supports_async(struct file *file, int rw)
2690 umode_t mode = file_inode(file)->i_mode;
2692 if (S_ISBLK(mode)) {
2693 if (io_bdev_nowait(file->f_inode->i_bdev))
2697 if (S_ISCHR(mode) || S_ISSOCK(mode))
2699 if (S_ISREG(mode)) {
2700 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2701 file->f_op != &io_uring_fops)
2706 /* any ->read/write should understand O_NONBLOCK */
2707 if (file->f_flags & O_NONBLOCK)
2710 if (!(file->f_mode & FMODE_NOWAIT))
2714 return file->f_op->read_iter != NULL;
2716 return file->f_op->write_iter != NULL;
2719 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2721 struct io_ring_ctx *ctx = req->ctx;
2722 struct kiocb *kiocb = &req->rw.kiocb;
2726 if (S_ISREG(file_inode(req->file)->i_mode))
2727 req->flags |= REQ_F_ISREG;
2729 kiocb->ki_pos = READ_ONCE(sqe->off);
2730 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2731 req->flags |= REQ_F_CUR_POS;
2732 kiocb->ki_pos = req->file->f_pos;
2734 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2735 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2736 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2740 ioprio = READ_ONCE(sqe->ioprio);
2742 ret = ioprio_check_cap(ioprio);
2746 kiocb->ki_ioprio = ioprio;
2748 kiocb->ki_ioprio = get_current_ioprio();
2750 /* don't allow async punt if RWF_NOWAIT was requested */
2751 if (kiocb->ki_flags & IOCB_NOWAIT)
2752 req->flags |= REQ_F_NOWAIT;
2754 if (ctx->flags & IORING_SETUP_IOPOLL) {
2755 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2756 !kiocb->ki_filp->f_op->iopoll)
2759 kiocb->ki_flags |= IOCB_HIPRI;
2760 kiocb->ki_complete = io_complete_rw_iopoll;
2761 req->iopoll_completed = 0;
2763 if (kiocb->ki_flags & IOCB_HIPRI)
2765 kiocb->ki_complete = io_complete_rw;
2768 req->rw.addr = READ_ONCE(sqe->addr);
2769 req->rw.len = READ_ONCE(sqe->len);
2770 req->buf_index = READ_ONCE(sqe->buf_index);
2774 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2780 case -ERESTARTNOINTR:
2781 case -ERESTARTNOHAND:
2782 case -ERESTART_RESTARTBLOCK:
2784 * We can't just restart the syscall, since previously
2785 * submitted sqes may already be in progress. Just fail this
2791 kiocb->ki_complete(kiocb, ret, 0);
2795 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2796 struct io_comp_state *cs)
2798 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2799 struct io_async_rw *io = req->async_data;
2801 /* add previously done IO, if any */
2802 if (io && io->bytes_done > 0) {
2804 ret = io->bytes_done;
2806 ret += io->bytes_done;
2809 if (req->flags & REQ_F_CUR_POS)
2810 req->file->f_pos = kiocb->ki_pos;
2811 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2812 __io_complete_rw(req, ret, 0, cs);
2814 io_rw_done(kiocb, ret);
2817 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2818 struct iov_iter *iter)
2820 struct io_ring_ctx *ctx = req->ctx;
2821 size_t len = req->rw.len;
2822 struct io_mapped_ubuf *imu;
2823 u16 index, buf_index = req->buf_index;
2827 if (unlikely(buf_index >= ctx->nr_user_bufs))
2829 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2830 imu = &ctx->user_bufs[index];
2831 buf_addr = req->rw.addr;
2834 if (buf_addr + len < buf_addr)
2836 /* not inside the mapped region */
2837 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2841 * May not be a start of buffer, set size appropriately
2842 * and advance us to the beginning.
2844 offset = buf_addr - imu->ubuf;
2845 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2849 * Don't use iov_iter_advance() here, as it's really slow for
2850 * using the latter parts of a big fixed buffer - it iterates
2851 * over each segment manually. We can cheat a bit here, because
2854 * 1) it's a BVEC iter, we set it up
2855 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2856 * first and last bvec
2858 * So just find our index, and adjust the iterator afterwards.
2859 * If the offset is within the first bvec (or the whole first
2860 * bvec, just use iov_iter_advance(). This makes it easier
2861 * since we can just skip the first segment, which may not
2862 * be PAGE_SIZE aligned.
2864 const struct bio_vec *bvec = imu->bvec;
2866 if (offset <= bvec->bv_len) {
2867 iov_iter_advance(iter, offset);
2869 unsigned long seg_skip;
2871 /* skip first vec */
2872 offset -= bvec->bv_len;
2873 seg_skip = 1 + (offset >> PAGE_SHIFT);
2875 iter->bvec = bvec + seg_skip;
2876 iter->nr_segs -= seg_skip;
2877 iter->count -= bvec->bv_len + offset;
2878 iter->iov_offset = offset & ~PAGE_MASK;
2885 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2888 mutex_unlock(&ctx->uring_lock);
2891 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2894 * "Normal" inline submissions always hold the uring_lock, since we
2895 * grab it from the system call. Same is true for the SQPOLL offload.
2896 * The only exception is when we've detached the request and issue it
2897 * from an async worker thread, grab the lock for that case.
2900 mutex_lock(&ctx->uring_lock);
2903 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2904 int bgid, struct io_buffer *kbuf,
2907 struct io_buffer *head;
2909 if (req->flags & REQ_F_BUFFER_SELECTED)
2912 io_ring_submit_lock(req->ctx, needs_lock);
2914 lockdep_assert_held(&req->ctx->uring_lock);
2916 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2918 if (!list_empty(&head->list)) {
2919 kbuf = list_last_entry(&head->list, struct io_buffer,
2921 list_del(&kbuf->list);
2924 idr_remove(&req->ctx->io_buffer_idr, bgid);
2926 if (*len > kbuf->len)
2929 kbuf = ERR_PTR(-ENOBUFS);
2932 io_ring_submit_unlock(req->ctx, needs_lock);
2937 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2940 struct io_buffer *kbuf;
2943 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2944 bgid = req->buf_index;
2945 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2948 req->rw.addr = (u64) (unsigned long) kbuf;
2949 req->flags |= REQ_F_BUFFER_SELECTED;
2950 return u64_to_user_ptr(kbuf->addr);
2953 #ifdef CONFIG_COMPAT
2954 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2957 struct compat_iovec __user *uiov;
2958 compat_ssize_t clen;
2962 uiov = u64_to_user_ptr(req->rw.addr);
2963 if (!access_ok(uiov, sizeof(*uiov)))
2965 if (__get_user(clen, &uiov->iov_len))
2971 buf = io_rw_buffer_select(req, &len, needs_lock);
2973 return PTR_ERR(buf);
2974 iov[0].iov_base = buf;
2975 iov[0].iov_len = (compat_size_t) len;
2980 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2983 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2987 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2990 len = iov[0].iov_len;
2993 buf = io_rw_buffer_select(req, &len, needs_lock);
2995 return PTR_ERR(buf);
2996 iov[0].iov_base = buf;
2997 iov[0].iov_len = len;
3001 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3004 if (req->flags & REQ_F_BUFFER_SELECTED) {
3005 struct io_buffer *kbuf;
3007 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3008 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3009 iov[0].iov_len = kbuf->len;
3014 else if (req->rw.len > 1)
3017 #ifdef CONFIG_COMPAT
3018 if (req->ctx->compat)
3019 return io_compat_import(req, iov, needs_lock);
3022 return __io_iov_buffer_select(req, iov, needs_lock);
3025 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3026 struct iovec **iovec, struct iov_iter *iter,
3029 void __user *buf = u64_to_user_ptr(req->rw.addr);
3030 size_t sqe_len = req->rw.len;
3034 opcode = req->opcode;
3035 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3037 return io_import_fixed(req, rw, iter);
3040 /* buffer index only valid with fixed read/write, or buffer select */
3041 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3044 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3045 if (req->flags & REQ_F_BUFFER_SELECT) {
3046 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3048 return PTR_ERR(buf);
3049 req->rw.len = sqe_len;
3052 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3054 return ret < 0 ? ret : sqe_len;
3057 if (req->flags & REQ_F_BUFFER_SELECT) {
3058 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3060 ret = (*iovec)->iov_len;
3061 iov_iter_init(iter, rw, *iovec, 1, ret);
3067 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3071 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3072 struct iovec **iovec, struct iov_iter *iter,
3075 struct io_async_rw *iorw = req->async_data;
3078 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3080 return iov_iter_count(&iorw->iter);
3083 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3085 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3089 * For files that don't have ->read_iter() and ->write_iter(), handle them
3090 * by looping over ->read() or ->write() manually.
3092 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3093 struct iov_iter *iter)
3098 * Don't support polled IO through this interface, and we can't
3099 * support non-blocking either. For the latter, this just causes
3100 * the kiocb to be handled from an async context.
3102 if (kiocb->ki_flags & IOCB_HIPRI)
3104 if (kiocb->ki_flags & IOCB_NOWAIT)
3107 while (iov_iter_count(iter)) {
3111 if (!iov_iter_is_bvec(iter)) {
3112 iovec = iov_iter_iovec(iter);
3114 /* fixed buffers import bvec */
3115 iovec.iov_base = kmap(iter->bvec->bv_page)
3117 iovec.iov_len = min(iter->count,
3118 iter->bvec->bv_len - iter->iov_offset);
3122 nr = file->f_op->read(file, iovec.iov_base,
3123 iovec.iov_len, io_kiocb_ppos(kiocb));
3125 nr = file->f_op->write(file, iovec.iov_base,
3126 iovec.iov_len, io_kiocb_ppos(kiocb));
3129 if (iov_iter_is_bvec(iter))
3130 kunmap(iter->bvec->bv_page);
3138 if (nr != iovec.iov_len)
3140 iov_iter_advance(iter, nr);
3146 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3147 const struct iovec *fast_iov, struct iov_iter *iter)
3149 struct io_async_rw *rw = req->async_data;
3151 memcpy(&rw->iter, iter, sizeof(*iter));
3152 rw->free_iovec = iovec;
3154 /* can only be fixed buffers, no need to do anything */
3155 if (iter->type == ITER_BVEC)
3158 unsigned iov_off = 0;
3160 rw->iter.iov = rw->fast_iov;
3161 if (iter->iov != fast_iov) {
3162 iov_off = iter->iov - fast_iov;
3163 rw->iter.iov += iov_off;
3165 if (rw->fast_iov != fast_iov)
3166 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3167 sizeof(struct iovec) * iter->nr_segs);
3169 req->flags |= REQ_F_NEED_CLEANUP;
3173 static inline int __io_alloc_async_data(struct io_kiocb *req)
3175 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3176 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3177 return req->async_data == NULL;
3180 static int io_alloc_async_data(struct io_kiocb *req)
3182 if (!io_op_defs[req->opcode].needs_async_data)
3185 return __io_alloc_async_data(req);
3188 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3189 const struct iovec *fast_iov,
3190 struct iov_iter *iter, bool force)
3192 if (!force && !io_op_defs[req->opcode].needs_async_data)
3194 if (!req->async_data) {
3195 if (__io_alloc_async_data(req))
3198 io_req_map_rw(req, iovec, fast_iov, iter);
3203 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3205 struct io_async_rw *iorw = req->async_data;
3206 struct iovec *iov = iorw->fast_iov;
3209 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3210 if (unlikely(ret < 0))
3213 iorw->bytes_done = 0;
3214 iorw->free_iovec = iov;
3216 req->flags |= REQ_F_NEED_CLEANUP;
3220 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3224 ret = io_prep_rw(req, sqe);
3228 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3231 /* either don't need iovec imported or already have it */
3232 if (!req->async_data)
3234 return io_rw_prep_async(req, READ);
3238 * This is our waitqueue callback handler, registered through lock_page_async()
3239 * when we initially tried to do the IO with the iocb armed our waitqueue.
3240 * This gets called when the page is unlocked, and we generally expect that to
3241 * happen when the page IO is completed and the page is now uptodate. This will
3242 * queue a task_work based retry of the operation, attempting to copy the data
3243 * again. If the latter fails because the page was NOT uptodate, then we will
3244 * do a thread based blocking retry of the operation. That's the unexpected
3247 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3248 int sync, void *arg)
3250 struct wait_page_queue *wpq;
3251 struct io_kiocb *req = wait->private;
3252 struct wait_page_key *key = arg;
3255 wpq = container_of(wait, struct wait_page_queue, wait);
3257 if (!wake_page_match(wpq, key))
3260 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3261 list_del_init(&wait->entry);
3263 init_task_work(&req->task_work, io_req_task_submit);
3264 percpu_ref_get(&req->ctx->refs);
3266 /* submit ref gets dropped, acquire a new one */
3267 refcount_inc(&req->refs);
3268 ret = io_req_task_work_add(req, true);
3269 if (unlikely(ret)) {
3270 struct task_struct *tsk;
3272 /* queue just for cancelation */
3273 init_task_work(&req->task_work, io_req_task_cancel);
3274 tsk = io_wq_get_task(req->ctx->io_wq);
3275 task_work_add(tsk, &req->task_work, 0);
3276 wake_up_process(tsk);
3282 * This controls whether a given IO request should be armed for async page
3283 * based retry. If we return false here, the request is handed to the async
3284 * worker threads for retry. If we're doing buffered reads on a regular file,
3285 * we prepare a private wait_page_queue entry and retry the operation. This
3286 * will either succeed because the page is now uptodate and unlocked, or it
3287 * will register a callback when the page is unlocked at IO completion. Through
3288 * that callback, io_uring uses task_work to setup a retry of the operation.
3289 * That retry will attempt the buffered read again. The retry will generally
3290 * succeed, or in rare cases where it fails, we then fall back to using the
3291 * async worker threads for a blocking retry.
3293 static bool io_rw_should_retry(struct io_kiocb *req)
3295 struct io_async_rw *rw = req->async_data;
3296 struct wait_page_queue *wait = &rw->wpq;
3297 struct kiocb *kiocb = &req->rw.kiocb;
3299 /* never retry for NOWAIT, we just complete with -EAGAIN */
3300 if (req->flags & REQ_F_NOWAIT)
3303 /* Only for buffered IO */
3304 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3308 * just use poll if we can, and don't attempt if the fs doesn't
3309 * support callback based unlocks
3311 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3314 wait->wait.func = io_async_buf_func;
3315 wait->wait.private = req;
3316 wait->wait.flags = 0;
3317 INIT_LIST_HEAD(&wait->wait.entry);
3318 kiocb->ki_flags |= IOCB_WAITQ;
3319 kiocb->ki_flags &= ~IOCB_NOWAIT;
3320 kiocb->ki_waitq = wait;
3324 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3326 if (req->file->f_op->read_iter)
3327 return call_read_iter(req->file, &req->rw.kiocb, iter);
3328 else if (req->file->f_op->read)
3329 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3334 static int io_read(struct io_kiocb *req, bool force_nonblock,
3335 struct io_comp_state *cs)
3337 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3338 struct kiocb *kiocb = &req->rw.kiocb;
3339 struct iov_iter __iter, *iter = &__iter;
3340 struct io_async_rw *rw = req->async_data;
3341 ssize_t io_size, ret, ret2;
3348 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3351 iov_count = iov_iter_count(iter);
3353 req->result = io_size;
3356 /* Ensure we clear previously set non-block flag */
3357 if (!force_nonblock)
3358 kiocb->ki_flags &= ~IOCB_NOWAIT;
3360 kiocb->ki_flags |= IOCB_NOWAIT;
3363 /* If the file doesn't support async, just async punt */
3364 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3368 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3372 ret = io_iter_do_read(req, iter);
3376 } else if (ret == -EIOCBQUEUED) {
3379 } else if (ret == -EAGAIN) {
3380 /* IOPOLL retry should happen for io-wq threads */
3381 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3383 /* no retry on NONBLOCK marked file */
3384 if (req->file->f_flags & O_NONBLOCK)
3386 /* some cases will consume bytes even on error returns */
3387 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3390 } else if (ret < 0) {
3391 /* make sure -ERESTARTSYS -> -EINTR is done */
3395 /* read it all, or we did blocking attempt. no retry. */
3396 if (!iov_iter_count(iter) || !force_nonblock ||
3397 (req->file->f_flags & O_NONBLOCK))
3402 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3409 rw = req->async_data;
3410 /* it's copied and will be cleaned with ->io */
3412 /* now use our persistent iterator, if we aren't already */
3415 rw->bytes_done += ret;
3416 /* if we can retry, do so with the callbacks armed */
3417 if (!io_rw_should_retry(req)) {
3418 kiocb->ki_flags &= ~IOCB_WAITQ;
3423 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3424 * get -EIOCBQUEUED, then we'll get a notification when the desired
3425 * page gets unlocked. We can also get a partial read here, and if we
3426 * do, then just retry at the new offset.
3428 ret = io_iter_do_read(req, iter);
3429 if (ret == -EIOCBQUEUED) {
3432 } else if (ret > 0 && ret < io_size) {
3433 /* we got some bytes, but not all. retry. */
3437 kiocb_done(kiocb, ret, cs);
3440 /* it's reportedly faster than delegating the null check to kfree() */
3446 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3450 ret = io_prep_rw(req, sqe);
3454 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3457 /* either don't need iovec imported or already have it */
3458 if (!req->async_data)
3460 return io_rw_prep_async(req, WRITE);
3463 static int io_write(struct io_kiocb *req, bool force_nonblock,
3464 struct io_comp_state *cs)
3466 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3467 struct kiocb *kiocb = &req->rw.kiocb;
3468 struct iov_iter __iter, *iter = &__iter;
3469 struct io_async_rw *rw = req->async_data;
3471 ssize_t ret, ret2, io_size;
3476 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3479 iov_count = iov_iter_count(iter);
3481 req->result = io_size;
3483 /* Ensure we clear previously set non-block flag */
3484 if (!force_nonblock)
3485 kiocb->ki_flags &= ~IOCB_NOWAIT;
3487 kiocb->ki_flags |= IOCB_NOWAIT;
3489 /* If the file doesn't support async, just async punt */
3490 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3493 /* file path doesn't support NOWAIT for non-direct_IO */
3494 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3495 (req->flags & REQ_F_ISREG))
3498 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3503 * Open-code file_start_write here to grab freeze protection,
3504 * which will be released by another thread in
3505 * io_complete_rw(). Fool lockdep by telling it the lock got
3506 * released so that it doesn't complain about the held lock when
3507 * we return to userspace.
3509 if (req->flags & REQ_F_ISREG) {
3510 __sb_start_write(file_inode(req->file)->i_sb,
3511 SB_FREEZE_WRITE, true);
3512 __sb_writers_release(file_inode(req->file)->i_sb,
3515 kiocb->ki_flags |= IOCB_WRITE;
3517 if (req->file->f_op->write_iter)
3518 ret2 = call_write_iter(req->file, kiocb, iter);
3519 else if (req->file->f_op->write)
3520 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3525 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3526 * retry them without IOCB_NOWAIT.
3528 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3530 /* no retry on NONBLOCK marked file */
3531 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3533 if (!force_nonblock || ret2 != -EAGAIN) {
3534 /* IOPOLL retry should happen for io-wq threads */
3535 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3538 kiocb_done(kiocb, ret2, cs);
3541 /* some cases will consume bytes even on error returns */
3542 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3543 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3548 /* it's reportedly faster than delegating the null check to kfree() */
3554 static int __io_splice_prep(struct io_kiocb *req,
3555 const struct io_uring_sqe *sqe)
3557 struct io_splice* sp = &req->splice;
3558 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3560 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3564 sp->len = READ_ONCE(sqe->len);
3565 sp->flags = READ_ONCE(sqe->splice_flags);
3567 if (unlikely(sp->flags & ~valid_flags))
3570 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3571 (sp->flags & SPLICE_F_FD_IN_FIXED));
3574 req->flags |= REQ_F_NEED_CLEANUP;
3576 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3578 * Splice operation will be punted aync, and here need to
3579 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3581 io_req_init_async(req);
3582 req->work.flags |= IO_WQ_WORK_UNBOUND;
3588 static int io_tee_prep(struct io_kiocb *req,
3589 const struct io_uring_sqe *sqe)
3591 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3593 return __io_splice_prep(req, sqe);
3596 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3598 struct io_splice *sp = &req->splice;
3599 struct file *in = sp->file_in;
3600 struct file *out = sp->file_out;
3601 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3607 ret = do_tee(in, out, sp->len, flags);
3609 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3610 req->flags &= ~REQ_F_NEED_CLEANUP;
3613 req_set_fail_links(req);
3614 io_req_complete(req, ret);
3618 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3620 struct io_splice* sp = &req->splice;
3622 sp->off_in = READ_ONCE(sqe->splice_off_in);
3623 sp->off_out = READ_ONCE(sqe->off);
3624 return __io_splice_prep(req, sqe);
3627 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3629 struct io_splice *sp = &req->splice;
3630 struct file *in = sp->file_in;
3631 struct file *out = sp->file_out;
3632 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3633 loff_t *poff_in, *poff_out;
3639 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3640 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3643 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3645 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3646 req->flags &= ~REQ_F_NEED_CLEANUP;
3649 req_set_fail_links(req);
3650 io_req_complete(req, ret);
3655 * IORING_OP_NOP just posts a completion event, nothing else.
3657 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3659 struct io_ring_ctx *ctx = req->ctx;
3661 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3664 __io_req_complete(req, 0, 0, cs);
3668 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3670 struct io_ring_ctx *ctx = req->ctx;
3675 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3677 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3680 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3681 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3684 req->sync.off = READ_ONCE(sqe->off);
3685 req->sync.len = READ_ONCE(sqe->len);
3689 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3691 loff_t end = req->sync.off + req->sync.len;
3694 /* fsync always requires a blocking context */
3698 ret = vfs_fsync_range(req->file, req->sync.off,
3699 end > 0 ? end : LLONG_MAX,
3700 req->sync.flags & IORING_FSYNC_DATASYNC);
3702 req_set_fail_links(req);
3703 io_req_complete(req, ret);
3707 static int io_fallocate_prep(struct io_kiocb *req,
3708 const struct io_uring_sqe *sqe)
3710 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3712 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3715 req->sync.off = READ_ONCE(sqe->off);
3716 req->sync.len = READ_ONCE(sqe->addr);
3717 req->sync.mode = READ_ONCE(sqe->len);
3721 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3725 /* fallocate always requiring blocking context */
3728 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3731 req_set_fail_links(req);
3732 io_req_complete(req, ret);
3736 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3738 const char __user *fname;
3741 if (unlikely(sqe->ioprio || sqe->buf_index))
3743 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3746 /* open.how should be already initialised */
3747 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3748 req->open.how.flags |= O_LARGEFILE;
3750 req->open.dfd = READ_ONCE(sqe->fd);
3751 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3752 req->open.filename = getname(fname);
3753 if (IS_ERR(req->open.filename)) {
3754 ret = PTR_ERR(req->open.filename);
3755 req->open.filename = NULL;
3758 req->open.nofile = rlimit(RLIMIT_NOFILE);
3759 req->flags |= REQ_F_NEED_CLEANUP;
3763 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3769 mode = READ_ONCE(sqe->len);
3770 flags = READ_ONCE(sqe->open_flags);
3771 req->open.how = build_open_how(flags, mode);
3772 return __io_openat_prep(req, sqe);
3775 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3777 struct open_how __user *how;
3781 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3783 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3784 len = READ_ONCE(sqe->len);
3785 if (len < OPEN_HOW_SIZE_VER0)
3788 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3793 return __io_openat_prep(req, sqe);
3796 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3798 struct open_flags op;
3805 ret = build_open_flags(&req->open.how, &op);
3809 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3813 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3816 ret = PTR_ERR(file);
3818 fsnotify_open(file);
3819 fd_install(ret, file);
3822 putname(req->open.filename);
3823 req->flags &= ~REQ_F_NEED_CLEANUP;
3825 req_set_fail_links(req);
3826 io_req_complete(req, ret);
3830 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3832 return io_openat2(req, force_nonblock);
3835 static int io_remove_buffers_prep(struct io_kiocb *req,
3836 const struct io_uring_sqe *sqe)
3838 struct io_provide_buf *p = &req->pbuf;
3841 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3844 tmp = READ_ONCE(sqe->fd);
3845 if (!tmp || tmp > USHRT_MAX)
3848 memset(p, 0, sizeof(*p));
3850 p->bgid = READ_ONCE(sqe->buf_group);
3854 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3855 int bgid, unsigned nbufs)
3859 /* shouldn't happen */
3863 /* the head kbuf is the list itself */
3864 while (!list_empty(&buf->list)) {
3865 struct io_buffer *nxt;
3867 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3868 list_del(&nxt->list);
3875 idr_remove(&ctx->io_buffer_idr, bgid);
3880 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3881 struct io_comp_state *cs)
3883 struct io_provide_buf *p = &req->pbuf;
3884 struct io_ring_ctx *ctx = req->ctx;
3885 struct io_buffer *head;
3888 io_ring_submit_lock(ctx, !force_nonblock);
3890 lockdep_assert_held(&ctx->uring_lock);
3893 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3895 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3897 io_ring_submit_lock(ctx, !force_nonblock);
3899 req_set_fail_links(req);
3900 __io_req_complete(req, ret, 0, cs);
3904 static int io_provide_buffers_prep(struct io_kiocb *req,
3905 const struct io_uring_sqe *sqe)
3907 struct io_provide_buf *p = &req->pbuf;
3910 if (sqe->ioprio || sqe->rw_flags)
3913 tmp = READ_ONCE(sqe->fd);
3914 if (!tmp || tmp > USHRT_MAX)
3917 p->addr = READ_ONCE(sqe->addr);
3918 p->len = READ_ONCE(sqe->len);
3920 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3923 p->bgid = READ_ONCE(sqe->buf_group);
3924 tmp = READ_ONCE(sqe->off);
3925 if (tmp > USHRT_MAX)
3931 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3933 struct io_buffer *buf;
3934 u64 addr = pbuf->addr;
3935 int i, bid = pbuf->bid;
3937 for (i = 0; i < pbuf->nbufs; i++) {
3938 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3943 buf->len = pbuf->len;
3948 INIT_LIST_HEAD(&buf->list);
3951 list_add_tail(&buf->list, &(*head)->list);
3955 return i ? i : -ENOMEM;
3958 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3959 struct io_comp_state *cs)
3961 struct io_provide_buf *p = &req->pbuf;
3962 struct io_ring_ctx *ctx = req->ctx;
3963 struct io_buffer *head, *list;
3966 io_ring_submit_lock(ctx, !force_nonblock);
3968 lockdep_assert_held(&ctx->uring_lock);
3970 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3972 ret = io_add_buffers(p, &head);
3977 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3980 __io_remove_buffers(ctx, head, p->bgid, -1U);
3985 io_ring_submit_unlock(ctx, !force_nonblock);
3987 req_set_fail_links(req);
3988 __io_req_complete(req, ret, 0, cs);
3992 static int io_epoll_ctl_prep(struct io_kiocb *req,
3993 const struct io_uring_sqe *sqe)
3995 #if defined(CONFIG_EPOLL)
3996 if (sqe->ioprio || sqe->buf_index)
3998 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4001 req->epoll.epfd = READ_ONCE(sqe->fd);
4002 req->epoll.op = READ_ONCE(sqe->len);
4003 req->epoll.fd = READ_ONCE(sqe->off);
4005 if (ep_op_has_event(req->epoll.op)) {
4006 struct epoll_event __user *ev;
4008 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4009 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4019 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4020 struct io_comp_state *cs)
4022 #if defined(CONFIG_EPOLL)
4023 struct io_epoll *ie = &req->epoll;
4026 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4027 if (force_nonblock && ret == -EAGAIN)
4031 req_set_fail_links(req);
4032 __io_req_complete(req, ret, 0, cs);
4039 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4041 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4042 if (sqe->ioprio || sqe->buf_index || sqe->off)
4044 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4047 req->madvise.addr = READ_ONCE(sqe->addr);
4048 req->madvise.len = READ_ONCE(sqe->len);
4049 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4056 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4058 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4059 struct io_madvise *ma = &req->madvise;
4065 ret = do_madvise(ma->addr, ma->len, ma->advice);
4067 req_set_fail_links(req);
4068 io_req_complete(req, ret);
4075 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4077 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4079 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4082 req->fadvise.offset = READ_ONCE(sqe->off);
4083 req->fadvise.len = READ_ONCE(sqe->len);
4084 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4088 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4090 struct io_fadvise *fa = &req->fadvise;
4093 if (force_nonblock) {
4094 switch (fa->advice) {
4095 case POSIX_FADV_NORMAL:
4096 case POSIX_FADV_RANDOM:
4097 case POSIX_FADV_SEQUENTIAL:
4104 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4106 req_set_fail_links(req);
4107 io_req_complete(req, ret);
4111 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4113 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4115 if (sqe->ioprio || sqe->buf_index)
4117 if (req->flags & REQ_F_FIXED_FILE)
4120 req->statx.dfd = READ_ONCE(sqe->fd);
4121 req->statx.mask = READ_ONCE(sqe->len);
4122 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4123 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4124 req->statx.flags = READ_ONCE(sqe->statx_flags);
4129 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4131 struct io_statx *ctx = &req->statx;
4134 if (force_nonblock) {
4135 /* only need file table for an actual valid fd */
4136 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4137 req->flags |= REQ_F_NO_FILE_TABLE;
4141 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4145 req_set_fail_links(req);
4146 io_req_complete(req, ret);
4150 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4153 * If we queue this for async, it must not be cancellable. That would
4154 * leave the 'file' in an undeterminate state, and here need to modify
4155 * io_wq_work.flags, so initialize io_wq_work firstly.
4157 io_req_init_async(req);
4158 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4160 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4162 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4163 sqe->rw_flags || sqe->buf_index)
4165 if (req->flags & REQ_F_FIXED_FILE)
4168 req->close.fd = READ_ONCE(sqe->fd);
4169 if ((req->file && req->file->f_op == &io_uring_fops))
4172 req->close.put_file = NULL;
4176 static int io_close(struct io_kiocb *req, bool force_nonblock,
4177 struct io_comp_state *cs)
4179 struct io_close *close = &req->close;
4182 /* might be already done during nonblock submission */
4183 if (!close->put_file) {
4184 ret = __close_fd_get_file(close->fd, &close->put_file);
4186 return (ret == -ENOENT) ? -EBADF : ret;
4189 /* if the file has a flush method, be safe and punt to async */
4190 if (close->put_file->f_op->flush && force_nonblock) {
4191 /* was never set, but play safe */
4192 req->flags &= ~REQ_F_NOWAIT;
4193 /* avoid grabbing files - we don't need the files */
4194 req->flags |= REQ_F_NO_FILE_TABLE;
4198 /* No ->flush() or already async, safely close from here */
4199 ret = filp_close(close->put_file, req->work.identity->files);
4201 req_set_fail_links(req);
4202 fput(close->put_file);
4203 close->put_file = NULL;
4204 __io_req_complete(req, ret, 0, cs);
4208 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4210 struct io_ring_ctx *ctx = req->ctx;
4215 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4217 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4220 req->sync.off = READ_ONCE(sqe->off);
4221 req->sync.len = READ_ONCE(sqe->len);
4222 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4226 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4230 /* sync_file_range always requires a blocking context */
4234 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4237 req_set_fail_links(req);
4238 io_req_complete(req, ret);
4242 #if defined(CONFIG_NET)
4243 static int io_setup_async_msg(struct io_kiocb *req,
4244 struct io_async_msghdr *kmsg)
4246 struct io_async_msghdr *async_msg = req->async_data;
4250 if (io_alloc_async_data(req)) {
4251 if (kmsg->iov != kmsg->fast_iov)
4255 async_msg = req->async_data;
4256 req->flags |= REQ_F_NEED_CLEANUP;
4257 memcpy(async_msg, kmsg, sizeof(*kmsg));
4261 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4262 struct io_async_msghdr *iomsg)
4264 iomsg->iov = iomsg->fast_iov;
4265 iomsg->msg.msg_name = &iomsg->addr;
4266 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4267 req->sr_msg.msg_flags, &iomsg->iov);
4270 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4272 struct io_async_msghdr *async_msg = req->async_data;
4273 struct io_sr_msg *sr = &req->sr_msg;
4276 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4279 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4280 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4281 sr->len = READ_ONCE(sqe->len);
4283 #ifdef CONFIG_COMPAT
4284 if (req->ctx->compat)
4285 sr->msg_flags |= MSG_CMSG_COMPAT;
4288 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4290 ret = io_sendmsg_copy_hdr(req, async_msg);
4292 req->flags |= REQ_F_NEED_CLEANUP;
4296 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4297 struct io_comp_state *cs)
4299 struct io_async_msghdr iomsg, *kmsg;
4300 struct socket *sock;
4304 sock = sock_from_file(req->file, &ret);
4305 if (unlikely(!sock))
4308 if (req->async_data) {
4309 kmsg = req->async_data;
4310 kmsg->msg.msg_name = &kmsg->addr;
4311 /* if iov is set, it's allocated already */
4313 kmsg->iov = kmsg->fast_iov;
4314 kmsg->msg.msg_iter.iov = kmsg->iov;
4316 ret = io_sendmsg_copy_hdr(req, &iomsg);
4322 flags = req->sr_msg.msg_flags;
4323 if (flags & MSG_DONTWAIT)
4324 req->flags |= REQ_F_NOWAIT;
4325 else if (force_nonblock)
4326 flags |= MSG_DONTWAIT;
4328 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4329 if (force_nonblock && ret == -EAGAIN)
4330 return io_setup_async_msg(req, kmsg);
4331 if (ret == -ERESTARTSYS)
4334 if (kmsg->iov != kmsg->fast_iov)
4336 req->flags &= ~REQ_F_NEED_CLEANUP;
4338 req_set_fail_links(req);
4339 __io_req_complete(req, ret, 0, cs);
4343 static int io_send(struct io_kiocb *req, bool force_nonblock,
4344 struct io_comp_state *cs)
4346 struct io_sr_msg *sr = &req->sr_msg;
4349 struct socket *sock;
4353 sock = sock_from_file(req->file, &ret);
4354 if (unlikely(!sock))
4357 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4361 msg.msg_name = NULL;
4362 msg.msg_control = NULL;
4363 msg.msg_controllen = 0;
4364 msg.msg_namelen = 0;
4366 flags = req->sr_msg.msg_flags;
4367 if (flags & MSG_DONTWAIT)
4368 req->flags |= REQ_F_NOWAIT;
4369 else if (force_nonblock)
4370 flags |= MSG_DONTWAIT;
4372 msg.msg_flags = flags;
4373 ret = sock_sendmsg(sock, &msg);
4374 if (force_nonblock && ret == -EAGAIN)
4376 if (ret == -ERESTARTSYS)
4380 req_set_fail_links(req);
4381 __io_req_complete(req, ret, 0, cs);
4385 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4386 struct io_async_msghdr *iomsg)
4388 struct io_sr_msg *sr = &req->sr_msg;
4389 struct iovec __user *uiov;
4393 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4394 &iomsg->uaddr, &uiov, &iov_len);
4398 if (req->flags & REQ_F_BUFFER_SELECT) {
4401 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4403 sr->len = iomsg->iov[0].iov_len;
4404 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4408 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4409 &iomsg->iov, &iomsg->msg.msg_iter,
4418 #ifdef CONFIG_COMPAT
4419 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4420 struct io_async_msghdr *iomsg)
4422 struct compat_msghdr __user *msg_compat;
4423 struct io_sr_msg *sr = &req->sr_msg;
4424 struct compat_iovec __user *uiov;
4429 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4430 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4435 uiov = compat_ptr(ptr);
4436 if (req->flags & REQ_F_BUFFER_SELECT) {
4437 compat_ssize_t clen;
4441 if (!access_ok(uiov, sizeof(*uiov)))
4443 if (__get_user(clen, &uiov->iov_len))
4447 sr->len = iomsg->iov[0].iov_len;
4450 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4451 UIO_FASTIOV, &iomsg->iov,
4452 &iomsg->msg.msg_iter, true);
4461 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4462 struct io_async_msghdr *iomsg)
4464 iomsg->msg.msg_name = &iomsg->addr;
4465 iomsg->iov = iomsg->fast_iov;
4467 #ifdef CONFIG_COMPAT
4468 if (req->ctx->compat)
4469 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4472 return __io_recvmsg_copy_hdr(req, iomsg);
4475 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4478 struct io_sr_msg *sr = &req->sr_msg;
4479 struct io_buffer *kbuf;
4481 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4486 req->flags |= REQ_F_BUFFER_SELECTED;
4490 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4492 return io_put_kbuf(req, req->sr_msg.kbuf);
4495 static int io_recvmsg_prep(struct io_kiocb *req,
4496 const struct io_uring_sqe *sqe)
4498 struct io_async_msghdr *async_msg = req->async_data;
4499 struct io_sr_msg *sr = &req->sr_msg;
4502 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4505 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4506 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4507 sr->len = READ_ONCE(sqe->len);
4508 sr->bgid = READ_ONCE(sqe->buf_group);
4510 #ifdef CONFIG_COMPAT
4511 if (req->ctx->compat)
4512 sr->msg_flags |= MSG_CMSG_COMPAT;
4515 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4517 ret = io_recvmsg_copy_hdr(req, async_msg);
4519 req->flags |= REQ_F_NEED_CLEANUP;
4523 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4524 struct io_comp_state *cs)
4526 struct io_async_msghdr iomsg, *kmsg;
4527 struct socket *sock;
4528 struct io_buffer *kbuf;
4530 int ret, cflags = 0;
4532 sock = sock_from_file(req->file, &ret);
4533 if (unlikely(!sock))
4536 if (req->async_data) {
4537 kmsg = req->async_data;
4538 kmsg->msg.msg_name = &kmsg->addr;
4539 /* if iov is set, it's allocated already */
4541 kmsg->iov = kmsg->fast_iov;
4542 kmsg->msg.msg_iter.iov = kmsg->iov;
4544 ret = io_recvmsg_copy_hdr(req, &iomsg);
4550 if (req->flags & REQ_F_BUFFER_SELECT) {
4551 kbuf = io_recv_buffer_select(req, !force_nonblock);
4553 return PTR_ERR(kbuf);
4554 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4555 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4556 1, req->sr_msg.len);
4559 flags = req->sr_msg.msg_flags;
4560 if (flags & MSG_DONTWAIT)
4561 req->flags |= REQ_F_NOWAIT;
4562 else if (force_nonblock)
4563 flags |= MSG_DONTWAIT;
4565 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4566 kmsg->uaddr, flags);
4567 if (force_nonblock && ret == -EAGAIN)
4568 return io_setup_async_msg(req, kmsg);
4569 if (ret == -ERESTARTSYS)
4572 if (req->flags & REQ_F_BUFFER_SELECTED)
4573 cflags = io_put_recv_kbuf(req);
4574 if (kmsg->iov != kmsg->fast_iov)
4576 req->flags &= ~REQ_F_NEED_CLEANUP;
4578 req_set_fail_links(req);
4579 __io_req_complete(req, ret, cflags, cs);
4583 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4584 struct io_comp_state *cs)
4586 struct io_buffer *kbuf;
4587 struct io_sr_msg *sr = &req->sr_msg;
4589 void __user *buf = sr->buf;
4590 struct socket *sock;
4593 int ret, cflags = 0;
4595 sock = sock_from_file(req->file, &ret);
4596 if (unlikely(!sock))
4599 if (req->flags & REQ_F_BUFFER_SELECT) {
4600 kbuf = io_recv_buffer_select(req, !force_nonblock);
4602 return PTR_ERR(kbuf);
4603 buf = u64_to_user_ptr(kbuf->addr);
4606 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4610 msg.msg_name = NULL;
4611 msg.msg_control = NULL;
4612 msg.msg_controllen = 0;
4613 msg.msg_namelen = 0;
4614 msg.msg_iocb = NULL;
4617 flags = req->sr_msg.msg_flags;
4618 if (flags & MSG_DONTWAIT)
4619 req->flags |= REQ_F_NOWAIT;
4620 else if (force_nonblock)
4621 flags |= MSG_DONTWAIT;
4623 ret = sock_recvmsg(sock, &msg, flags);
4624 if (force_nonblock && ret == -EAGAIN)
4626 if (ret == -ERESTARTSYS)
4629 if (req->flags & REQ_F_BUFFER_SELECTED)
4630 cflags = io_put_recv_kbuf(req);
4632 req_set_fail_links(req);
4633 __io_req_complete(req, ret, cflags, cs);
4637 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4639 struct io_accept *accept = &req->accept;
4641 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4643 if (sqe->ioprio || sqe->len || sqe->buf_index)
4646 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4647 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4648 accept->flags = READ_ONCE(sqe->accept_flags);
4649 accept->nofile = rlimit(RLIMIT_NOFILE);
4653 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4654 struct io_comp_state *cs)
4656 struct io_accept *accept = &req->accept;
4657 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4660 if (req->file->f_flags & O_NONBLOCK)
4661 req->flags |= REQ_F_NOWAIT;
4663 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4664 accept->addr_len, accept->flags,
4666 if (ret == -EAGAIN && force_nonblock)
4669 if (ret == -ERESTARTSYS)
4671 req_set_fail_links(req);
4673 __io_req_complete(req, ret, 0, cs);
4677 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4679 struct io_connect *conn = &req->connect;
4680 struct io_async_connect *io = req->async_data;
4682 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4684 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4687 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4688 conn->addr_len = READ_ONCE(sqe->addr2);
4693 return move_addr_to_kernel(conn->addr, conn->addr_len,
4697 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4698 struct io_comp_state *cs)
4700 struct io_async_connect __io, *io;
4701 unsigned file_flags;
4704 if (req->async_data) {
4705 io = req->async_data;
4707 ret = move_addr_to_kernel(req->connect.addr,
4708 req->connect.addr_len,
4715 file_flags = force_nonblock ? O_NONBLOCK : 0;
4717 ret = __sys_connect_file(req->file, &io->address,
4718 req->connect.addr_len, file_flags);
4719 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4720 if (req->async_data)
4722 if (io_alloc_async_data(req)) {
4726 io = req->async_data;
4727 memcpy(req->async_data, &__io, sizeof(__io));
4730 if (ret == -ERESTARTSYS)
4734 req_set_fail_links(req);
4735 __io_req_complete(req, ret, 0, cs);
4738 #else /* !CONFIG_NET */
4739 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4744 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4745 struct io_comp_state *cs)
4750 static int io_send(struct io_kiocb *req, bool force_nonblock,
4751 struct io_comp_state *cs)
4756 static int io_recvmsg_prep(struct io_kiocb *req,
4757 const struct io_uring_sqe *sqe)
4762 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4763 struct io_comp_state *cs)
4768 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4769 struct io_comp_state *cs)
4774 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4779 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4780 struct io_comp_state *cs)
4785 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4790 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4791 struct io_comp_state *cs)
4795 #endif /* CONFIG_NET */
4797 struct io_poll_table {
4798 struct poll_table_struct pt;
4799 struct io_kiocb *req;
4803 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4804 __poll_t mask, task_work_func_t func)
4809 /* for instances that support it check for an event match first: */
4810 if (mask && !(mask & poll->events))
4813 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4815 list_del_init(&poll->wait.entry);
4818 init_task_work(&req->task_work, func);
4819 percpu_ref_get(&req->ctx->refs);
4822 * If we using the signalfd wait_queue_head for this wakeup, then
4823 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4824 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4825 * either, as the normal wakeup will suffice.
4827 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4830 * If this fails, then the task is exiting. When a task exits, the
4831 * work gets canceled, so just cancel this request as well instead
4832 * of executing it. We can't safely execute it anyway, as we may not
4833 * have the needed state needed for it anyway.
4835 ret = io_req_task_work_add(req, twa_signal_ok);
4836 if (unlikely(ret)) {
4837 struct task_struct *tsk;
4839 WRITE_ONCE(poll->canceled, true);
4840 tsk = io_wq_get_task(req->ctx->io_wq);
4841 task_work_add(tsk, &req->task_work, 0);
4842 wake_up_process(tsk);
4847 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4848 __acquires(&req->ctx->completion_lock)
4850 struct io_ring_ctx *ctx = req->ctx;
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 struct poll_table_struct pt = { ._key = poll->events };
4855 req->result = vfs_poll(req->file, &pt) & poll->events;
4858 spin_lock_irq(&ctx->completion_lock);
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 add_wait_queue(poll->head, &poll->wait);
4867 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4869 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4870 if (req->opcode == IORING_OP_POLL_ADD)
4871 return req->async_data;
4872 return req->apoll->double_poll;
4875 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4877 if (req->opcode == IORING_OP_POLL_ADD)
4879 return &req->apoll->poll;
4882 static void io_poll_remove_double(struct io_kiocb *req)
4884 struct io_poll_iocb *poll = io_poll_get_double(req);
4886 lockdep_assert_held(&req->ctx->completion_lock);
4888 if (poll && poll->head) {
4889 struct wait_queue_head *head = poll->head;
4891 spin_lock(&head->lock);
4892 list_del_init(&poll->wait.entry);
4893 if (poll->wait.private)
4894 refcount_dec(&req->refs);
4896 spin_unlock(&head->lock);
4900 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4902 struct io_ring_ctx *ctx = req->ctx;
4904 io_poll_remove_double(req);
4905 req->poll.done = true;
4906 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4907 io_commit_cqring(ctx);
4910 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4912 struct io_ring_ctx *ctx = req->ctx;
4914 if (io_poll_rewait(req, &req->poll)) {
4915 spin_unlock_irq(&ctx->completion_lock);
4919 hash_del(&req->hash_node);
4920 io_poll_complete(req, req->result, 0);
4921 spin_unlock_irq(&ctx->completion_lock);
4923 *nxt = io_put_req_find_next(req);
4924 io_cqring_ev_posted(ctx);
4927 static void io_poll_task_func(struct callback_head *cb)
4929 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4930 struct io_ring_ctx *ctx = req->ctx;
4931 struct io_kiocb *nxt = NULL;
4933 io_poll_task_handler(req, &nxt);
4935 __io_req_task_submit(nxt);
4936 percpu_ref_put(&ctx->refs);
4939 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4940 int sync, void *key)
4942 struct io_kiocb *req = wait->private;
4943 struct io_poll_iocb *poll = io_poll_get_single(req);
4944 __poll_t mask = key_to_poll(key);
4946 /* for instances that support it check for an event match first: */
4947 if (mask && !(mask & poll->events))
4950 list_del_init(&wait->entry);
4952 if (poll && poll->head) {
4955 spin_lock(&poll->head->lock);
4956 done = list_empty(&poll->wait.entry);
4958 list_del_init(&poll->wait.entry);
4959 /* make sure double remove sees this as being gone */
4960 wait->private = NULL;
4961 spin_unlock(&poll->head->lock);
4963 __io_async_wake(req, poll, mask, io_poll_task_func);
4965 refcount_dec(&req->refs);
4969 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4970 wait_queue_func_t wake_func)
4974 poll->canceled = false;
4975 poll->events = events;
4976 INIT_LIST_HEAD(&poll->wait.entry);
4977 init_waitqueue_func_entry(&poll->wait, wake_func);
4980 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4981 struct wait_queue_head *head,
4982 struct io_poll_iocb **poll_ptr)
4984 struct io_kiocb *req = pt->req;
4987 * If poll->head is already set, it's because the file being polled
4988 * uses multiple waitqueues for poll handling (eg one for read, one
4989 * for write). Setup a separate io_poll_iocb if this happens.
4991 if (unlikely(poll->head)) {
4992 /* already have a 2nd entry, fail a third attempt */
4994 pt->error = -EINVAL;
4997 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4999 pt->error = -ENOMEM;
5002 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
5003 refcount_inc(&req->refs);
5004 poll->wait.private = req;
5011 if (poll->events & EPOLLEXCLUSIVE)
5012 add_wait_queue_exclusive(head, &poll->wait);
5014 add_wait_queue(head, &poll->wait);
5017 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5018 struct poll_table_struct *p)
5020 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5021 struct async_poll *apoll = pt->req->apoll;
5023 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5026 static void io_async_task_func(struct callback_head *cb)
5028 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5029 struct async_poll *apoll = req->apoll;
5030 struct io_ring_ctx *ctx = req->ctx;
5032 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5034 if (io_poll_rewait(req, &apoll->poll)) {
5035 spin_unlock_irq(&ctx->completion_lock);
5036 percpu_ref_put(&ctx->refs);
5040 /* If req is still hashed, it cannot have been canceled. Don't check. */
5041 if (hash_hashed(&req->hash_node))
5042 hash_del(&req->hash_node);
5044 io_poll_remove_double(req);
5045 spin_unlock_irq(&ctx->completion_lock);
5047 if (!READ_ONCE(apoll->poll.canceled))
5048 __io_req_task_submit(req);
5050 __io_req_task_cancel(req, -ECANCELED);
5052 percpu_ref_put(&ctx->refs);
5053 kfree(apoll->double_poll);
5057 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5060 struct io_kiocb *req = wait->private;
5061 struct io_poll_iocb *poll = &req->apoll->poll;
5063 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5066 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5069 static void io_poll_req_insert(struct io_kiocb *req)
5071 struct io_ring_ctx *ctx = req->ctx;
5072 struct hlist_head *list;
5074 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5075 hlist_add_head(&req->hash_node, list);
5078 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5079 struct io_poll_iocb *poll,
5080 struct io_poll_table *ipt, __poll_t mask,
5081 wait_queue_func_t wake_func)
5082 __acquires(&ctx->completion_lock)
5084 struct io_ring_ctx *ctx = req->ctx;
5085 bool cancel = false;
5087 io_init_poll_iocb(poll, mask, wake_func);
5088 poll->file = req->file;
5089 poll->wait.private = req;
5091 ipt->pt._key = mask;
5093 ipt->error = -EINVAL;
5095 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5097 spin_lock_irq(&ctx->completion_lock);
5098 if (likely(poll->head)) {
5099 spin_lock(&poll->head->lock);
5100 if (unlikely(list_empty(&poll->wait.entry))) {
5106 if (mask || ipt->error)
5107 list_del_init(&poll->wait.entry);
5109 WRITE_ONCE(poll->canceled, true);
5110 else if (!poll->done) /* actually waiting for an event */
5111 io_poll_req_insert(req);
5112 spin_unlock(&poll->head->lock);
5118 static bool io_arm_poll_handler(struct io_kiocb *req)
5120 const struct io_op_def *def = &io_op_defs[req->opcode];
5121 struct io_ring_ctx *ctx = req->ctx;
5122 struct async_poll *apoll;
5123 struct io_poll_table ipt;
5127 if (!req->file || !file_can_poll(req->file))
5129 if (req->flags & REQ_F_POLLED)
5133 else if (def->pollout)
5137 /* if we can't nonblock try, then no point in arming a poll handler */
5138 if (!io_file_supports_async(req->file, rw))
5141 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5142 if (unlikely(!apoll))
5144 apoll->double_poll = NULL;
5146 req->flags |= REQ_F_POLLED;
5148 INIT_HLIST_NODE(&req->hash_node);
5152 mask |= POLLIN | POLLRDNORM;
5154 mask |= POLLOUT | POLLWRNORM;
5156 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5157 if ((req->opcode == IORING_OP_RECVMSG) &&
5158 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5161 mask |= POLLERR | POLLPRI;
5163 ipt.pt._qproc = io_async_queue_proc;
5165 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5167 if (ret || ipt.error) {
5168 io_poll_remove_double(req);
5169 spin_unlock_irq(&ctx->completion_lock);
5170 kfree(apoll->double_poll);
5174 spin_unlock_irq(&ctx->completion_lock);
5175 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5176 apoll->poll.events);
5180 static bool __io_poll_remove_one(struct io_kiocb *req,
5181 struct io_poll_iocb *poll)
5183 bool do_complete = false;
5185 spin_lock(&poll->head->lock);
5186 WRITE_ONCE(poll->canceled, true);
5187 if (!list_empty(&poll->wait.entry)) {
5188 list_del_init(&poll->wait.entry);
5191 spin_unlock(&poll->head->lock);
5192 hash_del(&req->hash_node);
5196 static bool io_poll_remove_one(struct io_kiocb *req)
5200 io_poll_remove_double(req);
5202 if (req->opcode == IORING_OP_POLL_ADD) {
5203 do_complete = __io_poll_remove_one(req, &req->poll);
5205 struct async_poll *apoll = req->apoll;
5207 /* non-poll requests have submit ref still */
5208 do_complete = __io_poll_remove_one(req, &apoll->poll);
5211 kfree(apoll->double_poll);
5217 io_cqring_fill_event(req, -ECANCELED);
5218 io_commit_cqring(req->ctx);
5219 req_set_fail_links(req);
5220 io_put_req_deferred(req, 1);
5227 * Returns true if we found and killed one or more poll requests
5229 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5231 struct hlist_node *tmp;
5232 struct io_kiocb *req;
5235 spin_lock_irq(&ctx->completion_lock);
5236 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5237 struct hlist_head *list;
5239 list = &ctx->cancel_hash[i];
5240 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5241 if (io_task_match(req, tsk))
5242 posted += io_poll_remove_one(req);
5245 spin_unlock_irq(&ctx->completion_lock);
5248 io_cqring_ev_posted(ctx);
5253 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5255 struct hlist_head *list;
5256 struct io_kiocb *req;
5258 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5259 hlist_for_each_entry(req, list, hash_node) {
5260 if (sqe_addr != req->user_data)
5262 if (io_poll_remove_one(req))
5270 static int io_poll_remove_prep(struct io_kiocb *req,
5271 const struct io_uring_sqe *sqe)
5273 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5275 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5279 req->poll.addr = READ_ONCE(sqe->addr);
5284 * Find a running poll command that matches one specified in sqe->addr,
5285 * and remove it if found.
5287 static int io_poll_remove(struct io_kiocb *req)
5289 struct io_ring_ctx *ctx = req->ctx;
5293 addr = req->poll.addr;
5294 spin_lock_irq(&ctx->completion_lock);
5295 ret = io_poll_cancel(ctx, addr);
5296 spin_unlock_irq(&ctx->completion_lock);
5299 req_set_fail_links(req);
5300 io_req_complete(req, ret);
5304 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5307 struct io_kiocb *req = wait->private;
5308 struct io_poll_iocb *poll = &req->poll;
5310 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5313 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5314 struct poll_table_struct *p)
5316 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5318 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5321 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5323 struct io_poll_iocb *poll = &req->poll;
5326 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5328 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5333 events = READ_ONCE(sqe->poll32_events);
5335 events = swahw32(events);
5337 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5338 (events & EPOLLEXCLUSIVE);
5342 static int io_poll_add(struct io_kiocb *req)
5344 struct io_poll_iocb *poll = &req->poll;
5345 struct io_ring_ctx *ctx = req->ctx;
5346 struct io_poll_table ipt;
5349 INIT_HLIST_NODE(&req->hash_node);
5350 ipt.pt._qproc = io_poll_queue_proc;
5352 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5355 if (mask) { /* no async, we'd stolen it */
5357 io_poll_complete(req, mask, 0);
5359 spin_unlock_irq(&ctx->completion_lock);
5362 io_cqring_ev_posted(ctx);
5368 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5370 struct io_timeout_data *data = container_of(timer,
5371 struct io_timeout_data, timer);
5372 struct io_kiocb *req = data->req;
5373 struct io_ring_ctx *ctx = req->ctx;
5374 unsigned long flags;
5376 spin_lock_irqsave(&ctx->completion_lock, flags);
5377 list_del_init(&req->timeout.list);
5378 atomic_set(&req->ctx->cq_timeouts,
5379 atomic_read(&req->ctx->cq_timeouts) + 1);
5381 io_cqring_fill_event(req, -ETIME);
5382 io_commit_cqring(ctx);
5383 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5385 io_cqring_ev_posted(ctx);
5386 req_set_fail_links(req);
5388 return HRTIMER_NORESTART;
5391 static int __io_timeout_cancel(struct io_kiocb *req)
5393 struct io_timeout_data *io = req->async_data;
5396 ret = hrtimer_try_to_cancel(&io->timer);
5399 list_del_init(&req->timeout.list);
5401 req_set_fail_links(req);
5402 io_cqring_fill_event(req, -ECANCELED);
5403 io_put_req_deferred(req, 1);
5407 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5409 struct io_kiocb *req;
5412 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5413 if (user_data == req->user_data) {
5422 return __io_timeout_cancel(req);
5425 static int io_timeout_remove_prep(struct io_kiocb *req,
5426 const struct io_uring_sqe *sqe)
5428 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5430 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5432 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5435 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5440 * Remove or update an existing timeout command
5442 static int io_timeout_remove(struct io_kiocb *req)
5444 struct io_ring_ctx *ctx = req->ctx;
5447 spin_lock_irq(&ctx->completion_lock);
5448 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5450 io_cqring_fill_event(req, ret);
5451 io_commit_cqring(ctx);
5452 spin_unlock_irq(&ctx->completion_lock);
5453 io_cqring_ev_posted(ctx);
5455 req_set_fail_links(req);
5460 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5461 bool is_timeout_link)
5463 struct io_timeout_data *data;
5465 u32 off = READ_ONCE(sqe->off);
5467 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5469 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5471 if (off && is_timeout_link)
5473 flags = READ_ONCE(sqe->timeout_flags);
5474 if (flags & ~IORING_TIMEOUT_ABS)
5477 req->timeout.off = off;
5479 if (!req->async_data && io_alloc_async_data(req))
5482 data = req->async_data;
5485 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5488 if (flags & IORING_TIMEOUT_ABS)
5489 data->mode = HRTIMER_MODE_ABS;
5491 data->mode = HRTIMER_MODE_REL;
5493 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5497 static int io_timeout(struct io_kiocb *req)
5499 struct io_ring_ctx *ctx = req->ctx;
5500 struct io_timeout_data *data = req->async_data;
5501 struct list_head *entry;
5502 u32 tail, off = req->timeout.off;
5504 spin_lock_irq(&ctx->completion_lock);
5507 * sqe->off holds how many events that need to occur for this
5508 * timeout event to be satisfied. If it isn't set, then this is
5509 * a pure timeout request, sequence isn't used.
5511 if (io_is_timeout_noseq(req)) {
5512 entry = ctx->timeout_list.prev;
5516 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5517 req->timeout.target_seq = tail + off;
5520 * Insertion sort, ensuring the first entry in the list is always
5521 * the one we need first.
5523 list_for_each_prev(entry, &ctx->timeout_list) {
5524 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5527 if (io_is_timeout_noseq(nxt))
5529 /* nxt.seq is behind @tail, otherwise would've been completed */
5530 if (off >= nxt->timeout.target_seq - tail)
5534 list_add(&req->timeout.list, entry);
5535 data->timer.function = io_timeout_fn;
5536 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5537 spin_unlock_irq(&ctx->completion_lock);
5541 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5543 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5545 return req->user_data == (unsigned long) data;
5548 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5550 enum io_wq_cancel cancel_ret;
5553 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5554 switch (cancel_ret) {
5555 case IO_WQ_CANCEL_OK:
5558 case IO_WQ_CANCEL_RUNNING:
5561 case IO_WQ_CANCEL_NOTFOUND:
5569 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5570 struct io_kiocb *req, __u64 sqe_addr,
5573 unsigned long flags;
5576 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5577 if (ret != -ENOENT) {
5578 spin_lock_irqsave(&ctx->completion_lock, flags);
5582 spin_lock_irqsave(&ctx->completion_lock, flags);
5583 ret = io_timeout_cancel(ctx, sqe_addr);
5586 ret = io_poll_cancel(ctx, sqe_addr);
5590 io_cqring_fill_event(req, ret);
5591 io_commit_cqring(ctx);
5592 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5593 io_cqring_ev_posted(ctx);
5596 req_set_fail_links(req);
5600 static int io_async_cancel_prep(struct io_kiocb *req,
5601 const struct io_uring_sqe *sqe)
5603 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5605 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5607 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5610 req->cancel.addr = READ_ONCE(sqe->addr);
5614 static int io_async_cancel(struct io_kiocb *req)
5616 struct io_ring_ctx *ctx = req->ctx;
5618 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5622 static int io_files_update_prep(struct io_kiocb *req,
5623 const struct io_uring_sqe *sqe)
5625 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5627 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5629 if (sqe->ioprio || sqe->rw_flags)
5632 req->files_update.offset = READ_ONCE(sqe->off);
5633 req->files_update.nr_args = READ_ONCE(sqe->len);
5634 if (!req->files_update.nr_args)
5636 req->files_update.arg = READ_ONCE(sqe->addr);
5640 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5641 struct io_comp_state *cs)
5643 struct io_ring_ctx *ctx = req->ctx;
5644 struct io_uring_files_update up;
5650 up.offset = req->files_update.offset;
5651 up.fds = req->files_update.arg;
5653 mutex_lock(&ctx->uring_lock);
5654 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5655 mutex_unlock(&ctx->uring_lock);
5658 req_set_fail_links(req);
5659 __io_req_complete(req, ret, 0, cs);
5663 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5665 switch (req->opcode) {
5668 case IORING_OP_READV:
5669 case IORING_OP_READ_FIXED:
5670 case IORING_OP_READ:
5671 return io_read_prep(req, sqe);
5672 case IORING_OP_WRITEV:
5673 case IORING_OP_WRITE_FIXED:
5674 case IORING_OP_WRITE:
5675 return io_write_prep(req, sqe);
5676 case IORING_OP_POLL_ADD:
5677 return io_poll_add_prep(req, sqe);
5678 case IORING_OP_POLL_REMOVE:
5679 return io_poll_remove_prep(req, sqe);
5680 case IORING_OP_FSYNC:
5681 return io_prep_fsync(req, sqe);
5682 case IORING_OP_SYNC_FILE_RANGE:
5683 return io_prep_sfr(req, sqe);
5684 case IORING_OP_SENDMSG:
5685 case IORING_OP_SEND:
5686 return io_sendmsg_prep(req, sqe);
5687 case IORING_OP_RECVMSG:
5688 case IORING_OP_RECV:
5689 return io_recvmsg_prep(req, sqe);
5690 case IORING_OP_CONNECT:
5691 return io_connect_prep(req, sqe);
5692 case IORING_OP_TIMEOUT:
5693 return io_timeout_prep(req, sqe, false);
5694 case IORING_OP_TIMEOUT_REMOVE:
5695 return io_timeout_remove_prep(req, sqe);
5696 case IORING_OP_ASYNC_CANCEL:
5697 return io_async_cancel_prep(req, sqe);
5698 case IORING_OP_LINK_TIMEOUT:
5699 return io_timeout_prep(req, sqe, true);
5700 case IORING_OP_ACCEPT:
5701 return io_accept_prep(req, sqe);
5702 case IORING_OP_FALLOCATE:
5703 return io_fallocate_prep(req, sqe);
5704 case IORING_OP_OPENAT:
5705 return io_openat_prep(req, sqe);
5706 case IORING_OP_CLOSE:
5707 return io_close_prep(req, sqe);
5708 case IORING_OP_FILES_UPDATE:
5709 return io_files_update_prep(req, sqe);
5710 case IORING_OP_STATX:
5711 return io_statx_prep(req, sqe);
5712 case IORING_OP_FADVISE:
5713 return io_fadvise_prep(req, sqe);
5714 case IORING_OP_MADVISE:
5715 return io_madvise_prep(req, sqe);
5716 case IORING_OP_OPENAT2:
5717 return io_openat2_prep(req, sqe);
5718 case IORING_OP_EPOLL_CTL:
5719 return io_epoll_ctl_prep(req, sqe);
5720 case IORING_OP_SPLICE:
5721 return io_splice_prep(req, sqe);
5722 case IORING_OP_PROVIDE_BUFFERS:
5723 return io_provide_buffers_prep(req, sqe);
5724 case IORING_OP_REMOVE_BUFFERS:
5725 return io_remove_buffers_prep(req, sqe);
5727 return io_tee_prep(req, sqe);
5730 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5735 static int io_req_defer_prep(struct io_kiocb *req,
5736 const struct io_uring_sqe *sqe)
5740 if (io_alloc_async_data(req))
5742 return io_req_prep(req, sqe);
5745 static u32 io_get_sequence(struct io_kiocb *req)
5747 struct io_kiocb *pos;
5748 struct io_ring_ctx *ctx = req->ctx;
5749 u32 total_submitted, nr_reqs = 1;
5751 if (req->flags & REQ_F_LINK_HEAD)
5752 list_for_each_entry(pos, &req->link_list, link_list)
5755 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5756 return total_submitted - nr_reqs;
5759 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5761 struct io_ring_ctx *ctx = req->ctx;
5762 struct io_defer_entry *de;
5766 /* Still need defer if there is pending req in defer list. */
5767 if (likely(list_empty_careful(&ctx->defer_list) &&
5768 !(req->flags & REQ_F_IO_DRAIN)))
5771 seq = io_get_sequence(req);
5772 /* Still a chance to pass the sequence check */
5773 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5776 if (!req->async_data) {
5777 ret = io_req_defer_prep(req, sqe);
5781 io_prep_async_link(req);
5782 de = kmalloc(sizeof(*de), GFP_KERNEL);
5786 spin_lock_irq(&ctx->completion_lock);
5787 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5788 spin_unlock_irq(&ctx->completion_lock);
5790 io_queue_async_work(req);
5791 return -EIOCBQUEUED;
5794 trace_io_uring_defer(ctx, req, req->user_data);
5797 list_add_tail(&de->list, &ctx->defer_list);
5798 spin_unlock_irq(&ctx->completion_lock);
5799 return -EIOCBQUEUED;
5802 static void io_req_drop_files(struct io_kiocb *req)
5804 struct io_ring_ctx *ctx = req->ctx;
5805 unsigned long flags;
5807 spin_lock_irqsave(&ctx->inflight_lock, flags);
5808 list_del(&req->inflight_entry);
5809 if (waitqueue_active(&ctx->inflight_wait))
5810 wake_up(&ctx->inflight_wait);
5811 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5812 req->flags &= ~REQ_F_INFLIGHT;
5813 put_files_struct(req->work.identity->files);
5814 put_nsproxy(req->work.identity->nsproxy);
5815 req->work.flags &= ~IO_WQ_WORK_FILES;
5818 static void __io_clean_op(struct io_kiocb *req)
5820 if (req->flags & REQ_F_BUFFER_SELECTED) {
5821 switch (req->opcode) {
5822 case IORING_OP_READV:
5823 case IORING_OP_READ_FIXED:
5824 case IORING_OP_READ:
5825 kfree((void *)(unsigned long)req->rw.addr);
5827 case IORING_OP_RECVMSG:
5828 case IORING_OP_RECV:
5829 kfree(req->sr_msg.kbuf);
5832 req->flags &= ~REQ_F_BUFFER_SELECTED;
5835 if (req->flags & REQ_F_NEED_CLEANUP) {
5836 switch (req->opcode) {
5837 case IORING_OP_READV:
5838 case IORING_OP_READ_FIXED:
5839 case IORING_OP_READ:
5840 case IORING_OP_WRITEV:
5841 case IORING_OP_WRITE_FIXED:
5842 case IORING_OP_WRITE: {
5843 struct io_async_rw *io = req->async_data;
5845 kfree(io->free_iovec);
5848 case IORING_OP_RECVMSG:
5849 case IORING_OP_SENDMSG: {
5850 struct io_async_msghdr *io = req->async_data;
5851 if (io->iov != io->fast_iov)
5855 case IORING_OP_SPLICE:
5857 io_put_file(req, req->splice.file_in,
5858 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5860 case IORING_OP_OPENAT:
5861 case IORING_OP_OPENAT2:
5862 if (req->open.filename)
5863 putname(req->open.filename);
5866 req->flags &= ~REQ_F_NEED_CLEANUP;
5869 if (req->flags & REQ_F_INFLIGHT)
5870 io_req_drop_files(req);
5873 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5874 struct io_comp_state *cs)
5876 struct io_ring_ctx *ctx = req->ctx;
5879 switch (req->opcode) {
5881 ret = io_nop(req, cs);
5883 case IORING_OP_READV:
5884 case IORING_OP_READ_FIXED:
5885 case IORING_OP_READ:
5886 ret = io_read(req, force_nonblock, cs);
5888 case IORING_OP_WRITEV:
5889 case IORING_OP_WRITE_FIXED:
5890 case IORING_OP_WRITE:
5891 ret = io_write(req, force_nonblock, cs);
5893 case IORING_OP_FSYNC:
5894 ret = io_fsync(req, force_nonblock);
5896 case IORING_OP_POLL_ADD:
5897 ret = io_poll_add(req);
5899 case IORING_OP_POLL_REMOVE:
5900 ret = io_poll_remove(req);
5902 case IORING_OP_SYNC_FILE_RANGE:
5903 ret = io_sync_file_range(req, force_nonblock);
5905 case IORING_OP_SENDMSG:
5906 ret = io_sendmsg(req, force_nonblock, cs);
5908 case IORING_OP_SEND:
5909 ret = io_send(req, force_nonblock, cs);
5911 case IORING_OP_RECVMSG:
5912 ret = io_recvmsg(req, force_nonblock, cs);
5914 case IORING_OP_RECV:
5915 ret = io_recv(req, force_nonblock, cs);
5917 case IORING_OP_TIMEOUT:
5918 ret = io_timeout(req);
5920 case IORING_OP_TIMEOUT_REMOVE:
5921 ret = io_timeout_remove(req);
5923 case IORING_OP_ACCEPT:
5924 ret = io_accept(req, force_nonblock, cs);
5926 case IORING_OP_CONNECT:
5927 ret = io_connect(req, force_nonblock, cs);
5929 case IORING_OP_ASYNC_CANCEL:
5930 ret = io_async_cancel(req);
5932 case IORING_OP_FALLOCATE:
5933 ret = io_fallocate(req, force_nonblock);
5935 case IORING_OP_OPENAT:
5936 ret = io_openat(req, force_nonblock);
5938 case IORING_OP_CLOSE:
5939 ret = io_close(req, force_nonblock, cs);
5941 case IORING_OP_FILES_UPDATE:
5942 ret = io_files_update(req, force_nonblock, cs);
5944 case IORING_OP_STATX:
5945 ret = io_statx(req, force_nonblock);
5947 case IORING_OP_FADVISE:
5948 ret = io_fadvise(req, force_nonblock);
5950 case IORING_OP_MADVISE:
5951 ret = io_madvise(req, force_nonblock);
5953 case IORING_OP_OPENAT2:
5954 ret = io_openat2(req, force_nonblock);
5956 case IORING_OP_EPOLL_CTL:
5957 ret = io_epoll_ctl(req, force_nonblock, cs);
5959 case IORING_OP_SPLICE:
5960 ret = io_splice(req, force_nonblock);
5962 case IORING_OP_PROVIDE_BUFFERS:
5963 ret = io_provide_buffers(req, force_nonblock, cs);
5965 case IORING_OP_REMOVE_BUFFERS:
5966 ret = io_remove_buffers(req, force_nonblock, cs);
5969 ret = io_tee(req, force_nonblock);
5979 /* If the op doesn't have a file, we're not polling for it */
5980 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5981 const bool in_async = io_wq_current_is_worker();
5983 /* workqueue context doesn't hold uring_lock, grab it now */
5985 mutex_lock(&ctx->uring_lock);
5987 io_iopoll_req_issued(req);
5990 mutex_unlock(&ctx->uring_lock);
5996 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5998 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5999 struct io_kiocb *timeout;
6002 timeout = io_prep_linked_timeout(req);
6004 io_queue_linked_timeout(timeout);
6006 /* if NO_CANCEL is set, we must still run the work */
6007 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6008 IO_WQ_WORK_CANCEL) {
6014 ret = io_issue_sqe(req, false, NULL);
6016 * We can get EAGAIN for polled IO even though we're
6017 * forcing a sync submission from here, since we can't
6018 * wait for request slots on the block side.
6027 req_set_fail_links(req);
6028 io_req_complete(req, ret);
6031 return io_steal_work(req);
6034 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6037 struct fixed_file_table *table;
6039 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6040 return table->files[index & IORING_FILE_TABLE_MASK];
6043 static struct file *io_file_get(struct io_submit_state *state,
6044 struct io_kiocb *req, int fd, bool fixed)
6046 struct io_ring_ctx *ctx = req->ctx;
6050 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6052 fd = array_index_nospec(fd, ctx->nr_user_files);
6053 file = io_file_from_index(ctx, fd);
6055 req->fixed_file_refs = &ctx->file_data->node->refs;
6056 percpu_ref_get(req->fixed_file_refs);
6059 trace_io_uring_file_get(ctx, fd);
6060 file = __io_file_get(state, fd);
6066 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6071 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6072 if (unlikely(!fixed && io_async_submit(req->ctx)))
6075 req->file = io_file_get(state, req, fd, fixed);
6076 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6081 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6083 struct io_timeout_data *data = container_of(timer,
6084 struct io_timeout_data, timer);
6085 struct io_kiocb *req = data->req;
6086 struct io_ring_ctx *ctx = req->ctx;
6087 struct io_kiocb *prev = NULL;
6088 unsigned long flags;
6090 spin_lock_irqsave(&ctx->completion_lock, flags);
6093 * We don't expect the list to be empty, that will only happen if we
6094 * race with the completion of the linked work.
6096 if (!list_empty(&req->link_list)) {
6097 prev = list_entry(req->link_list.prev, struct io_kiocb,
6099 if (refcount_inc_not_zero(&prev->refs)) {
6100 list_del_init(&req->link_list);
6101 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6106 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6109 req_set_fail_links(prev);
6110 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6113 io_req_complete(req, -ETIME);
6115 return HRTIMER_NORESTART;
6118 static void __io_queue_linked_timeout(struct io_kiocb *req)
6121 * If the list is now empty, then our linked request finished before
6122 * we got a chance to setup the timer
6124 if (!list_empty(&req->link_list)) {
6125 struct io_timeout_data *data = req->async_data;
6127 data->timer.function = io_link_timeout_fn;
6128 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6133 static void io_queue_linked_timeout(struct io_kiocb *req)
6135 struct io_ring_ctx *ctx = req->ctx;
6137 spin_lock_irq(&ctx->completion_lock);
6138 __io_queue_linked_timeout(req);
6139 spin_unlock_irq(&ctx->completion_lock);
6141 /* drop submission reference */
6145 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6147 struct io_kiocb *nxt;
6149 if (!(req->flags & REQ_F_LINK_HEAD))
6151 if (req->flags & REQ_F_LINK_TIMEOUT)
6154 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6156 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6159 req->flags |= REQ_F_LINK_TIMEOUT;
6163 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6165 struct io_kiocb *linked_timeout;
6166 struct io_kiocb *nxt;
6167 const struct cred *old_creds = NULL;
6171 linked_timeout = io_prep_linked_timeout(req);
6173 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.identity->creds &&
6174 req->work.identity->creds != current_cred()) {
6176 revert_creds(old_creds);
6177 if (old_creds == req->work.identity->creds)
6178 old_creds = NULL; /* restored original creds */
6180 old_creds = override_creds(req->work.identity->creds);
6181 req->work.flags |= IO_WQ_WORK_CREDS;
6184 ret = io_issue_sqe(req, true, cs);
6187 * We async punt it if the file wasn't marked NOWAIT, or if the file
6188 * doesn't support non-blocking read/write attempts
6190 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6191 if (!io_arm_poll_handler(req)) {
6194 * Queued up for async execution, worker will release
6195 * submit reference when the iocb is actually submitted.
6197 io_queue_async_work(req);
6201 io_queue_linked_timeout(linked_timeout);
6205 if (unlikely(ret)) {
6206 /* un-prep timeout, so it'll be killed as any other linked */
6207 req->flags &= ~REQ_F_LINK_TIMEOUT;
6208 req_set_fail_links(req);
6210 io_req_complete(req, ret);
6214 /* drop submission reference */
6215 nxt = io_put_req_find_next(req);
6217 io_queue_linked_timeout(linked_timeout);
6222 if (req->flags & REQ_F_FORCE_ASYNC)
6228 revert_creds(old_creds);
6231 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6232 struct io_comp_state *cs)
6236 ret = io_req_defer(req, sqe);
6238 if (ret != -EIOCBQUEUED) {
6240 req_set_fail_links(req);
6242 io_req_complete(req, ret);
6244 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6245 if (!req->async_data) {
6246 ret = io_req_defer_prep(req, sqe);
6252 * Never try inline submit of IOSQE_ASYNC is set, go straight
6253 * to async execution.
6255 io_req_init_async(req);
6256 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6257 io_queue_async_work(req);
6260 ret = io_req_prep(req, sqe);
6264 __io_queue_sqe(req, cs);
6268 static inline void io_queue_link_head(struct io_kiocb *req,
6269 struct io_comp_state *cs)
6271 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6273 io_req_complete(req, -ECANCELED);
6275 io_queue_sqe(req, NULL, cs);
6278 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6279 struct io_kiocb **link, struct io_comp_state *cs)
6281 struct io_ring_ctx *ctx = req->ctx;
6285 * If we already have a head request, queue this one for async
6286 * submittal once the head completes. If we don't have a head but
6287 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6288 * submitted sync once the chain is complete. If none of those
6289 * conditions are true (normal request), then just queue it.
6292 struct io_kiocb *head = *link;
6295 * Taking sequential execution of a link, draining both sides
6296 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6297 * requests in the link. So, it drains the head and the
6298 * next after the link request. The last one is done via
6299 * drain_next flag to persist the effect across calls.
6301 if (req->flags & REQ_F_IO_DRAIN) {
6302 head->flags |= REQ_F_IO_DRAIN;
6303 ctx->drain_next = 1;
6305 ret = io_req_defer_prep(req, sqe);
6306 if (unlikely(ret)) {
6307 /* fail even hard links since we don't submit */
6308 head->flags |= REQ_F_FAIL_LINK;
6311 trace_io_uring_link(ctx, req, head);
6312 list_add_tail(&req->link_list, &head->link_list);
6314 /* last request of a link, enqueue the link */
6315 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6316 io_queue_link_head(head, cs);
6320 if (unlikely(ctx->drain_next)) {
6321 req->flags |= REQ_F_IO_DRAIN;
6322 ctx->drain_next = 0;
6324 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6325 req->flags |= REQ_F_LINK_HEAD;
6326 INIT_LIST_HEAD(&req->link_list);
6328 ret = io_req_defer_prep(req, sqe);
6330 req->flags |= REQ_F_FAIL_LINK;
6333 io_queue_sqe(req, sqe, cs);
6341 * Batched submission is done, ensure local IO is flushed out.
6343 static void io_submit_state_end(struct io_submit_state *state)
6345 if (!list_empty(&state->comp.list))
6346 io_submit_flush_completions(&state->comp);
6347 blk_finish_plug(&state->plug);
6348 io_state_file_put(state);
6349 if (state->free_reqs)
6350 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6354 * Start submission side cache.
6356 static void io_submit_state_start(struct io_submit_state *state,
6357 struct io_ring_ctx *ctx, unsigned int max_ios)
6359 blk_start_plug(&state->plug);
6361 INIT_LIST_HEAD(&state->comp.list);
6362 state->comp.ctx = ctx;
6363 state->free_reqs = 0;
6365 state->ios_left = max_ios;
6368 static void io_commit_sqring(struct io_ring_ctx *ctx)
6370 struct io_rings *rings = ctx->rings;
6373 * Ensure any loads from the SQEs are done at this point,
6374 * since once we write the new head, the application could
6375 * write new data to them.
6377 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6381 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6382 * that is mapped by userspace. This means that care needs to be taken to
6383 * ensure that reads are stable, as we cannot rely on userspace always
6384 * being a good citizen. If members of the sqe are validated and then later
6385 * used, it's important that those reads are done through READ_ONCE() to
6386 * prevent a re-load down the line.
6388 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6390 u32 *sq_array = ctx->sq_array;
6394 * The cached sq head (or cq tail) serves two purposes:
6396 * 1) allows us to batch the cost of updating the user visible
6398 * 2) allows the kernel side to track the head on its own, even
6399 * though the application is the one updating it.
6401 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6402 if (likely(head < ctx->sq_entries))
6403 return &ctx->sq_sqes[head];
6405 /* drop invalid entries */
6406 ctx->cached_sq_dropped++;
6407 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6411 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6413 ctx->cached_sq_head++;
6417 * Check SQE restrictions (opcode and flags).
6419 * Returns 'true' if SQE is allowed, 'false' otherwise.
6421 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6422 struct io_kiocb *req,
6423 unsigned int sqe_flags)
6425 if (!ctx->restricted)
6428 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6431 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6432 ctx->restrictions.sqe_flags_required)
6435 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6436 ctx->restrictions.sqe_flags_required))
6442 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6443 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6444 IOSQE_BUFFER_SELECT)
6446 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6447 const struct io_uring_sqe *sqe,
6448 struct io_submit_state *state)
6450 unsigned int sqe_flags;
6453 req->opcode = READ_ONCE(sqe->opcode);
6454 req->user_data = READ_ONCE(sqe->user_data);
6455 req->async_data = NULL;
6459 /* one is dropped after submission, the other at completion */
6460 refcount_set(&req->refs, 2);
6461 req->task = current;
6464 if (unlikely(req->opcode >= IORING_OP_LAST))
6467 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6470 sqe_flags = READ_ONCE(sqe->flags);
6471 /* enforce forwards compatibility on users */
6472 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6475 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6478 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6479 !io_op_defs[req->opcode].buffer_select)
6482 id = READ_ONCE(sqe->personality);
6484 struct io_identity *iod;
6486 io_req_init_async(req);
6487 iod = idr_find(&ctx->personality_idr, id);
6490 refcount_inc(&iod->count);
6491 io_put_identity(current->io_uring, req);
6492 get_cred(iod->creds);
6493 req->work.identity = iod;
6494 req->work.flags |= IO_WQ_WORK_CREDS;
6497 /* same numerical values with corresponding REQ_F_*, safe to copy */
6498 req->flags |= sqe_flags;
6500 if (!io_op_defs[req->opcode].needs_file)
6503 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6508 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6510 struct io_submit_state state;
6511 struct io_kiocb *link = NULL;
6512 int i, submitted = 0;
6514 /* if we have a backlog and couldn't flush it all, return BUSY */
6515 if (test_bit(0, &ctx->sq_check_overflow)) {
6516 if (!list_empty(&ctx->cq_overflow_list) &&
6517 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6521 /* make sure SQ entry isn't read before tail */
6522 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6524 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6527 atomic_long_add(nr, ¤t->io_uring->req_issue);
6528 refcount_add(nr, ¤t->usage);
6530 io_submit_state_start(&state, ctx, nr);
6532 for (i = 0; i < nr; i++) {
6533 const struct io_uring_sqe *sqe;
6534 struct io_kiocb *req;
6537 sqe = io_get_sqe(ctx);
6538 if (unlikely(!sqe)) {
6539 io_consume_sqe(ctx);
6542 req = io_alloc_req(ctx, &state);
6543 if (unlikely(!req)) {
6545 submitted = -EAGAIN;
6548 io_consume_sqe(ctx);
6549 /* will complete beyond this point, count as submitted */
6552 err = io_init_req(ctx, req, sqe, &state);
6553 if (unlikely(err)) {
6556 io_req_complete(req, err);
6560 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6561 true, io_async_submit(ctx));
6562 err = io_submit_sqe(req, sqe, &link, &state.comp);
6567 if (unlikely(submitted != nr)) {
6568 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6570 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6571 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6572 put_task_struct_many(current, nr - ref_used);
6575 io_queue_link_head(link, &state.comp);
6576 io_submit_state_end(&state);
6578 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6579 io_commit_sqring(ctx);
6584 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6586 /* Tell userspace we may need a wakeup call */
6587 spin_lock_irq(&ctx->completion_lock);
6588 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6589 spin_unlock_irq(&ctx->completion_lock);
6592 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6594 spin_lock_irq(&ctx->completion_lock);
6595 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6596 spin_unlock_irq(&ctx->completion_lock);
6599 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6600 int sync, void *key)
6602 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6605 ret = autoremove_wake_function(wqe, mode, sync, key);
6607 unsigned long flags;
6609 spin_lock_irqsave(&ctx->completion_lock, flags);
6610 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6611 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6622 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6623 unsigned long start_jiffies, bool cap_entries)
6625 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6626 struct io_sq_data *sqd = ctx->sq_data;
6627 unsigned int to_submit;
6631 if (!list_empty(&ctx->iopoll_list)) {
6632 unsigned nr_events = 0;
6634 mutex_lock(&ctx->uring_lock);
6635 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6636 io_do_iopoll(ctx, &nr_events, 0);
6637 mutex_unlock(&ctx->uring_lock);
6640 to_submit = io_sqring_entries(ctx);
6643 * If submit got -EBUSY, flag us as needing the application
6644 * to enter the kernel to reap and flush events.
6646 if (!to_submit || ret == -EBUSY || need_resched()) {
6648 * Drop cur_mm before scheduling, we can't hold it for
6649 * long periods (or over schedule()). Do this before
6650 * adding ourselves to the waitqueue, as the unuse/drop
6653 io_sq_thread_drop_mm();
6656 * We're polling. If we're within the defined idle
6657 * period, then let us spin without work before going
6658 * to sleep. The exception is if we got EBUSY doing
6659 * more IO, we should wait for the application to
6660 * reap events and wake us up.
6662 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6663 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6664 !percpu_ref_is_dying(&ctx->refs)))
6667 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6668 TASK_INTERRUPTIBLE);
6671 * While doing polled IO, before going to sleep, we need
6672 * to check if there are new reqs added to iopoll_list,
6673 * it is because reqs may have been punted to io worker
6674 * and will be added to iopoll_list later, hence check
6675 * the iopoll_list again.
6677 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6678 !list_empty_careful(&ctx->iopoll_list)) {
6679 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6683 to_submit = io_sqring_entries(ctx);
6684 if (!to_submit || ret == -EBUSY)
6688 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6689 io_ring_clear_wakeup_flag(ctx);
6691 /* if we're handling multiple rings, cap submit size for fairness */
6692 if (cap_entries && to_submit > 8)
6695 mutex_lock(&ctx->uring_lock);
6696 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6697 ret = io_submit_sqes(ctx, to_submit);
6698 mutex_unlock(&ctx->uring_lock);
6700 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6701 wake_up(&ctx->sqo_sq_wait);
6703 return SQT_DID_WORK;
6706 static void io_sqd_init_new(struct io_sq_data *sqd)
6708 struct io_ring_ctx *ctx;
6710 while (!list_empty(&sqd->ctx_new_list)) {
6711 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6712 init_wait(&ctx->sqo_wait_entry);
6713 ctx->sqo_wait_entry.func = io_sq_wake_function;
6714 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6715 complete(&ctx->sq_thread_comp);
6719 static int io_sq_thread(void *data)
6721 struct cgroup_subsys_state *cur_css = NULL;
6722 const struct cred *old_cred = NULL;
6723 struct io_sq_data *sqd = data;
6724 struct io_ring_ctx *ctx;
6725 unsigned long start_jiffies;
6727 start_jiffies = jiffies;
6728 while (!kthread_should_stop()) {
6729 enum sq_ret ret = 0;
6733 * Any changes to the sqd lists are synchronized through the
6734 * kthread parking. This synchronizes the thread vs users,
6735 * the users are synchronized on the sqd->ctx_lock.
6737 if (kthread_should_park())
6740 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6741 io_sqd_init_new(sqd);
6743 cap_entries = !list_is_singular(&sqd->ctx_list);
6745 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6746 if (current->cred != ctx->creds) {
6748 revert_creds(old_cred);
6749 old_cred = override_creds(ctx->creds);
6751 io_sq_thread_associate_blkcg(ctx, &cur_css);
6753 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6755 io_sq_thread_drop_mm();
6758 if (ret & SQT_SPIN) {
6761 } else if (ret == SQT_IDLE) {
6762 if (kthread_should_park())
6764 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6765 io_ring_set_wakeup_flag(ctx);
6767 start_jiffies = jiffies;
6768 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6769 io_ring_clear_wakeup_flag(ctx);
6776 io_sq_thread_unassociate_blkcg();
6778 revert_creds(old_cred);
6785 struct io_wait_queue {
6786 struct wait_queue_entry wq;
6787 struct io_ring_ctx *ctx;
6789 unsigned nr_timeouts;
6792 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6794 struct io_ring_ctx *ctx = iowq->ctx;
6797 * Wake up if we have enough events, or if a timeout occurred since we
6798 * started waiting. For timeouts, we always want to return to userspace,
6799 * regardless of event count.
6801 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6802 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6805 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6806 int wake_flags, void *key)
6808 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6811 /* use noflush == true, as we can't safely rely on locking context */
6812 if (!io_should_wake(iowq, true))
6815 return autoremove_wake_function(curr, mode, wake_flags, key);
6818 static int io_run_task_work_sig(void)
6820 if (io_run_task_work())
6822 if (!signal_pending(current))
6824 if (current->jobctl & JOBCTL_TASK_WORK) {
6825 spin_lock_irq(¤t->sighand->siglock);
6826 current->jobctl &= ~JOBCTL_TASK_WORK;
6827 recalc_sigpending();
6828 spin_unlock_irq(¤t->sighand->siglock);
6835 * Wait until events become available, if we don't already have some. The
6836 * application must reap them itself, as they reside on the shared cq ring.
6838 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6839 const sigset_t __user *sig, size_t sigsz)
6841 struct io_wait_queue iowq = {
6844 .func = io_wake_function,
6845 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6848 .to_wait = min_events,
6850 struct io_rings *rings = ctx->rings;
6854 if (io_cqring_events(ctx, false) >= min_events)
6856 if (!io_run_task_work())
6861 #ifdef CONFIG_COMPAT
6862 if (in_compat_syscall())
6863 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6867 ret = set_user_sigmask(sig, sigsz);
6873 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6874 trace_io_uring_cqring_wait(ctx, min_events);
6876 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6877 TASK_INTERRUPTIBLE);
6878 /* make sure we run task_work before checking for signals */
6879 ret = io_run_task_work_sig();
6884 if (io_should_wake(&iowq, false))
6888 finish_wait(&ctx->wait, &iowq.wq);
6890 restore_saved_sigmask_unless(ret == -EINTR);
6892 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6895 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6897 #if defined(CONFIG_UNIX)
6898 if (ctx->ring_sock) {
6899 struct sock *sock = ctx->ring_sock->sk;
6900 struct sk_buff *skb;
6902 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6908 for (i = 0; i < ctx->nr_user_files; i++) {
6911 file = io_file_from_index(ctx, i);
6918 static void io_file_ref_kill(struct percpu_ref *ref)
6920 struct fixed_file_data *data;
6922 data = container_of(ref, struct fixed_file_data, refs);
6923 complete(&data->done);
6926 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6928 struct fixed_file_data *data = ctx->file_data;
6929 struct fixed_file_ref_node *ref_node = NULL;
6930 unsigned nr_tables, i;
6935 spin_lock(&data->lock);
6936 if (!list_empty(&data->ref_list))
6937 ref_node = list_first_entry(&data->ref_list,
6938 struct fixed_file_ref_node, node);
6939 spin_unlock(&data->lock);
6941 percpu_ref_kill(&ref_node->refs);
6943 percpu_ref_kill(&data->refs);
6945 /* wait for all refs nodes to complete */
6946 flush_delayed_work(&ctx->file_put_work);
6947 wait_for_completion(&data->done);
6949 __io_sqe_files_unregister(ctx);
6950 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6951 for (i = 0; i < nr_tables; i++)
6952 kfree(data->table[i].files);
6954 percpu_ref_exit(&data->refs);
6956 ctx->file_data = NULL;
6957 ctx->nr_user_files = 0;
6961 static void io_put_sq_data(struct io_sq_data *sqd)
6963 if (refcount_dec_and_test(&sqd->refs)) {
6965 * The park is a bit of a work-around, without it we get
6966 * warning spews on shutdown with SQPOLL set and affinity
6967 * set to a single CPU.
6970 kthread_park(sqd->thread);
6971 kthread_stop(sqd->thread);
6978 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6980 struct io_ring_ctx *ctx_attach;
6981 struct io_sq_data *sqd;
6984 f = fdget(p->wq_fd);
6986 return ERR_PTR(-ENXIO);
6987 if (f.file->f_op != &io_uring_fops) {
6989 return ERR_PTR(-EINVAL);
6992 ctx_attach = f.file->private_data;
6993 sqd = ctx_attach->sq_data;
6996 return ERR_PTR(-EINVAL);
6999 refcount_inc(&sqd->refs);
7004 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7006 struct io_sq_data *sqd;
7008 if (p->flags & IORING_SETUP_ATTACH_WQ)
7009 return io_attach_sq_data(p);
7011 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7013 return ERR_PTR(-ENOMEM);
7015 refcount_set(&sqd->refs, 1);
7016 INIT_LIST_HEAD(&sqd->ctx_list);
7017 INIT_LIST_HEAD(&sqd->ctx_new_list);
7018 mutex_init(&sqd->ctx_lock);
7019 mutex_init(&sqd->lock);
7020 init_waitqueue_head(&sqd->wait);
7024 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7025 __releases(&sqd->lock)
7029 kthread_unpark(sqd->thread);
7030 mutex_unlock(&sqd->lock);
7033 static void io_sq_thread_park(struct io_sq_data *sqd)
7034 __acquires(&sqd->lock)
7038 mutex_lock(&sqd->lock);
7039 kthread_park(sqd->thread);
7042 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7044 struct io_sq_data *sqd = ctx->sq_data;
7049 * We may arrive here from the error branch in
7050 * io_sq_offload_create() where the kthread is created
7051 * without being waked up, thus wake it up now to make
7052 * sure the wait will complete.
7054 wake_up_process(sqd->thread);
7055 wait_for_completion(&ctx->sq_thread_comp);
7057 io_sq_thread_park(sqd);
7060 mutex_lock(&sqd->ctx_lock);
7061 list_del(&ctx->sqd_list);
7062 mutex_unlock(&sqd->ctx_lock);
7065 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7066 io_sq_thread_unpark(sqd);
7069 io_put_sq_data(sqd);
7070 ctx->sq_data = NULL;
7074 static void io_finish_async(struct io_ring_ctx *ctx)
7076 io_sq_thread_stop(ctx);
7079 io_wq_destroy(ctx->io_wq);
7084 #if defined(CONFIG_UNIX)
7086 * Ensure the UNIX gc is aware of our file set, so we are certain that
7087 * the io_uring can be safely unregistered on process exit, even if we have
7088 * loops in the file referencing.
7090 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7092 struct sock *sk = ctx->ring_sock->sk;
7093 struct scm_fp_list *fpl;
7094 struct sk_buff *skb;
7097 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7101 skb = alloc_skb(0, GFP_KERNEL);
7110 fpl->user = get_uid(ctx->user);
7111 for (i = 0; i < nr; i++) {
7112 struct file *file = io_file_from_index(ctx, i + offset);
7116 fpl->fp[nr_files] = get_file(file);
7117 unix_inflight(fpl->user, fpl->fp[nr_files]);
7122 fpl->max = SCM_MAX_FD;
7123 fpl->count = nr_files;
7124 UNIXCB(skb).fp = fpl;
7125 skb->destructor = unix_destruct_scm;
7126 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7127 skb_queue_head(&sk->sk_receive_queue, skb);
7129 for (i = 0; i < nr_files; i++)
7140 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7141 * causes regular reference counting to break down. We rely on the UNIX
7142 * garbage collection to take care of this problem for us.
7144 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7146 unsigned left, total;
7150 left = ctx->nr_user_files;
7152 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7154 ret = __io_sqe_files_scm(ctx, this_files, total);
7158 total += this_files;
7164 while (total < ctx->nr_user_files) {
7165 struct file *file = io_file_from_index(ctx, total);
7175 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7181 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7182 unsigned nr_tables, unsigned nr_files)
7186 for (i = 0; i < nr_tables; i++) {
7187 struct fixed_file_table *table = &file_data->table[i];
7188 unsigned this_files;
7190 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7191 table->files = kcalloc(this_files, sizeof(struct file *),
7195 nr_files -= this_files;
7201 for (i = 0; i < nr_tables; i++) {
7202 struct fixed_file_table *table = &file_data->table[i];
7203 kfree(table->files);
7208 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7210 #if defined(CONFIG_UNIX)
7211 struct sock *sock = ctx->ring_sock->sk;
7212 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7213 struct sk_buff *skb;
7216 __skb_queue_head_init(&list);
7219 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7220 * remove this entry and rearrange the file array.
7222 skb = skb_dequeue(head);
7224 struct scm_fp_list *fp;
7226 fp = UNIXCB(skb).fp;
7227 for (i = 0; i < fp->count; i++) {
7230 if (fp->fp[i] != file)
7233 unix_notinflight(fp->user, fp->fp[i]);
7234 left = fp->count - 1 - i;
7236 memmove(&fp->fp[i], &fp->fp[i + 1],
7237 left * sizeof(struct file *));
7244 __skb_queue_tail(&list, skb);
7254 __skb_queue_tail(&list, skb);
7256 skb = skb_dequeue(head);
7259 if (skb_peek(&list)) {
7260 spin_lock_irq(&head->lock);
7261 while ((skb = __skb_dequeue(&list)) != NULL)
7262 __skb_queue_tail(head, skb);
7263 spin_unlock_irq(&head->lock);
7270 struct io_file_put {
7271 struct list_head list;
7275 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7277 struct fixed_file_data *file_data = ref_node->file_data;
7278 struct io_ring_ctx *ctx = file_data->ctx;
7279 struct io_file_put *pfile, *tmp;
7281 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7282 list_del(&pfile->list);
7283 io_ring_file_put(ctx, pfile->file);
7287 spin_lock(&file_data->lock);
7288 list_del(&ref_node->node);
7289 spin_unlock(&file_data->lock);
7291 percpu_ref_exit(&ref_node->refs);
7293 percpu_ref_put(&file_data->refs);
7296 static void io_file_put_work(struct work_struct *work)
7298 struct io_ring_ctx *ctx;
7299 struct llist_node *node;
7301 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7302 node = llist_del_all(&ctx->file_put_llist);
7305 struct fixed_file_ref_node *ref_node;
7306 struct llist_node *next = node->next;
7308 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7309 __io_file_put_work(ref_node);
7314 static void io_file_data_ref_zero(struct percpu_ref *ref)
7316 struct fixed_file_ref_node *ref_node;
7317 struct io_ring_ctx *ctx;
7321 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7322 ctx = ref_node->file_data->ctx;
7324 if (percpu_ref_is_dying(&ctx->file_data->refs))
7327 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7329 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7331 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7334 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7335 struct io_ring_ctx *ctx)
7337 struct fixed_file_ref_node *ref_node;
7339 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7341 return ERR_PTR(-ENOMEM);
7343 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7346 return ERR_PTR(-ENOMEM);
7348 INIT_LIST_HEAD(&ref_node->node);
7349 INIT_LIST_HEAD(&ref_node->file_list);
7350 ref_node->file_data = ctx->file_data;
7354 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7356 percpu_ref_exit(&ref_node->refs);
7360 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7363 __s32 __user *fds = (__s32 __user *) arg;
7364 unsigned nr_tables, i;
7366 int fd, ret = -ENOMEM;
7367 struct fixed_file_ref_node *ref_node;
7368 struct fixed_file_data *file_data;
7374 if (nr_args > IORING_MAX_FIXED_FILES)
7377 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7380 file_data->ctx = ctx;
7381 init_completion(&file_data->done);
7382 INIT_LIST_HEAD(&file_data->ref_list);
7383 spin_lock_init(&file_data->lock);
7385 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7386 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7388 if (!file_data->table)
7391 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7392 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7395 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7397 ctx->file_data = file_data;
7399 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7400 struct fixed_file_table *table;
7403 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7407 /* allow sparse sets */
7417 * Don't allow io_uring instances to be registered. If UNIX
7418 * isn't enabled, then this causes a reference cycle and this
7419 * instance can never get freed. If UNIX is enabled we'll
7420 * handle it just fine, but there's still no point in allowing
7421 * a ring fd as it doesn't support regular read/write anyway.
7423 if (file->f_op == &io_uring_fops) {
7427 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7428 index = i & IORING_FILE_TABLE_MASK;
7429 table->files[index] = file;
7432 ret = io_sqe_files_scm(ctx);
7434 io_sqe_files_unregister(ctx);
7438 ref_node = alloc_fixed_file_ref_node(ctx);
7439 if (IS_ERR(ref_node)) {
7440 io_sqe_files_unregister(ctx);
7441 return PTR_ERR(ref_node);
7444 file_data->node = ref_node;
7445 spin_lock(&file_data->lock);
7446 list_add(&ref_node->node, &file_data->ref_list);
7447 spin_unlock(&file_data->lock);
7448 percpu_ref_get(&file_data->refs);
7451 for (i = 0; i < ctx->nr_user_files; i++) {
7452 file = io_file_from_index(ctx, i);
7456 for (i = 0; i < nr_tables; i++)
7457 kfree(file_data->table[i].files);
7458 ctx->nr_user_files = 0;
7460 percpu_ref_exit(&file_data->refs);
7462 kfree(file_data->table);
7464 ctx->file_data = NULL;
7468 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7471 #if defined(CONFIG_UNIX)
7472 struct sock *sock = ctx->ring_sock->sk;
7473 struct sk_buff_head *head = &sock->sk_receive_queue;
7474 struct sk_buff *skb;
7477 * See if we can merge this file into an existing skb SCM_RIGHTS
7478 * file set. If there's no room, fall back to allocating a new skb
7479 * and filling it in.
7481 spin_lock_irq(&head->lock);
7482 skb = skb_peek(head);
7484 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7486 if (fpl->count < SCM_MAX_FD) {
7487 __skb_unlink(skb, head);
7488 spin_unlock_irq(&head->lock);
7489 fpl->fp[fpl->count] = get_file(file);
7490 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7492 spin_lock_irq(&head->lock);
7493 __skb_queue_head(head, skb);
7498 spin_unlock_irq(&head->lock);
7505 return __io_sqe_files_scm(ctx, 1, index);
7511 static int io_queue_file_removal(struct fixed_file_data *data,
7514 struct io_file_put *pfile;
7515 struct fixed_file_ref_node *ref_node = data->node;
7517 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7522 list_add(&pfile->list, &ref_node->file_list);
7527 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7528 struct io_uring_files_update *up,
7531 struct fixed_file_data *data = ctx->file_data;
7532 struct fixed_file_ref_node *ref_node;
7537 bool needs_switch = false;
7539 if (check_add_overflow(up->offset, nr_args, &done))
7541 if (done > ctx->nr_user_files)
7544 ref_node = alloc_fixed_file_ref_node(ctx);
7545 if (IS_ERR(ref_node))
7546 return PTR_ERR(ref_node);
7549 fds = u64_to_user_ptr(up->fds);
7551 struct fixed_file_table *table;
7555 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7559 i = array_index_nospec(up->offset, ctx->nr_user_files);
7560 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7561 index = i & IORING_FILE_TABLE_MASK;
7562 if (table->files[index]) {
7563 file = table->files[index];
7564 err = io_queue_file_removal(data, file);
7567 table->files[index] = NULL;
7568 needs_switch = true;
7577 * Don't allow io_uring instances to be registered. If
7578 * UNIX isn't enabled, then this causes a reference
7579 * cycle and this instance can never get freed. If UNIX
7580 * is enabled we'll handle it just fine, but there's
7581 * still no point in allowing a ring fd as it doesn't
7582 * support regular read/write anyway.
7584 if (file->f_op == &io_uring_fops) {
7589 table->files[index] = file;
7590 err = io_sqe_file_register(ctx, file, i);
7592 table->files[index] = NULL;
7603 percpu_ref_kill(&data->node->refs);
7604 spin_lock(&data->lock);
7605 list_add(&ref_node->node, &data->ref_list);
7606 data->node = ref_node;
7607 spin_unlock(&data->lock);
7608 percpu_ref_get(&ctx->file_data->refs);
7610 destroy_fixed_file_ref_node(ref_node);
7612 return done ? done : err;
7615 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7618 struct io_uring_files_update up;
7620 if (!ctx->file_data)
7624 if (copy_from_user(&up, arg, sizeof(up)))
7629 return __io_sqe_files_update(ctx, &up, nr_args);
7632 static void io_free_work(struct io_wq_work *work)
7634 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7636 /* Consider that io_steal_work() relies on this ref */
7640 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7641 struct io_uring_params *p)
7643 struct io_wq_data data;
7645 struct io_ring_ctx *ctx_attach;
7646 unsigned int concurrency;
7649 data.user = ctx->user;
7650 data.free_work = io_free_work;
7651 data.do_work = io_wq_submit_work;
7653 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7654 /* Do QD, or 4 * CPUS, whatever is smallest */
7655 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7657 ctx->io_wq = io_wq_create(concurrency, &data);
7658 if (IS_ERR(ctx->io_wq)) {
7659 ret = PTR_ERR(ctx->io_wq);
7665 f = fdget(p->wq_fd);
7669 if (f.file->f_op != &io_uring_fops) {
7674 ctx_attach = f.file->private_data;
7675 /* @io_wq is protected by holding the fd */
7676 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7681 ctx->io_wq = ctx_attach->io_wq;
7687 static int io_uring_alloc_task_context(struct task_struct *task)
7689 struct io_uring_task *tctx;
7691 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7692 if (unlikely(!tctx))
7696 init_waitqueue_head(&tctx->wait);
7699 atomic_long_set(&tctx->req_issue, 0);
7700 atomic_long_set(&tctx->req_complete, 0);
7701 io_init_identity(&tctx->__identity);
7702 tctx->identity = &tctx->__identity;
7703 task->io_uring = tctx;
7707 void __io_uring_free(struct task_struct *tsk)
7709 struct io_uring_task *tctx = tsk->io_uring;
7711 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7712 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7713 if (tctx->identity != &tctx->__identity)
7714 kfree(tctx->identity);
7716 tsk->io_uring = NULL;
7719 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7720 struct io_uring_params *p)
7724 if (ctx->flags & IORING_SETUP_SQPOLL) {
7725 struct io_sq_data *sqd;
7728 if (!capable(CAP_SYS_ADMIN))
7731 sqd = io_get_sq_data(p);
7738 io_sq_thread_park(sqd);
7739 mutex_lock(&sqd->ctx_lock);
7740 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7741 mutex_unlock(&sqd->ctx_lock);
7742 io_sq_thread_unpark(sqd);
7744 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7745 if (!ctx->sq_thread_idle)
7746 ctx->sq_thread_idle = HZ;
7751 if (p->flags & IORING_SETUP_SQ_AFF) {
7752 int cpu = p->sq_thread_cpu;
7755 if (cpu >= nr_cpu_ids)
7757 if (!cpu_online(cpu))
7760 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7761 cpu, "io_uring-sq");
7763 sqd->thread = kthread_create(io_sq_thread, sqd,
7766 if (IS_ERR(sqd->thread)) {
7767 ret = PTR_ERR(sqd->thread);
7771 ret = io_uring_alloc_task_context(sqd->thread);
7774 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7775 /* Can't have SQ_AFF without SQPOLL */
7781 ret = io_init_wq_offload(ctx, p);
7787 io_finish_async(ctx);
7791 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7793 struct io_sq_data *sqd = ctx->sq_data;
7795 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7796 wake_up_process(sqd->thread);
7799 static inline void __io_unaccount_mem(struct user_struct *user,
7800 unsigned long nr_pages)
7802 atomic_long_sub(nr_pages, &user->locked_vm);
7805 static inline int __io_account_mem(struct user_struct *user,
7806 unsigned long nr_pages)
7808 unsigned long page_limit, cur_pages, new_pages;
7810 /* Don't allow more pages than we can safely lock */
7811 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7814 cur_pages = atomic_long_read(&user->locked_vm);
7815 new_pages = cur_pages + nr_pages;
7816 if (new_pages > page_limit)
7818 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7819 new_pages) != cur_pages);
7824 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7825 enum io_mem_account acct)
7828 __io_unaccount_mem(ctx->user, nr_pages);
7830 if (ctx->mm_account) {
7831 if (acct == ACCT_LOCKED)
7832 ctx->mm_account->locked_vm -= nr_pages;
7833 else if (acct == ACCT_PINNED)
7834 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7838 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7839 enum io_mem_account acct)
7843 if (ctx->limit_mem) {
7844 ret = __io_account_mem(ctx->user, nr_pages);
7849 if (ctx->mm_account) {
7850 if (acct == ACCT_LOCKED)
7851 ctx->mm_account->locked_vm += nr_pages;
7852 else if (acct == ACCT_PINNED)
7853 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7859 static void io_mem_free(void *ptr)
7866 page = virt_to_head_page(ptr);
7867 if (put_page_testzero(page))
7868 free_compound_page(page);
7871 static void *io_mem_alloc(size_t size)
7873 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7876 return (void *) __get_free_pages(gfp_flags, get_order(size));
7879 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7882 struct io_rings *rings;
7883 size_t off, sq_array_size;
7885 off = struct_size(rings, cqes, cq_entries);
7886 if (off == SIZE_MAX)
7890 off = ALIGN(off, SMP_CACHE_BYTES);
7898 sq_array_size = array_size(sizeof(u32), sq_entries);
7899 if (sq_array_size == SIZE_MAX)
7902 if (check_add_overflow(off, sq_array_size, &off))
7908 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7912 pages = (size_t)1 << get_order(
7913 rings_size(sq_entries, cq_entries, NULL));
7914 pages += (size_t)1 << get_order(
7915 array_size(sizeof(struct io_uring_sqe), sq_entries));
7920 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7924 if (!ctx->user_bufs)
7927 for (i = 0; i < ctx->nr_user_bufs; i++) {
7928 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7930 for (j = 0; j < imu->nr_bvecs; j++)
7931 unpin_user_page(imu->bvec[j].bv_page);
7933 if (imu->acct_pages)
7934 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7939 kfree(ctx->user_bufs);
7940 ctx->user_bufs = NULL;
7941 ctx->nr_user_bufs = 0;
7945 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7946 void __user *arg, unsigned index)
7948 struct iovec __user *src;
7950 #ifdef CONFIG_COMPAT
7952 struct compat_iovec __user *ciovs;
7953 struct compat_iovec ciov;
7955 ciovs = (struct compat_iovec __user *) arg;
7956 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7959 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7960 dst->iov_len = ciov.iov_len;
7964 src = (struct iovec __user *) arg;
7965 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7971 * Not super efficient, but this is just a registration time. And we do cache
7972 * the last compound head, so generally we'll only do a full search if we don't
7975 * We check if the given compound head page has already been accounted, to
7976 * avoid double accounting it. This allows us to account the full size of the
7977 * page, not just the constituent pages of a huge page.
7979 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7980 int nr_pages, struct page *hpage)
7984 /* check current page array */
7985 for (i = 0; i < nr_pages; i++) {
7986 if (!PageCompound(pages[i]))
7988 if (compound_head(pages[i]) == hpage)
7992 /* check previously registered pages */
7993 for (i = 0; i < ctx->nr_user_bufs; i++) {
7994 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7996 for (j = 0; j < imu->nr_bvecs; j++) {
7997 if (!PageCompound(imu->bvec[j].bv_page))
7999 if (compound_head(imu->bvec[j].bv_page) == hpage)
8007 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8008 int nr_pages, struct io_mapped_ubuf *imu,
8009 struct page **last_hpage)
8013 for (i = 0; i < nr_pages; i++) {
8014 if (!PageCompound(pages[i])) {
8019 hpage = compound_head(pages[i]);
8020 if (hpage == *last_hpage)
8022 *last_hpage = hpage;
8023 if (headpage_already_acct(ctx, pages, i, hpage))
8025 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8029 if (!imu->acct_pages)
8032 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8034 imu->acct_pages = 0;
8038 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8041 struct vm_area_struct **vmas = NULL;
8042 struct page **pages = NULL;
8043 struct page *last_hpage = NULL;
8044 int i, j, got_pages = 0;
8049 if (!nr_args || nr_args > UIO_MAXIOV)
8052 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8054 if (!ctx->user_bufs)
8057 for (i = 0; i < nr_args; i++) {
8058 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8059 unsigned long off, start, end, ubuf;
8064 ret = io_copy_iov(ctx, &iov, arg, i);
8069 * Don't impose further limits on the size and buffer
8070 * constraints here, we'll -EINVAL later when IO is
8071 * submitted if they are wrong.
8074 if (!iov.iov_base || !iov.iov_len)
8077 /* arbitrary limit, but we need something */
8078 if (iov.iov_len > SZ_1G)
8081 ubuf = (unsigned long) iov.iov_base;
8082 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8083 start = ubuf >> PAGE_SHIFT;
8084 nr_pages = end - start;
8087 if (!pages || nr_pages > got_pages) {
8090 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8092 vmas = kvmalloc_array(nr_pages,
8093 sizeof(struct vm_area_struct *),
8095 if (!pages || !vmas) {
8099 got_pages = nr_pages;
8102 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8109 mmap_read_lock(current->mm);
8110 pret = pin_user_pages(ubuf, nr_pages,
8111 FOLL_WRITE | FOLL_LONGTERM,
8113 if (pret == nr_pages) {
8114 /* don't support file backed memory */
8115 for (j = 0; j < nr_pages; j++) {
8116 struct vm_area_struct *vma = vmas[j];
8119 !is_file_hugepages(vma->vm_file)) {
8125 ret = pret < 0 ? pret : -EFAULT;
8127 mmap_read_unlock(current->mm);
8130 * if we did partial map, or found file backed vmas,
8131 * release any pages we did get
8134 unpin_user_pages(pages, pret);
8139 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8141 unpin_user_pages(pages, pret);
8146 off = ubuf & ~PAGE_MASK;
8148 for (j = 0; j < nr_pages; j++) {
8151 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8152 imu->bvec[j].bv_page = pages[j];
8153 imu->bvec[j].bv_len = vec_len;
8154 imu->bvec[j].bv_offset = off;
8158 /* store original address for later verification */
8160 imu->len = iov.iov_len;
8161 imu->nr_bvecs = nr_pages;
8163 ctx->nr_user_bufs++;
8171 io_sqe_buffer_unregister(ctx);
8175 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8177 __s32 __user *fds = arg;
8183 if (copy_from_user(&fd, fds, sizeof(*fds)))
8186 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8187 if (IS_ERR(ctx->cq_ev_fd)) {
8188 int ret = PTR_ERR(ctx->cq_ev_fd);
8189 ctx->cq_ev_fd = NULL;
8196 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8198 if (ctx->cq_ev_fd) {
8199 eventfd_ctx_put(ctx->cq_ev_fd);
8200 ctx->cq_ev_fd = NULL;
8207 static int __io_destroy_buffers(int id, void *p, void *data)
8209 struct io_ring_ctx *ctx = data;
8210 struct io_buffer *buf = p;
8212 __io_remove_buffers(ctx, buf, id, -1U);
8216 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8218 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8219 idr_destroy(&ctx->io_buffer_idr);
8222 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8224 io_finish_async(ctx);
8225 io_sqe_buffer_unregister(ctx);
8227 if (ctx->sqo_task) {
8228 put_task_struct(ctx->sqo_task);
8229 ctx->sqo_task = NULL;
8230 mmdrop(ctx->mm_account);
8231 ctx->mm_account = NULL;
8234 #ifdef CONFIG_BLK_CGROUP
8235 if (ctx->sqo_blkcg_css)
8236 css_put(ctx->sqo_blkcg_css);
8239 io_sqe_files_unregister(ctx);
8240 io_eventfd_unregister(ctx);
8241 io_destroy_buffers(ctx);
8242 idr_destroy(&ctx->personality_idr);
8244 #if defined(CONFIG_UNIX)
8245 if (ctx->ring_sock) {
8246 ctx->ring_sock->file = NULL; /* so that iput() is called */
8247 sock_release(ctx->ring_sock);
8251 io_mem_free(ctx->rings);
8252 io_mem_free(ctx->sq_sqes);
8254 percpu_ref_exit(&ctx->refs);
8255 free_uid(ctx->user);
8256 put_cred(ctx->creds);
8257 kfree(ctx->cancel_hash);
8258 kmem_cache_free(req_cachep, ctx->fallback_req);
8262 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8264 struct io_ring_ctx *ctx = file->private_data;
8267 poll_wait(file, &ctx->cq_wait, wait);
8269 * synchronizes with barrier from wq_has_sleeper call in
8273 if (!io_sqring_full(ctx))
8274 mask |= EPOLLOUT | EPOLLWRNORM;
8275 if (io_cqring_events(ctx, false))
8276 mask |= EPOLLIN | EPOLLRDNORM;
8281 static int io_uring_fasync(int fd, struct file *file, int on)
8283 struct io_ring_ctx *ctx = file->private_data;
8285 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8288 static int io_remove_personalities(int id, void *p, void *data)
8290 struct io_ring_ctx *ctx = data;
8291 struct io_identity *iod;
8293 iod = idr_remove(&ctx->personality_idr, id);
8295 put_cred(iod->creds);
8296 if (refcount_dec_and_test(&iod->count))
8302 static void io_ring_exit_work(struct work_struct *work)
8304 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8308 * If we're doing polled IO and end up having requests being
8309 * submitted async (out-of-line), then completions can come in while
8310 * we're waiting for refs to drop. We need to reap these manually,
8311 * as nobody else will be looking for them.
8315 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8316 io_iopoll_try_reap_events(ctx);
8317 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8318 io_ring_ctx_free(ctx);
8321 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8323 mutex_lock(&ctx->uring_lock);
8324 percpu_ref_kill(&ctx->refs);
8325 mutex_unlock(&ctx->uring_lock);
8327 io_kill_timeouts(ctx, NULL);
8328 io_poll_remove_all(ctx, NULL);
8331 io_wq_cancel_all(ctx->io_wq);
8333 /* if we failed setting up the ctx, we might not have any rings */
8335 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8336 io_iopoll_try_reap_events(ctx);
8337 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8340 * Do this upfront, so we won't have a grace period where the ring
8341 * is closed but resources aren't reaped yet. This can cause
8342 * spurious failure in setting up a new ring.
8344 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8347 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8349 * Use system_unbound_wq to avoid spawning tons of event kworkers
8350 * if we're exiting a ton of rings at the same time. It just adds
8351 * noise and overhead, there's no discernable change in runtime
8352 * over using system_wq.
8354 queue_work(system_unbound_wq, &ctx->exit_work);
8357 static int io_uring_release(struct inode *inode, struct file *file)
8359 struct io_ring_ctx *ctx = file->private_data;
8361 file->private_data = NULL;
8362 io_ring_ctx_wait_and_kill(ctx);
8366 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8368 struct files_struct *files = data;
8370 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8371 work->identity->files == files);
8375 * Returns true if 'preq' is the link parent of 'req'
8377 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8379 struct io_kiocb *link;
8381 if (!(preq->flags & REQ_F_LINK_HEAD))
8384 list_for_each_entry(link, &preq->link_list, link_list) {
8392 static bool io_match_link_files(struct io_kiocb *req,
8393 struct files_struct *files)
8395 struct io_kiocb *link;
8397 if (io_match_files(req, files))
8399 if (req->flags & REQ_F_LINK_HEAD) {
8400 list_for_each_entry(link, &req->link_list, link_list) {
8401 if (io_match_files(link, files))
8409 * We're looking to cancel 'req' because it's holding on to our files, but
8410 * 'req' could be a link to another request. See if it is, and cancel that
8411 * parent request if so.
8413 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8415 struct hlist_node *tmp;
8416 struct io_kiocb *preq;
8420 spin_lock_irq(&ctx->completion_lock);
8421 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8422 struct hlist_head *list;
8424 list = &ctx->cancel_hash[i];
8425 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8426 found = io_match_link(preq, req);
8428 io_poll_remove_one(preq);
8433 spin_unlock_irq(&ctx->completion_lock);
8437 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8438 struct io_kiocb *req)
8440 struct io_kiocb *preq;
8443 spin_lock_irq(&ctx->completion_lock);
8444 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8445 found = io_match_link(preq, req);
8447 __io_timeout_cancel(preq);
8451 spin_unlock_irq(&ctx->completion_lock);
8455 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8457 return io_match_link(container_of(work, struct io_kiocb, work), data);
8460 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8462 enum io_wq_cancel cret;
8464 /* cancel this particular work, if it's running */
8465 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8466 if (cret != IO_WQ_CANCEL_NOTFOUND)
8469 /* find links that hold this pending, cancel those */
8470 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8471 if (cret != IO_WQ_CANCEL_NOTFOUND)
8474 /* if we have a poll link holding this pending, cancel that */
8475 if (io_poll_remove_link(ctx, req))
8478 /* final option, timeout link is holding this req pending */
8479 io_timeout_remove_link(ctx, req);
8482 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8483 struct files_struct *files)
8485 struct io_defer_entry *de = NULL;
8488 spin_lock_irq(&ctx->completion_lock);
8489 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8490 if (io_match_link_files(de->req, files)) {
8491 list_cut_position(&list, &ctx->defer_list, &de->list);
8495 spin_unlock_irq(&ctx->completion_lock);
8497 while (!list_empty(&list)) {
8498 de = list_first_entry(&list, struct io_defer_entry, list);
8499 list_del_init(&de->list);
8500 req_set_fail_links(de->req);
8501 io_put_req(de->req);
8502 io_req_complete(de->req, -ECANCELED);
8508 * Returns true if we found and killed one or more files pinning requests
8510 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8511 struct files_struct *files)
8513 if (list_empty_careful(&ctx->inflight_list))
8516 io_cancel_defer_files(ctx, files);
8517 /* cancel all at once, should be faster than doing it one by one*/
8518 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8520 while (!list_empty_careful(&ctx->inflight_list)) {
8521 struct io_kiocb *cancel_req = NULL, *req;
8524 spin_lock_irq(&ctx->inflight_lock);
8525 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8526 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8527 req->work.identity->files != files)
8529 /* req is being completed, ignore */
8530 if (!refcount_inc_not_zero(&req->refs))
8536 prepare_to_wait(&ctx->inflight_wait, &wait,
8537 TASK_UNINTERRUPTIBLE);
8538 spin_unlock_irq(&ctx->inflight_lock);
8540 /* We need to keep going until we don't find a matching req */
8543 /* cancel this request, or head link requests */
8544 io_attempt_cancel(ctx, cancel_req);
8545 io_put_req(cancel_req);
8546 /* cancellations _may_ trigger task work */
8549 finish_wait(&ctx->inflight_wait, &wait);
8555 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8557 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8558 struct task_struct *task = data;
8560 return io_task_match(req, task);
8563 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8564 struct task_struct *task,
8565 struct files_struct *files)
8569 ret = io_uring_cancel_files(ctx, files);
8571 enum io_wq_cancel cret;
8573 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8574 if (cret != IO_WQ_CANCEL_NOTFOUND)
8577 /* SQPOLL thread does its own polling */
8578 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8579 while (!list_empty_careful(&ctx->iopoll_list)) {
8580 io_iopoll_try_reap_events(ctx);
8585 ret |= io_poll_remove_all(ctx, task);
8586 ret |= io_kill_timeouts(ctx, task);
8593 * We need to iteratively cancel requests, in case a request has dependent
8594 * hard links. These persist even for failure of cancelations, hence keep
8595 * looping until none are found.
8597 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8598 struct files_struct *files)
8600 struct task_struct *task = current;
8602 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8603 task = ctx->sq_data->thread;
8605 io_cqring_overflow_flush(ctx, true, task, files);
8607 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8614 * Note that this task has used io_uring. We use it for cancelation purposes.
8616 static int io_uring_add_task_file(struct file *file)
8618 struct io_uring_task *tctx = current->io_uring;
8620 if (unlikely(!tctx)) {
8623 ret = io_uring_alloc_task_context(current);
8626 tctx = current->io_uring;
8628 if (tctx->last != file) {
8629 void *old = xa_load(&tctx->xa, (unsigned long)file);
8633 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8642 * Remove this io_uring_file -> task mapping.
8644 static void io_uring_del_task_file(struct file *file)
8646 struct io_uring_task *tctx = current->io_uring;
8648 if (tctx->last == file)
8650 file = xa_erase(&tctx->xa, (unsigned long)file);
8655 static void __io_uring_attempt_task_drop(struct file *file)
8657 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8660 io_uring_del_task_file(file);
8664 * Drop task note for this file if we're the only ones that hold it after
8667 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8669 if (!current->io_uring)
8672 * fput() is pending, will be 2 if the only other ref is our potential
8673 * task file note. If the task is exiting, drop regardless of count.
8675 if (!exiting && atomic_long_read(&file->f_count) != 2)
8678 __io_uring_attempt_task_drop(file);
8681 void __io_uring_files_cancel(struct files_struct *files)
8683 struct io_uring_task *tctx = current->io_uring;
8685 unsigned long index;
8687 /* make sure overflow events are dropped */
8688 tctx->in_idle = true;
8690 xa_for_each(&tctx->xa, index, file) {
8691 struct io_ring_ctx *ctx = file->private_data;
8693 io_uring_cancel_task_requests(ctx, files);
8695 io_uring_del_task_file(file);
8699 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8701 return atomic_long_read(&tctx->req_issue) ==
8702 atomic_long_read(&tctx->req_complete);
8706 * Find any io_uring fd that this task has registered or done IO on, and cancel
8709 void __io_uring_task_cancel(void)
8711 struct io_uring_task *tctx = current->io_uring;
8715 /* make sure overflow events are dropped */
8716 tctx->in_idle = true;
8718 while (!io_uring_task_idle(tctx)) {
8719 /* read completions before cancelations */
8720 completions = atomic_long_read(&tctx->req_complete);
8721 __io_uring_files_cancel(NULL);
8723 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8726 * If we've seen completions, retry. This avoids a race where
8727 * a completion comes in before we did prepare_to_wait().
8729 if (completions != atomic_long_read(&tctx->req_complete))
8731 if (io_uring_task_idle(tctx))
8736 finish_wait(&tctx->wait, &wait);
8737 tctx->in_idle = false;
8740 static int io_uring_flush(struct file *file, void *data)
8742 struct io_ring_ctx *ctx = file->private_data;
8745 * If the task is going away, cancel work it may have pending
8747 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8750 io_uring_cancel_task_requests(ctx, data);
8751 io_uring_attempt_task_drop(file, !data);
8755 static void *io_uring_validate_mmap_request(struct file *file,
8756 loff_t pgoff, size_t sz)
8758 struct io_ring_ctx *ctx = file->private_data;
8759 loff_t offset = pgoff << PAGE_SHIFT;
8764 case IORING_OFF_SQ_RING:
8765 case IORING_OFF_CQ_RING:
8768 case IORING_OFF_SQES:
8772 return ERR_PTR(-EINVAL);
8775 page = virt_to_head_page(ptr);
8776 if (sz > page_size(page))
8777 return ERR_PTR(-EINVAL);
8784 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8786 size_t sz = vma->vm_end - vma->vm_start;
8790 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8792 return PTR_ERR(ptr);
8794 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8795 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8798 #else /* !CONFIG_MMU */
8800 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8802 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8805 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8807 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8810 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8811 unsigned long addr, unsigned long len,
8812 unsigned long pgoff, unsigned long flags)
8816 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8818 return PTR_ERR(ptr);
8820 return (unsigned long) ptr;
8823 #endif /* !CONFIG_MMU */
8825 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8830 if (!io_sqring_full(ctx))
8833 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8835 if (!io_sqring_full(ctx))
8839 } while (!signal_pending(current));
8841 finish_wait(&ctx->sqo_sq_wait, &wait);
8844 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8845 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8848 struct io_ring_ctx *ctx;
8855 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8856 IORING_ENTER_SQ_WAIT))
8864 if (f.file->f_op != &io_uring_fops)
8868 ctx = f.file->private_data;
8869 if (!percpu_ref_tryget(&ctx->refs))
8873 if (ctx->flags & IORING_SETUP_R_DISABLED)
8877 * For SQ polling, the thread will do all submissions and completions.
8878 * Just return the requested submit count, and wake the thread if
8882 if (ctx->flags & IORING_SETUP_SQPOLL) {
8883 if (!list_empty_careful(&ctx->cq_overflow_list))
8884 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8885 if (flags & IORING_ENTER_SQ_WAKEUP)
8886 wake_up(&ctx->sq_data->wait);
8887 if (flags & IORING_ENTER_SQ_WAIT)
8888 io_sqpoll_wait_sq(ctx);
8889 submitted = to_submit;
8890 } else if (to_submit) {
8891 ret = io_uring_add_task_file(f.file);
8894 mutex_lock(&ctx->uring_lock);
8895 submitted = io_submit_sqes(ctx, to_submit);
8896 mutex_unlock(&ctx->uring_lock);
8898 if (submitted != to_submit)
8901 if (flags & IORING_ENTER_GETEVENTS) {
8902 min_complete = min(min_complete, ctx->cq_entries);
8905 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8906 * space applications don't need to do io completion events
8907 * polling again, they can rely on io_sq_thread to do polling
8908 * work, which can reduce cpu usage and uring_lock contention.
8910 if (ctx->flags & IORING_SETUP_IOPOLL &&
8911 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8912 ret = io_iopoll_check(ctx, min_complete);
8914 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8919 percpu_ref_put(&ctx->refs);
8922 return submitted ? submitted : ret;
8925 #ifdef CONFIG_PROC_FS
8926 static int io_uring_show_cred(int id, void *p, void *data)
8928 const struct cred *cred = p;
8929 struct seq_file *m = data;
8930 struct user_namespace *uns = seq_user_ns(m);
8931 struct group_info *gi;
8936 seq_printf(m, "%5d\n", id);
8937 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8938 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8939 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8940 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8941 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8942 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8943 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8944 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8945 seq_puts(m, "\n\tGroups:\t");
8946 gi = cred->group_info;
8947 for (g = 0; g < gi->ngroups; g++) {
8948 seq_put_decimal_ull(m, g ? " " : "",
8949 from_kgid_munged(uns, gi->gid[g]));
8951 seq_puts(m, "\n\tCapEff:\t");
8952 cap = cred->cap_effective;
8953 CAP_FOR_EACH_U32(__capi)
8954 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8959 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8961 struct io_sq_data *sq = NULL;
8966 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8967 * since fdinfo case grabs it in the opposite direction of normal use
8968 * cases. If we fail to get the lock, we just don't iterate any
8969 * structures that could be going away outside the io_uring mutex.
8971 has_lock = mutex_trylock(&ctx->uring_lock);
8973 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8976 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8977 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8978 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8979 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8980 struct fixed_file_table *table;
8983 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8984 f = table->files[i & IORING_FILE_TABLE_MASK];
8986 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8988 seq_printf(m, "%5u: <none>\n", i);
8990 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8991 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8992 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8994 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8995 (unsigned int) buf->len);
8997 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8998 seq_printf(m, "Personalities:\n");
8999 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9001 seq_printf(m, "PollList:\n");
9002 spin_lock_irq(&ctx->completion_lock);
9003 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9004 struct hlist_head *list = &ctx->cancel_hash[i];
9005 struct io_kiocb *req;
9007 hlist_for_each_entry(req, list, hash_node)
9008 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9009 req->task->task_works != NULL);
9011 spin_unlock_irq(&ctx->completion_lock);
9013 mutex_unlock(&ctx->uring_lock);
9016 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9018 struct io_ring_ctx *ctx = f->private_data;
9020 if (percpu_ref_tryget(&ctx->refs)) {
9021 __io_uring_show_fdinfo(ctx, m);
9022 percpu_ref_put(&ctx->refs);
9027 static const struct file_operations io_uring_fops = {
9028 .release = io_uring_release,
9029 .flush = io_uring_flush,
9030 .mmap = io_uring_mmap,
9032 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9033 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9035 .poll = io_uring_poll,
9036 .fasync = io_uring_fasync,
9037 #ifdef CONFIG_PROC_FS
9038 .show_fdinfo = io_uring_show_fdinfo,
9042 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9043 struct io_uring_params *p)
9045 struct io_rings *rings;
9046 size_t size, sq_array_offset;
9048 /* make sure these are sane, as we already accounted them */
9049 ctx->sq_entries = p->sq_entries;
9050 ctx->cq_entries = p->cq_entries;
9052 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9053 if (size == SIZE_MAX)
9056 rings = io_mem_alloc(size);
9061 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9062 rings->sq_ring_mask = p->sq_entries - 1;
9063 rings->cq_ring_mask = p->cq_entries - 1;
9064 rings->sq_ring_entries = p->sq_entries;
9065 rings->cq_ring_entries = p->cq_entries;
9066 ctx->sq_mask = rings->sq_ring_mask;
9067 ctx->cq_mask = rings->cq_ring_mask;
9069 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9070 if (size == SIZE_MAX) {
9071 io_mem_free(ctx->rings);
9076 ctx->sq_sqes = io_mem_alloc(size);
9077 if (!ctx->sq_sqes) {
9078 io_mem_free(ctx->rings);
9087 * Allocate an anonymous fd, this is what constitutes the application
9088 * visible backing of an io_uring instance. The application mmaps this
9089 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9090 * we have to tie this fd to a socket for file garbage collection purposes.
9092 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9097 #if defined(CONFIG_UNIX)
9098 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9104 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9108 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9109 O_RDWR | O_CLOEXEC);
9113 ret = PTR_ERR(file);
9117 #if defined(CONFIG_UNIX)
9118 ctx->ring_sock->file = file;
9120 if (unlikely(io_uring_add_task_file(file))) {
9121 file = ERR_PTR(-ENOMEM);
9124 fd_install(ret, file);
9127 #if defined(CONFIG_UNIX)
9128 sock_release(ctx->ring_sock);
9129 ctx->ring_sock = NULL;
9134 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9135 struct io_uring_params __user *params)
9137 struct user_struct *user = NULL;
9138 struct io_ring_ctx *ctx;
9144 if (entries > IORING_MAX_ENTRIES) {
9145 if (!(p->flags & IORING_SETUP_CLAMP))
9147 entries = IORING_MAX_ENTRIES;
9151 * Use twice as many entries for the CQ ring. It's possible for the
9152 * application to drive a higher depth than the size of the SQ ring,
9153 * since the sqes are only used at submission time. This allows for
9154 * some flexibility in overcommitting a bit. If the application has
9155 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9156 * of CQ ring entries manually.
9158 p->sq_entries = roundup_pow_of_two(entries);
9159 if (p->flags & IORING_SETUP_CQSIZE) {
9161 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9162 * to a power-of-two, if it isn't already. We do NOT impose
9163 * any cq vs sq ring sizing.
9165 if (p->cq_entries < p->sq_entries)
9167 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9168 if (!(p->flags & IORING_SETUP_CLAMP))
9170 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9172 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9174 p->cq_entries = 2 * p->sq_entries;
9177 user = get_uid(current_user());
9178 limit_mem = !capable(CAP_IPC_LOCK);
9181 ret = __io_account_mem(user,
9182 ring_pages(p->sq_entries, p->cq_entries));
9189 ctx = io_ring_ctx_alloc(p);
9192 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9197 ctx->compat = in_compat_syscall();
9199 ctx->creds = get_current_cred();
9201 ctx->sqo_task = get_task_struct(current);
9204 * This is just grabbed for accounting purposes. When a process exits,
9205 * the mm is exited and dropped before the files, hence we need to hang
9206 * on to this mm purely for the purposes of being able to unaccount
9207 * memory (locked/pinned vm). It's not used for anything else.
9209 mmgrab(current->mm);
9210 ctx->mm_account = current->mm;
9212 #ifdef CONFIG_BLK_CGROUP
9214 * The sq thread will belong to the original cgroup it was inited in.
9215 * If the cgroup goes offline (e.g. disabling the io controller), then
9216 * issued bios will be associated with the closest cgroup later in the
9220 ctx->sqo_blkcg_css = blkcg_css();
9221 ret = css_tryget_online(ctx->sqo_blkcg_css);
9224 /* don't init against a dying cgroup, have the user try again */
9225 ctx->sqo_blkcg_css = NULL;
9232 * Account memory _before_ installing the file descriptor. Once
9233 * the descriptor is installed, it can get closed at any time. Also
9234 * do this before hitting the general error path, as ring freeing
9235 * will un-account as well.
9237 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9239 ctx->limit_mem = limit_mem;
9241 ret = io_allocate_scq_urings(ctx, p);
9245 ret = io_sq_offload_create(ctx, p);
9249 if (!(p->flags & IORING_SETUP_R_DISABLED))
9250 io_sq_offload_start(ctx);
9252 memset(&p->sq_off, 0, sizeof(p->sq_off));
9253 p->sq_off.head = offsetof(struct io_rings, sq.head);
9254 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9255 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9256 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9257 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9258 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9259 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9261 memset(&p->cq_off, 0, sizeof(p->cq_off));
9262 p->cq_off.head = offsetof(struct io_rings, cq.head);
9263 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9264 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9265 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9266 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9267 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9268 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9270 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9271 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9272 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9273 IORING_FEAT_POLL_32BITS;
9275 if (copy_to_user(params, p, sizeof(*p))) {
9281 * Install ring fd as the very last thing, so we don't risk someone
9282 * having closed it before we finish setup
9284 ret = io_uring_get_fd(ctx);
9288 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9291 io_ring_ctx_wait_and_kill(ctx);
9296 * Sets up an aio uring context, and returns the fd. Applications asks for a
9297 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9298 * params structure passed in.
9300 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9302 struct io_uring_params p;
9305 if (copy_from_user(&p, params, sizeof(p)))
9307 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9312 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9313 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9314 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9315 IORING_SETUP_R_DISABLED))
9318 return io_uring_create(entries, &p, params);
9321 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9322 struct io_uring_params __user *, params)
9324 return io_uring_setup(entries, params);
9327 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9329 struct io_uring_probe *p;
9333 size = struct_size(p, ops, nr_args);
9334 if (size == SIZE_MAX)
9336 p = kzalloc(size, GFP_KERNEL);
9341 if (copy_from_user(p, arg, size))
9344 if (memchr_inv(p, 0, size))
9347 p->last_op = IORING_OP_LAST - 1;
9348 if (nr_args > IORING_OP_LAST)
9349 nr_args = IORING_OP_LAST;
9351 for (i = 0; i < nr_args; i++) {
9353 if (!io_op_defs[i].not_supported)
9354 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9359 if (copy_to_user(arg, p, size))
9366 static int io_register_personality(struct io_ring_ctx *ctx)
9368 struct io_identity *id;
9371 id = kmalloc(sizeof(*id), GFP_KERNEL);
9375 io_init_identity(id);
9376 id->creds = get_current_cred();
9378 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9380 put_cred(id->creds);
9386 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9388 struct io_identity *iod;
9390 iod = idr_remove(&ctx->personality_idr, id);
9392 put_cred(iod->creds);
9393 if (refcount_dec_and_test(&iod->count))
9401 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9402 unsigned int nr_args)
9404 struct io_uring_restriction *res;
9408 /* Restrictions allowed only if rings started disabled */
9409 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9412 /* We allow only a single restrictions registration */
9413 if (ctx->restrictions.registered)
9416 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9419 size = array_size(nr_args, sizeof(*res));
9420 if (size == SIZE_MAX)
9423 res = memdup_user(arg, size);
9425 return PTR_ERR(res);
9429 for (i = 0; i < nr_args; i++) {
9430 switch (res[i].opcode) {
9431 case IORING_RESTRICTION_REGISTER_OP:
9432 if (res[i].register_op >= IORING_REGISTER_LAST) {
9437 __set_bit(res[i].register_op,
9438 ctx->restrictions.register_op);
9440 case IORING_RESTRICTION_SQE_OP:
9441 if (res[i].sqe_op >= IORING_OP_LAST) {
9446 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9448 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9449 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9451 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9452 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9461 /* Reset all restrictions if an error happened */
9463 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9465 ctx->restrictions.registered = true;
9471 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9473 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9476 if (ctx->restrictions.registered)
9477 ctx->restricted = 1;
9479 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9481 io_sq_offload_start(ctx);
9486 static bool io_register_op_must_quiesce(int op)
9489 case IORING_UNREGISTER_FILES:
9490 case IORING_REGISTER_FILES_UPDATE:
9491 case IORING_REGISTER_PROBE:
9492 case IORING_REGISTER_PERSONALITY:
9493 case IORING_UNREGISTER_PERSONALITY:
9500 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9501 void __user *arg, unsigned nr_args)
9502 __releases(ctx->uring_lock)
9503 __acquires(ctx->uring_lock)
9508 * We're inside the ring mutex, if the ref is already dying, then
9509 * someone else killed the ctx or is already going through
9510 * io_uring_register().
9512 if (percpu_ref_is_dying(&ctx->refs))
9515 if (io_register_op_must_quiesce(opcode)) {
9516 percpu_ref_kill(&ctx->refs);
9519 * Drop uring mutex before waiting for references to exit. If
9520 * another thread is currently inside io_uring_enter() it might
9521 * need to grab the uring_lock to make progress. If we hold it
9522 * here across the drain wait, then we can deadlock. It's safe
9523 * to drop the mutex here, since no new references will come in
9524 * after we've killed the percpu ref.
9526 mutex_unlock(&ctx->uring_lock);
9528 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9531 ret = io_run_task_work_sig();
9536 mutex_lock(&ctx->uring_lock);
9539 percpu_ref_resurrect(&ctx->refs);
9544 if (ctx->restricted) {
9545 if (opcode >= IORING_REGISTER_LAST) {
9550 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9557 case IORING_REGISTER_BUFFERS:
9558 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9560 case IORING_UNREGISTER_BUFFERS:
9564 ret = io_sqe_buffer_unregister(ctx);
9566 case IORING_REGISTER_FILES:
9567 ret = io_sqe_files_register(ctx, arg, nr_args);
9569 case IORING_UNREGISTER_FILES:
9573 ret = io_sqe_files_unregister(ctx);
9575 case IORING_REGISTER_FILES_UPDATE:
9576 ret = io_sqe_files_update(ctx, arg, nr_args);
9578 case IORING_REGISTER_EVENTFD:
9579 case IORING_REGISTER_EVENTFD_ASYNC:
9583 ret = io_eventfd_register(ctx, arg);
9586 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9587 ctx->eventfd_async = 1;
9589 ctx->eventfd_async = 0;
9591 case IORING_UNREGISTER_EVENTFD:
9595 ret = io_eventfd_unregister(ctx);
9597 case IORING_REGISTER_PROBE:
9599 if (!arg || nr_args > 256)
9601 ret = io_probe(ctx, arg, nr_args);
9603 case IORING_REGISTER_PERSONALITY:
9607 ret = io_register_personality(ctx);
9609 case IORING_UNREGISTER_PERSONALITY:
9613 ret = io_unregister_personality(ctx, nr_args);
9615 case IORING_REGISTER_ENABLE_RINGS:
9619 ret = io_register_enable_rings(ctx);
9621 case IORING_REGISTER_RESTRICTIONS:
9622 ret = io_register_restrictions(ctx, arg, nr_args);
9630 if (io_register_op_must_quiesce(opcode)) {
9631 /* bring the ctx back to life */
9632 percpu_ref_reinit(&ctx->refs);
9634 reinit_completion(&ctx->ref_comp);
9639 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9640 void __user *, arg, unsigned int, nr_args)
9642 struct io_ring_ctx *ctx;
9651 if (f.file->f_op != &io_uring_fops)
9654 ctx = f.file->private_data;
9656 mutex_lock(&ctx->uring_lock);
9657 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9658 mutex_unlock(&ctx->uring_lock);
9659 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9660 ctx->cq_ev_fd != NULL, ret);
9666 static int __init io_uring_init(void)
9668 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9669 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9670 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9673 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9674 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9675 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9676 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9677 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9678 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9679 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9680 BUILD_BUG_SQE_ELEM(8, __u64, off);
9681 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9682 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9683 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9684 BUILD_BUG_SQE_ELEM(24, __u32, len);
9685 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9686 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9687 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9688 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9689 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9690 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9691 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9692 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9693 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9694 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9695 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9696 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9697 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9698 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9699 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9700 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9701 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9702 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9703 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9705 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9706 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9707 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9710 __initcall(io_uring_init);