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>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
511 const struct iovec *free_iovec;
512 struct iov_iter iter;
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
544 REQ_F_BUFFER_SELECTED_BIT,
545 REQ_F_NO_FILE_TABLE_BIT,
546 REQ_F_WORK_INITIALIZED_BIT,
547 REQ_F_TASK_PINNED_BIT,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
559 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
563 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
568 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
569 /* fail rest of links */
570 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
571 /* on inflight list */
572 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
573 /* read/write uses file position */
574 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
575 /* must not punt to workers */
576 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* completion under lock */
582 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
584 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
585 /* already went through poll handler */
586 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
587 /* buffer already selected */
588 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
589 /* doesn't need file table for this request */
590 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
591 /* io_wq_work is initialized */
592 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
593 /* req->task is refcounted */
594 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
598 struct io_poll_iocb poll;
599 struct io_poll_iocb *double_poll;
603 * NOTE! Each of the iocb union members has the file pointer
604 * as the first entry in their struct definition. So you can
605 * access the file pointer through any of the sub-structs,
606 * or directly as just 'ki_filp' in this struct.
612 struct io_poll_iocb poll;
613 struct io_accept accept;
615 struct io_cancel cancel;
616 struct io_timeout timeout;
617 struct io_connect connect;
618 struct io_sr_msg sr_msg;
620 struct io_close close;
621 struct io_files_update files_update;
622 struct io_fadvise fadvise;
623 struct io_madvise madvise;
624 struct io_epoll epoll;
625 struct io_splice splice;
626 struct io_provide_buf pbuf;
627 struct io_statx statx;
628 /* use only after cleaning per-op data, see io_clean_op() */
629 struct io_completion compl;
632 struct io_async_ctx *io;
634 /* polled IO has completed */
640 struct io_ring_ctx *ctx;
643 struct task_struct *task;
646 struct list_head link_list;
649 * 1. used with ctx->iopoll_list with reads/writes
650 * 2. to track reqs with ->files (see io_op_def::file_table)
652 struct list_head inflight_entry;
654 struct percpu_ref *fixed_file_refs;
655 struct callback_head task_work;
656 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
657 struct hlist_node hash_node;
658 struct async_poll *apoll;
659 struct io_wq_work work;
662 struct io_defer_entry {
663 struct list_head list;
664 struct io_kiocb *req;
668 #define IO_IOPOLL_BATCH 8
670 struct io_comp_state {
672 struct list_head list;
673 struct io_ring_ctx *ctx;
676 struct io_submit_state {
677 struct blk_plug plug;
680 * io_kiocb alloc cache
682 void *reqs[IO_IOPOLL_BATCH];
683 unsigned int free_reqs;
686 * Batch completion logic
688 struct io_comp_state comp;
691 * File reference cache
695 unsigned int has_refs;
696 unsigned int ios_left;
700 /* needs req->io allocated for deferral/async */
701 unsigned async_ctx : 1;
702 /* needs current->mm setup, does mm access */
703 unsigned needs_mm : 1;
704 /* needs req->file assigned */
705 unsigned needs_file : 1;
706 /* don't fail if file grab fails */
707 unsigned needs_file_no_error : 1;
708 /* hash wq insertion if file is a regular file */
709 unsigned hash_reg_file : 1;
710 /* unbound wq insertion if file is a non-regular file */
711 unsigned unbound_nonreg_file : 1;
712 /* opcode is not supported by this kernel */
713 unsigned not_supported : 1;
714 /* needs file table */
715 unsigned file_table : 1;
717 unsigned needs_fs : 1;
718 /* set if opcode supports polled "wait" */
720 unsigned pollout : 1;
721 /* op supports buffer selection */
722 unsigned buffer_select : 1;
723 unsigned needs_fsize : 1;
726 static const struct io_op_def io_op_defs[] = {
727 [IORING_OP_NOP] = {},
728 [IORING_OP_READV] = {
732 .unbound_nonreg_file = 1,
736 [IORING_OP_WRITEV] = {
741 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_ADD] = {
762 .unbound_nonreg_file = 1,
764 [IORING_OP_POLL_REMOVE] = {},
765 [IORING_OP_SYNC_FILE_RANGE] = {
768 [IORING_OP_SENDMSG] = {
772 .unbound_nonreg_file = 1,
776 [IORING_OP_RECVMSG] = {
780 .unbound_nonreg_file = 1,
785 [IORING_OP_TIMEOUT] = {
789 [IORING_OP_TIMEOUT_REMOVE] = {},
790 [IORING_OP_ACCEPT] = {
793 .unbound_nonreg_file = 1,
797 [IORING_OP_ASYNC_CANCEL] = {},
798 [IORING_OP_LINK_TIMEOUT] = {
802 [IORING_OP_CONNECT] = {
806 .unbound_nonreg_file = 1,
809 [IORING_OP_FALLOCATE] = {
813 [IORING_OP_OPENAT] = {
817 [IORING_OP_CLOSE] = {
819 .needs_file_no_error = 1,
822 [IORING_OP_FILES_UPDATE] = {
826 [IORING_OP_STATX] = {
834 .unbound_nonreg_file = 1,
838 [IORING_OP_WRITE] = {
841 .unbound_nonreg_file = 1,
845 [IORING_OP_FADVISE] = {
848 [IORING_OP_MADVISE] = {
854 .unbound_nonreg_file = 1,
860 .unbound_nonreg_file = 1,
864 [IORING_OP_OPENAT2] = {
868 [IORING_OP_EPOLL_CTL] = {
869 .unbound_nonreg_file = 1,
872 [IORING_OP_SPLICE] = {
875 .unbound_nonreg_file = 1,
877 [IORING_OP_PROVIDE_BUFFERS] = {},
878 [IORING_OP_REMOVE_BUFFERS] = {},
882 .unbound_nonreg_file = 1,
886 enum io_mem_account {
891 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
892 struct io_comp_state *cs);
893 static void io_cqring_fill_event(struct io_kiocb *req, long res);
894 static void io_put_req(struct io_kiocb *req);
895 static void io_double_put_req(struct io_kiocb *req);
896 static void __io_double_put_req(struct io_kiocb *req);
897 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
898 static void __io_queue_linked_timeout(struct io_kiocb *req);
899 static void io_queue_linked_timeout(struct io_kiocb *req);
900 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
901 struct io_uring_files_update *ip,
903 static int io_prep_work_files(struct io_kiocb *req);
904 static void __io_clean_op(struct io_kiocb *req);
905 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
906 int fd, struct file **out_file, bool fixed);
907 static void __io_queue_sqe(struct io_kiocb *req,
908 const struct io_uring_sqe *sqe,
909 struct io_comp_state *cs);
910 static void io_file_put_work(struct work_struct *work);
912 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
913 struct iovec **iovec, struct iov_iter *iter,
915 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
916 const struct iovec *fast_iov,
917 struct iov_iter *iter, bool force);
919 static struct kmem_cache *req_cachep;
921 static const struct file_operations io_uring_fops;
923 struct sock *io_uring_get_socket(struct file *file)
925 #if defined(CONFIG_UNIX)
926 if (file->f_op == &io_uring_fops) {
927 struct io_ring_ctx *ctx = file->private_data;
929 return ctx->ring_sock->sk;
934 EXPORT_SYMBOL(io_uring_get_socket);
936 static void io_get_req_task(struct io_kiocb *req)
938 if (req->flags & REQ_F_TASK_PINNED)
940 get_task_struct(req->task);
941 req->flags |= REQ_F_TASK_PINNED;
944 static inline void io_clean_op(struct io_kiocb *req)
946 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
951 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
952 static void __io_put_req_task(struct io_kiocb *req)
954 if (req->flags & REQ_F_TASK_PINNED)
955 put_task_struct(req->task);
958 static void io_sq_thread_drop_mm(void)
960 struct mm_struct *mm = current->mm;
963 kthread_unuse_mm(mm);
968 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
971 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
972 !mmget_not_zero(ctx->sqo_mm)))
974 kthread_use_mm(ctx->sqo_mm);
980 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
981 struct io_kiocb *req)
983 if (!io_op_defs[req->opcode].needs_mm)
985 return __io_sq_thread_acquire_mm(ctx);
988 static inline void req_set_fail_links(struct io_kiocb *req)
990 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
991 req->flags |= REQ_F_FAIL_LINK;
995 * Note: must call io_req_init_async() for the first time you
996 * touch any members of io_wq_work.
998 static inline void io_req_init_async(struct io_kiocb *req)
1000 if (req->flags & REQ_F_WORK_INITIALIZED)
1003 memset(&req->work, 0, sizeof(req->work));
1004 req->flags |= REQ_F_WORK_INITIALIZED;
1007 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1009 return ctx->flags & IORING_SETUP_SQPOLL;
1012 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1014 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1016 complete(&ctx->ref_comp);
1019 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1021 return !req->timeout.off;
1024 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1026 struct io_ring_ctx *ctx;
1029 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1033 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1034 if (!ctx->fallback_req)
1038 * Use 5 bits less than the max cq entries, that should give us around
1039 * 32 entries per hash list if totally full and uniformly spread.
1041 hash_bits = ilog2(p->cq_entries);
1045 ctx->cancel_hash_bits = hash_bits;
1046 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1048 if (!ctx->cancel_hash)
1050 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1052 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1053 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1056 ctx->flags = p->flags;
1057 init_waitqueue_head(&ctx->sqo_wait);
1058 init_waitqueue_head(&ctx->cq_wait);
1059 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1060 init_completion(&ctx->ref_comp);
1061 init_completion(&ctx->sq_thread_comp);
1062 idr_init(&ctx->io_buffer_idr);
1063 idr_init(&ctx->personality_idr);
1064 mutex_init(&ctx->uring_lock);
1065 init_waitqueue_head(&ctx->wait);
1066 spin_lock_init(&ctx->completion_lock);
1067 INIT_LIST_HEAD(&ctx->iopoll_list);
1068 INIT_LIST_HEAD(&ctx->defer_list);
1069 INIT_LIST_HEAD(&ctx->timeout_list);
1070 init_waitqueue_head(&ctx->inflight_wait);
1071 spin_lock_init(&ctx->inflight_lock);
1072 INIT_LIST_HEAD(&ctx->inflight_list);
1073 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1074 init_llist_head(&ctx->file_put_llist);
1077 if (ctx->fallback_req)
1078 kmem_cache_free(req_cachep, ctx->fallback_req);
1079 kfree(ctx->cancel_hash);
1084 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1086 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1087 struct io_ring_ctx *ctx = req->ctx;
1089 return seq != ctx->cached_cq_tail
1090 + atomic_read(&ctx->cached_cq_overflow);
1096 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1098 struct io_rings *rings = ctx->rings;
1100 /* order cqe stores with ring update */
1101 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1103 if (wq_has_sleeper(&ctx->cq_wait)) {
1104 wake_up_interruptible(&ctx->cq_wait);
1105 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1110 * Returns true if we need to defer file table putting. This can only happen
1111 * from the error path with REQ_F_COMP_LOCKED set.
1113 static bool io_req_clean_work(struct io_kiocb *req)
1115 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1118 req->flags &= ~REQ_F_WORK_INITIALIZED;
1121 mmdrop(req->work.mm);
1122 req->work.mm = NULL;
1124 if (req->work.creds) {
1125 put_cred(req->work.creds);
1126 req->work.creds = NULL;
1129 struct fs_struct *fs = req->work.fs;
1131 if (req->flags & REQ_F_COMP_LOCKED)
1134 spin_lock(&req->work.fs->lock);
1137 spin_unlock(&req->work.fs->lock);
1140 req->work.fs = NULL;
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1202 static void io_queue_async_work(struct io_kiocb *req)
1204 struct io_kiocb *link;
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 link = __io_queue_async_work(req);
1211 io_queue_linked_timeout(link);
1214 static void io_kill_timeout(struct io_kiocb *req)
1218 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1220 atomic_set(&req->ctx->cq_timeouts,
1221 atomic_read(&req->ctx->cq_timeouts) + 1);
1222 list_del_init(&req->timeout.list);
1223 req->flags |= REQ_F_COMP_LOCKED;
1224 io_cqring_fill_event(req, 0);
1229 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1231 struct io_kiocb *req, *tmp;
1233 spin_lock_irq(&ctx->completion_lock);
1234 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1235 io_kill_timeout(req);
1236 spin_unlock_irq(&ctx->completion_lock);
1239 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1242 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1243 struct io_defer_entry, list);
1244 struct io_kiocb *link;
1246 if (req_need_defer(de->req, de->seq))
1248 list_del_init(&de->list);
1249 /* punt-init is done before queueing for defer */
1250 link = __io_queue_async_work(de->req);
1252 __io_queue_linked_timeout(link);
1253 /* drop submission reference */
1254 link->flags |= REQ_F_COMP_LOCKED;
1258 } while (!list_empty(&ctx->defer_list));
1261 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1263 while (!list_empty(&ctx->timeout_list)) {
1264 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1265 struct io_kiocb, timeout.list);
1267 if (io_is_timeout_noseq(req))
1269 if (req->timeout.target_seq != ctx->cached_cq_tail
1270 - atomic_read(&ctx->cq_timeouts))
1273 list_del_init(&req->timeout.list);
1274 io_kill_timeout(req);
1278 static void io_commit_cqring(struct io_ring_ctx *ctx)
1280 io_flush_timeouts(ctx);
1281 __io_commit_cqring(ctx);
1283 if (unlikely(!list_empty(&ctx->defer_list)))
1284 __io_queue_deferred(ctx);
1287 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1289 struct io_rings *rings = ctx->rings;
1292 tail = ctx->cached_cq_tail;
1294 * writes to the cq entry need to come after reading head; the
1295 * control dependency is enough as we're using WRITE_ONCE to
1298 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1301 ctx->cached_cq_tail++;
1302 return &rings->cqes[tail & ctx->cq_mask];
1305 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1309 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1311 if (!ctx->eventfd_async)
1313 return io_wq_current_is_worker();
1316 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1318 if (waitqueue_active(&ctx->wait))
1319 wake_up(&ctx->wait);
1320 if (waitqueue_active(&ctx->sqo_wait))
1321 wake_up(&ctx->sqo_wait);
1322 if (io_should_trigger_evfd(ctx))
1323 eventfd_signal(ctx->cq_ev_fd, 1);
1326 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1328 if (list_empty(&ctx->cq_overflow_list)) {
1329 clear_bit(0, &ctx->sq_check_overflow);
1330 clear_bit(0, &ctx->cq_check_overflow);
1331 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1335 /* Returns true if there are no backlogged entries after the flush */
1336 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1338 struct io_rings *rings = ctx->rings;
1339 struct io_uring_cqe *cqe;
1340 struct io_kiocb *req;
1341 unsigned long flags;
1345 if (list_empty_careful(&ctx->cq_overflow_list))
1347 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1348 rings->cq_ring_entries))
1352 spin_lock_irqsave(&ctx->completion_lock, flags);
1354 /* if force is set, the ring is going away. always drop after that */
1356 ctx->cq_overflow_flushed = 1;
1359 while (!list_empty(&ctx->cq_overflow_list)) {
1360 cqe = io_get_cqring(ctx);
1364 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1366 list_move(&req->compl.list, &list);
1368 WRITE_ONCE(cqe->user_data, req->user_data);
1369 WRITE_ONCE(cqe->res, req->result);
1370 WRITE_ONCE(cqe->flags, req->compl.cflags);
1372 WRITE_ONCE(ctx->rings->cq_overflow,
1373 atomic_inc_return(&ctx->cached_cq_overflow));
1377 io_commit_cqring(ctx);
1378 io_cqring_mark_overflow(ctx);
1380 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1381 io_cqring_ev_posted(ctx);
1383 while (!list_empty(&list)) {
1384 req = list_first_entry(&list, struct io_kiocb, compl.list);
1385 list_del(&req->compl.list);
1392 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1394 struct io_ring_ctx *ctx = req->ctx;
1395 struct io_uring_cqe *cqe;
1397 trace_io_uring_complete(ctx, req->user_data, res);
1400 * If we can't get a cq entry, userspace overflowed the
1401 * submission (by quite a lot). Increment the overflow count in
1404 cqe = io_get_cqring(ctx);
1406 WRITE_ONCE(cqe->user_data, req->user_data);
1407 WRITE_ONCE(cqe->res, res);
1408 WRITE_ONCE(cqe->flags, cflags);
1409 } else if (ctx->cq_overflow_flushed) {
1410 WRITE_ONCE(ctx->rings->cq_overflow,
1411 atomic_inc_return(&ctx->cached_cq_overflow));
1413 if (list_empty(&ctx->cq_overflow_list)) {
1414 set_bit(0, &ctx->sq_check_overflow);
1415 set_bit(0, &ctx->cq_check_overflow);
1416 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1420 req->compl.cflags = cflags;
1421 refcount_inc(&req->refs);
1422 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1426 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1428 __io_cqring_fill_event(req, res, 0);
1431 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1433 struct io_ring_ctx *ctx = req->ctx;
1434 unsigned long flags;
1436 spin_lock_irqsave(&ctx->completion_lock, flags);
1437 __io_cqring_fill_event(req, res, cflags);
1438 io_commit_cqring(ctx);
1439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1441 io_cqring_ev_posted(ctx);
1444 static void io_submit_flush_completions(struct io_comp_state *cs)
1446 struct io_ring_ctx *ctx = cs->ctx;
1448 spin_lock_irq(&ctx->completion_lock);
1449 while (!list_empty(&cs->list)) {
1450 struct io_kiocb *req;
1452 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1453 list_del(&req->compl.list);
1454 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1455 if (!(req->flags & REQ_F_LINK_HEAD)) {
1456 req->flags |= REQ_F_COMP_LOCKED;
1459 spin_unlock_irq(&ctx->completion_lock);
1461 spin_lock_irq(&ctx->completion_lock);
1464 io_commit_cqring(ctx);
1465 spin_unlock_irq(&ctx->completion_lock);
1467 io_cqring_ev_posted(ctx);
1471 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1472 struct io_comp_state *cs)
1475 io_cqring_add_event(req, res, cflags);
1480 req->compl.cflags = cflags;
1481 list_add_tail(&req->compl.list, &cs->list);
1483 io_submit_flush_completions(cs);
1487 static void io_req_complete(struct io_kiocb *req, long res)
1489 __io_req_complete(req, res, 0, NULL);
1492 static inline bool io_is_fallback_req(struct io_kiocb *req)
1494 return req == (struct io_kiocb *)
1495 ((unsigned long) req->ctx->fallback_req & ~1UL);
1498 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1500 struct io_kiocb *req;
1502 req = ctx->fallback_req;
1503 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1509 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1510 struct io_submit_state *state)
1512 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1513 struct io_kiocb *req;
1515 if (!state->free_reqs) {
1519 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1520 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1523 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1524 * retry single alloc to be on the safe side.
1526 if (unlikely(ret <= 0)) {
1527 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1528 if (!state->reqs[0])
1532 state->free_reqs = ret - 1;
1533 req = state->reqs[ret - 1];
1536 req = state->reqs[state->free_reqs];
1541 return io_get_fallback_req(ctx);
1544 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1548 percpu_ref_put(req->fixed_file_refs);
1553 static bool io_dismantle_req(struct io_kiocb *req)
1560 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1562 return io_req_clean_work(req);
1565 static void __io_free_req_finish(struct io_kiocb *req)
1567 struct io_ring_ctx *ctx = req->ctx;
1569 __io_put_req_task(req);
1570 if (likely(!io_is_fallback_req(req)))
1571 kmem_cache_free(req_cachep, req);
1573 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1574 percpu_ref_put(&ctx->refs);
1577 static void io_req_task_file_table_put(struct callback_head *cb)
1579 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1580 struct fs_struct *fs = req->work.fs;
1582 spin_lock(&req->work.fs->lock);
1585 spin_unlock(&req->work.fs->lock);
1588 req->work.fs = NULL;
1589 __io_free_req_finish(req);
1592 static void __io_free_req(struct io_kiocb *req)
1594 if (!io_dismantle_req(req)) {
1595 __io_free_req_finish(req);
1599 init_task_work(&req->task_work, io_req_task_file_table_put);
1600 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1601 if (unlikely(ret)) {
1602 struct task_struct *tsk;
1604 tsk = io_wq_get_task(req->ctx->io_wq);
1605 task_work_add(tsk, &req->task_work, 0);
1610 static bool io_link_cancel_timeout(struct io_kiocb *req)
1612 struct io_ring_ctx *ctx = req->ctx;
1615 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1617 io_cqring_fill_event(req, -ECANCELED);
1618 io_commit_cqring(ctx);
1619 req->flags &= ~REQ_F_LINK_HEAD;
1627 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1629 struct io_kiocb *link;
1632 if (list_empty(&req->link_list))
1634 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1635 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1638 list_del_init(&link->link_list);
1639 link->flags |= REQ_F_COMP_LOCKED;
1640 wake_ev = io_link_cancel_timeout(link);
1641 req->flags &= ~REQ_F_LINK_TIMEOUT;
1645 static void io_kill_linked_timeout(struct io_kiocb *req)
1647 struct io_ring_ctx *ctx = req->ctx;
1650 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1651 unsigned long flags;
1653 spin_lock_irqsave(&ctx->completion_lock, flags);
1654 wake_ev = __io_kill_linked_timeout(req);
1655 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1657 wake_ev = __io_kill_linked_timeout(req);
1661 io_cqring_ev_posted(ctx);
1664 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1666 struct io_kiocb *nxt;
1669 * The list should never be empty when we are called here. But could
1670 * potentially happen if the chain is messed up, check to be on the
1673 if (unlikely(list_empty(&req->link_list)))
1676 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1677 list_del_init(&req->link_list);
1678 if (!list_empty(&nxt->link_list))
1679 nxt->flags |= REQ_F_LINK_HEAD;
1684 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1686 static void __io_fail_links(struct io_kiocb *req)
1688 struct io_ring_ctx *ctx = req->ctx;
1690 while (!list_empty(&req->link_list)) {
1691 struct io_kiocb *link = list_first_entry(&req->link_list,
1692 struct io_kiocb, link_list);
1694 list_del_init(&link->link_list);
1695 trace_io_uring_fail_link(req, link);
1697 io_cqring_fill_event(link, -ECANCELED);
1698 link->flags |= REQ_F_COMP_LOCKED;
1699 __io_double_put_req(link);
1700 req->flags &= ~REQ_F_LINK_TIMEOUT;
1703 io_commit_cqring(ctx);
1704 io_cqring_ev_posted(ctx);
1707 static void io_fail_links(struct io_kiocb *req)
1709 struct io_ring_ctx *ctx = req->ctx;
1711 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1712 unsigned long flags;
1714 spin_lock_irqsave(&ctx->completion_lock, flags);
1715 __io_fail_links(req);
1716 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1718 __io_fail_links(req);
1721 io_cqring_ev_posted(ctx);
1724 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1726 req->flags &= ~REQ_F_LINK_HEAD;
1727 if (req->flags & REQ_F_LINK_TIMEOUT)
1728 io_kill_linked_timeout(req);
1731 * If LINK is set, we have dependent requests in this chain. If we
1732 * didn't fail this request, queue the first one up, moving any other
1733 * dependencies to the next request. In case of failure, fail the rest
1736 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1737 return io_req_link_next(req);
1742 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1744 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1746 return __io_req_find_next(req);
1749 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1752 struct task_struct *tsk = req->task;
1753 struct io_ring_ctx *ctx = req->ctx;
1756 if (tsk->flags & PF_EXITING)
1760 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1761 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1762 * processing task_work. There's no reliable way to tell if TWA_RESUME
1766 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1767 notify = TWA_SIGNAL;
1769 ret = task_work_add(tsk, cb, notify);
1771 wake_up_process(tsk);
1776 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1778 struct io_ring_ctx *ctx = req->ctx;
1780 spin_lock_irq(&ctx->completion_lock);
1781 io_cqring_fill_event(req, error);
1782 io_commit_cqring(ctx);
1783 spin_unlock_irq(&ctx->completion_lock);
1785 io_cqring_ev_posted(ctx);
1786 req_set_fail_links(req);
1787 io_double_put_req(req);
1790 static void io_req_task_cancel(struct callback_head *cb)
1792 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1793 struct io_ring_ctx *ctx = req->ctx;
1795 __io_req_task_cancel(req, -ECANCELED);
1796 percpu_ref_put(&ctx->refs);
1799 static void __io_req_task_submit(struct io_kiocb *req)
1801 struct io_ring_ctx *ctx = req->ctx;
1803 if (!__io_sq_thread_acquire_mm(ctx)) {
1804 mutex_lock(&ctx->uring_lock);
1805 __io_queue_sqe(req, NULL, NULL);
1806 mutex_unlock(&ctx->uring_lock);
1808 __io_req_task_cancel(req, -EFAULT);
1812 static void io_req_task_submit(struct callback_head *cb)
1814 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1815 struct io_ring_ctx *ctx = req->ctx;
1817 __io_req_task_submit(req);
1818 percpu_ref_put(&ctx->refs);
1821 static void io_req_task_queue(struct io_kiocb *req)
1825 init_task_work(&req->task_work, io_req_task_submit);
1826 percpu_ref_get(&req->ctx->refs);
1828 ret = io_req_task_work_add(req, &req->task_work, true);
1829 if (unlikely(ret)) {
1830 struct task_struct *tsk;
1832 init_task_work(&req->task_work, io_req_task_cancel);
1833 tsk = io_wq_get_task(req->ctx->io_wq);
1834 task_work_add(tsk, &req->task_work, 0);
1835 wake_up_process(tsk);
1839 static void io_queue_next(struct io_kiocb *req)
1841 struct io_kiocb *nxt = io_req_find_next(req);
1844 io_req_task_queue(nxt);
1847 static void io_free_req(struct io_kiocb *req)
1854 void *reqs[IO_IOPOLL_BATCH];
1857 struct task_struct *task;
1861 static inline void io_init_req_batch(struct req_batch *rb)
1868 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1869 struct req_batch *rb)
1871 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1872 percpu_ref_put_many(&ctx->refs, rb->to_free);
1876 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1877 struct req_batch *rb)
1880 __io_req_free_batch_flush(ctx, rb);
1882 put_task_struct_many(rb->task, rb->task_refs);
1887 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1889 if (unlikely(io_is_fallback_req(req))) {
1893 if (req->flags & REQ_F_LINK_HEAD)
1896 if (req->flags & REQ_F_TASK_PINNED) {
1897 if (req->task != rb->task) {
1899 put_task_struct_many(rb->task, rb->task_refs);
1900 rb->task = req->task;
1904 req->flags &= ~REQ_F_TASK_PINNED;
1907 WARN_ON_ONCE(io_dismantle_req(req));
1908 rb->reqs[rb->to_free++] = req;
1909 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1910 __io_req_free_batch_flush(req->ctx, rb);
1914 * Drop reference to request, return next in chain (if there is one) if this
1915 * was the last reference to this request.
1917 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1919 struct io_kiocb *nxt = NULL;
1921 if (refcount_dec_and_test(&req->refs)) {
1922 nxt = io_req_find_next(req);
1928 static void io_put_req(struct io_kiocb *req)
1930 if (refcount_dec_and_test(&req->refs))
1934 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1936 struct io_kiocb *nxt;
1939 * A ref is owned by io-wq in which context we're. So, if that's the
1940 * last one, it's safe to steal next work. False negatives are Ok,
1941 * it just will be re-punted async in io_put_work()
1943 if (refcount_read(&req->refs) != 1)
1946 nxt = io_req_find_next(req);
1947 return nxt ? &nxt->work : NULL;
1951 * Must only be used if we don't need to care about links, usually from
1952 * within the completion handling itself.
1954 static void __io_double_put_req(struct io_kiocb *req)
1956 /* drop both submit and complete references */
1957 if (refcount_sub_and_test(2, &req->refs))
1961 static void io_double_put_req(struct io_kiocb *req)
1963 /* drop both submit and complete references */
1964 if (refcount_sub_and_test(2, &req->refs))
1968 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1970 struct io_rings *rings = ctx->rings;
1972 if (test_bit(0, &ctx->cq_check_overflow)) {
1974 * noflush == true is from the waitqueue handler, just ensure
1975 * we wake up the task, and the next invocation will flush the
1976 * entries. We cannot safely to it from here.
1978 if (noflush && !list_empty(&ctx->cq_overflow_list))
1981 io_cqring_overflow_flush(ctx, false);
1984 /* See comment at the top of this file */
1986 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1989 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1991 struct io_rings *rings = ctx->rings;
1993 /* make sure SQ entry isn't read before tail */
1994 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1997 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1999 unsigned int cflags;
2001 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2002 cflags |= IORING_CQE_F_BUFFER;
2003 req->flags &= ~REQ_F_BUFFER_SELECTED;
2008 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2010 struct io_buffer *kbuf;
2012 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2013 return io_put_kbuf(req, kbuf);
2016 static inline bool io_run_task_work(void)
2019 * Not safe to run on exiting task, and the task_work handling will
2020 * not add work to such a task.
2022 if (unlikely(current->flags & PF_EXITING))
2024 if (current->task_works) {
2025 __set_current_state(TASK_RUNNING);
2033 static void io_iopoll_queue(struct list_head *again)
2035 struct io_kiocb *req;
2038 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2039 list_del(&req->inflight_entry);
2040 __io_complete_rw(req, -EAGAIN, 0, NULL);
2041 } while (!list_empty(again));
2045 * Find and free completed poll iocbs
2047 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2048 struct list_head *done)
2050 struct req_batch rb;
2051 struct io_kiocb *req;
2054 /* order with ->result store in io_complete_rw_iopoll() */
2057 io_init_req_batch(&rb);
2058 while (!list_empty(done)) {
2061 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2062 if (READ_ONCE(req->result) == -EAGAIN) {
2064 req->iopoll_completed = 0;
2065 list_move_tail(&req->inflight_entry, &again);
2068 list_del(&req->inflight_entry);
2070 if (req->flags & REQ_F_BUFFER_SELECTED)
2071 cflags = io_put_rw_kbuf(req);
2073 __io_cqring_fill_event(req, req->result, cflags);
2076 if (refcount_dec_and_test(&req->refs))
2077 io_req_free_batch(&rb, req);
2080 io_commit_cqring(ctx);
2081 if (ctx->flags & IORING_SETUP_SQPOLL)
2082 io_cqring_ev_posted(ctx);
2083 io_req_free_batch_finish(ctx, &rb);
2085 if (!list_empty(&again))
2086 io_iopoll_queue(&again);
2089 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2092 struct io_kiocb *req, *tmp;
2098 * Only spin for completions if we don't have multiple devices hanging
2099 * off our complete list, and we're under the requested amount.
2101 spin = !ctx->poll_multi_file && *nr_events < min;
2104 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2105 struct kiocb *kiocb = &req->rw.kiocb;
2108 * Move completed and retryable entries to our local lists.
2109 * If we find a request that requires polling, break out
2110 * and complete those lists first, if we have entries there.
2112 if (READ_ONCE(req->iopoll_completed)) {
2113 list_move_tail(&req->inflight_entry, &done);
2116 if (!list_empty(&done))
2119 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2123 /* iopoll may have completed current req */
2124 if (READ_ONCE(req->iopoll_completed))
2125 list_move_tail(&req->inflight_entry, &done);
2132 if (!list_empty(&done))
2133 io_iopoll_complete(ctx, nr_events, &done);
2139 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2140 * non-spinning poll check - we'll still enter the driver poll loop, but only
2141 * as a non-spinning completion check.
2143 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2146 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2149 ret = io_do_iopoll(ctx, nr_events, min);
2152 if (*nr_events >= min)
2160 * We can't just wait for polled events to come to us, we have to actively
2161 * find and complete them.
2163 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2165 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2168 mutex_lock(&ctx->uring_lock);
2169 while (!list_empty(&ctx->iopoll_list)) {
2170 unsigned int nr_events = 0;
2172 io_do_iopoll(ctx, &nr_events, 0);
2174 /* let it sleep and repeat later if can't complete a request */
2178 * Ensure we allow local-to-the-cpu processing to take place,
2179 * in this case we need to ensure that we reap all events.
2180 * Also let task_work, etc. to progress by releasing the mutex
2182 if (need_resched()) {
2183 mutex_unlock(&ctx->uring_lock);
2185 mutex_lock(&ctx->uring_lock);
2188 mutex_unlock(&ctx->uring_lock);
2191 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2193 unsigned int nr_events = 0;
2194 int iters = 0, ret = 0;
2197 * We disallow the app entering submit/complete with polling, but we
2198 * still need to lock the ring to prevent racing with polled issue
2199 * that got punted to a workqueue.
2201 mutex_lock(&ctx->uring_lock);
2204 * Don't enter poll loop if we already have events pending.
2205 * If we do, we can potentially be spinning for commands that
2206 * already triggered a CQE (eg in error).
2208 if (io_cqring_events(ctx, false))
2212 * If a submit got punted to a workqueue, we can have the
2213 * application entering polling for a command before it gets
2214 * issued. That app will hold the uring_lock for the duration
2215 * of the poll right here, so we need to take a breather every
2216 * now and then to ensure that the issue has a chance to add
2217 * the poll to the issued list. Otherwise we can spin here
2218 * forever, while the workqueue is stuck trying to acquire the
2221 if (!(++iters & 7)) {
2222 mutex_unlock(&ctx->uring_lock);
2224 mutex_lock(&ctx->uring_lock);
2227 ret = io_iopoll_getevents(ctx, &nr_events, min);
2231 } while (min && !nr_events && !need_resched());
2233 mutex_unlock(&ctx->uring_lock);
2237 static void kiocb_end_write(struct io_kiocb *req)
2240 * Tell lockdep we inherited freeze protection from submission
2243 if (req->flags & REQ_F_ISREG) {
2244 struct inode *inode = file_inode(req->file);
2246 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2248 file_end_write(req->file);
2251 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2252 struct io_comp_state *cs)
2254 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2257 if (kiocb->ki_flags & IOCB_WRITE)
2258 kiocb_end_write(req);
2260 if (res != req->result)
2261 req_set_fail_links(req);
2262 if (req->flags & REQ_F_BUFFER_SELECTED)
2263 cflags = io_put_rw_kbuf(req);
2264 __io_req_complete(req, res, cflags, cs);
2268 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2270 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2271 ssize_t ret = -ECANCELED;
2272 struct iov_iter iter;
2280 switch (req->opcode) {
2281 case IORING_OP_READV:
2282 case IORING_OP_READ_FIXED:
2283 case IORING_OP_READ:
2286 case IORING_OP_WRITEV:
2287 case IORING_OP_WRITE_FIXED:
2288 case IORING_OP_WRITE:
2292 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2298 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2301 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2309 req_set_fail_links(req);
2310 io_req_complete(req, ret);
2315 static bool io_rw_reissue(struct io_kiocb *req, long res)
2318 umode_t mode = file_inode(req->file)->i_mode;
2321 if (!S_ISBLK(mode) && !S_ISREG(mode))
2323 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2326 ret = io_sq_thread_acquire_mm(req->ctx, req);
2328 if (io_resubmit_prep(req, ret)) {
2329 refcount_inc(&req->refs);
2330 io_queue_async_work(req);
2338 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2339 struct io_comp_state *cs)
2341 if (!io_rw_reissue(req, res))
2342 io_complete_rw_common(&req->rw.kiocb, res, cs);
2345 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2347 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2349 __io_complete_rw(req, res, res2, NULL);
2352 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2354 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2356 if (kiocb->ki_flags & IOCB_WRITE)
2357 kiocb_end_write(req);
2359 if (res != -EAGAIN && res != req->result)
2360 req_set_fail_links(req);
2362 WRITE_ONCE(req->result, res);
2363 /* order with io_poll_complete() checking ->result */
2365 WRITE_ONCE(req->iopoll_completed, 1);
2369 * After the iocb has been issued, it's safe to be found on the poll list.
2370 * Adding the kiocb to the list AFTER submission ensures that we don't
2371 * find it from a io_iopoll_getevents() thread before the issuer is done
2372 * accessing the kiocb cookie.
2374 static void io_iopoll_req_issued(struct io_kiocb *req)
2376 struct io_ring_ctx *ctx = req->ctx;
2379 * Track whether we have multiple files in our lists. This will impact
2380 * how we do polling eventually, not spinning if we're on potentially
2381 * different devices.
2383 if (list_empty(&ctx->iopoll_list)) {
2384 ctx->poll_multi_file = false;
2385 } else if (!ctx->poll_multi_file) {
2386 struct io_kiocb *list_req;
2388 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2390 if (list_req->file != req->file)
2391 ctx->poll_multi_file = true;
2395 * For fast devices, IO may have already completed. If it has, add
2396 * it to the front so we find it first.
2398 if (READ_ONCE(req->iopoll_completed))
2399 list_add(&req->inflight_entry, &ctx->iopoll_list);
2401 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2403 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2404 wq_has_sleeper(&ctx->sqo_wait))
2405 wake_up(&ctx->sqo_wait);
2408 static void __io_state_file_put(struct io_submit_state *state)
2410 if (state->has_refs)
2411 fput_many(state->file, state->has_refs);
2415 static inline void io_state_file_put(struct io_submit_state *state)
2418 __io_state_file_put(state);
2422 * Get as many references to a file as we have IOs left in this submission,
2423 * assuming most submissions are for one file, or at least that each file
2424 * has more than one submission.
2426 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2432 if (state->fd == fd) {
2437 __io_state_file_put(state);
2439 state->file = fget_many(fd, state->ios_left);
2445 state->has_refs = state->ios_left;
2449 static bool io_bdev_nowait(struct block_device *bdev)
2452 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2459 * If we tracked the file through the SCM inflight mechanism, we could support
2460 * any file. For now, just ensure that anything potentially problematic is done
2463 static bool io_file_supports_async(struct file *file, int rw)
2465 umode_t mode = file_inode(file)->i_mode;
2467 if (S_ISBLK(mode)) {
2468 if (io_bdev_nowait(file->f_inode->i_bdev))
2472 if (S_ISCHR(mode) || S_ISSOCK(mode))
2474 if (S_ISREG(mode)) {
2475 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2476 file->f_op != &io_uring_fops)
2481 /* any ->read/write should understand O_NONBLOCK */
2482 if (file->f_flags & O_NONBLOCK)
2485 if (!(file->f_mode & FMODE_NOWAIT))
2489 return file->f_op->read_iter != NULL;
2491 return file->f_op->write_iter != NULL;
2494 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2495 bool force_nonblock)
2497 struct io_ring_ctx *ctx = req->ctx;
2498 struct kiocb *kiocb = &req->rw.kiocb;
2502 if (S_ISREG(file_inode(req->file)->i_mode))
2503 req->flags |= REQ_F_ISREG;
2505 kiocb->ki_pos = READ_ONCE(sqe->off);
2506 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2507 req->flags |= REQ_F_CUR_POS;
2508 kiocb->ki_pos = req->file->f_pos;
2510 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2511 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2512 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2516 ioprio = READ_ONCE(sqe->ioprio);
2518 ret = ioprio_check_cap(ioprio);
2522 kiocb->ki_ioprio = ioprio;
2524 kiocb->ki_ioprio = get_current_ioprio();
2526 /* don't allow async punt if RWF_NOWAIT was requested */
2527 if (kiocb->ki_flags & IOCB_NOWAIT)
2528 req->flags |= REQ_F_NOWAIT;
2530 if (kiocb->ki_flags & IOCB_DIRECT)
2531 io_get_req_task(req);
2534 kiocb->ki_flags |= IOCB_NOWAIT;
2536 if (ctx->flags & IORING_SETUP_IOPOLL) {
2537 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2538 !kiocb->ki_filp->f_op->iopoll)
2541 kiocb->ki_flags |= IOCB_HIPRI;
2542 kiocb->ki_complete = io_complete_rw_iopoll;
2543 req->iopoll_completed = 0;
2544 io_get_req_task(req);
2546 if (kiocb->ki_flags & IOCB_HIPRI)
2548 kiocb->ki_complete = io_complete_rw;
2551 req->rw.addr = READ_ONCE(sqe->addr);
2552 req->rw.len = READ_ONCE(sqe->len);
2553 req->buf_index = READ_ONCE(sqe->buf_index);
2557 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2563 case -ERESTARTNOINTR:
2564 case -ERESTARTNOHAND:
2565 case -ERESTART_RESTARTBLOCK:
2567 * We can't just restart the syscall, since previously
2568 * submitted sqes may already be in progress. Just fail this
2574 kiocb->ki_complete(kiocb, ret, 0);
2578 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2579 struct io_comp_state *cs)
2581 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2583 /* add previously done IO, if any */
2584 if (req->io && req->io->rw.bytes_done > 0) {
2586 ret = req->io->rw.bytes_done;
2588 ret += req->io->rw.bytes_done;
2591 if (req->flags & REQ_F_CUR_POS)
2592 req->file->f_pos = kiocb->ki_pos;
2593 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2594 __io_complete_rw(req, ret, 0, cs);
2596 io_rw_done(kiocb, ret);
2599 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2600 struct iov_iter *iter)
2602 struct io_ring_ctx *ctx = req->ctx;
2603 size_t len = req->rw.len;
2604 struct io_mapped_ubuf *imu;
2605 u16 index, buf_index;
2609 /* attempt to use fixed buffers without having provided iovecs */
2610 if (unlikely(!ctx->user_bufs))
2613 buf_index = req->buf_index;
2614 if (unlikely(buf_index >= ctx->nr_user_bufs))
2617 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2618 imu = &ctx->user_bufs[index];
2619 buf_addr = req->rw.addr;
2622 if (buf_addr + len < buf_addr)
2624 /* not inside the mapped region */
2625 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2629 * May not be a start of buffer, set size appropriately
2630 * and advance us to the beginning.
2632 offset = buf_addr - imu->ubuf;
2633 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2637 * Don't use iov_iter_advance() here, as it's really slow for
2638 * using the latter parts of a big fixed buffer - it iterates
2639 * over each segment manually. We can cheat a bit here, because
2642 * 1) it's a BVEC iter, we set it up
2643 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2644 * first and last bvec
2646 * So just find our index, and adjust the iterator afterwards.
2647 * If the offset is within the first bvec (or the whole first
2648 * bvec, just use iov_iter_advance(). This makes it easier
2649 * since we can just skip the first segment, which may not
2650 * be PAGE_SIZE aligned.
2652 const struct bio_vec *bvec = imu->bvec;
2654 if (offset <= bvec->bv_len) {
2655 iov_iter_advance(iter, offset);
2657 unsigned long seg_skip;
2659 /* skip first vec */
2660 offset -= bvec->bv_len;
2661 seg_skip = 1 + (offset >> PAGE_SHIFT);
2663 iter->bvec = bvec + seg_skip;
2664 iter->nr_segs -= seg_skip;
2665 iter->count -= bvec->bv_len + offset;
2666 iter->iov_offset = offset & ~PAGE_MASK;
2673 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2676 mutex_unlock(&ctx->uring_lock);
2679 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2682 * "Normal" inline submissions always hold the uring_lock, since we
2683 * grab it from the system call. Same is true for the SQPOLL offload.
2684 * The only exception is when we've detached the request and issue it
2685 * from an async worker thread, grab the lock for that case.
2688 mutex_lock(&ctx->uring_lock);
2691 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2692 int bgid, struct io_buffer *kbuf,
2695 struct io_buffer *head;
2697 if (req->flags & REQ_F_BUFFER_SELECTED)
2700 io_ring_submit_lock(req->ctx, needs_lock);
2702 lockdep_assert_held(&req->ctx->uring_lock);
2704 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2706 if (!list_empty(&head->list)) {
2707 kbuf = list_last_entry(&head->list, struct io_buffer,
2709 list_del(&kbuf->list);
2712 idr_remove(&req->ctx->io_buffer_idr, bgid);
2714 if (*len > kbuf->len)
2717 kbuf = ERR_PTR(-ENOBUFS);
2720 io_ring_submit_unlock(req->ctx, needs_lock);
2725 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2728 struct io_buffer *kbuf;
2731 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2732 bgid = req->buf_index;
2733 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2736 req->rw.addr = (u64) (unsigned long) kbuf;
2737 req->flags |= REQ_F_BUFFER_SELECTED;
2738 return u64_to_user_ptr(kbuf->addr);
2741 #ifdef CONFIG_COMPAT
2742 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2745 struct compat_iovec __user *uiov;
2746 compat_ssize_t clen;
2750 uiov = u64_to_user_ptr(req->rw.addr);
2751 if (!access_ok(uiov, sizeof(*uiov)))
2753 if (__get_user(clen, &uiov->iov_len))
2759 buf = io_rw_buffer_select(req, &len, needs_lock);
2761 return PTR_ERR(buf);
2762 iov[0].iov_base = buf;
2763 iov[0].iov_len = (compat_size_t) len;
2768 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2771 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2775 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2778 len = iov[0].iov_len;
2781 buf = io_rw_buffer_select(req, &len, needs_lock);
2783 return PTR_ERR(buf);
2784 iov[0].iov_base = buf;
2785 iov[0].iov_len = len;
2789 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2792 if (req->flags & REQ_F_BUFFER_SELECTED) {
2793 struct io_buffer *kbuf;
2795 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2796 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2797 iov[0].iov_len = kbuf->len;
2802 else if (req->rw.len > 1)
2805 #ifdef CONFIG_COMPAT
2806 if (req->ctx->compat)
2807 return io_compat_import(req, iov, needs_lock);
2810 return __io_iov_buffer_select(req, iov, needs_lock);
2813 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2814 struct iovec **iovec, struct iov_iter *iter,
2817 void __user *buf = u64_to_user_ptr(req->rw.addr);
2818 size_t sqe_len = req->rw.len;
2822 opcode = req->opcode;
2823 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2825 return io_import_fixed(req, rw, iter);
2828 /* buffer index only valid with fixed read/write, or buffer select */
2829 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2832 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2833 if (req->flags & REQ_F_BUFFER_SELECT) {
2834 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2836 return PTR_ERR(buf);
2837 req->rw.len = sqe_len;
2840 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2842 return ret < 0 ? ret : sqe_len;
2845 if (req->flags & REQ_F_BUFFER_SELECT) {
2846 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2848 ret = (*iovec)->iov_len;
2849 iov_iter_init(iter, rw, *iovec, 1, ret);
2855 #ifdef CONFIG_COMPAT
2856 if (req->ctx->compat)
2857 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2861 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2864 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2865 struct iovec **iovec, struct iov_iter *iter,
2869 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2871 return iov_iter_count(&req->io->rw.iter);
2874 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2876 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2880 * For files that don't have ->read_iter() and ->write_iter(), handle them
2881 * by looping over ->read() or ->write() manually.
2883 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2884 struct iov_iter *iter)
2889 * Don't support polled IO through this interface, and we can't
2890 * support non-blocking either. For the latter, this just causes
2891 * the kiocb to be handled from an async context.
2893 if (kiocb->ki_flags & IOCB_HIPRI)
2895 if (kiocb->ki_flags & IOCB_NOWAIT)
2898 while (iov_iter_count(iter)) {
2902 if (!iov_iter_is_bvec(iter)) {
2903 iovec = iov_iter_iovec(iter);
2905 /* fixed buffers import bvec */
2906 iovec.iov_base = kmap(iter->bvec->bv_page)
2908 iovec.iov_len = min(iter->count,
2909 iter->bvec->bv_len - iter->iov_offset);
2913 nr = file->f_op->read(file, iovec.iov_base,
2914 iovec.iov_len, io_kiocb_ppos(kiocb));
2916 nr = file->f_op->write(file, iovec.iov_base,
2917 iovec.iov_len, io_kiocb_ppos(kiocb));
2920 if (iov_iter_is_bvec(iter))
2921 kunmap(iter->bvec->bv_page);
2929 if (nr != iovec.iov_len)
2931 iov_iter_advance(iter, nr);
2937 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2938 const struct iovec *fast_iov, struct iov_iter *iter)
2940 struct io_async_rw *rw = &req->io->rw;
2942 memcpy(&rw->iter, iter, sizeof(*iter));
2943 rw->free_iovec = NULL;
2945 /* can only be fixed buffers, no need to do anything */
2946 if (iter->type == ITER_BVEC)
2949 unsigned iov_off = 0;
2951 rw->iter.iov = rw->fast_iov;
2952 if (iter->iov != fast_iov) {
2953 iov_off = iter->iov - fast_iov;
2954 rw->iter.iov += iov_off;
2956 if (rw->fast_iov != fast_iov)
2957 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2958 sizeof(struct iovec) * iter->nr_segs);
2960 rw->free_iovec = iovec;
2961 req->flags |= REQ_F_NEED_CLEANUP;
2965 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2967 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2968 return req->io == NULL;
2971 static int io_alloc_async_ctx(struct io_kiocb *req)
2973 if (!io_op_defs[req->opcode].async_ctx)
2976 return __io_alloc_async_ctx(req);
2979 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2980 const struct iovec *fast_iov,
2981 struct iov_iter *iter, bool force)
2983 if (!force && !io_op_defs[req->opcode].async_ctx)
2986 if (__io_alloc_async_ctx(req))
2989 io_req_map_rw(req, iovec, fast_iov, iter);
2994 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2995 bool force_nonblock)
2997 struct io_async_rw *iorw = &req->io->rw;
3001 iorw->iter.iov = iov = iorw->fast_iov;
3002 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock);
3003 if (unlikely(ret < 0))
3006 iorw->iter.iov = iov;
3007 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3011 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3012 bool force_nonblock)
3016 ret = io_prep_rw(req, sqe, force_nonblock);
3020 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3023 /* either don't need iovec imported or already have it */
3024 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3026 return io_rw_prep_async(req, READ, force_nonblock);
3030 * This is our waitqueue callback handler, registered through lock_page_async()
3031 * when we initially tried to do the IO with the iocb armed our waitqueue.
3032 * This gets called when the page is unlocked, and we generally expect that to
3033 * happen when the page IO is completed and the page is now uptodate. This will
3034 * queue a task_work based retry of the operation, attempting to copy the data
3035 * again. If the latter fails because the page was NOT uptodate, then we will
3036 * do a thread based blocking retry of the operation. That's the unexpected
3039 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3040 int sync, void *arg)
3042 struct wait_page_queue *wpq;
3043 struct io_kiocb *req = wait->private;
3044 struct wait_page_key *key = arg;
3047 wpq = container_of(wait, struct wait_page_queue, wait);
3049 if (!wake_page_match(wpq, key))
3052 list_del_init(&wait->entry);
3054 init_task_work(&req->task_work, io_req_task_submit);
3055 percpu_ref_get(&req->ctx->refs);
3057 /* submit ref gets dropped, acquire a new one */
3058 refcount_inc(&req->refs);
3059 ret = io_req_task_work_add(req, &req->task_work, true);
3060 if (unlikely(ret)) {
3061 struct task_struct *tsk;
3063 /* queue just for cancelation */
3064 init_task_work(&req->task_work, io_req_task_cancel);
3065 tsk = io_wq_get_task(req->ctx->io_wq);
3066 task_work_add(tsk, &req->task_work, 0);
3067 wake_up_process(tsk);
3073 * This controls whether a given IO request should be armed for async page
3074 * based retry. If we return false here, the request is handed to the async
3075 * worker threads for retry. If we're doing buffered reads on a regular file,
3076 * we prepare a private wait_page_queue entry and retry the operation. This
3077 * will either succeed because the page is now uptodate and unlocked, or it
3078 * will register a callback when the page is unlocked at IO completion. Through
3079 * that callback, io_uring uses task_work to setup a retry of the operation.
3080 * That retry will attempt the buffered read again. The retry will generally
3081 * succeed, or in rare cases where it fails, we then fall back to using the
3082 * async worker threads for a blocking retry.
3084 static bool io_rw_should_retry(struct io_kiocb *req)
3086 struct wait_page_queue *wait = &req->io->rw.wpq;
3087 struct kiocb *kiocb = &req->rw.kiocb;
3089 /* never retry for NOWAIT, we just complete with -EAGAIN */
3090 if (req->flags & REQ_F_NOWAIT)
3093 /* Only for buffered IO */
3094 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3098 * just use poll if we can, and don't attempt if the fs doesn't
3099 * support callback based unlocks
3101 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3104 wait->wait.func = io_async_buf_func;
3105 wait->wait.private = req;
3106 wait->wait.flags = 0;
3107 INIT_LIST_HEAD(&wait->wait.entry);
3108 kiocb->ki_flags |= IOCB_WAITQ;
3109 kiocb->ki_waitq = wait;
3111 io_get_req_task(req);
3115 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3117 if (req->file->f_op->read_iter)
3118 return call_read_iter(req->file, &req->rw.kiocb, iter);
3119 else if (req->file->f_op->read)
3120 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3125 static int io_read(struct io_kiocb *req, bool force_nonblock,
3126 struct io_comp_state *cs)
3128 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3129 struct kiocb *kiocb = &req->rw.kiocb;
3130 struct iov_iter __iter, *iter = &__iter;
3131 ssize_t io_size, ret, ret2;
3136 iter = &req->io->rw.iter;
3138 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3141 iov_count = iov_iter_count(iter);
3143 req->result = io_size;
3146 /* Ensure we clear previously set non-block flag */
3147 if (!force_nonblock)
3148 kiocb->ki_flags &= ~IOCB_NOWAIT;
3150 /* If the file doesn't support async, just async punt */
3151 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3155 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3159 ret = io_iter_do_read(req, iter);
3163 } else if (ret == -EIOCBQUEUED) {
3166 } else if (ret == -EAGAIN) {
3167 /* IOPOLL retry should happen for io-wq threads */
3168 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3170 /* no retry on NONBLOCK marked file */
3171 if (req->file->f_flags & O_NONBLOCK)
3173 /* some cases will consume bytes even on error returns */
3174 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3175 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3179 } else if (ret < 0) {
3180 /* make sure -ERESTARTSYS -> -EINTR is done */
3184 /* read it all, or we did blocking attempt. no retry. */
3185 if (!iov_iter_count(iter) || !force_nonblock ||
3186 (req->file->f_flags & O_NONBLOCK))
3191 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3198 /* it's copied and will be cleaned with ->io */
3200 /* now use our persistent iterator, if we aren't already */
3201 iter = &req->io->rw.iter;
3203 req->io->rw.bytes_done += ret;
3204 /* if we can retry, do so with the callbacks armed */
3205 if (!io_rw_should_retry(req)) {
3206 kiocb->ki_flags &= ~IOCB_WAITQ;
3211 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3212 * get -EIOCBQUEUED, then we'll get a notification when the desired
3213 * page gets unlocked. We can also get a partial read here, and if we
3214 * do, then just retry at the new offset.
3216 ret = io_iter_do_read(req, iter);
3217 if (ret == -EIOCBQUEUED) {
3220 } else if (ret > 0 && ret < io_size) {
3221 /* we got some bytes, but not all. retry. */
3225 kiocb_done(kiocb, ret, cs);
3228 /* it's reportedly faster than delegating the null check to kfree() */
3234 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3235 bool force_nonblock)
3239 ret = io_prep_rw(req, sqe, force_nonblock);
3243 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3246 /* either don't need iovec imported or already have it */
3247 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3249 return io_rw_prep_async(req, WRITE, force_nonblock);
3252 static int io_write(struct io_kiocb *req, bool force_nonblock,
3253 struct io_comp_state *cs)
3255 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3256 struct kiocb *kiocb = &req->rw.kiocb;
3257 struct iov_iter __iter, *iter = &__iter;
3259 ssize_t ret, ret2, io_size;
3262 iter = &req->io->rw.iter;
3264 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3267 iov_count = iov_iter_count(iter);
3269 req->result = io_size;
3271 /* Ensure we clear previously set non-block flag */
3272 if (!force_nonblock)
3273 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3275 /* If the file doesn't support async, just async punt */
3276 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3279 /* file path doesn't support NOWAIT for non-direct_IO */
3280 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3281 (req->flags & REQ_F_ISREG))
3284 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3289 * Open-code file_start_write here to grab freeze protection,
3290 * which will be released by another thread in
3291 * io_complete_rw(). Fool lockdep by telling it the lock got
3292 * released so that it doesn't complain about the held lock when
3293 * we return to userspace.
3295 if (req->flags & REQ_F_ISREG) {
3296 __sb_start_write(file_inode(req->file)->i_sb,
3297 SB_FREEZE_WRITE, true);
3298 __sb_writers_release(file_inode(req->file)->i_sb,
3301 kiocb->ki_flags |= IOCB_WRITE;
3303 if (req->file->f_op->write_iter)
3304 ret2 = call_write_iter(req->file, kiocb, iter);
3305 else if (req->file->f_op->write)
3306 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3311 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3312 * retry them without IOCB_NOWAIT.
3314 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3316 /* no retry on NONBLOCK marked file */
3317 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3319 if (!force_nonblock || ret2 != -EAGAIN) {
3320 /* IOPOLL retry should happen for io-wq threads */
3321 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3324 kiocb_done(kiocb, ret2, cs);
3327 /* some cases will consume bytes even on error returns */
3328 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3329 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3334 /* it's reportedly faster than delegating the null check to kfree() */
3340 static int __io_splice_prep(struct io_kiocb *req,
3341 const struct io_uring_sqe *sqe)
3343 struct io_splice* sp = &req->splice;
3344 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3347 if (req->flags & REQ_F_NEED_CLEANUP)
3349 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3353 sp->len = READ_ONCE(sqe->len);
3354 sp->flags = READ_ONCE(sqe->splice_flags);
3356 if (unlikely(sp->flags & ~valid_flags))
3359 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3360 (sp->flags & SPLICE_F_FD_IN_FIXED));
3363 req->flags |= REQ_F_NEED_CLEANUP;
3365 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3367 * Splice operation will be punted aync, and here need to
3368 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3370 io_req_init_async(req);
3371 req->work.flags |= IO_WQ_WORK_UNBOUND;
3377 static int io_tee_prep(struct io_kiocb *req,
3378 const struct io_uring_sqe *sqe)
3380 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3382 return __io_splice_prep(req, sqe);
3385 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3387 struct io_splice *sp = &req->splice;
3388 struct file *in = sp->file_in;
3389 struct file *out = sp->file_out;
3390 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3396 ret = do_tee(in, out, sp->len, flags);
3398 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3399 req->flags &= ~REQ_F_NEED_CLEANUP;
3402 req_set_fail_links(req);
3403 io_req_complete(req, ret);
3407 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3409 struct io_splice* sp = &req->splice;
3411 sp->off_in = READ_ONCE(sqe->splice_off_in);
3412 sp->off_out = READ_ONCE(sqe->off);
3413 return __io_splice_prep(req, sqe);
3416 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3418 struct io_splice *sp = &req->splice;
3419 struct file *in = sp->file_in;
3420 struct file *out = sp->file_out;
3421 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3422 loff_t *poff_in, *poff_out;
3428 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3429 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3432 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3434 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3435 req->flags &= ~REQ_F_NEED_CLEANUP;
3438 req_set_fail_links(req);
3439 io_req_complete(req, ret);
3444 * IORING_OP_NOP just posts a completion event, nothing else.
3446 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3448 struct io_ring_ctx *ctx = req->ctx;
3450 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3453 __io_req_complete(req, 0, 0, cs);
3457 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3459 struct io_ring_ctx *ctx = req->ctx;
3464 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3466 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3469 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3470 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3473 req->sync.off = READ_ONCE(sqe->off);
3474 req->sync.len = READ_ONCE(sqe->len);
3478 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3480 loff_t end = req->sync.off + req->sync.len;
3483 /* fsync always requires a blocking context */
3487 ret = vfs_fsync_range(req->file, req->sync.off,
3488 end > 0 ? end : LLONG_MAX,
3489 req->sync.flags & IORING_FSYNC_DATASYNC);
3491 req_set_fail_links(req);
3492 io_req_complete(req, ret);
3496 static int io_fallocate_prep(struct io_kiocb *req,
3497 const struct io_uring_sqe *sqe)
3499 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3501 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3504 req->sync.off = READ_ONCE(sqe->off);
3505 req->sync.len = READ_ONCE(sqe->addr);
3506 req->sync.mode = READ_ONCE(sqe->len);
3510 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3514 /* fallocate always requiring blocking context */
3517 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3520 req_set_fail_links(req);
3521 io_req_complete(req, ret);
3525 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3527 const char __user *fname;
3530 if (unlikely(sqe->ioprio || sqe->buf_index))
3532 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3535 /* open.how should be already initialised */
3536 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3537 req->open.how.flags |= O_LARGEFILE;
3539 req->open.dfd = READ_ONCE(sqe->fd);
3540 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3541 req->open.filename = getname(fname);
3542 if (IS_ERR(req->open.filename)) {
3543 ret = PTR_ERR(req->open.filename);
3544 req->open.filename = NULL;
3547 req->open.nofile = rlimit(RLIMIT_NOFILE);
3548 req->flags |= REQ_F_NEED_CLEANUP;
3552 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3556 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3558 if (req->flags & REQ_F_NEED_CLEANUP)
3560 mode = READ_ONCE(sqe->len);
3561 flags = READ_ONCE(sqe->open_flags);
3562 req->open.how = build_open_how(flags, mode);
3563 return __io_openat_prep(req, sqe);
3566 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3568 struct open_how __user *how;
3572 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3574 if (req->flags & REQ_F_NEED_CLEANUP)
3576 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3577 len = READ_ONCE(sqe->len);
3578 if (len < OPEN_HOW_SIZE_VER0)
3581 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3586 return __io_openat_prep(req, sqe);
3589 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3591 struct open_flags op;
3598 ret = build_open_flags(&req->open.how, &op);
3602 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3606 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3609 ret = PTR_ERR(file);
3611 fsnotify_open(file);
3612 fd_install(ret, file);
3615 putname(req->open.filename);
3616 req->flags &= ~REQ_F_NEED_CLEANUP;
3618 req_set_fail_links(req);
3619 io_req_complete(req, ret);
3623 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3625 return io_openat2(req, force_nonblock);
3628 static int io_remove_buffers_prep(struct io_kiocb *req,
3629 const struct io_uring_sqe *sqe)
3631 struct io_provide_buf *p = &req->pbuf;
3634 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3637 tmp = READ_ONCE(sqe->fd);
3638 if (!tmp || tmp > USHRT_MAX)
3641 memset(p, 0, sizeof(*p));
3643 p->bgid = READ_ONCE(sqe->buf_group);
3647 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3648 int bgid, unsigned nbufs)
3652 /* shouldn't happen */
3656 /* the head kbuf is the list itself */
3657 while (!list_empty(&buf->list)) {
3658 struct io_buffer *nxt;
3660 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3661 list_del(&nxt->list);
3668 idr_remove(&ctx->io_buffer_idr, bgid);
3673 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3674 struct io_comp_state *cs)
3676 struct io_provide_buf *p = &req->pbuf;
3677 struct io_ring_ctx *ctx = req->ctx;
3678 struct io_buffer *head;
3681 io_ring_submit_lock(ctx, !force_nonblock);
3683 lockdep_assert_held(&ctx->uring_lock);
3686 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3688 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3690 io_ring_submit_lock(ctx, !force_nonblock);
3692 req_set_fail_links(req);
3693 __io_req_complete(req, ret, 0, cs);
3697 static int io_provide_buffers_prep(struct io_kiocb *req,
3698 const struct io_uring_sqe *sqe)
3700 struct io_provide_buf *p = &req->pbuf;
3703 if (sqe->ioprio || sqe->rw_flags)
3706 tmp = READ_ONCE(sqe->fd);
3707 if (!tmp || tmp > USHRT_MAX)
3710 p->addr = READ_ONCE(sqe->addr);
3711 p->len = READ_ONCE(sqe->len);
3713 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3716 p->bgid = READ_ONCE(sqe->buf_group);
3717 tmp = READ_ONCE(sqe->off);
3718 if (tmp > USHRT_MAX)
3724 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3726 struct io_buffer *buf;
3727 u64 addr = pbuf->addr;
3728 int i, bid = pbuf->bid;
3730 for (i = 0; i < pbuf->nbufs; i++) {
3731 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3736 buf->len = pbuf->len;
3741 INIT_LIST_HEAD(&buf->list);
3744 list_add_tail(&buf->list, &(*head)->list);
3748 return i ? i : -ENOMEM;
3751 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3752 struct io_comp_state *cs)
3754 struct io_provide_buf *p = &req->pbuf;
3755 struct io_ring_ctx *ctx = req->ctx;
3756 struct io_buffer *head, *list;
3759 io_ring_submit_lock(ctx, !force_nonblock);
3761 lockdep_assert_held(&ctx->uring_lock);
3763 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3765 ret = io_add_buffers(p, &head);
3770 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3773 __io_remove_buffers(ctx, head, p->bgid, -1U);
3778 io_ring_submit_unlock(ctx, !force_nonblock);
3780 req_set_fail_links(req);
3781 __io_req_complete(req, ret, 0, cs);
3785 static int io_epoll_ctl_prep(struct io_kiocb *req,
3786 const struct io_uring_sqe *sqe)
3788 #if defined(CONFIG_EPOLL)
3789 if (sqe->ioprio || sqe->buf_index)
3791 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3794 req->epoll.epfd = READ_ONCE(sqe->fd);
3795 req->epoll.op = READ_ONCE(sqe->len);
3796 req->epoll.fd = READ_ONCE(sqe->off);
3798 if (ep_op_has_event(req->epoll.op)) {
3799 struct epoll_event __user *ev;
3801 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3802 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3812 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3813 struct io_comp_state *cs)
3815 #if defined(CONFIG_EPOLL)
3816 struct io_epoll *ie = &req->epoll;
3819 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3820 if (force_nonblock && ret == -EAGAIN)
3824 req_set_fail_links(req);
3825 __io_req_complete(req, ret, 0, cs);
3832 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3834 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3835 if (sqe->ioprio || sqe->buf_index || sqe->off)
3837 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3840 req->madvise.addr = READ_ONCE(sqe->addr);
3841 req->madvise.len = READ_ONCE(sqe->len);
3842 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3849 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3851 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3852 struct io_madvise *ma = &req->madvise;
3858 ret = do_madvise(ma->addr, ma->len, ma->advice);
3860 req_set_fail_links(req);
3861 io_req_complete(req, ret);
3868 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3870 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3872 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3875 req->fadvise.offset = READ_ONCE(sqe->off);
3876 req->fadvise.len = READ_ONCE(sqe->len);
3877 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3881 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3883 struct io_fadvise *fa = &req->fadvise;
3886 if (force_nonblock) {
3887 switch (fa->advice) {
3888 case POSIX_FADV_NORMAL:
3889 case POSIX_FADV_RANDOM:
3890 case POSIX_FADV_SEQUENTIAL:
3897 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3899 req_set_fail_links(req);
3900 io_req_complete(req, ret);
3904 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3906 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3908 if (sqe->ioprio || sqe->buf_index)
3910 if (req->flags & REQ_F_FIXED_FILE)
3913 req->statx.dfd = READ_ONCE(sqe->fd);
3914 req->statx.mask = READ_ONCE(sqe->len);
3915 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3916 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3917 req->statx.flags = READ_ONCE(sqe->statx_flags);
3922 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3924 struct io_statx *ctx = &req->statx;
3927 if (force_nonblock) {
3928 /* only need file table for an actual valid fd */
3929 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3930 req->flags |= REQ_F_NO_FILE_TABLE;
3934 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3938 req_set_fail_links(req);
3939 io_req_complete(req, ret);
3943 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3946 * If we queue this for async, it must not be cancellable. That would
3947 * leave the 'file' in an undeterminate state, and here need to modify
3948 * io_wq_work.flags, so initialize io_wq_work firstly.
3950 io_req_init_async(req);
3951 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3953 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3955 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3956 sqe->rw_flags || sqe->buf_index)
3958 if (req->flags & REQ_F_FIXED_FILE)
3961 req->close.fd = READ_ONCE(sqe->fd);
3962 if ((req->file && req->file->f_op == &io_uring_fops) ||
3963 req->close.fd == req->ctx->ring_fd)
3966 req->close.put_file = NULL;
3970 static int io_close(struct io_kiocb *req, bool force_nonblock,
3971 struct io_comp_state *cs)
3973 struct io_close *close = &req->close;
3976 /* might be already done during nonblock submission */
3977 if (!close->put_file) {
3978 ret = __close_fd_get_file(close->fd, &close->put_file);
3980 return (ret == -ENOENT) ? -EBADF : ret;
3983 /* if the file has a flush method, be safe and punt to async */
3984 if (close->put_file->f_op->flush && force_nonblock) {
3985 /* was never set, but play safe */
3986 req->flags &= ~REQ_F_NOWAIT;
3987 /* avoid grabbing files - we don't need the files */
3988 req->flags |= REQ_F_NO_FILE_TABLE;
3992 /* No ->flush() or already async, safely close from here */
3993 ret = filp_close(close->put_file, req->work.files);
3995 req_set_fail_links(req);
3996 fput(close->put_file);
3997 close->put_file = NULL;
3998 __io_req_complete(req, ret, 0, cs);
4002 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4004 struct io_ring_ctx *ctx = req->ctx;
4009 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4011 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4014 req->sync.off = READ_ONCE(sqe->off);
4015 req->sync.len = READ_ONCE(sqe->len);
4016 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4020 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4024 /* sync_file_range always requires a blocking context */
4028 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4031 req_set_fail_links(req);
4032 io_req_complete(req, ret);
4036 #if defined(CONFIG_NET)
4037 static int io_setup_async_msg(struct io_kiocb *req,
4038 struct io_async_msghdr *kmsg)
4042 if (io_alloc_async_ctx(req)) {
4043 if (kmsg->iov != kmsg->fast_iov)
4047 req->flags |= REQ_F_NEED_CLEANUP;
4048 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4052 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4053 struct io_async_msghdr *iomsg)
4055 iomsg->iov = iomsg->fast_iov;
4056 iomsg->msg.msg_name = &iomsg->addr;
4057 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4058 req->sr_msg.msg_flags, &iomsg->iov);
4061 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4063 struct io_sr_msg *sr = &req->sr_msg;
4064 struct io_async_ctx *io = req->io;
4067 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4070 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4071 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4072 sr->len = READ_ONCE(sqe->len);
4074 #ifdef CONFIG_COMPAT
4075 if (req->ctx->compat)
4076 sr->msg_flags |= MSG_CMSG_COMPAT;
4079 if (!io || req->opcode == IORING_OP_SEND)
4081 /* iovec is already imported */
4082 if (req->flags & REQ_F_NEED_CLEANUP)
4085 ret = io_sendmsg_copy_hdr(req, &io->msg);
4087 req->flags |= REQ_F_NEED_CLEANUP;
4091 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4092 struct io_comp_state *cs)
4094 struct io_async_msghdr iomsg, *kmsg;
4095 struct socket *sock;
4099 sock = sock_from_file(req->file, &ret);
4100 if (unlikely(!sock))
4104 kmsg = &req->io->msg;
4105 kmsg->msg.msg_name = &req->io->msg.addr;
4106 /* if iov is set, it's allocated already */
4108 kmsg->iov = kmsg->fast_iov;
4109 kmsg->msg.msg_iter.iov = kmsg->iov;
4111 ret = io_sendmsg_copy_hdr(req, &iomsg);
4117 flags = req->sr_msg.msg_flags;
4118 if (flags & MSG_DONTWAIT)
4119 req->flags |= REQ_F_NOWAIT;
4120 else if (force_nonblock)
4121 flags |= MSG_DONTWAIT;
4123 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4124 if (force_nonblock && ret == -EAGAIN)
4125 return io_setup_async_msg(req, kmsg);
4126 if (ret == -ERESTARTSYS)
4129 if (kmsg->iov != kmsg->fast_iov)
4131 req->flags &= ~REQ_F_NEED_CLEANUP;
4133 req_set_fail_links(req);
4134 __io_req_complete(req, ret, 0, cs);
4138 static int io_send(struct io_kiocb *req, bool force_nonblock,
4139 struct io_comp_state *cs)
4141 struct io_sr_msg *sr = &req->sr_msg;
4144 struct socket *sock;
4148 sock = sock_from_file(req->file, &ret);
4149 if (unlikely(!sock))
4152 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4156 msg.msg_name = NULL;
4157 msg.msg_control = NULL;
4158 msg.msg_controllen = 0;
4159 msg.msg_namelen = 0;
4161 flags = req->sr_msg.msg_flags;
4162 if (flags & MSG_DONTWAIT)
4163 req->flags |= REQ_F_NOWAIT;
4164 else if (force_nonblock)
4165 flags |= MSG_DONTWAIT;
4167 msg.msg_flags = flags;
4168 ret = sock_sendmsg(sock, &msg);
4169 if (force_nonblock && ret == -EAGAIN)
4171 if (ret == -ERESTARTSYS)
4175 req_set_fail_links(req);
4176 __io_req_complete(req, ret, 0, cs);
4180 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4181 struct io_async_msghdr *iomsg)
4183 struct io_sr_msg *sr = &req->sr_msg;
4184 struct iovec __user *uiov;
4188 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4189 &iomsg->uaddr, &uiov, &iov_len);
4193 if (req->flags & REQ_F_BUFFER_SELECT) {
4196 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4198 sr->len = iomsg->iov[0].iov_len;
4199 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4203 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4204 &iomsg->iov, &iomsg->msg.msg_iter);
4212 #ifdef CONFIG_COMPAT
4213 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4214 struct io_async_msghdr *iomsg)
4216 struct compat_msghdr __user *msg_compat;
4217 struct io_sr_msg *sr = &req->sr_msg;
4218 struct compat_iovec __user *uiov;
4223 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4224 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4229 uiov = compat_ptr(ptr);
4230 if (req->flags & REQ_F_BUFFER_SELECT) {
4231 compat_ssize_t clen;
4235 if (!access_ok(uiov, sizeof(*uiov)))
4237 if (__get_user(clen, &uiov->iov_len))
4241 sr->len = iomsg->iov[0].iov_len;
4244 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4246 &iomsg->msg.msg_iter);
4255 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4256 struct io_async_msghdr *iomsg)
4258 iomsg->msg.msg_name = &iomsg->addr;
4259 iomsg->iov = iomsg->fast_iov;
4261 #ifdef CONFIG_COMPAT
4262 if (req->ctx->compat)
4263 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4266 return __io_recvmsg_copy_hdr(req, iomsg);
4269 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4272 struct io_sr_msg *sr = &req->sr_msg;
4273 struct io_buffer *kbuf;
4275 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4280 req->flags |= REQ_F_BUFFER_SELECTED;
4284 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4286 return io_put_kbuf(req, req->sr_msg.kbuf);
4289 static int io_recvmsg_prep(struct io_kiocb *req,
4290 const struct io_uring_sqe *sqe)
4292 struct io_sr_msg *sr = &req->sr_msg;
4293 struct io_async_ctx *io = req->io;
4296 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4299 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4300 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4301 sr->len = READ_ONCE(sqe->len);
4302 sr->bgid = READ_ONCE(sqe->buf_group);
4304 #ifdef CONFIG_COMPAT
4305 if (req->ctx->compat)
4306 sr->msg_flags |= MSG_CMSG_COMPAT;
4309 if (!io || req->opcode == IORING_OP_RECV)
4311 /* iovec is already imported */
4312 if (req->flags & REQ_F_NEED_CLEANUP)
4315 ret = io_recvmsg_copy_hdr(req, &io->msg);
4317 req->flags |= REQ_F_NEED_CLEANUP;
4321 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4322 struct io_comp_state *cs)
4324 struct io_async_msghdr iomsg, *kmsg;
4325 struct socket *sock;
4326 struct io_buffer *kbuf;
4328 int ret, cflags = 0;
4330 sock = sock_from_file(req->file, &ret);
4331 if (unlikely(!sock))
4335 kmsg = &req->io->msg;
4336 kmsg->msg.msg_name = &req->io->msg.addr;
4337 /* if iov is set, it's allocated already */
4339 kmsg->iov = kmsg->fast_iov;
4340 kmsg->msg.msg_iter.iov = kmsg->iov;
4342 ret = io_recvmsg_copy_hdr(req, &iomsg);
4348 if (req->flags & REQ_F_BUFFER_SELECT) {
4349 kbuf = io_recv_buffer_select(req, !force_nonblock);
4351 return PTR_ERR(kbuf);
4352 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4353 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4354 1, req->sr_msg.len);
4357 flags = req->sr_msg.msg_flags;
4358 if (flags & MSG_DONTWAIT)
4359 req->flags |= REQ_F_NOWAIT;
4360 else if (force_nonblock)
4361 flags |= MSG_DONTWAIT;
4363 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4364 kmsg->uaddr, flags);
4365 if (force_nonblock && ret == -EAGAIN)
4366 return io_setup_async_msg(req, kmsg);
4367 if (ret == -ERESTARTSYS)
4370 if (req->flags & REQ_F_BUFFER_SELECTED)
4371 cflags = io_put_recv_kbuf(req);
4372 if (kmsg->iov != kmsg->fast_iov)
4374 req->flags &= ~REQ_F_NEED_CLEANUP;
4376 req_set_fail_links(req);
4377 __io_req_complete(req, ret, cflags, cs);
4381 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4382 struct io_comp_state *cs)
4384 struct io_buffer *kbuf;
4385 struct io_sr_msg *sr = &req->sr_msg;
4387 void __user *buf = sr->buf;
4388 struct socket *sock;
4391 int ret, cflags = 0;
4393 sock = sock_from_file(req->file, &ret);
4394 if (unlikely(!sock))
4397 if (req->flags & REQ_F_BUFFER_SELECT) {
4398 kbuf = io_recv_buffer_select(req, !force_nonblock);
4400 return PTR_ERR(kbuf);
4401 buf = u64_to_user_ptr(kbuf->addr);
4404 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4408 msg.msg_name = NULL;
4409 msg.msg_control = NULL;
4410 msg.msg_controllen = 0;
4411 msg.msg_namelen = 0;
4412 msg.msg_iocb = NULL;
4415 flags = req->sr_msg.msg_flags;
4416 if (flags & MSG_DONTWAIT)
4417 req->flags |= REQ_F_NOWAIT;
4418 else if (force_nonblock)
4419 flags |= MSG_DONTWAIT;
4421 ret = sock_recvmsg(sock, &msg, flags);
4422 if (force_nonblock && ret == -EAGAIN)
4424 if (ret == -ERESTARTSYS)
4427 if (req->flags & REQ_F_BUFFER_SELECTED)
4428 cflags = io_put_recv_kbuf(req);
4430 req_set_fail_links(req);
4431 __io_req_complete(req, ret, cflags, cs);
4435 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4437 struct io_accept *accept = &req->accept;
4439 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4441 if (sqe->ioprio || sqe->len || sqe->buf_index)
4444 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4445 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4446 accept->flags = READ_ONCE(sqe->accept_flags);
4447 accept->nofile = rlimit(RLIMIT_NOFILE);
4451 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4452 struct io_comp_state *cs)
4454 struct io_accept *accept = &req->accept;
4455 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4458 if (req->file->f_flags & O_NONBLOCK)
4459 req->flags |= REQ_F_NOWAIT;
4461 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4462 accept->addr_len, accept->flags,
4464 if (ret == -EAGAIN && force_nonblock)
4467 if (ret == -ERESTARTSYS)
4469 req_set_fail_links(req);
4471 __io_req_complete(req, ret, 0, cs);
4475 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4477 struct io_connect *conn = &req->connect;
4478 struct io_async_ctx *io = req->io;
4480 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4482 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4485 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4486 conn->addr_len = READ_ONCE(sqe->addr2);
4491 return move_addr_to_kernel(conn->addr, conn->addr_len,
4492 &io->connect.address);
4495 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4496 struct io_comp_state *cs)
4498 struct io_async_ctx __io, *io;
4499 unsigned file_flags;
4505 ret = move_addr_to_kernel(req->connect.addr,
4506 req->connect.addr_len,
4507 &__io.connect.address);
4513 file_flags = force_nonblock ? O_NONBLOCK : 0;
4515 ret = __sys_connect_file(req->file, &io->connect.address,
4516 req->connect.addr_len, file_flags);
4517 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4520 if (io_alloc_async_ctx(req)) {
4524 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4527 if (ret == -ERESTARTSYS)
4531 req_set_fail_links(req);
4532 __io_req_complete(req, ret, 0, cs);
4535 #else /* !CONFIG_NET */
4536 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4541 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4542 struct io_comp_state *cs)
4547 static int io_send(struct io_kiocb *req, bool force_nonblock,
4548 struct io_comp_state *cs)
4553 static int io_recvmsg_prep(struct io_kiocb *req,
4554 const struct io_uring_sqe *sqe)
4559 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4560 struct io_comp_state *cs)
4565 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4566 struct io_comp_state *cs)
4571 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4576 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4577 struct io_comp_state *cs)
4582 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4587 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4588 struct io_comp_state *cs)
4592 #endif /* CONFIG_NET */
4594 struct io_poll_table {
4595 struct poll_table_struct pt;
4596 struct io_kiocb *req;
4600 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4601 __poll_t mask, task_work_func_t func)
4606 /* for instances that support it check for an event match first: */
4607 if (mask && !(mask & poll->events))
4610 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4612 list_del_init(&poll->wait.entry);
4615 init_task_work(&req->task_work, func);
4616 percpu_ref_get(&req->ctx->refs);
4619 * If we using the signalfd wait_queue_head for this wakeup, then
4620 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4621 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4622 * either, as the normal wakeup will suffice.
4624 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4627 * If this fails, then the task is exiting. When a task exits, the
4628 * work gets canceled, so just cancel this request as well instead
4629 * of executing it. We can't safely execute it anyway, as we may not
4630 * have the needed state needed for it anyway.
4632 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4633 if (unlikely(ret)) {
4634 struct task_struct *tsk;
4636 WRITE_ONCE(poll->canceled, true);
4637 tsk = io_wq_get_task(req->ctx->io_wq);
4638 task_work_add(tsk, &req->task_work, 0);
4639 wake_up_process(tsk);
4644 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4645 __acquires(&req->ctx->completion_lock)
4647 struct io_ring_ctx *ctx = req->ctx;
4649 if (!req->result && !READ_ONCE(poll->canceled)) {
4650 struct poll_table_struct pt = { ._key = poll->events };
4652 req->result = vfs_poll(req->file, &pt) & poll->events;
4655 spin_lock_irq(&ctx->completion_lock);
4656 if (!req->result && !READ_ONCE(poll->canceled)) {
4657 add_wait_queue(poll->head, &poll->wait);
4664 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4666 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4667 if (req->opcode == IORING_OP_POLL_ADD)
4668 return (struct io_poll_iocb *) req->io;
4669 return req->apoll->double_poll;
4672 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4674 if (req->opcode == IORING_OP_POLL_ADD)
4676 return &req->apoll->poll;
4679 static void io_poll_remove_double(struct io_kiocb *req)
4681 struct io_poll_iocb *poll = io_poll_get_double(req);
4683 lockdep_assert_held(&req->ctx->completion_lock);
4685 if (poll && poll->head) {
4686 struct wait_queue_head *head = poll->head;
4688 spin_lock(&head->lock);
4689 list_del_init(&poll->wait.entry);
4690 if (poll->wait.private)
4691 refcount_dec(&req->refs);
4693 spin_unlock(&head->lock);
4697 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4699 struct io_ring_ctx *ctx = req->ctx;
4701 io_poll_remove_double(req);
4702 req->poll.done = true;
4703 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4704 io_commit_cqring(ctx);
4707 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4709 struct io_ring_ctx *ctx = req->ctx;
4711 if (io_poll_rewait(req, &req->poll)) {
4712 spin_unlock_irq(&ctx->completion_lock);
4716 hash_del(&req->hash_node);
4717 io_poll_complete(req, req->result, 0);
4718 req->flags |= REQ_F_COMP_LOCKED;
4719 *nxt = io_put_req_find_next(req);
4720 spin_unlock_irq(&ctx->completion_lock);
4722 io_cqring_ev_posted(ctx);
4725 static void io_poll_task_func(struct callback_head *cb)
4727 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4728 struct io_ring_ctx *ctx = req->ctx;
4729 struct io_kiocb *nxt = NULL;
4731 io_poll_task_handler(req, &nxt);
4733 __io_req_task_submit(nxt);
4734 percpu_ref_put(&ctx->refs);
4737 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4738 int sync, void *key)
4740 struct io_kiocb *req = wait->private;
4741 struct io_poll_iocb *poll = io_poll_get_single(req);
4742 __poll_t mask = key_to_poll(key);
4744 /* for instances that support it check for an event match first: */
4745 if (mask && !(mask & poll->events))
4748 if (poll && poll->head) {
4751 spin_lock(&poll->head->lock);
4752 done = list_empty(&poll->wait.entry);
4754 list_del_init(&poll->wait.entry);
4755 /* make sure double remove sees this as being gone */
4756 wait->private = NULL;
4757 spin_unlock(&poll->head->lock);
4759 __io_async_wake(req, poll, mask, io_poll_task_func);
4761 refcount_dec(&req->refs);
4765 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4766 wait_queue_func_t wake_func)
4770 poll->canceled = false;
4771 poll->events = events;
4772 INIT_LIST_HEAD(&poll->wait.entry);
4773 init_waitqueue_func_entry(&poll->wait, wake_func);
4776 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4777 struct wait_queue_head *head,
4778 struct io_poll_iocb **poll_ptr)
4780 struct io_kiocb *req = pt->req;
4783 * If poll->head is already set, it's because the file being polled
4784 * uses multiple waitqueues for poll handling (eg one for read, one
4785 * for write). Setup a separate io_poll_iocb if this happens.
4787 if (unlikely(poll->head)) {
4788 /* already have a 2nd entry, fail a third attempt */
4790 pt->error = -EINVAL;
4793 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4795 pt->error = -ENOMEM;
4798 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4799 refcount_inc(&req->refs);
4800 poll->wait.private = req;
4807 if (poll->events & EPOLLEXCLUSIVE)
4808 add_wait_queue_exclusive(head, &poll->wait);
4810 add_wait_queue(head, &poll->wait);
4813 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4814 struct poll_table_struct *p)
4816 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4817 struct async_poll *apoll = pt->req->apoll;
4819 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4822 static void io_async_task_func(struct callback_head *cb)
4824 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4825 struct async_poll *apoll = req->apoll;
4826 struct io_ring_ctx *ctx = req->ctx;
4828 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4830 if (io_poll_rewait(req, &apoll->poll)) {
4831 spin_unlock_irq(&ctx->completion_lock);
4832 percpu_ref_put(&ctx->refs);
4836 /* If req is still hashed, it cannot have been canceled. Don't check. */
4837 if (hash_hashed(&req->hash_node))
4838 hash_del(&req->hash_node);
4840 io_poll_remove_double(req);
4841 spin_unlock_irq(&ctx->completion_lock);
4843 if (!READ_ONCE(apoll->poll.canceled))
4844 __io_req_task_submit(req);
4846 __io_req_task_cancel(req, -ECANCELED);
4848 percpu_ref_put(&ctx->refs);
4849 kfree(apoll->double_poll);
4853 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4856 struct io_kiocb *req = wait->private;
4857 struct io_poll_iocb *poll = &req->apoll->poll;
4859 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4862 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4865 static void io_poll_req_insert(struct io_kiocb *req)
4867 struct io_ring_ctx *ctx = req->ctx;
4868 struct hlist_head *list;
4870 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4871 hlist_add_head(&req->hash_node, list);
4874 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4875 struct io_poll_iocb *poll,
4876 struct io_poll_table *ipt, __poll_t mask,
4877 wait_queue_func_t wake_func)
4878 __acquires(&ctx->completion_lock)
4880 struct io_ring_ctx *ctx = req->ctx;
4881 bool cancel = false;
4883 io_init_poll_iocb(poll, mask, wake_func);
4884 poll->file = req->file;
4885 poll->wait.private = req;
4887 ipt->pt._key = mask;
4889 ipt->error = -EINVAL;
4891 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4893 spin_lock_irq(&ctx->completion_lock);
4894 if (likely(poll->head)) {
4895 spin_lock(&poll->head->lock);
4896 if (unlikely(list_empty(&poll->wait.entry))) {
4902 if (mask || ipt->error)
4903 list_del_init(&poll->wait.entry);
4905 WRITE_ONCE(poll->canceled, true);
4906 else if (!poll->done) /* actually waiting for an event */
4907 io_poll_req_insert(req);
4908 spin_unlock(&poll->head->lock);
4914 static bool io_arm_poll_handler(struct io_kiocb *req)
4916 const struct io_op_def *def = &io_op_defs[req->opcode];
4917 struct io_ring_ctx *ctx = req->ctx;
4918 struct async_poll *apoll;
4919 struct io_poll_table ipt;
4923 if (!req->file || !file_can_poll(req->file))
4925 if (req->flags & REQ_F_POLLED)
4929 else if (def->pollout)
4933 /* if we can't nonblock try, then no point in arming a poll handler */
4934 if (!io_file_supports_async(req->file, rw))
4937 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4938 if (unlikely(!apoll))
4940 apoll->double_poll = NULL;
4942 req->flags |= REQ_F_POLLED;
4943 io_get_req_task(req);
4945 INIT_HLIST_NODE(&req->hash_node);
4949 mask |= POLLIN | POLLRDNORM;
4951 mask |= POLLOUT | POLLWRNORM;
4952 mask |= POLLERR | POLLPRI;
4954 ipt.pt._qproc = io_async_queue_proc;
4956 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4958 if (ret || ipt.error) {
4959 io_poll_remove_double(req);
4960 spin_unlock_irq(&ctx->completion_lock);
4961 kfree(apoll->double_poll);
4965 spin_unlock_irq(&ctx->completion_lock);
4966 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4967 apoll->poll.events);
4971 static bool __io_poll_remove_one(struct io_kiocb *req,
4972 struct io_poll_iocb *poll)
4974 bool do_complete = false;
4976 spin_lock(&poll->head->lock);
4977 WRITE_ONCE(poll->canceled, true);
4978 if (!list_empty(&poll->wait.entry)) {
4979 list_del_init(&poll->wait.entry);
4982 spin_unlock(&poll->head->lock);
4983 hash_del(&req->hash_node);
4987 static bool io_poll_remove_one(struct io_kiocb *req)
4991 io_poll_remove_double(req);
4993 if (req->opcode == IORING_OP_POLL_ADD) {
4994 do_complete = __io_poll_remove_one(req, &req->poll);
4996 struct async_poll *apoll = req->apoll;
4998 /* non-poll requests have submit ref still */
4999 do_complete = __io_poll_remove_one(req, &apoll->poll);
5002 kfree(apoll->double_poll);
5008 io_cqring_fill_event(req, -ECANCELED);
5009 io_commit_cqring(req->ctx);
5010 req->flags |= REQ_F_COMP_LOCKED;
5011 req_set_fail_links(req);
5018 static void io_poll_remove_all(struct io_ring_ctx *ctx)
5020 struct hlist_node *tmp;
5021 struct io_kiocb *req;
5024 spin_lock_irq(&ctx->completion_lock);
5025 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5026 struct hlist_head *list;
5028 list = &ctx->cancel_hash[i];
5029 hlist_for_each_entry_safe(req, tmp, list, hash_node)
5030 posted += io_poll_remove_one(req);
5032 spin_unlock_irq(&ctx->completion_lock);
5035 io_cqring_ev_posted(ctx);
5038 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5040 struct hlist_head *list;
5041 struct io_kiocb *req;
5043 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5044 hlist_for_each_entry(req, list, hash_node) {
5045 if (sqe_addr != req->user_data)
5047 if (io_poll_remove_one(req))
5055 static int io_poll_remove_prep(struct io_kiocb *req,
5056 const struct io_uring_sqe *sqe)
5058 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5060 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5064 req->poll.addr = READ_ONCE(sqe->addr);
5069 * Find a running poll command that matches one specified in sqe->addr,
5070 * and remove it if found.
5072 static int io_poll_remove(struct io_kiocb *req)
5074 struct io_ring_ctx *ctx = req->ctx;
5078 addr = req->poll.addr;
5079 spin_lock_irq(&ctx->completion_lock);
5080 ret = io_poll_cancel(ctx, addr);
5081 spin_unlock_irq(&ctx->completion_lock);
5084 req_set_fail_links(req);
5085 io_req_complete(req, ret);
5089 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5092 struct io_kiocb *req = wait->private;
5093 struct io_poll_iocb *poll = &req->poll;
5095 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5098 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5099 struct poll_table_struct *p)
5101 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5103 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5106 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5108 struct io_poll_iocb *poll = &req->poll;
5111 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5113 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5118 events = READ_ONCE(sqe->poll32_events);
5120 events = swahw32(events);
5122 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5123 (events & EPOLLEXCLUSIVE);
5125 io_get_req_task(req);
5129 static int io_poll_add(struct io_kiocb *req)
5131 struct io_poll_iocb *poll = &req->poll;
5132 struct io_ring_ctx *ctx = req->ctx;
5133 struct io_poll_table ipt;
5136 INIT_HLIST_NODE(&req->hash_node);
5137 ipt.pt._qproc = io_poll_queue_proc;
5139 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5142 if (mask) { /* no async, we'd stolen it */
5144 io_poll_complete(req, mask, 0);
5146 spin_unlock_irq(&ctx->completion_lock);
5149 io_cqring_ev_posted(ctx);
5155 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5157 struct io_timeout_data *data = container_of(timer,
5158 struct io_timeout_data, timer);
5159 struct io_kiocb *req = data->req;
5160 struct io_ring_ctx *ctx = req->ctx;
5161 unsigned long flags;
5163 spin_lock_irqsave(&ctx->completion_lock, flags);
5164 atomic_set(&req->ctx->cq_timeouts,
5165 atomic_read(&req->ctx->cq_timeouts) + 1);
5168 * We could be racing with timeout deletion. If the list is empty,
5169 * then timeout lookup already found it and will be handling it.
5171 if (!list_empty(&req->timeout.list))
5172 list_del_init(&req->timeout.list);
5174 io_cqring_fill_event(req, -ETIME);
5175 io_commit_cqring(ctx);
5176 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5178 io_cqring_ev_posted(ctx);
5179 req_set_fail_links(req);
5181 return HRTIMER_NORESTART;
5184 static int __io_timeout_cancel(struct io_kiocb *req)
5188 list_del_init(&req->timeout.list);
5190 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5194 req_set_fail_links(req);
5195 req->flags |= REQ_F_COMP_LOCKED;
5196 io_cqring_fill_event(req, -ECANCELED);
5201 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5203 struct io_kiocb *req;
5206 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5207 if (user_data == req->user_data) {
5216 return __io_timeout_cancel(req);
5219 static int io_timeout_remove_prep(struct io_kiocb *req,
5220 const struct io_uring_sqe *sqe)
5222 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5224 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5226 if (sqe->ioprio || sqe->buf_index || sqe->len)
5229 req->timeout.addr = READ_ONCE(sqe->addr);
5230 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5231 if (req->timeout.flags)
5238 * Remove or update an existing timeout command
5240 static int io_timeout_remove(struct io_kiocb *req)
5242 struct io_ring_ctx *ctx = req->ctx;
5245 spin_lock_irq(&ctx->completion_lock);
5246 ret = io_timeout_cancel(ctx, req->timeout.addr);
5248 io_cqring_fill_event(req, ret);
5249 io_commit_cqring(ctx);
5250 spin_unlock_irq(&ctx->completion_lock);
5251 io_cqring_ev_posted(ctx);
5253 req_set_fail_links(req);
5258 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5259 bool is_timeout_link)
5261 struct io_timeout_data *data;
5263 u32 off = READ_ONCE(sqe->off);
5265 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5267 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5269 if (off && is_timeout_link)
5271 flags = READ_ONCE(sqe->timeout_flags);
5272 if (flags & ~IORING_TIMEOUT_ABS)
5275 req->timeout.off = off;
5277 if (!req->io && io_alloc_async_ctx(req))
5280 data = &req->io->timeout;
5283 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5286 if (flags & IORING_TIMEOUT_ABS)
5287 data->mode = HRTIMER_MODE_ABS;
5289 data->mode = HRTIMER_MODE_REL;
5291 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5295 static int io_timeout(struct io_kiocb *req)
5297 struct io_ring_ctx *ctx = req->ctx;
5298 struct io_timeout_data *data = &req->io->timeout;
5299 struct list_head *entry;
5300 u32 tail, off = req->timeout.off;
5302 spin_lock_irq(&ctx->completion_lock);
5305 * sqe->off holds how many events that need to occur for this
5306 * timeout event to be satisfied. If it isn't set, then this is
5307 * a pure timeout request, sequence isn't used.
5309 if (io_is_timeout_noseq(req)) {
5310 entry = ctx->timeout_list.prev;
5314 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5315 req->timeout.target_seq = tail + off;
5318 * Insertion sort, ensuring the first entry in the list is always
5319 * the one we need first.
5321 list_for_each_prev(entry, &ctx->timeout_list) {
5322 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5325 if (io_is_timeout_noseq(nxt))
5327 /* nxt.seq is behind @tail, otherwise would've been completed */
5328 if (off >= nxt->timeout.target_seq - tail)
5332 list_add(&req->timeout.list, entry);
5333 data->timer.function = io_timeout_fn;
5334 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5335 spin_unlock_irq(&ctx->completion_lock);
5339 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5341 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5343 return req->user_data == (unsigned long) data;
5346 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5348 enum io_wq_cancel cancel_ret;
5351 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5352 switch (cancel_ret) {
5353 case IO_WQ_CANCEL_OK:
5356 case IO_WQ_CANCEL_RUNNING:
5359 case IO_WQ_CANCEL_NOTFOUND:
5367 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5368 struct io_kiocb *req, __u64 sqe_addr,
5371 unsigned long flags;
5374 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5375 if (ret != -ENOENT) {
5376 spin_lock_irqsave(&ctx->completion_lock, flags);
5380 spin_lock_irqsave(&ctx->completion_lock, flags);
5381 ret = io_timeout_cancel(ctx, sqe_addr);
5384 ret = io_poll_cancel(ctx, sqe_addr);
5388 io_cqring_fill_event(req, ret);
5389 io_commit_cqring(ctx);
5390 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5391 io_cqring_ev_posted(ctx);
5394 req_set_fail_links(req);
5398 static int io_async_cancel_prep(struct io_kiocb *req,
5399 const struct io_uring_sqe *sqe)
5401 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5403 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5405 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5408 req->cancel.addr = READ_ONCE(sqe->addr);
5412 static int io_async_cancel(struct io_kiocb *req)
5414 struct io_ring_ctx *ctx = req->ctx;
5416 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5420 static int io_files_update_prep(struct io_kiocb *req,
5421 const struct io_uring_sqe *sqe)
5423 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5425 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5427 if (sqe->ioprio || sqe->rw_flags)
5430 req->files_update.offset = READ_ONCE(sqe->off);
5431 req->files_update.nr_args = READ_ONCE(sqe->len);
5432 if (!req->files_update.nr_args)
5434 req->files_update.arg = READ_ONCE(sqe->addr);
5438 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5439 struct io_comp_state *cs)
5441 struct io_ring_ctx *ctx = req->ctx;
5442 struct io_uring_files_update up;
5448 up.offset = req->files_update.offset;
5449 up.fds = req->files_update.arg;
5451 mutex_lock(&ctx->uring_lock);
5452 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5453 mutex_unlock(&ctx->uring_lock);
5456 req_set_fail_links(req);
5457 __io_req_complete(req, ret, 0, cs);
5461 static int io_req_defer_prep(struct io_kiocb *req,
5462 const struct io_uring_sqe *sqe)
5469 if (io_alloc_async_ctx(req))
5471 ret = io_prep_work_files(req);
5475 io_prep_async_work(req);
5477 switch (req->opcode) {
5480 case IORING_OP_READV:
5481 case IORING_OP_READ_FIXED:
5482 case IORING_OP_READ:
5483 ret = io_read_prep(req, sqe, true);
5485 case IORING_OP_WRITEV:
5486 case IORING_OP_WRITE_FIXED:
5487 case IORING_OP_WRITE:
5488 ret = io_write_prep(req, sqe, true);
5490 case IORING_OP_POLL_ADD:
5491 ret = io_poll_add_prep(req, sqe);
5493 case IORING_OP_POLL_REMOVE:
5494 ret = io_poll_remove_prep(req, sqe);
5496 case IORING_OP_FSYNC:
5497 ret = io_prep_fsync(req, sqe);
5499 case IORING_OP_SYNC_FILE_RANGE:
5500 ret = io_prep_sfr(req, sqe);
5502 case IORING_OP_SENDMSG:
5503 case IORING_OP_SEND:
5504 ret = io_sendmsg_prep(req, sqe);
5506 case IORING_OP_RECVMSG:
5507 case IORING_OP_RECV:
5508 ret = io_recvmsg_prep(req, sqe);
5510 case IORING_OP_CONNECT:
5511 ret = io_connect_prep(req, sqe);
5513 case IORING_OP_TIMEOUT:
5514 ret = io_timeout_prep(req, sqe, false);
5516 case IORING_OP_TIMEOUT_REMOVE:
5517 ret = io_timeout_remove_prep(req, sqe);
5519 case IORING_OP_ASYNC_CANCEL:
5520 ret = io_async_cancel_prep(req, sqe);
5522 case IORING_OP_LINK_TIMEOUT:
5523 ret = io_timeout_prep(req, sqe, true);
5525 case IORING_OP_ACCEPT:
5526 ret = io_accept_prep(req, sqe);
5528 case IORING_OP_FALLOCATE:
5529 ret = io_fallocate_prep(req, sqe);
5531 case IORING_OP_OPENAT:
5532 ret = io_openat_prep(req, sqe);
5534 case IORING_OP_CLOSE:
5535 ret = io_close_prep(req, sqe);
5537 case IORING_OP_FILES_UPDATE:
5538 ret = io_files_update_prep(req, sqe);
5540 case IORING_OP_STATX:
5541 ret = io_statx_prep(req, sqe);
5543 case IORING_OP_FADVISE:
5544 ret = io_fadvise_prep(req, sqe);
5546 case IORING_OP_MADVISE:
5547 ret = io_madvise_prep(req, sqe);
5549 case IORING_OP_OPENAT2:
5550 ret = io_openat2_prep(req, sqe);
5552 case IORING_OP_EPOLL_CTL:
5553 ret = io_epoll_ctl_prep(req, sqe);
5555 case IORING_OP_SPLICE:
5556 ret = io_splice_prep(req, sqe);
5558 case IORING_OP_PROVIDE_BUFFERS:
5559 ret = io_provide_buffers_prep(req, sqe);
5561 case IORING_OP_REMOVE_BUFFERS:
5562 ret = io_remove_buffers_prep(req, sqe);
5565 ret = io_tee_prep(req, sqe);
5568 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5577 static u32 io_get_sequence(struct io_kiocb *req)
5579 struct io_kiocb *pos;
5580 struct io_ring_ctx *ctx = req->ctx;
5581 u32 total_submitted, nr_reqs = 1;
5583 if (req->flags & REQ_F_LINK_HEAD)
5584 list_for_each_entry(pos, &req->link_list, link_list)
5587 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5588 return total_submitted - nr_reqs;
5591 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5593 struct io_ring_ctx *ctx = req->ctx;
5594 struct io_defer_entry *de;
5598 /* Still need defer if there is pending req in defer list. */
5599 if (likely(list_empty_careful(&ctx->defer_list) &&
5600 !(req->flags & REQ_F_IO_DRAIN)))
5603 seq = io_get_sequence(req);
5604 /* Still a chance to pass the sequence check */
5605 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5609 ret = io_req_defer_prep(req, sqe);
5613 io_prep_async_link(req);
5614 de = kmalloc(sizeof(*de), GFP_KERNEL);
5618 spin_lock_irq(&ctx->completion_lock);
5619 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5620 spin_unlock_irq(&ctx->completion_lock);
5622 io_queue_async_work(req);
5623 return -EIOCBQUEUED;
5626 trace_io_uring_defer(ctx, req, req->user_data);
5629 list_add_tail(&de->list, &ctx->defer_list);
5630 spin_unlock_irq(&ctx->completion_lock);
5631 return -EIOCBQUEUED;
5634 static void __io_clean_op(struct io_kiocb *req)
5636 struct io_async_ctx *io = req->io;
5638 if (req->flags & REQ_F_BUFFER_SELECTED) {
5639 switch (req->opcode) {
5640 case IORING_OP_READV:
5641 case IORING_OP_READ_FIXED:
5642 case IORING_OP_READ:
5643 kfree((void *)(unsigned long)req->rw.addr);
5645 case IORING_OP_RECVMSG:
5646 case IORING_OP_RECV:
5647 kfree(req->sr_msg.kbuf);
5650 req->flags &= ~REQ_F_BUFFER_SELECTED;
5653 if (req->flags & REQ_F_NEED_CLEANUP) {
5654 switch (req->opcode) {
5655 case IORING_OP_READV:
5656 case IORING_OP_READ_FIXED:
5657 case IORING_OP_READ:
5658 case IORING_OP_WRITEV:
5659 case IORING_OP_WRITE_FIXED:
5660 case IORING_OP_WRITE:
5661 if (io->rw.free_iovec)
5662 kfree(io->rw.free_iovec);
5664 case IORING_OP_RECVMSG:
5665 case IORING_OP_SENDMSG:
5666 if (io->msg.iov != io->msg.fast_iov)
5669 case IORING_OP_SPLICE:
5671 io_put_file(req, req->splice.file_in,
5672 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5675 req->flags &= ~REQ_F_NEED_CLEANUP;
5678 if (req->flags & REQ_F_INFLIGHT) {
5679 struct io_ring_ctx *ctx = req->ctx;
5680 unsigned long flags;
5682 spin_lock_irqsave(&ctx->inflight_lock, flags);
5683 list_del(&req->inflight_entry);
5684 if (waitqueue_active(&ctx->inflight_wait))
5685 wake_up(&ctx->inflight_wait);
5686 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5687 req->flags &= ~REQ_F_INFLIGHT;
5691 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5692 bool force_nonblock, struct io_comp_state *cs)
5694 struct io_ring_ctx *ctx = req->ctx;
5697 switch (req->opcode) {
5699 ret = io_nop(req, cs);
5701 case IORING_OP_READV:
5702 case IORING_OP_READ_FIXED:
5703 case IORING_OP_READ:
5705 ret = io_read_prep(req, sqe, force_nonblock);
5709 ret = io_read(req, force_nonblock, cs);
5711 case IORING_OP_WRITEV:
5712 case IORING_OP_WRITE_FIXED:
5713 case IORING_OP_WRITE:
5715 ret = io_write_prep(req, sqe, force_nonblock);
5719 ret = io_write(req, force_nonblock, cs);
5721 case IORING_OP_FSYNC:
5723 ret = io_prep_fsync(req, sqe);
5727 ret = io_fsync(req, force_nonblock);
5729 case IORING_OP_POLL_ADD:
5731 ret = io_poll_add_prep(req, sqe);
5735 ret = io_poll_add(req);
5737 case IORING_OP_POLL_REMOVE:
5739 ret = io_poll_remove_prep(req, sqe);
5743 ret = io_poll_remove(req);
5745 case IORING_OP_SYNC_FILE_RANGE:
5747 ret = io_prep_sfr(req, sqe);
5751 ret = io_sync_file_range(req, force_nonblock);
5753 case IORING_OP_SENDMSG:
5754 case IORING_OP_SEND:
5756 ret = io_sendmsg_prep(req, sqe);
5760 if (req->opcode == IORING_OP_SENDMSG)
5761 ret = io_sendmsg(req, force_nonblock, cs);
5763 ret = io_send(req, force_nonblock, cs);
5765 case IORING_OP_RECVMSG:
5766 case IORING_OP_RECV:
5768 ret = io_recvmsg_prep(req, sqe);
5772 if (req->opcode == IORING_OP_RECVMSG)
5773 ret = io_recvmsg(req, force_nonblock, cs);
5775 ret = io_recv(req, force_nonblock, cs);
5777 case IORING_OP_TIMEOUT:
5779 ret = io_timeout_prep(req, sqe, false);
5783 ret = io_timeout(req);
5785 case IORING_OP_TIMEOUT_REMOVE:
5787 ret = io_timeout_remove_prep(req, sqe);
5791 ret = io_timeout_remove(req);
5793 case IORING_OP_ACCEPT:
5795 ret = io_accept_prep(req, sqe);
5799 ret = io_accept(req, force_nonblock, cs);
5801 case IORING_OP_CONNECT:
5803 ret = io_connect_prep(req, sqe);
5807 ret = io_connect(req, force_nonblock, cs);
5809 case IORING_OP_ASYNC_CANCEL:
5811 ret = io_async_cancel_prep(req, sqe);
5815 ret = io_async_cancel(req);
5817 case IORING_OP_FALLOCATE:
5819 ret = io_fallocate_prep(req, sqe);
5823 ret = io_fallocate(req, force_nonblock);
5825 case IORING_OP_OPENAT:
5827 ret = io_openat_prep(req, sqe);
5831 ret = io_openat(req, force_nonblock);
5833 case IORING_OP_CLOSE:
5835 ret = io_close_prep(req, sqe);
5839 ret = io_close(req, force_nonblock, cs);
5841 case IORING_OP_FILES_UPDATE:
5843 ret = io_files_update_prep(req, sqe);
5847 ret = io_files_update(req, force_nonblock, cs);
5849 case IORING_OP_STATX:
5851 ret = io_statx_prep(req, sqe);
5855 ret = io_statx(req, force_nonblock);
5857 case IORING_OP_FADVISE:
5859 ret = io_fadvise_prep(req, sqe);
5863 ret = io_fadvise(req, force_nonblock);
5865 case IORING_OP_MADVISE:
5867 ret = io_madvise_prep(req, sqe);
5871 ret = io_madvise(req, force_nonblock);
5873 case IORING_OP_OPENAT2:
5875 ret = io_openat2_prep(req, sqe);
5879 ret = io_openat2(req, force_nonblock);
5881 case IORING_OP_EPOLL_CTL:
5883 ret = io_epoll_ctl_prep(req, sqe);
5887 ret = io_epoll_ctl(req, force_nonblock, cs);
5889 case IORING_OP_SPLICE:
5891 ret = io_splice_prep(req, sqe);
5895 ret = io_splice(req, force_nonblock);
5897 case IORING_OP_PROVIDE_BUFFERS:
5899 ret = io_provide_buffers_prep(req, sqe);
5903 ret = io_provide_buffers(req, force_nonblock, cs);
5905 case IORING_OP_REMOVE_BUFFERS:
5907 ret = io_remove_buffers_prep(req, sqe);
5911 ret = io_remove_buffers(req, force_nonblock, cs);
5915 ret = io_tee_prep(req, sqe);
5919 ret = io_tee(req, force_nonblock);
5929 /* If the op doesn't have a file, we're not polling for it */
5930 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5931 const bool in_async = io_wq_current_is_worker();
5933 /* workqueue context doesn't hold uring_lock, grab it now */
5935 mutex_lock(&ctx->uring_lock);
5937 io_iopoll_req_issued(req);
5940 mutex_unlock(&ctx->uring_lock);
5946 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5948 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5949 struct io_kiocb *timeout;
5952 timeout = io_prep_linked_timeout(req);
5954 io_queue_linked_timeout(timeout);
5956 /* if NO_CANCEL is set, we must still run the work */
5957 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5958 IO_WQ_WORK_CANCEL) {
5964 ret = io_issue_sqe(req, NULL, false, NULL);
5966 * We can get EAGAIN for polled IO even though we're
5967 * forcing a sync submission from here, since we can't
5968 * wait for request slots on the block side.
5977 req_set_fail_links(req);
5978 io_req_complete(req, ret);
5981 return io_steal_work(req);
5984 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5987 struct fixed_file_table *table;
5989 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5990 return table->files[index & IORING_FILE_TABLE_MASK];
5993 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5994 int fd, struct file **out_file, bool fixed)
5996 struct io_ring_ctx *ctx = req->ctx;
6000 if (unlikely(!ctx->file_data ||
6001 (unsigned) fd >= ctx->nr_user_files))
6003 fd = array_index_nospec(fd, ctx->nr_user_files);
6004 file = io_file_from_index(ctx, fd);
6006 req->fixed_file_refs = ctx->file_data->cur_refs;
6007 percpu_ref_get(req->fixed_file_refs);
6010 trace_io_uring_file_get(ctx, fd);
6011 file = __io_file_get(state, fd);
6014 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6021 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6026 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6027 if (unlikely(!fixed && io_async_submit(req->ctx)))
6030 return io_file_get(state, req, fd, &req->file, fixed);
6033 static int io_grab_files(struct io_kiocb *req)
6036 struct io_ring_ctx *ctx = req->ctx;
6038 io_req_init_async(req);
6040 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6042 if (!ctx->ring_file)
6046 spin_lock_irq(&ctx->inflight_lock);
6048 * We use the f_ops->flush() handler to ensure that we can flush
6049 * out work accessing these files if the fd is closed. Check if
6050 * the fd has changed since we started down this path, and disallow
6051 * this operation if it has.
6053 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
6054 list_add(&req->inflight_entry, &ctx->inflight_list);
6055 req->flags |= REQ_F_INFLIGHT;
6056 req->work.files = current->files;
6059 spin_unlock_irq(&ctx->inflight_lock);
6065 static inline int io_prep_work_files(struct io_kiocb *req)
6067 if (!io_op_defs[req->opcode].file_table)
6069 return io_grab_files(req);
6072 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6074 struct io_timeout_data *data = container_of(timer,
6075 struct io_timeout_data, timer);
6076 struct io_kiocb *req = data->req;
6077 struct io_ring_ctx *ctx = req->ctx;
6078 struct io_kiocb *prev = NULL;
6079 unsigned long flags;
6081 spin_lock_irqsave(&ctx->completion_lock, flags);
6084 * We don't expect the list to be empty, that will only happen if we
6085 * race with the completion of the linked work.
6087 if (!list_empty(&req->link_list)) {
6088 prev = list_entry(req->link_list.prev, struct io_kiocb,
6090 if (refcount_inc_not_zero(&prev->refs)) {
6091 list_del_init(&req->link_list);
6092 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6097 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6100 req_set_fail_links(prev);
6101 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6104 io_req_complete(req, -ETIME);
6106 return HRTIMER_NORESTART;
6109 static void __io_queue_linked_timeout(struct io_kiocb *req)
6112 * If the list is now empty, then our linked request finished before
6113 * we got a chance to setup the timer
6115 if (!list_empty(&req->link_list)) {
6116 struct io_timeout_data *data = &req->io->timeout;
6118 data->timer.function = io_link_timeout_fn;
6119 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6124 static void io_queue_linked_timeout(struct io_kiocb *req)
6126 struct io_ring_ctx *ctx = req->ctx;
6128 spin_lock_irq(&ctx->completion_lock);
6129 __io_queue_linked_timeout(req);
6130 spin_unlock_irq(&ctx->completion_lock);
6132 /* drop submission reference */
6136 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6138 struct io_kiocb *nxt;
6140 if (!(req->flags & REQ_F_LINK_HEAD))
6142 if (req->flags & REQ_F_LINK_TIMEOUT)
6145 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6147 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6150 req->flags |= REQ_F_LINK_TIMEOUT;
6154 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6155 struct io_comp_state *cs)
6157 struct io_kiocb *linked_timeout;
6158 struct io_kiocb *nxt;
6159 const struct cred *old_creds = NULL;
6163 linked_timeout = io_prep_linked_timeout(req);
6165 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6166 req->work.creds != current_cred()) {
6168 revert_creds(old_creds);
6169 if (old_creds == req->work.creds)
6170 old_creds = NULL; /* restored original creds */
6172 old_creds = override_creds(req->work.creds);
6175 ret = io_issue_sqe(req, sqe, true, cs);
6178 * We async punt it if the file wasn't marked NOWAIT, or if the file
6179 * doesn't support non-blocking read/write attempts
6181 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6182 if (!io_arm_poll_handler(req)) {
6184 ret = io_prep_work_files(req);
6188 * Queued up for async execution, worker will release
6189 * submit reference when the iocb is actually submitted.
6191 io_queue_async_work(req);
6195 io_queue_linked_timeout(linked_timeout);
6199 if (unlikely(ret)) {
6201 /* un-prep timeout, so it'll be killed as any other linked */
6202 req->flags &= ~REQ_F_LINK_TIMEOUT;
6203 req_set_fail_links(req);
6205 io_req_complete(req, ret);
6209 /* drop submission reference */
6210 nxt = io_put_req_find_next(req);
6212 io_queue_linked_timeout(linked_timeout);
6217 if (req->flags & REQ_F_FORCE_ASYNC)
6223 revert_creds(old_creds);
6226 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6227 struct io_comp_state *cs)
6231 ret = io_req_defer(req, sqe);
6233 if (ret != -EIOCBQUEUED) {
6235 req_set_fail_links(req);
6237 io_req_complete(req, ret);
6239 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6241 ret = io_req_defer_prep(req, sqe);
6247 * Never try inline submit of IOSQE_ASYNC is set, go straight
6248 * to async execution.
6250 io_req_init_async(req);
6251 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6252 io_queue_async_work(req);
6254 __io_queue_sqe(req, sqe, cs);
6258 static inline void io_queue_link_head(struct io_kiocb *req,
6259 struct io_comp_state *cs)
6261 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6263 io_req_complete(req, -ECANCELED);
6265 io_queue_sqe(req, NULL, cs);
6268 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6269 struct io_kiocb **link, struct io_comp_state *cs)
6271 struct io_ring_ctx *ctx = req->ctx;
6275 * If we already have a head request, queue this one for async
6276 * submittal once the head completes. If we don't have a head but
6277 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6278 * submitted sync once the chain is complete. If none of those
6279 * conditions are true (normal request), then just queue it.
6282 struct io_kiocb *head = *link;
6285 * Taking sequential execution of a link, draining both sides
6286 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6287 * requests in the link. So, it drains the head and the
6288 * next after the link request. The last one is done via
6289 * drain_next flag to persist the effect across calls.
6291 if (req->flags & REQ_F_IO_DRAIN) {
6292 head->flags |= REQ_F_IO_DRAIN;
6293 ctx->drain_next = 1;
6295 ret = io_req_defer_prep(req, sqe);
6296 if (unlikely(ret)) {
6297 /* fail even hard links since we don't submit */
6298 head->flags |= REQ_F_FAIL_LINK;
6301 trace_io_uring_link(ctx, req, head);
6302 io_get_req_task(req);
6303 list_add_tail(&req->link_list, &head->link_list);
6305 /* last request of a link, enqueue the link */
6306 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6307 io_queue_link_head(head, cs);
6311 if (unlikely(ctx->drain_next)) {
6312 req->flags |= REQ_F_IO_DRAIN;
6313 ctx->drain_next = 0;
6315 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6316 req->flags |= REQ_F_LINK_HEAD;
6317 INIT_LIST_HEAD(&req->link_list);
6319 ret = io_req_defer_prep(req, sqe);
6321 req->flags |= REQ_F_FAIL_LINK;
6324 io_queue_sqe(req, sqe, cs);
6332 * Batched submission is done, ensure local IO is flushed out.
6334 static void io_submit_state_end(struct io_submit_state *state)
6336 if (!list_empty(&state->comp.list))
6337 io_submit_flush_completions(&state->comp);
6338 blk_finish_plug(&state->plug);
6339 io_state_file_put(state);
6340 if (state->free_reqs)
6341 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6345 * Start submission side cache.
6347 static void io_submit_state_start(struct io_submit_state *state,
6348 struct io_ring_ctx *ctx, unsigned int max_ios)
6350 blk_start_plug(&state->plug);
6352 state->plug.nowait = true;
6355 INIT_LIST_HEAD(&state->comp.list);
6356 state->comp.ctx = ctx;
6357 state->free_reqs = 0;
6359 state->ios_left = max_ios;
6362 static void io_commit_sqring(struct io_ring_ctx *ctx)
6364 struct io_rings *rings = ctx->rings;
6367 * Ensure any loads from the SQEs are done at this point,
6368 * since once we write the new head, the application could
6369 * write new data to them.
6371 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6375 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6376 * that is mapped by userspace. This means that care needs to be taken to
6377 * ensure that reads are stable, as we cannot rely on userspace always
6378 * being a good citizen. If members of the sqe are validated and then later
6379 * used, it's important that those reads are done through READ_ONCE() to
6380 * prevent a re-load down the line.
6382 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6384 u32 *sq_array = ctx->sq_array;
6388 * The cached sq head (or cq tail) serves two purposes:
6390 * 1) allows us to batch the cost of updating the user visible
6392 * 2) allows the kernel side to track the head on its own, even
6393 * though the application is the one updating it.
6395 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6396 if (likely(head < ctx->sq_entries))
6397 return &ctx->sq_sqes[head];
6399 /* drop invalid entries */
6400 ctx->cached_sq_dropped++;
6401 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6405 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6407 ctx->cached_sq_head++;
6410 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6411 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6412 IOSQE_BUFFER_SELECT)
6414 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6415 const struct io_uring_sqe *sqe,
6416 struct io_submit_state *state)
6418 unsigned int sqe_flags;
6421 req->opcode = READ_ONCE(sqe->opcode);
6422 req->user_data = READ_ONCE(sqe->user_data);
6427 /* one is dropped after submission, the other at completion */
6428 refcount_set(&req->refs, 2);
6429 req->task = current;
6432 if (unlikely(req->opcode >= IORING_OP_LAST))
6435 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6438 sqe_flags = READ_ONCE(sqe->flags);
6439 /* enforce forwards compatibility on users */
6440 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6443 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6444 !io_op_defs[req->opcode].buffer_select)
6447 id = READ_ONCE(sqe->personality);
6449 io_req_init_async(req);
6450 req->work.creds = idr_find(&ctx->personality_idr, id);
6451 if (unlikely(!req->work.creds))
6453 get_cred(req->work.creds);
6456 /* same numerical values with corresponding REQ_F_*, safe to copy */
6457 req->flags |= sqe_flags;
6459 if (!io_op_defs[req->opcode].needs_file)
6462 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6465 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6466 struct file *ring_file, int ring_fd)
6468 struct io_submit_state state;
6469 struct io_kiocb *link = NULL;
6470 int i, submitted = 0;
6472 /* if we have a backlog and couldn't flush it all, return BUSY */
6473 if (test_bit(0, &ctx->sq_check_overflow)) {
6474 if (!list_empty(&ctx->cq_overflow_list) &&
6475 !io_cqring_overflow_flush(ctx, false))
6479 /* make sure SQ entry isn't read before tail */
6480 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6482 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6485 io_submit_state_start(&state, ctx, nr);
6487 ctx->ring_fd = ring_fd;
6488 ctx->ring_file = ring_file;
6490 for (i = 0; i < nr; i++) {
6491 const struct io_uring_sqe *sqe;
6492 struct io_kiocb *req;
6495 sqe = io_get_sqe(ctx);
6496 if (unlikely(!sqe)) {
6497 io_consume_sqe(ctx);
6500 req = io_alloc_req(ctx, &state);
6501 if (unlikely(!req)) {
6503 submitted = -EAGAIN;
6507 err = io_init_req(ctx, req, sqe, &state);
6508 io_consume_sqe(ctx);
6509 /* will complete beyond this point, count as submitted */
6512 if (unlikely(err)) {
6515 io_req_complete(req, err);
6519 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6520 true, io_async_submit(ctx));
6521 err = io_submit_sqe(req, sqe, &link, &state.comp);
6526 if (unlikely(submitted != nr)) {
6527 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6529 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6532 io_queue_link_head(link, &state.comp);
6533 io_submit_state_end(&state);
6535 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6536 io_commit_sqring(ctx);
6541 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6543 /* Tell userspace we may need a wakeup call */
6544 spin_lock_irq(&ctx->completion_lock);
6545 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6546 spin_unlock_irq(&ctx->completion_lock);
6549 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6551 spin_lock_irq(&ctx->completion_lock);
6552 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6553 spin_unlock_irq(&ctx->completion_lock);
6556 static int io_sq_thread(void *data)
6558 struct io_ring_ctx *ctx = data;
6559 const struct cred *old_cred;
6561 unsigned long timeout;
6564 complete(&ctx->sq_thread_comp);
6566 old_cred = override_creds(ctx->creds);
6568 timeout = jiffies + ctx->sq_thread_idle;
6569 while (!kthread_should_park()) {
6570 unsigned int to_submit;
6572 if (!list_empty(&ctx->iopoll_list)) {
6573 unsigned nr_events = 0;
6575 mutex_lock(&ctx->uring_lock);
6576 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6577 io_do_iopoll(ctx, &nr_events, 0);
6579 timeout = jiffies + ctx->sq_thread_idle;
6580 mutex_unlock(&ctx->uring_lock);
6583 to_submit = io_sqring_entries(ctx);
6586 * If submit got -EBUSY, flag us as needing the application
6587 * to enter the kernel to reap and flush events.
6589 if (!to_submit || ret == -EBUSY || need_resched()) {
6591 * Drop cur_mm before scheduling, we can't hold it for
6592 * long periods (or over schedule()). Do this before
6593 * adding ourselves to the waitqueue, as the unuse/drop
6596 io_sq_thread_drop_mm();
6599 * We're polling. If we're within the defined idle
6600 * period, then let us spin without work before going
6601 * to sleep. The exception is if we got EBUSY doing
6602 * more IO, we should wait for the application to
6603 * reap events and wake us up.
6605 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6606 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6607 !percpu_ref_is_dying(&ctx->refs))) {
6613 prepare_to_wait(&ctx->sqo_wait, &wait,
6614 TASK_INTERRUPTIBLE);
6617 * While doing polled IO, before going to sleep, we need
6618 * to check if there are new reqs added to iopoll_list,
6619 * it is because reqs may have been punted to io worker
6620 * and will be added to iopoll_list later, hence check
6621 * the iopoll_list again.
6623 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6624 !list_empty_careful(&ctx->iopoll_list)) {
6625 finish_wait(&ctx->sqo_wait, &wait);
6629 io_ring_set_wakeup_flag(ctx);
6631 to_submit = io_sqring_entries(ctx);
6632 if (!to_submit || ret == -EBUSY) {
6633 if (kthread_should_park()) {
6634 finish_wait(&ctx->sqo_wait, &wait);
6637 if (io_run_task_work()) {
6638 finish_wait(&ctx->sqo_wait, &wait);
6639 io_ring_clear_wakeup_flag(ctx);
6642 if (signal_pending(current))
6643 flush_signals(current);
6645 finish_wait(&ctx->sqo_wait, &wait);
6647 io_ring_clear_wakeup_flag(ctx);
6651 finish_wait(&ctx->sqo_wait, &wait);
6653 io_ring_clear_wakeup_flag(ctx);
6656 mutex_lock(&ctx->uring_lock);
6657 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6658 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6659 mutex_unlock(&ctx->uring_lock);
6660 timeout = jiffies + ctx->sq_thread_idle;
6665 io_sq_thread_drop_mm();
6666 revert_creds(old_cred);
6673 struct io_wait_queue {
6674 struct wait_queue_entry wq;
6675 struct io_ring_ctx *ctx;
6677 unsigned nr_timeouts;
6680 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6682 struct io_ring_ctx *ctx = iowq->ctx;
6685 * Wake up if we have enough events, or if a timeout occurred since we
6686 * started waiting. For timeouts, we always want to return to userspace,
6687 * regardless of event count.
6689 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6690 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6693 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6694 int wake_flags, void *key)
6696 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6699 /* use noflush == true, as we can't safely rely on locking context */
6700 if (!io_should_wake(iowq, true))
6703 return autoremove_wake_function(curr, mode, wake_flags, key);
6707 * Wait until events become available, if we don't already have some. The
6708 * application must reap them itself, as they reside on the shared cq ring.
6710 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6711 const sigset_t __user *sig, size_t sigsz)
6713 struct io_wait_queue iowq = {
6716 .func = io_wake_function,
6717 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6720 .to_wait = min_events,
6722 struct io_rings *rings = ctx->rings;
6726 if (io_cqring_events(ctx, false) >= min_events)
6728 if (!io_run_task_work())
6733 #ifdef CONFIG_COMPAT
6734 if (in_compat_syscall())
6735 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6739 ret = set_user_sigmask(sig, sigsz);
6745 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6746 trace_io_uring_cqring_wait(ctx, min_events);
6748 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6749 TASK_INTERRUPTIBLE);
6750 /* make sure we run task_work before checking for signals */
6751 if (io_run_task_work())
6753 if (signal_pending(current)) {
6754 if (current->jobctl & JOBCTL_TASK_WORK) {
6755 spin_lock_irq(¤t->sighand->siglock);
6756 current->jobctl &= ~JOBCTL_TASK_WORK;
6757 recalc_sigpending();
6758 spin_unlock_irq(¤t->sighand->siglock);
6764 if (io_should_wake(&iowq, false))
6768 finish_wait(&ctx->wait, &iowq.wq);
6770 restore_saved_sigmask_unless(ret == -EINTR);
6772 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6775 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6777 #if defined(CONFIG_UNIX)
6778 if (ctx->ring_sock) {
6779 struct sock *sock = ctx->ring_sock->sk;
6780 struct sk_buff *skb;
6782 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6788 for (i = 0; i < ctx->nr_user_files; i++) {
6791 file = io_file_from_index(ctx, i);
6798 static void io_file_ref_kill(struct percpu_ref *ref)
6800 struct fixed_file_data *data;
6802 data = container_of(ref, struct fixed_file_data, refs);
6803 complete(&data->done);
6806 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6808 struct fixed_file_data *data = ctx->file_data;
6809 struct fixed_file_ref_node *ref_node = NULL;
6810 unsigned nr_tables, i;
6815 spin_lock(&data->lock);
6816 if (!list_empty(&data->ref_list))
6817 ref_node = list_first_entry(&data->ref_list,
6818 struct fixed_file_ref_node, node);
6819 spin_unlock(&data->lock);
6821 percpu_ref_kill(&ref_node->refs);
6823 percpu_ref_kill(&data->refs);
6825 /* wait for all refs nodes to complete */
6826 flush_delayed_work(&ctx->file_put_work);
6827 wait_for_completion(&data->done);
6829 __io_sqe_files_unregister(ctx);
6830 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6831 for (i = 0; i < nr_tables; i++)
6832 kfree(data->table[i].files);
6834 percpu_ref_exit(&data->refs);
6836 ctx->file_data = NULL;
6837 ctx->nr_user_files = 0;
6841 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6843 if (ctx->sqo_thread) {
6844 wait_for_completion(&ctx->sq_thread_comp);
6846 * The park is a bit of a work-around, without it we get
6847 * warning spews on shutdown with SQPOLL set and affinity
6848 * set to a single CPU.
6850 kthread_park(ctx->sqo_thread);
6851 kthread_stop(ctx->sqo_thread);
6852 ctx->sqo_thread = NULL;
6856 static void io_finish_async(struct io_ring_ctx *ctx)
6858 io_sq_thread_stop(ctx);
6861 io_wq_destroy(ctx->io_wq);
6866 #if defined(CONFIG_UNIX)
6868 * Ensure the UNIX gc is aware of our file set, so we are certain that
6869 * the io_uring can be safely unregistered on process exit, even if we have
6870 * loops in the file referencing.
6872 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6874 struct sock *sk = ctx->ring_sock->sk;
6875 struct scm_fp_list *fpl;
6876 struct sk_buff *skb;
6879 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6883 skb = alloc_skb(0, GFP_KERNEL);
6892 fpl->user = get_uid(ctx->user);
6893 for (i = 0; i < nr; i++) {
6894 struct file *file = io_file_from_index(ctx, i + offset);
6898 fpl->fp[nr_files] = get_file(file);
6899 unix_inflight(fpl->user, fpl->fp[nr_files]);
6904 fpl->max = SCM_MAX_FD;
6905 fpl->count = nr_files;
6906 UNIXCB(skb).fp = fpl;
6907 skb->destructor = unix_destruct_scm;
6908 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6909 skb_queue_head(&sk->sk_receive_queue, skb);
6911 for (i = 0; i < nr_files; i++)
6922 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6923 * causes regular reference counting to break down. We rely on the UNIX
6924 * garbage collection to take care of this problem for us.
6926 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6928 unsigned left, total;
6932 left = ctx->nr_user_files;
6934 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6936 ret = __io_sqe_files_scm(ctx, this_files, total);
6940 total += this_files;
6946 while (total < ctx->nr_user_files) {
6947 struct file *file = io_file_from_index(ctx, total);
6957 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6963 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6968 for (i = 0; i < nr_tables; i++) {
6969 struct fixed_file_table *table = &ctx->file_data->table[i];
6970 unsigned this_files;
6972 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6973 table->files = kcalloc(this_files, sizeof(struct file *),
6977 nr_files -= this_files;
6983 for (i = 0; i < nr_tables; i++) {
6984 struct fixed_file_table *table = &ctx->file_data->table[i];
6985 kfree(table->files);
6990 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6992 #if defined(CONFIG_UNIX)
6993 struct sock *sock = ctx->ring_sock->sk;
6994 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6995 struct sk_buff *skb;
6998 __skb_queue_head_init(&list);
7001 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7002 * remove this entry and rearrange the file array.
7004 skb = skb_dequeue(head);
7006 struct scm_fp_list *fp;
7008 fp = UNIXCB(skb).fp;
7009 for (i = 0; i < fp->count; i++) {
7012 if (fp->fp[i] != file)
7015 unix_notinflight(fp->user, fp->fp[i]);
7016 left = fp->count - 1 - i;
7018 memmove(&fp->fp[i], &fp->fp[i + 1],
7019 left * sizeof(struct file *));
7026 __skb_queue_tail(&list, skb);
7036 __skb_queue_tail(&list, skb);
7038 skb = skb_dequeue(head);
7041 if (skb_peek(&list)) {
7042 spin_lock_irq(&head->lock);
7043 while ((skb = __skb_dequeue(&list)) != NULL)
7044 __skb_queue_tail(head, skb);
7045 spin_unlock_irq(&head->lock);
7052 struct io_file_put {
7053 struct list_head list;
7057 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7059 struct fixed_file_data *file_data = ref_node->file_data;
7060 struct io_ring_ctx *ctx = file_data->ctx;
7061 struct io_file_put *pfile, *tmp;
7063 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7064 list_del(&pfile->list);
7065 io_ring_file_put(ctx, pfile->file);
7069 spin_lock(&file_data->lock);
7070 list_del(&ref_node->node);
7071 spin_unlock(&file_data->lock);
7073 percpu_ref_exit(&ref_node->refs);
7075 percpu_ref_put(&file_data->refs);
7078 static void io_file_put_work(struct work_struct *work)
7080 struct io_ring_ctx *ctx;
7081 struct llist_node *node;
7083 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7084 node = llist_del_all(&ctx->file_put_llist);
7087 struct fixed_file_ref_node *ref_node;
7088 struct llist_node *next = node->next;
7090 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7091 __io_file_put_work(ref_node);
7096 static void io_file_data_ref_zero(struct percpu_ref *ref)
7098 struct fixed_file_ref_node *ref_node;
7099 struct io_ring_ctx *ctx;
7103 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7104 ctx = ref_node->file_data->ctx;
7106 if (percpu_ref_is_dying(&ctx->file_data->refs))
7109 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7111 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7113 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7116 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7117 struct io_ring_ctx *ctx)
7119 struct fixed_file_ref_node *ref_node;
7121 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7123 return ERR_PTR(-ENOMEM);
7125 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7128 return ERR_PTR(-ENOMEM);
7130 INIT_LIST_HEAD(&ref_node->node);
7131 INIT_LIST_HEAD(&ref_node->file_list);
7132 ref_node->file_data = ctx->file_data;
7136 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7138 percpu_ref_exit(&ref_node->refs);
7142 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7145 __s32 __user *fds = (__s32 __user *) arg;
7150 struct fixed_file_ref_node *ref_node;
7156 if (nr_args > IORING_MAX_FIXED_FILES)
7159 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7160 if (!ctx->file_data)
7162 ctx->file_data->ctx = ctx;
7163 init_completion(&ctx->file_data->done);
7164 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7165 spin_lock_init(&ctx->file_data->lock);
7167 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7168 ctx->file_data->table = kcalloc(nr_tables,
7169 sizeof(struct fixed_file_table),
7171 if (!ctx->file_data->table) {
7172 kfree(ctx->file_data);
7173 ctx->file_data = NULL;
7177 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7178 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7179 kfree(ctx->file_data->table);
7180 kfree(ctx->file_data);
7181 ctx->file_data = NULL;
7185 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7186 percpu_ref_exit(&ctx->file_data->refs);
7187 kfree(ctx->file_data->table);
7188 kfree(ctx->file_data);
7189 ctx->file_data = NULL;
7193 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7194 struct fixed_file_table *table;
7198 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7200 /* allow sparse sets */
7206 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7207 index = i & IORING_FILE_TABLE_MASK;
7215 * Don't allow io_uring instances to be registered. If UNIX
7216 * isn't enabled, then this causes a reference cycle and this
7217 * instance can never get freed. If UNIX is enabled we'll
7218 * handle it just fine, but there's still no point in allowing
7219 * a ring fd as it doesn't support regular read/write anyway.
7221 if (file->f_op == &io_uring_fops) {
7226 table->files[index] = file;
7230 for (i = 0; i < ctx->nr_user_files; i++) {
7231 file = io_file_from_index(ctx, i);
7235 for (i = 0; i < nr_tables; i++)
7236 kfree(ctx->file_data->table[i].files);
7238 percpu_ref_exit(&ctx->file_data->refs);
7239 kfree(ctx->file_data->table);
7240 kfree(ctx->file_data);
7241 ctx->file_data = NULL;
7242 ctx->nr_user_files = 0;
7246 ret = io_sqe_files_scm(ctx);
7248 io_sqe_files_unregister(ctx);
7252 ref_node = alloc_fixed_file_ref_node(ctx);
7253 if (IS_ERR(ref_node)) {
7254 io_sqe_files_unregister(ctx);
7255 return PTR_ERR(ref_node);
7258 ctx->file_data->cur_refs = &ref_node->refs;
7259 spin_lock(&ctx->file_data->lock);
7260 list_add(&ref_node->node, &ctx->file_data->ref_list);
7261 spin_unlock(&ctx->file_data->lock);
7262 percpu_ref_get(&ctx->file_data->refs);
7266 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7269 #if defined(CONFIG_UNIX)
7270 struct sock *sock = ctx->ring_sock->sk;
7271 struct sk_buff_head *head = &sock->sk_receive_queue;
7272 struct sk_buff *skb;
7275 * See if we can merge this file into an existing skb SCM_RIGHTS
7276 * file set. If there's no room, fall back to allocating a new skb
7277 * and filling it in.
7279 spin_lock_irq(&head->lock);
7280 skb = skb_peek(head);
7282 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7284 if (fpl->count < SCM_MAX_FD) {
7285 __skb_unlink(skb, head);
7286 spin_unlock_irq(&head->lock);
7287 fpl->fp[fpl->count] = get_file(file);
7288 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7290 spin_lock_irq(&head->lock);
7291 __skb_queue_head(head, skb);
7296 spin_unlock_irq(&head->lock);
7303 return __io_sqe_files_scm(ctx, 1, index);
7309 static int io_queue_file_removal(struct fixed_file_data *data,
7312 struct io_file_put *pfile;
7313 struct percpu_ref *refs = data->cur_refs;
7314 struct fixed_file_ref_node *ref_node;
7316 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7320 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7322 list_add(&pfile->list, &ref_node->file_list);
7327 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7328 struct io_uring_files_update *up,
7331 struct fixed_file_data *data = ctx->file_data;
7332 struct fixed_file_ref_node *ref_node;
7337 bool needs_switch = false;
7339 if (check_add_overflow(up->offset, nr_args, &done))
7341 if (done > ctx->nr_user_files)
7344 ref_node = alloc_fixed_file_ref_node(ctx);
7345 if (IS_ERR(ref_node))
7346 return PTR_ERR(ref_node);
7349 fds = u64_to_user_ptr(up->fds);
7351 struct fixed_file_table *table;
7355 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7359 i = array_index_nospec(up->offset, ctx->nr_user_files);
7360 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7361 index = i & IORING_FILE_TABLE_MASK;
7362 if (table->files[index]) {
7363 file = table->files[index];
7364 err = io_queue_file_removal(data, file);
7367 table->files[index] = NULL;
7368 needs_switch = true;
7377 * Don't allow io_uring instances to be registered. If
7378 * UNIX isn't enabled, then this causes a reference
7379 * cycle and this instance can never get freed. If UNIX
7380 * is enabled we'll handle it just fine, but there's
7381 * still no point in allowing a ring fd as it doesn't
7382 * support regular read/write anyway.
7384 if (file->f_op == &io_uring_fops) {
7389 table->files[index] = file;
7390 err = io_sqe_file_register(ctx, file, i);
7392 table->files[index] = NULL;
7403 percpu_ref_kill(data->cur_refs);
7404 spin_lock(&data->lock);
7405 list_add(&ref_node->node, &data->ref_list);
7406 data->cur_refs = &ref_node->refs;
7407 spin_unlock(&data->lock);
7408 percpu_ref_get(&ctx->file_data->refs);
7410 destroy_fixed_file_ref_node(ref_node);
7412 return done ? done : err;
7415 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7418 struct io_uring_files_update up;
7420 if (!ctx->file_data)
7424 if (copy_from_user(&up, arg, sizeof(up)))
7429 return __io_sqe_files_update(ctx, &up, nr_args);
7432 static void io_free_work(struct io_wq_work *work)
7434 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7436 /* Consider that io_steal_work() relies on this ref */
7440 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7441 struct io_uring_params *p)
7443 struct io_wq_data data;
7445 struct io_ring_ctx *ctx_attach;
7446 unsigned int concurrency;
7449 data.user = ctx->user;
7450 data.free_work = io_free_work;
7451 data.do_work = io_wq_submit_work;
7453 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7454 /* Do QD, or 4 * CPUS, whatever is smallest */
7455 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7457 ctx->io_wq = io_wq_create(concurrency, &data);
7458 if (IS_ERR(ctx->io_wq)) {
7459 ret = PTR_ERR(ctx->io_wq);
7465 f = fdget(p->wq_fd);
7469 if (f.file->f_op != &io_uring_fops) {
7474 ctx_attach = f.file->private_data;
7475 /* @io_wq is protected by holding the fd */
7476 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7481 ctx->io_wq = ctx_attach->io_wq;
7487 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7488 struct io_uring_params *p)
7492 if (ctx->flags & IORING_SETUP_SQPOLL) {
7494 if (!capable(CAP_SYS_ADMIN))
7497 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7498 if (!ctx->sq_thread_idle)
7499 ctx->sq_thread_idle = HZ;
7501 if (p->flags & IORING_SETUP_SQ_AFF) {
7502 int cpu = p->sq_thread_cpu;
7505 if (cpu >= nr_cpu_ids)
7507 if (!cpu_online(cpu))
7510 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7514 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7517 if (IS_ERR(ctx->sqo_thread)) {
7518 ret = PTR_ERR(ctx->sqo_thread);
7519 ctx->sqo_thread = NULL;
7522 wake_up_process(ctx->sqo_thread);
7523 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7524 /* Can't have SQ_AFF without SQPOLL */
7529 ret = io_init_wq_offload(ctx, p);
7535 io_finish_async(ctx);
7539 static inline void __io_unaccount_mem(struct user_struct *user,
7540 unsigned long nr_pages)
7542 atomic_long_sub(nr_pages, &user->locked_vm);
7545 static inline int __io_account_mem(struct user_struct *user,
7546 unsigned long nr_pages)
7548 unsigned long page_limit, cur_pages, new_pages;
7550 /* Don't allow more pages than we can safely lock */
7551 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7554 cur_pages = atomic_long_read(&user->locked_vm);
7555 new_pages = cur_pages + nr_pages;
7556 if (new_pages > page_limit)
7558 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7559 new_pages) != cur_pages);
7564 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7565 enum io_mem_account acct)
7568 __io_unaccount_mem(ctx->user, nr_pages);
7571 if (acct == ACCT_LOCKED)
7572 ctx->sqo_mm->locked_vm -= nr_pages;
7573 else if (acct == ACCT_PINNED)
7574 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7578 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7579 enum io_mem_account acct)
7583 if (ctx->limit_mem) {
7584 ret = __io_account_mem(ctx->user, nr_pages);
7590 if (acct == ACCT_LOCKED)
7591 ctx->sqo_mm->locked_vm += nr_pages;
7592 else if (acct == ACCT_PINNED)
7593 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7599 static void io_mem_free(void *ptr)
7606 page = virt_to_head_page(ptr);
7607 if (put_page_testzero(page))
7608 free_compound_page(page);
7611 static void *io_mem_alloc(size_t size)
7613 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7616 return (void *) __get_free_pages(gfp_flags, get_order(size));
7619 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7622 struct io_rings *rings;
7623 size_t off, sq_array_size;
7625 off = struct_size(rings, cqes, cq_entries);
7626 if (off == SIZE_MAX)
7630 off = ALIGN(off, SMP_CACHE_BYTES);
7638 sq_array_size = array_size(sizeof(u32), sq_entries);
7639 if (sq_array_size == SIZE_MAX)
7642 if (check_add_overflow(off, sq_array_size, &off))
7648 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7652 pages = (size_t)1 << get_order(
7653 rings_size(sq_entries, cq_entries, NULL));
7654 pages += (size_t)1 << get_order(
7655 array_size(sizeof(struct io_uring_sqe), sq_entries));
7660 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7664 if (!ctx->user_bufs)
7667 for (i = 0; i < ctx->nr_user_bufs; i++) {
7668 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7670 for (j = 0; j < imu->nr_bvecs; j++)
7671 unpin_user_page(imu->bvec[j].bv_page);
7673 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7678 kfree(ctx->user_bufs);
7679 ctx->user_bufs = NULL;
7680 ctx->nr_user_bufs = 0;
7684 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7685 void __user *arg, unsigned index)
7687 struct iovec __user *src;
7689 #ifdef CONFIG_COMPAT
7691 struct compat_iovec __user *ciovs;
7692 struct compat_iovec ciov;
7694 ciovs = (struct compat_iovec __user *) arg;
7695 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7698 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7699 dst->iov_len = ciov.iov_len;
7703 src = (struct iovec __user *) arg;
7704 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7709 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7712 struct vm_area_struct **vmas = NULL;
7713 struct page **pages = NULL;
7714 int i, j, got_pages = 0;
7719 if (!nr_args || nr_args > UIO_MAXIOV)
7722 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7724 if (!ctx->user_bufs)
7727 for (i = 0; i < nr_args; i++) {
7728 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7729 unsigned long off, start, end, ubuf;
7734 ret = io_copy_iov(ctx, &iov, arg, i);
7739 * Don't impose further limits on the size and buffer
7740 * constraints here, we'll -EINVAL later when IO is
7741 * submitted if they are wrong.
7744 if (!iov.iov_base || !iov.iov_len)
7747 /* arbitrary limit, but we need something */
7748 if (iov.iov_len > SZ_1G)
7751 ubuf = (unsigned long) iov.iov_base;
7752 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7753 start = ubuf >> PAGE_SHIFT;
7754 nr_pages = end - start;
7756 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7761 if (!pages || nr_pages > got_pages) {
7764 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7766 vmas = kvmalloc_array(nr_pages,
7767 sizeof(struct vm_area_struct *),
7769 if (!pages || !vmas) {
7771 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7774 got_pages = nr_pages;
7777 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7781 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7786 mmap_read_lock(current->mm);
7787 pret = pin_user_pages(ubuf, nr_pages,
7788 FOLL_WRITE | FOLL_LONGTERM,
7790 if (pret == nr_pages) {
7791 /* don't support file backed memory */
7792 for (j = 0; j < nr_pages; j++) {
7793 struct vm_area_struct *vma = vmas[j];
7796 !is_file_hugepages(vma->vm_file)) {
7802 ret = pret < 0 ? pret : -EFAULT;
7804 mmap_read_unlock(current->mm);
7807 * if we did partial map, or found file backed vmas,
7808 * release any pages we did get
7811 unpin_user_pages(pages, pret);
7812 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7817 off = ubuf & ~PAGE_MASK;
7819 for (j = 0; j < nr_pages; j++) {
7822 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7823 imu->bvec[j].bv_page = pages[j];
7824 imu->bvec[j].bv_len = vec_len;
7825 imu->bvec[j].bv_offset = off;
7829 /* store original address for later verification */
7831 imu->len = iov.iov_len;
7832 imu->nr_bvecs = nr_pages;
7834 ctx->nr_user_bufs++;
7842 io_sqe_buffer_unregister(ctx);
7846 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7848 __s32 __user *fds = arg;
7854 if (copy_from_user(&fd, fds, sizeof(*fds)))
7857 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7858 if (IS_ERR(ctx->cq_ev_fd)) {
7859 int ret = PTR_ERR(ctx->cq_ev_fd);
7860 ctx->cq_ev_fd = NULL;
7867 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7869 if (ctx->cq_ev_fd) {
7870 eventfd_ctx_put(ctx->cq_ev_fd);
7871 ctx->cq_ev_fd = NULL;
7878 static int __io_destroy_buffers(int id, void *p, void *data)
7880 struct io_ring_ctx *ctx = data;
7881 struct io_buffer *buf = p;
7883 __io_remove_buffers(ctx, buf, id, -1U);
7887 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7889 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7890 idr_destroy(&ctx->io_buffer_idr);
7893 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7895 io_finish_async(ctx);
7896 io_sqe_buffer_unregister(ctx);
7898 mmdrop(ctx->sqo_mm);
7902 io_sqe_files_unregister(ctx);
7903 io_eventfd_unregister(ctx);
7904 io_destroy_buffers(ctx);
7905 idr_destroy(&ctx->personality_idr);
7907 #if defined(CONFIG_UNIX)
7908 if (ctx->ring_sock) {
7909 ctx->ring_sock->file = NULL; /* so that iput() is called */
7910 sock_release(ctx->ring_sock);
7914 io_mem_free(ctx->rings);
7915 io_mem_free(ctx->sq_sqes);
7917 percpu_ref_exit(&ctx->refs);
7918 free_uid(ctx->user);
7919 put_cred(ctx->creds);
7920 kfree(ctx->cancel_hash);
7921 kmem_cache_free(req_cachep, ctx->fallback_req);
7925 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7927 struct io_ring_ctx *ctx = file->private_data;
7930 poll_wait(file, &ctx->cq_wait, wait);
7932 * synchronizes with barrier from wq_has_sleeper call in
7936 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7937 ctx->rings->sq_ring_entries)
7938 mask |= EPOLLOUT | EPOLLWRNORM;
7939 if (io_cqring_events(ctx, false))
7940 mask |= EPOLLIN | EPOLLRDNORM;
7945 static int io_uring_fasync(int fd, struct file *file, int on)
7947 struct io_ring_ctx *ctx = file->private_data;
7949 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7952 static int io_remove_personalities(int id, void *p, void *data)
7954 struct io_ring_ctx *ctx = data;
7955 const struct cred *cred;
7957 cred = idr_remove(&ctx->personality_idr, id);
7963 static void io_ring_exit_work(struct work_struct *work)
7965 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7969 * If we're doing polled IO and end up having requests being
7970 * submitted async (out-of-line), then completions can come in while
7971 * we're waiting for refs to drop. We need to reap these manually,
7972 * as nobody else will be looking for them.
7976 io_cqring_overflow_flush(ctx, true);
7977 io_iopoll_try_reap_events(ctx);
7978 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7979 io_ring_ctx_free(ctx);
7982 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7984 mutex_lock(&ctx->uring_lock);
7985 percpu_ref_kill(&ctx->refs);
7986 mutex_unlock(&ctx->uring_lock);
7988 io_kill_timeouts(ctx);
7989 io_poll_remove_all(ctx);
7992 io_wq_cancel_all(ctx->io_wq);
7994 /* if we failed setting up the ctx, we might not have any rings */
7996 io_cqring_overflow_flush(ctx, true);
7997 io_iopoll_try_reap_events(ctx);
7998 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8001 * Do this upfront, so we won't have a grace period where the ring
8002 * is closed but resources aren't reaped yet. This can cause
8003 * spurious failure in setting up a new ring.
8005 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8008 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8010 * Use system_unbound_wq to avoid spawning tons of event kworkers
8011 * if we're exiting a ton of rings at the same time. It just adds
8012 * noise and overhead, there's no discernable change in runtime
8013 * over using system_wq.
8015 queue_work(system_unbound_wq, &ctx->exit_work);
8018 static int io_uring_release(struct inode *inode, struct file *file)
8020 struct io_ring_ctx *ctx = file->private_data;
8022 file->private_data = NULL;
8023 io_ring_ctx_wait_and_kill(ctx);
8027 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8029 struct files_struct *files = data;
8031 return work->files == files;
8035 * Returns true if 'preq' is the link parent of 'req'
8037 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8039 struct io_kiocb *link;
8041 if (!(preq->flags & REQ_F_LINK_HEAD))
8044 list_for_each_entry(link, &preq->link_list, link_list) {
8052 static inline bool io_match_files(struct io_kiocb *req,
8053 struct files_struct *files)
8055 return (req->flags & REQ_F_WORK_INITIALIZED) && req->work.files == files;
8058 static bool io_match_link_files(struct io_kiocb *req,
8059 struct files_struct *files)
8061 struct io_kiocb *link;
8063 if (io_match_files(req, files))
8065 if (req->flags & REQ_F_LINK_HEAD) {
8066 list_for_each_entry(link, &req->link_list, link_list) {
8067 if (io_match_files(link, files))
8075 * We're looking to cancel 'req' because it's holding on to our files, but
8076 * 'req' could be a link to another request. See if it is, and cancel that
8077 * parent request if so.
8079 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8081 struct hlist_node *tmp;
8082 struct io_kiocb *preq;
8086 spin_lock_irq(&ctx->completion_lock);
8087 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8088 struct hlist_head *list;
8090 list = &ctx->cancel_hash[i];
8091 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8092 found = io_match_link(preq, req);
8094 io_poll_remove_one(preq);
8099 spin_unlock_irq(&ctx->completion_lock);
8103 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8104 struct io_kiocb *req)
8106 struct io_kiocb *preq;
8109 spin_lock_irq(&ctx->completion_lock);
8110 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8111 found = io_match_link(preq, req);
8113 __io_timeout_cancel(preq);
8117 spin_unlock_irq(&ctx->completion_lock);
8121 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8123 return io_match_link(container_of(work, struct io_kiocb, work), data);
8126 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8128 enum io_wq_cancel cret;
8130 /* cancel this particular work, if it's running */
8131 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8132 if (cret != IO_WQ_CANCEL_NOTFOUND)
8135 /* find links that hold this pending, cancel those */
8136 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8137 if (cret != IO_WQ_CANCEL_NOTFOUND)
8140 /* if we have a poll link holding this pending, cancel that */
8141 if (io_poll_remove_link(ctx, req))
8144 /* final option, timeout link is holding this req pending */
8145 io_timeout_remove_link(ctx, req);
8148 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8149 struct files_struct *files)
8151 struct io_defer_entry *de = NULL;
8154 spin_lock_irq(&ctx->completion_lock);
8155 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8156 if (io_match_link_files(de->req, files)) {
8157 list_cut_position(&list, &ctx->defer_list, &de->list);
8161 spin_unlock_irq(&ctx->completion_lock);
8163 while (!list_empty(&list)) {
8164 de = list_first_entry(&list, struct io_defer_entry, list);
8165 list_del_init(&de->list);
8166 req_set_fail_links(de->req);
8167 io_put_req(de->req);
8168 io_req_complete(de->req, -ECANCELED);
8173 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8174 struct files_struct *files)
8176 if (list_empty_careful(&ctx->inflight_list))
8179 io_cancel_defer_files(ctx, files);
8180 /* cancel all at once, should be faster than doing it one by one*/
8181 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8183 while (!list_empty_careful(&ctx->inflight_list)) {
8184 struct io_kiocb *cancel_req = NULL, *req;
8187 spin_lock_irq(&ctx->inflight_lock);
8188 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8189 if (req->work.files != files)
8191 /* req is being completed, ignore */
8192 if (!refcount_inc_not_zero(&req->refs))
8198 prepare_to_wait(&ctx->inflight_wait, &wait,
8199 TASK_UNINTERRUPTIBLE);
8200 spin_unlock_irq(&ctx->inflight_lock);
8202 /* We need to keep going until we don't find a matching req */
8205 /* cancel this request, or head link requests */
8206 io_attempt_cancel(ctx, cancel_req);
8207 io_put_req(cancel_req);
8208 /* cancellations _may_ trigger task work */
8211 finish_wait(&ctx->inflight_wait, &wait);
8215 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8217 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8218 struct task_struct *task = data;
8220 return req->task == task;
8223 static int io_uring_flush(struct file *file, void *data)
8225 struct io_ring_ctx *ctx = file->private_data;
8227 io_uring_cancel_files(ctx, data);
8230 * If the task is going away, cancel work it may have pending
8232 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8233 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8238 static void *io_uring_validate_mmap_request(struct file *file,
8239 loff_t pgoff, size_t sz)
8241 struct io_ring_ctx *ctx = file->private_data;
8242 loff_t offset = pgoff << PAGE_SHIFT;
8247 case IORING_OFF_SQ_RING:
8248 case IORING_OFF_CQ_RING:
8251 case IORING_OFF_SQES:
8255 return ERR_PTR(-EINVAL);
8258 page = virt_to_head_page(ptr);
8259 if (sz > page_size(page))
8260 return ERR_PTR(-EINVAL);
8267 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8269 size_t sz = vma->vm_end - vma->vm_start;
8273 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8275 return PTR_ERR(ptr);
8277 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8278 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8281 #else /* !CONFIG_MMU */
8283 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8285 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8288 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8290 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8293 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8294 unsigned long addr, unsigned long len,
8295 unsigned long pgoff, unsigned long flags)
8299 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8301 return PTR_ERR(ptr);
8303 return (unsigned long) ptr;
8306 #endif /* !CONFIG_MMU */
8308 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8309 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8312 struct io_ring_ctx *ctx;
8319 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8327 if (f.file->f_op != &io_uring_fops)
8331 ctx = f.file->private_data;
8332 if (!percpu_ref_tryget(&ctx->refs))
8336 * For SQ polling, the thread will do all submissions and completions.
8337 * Just return the requested submit count, and wake the thread if
8341 if (ctx->flags & IORING_SETUP_SQPOLL) {
8342 if (!list_empty_careful(&ctx->cq_overflow_list))
8343 io_cqring_overflow_flush(ctx, false);
8344 if (flags & IORING_ENTER_SQ_WAKEUP)
8345 wake_up(&ctx->sqo_wait);
8346 submitted = to_submit;
8347 } else if (to_submit) {
8348 mutex_lock(&ctx->uring_lock);
8349 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8350 mutex_unlock(&ctx->uring_lock);
8352 if (submitted != to_submit)
8355 if (flags & IORING_ENTER_GETEVENTS) {
8356 min_complete = min(min_complete, ctx->cq_entries);
8359 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8360 * space applications don't need to do io completion events
8361 * polling again, they can rely on io_sq_thread to do polling
8362 * work, which can reduce cpu usage and uring_lock contention.
8364 if (ctx->flags & IORING_SETUP_IOPOLL &&
8365 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8366 ret = io_iopoll_check(ctx, min_complete);
8368 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8373 percpu_ref_put(&ctx->refs);
8376 return submitted ? submitted : ret;
8379 #ifdef CONFIG_PROC_FS
8380 static int io_uring_show_cred(int id, void *p, void *data)
8382 const struct cred *cred = p;
8383 struct seq_file *m = data;
8384 struct user_namespace *uns = seq_user_ns(m);
8385 struct group_info *gi;
8390 seq_printf(m, "%5d\n", id);
8391 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8392 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8393 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8394 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8395 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8396 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8397 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8398 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8399 seq_puts(m, "\n\tGroups:\t");
8400 gi = cred->group_info;
8401 for (g = 0; g < gi->ngroups; g++) {
8402 seq_put_decimal_ull(m, g ? " " : "",
8403 from_kgid_munged(uns, gi->gid[g]));
8405 seq_puts(m, "\n\tCapEff:\t");
8406 cap = cred->cap_effective;
8407 CAP_FOR_EACH_U32(__capi)
8408 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8413 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8417 mutex_lock(&ctx->uring_lock);
8418 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8419 for (i = 0; i < ctx->nr_user_files; i++) {
8420 struct fixed_file_table *table;
8423 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8424 f = table->files[i & IORING_FILE_TABLE_MASK];
8426 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8428 seq_printf(m, "%5u: <none>\n", i);
8430 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8431 for (i = 0; i < ctx->nr_user_bufs; i++) {
8432 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8434 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8435 (unsigned int) buf->len);
8437 if (!idr_is_empty(&ctx->personality_idr)) {
8438 seq_printf(m, "Personalities:\n");
8439 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8441 seq_printf(m, "PollList:\n");
8442 spin_lock_irq(&ctx->completion_lock);
8443 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8444 struct hlist_head *list = &ctx->cancel_hash[i];
8445 struct io_kiocb *req;
8447 hlist_for_each_entry(req, list, hash_node)
8448 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8449 req->task->task_works != NULL);
8451 spin_unlock_irq(&ctx->completion_lock);
8452 mutex_unlock(&ctx->uring_lock);
8455 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8457 struct io_ring_ctx *ctx = f->private_data;
8459 if (percpu_ref_tryget(&ctx->refs)) {
8460 __io_uring_show_fdinfo(ctx, m);
8461 percpu_ref_put(&ctx->refs);
8466 static const struct file_operations io_uring_fops = {
8467 .release = io_uring_release,
8468 .flush = io_uring_flush,
8469 .mmap = io_uring_mmap,
8471 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8472 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8474 .poll = io_uring_poll,
8475 .fasync = io_uring_fasync,
8476 #ifdef CONFIG_PROC_FS
8477 .show_fdinfo = io_uring_show_fdinfo,
8481 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8482 struct io_uring_params *p)
8484 struct io_rings *rings;
8485 size_t size, sq_array_offset;
8487 /* make sure these are sane, as we already accounted them */
8488 ctx->sq_entries = p->sq_entries;
8489 ctx->cq_entries = p->cq_entries;
8491 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8492 if (size == SIZE_MAX)
8495 rings = io_mem_alloc(size);
8500 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8501 rings->sq_ring_mask = p->sq_entries - 1;
8502 rings->cq_ring_mask = p->cq_entries - 1;
8503 rings->sq_ring_entries = p->sq_entries;
8504 rings->cq_ring_entries = p->cq_entries;
8505 ctx->sq_mask = rings->sq_ring_mask;
8506 ctx->cq_mask = rings->cq_ring_mask;
8508 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8509 if (size == SIZE_MAX) {
8510 io_mem_free(ctx->rings);
8515 ctx->sq_sqes = io_mem_alloc(size);
8516 if (!ctx->sq_sqes) {
8517 io_mem_free(ctx->rings);
8526 * Allocate an anonymous fd, this is what constitutes the application
8527 * visible backing of an io_uring instance. The application mmaps this
8528 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8529 * we have to tie this fd to a socket for file garbage collection purposes.
8531 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8536 #if defined(CONFIG_UNIX)
8537 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8543 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8547 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8548 O_RDWR | O_CLOEXEC);
8551 ret = PTR_ERR(file);
8555 #if defined(CONFIG_UNIX)
8556 ctx->ring_sock->file = file;
8558 fd_install(ret, file);
8561 #if defined(CONFIG_UNIX)
8562 sock_release(ctx->ring_sock);
8563 ctx->ring_sock = NULL;
8568 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8569 struct io_uring_params __user *params)
8571 struct user_struct *user = NULL;
8572 struct io_ring_ctx *ctx;
8578 if (entries > IORING_MAX_ENTRIES) {
8579 if (!(p->flags & IORING_SETUP_CLAMP))
8581 entries = IORING_MAX_ENTRIES;
8585 * Use twice as many entries for the CQ ring. It's possible for the
8586 * application to drive a higher depth than the size of the SQ ring,
8587 * since the sqes are only used at submission time. This allows for
8588 * some flexibility in overcommitting a bit. If the application has
8589 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8590 * of CQ ring entries manually.
8592 p->sq_entries = roundup_pow_of_two(entries);
8593 if (p->flags & IORING_SETUP_CQSIZE) {
8595 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8596 * to a power-of-two, if it isn't already. We do NOT impose
8597 * any cq vs sq ring sizing.
8599 if (p->cq_entries < p->sq_entries)
8601 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8602 if (!(p->flags & IORING_SETUP_CLAMP))
8604 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8606 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8608 p->cq_entries = 2 * p->sq_entries;
8611 user = get_uid(current_user());
8612 limit_mem = !capable(CAP_IPC_LOCK);
8615 ret = __io_account_mem(user,
8616 ring_pages(p->sq_entries, p->cq_entries));
8623 ctx = io_ring_ctx_alloc(p);
8626 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8631 ctx->compat = in_compat_syscall();
8633 ctx->creds = get_current_cred();
8635 mmgrab(current->mm);
8636 ctx->sqo_mm = current->mm;
8639 * Account memory _before_ installing the file descriptor. Once
8640 * the descriptor is installed, it can get closed at any time. Also
8641 * do this before hitting the general error path, as ring freeing
8642 * will un-account as well.
8644 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8646 ctx->limit_mem = limit_mem;
8648 ret = io_allocate_scq_urings(ctx, p);
8652 ret = io_sq_offload_start(ctx, p);
8656 memset(&p->sq_off, 0, sizeof(p->sq_off));
8657 p->sq_off.head = offsetof(struct io_rings, sq.head);
8658 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8659 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8660 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8661 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8662 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8663 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8665 memset(&p->cq_off, 0, sizeof(p->cq_off));
8666 p->cq_off.head = offsetof(struct io_rings, cq.head);
8667 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8668 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8669 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8670 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8671 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8672 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8674 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8675 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8676 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8677 IORING_FEAT_POLL_32BITS;
8679 if (copy_to_user(params, p, sizeof(*p))) {
8685 * Install ring fd as the very last thing, so we don't risk someone
8686 * having closed it before we finish setup
8688 ret = io_uring_get_fd(ctx);
8692 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8695 io_ring_ctx_wait_and_kill(ctx);
8700 * Sets up an aio uring context, and returns the fd. Applications asks for a
8701 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8702 * params structure passed in.
8704 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8706 struct io_uring_params p;
8709 if (copy_from_user(&p, params, sizeof(p)))
8711 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8716 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8717 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8718 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8721 return io_uring_create(entries, &p, params);
8724 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8725 struct io_uring_params __user *, params)
8727 return io_uring_setup(entries, params);
8730 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8732 struct io_uring_probe *p;
8736 size = struct_size(p, ops, nr_args);
8737 if (size == SIZE_MAX)
8739 p = kzalloc(size, GFP_KERNEL);
8744 if (copy_from_user(p, arg, size))
8747 if (memchr_inv(p, 0, size))
8750 p->last_op = IORING_OP_LAST - 1;
8751 if (nr_args > IORING_OP_LAST)
8752 nr_args = IORING_OP_LAST;
8754 for (i = 0; i < nr_args; i++) {
8756 if (!io_op_defs[i].not_supported)
8757 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8762 if (copy_to_user(arg, p, size))
8769 static int io_register_personality(struct io_ring_ctx *ctx)
8771 const struct cred *creds = get_current_cred();
8774 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8775 USHRT_MAX, GFP_KERNEL);
8781 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8783 const struct cred *old_creds;
8785 old_creds = idr_remove(&ctx->personality_idr, id);
8787 put_cred(old_creds);
8794 static bool io_register_op_must_quiesce(int op)
8797 case IORING_UNREGISTER_FILES:
8798 case IORING_REGISTER_FILES_UPDATE:
8799 case IORING_REGISTER_PROBE:
8800 case IORING_REGISTER_PERSONALITY:
8801 case IORING_UNREGISTER_PERSONALITY:
8808 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8809 void __user *arg, unsigned nr_args)
8810 __releases(ctx->uring_lock)
8811 __acquires(ctx->uring_lock)
8816 * We're inside the ring mutex, if the ref is already dying, then
8817 * someone else killed the ctx or is already going through
8818 * io_uring_register().
8820 if (percpu_ref_is_dying(&ctx->refs))
8823 if (io_register_op_must_quiesce(opcode)) {
8824 percpu_ref_kill(&ctx->refs);
8827 * Drop uring mutex before waiting for references to exit. If
8828 * another thread is currently inside io_uring_enter() it might
8829 * need to grab the uring_lock to make progress. If we hold it
8830 * here across the drain wait, then we can deadlock. It's safe
8831 * to drop the mutex here, since no new references will come in
8832 * after we've killed the percpu ref.
8834 mutex_unlock(&ctx->uring_lock);
8835 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8836 mutex_lock(&ctx->uring_lock);
8838 percpu_ref_resurrect(&ctx->refs);
8845 case IORING_REGISTER_BUFFERS:
8846 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8848 case IORING_UNREGISTER_BUFFERS:
8852 ret = io_sqe_buffer_unregister(ctx);
8854 case IORING_REGISTER_FILES:
8855 ret = io_sqe_files_register(ctx, arg, nr_args);
8857 case IORING_UNREGISTER_FILES:
8861 ret = io_sqe_files_unregister(ctx);
8863 case IORING_REGISTER_FILES_UPDATE:
8864 ret = io_sqe_files_update(ctx, arg, nr_args);
8866 case IORING_REGISTER_EVENTFD:
8867 case IORING_REGISTER_EVENTFD_ASYNC:
8871 ret = io_eventfd_register(ctx, arg);
8874 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8875 ctx->eventfd_async = 1;
8877 ctx->eventfd_async = 0;
8879 case IORING_UNREGISTER_EVENTFD:
8883 ret = io_eventfd_unregister(ctx);
8885 case IORING_REGISTER_PROBE:
8887 if (!arg || nr_args > 256)
8889 ret = io_probe(ctx, arg, nr_args);
8891 case IORING_REGISTER_PERSONALITY:
8895 ret = io_register_personality(ctx);
8897 case IORING_UNREGISTER_PERSONALITY:
8901 ret = io_unregister_personality(ctx, nr_args);
8908 if (io_register_op_must_quiesce(opcode)) {
8909 /* bring the ctx back to life */
8910 percpu_ref_reinit(&ctx->refs);
8912 reinit_completion(&ctx->ref_comp);
8917 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8918 void __user *, arg, unsigned int, nr_args)
8920 struct io_ring_ctx *ctx;
8929 if (f.file->f_op != &io_uring_fops)
8932 ctx = f.file->private_data;
8934 mutex_lock(&ctx->uring_lock);
8935 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8936 mutex_unlock(&ctx->uring_lock);
8937 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8938 ctx->cq_ev_fd != NULL, ret);
8944 static int __init io_uring_init(void)
8946 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8947 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8948 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8951 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8952 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8953 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8954 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8955 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8956 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8957 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8958 BUILD_BUG_SQE_ELEM(8, __u64, off);
8959 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8960 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8961 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8962 BUILD_BUG_SQE_ELEM(24, __u32, len);
8963 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8964 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8965 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8966 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8967 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8968 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8969 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8970 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8971 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8972 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8973 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8974 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8975 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8976 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8977 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8978 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8979 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8980 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8981 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8983 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8984 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8985 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8988 __initcall(io_uring_init);