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 * ->poll_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 poll_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;
402 /* NOTE: kiocb has the file as the first member, so don't do it here */
410 struct sockaddr __user *addr;
417 struct user_msghdr __user *umsg;
423 struct io_buffer *kbuf;
429 struct filename *filename;
431 unsigned long nofile;
434 struct io_files_update {
460 struct epoll_event event;
464 struct file *file_out;
465 struct file *file_in;
472 struct io_provide_buf {
486 const char __user *filename;
487 struct statx __user *buffer;
490 struct io_async_connect {
491 struct sockaddr_storage address;
494 struct io_async_msghdr {
495 struct iovec fast_iov[UIO_FASTIOV];
497 struct sockaddr __user *uaddr;
499 struct sockaddr_storage addr;
503 struct iovec fast_iov[UIO_FASTIOV];
507 struct wait_page_queue wpq;
510 struct io_async_ctx {
512 struct io_async_rw rw;
513 struct io_async_msghdr msg;
514 struct io_async_connect connect;
515 struct io_timeout_data timeout;
520 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
521 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
522 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
523 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
524 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
525 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
532 REQ_F_LINK_TIMEOUT_BIT,
534 REQ_F_COMP_LOCKED_BIT,
535 REQ_F_NEED_CLEANUP_BIT,
538 REQ_F_BUFFER_SELECTED_BIT,
539 REQ_F_NO_FILE_TABLE_BIT,
540 REQ_F_WORK_INITIALIZED_BIT,
541 REQ_F_TASK_PINNED_BIT,
543 /* not a real bit, just to check we're not overflowing the space */
549 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
550 /* drain existing IO first */
551 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
553 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
554 /* doesn't sever on completion < 0 */
555 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
557 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
558 /* IOSQE_BUFFER_SELECT */
559 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
562 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
563 /* fail rest of links */
564 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
565 /* on inflight list */
566 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
567 /* read/write uses file position */
568 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
569 /* must not punt to workers */
570 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
571 /* has linked timeout */
572 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
574 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
575 /* completion under lock */
576 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
578 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
579 /* in overflow list */
580 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
581 /* already went through poll handler */
582 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
583 /* buffer already selected */
584 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
585 /* doesn't need file table for this request */
586 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
587 /* io_wq_work is initialized */
588 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
589 /* req->task is refcounted */
590 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
594 struct io_poll_iocb poll;
595 struct io_poll_iocb *double_poll;
596 struct io_wq_work work;
600 * NOTE! Each of the iocb union members has the file pointer
601 * as the first entry in their struct definition. So you can
602 * access the file pointer through any of the sub-structs,
603 * or directly as just 'ki_filp' in this struct.
609 struct io_poll_iocb poll;
610 struct io_accept accept;
612 struct io_cancel cancel;
613 struct io_timeout timeout;
614 struct io_connect connect;
615 struct io_sr_msg sr_msg;
617 struct io_close close;
618 struct io_files_update files_update;
619 struct io_fadvise fadvise;
620 struct io_madvise madvise;
621 struct io_epoll epoll;
622 struct io_splice splice;
623 struct io_provide_buf pbuf;
624 struct io_statx statx;
627 struct io_async_ctx *io;
630 /* polled IO has completed */
635 struct io_ring_ctx *ctx;
636 struct list_head list;
639 struct task_struct *task;
645 struct list_head link_list;
647 struct list_head inflight_entry;
649 struct percpu_ref *fixed_file_refs;
653 * Only commands that never go async can use the below fields,
654 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
655 * async armed poll handlers for regular commands. The latter
656 * restore the work, if needed.
659 struct hlist_node hash_node;
660 struct async_poll *apoll;
662 struct io_wq_work work;
664 struct callback_head task_work;
667 #define IO_IOPOLL_BATCH 8
669 struct io_comp_state {
671 struct list_head list;
672 struct io_ring_ctx *ctx;
675 struct io_submit_state {
676 struct blk_plug plug;
679 * io_kiocb alloc cache
681 void *reqs[IO_IOPOLL_BATCH];
682 unsigned int free_reqs;
685 * Batch completion logic
687 struct io_comp_state comp;
690 * File reference cache
694 unsigned int has_refs;
695 unsigned int used_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;
725 static const struct io_op_def io_op_defs[] = {
726 [IORING_OP_NOP] = {},
727 [IORING_OP_READV] = {
731 .unbound_nonreg_file = 1,
735 [IORING_OP_WRITEV] = {
740 .unbound_nonreg_file = 1,
743 [IORING_OP_FSYNC] = {
746 [IORING_OP_READ_FIXED] = {
748 .unbound_nonreg_file = 1,
751 [IORING_OP_WRITE_FIXED] = {
754 .unbound_nonreg_file = 1,
757 [IORING_OP_POLL_ADD] = {
759 .unbound_nonreg_file = 1,
761 [IORING_OP_POLL_REMOVE] = {},
762 [IORING_OP_SYNC_FILE_RANGE] = {
765 [IORING_OP_SENDMSG] = {
769 .unbound_nonreg_file = 1,
773 [IORING_OP_RECVMSG] = {
777 .unbound_nonreg_file = 1,
782 [IORING_OP_TIMEOUT] = {
786 [IORING_OP_TIMEOUT_REMOVE] = {},
787 [IORING_OP_ACCEPT] = {
790 .unbound_nonreg_file = 1,
794 [IORING_OP_ASYNC_CANCEL] = {},
795 [IORING_OP_LINK_TIMEOUT] = {
799 [IORING_OP_CONNECT] = {
803 .unbound_nonreg_file = 1,
806 [IORING_OP_FALLOCATE] = {
809 [IORING_OP_OPENAT] = {
813 [IORING_OP_CLOSE] = {
815 .needs_file_no_error = 1,
818 [IORING_OP_FILES_UPDATE] = {
822 [IORING_OP_STATX] = {
830 .unbound_nonreg_file = 1,
834 [IORING_OP_WRITE] = {
837 .unbound_nonreg_file = 1,
840 [IORING_OP_FADVISE] = {
843 [IORING_OP_MADVISE] = {
849 .unbound_nonreg_file = 1,
855 .unbound_nonreg_file = 1,
859 [IORING_OP_OPENAT2] = {
863 [IORING_OP_EPOLL_CTL] = {
864 .unbound_nonreg_file = 1,
867 [IORING_OP_SPLICE] = {
870 .unbound_nonreg_file = 1,
872 [IORING_OP_PROVIDE_BUFFERS] = {},
873 [IORING_OP_REMOVE_BUFFERS] = {},
877 .unbound_nonreg_file = 1,
881 enum io_mem_account {
886 static bool io_rw_reissue(struct io_kiocb *req, long res);
887 static void io_cqring_fill_event(struct io_kiocb *req, long res);
888 static void io_put_req(struct io_kiocb *req);
889 static void io_double_put_req(struct io_kiocb *req);
890 static void __io_double_put_req(struct io_kiocb *req);
891 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
892 static void io_queue_linked_timeout(struct io_kiocb *req);
893 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
894 struct io_uring_files_update *ip,
896 static int io_grab_files(struct io_kiocb *req);
897 static void io_complete_rw_common(struct kiocb *kiocb, long res,
898 struct io_comp_state *cs);
899 static void io_cleanup_req(struct io_kiocb *req);
900 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
901 int fd, struct file **out_file, bool fixed);
902 static void __io_queue_sqe(struct io_kiocb *req,
903 const struct io_uring_sqe *sqe,
904 struct io_comp_state *cs);
905 static void io_file_put_work(struct work_struct *work);
907 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
908 struct iovec **iovec, struct iov_iter *iter,
910 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
911 struct iovec *iovec, struct iovec *fast_iov,
912 struct iov_iter *iter);
914 static struct kmem_cache *req_cachep;
916 static const struct file_operations io_uring_fops;
918 struct sock *io_uring_get_socket(struct file *file)
920 #if defined(CONFIG_UNIX)
921 if (file->f_op == &io_uring_fops) {
922 struct io_ring_ctx *ctx = file->private_data;
924 return ctx->ring_sock->sk;
929 EXPORT_SYMBOL(io_uring_get_socket);
931 static void io_get_req_task(struct io_kiocb *req)
933 if (req->flags & REQ_F_TASK_PINNED)
935 get_task_struct(req->task);
936 req->flags |= REQ_F_TASK_PINNED;
939 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
940 static void __io_put_req_task(struct io_kiocb *req)
942 if (req->flags & REQ_F_TASK_PINNED)
943 put_task_struct(req->task);
946 static void io_sq_thread_drop_mm(void)
948 struct mm_struct *mm = current->mm;
951 kthread_unuse_mm(mm);
956 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
959 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
961 kthread_use_mm(ctx->sqo_mm);
967 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
968 struct io_kiocb *req)
970 if (!io_op_defs[req->opcode].needs_mm)
972 return __io_sq_thread_acquire_mm(ctx);
975 static inline void req_set_fail_links(struct io_kiocb *req)
977 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
978 req->flags |= REQ_F_FAIL_LINK;
982 * Note: must call io_req_init_async() for the first time you
983 * touch any members of io_wq_work.
985 static inline void io_req_init_async(struct io_kiocb *req)
987 if (req->flags & REQ_F_WORK_INITIALIZED)
990 memset(&req->work, 0, sizeof(req->work));
991 req->flags |= REQ_F_WORK_INITIALIZED;
994 static inline bool io_async_submit(struct io_ring_ctx *ctx)
996 return ctx->flags & IORING_SETUP_SQPOLL;
999 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1001 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1003 complete(&ctx->ref_comp);
1006 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1008 return !req->timeout.off;
1011 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1013 struct io_ring_ctx *ctx;
1016 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1020 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1021 if (!ctx->fallback_req)
1025 * Use 5 bits less than the max cq entries, that should give us around
1026 * 32 entries per hash list if totally full and uniformly spread.
1028 hash_bits = ilog2(p->cq_entries);
1032 ctx->cancel_hash_bits = hash_bits;
1033 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1035 if (!ctx->cancel_hash)
1037 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1039 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1040 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1043 ctx->flags = p->flags;
1044 init_waitqueue_head(&ctx->sqo_wait);
1045 init_waitqueue_head(&ctx->cq_wait);
1046 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1047 init_completion(&ctx->ref_comp);
1048 init_completion(&ctx->sq_thread_comp);
1049 idr_init(&ctx->io_buffer_idr);
1050 idr_init(&ctx->personality_idr);
1051 mutex_init(&ctx->uring_lock);
1052 init_waitqueue_head(&ctx->wait);
1053 spin_lock_init(&ctx->completion_lock);
1054 INIT_LIST_HEAD(&ctx->poll_list);
1055 INIT_LIST_HEAD(&ctx->defer_list);
1056 INIT_LIST_HEAD(&ctx->timeout_list);
1057 init_waitqueue_head(&ctx->inflight_wait);
1058 spin_lock_init(&ctx->inflight_lock);
1059 INIT_LIST_HEAD(&ctx->inflight_list);
1060 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1061 init_llist_head(&ctx->file_put_llist);
1064 if (ctx->fallback_req)
1065 kmem_cache_free(req_cachep, ctx->fallback_req);
1066 kfree(ctx->cancel_hash);
1071 static inline bool req_need_defer(struct io_kiocb *req)
1073 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1074 struct io_ring_ctx *ctx = req->ctx;
1076 return req->sequence != ctx->cached_cq_tail
1077 + atomic_read(&ctx->cached_cq_overflow);
1083 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1085 struct io_rings *rings = ctx->rings;
1087 /* order cqe stores with ring update */
1088 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1090 if (wq_has_sleeper(&ctx->cq_wait)) {
1091 wake_up_interruptible(&ctx->cq_wait);
1092 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1096 static void io_req_work_grab_env(struct io_kiocb *req)
1098 const struct io_op_def *def = &io_op_defs[req->opcode];
1100 io_req_init_async(req);
1102 if (!req->work.mm && def->needs_mm) {
1103 mmgrab(current->mm);
1104 req->work.mm = current->mm;
1106 if (!req->work.creds)
1107 req->work.creds = get_current_cred();
1108 if (!req->work.fs && def->needs_fs) {
1109 spin_lock(¤t->fs->lock);
1110 if (!current->fs->in_exec) {
1111 req->work.fs = current->fs;
1112 req->work.fs->users++;
1114 req->work.flags |= IO_WQ_WORK_CANCEL;
1116 spin_unlock(¤t->fs->lock);
1120 static inline void io_req_work_drop_env(struct io_kiocb *req)
1122 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1126 mmdrop(req->work.mm);
1127 req->work.mm = NULL;
1129 if (req->work.creds) {
1130 put_cred(req->work.creds);
1131 req->work.creds = NULL;
1134 struct fs_struct *fs = req->work.fs;
1136 spin_lock(&req->work.fs->lock);
1139 spin_unlock(&req->work.fs->lock);
1145 static void io_prep_async_work(struct io_kiocb *req)
1147 const struct io_op_def *def = &io_op_defs[req->opcode];
1149 io_req_init_async(req);
1151 if (req->flags & REQ_F_ISREG) {
1152 if (def->hash_reg_file)
1153 io_wq_hash_work(&req->work, file_inode(req->file));
1155 if (def->unbound_nonreg_file)
1156 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 io_req_work_grab_env(req);
1162 static void io_prep_async_link(struct io_kiocb *req)
1164 struct io_kiocb *cur;
1166 io_prep_async_work(req);
1167 if (req->flags & REQ_F_LINK_HEAD)
1168 list_for_each_entry(cur, &req->link_list, link_list)
1169 io_prep_async_work(cur);
1172 static void __io_queue_async_work(struct io_kiocb *req)
1174 struct io_ring_ctx *ctx = req->ctx;
1175 struct io_kiocb *link = io_prep_linked_timeout(req);
1177 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1178 &req->work, req->flags);
1179 io_wq_enqueue(ctx->io_wq, &req->work);
1182 io_queue_linked_timeout(link);
1185 static void io_queue_async_work(struct io_kiocb *req)
1187 /* init ->work of the whole link before punting */
1188 io_prep_async_link(req);
1189 __io_queue_async_work(req);
1192 static void io_kill_timeout(struct io_kiocb *req)
1196 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1198 atomic_inc(&req->ctx->cq_timeouts);
1199 list_del_init(&req->list);
1200 req->flags |= REQ_F_COMP_LOCKED;
1201 io_cqring_fill_event(req, 0);
1206 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1208 struct io_kiocb *req, *tmp;
1210 spin_lock_irq(&ctx->completion_lock);
1211 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1212 io_kill_timeout(req);
1213 spin_unlock_irq(&ctx->completion_lock);
1216 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1219 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1220 struct io_kiocb, list);
1222 if (req_need_defer(req))
1224 list_del_init(&req->list);
1225 /* punt-init is done before queueing for defer */
1226 __io_queue_async_work(req);
1227 } while (!list_empty(&ctx->defer_list));
1230 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1232 while (!list_empty(&ctx->timeout_list)) {
1233 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1234 struct io_kiocb, list);
1236 if (io_is_timeout_noseq(req))
1238 if (req->timeout.target_seq != ctx->cached_cq_tail
1239 - atomic_read(&ctx->cq_timeouts))
1242 list_del_init(&req->list);
1243 io_kill_timeout(req);
1247 static void io_commit_cqring(struct io_ring_ctx *ctx)
1249 io_flush_timeouts(ctx);
1250 __io_commit_cqring(ctx);
1252 if (unlikely(!list_empty(&ctx->defer_list)))
1253 __io_queue_deferred(ctx);
1256 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1258 struct io_rings *rings = ctx->rings;
1261 tail = ctx->cached_cq_tail;
1263 * writes to the cq entry need to come after reading head; the
1264 * control dependency is enough as we're using WRITE_ONCE to
1267 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1270 ctx->cached_cq_tail++;
1271 return &rings->cqes[tail & ctx->cq_mask];
1274 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1278 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1280 if (!ctx->eventfd_async)
1282 return io_wq_current_is_worker();
1285 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1287 if (waitqueue_active(&ctx->wait))
1288 wake_up(&ctx->wait);
1289 if (waitqueue_active(&ctx->sqo_wait))
1290 wake_up(&ctx->sqo_wait);
1291 if (io_should_trigger_evfd(ctx))
1292 eventfd_signal(ctx->cq_ev_fd, 1);
1295 /* Returns true if there are no backlogged entries after the flush */
1296 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1298 struct io_rings *rings = ctx->rings;
1299 struct io_uring_cqe *cqe;
1300 struct io_kiocb *req;
1301 unsigned long flags;
1305 if (list_empty_careful(&ctx->cq_overflow_list))
1307 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1308 rings->cq_ring_entries))
1312 spin_lock_irqsave(&ctx->completion_lock, flags);
1314 /* if force is set, the ring is going away. always drop after that */
1316 ctx->cq_overflow_flushed = 1;
1319 while (!list_empty(&ctx->cq_overflow_list)) {
1320 cqe = io_get_cqring(ctx);
1324 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1326 list_move(&req->list, &list);
1327 req->flags &= ~REQ_F_OVERFLOW;
1329 WRITE_ONCE(cqe->user_data, req->user_data);
1330 WRITE_ONCE(cqe->res, req->result);
1331 WRITE_ONCE(cqe->flags, req->cflags);
1333 WRITE_ONCE(ctx->rings->cq_overflow,
1334 atomic_inc_return(&ctx->cached_cq_overflow));
1338 io_commit_cqring(ctx);
1340 clear_bit(0, &ctx->sq_check_overflow);
1341 clear_bit(0, &ctx->cq_check_overflow);
1342 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1344 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1345 io_cqring_ev_posted(ctx);
1347 while (!list_empty(&list)) {
1348 req = list_first_entry(&list, struct io_kiocb, list);
1349 list_del(&req->list);
1356 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1358 struct io_ring_ctx *ctx = req->ctx;
1359 struct io_uring_cqe *cqe;
1361 trace_io_uring_complete(ctx, req->user_data, res);
1364 * If we can't get a cq entry, userspace overflowed the
1365 * submission (by quite a lot). Increment the overflow count in
1368 cqe = io_get_cqring(ctx);
1370 WRITE_ONCE(cqe->user_data, req->user_data);
1371 WRITE_ONCE(cqe->res, res);
1372 WRITE_ONCE(cqe->flags, cflags);
1373 } else if (ctx->cq_overflow_flushed) {
1374 WRITE_ONCE(ctx->rings->cq_overflow,
1375 atomic_inc_return(&ctx->cached_cq_overflow));
1377 if (list_empty(&ctx->cq_overflow_list)) {
1378 set_bit(0, &ctx->sq_check_overflow);
1379 set_bit(0, &ctx->cq_check_overflow);
1380 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1382 req->flags |= REQ_F_OVERFLOW;
1383 refcount_inc(&req->refs);
1385 req->cflags = cflags;
1386 list_add_tail(&req->list, &ctx->cq_overflow_list);
1390 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1392 __io_cqring_fill_event(req, res, 0);
1395 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1397 struct io_ring_ctx *ctx = req->ctx;
1398 unsigned long flags;
1400 spin_lock_irqsave(&ctx->completion_lock, flags);
1401 __io_cqring_fill_event(req, res, cflags);
1402 io_commit_cqring(ctx);
1403 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1405 io_cqring_ev_posted(ctx);
1408 static void io_submit_flush_completions(struct io_comp_state *cs)
1410 struct io_ring_ctx *ctx = cs->ctx;
1412 spin_lock_irq(&ctx->completion_lock);
1413 while (!list_empty(&cs->list)) {
1414 struct io_kiocb *req;
1416 req = list_first_entry(&cs->list, struct io_kiocb, list);
1417 list_del(&req->list);
1418 __io_cqring_fill_event(req, req->result, req->cflags);
1419 if (!(req->flags & REQ_F_LINK_HEAD)) {
1420 req->flags |= REQ_F_COMP_LOCKED;
1423 spin_unlock_irq(&ctx->completion_lock);
1425 spin_lock_irq(&ctx->completion_lock);
1428 io_commit_cqring(ctx);
1429 spin_unlock_irq(&ctx->completion_lock);
1431 io_cqring_ev_posted(ctx);
1435 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1436 struct io_comp_state *cs)
1439 io_cqring_add_event(req, res, cflags);
1443 req->cflags = cflags;
1444 list_add_tail(&req->list, &cs->list);
1446 io_submit_flush_completions(cs);
1450 static void io_req_complete(struct io_kiocb *req, long res)
1452 __io_req_complete(req, res, 0, NULL);
1455 static inline bool io_is_fallback_req(struct io_kiocb *req)
1457 return req == (struct io_kiocb *)
1458 ((unsigned long) req->ctx->fallback_req & ~1UL);
1461 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1463 struct io_kiocb *req;
1465 req = ctx->fallback_req;
1466 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1472 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1473 struct io_submit_state *state)
1475 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1476 struct io_kiocb *req;
1478 if (!state->free_reqs) {
1482 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1483 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1486 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1487 * retry single alloc to be on the safe side.
1489 if (unlikely(ret <= 0)) {
1490 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1491 if (!state->reqs[0])
1495 state->free_reqs = ret - 1;
1496 req = state->reqs[ret - 1];
1499 req = state->reqs[state->free_reqs];
1504 return io_get_fallback_req(ctx);
1507 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1511 percpu_ref_put(req->fixed_file_refs);
1516 static void io_dismantle_req(struct io_kiocb *req)
1518 if (req->flags & REQ_F_NEED_CLEANUP)
1519 io_cleanup_req(req);
1524 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1525 __io_put_req_task(req);
1526 io_req_work_drop_env(req);
1528 if (req->flags & REQ_F_INFLIGHT) {
1529 struct io_ring_ctx *ctx = req->ctx;
1530 unsigned long flags;
1532 spin_lock_irqsave(&ctx->inflight_lock, flags);
1533 list_del(&req->inflight_entry);
1534 if (waitqueue_active(&ctx->inflight_wait))
1535 wake_up(&ctx->inflight_wait);
1536 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1540 static void __io_free_req(struct io_kiocb *req)
1542 struct io_ring_ctx *ctx;
1544 io_dismantle_req(req);
1546 if (likely(!io_is_fallback_req(req)))
1547 kmem_cache_free(req_cachep, req);
1549 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1550 percpu_ref_put(&ctx->refs);
1553 static bool io_link_cancel_timeout(struct io_kiocb *req)
1555 struct io_ring_ctx *ctx = req->ctx;
1558 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1560 io_cqring_fill_event(req, -ECANCELED);
1561 io_commit_cqring(ctx);
1562 req->flags &= ~REQ_F_LINK_HEAD;
1570 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1572 struct io_kiocb *link;
1575 if (list_empty(&req->link_list))
1577 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1578 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1581 list_del_init(&link->link_list);
1582 wake_ev = io_link_cancel_timeout(link);
1583 req->flags &= ~REQ_F_LINK_TIMEOUT;
1587 static void io_kill_linked_timeout(struct io_kiocb *req)
1589 struct io_ring_ctx *ctx = req->ctx;
1592 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1593 unsigned long flags;
1595 spin_lock_irqsave(&ctx->completion_lock, flags);
1596 wake_ev = __io_kill_linked_timeout(req);
1597 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1599 wake_ev = __io_kill_linked_timeout(req);
1603 io_cqring_ev_posted(ctx);
1606 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1608 struct io_kiocb *nxt;
1611 * The list should never be empty when we are called here. But could
1612 * potentially happen if the chain is messed up, check to be on the
1615 if (unlikely(list_empty(&req->link_list)))
1618 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1619 list_del_init(&req->link_list);
1620 if (!list_empty(&nxt->link_list))
1621 nxt->flags |= REQ_F_LINK_HEAD;
1626 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1628 static void __io_fail_links(struct io_kiocb *req)
1630 struct io_ring_ctx *ctx = req->ctx;
1632 while (!list_empty(&req->link_list)) {
1633 struct io_kiocb *link = list_first_entry(&req->link_list,
1634 struct io_kiocb, link_list);
1636 list_del_init(&link->link_list);
1637 trace_io_uring_fail_link(req, link);
1639 io_cqring_fill_event(link, -ECANCELED);
1640 __io_double_put_req(link);
1641 req->flags &= ~REQ_F_LINK_TIMEOUT;
1644 io_commit_cqring(ctx);
1645 io_cqring_ev_posted(ctx);
1648 static void io_fail_links(struct io_kiocb *req)
1650 struct io_ring_ctx *ctx = req->ctx;
1652 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1653 unsigned long flags;
1655 spin_lock_irqsave(&ctx->completion_lock, flags);
1656 __io_fail_links(req);
1657 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1659 __io_fail_links(req);
1662 io_cqring_ev_posted(ctx);
1665 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1667 req->flags &= ~REQ_F_LINK_HEAD;
1668 if (req->flags & REQ_F_LINK_TIMEOUT)
1669 io_kill_linked_timeout(req);
1672 * If LINK is set, we have dependent requests in this chain. If we
1673 * didn't fail this request, queue the first one up, moving any other
1674 * dependencies to the next request. In case of failure, fail the rest
1677 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1678 return io_req_link_next(req);
1683 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1685 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1687 return __io_req_find_next(req);
1690 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1692 struct task_struct *tsk = req->task;
1693 struct io_ring_ctx *ctx = req->ctx;
1694 int ret, notify = TWA_RESUME;
1697 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1698 * If we're not using an eventfd, then TWA_RESUME is always fine,
1699 * as we won't have dependencies between request completions for
1700 * other kernel wait conditions.
1702 if (ctx->flags & IORING_SETUP_SQPOLL)
1704 else if (ctx->cq_ev_fd)
1705 notify = TWA_SIGNAL;
1707 ret = task_work_add(tsk, cb, notify);
1709 wake_up_process(tsk);
1713 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1715 struct io_ring_ctx *ctx = req->ctx;
1717 spin_lock_irq(&ctx->completion_lock);
1718 io_cqring_fill_event(req, error);
1719 io_commit_cqring(ctx);
1720 spin_unlock_irq(&ctx->completion_lock);
1722 io_cqring_ev_posted(ctx);
1723 req_set_fail_links(req);
1724 io_double_put_req(req);
1727 static void io_req_task_cancel(struct callback_head *cb)
1729 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1731 __io_req_task_cancel(req, -ECANCELED);
1734 static void __io_req_task_submit(struct io_kiocb *req)
1736 struct io_ring_ctx *ctx = req->ctx;
1738 if (!__io_sq_thread_acquire_mm(ctx)) {
1739 mutex_lock(&ctx->uring_lock);
1740 __io_queue_sqe(req, NULL, NULL);
1741 mutex_unlock(&ctx->uring_lock);
1743 __io_req_task_cancel(req, -EFAULT);
1747 static void io_req_task_submit(struct callback_head *cb)
1749 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1751 __io_req_task_submit(req);
1754 static void io_req_task_queue(struct io_kiocb *req)
1758 init_task_work(&req->task_work, io_req_task_submit);
1760 ret = io_req_task_work_add(req, &req->task_work);
1761 if (unlikely(ret)) {
1762 struct task_struct *tsk;
1764 init_task_work(&req->task_work, io_req_task_cancel);
1765 tsk = io_wq_get_task(req->ctx->io_wq);
1766 task_work_add(tsk, &req->task_work, 0);
1767 wake_up_process(tsk);
1771 static void io_queue_next(struct io_kiocb *req)
1773 struct io_kiocb *nxt = io_req_find_next(req);
1776 io_req_task_queue(nxt);
1779 static void io_free_req(struct io_kiocb *req)
1786 void *reqs[IO_IOPOLL_BATCH];
1790 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1791 struct req_batch *rb)
1793 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1794 percpu_ref_put_many(&ctx->refs, rb->to_free);
1798 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1799 struct req_batch *rb)
1802 __io_req_free_batch_flush(ctx, rb);
1805 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1807 if (unlikely(io_is_fallback_req(req))) {
1811 if (req->flags & REQ_F_LINK_HEAD)
1814 io_dismantle_req(req);
1815 rb->reqs[rb->to_free++] = req;
1816 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1817 __io_req_free_batch_flush(req->ctx, rb);
1821 * Drop reference to request, return next in chain (if there is one) if this
1822 * was the last reference to this request.
1824 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1826 struct io_kiocb *nxt = NULL;
1828 if (refcount_dec_and_test(&req->refs)) {
1829 nxt = io_req_find_next(req);
1835 static void io_put_req(struct io_kiocb *req)
1837 if (refcount_dec_and_test(&req->refs))
1841 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1843 struct io_kiocb *nxt;
1846 * A ref is owned by io-wq in which context we're. So, if that's the
1847 * last one, it's safe to steal next work. False negatives are Ok,
1848 * it just will be re-punted async in io_put_work()
1850 if (refcount_read(&req->refs) != 1)
1853 nxt = io_req_find_next(req);
1854 return nxt ? &nxt->work : NULL;
1858 * Must only be used if we don't need to care about links, usually from
1859 * within the completion handling itself.
1861 static void __io_double_put_req(struct io_kiocb *req)
1863 /* drop both submit and complete references */
1864 if (refcount_sub_and_test(2, &req->refs))
1868 static void io_double_put_req(struct io_kiocb *req)
1870 /* drop both submit and complete references */
1871 if (refcount_sub_and_test(2, &req->refs))
1875 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1877 struct io_rings *rings = ctx->rings;
1879 if (test_bit(0, &ctx->cq_check_overflow)) {
1881 * noflush == true is from the waitqueue handler, just ensure
1882 * we wake up the task, and the next invocation will flush the
1883 * entries. We cannot safely to it from here.
1885 if (noflush && !list_empty(&ctx->cq_overflow_list))
1888 io_cqring_overflow_flush(ctx, false);
1891 /* See comment at the top of this file */
1893 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1896 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1898 struct io_rings *rings = ctx->rings;
1900 /* make sure SQ entry isn't read before tail */
1901 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1904 static int io_put_kbuf(struct io_kiocb *req)
1906 struct io_buffer *kbuf;
1909 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1910 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1911 cflags |= IORING_CQE_F_BUFFER;
1917 static inline bool io_run_task_work(void)
1919 if (current->task_works) {
1920 __set_current_state(TASK_RUNNING);
1928 static void io_iopoll_queue(struct list_head *again)
1930 struct io_kiocb *req;
1933 req = list_first_entry(again, struct io_kiocb, list);
1934 list_del(&req->list);
1935 if (!io_rw_reissue(req, -EAGAIN))
1936 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1937 } while (!list_empty(again));
1941 * Find and free completed poll iocbs
1943 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1944 struct list_head *done)
1946 struct req_batch rb;
1947 struct io_kiocb *req;
1950 /* order with ->result store in io_complete_rw_iopoll() */
1954 while (!list_empty(done)) {
1957 req = list_first_entry(done, struct io_kiocb, list);
1958 if (READ_ONCE(req->result) == -EAGAIN) {
1959 req->iopoll_completed = 0;
1960 list_move_tail(&req->list, &again);
1963 list_del(&req->list);
1965 if (req->flags & REQ_F_BUFFER_SELECTED)
1966 cflags = io_put_kbuf(req);
1968 __io_cqring_fill_event(req, req->result, cflags);
1971 if (refcount_dec_and_test(&req->refs))
1972 io_req_free_batch(&rb, req);
1975 io_commit_cqring(ctx);
1976 if (ctx->flags & IORING_SETUP_SQPOLL)
1977 io_cqring_ev_posted(ctx);
1978 io_req_free_batch_finish(ctx, &rb);
1980 if (!list_empty(&again))
1981 io_iopoll_queue(&again);
1984 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1987 struct io_kiocb *req, *tmp;
1993 * Only spin for completions if we don't have multiple devices hanging
1994 * off our complete list, and we're under the requested amount.
1996 spin = !ctx->poll_multi_file && *nr_events < min;
1999 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
2000 struct kiocb *kiocb = &req->rw.kiocb;
2003 * Move completed and retryable entries to our local lists.
2004 * If we find a request that requires polling, break out
2005 * and complete those lists first, if we have entries there.
2007 if (READ_ONCE(req->iopoll_completed)) {
2008 list_move_tail(&req->list, &done);
2011 if (!list_empty(&done))
2014 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2018 /* iopoll may have completed current req */
2019 if (READ_ONCE(req->iopoll_completed))
2020 list_move_tail(&req->list, &done);
2027 if (!list_empty(&done))
2028 io_iopoll_complete(ctx, nr_events, &done);
2034 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2035 * non-spinning poll check - we'll still enter the driver poll loop, but only
2036 * as a non-spinning completion check.
2038 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2041 while (!list_empty(&ctx->poll_list) && !need_resched()) {
2044 ret = io_do_iopoll(ctx, nr_events, min);
2047 if (*nr_events >= min)
2055 * We can't just wait for polled events to come to us, we have to actively
2056 * find and complete them.
2058 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2060 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2063 mutex_lock(&ctx->uring_lock);
2064 while (!list_empty(&ctx->poll_list)) {
2065 unsigned int nr_events = 0;
2067 io_do_iopoll(ctx, &nr_events, 0);
2069 /* let it sleep and repeat later if can't complete a request */
2073 * Ensure we allow local-to-the-cpu processing to take place,
2074 * in this case we need to ensure that we reap all events.
2075 * Also let task_work, etc. to progress by releasing the mutex
2077 if (need_resched()) {
2078 mutex_unlock(&ctx->uring_lock);
2080 mutex_lock(&ctx->uring_lock);
2083 mutex_unlock(&ctx->uring_lock);
2086 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2088 unsigned int nr_events = 0;
2089 int iters = 0, ret = 0;
2092 * We disallow the app entering submit/complete with polling, but we
2093 * still need to lock the ring to prevent racing with polled issue
2094 * that got punted to a workqueue.
2096 mutex_lock(&ctx->uring_lock);
2099 * Don't enter poll loop if we already have events pending.
2100 * If we do, we can potentially be spinning for commands that
2101 * already triggered a CQE (eg in error).
2103 if (io_cqring_events(ctx, false))
2107 * If a submit got punted to a workqueue, we can have the
2108 * application entering polling for a command before it gets
2109 * issued. That app will hold the uring_lock for the duration
2110 * of the poll right here, so we need to take a breather every
2111 * now and then to ensure that the issue has a chance to add
2112 * the poll to the issued list. Otherwise we can spin here
2113 * forever, while the workqueue is stuck trying to acquire the
2116 if (!(++iters & 7)) {
2117 mutex_unlock(&ctx->uring_lock);
2119 mutex_lock(&ctx->uring_lock);
2122 ret = io_iopoll_getevents(ctx, &nr_events, min);
2126 } while (min && !nr_events && !need_resched());
2128 mutex_unlock(&ctx->uring_lock);
2132 static void kiocb_end_write(struct io_kiocb *req)
2135 * Tell lockdep we inherited freeze protection from submission
2138 if (req->flags & REQ_F_ISREG) {
2139 struct inode *inode = file_inode(req->file);
2141 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2143 file_end_write(req->file);
2146 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2147 struct io_comp_state *cs)
2149 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2152 if (kiocb->ki_flags & IOCB_WRITE)
2153 kiocb_end_write(req);
2155 if (res != req->result)
2156 req_set_fail_links(req);
2157 if (req->flags & REQ_F_BUFFER_SELECTED)
2158 cflags = io_put_kbuf(req);
2159 __io_req_complete(req, res, cflags, cs);
2163 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2165 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2166 ssize_t ret = -ECANCELED;
2167 struct iov_iter iter;
2175 switch (req->opcode) {
2176 case IORING_OP_READV:
2177 case IORING_OP_READ_FIXED:
2178 case IORING_OP_READ:
2181 case IORING_OP_WRITEV:
2182 case IORING_OP_WRITE_FIXED:
2183 case IORING_OP_WRITE:
2187 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2192 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2195 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2200 req_set_fail_links(req);
2201 io_req_complete(req, ret);
2205 static void io_rw_resubmit(struct callback_head *cb)
2207 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2208 struct io_ring_ctx *ctx = req->ctx;
2211 err = io_sq_thread_acquire_mm(ctx, req);
2213 if (io_resubmit_prep(req, err)) {
2214 refcount_inc(&req->refs);
2215 io_queue_async_work(req);
2220 static bool io_rw_reissue(struct io_kiocb *req, long res)
2225 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2228 init_task_work(&req->task_work, io_rw_resubmit);
2229 ret = io_req_task_work_add(req, &req->task_work);
2236 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2237 struct io_comp_state *cs)
2239 if (!io_rw_reissue(req, res))
2240 io_complete_rw_common(&req->rw.kiocb, res, cs);
2243 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2245 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2247 __io_complete_rw(req, res, res2, NULL);
2250 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2252 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2254 if (kiocb->ki_flags & IOCB_WRITE)
2255 kiocb_end_write(req);
2257 if (res != -EAGAIN && res != req->result)
2258 req_set_fail_links(req);
2260 WRITE_ONCE(req->result, res);
2261 /* order with io_poll_complete() checking ->result */
2263 WRITE_ONCE(req->iopoll_completed, 1);
2267 * After the iocb has been issued, it's safe to be found on the poll list.
2268 * Adding the kiocb to the list AFTER submission ensures that we don't
2269 * find it from a io_iopoll_getevents() thread before the issuer is done
2270 * accessing the kiocb cookie.
2272 static void io_iopoll_req_issued(struct io_kiocb *req)
2274 struct io_ring_ctx *ctx = req->ctx;
2277 * Track whether we have multiple files in our lists. This will impact
2278 * how we do polling eventually, not spinning if we're on potentially
2279 * different devices.
2281 if (list_empty(&ctx->poll_list)) {
2282 ctx->poll_multi_file = false;
2283 } else if (!ctx->poll_multi_file) {
2284 struct io_kiocb *list_req;
2286 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2288 if (list_req->file != req->file)
2289 ctx->poll_multi_file = true;
2293 * For fast devices, IO may have already completed. If it has, add
2294 * it to the front so we find it first.
2296 if (READ_ONCE(req->iopoll_completed))
2297 list_add(&req->list, &ctx->poll_list);
2299 list_add_tail(&req->list, &ctx->poll_list);
2301 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2302 wq_has_sleeper(&ctx->sqo_wait))
2303 wake_up(&ctx->sqo_wait);
2306 static void __io_state_file_put(struct io_submit_state *state)
2308 int diff = state->has_refs - state->used_refs;
2311 fput_many(state->file, diff);
2315 static inline void io_state_file_put(struct io_submit_state *state)
2318 __io_state_file_put(state);
2322 * Get as many references to a file as we have IOs left in this submission,
2323 * assuming most submissions are for one file, or at least that each file
2324 * has more than one submission.
2326 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2332 if (state->fd == fd) {
2337 __io_state_file_put(state);
2339 state->file = fget_many(fd, state->ios_left);
2344 state->has_refs = state->ios_left;
2345 state->used_refs = 1;
2350 static bool io_bdev_nowait(struct block_device *bdev)
2353 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2360 * If we tracked the file through the SCM inflight mechanism, we could support
2361 * any file. For now, just ensure that anything potentially problematic is done
2364 static bool io_file_supports_async(struct file *file, int rw)
2366 umode_t mode = file_inode(file)->i_mode;
2368 if (S_ISBLK(mode)) {
2369 if (io_bdev_nowait(file->f_inode->i_bdev))
2373 if (S_ISCHR(mode) || S_ISSOCK(mode))
2375 if (S_ISREG(mode)) {
2376 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2377 file->f_op != &io_uring_fops)
2382 /* any ->read/write should understand O_NONBLOCK */
2383 if (file->f_flags & O_NONBLOCK)
2386 if (!(file->f_mode & FMODE_NOWAIT))
2390 return file->f_op->read_iter != NULL;
2392 return file->f_op->write_iter != NULL;
2395 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2396 bool force_nonblock)
2398 struct io_ring_ctx *ctx = req->ctx;
2399 struct kiocb *kiocb = &req->rw.kiocb;
2403 if (S_ISREG(file_inode(req->file)->i_mode))
2404 req->flags |= REQ_F_ISREG;
2406 kiocb->ki_pos = READ_ONCE(sqe->off);
2407 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2408 req->flags |= REQ_F_CUR_POS;
2409 kiocb->ki_pos = req->file->f_pos;
2411 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2412 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2413 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2417 ioprio = READ_ONCE(sqe->ioprio);
2419 ret = ioprio_check_cap(ioprio);
2423 kiocb->ki_ioprio = ioprio;
2425 kiocb->ki_ioprio = get_current_ioprio();
2427 /* don't allow async punt if RWF_NOWAIT was requested */
2428 if (kiocb->ki_flags & IOCB_NOWAIT)
2429 req->flags |= REQ_F_NOWAIT;
2431 if (kiocb->ki_flags & IOCB_DIRECT)
2432 io_get_req_task(req);
2435 kiocb->ki_flags |= IOCB_NOWAIT;
2437 if (ctx->flags & IORING_SETUP_IOPOLL) {
2438 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2439 !kiocb->ki_filp->f_op->iopoll)
2442 kiocb->ki_flags |= IOCB_HIPRI;
2443 kiocb->ki_complete = io_complete_rw_iopoll;
2444 req->iopoll_completed = 0;
2445 io_get_req_task(req);
2447 if (kiocb->ki_flags & IOCB_HIPRI)
2449 kiocb->ki_complete = io_complete_rw;
2452 req->rw.addr = READ_ONCE(sqe->addr);
2453 req->rw.len = READ_ONCE(sqe->len);
2454 req->buf_index = READ_ONCE(sqe->buf_index);
2458 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2464 case -ERESTARTNOINTR:
2465 case -ERESTARTNOHAND:
2466 case -ERESTART_RESTARTBLOCK:
2468 * We can't just restart the syscall, since previously
2469 * submitted sqes may already be in progress. Just fail this
2475 kiocb->ki_complete(kiocb, ret, 0);
2479 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2480 struct io_comp_state *cs)
2482 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2484 if (req->flags & REQ_F_CUR_POS)
2485 req->file->f_pos = kiocb->ki_pos;
2486 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2487 __io_complete_rw(req, ret, 0, cs);
2489 io_rw_done(kiocb, ret);
2492 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2493 struct iov_iter *iter)
2495 struct io_ring_ctx *ctx = req->ctx;
2496 size_t len = req->rw.len;
2497 struct io_mapped_ubuf *imu;
2498 u16 index, buf_index;
2502 /* attempt to use fixed buffers without having provided iovecs */
2503 if (unlikely(!ctx->user_bufs))
2506 buf_index = req->buf_index;
2507 if (unlikely(buf_index >= ctx->nr_user_bufs))
2510 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2511 imu = &ctx->user_bufs[index];
2512 buf_addr = req->rw.addr;
2515 if (buf_addr + len < buf_addr)
2517 /* not inside the mapped region */
2518 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2522 * May not be a start of buffer, set size appropriately
2523 * and advance us to the beginning.
2525 offset = buf_addr - imu->ubuf;
2526 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2530 * Don't use iov_iter_advance() here, as it's really slow for
2531 * using the latter parts of a big fixed buffer - it iterates
2532 * over each segment manually. We can cheat a bit here, because
2535 * 1) it's a BVEC iter, we set it up
2536 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2537 * first and last bvec
2539 * So just find our index, and adjust the iterator afterwards.
2540 * If the offset is within the first bvec (or the whole first
2541 * bvec, just use iov_iter_advance(). This makes it easier
2542 * since we can just skip the first segment, which may not
2543 * be PAGE_SIZE aligned.
2545 const struct bio_vec *bvec = imu->bvec;
2547 if (offset <= bvec->bv_len) {
2548 iov_iter_advance(iter, offset);
2550 unsigned long seg_skip;
2552 /* skip first vec */
2553 offset -= bvec->bv_len;
2554 seg_skip = 1 + (offset >> PAGE_SHIFT);
2556 iter->bvec = bvec + seg_skip;
2557 iter->nr_segs -= seg_skip;
2558 iter->count -= bvec->bv_len + offset;
2559 iter->iov_offset = offset & ~PAGE_MASK;
2566 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2569 mutex_unlock(&ctx->uring_lock);
2572 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2575 * "Normal" inline submissions always hold the uring_lock, since we
2576 * grab it from the system call. Same is true for the SQPOLL offload.
2577 * The only exception is when we've detached the request and issue it
2578 * from an async worker thread, grab the lock for that case.
2581 mutex_lock(&ctx->uring_lock);
2584 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2585 int bgid, struct io_buffer *kbuf,
2588 struct io_buffer *head;
2590 if (req->flags & REQ_F_BUFFER_SELECTED)
2593 io_ring_submit_lock(req->ctx, needs_lock);
2595 lockdep_assert_held(&req->ctx->uring_lock);
2597 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2599 if (!list_empty(&head->list)) {
2600 kbuf = list_last_entry(&head->list, struct io_buffer,
2602 list_del(&kbuf->list);
2605 idr_remove(&req->ctx->io_buffer_idr, bgid);
2607 if (*len > kbuf->len)
2610 kbuf = ERR_PTR(-ENOBUFS);
2613 io_ring_submit_unlock(req->ctx, needs_lock);
2618 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2621 struct io_buffer *kbuf;
2624 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2625 bgid = req->buf_index;
2626 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2629 req->rw.addr = (u64) (unsigned long) kbuf;
2630 req->flags |= REQ_F_BUFFER_SELECTED;
2631 return u64_to_user_ptr(kbuf->addr);
2634 #ifdef CONFIG_COMPAT
2635 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2638 struct compat_iovec __user *uiov;
2639 compat_ssize_t clen;
2643 uiov = u64_to_user_ptr(req->rw.addr);
2644 if (!access_ok(uiov, sizeof(*uiov)))
2646 if (__get_user(clen, &uiov->iov_len))
2652 buf = io_rw_buffer_select(req, &len, needs_lock);
2654 return PTR_ERR(buf);
2655 iov[0].iov_base = buf;
2656 iov[0].iov_len = (compat_size_t) len;
2661 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2664 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2668 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2671 len = iov[0].iov_len;
2674 buf = io_rw_buffer_select(req, &len, needs_lock);
2676 return PTR_ERR(buf);
2677 iov[0].iov_base = buf;
2678 iov[0].iov_len = len;
2682 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2685 if (req->flags & REQ_F_BUFFER_SELECTED) {
2686 struct io_buffer *kbuf;
2688 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2689 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2690 iov[0].iov_len = kbuf->len;
2695 else if (req->rw.len > 1)
2698 #ifdef CONFIG_COMPAT
2699 if (req->ctx->compat)
2700 return io_compat_import(req, iov, needs_lock);
2703 return __io_iov_buffer_select(req, iov, needs_lock);
2706 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2707 struct iovec **iovec, struct iov_iter *iter,
2710 void __user *buf = u64_to_user_ptr(req->rw.addr);
2711 size_t sqe_len = req->rw.len;
2715 opcode = req->opcode;
2716 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2718 return io_import_fixed(req, rw, iter);
2721 /* buffer index only valid with fixed read/write, or buffer select */
2722 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2725 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2726 if (req->flags & REQ_F_BUFFER_SELECT) {
2727 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2730 return PTR_ERR(buf);
2732 req->rw.len = sqe_len;
2735 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2737 return ret < 0 ? ret : sqe_len;
2741 struct io_async_rw *iorw = &req->io->rw;
2744 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2745 if (iorw->iov == iorw->fast_iov)
2750 if (req->flags & REQ_F_BUFFER_SELECT) {
2751 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2753 ret = (*iovec)->iov_len;
2754 iov_iter_init(iter, rw, *iovec, 1, ret);
2760 #ifdef CONFIG_COMPAT
2761 if (req->ctx->compat)
2762 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2766 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2770 * For files that don't have ->read_iter() and ->write_iter(), handle them
2771 * by looping over ->read() or ->write() manually.
2773 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2774 struct iov_iter *iter)
2779 * Don't support polled IO through this interface, and we can't
2780 * support non-blocking either. For the latter, this just causes
2781 * the kiocb to be handled from an async context.
2783 if (kiocb->ki_flags & IOCB_HIPRI)
2785 if (kiocb->ki_flags & IOCB_NOWAIT)
2788 while (iov_iter_count(iter)) {
2792 if (!iov_iter_is_bvec(iter)) {
2793 iovec = iov_iter_iovec(iter);
2795 /* fixed buffers import bvec */
2796 iovec.iov_base = kmap(iter->bvec->bv_page)
2798 iovec.iov_len = min(iter->count,
2799 iter->bvec->bv_len - iter->iov_offset);
2803 nr = file->f_op->read(file, iovec.iov_base,
2804 iovec.iov_len, &kiocb->ki_pos);
2806 nr = file->f_op->write(file, iovec.iov_base,
2807 iovec.iov_len, &kiocb->ki_pos);
2810 if (iov_iter_is_bvec(iter))
2811 kunmap(iter->bvec->bv_page);
2819 if (nr != iovec.iov_len)
2821 iov_iter_advance(iter, nr);
2827 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2828 struct iovec *iovec, struct iovec *fast_iov,
2829 struct iov_iter *iter)
2831 req->io->rw.nr_segs = iter->nr_segs;
2832 req->io->rw.size = io_size;
2833 req->io->rw.iov = iovec;
2834 if (!req->io->rw.iov) {
2835 req->io->rw.iov = req->io->rw.fast_iov;
2836 if (req->io->rw.iov != fast_iov)
2837 memcpy(req->io->rw.iov, fast_iov,
2838 sizeof(struct iovec) * iter->nr_segs);
2840 req->flags |= REQ_F_NEED_CLEANUP;
2844 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2846 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2847 return req->io == NULL;
2850 static int io_alloc_async_ctx(struct io_kiocb *req)
2852 if (!io_op_defs[req->opcode].async_ctx)
2855 return __io_alloc_async_ctx(req);
2858 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2859 struct iovec *iovec, struct iovec *fast_iov,
2860 struct iov_iter *iter)
2862 if (!io_op_defs[req->opcode].async_ctx)
2865 if (__io_alloc_async_ctx(req))
2868 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2873 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2874 bool force_nonblock)
2876 struct io_async_ctx *io;
2877 struct iov_iter iter;
2880 ret = io_prep_rw(req, sqe, force_nonblock);
2884 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2887 /* either don't need iovec imported or already have it */
2888 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2892 io->rw.iov = io->rw.fast_iov;
2894 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2899 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2903 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2904 int sync, void *arg)
2906 struct wait_page_queue *wpq;
2907 struct io_kiocb *req = wait->private;
2908 struct wait_page_key *key = arg;
2911 wpq = container_of(wait, struct wait_page_queue, wait);
2913 ret = wake_page_match(wpq, key);
2917 list_del_init(&wait->entry);
2919 init_task_work(&req->task_work, io_req_task_submit);
2920 /* submit ref gets dropped, acquire a new one */
2921 refcount_inc(&req->refs);
2922 ret = io_req_task_work_add(req, &req->task_work);
2923 if (unlikely(ret)) {
2924 struct task_struct *tsk;
2926 /* queue just for cancelation */
2927 init_task_work(&req->task_work, io_req_task_cancel);
2928 tsk = io_wq_get_task(req->ctx->io_wq);
2929 task_work_add(tsk, &req->task_work, 0);
2930 wake_up_process(tsk);
2935 static bool io_rw_should_retry(struct io_kiocb *req)
2937 struct kiocb *kiocb = &req->rw.kiocb;
2940 /* never retry for NOWAIT, we just complete with -EAGAIN */
2941 if (req->flags & REQ_F_NOWAIT)
2944 /* already tried, or we're doing O_DIRECT */
2945 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2948 * just use poll if we can, and don't attempt if the fs doesn't
2949 * support callback based unlocks
2951 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2955 * If request type doesn't require req->io to defer in general,
2956 * we need to allocate it here
2958 if (!req->io && __io_alloc_async_ctx(req))
2961 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2962 io_async_buf_func, req);
2964 io_get_req_task(req);
2971 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2973 if (req->file->f_op->read_iter)
2974 return call_read_iter(req->file, &req->rw.kiocb, iter);
2975 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2978 static int io_read(struct io_kiocb *req, bool force_nonblock,
2979 struct io_comp_state *cs)
2981 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2982 struct kiocb *kiocb = &req->rw.kiocb;
2983 struct iov_iter iter;
2985 ssize_t io_size, ret;
2987 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2991 /* Ensure we clear previously set non-block flag */
2992 if (!force_nonblock)
2993 kiocb->ki_flags &= ~IOCB_NOWAIT;
2996 req->result = io_size;
2998 /* If the file doesn't support async, just async punt */
2999 if (force_nonblock && !io_file_supports_async(req->file, READ))
3002 iov_count = iov_iter_count(&iter);
3003 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3005 unsigned long nr_segs = iter.nr_segs;
3008 ret2 = io_iter_do_read(req, &iter);
3010 /* Catch -EAGAIN return for forced non-blocking submission */
3011 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3012 kiocb_done(kiocb, ret2, cs);
3014 iter.count = iov_count;
3015 iter.nr_segs = nr_segs;
3017 ret = io_setup_async_rw(req, io_size, iovec,
3018 inline_vecs, &iter);
3021 /* if we can retry, do so with the callbacks armed */
3022 if (io_rw_should_retry(req)) {
3023 ret2 = io_iter_do_read(req, &iter);
3024 if (ret2 == -EIOCBQUEUED) {
3026 } else if (ret2 != -EAGAIN) {
3027 kiocb_done(kiocb, ret2, cs);
3031 kiocb->ki_flags &= ~IOCB_WAITQ;
3036 if (!(req->flags & REQ_F_NEED_CLEANUP))
3041 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3042 bool force_nonblock)
3044 struct io_async_ctx *io;
3045 struct iov_iter iter;
3048 ret = io_prep_rw(req, sqe, force_nonblock);
3052 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3055 req->fsize = rlimit(RLIMIT_FSIZE);
3057 /* either don't need iovec imported or already have it */
3058 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3062 io->rw.iov = io->rw.fast_iov;
3064 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3069 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3073 static int io_write(struct io_kiocb *req, bool force_nonblock,
3074 struct io_comp_state *cs)
3076 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3077 struct kiocb *kiocb = &req->rw.kiocb;
3078 struct iov_iter iter;
3080 ssize_t ret, io_size;
3082 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3086 /* Ensure we clear previously set non-block flag */
3087 if (!force_nonblock)
3088 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3091 req->result = io_size;
3093 /* If the file doesn't support async, just async punt */
3094 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3097 /* file path doesn't support NOWAIT for non-direct_IO */
3098 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3099 (req->flags & REQ_F_ISREG))
3102 iov_count = iov_iter_count(&iter);
3103 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3105 unsigned long nr_segs = iter.nr_segs;
3109 * Open-code file_start_write here to grab freeze protection,
3110 * which will be released by another thread in
3111 * io_complete_rw(). Fool lockdep by telling it the lock got
3112 * released so that it doesn't complain about the held lock when
3113 * we return to userspace.
3115 if (req->flags & REQ_F_ISREG) {
3116 __sb_start_write(file_inode(req->file)->i_sb,
3117 SB_FREEZE_WRITE, true);
3118 __sb_writers_release(file_inode(req->file)->i_sb,
3121 kiocb->ki_flags |= IOCB_WRITE;
3123 if (!force_nonblock)
3124 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3126 if (req->file->f_op->write_iter)
3127 ret2 = call_write_iter(req->file, kiocb, &iter);
3129 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3131 if (!force_nonblock)
3132 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3135 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3136 * retry them without IOCB_NOWAIT.
3138 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3140 if (!force_nonblock || ret2 != -EAGAIN) {
3141 kiocb_done(kiocb, ret2, cs);
3143 iter.count = iov_count;
3144 iter.nr_segs = nr_segs;
3146 ret = io_setup_async_rw(req, io_size, iovec,
3147 inline_vecs, &iter);
3154 if (!(req->flags & REQ_F_NEED_CLEANUP))
3159 static int __io_splice_prep(struct io_kiocb *req,
3160 const struct io_uring_sqe *sqe)
3162 struct io_splice* sp = &req->splice;
3163 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3166 if (req->flags & REQ_F_NEED_CLEANUP)
3168 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3172 sp->len = READ_ONCE(sqe->len);
3173 sp->flags = READ_ONCE(sqe->splice_flags);
3175 if (unlikely(sp->flags & ~valid_flags))
3178 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3179 (sp->flags & SPLICE_F_FD_IN_FIXED));
3182 req->flags |= REQ_F_NEED_CLEANUP;
3184 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3186 * Splice operation will be punted aync, and here need to
3187 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3189 io_req_init_async(req);
3190 req->work.flags |= IO_WQ_WORK_UNBOUND;
3196 static int io_tee_prep(struct io_kiocb *req,
3197 const struct io_uring_sqe *sqe)
3199 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3201 return __io_splice_prep(req, sqe);
3204 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3206 struct io_splice *sp = &req->splice;
3207 struct file *in = sp->file_in;
3208 struct file *out = sp->file_out;
3209 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3215 ret = do_tee(in, out, sp->len, flags);
3217 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3218 req->flags &= ~REQ_F_NEED_CLEANUP;
3221 req_set_fail_links(req);
3222 io_req_complete(req, ret);
3226 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3228 struct io_splice* sp = &req->splice;
3230 sp->off_in = READ_ONCE(sqe->splice_off_in);
3231 sp->off_out = READ_ONCE(sqe->off);
3232 return __io_splice_prep(req, sqe);
3235 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3237 struct io_splice *sp = &req->splice;
3238 struct file *in = sp->file_in;
3239 struct file *out = sp->file_out;
3240 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3241 loff_t *poff_in, *poff_out;
3247 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3248 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3251 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3253 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3254 req->flags &= ~REQ_F_NEED_CLEANUP;
3257 req_set_fail_links(req);
3258 io_req_complete(req, ret);
3263 * IORING_OP_NOP just posts a completion event, nothing else.
3265 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3267 struct io_ring_ctx *ctx = req->ctx;
3269 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3272 __io_req_complete(req, 0, 0, cs);
3276 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3278 struct io_ring_ctx *ctx = req->ctx;
3283 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3285 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3288 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3289 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3292 req->sync.off = READ_ONCE(sqe->off);
3293 req->sync.len = READ_ONCE(sqe->len);
3297 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3299 loff_t end = req->sync.off + req->sync.len;
3302 /* fsync always requires a blocking context */
3306 ret = vfs_fsync_range(req->file, req->sync.off,
3307 end > 0 ? end : LLONG_MAX,
3308 req->sync.flags & IORING_FSYNC_DATASYNC);
3310 req_set_fail_links(req);
3311 io_req_complete(req, ret);
3315 static int io_fallocate_prep(struct io_kiocb *req,
3316 const struct io_uring_sqe *sqe)
3318 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3320 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3323 req->sync.off = READ_ONCE(sqe->off);
3324 req->sync.len = READ_ONCE(sqe->addr);
3325 req->sync.mode = READ_ONCE(sqe->len);
3326 req->fsize = rlimit(RLIMIT_FSIZE);
3330 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3334 /* fallocate always requiring blocking context */
3338 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3339 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3341 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3343 req_set_fail_links(req);
3344 io_req_complete(req, ret);
3348 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3350 const char __user *fname;
3353 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3355 if (unlikely(sqe->ioprio || sqe->buf_index))
3357 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3360 /* open.how should be already initialised */
3361 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3362 req->open.how.flags |= O_LARGEFILE;
3364 req->open.dfd = READ_ONCE(sqe->fd);
3365 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3366 req->open.filename = getname(fname);
3367 if (IS_ERR(req->open.filename)) {
3368 ret = PTR_ERR(req->open.filename);
3369 req->open.filename = NULL;
3372 req->open.nofile = rlimit(RLIMIT_NOFILE);
3373 req->flags |= REQ_F_NEED_CLEANUP;
3377 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3381 if (req->flags & REQ_F_NEED_CLEANUP)
3383 mode = READ_ONCE(sqe->len);
3384 flags = READ_ONCE(sqe->open_flags);
3385 req->open.how = build_open_how(flags, mode);
3386 return __io_openat_prep(req, sqe);
3389 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3391 struct open_how __user *how;
3395 if (req->flags & REQ_F_NEED_CLEANUP)
3397 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3398 len = READ_ONCE(sqe->len);
3399 if (len < OPEN_HOW_SIZE_VER0)
3402 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3407 return __io_openat_prep(req, sqe);
3410 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3412 struct open_flags op;
3419 ret = build_open_flags(&req->open.how, &op);
3423 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3427 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3430 ret = PTR_ERR(file);
3432 fsnotify_open(file);
3433 fd_install(ret, file);
3436 putname(req->open.filename);
3437 req->flags &= ~REQ_F_NEED_CLEANUP;
3439 req_set_fail_links(req);
3440 io_req_complete(req, ret);
3444 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3446 return io_openat2(req, force_nonblock);
3449 static int io_remove_buffers_prep(struct io_kiocb *req,
3450 const struct io_uring_sqe *sqe)
3452 struct io_provide_buf *p = &req->pbuf;
3455 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3458 tmp = READ_ONCE(sqe->fd);
3459 if (!tmp || tmp > USHRT_MAX)
3462 memset(p, 0, sizeof(*p));
3464 p->bgid = READ_ONCE(sqe->buf_group);
3468 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3469 int bgid, unsigned nbufs)
3473 /* shouldn't happen */
3477 /* the head kbuf is the list itself */
3478 while (!list_empty(&buf->list)) {
3479 struct io_buffer *nxt;
3481 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3482 list_del(&nxt->list);
3489 idr_remove(&ctx->io_buffer_idr, bgid);
3494 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3495 struct io_comp_state *cs)
3497 struct io_provide_buf *p = &req->pbuf;
3498 struct io_ring_ctx *ctx = req->ctx;
3499 struct io_buffer *head;
3502 io_ring_submit_lock(ctx, !force_nonblock);
3504 lockdep_assert_held(&ctx->uring_lock);
3507 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3509 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3511 io_ring_submit_lock(ctx, !force_nonblock);
3513 req_set_fail_links(req);
3514 __io_req_complete(req, ret, 0, cs);
3518 static int io_provide_buffers_prep(struct io_kiocb *req,
3519 const struct io_uring_sqe *sqe)
3521 struct io_provide_buf *p = &req->pbuf;
3524 if (sqe->ioprio || sqe->rw_flags)
3527 tmp = READ_ONCE(sqe->fd);
3528 if (!tmp || tmp > USHRT_MAX)
3531 p->addr = READ_ONCE(sqe->addr);
3532 p->len = READ_ONCE(sqe->len);
3534 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3537 p->bgid = READ_ONCE(sqe->buf_group);
3538 tmp = READ_ONCE(sqe->off);
3539 if (tmp > USHRT_MAX)
3545 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3547 struct io_buffer *buf;
3548 u64 addr = pbuf->addr;
3549 int i, bid = pbuf->bid;
3551 for (i = 0; i < pbuf->nbufs; i++) {
3552 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3557 buf->len = pbuf->len;
3562 INIT_LIST_HEAD(&buf->list);
3565 list_add_tail(&buf->list, &(*head)->list);
3569 return i ? i : -ENOMEM;
3572 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3573 struct io_comp_state *cs)
3575 struct io_provide_buf *p = &req->pbuf;
3576 struct io_ring_ctx *ctx = req->ctx;
3577 struct io_buffer *head, *list;
3580 io_ring_submit_lock(ctx, !force_nonblock);
3582 lockdep_assert_held(&ctx->uring_lock);
3584 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3586 ret = io_add_buffers(p, &head);
3591 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3594 __io_remove_buffers(ctx, head, p->bgid, -1U);
3599 io_ring_submit_unlock(ctx, !force_nonblock);
3601 req_set_fail_links(req);
3602 __io_req_complete(req, ret, 0, cs);
3606 static int io_epoll_ctl_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 #if defined(CONFIG_EPOLL)
3610 if (sqe->ioprio || sqe->buf_index)
3612 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3615 req->epoll.epfd = READ_ONCE(sqe->fd);
3616 req->epoll.op = READ_ONCE(sqe->len);
3617 req->epoll.fd = READ_ONCE(sqe->off);
3619 if (ep_op_has_event(req->epoll.op)) {
3620 struct epoll_event __user *ev;
3622 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3623 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3633 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3634 struct io_comp_state *cs)
3636 #if defined(CONFIG_EPOLL)
3637 struct io_epoll *ie = &req->epoll;
3640 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3641 if (force_nonblock && ret == -EAGAIN)
3645 req_set_fail_links(req);
3646 __io_req_complete(req, ret, 0, cs);
3653 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3655 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3656 if (sqe->ioprio || sqe->buf_index || sqe->off)
3658 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3661 req->madvise.addr = READ_ONCE(sqe->addr);
3662 req->madvise.len = READ_ONCE(sqe->len);
3663 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3670 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3672 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3673 struct io_madvise *ma = &req->madvise;
3679 ret = do_madvise(ma->addr, ma->len, ma->advice);
3681 req_set_fail_links(req);
3682 io_req_complete(req, ret);
3689 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3691 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3693 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3696 req->fadvise.offset = READ_ONCE(sqe->off);
3697 req->fadvise.len = READ_ONCE(sqe->len);
3698 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3702 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3704 struct io_fadvise *fa = &req->fadvise;
3707 if (force_nonblock) {
3708 switch (fa->advice) {
3709 case POSIX_FADV_NORMAL:
3710 case POSIX_FADV_RANDOM:
3711 case POSIX_FADV_SEQUENTIAL:
3718 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3720 req_set_fail_links(req);
3721 io_req_complete(req, ret);
3725 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3729 if (sqe->ioprio || sqe->buf_index)
3731 if (req->flags & REQ_F_FIXED_FILE)
3734 req->statx.dfd = READ_ONCE(sqe->fd);
3735 req->statx.mask = READ_ONCE(sqe->len);
3736 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3737 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3738 req->statx.flags = READ_ONCE(sqe->statx_flags);
3743 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3745 struct io_statx *ctx = &req->statx;
3748 if (force_nonblock) {
3749 /* only need file table for an actual valid fd */
3750 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3751 req->flags |= REQ_F_NO_FILE_TABLE;
3755 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3759 req_set_fail_links(req);
3760 io_req_complete(req, ret);
3764 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 * If we queue this for async, it must not be cancellable. That would
3768 * leave the 'file' in an undeterminate state, and here need to modify
3769 * io_wq_work.flags, so initialize io_wq_work firstly.
3771 io_req_init_async(req);
3772 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3774 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3776 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3777 sqe->rw_flags || sqe->buf_index)
3779 if (req->flags & REQ_F_FIXED_FILE)
3782 req->close.fd = READ_ONCE(sqe->fd);
3783 if ((req->file && req->file->f_op == &io_uring_fops) ||
3784 req->close.fd == req->ctx->ring_fd)
3787 req->close.put_file = NULL;
3791 static int io_close(struct io_kiocb *req, bool force_nonblock,
3792 struct io_comp_state *cs)
3794 struct io_close *close = &req->close;
3797 /* might be already done during nonblock submission */
3798 if (!close->put_file) {
3799 ret = __close_fd_get_file(close->fd, &close->put_file);
3801 return (ret == -ENOENT) ? -EBADF : ret;
3804 /* if the file has a flush method, be safe and punt to async */
3805 if (close->put_file->f_op->flush && force_nonblock) {
3806 /* was never set, but play safe */
3807 req->flags &= ~REQ_F_NOWAIT;
3808 /* avoid grabbing files - we don't need the files */
3809 req->flags |= REQ_F_NO_FILE_TABLE;
3813 /* No ->flush() or already async, safely close from here */
3814 ret = filp_close(close->put_file, req->work.files);
3816 req_set_fail_links(req);
3817 fput(close->put_file);
3818 close->put_file = NULL;
3819 __io_req_complete(req, ret, 0, cs);
3823 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3825 struct io_ring_ctx *ctx = req->ctx;
3830 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3832 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3835 req->sync.off = READ_ONCE(sqe->off);
3836 req->sync.len = READ_ONCE(sqe->len);
3837 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3841 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3845 /* sync_file_range always requires a blocking context */
3849 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3852 req_set_fail_links(req);
3853 io_req_complete(req, ret);
3857 #if defined(CONFIG_NET)
3858 static int io_setup_async_msg(struct io_kiocb *req,
3859 struct io_async_msghdr *kmsg)
3863 if (io_alloc_async_ctx(req)) {
3864 if (kmsg->iov != kmsg->fast_iov)
3868 req->flags |= REQ_F_NEED_CLEANUP;
3869 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3873 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3874 struct io_async_msghdr *iomsg)
3876 iomsg->iov = iomsg->fast_iov;
3877 iomsg->msg.msg_name = &iomsg->addr;
3878 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3879 req->sr_msg.msg_flags, &iomsg->iov);
3882 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3884 struct io_sr_msg *sr = &req->sr_msg;
3885 struct io_async_ctx *io = req->io;
3888 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3891 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3892 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3893 sr->len = READ_ONCE(sqe->len);
3895 #ifdef CONFIG_COMPAT
3896 if (req->ctx->compat)
3897 sr->msg_flags |= MSG_CMSG_COMPAT;
3900 if (!io || req->opcode == IORING_OP_SEND)
3902 /* iovec is already imported */
3903 if (req->flags & REQ_F_NEED_CLEANUP)
3906 ret = io_sendmsg_copy_hdr(req, &io->msg);
3908 req->flags |= REQ_F_NEED_CLEANUP;
3912 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3913 struct io_comp_state *cs)
3915 struct io_async_msghdr *kmsg = NULL;
3916 struct socket *sock;
3919 sock = sock_from_file(req->file, &ret);
3921 struct io_async_msghdr iomsg;
3925 kmsg = &req->io->msg;
3926 kmsg->msg.msg_name = &req->io->msg.addr;
3927 /* if iov is set, it's allocated already */
3929 kmsg->iov = kmsg->fast_iov;
3930 kmsg->msg.msg_iter.iov = kmsg->iov;
3932 ret = io_sendmsg_copy_hdr(req, &iomsg);
3938 flags = req->sr_msg.msg_flags;
3939 if (flags & MSG_DONTWAIT)
3940 req->flags |= REQ_F_NOWAIT;
3941 else if (force_nonblock)
3942 flags |= MSG_DONTWAIT;
3944 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3945 if (force_nonblock && ret == -EAGAIN)
3946 return io_setup_async_msg(req, kmsg);
3947 if (ret == -ERESTARTSYS)
3951 if (kmsg && kmsg->iov != kmsg->fast_iov)
3953 req->flags &= ~REQ_F_NEED_CLEANUP;
3955 req_set_fail_links(req);
3956 __io_req_complete(req, ret, 0, cs);
3960 static int io_send(struct io_kiocb *req, bool force_nonblock,
3961 struct io_comp_state *cs)
3963 struct socket *sock;
3966 sock = sock_from_file(req->file, &ret);
3968 struct io_sr_msg *sr = &req->sr_msg;
3973 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3978 msg.msg_name = NULL;
3979 msg.msg_control = NULL;
3980 msg.msg_controllen = 0;
3981 msg.msg_namelen = 0;
3983 flags = req->sr_msg.msg_flags;
3984 if (flags & MSG_DONTWAIT)
3985 req->flags |= REQ_F_NOWAIT;
3986 else if (force_nonblock)
3987 flags |= MSG_DONTWAIT;
3989 msg.msg_flags = flags;
3990 ret = sock_sendmsg(sock, &msg);
3991 if (force_nonblock && ret == -EAGAIN)
3993 if (ret == -ERESTARTSYS)
3998 req_set_fail_links(req);
3999 __io_req_complete(req, ret, 0, cs);
4003 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4004 struct io_async_msghdr *iomsg)
4006 struct io_sr_msg *sr = &req->sr_msg;
4007 struct iovec __user *uiov;
4011 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4012 &iomsg->uaddr, &uiov, &iov_len);
4016 if (req->flags & REQ_F_BUFFER_SELECT) {
4019 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4021 sr->len = iomsg->iov[0].iov_len;
4022 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4026 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4027 &iomsg->iov, &iomsg->msg.msg_iter);
4035 #ifdef CONFIG_COMPAT
4036 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4037 struct io_async_msghdr *iomsg)
4039 struct compat_msghdr __user *msg_compat;
4040 struct io_sr_msg *sr = &req->sr_msg;
4041 struct compat_iovec __user *uiov;
4046 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4047 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4052 uiov = compat_ptr(ptr);
4053 if (req->flags & REQ_F_BUFFER_SELECT) {
4054 compat_ssize_t clen;
4058 if (!access_ok(uiov, sizeof(*uiov)))
4060 if (__get_user(clen, &uiov->iov_len))
4064 sr->len = iomsg->iov[0].iov_len;
4067 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4069 &iomsg->msg.msg_iter);
4078 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4079 struct io_async_msghdr *iomsg)
4081 iomsg->msg.msg_name = &iomsg->addr;
4082 iomsg->iov = iomsg->fast_iov;
4084 #ifdef CONFIG_COMPAT
4085 if (req->ctx->compat)
4086 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4089 return __io_recvmsg_copy_hdr(req, iomsg);
4092 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4093 int *cflags, bool needs_lock)
4095 struct io_sr_msg *sr = &req->sr_msg;
4096 struct io_buffer *kbuf;
4098 if (!(req->flags & REQ_F_BUFFER_SELECT))
4101 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4106 req->flags |= REQ_F_BUFFER_SELECTED;
4108 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4109 *cflags |= IORING_CQE_F_BUFFER;
4113 static int io_recvmsg_prep(struct io_kiocb *req,
4114 const struct io_uring_sqe *sqe)
4116 struct io_sr_msg *sr = &req->sr_msg;
4117 struct io_async_ctx *io = req->io;
4120 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4123 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4124 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4125 sr->len = READ_ONCE(sqe->len);
4126 sr->bgid = READ_ONCE(sqe->buf_group);
4128 #ifdef CONFIG_COMPAT
4129 if (req->ctx->compat)
4130 sr->msg_flags |= MSG_CMSG_COMPAT;
4133 if (!io || req->opcode == IORING_OP_RECV)
4135 /* iovec is already imported */
4136 if (req->flags & REQ_F_NEED_CLEANUP)
4139 ret = io_recvmsg_copy_hdr(req, &io->msg);
4141 req->flags |= REQ_F_NEED_CLEANUP;
4145 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4146 struct io_comp_state *cs)
4148 struct io_async_msghdr *kmsg = NULL;
4149 struct socket *sock;
4150 int ret, cflags = 0;
4152 sock = sock_from_file(req->file, &ret);
4154 struct io_buffer *kbuf;
4155 struct io_async_msghdr iomsg;
4159 kmsg = &req->io->msg;
4160 kmsg->msg.msg_name = &req->io->msg.addr;
4161 /* if iov is set, it's allocated already */
4163 kmsg->iov = kmsg->fast_iov;
4164 kmsg->msg.msg_iter.iov = kmsg->iov;
4166 ret = io_recvmsg_copy_hdr(req, &iomsg);
4172 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4174 return PTR_ERR(kbuf);
4176 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4177 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4178 1, req->sr_msg.len);
4181 flags = req->sr_msg.msg_flags;
4182 if (flags & MSG_DONTWAIT)
4183 req->flags |= REQ_F_NOWAIT;
4184 else if (force_nonblock)
4185 flags |= MSG_DONTWAIT;
4187 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4188 kmsg->uaddr, flags);
4189 if (force_nonblock && ret == -EAGAIN) {
4190 ret = io_setup_async_msg(req, kmsg);
4195 if (ret == -ERESTARTSYS)
4201 if (kmsg && kmsg->iov != kmsg->fast_iov)
4203 req->flags &= ~REQ_F_NEED_CLEANUP;
4205 req_set_fail_links(req);
4206 __io_req_complete(req, ret, cflags, cs);
4210 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4211 struct io_comp_state *cs)
4213 struct io_buffer *kbuf = NULL;
4214 struct socket *sock;
4215 int ret, cflags = 0;
4217 sock = sock_from_file(req->file, &ret);
4219 struct io_sr_msg *sr = &req->sr_msg;
4220 void __user *buf = sr->buf;
4225 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4227 return PTR_ERR(kbuf);
4229 buf = u64_to_user_ptr(kbuf->addr);
4231 ret = import_single_range(READ, buf, sr->len, &iov,
4238 req->flags |= REQ_F_NEED_CLEANUP;
4239 msg.msg_name = NULL;
4240 msg.msg_control = NULL;
4241 msg.msg_controllen = 0;
4242 msg.msg_namelen = 0;
4243 msg.msg_iocb = NULL;
4246 flags = req->sr_msg.msg_flags;
4247 if (flags & MSG_DONTWAIT)
4248 req->flags |= REQ_F_NOWAIT;
4249 else if (force_nonblock)
4250 flags |= MSG_DONTWAIT;
4252 ret = sock_recvmsg(sock, &msg, flags);
4253 if (force_nonblock && ret == -EAGAIN)
4255 if (ret == -ERESTARTSYS)
4260 req->flags &= ~REQ_F_NEED_CLEANUP;
4262 req_set_fail_links(req);
4263 __io_req_complete(req, ret, cflags, cs);
4267 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4269 struct io_accept *accept = &req->accept;
4271 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4273 if (sqe->ioprio || sqe->len || sqe->buf_index)
4276 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4277 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4278 accept->flags = READ_ONCE(sqe->accept_flags);
4279 accept->nofile = rlimit(RLIMIT_NOFILE);
4283 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4284 struct io_comp_state *cs)
4286 struct io_accept *accept = &req->accept;
4287 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4290 if (req->file->f_flags & O_NONBLOCK)
4291 req->flags |= REQ_F_NOWAIT;
4293 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4294 accept->addr_len, accept->flags,
4296 if (ret == -EAGAIN && force_nonblock)
4299 if (ret == -ERESTARTSYS)
4301 req_set_fail_links(req);
4303 __io_req_complete(req, ret, 0, cs);
4307 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4309 struct io_connect *conn = &req->connect;
4310 struct io_async_ctx *io = req->io;
4312 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4314 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4317 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4318 conn->addr_len = READ_ONCE(sqe->addr2);
4323 return move_addr_to_kernel(conn->addr, conn->addr_len,
4324 &io->connect.address);
4327 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4328 struct io_comp_state *cs)
4330 struct io_async_ctx __io, *io;
4331 unsigned file_flags;
4337 ret = move_addr_to_kernel(req->connect.addr,
4338 req->connect.addr_len,
4339 &__io.connect.address);
4345 file_flags = force_nonblock ? O_NONBLOCK : 0;
4347 ret = __sys_connect_file(req->file, &io->connect.address,
4348 req->connect.addr_len, file_flags);
4349 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4352 if (io_alloc_async_ctx(req)) {
4356 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4359 if (ret == -ERESTARTSYS)
4363 req_set_fail_links(req);
4364 __io_req_complete(req, ret, 0, cs);
4367 #else /* !CONFIG_NET */
4368 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4373 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4374 struct io_comp_state *cs)
4379 static int io_send(struct io_kiocb *req, bool force_nonblock,
4380 struct io_comp_state *cs)
4385 static int io_recvmsg_prep(struct io_kiocb *req,
4386 const struct io_uring_sqe *sqe)
4391 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4392 struct io_comp_state *cs)
4397 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4398 struct io_comp_state *cs)
4403 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4408 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4409 struct io_comp_state *cs)
4414 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4419 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4420 struct io_comp_state *cs)
4424 #endif /* CONFIG_NET */
4426 struct io_poll_table {
4427 struct poll_table_struct pt;
4428 struct io_kiocb *req;
4432 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4433 __poll_t mask, task_work_func_t func)
4437 /* for instances that support it check for an event match first: */
4438 if (mask && !(mask & poll->events))
4441 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4443 list_del_init(&poll->wait.entry);
4446 init_task_work(&req->task_work, func);
4448 * If this fails, then the task is exiting. When a task exits, the
4449 * work gets canceled, so just cancel this request as well instead
4450 * of executing it. We can't safely execute it anyway, as we may not
4451 * have the needed state needed for it anyway.
4453 ret = io_req_task_work_add(req, &req->task_work);
4454 if (unlikely(ret)) {
4455 struct task_struct *tsk;
4457 WRITE_ONCE(poll->canceled, true);
4458 tsk = io_wq_get_task(req->ctx->io_wq);
4459 task_work_add(tsk, &req->task_work, 0);
4460 wake_up_process(tsk);
4465 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4466 __acquires(&req->ctx->completion_lock)
4468 struct io_ring_ctx *ctx = req->ctx;
4470 if (!req->result && !READ_ONCE(poll->canceled)) {
4471 struct poll_table_struct pt = { ._key = poll->events };
4473 req->result = vfs_poll(req->file, &pt) & poll->events;
4476 spin_lock_irq(&ctx->completion_lock);
4477 if (!req->result && !READ_ONCE(poll->canceled)) {
4478 add_wait_queue(poll->head, &poll->wait);
4485 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4487 struct io_poll_iocb *poll = data;
4489 lockdep_assert_held(&req->ctx->completion_lock);
4491 if (poll && poll->head) {
4492 struct wait_queue_head *head = poll->head;
4494 spin_lock(&head->lock);
4495 list_del_init(&poll->wait.entry);
4496 if (poll->wait.private)
4497 refcount_dec(&req->refs);
4499 spin_unlock(&head->lock);
4503 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4505 struct io_ring_ctx *ctx = req->ctx;
4507 io_poll_remove_double(req, req->io);
4508 req->poll.done = true;
4509 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4510 io_commit_cqring(ctx);
4513 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4515 struct io_ring_ctx *ctx = req->ctx;
4517 if (io_poll_rewait(req, &req->poll)) {
4518 spin_unlock_irq(&ctx->completion_lock);
4522 hash_del(&req->hash_node);
4523 io_poll_complete(req, req->result, 0);
4524 req->flags |= REQ_F_COMP_LOCKED;
4525 *nxt = io_put_req_find_next(req);
4526 spin_unlock_irq(&ctx->completion_lock);
4528 io_cqring_ev_posted(ctx);
4531 static void io_poll_task_func(struct callback_head *cb)
4533 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4534 struct io_kiocb *nxt = NULL;
4536 io_poll_task_handler(req, &nxt);
4538 __io_req_task_submit(nxt);
4541 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4542 int sync, void *key)
4544 struct io_kiocb *req = wait->private;
4545 struct io_poll_iocb *poll = req->apoll->double_poll;
4546 __poll_t mask = key_to_poll(key);
4548 /* for instances that support it check for an event match first: */
4549 if (mask && !(mask & poll->events))
4552 if (poll && poll->head) {
4555 spin_lock(&poll->head->lock);
4556 done = list_empty(&poll->wait.entry);
4558 list_del_init(&poll->wait.entry);
4559 spin_unlock(&poll->head->lock);
4561 __io_async_wake(req, poll, mask, io_poll_task_func);
4563 refcount_dec(&req->refs);
4567 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4568 wait_queue_func_t wake_func)
4572 poll->canceled = false;
4573 poll->events = events;
4574 INIT_LIST_HEAD(&poll->wait.entry);
4575 init_waitqueue_func_entry(&poll->wait, wake_func);
4578 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4579 struct wait_queue_head *head,
4580 struct io_poll_iocb **poll_ptr)
4582 struct io_kiocb *req = pt->req;
4585 * If poll->head is already set, it's because the file being polled
4586 * uses multiple waitqueues for poll handling (eg one for read, one
4587 * for write). Setup a separate io_poll_iocb if this happens.
4589 if (unlikely(poll->head)) {
4590 /* already have a 2nd entry, fail a third attempt */
4592 pt->error = -EINVAL;
4595 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4597 pt->error = -ENOMEM;
4600 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4601 refcount_inc(&req->refs);
4602 poll->wait.private = req;
4609 if (poll->events & EPOLLEXCLUSIVE)
4610 add_wait_queue_exclusive(head, &poll->wait);
4612 add_wait_queue(head, &poll->wait);
4615 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4616 struct poll_table_struct *p)
4618 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4619 struct async_poll *apoll = pt->req->apoll;
4621 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4624 static void io_async_task_func(struct callback_head *cb)
4626 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4627 struct async_poll *apoll = req->apoll;
4628 struct io_ring_ctx *ctx = req->ctx;
4630 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4632 if (io_poll_rewait(req, &apoll->poll)) {
4633 spin_unlock_irq(&ctx->completion_lock);
4637 /* If req is still hashed, it cannot have been canceled. Don't check. */
4638 if (hash_hashed(&req->hash_node))
4639 hash_del(&req->hash_node);
4641 io_poll_remove_double(req, apoll->double_poll);
4642 spin_unlock_irq(&ctx->completion_lock);
4644 /* restore ->work in case we need to retry again */
4645 if (req->flags & REQ_F_WORK_INITIALIZED)
4646 memcpy(&req->work, &apoll->work, sizeof(req->work));
4648 if (!READ_ONCE(apoll->poll.canceled))
4649 __io_req_task_submit(req);
4651 __io_req_task_cancel(req, -ECANCELED);
4653 kfree(apoll->double_poll);
4657 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4660 struct io_kiocb *req = wait->private;
4661 struct io_poll_iocb *poll = &req->apoll->poll;
4663 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4666 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4669 static void io_poll_req_insert(struct io_kiocb *req)
4671 struct io_ring_ctx *ctx = req->ctx;
4672 struct hlist_head *list;
4674 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4675 hlist_add_head(&req->hash_node, list);
4678 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4679 struct io_poll_iocb *poll,
4680 struct io_poll_table *ipt, __poll_t mask,
4681 wait_queue_func_t wake_func)
4682 __acquires(&ctx->completion_lock)
4684 struct io_ring_ctx *ctx = req->ctx;
4685 bool cancel = false;
4687 io_init_poll_iocb(poll, mask, wake_func);
4688 poll->file = req->file;
4689 poll->wait.private = req;
4691 ipt->pt._key = mask;
4693 ipt->error = -EINVAL;
4695 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4697 spin_lock_irq(&ctx->completion_lock);
4698 if (likely(poll->head)) {
4699 spin_lock(&poll->head->lock);
4700 if (unlikely(list_empty(&poll->wait.entry))) {
4706 if (mask || ipt->error)
4707 list_del_init(&poll->wait.entry);
4709 WRITE_ONCE(poll->canceled, true);
4710 else if (!poll->done) /* actually waiting for an event */
4711 io_poll_req_insert(req);
4712 spin_unlock(&poll->head->lock);
4718 static bool io_arm_poll_handler(struct io_kiocb *req)
4720 const struct io_op_def *def = &io_op_defs[req->opcode];
4721 struct io_ring_ctx *ctx = req->ctx;
4722 struct async_poll *apoll;
4723 struct io_poll_table ipt;
4726 if (!req->file || !file_can_poll(req->file))
4728 if (req->flags & REQ_F_POLLED)
4730 if (!def->pollin && !def->pollout)
4733 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4734 if (unlikely(!apoll))
4736 apoll->double_poll = NULL;
4738 req->flags |= REQ_F_POLLED;
4739 if (req->flags & REQ_F_WORK_INITIALIZED)
4740 memcpy(&apoll->work, &req->work, sizeof(req->work));
4742 io_get_req_task(req);
4744 INIT_HLIST_NODE(&req->hash_node);
4748 mask |= POLLIN | POLLRDNORM;
4750 mask |= POLLOUT | POLLWRNORM;
4751 mask |= POLLERR | POLLPRI;
4753 ipt.pt._qproc = io_async_queue_proc;
4755 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4758 io_poll_remove_double(req, apoll->double_poll);
4759 spin_unlock_irq(&ctx->completion_lock);
4760 if (req->flags & REQ_F_WORK_INITIALIZED)
4761 memcpy(&req->work, &apoll->work, sizeof(req->work));
4762 kfree(apoll->double_poll);
4766 spin_unlock_irq(&ctx->completion_lock);
4767 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4768 apoll->poll.events);
4772 static bool __io_poll_remove_one(struct io_kiocb *req,
4773 struct io_poll_iocb *poll)
4775 bool do_complete = false;
4777 spin_lock(&poll->head->lock);
4778 WRITE_ONCE(poll->canceled, true);
4779 if (!list_empty(&poll->wait.entry)) {
4780 list_del_init(&poll->wait.entry);
4783 spin_unlock(&poll->head->lock);
4784 hash_del(&req->hash_node);
4788 static bool io_poll_remove_one(struct io_kiocb *req)
4792 if (req->opcode == IORING_OP_POLL_ADD) {
4793 io_poll_remove_double(req, req->io);
4794 do_complete = __io_poll_remove_one(req, &req->poll);
4796 struct async_poll *apoll = req->apoll;
4798 io_poll_remove_double(req, apoll->double_poll);
4800 /* non-poll requests have submit ref still */
4801 do_complete = __io_poll_remove_one(req, &apoll->poll);
4805 * restore ->work because we will call
4806 * io_req_work_drop_env below when dropping the
4809 if (req->flags & REQ_F_WORK_INITIALIZED)
4810 memcpy(&req->work, &apoll->work,
4812 kfree(apoll->double_poll);
4818 io_cqring_fill_event(req, -ECANCELED);
4819 io_commit_cqring(req->ctx);
4820 req->flags |= REQ_F_COMP_LOCKED;
4827 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4829 struct hlist_node *tmp;
4830 struct io_kiocb *req;
4833 spin_lock_irq(&ctx->completion_lock);
4834 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4835 struct hlist_head *list;
4837 list = &ctx->cancel_hash[i];
4838 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4839 posted += io_poll_remove_one(req);
4841 spin_unlock_irq(&ctx->completion_lock);
4844 io_cqring_ev_posted(ctx);
4847 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4849 struct hlist_head *list;
4850 struct io_kiocb *req;
4852 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4853 hlist_for_each_entry(req, list, hash_node) {
4854 if (sqe_addr != req->user_data)
4856 if (io_poll_remove_one(req))
4864 static int io_poll_remove_prep(struct io_kiocb *req,
4865 const struct io_uring_sqe *sqe)
4867 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4869 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4873 req->poll.addr = READ_ONCE(sqe->addr);
4878 * Find a running poll command that matches one specified in sqe->addr,
4879 * and remove it if found.
4881 static int io_poll_remove(struct io_kiocb *req)
4883 struct io_ring_ctx *ctx = req->ctx;
4887 addr = req->poll.addr;
4888 spin_lock_irq(&ctx->completion_lock);
4889 ret = io_poll_cancel(ctx, addr);
4890 spin_unlock_irq(&ctx->completion_lock);
4893 req_set_fail_links(req);
4894 io_req_complete(req, ret);
4898 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4901 struct io_kiocb *req = wait->private;
4902 struct io_poll_iocb *poll = &req->poll;
4904 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4907 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4908 struct poll_table_struct *p)
4910 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4912 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4915 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4917 struct io_poll_iocb *poll = &req->poll;
4920 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4922 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4927 events = READ_ONCE(sqe->poll32_events);
4929 events = swahw32(events);
4931 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4932 (events & EPOLLEXCLUSIVE);
4934 io_get_req_task(req);
4938 static int io_poll_add(struct io_kiocb *req)
4940 struct io_poll_iocb *poll = &req->poll;
4941 struct io_ring_ctx *ctx = req->ctx;
4942 struct io_poll_table ipt;
4945 /* ->work is in union with hash_node and others */
4946 io_req_work_drop_env(req);
4947 req->flags &= ~REQ_F_WORK_INITIALIZED;
4949 INIT_HLIST_NODE(&req->hash_node);
4950 INIT_LIST_HEAD(&req->list);
4951 ipt.pt._qproc = io_poll_queue_proc;
4953 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4956 if (mask) { /* no async, we'd stolen it */
4958 io_poll_complete(req, mask, 0);
4960 spin_unlock_irq(&ctx->completion_lock);
4963 io_cqring_ev_posted(ctx);
4969 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4971 struct io_timeout_data *data = container_of(timer,
4972 struct io_timeout_data, timer);
4973 struct io_kiocb *req = data->req;
4974 struct io_ring_ctx *ctx = req->ctx;
4975 unsigned long flags;
4977 atomic_inc(&ctx->cq_timeouts);
4979 spin_lock_irqsave(&ctx->completion_lock, flags);
4981 * We could be racing with timeout deletion. If the list is empty,
4982 * then timeout lookup already found it and will be handling it.
4984 if (!list_empty(&req->list))
4985 list_del_init(&req->list);
4987 io_cqring_fill_event(req, -ETIME);
4988 io_commit_cqring(ctx);
4989 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4991 io_cqring_ev_posted(ctx);
4992 req_set_fail_links(req);
4994 return HRTIMER_NORESTART;
4997 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4999 struct io_kiocb *req;
5002 list_for_each_entry(req, &ctx->timeout_list, list) {
5003 if (user_data == req->user_data) {
5004 list_del_init(&req->list);
5013 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5017 req_set_fail_links(req);
5018 io_cqring_fill_event(req, -ECANCELED);
5023 static int io_timeout_remove_prep(struct io_kiocb *req,
5024 const struct io_uring_sqe *sqe)
5026 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5028 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5030 if (sqe->ioprio || sqe->buf_index || sqe->len)
5033 req->timeout.addr = READ_ONCE(sqe->addr);
5034 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5035 if (req->timeout.flags)
5042 * Remove or update an existing timeout command
5044 static int io_timeout_remove(struct io_kiocb *req)
5046 struct io_ring_ctx *ctx = req->ctx;
5049 spin_lock_irq(&ctx->completion_lock);
5050 ret = io_timeout_cancel(ctx, req->timeout.addr);
5052 io_cqring_fill_event(req, ret);
5053 io_commit_cqring(ctx);
5054 spin_unlock_irq(&ctx->completion_lock);
5055 io_cqring_ev_posted(ctx);
5057 req_set_fail_links(req);
5062 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5063 bool is_timeout_link)
5065 struct io_timeout_data *data;
5067 u32 off = READ_ONCE(sqe->off);
5069 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5071 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5073 if (off && is_timeout_link)
5075 flags = READ_ONCE(sqe->timeout_flags);
5076 if (flags & ~IORING_TIMEOUT_ABS)
5079 req->timeout.off = off;
5081 if (!req->io && io_alloc_async_ctx(req))
5084 data = &req->io->timeout;
5087 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5090 if (flags & IORING_TIMEOUT_ABS)
5091 data->mode = HRTIMER_MODE_ABS;
5093 data->mode = HRTIMER_MODE_REL;
5095 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5099 static int io_timeout(struct io_kiocb *req)
5101 struct io_ring_ctx *ctx = req->ctx;
5102 struct io_timeout_data *data = &req->io->timeout;
5103 struct list_head *entry;
5104 u32 tail, off = req->timeout.off;
5106 spin_lock_irq(&ctx->completion_lock);
5109 * sqe->off holds how many events that need to occur for this
5110 * timeout event to be satisfied. If it isn't set, then this is
5111 * a pure timeout request, sequence isn't used.
5113 if (io_is_timeout_noseq(req)) {
5114 entry = ctx->timeout_list.prev;
5118 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5119 req->timeout.target_seq = tail + off;
5122 * Insertion sort, ensuring the first entry in the list is always
5123 * the one we need first.
5125 list_for_each_prev(entry, &ctx->timeout_list) {
5126 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5128 if (io_is_timeout_noseq(nxt))
5130 /* nxt.seq is behind @tail, otherwise would've been completed */
5131 if (off >= nxt->timeout.target_seq - tail)
5135 list_add(&req->list, entry);
5136 data->timer.function = io_timeout_fn;
5137 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5138 spin_unlock_irq(&ctx->completion_lock);
5142 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5144 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5146 return req->user_data == (unsigned long) data;
5149 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5151 enum io_wq_cancel cancel_ret;
5154 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5155 switch (cancel_ret) {
5156 case IO_WQ_CANCEL_OK:
5159 case IO_WQ_CANCEL_RUNNING:
5162 case IO_WQ_CANCEL_NOTFOUND:
5170 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5171 struct io_kiocb *req, __u64 sqe_addr,
5174 unsigned long flags;
5177 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5178 if (ret != -ENOENT) {
5179 spin_lock_irqsave(&ctx->completion_lock, flags);
5183 spin_lock_irqsave(&ctx->completion_lock, flags);
5184 ret = io_timeout_cancel(ctx, sqe_addr);
5187 ret = io_poll_cancel(ctx, sqe_addr);
5191 io_cqring_fill_event(req, ret);
5192 io_commit_cqring(ctx);
5193 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5194 io_cqring_ev_posted(ctx);
5197 req_set_fail_links(req);
5201 static int io_async_cancel_prep(struct io_kiocb *req,
5202 const struct io_uring_sqe *sqe)
5204 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5206 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5208 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5211 req->cancel.addr = READ_ONCE(sqe->addr);
5215 static int io_async_cancel(struct io_kiocb *req)
5217 struct io_ring_ctx *ctx = req->ctx;
5219 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5223 static int io_files_update_prep(struct io_kiocb *req,
5224 const struct io_uring_sqe *sqe)
5226 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5228 if (sqe->ioprio || sqe->rw_flags)
5231 req->files_update.offset = READ_ONCE(sqe->off);
5232 req->files_update.nr_args = READ_ONCE(sqe->len);
5233 if (!req->files_update.nr_args)
5235 req->files_update.arg = READ_ONCE(sqe->addr);
5239 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5240 struct io_comp_state *cs)
5242 struct io_ring_ctx *ctx = req->ctx;
5243 struct io_uring_files_update up;
5249 up.offset = req->files_update.offset;
5250 up.fds = req->files_update.arg;
5252 mutex_lock(&ctx->uring_lock);
5253 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5254 mutex_unlock(&ctx->uring_lock);
5257 req_set_fail_links(req);
5258 __io_req_complete(req, ret, 0, cs);
5262 static int io_req_defer_prep(struct io_kiocb *req,
5263 const struct io_uring_sqe *sqe)
5270 if (io_op_defs[req->opcode].file_table) {
5271 io_req_init_async(req);
5272 ret = io_grab_files(req);
5277 switch (req->opcode) {
5280 case IORING_OP_READV:
5281 case IORING_OP_READ_FIXED:
5282 case IORING_OP_READ:
5283 ret = io_read_prep(req, sqe, true);
5285 case IORING_OP_WRITEV:
5286 case IORING_OP_WRITE_FIXED:
5287 case IORING_OP_WRITE:
5288 ret = io_write_prep(req, sqe, true);
5290 case IORING_OP_POLL_ADD:
5291 ret = io_poll_add_prep(req, sqe);
5293 case IORING_OP_POLL_REMOVE:
5294 ret = io_poll_remove_prep(req, sqe);
5296 case IORING_OP_FSYNC:
5297 ret = io_prep_fsync(req, sqe);
5299 case IORING_OP_SYNC_FILE_RANGE:
5300 ret = io_prep_sfr(req, sqe);
5302 case IORING_OP_SENDMSG:
5303 case IORING_OP_SEND:
5304 ret = io_sendmsg_prep(req, sqe);
5306 case IORING_OP_RECVMSG:
5307 case IORING_OP_RECV:
5308 ret = io_recvmsg_prep(req, sqe);
5310 case IORING_OP_CONNECT:
5311 ret = io_connect_prep(req, sqe);
5313 case IORING_OP_TIMEOUT:
5314 ret = io_timeout_prep(req, sqe, false);
5316 case IORING_OP_TIMEOUT_REMOVE:
5317 ret = io_timeout_remove_prep(req, sqe);
5319 case IORING_OP_ASYNC_CANCEL:
5320 ret = io_async_cancel_prep(req, sqe);
5322 case IORING_OP_LINK_TIMEOUT:
5323 ret = io_timeout_prep(req, sqe, true);
5325 case IORING_OP_ACCEPT:
5326 ret = io_accept_prep(req, sqe);
5328 case IORING_OP_FALLOCATE:
5329 ret = io_fallocate_prep(req, sqe);
5331 case IORING_OP_OPENAT:
5332 ret = io_openat_prep(req, sqe);
5334 case IORING_OP_CLOSE:
5335 ret = io_close_prep(req, sqe);
5337 case IORING_OP_FILES_UPDATE:
5338 ret = io_files_update_prep(req, sqe);
5340 case IORING_OP_STATX:
5341 ret = io_statx_prep(req, sqe);
5343 case IORING_OP_FADVISE:
5344 ret = io_fadvise_prep(req, sqe);
5346 case IORING_OP_MADVISE:
5347 ret = io_madvise_prep(req, sqe);
5349 case IORING_OP_OPENAT2:
5350 ret = io_openat2_prep(req, sqe);
5352 case IORING_OP_EPOLL_CTL:
5353 ret = io_epoll_ctl_prep(req, sqe);
5355 case IORING_OP_SPLICE:
5356 ret = io_splice_prep(req, sqe);
5358 case IORING_OP_PROVIDE_BUFFERS:
5359 ret = io_provide_buffers_prep(req, sqe);
5361 case IORING_OP_REMOVE_BUFFERS:
5362 ret = io_remove_buffers_prep(req, sqe);
5365 ret = io_tee_prep(req, sqe);
5368 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5377 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5379 struct io_ring_ctx *ctx = req->ctx;
5382 /* Still need defer if there is pending req in defer list. */
5383 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5387 if (io_alloc_async_ctx(req))
5389 ret = io_req_defer_prep(req, sqe);
5393 io_prep_async_link(req);
5395 spin_lock_irq(&ctx->completion_lock);
5396 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5397 spin_unlock_irq(&ctx->completion_lock);
5401 trace_io_uring_defer(ctx, req, req->user_data);
5402 list_add_tail(&req->list, &ctx->defer_list);
5403 spin_unlock_irq(&ctx->completion_lock);
5404 return -EIOCBQUEUED;
5407 static void io_cleanup_req(struct io_kiocb *req)
5409 struct io_async_ctx *io = req->io;
5411 switch (req->opcode) {
5412 case IORING_OP_READV:
5413 case IORING_OP_READ_FIXED:
5414 case IORING_OP_READ:
5415 if (req->flags & REQ_F_BUFFER_SELECTED)
5416 kfree((void *)(unsigned long)req->rw.addr);
5418 case IORING_OP_WRITEV:
5419 case IORING_OP_WRITE_FIXED:
5420 case IORING_OP_WRITE:
5421 if (io->rw.iov != io->rw.fast_iov)
5424 case IORING_OP_RECVMSG:
5425 if (req->flags & REQ_F_BUFFER_SELECTED)
5426 kfree(req->sr_msg.kbuf);
5428 case IORING_OP_SENDMSG:
5429 if (io->msg.iov != io->msg.fast_iov)
5432 case IORING_OP_RECV:
5433 if (req->flags & REQ_F_BUFFER_SELECTED)
5434 kfree(req->sr_msg.kbuf);
5436 case IORING_OP_OPENAT:
5437 case IORING_OP_OPENAT2:
5439 case IORING_OP_SPLICE:
5441 io_put_file(req, req->splice.file_in,
5442 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5446 req->flags &= ~REQ_F_NEED_CLEANUP;
5449 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5450 bool force_nonblock, struct io_comp_state *cs)
5452 struct io_ring_ctx *ctx = req->ctx;
5455 switch (req->opcode) {
5457 ret = io_nop(req, cs);
5459 case IORING_OP_READV:
5460 case IORING_OP_READ_FIXED:
5461 case IORING_OP_READ:
5463 ret = io_read_prep(req, sqe, force_nonblock);
5467 ret = io_read(req, force_nonblock, cs);
5469 case IORING_OP_WRITEV:
5470 case IORING_OP_WRITE_FIXED:
5471 case IORING_OP_WRITE:
5473 ret = io_write_prep(req, sqe, force_nonblock);
5477 ret = io_write(req, force_nonblock, cs);
5479 case IORING_OP_FSYNC:
5481 ret = io_prep_fsync(req, sqe);
5485 ret = io_fsync(req, force_nonblock);
5487 case IORING_OP_POLL_ADD:
5489 ret = io_poll_add_prep(req, sqe);
5493 ret = io_poll_add(req);
5495 case IORING_OP_POLL_REMOVE:
5497 ret = io_poll_remove_prep(req, sqe);
5501 ret = io_poll_remove(req);
5503 case IORING_OP_SYNC_FILE_RANGE:
5505 ret = io_prep_sfr(req, sqe);
5509 ret = io_sync_file_range(req, force_nonblock);
5511 case IORING_OP_SENDMSG:
5512 case IORING_OP_SEND:
5514 ret = io_sendmsg_prep(req, sqe);
5518 if (req->opcode == IORING_OP_SENDMSG)
5519 ret = io_sendmsg(req, force_nonblock, cs);
5521 ret = io_send(req, force_nonblock, cs);
5523 case IORING_OP_RECVMSG:
5524 case IORING_OP_RECV:
5526 ret = io_recvmsg_prep(req, sqe);
5530 if (req->opcode == IORING_OP_RECVMSG)
5531 ret = io_recvmsg(req, force_nonblock, cs);
5533 ret = io_recv(req, force_nonblock, cs);
5535 case IORING_OP_TIMEOUT:
5537 ret = io_timeout_prep(req, sqe, false);
5541 ret = io_timeout(req);
5543 case IORING_OP_TIMEOUT_REMOVE:
5545 ret = io_timeout_remove_prep(req, sqe);
5549 ret = io_timeout_remove(req);
5551 case IORING_OP_ACCEPT:
5553 ret = io_accept_prep(req, sqe);
5557 ret = io_accept(req, force_nonblock, cs);
5559 case IORING_OP_CONNECT:
5561 ret = io_connect_prep(req, sqe);
5565 ret = io_connect(req, force_nonblock, cs);
5567 case IORING_OP_ASYNC_CANCEL:
5569 ret = io_async_cancel_prep(req, sqe);
5573 ret = io_async_cancel(req);
5575 case IORING_OP_FALLOCATE:
5577 ret = io_fallocate_prep(req, sqe);
5581 ret = io_fallocate(req, force_nonblock);
5583 case IORING_OP_OPENAT:
5585 ret = io_openat_prep(req, sqe);
5589 ret = io_openat(req, force_nonblock);
5591 case IORING_OP_CLOSE:
5593 ret = io_close_prep(req, sqe);
5597 ret = io_close(req, force_nonblock, cs);
5599 case IORING_OP_FILES_UPDATE:
5601 ret = io_files_update_prep(req, sqe);
5605 ret = io_files_update(req, force_nonblock, cs);
5607 case IORING_OP_STATX:
5609 ret = io_statx_prep(req, sqe);
5613 ret = io_statx(req, force_nonblock);
5615 case IORING_OP_FADVISE:
5617 ret = io_fadvise_prep(req, sqe);
5621 ret = io_fadvise(req, force_nonblock);
5623 case IORING_OP_MADVISE:
5625 ret = io_madvise_prep(req, sqe);
5629 ret = io_madvise(req, force_nonblock);
5631 case IORING_OP_OPENAT2:
5633 ret = io_openat2_prep(req, sqe);
5637 ret = io_openat2(req, force_nonblock);
5639 case IORING_OP_EPOLL_CTL:
5641 ret = io_epoll_ctl_prep(req, sqe);
5645 ret = io_epoll_ctl(req, force_nonblock, cs);
5647 case IORING_OP_SPLICE:
5649 ret = io_splice_prep(req, sqe);
5653 ret = io_splice(req, force_nonblock);
5655 case IORING_OP_PROVIDE_BUFFERS:
5657 ret = io_provide_buffers_prep(req, sqe);
5661 ret = io_provide_buffers(req, force_nonblock, cs);
5663 case IORING_OP_REMOVE_BUFFERS:
5665 ret = io_remove_buffers_prep(req, sqe);
5669 ret = io_remove_buffers(req, force_nonblock, cs);
5673 ret = io_tee_prep(req, sqe);
5677 ret = io_tee(req, force_nonblock);
5687 /* If the op doesn't have a file, we're not polling for it */
5688 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5689 const bool in_async = io_wq_current_is_worker();
5691 /* workqueue context doesn't hold uring_lock, grab it now */
5693 mutex_lock(&ctx->uring_lock);
5695 io_iopoll_req_issued(req);
5698 mutex_unlock(&ctx->uring_lock);
5704 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5706 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5707 struct io_kiocb *timeout;
5710 timeout = io_prep_linked_timeout(req);
5712 io_queue_linked_timeout(timeout);
5714 /* if NO_CANCEL is set, we must still run the work */
5715 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5716 IO_WQ_WORK_CANCEL) {
5722 ret = io_issue_sqe(req, NULL, false, NULL);
5724 * We can get EAGAIN for polled IO even though we're
5725 * forcing a sync submission from here, since we can't
5726 * wait for request slots on the block side.
5735 req_set_fail_links(req);
5736 io_req_complete(req, ret);
5739 return io_steal_work(req);
5742 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5745 struct fixed_file_table *table;
5747 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5748 return table->files[index & IORING_FILE_TABLE_MASK];
5751 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5752 int fd, struct file **out_file, bool fixed)
5754 struct io_ring_ctx *ctx = req->ctx;
5758 if (unlikely(!ctx->file_data ||
5759 (unsigned) fd >= ctx->nr_user_files))
5761 fd = array_index_nospec(fd, ctx->nr_user_files);
5762 file = io_file_from_index(ctx, fd);
5764 req->fixed_file_refs = ctx->file_data->cur_refs;
5765 percpu_ref_get(req->fixed_file_refs);
5768 trace_io_uring_file_get(ctx, fd);
5769 file = __io_file_get(state, fd);
5772 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5779 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5784 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5785 if (unlikely(!fixed && io_async_submit(req->ctx)))
5788 return io_file_get(state, req, fd, &req->file, fixed);
5791 static int io_grab_files(struct io_kiocb *req)
5794 struct io_ring_ctx *ctx = req->ctx;
5796 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5798 if (!ctx->ring_file)
5802 spin_lock_irq(&ctx->inflight_lock);
5804 * We use the f_ops->flush() handler to ensure that we can flush
5805 * out work accessing these files if the fd is closed. Check if
5806 * the fd has changed since we started down this path, and disallow
5807 * this operation if it has.
5809 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5810 list_add(&req->inflight_entry, &ctx->inflight_list);
5811 req->flags |= REQ_F_INFLIGHT;
5812 req->work.files = current->files;
5815 spin_unlock_irq(&ctx->inflight_lock);
5821 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5823 struct io_timeout_data *data = container_of(timer,
5824 struct io_timeout_data, timer);
5825 struct io_kiocb *req = data->req;
5826 struct io_ring_ctx *ctx = req->ctx;
5827 struct io_kiocb *prev = NULL;
5828 unsigned long flags;
5830 spin_lock_irqsave(&ctx->completion_lock, flags);
5833 * We don't expect the list to be empty, that will only happen if we
5834 * race with the completion of the linked work.
5836 if (!list_empty(&req->link_list)) {
5837 prev = list_entry(req->link_list.prev, struct io_kiocb,
5839 if (refcount_inc_not_zero(&prev->refs)) {
5840 list_del_init(&req->link_list);
5841 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5846 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5849 req_set_fail_links(prev);
5850 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5853 io_req_complete(req, -ETIME);
5855 return HRTIMER_NORESTART;
5858 static void io_queue_linked_timeout(struct io_kiocb *req)
5860 struct io_ring_ctx *ctx = req->ctx;
5863 * If the list is now empty, then our linked request finished before
5864 * we got a chance to setup the timer
5866 spin_lock_irq(&ctx->completion_lock);
5867 if (!list_empty(&req->link_list)) {
5868 struct io_timeout_data *data = &req->io->timeout;
5870 data->timer.function = io_link_timeout_fn;
5871 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5874 spin_unlock_irq(&ctx->completion_lock);
5876 /* drop submission reference */
5880 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5882 struct io_kiocb *nxt;
5884 if (!(req->flags & REQ_F_LINK_HEAD))
5886 if (req->flags & REQ_F_LINK_TIMEOUT)
5889 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5891 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5894 req->flags |= REQ_F_LINK_TIMEOUT;
5898 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5899 struct io_comp_state *cs)
5901 struct io_kiocb *linked_timeout;
5902 struct io_kiocb *nxt;
5903 const struct cred *old_creds = NULL;
5907 linked_timeout = io_prep_linked_timeout(req);
5909 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5910 req->work.creds != current_cred()) {
5912 revert_creds(old_creds);
5913 if (old_creds == req->work.creds)
5914 old_creds = NULL; /* restored original creds */
5916 old_creds = override_creds(req->work.creds);
5919 ret = io_issue_sqe(req, sqe, true, cs);
5922 * We async punt it if the file wasn't marked NOWAIT, or if the file
5923 * doesn't support non-blocking read/write attempts
5925 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5926 if (io_arm_poll_handler(req)) {
5928 io_queue_linked_timeout(linked_timeout);
5932 io_req_init_async(req);
5934 if (io_op_defs[req->opcode].file_table) {
5935 ret = io_grab_files(req);
5941 * Queued up for async execution, worker will release
5942 * submit reference when the iocb is actually submitted.
5944 io_queue_async_work(req);
5948 if (unlikely(ret)) {
5950 /* un-prep timeout, so it'll be killed as any other linked */
5951 req->flags &= ~REQ_F_LINK_TIMEOUT;
5952 req_set_fail_links(req);
5954 io_req_complete(req, ret);
5958 /* drop submission reference */
5959 nxt = io_put_req_find_next(req);
5961 io_queue_linked_timeout(linked_timeout);
5966 if (req->flags & REQ_F_FORCE_ASYNC)
5972 revert_creds(old_creds);
5975 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5976 struct io_comp_state *cs)
5980 ret = io_req_defer(req, sqe);
5982 if (ret != -EIOCBQUEUED) {
5984 req_set_fail_links(req);
5986 io_req_complete(req, ret);
5988 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5991 if (io_alloc_async_ctx(req))
5993 ret = io_req_defer_prep(req, sqe);
5994 if (unlikely(ret < 0))
5999 * Never try inline submit of IOSQE_ASYNC is set, go straight
6000 * to async execution.
6002 io_req_init_async(req);
6003 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6004 io_queue_async_work(req);
6006 __io_queue_sqe(req, sqe, cs);
6010 static inline void io_queue_link_head(struct io_kiocb *req,
6011 struct io_comp_state *cs)
6013 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6015 io_req_complete(req, -ECANCELED);
6017 io_queue_sqe(req, NULL, cs);
6020 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6021 struct io_kiocb **link, struct io_comp_state *cs)
6023 struct io_ring_ctx *ctx = req->ctx;
6027 * If we already have a head request, queue this one for async
6028 * submittal once the head completes. If we don't have a head but
6029 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6030 * submitted sync once the chain is complete. If none of those
6031 * conditions are true (normal request), then just queue it.
6034 struct io_kiocb *head = *link;
6037 * Taking sequential execution of a link, draining both sides
6038 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6039 * requests in the link. So, it drains the head and the
6040 * next after the link request. The last one is done via
6041 * drain_next flag to persist the effect across calls.
6043 if (req->flags & REQ_F_IO_DRAIN) {
6044 head->flags |= REQ_F_IO_DRAIN;
6045 ctx->drain_next = 1;
6047 if (io_alloc_async_ctx(req))
6050 ret = io_req_defer_prep(req, sqe);
6052 /* fail even hard links since we don't submit */
6053 head->flags |= REQ_F_FAIL_LINK;
6056 trace_io_uring_link(ctx, req, head);
6057 io_get_req_task(req);
6058 list_add_tail(&req->link_list, &head->link_list);
6060 /* last request of a link, enqueue the link */
6061 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6062 io_queue_link_head(head, cs);
6066 if (unlikely(ctx->drain_next)) {
6067 req->flags |= REQ_F_IO_DRAIN;
6068 ctx->drain_next = 0;
6070 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6071 req->flags |= REQ_F_LINK_HEAD;
6072 INIT_LIST_HEAD(&req->link_list);
6074 if (io_alloc_async_ctx(req))
6077 ret = io_req_defer_prep(req, sqe);
6079 req->flags |= REQ_F_FAIL_LINK;
6082 io_queue_sqe(req, sqe, cs);
6090 * Batched submission is done, ensure local IO is flushed out.
6092 static void io_submit_state_end(struct io_submit_state *state)
6094 if (!list_empty(&state->comp.list))
6095 io_submit_flush_completions(&state->comp);
6096 blk_finish_plug(&state->plug);
6097 io_state_file_put(state);
6098 if (state->free_reqs)
6099 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6103 * Start submission side cache.
6105 static void io_submit_state_start(struct io_submit_state *state,
6106 struct io_ring_ctx *ctx, unsigned int max_ios)
6108 blk_start_plug(&state->plug);
6110 state->plug.nowait = true;
6113 INIT_LIST_HEAD(&state->comp.list);
6114 state->comp.ctx = ctx;
6115 state->free_reqs = 0;
6117 state->ios_left = max_ios;
6120 static void io_commit_sqring(struct io_ring_ctx *ctx)
6122 struct io_rings *rings = ctx->rings;
6125 * Ensure any loads from the SQEs are done at this point,
6126 * since once we write the new head, the application could
6127 * write new data to them.
6129 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6133 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6134 * that is mapped by userspace. This means that care needs to be taken to
6135 * ensure that reads are stable, as we cannot rely on userspace always
6136 * being a good citizen. If members of the sqe are validated and then later
6137 * used, it's important that those reads are done through READ_ONCE() to
6138 * prevent a re-load down the line.
6140 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6142 u32 *sq_array = ctx->sq_array;
6146 * The cached sq head (or cq tail) serves two purposes:
6148 * 1) allows us to batch the cost of updating the user visible
6150 * 2) allows the kernel side to track the head on its own, even
6151 * though the application is the one updating it.
6153 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6154 if (likely(head < ctx->sq_entries))
6155 return &ctx->sq_sqes[head];
6157 /* drop invalid entries */
6158 ctx->cached_sq_dropped++;
6159 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6163 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6165 ctx->cached_sq_head++;
6168 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6169 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6170 IOSQE_BUFFER_SELECT)
6172 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6173 const struct io_uring_sqe *sqe,
6174 struct io_submit_state *state)
6176 unsigned int sqe_flags;
6180 * All io need record the previous position, if LINK vs DARIN,
6181 * it can be used to mark the position of the first IO in the
6184 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6185 req->opcode = READ_ONCE(sqe->opcode);
6186 req->user_data = READ_ONCE(sqe->user_data);
6191 /* one is dropped after submission, the other at completion */
6192 refcount_set(&req->refs, 2);
6193 req->task = current;
6196 if (unlikely(req->opcode >= IORING_OP_LAST))
6199 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6202 sqe_flags = READ_ONCE(sqe->flags);
6203 /* enforce forwards compatibility on users */
6204 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6207 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6208 !io_op_defs[req->opcode].buffer_select)
6211 id = READ_ONCE(sqe->personality);
6213 io_req_init_async(req);
6214 req->work.creds = idr_find(&ctx->personality_idr, id);
6215 if (unlikely(!req->work.creds))
6217 get_cred(req->work.creds);
6220 /* same numerical values with corresponding REQ_F_*, safe to copy */
6221 req->flags |= sqe_flags;
6223 if (!io_op_defs[req->opcode].needs_file)
6226 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6229 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6230 struct file *ring_file, int ring_fd)
6232 struct io_submit_state state;
6233 struct io_kiocb *link = NULL;
6234 int i, submitted = 0;
6236 /* if we have a backlog and couldn't flush it all, return BUSY */
6237 if (test_bit(0, &ctx->sq_check_overflow)) {
6238 if (!list_empty(&ctx->cq_overflow_list) &&
6239 !io_cqring_overflow_flush(ctx, false))
6243 /* make sure SQ entry isn't read before tail */
6244 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6246 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6249 io_submit_state_start(&state, ctx, nr);
6251 ctx->ring_fd = ring_fd;
6252 ctx->ring_file = ring_file;
6254 for (i = 0; i < nr; i++) {
6255 const struct io_uring_sqe *sqe;
6256 struct io_kiocb *req;
6259 sqe = io_get_sqe(ctx);
6260 if (unlikely(!sqe)) {
6261 io_consume_sqe(ctx);
6264 req = io_alloc_req(ctx, &state);
6265 if (unlikely(!req)) {
6267 submitted = -EAGAIN;
6271 err = io_init_req(ctx, req, sqe, &state);
6272 io_consume_sqe(ctx);
6273 /* will complete beyond this point, count as submitted */
6276 if (unlikely(err)) {
6279 io_req_complete(req, err);
6283 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6284 true, io_async_submit(ctx));
6285 err = io_submit_sqe(req, sqe, &link, &state.comp);
6290 if (unlikely(submitted != nr)) {
6291 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6293 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6296 io_queue_link_head(link, &state.comp);
6297 io_submit_state_end(&state);
6299 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6300 io_commit_sqring(ctx);
6305 static int io_sq_thread(void *data)
6307 struct io_ring_ctx *ctx = data;
6308 const struct cred *old_cred;
6310 unsigned long timeout;
6313 complete(&ctx->sq_thread_comp);
6315 old_cred = override_creds(ctx->creds);
6317 timeout = jiffies + ctx->sq_thread_idle;
6318 while (!kthread_should_park()) {
6319 unsigned int to_submit;
6321 if (!list_empty(&ctx->poll_list)) {
6322 unsigned nr_events = 0;
6324 mutex_lock(&ctx->uring_lock);
6325 if (!list_empty(&ctx->poll_list) && !need_resched())
6326 io_do_iopoll(ctx, &nr_events, 0);
6328 timeout = jiffies + ctx->sq_thread_idle;
6329 mutex_unlock(&ctx->uring_lock);
6332 to_submit = io_sqring_entries(ctx);
6335 * If submit got -EBUSY, flag us as needing the application
6336 * to enter the kernel to reap and flush events.
6338 if (!to_submit || ret == -EBUSY || need_resched()) {
6340 * Drop cur_mm before scheduling, we can't hold it for
6341 * long periods (or over schedule()). Do this before
6342 * adding ourselves to the waitqueue, as the unuse/drop
6345 io_sq_thread_drop_mm();
6348 * We're polling. If we're within the defined idle
6349 * period, then let us spin without work before going
6350 * to sleep. The exception is if we got EBUSY doing
6351 * more IO, we should wait for the application to
6352 * reap events and wake us up.
6354 if (!list_empty(&ctx->poll_list) || need_resched() ||
6355 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6356 !percpu_ref_is_dying(&ctx->refs))) {
6362 prepare_to_wait(&ctx->sqo_wait, &wait,
6363 TASK_INTERRUPTIBLE);
6366 * While doing polled IO, before going to sleep, we need
6367 * to check if there are new reqs added to poll_list, it
6368 * is because reqs may have been punted to io worker and
6369 * will be added to poll_list later, hence check the
6372 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6373 !list_empty_careful(&ctx->poll_list)) {
6374 finish_wait(&ctx->sqo_wait, &wait);
6378 /* Tell userspace we may need a wakeup call */
6379 spin_lock_irq(&ctx->completion_lock);
6380 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6381 spin_unlock_irq(&ctx->completion_lock);
6383 to_submit = io_sqring_entries(ctx);
6384 if (!to_submit || ret == -EBUSY) {
6385 if (kthread_should_park()) {
6386 finish_wait(&ctx->sqo_wait, &wait);
6389 if (io_run_task_work()) {
6390 finish_wait(&ctx->sqo_wait, &wait);
6393 if (signal_pending(current))
6394 flush_signals(current);
6396 finish_wait(&ctx->sqo_wait, &wait);
6398 spin_lock_irq(&ctx->completion_lock);
6399 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6400 spin_unlock_irq(&ctx->completion_lock);
6404 finish_wait(&ctx->sqo_wait, &wait);
6406 spin_lock_irq(&ctx->completion_lock);
6407 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6408 spin_unlock_irq(&ctx->completion_lock);
6411 mutex_lock(&ctx->uring_lock);
6412 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6413 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6414 mutex_unlock(&ctx->uring_lock);
6415 timeout = jiffies + ctx->sq_thread_idle;
6420 io_sq_thread_drop_mm();
6421 revert_creds(old_cred);
6428 struct io_wait_queue {
6429 struct wait_queue_entry wq;
6430 struct io_ring_ctx *ctx;
6432 unsigned nr_timeouts;
6435 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6437 struct io_ring_ctx *ctx = iowq->ctx;
6440 * Wake up if we have enough events, or if a timeout occurred since we
6441 * started waiting. For timeouts, we always want to return to userspace,
6442 * regardless of event count.
6444 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6445 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6448 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6449 int wake_flags, void *key)
6451 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6454 /* use noflush == true, as we can't safely rely on locking context */
6455 if (!io_should_wake(iowq, true))
6458 return autoremove_wake_function(curr, mode, wake_flags, key);
6462 * Wait until events become available, if we don't already have some. The
6463 * application must reap them itself, as they reside on the shared cq ring.
6465 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6466 const sigset_t __user *sig, size_t sigsz)
6468 struct io_wait_queue iowq = {
6471 .func = io_wake_function,
6472 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6475 .to_wait = min_events,
6477 struct io_rings *rings = ctx->rings;
6481 if (io_cqring_events(ctx, false) >= min_events)
6483 if (!io_run_task_work())
6488 #ifdef CONFIG_COMPAT
6489 if (in_compat_syscall())
6490 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6494 ret = set_user_sigmask(sig, sigsz);
6500 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6501 trace_io_uring_cqring_wait(ctx, min_events);
6503 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6504 TASK_INTERRUPTIBLE);
6505 /* make sure we run task_work before checking for signals */
6506 if (io_run_task_work())
6508 if (signal_pending(current)) {
6509 if (current->jobctl & JOBCTL_TASK_WORK) {
6510 spin_lock_irq(¤t->sighand->siglock);
6511 current->jobctl &= ~JOBCTL_TASK_WORK;
6512 recalc_sigpending();
6513 spin_unlock_irq(¤t->sighand->siglock);
6519 if (io_should_wake(&iowq, false))
6523 finish_wait(&ctx->wait, &iowq.wq);
6525 restore_saved_sigmask_unless(ret == -EINTR);
6527 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6530 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6532 #if defined(CONFIG_UNIX)
6533 if (ctx->ring_sock) {
6534 struct sock *sock = ctx->ring_sock->sk;
6535 struct sk_buff *skb;
6537 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6543 for (i = 0; i < ctx->nr_user_files; i++) {
6546 file = io_file_from_index(ctx, i);
6553 static void io_file_ref_kill(struct percpu_ref *ref)
6555 struct fixed_file_data *data;
6557 data = container_of(ref, struct fixed_file_data, refs);
6558 complete(&data->done);
6561 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6563 struct fixed_file_data *data = ctx->file_data;
6564 struct fixed_file_ref_node *ref_node = NULL;
6565 unsigned nr_tables, i;
6570 spin_lock(&data->lock);
6571 if (!list_empty(&data->ref_list))
6572 ref_node = list_first_entry(&data->ref_list,
6573 struct fixed_file_ref_node, node);
6574 spin_unlock(&data->lock);
6576 percpu_ref_kill(&ref_node->refs);
6578 percpu_ref_kill(&data->refs);
6580 /* wait for all refs nodes to complete */
6581 flush_delayed_work(&ctx->file_put_work);
6582 wait_for_completion(&data->done);
6584 __io_sqe_files_unregister(ctx);
6585 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6586 for (i = 0; i < nr_tables; i++)
6587 kfree(data->table[i].files);
6589 percpu_ref_exit(&data->refs);
6591 ctx->file_data = NULL;
6592 ctx->nr_user_files = 0;
6596 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6598 if (ctx->sqo_thread) {
6599 wait_for_completion(&ctx->sq_thread_comp);
6601 * The park is a bit of a work-around, without it we get
6602 * warning spews on shutdown with SQPOLL set and affinity
6603 * set to a single CPU.
6605 kthread_park(ctx->sqo_thread);
6606 kthread_stop(ctx->sqo_thread);
6607 ctx->sqo_thread = NULL;
6611 static void io_finish_async(struct io_ring_ctx *ctx)
6613 io_sq_thread_stop(ctx);
6616 io_wq_destroy(ctx->io_wq);
6621 #if defined(CONFIG_UNIX)
6623 * Ensure the UNIX gc is aware of our file set, so we are certain that
6624 * the io_uring can be safely unregistered on process exit, even if we have
6625 * loops in the file referencing.
6627 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6629 struct sock *sk = ctx->ring_sock->sk;
6630 struct scm_fp_list *fpl;
6631 struct sk_buff *skb;
6634 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6638 skb = alloc_skb(0, GFP_KERNEL);
6647 fpl->user = get_uid(ctx->user);
6648 for (i = 0; i < nr; i++) {
6649 struct file *file = io_file_from_index(ctx, i + offset);
6653 fpl->fp[nr_files] = get_file(file);
6654 unix_inflight(fpl->user, fpl->fp[nr_files]);
6659 fpl->max = SCM_MAX_FD;
6660 fpl->count = nr_files;
6661 UNIXCB(skb).fp = fpl;
6662 skb->destructor = unix_destruct_scm;
6663 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6664 skb_queue_head(&sk->sk_receive_queue, skb);
6666 for (i = 0; i < nr_files; i++)
6677 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6678 * causes regular reference counting to break down. We rely on the UNIX
6679 * garbage collection to take care of this problem for us.
6681 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6683 unsigned left, total;
6687 left = ctx->nr_user_files;
6689 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6691 ret = __io_sqe_files_scm(ctx, this_files, total);
6695 total += this_files;
6701 while (total < ctx->nr_user_files) {
6702 struct file *file = io_file_from_index(ctx, total);
6712 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6718 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6723 for (i = 0; i < nr_tables; i++) {
6724 struct fixed_file_table *table = &ctx->file_data->table[i];
6725 unsigned this_files;
6727 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6728 table->files = kcalloc(this_files, sizeof(struct file *),
6732 nr_files -= this_files;
6738 for (i = 0; i < nr_tables; i++) {
6739 struct fixed_file_table *table = &ctx->file_data->table[i];
6740 kfree(table->files);
6745 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6747 #if defined(CONFIG_UNIX)
6748 struct sock *sock = ctx->ring_sock->sk;
6749 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6750 struct sk_buff *skb;
6753 __skb_queue_head_init(&list);
6756 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6757 * remove this entry and rearrange the file array.
6759 skb = skb_dequeue(head);
6761 struct scm_fp_list *fp;
6763 fp = UNIXCB(skb).fp;
6764 for (i = 0; i < fp->count; i++) {
6767 if (fp->fp[i] != file)
6770 unix_notinflight(fp->user, fp->fp[i]);
6771 left = fp->count - 1 - i;
6773 memmove(&fp->fp[i], &fp->fp[i + 1],
6774 left * sizeof(struct file *));
6781 __skb_queue_tail(&list, skb);
6791 __skb_queue_tail(&list, skb);
6793 skb = skb_dequeue(head);
6796 if (skb_peek(&list)) {
6797 spin_lock_irq(&head->lock);
6798 while ((skb = __skb_dequeue(&list)) != NULL)
6799 __skb_queue_tail(head, skb);
6800 spin_unlock_irq(&head->lock);
6807 struct io_file_put {
6808 struct list_head list;
6812 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6814 struct fixed_file_data *file_data = ref_node->file_data;
6815 struct io_ring_ctx *ctx = file_data->ctx;
6816 struct io_file_put *pfile, *tmp;
6818 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6819 list_del(&pfile->list);
6820 io_ring_file_put(ctx, pfile->file);
6824 spin_lock(&file_data->lock);
6825 list_del(&ref_node->node);
6826 spin_unlock(&file_data->lock);
6828 percpu_ref_exit(&ref_node->refs);
6830 percpu_ref_put(&file_data->refs);
6833 static void io_file_put_work(struct work_struct *work)
6835 struct io_ring_ctx *ctx;
6836 struct llist_node *node;
6838 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6839 node = llist_del_all(&ctx->file_put_llist);
6842 struct fixed_file_ref_node *ref_node;
6843 struct llist_node *next = node->next;
6845 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6846 __io_file_put_work(ref_node);
6851 static void io_file_data_ref_zero(struct percpu_ref *ref)
6853 struct fixed_file_ref_node *ref_node;
6854 struct io_ring_ctx *ctx;
6858 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6859 ctx = ref_node->file_data->ctx;
6861 if (percpu_ref_is_dying(&ctx->file_data->refs))
6864 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6866 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6868 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6871 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6872 struct io_ring_ctx *ctx)
6874 struct fixed_file_ref_node *ref_node;
6876 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6878 return ERR_PTR(-ENOMEM);
6880 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6883 return ERR_PTR(-ENOMEM);
6885 INIT_LIST_HEAD(&ref_node->node);
6886 INIT_LIST_HEAD(&ref_node->file_list);
6887 ref_node->file_data = ctx->file_data;
6891 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6893 percpu_ref_exit(&ref_node->refs);
6897 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6900 __s32 __user *fds = (__s32 __user *) arg;
6905 struct fixed_file_ref_node *ref_node;
6911 if (nr_args > IORING_MAX_FIXED_FILES)
6914 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6915 if (!ctx->file_data)
6917 ctx->file_data->ctx = ctx;
6918 init_completion(&ctx->file_data->done);
6919 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6920 spin_lock_init(&ctx->file_data->lock);
6922 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6923 ctx->file_data->table = kcalloc(nr_tables,
6924 sizeof(struct fixed_file_table),
6926 if (!ctx->file_data->table) {
6927 kfree(ctx->file_data);
6928 ctx->file_data = NULL;
6932 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6933 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6934 kfree(ctx->file_data->table);
6935 kfree(ctx->file_data);
6936 ctx->file_data = NULL;
6940 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6941 percpu_ref_exit(&ctx->file_data->refs);
6942 kfree(ctx->file_data->table);
6943 kfree(ctx->file_data);
6944 ctx->file_data = NULL;
6948 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6949 struct fixed_file_table *table;
6953 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6955 /* allow sparse sets */
6961 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6962 index = i & IORING_FILE_TABLE_MASK;
6970 * Don't allow io_uring instances to be registered. If UNIX
6971 * isn't enabled, then this causes a reference cycle and this
6972 * instance can never get freed. If UNIX is enabled we'll
6973 * handle it just fine, but there's still no point in allowing
6974 * a ring fd as it doesn't support regular read/write anyway.
6976 if (file->f_op == &io_uring_fops) {
6981 table->files[index] = file;
6985 for (i = 0; i < ctx->nr_user_files; i++) {
6986 file = io_file_from_index(ctx, i);
6990 for (i = 0; i < nr_tables; i++)
6991 kfree(ctx->file_data->table[i].files);
6993 percpu_ref_exit(&ctx->file_data->refs);
6994 kfree(ctx->file_data->table);
6995 kfree(ctx->file_data);
6996 ctx->file_data = NULL;
6997 ctx->nr_user_files = 0;
7001 ret = io_sqe_files_scm(ctx);
7003 io_sqe_files_unregister(ctx);
7007 ref_node = alloc_fixed_file_ref_node(ctx);
7008 if (IS_ERR(ref_node)) {
7009 io_sqe_files_unregister(ctx);
7010 return PTR_ERR(ref_node);
7013 ctx->file_data->cur_refs = &ref_node->refs;
7014 spin_lock(&ctx->file_data->lock);
7015 list_add(&ref_node->node, &ctx->file_data->ref_list);
7016 spin_unlock(&ctx->file_data->lock);
7017 percpu_ref_get(&ctx->file_data->refs);
7021 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7024 #if defined(CONFIG_UNIX)
7025 struct sock *sock = ctx->ring_sock->sk;
7026 struct sk_buff_head *head = &sock->sk_receive_queue;
7027 struct sk_buff *skb;
7030 * See if we can merge this file into an existing skb SCM_RIGHTS
7031 * file set. If there's no room, fall back to allocating a new skb
7032 * and filling it in.
7034 spin_lock_irq(&head->lock);
7035 skb = skb_peek(head);
7037 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7039 if (fpl->count < SCM_MAX_FD) {
7040 __skb_unlink(skb, head);
7041 spin_unlock_irq(&head->lock);
7042 fpl->fp[fpl->count] = get_file(file);
7043 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7045 spin_lock_irq(&head->lock);
7046 __skb_queue_head(head, skb);
7051 spin_unlock_irq(&head->lock);
7058 return __io_sqe_files_scm(ctx, 1, index);
7064 static int io_queue_file_removal(struct fixed_file_data *data,
7067 struct io_file_put *pfile;
7068 struct percpu_ref *refs = data->cur_refs;
7069 struct fixed_file_ref_node *ref_node;
7071 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7075 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7077 list_add(&pfile->list, &ref_node->file_list);
7082 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7083 struct io_uring_files_update *up,
7086 struct fixed_file_data *data = ctx->file_data;
7087 struct fixed_file_ref_node *ref_node;
7092 bool needs_switch = false;
7094 if (check_add_overflow(up->offset, nr_args, &done))
7096 if (done > ctx->nr_user_files)
7099 ref_node = alloc_fixed_file_ref_node(ctx);
7100 if (IS_ERR(ref_node))
7101 return PTR_ERR(ref_node);
7104 fds = u64_to_user_ptr(up->fds);
7106 struct fixed_file_table *table;
7110 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7114 i = array_index_nospec(up->offset, ctx->nr_user_files);
7115 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7116 index = i & IORING_FILE_TABLE_MASK;
7117 if (table->files[index]) {
7118 file = io_file_from_index(ctx, index);
7119 err = io_queue_file_removal(data, file);
7122 table->files[index] = NULL;
7123 needs_switch = true;
7132 * Don't allow io_uring instances to be registered. If
7133 * UNIX isn't enabled, then this causes a reference
7134 * cycle and this instance can never get freed. If UNIX
7135 * is enabled we'll handle it just fine, but there's
7136 * still no point in allowing a ring fd as it doesn't
7137 * support regular read/write anyway.
7139 if (file->f_op == &io_uring_fops) {
7144 table->files[index] = file;
7145 err = io_sqe_file_register(ctx, file, i);
7157 percpu_ref_kill(data->cur_refs);
7158 spin_lock(&data->lock);
7159 list_add(&ref_node->node, &data->ref_list);
7160 data->cur_refs = &ref_node->refs;
7161 spin_unlock(&data->lock);
7162 percpu_ref_get(&ctx->file_data->refs);
7164 destroy_fixed_file_ref_node(ref_node);
7166 return done ? done : err;
7169 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7172 struct io_uring_files_update up;
7174 if (!ctx->file_data)
7178 if (copy_from_user(&up, arg, sizeof(up)))
7183 return __io_sqe_files_update(ctx, &up, nr_args);
7186 static void io_free_work(struct io_wq_work *work)
7188 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7190 /* Consider that io_steal_work() relies on this ref */
7194 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7195 struct io_uring_params *p)
7197 struct io_wq_data data;
7199 struct io_ring_ctx *ctx_attach;
7200 unsigned int concurrency;
7203 data.user = ctx->user;
7204 data.free_work = io_free_work;
7205 data.do_work = io_wq_submit_work;
7207 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7208 /* Do QD, or 4 * CPUS, whatever is smallest */
7209 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7211 ctx->io_wq = io_wq_create(concurrency, &data);
7212 if (IS_ERR(ctx->io_wq)) {
7213 ret = PTR_ERR(ctx->io_wq);
7219 f = fdget(p->wq_fd);
7223 if (f.file->f_op != &io_uring_fops) {
7228 ctx_attach = f.file->private_data;
7229 /* @io_wq is protected by holding the fd */
7230 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7235 ctx->io_wq = ctx_attach->io_wq;
7241 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7242 struct io_uring_params *p)
7246 if (ctx->flags & IORING_SETUP_SQPOLL) {
7247 mmgrab(current->mm);
7248 ctx->sqo_mm = current->mm;
7251 if (!capable(CAP_SYS_ADMIN))
7254 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7255 if (!ctx->sq_thread_idle)
7256 ctx->sq_thread_idle = HZ;
7258 if (p->flags & IORING_SETUP_SQ_AFF) {
7259 int cpu = p->sq_thread_cpu;
7262 if (cpu >= nr_cpu_ids)
7264 if (!cpu_online(cpu))
7267 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7271 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7274 if (IS_ERR(ctx->sqo_thread)) {
7275 ret = PTR_ERR(ctx->sqo_thread);
7276 ctx->sqo_thread = NULL;
7279 wake_up_process(ctx->sqo_thread);
7280 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7281 /* Can't have SQ_AFF without SQPOLL */
7286 ret = io_init_wq_offload(ctx, p);
7292 io_finish_async(ctx);
7294 mmdrop(ctx->sqo_mm);
7300 static inline void __io_unaccount_mem(struct user_struct *user,
7301 unsigned long nr_pages)
7303 atomic_long_sub(nr_pages, &user->locked_vm);
7306 static inline int __io_account_mem(struct user_struct *user,
7307 unsigned long nr_pages)
7309 unsigned long page_limit, cur_pages, new_pages;
7311 /* Don't allow more pages than we can safely lock */
7312 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7315 cur_pages = atomic_long_read(&user->locked_vm);
7316 new_pages = cur_pages + nr_pages;
7317 if (new_pages > page_limit)
7319 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7320 new_pages) != cur_pages);
7325 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7326 enum io_mem_account acct)
7329 __io_unaccount_mem(ctx->user, nr_pages);
7332 if (acct == ACCT_LOCKED)
7333 ctx->sqo_mm->locked_vm -= nr_pages;
7334 else if (acct == ACCT_PINNED)
7335 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7339 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7340 enum io_mem_account acct)
7344 if (ctx->limit_mem) {
7345 ret = __io_account_mem(ctx->user, nr_pages);
7351 if (acct == ACCT_LOCKED)
7352 ctx->sqo_mm->locked_vm += nr_pages;
7353 else if (acct == ACCT_PINNED)
7354 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7360 static void io_mem_free(void *ptr)
7367 page = virt_to_head_page(ptr);
7368 if (put_page_testzero(page))
7369 free_compound_page(page);
7372 static void *io_mem_alloc(size_t size)
7374 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7377 return (void *) __get_free_pages(gfp_flags, get_order(size));
7380 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7383 struct io_rings *rings;
7384 size_t off, sq_array_size;
7386 off = struct_size(rings, cqes, cq_entries);
7387 if (off == SIZE_MAX)
7391 off = ALIGN(off, SMP_CACHE_BYTES);
7399 sq_array_size = array_size(sizeof(u32), sq_entries);
7400 if (sq_array_size == SIZE_MAX)
7403 if (check_add_overflow(off, sq_array_size, &off))
7409 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7413 pages = (size_t)1 << get_order(
7414 rings_size(sq_entries, cq_entries, NULL));
7415 pages += (size_t)1 << get_order(
7416 array_size(sizeof(struct io_uring_sqe), sq_entries));
7421 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7425 if (!ctx->user_bufs)
7428 for (i = 0; i < ctx->nr_user_bufs; i++) {
7429 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7431 for (j = 0; j < imu->nr_bvecs; j++)
7432 unpin_user_page(imu->bvec[j].bv_page);
7434 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7439 kfree(ctx->user_bufs);
7440 ctx->user_bufs = NULL;
7441 ctx->nr_user_bufs = 0;
7445 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7446 void __user *arg, unsigned index)
7448 struct iovec __user *src;
7450 #ifdef CONFIG_COMPAT
7452 struct compat_iovec __user *ciovs;
7453 struct compat_iovec ciov;
7455 ciovs = (struct compat_iovec __user *) arg;
7456 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7459 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7460 dst->iov_len = ciov.iov_len;
7464 src = (struct iovec __user *) arg;
7465 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7470 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7473 struct vm_area_struct **vmas = NULL;
7474 struct page **pages = NULL;
7475 int i, j, got_pages = 0;
7480 if (!nr_args || nr_args > UIO_MAXIOV)
7483 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7485 if (!ctx->user_bufs)
7488 for (i = 0; i < nr_args; i++) {
7489 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7490 unsigned long off, start, end, ubuf;
7495 ret = io_copy_iov(ctx, &iov, arg, i);
7500 * Don't impose further limits on the size and buffer
7501 * constraints here, we'll -EINVAL later when IO is
7502 * submitted if they are wrong.
7505 if (!iov.iov_base || !iov.iov_len)
7508 /* arbitrary limit, but we need something */
7509 if (iov.iov_len > SZ_1G)
7512 ubuf = (unsigned long) iov.iov_base;
7513 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7514 start = ubuf >> PAGE_SHIFT;
7515 nr_pages = end - start;
7517 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7522 if (!pages || nr_pages > got_pages) {
7525 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7527 vmas = kvmalloc_array(nr_pages,
7528 sizeof(struct vm_area_struct *),
7530 if (!pages || !vmas) {
7532 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7535 got_pages = nr_pages;
7538 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7542 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7547 mmap_read_lock(current->mm);
7548 pret = pin_user_pages(ubuf, nr_pages,
7549 FOLL_WRITE | FOLL_LONGTERM,
7551 if (pret == nr_pages) {
7552 /* don't support file backed memory */
7553 for (j = 0; j < nr_pages; j++) {
7554 struct vm_area_struct *vma = vmas[j];
7557 !is_file_hugepages(vma->vm_file)) {
7563 ret = pret < 0 ? pret : -EFAULT;
7565 mmap_read_unlock(current->mm);
7568 * if we did partial map, or found file backed vmas,
7569 * release any pages we did get
7572 unpin_user_pages(pages, pret);
7573 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7578 off = ubuf & ~PAGE_MASK;
7580 for (j = 0; j < nr_pages; j++) {
7583 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7584 imu->bvec[j].bv_page = pages[j];
7585 imu->bvec[j].bv_len = vec_len;
7586 imu->bvec[j].bv_offset = off;
7590 /* store original address for later verification */
7592 imu->len = iov.iov_len;
7593 imu->nr_bvecs = nr_pages;
7595 ctx->nr_user_bufs++;
7603 io_sqe_buffer_unregister(ctx);
7607 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7609 __s32 __user *fds = arg;
7615 if (copy_from_user(&fd, fds, sizeof(*fds)))
7618 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7619 if (IS_ERR(ctx->cq_ev_fd)) {
7620 int ret = PTR_ERR(ctx->cq_ev_fd);
7621 ctx->cq_ev_fd = NULL;
7628 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7630 if (ctx->cq_ev_fd) {
7631 eventfd_ctx_put(ctx->cq_ev_fd);
7632 ctx->cq_ev_fd = NULL;
7639 static int __io_destroy_buffers(int id, void *p, void *data)
7641 struct io_ring_ctx *ctx = data;
7642 struct io_buffer *buf = p;
7644 __io_remove_buffers(ctx, buf, id, -1U);
7648 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7650 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7651 idr_destroy(&ctx->io_buffer_idr);
7654 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7656 io_finish_async(ctx);
7658 mmdrop(ctx->sqo_mm);
7662 io_sqe_buffer_unregister(ctx);
7663 io_sqe_files_unregister(ctx);
7664 io_eventfd_unregister(ctx);
7665 io_destroy_buffers(ctx);
7666 idr_destroy(&ctx->personality_idr);
7668 #if defined(CONFIG_UNIX)
7669 if (ctx->ring_sock) {
7670 ctx->ring_sock->file = NULL; /* so that iput() is called */
7671 sock_release(ctx->ring_sock);
7675 io_mem_free(ctx->rings);
7676 io_mem_free(ctx->sq_sqes);
7678 percpu_ref_exit(&ctx->refs);
7679 free_uid(ctx->user);
7680 put_cred(ctx->creds);
7681 kfree(ctx->cancel_hash);
7682 kmem_cache_free(req_cachep, ctx->fallback_req);
7686 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7688 struct io_ring_ctx *ctx = file->private_data;
7691 poll_wait(file, &ctx->cq_wait, wait);
7693 * synchronizes with barrier from wq_has_sleeper call in
7697 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7698 ctx->rings->sq_ring_entries)
7699 mask |= EPOLLOUT | EPOLLWRNORM;
7700 if (io_cqring_events(ctx, false))
7701 mask |= EPOLLIN | EPOLLRDNORM;
7706 static int io_uring_fasync(int fd, struct file *file, int on)
7708 struct io_ring_ctx *ctx = file->private_data;
7710 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7713 static int io_remove_personalities(int id, void *p, void *data)
7715 struct io_ring_ctx *ctx = data;
7716 const struct cred *cred;
7718 cred = idr_remove(&ctx->personality_idr, id);
7724 static void io_ring_exit_work(struct work_struct *work)
7726 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7730 * If we're doing polled IO and end up having requests being
7731 * submitted async (out-of-line), then completions can come in while
7732 * we're waiting for refs to drop. We need to reap these manually,
7733 * as nobody else will be looking for them.
7737 io_cqring_overflow_flush(ctx, true);
7738 io_iopoll_try_reap_events(ctx);
7739 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7740 io_ring_ctx_free(ctx);
7743 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7745 mutex_lock(&ctx->uring_lock);
7746 percpu_ref_kill(&ctx->refs);
7747 mutex_unlock(&ctx->uring_lock);
7749 io_kill_timeouts(ctx);
7750 io_poll_remove_all(ctx);
7753 io_wq_cancel_all(ctx->io_wq);
7755 /* if we failed setting up the ctx, we might not have any rings */
7757 io_cqring_overflow_flush(ctx, true);
7758 io_iopoll_try_reap_events(ctx);
7759 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7762 * Do this upfront, so we won't have a grace period where the ring
7763 * is closed but resources aren't reaped yet. This can cause
7764 * spurious failure in setting up a new ring.
7766 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7769 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7770 queue_work(system_wq, &ctx->exit_work);
7773 static int io_uring_release(struct inode *inode, struct file *file)
7775 struct io_ring_ctx *ctx = file->private_data;
7777 file->private_data = NULL;
7778 io_ring_ctx_wait_and_kill(ctx);
7782 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7784 struct files_struct *files = data;
7786 return work->files == files;
7789 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7790 struct files_struct *files)
7792 if (list_empty_careful(&ctx->inflight_list))
7795 /* cancel all at once, should be faster than doing it one by one*/
7796 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7798 while (!list_empty_careful(&ctx->inflight_list)) {
7799 struct io_kiocb *cancel_req = NULL, *req;
7802 spin_lock_irq(&ctx->inflight_lock);
7803 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7804 if (req->work.files != files)
7806 /* req is being completed, ignore */
7807 if (!refcount_inc_not_zero(&req->refs))
7813 prepare_to_wait(&ctx->inflight_wait, &wait,
7814 TASK_UNINTERRUPTIBLE);
7815 spin_unlock_irq(&ctx->inflight_lock);
7817 /* We need to keep going until we don't find a matching req */
7821 if (cancel_req->flags & REQ_F_OVERFLOW) {
7822 spin_lock_irq(&ctx->completion_lock);
7823 list_del(&cancel_req->list);
7824 cancel_req->flags &= ~REQ_F_OVERFLOW;
7825 if (list_empty(&ctx->cq_overflow_list)) {
7826 clear_bit(0, &ctx->sq_check_overflow);
7827 clear_bit(0, &ctx->cq_check_overflow);
7828 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7830 spin_unlock_irq(&ctx->completion_lock);
7832 WRITE_ONCE(ctx->rings->cq_overflow,
7833 atomic_inc_return(&ctx->cached_cq_overflow));
7836 * Put inflight ref and overflow ref. If that's
7837 * all we had, then we're done with this request.
7839 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7840 io_free_req(cancel_req);
7841 finish_wait(&ctx->inflight_wait, &wait);
7845 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7846 io_put_req(cancel_req);
7850 finish_wait(&ctx->inflight_wait, &wait);
7854 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7856 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7857 struct task_struct *task = data;
7859 return req->task == task;
7862 static int io_uring_flush(struct file *file, void *data)
7864 struct io_ring_ctx *ctx = file->private_data;
7866 io_uring_cancel_files(ctx, data);
7869 * If the task is going away, cancel work it may have pending
7871 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7872 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7877 static void *io_uring_validate_mmap_request(struct file *file,
7878 loff_t pgoff, size_t sz)
7880 struct io_ring_ctx *ctx = file->private_data;
7881 loff_t offset = pgoff << PAGE_SHIFT;
7886 case IORING_OFF_SQ_RING:
7887 case IORING_OFF_CQ_RING:
7890 case IORING_OFF_SQES:
7894 return ERR_PTR(-EINVAL);
7897 page = virt_to_head_page(ptr);
7898 if (sz > page_size(page))
7899 return ERR_PTR(-EINVAL);
7906 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7908 size_t sz = vma->vm_end - vma->vm_start;
7912 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7914 return PTR_ERR(ptr);
7916 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7917 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7920 #else /* !CONFIG_MMU */
7922 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7924 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7927 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7929 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7932 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7933 unsigned long addr, unsigned long len,
7934 unsigned long pgoff, unsigned long flags)
7938 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7940 return PTR_ERR(ptr);
7942 return (unsigned long) ptr;
7945 #endif /* !CONFIG_MMU */
7947 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7948 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7951 struct io_ring_ctx *ctx;
7958 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7966 if (f.file->f_op != &io_uring_fops)
7970 ctx = f.file->private_data;
7971 if (!percpu_ref_tryget(&ctx->refs))
7975 * For SQ polling, the thread will do all submissions and completions.
7976 * Just return the requested submit count, and wake the thread if
7980 if (ctx->flags & IORING_SETUP_SQPOLL) {
7981 if (!list_empty_careful(&ctx->cq_overflow_list))
7982 io_cqring_overflow_flush(ctx, false);
7983 if (flags & IORING_ENTER_SQ_WAKEUP)
7984 wake_up(&ctx->sqo_wait);
7985 submitted = to_submit;
7986 } else if (to_submit) {
7987 mutex_lock(&ctx->uring_lock);
7988 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7989 mutex_unlock(&ctx->uring_lock);
7991 if (submitted != to_submit)
7994 if (flags & IORING_ENTER_GETEVENTS) {
7995 min_complete = min(min_complete, ctx->cq_entries);
7998 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7999 * space applications don't need to do io completion events
8000 * polling again, they can rely on io_sq_thread to do polling
8001 * work, which can reduce cpu usage and uring_lock contention.
8003 if (ctx->flags & IORING_SETUP_IOPOLL &&
8004 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8005 ret = io_iopoll_check(ctx, min_complete);
8007 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8012 percpu_ref_put(&ctx->refs);
8015 return submitted ? submitted : ret;
8018 #ifdef CONFIG_PROC_FS
8019 static int io_uring_show_cred(int id, void *p, void *data)
8021 const struct cred *cred = p;
8022 struct seq_file *m = data;
8023 struct user_namespace *uns = seq_user_ns(m);
8024 struct group_info *gi;
8029 seq_printf(m, "%5d\n", id);
8030 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8031 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8032 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8033 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8034 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8035 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8036 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8037 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8038 seq_puts(m, "\n\tGroups:\t");
8039 gi = cred->group_info;
8040 for (g = 0; g < gi->ngroups; g++) {
8041 seq_put_decimal_ull(m, g ? " " : "",
8042 from_kgid_munged(uns, gi->gid[g]));
8044 seq_puts(m, "\n\tCapEff:\t");
8045 cap = cred->cap_effective;
8046 CAP_FOR_EACH_U32(__capi)
8047 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8052 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8056 mutex_lock(&ctx->uring_lock);
8057 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8058 for (i = 0; i < ctx->nr_user_files; i++) {
8059 struct fixed_file_table *table;
8062 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8063 f = table->files[i & IORING_FILE_TABLE_MASK];
8065 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8067 seq_printf(m, "%5u: <none>\n", i);
8069 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8070 for (i = 0; i < ctx->nr_user_bufs; i++) {
8071 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8073 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8074 (unsigned int) buf->len);
8076 if (!idr_is_empty(&ctx->personality_idr)) {
8077 seq_printf(m, "Personalities:\n");
8078 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8080 seq_printf(m, "PollList:\n");
8081 spin_lock_irq(&ctx->completion_lock);
8082 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8083 struct hlist_head *list = &ctx->cancel_hash[i];
8084 struct io_kiocb *req;
8086 hlist_for_each_entry(req, list, hash_node)
8087 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8088 req->task->task_works != NULL);
8090 spin_unlock_irq(&ctx->completion_lock);
8091 mutex_unlock(&ctx->uring_lock);
8094 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8096 struct io_ring_ctx *ctx = f->private_data;
8098 if (percpu_ref_tryget(&ctx->refs)) {
8099 __io_uring_show_fdinfo(ctx, m);
8100 percpu_ref_put(&ctx->refs);
8105 static const struct file_operations io_uring_fops = {
8106 .release = io_uring_release,
8107 .flush = io_uring_flush,
8108 .mmap = io_uring_mmap,
8110 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8111 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8113 .poll = io_uring_poll,
8114 .fasync = io_uring_fasync,
8115 #ifdef CONFIG_PROC_FS
8116 .show_fdinfo = io_uring_show_fdinfo,
8120 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8121 struct io_uring_params *p)
8123 struct io_rings *rings;
8124 size_t size, sq_array_offset;
8126 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8127 if (size == SIZE_MAX)
8130 rings = io_mem_alloc(size);
8135 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8136 rings->sq_ring_mask = p->sq_entries - 1;
8137 rings->cq_ring_mask = p->cq_entries - 1;
8138 rings->sq_ring_entries = p->sq_entries;
8139 rings->cq_ring_entries = p->cq_entries;
8140 ctx->sq_mask = rings->sq_ring_mask;
8141 ctx->cq_mask = rings->cq_ring_mask;
8142 ctx->sq_entries = rings->sq_ring_entries;
8143 ctx->cq_entries = rings->cq_ring_entries;
8145 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8146 if (size == SIZE_MAX) {
8147 io_mem_free(ctx->rings);
8152 ctx->sq_sqes = io_mem_alloc(size);
8153 if (!ctx->sq_sqes) {
8154 io_mem_free(ctx->rings);
8163 * Allocate an anonymous fd, this is what constitutes the application
8164 * visible backing of an io_uring instance. The application mmaps this
8165 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8166 * we have to tie this fd to a socket for file garbage collection purposes.
8168 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8173 #if defined(CONFIG_UNIX)
8174 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8180 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8184 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8185 O_RDWR | O_CLOEXEC);
8188 ret = PTR_ERR(file);
8192 #if defined(CONFIG_UNIX)
8193 ctx->ring_sock->file = file;
8195 fd_install(ret, file);
8198 #if defined(CONFIG_UNIX)
8199 sock_release(ctx->ring_sock);
8200 ctx->ring_sock = NULL;
8205 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8206 struct io_uring_params __user *params)
8208 struct user_struct *user = NULL;
8209 struct io_ring_ctx *ctx;
8215 if (entries > IORING_MAX_ENTRIES) {
8216 if (!(p->flags & IORING_SETUP_CLAMP))
8218 entries = IORING_MAX_ENTRIES;
8222 * Use twice as many entries for the CQ ring. It's possible for the
8223 * application to drive a higher depth than the size of the SQ ring,
8224 * since the sqes are only used at submission time. This allows for
8225 * some flexibility in overcommitting a bit. If the application has
8226 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8227 * of CQ ring entries manually.
8229 p->sq_entries = roundup_pow_of_two(entries);
8230 if (p->flags & IORING_SETUP_CQSIZE) {
8232 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8233 * to a power-of-two, if it isn't already. We do NOT impose
8234 * any cq vs sq ring sizing.
8236 if (p->cq_entries < p->sq_entries)
8238 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8239 if (!(p->flags & IORING_SETUP_CLAMP))
8241 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8243 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8245 p->cq_entries = 2 * p->sq_entries;
8248 user = get_uid(current_user());
8249 limit_mem = !capable(CAP_IPC_LOCK);
8252 ret = __io_account_mem(user,
8253 ring_pages(p->sq_entries, p->cq_entries));
8260 ctx = io_ring_ctx_alloc(p);
8263 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8268 ctx->compat = in_compat_syscall();
8270 ctx->creds = get_current_cred();
8272 ret = io_allocate_scq_urings(ctx, p);
8276 ret = io_sq_offload_start(ctx, p);
8280 memset(&p->sq_off, 0, sizeof(p->sq_off));
8281 p->sq_off.head = offsetof(struct io_rings, sq.head);
8282 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8283 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8284 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8285 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8286 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8287 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8289 memset(&p->cq_off, 0, sizeof(p->cq_off));
8290 p->cq_off.head = offsetof(struct io_rings, cq.head);
8291 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8292 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8293 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8294 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8295 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8296 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8298 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8299 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8300 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8301 IORING_FEAT_POLL_32BITS;
8303 if (copy_to_user(params, p, sizeof(*p))) {
8308 * Install ring fd as the very last thing, so we don't risk someone
8309 * having closed it before we finish setup
8311 ret = io_uring_get_fd(ctx);
8315 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8316 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8318 ctx->limit_mem = limit_mem;
8321 io_ring_ctx_wait_and_kill(ctx);
8326 * Sets up an aio uring context, and returns the fd. Applications asks for a
8327 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8328 * params structure passed in.
8330 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8332 struct io_uring_params p;
8335 if (copy_from_user(&p, params, sizeof(p)))
8337 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8342 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8343 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8344 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8347 return io_uring_create(entries, &p, params);
8350 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8351 struct io_uring_params __user *, params)
8353 return io_uring_setup(entries, params);
8356 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8358 struct io_uring_probe *p;
8362 size = struct_size(p, ops, nr_args);
8363 if (size == SIZE_MAX)
8365 p = kzalloc(size, GFP_KERNEL);
8370 if (copy_from_user(p, arg, size))
8373 if (memchr_inv(p, 0, size))
8376 p->last_op = IORING_OP_LAST - 1;
8377 if (nr_args > IORING_OP_LAST)
8378 nr_args = IORING_OP_LAST;
8380 for (i = 0; i < nr_args; i++) {
8382 if (!io_op_defs[i].not_supported)
8383 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8388 if (copy_to_user(arg, p, size))
8395 static int io_register_personality(struct io_ring_ctx *ctx)
8397 const struct cred *creds = get_current_cred();
8400 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8401 USHRT_MAX, GFP_KERNEL);
8407 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8409 const struct cred *old_creds;
8411 old_creds = idr_remove(&ctx->personality_idr, id);
8413 put_cred(old_creds);
8420 static bool io_register_op_must_quiesce(int op)
8423 case IORING_UNREGISTER_FILES:
8424 case IORING_REGISTER_FILES_UPDATE:
8425 case IORING_REGISTER_PROBE:
8426 case IORING_REGISTER_PERSONALITY:
8427 case IORING_UNREGISTER_PERSONALITY:
8434 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8435 void __user *arg, unsigned nr_args)
8436 __releases(ctx->uring_lock)
8437 __acquires(ctx->uring_lock)
8442 * We're inside the ring mutex, if the ref is already dying, then
8443 * someone else killed the ctx or is already going through
8444 * io_uring_register().
8446 if (percpu_ref_is_dying(&ctx->refs))
8449 if (io_register_op_must_quiesce(opcode)) {
8450 percpu_ref_kill(&ctx->refs);
8453 * Drop uring mutex before waiting for references to exit. If
8454 * another thread is currently inside io_uring_enter() it might
8455 * need to grab the uring_lock to make progress. If we hold it
8456 * here across the drain wait, then we can deadlock. It's safe
8457 * to drop the mutex here, since no new references will come in
8458 * after we've killed the percpu ref.
8460 mutex_unlock(&ctx->uring_lock);
8461 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8462 mutex_lock(&ctx->uring_lock);
8464 percpu_ref_resurrect(&ctx->refs);
8471 case IORING_REGISTER_BUFFERS:
8472 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8474 case IORING_UNREGISTER_BUFFERS:
8478 ret = io_sqe_buffer_unregister(ctx);
8480 case IORING_REGISTER_FILES:
8481 ret = io_sqe_files_register(ctx, arg, nr_args);
8483 case IORING_UNREGISTER_FILES:
8487 ret = io_sqe_files_unregister(ctx);
8489 case IORING_REGISTER_FILES_UPDATE:
8490 ret = io_sqe_files_update(ctx, arg, nr_args);
8492 case IORING_REGISTER_EVENTFD:
8493 case IORING_REGISTER_EVENTFD_ASYNC:
8497 ret = io_eventfd_register(ctx, arg);
8500 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8501 ctx->eventfd_async = 1;
8503 ctx->eventfd_async = 0;
8505 case IORING_UNREGISTER_EVENTFD:
8509 ret = io_eventfd_unregister(ctx);
8511 case IORING_REGISTER_PROBE:
8513 if (!arg || nr_args > 256)
8515 ret = io_probe(ctx, arg, nr_args);
8517 case IORING_REGISTER_PERSONALITY:
8521 ret = io_register_personality(ctx);
8523 case IORING_UNREGISTER_PERSONALITY:
8527 ret = io_unregister_personality(ctx, nr_args);
8534 if (io_register_op_must_quiesce(opcode)) {
8535 /* bring the ctx back to life */
8536 percpu_ref_reinit(&ctx->refs);
8538 reinit_completion(&ctx->ref_comp);
8543 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8544 void __user *, arg, unsigned int, nr_args)
8546 struct io_ring_ctx *ctx;
8555 if (f.file->f_op != &io_uring_fops)
8558 ctx = f.file->private_data;
8560 mutex_lock(&ctx->uring_lock);
8561 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8562 mutex_unlock(&ctx->uring_lock);
8563 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8564 ctx->cq_ev_fd != NULL, ret);
8570 static int __init io_uring_init(void)
8572 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8573 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8574 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8577 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8578 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8579 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8580 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8581 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8582 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8583 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8584 BUILD_BUG_SQE_ELEM(8, __u64, off);
8585 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8586 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8587 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8588 BUILD_BUG_SQE_ELEM(24, __u32, len);
8589 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8590 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8591 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8592 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8593 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8594 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8595 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8596 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8597 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8598 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8599 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8600 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8601 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8602 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8603 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8604 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8605 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8606 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8607 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8609 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8610 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8611 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8614 __initcall(io_uring_init);